Datasets:

kilian-group
/

arxiv-classifier

Dataset card Data Studio Files Files and versions Community

Subset (4)

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

Split (2)

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

paper_id stringlengths 9 16	version stringclasses 26 values	yymm stringclasses 311 values	created timestamp[s]	title stringlengths 6 335	secondary_subfield sequencelengths 1 8	abstract stringlengths 25 3.93k	primary_subfield stringclasses 124 values	field stringclasses 20 values	fulltext stringlengths 0 2.84M
0906.1858	3	0906	2015-07-18T00:17:01	Tensor products of function systems revisited	[ "math.OA", "math.FA" ]	Based on the Archimedeanization developed by Paulsen and Tomforde, we give an explicit description for the positive cones of maximal tensor products of function systems. From this description, we obtain an approximation theorem for nuclear maps between function systems. As an application, we give elementary proofs on several characterizations of nuclear function systems that are already known.	math.OA	math	TENSOR PRODUCTS OF FUNCTION SYSTEMS REVISITED KYUNG HOON HAN Abstract. Based on the Archimedeanization developed by Paulsen and Tomforde, we give an explicit description for the positive cones of maximal tensor products of function systems. From this description, we obtain an approximation theorem for nuclear maps between function systems. As an application, we give elementary proofs on several characterizations of nuclear function systems that are already known. 1. Introduction A real vector space V is called an ordered real vector space if there exists a cone V ` Ă V such that V ` X ´V ` " t0u. The cone V ` induces a partial order by v ě w if and only if v ´ w P V `. For an ordered real vector space pV, V `q, an element e in V is called an order unit if for each v in V , there exists a real number r ą 0 such that re ě v. We call an order unit e an Archimedean order unit if εe ` v P V ` for any ε ą 0 implies that v P V `. The order norm of an ordered real vector space with an Archimedean order unit is defined by }v} " inftr ą 0 : ´re ď v ď reu. An ordered real vector space equipped with an Archimedean order unit is called an Archimedean (partially) ordered (vector) space [Ka, PT] or a function system [Ef, NP]. Kadison proved that every Archimedean ordered space can be embedded into a real continuous function algebra on a compact Hausdorff space through a unital order iso- morphism that is also isometric with respect to the order norm [Ka]. Kadison's rep- resentation theorem for an Archimedean ordered space justifies the alternative term function system. The tensor products of function systems have been studied in [NP] and [Ef]. Two extremal tensor products of function systems, the minimal tensor product and the maximal tensor product were considered. A function system V is called nuclear if V bmin W " V bmax W for any function system W . Along the line of the duality between function systems and compact convex sets, Namioka and Phelps proved that the nuclearity corresponds to the Choquet simplex [NP, Theorem 2.2]. Recently, Paulsen and Tomforde introduced the Archimedeanization [PT] and this has been applied to operator system theory in a series of papers [PTT, KPTT1, KPTT2]. In this paper, we focus on its application to function systems. Section 3 gives an explicit description for the positive cones of maximal tensor products of function systems applying the Archimedeanization. Based on this, we prove that the maximal tensor product is projective as a bifunctor on the category consisting of function systems 2000 Mathematics Subject Classification. 46B40, 46B28. Key words and phrases. function system, Archimedean ordered space, Archimedean ordered -vector space, tensor product, nuclear. This work was supported by the National Research Foundation of Korea Grant funded by the Korean Government (2012R1A1A1012190). 1 2 KYUNG HOON HAN and unital positive maps. Section 4 applies the ideas of [HP] to function systems and obtains an approximation theorem for nuclear maps between function systems. As an application, we give elementary proofs for several characterizations of nuclear function systems that are already known [Ef, Theorem 7.4], [NP, Theorem 1.4]. The referee kindly pointed out to the author that there has been another line of research on tensor products of Archimedean partially ordered vector spaces with order unit and Archimedeanization procedures, mainly influenced by Fremlin and extended by others, for which we refer to [El, F, GL, BR, GK]. 2. Preliminaries Two extremal tensor products of function systems have been studied in [NP] and [Ef, Section 7]. Suppose that V and W are function systems. Given faithful representations ϕ : V Ñ CpXq and ψ : W Ñ CpY q, their tensor product ϕ b ψ : V b W Ñ CpX Y q is also faithful [Ef, Section 7]. The minimal tensor product V bmin W is defined as the function system structure on V b W induced by ϕ b ψ. The minimal tensor product is independent of the choice of the faithful representations ϕ and ψ [Ef, Lemma 7.1]. The minimal tensor products of function systems have an alternative description. The positive cone of the minimal tensor product V bmin W is given by pV bmin W q` " tz P V b W : pf b gqpzq ě 0, f P SpV q, g P SpW qu, where SpV q denotes the state space on V [NP]. Let pV bW qd denote the algebraic dual of V bW . The maximal state space is defined as SmaxpV b W q :" tr P pV b W qd : rV `bW ` ě 0u which is a weakd compact convex subset of pV b W qd. We can regard elements of V b W as continuous affine functions on SmaxpV b W q. This is a faithful realization. The maximal tensor product V bmax W is defined as the function system structure on V b W induced by the inclusion V b W Ă CpSmaxpV b W qq [Ef, Section 7]. If ϕ : V1 Ñ V2 and ψ : W1 Ñ W2 are unital positive maps, then their tensor products ϕ b ψ : V1 bmin W1 Ñ V2 bmin W2 and ϕ b ψ : V1 bmax W1 Ñ V2 bmax W2 are unital positive. In particular, ϕ b ψ : V1 bmin W1 Ñ V2 bmin W2 is an order embedding if ϕ and ψ are order embeddings. This is not true for the maximal tensor product. There is a natural duality between compact convex sets and function systems. For a compact convex set K, the space ApKq of real continuous affine functions on K is a function system. For a function system V , the state space SpV q equipped with a weak-topology is a compact convex set. By the barycenter formula [A, Proposition 1.2.2], the state space SpApKqq consists entirely of evaluations at points in K. In fact, the state space SpApKqq is affinely homeomorphic to K. Conversely, the continuous affine functions ApSpV qq consists entirely of evaluations at elements in V . The affine function system ApSpV qq is unitally order isomorphic to V . Let P pKq denote the cone of real continuous convex functions on a compact convex set K. For µ, ν P MRpKq, the Choquet order is defined as µ ă ν ô µpf q ď νpf q, @f P P pKq. Roughly speaking, the Choquet order measures how far the mass of a measure is dis- tributed to the outside. A complex measure µ on a compact convex set K is said to TENSOR PRODUCTS OF FUNCTION SYSTEMS REVISITED 3 be a boundary measure if µ is a maximal element of M `pKq with respect to the Cho- quet order. Every point x in a compact convex set K can be represented by a positive normalized boundary measure µ; that is, apxq " żK apyqdµpyq, @a P ApKq. The boundary measure associated with each point in K is unique if and only if the dual space ApKq is lattice ordered. In this case, we call K a Choquet simplex. A function system V is called nuclear if V bmin W " V bmax W for any function system W . Namioka and Phelps proved that a compact convex set K is a Choquet simplex if and only if the continuous affine function system ApKq on K is nuclear. Dually, a function system V is nuclear if and only if its state space SpV q is a Choquet simplex. Paulsen and Tomforde introduced a functorial process, called Archimedeanizaton, for forming an Archimedean ordered space from an ordered real vector space with an order unit. Given an ordered real vector space pV, V `q with an order unit e, we let D :" tv P V : εe ` v P V ` for all ε ą 0u and N :" D X ´D. The Archimedeanization VArch of V is defined as an ordered real vector space pV { N, D ` Nq with an order unit e ` N. Then, VArch is an Archimedean ordered space. The Archimedeanization is characterized by the universal property that it satisfies: for an Archimedean ordered space W and a unital positive map ϕ : V Ñ W , there exists a unique unital positive linear map ϕ : VArch Ñ W with ϕ " ϕ q. V ❆❆❆❆❆❆❆❆ ϕ q W VArch ②②②②②②②② ϕ We say that a subspace J of an Archimedean ordered space V is an order ideal of V if p P J and 0 ď q ď p imply that q P J. The Archimedean quotient of V by J is defined as the Archimedeanization of pV {J, V ` ` J, e ` Jq. Given Archimedean ordered spaces V, W and a unital positive linear map ϕ : V Ñ W , the Archimedean quotient by ker ϕ is unitally order isomorphic to pV { ker ϕ, pV { ker ϕq`, e ` ker ϕq, where pV { ker ϕq` :" tv ` ker ϕ : @ε ą 0, Dj P ker ϕ such that j ` εe ` v P V `u. The map ϕ : V { ker ϕ Ñ W given by ϕpv ` ker ϕq " ϕpvq is a unital positive linear map. V ❅❅❅❅❅❅❅❅ ϕ q V { ker ϕ z✈✈✈✈✈✈✈✈✈ ϕ W 3. Maximal tensor products of function systems First, we give an explicit description for the positive elements in the maximal tensor products of function systems applying Archimedeanization. We define n V ` b W ` :" t ÿi"1 vi b wi P V b W : n P N, vi P V `, wi P W `u / / / / z 4 and KYUNG HOON HAN D :" tz P V b W : @ε ą 0, z ` εeV b eW P V ` b W `u. For v P V and w P W , we have }v}}w}eV b eW v b w 1 2 " p}v}eV vq b p}w}eW ` wq ` 1 2 p}v}eV ¯ vq b p}w}eW ´ wq P V ` b W `. Hence, pV b W, V ` b W `, eV b eW q is an ordered real vector space with an order unit. Given z P D and f P SpV q, g P SpW q, we have 0 ď pf b gqpz ` εeV b eW q ď pf b gqpzq ` ε for any ε ą 0. It follows that D Ă pV bmin W q`, so N :" D X ´D " t0u. Hence, the triple pV b W, D, eV b eW q is an Archimedeanization of pV b W, V ` b W `, eV b eW q. Theorem 3.1. Suppose that pV, V `, eV q and pW, W `, eW q are function systems. Then the maximal tensor product V bmax W coincides with the Archimedeanization pV b W, D, eV b eW q of pV b W, V ` b W `, eV b eW q. Proof. Ă) Let z P pV bmax W q` and f be a state on pV b W, D, eV b eW q. Since V ` b W ` Ă D, we have f V `bW ` ě 0, so f P SmaxpV b W q. It follows that f pzq ě 0 for all states f on pV b W, D, eV b eW q. By [PT, Proposition 2.20], z belongs to D. Ą) Let z P D and f P SmaxpV b W q. Since z ` εeV b eW P V ` b W ` for ε ą 0, we have 0 ď f pz ` εeV b eW q " f pzq ` ε. It follows that f pzq ě 0, so z P pV bmax W q`. (cid:3) The maximal tensor products of function systems are characterized by the following universal property. Proposition 3.2. Suppose that V, W and Z are function systems and Φ : V W Ñ Z is a bilinear map such that Φpv, wq P Z ` for all v P V ` and w P W `. Then there exists a unique positive linear map Φ : V bmax W Ñ Z such that Φpv, wq " Φpv b wq. V W #●●●●●●●●● Φ V bmax W ytttttttttt Φ Z Proof. Suppose that z P pV bmax W q`. Then, we have z ` εeV b eW P V ` b W ` for any ε ą 0. It follows that 0 ď Φpz ` εeV b eW q " Φpzq ` εΦpeV , eW q ď Φpzq ` ε}ΦpeV , eW q}eZ for any ε ą 0. Since eZ is Archimedean, Φpzq belongs to Z `. (cid:3) Proposition 3.3. Suppose that V and W are function systems. A functional ϕ : V bmax W Ñ R is positive if and only if its associated linear map Lϕ : V Ñ W is positive. Proof. Since xLϕpvq, wy " ϕpv b wq, v P V, w P W, Lϕ : V Ñ W is positive if and only if ϕ : V W Ñ R is a positive bilinear map. By Proposition 3.2, this is equivalent to the positivity of ϕ : V bmax W Ñ R. (cid:3) / / # y TENSOR PRODUCTS OF FUNCTION SYSTEMS REVISITED 5 The second dual of a function system is also a function system [Ef, Proposition 3.7]. Proposition 3.4. Suppose that V, W are function systems and ιW : W Ñ W is a canonical inclusion. Then the map idV b ιW : V bmax W Ñ V bmax W is an order embedding. Proof. Suppose that idV bιW pzq P pV bmaxW q` and ϕ P SpV bmaxW q. By Proposition 3.3, the associated linear map Lϕ : V Ñ W is positive. Since the composition ιW Lϕ : V Ñ W is also positive, its associated functional ϕ^ on V bmax W is a state by Proposition 3.3 again. From ϕpzq " ϕ^pidV b ιW pzqq ě 0, ϕ P SpV bmax W q, we see that z P pV bmax W q`. (cid:3) Proposition 3.5. Suppose that pV, V `, eV q and pW, W `, eW q are function systems. The order norms } }V bminW and } }V bmaxW are cross norms with respect to the order norms of V and W . In addition, the inequality } }V bminW ď } }V bmaxW holds. Proof. Let v P V and w P W . From }v}}w}eV b eW v b w 1 2 p}v}eV vq b p}w}eW ` wq ` 1 2 " p}v}eV ¯ vq b p}w}eW ´ wq P V ` b W `, we see that } }V bmaxW is a subcross norm. By the definition of the order norm, the inclusion pV bmax W q` Ă pV bmin W q` implies the inequality } }V bminW ď } }V bmaxW . It follows that }v}}w} " suptpf b gqpv b wq : f P SpV q, g P SpW qu ď }v b w}V bminW ď }v b w}V bmaxW ď }v}}w}, because f b g is a state on V bmin W . (cid:3) Let A P MmpRq and B P MnpRq. We denote the functional X P MmpRq ÞÑ trpXAq P R by trp Aq. From AA b BB " pA b BqpA b Bq and trp Aq b trp Bq " trp pA b Bqq, we see that pMmpRq bmax MnpRqq` Ă MmnpRq` Ă pMmpRq bmin MnpRqq`. Since the transpose map t on M2pRq is a unital positive map and idM2pRq b tp 1 0 0 1 0 0 0 0 0 0 0 0 1 0 0 1 q " ‹‹‚ 1 0 0 0 0 0 1 0 0 1 0 0 0 0 0 1 ‹‹‚ R M4pRq`, 6 KYUNG HOON HAN we see that and 1 0 0 1 0 0 0 0 0 0 0 0 1 0 0 1 1 0 0 0 0 0 1 0 0 1 0 0 0 0 0 1 ‹‹‚ ‹‹‚ P M4pRq` z pM2pRq bmax M2pRqq` P pM2pRq bmin M2pRqq` z M4pRq`. Definition 3.6. Suppose that T : V Ñ W is a unital positive surjective linear map for function systems V and W . We call T : V Ñ W an order quotient map if for any w in W ` and ε ą 0, we can find an element v in V such that v ` εeV P V ` and T pvq " w. The key point of the above definition is that the lifting v depends on the choice of ε ą 0. By slightly modifying [PT, Theorem 2.45], we obtain the following proposition, which justifies the term order quotient map. Proposition 3.7. Suppose that T : V Ñ W is a unital positive surjective linear map for function systems V and W . Then T : V Ñ W is an order quotient map if and only if T : V { ker T Ñ W is an order isomorphism. Proof. T : V Ñ W is an order quotient map ô @w P W `, @ε ą 0, Dv P V, v ` εeV P V ` and T pvq " w ô @w P W `, Dv P V, v ` ker T P pV { ker T q` and T pv ` ker T q " w ô T : V { ker T Ñ W is an order isomorphism. Recall that a bounded linear map T : V Ñ W for normed spaces V and W is called a quotient map if it maps the open unit ball of V onto the open unit ball of W . Proposition 3.8. Suppose that T : V Ñ W is a unital positive linear map for function systems V and W . If T : V Ñ W is a quotient map, then it is an order quotient map. (cid:3) Proof. Let w P W ` and ε ą 0. Then we have ´ 1 2 }w}eW ď w ´ 1 2 }w}eW ď 1 2 }w}eW . There exists an element v in V such that It follows that T pvq " w ´ 1 2 }w}eW and }v} ď v ` 1 2 }w}eV ` εeV P V ` and T pv ` 1 2 1 2 }w} ` ε. }w}eV q " w. The following theorem shows the projectivity of the maximal tensor product. Theorem 3.9. For function systems V1, V2, W and an order quotient map Q : V1 Ñ V2, the linear map Q b idW : V1 bmax W Ñ V2 bmax W is an order quotient map. (cid:3) TENSOR PRODUCTS OF FUNCTION SYSTEMS REVISITED 7 Proof. Let z P pV2 bmax W q` and ε ą 0. Then we have We write z ` ε 2 eV2 b eW P V ` 2 b W `. z " n ÿi"1 vi b wi ´ ε 2 eV2 b eW for vi P V ` 2 and wi P W `. There exists ui in V1 such that Qpuiq " vi and ui ` ε 2n}wi} eV1 P V ` 1 for each 1 ď i ď n. It follows that pQ b idW qp n ÿi"1 ui b wi ´ ε 2 eV1 b eW q " n ÿi"1 vi b wi ´ ε 2 eV2 b eW " z and n p ui b wi ´ ε 2 eV1 b eW q ` εeV1 b eW n ÿi"1 ÿi"1 1 b W ` P V ` pui ` " ε 2n}wi} eV1q b wi ` n ÿi"1 ε 2n eV1 b peW ´ 1 }wi} wiq (cid:3) 4. Nuclear function systems A function system V is called nuclear if the identity V bmin W " V bmax W holds for any function system W . Proposition 4.1. Suppose that E is a finite-dimensional function system and s is a faithful state on E. Then, pE, pEq`, sq is a function system. Proof. Let f be a linear functional on E. Since the set K :" tx P E` : }v} " 1u is compact and spxq ą 0 for any x P K, the continuous function f s has a maximum M on K. Then we have ´M s ď f ď M s. Suppose that f ` εs ě 0 for any ε ą 0. Then 0 ď pf ` εsqpxq " f pxq ` εspxq for all ε ą 0 and x P E`, so f ě 0. (cid:3) Proposition 4.2. Suppose that E and W are function systems with E finite dimen- sional. Then z P pE bmin W q` if and only if its associated linear map Tz : E Ñ W is positive. Proof. Since E is finite dimensional, every positive functional on E is an evaluation v for some v P E`. We have z P pE bmin W q` if and only if 0 ď pv b f qpzq " f pTzpvqq for all v P E` and f P pW q`, which is equivalent to the positivity of Tz : E Ñ W . (cid:3) We apply the idea of [HP, Theorem 3.1] to function systems. 8 KYUNG HOON HAN Theorem 4.3. Suppose that Φ : V Ñ W is a unital positive map for function systems V and W . The following are equivalent: (i) the map idA b Φ : A bmin V Ñ A bmax W is positive for any function system A; (ii) the map idE b Φ : E bmin V Ñ E bmax W is positive for any finite dimensional function system E; (iii) there exist nets of unital positive maps ϕλ : V Ñ ℓ8 nλ and ψλ : ℓ8 nλ Ñ W such that ψλ ϕλ converges to the map Φ in the point-norm topology. V ❆❆❆❆❆❆❆❆ ϕλ Φ ℓ8 nλ W >⑤⑤⑤⑤⑤⑤⑤⑤ ψλ Proof. Clearly, (i) implies (ii). piiiq ñ piq. First, let us show that ℓ8 in pA bmin ℓ8 n q`. We have n is nuclear. We choose an element řn ÿi"1 xi b eiq " f pxjq jqp n i"1 xi b ei 0 ď pf b e1 for all f P SpAq and 1 ď j ď n. We see that xj P A`, thus n ÿi"1 xi b ei P A` b ℓ8 n `. Hence, ℓ8 n is nuclear. The map idA b ψλ ϕλ : A bmin V Ñ A bmin ℓ8 nλ " A bmax ℓ8 nλ Ñ A bmax W is unital positive. Since } }AbmaxW is a cross norm by Proposition 3.5, pidA b ψλ ϕλqpzq converges to idA b Φpzq for each z P A b V . It follows that z P pA bmin V q` implies idA b Φpzq P pA bmax W q`. piiq ñ piiiq. Let E be a finite dimensional function subsystem of V and s be a faithful state on E. By Proposition 4.1, pE, pEq`, sq is a function system. By Proposition 4.2, we can regard the inclusion ι : E Ă V as an element in pE bmin V q`. The restriction ΦE : E Ñ W can be identified with an element pidE b Φqpιq. By assumption, this belongs to pE bmax W q`. We consider the directed set Ω " tpE, εq : E is a finite dimensional function subsystem of V , ε ą 0u with the standard partial order. Let λ " pE, εq. For any ε ą 0, the restriction ΦE can be written as nλ ΦE ` εs b eW " fk b wk ÿk"1 for nonzero fk P pEq` and wk P W ` (1 ď k ď nλ). The map f : x P E ÞÑ pf1pxq, , fnλpxqq P ℓ8 nλ / / > TENSOR PRODUCTS OF FUNCTION SYSTEMS REVISITED 9 is positive, and we may assume that f peV q " pc1, , cmλ, 0, , 0q for ck ą 0 by rearrangement. For x P E`, we have 0 ď f pxq ď }x}f peV q " }x}pc1, , cmλ, 0, , 0q. Since every element in E can be written as a difference of positive elements in E and each fk is assumed to be nonzero, we have that mλ " nλ and f peV q is invertible. Because nλ nλ fk b wk " ÿk"1 ÿk"1 fk fkpeV q b fkpeV qwk, we may assume that f is a unital positive map. By the Krein theorem, f : E Ñ ℓ8 nλ extends to a unital positive map ϕλ : V Ñ ℓ8 nλ Ñ W by nλ. We define a positive map ψ1 λ : ℓ8 ψ1 λpc1, , cnλq " nλ ÿk"1 ckwk. For x P E, we have }Φpxq ´ ψ1 λ ϕλpxq} " }Φpxq ´ nλ ÿk"1 fkpxqwk} " ε}spxqeW } ď ε}x}. Hence, we can take nets of unital positive maps ϕλ : V Ñ ℓ8 nλ Ñ W such that ψ1 ψ1 λ : ℓ8 Since each ϕλ is unital, ψ1 set n and positive maps λ ϕλ converges to the map Φ in the point-norm topology. λp1nλq converges to eW . Let us choose a state ωλ on ℓ8 nλ and ψλpcq :" ψ1 λpcq ` ωλpcqpeW ´ ψ1 λp1nλqq. 1 }ψ1 λ} 1 }ψ1 λ} Then ψλ : ℓ8 in the point-norm topology. nλ Ñ W is a unital positive map such that ψλ ϕλ converges to the map Φ (cid:3) As an application of Theorem 4.3, we give elementary proofs of known characteriza- tions of nuclear function systems [Ef, Theorem 7.4], [NP, Theorem 2.2] except for the implication from the last. Theorem 4.4. Let V be a function system. The following are equivalent: (i) V is nuclear; (ii) we have V bmin E " V bmax E for any finite dimensional function system E; (iii) there exist nets of unital positive maps ϕλ : V Ñ ℓ8 nλ and ψλ : ℓ8 nλ Ñ V such that ψλ ϕλ converges to idV in the point-norm topology; (iv) there exist nets of weak-continuous unital positive maps ϕλ : V Ñ ℓ8 nλ and nλ Ñ V such that ψλ ϕλ converges to idV in the point-weak topology; ψλ : ℓ8 (v) the second dual V is injective; (vi) given V Ă V Ă CpKq with a compact set K, there exists a unital positive map Φ : CpKq Ñ V such that ΦV " idV ; (vii) given both any function system W1 containing V and a function system W2, the inclusion V bmax W2 Ă W1 bmax W2 is an order embedding; (viii) given both any function system W containing V and a finite dimensional func- tion system E, the inclusion V bmax E Ă W bmax E is an order embedding; 10 KYUNG HOON HAN (ix) given both any function systems W1 Ă W2, the inclusion V bmax W1 Ă V bmax W2 is an order embedding; (x) the state space SpV q is a Choquet simplex. Proof. Putting V " W and Φ " idV in Theorem 4.3, we obtain the equivalences piq ô piiq ô piiiq. For a compact set K, we can show that CpKq satisfies (iii) using the partition of unity. For the details, see the proof of [L, Theorem 2.3.7]. The implications pvq ñ pviq, pviiq ñ pviiiq and piq ñ pixq are trivial. piiiq ñ pivq. The proof consists of the idea of [EOR, Thoerem 4.5] and the pertur- bation argument. Take ε ą 0 and a finite dimensional function subsystem E of V and a finite dimensional subspace F of V . By the principle of local reflexivity, we can choose a linear map θ : E Ñ V such that }θ} ă 1 ` ε, θpvq " v for v P E X V and xθpxq, f y " xx, f y for x P E, f P F . By the second condition, the map θ : E Ñ V is unital. By assumption, there exist unital positive maps ϕ : V Ñ ℓ8 n Ñ V such that }ψ ϕpθpxqq ´ θpxq} ď ε}θpxq} for all x P E. Since the inclusion ι : E Ă V is the dual map of ιV : V Ñ E, the map ιV : V Ñ E is a quotient map. Applying this to each coordinate of ϕ θ : E Ñ ℓ8 n , we obtain its weak-continuous extension ϕ1 : V Ñ ℓ8 n and ψ : ℓ8 n with }ϕ1} ă 1 ` ε. V ϕ1 ℓ8 n idV ϕ =⑤⑤⑤⑤⑤⑤⑤⑤ E θ V V ψ "❊❊❊❊❊❊❊❊❊ / V Suppose that V Ă CpKq for a compact set K. By the Hahn-Banach theorem and the Riesz representation theorem, ϕ1V : V Ñ ℓ8 n is the restriction of pµ1, , µnq : CpKq Ñ ℓ8 n for some µk P MRpKq with }µk} ă 1 ` ε. CpKq V pµ1, ,µnq '❖❖❖❖❖❖❖❖❖❖❖❖❖ ϕ1V / ℓ8 n Let µk " µ` k ´ µ´ k be the Jordan decomposition. Then, we have µ` k p1q ě µ` k p1q ´ µ´ k p1q " µkp1q " 1. It follows that }µk ´ µ` k k p1q µ` } µ` k ´ k ´ µ´ µ` k k p1q k p1q ´ 1 µ` k p1q k } ` }µ` k } ´ 1 k } ` µ` }µ` k } } " }µ` ď }µ´ " }µ´ " }µk} ´ 1 ă ε. / / / / " = / ? O O ' ? O O / TENSOR PRODUCTS OF FUNCTION SYSTEMS REVISITED 11 Let ϕ2 : V Ñ ℓ8 V . Then ϕ2 : V Ñ ℓ8 n be the second dual of the restriction of p 1 µ` n p1q n is a weak-continuous unital positive map satisfying 1 , , 1 µ` 1 p1q µ` µ` n q on }ϕ1 ´ ϕ2} " }ϕ1V ´ ϕ2V } ď maxt}µk ´ µ` k k p1q µ` } : 1 ď k ď nu ă ε. For all x P E and f P F , we have xψ ϕ2pxq ´ x, f y ďxψ ϕ2pxq ´ ψ ϕ1pxq, f y ` xψ ϕ θpxq ´ θpxq, f y ` xθpxq ´ x, f y ăp2ε ` ε2q}x}}f }. The index set tpε, E, F qu is directed in such a way that pε1, E1, F1q ď pε2, E2, F2q if and only if ε1 ě ε2 and E1 Ă E2, F1 Ă F2. pivq ñ pvq. Suppose that W1 Ă W2 are function systems and Φ : W1 Ñ V is a unital positive map. Let Φλ : W2 Ñ ℓ8 nλ be a unital positive extension of ϕλ Φ : W1 Ñ ℓ8 nλ. Then the point-weak cluster point of ψλΦλ : W2 Ñ V is the unital positive extension of Φ : W1 Ñ V . W2 Φλ W1 Φ / V / ℓ8 nλ id V ϕλ =④④④④④④④④ ψλ !❈❈❈❈❈❈❈❈ V pviq ñ piq. Suppose that V Ă CpKq for a compact set K. By Proposition 3.4, we have V bmin W / CpKq bmin W " CpKq bmax W ΦbidW / V bmax W pvq ñ pviiq. Since there exists a positive projection from W 1 onto V , the inclusion V bmax W2 Ă W 1 bmax W2 is an order embedding. By Proposition 3.4, we have V bmax W. W1 bmax W2 / W 1 bmax W2 V bmax W2 / V bmax W2. pviiiq ñ piiq. Suppose that V Ă CpKq for a compact set K. The conclusion follows from V bmax E / CpKq bmax E V bmin E / CpKq bmin E. pixq ñ piq. Suppose that W Ă CpKq for a compact set K. The conclusion follows from V bmax W / V bmax CpKq V bmin W / V bmin CpKq. / ! ? O O / / / = / / ? O O / / ? O O / / / / 12 KYUNG HOON HAN piiiq ñ pxq. Suppose that f and g are linear functionals on V . The net ϕ λ ψ λ : V Ñ V converges to idV in the point-weak topology. V ❇❇❇❇❇❇❇❇ ψ λ id V ℓ1 nλ V >⑤⑤⑤⑤⑤⑤⑤⑤ ϕ λ λpψ λpf q _ ψ nλ " pℓ8 λpgqq. Applying ϕ The ordered space ℓ1 point of ϕ limit, we see that f, g ď h. If f, g ď k, then ψ λpf q _ ψ ψ For pxq ñ piq, see [NP, Theorem 2.2]. Alternatively, see [Ef, Lemma 6.3, Corollary (cid:3) nλq is lattice ordered. Let h P V be a weak cluster λpgq and taking λ to λpgq ď ψ λpkq and taking limit, we also see that h ď k. λpf q, ψ λpkq. Applying ϕ λpgq ď ψ λ to ψ λpf q, ψ λpgq ď ψ λpf q _ ψ 6.4] for pxq ñ pvq. 5. The complex case An ordered -vector space pV, V `q is a pair consisting of a -vector space V and a proper cone V ` in Vh. We may define a partial ordering on Vh by defining v ě w if and only if v ´ w P V `. If pV, V `q is an ordered -vector space, an element e P Vh is called an order unit for V if, for all v P Vh, there exists a real number r ą 0 such that re ` v ě 0. If pV, V `q is an ordered -vector space with an order unit e, then we say that e is an Archimedean order unit if v P V ` whenever v P V and εe ` v ě 0 for any ε ą 0. In this case, we call the triple pV, V `, eq an Archimedean ordered -vector space. Paulsen and Tomforde proved that every Archimedean ordered -vector space can be embedded into a complex continuous function algebra on a compact Hausdorff space through a unital order isomorphism that is also isometric with respect to the minimal order norm [PT, Theorem 5.2]. In this section, we consider the tensor products of Archimedean ordered -vector spaces. For -vector spaces V and W , the involution on V b W is defined as pv b wq " v b w for v P V and w P W . Definition 5.1. Suppose that pV, V `, eV q and pW, W `, eW q are Archimedean ordered -vector spaces. (1) We define a minimal tensor product V bmin W as pV bW, pV bmin W q`, eV beW q, where pV bmin W q` " tz P V b W : pf b gqpzq ě 0 for all f P SpV q, g P SpW qu. (2) We define a maximal tensor product V bmaxW as pV bW, pV bmaxW q`, eV beW q, where pV bmax W q` " tz P V b W : z ` εeV b eW P V ` b W ` for all ε ą 0u. If a norm } } on V satisfies }v} " }v} for all v P V and extends the order norm on Vh, then we call it an order norm on V . Two extremal order norms on Archimedean ordered -vector spaces are defined in [PT, Definition 4.4] and [PT, Definition 4.6]. The minimal order norm } }m on V is defined by The maximal order norm } }M on V is defined by }v}m :" suptf pvq : f : V Ñ C is a stateu. n }v}M :" inft ÿi"1 λi}vi} : v " n ÿi"1 λivi with vi P Vh and λi P Cu. / / > TENSOR PRODUCTS OF FUNCTION SYSTEMS REVISITED 13 Theorem 5.2. Suppose that pV, V `, eV q and pW, W `, eW q are Archimedean ordered -vector spaces. (1) We have pV bmin W q` " pVh bmin Whq`. Hence, the minimal tensor product V bmin W is an Archimedean ordered -vector space. (2) We have pV bmax W q` " pVh bmax Whq`. Hence, the maximal tensor product V bmax W is an Archimedean ordered -vector space. (3) The minimal order norm induced by the minimal tensor product is a cross norm with respect to the minimal order norms of V and W . (4) The maximal order norm induced by the maximal tensor product is a subcross norm with respect to the maximal order norms of V and W . Proof. (1) First, let us show that pV bmin W q` Ă Vh b Wh. Let z " řn pV bmin W q`. We may assume that each vk is Hermitian and tvkun independent. For f P SpV q and g P SpW q, we have k"1 vk b wk P k"1 is R-linearly n ÿk"1 f pvkqgpw kq " n ÿk"1 It follows that f pvkqgpwkq " pf b gqpzq " pf b gqpzq " n ÿk"1 f pvkqgpwkq. f p n ÿk"1 gpwk ´ w k qvkq " 0 for all f P SpV q. By [PT, Proposition 3.12], we have řn that k"1 gpwk ´ w k qvk " 0. We see n pRegpwk´w kqqvk " ÿk"1 n ÿk"1 Repgpwk´w kqvkq " 0 " n ÿk"1 Impgpwk´w kqvkq " n pImgpwk´w kqqvk. ÿk"1 Since tvkun 1 ď k ď n. By [PT, Proposition 3.12] again, wk is Hermitian. k"1 is R-linearly independent, we have gpwk ´ w k q " 0 for all g P SpW q and For a real functional f : Vh Ñ R, the complexification f : V Ñ C is defined by f pvq " f p v ` v 2 q ` if p v ´ v 2i q [PT, Definition 3.9]. Then, f : Vh Ñ R is a state if and only if f : V Ñ C is a state [PT, Proposition 3.10]. Moreover, every state on V is realized in this form [PT, Proposition 3.11]. It follows that pV bmin W q` " pVh bmin Whq`. (3) We denote by } }V bminW,m the minimal order norm induced by the minimal tensor product V bmin W . For v P V and w P W , we have }v}m}w}m " suptpf b gqpv b wq : f P SpV q, g P SpW qu ď }v b w}V bminW,m because f b g P SpV bmin W q. For a state F on V bmin W , we let F pv b wq " eiθF pv b wq and u " e´iθv. 14 KYUNG HOON HAN Then, we have F pv b wq "F pu b wq "F p "F pp u b w ` 1 2 u ` u 1 2 q b p u b wq w ` w u ´ u w ´ w q ´ p q b p qq 2i w ´ w 2i 2 u ` u 2 " suptf p 2 w ` w 2 qgp 2i u ´ u 2i q ´ f p 2 u ` u 2 u ´ u qqpgp ď suptpf p q ` if p 2 2i 2 " supt f puqgpwq : f P SpVhq, g P SpWhqu "}u}m}w}m "}v}m}w}m. ď}p q b p q ´ p q b p q}VhbminWh u ` u w ` w u ´ u w ´ w qgp 2i w ` w 2i q ` igp q : f P SpVhq, g P SpWhqu w ´ w 2i qq : f P SpVhq, g P SpWhqu (4) We denote by } }V bmaxW,M the maximal order norm induced by the maximal tensor product V bmax W . For v P V and w P W , we write v " m ÿk"1 λkvk and w " n ÿl"1 µlwl for λk, µl P C and vk P Vh, wl P Wh. Then we have v b w " ÿ1ďkďm 1ďlďn λkµl vk b wl and It follows that m p ÿk"1 λk}vk}qp n ÿl"1 µl}wl}q " ÿ1ďkďm 1ďlďn λkµl}vk b wl}VhbmaxWh. }v b w}V bmaxW,M ď }v}M }w}M . (cid:3) For the definitions of OMIN and OMAX in the following proposition, we refer to [PTT, Definition 3.3] and [PTT, Definition 3.12]. Proposition 5.3. For an Archimedean ordered -vector space V , we have MnpOMINpV qq` " pMnpCq bmin V q` and MnpOMAXpV qq` " pMnpCq bmax V q`. Proof. For f P SpMnpCqq and g P SpV q, we have f prgpvijqspi,jqq " f p n ÿi,j"1 gpvijqeijq " n ÿi,j"1 f peijqgpvijq " pf b gqp n ÿi,j"1 eij b vijq. The first identity follows from [PTT, Theorem 3.2]. (cid:3) TENSOR PRODUCTS OF FUNCTION SYSTEMS REVISITED 15 Hereafter, we list the complex versions of the statements in Sections 3 and 4. Their proofs are similar to those of real cases, or the restriction and the complexification [PT, Remark 3.14] enable us to reduce them to the real cases. Hence, most proofs will be omitted. Proposition 5.4. Suppose that V, W and Z are Archimedean ordered -vector spaces and Φ : V W Ñ Z is a bilinear map such that Φpv, wq P Z ` for all v P V ` and w P W `. Then, there exists a unique positive linear map Φ : V bmax W Ñ Z such that Φpv, wq " Φpv b wq. Proposition 5.5. Suppose that S : V1 Ñ V2 and T : W1 Ñ W2 are unital positive linear maps for Archimedean ordered -vector spaces V1, V2, W1, W2. Then (1) S b T : V1 bmin W1 Ñ V2 bmin W2 is a unital positive linear map. (2) S b T : V1 bmax W1 Ñ V2 bmax W2 is a unital positive linear map. Definition 5.6. Suppose that T : V Ñ W is a unital positive surjective linear map for Archimedean ordered -vector spaces V and W . We call T : V Ñ W an order quotient map if for any w in W ` and ε ą 0, we can find an element v in V such that v ` εeV P V ` and T pvq " w. Proposition 5.7. Suppose that T : V Ñ W is a unital positive surjective linear map for Archimedean ordered -vector spaces V and W . Then T : V Ñ W is an order quotient map if and only if T : V { ker T Ñ W is an order isomorphism. Proposition 5.8. Suppose that T : V Ñ W is a unital positive linear map for Archimedean ordered -vector spaces V and W . Then (1) T : V Ñ W is an order embedding if and only if it is an isometry with respect to the minimal order norms, and (2) T : V Ñ W is an order quotient map if it is a quotient map with respect to the order norms. Proof. (1) This follows from [PT, Theorem 4.22]. (2) In the proof of Proposition 3.8, we consider the Hermitian lifting 1 2pv ` vq. (cid:3) Theorem 5.9. (1) For Archimedean ordered -vector spaces V1, V2, W and a unital order embedding ι : V1 Ñ V2, the linear map ι b idW : V1 bmin W Ñ V2 bmin W is a unital order embedding. (2) For Archimedean ordered -vector spaces V1, V2, W and an order quotient map Q : V1 Ñ V2, the linear map Q b idW : V1 bmax W Ñ V2 bmax W is an order quotient map. Proof. (1) Combining [PT, Corollary 2.15] with [PT, Proposition 3.11], we obtain a Hahn-Banach type theorem for a state on an Archimedean ordered -vector space. (cid:3) Definition 5.10. An Archimedean ordered -vector space V is called nuclear if the identity V bmin W " V bmax W holds for any Archimedean ordered -vector space W . Proposition 5.11. An Archimedean ordered -vector space V is nuclear if and only if the Archimedean ordered space Vh is nuclear. 16 KYUNG HOON HAN Proof. Let W be an Archimedean ordered space and W C :" W ' iW be its complexi- fication [PT, Remark 3.14]. The conclusion follows from pVh bmin W q` " pV bmin W Cq` and pVh bmax W q` " pV bmax W Cq`. (cid:3) Theorem 5.12. Suppose that Φ : V Ñ W is a unital positive map for Archimedean ordered -vector spaces V and W . The following are equivalent: (i) the map idA b Φ : A bmin V Ñ A bmax W is positive for any Archimedean ordered -vector space A; (ii) the map idE b Φ : E bmin V Ñ E bmax W is positive for any finite dimensional Archimedean ordered -vector space E; (iii) there exist nets of unital positive maps ϕλ : V Ñ ℓ8 nλpCq and ψλ : ℓ8 such that ψλ ϕλ converges to the map Φ in the point-norm topology. nλpCq Ñ W Proof. piiq ñ piiiq. The map idE b ΦVh : E bmin Vh Ñ E bmax Wh is positive for any finite dimensional Archimedean ordered space E. By Theorem 4.3, there exist nets of unital positive maps ϕλ : Vh Ñ ℓ8 nλpRq Ñ Wh such that ψλ ϕλ converges to ΦVh in the point-norm topology. Let ϕλ : V Ñ ℓ8 nλpCq and nλpCq Ñ W be the complexifications of ϕ and ψ, respectively. Then, ψλ ϕλ ψλ : ℓ8 converges to the map Φ in the point-norm topology. nλpRq and ψλ : ℓ8 (cid:3) Theorem 5.13. Let V be an Archimedean ordered -vector space. The following are equivalent: (i) V is nuclear; (ii) we have for any finite dimensional Archimedean ordered -vector space E; V bmin E " V bmax E (iii) there exist nets of unital positive maps ϕλ : V Ñ ℓ8 nλpCq and ψλ : ℓ8 nλpCq Ñ V such that ψλ ϕλ converges to idV in the point-norm topology; (iv) there exist nets of weak-continuous unital positive maps ϕλ : V Ñ ℓ8 nλpCq nλpCq Ñ V such that ψλ ϕλ converges to idV in the point-weak and ψλ : ℓ8 topology; (v) the second dual V is injective; (vi) given V Ă V Ă CpKq with a compact set K, there exists a unital positive map Φ : CpKq Ñ V such that ΦV " idV ; (vii) given both any Archimedean ordered -vector space W1 containing V and an Archimedean ordered -vector space W2, the inclusion V bmax W2 Ă W1 bmax W2 is an order embedding; (viii) given both any Archimedean ordered -vector space W containing V and a finite dimensional Archimedean ordered -vector space E, the inclusion V bmax E Ă W bmax E is an order embedding; (ix) given both any Archimedean ordered -vector spaces W1 Ă W2, the inclusion V bmax W1 Ă V bmax W2 is an order embedding; TENSOR PRODUCTS OF FUNCTION SYSTEMS REVISITED 17 (x) the state space SpVhq is a Choquet simplex. Acknowledgments The author is grateful to the referee for careful reading and bringing his attention to Ref. [El, F, GL, BR, GK]. References [A] E.M. Alfsen, Compact convex sets and boundary integrals, Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 57. Springer-Verlag, New York-Heidelberg, 1971. [BR] G.J.H.M. Buskes and A.C.M. van Rooij, The Archimedean l-group tensor product, Order 10 [Ef] [El] (1993), no. 1, 93-102. E.G. Effros, Injectives and Tensor Products for Convex Sets and C -Algebras, NATO Advanced Study Institute, University College of Swansea, 1972. A.J. Ellis, Linear operators in partially ordered normed vector spaces, J. London Math. Soc. 41 (1966) 323-332. [EOR] E.G. Effros, N. Ozawa and Z.-J. Ruan, On injectivity and nuclearity for operator spaces, Duke [F] Math. J. 110 (2001) 489 -- 521. D.H. Fremlin, Tensor products of Archimedean vector lattices, Amer. J. Math. 94 (1972), 777- 798. [GK] O. van Gaans and A. Kalauch, Tensor products of Archimedean partially ordered vector spaces, [GL] Positivity 14 (2010), no. 4, 705-714. J.J. Grobler and C.C.A. Labuschagne, The tensor product of Archimedean ordered vector spaces, Math. Proc. Cambridge Philos. Soc. 104 (1988), no. 2, 331-345. [HP] K.H. Han and V.I. Paulsen, An approximation theorem for nuclear operator systems, J. Funct. Anal. 261 (2011) 999 -- 1009. [Ka] R. V. Kadison, A representation theory for commutative topological algebra. Mem. Amer. Math. Soc. 1951 (1951) no. 7. [KPTT1] A. Kavruk, V.I. Paulsen, I.G. Todorov and M. Tomforde, Tensor products of operator sys- tems, J. Funct. Anal. 261 (2011) no.2 267 -- 299. [KPTT2] A. Kavruk, V.I. Paulsen, I.G. Todorov and M. Tomforde, Quotients, exactness and WEP in [L] [NP] the operator systems category, Adv. Math. 235 (2013) 321-360. H. Lin, An introduction to the classification of amenable C -algebras, World Scientific Publish- ing Co., Inc., River Edge, NJ, 2001. I. Namioka and R. R. Phelps, Tensor products of compact convex sets, Pacific J. Math. 31 (1969) 469-480. [PT] V.I. Paulsen and M. Tomforde, Vector spaces with an order unit, Indiana Univ. Math. J. 58 (2009) no.3 1319 -- 1359. [PTT] V.I. Paulsen, I.G. Todorov and M. Tomforde, Operator system structures on ordered spaces, Proc. London Math. Soc. (3) 102 (2011) 25 -- 49. Department of Mathematics, The University of Suwon, Gyeonggi-do 445-743, Korea E-mail address: [email protected]
1806.02443	6	1806	2019-12-30T11:42:52	Equilibrium states and entropy theory for Nica-Pimsner algebras	[ "math.OA", "math.FA" ]	We study the equilibrium simplex of Nica-Pimsner algebras arising from product systems of finite rank on the free abelian semigroup. First we show that every equilibrium state has a convex decomposition into parts parametrized by ideals on the unit hypercube. Secondly we associate every gauge-invariant part to a sub-simplex of tracial states of the diagonal algebra. We show how this parametrization lifts to the full equilibrium simplices of non-infinite type. The finite rank entails an entropy theory for identifying the two critical inverse temperatures: (a) the least upper bound for existence of non finite-type equilibrium states, and (b) the least positive inverse temperature below which there are no equilibrium states at all. We show that the first one can be at most the strong entropy of the product system whereas the second is the infimum of the tracial entropies (modulo negative values). Thus phase transitions can happen only in-between these two critical points and possibly at zero temperature.	math.OA	math	EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS EVGENIOS T.A. KAKARIADIS Abstract. We study the equilibrium simplex of Nica-Pimsner algebras arising from product systems of finite rank on the free abelian semigroup. First we show that every equilibrium state has a convex decomposition into parts parametrized by ideals on the unit hypercube. Secondly we associate every gauge-invariant part to a sub-simplex of tracial states of the diagonal algebra. We show how this parametrization lifts to the full equilibrium simplices of non-infinite type. The finite rank entails an entropy theory for identifying the two critical inverse temper- atures: (a) the least upper bound for existence of non finite-type equilibrium states, and (b) the least positive inverse temperature below which there are no equilibrium states at all. We show that the first one can be at most the strong entropy of the product system whereas the second is the infimum of the tracial entropies (modulo negative values). Thus phase transitions can happen only in-between these two critical points and possibly at zero temperature. 9 1 0 2 c e D 0 3 ] . A O h t a m [ 6 v 3 4 4 2 0 . 6 0 8 1 : v i X r a 1. Introduction Kubo-Martin-Schwinger states encapture the properties of a quantum system at thermal equilibrium. Taking motivation from the grand canonical form of Gibbs states, by now there is a well established theory of KMS-states for C-dynamical systems over R. Equilibrium states form an effective tool for R-equivariant isomorphisms and the community has focused on their parametrization. It has become evident that they are tractable in particular for algebras arising from Fock type constructions. Complete results have been established for Z+-systems and algebras of semigroups of particular type, while steps have been taken forward in specific multivariable contexts. In these cases the goal is to attain a full decomposition and parametrization of the equilibrium E-simplex. Surprisingly however much less is known for ZN + , even when finite unit decompositions are available. The overarching aim of this paper is to tackle this class. Namely, we establish the multivariable Wold decomposition and parametrization combined with (several notions of) entropy for Nica-Pimsner algebras of finite rank ZN + -product systems. 1.1. Background. In the early 2000s Laca-Neshveyev [25] identified the scaling property that characterizes the E-simplex for a large class of C-algebras, namely the Pimsner algebras. Introduced in [30] and finessed by Katsura [24], Pimsner algebras are naturally constructed from a single Hilbert bimodule X over a C-algebra A, better known as a C-correspondence. They encompass a great variety of constructs, e.g., dynamical systems, transfer operators, graphs, and there are two main variants. The Toeplitz-Pimsner algebra T (X) is generated by the left creation operators, while the Cuntz-Pimsner algebra O(X) is a quotient by an appropriate sub-ideal of the compacts K(FX), the smallest possible so that O(X) attains an isometric copy of X and A. One of the breakthrough findings in [25] is that the simplex of equilibrium states Eβ(T (X)) at inverse temperature β has a rather rich structure that stabilizes after a critical temperature βc. Moreover, every equilibrium state of T (X) has a unique convex decomposition in a finite part (arising from K(FX)) and in an infinite part (inducing on T (X)/K(FX)). The latter quotient is not always the anticipated O(X), but they do coincide when X is regular. 2010 Mathematics Subject Classification. 46L30, 46L55, 46L08, 58B34. Key words and phrases: equilibrium states, product systems, Nica-Pimsner algebras. 1 2 E.T.A. KAKARIADIS Following the work of Bost-Connes [3], Laca-Raeburn [26] applied the framework of [25] to study C-algebras of the (ax + b)-semigroup. In the process Laca-Raeburn [26] formalized an algorithm for parametrizing equilibrium states of finite type that has been widely applicable. A great volume of subsequent works for specific one-variable constructs emerged from [26]. One of the common characteristics is the identification of two critical inverse temperatures c and βc: there are no equlibrium states below β′ β′ c while it is just the finite part that survives above βc. Hence phase transitions can only happen in-between β′ c and βc. A second type of phase transition may occur at β = ∞ between ground states and KMS∞-states. Although the C-algebra T (N ⋊ N×) of [26] does not fit in the Pimsner class, it accom- modates several of the arguments of [25]. As shown by Brownlowe-an Huef-Laca-Raeburn [7] it is a Toeplitz-Nica-Pimsner algebra in the sense of Fowler [15]. In this case the product system X = {Xp}p∈P is a family of C-correspondences (along with multiplication rules) over a quasi-lattice group (G, P ). In addition to Fowler's Toeplitz-Nica-Pimsner algebra N T (X) of the Fock representation, Carlsen-Larsen-Sims-Vittadello [8] proved the existence of a Cuntz-Nica-Pimsner algebra N O(X) as co-universal with respect to the Gauge-Invariant- Uniqueness-Theorem, for rather general pairs (G, P ). Nica-Pimsner algebras encode in turn a variety of semigroup transformations. Conse- quently the community envisaged a program that builds on [25, 26] and aspires to under- stand the E-structure for Nica-Pimsner algebras and their quotients. There is a number of worked out cases that shapes the framework in the following sense: (Q.1) What is a canonical Wold decomposition for (G, P )? (Q.2) How does it characterize the types of the equilibrium states? (Q.3) How does the product system data affect the critical temperatures? Results in this direction can then be used for rigidity or classification purposes. For example equilibrium states have been used recently for reconstruction of graphs [11]. Let us review some of the obtained parametrizations. Brownlowe-an Huef-Laca-Raeburn [7] studied the E-structure for quotients of T (N ⋊ N×) of [26] in conjuction with previous work of Larsen [28] on multivariable transfer operators. These constructs of product systems attain orthonormal bases and Hong-Larsen-Szymanski [16, 17] also studied equilibrium states in such generality. Under some extra conditions they identified a finite-type only structure at suitably high inverse temperatures. However the critical points β′ c and βc, or states of non-finite type are not computed in [17]. Afsar-Brownlowe-Larsen-Stammeier [1] recently extended this by identifying the equilibrium simplex for the Toeplitz-type C-algebras of right LCM semigroups, with a description for the critical interval. In the particular case where the product system comes from a ZN + -dynamical system it has been shown in [22] that β′ c = βc = log N and the states are either finite or infinite. Moving beyond orthonormal bases, Afsar-an Huef-Raeburn [2] considered product systems arising from local homeomorphisms. An upper bound for βc is given but it is not shown if it is optimal. Considerably more has been achieved for higher-rank graphs that admit orthogonal but not orthonormal bases. If the graph is irreducible then it is shown in [19] that β′ c and βc coincide with the logarithm of the common Perron-Frobenius eigenvalue. Fletcher-an Huef-Raeburn [14] later formalized an algorithm for computing the E-structure of their Toeplitz algebras. However the lowest inverse temperature β′ c is not computed. Very recently Christensen [9] answered questions (Q.1) and (Q.2) and provided the full parametrization for higher rank graphs without any of the assumptions of [14]. His independent approach fits exactly the framework we have been developing for the current paper. There are two additional works of the author that inform our direction, in particular with respect to (Q.3). In [23] we revisited the description of [25] for finite rank C-corres- pondences. By using several notions of entropy inspired by Pinzari-Watatani-Yonetani [31] we identified the critical values through the structural data of X. Moreover we identified the E-structure not just for T (X) but also for all relative Cuntz-Toeplitz algebras, including a full parametrization for O(X) itself (without assuming injectivity of X). At the same time EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 3 finite rank product systems are strong compactly aligned product systems as in [13]. We can thus use the analysis of [13] for the realization of ideals of generalized compacts and of N O(X). 1.2. Main results. The class of all product systems and quasi-lattices is extremely vast to hope to treat all cases in considerable depth in one stroke. In this paper we wish to answer Questions (1) -- (3) for finite rank product systems over ZN + , a class that encompasses a number of cases, e.g., [2, 9, 14, 18, 19, 20, 22, 28]. The reader is advised to consult Section 2.1 for the ZN + -notation we use. Every Xi is of finite rank, i.e., it admits a family {xi,j j = 1, . . . , di} of vectors in the unit ball such that (1.1) ξi = diXj=1 xi,jhxi,j, ξii for all ξi ∈ Xi. Thus so does every Xn for the family {xµ ℓ(µ) = n}, or equivalently KXn = LXn for all n ∈ ZN + . Note that we do not assume orthogonality of the xi,j and the decomposition may not be unique. Such families are sometimes referred to as Parseval frames, e.g., [17], but here we will term x = {xi,j j = 1, . . . , di, i = 1, . . . , N } as a unit decomposition of X. The Toeplitz-Nica-Pimsner algebra is the C-algebra generated by the Fock left creation operators {π(a), t(ξn) a ∈ A, ξn ∈ Xn, n ∈ ZN m Xm. The rotation along the TN -gauge action corresponds to the multivariable Quantum Harmonic Oscillator as with the one-dimensional case, cf., [23]. By passing to the unitization we may assume that X is unital. Since X is of finite type, the projections + } on the full Fock space FX =P⊕ (1.2) Pi := diXj=1 t(xi,j)t(xi,j)∗ and QF :=Yi∈F (1 − Pi) are in N T (X) and commute with π(A). Suppose that {IF ∅ 6= F ⊆ {1, . . . , N }} is a lattice of ⊥-invariant ideals in the sense that IF ⊆ IF ′ when F ⊆ F ′ ⊆ {1, . . . , N } and (1.3) hXn, IF Xni ⊆ IF for all n ⊥ F. Then we define N O(I, X) be the quotient of N T (X) by the ideal (1.4) KI := hπ(a)QF a ∈ IF , ∅ 6= F ⊆ {1, . . . , N }i. By [13] the Cuntz-Nica-Pimsner algebra N O(X) corresponds to N O(I, X) for (1.5) IF := {a ∈ JF hXn, JF Xni ⊆ JF for all n ⊥ F } with JF := (\i∈F ker φi)⊥. The Gauge-Invariant-Uniqueness-Theorem [8] asserts that N O(X) is the smallest Nica- Pimsner algebra that attains a faithful copy of A. In particular we define (1.6) N O(F, A, X) := N T (X)/hQi i ∈ F i and N O(A, X) := N T (X)/hQ1, . . . , QNi. It follows that N O(A, X) is N O(X) when every Xi is injective. For fixed β > 0 the F -parts of the Wold decomposition of an equilibrium state relate to the F -projections (1.7) Qn F := Xℓ(µ)=n t(xµ)QF t(xµ)∗, so that Qn {1,...,N } ≡ pn : FX → Xn. For every F ⊆ {1, . . . , N } we define (1.8) EF β (N T (X)) := {ϕ ∈ Eβ(N T (X)) Xn∈F ϕ(Qn F ) = 1 and ϕ(Qi) = 0 for all i /∈ F }, 4 E.T.A. KAKARIADIS with the understanding that for F = {1, . . . , N } we write (1.9) E{1,...,N } β (N T (X)) ≡ Efin β (N T (X)) := {ϕ ∈ Eβ(N T (X)) Xn∈ZN + ϕ(pn) = 1}, and for F = ∅ we write (1.10) E∅ β(N T (X)) ≡ E∞ β (N T (X)) := {ϕ ∈ Eβ(N T (X)) ϕ(Qi) = 0 for all i = 1, . . . , N }. We also consider the gauge invariant equilibrium states and we write (1.11) G-EF β (N T (X)) := {ϕ ∈ EF β (N T (X)) ϕ = ϕE}, where E : N T (X) → N T (X)γ is the conditional expectation of the gauge action. In Propo- sition 3.4 we show that any equilibrium state defines a gauge-invariant equilibrium state. In particular we show that the finite-type equilibrium states have to be gauge-invariant. Equivariance of the Qn F allows to define EF β (N O(I, X)) in a similar manner. Due to the KMS condition the infinite-type states are exactly the equilibrium states of N O(A, X) and the finite-type equilibrium states correspond to unique extensions of states on K(FX). The EF β -states correspond to states of the ideal hQF i that annihilate hQi i /∈ F i. Our first main result is that the F -parts are the building blocks of the Eβ-simplex. Theorem A. (Theorem 3.5) Let X be a product system of finite rank over A and let β > 0. Then every ϕ ∈ Eβ(N T (X)) admits a unique decomposition in a convex combination of ϕF ∈ EF β (N T (X)). Moreover the F -part ϕF is non-trivial if and only if ϕ(QF ) 6= 0. The same decomposition passes to the gauge-invariant equilibrium states and also to relative Cuntz-Nica-Pimsner algebras. Our next step is to parametrize every G-EF β (N T (X)) by specific subsets of tracial states of A. For every ∅ 6= F ⊆ {1, . . . , N } and τ ∈ T(A) we define (1.12) cF τ,β :=X{e−µβτ (hxµ, xµi) ℓ(µ) ∈ F }. Then the associated F -set of tracial states of A is given by (1.13) TF β (A) := {τ ∈ T(A) cF τ,β < ∞ and eβτ (a) = hxi,j, axi,ji for all i /∈ F }. diXj=1 In particular for F = {1, . . . , N } we write (1.14) β (A) := {τ ∈ T(A) c{1,...,N } Tfin τ,β = Xℓ(µ)∈ZN + e−µβτ (hxµ, xµi) < ∞}. The case of F = ∅ is captured in the averaging traces, namely (1.15) AVTβ(A) := {τ ∈ T(A) eβτ (a) = τ (hxi,j, axi,ji) for all i = 1, . . . , N }. diXj=1 Nevertheless these parts contain more states than what we want and we need to restrict further to traces that annihilate the kernel (1.16) IF c := ker{A → N O(F c, A, X)} = {a ∈ A lim k ϕk·1F c (a) = 0}. When Xi is injective for every i ∈ F c then IF c = (0). Passing further to states that annihilate IF we obtain the full E-structure for N O(I, X) at inverse temperature β > 0. By setting IF = (0) for every F gives the E-structure of N T (X). Setting IF = IF for every F we obtain the E-structure for the important quotient N O(X). EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 5 Theorem B. (Theorem 4.4 and Theorem 5.2) Let X be a product system of finite rank over A and β > 0. Let {IF ∅ 6= F ⊆ {1, . . . , N }} be a lattice of ⊥-invariant ideals of A. Then: (1) The infinite-type simplex G-E∞ β (N O(I, X)) is weak-homeomorphic onto {τ ∈ AVTβ(A) τ I{1,...,N} = 0}. (2) For every F 6= ∅ there is a continuous bijection from G-EF β (N O(I, X)) onto {τ ∈ TF β (A) τ IF c +IF = 0}. (3) These parametrizations respect convex combinations and thus the extreme points of the simplices. If TF β (A) or G-EF β (N O(I, X)) is weak-closed then the map in item (3) is a weak- homeomorphism. The proof is constructive and provides explicit formulas for these parametri- zations. To this end we use an induced product system Z = {Zn}n∈F over BF c in the Fock space representation that is supported on F and is given by BF c = span{t(Xk)t(Xw)∗ k, w ⊥ F } and Zn = t(Xn)BF c (1.17) First we use a direct limit method to lift a τ ∈ TF an equilibrium state on N T (X) from hQF i. Finally we use the ⊥-invariance of the IF to ensure this two-step process factorizes through N O(I, X). Gauge-invariance is used only to β (A) to a KMS-state eτ on BF c. Then we use the finite-type statistical approximation on the GNS-representation ofeτ to construct move from τ to eτ . Nevertheless with that given we can apply the same method to extend the parametrization to the entire EF Theorem C. (Theorem 6.1) Let X be a product system of finite rank over A and β > 0. Then ϕ ∈ E∞ β (N T (X)) if and only if β (N T (X)). for n ∈ F. ϕ(f ) = e−β ϕ(t(xi,j)∗f t(xi,j)) for all f ∈ N T (X), i = 1, . . . , N. diXj=1 For ∅ 6= F ⊆ {1, . . . , N } the parametrization of Theorem 4.4 lifts to a parametrization where {eτ ∈ Eβ(BF c) eτ π ∈ TF β (A),eτ I′ I′ F c := ker{BF c → N O(F c, A, X)}. F c = 0} → EF β (N T (X)) Likewise the parametrization of the relative Cuntz-Nica-Pimsner algebras from Theorem 5.2 lifts to where {eτ ∈ Eβ(BF c) eτ π ∈ TF β (A),eτ I′ I ′ F := span{t(Xm)π(IF )t(Xw)∗ m, w ⊥ F }. F c +I ′ F = 0} → EF β (N O(I, X)) In all results we consider the rotations to have the same (constant) rate. In Remark 6.2 we show that our methods extend to cover different weights as well. When all weights are non-zero then this is just a scaling argument. When some weights are zero then we just need to apply the general results to the Toeplitz-Nica-Pimsner algebra on the remaining fibers1. Next we turn our focus to the inverse temperatures β ≥ 0 for which equilibrium states exist. The existence of a unit decomposition entails a notion of entropy; the interested reader may consult the monograph of Neshveyev-Størmer [29] for a detailed discussion on the subject. Such an idea is developed by Pinzari-Watatani-Yonetani [31] for Cuntz-Pimsner algebras of a single C-correspondence. Assuming the existence of both a left and a right unit decomposition they identify the critical minimal and maximal inverse temperatures of Cuntz-Pimsner algebras. This is achieved by considering a Perron-Frobenius type theory 1 It must be noted that this process is not class-preserving. Nevertheless, Christensen [9] remarkably tackles the problem by remaining within the same class of higher rank graphs. (1.18) hτ X := lim sup k and for F ⊆ {1, . . . , N } we define the F -tracial entropy (1.19) hτ,F X := lim sup k 1 k τ (hxµ, xµi)(cid:21), τ (hxµ, xµi)(cid:21). 1 k log(cid:20) Xµ=k log(cid:20) Xµ=k,ℓ(µ)∈F log k Xµ=k 1 k 6 E.T.A. KAKARIADIS for the induced completely positive maps which is applied to Cuntz-Krieger algebras. Their notion of entropy is then compared to the Brown-Voiculescu topological entropy [34, 6] of the arising subshifts, which in turn finds its roots in the non-commutative entropy of Connes-Størmer [10]. Recently it has been discovered by an Huef-Laca-Raeburn-Sims [21] that not just the entropy of the adjacency matrix but also the entropies of the sink subgraphs are necessary to identify the phase transitions for the Toeplitz-Cuntz-Krieger algebras, and therefore of the Cuntz-Krieger algebras by descending appropriately; see also [23]. The key idea is that the convergence of the statistical approximation over a state depends on the irreducible components it visits forwards. This route also removes the assumption of a right unit decomposition from [31]. In order to define the multivariable analogue in our context we need to consider all possible β -parts, and then discover their relations. For every F -statistical approximations for the EF τ ∈ T(A) we define the tracial entropy We define the entropy of a unit decomposition x = {xi,j j = 1, . . . , di, i = 1, . . . , N } by (1.20) hx X := lim sup k hxµ, xµikA, and then the strong entropy of X by (1.21) X := inf{hx hs X x is a unit decomposition of X}. Likewise for F ⊆ {1, . . . , N } we define the F -entropy of a unit decomposition x by (1.22) hx,F X := lim sup k 1 k log k Xµ=k,ℓ(µ)∈F hxµ, xµikA, and the F -strong entropy of X by (1.23) hs,F X := inf{hx,F X x is a unit decomposition of X}. Moreover we define the entropy of X by (1.24) hX := inf{β > 0 Eβ(N T (X)) 6= ∅}. Due to our parametrization we are able to deduce hX from the tracial entropies. This is quite pleasing as for the first time we are able to recognize the critical inverse temperatures for ZN + -product systems. Theorem D. (Proposition 7.3, Proposition 7.4, Theorem 7.5 and Proposition 7.6) Let X be a product system of finite rank over A. (1) If τ ∈ T(A) then X ≤ hs hτ X i = 1, . . . , N } ≤ max{log di i = 1, . . . , N }. X = max{hs,i (2) If τ ∈ TF β (A) for some F ⊆ {1, . . . , N } and β > 0 then hτ,F X ≤ hτ X ≤ β. (3) The entropy of X is the infimum of the tracial entropies modulo negative values, i.e., hX = max{0, inf{hτ X τ ∈ T(A)}}. EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 7 (4) If in addition hX > 0 then there exists a τ ∈ T(A) such that hX = hτ . (5) If β > hs,F Eβ(N T (X)) = Efin β (N T (X)) = ∅ for all C ( F . X then EC β (N T (X)) ≃ T(A). In particular, if β > hs X then The strong entropy can be the greatest lower bound for the non-finite parts depending on the structure of the product system. For example hs X = log λ for the Perron-Frobenius eigenvalue of a higher-rank graph [19]. Proposition 7.6 also yields existence of KMS∞-states as limits of finite-type states. Thus another phase transition may occur at β = ∞. Theorem E. (Theorem 7.7) Let X be a product system of finite rank over A. Then there exists an affine weak-homeomoprhism Ψ between the states τ ∈ S(A) and the ground states of N T (X) such that Ψτ (π(a)) = τ (a) for all a ∈ A and Ψτ (t(ξn)t(ηm)∗) = 0 when n + m 6= 0. The restriction of Ψ to the tracial states T(A) induces a weak-homeomorphism onto the KMS∞-states of N T (X). If {IF ∅ 6= F ⊆ {1, . . . , N }} is a lattice of ⊥-invariant ideals of A then the corresponding weak-homeomorphisms for N O(I, X) arise by restricting on states that annihilate the ideal I{1,...,N }. We close with some examples to show how our theory covers some known results. Apart + -dynamics of [22] and irreducible higher rank graphs of [19] we apply the entropy from the ZN theory to product systems of multivariable factorial languages from [13]. 1.3. Organization of paper. In Section 2 we fix notation and set up the terminology for Nica-Pimsner algebras of finite rank product systems. In Section 3 we provide the convex decomposition of an equilibrium state on N T (X) in its F -parts. Then we proceed in Section 4 to the parametrization of the F -simplices. In Section 5 we provide the full decomposi- tion/parametrization theorem for Nica-Pimsner algebras by using factorization through the canonical quotient map. In Section 7 we make the connection with the entropies: we show how they affect the E-structure and we give the parametrization of the limit states. The examples are presented in Section 8. 2.1. Notation. The free generators of ZN for N < ∞ will be denoted by 1, . . . , N. We write 2. Preliminaries for the length of n ∈ ZN write + . We fix 1 := (1, . . . , 1). For ∅ 6= F ⊆ {1, . . . , N } and n ∈ ZN + we n ≡ X{nii i ∈ {1, . . . , N }} := ni NXi=1 nF :=Xi∈F ni · i. In this sense we write 1F =P{i i ∈ F }. We consider the usual lattice structure on ZN We denote the support of n by + . supp n := {i ∈ {1, . . . , N } ni 6= 0} and we write n ⊥ m if and only if Thus n ⊥ F means that supp nT F = ∅. For simplicity we write n ∈ F if and only if supp n ⊆ F. supp n\ supp m = ∅. 8 E.T.A. KAKARIADIS We will be making use of the alternating sums, i.e., for F ⊆ {1, . . . , N } and {f1, . . . , fN } commuting elements we have Yi∈F (1 − fi) =X{(−1)CYi∈C fi C ⊆ F } = 0. We will consider multivariable words on different sets of symbols. Let {d1, . . . , dN } be a set of finite positive integers. We write µ = (µ1, . . . , µN ) ∈ Fd1 + × · · · × FdN + for the tuple of N words where each µi is a word over the symbol set {1, . . . , di}. denotes the length of each word then we define the multi-length ℓ(µ) and the length µ by If µi ℓ(µ) := (µ1, . . . , µN ) and µ := ℓ(µ) = µi. NXi=1 We will be considering limits over ZN family of finite sets in ZN + with the understanding of convergence over the directed + . Therefore for a non-negative sequence (an)n∈ZN we have + lim n∈ZN + an = lim n=k→∞ an = lim n≤k·1,k→∞ an. 2.2. C-correspondences. The reader should be well acquainted with the general theory of Hilbert modules and C-correspondences. For example, one may consult [24] and [27] which we follow for terminology. Here we just wish to fix notation. A C-correspondence X over A is a right Hilbert module over A with a left action given by a ∗-homomorphism φX : A → LX. We write LX and KX for the adjointable operators and the compact operators of X, respectively. We will write aξ for φX (a)ξ when it is clear from the context which left action we use. The "rank one compact operators" ζ 7→ ξhη, ζi are denoted by θX ξ,η. For two C-correspondences X, Y over the same A we write X ⊗A Y for the stabilized tensor product over A. Moreover we say that X is unitarily equivalent to Y if there is a surjective adjointable U ∈ L(X, Y ). A representation (ρ, v) of a C-correspondence is a left module map that preserves the inner product. Then (ρ, v) is automatically a bimodule map. Moreover there exists a ∗- homomorphism ψ on KX such that ψ(θX ξ,η) = v(ξ)v(η)∗. If ρ is injective then so is ψ. 2.3. Product systems. Fix a set {Xi one for each generator of ZN correspondences over A such that + . A product system X is a family {Xn n ∈ ZN i ∈ {1, . . . , N }} of C-correspondences over A, + } of C- X0 = A and X⊗n1 i1 ⊗A · · · ⊗A X⊗nk ik ≃ Xn whenever n =X njij and n1 6= 0. We require n1 6= 0 so that these equivalences do not force non-degeneracy of the fibers. Consequently X comes with a family of product rules in the form of unitary equivalences un,m : Xn ⊗A Xm → Xn+m. We will suppress the use of the un,m as much as possible by writing ξnξm ∈ Xn+m for the element un,m(ξn⊗ξm). Along with the system we have some canonical operations that respect these equivalences. To this end we define the maps : LXn → LXn+m such that in+m and thus in+m It is clear that in+m+r work [15], a product system is called compactly aligned if it has the property: (φn(a)) = φn+m(a). Following Fowler's (S) = un,m(S ⊗ idXm)u∗ n+m in+m n = in+m+r n in+m n n,m. n n in∨m n (S)in∨m m (T ) ∈ KXn∨m whenever S ∈ KXn, T ∈ KXm. If a compactly aligned product system satisfies in+i is called strong compactly aligned [13]. n (KXn) ⊆ KXn+i whenever n ⊥ i then it EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 9 2.4. Finite rank. A product system will be said to have finite rank {d1, . . . , dN } if for every i = 1, . . . , N there exists a family of vectors {xi,j j = 1, . . . , di} such that 1Xi = θXi xi,j,xi,j . diXj=1 We will say that a family x = {xi,j j = 1, . . . , di, i = 1, . . . , N } is a unit decomposition for the C-correspondence. If µi = j1 · · · jr is a word on the symbols {1, . . . , di} then we write xi,µi = xi,j1 ⊗ · · · ⊗ xi,jr . We may extend this notation to the entire ZN + . Let µ = (µ1, . . . , µN ) be a multivariable word such that each µi is a word on the symbols {1, . . . , di}. If µi1, . . . , µik 6= ∅ for i1 < · · · < ik then we write xµ = xi1,µi1 · · · xik,µik ∈ Xℓ(µ). Then the family {xµ ℓ(µ) = n} is a unit decomposition for Xn. Notice here that a re- ordering of the xi,µi in xµ gives another unit decomposition (of the same cardinality). We have that X is of finite rank if and only if KXn = LXn for all n ∈ ZN + , and thus a product system of finite rank is automatically strong compactly aligned. 2.5. Representations. A Nica-covariant representation (ρ, v) of a product system X = {Xn n ∈ ZN + } consists of a family of representations (ρ, vn) of Xn that satisfy the product rule: vn+m(ξnξm) = vn(ξn)vm(ξm) for all ξn ∈ Xn, ξm ∈ Xm, and the Nica-covariance: ψn(S)ψm(T ) = ψn∨m(in∨m n (S)in∨m m (T )) whenever S ∈ KXn, T ∈ KXm. We write ρ × v for the induced representation of a Nica-covariant pair (ρ, v). Henceforth we will suppress the use of the indices and write v instead of vn. The Toeplitz-Nica-Pimsner algebra N T (X) is the universal C-algebra generated by A and X with respect to the representations of X. The Fock space provides an essential example of an injective Nica-covariant representation for N T (X). In short, let F(X) =P ⊕{Xm m ∈ ZN + }. For a ∈ A and ξn ∈ Xn define π(a)ηm = φm(a)ηm and t(ξn)ηm = ξnηm for all ηm ∈ Xm. Then (π, t) is Nica-covariant and it is called the Fock representation of X [15]. By taking the compression at the (0, 0)-entry we see that π, and thus t, is injective. Given a Nica-covariant representation (ρ, v) and m, m′ ∈ ZN + we define the cores of the representation (ρ, v) by B[m,m+m′] := span{ψn(kn) kn ∈ KXn, m ≤ n ≤ m + m′}. These ∗-algebras are closed in C∗(ρ, v), e.g., [8, Lemma 36]. It follows from the work of Fowler [15, Proposition 5.10] that if (ρ, v) is a Nica-covariant representation of a compactly aligned product system X then (2.1) v(Xm)∗v(Xn) ⊆ v(Xn−n∧m)v(Xm−n∧m)∗. Therefore the cores are ⊥-stable in the sense that tn(Xn)∗ · B[m,m+m′] · tn(Xn) ⊆ B[m,m+m′] for all n ⊥ m + m′. A Nica-covariant representation (ρ, v) admits a gauge action if there is a point-norm contin- uous family of ∗-automorphisms {γz}z∈TN such that γz(ρ(a)) = ρ(a) for all a ∈ A and γz(v(ξn)) = zn v(ξn) for all ξn ∈ Xn. In this case B[0,∞] = ∪nB[0,n] is the fixed point algebra of C∗(ρ, v). By universality N T (X) (as well as any of the covariant C-algebras we will define below) admits a gauge action. 10 E.T.A. KAKARIADIS The appropriate Cuntz-analogue of a C-algebra should attain an isometric copy of A and be co-universal with respect to a Gauge-Invariant-Uniqueness-Theorem. Its existence has been established for compactly aligned product systems in [8]. The form of the representa- tions makes use of the augmented Fock space construction of [33]. To get the Cuntz-Nica- Pimsner algebra for strong compactly aligned we require a little bit less work [13]. For a finite ∅ 6= F ⊆ {1, . . . , N } we form the ideal n (KXn) n ≤ 1}), JF := (\i∈F ker φi)⊥ ∩ (\{φ−1 n (KXn) n ≤ 1} =\{φ−1 i \{φ−1 with the understanding that φ0 = idA. In particular when X is strong compactly aligned, we have that (KXi) i ∈ {1, . . . , N }}. Furthermore we define the ideal IF := {a ∈ JF hXn, aXni ⊆ JF for all n ⊥ F }. Hence IF is the biggest ideal in JF that remains invariant under the "action" of F ⊥. A Nica- covariant representation (ρ, v) of X is called Cuntz-Nica-Pimsner (or a CNP-representation) if it satisfies X{(−1)nψn(φn(a)) n ≤ 1F } = 0 for all a ∈ IF , where ψ0(φ0(a)) = ρ(a). The Cuntz-Nica-Pimsner algebra N O(X) is the universal C- algebra with respect to the CNP-representations. One of the main results of [13] is that this is the ∗-algebraic description of Cuntz-Nica-Pimsner algebra, i.e., this N O(X) coincides with the one in [8]. From now on we pass to product systems of finite rank. Notice that in this case φn(A) ⊆ KXn for all n ∈ ZN + . Definition 2.1. Let X be a product system of finite rank over A and fix ∅ 6= F ⊆ {1, . . . , N }. A representation (ρ, v) will be called F -covariant if ρ(a) = ψi(φi(a)) for all a ∈ A and i ∈ F. The F -Cuntz-Nica-Pimsner algebra N O(F, A, X) is the universal C-algebra with respect to the F -covariant representations. It follows that N O(F, A, X) = N T (X)/hπ(a)Qi a ∈ A, i ∈ F i. To this end we will write IF := ker {A → N T (X)/hπ(a)Qi a ∈ A, i ∈ F i} . Following the arguments of [13, Corollary 5.1] it can be shown that (ρ, v) is F -covariant if and only if ρ(a) = ψn(φn(a)) for all a ∈ A and supp n ⊆ F. We write N O(A, X) for the algebra related to the {1, . . . , N }-covariant pairs. Sometimes N O(A, X) appears as N O(X)cov in the literature, but here we see it as an example in the bigger class of relative Nica-Pimsner algebras. The latter is not always the Cuntz-analogue for product systems as it may not attain an isometric copy of A. By the Gauge-Invariant- Uniqueness-Theorem for regular product systems of [33] we get that N O(A, X) = N O(X) if every φi is injective. 2.6. Kubo-Martin-Schwinger states. Let σ : R → Aut(A) be an action on a C-algebra A. Then there exists a norm-dense σ-invariant ∗-subalgebra Aan of A such that for every f ∈ Aan the function R ∋ r 7→ σr(f ) ∈ A is analytically continued to an entire function C ∋ z 7→ σz(f ) ∈ A [4, Proposition 2.5.22]. If β > 0, then a state ϕ of A is called a (σ, β)-KMS state (or equilibrium state at β) if it satisfies the KMS-condition: ϕ(f g) = ϕ(gσiβ (f )) for all f, g in a norm-dense σ-invariant ∗-subalgebra of Aan. If β = 0 or if the action is trivial then a KMS-state is a tracial state on A. The KMS-condition follows as an equivalent for the existence of particular continuous functions [5, Proposition 5.3.7]. More precisely, a state ϕ is an equilibrium state at β > 0 if and only if for any pair EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 11 f, g ∈ A there exists a complex function Ff,g that is analytic on D = {z ∈ C 0 < Im(z) < β} and continuous (hence bounded) on D such that Ff,g(r) = ϕ(f σr(g)) and Ff,g(r + iβ) = ϕ(σr(g)f ) for all r ∈ R. A state ϕ of A is called a KMS∞-state if it is the weak-limit of (σ, β)-KMS states as β ↑ ∞. A state ϕ of a C-algebra A is called a ground state if the function z 7→ ϕ(f σz(g)) is bounded on {z ∈ C Imz > 0} for all f, g inside a dense analytic subset of A. The distinction between ground states and KMS∞-states is not apparent in [5] and is coined in [26]. Remark 2.2. Henceforth we will focus on rotational actions in direct relation to Gibbs states. Recall that in the classical case the R-action is induced by r 7→ eir(H−κN ) where H is the selfadjoint Hamiltonian, N is the number operator and κ is the chemical potential. When H is the Quantum Harmonic Oscillator then H − κN = hω/2 + (hω − κ)N where hω/2 is the zero point energy. The Spectral Theorem then materializes the R-action as induced by the rotational number unitaries r 7→ ueirs for s = hω − κ on the related Fock space, for example see [23, Proposition 2.1]. The same arguments apply for the multivariable version of the Quantum Harmonic Oscillator, i.e., the tensor product of number operators fiberwise. We will be scaling hω − κ = 1; substituting β with (hω − κ)β unravels the general case. Let {γz}z∈Tn be the gauge action on N T (X) and define σ : R → Aut(N T (X)) : r 7→ γ(exp(ir),...,exp(ir)). The monomials of the form t(ξn)t(ηm)∗ span a dense ∗-subalgebra of analytic elements of N T (X) since the function R → N T (X) : r 7→ σr(t(ξn)t(ηm)∗) = ein−mrt(ξn)t(ηm)∗ is analytically extended to the entire function C → N T (X) : z 7→ ein−mzt(ξn)t(ηm)∗. For β ∈ R the (σ, β)-KMS condition is then equivalent to having (2.2) ϕ(t(ξn)t(ηm)∗ · t(ξk)t(ηw)∗) = e−n−mβϕ(t(ξk)t(ηw)∗ · t(ξn)t(ηm)∗) for all n, m, k, w ∈ ZN + . Definition 2.3. Let X be a compactly aligned product system. For a fixed β > 0 we write Eβ(N T (X)) for the (σ, β)-KMS states on N T (X) with respect to the action σ : R → Aut(N T (X)) : r 7→ γ(exp(ir),...,exp(ir)). If E : N T (X) → N T (X)γ is the conditional expectation of the gauge action, then we write G-Eβ(N T (X)) := {ϕ ∈ Eβ(N T (X)) ϕ = ϕE} for the sub-simplex of the gauge-invariant equilibrium states. We note that it suffices to consider the unital case. If there is at least one Xi0 that is not unital then consider the unitization A1 = A + C of A and extend the operations φn(1)ξn = ξn = ξn · 1. Note here that A1 = A ⊕ C if A is unital but φn(1A) 6= 1Xn. We will write X 1 = {X 1 + n}n∈ZN . Proposition 2.4. Let X be a compactly aligned product system over A. Then ϕ is a (σ, β)- KMS state for N T (X 1) if and only if it restricts to a (σ, β)-KMS state on N T (X). Proof. As σ is the same action on both N T (X 1) and N T (X) the proof follows by noting that N T (X 1) is the unitization of N T (X). See also [23, Proposition 3.2] for more details from the Z+-case. 12 E.T.A. KAKARIADIS Remark 2.5. Due to Proposition 2.4 we assume that the product system is unital for our proofs and exposition. However the results cover also non-unital product systems. We will use the same symbol for the extension of a state from N T (X) to N T (X 1) from Proposition 2.4. We have an easier version of the KMS-condition that will be handy for our computations. Proposition 2.6. Let X be a compactly aligned product system over A and let β > 0. Then a positive functional ϕ of N T (X) satisfies the KMS-condition (2.2) if and only if ϕ(t(ξn) · t(ξk)t(ηw)∗) = e−nβϕ(t(ξk)t(ηw)∗ · t(ξn)) for all n, k, w ∈ ZN + . (2.3) If in addition ϕ = ϕE, then ϕ satisfies the KMS-condition (2.2) if and only if (2.4) ϕ(t(ξn)t(ηm)∗) = δn,me−nβϕ(t(ηm)∗t(ξn)) for all ξn ∈ Xn, ηm ∈ Xm. Consequently, two gauge-invariant (σ, β)-KMS states coincide if and only if they agree on π(A). Proof. For the first part, equation (2.2) implies equation (2.3) for ηm = 1A. For the converse, we may use continuity to deduce that ϕ(t(ξn)f ) = e−nβϕ(f t(ξn)) for all f ∈ t(Xk)t(Xw)∗ and for all n, k, w ∈ ZN + . Since ϕ is positive, by applying adjoints in equation (2.3) we get ϕ(t(ξn)∗f ) = enβϕ(f t(ξn)∗) for all f ∈ t(Xk)t(Xw)∗ and for all n, k, w ∈ ZN + . Now for m ∈ ZN + we have that t(ηm)∗t(ξk)t(ξw)∗ ∈ t(Xm)∗t(Xk)t(Xw)∗ ⊆ t(Xk−k∧m)t(Xm−k∧m+w)∗. At the same time we have that t(ξk)t(ηw)∗t(ξn) ∈ t(Xk)t(Xw)∗t(Xn) ⊆ t(Xk+n−n∧w)t(Xw−n∧w)∗. Therefore we have that ϕ(t(ξn)t(ηm)∗ · t(ξk)t(ηw)∗) = ϕ(t(ξn) · t(ηm)∗t(ξk)t(ηw)∗) = e−nβϕ(t(ηm)∗ · t(ξk)t(ηw)∗t(ξn)) = e−nβemβϕ(t(ξk)t(ηw)∗t(ξn)t(ηm)∗) which gives equation (2.2). Now suppose that in addition ϕ = ϕE. If ϕ satisfies equation (2.2) then it also implies equation (2.4). Conversely we will show that if ϕ satisfies equation (2.4) then it satisfies equation (2.3), and thus equation (2.2). To this end we consider two cases. Suppose first that n + k 6= w. Since ϕ = ϕE we thus have ϕ(t(ξn)t(ξk)t(ηw)∗) = ϕE(t(ξn)t(ξk)t(ηw)∗) = 0. At the same time we have that t(ξk)t(ηw)∗t(ξn) ∈ t(Xk+n−n∧w)t(Xw−n∧w)∗). Since k + n − n ∧ w 6= w − n ∧ w we get that e−nβϕ(t(ξk)t(ηw)∗t(ξn)) = e−nβϕE(t(ξk)t(ηw)∗t(ξn)) = 0, so that equation (2.3) holds trivially. Now suppose that n + k = w so that w ≥ n. Then we have that t(ηw)∗t(ξn) ∈ t(Xw−n)∗ = t(Xk)∗ and thus ϕ(t(ξn) · t(ξk)t(ηw)∗) = ϕ(t(ξn)t(ξk) · t(ηw)∗) = e−n+kβϕ(t(ηw)∗t(ξn) · t(ξk)) = e−n+kβekβϕ(t(ξk)t(ηw)∗t(ξn)) = e−nβϕ(t(ξk)t(ηw)∗ · t(ξn)), and the proof is complete. EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 13 3. Wold decomposition Suppose that X is a unital product system with a unit decomposition x = {xi,j j = 1, . . . , di, i = 1, . . . , N }. For every i ∈ {1, . . . , N } we let the projections For F ⊆ {1, . . . , N } we define the projections Pi := diXj=1 PF :=Yi∈F Pi t(xi,j)t(xi,j)∗ ∈ N T (X). and QF :=Yi∈F (1 − Pi). where 1 ≡ 1F X . It is clear that all these projections are in N T (X). For every n ∈ ZN the projection pn : FX → Xn. It is straightforward to check that + let PF =X{pn n ≥ 1F } = Xℓ(µ)=1F t(xµ)t(xµ)∗, and therefore the PF and QF all commute. In particular they are in the center of the fixed point algebra N T (X)γ . We will also write Pk·i :=X{pn n ≥ k · i} = Xℓ(µ)=k·i t(xµ)t(xµ)∗. For every n ∈ ZN + we define the n-th "translate" of QF ≡ Q0 F by If in addition n ∈ F then we can directly verify that and therefore (3.1) Qn xµhxµ, ξwi = δn,wF ξw, t(xµ)QF t(xµ)∗. Qn F := Xℓ(µ)=n F ξw = δn,wF Xℓ(µ)=n F t(ξm) =(t(ξm)Qn−mF 0 F Qn if n ≥ mF , otherwise. Lemma 3.1. With the aforementioned notation we have that Qn F · Qm C = δn,mF · Qm C for all n ∈ F, m ∈ C with ∅ 6= F ⊆ C. Proof. Let ℓ(ν) = n ∈ F and ℓ(µ) = m ∈ C. If n − n ∧ m 6= 0 then it has an intersection with F , and thus with C, and so equation (3.1) yields QF t(xν)∗t(xµ)QC ∈ QF t(Xm−n∧m) · t(Xn−n∧m)∗QC = (0). If n = n ∧ m, i.e., if n ≤ m, but mF 6= n then m − n has an intersection with F and so again equation (3.1) yields QF t(xν)∗t(xµ)QC ∈ QF t(Xm−n)QC = (0). If n = mF then t(xν)∗t(xµ) ∈ t(Xm−mF ) and so equation (3.1) gives again QF t(xν)∗t(xµ) = t(xν )∗t(xµ)QF . 14 E.T.A. KAKARIADIS Therefore we get Qn F · Qm t(xν )QF t(xν)∗t(xµ)QC t(xµ)∗ t(xν)t(xν )∗t(xµ)QF QCt(xµ)∗ C = Xℓ(ν)=n Xℓ(µ)=m = δn,mF Xℓ(ν)=n Xℓ(µ)=m = δn,mF Xℓ(µ′′)=mF c Xℓ(µ′)=mF Xℓ(ν)=mF = δn,mF Xℓ(µ)=m t(xµ)QCt(xµ)∗ = δn,mF · Qm C . [t(xν )t(xν )∗t(xµ′)]t(xµ′′)QF QCt(xµ′′)∗t(xµ′)∗ Definition 3.2. Let X be a product system of finite rank over A and fix β > 0. For every F ⊆ {1, . . . , N } we define EF β (N T (X)) := {ϕ ∈ Eβ(N T (X)) Xn∈F ϕ(Qn F ) = 1 and ϕ(Qi) = 0 for all i /∈ F }, with the understanding that for F = {1, . . . , N } we write E{1,...,N } β (N T (X)) ≡ Efin β (N T (X)) := {ϕ ∈ Eβ(N T (X)) Xn∈ZN + ϕ(pn) = 1}, and for F = ∅ we write E∅ β(N T (X)) ≡ E∞ β (N T (X)) := {ϕ ∈ Eβ(N T (X)) ϕ(Qi) = 0 for all i = 1, . . . , N }. Likewise we define the F -parts for the gauge-invariant equilibrium states by G-EF β (N T (X)) := {ϕ ∈ EF β (N T (X)) ϕ = ϕE}. The F -equilibria are connected with specific ideals in N T (X). We provide this description in the next proposition. Proposition 3.3. Let X be a product system of finite rank over A and fix β > 0. Then a state ϕ is in EF β (N T (X)) if and only if it restricts to a state on hQF i that satisfies the KMS-condition and annihilates hQi i /∈ F i. Proof. First we claim that {Pn≤k·1F Indeed by [13, Proposition 4.6] and equation (3.1) we have that Qn F k ∈ Z+} defines an approximate unit for hQF i. hQF i = span{t(ξn)QF t(ηm)∗ n, m ∈ ZN + } = span{QnF = C∗(Qn F t(ξn)t(ηm)∗QmF n, m ∈ ZN F n, m ∈ F, f ∈ N T (X)). F f Qm F + } Then, by using the KMS-condition and orthogonality of Qn identified by F , we have that ϕ is uniquely ϕ(f ) = lim ϕ(Qn F f Qm F ) = lim ϕ(Qn F f Qn F ). k Xn,m≤k·1F k Xn≤k·1F Secondly we have to show that if ϕ(Qi) = 0 for all i /∈ F then ϕ annihilates the ideal the Qi generate. By the Cauchy-Schwartz inequality we have that ϕ(Qig) = 0 for all g ∈ N T (X). In particular we can use the KMS-condition (as all elements below are analytical) to obtain ϕ(t(ξn)t(ηm)∗ · Qi · t(ξk)t(ηw)∗) = e−n−mβϕ(Qi · t(ξk)t(ηw)∗t(ξn)t(ηm)∗) = 0. Therefore by continuity ϕ(f Qig) = 0 for all f, g ∈ N T (X). Moreover the F -type equilibrium states are F -invariant in the following sense. EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 15 Proposition 3.4. Let X be a product system of finite rank over A and fix β > 0. For F ⊆ {1, . . . , N } we have that EF β (N T (X)) 6= ∅. Moreover for every ϕ ∈ G-EF β (N T (X)) 6= ∅ if and only if G-EF β (N T (X)) we have that In particular the equilibrium states of finite type are gauge-invariant. ϕ(t(ξn)t(ηm)∗) = δnF ,mF ϕ(t(ξn)t(ηm)∗). Proof. For the first assertion, let ϕ ∈ Eβ(N T (X)) and set ϕ′ = ϕE. Then ϕ′E = ϕE2 = ϕ′, so that ϕ′ is gauge-invariant. Moreover we have that ϕ′ satisfies equation (2.4) since ϕ′(t(ξn)t(ηm)∗) = δn,mϕ(t(ξn)t(ηm)∗) = δn,me−nβϕ(t(ηm)∗t(ξn)) = δn,me−nβϕ′(t(ηm)∗t(ξn)). By Proposition 2.6 we thus get that ϕ′ is a gauge-invariant equilibrium state at β. Since the Qn F are gauge-invariant we get that ϕ′(Qn F ) = ϕE(Qn F ) = ϕ(Qn F ) for all n ∈ F. Therefore ϕE ∈ EF β (N T (X)) whenever ϕ ∈ EF β (N T (X)). For the second assertion we may use equation (3.1) and orthogonality of Qw F for w ∈ F from Lemma 3.1 to obtain Qw F t(ξn)t(ηm)∗Qw F =(Qw 0 F t(ξn)Qw−nF F Qw−mF F t(ηm)∗Qw F if w ≥ nF , mF , otherwise, = δnF ,mF Qw F t(ξn)t(ηm)∗Qw F . If ϕ ∈ EF β (N T (X)) then ϕ(f ) =Pw∈F ϕ(Qw ϕ(t(ξn)t(ηm)∗) = δnF ,mF lim ϕ(Qw F f Qw F t(ξn)t(ηm)∗Qw F ) for all f ∈ N T (X). Therefore F ) = δnF ,mF ϕ(t(ξn)t(ηm)∗). k Xw≤k·1F For the last assertion just notice that Qw {1,...,N } = pw. The main theorem of this section is the following convex decomposition. Theorem 3.5. Let X be a product system of finite rank over A and let β > 0. Then every ϕ in Eβ(N T (X)) (resp. in G-Eβ(N T (X))) admits a unique decomposition in a convex combination of ϕF in EF β (N T (X))). Moreover the F -part ϕF is non-trivial if and only if ϕ(QF ) 6= 0. β (N T (X)) (resp. in G-EF The proof will follow from a number of lemmas. The only place we require X to have finite rank is to ensure that the projections we use are in N T (X) (in fact in π(A)′). In what follows fix ϕ be a positive functional that satisfies the KMS-condition at β. For fixed ∅ 6= C ⊆ {1, . . . , N } we define the auxiliary positive functionals (3.2) ϕ(Qn C f Qn C) for all f ∈ N T (X). Qn C(cid:1) ≤ kf k, C ) = Xn,m∈C Since the Qn C are orthogonal projections by Lemma 3.1, we have that ϕ0,C(f ) :=Xn∈C Xfinite n∈C ϕ(Qn C f Qn C) ≤ kf k · ϕ(cid:0) Xfinite n∈C and so indeed ϕ0,C defines a positive functional. Due to the KMS-condition we also have that ϕ0,C (f ) =Xn∈C ϕ(Qn C f Qn C) =Xn∈C ϕ(Qn C f ) =Xn∈C ϕ(f Qn ϕ(Qn C f Qm C ). Lemma 3.6. With the aforementioned notation, we have that the functional ϕ0,C satisfies the KMS-condition and ϕ0,C (QC) = ϕ(QC ). If ϕ = ϕE then ϕ0,C = ϕ0,C E as well. 16 E.T.A. KAKARIADIS Proof. By Proposition 2.6 it suffices to check that it satisfies equation (2.3). The KMS- condition on ϕ and equation (3.1) yield ϕ0,C (t(ξm)t(ξk)t(ηw)∗) = e−mβXn∈C ϕ(t(ξk)t(ηw)∗Qn Ct(ξm)) Moreover we have that ) n ≥ mC, n ∈ C} C ) n ∈ C} C = e−mβX{ϕ(t(ξk)t(ηw)∗t(ξm)Qn−mC = e−mβX{ϕ(t(ξk)t(ηw)∗t(ξm)Qn ϕ0,C(QC ) =Xn∈C = e−mβϕ0,C (t(ξk)t(ηw)∗t(ξm)). C) = ϕ(QC). C QCQn ϕ(Qn The last claim is immediate and the proof is complete. Lemma 3.7. With the aforementioned notation we have that ϕ0,C (QD) = ϕ0,C\D(QD) for every C, D ⊆ {1, . . . , N }, with the understanding that ϕ0,∅(QD) = ϕ0,D(QD) = ϕ(QD). Proof. For n ∈ C we directly compute if n ⊥ D, otherwise, if n ⊥ D, otherwise, if n ⊥ D, otherwise, Qn CQD = Xℓ(µ)=n t(xµ)QC\DQD∩Ct(xµ)∗QD 0 0 =(Pℓ(µ)=n t(xµ)QC\DQD∩CQDt(xµ)∗QD =(Pℓ(µ)=n t(xµ)QC\DQDt(xµ)∗QD =(Pℓ(µ)=n t(xµ)QC\Dt(xµ)∗QD =(Qn ϕ0,C (QD) =Xn∈C ϕF (f ) := XC⊇F C QD) = Xn∈C\D if n ∈ C \ D, otherwise. C\DQD ϕ(Qn ϕ(Qn 0 0 Now for a fixed F 6= ∅ we define the functionals (3.3) (−1)C\F ϕ0,C (f ) for all f ∈ N T (X). Hence we get that C\DQD) = ϕ0,C\D(QD). Lemma 3.8. With the aforementioned notation and for F 6= ∅, the functional ϕF of N T (X) is positive, and ϕF = 0 if ϕ(QF ) = 0. If ϕ is gauge-invariant then so is ϕF . Moreover we have that ϕF ≤ ϕ and the state ϕF (1)−1ϕF is in EF Proof. It is clear that ϕF satisfies the KMS-condition, since so does every summand ϕ0,C . Likewise if in addition ϕ = ϕE then so it holds for ϕF . β (N T (X)). If ϕ(QF ) = 0 then ϕ(QC ) = 0 for every C ⊇ F . Thus for n ∈ C we get that ϕ(Qn C) = Xℓ(µ)=n e−nβϕ(QC t(xµ)∗t(xµ)QC) ≤ Xℓ(µ)=n e−nβϕ(QC ) = 0. Therefore for every 0 ≤ f ∈ N T (X) we obtain Cf Qn ϕ(Qn ϕ0,C(f ) =Xn∈C C) ≤ kf kXn∈C ϕ(Qn C) = 0. EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 17 As N T (X) is spanned by its positive elements, we would get that ϕF = 0. To show positivity, we may use the alternating sums for every n-level with n ≤ k · 1F and the alternating sums expansion for the projection to deduce that XC⊇F Xn≤k·1C (−1)C\F Qn where we write Rm F := Xℓ(µ)=m Yi /∈F (1 − (1 − P(k+1)·i)) =Yi /∈F C = Xm≤k·1F(cid:20)X{(−1)C\F Qm+w = Xm≤k·1F(cid:20) Xℓ(µ)=m t(xµ)(cid:20)Yi /∈F P(k+1)·i(cid:21)QF t(xµ)∗ = Xℓ(µ)=m C t(xµ)(cid:20)Yi /∈F F =X{(−1)C\F Qm+w P(k+1)·i(cid:21)QF t(xµ)∗ζn =(ζn 0 C Rm Indeed since supp m ⊆ F we have Xℓ(µ)=m t(xµ)(cid:20)Yi /∈F Alternatively, for a fixed m ≤ k · 1F , we will show that (3.4) w ≤ k · 1C\F , C ⊇ F }. P(k+1)·i w ≤ k · 1C\F , C ⊇ F }(cid:21) P(k+1)·i(cid:21)QF t(xµ)∗(cid:21) ≡ Xm≤k·1F (cid:20)Yi /∈F P(k+1)·i(cid:21)QF t(xµ)∗2. Rm F , if nF = m, nF c ≥ (k + 1) · 1F c, otherwise. On the other hand if 0 6= w ∈ C \F then Qm+w otherwise it is zero. Equivalently we have C ζn = ζn exactly when nC = mC +wC = m+w, Qm+w C ζn =(ζn 0 if nF = m, nC\F = w, otherwise. Therefore we have the following cases that verify equation (3.4): • Case 1, nF = m and nF c ≥ (k + 1) · 1F c. In this case we have that nC\F ≥ (k + 1) · 1C\F > w for every C ⊇ F and w ≤ k · 1C\F . Hence we have • Case 2, nF 6= m. In this case we directly verify that X{(−1)C\F Qm+w C X{(−1)C\F Qm+w C ζn w ≤ k · 1C\F , C ⊇ F } = Qm F ζn = ζn. ζn w ≤ k · 1C\F , C ⊇ F } = 0. • Case 3, nF = m and ni ≤ k for some i /∈ F . In this case set G := {i /∈ F ni ≤ k} 6= ∅. For every C ⊇ F with C \ F 6⊆ G and w ≤ k · 1C\F we have that wC\F 6= nC\F . On the other hand for every C ⊇ F with C \ F ⊆ G there exists a unique w ≤ k · 1C\F such that w = nC\F . Therefore we deduce X{(−1)C\F Qm+w C ζn w ≤ k · 1C\F , C ⊇ F } =X{(−1)C\F ζn C \ F ⊆ G} (−1)Cζn = 0. = XC⊆G 18 E.T.A. KAKARIADIS Now we return to the proof. As in Lemma 3.1 we have that {Rm F }m≤k·1F is a family of pairwise orthogonal projections. Thus the KMS-condition yields ϕ(Qn Cf ) = lim k XC⊇F (−1)C\F Xn≤k·1C ϕ(Qn C f ) F f ) = lim F ) = lim k Xn,m≤k·1F ϕ(Rn F f Rm F ). Therefore ϕF is a positive functional and moreover ϕF ≤ ϕ. In particular Lemma 3.1 yields ϕF (f ) = XC⊇F = lim ϕ(Rm (−1)C\F Xn∈C k Xm≤k·1F Xn∈F ϕF (Qn F Qm C ) ϕ(Qn ϕ(Rm F f Rm k Xm≤k·1F (−1)C\F Xn∈F Xm∈C (−1)C\F Xm′′∈C\F Xn∈F Xm′∈F (−1)C\F Xm′′∈C\F Xm′∈F (−1)C\F Xm∈C ϕ(Qm C ) = ϕF (1). F ) = XC⊇F = XC⊇F = XC⊇F = XC⊇F ϕ(Qn F Qm′+m′′ C ) ϕ(Qm′+m′′ ) C To finish the proof we have to show that ϕF (Qi) = 0 for every i /∈ F . To this end we directly compute ϕF (Qi) = XC⊇F ∪{i} = Xi /∈C⊇F = − Xi /∈C⊇F (−1)C\F ϕ0,C(Qi) + Xi /∈C⊇F (−1)C\F +1ϕ0,C∪{i}(Qi) + Xi /∈C⊇F (−1)C\F ϕ0,C (Qi) + Xi /∈C⊇F (−1)C\F ϕ0,C (Qi) (−1)C\F ϕ0,C (Qi) (−1)C\F ϕ0,C(Qi) = 0, where we used that ϕ0,C∪{i}(Qi) = ϕ0,C(Qi) when i /∈ C from Lemma 3.7. Lemma 3.9. With the aforementioned notation and for ∅ 6= F 6= {1, . . . , N }, we have that (ϕC )F = 0 whenever C 6= F with C ≥ F . Proof. We first show that (ϕC )0,D = 0 whenever C 6⊇ D. Indeed in this case there exists an i ∈ D \ C and so ϕC(QD) = 0. In particular we have ϕC(Qn Df ) = 0 for all n ∈ D and f ∈ N T (X). Therefore we get that (ϕC )0,D(f ) = Xn∈D ϕC (Qn Df ) = 0. We will consider two cases for C and F . First suppose that C 6⊇ F . If D ⊇ F then C 6⊇ D, and so by the above we have that (ϕC )F = XD⊇F (−1)D\F (ϕC )0,D = 0. Secondly suppose that C ) F . If C 6⊇ D then as before we have (ϕC )0,D = 0. If C ⊇ D then Lemma 3.1 gives (ϕ0,G)0,D = ϕ0,G for all G ⊇ C since (ϕ0,G)0,D(f ) = Xn∈D Xm∈G ϕ(Qm G Qn Df ) = Xn∈D Xm′∈D Xm′′∈G\D ϕ(Qm′+m′′ G f ) = ϕ0,G(f ). = Xm′∈D Xm′′∈G\D ϕ(Qm′+m′′ G Qn Df ) EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 19 Thus we have (ϕC )F = XD⊇F = XC⊇D⊇F (−1)D\F XG⊇C (−1)D\F XG⊇C (−1)G\C(ϕ0,G)0,D = XC⊇D⊇F (−1)G\Cϕ0,G = δC,{1,...,N } XD⊇F (−1)D\F XG⊇C since F 6= {1, . . . , N }. (−1)G\C(ϕ0,G)0,D (−1)D\F ϕ0,{1,...,N } = 0, Proof of Theorem 3.5. For every ∅ 6= F ⊆ {1, . . . , N } let the functionals ϕF as defined above and set ϕ∞ := ϕ −X{ϕF ∅ 6= F ⊆ {1, . . . , N }}. As we have observed every ϕF satisfies the KMS-condition and its normalization will provide the F -component for the convex decomposition of ϕ. Consequently the normalization of ϕ∞ will give the infinite part of ϕ. To see that ϕ∞ is positive we proceed inductively. First consider F = {1, . . . , N }. By Lemma 3.7 we get that ϕ{1,...,N }(Q{1,...,N }) = ϕ0,{1,...,N }(Q{1,...,N }) = ϕ(Q{1,...,N }). The functional ϕ − ϕ{1,...,N } satisfies the KMS-condition and it is positive by Lemma 3.8. It annihilates Q{1,...,N } and thus as in Proposition 3.3, it annihilates the ideal that Q{1,...,N } generates. Hence ϕ − ϕ{1,...,N } induces a positive functional on the quotient of N T (X) by the ideal hQ{1,...,N }i. Now let F = {2, . . . , N }. By Lemma 3.9 and Lemma 3.8 we also have that ϕ{2,...,N } = (ϕ − ϕ{1,...,N }){2,...,N } ≤ ϕ − ϕ{1,...,N } and therefore the functional ϕ − ϕ{1,...,N } − ϕ{2,...,N } is positive. By Lemma 3.7 we have that ϕ{1,...,N }(Q{2,...,N }) + ϕ{2,...,N }(Q{2,...,N }) = = ϕ0,{1,...,N }(Q{2,...,N }) + ϕ0,{2,...,N }(Q{2,...,N }) − ϕ0,{1,...,N }(Q{2,...,N }) = ϕ(Q{2,...,N }). Moreover ϕ{2,...,N }(Q1) = 0 by Lemma 3.8, and so ϕ{2,...,N }(Q{1,...,N }) = 0. Therefore the positive functional ϕ − ϕ{1,...,N } − ϕ{2,...,N } annihilates both Q{1,...,N } and Q{2,...,N } and satisfies the KMS-condition. Hence, as in Proposition 3.3, it defines a positive functional on the quotient of N T (X) by the ideal hQ{1,...,N }, Q{2,...,N }i. For the inductive hypothesis let F1 = · · · = Fk = n and suppose that ϕ − XF ≥n+1 ϕF − ϕFl kXl=1 defines a positive functional on the ideal Let D 6= F1, . . . , Fk such that D = n. We wish to show that the functional hQF , QFl F ≥ n + 1, l = 1, . . . , ki. ϕ − XF ≥n+1 ϕF − kXl=1 ϕFl − ϕD is positive and annihilates the projections QD and {QF , QFl F ≥ n + 1, l = 1, . . . , k}. 20 E.T.A. KAKARIADIS An application of Lemma 3.9 and Lemma 3.8 gives positivity since ϕD = (ϕ − XF ≥n+1 ϕF − kXl=1 ϕFl)D ≤ ϕ − XF ≥n+1 ϕF − ϕFl. kXl=1 For G such that D ∩ Gc 6= ∅ and i ∈ D \ G we have ϕG(QD) ≤ ϕG(Qi) = 0. In particular ϕFl(QD) = 0 for all l = 1, . . . , k. Therefore we obtain XF ≥n+1 kXl=1 ϕF (QD) + ϕFl(QD) + ϕD(QD) = = XF :D⊆F(cid:20) XC:F ⊆C = XC:D⊆C ϕ0,C(QD)(cid:20) XF :D⊆F ⊆C (−1)C\F ϕ0,C(QD)(cid:21) = XC:D⊆C(cid:20) XF :D⊆F ⊆C (−1)C\F ϕ0,C(QD)(cid:21) (−1)C\F (cid:21) = ϕ0,D(QD) = ϕ(QD), since PF :D⊆F ⊆C(−1)C\F = 0 unless D = C. Furthermore we have that Fl ∩ Dc 6= ∅ as D 6= Fl and F ∩ Dc 6= ∅ as D < F . Therefore G ∩ Dc 6= ∅ for any G ∈ {F, Fl F ≥ k + 1, l = 1, . . . , k}. By choosing i ∈ G \ D we get that Finally note that the positive functional ϕD(QG) ≤ ϕD(Qi) = 0. ϕ − XF ≥n+1 ϕF − kXl=1 ϕFl − ϕD satisfies the KMS condition as so does every summand. Therefore it defines a positive func- tional on the quotient N T (X)/hQF , QFl, QD F ≥ n + 1, l = 1, . . . ki. Inducting on F = N, N −1, . . . , 1 then gives that ϕ∞ is positive on the quotient of N T (X) by the ideal hQF ∅ 6= F i = hQi i = 1, . . . , N i which is exactly N O(A, X). This includes that ϕ∞(Qi) = 0 for all i ∈ {1, . . . , N }. To show uniqueness of the decomposition suppose that there are ϕ′ F ∈ EF β (N T (X)) with ϕ∞ + X∅6=F ⊆{1,...,N } ϕF = ϕ′ ∞ + X∅6=F ⊆{1,...,N } ϕ′ F . By Proposition 3.3 every ϕF , ϕ′ hQF i while F ∈ EF β (N T (X)) is uniquely identified by its restriction on Applying on hQ{1,...,N }i yields ϕF hQDi = 0 = ϕ′ F hQDi if D ∩ F c 6= ∅. ϕ{1,...,N }hQ{1,...,N}i = ϕ′ {1,...,N }hQ{1,...,N}i and thus so do their unique extensions on N T (X). We may thus remove those from the sums and proceed with {2, . . . , N }. Inducting on the sets F = N − 1 identifies ϕF with ϕ′ F whenever F = N, N − 1. Inducting on F concludes that so is true for all F leaving ϕ∞ = ϕ′ ∞. EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 21 4. Parametrization of the gauge-invariant equilibria We now proceed to the parametrization of the F -components of the gauge-invariant equi- librium states. We need to identify the corresponding F -parts in the simplex T(A) of tracial states of A. Definition 4.1. Let X be a product system over A with a unit decomposition x = {xi,j j = 1, . . . , di, i = 1, . . . , N } and let β > 0. For every ∅ 6= F ⊆ {1, . . . , N } and τ ∈ T(A) we define Moreover we define the F -set of tracial states of A by cF τ,β :=X{e−µβτ (hxµ, xµi) ℓ(µ) ∈ F }. TF β (A) := {τ ∈ T(A) cF τ,β < ∞ and eβτ (a) = hxi,j, axi,ji for all i /∈ F }. diXj=1 In particular for F = {1, . . . , N } we write β (A) := {τ ∈ T(A) c{1,...,N } Tfin τ,β = Xµ=k e−kβτ (hxµ, xµi) < ∞}. The case of F = ∅ is recaptured in the averaging traces, namely AVTβ(A) := {τ ∈ T(A) eβτ (a) = diXj=1 τ (hxi,j, axi,ji) for all i = 1, . . . , N }. Due to Remark 4.7 that will follow, the above definitions are independent of the choice of the unit decomposition. Proposition 4.2. Let X be a product system of finite rank over A and let β > 0. Then β (A) ∩ TF ′ TF Proof. Without loss of generality let i ∈ F ′ \ F and suppose there is a τ ∈ TF β (A) = ∅ whenever F 6= F ′. β (A) ∩ TF ′ β (A). Then ekβ =Pµi=k τ (hxi,µi, xi,µii) as i /∈ F and so we reach the contradiction e−kβτ (hxi,µi, xi,µii) = 1 = ∞. ∞ > cF ′ τ,β ≥ ∞Xk=0 Xµi=k ∞Xk=0 We wish to establish a parametrization of G-EF β (N T (X)) by an appropriate sub-simplex β (A). In order to achieve this we need a characterization of the ideals given by IF = of TF ker{A → N O(F, A, X)} from Definition 2.1. Proposition 4.3. Let X be a product system of finite rank over A and fix a set ∅ 6= F ⊆ {1, . . . , N }. Then IF = {a ∈ A lim k ϕk·1F (a) = 0}. Proof. Recall that by definition we have N O(F, A, X) = N T (X)/hQi i ∈ F i and [13, Proposition 4.6] yields hQi i ∈ F i = span{t(Xn)QC t(Xm)∗ ∅ 6= C ⊆ F, n, m ∈ ZN + }. Equation (3.1) then implies QF chQi i ∈ F iQF c = span{t(Xn)QC∪F ct(Xm)∗ ∅ 6= C ⊆ F, n, m ∈ F }. Let the projections p(k) := pk·1F for k ∈ Z+. It is clear that if n, m ∈ F and k is large enough so that k · 1C > nC , mC then p(k)t(ξn)QC∪F ct(ξm)∗p(k) = 0. 22 E.T.A. KAKARIADIS Therefore if π(a) ∈ IF then QF cπ(a)QF c ∈ QF chQi i ∈ F iQF c and so lim k φk·1F (a) = lim k p(k)π(a)p(k) = lim k p(k)QF cπ(a)QF cp(k) = 0. Conversely, fix ε > 0 and let k ∈ Z+ such that kφk·1F (a)k < ε. If i ∈ F then 1 − Pk·i =X{pn ni = 0, . . . , k − 1} = k−1Xl=0 Ql·i i ∈ IF , f := X∅6=C⊆F (−1)Cπ(a)Yi∈C (1 − Pk·i) ∈ IF . and therefore Then we get kπ(a) + IF k ≤ kπ(a) + f k = kXC⊆F (−1)Cπ(a)Yi∈C (1 − (1 − Pk·i))k = kπ(a)Yi∈F (1 − Pk·i)k Pk·ik = kπ(a)Yi∈F = k X ⊕ n≥k·1F φn(a)k = sup n∈ZN + kφk·1F (a) ⊗ idXnk ≤ kφk·1F k < ε. As ε was arbitrary we derive that π(a) ∈ IF . Theorem 4.4. Let X be a product system of finite rank over A and β > 0. Then we have the following parametrization: (1) For F = ∅ there is a bijection Φ∞ : {τ ∈ AVTβ(A) τ I{1,...,N} = 0} → G-E∞ β (N T (X)), such that Φ∞ (2) For F 6= ∅ there is a bijection τ (π(a)) = τ (a) for all a ∈ A. ΦF : {τ ∈ TF β (A) τ IF c = 0} → G-EF β (N T (X)), such that ΦF τ (QF ) · cF τ,β = 1 and ΦF τ (QF π(a)QF ) = ΦF τ (QF ) · τ (a) for all a ∈ A. If x = {xi,j j = 1, . . . , di, i = 1, . . . , N } is a unit decomposition for X then ΦF τ (t(ξn)t(ηm)∗) = δn,m (cF τ,β)−1 Xℓ(µ)∈F e−(n+µ)βτ (hηm ⊗ xµ, ξn ⊗ xµi), This description is independent of the choice of the decomposition. (3) Every ΦF for F 6= ∅ respects convex combinations in the sense that if λ ∈ (0, 1), then cF τ1,β cF τ,β for τ = λτ1 + (1 − λ)τ2 with τ1, τ2 ∈ TF ΦF (τ ) = λ β (A), and ΦF (τ1) + (1 − λ) cF τ2,β cF τ,β ΦF (τ2) (ΦF )−1(ϕ) = λ ϕ1(QF ) ϕ(QF ) (ΦF )−1(ϕ1) + (1 − λ) ϕ2(QF ) ϕ(QF ) (ΦF )−1(ϕ2) for ϕ = λϕ1 + (1 − λ)ϕ2 with ϕ1, ϕ2 ∈ G-EF preserve the extreme points of the simplices. β (N T (X)). Therefore the parametrizations The proof follows from a number of steps. Henceforth we write Φ ≡ ΦF . EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 23 Lemma 4.5. Let τ ∈ TF β (A) with τ IF c = 0. Consider the C-subalgebra BF c := span{t(Xk)t(Xw)∗ k, w ⊥ F } satisfies the KMS-condition (2.4). of the fixed point algebra N T (X)γ . Then τ extends to a gauge invariant state eτ on BF c that Proof. It suffices to extend τ on E(BF c). Let q : N T (X) → N O(F c, A, X) be the canonical quotient map. For convenience set σ := qπ, s := qt and ψ′ = qψ. Let B = σ(A) and Yn = s(Xn). Now we see that Y = {Yn}n⊥F is a product system of finite rank. Moreover it is injective. Indeed for i /∈ F the covariance on N O(F c, A, X) gives that s(xi,j)s(xi,j)∗ = ψ′ i(φX (1A)) = σ(1A) = 1. diXj=1 Therefore if σ(a) ∈ ker φY,i then σ(a) = σ(a)Pdi We claim that the identity representation id : Y → N O(F c, A, X) gives the inclusion N O(B, Y ) ⊆ N O(F c, A, X). As Y is regular we have that N O(B, Y ) = N O(Y ). The identity representation is trivially injective on B and admits a gauge action. Moreover for every i ⊥ F we have j=1 s(xi,j)s(xi,j)∗ = 0. id(b)(I − diXj=1 id(s(xi,j))id(s(xi,j))∗) = bq(Qi) = 0. By [13] then (idB, idY ) is covariant along all directions and thus by [33] it lifts to a faithful representation of N O(B, Y ). Similar to the one-variable case for regular C-correspondences [30] we obtain that the fixed point algebra of N O(Y ) can be written as a direct limit. Namely, we have that lim−→(KYn, ⊗id) ≃ N O(Y )γ = span{ψ′ n(KXn) n ⊥ F } = E(BF c), where ⊗idYm : KYn → KYn+m : θ Xn ξn,ηn 7→ θ Xn ξn,ηn Therefore we obtain the diagram ⊗ idYm = Xℓ(µ)=m θ Xn ξnxµ ,ηnxµ . KYn ≃ ⊗idYm / KYn+m ≃ ψ′ n(KXn) ιn+m n /❴❴❴❴❴❴❴❴❴❴ ψ′ n+m(KXn+m) where the induced map ιn+m n is given by ιn+m n (s(ξn)s(ηn)∗) = Xℓ(µ)=m s(ξn)s(xµ)s(xµ)∗s(ηn)∗. In order to extend τ to a state on E(BF c) we have to find states τ ′ n ⊥ F that are compatible with the direct limit connecting maps, i.e., that they satisfy n(KXn) → C for every n : ψ′ n = τ ′ For the first step τ defines a tracial state τ ′ τ ′ n+iιn+i n for all i /∈ F. 0 on σ(A) with τ = τ ′ Moreover it is clear that if n ⊥ F then 0σ as it factors through q. τ ′ 0σ(a) = e−nβ Xℓ(µ)=n τ (hxµ, axµi). / / O O 24 E.T.A. KAKARIADIS For every n ⊥ F we define the functional τ ′ n on ψ′ n(KXn) by nψ′ τ ′ n(kn) := e−nβ Xℓ(µ)=n τ (hxµ, knxµi) for kn ∈ KXn. If ψn(kn) ∈ ker q then π(hxµ, knxµi) = t(xµ)∗ψn(kn)t(xµ) ∈ ker q ∩ π(A) and thus τ ′ well defined. Note also that τ ′ {yν ℓ(ν) = n} is another decomposition of Xn then n is n does not depend on the choice of the decomposition. For if Xℓ(µ)=n τ (hxµ, knxµi) = Xℓ(µ)=n Xℓ(ν)=n = Xℓ(ν)=n Xℓ(µ)=n τ (hxµ, knyνihyν , xµi) τ (hyν, xµihxµ, knyνi) = Xℓ(ν)=n τ (hyν , knyνi). Every τ ′ n is a state since for every positive contraction kn we obtain e−nβ Xℓ(µ)=n τ (hxµ, knxµi) ≤ e−nβ Xℓ(µ)=n τ (hxµ, xµi) = τ (1) = 1. Now we see that every τ ′ n satisfies τ ′ n(s(ξn)s(ηn)∗) = e−nβ Xℓ(µ)=n = e−nβ Xℓ(µ)=n τ (hxµ, θ Xn ξn,ηn xµi) τ (hηn, xµihxµ, ξni) = e−nβτ (hηn, ξni). Now we can verify that τ ′ n+iιn+i n = τ ′ n for i /∈ F by computing on rank one operators: τ ′ n+iιn+i n (s(ξn)s(ηn)∗) = diXj=1 τ ′ n+i(s(ξn)s(xi,j)s(xi,j)∗s(ηn)∗) = e−(n+1)β = e−(n+1)β τ (hηn ⊗ xi,j, ξn ⊗ xi,ji) τ (hxi,j, hηn, ξnixi,ji) diXj=1 diXj=1 n(s(ξn)s(ηn)∗). n for the induced state on E(BF c) that extends every τ ′ = e−nβτ (hηn, ξni) = τ ′ satisfies the KMS-condition since Write eτ ′ := lim−→ τ ′ n. We see that eτ ′ eτ ′(s(ξn)s(ηn)∗) = τ ′ n(s(ξn)s(ηn)∗) = e−nβτ (hηn, ξni) = e−nβeτ ′(s(ηn)∗s(ξn)). Therefore we get that the functional (4.1) τ . defines a gauge invariant state on BF c that satisfies the KMS-condition, and clearly eτ π = Proof of Theorem 4.4 (1). It is clear that eτ (Qi) = 0 for every i /∈ F . If F = ∅ we stop the construction here and deduce the weak-homeomorphism Φ∞ : AVTβ(A) ∩ {τ ∈ T(A) τ I{1,...,N} = 0} → G-E∞ eτ :=eτ ′qE β (N T (X)) : τ 7→eτ . EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 25 If F 6= ∅ then we consider the projection QF = Qi∈F (1 − Pi) and we will construct the equilibrium state by using the statistical approximations on QF . For every n ∈ F we define the C-correspondence Zn := t(Xn)BF c over BF c := span{t(Xk)t(Xw)∗ k, w ⊥ F }, where the bimodule structure is induced from N T (X). Since BF c is unital we have that t(Xn) ⊆ t(Xn)BF c and thus for n, m ∈ F we derive t(Xn)t(Xm)BF c ⊆ t(Xn)BF ct(Xm)BF c ⊆ t(Xn+m)BF c. Thus {Zn}n∈F defines a product system. By the Gauge-Invariant-Uniqueness-Theorem we have N T (X) = N T (Z). Indeed first notice that the map id : Z → N T (X) defines a Nica-covariant representation with a gauge action. Moreover C∗(idBF c , idZ) = N T (X) since t(Xn)t(Xm)∗ ⊆ t(XnF )BF ct(XmF )∗ for all n, m ∈ ZN + . Finally we need to verify that BF c ∩ BZ (0,∞] = (0) where (0,∞] := span{t(ξn)bt(ηn)∗ 0 6= n ∈ F, b ∈ BF c}. BZ To reach contradiction let g ∈ BF c ∩ BZ (0,∞] so that QF gQF ∈ QF BZ (0,∞]QF = (0). Let k, w ⊥ F such that g = ψ(kk,w) + g′ and k or w is minimal with 0 6= kk,w ∈ K(Xw, Xk). We then compute K(Xw, Xk) ∋ kk,w = pkQF ψ(kk,w)QF pw = pkQF gQF pw = 0. However this gives the contradiction kψ(kk,w)k = sup m∈ZN + kkk,w ⊗ idXmk = 0. ιn : Zn → FZ, of the inclusions by writing Consequently N T (X) acts on the Fock space FZ =P ⊕{Zn n ∈ F }. Let us keep track and write (eπ,et) for the representation of N T (X) with eπ(a)ιm(t(ξm)b′) = ιm(π(a)t(ξm)b′) and et(t(ξn))ιm(t(ξm)b) = ιnF +m(t(ξn)t(ξm)b). Consider the GNS-representation (Heτ , ρeτ , xeτ ) related to the constructed eτ and define ϕµ(f ) := hιℓ(µ)(t(xµ)) ⊗ xeτ , (ρ × v)(f )(cid:2)ιℓ(µ)(t(xµ)) ⊗ xeτ )(cid:3)iH for f ∈ N T (X). (ρ, v) := (eπ ⊗ I,et ⊗ I) acting on H := FZ ⊗ρeτ Heτ . For every µ with ℓ(µ) ∈ F let the vector state ϕµ given by We use the ϕµ to define the functional (4.2) For f = 1A we have and therefore Φτ (f ) := (cF τ,β)−1X{e−µβϕµ(f ) ℓ(µ) ∈ F }. ϕµ(π(1A)) =eτ (t(xµ)∗t(xµ)) = τ (hxµ, xµi), τ,β)−1X{e−µβτ (hxµ, xµi) ℓ(µ) ∈ F } = 1. Φτ (π(1A)) = (cF This guarantees that Φτ is a well defined state on N T (X). G-EF β (N T (X)) we require the following properties for the ϕµ: In order to show that Φτ ∈ 26 E.T.A. KAKARIADIS • (i) For every m ∈ F and ℓ(µ) ∈ F we have that • (ii) For every 0 6= m ∈ F we have et(t(ξm)∗)ιℓ(µ)(t(xµ)) = δm,m∧ℓ(µ) · ιℓ(µ)−m(t(ξm)∗t(xµ)). (eπ ×et)(QF )ιm(t(ζm)b) = 0, while (eπ ×et)(QF )ι0(π(1)) = ι0(π(1)). Therefore we have • (iii) Recall thateτ =eτ E on BF c. For every n, m ∈ ZN ϕµ(QF bQF ) = δℓ(µ),0eτ (b) for all b ∈ BF c. ϕµ(t(ξn)t(ηm)∗) = + we get that = ϕµ(t(ξnF )(cid:2)t(ξnF c )t(ηmF c )∗(cid:3)t(ηmF )∗) = δnF ,nF ∧ℓ(µ) · δmF ,mF ∧ℓ(µ) ·eτ E(t(xµ)∗t(ξnF )(cid:2)t(ξnF c )t(ηmF c )∗(cid:3)t(ηmF )∗t(xµ)) = δn,m · δnF ,nF ∧ℓ(µ)eτ (t(xµ)∗t(ξn)t(ηm)∗t(xµ)). Xℓ(µ)=r ϕµ(t(ξn)t(ηm)∗) = δn,m · δnF ,nF ∧r · e−nF c β Xℓ(ν)=r−nF ϕν (t(ηm)∗t(ξn)). + then Lemma 4.6. With the aforementioned notation, if r ∈ F and n, m ∈ ZN (4.3) Consequently Φτ attains the stated form with respect to the unit decomposition x = {xi,j j = 1, . . . , di, i = 1, . . . , N } and thus it satisfies the KMS-condition. Proof. Property (iii) for ϕµ implies that ϕµ = ϕµE and so Φτ = Φτ E. Therefore let If nF 6≤ ℓ(µ) ∈ F then again property (iii) yields us consider the case where n = m. ϕµ(t(ξn)t(ηn)∗) = 0. Now suppose that nF ≤ ℓ(µ) ∈ F with ℓ(µ) = r. Since r − nF ∈ F we get (r − nF ) ∧ nF c = 0 and so t(ηn)∗t(xµ) ∈ t(XnF c )∗t(Xr−nF ) ⊆ t(Xr−nF )t(XnF c )∗. Therefore by using the approximate identity of ψ(KXr−nF ) on t(Xr−nF ) we get t(xν)t(xν )∗t(ηn)∗t(xµ) = t(ηn)∗t(xµ). Xℓ(ν)=r−nF Xℓ(µ)=r Now for ℓ(ν) = r − nF we have t(ξn)t(xν) ∈ t(Xn+r−nF ) = t(Xr)t(XnF c ) and so t(xµ)t(xµ)∗t(ξn)t(xν ) = t(ξn)t(xν ). Finally we have that t(xµ)∗t(ξn)t(xν ) ∈ t(Xr−nF )∗t(XnF c )t(Xr−nF ) ⊆ t(XnF c ) ⊆ BF c. Therefore we can use the KMS-condition foreτ and nF c to obtain Xℓ(µ)=r ϕµ(t(ξn)t(ηn)∗) = Xℓ(µ)=r Xℓ(ν)=r−nFeτ (t(xµ)∗t(ξn)t(xν ) } ∈ t(XnF c ) · t(xν )∗t(ηn)∗t(xµ) ) } {z ∈ t(XnF c )∗ {z = e−nF c β Xℓ(ν)=r−nF Xℓ(µ)=reτ (t(xν)∗t(ηn)∗ · t(xµ)t(xµ)∗t(ξn)t(xν )) = e−nF c β Xℓ(ν)=r−nFeτ (t(xν)∗t(ηn)∗t(ξn)t(xν )) = e−nF c β Xℓ(ν)=r−nF ϕν(t(ηn)∗t(ξn)). EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 27 Consequently we derive τ,β · Φτ (t(ξn)t(ηn)∗) = cF =X{e−µβϕµ(t(ξn)t(ηn)∗) ℓ(µ) ∈ F } = e−nF c βX{e−µβϕν (t(ηn)∗t(ξn)) ℓ(ν) = ℓ(µ) − nF , nF ≤ ℓ(µ) ∈ F } = e−nF c βX{e−(µ+nF )βϕµ(t(ηn)∗t(ξn)) ℓ(µ) ∈ F } τ,β · e−nβΦτ (t(ηn)∗t(ξn)). = cF It follows now that Φτ both has the required form and that it satisfies the KMS-condition. Remark 4.7. We note that the form of Φτ does not depend on the choice of the decompo- sitions. Indeed if y = {yi,j j = 1, . . . , d′ i, i = 1, . . . , N } defines a second decomposition for X then for any n ∈ ZN + (and not just for n ∈ F ) we get Xℓ(µ)=n τ (hηn ⊗ xµ, ξn ⊗ xµi) = Xℓ(µ)=n Xℓ(ν)=n = Xℓ(ν)=n Xℓ(µ)=n = Xℓ(ν)=n τ (hηn ⊗ xµ, ξn ⊗ yνihyν, xµi) τ (hyν , xµihxµ, hηn, ξniyνi) τ (hyν , hηn, ξniyνi) = Xℓ(ν)=n τ (hηn ⊗ yν, ξn ⊗ yνi). In particular we have thatPℓ(µ)=n τ (hxµ, xµi) =Pℓ(ν)=n τ (hyν, yνi). Moving one step further let i /∈ F and n ∈ F so that both families {xµxi,j ℓ(µ) = n, j = 1, . . . , di} and {xi,jxµ ℓ(µ) = n, j = 1, . . . , di} define unit decompositions for Xn+i. Since τ ∈ TF β (A) we conclude that diXj=1 Xℓ(µ)=n τ (hxi,j ⊗ xµ, xi,j ⊗ xµi) = = diXj=1 Xℓ(µ)=n diXj=1 Xℓ(µ)=n diXj=1 = Xℓ(µ)=n τ (hxi,jxµ, xi,jxµi) τ (hxµxi,j, xµxi,ji) τ (hxi,j, hxµ, xµixi,ji) = Xℓ(µ)=n eβτ (hxµ, xµi). Proof of Theorem 4.4 (2 -- 4). First we show that Φτ is indeed in G-EF using the KMS-condition and property (ii), for ℓ(ν) = n ∈ F we have that β (N T (X)). By τ,β · Φτ (Qn cF F ) = cF e−nβΦτ (QF t(xν )∗t(xν)QF ) τ,β Xℓ(ν)=n = Xℓ(ν)=n = Xℓ(ν)=n e−nβX{e−µβϕµ(QF π(hxν , xν i)QF ) ℓ(µ) ∈ F } e−nβτ (hxν, xν i). Applying in particular for n = 0 we get Φτ (QF )−1 = cF conclude τ,β. Summing over all n ∈ F we Xn∈F Φτ (Qn F ) = (cF τ,β)−1X{e−νβτ (hxν , xν i) ℓ(ν) ∈ F } = Φτ (π(1A)) = 1. 28 E.T.A. KAKARIADIS Now let i /∈ F and we will show that Φτ (Pi) = 1. By Remark 4.7 we derive that Φτ (Pi) = Φτ (t(xi,j)t(xi,j)∗) diXj=1 τ,β)−1 Xw∈ZN τ,β)−1 Xw∈ZN + + = (cF = (cF e−(w+1)β diXj=1 Xℓ(µ)=w e−(w+1)βeβ Xℓ(µ)=w τ (hxi,j ⊗ xµ, xi,j ⊗ xµi) τ (hxµ, xµi) = Φτ (1) = 1. Next we show that Φ is a bijection. For injectivity notice that property (ii) yields Therefore τ is uniquely determined by Φτ . For surjectivity let ϕ ∈ G-EF ϕ(QF ) = 0 then we would have that β (N T (X)). If Φτ (QF π(a)QF ) = (cF τ,β)−1eτ (π(a)) = Φτ (QF )τ (a). ϕ(QF t(xν )∗t(xν )QF ) ≤ e−nβ Xℓ(µ)=n ϕ(QF ) = 0 ϕ(Qn F ) = e−nβ Xℓ(µ)=n which contradicts thatPn∈F ϕ(Qn F ) = 1. Hence we can define the state τϕ on A given by τϕ(a) := ϕ(QF )−1ϕ(QF π(a)QF ) for all a ∈ A. It is clear that τϕ ∈ T(A) as ϕ is an equilibrium state and QF ∈ π(A)′. Since ϕ(Qi) = 0 for all i /∈ F we get ϕIF c = 0 and thus τϕIF c = 0. Furthermore we use the KMS-condition on ϕ to get ϕ(QF ) Xℓ(µ)=n τϕ(hxµ, xµi) = enβ Xℓ(µ)=n ϕ(t(xµ)QF t(xµ)∗) = enβϕ(Qn F ), and therefore cF τ,β =X{e−µβτϕ(hxµ, xµi) ℓ(µ) ∈ F } = ϕ(QF )−1Xn∈F ϕ(Qn F ) = ϕ(QF )−1. Furthermore as ϕ(Qi) = 0 for i /∈ F we get that ϕ(Pif ) = ϕ(f ) for all f ∈ N T (X). Hence by the KMS-condition on ϕ and equation (3.1) we derive eβτϕ(π(a)) = eβϕ(QF )−1ϕ(QF π(a)QF ) = eβϕ(QF )−1ϕ(PiQF π(a)QF ) = ϕ(QF )−1 ϕ(t(xi,j )∗QF π(a)QF t(xi,j)) diXj=1 ϕ(QF )−1ϕ(QF t(xi,j)∗π(a)t(xi,j)QF ) = = diXj=1 τϕ(hxi,j, axi,ji). diXj=1 Consequently τϕ ∈ TF β (A) and we can now form the induced state Φτϕ. We wish to show that ϕ = Φτϕ and conclude surjectivity. Since they are both gauge-invariant, Proposition 2.6 F ) = 1 we get that asserts that it suffices to show that they agree on π(A). SincePn∈F ϕ(Qn EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 29 ϕ(π(a)) =Pn∈F ϕ(Qn ϕ(π(a)) = lim m F π(a)) and we can deduce mXk=0(cid:20)X{ϕ(t(xµ)QF t(xµ)∗π(a)) ℓ(µ) ∈ F, µ = k}(cid:21) mXk=0(cid:20)X{e−kβϕ(QF t(xµ)∗π(a)t(xµ)QF ) ℓ(µ) ∈ F, µ = k}(cid:21) mXk=0(cid:20)X{e−kβτϕ(hxµ, axµi) ℓ(µ) ∈ F, µ = k}(cid:21) τ,β)−1X{e−µβτϕ(hxµ, axµi) ℓ(µ) ∈ F } = Φτϕ(π(a)). = lim m = ϕ(QF ) lim m = (cF The last part on convexity follows in the same way as in [23, Corollary 6.3] and it is omitted. Remark 4.8. The inverse correspondence (ΦF )−1 : G-EF β (A) is continuous for all F . This follows by the way we retrieve τϕ when F 6= ∅, and because Φ(τ ) is an extension of τ when F = ∅. Therefore Φ is a homeomoprhism when TF β (A) is closed in T(A). On the other hand Φ∞ is always a weak-homeomorphism as AVTβ(A) is weak-closed. β (N T (X)) → TF 5. Relative Nica-Pimnser algebras We next apply the obtained parametrization to relative Nica-Pimsner algebras. We say that a family {IF ∅ 6= F ⊆ {1, . . . , N }} of ideals of A is a lattice of ⊥-invariant ideals if: (i) IF ⊆ IF ′ when F ⊆ F ′; and (ii) hXn, IF Xni ⊆ IF when n ⊥ F . We then define the TN -equivariant ideal KI := hπ(a)QF a ∈ IF , F ⊆ {1, . . . , N }i = span{t(Xn)π(a)QF t(Xm)∗ n, m ∈ ZN + , a ∈ IF , ∅ 6= F ⊆ {1, . . . , N }}. The fact that the ideal KI attains this form as a linear space follows in the same way as in [13, Proposition 4.6]. The only that is required is ⊥-invariance and that IF ⊆ I ′ F when F ⊆ F ′. Definition 5.1. Let X be a product system of finite rank. Let {IF ∅ 6= F ⊆ {1, . . . , N }} be a lattice of ⊥-invariant ideals of A. We define the relative Nica-Pimsner algebra N O(I, X) be the quotient of N T (X) by the TN -equivariant ideal KI . Similar to N T (X), we use the projections QF + KI to make sense of the EF β (N O(I, X)) simplices. Since N O(A, X) is a quotient of N O(I, X), we have the same simplex of infinite- type states for N O(I, X). The only restriction for the F -parametrization is to have the states in TF β (A) to annihilate both IF c and IF . Theorem 5.2. Let X be a product system of finite rank over A and β > 0. Let {IF ∅ 6= F ⊆ {1, . . . , N }} be a lattice of ⊥-invariant ideals of A. Then: (1) Every state in G-Eβ(N O(I, X)) admits a unique convex decomposition over the F - simplices G-EF (2) For F = ∅ there is a bijection β (N O(I, X)). Φ∞ : {τ ∈ AVTβ(A) τ I{1,...,N} = 0} → G-E∞ β (N O(I, X)) such that Φ∞ τ π = τ. (3) For F 6= ∅ there is a bijection ΦF : {τ ∈ TF β (A) τ IF c +IF = 0} → G-EF β (N O(I, X)), that factors through qKI : N T (X) → N O(I, X), arising from Theorem 4.4. (4) Every ΦF for F 6= ∅ respects convex combinations and thus preserves the extreme points of the simplices. 30 E.T.A. KAKARIADIS Proof. Let us write K ≡ KI for simplicity. Since qK : N T (X) → N O(I, X) is gauge-invariant we have that ϕ′ ∈ G-EF β (N T (X)) such that ϕ′qK = ϕ. The case F = ∅ gives the same infinite-type states. Hence we consider the case where F 6= ∅. β (N O(I, X)) if and only if there exists a ϕ ∈ G-EF If ϕ′ ∈ G-EF β (N T (X)) defines a unique state τϕ ∈ β (A) such that τϕIF c = 0 from Theorem 4.4. We need to show that τϕIF = 0. Indeed if β (N O(I, X)) then ϕ = ϕ′qK ∈ G-EF TF a ∈ IF then π(a)QF ∈ K and so τϕ(a) = [ϕ′qK(QF )]−1ϕ′qK(QF π(a)QF ) = 0. Conversely let τ ∈ TF β (A) that annihilates both IF c and IF . Construct the induced ϕ ≡ ΦF τ from Theorem 4.4. We need to show that ϕK = 0 and this will induce the required ϕ′ ∈ G-EF β (N O(I, X)). By using the KMS-condition as in Proposition 3.3 it suffices to show that ϕ(π(a)QC ) = 0 for all a ∈ IC and ∅ 6= C ⊆ {1, . . . , N }. First suppose that C ∩ F c 6= ∅. By construction we already have that ϕ(Qi) = 0 for any i ∈ C \ F . Hence ϕ(QC ) = 0 and therefore ϕ(π(a)QC ) = 0 for all a ∈ IC . Next let C ⊆ F . For any D ⊆ C we get ϕ(π(a)PD) = Xℓ(ν)=1D e−Dβϕ(t(xν )∗π(a)t(xν )) = (cF τ,β)−1X{e−(µ+D)βτ (hxν ⊗ xµ, axν ⊗ xµi) ℓ(ν) = 1D, ℓ(µ) ∈ F }. Thus by using the alternating sums we get ϕ(π(a)QC ) = XD⊆C (−1)Dϕ(π(a)PD ) = (cF = (cF (−1)D(cid:20)X{e−µβτ (hxµ, axµi) 1D ≤ ℓ(µ) ∈ F }(cid:21) τ,β)−1 XD⊆C τ,β)−1X{e−µβτ (hxµ, axµi) ℓ(µ) ⊥ C}. For the last equality, if ℓ(µ) 6= 0 with ℓ(µ) ∩ C 6= ∅ then {D ⊆ C 1D ≤ ℓ(µ)} = {D D ⊆ supp ℓ(µ) ∩ C}. Thus the part corresponding to this ℓ(µ) contributes zero to the sum, namely XD⊆C,1D≤supp ℓ(µ) (−1)De−µβτ (hxµ, axµi) = e−µβτ (hxµ, axµi) XD⊆supp ℓ(µ)∩C (−1)D = 0. Hence the only part that survives is for D = ∅ and ℓ(µ) ⊥ C. However a ∈ IC and so for ℓ(µ) ⊥ C we get that hxµ, axµi ∈ IC ⊆ IF , giving τ (hxµ, axµi) = 0. Therefore we conclude that ϕ(π(a)QC ) = 0 also in this case, and the proof is complete. 6. Extending to the full parametrizations We may now obtain the parametrizations of the full simplices of the F -equilibrium states. First we note that ϕ is of infinite type if and only if ϕ(Qif ) = 0 for all f ∈ N T (X) and i = 1, . . . , N , if and only if ϕ(f ) = ϕ(Pif ) = e−β diXj=1 ϕ(t(xi,j)∗f t(xi,j)) for all f ∈ N T (X), i = 1, . . . , N. EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 31 This the furthest we can go at this generality and thus we now pass to the non-infinite types. Now each F -part will depend on the product system we constructed for the proof of Theorem 4.4. Theorem 6.1. Let X be a product system of finite rank over A and β > 0. For ∅ 6= F ⊆ {1, . . . , N } let the F -product system Z over BF c constructed concretely in N T (X) by Zn = t(Xn)BF c for n ∈ F and BF c = span{t(Xk)t(Xw)∗ k, w ⊥ F }. Then the parametrization of Theorem 4.4 lifts to a parametrization where {eτ ∈ Eβ(BF c) eτ π ∈ TF β (A),eτ I′ I′ F c := ker{BF c → N O(F c, A, X)}. F c = 0} → EF β (N T (X)), Likewise the parametrization of the relative Cuntz-Nica-Pimsner algebras from Theorem 5.2 lifts to where {eτ ∈ Eβ(BF c) eτ π ∈ TF β (a),eτ I′ I ′ F := span{t(Xk)π(IF )t(Xw)∗ k, w ⊥ F }. F c +I ′ F = 0} → EF β (N O(I, X)), Proof. The construction is the same with that of Theorem 4.4 and Theorem 5.2 as long as we substitute A with BF c and we verify the relevant points. By Proposition 3.4 we can show that ϕ(t(ξn)t(ξk)t(ηw)∗t(ηm)∗) = δn,mϕ(t(ξn)t(ξk)t(ηw)∗t(ηm)∗) for all n, m ∈ F, k, w ⊥ F. Hence by taking linear combinations and limits we deduce that ϕ(t(ξn)bt(ηm)∗) = δn,mϕ(t(ξn)bt(ηm)∗) for all n, m ∈ F and b ∈ BF c. Let eτ such that eτ π ∈ TF F c = 0. We need to check that the constructed ϕ := Φeτ of Theorem 4.4 is an equilibrium state. For convenience let us say that an element b ∈ BF c is simple if b ∈ t(Xk)t(Xw)∗ for some k, w ⊥ F . In this case we say that b has degree deg(b) = k − w. We may proceed in the same way as in Lemma 4.6 for n, m ∈ F and b1, b2 simple elements to obtain β (A) and eτ I′ (6.1) ϕ(t(ξn)b1b2t(ηm)∗) = δn,me−nF βe− deg(b1)βϕ(b2t(ηm)∗t(ξn)b1). taking linear combinations and limits. In particular ϕ inherits the KMS condition on BF c fromeτ . We may extend this relation by We shall show that ϕ satisfies the KMS condition. Let n, k, w ∈ ZN + and we can apply for t(ξn)t(ξk) ∈ t(XnF +kF ) · t(XnF c +kF c ) and t(ηw) ∈ t(XwF ) · t(XwF c ), to get that ϕ(t(ξn)t(ξk)t(ηw)∗) = δnF +kF ,wF e−n+kβϕ(t(ηw)∗t(ξn)t(ξk)). As long as wF = nF + kF ≥ nF we have t(ηw)∗t(ξn) ∈ t(XwF −nF )∗t(XwF c )∗t(XnF c ) ⊆ t(XwF −nF )∗t(XnF c −nF c ∧wF c )t(XwF c −nF c ∧wF c )∗ ⊆ t(XnF c −nF c ∧wF c )t(XwF c −nF c ∧wF c )∗t(XwF −nF )∗. Now we apply to the adjoint of equation (6.1) and thus get ϕ(t(ξn)t(ξk)t(ηw)∗) = δnF +kF ,wF e−n+kβϕ(t(ηw)∗t(ξn) · t(ξk)) = δnF +kF ,wF e−n+kβe−kβϕ(t(ξk)t(ηw)∗t(ξn)) = δnF +kF ,wF e−nβϕ(t(ξk)t(ηw)∗t(ξn)). 32 Notice that and hence giving that E.T.A. KAKARIADIS t(ξk)t(ηw)∗t(ξn) ∈ t(Xk+n−n∧w)t(Xw−n∧w)∗ ϕ(t(ξk)t(ηw)∗t(ξn)) = δ(k+n−n∧w)F ,(w−n∧w)F ϕ(t(ξk)t(ηw)∗t(ξn)) = δnF +kF ,wF ϕ(t(ξk)t(ηw)∗t(ξn)), ϕ(t(ξn)t(ξk)t(ηw)∗) = e−nβϕ(t(ξk)t(ηw)∗t(ξn)). Proposition 2.6 then implies that ϕ satisfies the KMS condition on N T (X). Injectivity of the parametrization is immediate as τ is completely identified by Φ(τ ). No- tice that QF commutes with BF c due to equation (3.1). Equation (6.1) implies that two equilibrium states that are gauge invariant along F coincide if and only if they coincide on BF c. This yields surjectivity of the parametrization. To see that the parametrization is inherited by N O(I, X) we just need to check that I ′ F is the ideal generated by π(IF ) in BF c and that it is ⊥-invariant. We refer to [13, Proposition 4.6] which gives that π(IF )t(Xm) ⊆ t(Xm)π(IF ) whenever m ⊥ F . Hence indeed I ′ F = BF cπ(IF )BF c. For invariance let m, k, w ⊥ F . Then t(Xm)∗t(Xk)π(IF )t(Xw)∗t(Xm) ⊆ t(Xk−k∧m)t(Xm−k∧m)∗π(IF )t(Xm−m∧w)t(Xw−m∧w)∗ ⊆ t(Xk−k∧m)t(Xx)∗π(IF )t(Xy)t(Xw−m∧w)∗ ⊆ I ′ F , where we have set x = m − k ∧ m − r, y = m − m ∧ w − r, r = (m − k ∧ m) ∧ (m − m ∧ w), and we have used invariance of IF for r ⊥ F , so that t(Xr)∗π(IF )t(Xr) ⊆ π(IF ). Remark 6.2. Recently, Christensen [9] parametrized the equilibrium states for higher rank graphs, without using any of the assumptions of [14]. The decomposition/parametrization for the Toeplitz-Nica-Pimsner algebra that he obtains has the same form with Theorem 3.5 and Theorem 4.4. Nevertheless he achieves more by considering the action to be weighted on different fibers. We shall show here that our main theorems accommodate this setting with a small cali- bration. Fix s = (s1, . . . , sN ) ≥ 0 and let the action σ(s) : r → Aut(N T (X)) : r 7→ γ(exp(is1r),...,exp(isN r)). Then we obtain r (t(ξn)t(ηm)∗) = e−ihn−m,sirt(ξn)t(ηm)∗ σ(s) for hw, si = wisi for all w ∈ ZN . NXi=1 In particular we have µ = hℓ(µ), 1i. Notice that the projections PF , QF are invariant under this new action σ(s). If si 6= 0 for all i ∈ {1, . . . , N } then our arguments apply verbatim by substituting · with h·, si; the only that is required is linearity of h·, si. Now suppose that si = 0 for some i, and without loss of generality let G = {i si = 0}. By using the Fock representation let and X′ n := t(Xn)BG for all n ⊥ G, BG := span{t(ξk)t(ηw)∗ k, w ∈ G} and consider the product system X′ = {X′ n}n⊥G. Now we may apply Theorem 3.5 and Theorem 4.4 to obtain the F -parts for N T (X′). However as shown in the proof of Theorem 4.4 we get that N T (X′) = N T (X) in a canonical way, i.e., the F -projections are the same for any F ⊥ G. Consequently if F ⊥ G then N O(F c, BG, X′) = N O(F c, A, X). Thus the required parametrization for N T (X) is the same with Theorem 4.4 with the only difference that we need to consider traces on BG rather than just on A. From there we can achieve the full parametrizations as in Theorem 6.1. In particular for F ⊆ {1, . . . , N } we define the F -tracial entropy hτ X := lim sup k hτ,F X := lim sup k 1 k τ (hxµ, xµi)(cid:21). τ (hxµ, xµi)(cid:21). 1 k log(cid:20) Xµ=k log(cid:20) Xµ=k,ℓ(µ)∈F log k Xµ=k 1 k EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 33 7. Entropy Now we turn our attention to the entropy of the product system. This will be connected to existence of equilibrium states and tracial states in TF β (A). Definition 7.1. Let X be a product system of finite rank over A. Let x = {xi,j j = 1, . . . , di, i = 1, . . . , N } be a unit decomposition of X. For every τ ∈ T(A) we define the tracial entropy We define the entropy of the unit decomposition x by (7.1) hx X := lim sup k hxµ, xµikA, and then the strong entropy of X by (7.2) X := inf{hx hs X x is a unit decomposition of X}. Likewise for F ⊆ {1, . . . , N } we define the F -entropy of a unit decomposition x by (7.3) hx,F X := lim sup k 1 k log k Xµ=k,ℓ(µ)∈F hxµ, xµikA, and the F -strong entropy of X by (7.4) hs,F X := inf{hx,F X x is a unit decomposition of X}. Moreover we define the entropy of X by hX := inf{β > 0 Eβ(N T (X)) 6= ∅}. Remark 7.2. Using Remark 4.7 we can see that the tracial entropies do not depend on the choice of the decomposition. Hence, if A is abelian then hs,F X for every unit decomposition x. It is also clear that if τ ∈ T(A) with hτ,F X = hx,F X < β then cF τ,β < ∞. X is actually a limit of a decreasing sequence. Indeed for Furthermore the lim sup for hx every k′ ≤ k we get that Xµ=k hxµ, xµi = Xµ=k−k′ hxµ,(cid:0) Xν=k′ hxν , xν i(cid:1)xµi ≤ k Xν=k′ hxν , xνikA Xµ=k−k′ hxµ, xµi, showing that the sequence(cid:0)kPµ=khxµ, xµikA(cid:1)k is submultiplicative. Proposition 7.3. Let X be a product system of finite rank over A. Let x = {xi,j i = 1, . . . , N, j = 1, . . . , di} be a unit decomposition of X. Then for any τ ∈ T(A) we have hτ X ≤ hx X = max{hx,i X i = 1, . . . , N } ≤ max{log di i = 1, . . . , N }, so that X = max{hs,i and likewise for their F -analogues. hτ X ≤ hs X i = 1, . . . , N } ≤ max{log di i = 1, . . . , N }, 34 E.T.A. KAKARIADIS Proof. It suffices to show the first part. It is clear that hτ X ≤ hx X since Xℓ(µ)=n τ (hxµ, xµi) ≤ k Xℓ(µ)=n hxµ, xµikA. Moreover it follows directly that hx,i X = lim sup k hxi,µi, xi,µiikA ≤ lim sup 1 = log di. At the same time we have that hx,F X = lim sup k 1 k hxµ, xµikA ≤ lim sup k 1 k hxµ, xµikA = hx X . 1 k log k Xµi=k log k Xµ=k,ℓ(µ)∈F k 1 k log Xµi=k log k Xµ=k Now set c > 0 such that For ε > 0 let ki such that log c := max{hx,i X i = 1, . . . , N } ≤ hx X . k Xµi=k hxi,µi, xi,µiikA ≤ (c + ε)k for all k ≥ ki. Set k0 := max{k1, . . . , kN }. We claim that for every F there exists a polynomial pF (of degree F − 1) with positive co-efficients such that (7.5) k Xµ=k,ℓ(µ)∈F hxµ, xµikA ≤ pF (k)(c + ε)k, for all k ≥ N · k0. Applying then for F = {1, . . . , N } and the induced polynomial p(k) ≡ p{1,...,N }(k) = aN −1kN −1 + · · · + a0 will give that log c ≤ hx X ≤ lim k 1 k log[p(k)(c + ε)k] = lim k 1 k log[kN −1(c + ε)k] = log(c + ε), since aN −1kN −1 ≤ p(k) ≤ (aN −1 + · · · + a0)kN −1. Taking ε → 0 yields the required hx limit of a decreasing sequence. Set X = log c. To prove the claim, recall that every hx,F X is a M := sup{k Xµ=k,ℓ(µ)∈F hxµ, xµik1/k A F ⊆ {1, . . . , N }, k ∈ Z+}. We have already chosen k0 so that the claim holds for F = {i} with p{i} = 1. Let F = {i, j} and k ≥ N · k0. Then k − n ≥ k0 for every n ≤ k0 and we get k0−1Xn=0 k Xµi=n hxi,µi, xi,µiikA · k Xµj =k−n k0−1Xn=0 ≤ hxj,µj , xj,µj ikA ≤ M n(c + ε)k−n =(cid:2) k0−1Xn=0 M n(c + ε)−n(cid:3)(c + ε)k. EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 35 By a change of variable we derive the symmetrical kXn=k−k0+1 k Xµi=n hxi,µi, xi,µiikA · k Xµj =k−n k0−1Xn=0 ≤(cid:2) k0−1Xn=0 ≤ hxj,µj , xj,µj ikA ≤ hxi,µi, xi,µiikA · k Xµj =n k Xµi=k−n M n(c + ε)−n(cid:3)(c + ε)k. hxj,µj , xj,µj ikA Moreover for the part in-between we have k−k0Xn=k0 k Xµi=n hxi,µi, xi,µiikA · k Xµj =k−n k−k0Xn=k0 ≤ hxj,µj , xj,µj ikA ≤ (c + ε)n(c + ε)k−n ≤ 2k(c + ε)k. Putting these together and using submultiplicativity for the terms of hx,{i,j} required X we derive the k Xµi+µj=k hxi,µi ⊗ xj,µj , xi,µi ⊗ xj,µj ikA ≤ hxi,µi ⊗ xj,µj , xi,µi ⊗ xj,µj ikA ≤ ≤ kXn=0 k Xµi=n,µj=k−n k0−1Xn=0 k Xµi=n k−k0Xn=k0 + hxi,µi, xi,µiikA · k Xµj =k−n k Xµi=n kXn=k−k0+1 k0−1Xn=0 M n(c + ε)−n(cid:3)(c + ε)k. hxi,µi, xi,µiikA · k Xµj =k−n k Xµi=n + ≤(cid:2)2k + 2 hxj,µj , xj,µj ikA+ hxj,µj , xj,µj ikA+ hxi,µi, xi,µiikA · k Xµj =k−n hxj,µj , xj,µj ikA Now suppose that equation (7.5) holds for some F and we will show that it holds for F ∪ {i} for i /∈ F . As before for k ≥ N · k0 and n ≤ k0 we have k0−1Xn=0 k Xµi=n hxi,µi, xi,µiikA · k Xµ=k−n,ℓ(µ)∈F hxµ, xµikA ≤ ≤ k0−1Xn=0 M npF (k − n)(c + ε)k−n ≤(cid:2) k0−1Xn=0 M n(c + ε)−n(cid:3)pF (k)(c + ε)k. 36 E.T.A. KAKARIADIS Here we use that pF has positive co-efficients so that pF (k − n) ≤ pF (k) for n ≤ k0 ≤ k. Its symmetrical is given by kXn=k−k0+1 Moreover we have k Xµi=n ≤ hxi,µi, xi,µiikA · k Xµ=k−n,ℓ(µ)∈F k0−1Xn=0 ≤(cid:2) k0−1Xn=0 k Xµi=k−n M n(c + ε)−n(cid:3)(c + ε)k. hxµ, xµikA ≤ hxi,µi, xi,µiikA · k Xµ=n,ℓ(µ)∈F hxµ, xµikA hxµ, xµikA ≤ k−k0Xn=k0 (c + ε)npF (k − n)(c + ε)k−n ≤ ≤ k Xµi=n hxi,µi, xi,µiikA · k Xµ=k−n,ℓ(µ)∈F k−k0Xn=k0 k Xµi+µ=k,ℓ(µ)∈F hxi,µi ⊗ xµ, xi,µi ⊗ xµikA ≤ We thus conclude pF (k)(c + ε)k ≤ 2kpF (k)(c + ε)k. k−k0Xn=k0 ≤(cid:20)2kpF (k) + (1 + pF (k)) M n(c + ε)−n(cid:21)(c + ε)k. We have that the polynomial pF ∪{i}(k) := 2kpF (k) + (1 + pF (k)) M n(c + ε)−n k0−1Xn=0 k0−1Xn=0 has positive co-efficients (and degree 1 + deg pF = F ). This proves the inductive step and the proof is complete. Proposition 7.4. Let X be a product system of finite rank over A and let β > 0. If τ ∈ β (A) for some F ⊆ {1, . . . , N } then hτ,F TF X ≤ hτ X ≤ β. Proof. Let τ ∈ TF If F = ∅ then it is immediate that β (A) for β > 0. As we are adding positive reals it follows that hτ,F X ≤ hτ X . hτ X = lim sup k 1 k log(cid:20) Xµ=k τ (hxµ, xµi)(cid:21) = lim sup k 1 k log ekβ = β. Let F 6= ∅ and fix ε > 0. Since cF τ,β < ∞ there exists a k0 > 0 such that (cid:20)e−kβ Xµ=k,ℓ(µ)∈F τ (hxµ, xµi)(cid:21)1/k ≤ (1 + ε) for all k ≥ k0. Set also M := max{1,(cid:20)e−kβ Xµ=k,ℓ(µ)∈F τ (hxµ, xµi)(cid:21)1/k k = 0, . . . , k0 − 1}. Therefore for k ≥ k0 we get (cid:20) Xµ=k τ (hxµ, xµi)(cid:21)1/k Xµ=k,ℓ(µ)F =m =(cid:20) k0−1Xm=0 Xµ=k,ℓ(µ)F =m ≤(cid:2)k0M k0 + 2k(1 + ε)k(cid:3)1/keβ. τ (hxµ, xµi) + kXm=k0 Xµ=k,ℓ(µ)F =m τ (hxµ, xµi)(cid:21)1/k EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 37 For k ≥ k0 and m ≤ k0 − 1 we have Xµ=k,ℓ(µ)F =m τ (hxµ, xµi) = Xµ=m,ℓ(µ)∈F(cid:20) Xν=k−m,ℓ(ν)⊥F τ (hxν, hxµ, xµixνi)(cid:21) = e(k−m)β Xµ=m,ℓ(µ)∈F τ (hxµ, xµi) ≤ ekβM m ≤ ekβM k0. A similar computation as above (with 1 + ε in place of M ) for k ≥ m ≥ k0 yields τ (hxµ, xµi) ≤ ekβ(1 + ε)m ≤ ekβ(1 + ε)k. The dominating sequence converges to (1 + ε)eβ. As ε > 0 was arbitrary we obtain the required hτ X ≤ β. Theorem 7.5. Let X be a product system of finite rank over A. Then hX = max{0, inf{hτ X τ ∈ T(A)}}. If in addition hX > 0 then there exists a τ ∈ T(A) such that hX = hτ . Proof. Let β > 0 such that Eβ(N T (X)) 6= ∅. By Proposition 3.4 we moreover have that G-Eβ(N T (X)) 6= ∅. Then by Theorem 3.5 and Theorem 4.4 there exists a τ ∈ TF β (A) for some F . Proposition 7.4 then implies that hτ X ≤ β. Therefore inf{hτ X τ ∈ T(A)} ≤ hX . If hX = 0 then there is nothing to show. Suppose that hX > 0 and assume that there exists a τ ∈ T(A) with 0 ≤ hτ X , hX ), the root test implies that c{1,...,N } β (N T (X)). This is a contradiction as it should be that Eβ(N T (X)) = ∅ by the choice of β. Hence hX is the infimum of the tracial entropies. β (A) and by Theorem 4.4 it induces a Φτ in Efin X < hX . Then for a positive β ∈ (hτ < ∞. Hence τ ∈ Tfin τ,β Furthermore, by weak-continuity there exists an equilibrium state at β = hX for N T (X). (A) for some F , X by the first part, thus Assuming hX > 0, by Theorem 3.5 and Theorem 4.4 there is a τ ∈ TF hX and Proposition 7.4 gives hτ obtaining equality. X ≤ hX . However we also have hX ≤ hτ We now turn our attention to ground states and KMS∞-states. First let us show that KMS∞-states do exist and can only come as limits of equilibrium states of finite type. Proposition 7.6. Let X be a product system of finite rank over A and let F ⊆ {1, . . . , N }. Then β (N T (X)) = ∅ for all C ( F and β > hs,F EC X . Consequently, we have that Eβ(N T (X)) = Efin β (N T (X)) ≃ T(A) for all β > hs X . Proof. First consider the case for C = ∅ and let ϕ ∈ E∞ β (N T (X)). By Proposition 3.4 we can substitute ϕ with a gauge-invariant one. Hence without loss of generality assume Since j ∈ supp n ⊥ C then ϕ(Qj) = 0 and thus ϕ(1 − Pnj ·j) = 0. The Cauchy-Schwartz inequality then yields 1 − Pnj ·j = nj−1Xk=0 Xℓ(µ)=k·j (1 − Pnj·j)) =(1 0 ϕ(Yj∈D t(xµ)Qjt(xµ)∗. if D = ∅, if ∅ 6= D ⊆ supp n, 38 E.T.A. KAKARIADIS that ϕ = ϕE and let τ ∈ AVTβ(A) such that ϕ = Φ∞(τ ) ∈ E∞ F ⊆ {1, . . . , N } and a unit decomposition x = {xi,j j = 1, . . . , di, i = 1, . . . , N } we have β (N T (X)). Then for any β = lim sup k 1 k Xµ=k,ℓ(µ)∈F τ (hxµ, xµi) ≤ lim sup k 1 k k Xµ=k,ℓ(µ)∈F hxµ, xµikA = hx,F X . Taking infimum over x shows that β ≤ hs,F X . Now, fix ∅ 6= C ( F and let ϕ ∈ EC β (N T (X)). Once more by Proposition 3.4 we may take ϕ to be gauge-invariant. For n ∈ F \ C take the projection Pn :=X{pm m ≥ n} = Yi∈supp n = Yi∈supp n (1 − (1 − Pni·i)) = XD⊆supp n Pni·i (−1)DYj∈D (1 − Pnj ·j). However we have that and thus ϕ(Pn) = 1. However Pn =Pℓ(µ)=n t(xµ)t(xµ)∗ and therefore for n = k we get ϕπ(hxµ, xµi) 1 = ϕ(Pn) = e−kβ Xℓ(µ)=n ≤ e−kβ Xµ=k,ℓ(µ)∈F ϕπ(hxµ, xµi) ≤ e−kβk Xµ=k,ℓ(µ)∈F hxµ, xµikA, where we used that ϕπ ∈ T(A). From the latter it follows that β ≤ hx,F that β ≤ hs,F X . X and consequently In particular we have Eβ(N T (X)) = Efin that T(A) = T{1,...,N } (A) when β > hs decomposition of X such that β > hx and thus β β (N T (X)) for all β > hs X . It remains to show X . Let x = {xi,j j = 1, . . . , di, i = 1, . . . , N } be a unit X ≥ hs X ≤ hx X X . For every τ ∈ T(A) we have hτ X ≤ hs lim sup k (cid:20)e−kβ Xµ=k τ (hxµ, xµi)(cid:21)1/k ≤ e−βehx X < 1. Therefore c{1,...,N } τ,β < ∞, and the proof is complete. Now we can provide the characterization of the limit states. Theorem 7.7. Let X be a product system of finite rank over A. Then there exists an affine weak-homeomoprhism Ψ between the states τ ∈ S(A) and the ground states of N T (X) such that Ψτ (π(a)) = τ (a) for all a ∈ A and Ψτ (t(ξn)t(ηm)∗) = 0 when n + m 6= 0. The restriction of Ψ to the tracial states T(A) induces a weak-homeomorphism onto the KMS∞-states of N T (X). If {IF ∅ 6= F ⊆ {1, . . . , N }} is a lattice of ⊥-invariant ideals of A then the corresponding weak-homeomorphisms for N O(I, X) arise by restricting on states that annihilate the ideal I{1,...,N }. EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 39 Proof. For N T (X) let τ ∈ S(A) and consider its GNS-representations (xτ , Hτ , ρτ ). Let (ρ, v) = (π ⊗ I, t ⊗ I) be the induced representation on H = FX ⊗ρτ Hτ and define the vector state Ψτ (f ) := h1A ⊗ xτ , (ρ × v)(f )1A ⊗ xτ iH = τ (Qf Q) for Q := Qi. NYi=1 This map is clearly continuous and injective. Surjectivity and restriction to KMS∞-states follows verbatim from [23, Proof of Theorem 9.1]. Here we need to use Proposition 7.6 so that every KMS∞-state is a limit of finite-type states. For N O(I, X) we notice that Ψτ (π(a)QF ) = τ (Qπ(a)QF Q) = τ (Qπ(a)Q) = τ (a). Therefore Ψτ factors through qKI : N T (X) → N O(I, X) if and only if τ IF for all F , if and only if τ I{1,...,N} , since IF ⊆ I{1,...,N } for all F . 8. Examples 8.1. ZN + -dynamics. When every Xi admits an orthonormal basis then it is easy to deduce that hX = hτ β (A) 6= ∅ then β = log di for all i /∈ F . Hence the only possible points of breaking symmetry is where some of the di's coincide. X = maxi log di. A moment's thought also indicates that if TF X = hs An example in this respect is studied in [23] for ZN + -dynamical systems in the sense of [12]. + → End(A) be an action by unital endomorphisms on a C-algebra A. In this case let α : ZN The associated product system is given by Xn = AA with a · ξn = αn(a)ξn where the identification ξn ⊗ ξm → ξnξm is given by multiplication in A. Then N T (X) is the universal C-algebra with respect to (π, V1, . . . , VN ) such that the Vi are doubly commuting isometries and π(a)Vi = Viπαi(a). By [12] we have in particular that N O(X) is the quotient by the relations π(a)Yi∈F (I − ViV ∗ i ) for all a ∈ IF = \n⊥F α−1 n ((\i∈F ker αi)⊥). In this case di = 1 for all i and all equilibrium states are gauge-invariant. Therefore Eβ(N T (X)) = Efin β (N T (X)) for all β > 0. The equilibrium states of infinite type correspond to tracial states of the universal C-algebra subject to the relations above for (doubly) com- muting unitaries Ui. We may intrinsically identify N O(A, X) with the C-crossed product of the system given by adUi ∈ Aut(qπ(A)) where q : N T (X) → N O(A, X). 8.2. Higher-rank graphs. Let X arise from an irreducible higher-rank graph Λ, so that the single-coloured subgraphs are irreducible. We will not repeat the construction of higher-rank graphs here and the reader may refer to [32] for the details. By [23, Theorem 8.2] and Proposition 7.3 we have that log λ = hτ,i X ≤ hτ X ≤ hs X = max i hs,i X = log λ, where λ is the common Perron-Frobenius eigenvalue of the commuting adjacency matrices. Thus hX = log λ. Moreover this shows that all states above log λ are of finite type and thus gauge-invariant. Now at β = log λ there exists a unique infinite-type state arising from the common Perron-Frobenius eigenvector wPF. The proof is similar to that of [23, Proof of Theorem 8.2]. Hence the E-structure for N T (X) is Eβ(N T (X)) =(Efin E∞ β (N T (X)) β (N T (X)) if β > log λ, if β = log λ, ≃(Pn {wPF} if β > log λ, if β = log λ, where Pn is the probability simplex on the n number of vertices (with dimension n − 1). In this way we fully recover the results of [19]. 40 E.T.A. KAKARIADIS 8.3. Multivariable factorial languages. In [13] we consider product systems that arise from multivariable factorial languages (m-FL). Here we show that the notion of entropy cor- responds to that of allowable words. But first let us recall some elements of the construction from [13]. Fix d symbols and N colours. The elements of (Fd +)N are denoted by µ and consist of N -tuples of sequences on d elements. We write δi(k) for the element that has the symbol k at the i-th position and the empty word in all other places. We use the operation of co-ordinatewise concatenation. Then ((Fd)N , (Fd +)N ) becomes a quasi-lattice by using the partial order coordinate-wise. A set Λ∗ ⊆ (Fd +)N is said to be a multivariable factorial language if (i) (ii) for every i ∈ [N ] there exists at least one k ∈ [d] such that δi(k) ∈ Λ∗; if µ ∈ Λ∗ and ν ∈ µ then ν ∈ Λ∗. As with the one variable case, there is here an alternative definition via forbidden words. We can use the quantization on ℓ2(Λ∗) given by Tδi(k)eµ =(eδi(k)∗µ 0 if δi(k) ∗ µ ∈ Λ∗, otherwise. The product system related to Λ∗ is constructed concretely and is given by Xn = span{Tµa a ∈ A, ℓ(µ) = n} for A = C∗(T ∗ µ Tµ µ ∈ Λ∗). In particular we get an anti-homomorphism α : Λ∗ → End(A) given by αδi(k) = adT ∗ N T (X) is the universal C-algebra with respect to (π, V ) such that: δi(k) . Then (i) V : Λ∗ → B(H) is a Nica-covariant representation for Λ∗, in the sense that VµV ∗ µ VνV ∗ ν =(Vµ∨νV ∗ 0 µ∨ν if µ ∨ ν ∈ Λ∗, otherwise, δi(k)Vδi(l) = δk,lπ(T ∗ (ii) π(a)Vδi(k) = Vδi(k)παδi(k)(a) for all (i, k) ∈ [N ] × [d] and a ∈ A, (iii) V ∗ We wish to show that all equilibrium states in this case are gauge-invariant. To this end let ϕ ∈ Eβ(N T (X) and fix a ∈ A and µ, ν ∈ Λ∗. Without loss of generality we assume that ℓ(µ) ⊥ ℓ(ν). If µ ∨ ν = ∞ then δi(k)Tδi(k)) for all i ∈ [N ] and k, l ∈ [d]. ν ) = e−µβϕ(V ∗ ν Vµπ(a)) = 0. ϕ(Vµπ(a)V ∗ If µ ∨ ν < ∞ and is in Λ∗ then we have ν ) = e−µβϕ(V ∗ = e−2µβϕ(V ∗ ϕ(Vµπ(a)V ∗ ν Vµπ(a)) = e−µβϕ(VµV ∗ ν Vµπ(T ∗ µ T ∗ ν π(T ∗ µ∨νTµ∨νa)) µ∨νTµ∨νa)Tµ). ν ) ≤ e−nµβkak for all n ∈ Z+. This shows Applying repeatedly we thus get that ϕ(Vµπ(a)V ∗ that ϕ(Vµπ(a)V ∗ ν ) = 0. A similar argument applies when ℓ(µ) ∧ ℓ(ν) 6= 0. It is clear that XΛ∗ is of finite rank with the decomposition given by {Tδi(k) k = 1, . . . , d} for every i = 1, . . . , N . Fix now F ⊆ {1, . . . , N } and let the F -projections of the allowable words by cF (µ) := ∗i∈F µi. By definition of Λ∗ we have that cF (Λ∗) ⊆ Λ∗ as cF (µ) ∈ µ. Notice that the T ∗ orthogonal and therefore the F -strong entropy is given by µ Tµ are pairwise hs,F X = lim k 1 k log(cid:0)BF k (Λ∗)(cid:1) for BF k (Λ∗) := {µ ∈ Λ∗ cF (µ) = µ, µ = k}. Hence hs,F entropy hs X measures the entropy of the allowable part supported on F . Likewise the strong X coincides with the entropy of the allowable words. EQUILIBRIUM STATES AND ENTROPY THEORY FOR NICA-PIMSNER ALGEBRAS 41 Acknowledgements. The author would like to thank Sergey Neshveyev for pointing out an error in a preprint of this paper. The author would like to dedicate this paper to Aimilia, on the occasion of her first xx-th birthday. Many happy returns. References [1] Z. Afsar, N. Brownlowe, N.S. Larsen and N. Stammeier, Equilibrium states on right LCM semigroup C-algebras, Int. Math. Res. Not. 6 (2019), 1642 -- 1698. [2] Z. Afsar, A. an Huef and I. Raeburn, KMS states on C-algebras associated to a family of ∗-commuting local homeomorphisms, J. Math. Anal. Appl. 464:2 (2018), 965 -- 1009. [3] J.-B. Bost and A. Connes, Hecke algebras, type III factors and phase transitions with spontaneous sym- metry breaking in number theory, Selecta Math. (N.S.) 1:3 (1995), 411 -- 457. [4] O. Bratteli and D.W. Robinson, Operator algebras and quantum statistical mechanics: C- and W- algebras. Symmetry groups. Decomposition of states, Second Edition, Texts and Monographs in Physics, vol. 1, Springer, New York, 1987. [5] O. Bratteli and D.W. Robinson, Operator algebras and quantum statistical mechanics: Equilibrium states. Models in quantum statistical mechanics, Second Edition, Texts and Monographs in Physics, vol. 2, Springer, Berlin, 1997. [6] N.P. Brown, Topological entropy in exact C-algebras, Math. Ann. 314 (1999), 347 -- 367. [7] N. Brownlowe, A. an Huef, M. Laca and I. Raeburn, Boundary quotients of the Toeplitz algebra of the affine semigroup over the natural numbers, Ergodic Theory Dynam. Systems 32:1 (2012), 35 -- 62. [8] T.M. Carlsen, N.S. Larsen, A. Sims and S.T. Vittadello, Co-universal algebras associated to product systems, and gauge-invariant uniqueness theorems, Proc. London Math. Soc. (3) 103:4 (2011), 563 -- 600. [9] J. Christensen, KMS states on the Toeplitz algebras of higher-rank graphs, preprint at arXiv:1805.09010. [10] A. Connes and E. Størmer, Entropy of I I 1 von Neumann algebras, Acta Math. 134 (1975), 289 -- 306. [11] G. Cornelissen and M. Marcolli, Graphs reconstruction and quantum statistical mechanics, J. Geom. Phys. 72 (2013), 110 -- 117. [12] K.R. Davidson, A.H. Fuller and E.T.A. Kakariadis, Semicrossed products of operator algebras by semi- groups, Mem. Amer. Math. Soc. 247 (2017), no. 1168, (97 + v pages). [13] A. Dor-On and E.T.A. Kakariadis, Operator algebras for higher rank analysis, to appear at Journal d'Analyse Math´ematique (preprint at arXiv:1803.11260). [14] J. Fletcher, A. an Huef and I. Raeburn, A program for finding all KMS states on the Toeplitz algebra of a higher-rank graph, preprint at arXiv:1801.03189. [15] N.J. Fowler, Discrete product systems of Hilbert bimodules, Pacific J. Math. 204:2 (2002), 335 -- 375. [16] J.H. Hong, N.S. Larsen and W. Szymaski, The Cuntz algebra QN and C-algebras of product systems, Progress in operator algebras, noncommutative geometry, and their applications, 97 -- 109, Theta Ser. Adv. Math., 15, Theta, Bucharest, 2012. [17] J.H. Hong, N.S. Larsen and W. Szymaski, KMS states on Nica-Toeplitz algebras of product systems, Internat. J. Math. 23:12 (2012), 1250123, 38 pp. [18] A. an Huef, S. Kang and I. Raeburn, KMS states on the operator algebras of reducible higher-rank graphs, Integral Equations & Operator Theory 88 (2017), 91 -- 126. [19] A. an Huef, M. Laca, I. Raeburn and A. Sims, KMS states on C-algebras associated to higher-rank graphs, J. Funct. Anal. 266:1 (2014), 265 -- 283. [20] A. an Huef, M. Laca, I. Raeburn and A. Sims, KMS states on the C-algebra of a higher-rank graph and periodicity in the path space, J. Funct. Anal. 268:7 (2015), 1840 -- 1875. [21] A. an Huef, M. Laca, I. Raeburn and A. Sims, KMS states on the C-algebras of reducible graphs, Ergodic Theory Dynam. Systems 35:8 (2015), 2535 -- 2558. [22] E.T.A. Kakariadis, On Nica-Pimsner algebras of C-dynamical systems over Zn +, Int. Math. Res. Not. IMRN 4 (2017), 1013 -- 1065. [23] E.T.A. Kakariadis, Entropy theory for the parametrization of the equlibrium states of Pimsner algebras, preprint at arXiv:1712.08795. [24] T. Katsura, On C-algebras associated with C-correspondences, J. Funct. Anal. 217:2 (2004), 366 -- 401. [25] M. Laca and S. Neshveyev, KMS states of quasi-free dynamics on Pimsner algebras, J. Funct. Anal. 211:2 (2004), 457 -- 482. [26] M. Laca and I. Raeburn, Phase transition on the Toeplitz algebra of the affine semigroup over the natural numbers, Adv. Math. 225:2 (2010), 643 -- 688. [27] E.C. Lance, Hilbert C-modules. A toolkit for operator algebraists, London Mathematical Society Lecture Note Series, 210. Cambridge University Press, Cambridge, 1995. [28] N.S. Larsen, Crossed products by abelian semigroups via transfer operators, Ergodic Theory Dynam. Systems 30:4 (2010), 1147 -- 1164. [29] S. Neshveyev and E. Størmer, Dynamical entropy in operator algebras, Springer-Verlag (2006). 42 E.T.A. KAKARIADIS [30] M.V. Pimsner, A class of C-algebras generalizing both Cuntz-Krieger algebras and crossed products by Z, Free probability theory (Waterloo, ON, 1995), 189 -- 212, Fields Inst. Commun., 12, Amer. Math. Soc., Providence, RI, 1997. [31] C. Pinzari, Y. Watatani and K. Yonetani, KMS States, entropy and the variational principle in full C-dynamical systems, Comm. Math. Phys. 213:2 (2000), 331 -- 379. [32] I. Raeburn, A. Sims, and T. Yeend, The C-algebras of finitely aligned higher-rank graphs, J. Funct. Anal. 213:1 (2004), 206 -- 240. [33] A. Sims and T. Yeend, Cuntz-Nica-Pimsner algebras associated to product systems of Hilbert bimodules, J. Operator Theory 64:2 (2010), 349 -- 376. [34] D.-V. Voiculescu, Dynamical approximation entropies and topological entropy in operator algebras, Comm. Math. Phys. 170 (1995), 249 -- 281. School of Mathematics, Statistics and Physics, Newcastle University, Newcastle upon Tyne, NE1 7RU, UK E-mail address: [email protected]
1010.0879	1	1010	2010-10-05T13:30:50	On Bost-Connes type systems and Complex Multiplication	[ "math.OA", "math.NT" ]	By using the theory of Complex Multiplication for general Siegel modular varieties we construct arithmetic subalgebras for BC-type systems attached to number fields containing a CM field. Our approach extends the construction of Connes, Marcolli and Ramachandran given in the case of imaginary quadratic fields.	math.OA	math	On Bost-Connes type systems and Complex Multiplication Bora Yalkinoglu Abstract By using the theory of Complex Multiplication for general Siegel mod- ular varieties we construct arithmetic subalgebras for BC-type systems attached to number fields containing a CM field. The abelian exten- sions obtained in this way are characterized by results of [Wei]. Our approach is based on a general construction of BC-type systems of Ha and Paugam [HP05] and extends the construction of the arithmetic sub- algebra of Connes, Marcolli and Ramachandran [CMR05] for imaginary quadratic fields. Contents 1 Introduction 2 On the arithmetic subalgebra for K = (i) 3 Two Shimura data and a map 4 About arithmetic Modular functions 5 On Bost-Connes-Marcolli systems 6 Two BCM pairs and a map 7 Construction of our arithmetic subalgebra A Algebraic Groups B CM fields C The Serre group D Shimura Varieties 2 8 17 20 25 28 33 35 37 38 41 2010 Mathematics Subject Classification: 58B34, 11R37, 14G35 Key words: Bost-Connes type systems, Complex Multiplication, Shimura Va- rieties 1 1 Introduction In their fundamental article [BC95] Bost and Connes constructed a quantum statistical mechanical system, the so called Bost-Connes or BC system, that recovers (among other arithmetic properties of ) the explicit class field theory of . It is a natural question to ask for other quantum statistical mechanical systems that recover - at least partially - the (explicit) class field theory of number fields different from . This paper is a contribution to this question in the case of CM fields. In nature a CM field arises as (rational) endomorphism ring of an Abelian variety with Complex Multiplication. Background A quantum statistical mechanical system A = (A, (σt)t∈R) is a C∗-algebra A together with a one-parameter group of automorphisms (σt)t∈R. Sometimes we call A simply a C∗-dynamical system. One should think of A as the alge- bra of observables of a quantum physical system with the one-parameter group (σt)t∈R implementing the time evolution of the observables. A symmetry of A is a C∗-automorphism of A which commutes with the time evolution. Inter- esting properties of A can be read off its equilibrium states at a temperature T ∈ [0, ∞). An equilibrium state at inverse temperature β = 1 T is given by a so called KM Sβ-state. See [BR79] and [BR81] for more information. Now a first step towards generalizing the Bost-Connes system to other num- ber fields is given by a system of the following kind (see [CMR06] and [LLN09]) A BC-type system for a number field K is a quantum statistical mechani- cal system A = (A, (σt)t∈R) with the following properties (i) The partition function of A is given by the Dedekind zeta function of K. (ii) The quotient of the idele class group CK by the connected component DK of the identity of CK acts as symmetries on A. (iii) For each inverse temperature 0 < β ≤ 1 there is a unique KM Sβ-state. (iv) For each β > 1 the action of the symmetry group CK/DK on the set of extremal KM Sβ-states is free and transitive. Remark 1. i) Using class field theory we see that property (2) simply states that the Galois group Gal(K ab/K) ∼= CK/DK of the maximal abelian extension K ab of K is acting by symmetries on A. ii) Properties (3) and (4) say that there is a spontaneous symmetry breaking phenomenon at β = 1 (see [BC95] or pp. 400 in the book [CM08] of Connes and Marcolli). 2 Thanks to the work of Ha and Paugam [HP05] BC-type systems are known to exist for an arbitrary number field K, we denote their solution (see the next section for the definition) by AK = (AK , (σt)t∈R). (1) The paper of Laca, Larsen and Neshveyev [LLN09] gives a different description of AK. In fact [HP05] yields that AK fulfills the first two properties and [LLN09] verifies the last two properties. We should mention that the approach of [HP05] works much more general, it al- lows to attach a quantum statistical mechanical system to an arbitrary Shimura variety; the above construction being a special case. The Shimura-theoretic approach of [HP05] is based on the foundational work of Connes, Marcolli and Ramachandran (see [CMR05] and [CMR06]). The notion of an arithmetic subalgebra for a BC-type system, first encoun- tered in [BC95], hints to a possible relation between seemingly unrelated areas, namely it asks for a connection between quantum statistical mechanical systems (physics/operator algebras) and class field theory (number theory). More pre- cisely (see [CMR06]) A BC-system for a number field K is a BC-type system A = (A, (σt)t∈R) such that (v) There is a K-rational subalgebra Aarith of A, called arithmetic subalgebra of A, such that for every extremal KM S∞-state and every f ∈ Aarith we have (f ) ∈ K ab (2) and further K ab is generated over K in this manner, i.e. K ab = K((f ) extremal KM S∞-state, f ∈ Aarith). (3) (vi) Let ν ∈ CK be a symmetry of A, an extremal KM S∞-state of A and f ∈ Aarith and denote further by [ν] ∈ Gal(K ab/K) the image of ν under Artin's reciprocity morphism CK → Gal(K ab/K). If we denote the action (given by pull-back) of ν on by ν, we have the following compatibility relation ν(f ) = [ν]−1((f )) ∈ K ab. (4) Now the difficulty of constructing BC-systems comes from its relation with Hilbert's 12th problem, which asks for an explicit class field theory of a number field K. This problem is completely solved only in two cases, namely the case of K = and the case of K equal to an imaginary quadratic field, like K = (i). It does not come as a surprise that BC-systems are so far known to exist only 3 in these two cases. For K = see [BC95] and for the imaginary quadratic case we refer to [CMR05] and the very detailed exposition given in [CM08] pp. 551. The construction in [CMR05] is based on the theory of Complex Multiplication (see section 4) which is a part of the arithmetic theory of Shimura Varieties. This theory allows to construct explicitly abelian extensions of higher dimen- sional generalizations of imaginary quadratic number fields, so called CM fields. A CM field E is a totally imaginary quadratic extension of a totally real number field. So for example cyclotomic number fields (ζn), ζn being a primitive n-th root of unity, are seen to be CM-fields. Abelian extensions of a CM field E are obtained by evaluating arithmetic Mod- ular functions on so called CM-points on a Siegel upper half plane (see section 4.1 for more information). Except for the case of an imaginary quadratic field, it is unfortunately not pos- sible to generate the maximal abelian extension Eab of a CM field E in this way, but still a non trivial abelian extension of infinite degree over E. We denote the latter abelian extension of E by Ec ⊂ Eab. (5) See Thm. 4.2 for the characterization of Ec given in [Wei]. As we remarked above for E equal to an imaginary quadratic field there is an equality Ec = Eab. Due to the lack of a general knowledge of the explicit class field theory of an arbitrary number field K we content ourself with the following weaking of a BC-system: Let F ⊂ K ab be an arbitrary abelian extension of K. A partial BC-system for the extension F/K is defined like a BC system A = (A, (σt)t∈R) for K except that we do not demand the arithmetic subalgebra Aarith of A to generate the maximal abelian extension K ab of K but instead the abelian extension of F , i.e. in property (v) we replace K ab by F and call the corresponding K-rational subalgebra Aarith a (partial) arithmetic subalgebra of A. Statement of our result Now the aim of our paper is to prove the following Theorem 1.1. Let K be a number field containing a CM field. Denote by E the maximal CM field contained in K and define the abelian extension K c of K to be the compositum K c = K · Ec (6) (with the notation from (5)). Then the BC-type system AK of [HP05] (see (6.1)) is a partial BC-system for the extension K c/K, i.e. there exists a (partial) arithmetic subalgebra Aarith for AK . K 4 We want to explain our construction of Aarith (cf. section 7). It is inspired by the construction given in [CMR05] and [CMR06] (see also pp. 551 [CM08]). K Idea of our construction Let (G, X, h) be a Shimura datum and denote by Sh(G, X, h) the associated Shimura variety. In [HP05] the authors associate to the datum (G, X, h), the variety Sh(G, X, h) and some additional data a quotient map U −→ Z (7) between a topological groupoid U and a quotient Z = Γ\U for a group Γ. Out of this data they construct a C∗-dynamical system A = (A, (σt)t∈R). A dense ∗-subalgebra H of A is thereby given by the compactly supported, continuous functions H = Cc(Z) (8) on Z, where the groupoid structure of U induces the ∗-algebra structure on H. Moreover there are two variations of the above quotient map which can be put into the following (commutative) diagram (9) U U + U ad / Z Z + / Z ad. For more information see section 5. We will apply this general procedure in two special cases. For this we fix a number field K together with a maximal CM subfield E. I) The BC-type system AK The (0-dimensional) Shimura datum SK = (T K, XK, hK) gives rise to a quotient map denoted by UK −→ ZK. In this case the groupoid UK is of the form (cf. (13)) The associated C∗-dynamical system is denoted by UK = T K(Af ) ⊞ (bOK × Sh(T K, XK, hK)). AK = (AK , (σt)r∈R). (10) 5 / O O / / O O / It gives rise to a BC-type system for K. For the precise definition of AK and its properties (e.g. symmetries, extremal KM S∞-states) we refer the reader to section 3.1 and 6.1. Moreover we denote by HK the dense subalgebra of AK given by HK = Cc(ZK). (11) II) The Shimura system ASh To the CM field E we associate the Shimura datum SSh = (GSp(VE, ψE), H± g , hcm) where in fact the construction of the morphism hcm takes some time (see 3.2). This is due to two difficulties that arise in the case of a general CM field E which are not visible in the case of imaginary quadratic fields. On the one hand one has to use the Serre group SE and on the other hand in general the reflex field E∗ of a CM field E is not anymore equal to E (see B.3). We denote the associated quotient map by USh −→ ZSh and analogously its variations (see 6.2.2). Here the relevant groupoids in hand are of the following form USh = GSp(Af ) ⊞ (ΓSh,M × Sh(GSp(VE, ψE), H± g , hcm)) and U ad Sh = GSpad( )+ ⊞ (Γad Sh,M × Hg). We denote the resulting C∗-dynamical system by ASh and call it Shimura system (cf. section 5.3). Remark 2. In the case of an imaginary quadratic field K the Shimura system ASh gives rise to the GL2-system of Connes and Marcolli [CM08]. The second system is of great importance for us because of the following: Denote by xcm ∈ Hg the CM-point associated with hcm and denote by Mcm the ring of arithmetic Modular functions on Hg defined at xcm. By the theory of Complex Multiplication we know that for every f ∈ Mcm we have (cf. (6) and 4.3.1) f (xcm) ∈ K c ⊂ K ab and moreover K c is generated in this way. Our idea is now that Mcm gives rise to the arithmetic subalgebra Aarith K . More precisely we will construct a (commutative) diagram (see 6.3) (12) UK USh ZK / ZSh 6 / / / which is induced by a morphism of Shimura data SK → SSh constructed in section 3.3. Then, using criterion 5.1, we see that the morphism Z + diagram) is invertible and obtain in this way a continuous map Sh → ZSh (from the above Θ : ZK −→ ZSh −→ Z + Sh −→ Z ad Sh. Using easy properties of the automorphism group of Mcm, we can model each Sh (which might have singularities). f in Mcm as a function ef on the space Z ad Nevertheless in Prop. 7.1 we see that for every f ∈ Mcm the pull back ef ◦ Θ as the K-algebra generated by lies in HK = Cc(ZK) and we can define Aarith these elements, i.e. K Aarith K = < ef ◦ Θ f ∈ Mcm >K . Now, using the classification of extremal KM S∞-states of AK (see section 6.1.5), the verification of property (5′) is an immediate consequence of our con- struction and property (6) follows by using Shimura's reciprocity law and the observation made in Prop. 4.4 (see section 7.1 for the details). Our paper is organized along the lines of this section, recalling on the way the necessary background. In addition we put some effort in writing a long Appendix which covers hopefully enough information to make this paper "read- able" for a person which has little beforehand knowledge of the arithmetic theory of Shimura varieties. Acknowledgment The author would like to thank James Milne, Sergey Neshveyev and Frederic Paugam for helpful and valuable comments and his advisor Eric Leichtnam for his guidance through this first project of the author's Phd thesis. This work has been supported by the Marie Curie Research Training Network in Noncommutative Geometry, EU-NCG. Notations and conventions We use the commen notations N, Z, , R, C. If A denotes a ring or monoid, we denote its group of multiplicative units by A×. A number field is a finite extension of . The ring of integers of a number field K is denoted by OK. We denote by AK = AK,f × AK,∞ the adele ring of K (with its usual topology), where AK,f denotes the finite adeles and AK,∞ the infinite adeles of K. AK contains K by the usual diagonal embedding and by bOK we denote the closure of OK in AK,f . Invertible adeles are called ideles. The idele class group A× of K is denoted by CK, its connected component of the identity by DK. We fix an algebraic closure Q of in C. Usually we think of a number field K as lying in C by an embedding τ : K → Q ⊂ C. Complex conjugation on C is denoted by ι. Sometimes we write zι for the complex conjugate of a complex K/K × 7 number z. Artin's reciprocity map A× K → Gal(K ab/K) : ν 7→ [ν] is normalized such that an uniformizing parameter maps to the arithmetic Frobenius element. Further given a group G acting partially on a set X we denote by G ⊞ X = {(g, x) ∈ G × X gx ∈ X} (13) the corresponding groupoid (see p. 327 [LLN09]). If X denotes a topological space we write π0(X) for its set of connected compo- nents. 2 On the arithmetic subalgebra for K = (i) Before we describe our general construction we will explain the easiest case K = (i), where many simplifications occur, in some detail and point out the modifications necessary for the general case. For the reminder of this section K always denotes (i), although many of the definitions work in general. For the convenience of the reader we will try to make the following section as self- contained as possible. 2.1 The quotient map UK → ZK We denote by T K the -algebraic torus given by the Weil restriction T K = ResK/ (Gm,K) of the multiplicative group Gm,K, i.e. for a -algebra R the R-points of T K are given by T K(R) = (R ⊗ K)×. In particular we see that T K( ) = K ×, T K(Af ) = A× K,∞. In our special case we obtain that after extending scalars to R the R-algebraic group T K is isomor- phic to S = ResC/R(Gm,R). Further the finite set XK = T K(R)/T K(R)+ = π0(T K(R)) consists in our case of only one point. With this in mind we consider the 0-dimensional Shimura datum (see D.5) K,f and T K(R) = A× R SK = (T K, XK, hK) (14) where the morphism hK : S → T K T K The (0-dimensional) Shimura variety Sh(SK) is in our case of the simple form ∼= S). (In the general case hK is chosen accordingly to Lemma 3.1.) is simply given by the identity (thanks to R R Sh(SK) = T K( )\(XK × T K(Af )) = K ×\A× K,f . (15) We write [z, l] for an element in Sh(SK) meaning that z ∈ XK and l ∈ T K(Af ). (For general number fields the description of Sh(SK) is less explicit but no difficulty occurs.) Remark 3. The reader should notice that by class field theory we can identify Sh(SK) = K ×\A× K,f with the Galois group Gal(K ab/K) of the maximal abelian extension K ab of K. This is true in general, see section 6.1.4. 8 The (topological) groupoid UK underlying the BC-type system AK = (AK, (σt)t∈R) is now of the form (see (13) for the notation) UK = T K(Af ) ⊞ (bOK × Sh(SK)) with the natural action of T K(Af ) on Sh(SK) (see D.2) and the partial action of T K(Af ) = A× The group K,f on the multiplicative semigroup bOK ⊂ AK,f by multiplication. (16) (17) is acting on UK as follows Γ2 K = bO× K × bO× K (γ1, γ2)(g, ρ, [z, l]) = (γ−1 1 gγ2, γ2ρ, [z, lγ−1 2 ]), (18) where γ1, γ2 ∈ bO× quotient map K, g, l ∈ T K(Af ), ρ ∈ bOK and z ∈ XK and we obtain the UK −→ ZK = Γ2 K\UK. (19) In the end of this section we will construct the arithmetic subalgebra Aarith of the BC-type system AK which is contained in HK = Cc(ZK) ⊂ AK. For this we will need K 2.2 The quotient map USh −→ ZSh In our case of K = (i) the maximal CM subfield E of K is equal to K. The Shimura datum SSh associated with E is of the form (see 3.2) SSh = (GSp(VE, ψE), H±, hcm). (20) Here GSp = GSp(VE, ψE) is the general symplectic group (cf. D.3) associated with the symplectic vector space (VE, ψE). The latter is in general chosen accordingly to (55). Due to the fact that the reflex field E∗ (cf. B.3) is equal to E and the Serre group SE is equal to T E = T K we can simply choose the -vector space VE to be the -vector space E and the symplectic form ψE : E × E → to be the map (x, y) 7→ T rE/ (ixyι). A simple calculation shows that ψE(f (x), f (y)) = det(f )ψE(x, y), for all f ∈ End (VE) and all x, y ∈ VE, therefore we can identify GSp with GL2 = GL(VE). Now again using the fact that the Serre group SE equals T E we see that the general construction of hcm : S = T E → GSpR = GL2,R (see (59)) is given on the R-points by a+ib ∈ C× = S(R) →(cid:18) a −b a (cid:19) ∈ GL2(R). Each α ∈ GSp(R) defines a map α−1hcmα : S → GSpR given on the R-points by a + ib ∈ C× 7→ α−1hcm(a + ib)α ∈ GL2(R) and the GSp(R)-conjugacy class X = {α−1hcmα α ∈ GSpR(R)} of hcm can be identified with the Siegel upper lower half space H± = C − R by the map b R α−1hcmα ∈ X 7→ (α−1hcm(i)α) · i ∈ H±, 9 where the latter action · denotes Moebius transformation. Under this identifi- cation the morphism hcm corresponds to the point xcm = i on the upper half plane H. The point xcm ∈ H is a so called CM-point (see D.6). Remark 4. The definition of a CM-point and the observation made in (58) explain the need of using the Serre group in the general construction of hcm. The explanation given in 4.1 shows in particular why we have to define the vector space VE in general accordingly to (55). The Shimura variety Sh(SSh) is of the nice form (cf. (136)) Sh(SSh) = GSp( )\(H± × GSp(Af )). Again we write elements as [z, l] ∈ Sh(SSh) with z ∈ H± and g ∈ GSp(Af ). In our case the topological groupoid USh underlying the Shimura system ASh is given by (cf. 6.2.1) USh = GSp(Af ) ⊞ (M2(bZ) × Sh(SSh)), (21) where GSp(Af ) = GL2(Af ) is acting in the natural way on Sh(SSh) and par- tially on the multiplicative monoid of 2 × 2-matrices M2(bZ) ⊂ M2(Af ) with entries in bZ. The group (22) Γ2 Sh = GL2(bZ) × GL2(bZ) is acting on USh exactly like in (18) and induces the quotient map USh −→ ZSh = Γ2 Sh\USh. (23) Remark 5. Note that the quotient ZSh is not a groupoid anymore (see top of p. 251 [HP05]). In our example it is sufficient to consider the positive groupoid U + 6.2.2) associated with USh. (In the general case the adjoint groupoid U ad to be more appropriate.) It is given by Sh (see Sh seems together with the group U + Sh = GSp( )+ ⊞ (M2(bZ) × H) (Γ+ Sh)2 = GL2(Z)+ × GL2(Z)+ = SL2(Z) × SL2(Z) acting by (γ1, γ2)(g, ρ, z) = (γ1gγ−1 2 , γ2ρ, γ2z) and inducing the quotient map U + Sh −→ Z + Sh = (Γ+ Sh)2\U + Sh. 10 (24) (25) (26) (27) By construction GSp(R) is acting (free and transitively) on H± and GSp(R)+, the connected component of the identity, can be thought of as stabilizer of the upper half plane H+ = H which explains the action of GSp( )+ = GSp( ) ∩ GSp(R)+ on H. Thanks to criterion 5.1 we know that the natural (equivariant) morphism of topological groupoids U + Sh → USh given by (g, ρ, z) 7→ (g, ρ, [z, 1]) induces a homeomorphism on the quotient spaces Z + Sh −→ ZSh in the commu- tative diagram USh U + Sh / ZSh ∼= Z + Sh. (28) Remark 6. The groupoid U + Marcolli (see [CM08] and 5.8 [HP05]) Sh corresponds to the GL2-system of Connes and 2.3 A map relating UK and USh We want to define an equivariant morphism of topological groupoids UK −→ USh, where equivariance is meant with respect to the actions of ΓK on UK and ΓSh on USh. For this it is necessary (and more or less sufficient) to con- struct a morphism of Shimura data between SK = (T K, XK, hK) and SSh = (GSp(VE, ψE)), H±, hcm), which is given by a morphism of algebraic groups ϕ : T K −→ GSp (29) such that hcm = ϕR ◦ hK. In our case the general construction of ϕ, stated in (60), reduces to the simple map (on the -points) a + ib ∈ K × = T K( ) 7→(cid:18) a −b a (cid:19) ∈ GL2( ) = GSp( ) b and we see in fact that after extending scalars to R the morphism ϕR : T K = S → GSpR is already equal to hcm : S → GSpR. The simplicity of our example comes again from the fact that we don't have to bother about the Serre group, which makes things less explicit, although the map ϕ : T K → GSp still has a quite explicit description even in the general case thanks to Lemma 3.3. R Now by functoriality (see section D.5) we obtain a morphism of Shimura va- rieties Sh(ϕ) : Sh(SK) → Sh(SSh) which can be explicitly described by [z, l] ∈ K ×\(XK × T K(Af )) 7→ [xcm, ϕ(Af )(l)] ∈ GSp( )\(H± × GSp(Af )). In the general case we have essentially the same description (see (105)), the point being that every element z in XK is mapped to xcm ∈ H±, as in the 11 / / / O O O O general case. Using bOK = bZ ⊗Z OK we can continue the map ϕ(Af ) to a morphism of (topological) semigroups M (ϕ)(Af ) : bOK → M2(bZ) by setting n ⊗ (a + ib) 7→ (cid:18) an −bn an (cid:19). By continuation we mean that ϕ(Af ) and M (ϕ)(Af ) agree on the intersection of T K(Af ) ∩ bOK ⊂ AK,f . In the general case the explicit description of ϕ given in (62) is used to continue ϕ to M (ϕ) (see 6.3.1), the above example being a special case. Now it can easily be checked that bn (g, ρ, [z, l]) ∈ UK 7→ (ϕ(Af )(g), M (ϕ)(Af )(ρ), [xcm, ϕ(Af )(l)]) ∈ USh (30) defines the desired equivariant morphism of topological groupoids Summarizing we obtain the following commutative diagram (using (28)) UK −→ USh. UK USh U + Sh ZK / ZSh ∼= / Z + Sh (31) (32) which gives us the desired morphism of topological spaces Θ : ZK −→ ZSh −→ Z + Sh. (33) In general we have to go one step further and use the adjoint groupoid Z ad Sh (cf. 6.2.2) but this only due to the general description of the automorphism group Aut (M) of the field of arithmetic automorphic functions M, see 4.3.1 and the next section for explanations. 2.4 Interlude: Theory of Complex Multiplication In this section we will provide the number theoretic background which is nec- essary to understand the constructions done so far. We are interested in constructing the maximal abelian extension Eab of E = K = (i) and there are in general two known approaches to this problem. I) The elliptic curve A : y2 = x3 + x Let us denote by A the elliptic curve defined by the equation A : y2 = x3 + x. (34) 12 / / / O O / O O It is known that the field of definition of A and its torsion points generate the maximal abelian extension Eab of E. (By field of definition of the torsion points of A we mean the coordinates of the torsion points.) Notice that in our case the complex points of our elliptic curve A are given by A(C) = C/OE and the rational ring of endomorphisms of A turns out to be End(A(C)) = ⊗Z OE = E,we say that A has complex multiplication by E. Remember that OE = Z[i]. Remark 7. To obtain the abelian extensions of K provided by A explicitly one may use for example the Weierstrass p-function associated with A (see [Sil94]) but as we will use another approach we don't want to dive into this beautiful part of explicit class field theory. II) The (Siegel) Modular curve Sh(GSp, H±, hcm) We want to interpret the Shimura variety Sh(SSh) = Sh(GSp, H±, hcm) con- structed in the last section as moduli space of elliptic curves with torsion data. The moduli theoretic picture For this we consider the connected component Sho = Sh(SSh)o of our Shimura variety which is described by the projective system (see D.4 and [Mil04]) Sho = lim←− N Γ(N )\H, (35) where Γ(N ), for N ≥ 1, denotes the subgroup of Γ = Γ(1) = SL2(Z) defined by Γ(N ) = {g ∈ Γ g ≡(cid:18) 1 0 0 1 (cid:19) mod N }. We can view the quotient Γ(N )\H as a complex analytic space, but thanks to the work of Baily and Borel, it carries also a unique structure of an algebraic variety over C (see [Mil04]). We will use both viewpoints. Seen as an analytic space we write H(N ) = Γ(N )\H and for the algebraic space we write Sho N = Γ(N )\H. Observe now that the space H(N ) classifies isomorphism classes of pairs (A, t) given by an elliptic curve A over C together with a N -torsion point t of A. In particular H(1) = Γ\H classifies isomorphism classes of elliptic curves over C. In this picture our CM-point [xcm]1 = [i]1 ∈ H(1) corresponds to the iso- morphism class of the elliptic curve A from (34) and more general the points [xcm]N ∈ H(N ) capture the field of definition of A and its various torsion points and recover the maximal abelian extension Eab of E in this way! By [z]N we denote the image of z ∈ H in H(N ) under the natural quotient map H → H(N ). Remark 8. For the relation between Sh and Sho we refer the reader to pp. 51 [Mil04]. 13 The field of arithmetic Modular functions M To construct the abelian extensions provided by the various points [xcm]N ex- plicitly we proceed as follows: We consider the connected canonical model M o of Sho (see D.7) which provides us with an algebraic model M o N = Γ(N )\M o of the algebraic variety Sho In general we N is an algebraic obtain algebraic models over subfields of variety defined over the cyclotomic field (ζN ) and after scalar extension to C it becomes isomorphic to the complex algebraic variety Sho N . Let us denote by k(M o N , in particular this means elements in k(M o N ) are rational over (ζN ). It makes sense to view the point [xcm]N as a point on M o N ) is defined at [xcm]N , then we know (cf. 4.3.1) N over the cyclotomic field (ζN ). ab. This means M o N ) the field of rational functions on M o N and if a function f ∈ k(M o f ([xcm]N ) ∈ Eab. (36) N ) on the complex algebraic variety Sho In particular varying over the various N and the rational functions in k(M o N ) the values f ([xcm]N ) generate Eab over E. The next step is to realize that the function field k(M o N ) can be seen as subset of the field of rational functions k(Sho N (cf. 4.3.1). As rational functions in k(Sho N ) correspond to meromorphic functions on H(N ) and meromorphic functions on H(N ) are nothing else than meromorphic functions on H that are invariant under the action of Γ(N ) we can view each rational function in k(M o N ) as a meromorphic function on H which is invariant under Γ(N ). If we denote by k(M o N ) that give rise to meromorphic functions on H that are meromorphic at the cusps (see 4.3.1), it makes sense to define the field of meromorphic functions M on H given by the union N ) consisting of functions f ∈ k(M o N )cusp the subfield of k(M o M =[N k(M o N )cusp. (37) Due to (36) we know furthermore that for every f ∈ M, which is defined in xcm, we have f (xcm) ∈ K ab = Eab (38) and K ab is generated in this way. Therefore we call the field M the field of arithmetic Modular functions. Explicitly M is described for example in [Shi00] or [CM08] (Def. 3.60). A very famous arithmetic Modular function is given by the j-function which generates the Hilbert class field of an arbitrary imaginary quadratic field. Remark 9. 1) In light of 4.1 and 3.2 2) we mention that the field of definition E(xcm) of xcm is in our example equal to E = (i), which is the reason why our example is especially simple. 2) If we take a generic meromorphic function g on H(N ) that is defined in [xcm]N then the value g([xcm]N ) ∈ C will not even be algebraic. This is the reason why we need the canonical model M o which provides an artithmetic structure for the field of meromorphic functions on H(N ). 14 In the general case the construction of M is quite similar to the construction above (see section 4), the only main difference being that in general the theory is much less explicit (e.g. the description of M). Automorphisms of M and Shimura's reciprocity law In our example K = E = (i), using the notation from 4.3, we have the equality E = GSp(Af ) and obtain a group homomorphism × GSp(Af ) −→ E −→ Aut (M), (39) where the first arrow is simply the projection and the second arrow comes from 4.3.2. We denote the action of α ∈ GSp(Af ) on a function f ∈ M by αf . In particular we see that α ∈ GSp( )+ = SL2( ) is acting by αf = f ◦ α−1, (40) where α−1 acts on H by Mobius transformation. The adjoint system, which is well suited for the general case (cf. 6.2.2), is not needed in our special ex- ample. We note that the group GSpad( )+ occuring in (6.2.2) is given by SL2( )/{±I}, where I denotes the unit in SL2( ). Due to the fact that {±I} acts trivially on H we can lift the action to GSp( )+. Further there is a mor- phism of algebraic groups η : T K → GSp which induces a group homomorphism denoted by (compare (75)) η = η(Af ) : T K(Af ) → GSp(Af ). (41) (42) The reciprocity law of Shimura can be stated in our special case as follows: Let ν be in A× K,f , denote by [ν] ∈ Gal(K ab/K) its image under Artin's reci- procity map and let f ∈ M be defined in xcm ∈ H. Then η(ν)f is also defined in xcm ∈ H and η(ν)f (xcm) = [ν]−1f (xcm) ∈ K ab. (43) The formulation in the general case concentrates on the group E (see 4.4). Now we are ready for the 2.5 Construction of the arithmetic subalgebra Define Mcm to be the subring of functions in M which are defined at xcm ∈ H. (44) For every f ∈ Mcm we define a function ef on the groupoid U + ρ ∈ M2(bZ) − GSp(bZ) ef (g, ρ, z) =( ρf (z) ρ ∈ GSp(bZ) Sh by 0 15 Thanks to (40) ef is invariant under the action of Γ+ and therefore ef descends to the quotient Z + further that the pull-back ef ◦ Θ (see (33)) defines a compactly supported, con- tinuous function on ZK, i.e. we have ef ◦ Θ ∈ HK = Cc(ZK) ⊂ AK. Therefore Sh = SL2(Z) × SL2(Z) Sh (cf. (27)). Proposition 7.1 shows we can define the K-subalgebra Aarith of HK generated by these elements K Aarith K =< ef ◦ Θ f ∈ Mcm >K . Now we want to show that Aarith (see (2)). The set E∞ of extremal KM S∞-states of AK is indexed by the set Sh(SK) and for every ω ∈ Sh(SK) the corresponding KM S∞-state ω is given on an element f ∈ HK = Cc(ZK) by (cf. 6.1.5) is indeed an arithmetic subalgebra for AK K ω(f ) = f (1, 1, ω). (46) Using remark 3 we write [ω] for an element ω ∈ Sh(SK) when regarded as element in Gal(K ab/K). Now if we take a function f ∈ Mcm and ω ∈ Sh(SK) we immediately see ω(ef ◦ Θ) = [ω]−1(f (xcm)) ∈ K ab and property (v) follows (in general we will only show property (v') of course). To show property (vi) we take a symmetry ν ∈ CK = A× K/K × (see 6.1.4) of AK and denote by [ν] ∈ Gal(K ab/K) its image under Artin's reciprocity homomorphism and let f and ω be as above. We denote the action (pull-back) of ν on ω by νω and obtain (45) (47) (48) (49) (50) But thanks to our observation Proposition 4.4 we know that νω(ef ◦ Θ) = ϕ(Af )(ν)f (xcm). ϕ(Af )(ν) = η(ν) ∈ GSp(Af ) and now thanks to Shimura's (43) we can conclude νω(ef ◦ Θ) = η(ν)f (xcm) = [ν]−1(f (xcm)) = [ν]−1(ω(ef ◦ Θ)) ∈ K ab which proves property (vi). For the general case and more details we refer the reader to 7.1. Remark 10. Our arithmetic subalgebra AK in the case of K = (i) is essentially the same as in [CMR05]. Remark 11. In a fancy (and very sketchy) way we might say that the two different pictures, one concentrating on the single elliptic curve A the other on the moduli space of elliptic curves (see the beginning of section 2.4), are related via the Langlands correspondence. In terms of Langlands correspondence, the single elliptic curve A lives on the motivic side whereas the moduli space of 16 elliptic curves lives (partly) on the automorphic side. As we used the second picture for our construction of an arithmetic subalgebra, we might say that our construction is automorphic in nature. This explains the fact that we have a "natural" action of the idele class group on our arithmetic subalgebra (see above). Using the recent theory of endomotives (see chapter 4 and in particular p. 551 [CM08]) one can recover the arithmetic subalgebra Aarth K by only using the single elliptic curve A. In particular one obtains a natural action of the Galois group. This and more will be elaborated in another paper. We now concentrate on the general case. 3 Two Shimura data and a map As throughout the paper let K denote a number field and E its maximal CM subfield. We fix an embedding τ : K → Q → C and denote complex conjugation on C by ι. To K, resp. E, we will attach a Shimura datum SK, resp. SSh, and show how to construct a morphism ϕ : SK → SSh between them. We will freely use the Appendix: every object not defined in the following can be found there or in the references given therein. Recall that the Serre group attached to K is denoted by SK (cf. C.1), it is a quotient of the algebraic torus T K (defined below), the corresponding quotient map is denoted by πK : T K → SK. 3.1 Protagonist I: SK The 0-dimensional Shimura datum SK = (T K, XK, hK) (see section D.5) is given by the Weil restriction T K = ResK/ (Gm,K), the discrete space XK = T K(R)/T K(R)+ and a morphism hK : S = ResC/R(Gm,C) → T K which is chosen accordingly to the next R Lemma 3.1. There is a morphism of algebraic groups hK : S → T K the diagram R such that S hK / hK ???????? T K R πK R SK (51) commutes. R Proof. Remember that hK : S → SK R is defined as the composition ResC/R(µK ) S / ResC/R(SK C ) N mC/R / SK R , (52) 17 / / / where µK : Gm,C → T K T E simply by C is defined by µK = πK C ◦ µτ (cf. C.2). Define hK : S → R ResC/R(µτ ) S / ResC/R(T K C ) N mC/R / T K R . (53) For proving our claim it is enough to show that the following diagram ResC/R(µτ ) S +VVVVVVVVVVVVVVVVVVVVVVVV ResC/R(µK ) / ResC/R(T K C ) N mC/R T K R (54) ResC/R(πK C ) ResC/R(SK C ) N mC/R πK R / SK R is everywhere commutative. The triangle on the left is commutative, because ResC/R is a functor. Thanks to theorem A.1 it is enough to show that rectangle on the right is commutative after applying the functor X ∗ (cf. A.3). Since πK : T K → SK is defined to be the inclusion X ∗(SK) ⊂ X ∗(T K) on the level of characters (cf. C.1), we see that X ∗(ResC/R(πK ) are inclusions as well and the commutativity follows. C )) and X ∗(πK R 3.2 Protagonist II: SSh The construction of the Shimura datum SSh in this section goes back to Shimura [Shi00], see also [Wei]. It is of the form SSh = (GSp(VE, ψE), H± g , hcm). The symplectic -vector space (VE, ψE) is defined as follows. Choose a finite collection of primitive CM types (Ei, Φi), 1 ≤ i ≤ r, such that i) for all i the reflex field E∗ i ii) the natural map (take (119) and apply the universal property from C.1) is contained in E, i.e. ∀i E∗ i ⊂ E, and Q NE/E∗ i SE i=1 SE ∗ i /Qr Q ρΦi i=1 T Ei /Qr (55) is injective. (Proposition 1.5.1 [Wei] shows that this is allways possible.) For every i ∈ {1, .., r} we define a symplectic form ψi : Ei × Ei → on Ei by choosing a totally imaginary generator ξi of Ei (over ) and setting ψi(x, y) = T rEi/ (ξixyι). (56) Now we define (VE, ψE) as the direct sum of the symplectic spaces (Ei, ψi). Instead of GSp(VE, ψE) we will sometimes simply write GSp. To define the morphism hcm the essential step is to observe (see Remark 9.2 : SE → T Ei is contained in the [Mil98]) that the image of the map ρΦi ◦ NE/E ∗ subtorus T Ei of T Ei, which is defined on the level of -points by i T Ei( ) = {x ∈ E× i xxι ∈ ×} (57) 18 / / + / / / / / / and analogously T Ei(R) is defined for an arbitrary -algebra R. This is an important observation, because there is an obvious inclusion of algebraic groups (cf. A.2) i : T Ei → GSp(VE, ψE), (58) rYi=1 With this in mind we define hcm as the composition whereas there is in general no embeddingQ T Ei → GSp. /Qr i=1 SE ∗ /Qr Q NE/E∗ i Q ρΦi ,R / SE hE ,R i S R R R cm : S →Qr i=1 T Ei for the composition of the first three arrows. Write h′ Remark 12. 1) By construction hcm is a CM point (cf. D.6) which is needed later to construct explicitly abelian extensions K. See 4.1. 2) Viewed as a point on the complex analytic space H± of hcm. Further we denote the connected component of H± Hg, i.e. xcm ∈ Hg. g we write xcm instead g containing xcm by Our CM point hcm enjoys the following properties i=1 T Ei R i R / GSpR. (59) 2) The field of definition E(xcm) of xcm is equal to the composite of the reflex r ⊂ E, i.e. the associated cocharacter µcm of hcm is defined i=1 hΦi (see (118)). Lemma 3.2. 1) We have hcm = i ◦Qr fields eE = E∗ over eE (see D.6). 1 · · · E∗ 3) The GSp(R)-conjugacy classes of hcm can be identified with the Siegel upper- lower half plane H± g , for some g ∈ N depending on E. Proof. 1) This follows immediately from (131) and (135). 2) This follows from p. 105 [Mil04] and 1). 3) For this we refer to the proof of lemma 3.11 [Wei]. 3.3 The map ϕ : SK → SSh On the level of algebraic groups ϕ : T K → GSp is simply defined as the com- position T K πK / SK NK/E / SE Q NE/E∗ i i=1 SE ∗ i Q ρΦi / i=1 T Ei /Qr /Qr i / GSp. (60) For ϕ being a map between SK and SSh we have to check that the diagram (61) hK / T K S !DDDDDDDDD hcm R ϕR GSpR 19 / / / / / / / / / ! commutes, but this compatibility is built into the construction of hK. Using the reflex norm (cf. B.3 and (133)) we can describe ϕ as follows Lemma 3.3. The map ϕ : T K → GSp is equal to the composition Q NK/E∗ i T K i /Q T E ∗ Q NΦi i=1 T Ei /Qr i / GSp. (62) 4 About arithmetic Modular functions 4.1 Introduction We follow closely our references [Del79], [MS81] and [Wei]. See also [Hid04]. The reader should be aware of the fact that we are using a different normal- ization of Artin's reciprocity map than in [MS81] and have to correct a "sign error" in [Del79] as pointed out on p. 106 [Mil04]. As usual we denote by K a number field containing a CM subfield and de- note by E the maximal CM subfield of K. In this section we want to explain how the theory of Complex Multiplication provides (explicit) abelian extensions of K. In general one looks at a CM-point x ∈ X on a Shimura variety Sh(G, X) and by the theory of canonical models one knows that the point [x, 1] on the canonical model M (G, X) of Sh(G, X) is rational over the maximal abelian ex- tension E(x)ab of the field of definition E(x) of x (see D.6). In our case we look at Siegel modular varieties Sh(GSp, H± g ) which can be considered as (fine) moduli spaces of Abelian varieties over C with additional data (level structure, torsion data and polarization). See chap. 6 [Mil04] for an explanation of this. Each point x ∈ Hg corresponds to an Abelian variety Ax. In opposite to the case of imaginary quadratic fields, in general the field of defi- nition E(x) is neither contained in E nor in K. Therefore, in order to construct abelian extensions of K, we have to find an Abelian variety Ax such that E(x) ⊂ K. (63) This is exactly the reason for our choice of xcm ∈ Hg because we know (see 3.2 2)) that 1 · · · E∗ r ⊂ E ⊂ K. (64) E(xcm) = eE = E∗ Here xcm corresponds to a product Acm = A1 × · · · × Ar of simple Abelian vari- eties Ai with complex multiplication given by Ei. This construction is the best one can do to generate abelian extensions of K using the theory of Complex Multiplication. The miracle here is again that the field of definition of Acm and of its torsion points generate abelian extensions of E(xcm). Now to obtain these abelian extensions explicitly one proceeds in complete analogy with the case of (i) explained in 2.4, namely rational functions on the connected canonical 20 / / / model M o of the connected Shimura variety Sh(GSp, H± g )o give rise to arith- metic Modular functions on Hg which generate the desired abelian extensions when evaluated at xcm. This will be explained in detail in the following. 4.2 Working over Q 4.2.1 The field F of arithmetic automorphic functions We start with the remark that the reflex field of (GSp, H± to (see remark 32). g ) (cf. D.6) is equal Remark 13. This is the second notion of "reflex field". But the reader shouldn't get confused. Γ\Hg (cf. D.4). Denote by Σ the set of arithmetic subgroups Γ of GSpad( )+ which contain the image of a congruence subgroup of GSpder( ). The connected component of the identity Sho of Sh = Sh(GSp, H± g ) is then given by the inverse limit Sho = lim←−Γ∈Σ g ) the canonical model of Sho in the sense of Denote by M o = M o(GSp, H± 2.7.10 [Del79], i.e. M o is defined over Q. For every Γ ∈ Σ the space Γ\Hg is an algebraic variety over C and Γ\M o a model over . The field of rational functions k(Γ\M o) on Γ\M o is contained in the field of rational functions k(Γ\Hg). Elements in the latter field correspond to mero- morphic functions on Hg (now viewed as a complex analytic space) that are invariant under Γ ∈ Σ. Following [MS81] we call the field F =SΓ∈Σ k(Γ\M o) the field of arithmetic automorphic functions on Hg. 4.2.2 About Aut (F ) The (topological) group E defined by the extension (see 2.5.9 [Del79]) 1 / GSpad( )+ σ / E / Gal( / ) / 1 (65) is acting continuously on M o (2.7.10 [Del79]) and induces an action on F by αf = σ(α) · (f ◦ α−1) = (σ(α)f ) ◦ (σ(α)α−1) (66) (see [MS81] 3.2). This is meaningful because f and α−1 are both defined over Q. Using this action one can proove Theorem 4.1 (3.3 [MS81]). The map E → Aut (F ) given by (66) identifies E with an open subgroup of Aut (F ). 4.3 Going down to ab We said that M o is defined over Q but it is already defined over a subfield k of ab. More precisely k is the fixed field of the kernel of the map Gal(Q/ )ab → 21 / / / / π0π(GSp) defined in 2.6.2.1 [Del79]. Therefore the action of E on M o factors through the quotient E of E defined by the following the commutative diagram with exact rows (see 4.2 and 4.12 [MS81] or 2.5.3 [Del79]) 1 1 1 / GSpad( )+ id / GSpad( )+ id σ σ E E pr τ Gal(Q/ ) res Gal(k/ ) l (67) 1 1 / GSpad( )+ / GSp(Af ) C( ) / π0π(GSp) / 1, here C denotes the center of GSp. × is discrete in GSp(Af ) (cf. [Hid04]). Remark 14. 1) We know that C( ) = 2) Because l is injective, τ is injective as well and we can identify E with an (open) subgroup of GSp(Af ) 3) E is of course depending on K but we suppress this dependence in our nota- tion. . × 4.3.1 The field M of arithmetic Modular functions Let f ∈ F be a rational function, i.e. f is a rational function on Γ\M o, for some Γ ∈ Σ. We call f an arithmetic Modular function if it is rational over k, and meromorphic at the cusps (when viewed on the corresponding complex analytic space). Compare this to 3.4 [Wei] or pp. 35 [Mil98]. Definition 4.1. The subfield of F generated by all arithmetic Modular func- tions is denoted by M. Further we denote by Mcm the subring of M of all arithmetic Modular functions which are defined in xcm. The importance of Mcm for our purposes is explained (see 14.4 [Mil98] and 3.11 [Wei]) by Theorem 4.2. Let Acm denote the abelian variety corresponding to xcm (cf. 4.1). Denote by K Acm the field extension of K obtained by adjoining the field of definition of Acm and all of its torsion points. Further denote by KM the field extension of K obtained by adjoining the values f (xcm), for f ∈ Mcm. Finally denote by K c the composition of K with the fixed field of the image of the Verlagerungsmap V er : Gal(F ab/F ) → Gal(Eab/E), where F is the maximal totally real subfield of E. Then we have the equality K c = K Acm = KM. (68) 22 / / / / / / / / / / / / / / / / / / Remark 15. Notice we are not simply using the field of arithmetic automorphic functions as considered by Shimura, see 4.8 [MS81] for his definition, because the exact size of the abelian extension obtained by using these functions is not clear (at least to the author). It is clear that the field of Shimura is contained in KM and we guess that it should generate the same extension K c of K. 4.3.2 About Aut (M) It is clear that M is closed under the action of E (see 3.2 or 4.4 [MS81]) and therefore we obtain a continuous map given like above by E → Aut (M) αf = σ(α) · (f ◦ α−1) = (σ(α)f ) ◦ (σ(α)α−1). In particular GSpad( )+ is acting on M by αf = f ◦ α−1. (69) (70) (71) 4.4 The reciprocity law at xcm Write µcm : Gm,C / SE C (h′ cm)C i=1 T Ei,C i / GSpC (72) /Qr E∗ i the cocharacter µ′ for the associated cocharacter of hcm (cf. D.6). From Lemma 3.2 1) we know that hcm = i ◦Q hφi and therefore µcm = i ◦Q µφi . Because µφi is defined over simplify the notation set T =Qr cm =Q µφi is defined over eE = E∗ 1 · · · E∗ i=1 T Ei. Define the morphism η : T K → GSp r ⊂ E ⊂ K. To (73) as composition of ResK/ (µ′ cm) T K / ResK/ (TK) N mK/ / T i / GSp. (74) If we identify E with an (open) subset of GSp(Af ) show (see 4.5 [MS81] or 2.6.3 [Del79]) that η(A) : A× ν 7→ η(Af )(ν) mod ×, a group homomorphism × using τ (see (67)), one can K → GSp(A) induces, by η : A× K → E. (75) If we denote by [ν] ∈ Gal(K ab/K) the image of ν ∈ A× map one can show (cf. 4.5 [MS81]) that (cf. (67)) K under Artin's reciprocity σ(η(ν)) = [ν]−1k. (76) 23 / / / / / / Remark 16. The careful reader will ask why it is allowed to define η using the extension K of the field of definition of the cocharacter µcm given by eE, because our reference [MS81] uses eE to define η. The explanation for this is given by lemma 4.5 and standard class field theory. Now we are able to state the reciprocity law Theorem 4.3 (see 4.6 and 4.10 [MS81]). Let ν ∈ A× f (xcm) is rational over K ab. Further η(ν)f is defined in xcm and K and f ∈ Mcm. Then η(ν)f (xcm) = [ν]−1(f (xcm)). (77) Proof. We simply reproduce the argument given in the proof of Thm. 4.6 [MS81]. The first assertion is clear by the definition of the canonical model (cf. D.6) and the other two assertions follow from the following calculation. Regard the special point xcm as a point on the canonical model [xcm, 1] ∈ M o. The action of η(ν)−1 is given by η(ν)−1[xcm, 1] = σ(η(ν)−1)[xcm, η(ν)] and fur- ther we know [xcm, η(ν)] = [ν]−1[xcm, 1] (by (142)). Therefore we obtain η(ν)f (xcm) = σ(η(ν)) · (f ◦ η(ν)−1)([xcm, 1]) = (σ(η(ν)f ) ◦ (σ(η(ν))η(ν)−1)([xcm, 1]) = (σ(η(ν)f ) ◦ (σ(η(ν))σ(η(ν))−1)([xcm, η(ν)]) (76) = ([ν]−1kf )([ν]−1[xcm, 1]) = [ν]−1(f ([xcm, 1])) = [ν]−1(f (xcm)). The next observation is one of the key ingredients in our construction of the arithmetic subalgebra, which we call therefore Proposition 4.4. The two maps of algebraic groups ϕ and η are equal. Proof. This is an immediate corollary of Prop. C.1, Lemma 3.3 , the compati- bility properties of the norm map and the next simple Lemma. Lemma 4.5. Let (L, φ) a CM type and L′ a finite extension of the reflex field L∗. Then the following diagram ResL′ / µL′ T L′ NL′/L∗ ResL∗ / µL∗ T L∗ / ResL∗/ (T L L∗) ResL′/ (T L L′) N mL′/ 6mmmmmmmmmmmmmmm N mL∗/ / T L (78) is commutative. Now having all the number-theoretic ingredients we need in hand, we can move on to the "operator-theoretic" part of this paper. 24 / / / / 6 5 On Bost-Connes-Marcolli systems We review very briefly the general construction of C∗-dynamical systems, named Bost-Connes-Marcolli systems, as given in [HP05]. 5.1 BCM pairs A BCM pair (D, L) is a pair consisting of a BCM datum D = (G, X, V, M ) together with a level structure L = (L, Γ, ΓM ) of D. A BCM datum is a Shimura datum (G, X) together with an enveloping algebraic semigroup M and a faithful representation φ : G → GL(V ) such that φ(G) ⊂ M ⊂ End(V ). Here V denotes a -vector space of finite dimension. A level structure L of D consists of a lattice L ⊂ V , a compact open subgroup Γ ⊂ G(Af ) and a compact open semigroup ΓM ⊂ M (Af ) such that φ(Γ) ⊂ ΓM and ΓM stabilizes L ⊗Z Remark 17. In 3.1 [HP05] a more general notion of Shimura datum is allowed than ours given in the Appendix. Z. To every BCM datum D and lattice L ⊂ V one can associate the fol- lowing so called maximal level structure to obtain a BCM pair by setting ΓM = M (Af ) ∩ End(L ⊗ZbZ) and Γ = φ−1(Γ× M ). The level structure L is called fine if Γ is acting freely on G( )\(X × G(Af )). Remark 18. For the definition of the topology of G(Af ) and M (Af ) we refer the reader to [PR94]. Especially one can show that φ(Af ) : G(Af ) → M (Af ) is a continuous map (cf. lemma 5.2 [PR94]). 5.2 Quotient maps attached to BCM pairs Let (D, L) a BCM pair. 5.2.1 The BCM groupoid There is a partially defined action of G(Af ) on the direct product ΓM ×Sh(G, X) given by g(ρ, [z, l]) = (gρ, [z, lg−1]), (79) where we suppressed the mophism φ. Using this the BCM groupoid U is the topological groupoid (using the notation given in (13)) defined by U = G(Af ) ⊞ (ΓM × Sh(G, X)). There is an action of the group Γ × Γ on U given by (γ1, γ2)(g, ρ, [z, l])(γ1gγ−1 2 , γ2ρ, [z, lγ−1 2 ]). (80) (81) 25 We denote the quotient by Z = (Γ × Γ)\U and obtain a natural quotient map U / Z. (82) We denote elements in Z by [g, ρ, [z, l]]. Remark 19. In general the quotient Z is not a groupoid anymore, see 4.2.1 [HP05]. 5.2.2 The positive BCM groupoid Assume that the Shimura datum (G, X) of our BCM pair satisfies (SV 5) (cf. D.1). Moreover we choose a connected component X + of X and set G( )+ = G( ) ∩ G(R)+, where G(R)+ denotes the connected component of the identity of G(R). Then G( )+ is acting naturally on X +, because X + can be regarded as a G(R)+-conjugacy class (see D.4). Now we can consider the positive (BCM) groupoid U + which is the topological groupoid given by U + = G( )+ ⊞ (ΓM × X +). (83) If we set Γ+ = Γ ∩ G( )+ we see further that Γ+ × Γ+ is acting on U + by (γ1, γ2)(g, ρ, z) = (γ1gγ−1 2 , γ2ρ, γ2z). (84) We denote the quotient by Z + = (Γ+ × Γ+)\U + and obtain another quotient map There is a natural equivariant morphism of topological groupoids U + / Z +. given by (g, ρ, z) 7→ (g, ρ, [z, 1]), inducing a commutative diagram U + / U U / Z U + / Z + (85) (86) (87) The following criterion given in 5.1 [HP05] will be crucial for our approach Criterion 5.1. If the natural map G( ) ∩ Γ → G( )/G( )+ is surjective and G( )\G(Af )/Γ = 1, then the natural morphism (86) induces a homeomor- phism of topological spaces Z + / Z. (88) Remark 20. 1) In other words the two conditions of the criterion simply mean that we have a decomposition of the form G(Af ) = G( )+ · Γ. 2) The inverse Z −→ Z + of the above homeomorphism is given explicitly as follows. By using the first remark we can write every l ∈ G(Af ) as a product l = αβ with α ∈ G( )+ and β ∈ Γ (this decomposition is unique up to an element in Γ+ = Γ ∩ G( )+). In particular every element [g, ρ, [z, l]] ∈ Z can be written as [g, ρ, [z, l]] = [gβ−1, βρ, [α−1z, 1]] and under the inverse of the above homeomorphism this element is sent to [gβ−1, βρ, α−1z] ∈ Z +. 26 / / / / O O / O O / 5.2.3 The adjoint BCM algebra Let us denote by C the center of G and assume further that φ(C( )) is a normal subsemigroup of M (Af ). The adjoint group Gad of G is the quotient of G by its center C (in the sense of algebraic groups, see [Wat79]). Let us define the semigroup Γad M to be the quotient of ΓM by the normal subsemigroup φ(C( ))∩ΓM and remember that X + can be naturally regarded as a Gad(R)+- conjugacy class (see D.4). With this in hand we define the adjoint (BCM) groupoid U ad to be the topological groupoid U ad = Gad( )+ ⊞ (Γad M × X +). (89) It is known that the projection G −→ Gad induces a surjective group homomor- phism πad : G( )+ −→ Gad( )+ (see 5.1 [Mil04]). Setting Γad = πad(Γ+) we see immediately that Γad × Γad is acting on U ad exactly as in 84. We obtain yet another quotient map U ad / Z ad. (90) Using the two projections ΓM → Γad equivariant morphism of topological groupoids M and πad there is by construction an obvious which induces (together with (87)) a commutative diagram U + −→ U ad U / Z U + Z + U ad / Z ad. (91) (92) 5.3 BCM algebras and systems Let (D, L) be a BCM pair. The BCM algebra H = HD,L is defined to be the set of compactly supported, continuous function on the quotient Z = (Γ × Γ)\U of the BCM groupoid U , i.e. H = Cc(Z). (93) By viewing functions in H as Γ × Γ-invariant functions on the groupoid U , we can equip H with the structure of a ∗-algebra by using the usual convolution and involution on U (like in the construction of groupoid C∗-algbras). We refer to 4.3.2 [HP05] for the details. After completing H in a suitable norm we obtain a C∗-algebra A (see 6.2 [HP05]). Further there is a time evolution (σt)t∈R on 27 / / O O / / O O / H resp. A so that we end up with the BCM system A = AD,L given by the C∗-dynamical system A = (A, (σt)t∈R) (94) associated with the BCM pair (L, D). For the general definition of the time evolution we refer to 4.4 [HP05]. We will state the time evolution (σt)t∈R only in the case of our BC-type systems AK (cf. 6.1.3). Remark 21. In complete analogy one might construct a positive respectively adjoint BCM system, but we don't need this. 5.4 On Symmetries of BCM algebras In section 4.5 [HP05] the authors define symmetries of BCM algebras but for our purpose we need to deviate from their definition in order to be in accordance with the definition of symmetries for BC-type systems given in [LLN09]. Let (D, L) be a BCM pair with fine level structure (see 5.1) and recall (D.2) that there is a natural right action of G(Af ) on the Shimura variety Sh(G, X) which is denoted by m[z, l] = [z, lm]. Define the subgroup GΓ(Af ) = {g ∈ G(Af ) gγ = γg ∀γ ∈ Γ}. Further, if we denote by C the center of G, the group C(R) is acting on Sh(G, X) by c[z, l] = [cz, l]. We end up with a right action of GΓ(Af ) × C(R) as symmetries on the BCM algebra HD,L given on a function f ∈ Cc(Z) by (m,c)f (g, ρ, [z, l]) = f (g, ρ, [cz, lm]). (95) Remark 22. If G(Af ) is a commutative group and we have a decomposition G(Af ) = G( ) · Γ then it is immediate that our symmetries agree with the ones defined in 4.5 [HP05]. 6 Two BCM pairs and a map In this section we will apply the constructions from the last section to our Shimura data SK = (T K, XK, hK) and SSh = (GSp(VE, ψE), H± g , hcm) from section 3 and show how the two resulting systems can be related. 6.1 Costume I: (DK, LK) and AK This section is valid for an arbitrary number field K. 6.1.1 The BCM groupoid (DK , LK) Let us recall the BCM pair (DK, LK) from 5.5 [HP05] attached to SK . It is given by (DK , LK) = ((SK , K, M K), (OK , bO× K, bOK)), 28 (96) where the algebraic semigroup M K is represented by the functor which assigns to a -algebra R the semigroup of -algebra homomorphisms Hom(K[X], K ⊗ R). By definition we have that M K(R)× = T K(R) for every -algebra R, which gives an embedding φ : T K → M K. (It will be convenient to set ΓK = bO× 6.1.2 The quotient map UK → ZK K.) The corresponding BCM groupoid, denoted by UK, is given by UK = T K(Af ) ⊞ (bOK × Sh(SK)) K acting as in (81). We denote the quotient of this action (97) with Γ2 by K = bO× K × bO× ZK = Γ2 K\UK. (98) 6.1.3 The time evolution Following 7.1 [HP05] the time evolution (σt)t∈R on the BCM algebra HK = Cc(ZK) is given as follows. Denote by N = NK/ : A× K,f → R the usual idele norm. Let f ∈ HK be a function, then we have σt(f )(g, ρ, [z, l]) = N(g)itf (g, ρ, [z, l]). (99) 6.1.4 On symmetries First, from 2.2.3 [Del79] we know that there is an isomorphism between Sh = Sh(T K, XK, hK) and π0(CK ). By class field theory the latter space π0(CK) = CK/DK is identified with the Galois group Gal(K ab/K) of the maximal abelian extension of K using the Artin reciprocity homomorphism. Under this identifi- cation the natural action of T K(Af ) = A× K,f on Sh corresponds simply to the if ν is a finite idele in T (Af ) and ω1 = [g, l] ∈ Sh Artin reciprocity map, i.e. corresponds to the identity in Gal(K ab/K), then νω1 = [g, lν] corresponds to the image [ν] in Gal(K ab/K) of ν under Artin's reciprocity map. Now, because T K is commutative we see that C(Af ) × C(R) = T K(A) = A× K is acting by symmetries on HK. By what we just said this action is simply given by the natural map A× K → π0(CK ) = CK/DK so that we obtain (tautologically) the desired action of CK /DK ∼= Gal(K ab/K) on HK. Remark 23. The reader should notice that in the case of an imaginary quadratic number field K our symmetries do not agree with the symmetries defined [CMR05] (except when the class number of K is equal to one, where the two definitions agree). For a short discussion on this matter we refer the reader to remark 25. 29 6.1.5 About extremal KM S∞-states of AK We refer to pp. 445 [CM08] or [BR81] for the notion of extremal KM S∞-states. Let AK = (AK , (σt)t∈R) denote the corresponding BCM system (cf. 5.3). In Theorem 2.1 (vi) [LLN09] it is shown that the set E∞ of extremal KM S∞-states of AK is indexed by the set Sh = Sh(T K, XK, hK) and the extremal KM S∞- state ω associated with ω ∈ Sh is given on a function f ∈ HK by evaluation, namely ω(f ) = f (1, 1, ω). (100) Remark 24. 1) It follows immediately that the symmetry group CK /DK is acting free and transitively on the set of extremal KM S∞-states. 2) Using the definition of symmetries given in [HP05] we obtain another action of CK /DK on HK but in general this will not induce a free and transitive action on the extremal KM S∞-states. In fact the two different actions of CK/DK on HK given in [LLN09] resp. [HP05] are equivalent if and only if the class number hK of K is equal to 1. This follows directly from the fact that if (and only if) hK = 1 then already bO× All put together we have K surjects onto Gal(K ab/K). Theorem 6.1 ([HP05] and [LLN09]). Let K be an arbitrary number field. Then the BCM system AK = (AK , (σt)t∈R) is a BC-type system (cf. page 2). 6.2 Costume II: (DSh, LSh) 6.2.1 The BCM groupoid (DSh, LSh) Recall the construction of the symplectic vector space (VE, ψE) (see 3.2). We still have some freedom in specifying the totally imaginary generators ξi of the (primitive) CM fields Ei, which in turn define the symplectic form ψE (cf. (56)). Let us denote by LE the lattice LE = ⊕r i=1Ei. We now fix generators ξi according to i=1OEi ⊂ VE = ⊕r Lemma 6.2. For each i ∈ {1, .., r} there exists a totally imaginary generator ξi ∈ Ei, such that the associated symplectic vector space (VE, ψE) is integral with respect to LE, i.e. there exists a symplectic basis {ej} for (VE, ψE) such that ej ∈ LE, for each j. Proof. For each i, choose any totally imaginary generator eξi ∈ Ei and regard the associated symplectic form eψE on VE (see (56)). exists a symplectic basis {eej} for (VE,eψE). Now for each j there exists a qj ∈ N . Set q = Qj qj and such that ej = qjeej ∈ LE, because Ei = OEi ⊗Z ξi = q−2eξi ∈ Ei, for every i. By construction it is now clear that {ej} is an define the symplectic form ψE on VE by using the totally imaginary generators integral symplectic basis for (VE, ψE). It is known that there 30 Now having fixed our Shimura datum SSh = (GSp(VE, ψE), H± g , hcm) we define the BCM pair (DSh, LSh) equipped with the maximal level structure (cf. 5.1) with respect to the lattice LE by (DSh, LSh) = ((SSh, VE, M Sp), (LE, ΓSh, ΓSh,M )), (101) where the algebraic semigroup M Sp = M Sp(VE, ψE) is represented by the functor which assigns to a -algebra R the semigroup M Sp(R) = {f ∈ EndR(VE ⊗ R) ∃ ν(f ) ∈ R : ψE,R(f (x), f (y)) = ν(f )ψE,R(x, y) ∀x, y}. It is clear by definition (compare A.2) that M Sp(R)× = GSp(R) which defines a natural injection φ : GSp → M Sp. 6.2.2 Some quotient maps We denote the corresponding BCM groupoid by USh = GSp(Af ) ⊞ (ΓSh,M × Sh(SSh)), (102) where the group Γ2 of USh by this action by Sh = ΓSh × ΓSh is acting as usual. We denote the quotient ZSh = Γ2 Sh\USh. (103) Thanks to D.3 and Remark 14 1) we are allowed to consider the positive and adjoint BCM groupoid, which we denote by U + Sh respectively. The corresponding quotients are denoted analogously by Z + Sh = (Γ+ Sh and Z ad Sh and U ad Sh)2\U + Sh = (Γad Sh)2\U ad Sh. 6.3 The map Θ : ZK → Z ad Sh The aim in this section is to construct a continuous map Θ : ZK −→ Z ad Sh. 6.3.1 Relating ZK and ZSh Recall that the morphism of Shimura data ϕ : SK → SSh constructed in 3.3 is a morphism of algebraic groups ϕ : T K → GSp, inducing a morphism Sh(ϕ) : Sh(SK) → Sh(SSh) of Shimura varieties. Moreover there is a natural continuation of ϕ to a morphism of algebraic semigroups M (ϕ) : M K → M Sp due to the following. We know that ϕ : T K → GSp can be expressed in terms of reflex norms (see lemma 3.3 and B.3), which are given by determinants and this definition makes still sense if we replace T K and GSp by their envelop- ing semigroups M K and M Sp, respectively. Now we can define an equivariant morphism of topological groupoids Ω : UK → USh (104) 31 by (g, m, z) ∈ UK 7→ (ϕ(Af )(g), M (ϕ)(Af )(z), Sh(ϕ)(z)) ∈ USh. (105) To show the equivariance of Ω use ϕ(Af )(ΓK ) ⊂ ΓSh and the equivariance of Sh(ϕ) (cf. D.2). We obtain a continuous map Ω : ZK → ZSh. (106) 6.3.2 Relating ZSh and Z ad Sh In order to relate ZSh and Z ad Criterion 5.1 by proving the following two lemmata. Sh we will show that we are allowed to apply Lemma 6.3. We have GSp(VE, ψE)(Af ) = GSp(VE, ψE)( ) · ΓSh. Proof. Let {ej} be an integral symplectic basis of VE with respect to LE (cf. Lemma 6.2). Each f ∈ GSp(Af ) is Af -linear and therefore determined by the values on ej ⊗ 1 ⊗ 1 ∈ LE ⊗Z ⊗ZbZ given by ak,j ⊗ bk,j ⊗ ck,j ∈ LE ⊗Z f (ej ⊗ 1 ⊗ 1) =Xk ⊗ZbZ. (107) Let dk,j ∈ N be the denominator of bk,j and define c(f ) =Qk,j dk,j ∈ N. Now observe that the map Mf : ei 7→ c(f )ei is a map in GSp( ) ⊂ GSp(Af ) i.e. Mf is compatible with the symplectic struture ψE. Obviously and thus we obtain the desired decomposition Mf ◦ f ∈ ΓSh f = M −1 f ◦ (Mf ◦ f ) ∈ GSp( ) · ΓSh. Lemma 6.4. The map GSp( ) ∩ ΓSh → GSp( )/GSp( )+ is surjective. Proof. We know that GSp(R)+ = {f ∈ GSp(R)ν(f ) > 0} (see A.2). From this we get GSp( )+ = GSp( ) ∩ GSp(R)+ = {f ∈ GSp( )ν(f ) > 0}. Let f be an element in GSp( ). If we define Mf exactly like in the proof above, we see that Mf ∈ GSp( )+ and can conclude Mf ◦ f ∈ GSp( ) ∩ ΓSh. Thus we see that the natural morphism Z + Sh → ZSh (see 86) is a homeo- morphism so that we can invert this map and compose it with the natural map Z + Sh (cf. 91) to obtain a continuous map Sh → Z ad ZSh / Z ad Sh. (108) 32 / Finally if we compose the last map with Ω from above we obtain a continuous morphism denoted Θ : ZK / Z ad Sh. (109) One crucial property of Θ is that every element z ∈ ZK will be sent to an element of the form Θ(z) = [gβ−1, βρ, α−1xcm] ∈ Z ad Sh, (110) where g ∈ Gad( )+, ρ ∈ Γad πad(ϕ(T K(Af ))) ⊂ Gad(Af ). Sh,M , α ∈ Gad( ) and β ∈ Γad Sh, such that αβ ∈ Now we are ready for the 7 Construction of our arithmetic subalgebra The idea of the construction to follow goes back to [CMR05] and [CMR06]. We constructed the ring Mcm of arithmetic Modular functions on Hg that are defined in xcm (see 4.3.1). Further the group E acts by automophisms on Mcm according to 4.3.2. (Recall that we use the notation αf to denote the action of an automorphism α on a function f ∈ Mcm.) Thanks to the embeddings (cf. 4.3) E / GSp(Af ) × / MSp(Af ) × Γad Sh,M (111) the intersection E ∩ Γad Sh,M is meaningful and thus we can define, for each f ∈ Mcm, a function ef on U ad Sh by ef (g, ρ, z) =(cid:26) ρf (z), 0 if ρ ∈ E ∩ Γad Sh,M else. (112) (71) = γ−1 (113) Sh × Γad Sh because 2 , γ2m, γ2z) def= γ2mf (γ2z) By construction ef is invariant under the action of (γ1, γ2) ∈ Γad ef (γ1gγ−1 Therefore we can regard ef as function on the quotient Z ad K × bOK × Sh(T K, XK), which is a compact and clopen subset We set WK = bO× Proposition 7.1. Let f be a function in Mcm, then ef ◦ Θ is contained in of UK and invariant under the action of Γ2 preliminaries we have the following 2 γ2mf (z) = ef (g, m, z). Sh = (Γad K · WK ⊂ WK. With these Sh)2\U ad Sh. Cc(ZK), i.e. K, i.e. Γ2 ef ◦ Θ ∈ HK ⊂ AK. 33 (114) / / / o o Proof. As we already remarked in (110) the image of an element z ∈ ZK under Θ is of the form [gβ−1, βρ, α−1xcm] ∈ Z ad because Θ is continuous, the action of E is continuous and does not produce singularities at special points (see Theorem 4.3). K-invariant function on UK. Thanks to E ⊂ GSp(Af ) Let us now regard fK as Γ2 and (112) we see that the support of our function fK is contained in the clopen K × Sh(T K, XK) ⊂ UK. Using the compact subset WK ⊂ UK, Sh. Therefore fK = ef ◦ Θ is continous, × the next easy lemma finishes the proof. subset bO× K × bO× Lemma 7.2. Let G be a topological group, X be a topological G-space and Y ⊂ X a compact, clopen subset such that GY ⊂ Y . If we have a continuous, G-invariant function f ∈ CG(X) then f Y ∈ Cc(G\X). Now we can define our arithmetic subalgebra of AK = (AK, σt) (cf. 6.1). Definition 7.1. Denote by Aarith K the K-rational subalgebra of AK generated 7.1 Proof of Theorem 1.1 by the set of functions {ef ◦ Θ f ∈ Mcm}. Let f ∈ Mcm and denote fK = ef ◦ Θ ∈ Aarith K . Further denote by ω the extremal KM S∞-state of AK corresponding to ω ∈ Sh = Sh(T K, XK, hK) (see 6.1.5). Recall the isomorphism Gal(K ab/K) ∼= π0(CK ) = Sh given by Artin reciprocity. Considered as element in Gal(K ab/K) we write [ω] for ω. Property vi) Let ν ∈ A× K be a symmetry of AK (see 6.1.4). Thanks to Lemma 6.2 and 6.3, we can write ϕ(Af )(ν) = αβ ∈ GSp(Af ) with α ∈ GSp( )+ and β ∈ ΓSh. By α resp. β we will denote their images in GSpad( )+ resp. Γad Sh under the map πad ◦ ϕ(Af ). Moreover we denote the image of ν under Artin reciprocity by [ν] ∈ Gal(K ab/K). The action of the symmetries on the extremal KM S∞-states is given by pull back and because this action is free and transitive it is enough to restrict to the case of the extremal KM S∞-state 1 corresponding to the identity in Gal(K ab/K). Using Proposition 4.4 and the reciprocity law (77) we can calculate the action of ν on ω(fK) as follows ν1(fK) def= 1(νfK) −1 , β, α−1xcm) 4.4= η(β)f (α−1xcm) (110) = ef (β (71) = η(α)η(β)f (xcm) = η(ν)f (xcm) = [ν]−1(1(fK)) (77) = [ν]−1(f (xcm)) This is precisely the intertwining property we wanted to prove. 34 Property v) Using the notation from above, we conclude immediately that by construction and Theorem 4.2 we have and the above calculation shows further that 1(fK) = ef (1, 1, xxm) = f (xcm) ∈ K ab (115) ω(fK) = [ω]1(fK) = [ω]−1(f (xcm)) ∈ K ab (116) finishing our proof. Remark 25. We want to conclude our paper with a short discussion comparing our construction with the original construction of Connes, Marcolli and Ra- machdran (see [CMR05]) in the case of an imaginary quadratic field K. Apart from the fact that we are not dealing with the "K-lattice" picture as done in [CMR05] the main difference lies in the different definitions of symmetries. If the class number hK of K is equal to one, it is immediate that the two defini- tions agree, however for hK > 1 their symmetries contain endomorphisms (see Prop. 2.17 [CMR05]) whereas our symmetries are always given by automor- phisms. We want to mention that it is no problem to generalize (this is already contained in [CM08]) their definition to the context of a BC-type system for an arbitrary number field and, without changing the definition of our arithmetic subalgebra, we could have proved Theorem 1.1 by using the new definition of symmetries (now containing endomorphisms). This might look odd at first sight but in the context of endomotives (see [CM08] for a reference) the relationship between the two different definitions will be- come transparent. More precisely in a forthcoming article we will show that every BC-type system (for an arbitrary number field) is an endomotive and the precise relationship between the two definitions of symmetries (and their actions on (extremal) KM Sβ-states) will become clear. A Algebraic Groups Our references are [Wat79] and [Mil06]. Let k denote a field of characteristic zero, K a finite field extension of k and k an algebraic closure of k. Further we denote by R a k-algebra. A.1 Functorial definition and basic constructions An (affine) algebraic group G (over k) is a representable functor from (com- mutative, unital) k-algebras to groups. We denote by k[G] its representing algebra, i.e. for any R we have G(R) = Homk−alg(k[G], R). A homomorphism F : G → H between two algebraic groups G and H (over k) is given by a natural transformation of functors. 35 Let G and H be two algebraic groups over k, then their direct product G × H is the algebraic group (over k) given by R 7→ G(R) × H(R). Let G be an algebraic group over k and K and extension of k. Then by ex- tension of scalars we obtain an algebraic group GK over K represented by K ⊗k k[G]. Now let G be an algebraic group over K. The Weil restriction ResK/k(G) is an algebraic group over k defined by ResK/k(G)(R) = G(K ⊗k R). Remark 26. All three constructions are functorial. A.2 Examples 1) The multiplicative group Gm,k (over k) is represented by k[x, x−1] = k[x, y]/(xy − 1), i.e. Gm,k(R) = R×. 2) Define S = ResC/R(Gm,C). We have S(R) = C× and S(C) ∼= C× × C×. In particular we have SC ∼= Gm,C × Gm,C. 3) More general an algebraic group T over k is called a torus if Tk is isomorphic to a product of copies of Gm,k. 4) The general symplectic group GSp(V, ψ) attached to a symplectic - vector space (V, ψ) is an algebraic group over defined on a -algebra R by GSp(R) = {f ∈ EndR(V ⊗ R) ∃ ν(f ) ∈ R× : ψR(f (x), f (y)) = ν(f )ψR(x, y) ∀x, y ∈ V ⊗ R}. A.3 Characters Let G be an algebraic group over k and set Λ = Z[Gal(k/k)]. The character group X ∗(G) of G is defined by Hom(Gk, Gm,k). There is a natural action of Gal(k/k) on X ∗(G), i.e. X ∗(G) is a Λ-module. Analogously the cocharacter group X∗(G) of G is the Λ-module Hom(Gm,k, Gk). We denote the action of σ ∈ Λ on a (co)character f by σf or f σ. There is the following important Theorem A.1 (7.3 [Wat79]). The functor G 7→ X ∗(G) is a contravariant equivalance from the category of algebraic groups of multiplicative type over k and the category of finitely generated abelian groups with a continuous action of Gal(k/k). Remark 27. See 7.2 [Wat79] for the definition of groups of multiplicative type. We only have to know that algebraic tori are of multiplicative type. There is a natural bi-additive and Gal(k/k)-invariant pairing < ·, · >: X∗(G)× X ∗(G) → Z given by < χ, µ >= µ ◦ χ ∈ Hom(Gm,k, Gm,k) ∼= Z. If G is of mul- tiplicative type the pairing is perfect, i.e. there is an isomorphism of Λ-modules X∗(G) ∼= Hom(X ∗(G), Z). A.4 Norm maps Let L be a finite field extension of K, i.e. we have a tower k ⊂ K ⊂ L, and let T be a torus over k. Then there are two types of morphisms of algebraic groups 36 which we call norm maps. The first one N mL/K : ResL/K(TL) → TK (117) is induced by the usual norm map of algebras R⊗K L → R, for R a K-algebra. In applying the Weil restriction functor ResK/k we obtain the second one, namely NL/K : ResL/k(TL) → ResK/k(TK). A.5 The case of number fields Let K be a number field. We are interested in the algebraic group T K = ResK/ (Gm,K) (over ). We have T K(R) = (K ⊗ R)×. It is easy to see that the isomorphism K ⊗ isomorphism of algebraic groups T K X ∗(T K) ∼= Z ∼=Qρ∈Hom(K, ) induces an ∼= Qρ∈Hom(K, ) Gm, Hom(K, ) with Gal(Q/ ) acting as follows. It follows that . For f =Pρ∈Hom(K,Q) aρ[ρ] ∈ Z σf = Xρ∈Hom(K,Q) Hom(K,Q) and σ ∈ Gal(Q/ ) we have aρ[σ ◦ ρ] = Xρ∈Hom(K,Q) aσ−1◦ρ[ρ]. For any inclusion K ⊂ L of number fields the norm map NL/K : T L → T K is defined by saying that a character f =Pρ∈Hom(K,Q) aρ[ρ] ∈ X ∗(T K) is mapped to the character fL =Pρ′∈Hom(L,Q)(aρ′K)[ρ′] ∈ X ∗(T L). B CM fields We follow [Mil06] and [Mil04]. By ι we denote the complex conjugation of C. B.1 CM fields and CM types Let E denote a number field. If E is a totally imaginary quadratic extension of a totally real field, we call E a CM field. In particular the degree of a CM field is always even. A CM type (E, Φ) is a CM field E together with a subset Φ ⊂ Hom(E, C) such that Φ ∪ ιΦ = Hom(E, C) and Φ ∩ ιΦ = ∅. B.2 About hφ and µφ Let (E, Φ) be a CM type. Then there are natural isomorphisms T E resp. T E C ∼=Qφ∈Φ S ∼=Qφ∈Φ Gm,C ×Qφ∈ιΦ Gm,C, where the first one is induced by E ⊗ ∼=Qφ∈Φ C and the second one by E ⊗ C ∼=Qφ∈Φ C ×Qφ∈ιΦ C. Thus we obtain natural morphisms R R hΦ : S → T E R ; z 7→ (z)φ∈Φ (118) 37 and µΦ : Gm,C → T E C ; z 7→ (z)φ∈Φ × (1)φ∈ιΦ. (119) If we take µΦ for granted we could have defined hΦ by the composition ResC/R(Gm,C) ResC/R(µΦ) / ResC/R(T E C ) N mC/R / T E R . (120) In particular we see that hΦ and µΦ are related by hΦ,C(z, 1) = µΦ(z). (121) Remark 28. In the last two sections one might have replaced C by Q. B.3 The reflex field and reflex norm Let (E, Φ) be a CM type. The reflex field E∗ of (E, Φ) is the subfield of defined by any one of the following conditions: a) σ ∈ Gal( / ) fixes E∗ if and only if σΦ = Φ ; b) E∗ is the field generated over by the elementsPφ∈Φ φ(e), e ∈ E ; c) E∗ is the smallest subfield of such that there exists a E ⊗ E∗-module V such that T rE ∗ (eV ) =Xφ∈Φ φ(e), for all e ∈ E. (122) The reflex norm of (E, Φ) is the morphism of algebraic groups NΦ : T E ∗ given, for R a -algebra, by → T E a ∈ T E ∗ (R) 7→ detE⊗ R(aV ⊗ R) ∈ T E(R). (123) C The Serre group Our references are [Mil98], [Mil06] and [Wei]. Let K be a number field. We fix an embedding τ : K → Q → C and denote by ι complex conjugation on C. C.1 Definition of the Serre group The following are equivalent: (1) The Serre group attached to K is a pair (SK, µK) consisting of a - algebraic torus SK and a cocharacter µK ∈ X∗(SK) defined by the following universal property. For every pair (T, µ) consisting of a -algebraic torus T and a cocharacter µ ∈ X∗(T ) defined over K satisfying the Serre condition (ι + 1)(σ − 1)µ = 0 = (σ − 1)(ι + 1)µ ∀σ ∈ Gal( / ) (124) 38 / / there exists a unique morphism ρµ : SK → T such that the diagram ρµ,Q SK Q aCCCCCCCC µK / TQ µ ={{{{{{{{ commutes. Gm,Q (2) The Serre group SK is defined to be the quotient of T K such that X ∗(SK) is the subgroup of X ∗(T K) given by all elements f ∈ X ∗(T K) which satisfy the Serre condition (σ − 1)(ι + 1)f = 0 = (ι + 1)(σ − 1)f ∀σ ∈ Gal(Q/ ). The cocharacter µK is induced by the cocharacter µτ ∈ X∗(T K) defined by < µτ , Σnσ[σ] >= nτ , ∀ Σnσ[σ] ∈ Z Hom(K,Q) ∼= X ∗(T K). (3) If K does not contain a CM subfield, we set E = , otherwise E denotes the maximal CM subfield of K and F the maximal totally real subfield of E. Then there is an exact sequence of -algebraic groups 1 / ker(NF/ : T F → T ) i / T K πK / SK / 1, (125) where i is the obvious inclusion. The cocharacter µτ of T K, defined as in (2), induces µK, i.e. µK = πK ◦ µτ . Remark 29. For K = or K an imaginary quadratic field there is the obvious equality SK = T K. C.2 About µK and hK The cocharacter µK = πK ◦ µτ : Gm,C → SK natural morphism C from the last section induces a hK : S → SK R (126) defined by ResC/R(Gm,C) ResC/R(µK ) / ResC/R(SK C ) N mC/R / SK R . (127) We see that µK and hK are related by or in other words, for z ∈ C×, we have hK(z) = µK(z)µK(z)ι. C (z, 1) = µK(z) hK (128) 39 / a = / / / / / / C.3 About ρΦ and the reflex norm NΦ Let (E, Φ) be a CM type. The natural morphism µΦ ∈ X∗(T E) (cf. 119) is defined over the reflex field E∗ and an easy calculation shows that it satisfies the Serre condition (124). By the universal property of the Serre group we obtain a -rational morphism such that Also, we see immediately that ρΦ : SE ∗ −→ T E µΦ = ρΦ,C ◦ µE ∗ . hΦ = ρΦ,R ◦ hE ∗ . (129) (130) (131) Moreover we can relate ρΦ and µΦ by the following commutative diagram / T E (132) T E ∗ πE∗ !DDDDDDDD SE ∗ Res(µΦ,E∗ ) NE∗/ E ∗ ) ResE ∗/ (T E 3ggggggggggggggggggggggggggggggg ρΦ which can be seen on the level of characters. The relation with the reflex norm NΦ : T E ∗ → T E (see B.3) is given by the following important Proposition C.1 ([Mil06]). We have the equality NΦ = NE ∗/ ◦ Res(µΦ,E ∗ ). (133) C.4 More properties of the Serre group The following properties are all taken from [Mil06]. Proposition C.2. Let E ⊂ K denote two number fields. 1. The norm map NK/E : T K → T E induces a commutative diagram NK/E T K T E πK πE SK / SE. (134) We call the induced morphism NK/E : SK → SE. 40 / / ! / 3 / / / 2. There is a commutative diagram S SK hK hE @@@@@@@@ NK/E SE (135) 3. Let E denote the maximal CM field contained in K, if there is no such subfield we set E = . Then NK/E : SK → SE is an isomorphism. 4. Let (E, Φ) be a CM type and K1 ⊂ K2 two number fields, such that E∗ ⊂ K1, and let ρΦ,i : SKi → T E be the corresponding maps from the universal property of the Serre group. Then we have ρΦ,1 ◦ NK2/K1 = ρΦ,2. D Shimura Varieties Our references are Deligne's foundational [Del79], Milne's [Mil04] and Hida's [Hid04]. Let G be an algebraic group over . Then the adjoint group Gad of G is defined to be the quotient of G by its center C. The derived group Gder of G is defined to be the intersection of the normal algebraic subgroups of G such that G/N is commutative. By G(R)+ we denote the identity component of G(R) relative to its real topology and set G( )+ = G( ) ∩ G(R)+. If G is reductive, we denote by G(R)+ the group of elements of G(R) whose image in Gad(R) lies in its identity component and set G( )+ = G( ) ∩ G(R)+. D.1 Shimura datum A Shimura datum is a pair (G, X) consisting of a reductive group G (over ) and a G(R)-conjugacy class X of homomorphisms h : S → GR, such that the following (three) axioms are satisfied (SV 1): For each h ∈ X, the representation Lie(GR) defined by h is of type {(−1, 1), (0, 0), (1, −1)}. (SV 2): For each h ∈ X, ad(h(i)) is a Cartan involution on Gad (SV 3): Gad has no -factors on which the projection of h is trivial. . R Because G(R) is acting transitively on X it is enough to give a morphism h0 : S → GR to specify a Shimura datum. Therefore a Shimura datum is sometimes written as triple (G, X, h0) or simply by (G, h0). Further in our case of interest the following axioms are satisfied and (simplify the situation enormously). 41 / / (SV 4): The weight homomorphism ωX : Gm,R → GR is defined over . (SV 5): The group C( ) is discrete in C(Af ). (SV 6): The identity component of the center Co splits over a CM-field. (SC): The derived group Gder is simply connected. (CT ): The center C is a cohomologically trivial torus. Remark 30. 1) Axioms (SV 1 − 6) are taken from [Mil04], the other two axioms are taken from [Hid04]. 2) The axioms of a Shimura variety (SV 1-3) imply, for example, that X is a finite union of hermitian symmetric domains. When viewed as an analytic space we sometimes write x instead of h for points in X and hx for the associated morphism hx : S → GR. 3) In 3.1 [HP05] a more general definition of a Shimura datum is given. For our purpose Deligne's original definition, as given above, and so called 0-dimensional Shimura varieties are sufficient. A morphism of Shimura data (G, X) → (G′, X ′) is a morphism G → G′ of algebraic groups which induces a map X → X ′. D.2 Shimura varieties Let (G, X) be a Shimura datum and let K be a compact open subgroup of G(Af ). Set ShK = ShK(G, X) = G( )\X × G(Af )/K, where G( ) is acting on X and G(Af ) on the left, and K is acting on G(Af ) on the right. On can show (see 5.13 [Mil04]) that there is a homeomorphism ShK ∼=F Γg\X +. Here X + is a connected component of X and Γg is the subgroup gKg−1 ∩ G( )+ where g runs runs over a set of representatives of G( )+\G(Af )/K. When K is chosen sufficiently small, then Γg\X + is an arithmetic locally symmetric variety. For an inclusion K ′ ⊂ K we obtain a natural map ShK ′ → ShK and in this way an inverse system (ShK)K. There is a natural right action of G(Af ) on this system (cf. p 55 [Mil04]). The Shimura variety Sh(G, X) associated with the Shimura datum (G, X) is defined to be the inverse limit of varieties lim←−K ShK(G, X) together with the natural action of G(Af ). Here K runs through sufficiently small compact open subgroups of G(Af ). Sh(G, X) can be regarded as a scheme over C. Let (G, X) be a Shimura datum such that (SV 5) holds, then one has Sh(G, X) = lim←− K ShK(G, X) = G( )\X × G(Af ). (136) In this case we write [x, l] for an element in Sh(G, X) and the (right) action of an element g ∈ G(Af ) is given by g[x, l] = [x, lg]. (137) 42 In the general case, wenn (SV 5) is not holding we use the same notation, un- derstanding that [x, l] stands for a family (xK , lK)K indexed by compact open subgroups K of G(Af ). A morphism of Shimura varieties Sh(G, X) → Sh(G′, X ′) is an inverse sys- tem of regular maps of algebraic varieties compatible with the action of G(Af ). We have the following functorial property: A morphism ϕ : (G, X) → (G′, X ′) of Shimura data defines an equivariant mor- phism Sh(ϕ) : Sh(G, X) → Sh(G′, X ′) of Shimura varieties, which is a closed immersion if G → G′ is injective (Thm 5.16 [Mil04]). D.3 Example We want to give some details about the Shimura varieties attached to the data SSh constructed in 3.2. For the identification of the GSp(R)-conjugacy class of hcm with the higher Siegel upper lower half space H± g = {M = A + iB ∈ Mg(C) A = At, B positive or negative definitive } we refer further to exercise 6.2 [Mil04]. In addition the data SSh fulfill all the axioms stated in D.1. The validity of (SV 1-6) is shown on p. 67 [Mil04] and the validity of (SC) and (CT ) in [Hid04]. The latter two axioms are important for making the arguments in [MS81] in this case. From (SV 5) follows in particular that we don't have to bother about the limits in the definition of Sh(GSp, H± g ) because we have Sh(GSp, H± g ) = GSp( )\(H± g × GSp(Af )). D.4 Connected Shimura varieties A connected Shimura datum is a pair (G, X +) consisting of a semisimple algebraic group G over and a Gad(R)+-conjugacy class of homomorphisms h : S → Gad The connected Shimura variety Sho(G, X +) associated with a connected Shimura datum (G, X +) is defined by the inverse limit satisfying axioms (SV 1-3). R Sho(G, X +) = lim←− Γ Γ\X + (138) where Γ runs over the torsion-free arithmetic subgroups of Gad( )+ whose in- verse image in G( )+ is a congruence subgroup. If we start with a Shimura datum (G, X) and choose a connected component X + of X, we can view X + as a Gad(R)+-conjugacy class of morphisms h : S → Gad by projecting elements in X + to Gad . One can show that (Gder, X +) is a connected Shimura datum. Further if we choose the connected component Sh(G, X)o of Sh(G, X) containing X + × 1, one has the following compatibility relation R R Sh(G, X)o = Sho(Gder, X +). (139) 43 D.5 0-dimensional Shimura varieties In section 3.1 we defined a "Shimura datum" SK = (T K, XK) which is not a Shimura datum in the above sense because XK has more than one conjugacy class (recall that T K is commutative). Rather SK is a Shimura datum in the generalized sense of Pink [Pin90] which we don't want to recall here. Instead we define the notion of a 0-dimensional Shimura varieties following [Mil04] which covers all exceptional Shimura data we consider. We define a 0-dimensional Shimura datum to be a triple (T, Y, h), where T is a torus over , Y a finite set on which T (R)/T (R)+ acts transitively and h : S → TR a morphism of algebraic groups. We view Y as a finite cover of {h}. We remark that the axioms (SV 1 − 3) are automatically satisfied in this setup. The associated 0-dimensional Shimura variety Sh(T, Y, h) is defined to be the inverse system of finite sets T ( )\Y × T (Af )/K with K running over the compact open subgroups of T (Af ). A morphism (T, Y, h) → (H, h0) from a 0-dimensional Shimura datum to a Shimura datum, with H an algebraic torus, is given by a morphism of algebraic groups ϕ : T → H such that h = ϕR ◦ h0. If ϕ is such a morphism it defines a morphism Sh(ϕ) : Sh(T, Y, h) → Sh(H, h0) of Shimura varieties. Remark 31. We have that SK fulfills axiom (SV 5) if and only if K = or K an imaginary quadratic field (see 3.2 [HP05]). D.6 Canonical model of Shimura varieties Let (G, X) be a Shimura datum. A point x ∈ X is called a special point if there exists a torus T ⊂ G sucht that hx factors through TR. The pair (T, x) or (T, hx) is called special pair. If (G, X) satisfies the axioms (SV 4) and (SV 6), then a special point is called CM point and a special pair is called CM pair. Now given a special pair (x, T ) we can consider the cocharacter µx of GC de- fined by µx(z) = hx,C(z, 1). Denote by E(x) the field of definition of µx, i.e. the smallest subfield k of C such that µx : Gm,k → Gk is defined. Let Rx denote the composition T E(x) ResE(x)/ (µx) / ResE(x)/ (TE(x)) N mE(x)/ / / T and define the reciprocity morphism rx = Rx(Af ) : A× E(x),f → T (Af ). (140) (141) Moreover every datum (G, X) defines an algebraic number field E(G, X), the reflex field of (G, X). For the definition we refer the reader to 12.2 [Mil04]. Remark 32. 1) For the Shimura datum SSh = (GSp, H± E(SK) = (cf. p. 112 [Mil04]). 2) For explanations to relations with the reflex field of a CM field (cf. B.3), see example 12.4 b) of pp. 105 [Mil04]. g ) (see 3.2) we have 44 / A model M o(G, X) of Sh(G, X) over the reflex field E(G, X) is called canonical if 1)M o(G, X) is equipped with a right action of G(Af ) that induces an equivari- ant isomorphism M o(G, X)C ∼= Sh(G, X), and 2) for every special pair (T, x) ⊂ (G, X) and g ∈ G(Af ) the point [x, g] ∈ M o(G, X) is rational over E(x)ab and the action of σ ∈ Gal(E(x)ab/E(x)) is given by σ[x, g] = [x, rx(ν)g] (142) E(x),f is such that [ν] = σ−1 under Artin's reciprocity map. where ν ∈ A× In particular, for every compact open subgroup K ⊂ G(Af ), it follows that M o K(G, X) = M o(G, X)/K is a model of ShK(G, X) over E(G, X). Remark 33. Canonical models are known to exist for all Shimura varieties (see [Mil04]). D.7 Canonical model of connected Shimura varieties We refer to 2.7.10 [Del79] for the precise definition of the canonical model M o(G, X +) of a connected Shimura variety Sh(G, X +). Here we just want to mention the compatibility M o(Gder, X +) = M o(G, X)o (143) where the latter denotes a correctly chosen connected component of the canon- ical model M o(G, X). References [BC95] J.-B. Bost and A. Connes. Hecke algebras, type III factors and phase transitions with spontaneous symmetry breaking in number theory. Selecta Math. (N.S.), 1(3):411 -- 457, 1995. [BR79] Ola Bratteli and Derek W. Robinson. Operator algebras and quan- tum statistical mechanics. Vol. 1. Springer-Verlag, New York, 1979. C∗- and W ∗-algebras, algebras, symmetry groups, decomposition of states, Texts and Monographs in Physics. [BR81] Ola Bratteli and Derek W. Robinson. Operator algebras and quantum- statistical mechanics. II. Springer-Verlag, New York, 1981. Equi- librium states. Models in quantum-statistical mechanics, Texts and Monographs in Physics. [CM08] Alain Connes and Matilde Marcolli. Noncommutative geometry, quan- tum fields and motives, volume 55 of American Mathematical Soci- ety Colloquium Publications. American Mathematical Society, Provi- dence, RI, 2008. 45 [CMR05] Alain Connes, Matilde Marcolli, and Niranjan Ramachandran. KMS states and complex multiplication. Selecta Math. (N.S.), 11(3-4):325 -- 347, 2005. [CMR06] Alain Connes, Matilde Marcolli, and Niranjan Ramachandran. KMS states and complex multiplication. II. In Operator Algebras: The Abel Symposium 2004, volume 1 of Abel Symp., pages 15 -- 59. Springer, Berlin, 2006. [Del79] interprétation modulaire, et Pierre Deligne. Variétés de Shimura: techniques de construction de modèles canoniques. In Automorphic forms, representations and L-functions (Proc. Sympos. Pure Math., Oregon State Univ., Corvallis, Ore., 1977), Part 2, Proc. Sympos. Pure Math., XXXIII, pages 247 -- 289. Amer. Math. Soc., Providence, R.I., 1979. [Hid04] Haruzo Hida. p-adic automorphic forms on Shimura varieties. Springer Monographs in Mathematics. Springer-Verlag, New York, 2004. [HP05] Eugene Ha and Frédéric Paugam. Bost-Connes-Marcolli systems for Shimura varieties. I. Definitions and formal analytic properties. IMRP Int. Math. Res. Pap., (5):237 -- 286, 2005. [LLN09] Marcelo Laca, Nadia S. Larsen, and Sergey Neshveyev. On Bost- J. Number Theory, Connes types systems for number fields. 129(2):325 -- 338, 2009. [Mil98] [Mil04] [Mil06] [MS81] Joseph S. Milne. Abelian varieties with Complex Multiplication (for Pedestrians). http://arxiv.org/pdf/math/9806172v1, 1998. Joseph http://www.jmilne.org/math/xnotes/svi.pdf, 2004. Introduction S. Milne. to Shimura varieties. Joseph http://www.jmilne.org/math/CourseNotes/CM.pdf, 2006. Complex Milne. S. Multiplication. J. S. Milne and Kuang-yen Shih. Automorphism groups of Shimura varieties and reciprocity laws. Amer. J. Math., 103(5):911 -- 935, 1981. [Pin90] Richard Pink. Arithmetical compactification of mixed Shimura vari- eties. Bonner Mathematische Schriften [Bonn Mathematical Publica- tions], 209. Universität Bonn Mathematisches Institut, Bonn, 1990. Dissertation, Rheinische Friedrich-Wilhelms-Universität Bonn, Bonn, 1989. [PR94] Vladimir Platonov and Andrei Rapinchuk. Algebraic groups and num- ber theory, volume 139 of Pure and Applied Mathematics. Academic Press Inc., Boston, MA, 1994. Translated from the 1991 Russian original by Rachel Rowen. 46 [Shi00] Goro Shimura. Arithmeticity in the theory of automorphic forms, vol- ume 82 of Mathematical Surveys and Monographs. American Mathe- matical Society, Providence, RI, 2000. [Sil94] Joseph H. Silverman. Advanced topics in the arithmetic of elliptic curves, volume 151 of Graduate Texts in Mathematics. Springer- Verlag, New York, 1994. [Wat79] William C. Waterhouse. Introduction to affine group schemes, vol- ume 66 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1979. [Wei] Wafa Wei. Hilbert's bielefeld.de/sfb343/preprints/pr94070.ps.gz. problem. Moduli fields 12th of CM Motives to http://www.mathematik.uni- applied 47
1011.5400	2	1011	2011-05-03T17:24:48	Quantum isometry groups of duals of free powers of cyclic groups	[ "math.OA", "math.FA", "math.QA" ]	We study the quantum isometry groups of the noncommutative Riemannian manifolds associated to discrete group duals. The basic representation theory problem is to compute the law of the main character of the relevant quantum group, and our main result here is as follows: for the group Z_s^{*n}, with s>4 and n>1, half of the character follows the compound free Poisson law with respect to the measure $\underline{\epsilon}$/2, where $\epsilon$ is the uniform measure on the s-th roots of unity, and $\epsilon\to\underline{\epsilon}$ is the canonical projection map from complex to real measures. We discuss as well a number of technical versions of this result, notably with the construction of a new quantum group, which appears as a "representation-theoretic limit", at s equal to infinity.	math.OA	math	1 QUANTUM ISOMETRY GROUPS OF DUALS OF FREE POWERS OF 1 0 2 CYCLIC GROUPS y a M 3 ] . A O h t a m [ TEODOR BANICA AND ADAM SKALSKI Abstract. We study the quantum isometry groups G+(bΓ) of the noncommutative Rie- mannian manifolds associated to discrete group dualsbΓ. The basic representation theory problem is to compute the law of the main character χ : G+(bΓ) → C, and our main result here is as follows: for Γ = Z∗n s , with s ≥ 5 and n ≥ 2, the variable χ/2 follows the com- pound free Poisson law πε/2, where ε is the uniform measure on the s-th roots of unity, and ε → ε is the canonical projection map from complex to real measures. We discuss as well a number of technical versions of this result, notably with the construction of a new quantum group, which appears as a "representation-theoretic limit", at s = ∞. 2 v 0 0 4 5 . 1 1 0 1 : v i X r a Introduction The notion of compact quantum group was introduced by Woronowicz in [35], [36]. Woronowicz's formalism, which is both quite general, and remarkably easy to handle, allowed Wang to construct in [33], [34] a number of universal quantum groups, namely the free analogues of On, Un, Sn. The next step, developed by Bichon in [16] and by the first-named author in [1], was the construction of quantum automorphism groups of various discrete structures (finite graphs, finite metric spaces). This allowed a considerable extension of Wang's original list of free quantum groups, notably with a free analogue of the hyperoctahedral group Hn [3], and a number of technical versions of it [2], [5], [9]. Another important generalization of Wang's original constructions comes from the work of Goswami [23]. The idea is that Connes' noncommutative geometry theory [20], [21] has shown that a number of important situations, mainly coming from particle physics or number theory, are best described by a certain spectral triple X = (A, H, D). It is therefore natural to ask for the computation of the quantum isometry group G+(X) of such a spectral triple, constructed in [23]. The theory here has been quickly developed in the last few years, first with a number of foundational papers by Bhowmick and Goswami [11], [12], [13]. The earlier quantum automorphism constructions in [1], [16] were refor- mulated, unified and generalized in [14], by using the discrete spectral triples constructed 2000 Mathematics Subject Classification. 46L65 (16W30, 46L54, 58J42). Key words and phrases. Spectral triple, Quantum isometry, Noncrossing partition. 1 2 TEODOR BANICA AND ADAM SKALSKI by Christensen and Ivan in [18]. Another important computation, with several poten- tial applications, is that of the quantum isometry group of Connes' spectral triple of the standard model: this was recently done by Bhowmick, D'Andrea and Dabrowski [10]. One promising direction in view of the general understanding of the quantum isometry groups is that of the explicit computation of G+(X), in the case where X = bΓ is a discrete group dual: indeed, a number of very concrete tools, coming from combinatorics, subfactors, or free probability, are available here. The general theory, as well as a number of examples, including an explicit presentation result in the free group case Γ = Fn, were worked out in [15]. This latter free group case was studied in much detail in [8], where a link with the hyperoctahedral quantum group in [3], with the Fuss-Catalan combinatorics of Bisch and Jones [17], and with the free spheres constructed in [7] was obtained. More generally, it was shown in [8] that the quantum isometry group G+(bFn) belongs in fact to a general class of "two-parameter quantum symmetry groups", and this makes a link with the recent classification program for the easy quantum groups, started in [5], [9]. The purpose of the present paper is to push one step forward the previous combinatorial and probabilistic considerations in [8], [15]. These considerations basically concern the computation of the law of the main character, and our main result here is as follows. Theorem. For Γ = Z∗n follows the compound free Poisson law πε/2, where ε is the uniform measure on the s-roots of unity, and ε → ε is the canonical projection map from complex to real measures. s with s ≥ 5 and n ≥ 2, the half of the main character of G+(bΓ) The proof of this result uses the general diagrammatic methods in [2], [8], Voiculescu's R-transform from [31], [32] and Speicher's notion of free cumulants from [27], [28]. As in the previous paper [2], we end up with a measure which is a compound free Poisson law, in the sense of [24], [29]. However, the situation here is more complicated, combinatorially speaking, because the projection map ε → ε appears. The above theorem holds in fact as well at s = 3. However, at s = 2, 4 the situation is quite different, and we will present here some results in this direction. The main point here is that the case s = 4 is truly special, as already observed in [1], [15]. The limiting case s → ∞ is quite interesting as well, because the quantum groups in the above theorem "converge", in a certain representation theory sense, to a compact quantum group which is different from the quantum isometry group of the dual of the free group Fn = Z∗n studied in [8]. The construction of this new quantum group, for which we refer to section 6 below, will be actually our second main result in this paper. Finally, let us mention that there are many similarities with the "easy quantum group" results in [2], [5], [6]. One can expect that some further generalizations of the above theorem might provide some answers to the questions raised in [5]. We do not know if it is so, but we will make some comments in this direction, at the end of the paper. The paper is organized as follows: 1 is a preliminary section, in 2 we discuss the s ), and in 3-5 we compute diagrams, discuss the compound general properties of G+(dZ∗n QUANTUM ISOMETRY GROUPS 3 free Poisson laws, and prove the above theorem. The final sections, 6-7, contain some technical versions of that theorem, and a few concluding remarks. Acknowledgements. T.B. would like to thank P. Hajac and the IMPAN for the warm hospitality, during a visit in October 2010, when part of this work was done. The work of T.B. was supported by the ANR grants "Galoisint" and "Granma". We also want to thank the referees for the comments that have led to an improvement of the paper. 1. Quantum isometries noncommutative geometry [20], [21], where the basic definition is as follows. The natural framework for the study of noncommutative objects like bΓ is Connes' Definition 1.1. A compact spectral triple (A, H, D) consists of the following: (1) a unital C ∗-algebra A; (2) a Hilbert space H, on which A acts; (3) an unbounded self-adjoint operator D on H, with compact resolvents, such that [D, x] has a bounded extension, for any x in a dense ∗-subalgebra A ⊂ A. This definition is of course over-simplified, as to best fit with the purposes of the present paper. We refer to [20], [21] for the precise formulation of the axioms. Now let Γ =< g1, . . . , gm > be a discrete group, with the generating set S = {g1, . . . , gm} assumed to satisfy S = S−1 and 1 /∈ S. The generating set S is of course taken to be a set without repetitions. In most of the examples below, m will be actually minimal. The length on Γ is given by l(g) = min{r ∈ N∃ h1 . . . hr ∈ S, g = h1 . . . hr}, and the Consider the group algebra A = CΓ, with involution given by g∗ = g−1 and antilinearity, distance on Γ is given by d(g, h) = l(g−1h). Observe that S = {g ∈ Gl(g) = 1}. and let A = C ∗(Γ) be the completion of A with respect to the maximal C ∗-norm. Definition 1.2. Associated to any discrete group Γ =< g1, . . . , gm > as above is a compact spectral triplebΓ = (A, H, D), as follows: (1) A = C ∗(Γ) is the full group algebra of Γ. (2) H = l2(Γ), with A acting on it by gδh = δgh. (3) D(δg) = l(g)δg, on the standard basis of H. by Bhowmick and Goswami in Section 2 of [12]. We recall that the standard trace of CΓ ⊂ C ∗(Γ) is given by tr(g) = δg1 and also We can think ofbΓ as being a "noncommutative Riemannian manifold", and then con- sider the quantum isometry group of orientation preserving isometries G+(bΓ), introduced Proposition 1.3. G+(bΓ) is the universal compact quantum group acting on C ∗(Γ), such that all the eigenspaces of D and the standard trace on CΓ are invariant under this action. identify in a natural way CΓ with a vector subspace of l2(Γ). 4 TEODOR BANICA AND ADAM SKALSKI Proof. This is more or less a consequence of a simpler reformulation of the general defini- tion of the quantum isometry groups of orientation preserving isometries given in [12], in the case where the representation of A on H allows a cyclic and separating vector which spans the kernel of D and satisfies natural conditions with respect to the 'smooth alge- bra' A (Corollary 2.27 of [12]). The conditions above are satisfied in the case of manifolds coming from group duals; this, together with the consequences for the computations of G+(bΓ), is discussed in detail in Section 2 of [15]. Note that the fact that the eigenspaces of D are preserved by the action α implies that α : CΓ → CΓ ⊗alg C(G+(bΓ)), so that the trace preservation condition makes sense. Let us look now at the eigenspaces of D. There is one such eigenspace for each positive (cid:3) integer r ∈ N, namely the span of the set Γr = {g ∈ Γl(g) = r} of words of length r. span of words of Γ of length r. Moreover, the representation at r = 1 is faithful. of the trace shows that the corresponding representation is unitary. The faithfulness (cid:3) Proposition 1.4. For any r ∈ N, G+(bΓ) has a unitary representation on span(Γr), the Proof. We know from Proposition 1.3 that G+(bΓ) acts on span(Γr), and the invariance property at r = 1 follows from the universal property of G+(bΓ). See [15]. The above two statements make it clear how to explicitly construct G+(bΓ): since this quantum group has to act faithfully on the span of Γ1 = S, we just have to consider the universal quantum group acting on span(S), and then divide by a suitable ideal. So, let Au(m) be the universal C ∗-algebra generated by the entries of a m × m matrix u, such that both u = (uij) and ut = (uji) are unitaries. This is a Hopf C ∗-algebra in the sense of Woronowicz [35], with comultiplication ∆(uij) = Σuik ⊗ ukj, counit ε(uij) = δij, and antipode S(uij) = u∗ According to the general results of Woronowicz in [35], we can think of Au(m) as being the algebra of continuous functions on a certain compact quantum group. We denote this quantum group by U + ji. Observe that we have S2 = id. See Wang [33]. m is the abstract object given by Au(m) = C(U + m. That is, U + m). Theorem 1.5. G+(bΓ) is the subgroup of U + (1) Those making α(gi) = Σgj ⊗ uji a morphism of ∗-algebras. (2) Those making span(Γr) invariant under α, for any r ≥ 2. m presented by the following relations: Proof. This result is from [15], where it appears in a slightly different formulation. We present below an explanation of the present formulation, along with a short proof. fully on span(S), and that the corresponding representation is unitary. In terms of Hopf We know from Proposition 1.3 and Proposition 1.4 that the quantum G+(bΓ) acts faith- algebras, this means that we have a coaction map α : span(S) → span(S)⊗alg C(G+(bΓ)), is unitary, and its coefficients uij generate the C ∗-algebra C(G+(bΓ)). and that if we denote this map by α(gi) = Σgj ⊗ uji, then the corepresentation u = (uij) QUANTUM ISOMETRY GROUPS 5 So, we obtain in this way a surjective morphism of Hopf C ∗-algebras Φ : Au(m) → m. Let us try to understand now what the kernel of Φ is. According to the universal property of Γ, from Proposition 1.3, and to Woronowicz's Tannakian results in [36], this kernel should be exactly the ideal generated by the following relations: C(G+(bΓ)), which corresponds to an embedding of compact quantum groups G+(bΓ) ⊂ U + (1) Those making α a morphism of ∗-algebras. This means that for any hi, ki ∈ S satisfying h1 . . . hp = k1 . . . kq we write the equality α(h1 . . . hp) = α(k1 . . . kq) as a formula of type Σg∈Γg ⊗ Eg = Σg∈Γg ⊗ Fg, and Eg = Fg, with g ∈ Γ, are our relations. (2) Those making span(Γr) invariant under α, for any r ≥ 2. Once again, an explanation is needed here. The idea is that if we denote by Γr the set of words of length ≤ r, then span(Γr) is invariant under α. Now since we have Γr ⊂ Γr, the invariance condition for the subspaces span(Γr) ⊂ span(Γr) corresponds indeed to certain relations. Note that the condition in (2) implies in particular that the action α preserves the trace (as tr(Γr) = {0} for r 6= 0), and the proof is completed. (cid:3) Note that in all the cases computed so far in [15] and [8] the relations in (2) actually were a consequence of these imposed by (1) -- this will also be the case in this paper. We do not know if this is true for arbitrary Γ. 2. Free products Γ = Z∗n Let us first recall a basic result from [15], solving the problem at n = 1. s . We will usually assume s ∈ {2, 3, . . . ,∞}, with the convention Z∞ = Z. In this section we discuss the presentation of the quantum group G+(bΓ), in the case Proposition 2.1. The quantum groups G+(cZs) are as follows: (1) At s = ∞ we havecZs = T, and G+(T) = O2 = T ⋊ Z2. (2) More generally, at any s 6= 4 we have G+(cZs) = Zs ⋊ Z2. (3) At s = 4 the quantum group G+(cZs) is non-classical, and infinite. Proof. These results can be deduced from Theorem 1.5, the idea being that for s 6= 4, the relations there make the coefficients uij commute. See [15]. (cid:3) In the general case now n ∈ N, let g1, . . . , gn be the standard generators of Z∗n . The fundamental coaction is denoted as follows: endow this group with the generating set S = {gi1, gi2}, where gi1 = gi, gi2 = g−1 i s , and α(gia) =Xjb gjb ⊗ ujb,ia The simplest case is when s = ∞. Here our group Γ = Z∗n we have the following result, from [8]. s is the free group Fn, and Proposition 2.2. H + tions, between the generators denoted uia,jb with i, j = 1, . . . , n and a, b = 1, 2: n0 = G+(cFn) is the subgroup of U + 2n presented by the following rela- α(gia)α(gi¯a) = 1 ⊗ 1 ⇐⇒ Xjbkc ⇐⇒ Xjbkc gjbgkc ⊗ ujb,iaukc,i¯a = 1 ⊗ 1 gjbgk¯c ⊗ ujb,iau∗ kc,ia = 1 ⊗ 1 6 TEODOR BANICA AND ADAM SKALSKI (1) The entries uia,jb are partial isometries. (2) We have u∗ ia,jb = ui¯a,j¯b, for any i, j, a, b. Proof. This is done in [8], the idea being as follows. First, we know from Theorem 1.5 that α must preserve the involution, i.e. that we must have α(gia)∗ = α(gi¯a) for any i, a, and this gives u∗ ia,jb = ui¯a,j¯b, for any i, j, a, b. Next, the other condition that we get from Theorem 1.5 is: Thus we must have ujb,iau∗ jb,ia = 1. By multiply- ing this latter equality at right by ukc,ia, we obtain ukc,iau∗ kc,iaukc,ia = ukc,ia, so ukc,ia must be a partial isometry. This leads to the conclusion in the statement, as the conditions in (2) in Theorem 1.5 are automatically satisfied, see the proof of Theorem 5.1 in [15]. (cid:3) kc,ia = 0 for (jb) 6= (kc), andPjb ujb,iau∗ In the above statement, the notation H + under consideration is part of a certain two-parameter series H + H + m = H + Let us record as well the following useful reformulation of the above result. 0m is the hyperoctahedral quantum group in [3]. See [8]. n0 comes from the fact that the quantum group nm, which is such that Proposition 2.3. C(H + 1, . . . , n and a, b = 1, 2, subject to the following relations: n0) is the universal algebra generated by variables uia,jb with i, j = (1) uia,jb are partial isometries, satisfying u∗ (2) ujb,iau∗ kc,ia = 0 for (jb) 6= (kc), andPjb ujb,iau∗ ia,jb = ui¯a,j¯b. jb,ia = 1. Proof. This is indeed just a technical reformulation of the above result, the idea being that the relations in (2) correspond to the fact that the matrices u, ut are unitaries. (cid:3) s ) is the subgroup In the general case now, we have the following statements. Lemma 2.4. For s ≥ 5 and n ≥ 2, the quantum group H s+ of H + n0 presented by the relations coming from α(gia)s = 1 ⊗ 1, for any i, a. n0 = G+(dZ∗n Proof. First, let us mention the fact that the result holds as well at s = 3, with the same proof as below, and at s = ∞ too, with the convention that the extra relations do not exist in this case. These special cases will be discussed in section 6 below. Consider the standard coaction, α(gia) =Pjb gjb ⊗ ujb,ia. Our assumption s 6= 1, 2, 4 shows that the elements gjbgk¯c are distinct and different from 1 for (jb) 6= (kc), so the computations in the proof of Proposition 2.2 apply, and show that the generators uia,jb must satisfy the conditions found there. That is, we have indeed H s+ According now to Theorem 1.5, the conditions which are left are simply those coming from the equalities α(gia)s = 1⊗ 1, for any i, a, and we are done. The conditions in (2) in n0 ⊂ H + n0. QUANTUM ISOMETRY GROUPS 7 Theorem 1.5 are satisfied due to the fact that we are dealing with a quantum subgroup of H + s are those of the form gs = e -- compare with the proof of Theorem 5.1 in [15]. (cid:3) n0 and the only extra relations in Z∗n Theorem 2.5. For s ≥ 5 and n ≥ 2, the quantum group H s+ n0 presented by the relation ξ ∈ F ix(u⊗s), where ξ =Pia e⊗s of H + n0 = G+(dZ∗n ia . Proof. According to Lemma 2.4, we just have to understand the meaning of the relations coming from the conditions α(gia)s = 1 ⊗ 1, for any i, a. We have: gj1b1 . . . gjsbs ⊗ uj1b1,ia . . . ujsbs,ia s ) is the subgroup α(gia)s = Xj1...js Xb1...bs = Xγ γ ⊗ Xj1...js Xb1...bs δγ,gj1b1 ...gjsbs uj1b1,ia . . . ujsbs,ia! We know from Proposition 2.3 that we have ujb,iaukc,ia = 0 for (jb) 6= (k¯c), so we can erase from the above formula all the decompositions γ = gj1b1 . . . gjsbs which are not reduced with respect to the rules giagi¯a = 1. So, if we denote by Ws the set of reduced words of length s, we have the following formula: α(gia)s = 1 ⊗ Xjb us jb,ia! + Xγ∈Ws γ ⊗ Xj1...js Xb1...bs δγ,gj1b1 ...gjsbs uj1b1,ia . . . ujsbs,ia! Let us look now at an equality of type γ = γ′, with γ, γ′ ∈ Ws. This can happen only s ⊂ Ws in the situation gp the set of words which are not of this type, we have the following formula: k¯c , with j 6= k and p + q = s. So, if we denote by W ′ kc = gs−p j¯b gs−q jbgq α(gia)s = 1 ⊗ Xjb + Xgj1b1 ...gjsbs ∈W ′ us s jb,ia! +Xj6=kXc Xs=p+q gj1b1 . . . gjsbs ⊗ uj1b1,ia . . . ujsbs,ia gp j1gq kc ⊗ (up j1,iauq kc,ia + us−p j2,iaus−q k¯c,ia) Thus the equality α(gia)s = 1 ⊗ 1 is equivalent to the following conditions: jb,ia = 1 kc,ia + us−p up j1,iauq j2,iaus−q uj1b1,ia . . . ujsbs,ia = 0  Pjb us ∀ i,∀ a k¯c,ia = 0 ∀ j 6= k,∀ c,∀ s = p + q ∀ gj1b1 . . . gjsbs ∈ W ′ s (1) 8 TEODOR BANICA AND ADAM SKALSKI ia ⊗ 1! e⊗s Let us look now at the condition ξ ∈ F ix(u⊗s). We have: u⊗s(ξ ⊗ 1) = Xirjrarbr = Xj1...js Xb1...bs u⊗s(ξ ⊗ 1) = Xjb uj1b1,ia . . . ujsbs,ia! ej1b1,i1a1 ⊗ . . . ⊗ ejsbs,isas ⊗ uj1b1,i1a1 . . . ujsbs,isas! Xia ej1b1 ⊗ . . . ejsbs ⊗ Xia jb,ia! jb ⊗ Xia ej1b1 ⊗ . . . ejsbs ⊗ Xia + Xgj1b1 ...gjsbs ∈Ws uj1b1,ia . . . ujsbs,ia! e⊗s us By using now the formula ujb,iaukc,ia = 0, valid for any (jb) 6= (k¯c), we get: Thus the condition ξ ∈ F ix(u⊗s) is equivalent to the following conditions: jb,ia = 1 (Pjb us Pia uj1b1,ia . . . ujsbs,ia = 0 ∀ gj1b1 . . . gjsbs ∈ Ws ∀ i,∀ a (2) gj1b1 . . . gjsbs ∈ Ws − W ′ must be of the form gp s. Now by recalling the definition of Ws, W ′ jbgq the assertions are clear, except perhaps for the equality Pia uj1b1,ia . . . ujsbs,ia = 0, with Pia up Let us prove now that we have (1) ⇐⇒ (2). We begin with (1) =⇒ (2). Here all s, our word gj1b1 . . . gjsbs kc with j 6= k and p + q = s, and we must prove that we have jb,iauq kc,ia = 0 in this situation. For this purpose, we use the middle condition in (1), j1,iauq namely up k¯c,ia = 0. By multiplying by u∗ kc,ia = 0. By multiplying by uj1,ia we obtain uj1,iau∗ kc,ia = 0. But now, since uj1,ia is a partial isometry, and for each partial isometry v we have vv∗v = v, we can remove the uj1,iau∗ kc,ia = 0, and this gives the result, since all j1,ia term, and we are left with up j1,ia we obtain u∗ kc,ia+us−p j2,iaus−q j1,iaup j1,iauq j1,iaup j1,iauq j1,iauq jb,iauq kc,ia are now equal to 0. In the other direction now, (2) =⇒ (1), all the assertions are clear as well: indeed, the third relations in (1) can be obtained from the second relations in (2) by multiplying by ujb,ia, known to be a partial isometry, and the second relations in (1) can be obtained from the second relations in (2), for the words of type gj1b1 . . . gjsbs ∈ Ws − W ′ s. (cid:3) terms in the sumPia up 3. Noncrossing partitions In this section we study the representation theory of G+(dZ∗n s ). According to a well- known principle, going back to Woronowicz's work in [36], most of the representation theory invariants are encoded in the structure of the spaces Hom(u⊗k, u⊗l), which form a tensor category. So, it is this tensor category that we want to compute. QUANTUM ISOMETRY GROUPS 9 In addition, another well-known principle, once again going back to the considerations in [36], tells us that we should look for an explicit basis for the elements of Hom(u⊗k, u⊗l), indexed by "diagrams". This latter principle has been proved to be fruitful in a number of situations, see e.g. [5], [8], and we will prove that it applies once again here. Already in [8] we noticed that due to the fact that the quantum isometry groups we study have non- selfadjoint entries in the fundamental representations, it is natural to consider diagrams with vertices of two different colors. So, let us first construct a candidate for such a basis of Hom(u⊗k, u⊗l). Definition 3.1. We let Ds(k, l) be the set of noncrossing partitions between k upper points and l lower points, with each leg colored black or white, with the following rules: (1) In each block, the signed number of black legs equals the signed number of white (2) We agree to identify any colored partition with all the colored partitions obtained legs, modulo s (the sign being + for an upper leg, and − for a lower leg). by interchanging the two colors, independently in each block. In this definition the parameter s ranges in the set {1, 2, 3, . . . ,∞}, with the usual convention that a = b modulo ∞ means a = b. The condition in (1) means to have the equality bup− bdown = wup− wdown modulo s, for each block of our colored partition, where bup, bdown, wup, wdown count the black/white, up/down legs of the block. Observe that we have inclusions D∞(k, l) ⊂ Ds(k, l) ⊂ D1(k, l), for any s. More Here are a few examples of such sets of partitions. generally, we have an inclusion Ds(k, l) ⊂ Dt(k, l), whenever ts. Proposition 3.2. At s = 1, 2,∞, the sets Ds(k, l) are as follows: (1) D1(k, l) is the set of noncrossing partitions NC(k, l), with the legs arbitrarily bi- colored black or white, taken modulo the block-interchanging of colors. (2) D2(k, l) is the set of noncrossing partitions with even blocks NCeven(k, l), with the legs arbitrarily bicolored black or white, modulo the block-interchanging of colors. (3) D∞(k, l) is once again NCeven(k, l), this time colored such that the signed numbers of black and white legs are the same, modulo the block-interchanging of colors. Proof. This is clear from the fact that the numbers bup, bdown, wup, wdown, which must satisfy bup − bdown = wup − wdown modulo s, sum up to the size of the block. (cid:3) Consider now the Hilbert space H = C2n, with standard basis denoted {eI} = {eia}. Associated to any partition π ∈ Ds(k, l) is a linear map Tπ : H ⊗k → H ⊗l, as follows: Here the symbol δπ ∈ {0, 1} is defined as follows: we put the indices on the points of π, and we set δπ = 1 when for each block there exists an index I = ia such that all black indices are I = ia, and all white indices are ¯I = i¯a. Otherwise, we set δπ = 0. Tπ(eI1 ⊗ . . . ⊗ eIk) = XJ1...Jl . . . Ik . . . Jl(cid:19) eJ1 ⊗ . . . ⊗ eJl δπ(cid:18)I1 J1 10 TEODOR BANICA AND ADAM SKALSKI Proposition 3.3. The application π → Tπ transforms the horizontal concatenation, ver- tical concatenation and upside-down turning of colored diagrams into the tensor product, composition, and involution of linear maps (with a 2n factor for the circles). Proof. All the assertions are clear from definitions, except perhaps for the last assertion, concerning the value of the circles. So, consider the partition ∩, with legs colored bw. This partition belongs to Ds(0, 2) for any s, and the associated linear map is by definition for any s, and the associated linear map is T∪(eia ⊗ ejb) = δia,j¯b. Now since we have T∩(1) =Pia eia ⊗ ei¯a. Similarly, the partition ∪, with legs colored bw, belongs to Ds(2, 0) T∪T∩(1) = T∪(Pia eia ⊗ ei¯a) = #{ia} = 2n, this gives the result. In order to make the link with the quantum groups considered in the previous section, (cid:3) we will need a supplementary category of partitions, constructed in [8]: Definition 3.4. We let D∞(k, l) be the set of partitions in NCeven(k, l), with all legs colored black or white, with the following rules: (1) In each block, the top and bottom legs either both follow the pattern bwbw . . ., or both follow the pattern wbwb . . . (2) We agree to identify any colored partition with all the colored partitions obtained by interchanging the two colors, independently in each block. Once again, we can associate linear maps to such partitions, by the same formula as the one given before Proposition 3.3, and an analogue of that proposition holds. See [8]. Observe that for any s we have inclusions as follows: D∞(k, l) ⊂ D∞(k, l) ⊂ Ds(k, l) ⊂ D1(k, l) Here the first inclusion follows by comparing Definition 3.4 with Proposition 3.2 (3), and the other two inclusions were already noticed, before Proposition 3.2. As we will see in the proof below, the above inclusions correspond, via Woronowicz's Tannakian duality [36], to inclusions between the quantum groups we are interested in. Theorem 3.5. For the quantum group G+(dZ∗n Hom(u⊗k, u⊗l) = span(Tππ ∈ Ds(k, l)) s ), with 5 ≤ s < ∞ and n ≥ 2, we have: In addition, the linear maps Tπ appearing on the right are linearly independent. Proof. Let us first mention the fact that the result holds as well at s = 3. Regarding this case, or more generally all the special cases (s = 2, 3, 4,∞) we refer to section 6 below. So, let G be the quantum isometry group in the statement. Step 1. We first find a purely combinatorial formulation of the problem. We know from Theorem 2.5 that G ⊂ H + relations ξ ∈ F ix(u⊗k), where ξ =Pia e⊗s n0 appears as the subgroup presented by the ia . We also know from [8] that the category of QUANTUM ISOMETRY GROUPS 11 partitions for H + 1-block partition having s legs, we conclude that the category of partitions for G is: n0 is the above category D∞. Now since we have ξ = T1s, where 1s is the D′ s =< D∞, 1s > Summarizing, we have to prove that we have Ds = D′ s. Since the inclusion D′ clear from definitions, we are left with proving the inclusion Ds ⊂ D′ s. s =< D∞, 1s >. Step 2. We construct some useful elements in D′ Consider the 1-block partition 1s ∈ NC(0, s), with all legs colored black. By capping it , we obtain a 1-block partition in NC(1, s−1), ◦ at right, from below, with the partition • ??? with the upper leg white, and all the lower legs black (in the example s = 5): s ⊂ Ds is ??? • RRRRR lllll • • • • = • ◦ ??? ◦ HHHHHHHHH 666666 vvvvvvvvv • • • • We can repeat this process, and for any x ∈ {1, . . . , s} we obtain a partition in NC(x, s− x), having all upper legs white, and all the lower legs black. We call these partitions "forks". ) we obtain the partition • Consider now once again the 1-block partition 1s ∈ NC(0, s), with all legs colored black. By capping it from below in the center with a 'colour inverted' fork in NC(s− 2, 2) • • (so for s = 5 it is • ??? ??? ??? • OOOOO , which we call a "comb". ooooo ◦ ◦ ◦ ◦ Step 3. We prove first that we have D∞ ⊂ D′ s. Since D∞ is generated by its blocks, it is enough to show that any 1-block partition in D∞ belongs to D′ s. Also, by Frobenius duality, which diagramatically corresponds to rotating the legs, we can restrict attention to the partitions having no upper legs. So, what we want to prove is that any 1-block partition π ∈ D∞(0, k) is in D′ s. Now recall the definition of D∞: these are the partitions such that in each block, the signed number of black legs equals the signed number of white legs. Now since our partition π has just one block, and no upper legs, the condition π ∈ D∞(0, k) simply tells us that the number of black legs should equal the number of white legs. We proceed by recurrence on the number of legs (note that if k = 2 then π ∈ D∞, so we can start from k = 4). First, by using a rotation, we can assume that the last 2 legs at right have different colors, say bw. Depending now on the color of the leg preceding these 2 last legs, we have two cases: (1) Case π = Rbbw, with R word in b, w. supposed by recurrence to be in D′ in D′ are done. ◦ s by Step 2, and we cap in the middle with • ??? s, the colour inverted comb ◦ In this case we put at right of π′ = Rb, ??? ◦ OOOOO , which is as well ooooo• • from below. We obtain π, and we (2) Case π = Rwbw, with R word in b, w. In this case we put at right of π′ = Rw, s, and from below. We obtain once again π, and we are done. supposed by recurrence to be in D′ • we cap in the middle with ◦ ??? ??? ◦ OOOOO , which is in D∞ ⊂ D′ ooooo◦ • s, the partition • 12 TEODOR BANICA AND ADAM SKALSKI Step 4. We prove now that we have indeed Ds ⊂ D′ s. According to the result in Step 3, it suffices to show that we have Ds ⊂< D∞, 1s >. In order to do this, we proceed as in Step 3: since Ds is generated by its blocks, and by Frobenius duality, it is enough to check that any 1-block partition π ∈ D∞(0, k) is in < D∞, 1s >. That is, we have to show that any 1-block partition π ∈ D∞(0, k) whose number of black legs equals the number of white legs modulo s is in < D∞, 1s >. Let then π ∈ Ds be a 1-block element, say in NC(2p + ks), where p, k ∈ N and p denotes the number of black legs in π. It suffices to observe that it can be always realised as a capping of a 1-block partition π′ ∈ D∞ ∩ NC(2p + 2ks), given by π followed by ks black legs, by k copies of our partition 1s with all black legs from below and on the right. A simple example is given below: RRRRRRRRRRR LLLLLLLL 999999 lllllllllll • rrrrrrrr ◦ ◦ ◦ ◦ ◦ ◦ = WWWWWWWWWWWWWWWWWWWWW TTTTTTTTTTTTTTT VVVVVVVVVVVVVVVVVV PPPPPPPPPPPP HHHHHHHHH 666666 ggggggggggggggggggggg • hhhhhhhhhhhhhhhhhh ◦ jjjjjjjjjjjjjjj ◦ nnnnnnnnnnnn ◦ vvvvvvvvv ◦ • • • • • ◦ ◦ • • • • •lllll RRRRR ??? Step 5. It remains to prove that the maps Tπ are linearly independent. The linear independence is a well-known issue, and can be solved by using a standard trick, namely a positivity argument involving a trace, going back to Jones' paper [25]. The idea here is to solve first the problem in the case k = l, where the span of the diagrams has a natural structure of C ∗-algebra, and then to use Frobenius duality and the natural embeddings Ds(k, l) → Ds(k, l + s), in order to get the result for any k, l. namely the partitions in NC, bicolored, modulo the block-interchanging of colors. However, in our case we can simply use some previously known results, as follows. Since we have Ds(k, l) ⊂ D1(k, l), our problem is actually about the diagrams in D1, So, let us recall from [8] that if S+ n0 denotes the quantum group presented by the relations making all the standard generators uia,jb projections, then we have Hom(u⊗k, u⊗l) = span(Tππ ∈ D1(k, l)). We refer as well to [8] for the precise relation between S+ n is the hyperoctahedral quantum group constructed in [3], taken with its sudoku representation. indeed, the dimension of Hom(u⊗k, u⊗l) is known from [3], and since a basic diagram count, performed in [8], shows that this is exactly the number of diagrams in D1(k, l), this gives the result. (cid:3) Let us also recall from [8] that we have an isomorphism S+ n0 and the quantum permutation group S+ n , constructed by Wang in [34]. With these results in hand, the linear independence is clear: n0 ⊂ H + n0 ≃ H + n , where H + As a first comment, there should be as well a third proof for the linear independence, just by showing that the Gram determinant of the linear maps Tπ is nonzero. Indeed, two key "Temperley-Lieb" determinants were computed in [22], [30], and, according to the results in [4], these computations can be usually extended to objects of type D1. Note however that the computation of the Gram determinant for Ds with s ≥ 3 is probably quite a difficult task, because the size of the matrix, which will be shown to be a moment of a certain compound Poisson law, is a quite complicated combinatorial object. So, computing this Gram determinant is a question that we would like to raise here. QUANTUM ISOMETRY GROUPS 13 Finally, let us mention that these Gram determinant questions are of particular interest in relation with the Weingarten formula [19]. See [6] for some potential applications. 4. Poisson laws In the remainder of this paper we present a concrete application of Theorem 3.5, namely fundamental representation, χ = Σuii, with respect to the Haar functional. s ). This measure is by definition the spectral measure of the character of the the computation of the Kesten measure of the discrete quantum group Λ = bG, where G = G+(dZ∗n For the moment, let us start with some probabilistic preliminaries. It is known from [2], [5], [6], [8] that what we can expect to find as Kesten measures for our quantum groups should be some versions of the free Poisson law. So, in this section we will study in detail the free Poisson law, and its "compound" versions, first constructed in [24], [29]. Definition 4.1. The Poisson law pt and the free Poisson law πt are the real probability measures appearing when performing a Poisson limit, respectively a free Poisson limit pt = lim n→∞(cid:18)(cid:18)1 − 1 n(cid:19) δ0 + 1 n δt(cid:19)∗n πt = lim n→∞(cid:18)(cid:18)1 − 1 n(cid:19) δ0 + 1 n δt(cid:19)⊞n A well-known computation shows that at t = 1 the Poisson law is p1 = 1 where ∗ is the usual convolution of measures, and ⊞ is Voiculescu's free convolution [31]. δk k! , √1 − 4x−1 dx. Note that π1 is the same as the and that the free Poisson law is π1 = 1 2π Marchenko-Pastur law, discovered in a slightly different context in [26]. See [27]. eP∞ k=1 More generally, let ρ be a compactly supported positive measure on R. By replacing in the above formulae δt by ρ, we obtain measures called compound Poisson/free Poisson laws. Definition 4.2. Associated to any compactly supported positive measure ρ on R are the probability measures pρ = lim n→∞(cid:18)(cid:16)1 − c n(cid:17) δ0 + 1 n ρ(cid:19)∗n πρ = lim n→∞(cid:18)(cid:16)1 − c n(cid:17) δ0 + 1 n ρ(cid:19)⊞n where c = mass(ρ), called compound Poisson and compound free Poisson laws. In what follows we will be interested in the case where ρ is discrete, as is for instance the case for ρ = δt with t > 0, which produces the Poisson and free Poisson laws. We recall that the usual convolution operation ∗ is linearized by log F , where F is the Fourier transform. There are several ways of defining and normalizing F , and for the purposes of this paper, we will take as definition Fδz (y) = e−iyz, on the Dirac masses. We recall also from Voiculescu [31] that the free convolution operation ⊞ is linearized by the R-transform, constructed as follows: first, we let f (y) = 1 + m1y + m2y2 + . . . be the moment generating function of our measure, so that G(ξ) = ξ−1f (ξ−1) is the Cauchy transform; then we set R(y) = K(y) − y−1, where K(y) is such that G(K(y)) = y. In the free case now, we use a similar method. First, we have: fµn(y) =(cid:16)1 − c n(cid:17) + 1 n sXi=1 ci 1 − ziy cie−iyzi!n c 1 n n (y) = (cid:16)1 − n(cid:17) + sXi=1 =⇒ Fpρ(y) = exp sXi=1 ci(e−iyzi − 1)! n(cid:17) 1 =⇒ Gµn(ξ) =(cid:16)1 − n(cid:17) =⇒ y =(cid:16)1 − sXi=1 sXi=1 Kµn(y) 1 n 1 n + + ξ ci c 1 c ξ − zi ci Kµn(y) − zi 14 TEODOR BANICA AND ADAM SKALSKI The following result allows one to detect compound Poisson/free Poisson laws. Lemma 4.3. For ρ =Ps i=1 ciδzi with ci > 0 and zi ∈ R, we have Fpρ(y) = exp sXi=1 ci(e−iyzi − 1)! Rπρ(y) = sXi=1 cizi 1 − yzi where F, R denote respectively the Fourier transform, and Voiculescu's R-transform. Proof. Let µn be the measure appearing in Definition 4.2, under the convolution signs. In the classical case, we have the following well-known computation: Fµn(y) =(cid:16)1 − c n(cid:17) + 1 n sXi=1 cie−iyzi =⇒ Fµ∗n Now since Kµn(y) = y−1 + Rµn(y) = y−1 + R/n, where R = Rµ⊞n n (y), we get: y =(cid:16)1 − =⇒ 1 =(cid:16)1 − c n(cid:17) n(cid:17) c 1 y−1 + R/n + 1 n 1 1 + yR/n + 1 n sXi=1 sXi=1 ci y−1 + R/n − zi ci 1 + yR/n − yzi Now multiplying by n, rearranging the terms, and letting n → ∞, we get: ci c + yR ci 1 + yR/n 1 + yR/n − yzi =⇒ c + yRπρ(y) = = sXi=1 This finishes the proof in the free case, and we are done. =⇒ Rπρ(y) = sXi=1 1 − yzi sXi=1 cizi 1 − yzi (cid:3) We have as well the following result, which provides an alternative to Definition 4.2. QUANTUM ISOMETRY GROUPS 15 Theorem 4.4. For ρ =Ps i=1 ciδzi with ci > 0 and zi ∈ R, we have pρ/πρ = law sXi=1 ziαi! where the variables αi are Poisson/free Poisson(ci), independent/free. Proof. Observe first that the result holds for ρ = δt, with t > 0. Also, it was shown in [2] that an analogue of this result holds for ρ = uniform measure on the s-th roots of unity (note however that the latter measure is not supported on R). In the general case, let α be the sum of Poisson/free Poisson variables in the statement. We will show that the Fourier/R-transform of α is given by the formulae in Lemma 4.3. Indeed, by using some well-known Fourier transform formulae, we have: Fαi(y) = exp(ci(e−iy − 1)) =⇒ Fziαi(y) = exp(ci(e−iyzi − 1)) ci(e−iyzi − 1)! =⇒ Fα(y) = exp sXi=1 Also, by using some well-known R-transform formulae, we have: Rαi(y) = ci 1 − y =⇒ Rziαi(y) = =⇒ Rα(y) = cizi 1 − yzi cizi 1 − yzi sXi=1 Thus we have indeed the same formulae as those in Lemma 4.3, and we are done. (cid:3) Remark 4.5. It is tempting to replace in the above results compactly supported positive measures on R by their counterparts supported on C. It is easy to see that the classical parts pass through without any modifications; it is the free case that offers a challenge. The free convolution is defined in terms of the distribution of the sum of two free noncom- mutative random variables. If the variables in question are selfadjoint (which corresponds to the measures being supported on R), then so is their sum. There is however no reason why the sum of normal operators should be normal, so, although the formal computations do not change, it is not clear how they should be interpreted. We thank the referee for pointing this out. 5. Spectral measures Let us go back now to our quantum group problematics. We recall from the beginning of section 3 that most of the representation theory invariants of a quantum group (G, u) are known to be encoded in the structure of the linear spaces Hom(u⊗k, u⊗l). Here u denotes the fundamental corepresentation of C(G), and we assume that we have n0 from [8], u = ¯u as corepresentations. Note that this is true for the quantum group H + [15], hence is true as well for the quantum groups investigated in this paper. 16 TEODOR BANICA AND ADAM SKALSKI The main representation theory invariant is the dimension of the above spaces. By Woronowicz's general results in [35] we have dim(Hom(u⊗k, u⊗l)) = h(χk+l), where χ = Σuii is the character of u, and where h : C(G) → C is the Haar functional. more convenient to specify directly the law of χ. So, the main problem is to compute the moments of χ. For several reasons, it is actually Definition 5.1. The spectral measure of a compact quantum group (G, u) with u = ¯u is the real probability measure µ given by R ϕ(x) dµ(x) = h(ϕ(χ)), where χ = Σuii is the character of the fundamental representation, and h : C(G) → C is the Haar functional. For a number of other interpretations of this key measure, and a number of motivations for its explicit computation, we refer for instance to [1], [5]. We are now in position to state and prove our main result. Let ρ → ρ the projection map from complex to real measures, given on Dirac masses by δz → δRe(z). Theorem 5.2. The spectral measure of G+(dZ∗n law(χ/2) = πε/2 s ) with 5 ≤ s < ∞ and n ≥ 2 is given by where ε is the uniform measure on the s-roots of unity. Proof. According to the general theory developed by Woronowicz in [35], the moments of the character χ = Σuii are given by: Z χr = dim(F ix(u⊗r)) So, we have to count the diagrams in Theorem 3.5. If we denote by κr the number of ways of coloring an r-block, then the number of diagrams in Ds(0, r) is: mr = Xπ∈N C(r)Yb∈π κ#b According now to Speicher's results in [28] (see Proposition 11.4 (3) in [27]), we can con- clude that the free cumulants of the measure under consideration are the above numbers κr. We will write p = q(s) to indicate that p = q modulo s. Since for coloring a r-block, as in Definition 3.1, what we have to do is just to pick two numbers p, q satisfying p + q = r and p = q(s), then choose p legs of our block and color them black, and color the remaining q legs white, remembering that the colors can be also exchanged, the numbers κr are given by: κr = 1 p(cid:19) 2 Xp+q=r, p=q(s)(cid:18)r Consider now the root of unity w = e2πi/s. Our claim is that we have: QUANTUM ISOMETRY GROUPS 17 Indeed, the equality between the above two quantities can be checked as follows: (wk + w−k)r = 1 s sXk=1 κr = 1 2s 1 s (wk + w−k)r sXk=1 p(cid:19)(wk)p(w−k)r−p rXp=0(cid:18)r sXk=1 (w2p−r)k! p(cid:19) 1 rXp=0(cid:18)r sXk=1 p(cid:19) = Xp+q=r, p=q(s)(cid:18)r = s We use now the fact, once again originally proved in Speicher's paper [28], that the free cumulants are the coefficients of the R-transform (actually in [27] the R-transform is first introduced as the power series with the coefficients given by the free cumulants and only later related to the functional equations). So, by using the above formula, we conclude that the R-transform Rχ(y) = κ1 + κ2y + κ3y2 + . . . of our measure is given by: Now by using the well-known dilation formula Raχ(y) = aRχ(ay), we obtain: Rχ(y) = 1 2s wk + w−k 1 − (wk + w−k)y Rχ/2(y) = 1 4s wk + w−k 1 − (wk + w−k)y/2 sXk=1 sXk=1 1 2s sXk=1 sXk=1 sXk=1 Consider now the uniform measure on the s-th roots of unity, ε = 1 k=1 δwk. Its image via the projection map from complex to real measures, δz → δRe(z), is given by: sPs ε = 1 s δ(wk+w−k)/2 Thus the real measure appearing in the statement is: By using now Lemma 4.3, we have: Rπε/2(y) = 1 4s ε/2 = δ(wk+w−k)/2 wk + w−k 1 − (wk + w−k)y/2 18 TEODOR BANICA AND ADAM SKALSKI Thus we have Rχ/2(y) = Rπε/2(y), and this gives the conclusion in the statement. (cid:3) We can reformulate the above statement in the following way. s with 5 ≤ s < ∞ and n ≥ 2, the main character of the Proposition 5.3. For Γ = Z∗n quantum isometry group G+(bΓ) decomposes as χ = 2 cos(cid:18)2kπ s (cid:19) αk sXk=1 where α1, . . . , αs are free Poisson variables of parameter 1/(2s), free. Proof. This follows from Theorem 4.4 and from Theorem 5.2. Indeed, if α1, . . . , αk are free Poisson variables as in the statement, we have: law(χ/2) = πε/2 = π{ 1 2s Ps k=1 δ(wk +w−k )/2} = law sXk=1 wk + w−k 2 αk! Thus χ has the same law as twice the variable on the right, and we are done. (cid:3) As a first observation, the above results remind those previously found in [2], [5], where the free Poisson law, and its compound versions, also play a central role. n n, H s+ n , H (s) More precisely, for the quantum groups H s considered in [2], [5], the cor- responding asymptotic spectral measures are pε, πε, pε respectively, where ε → ε is the "squeezing" operation δz → δz, from the complex to the real probability measures. s ), while being not exactly "easy" in the sense of [2], [5], seems to belong to the same circle of ideas as the general "hyperoctahedral series" of easy quantum groups, whose existence was conjectured in [5]. This suggests that the hyperoctahedral series, or at least a big part of it, might appear via some simple algebraic So, the quantum group G+(dZ∗n manipulations from the quantum isometry groups of type G+(bΓ), with Γ of the form Zs1 ∗ . . . ∗ Zsn, or perhaps a bit more general. We do not know if it is so. In this section we discuss the special cases s = 2, 3, 4,∞. We begin with some results at s = 2, 3, 4, refining and generalizing those found in [15], at n = 1. 6. Special cases 2 ) is the hyperoctahedral quantum group H + n from [3]. Proposition 6.1. G+(dZ∗n Proof. Let g1, . . . , gn be the standard generators of Z∗n 2 . We write the coaction of the quantum isometry group in the statement as α(gi) = Σjgj ⊗ uji. Then, according to the general results in section 1, the relations between the generators uij are as follows: (1) First, we have the relations uij = u∗ (2) Second, we have ujiuki = 0 for j 6= k, and Σju2 ij (coming from gi = g∗ i ). ji = 1 (coming from g2 i = 1). QUANTUM ISOMETRY GROUPS 19 From (2) we get uki = Pj u2 ki, and together with (1), this shows that the elements uij are partial isometries. With this observation in hand, the above relations are simply the "cubic" relations in [3], for partial isometries, and this gives the result. (cid:3) jiuki = u3 We refer to [3] for a full discussion of the representation theory invariants of H + n , and to [2], [6] for some technical versions and generalizations of these computations. Proposition 6.2. Theorems 2.5, 3.5, 5.2 hold as well at s = 3, and give the presentation, diagrams, and spectral measure for the quantum group G+(dZ∗n Proof. This is clear from a careful examination of the proofs of the above results: the beginning of the proof of Theorem 2.5 holds as well at s = 3, because here we do not have length 2 relations either, and the rest of the proofs simply hold, unchanged. (cid:3) 3 ). At s = 4 now, the situation is quite complicated, and we currently do not have general results. Let us just state and prove the following result, slightly improving a result in [15]. Proposition 6.3. We have a ∗-algebra isomorphism C(G+(cZ4)) ≃ C ∗(D∞ × Z2). Proof. We use the notations in [15]. According to the results there, the fundamental corepresentation of our quantum isometry group must be of the following type: The generators must satisfy the defining relations for SU −1 2 , namely: AA∗ + BB∗ = 1, AB + BA = 0, AB∗ + B∗A = 0 In addition, the operator T = A2 + B2 must satisfy the following relations: T = T ∗, T 2 = 1, T A = A∗, T B = B∗ See [15]. The point now is that, since T commutes with A, B, A∗, B∗, we can write: B∗ A∗(cid:19) u =(cid:18) A B T =(cid:18)1 0 −1(cid:19) 0 Thus the quantum isometry algebra in the statement decomposes naturally as a direct sum of two ∗-algebras, and by passing to the generators A + B, A − B, we see that each of these ∗-algebras is isomorphic to the group algebra C ∗(D∞). This gives the result. (cid:3) is probably true, but unfortunately we do not see an analogue of this result, for n > 1. Let us discuss now the case s = ∞. We have two statements here. The above result suggests that G+(cZ4) might be a twist of the dual of D∞ × Z2. This Proposition 6.4. G+(bFn) is the quantum group H + Proof. This result, and the precise relation between H + n0 from [8]. n0 and H + n , are explained in [8]. (cid:3) 20 TEODOR BANICA AND ADAM SKALSKI Our second statement concerns the "representation-theoretic" limit, with s → ∞, of s ). Quite surprisingly, this limiting quantum group is not the n0 from [8]. Its main properties can be summarized as follows. quantum group H + the quantum groups G+(dZ∗n Theorem 6.5. Let Kn be the quantum subgroup of H + all the standard generators uij normal. Then: n0 presented by the relations making (1) Hom(u⊗k, u⊗l) = span(Tππ ∈ D∞(k, l)), for any k, l ∈ N. (2) law(χ/2) = πε/2, where ε is the uniform measure on the unit circle. Proof. We recall from section 3 above that we have inclusions D∞(k, l) ⊂ D∞(k, l), for any k, l. The idea will be to find first a categorical generation result of type D∞ =< D∞, π >, and then to find the quantum group and probabilistic interpretations of this result. So, let π ∈ D∞(0, 4) be the 1-block partition colored bbww. Our claim is that we have D∞ =< D∞, π >. Since the inclusion ⊃ is clear, we just have to prove the inclusion ⊂. In order to do this, let us look at the category D =< D∞, π >. By using Frobenius duality, i.e. by using rotations, we can deduce from π ∈ D the fact that D contains all the 1-block partitions having 0 upper legs and 4 lower legs, with the 4 lower legs colored half-black, half-white. So, in other words, we have D(0, 4) = D∞(0, 4). Now by using once again Frobenius duality, i.e. by using rotations, we deduce that we have D(2, 2) = D∞(2, 2). In particular, if σ ∈ D∞(2, 2) denotes the 1-block partition with the upper legs colored bw and the lower legs colored wb, then σ ∈ D(2, 2). With this observation in hand, we can finish the proof of the above claim: indeed, by using σ we can exchange the position of two consecutive black and white colors in any partition in D, and it is easy to see that this operation allows one to construct any element in D∞ starting from the elements of D∞. So, we have D∞ =< D∞, π > as claimed. (1) Let us try to compute the quantum group K ′ n0 having as Hom spaces the linear spaces span(Tππ ∈ D∞(k, l)). Since we have D∞ =< D∞, π >, this quantum n0 is simply the one presented by the relations coming from Tπ ∈ F ix(u⊗4). group K ′ Now, according to the definitions in section 3 above, we have: n ⊂ H + n ⊂ H + Tπ =Xia eia ⊗ eia ⊗ ei¯a ⊗ ei¯a By using the relations in Proposition 2.3, we see that this vector is fixed by u⊗4 if and only if the following condition is satisfied, where pia,jb = uia,jbu∗ ia,jb, qia,jb = u∗ ia,jbuia,jb: qia,jbpia,kcqia,ld =(qia,jb 0 if (jb) = (kc) = (ld) otherwise This condition is easily seen to be equivalent to the "normality" relations pia,jb = qia,jb, n that we are currently computing is nothing but for any i, j, a, b, so the quantum group K ′ the quantum group Kn appearing in the statement, and we are done. QUANTUM ISOMETRY GROUPS 21 (2) This follows from (1), as in the proof of Theorem 5.2, after of course performing (cid:3) some obvious modifications in the preliminary material in sections 4 and 5. One question arising from the above result is that of finding a suitable geometric inter- pretation of Kn. Since this is a subgroup of H + n0 = G+(bFn), we can expect Kn to consist of the "quantum isometries" of bFn preserving not only the length, but also some "extra structure". A careful study here leads to the following informal answer: "under the action of Kn a word in Fn = Z∗· · ·∗ Z should get sent to the words of the same length, with the switches between the generators coming from various copies of Z appearing at the same places". This will be explained in detail in a future paper from the present series. 7. Concluding remarks We have seen in this paper that the representation theory invariants of G+(bΓ) can be explicitly computed in the case Γ = Z∗n s , by using diagrammatic techniques. The answer that we obtain -- an explicit formula for the spectral measure, as a compound free Poisson law πε -- appears to be quite interesting from the point of view of free probability. There are several questions arising from the present work. The main one is probably the computation of the invariants for general free products of cyclic groups, Γ = Zs1 ∗. . .∗ Zsn. Here we can definitely expect to have some very interesting, new combinatorics, ultimately coming from the arithmetic properties of the sequence of indices s1, . . . , sn. Understanding this next-step combinatorics looks like an important task towards the general understanding of the quantum isometry groups introduced in [23], and of the easy quantum groups introduced in [9]. We intend to come back to this fundamental question, at least with partial results, in some future work. References [1] T. Banica, Quantum automorphism groups of homogeneous graphs, J. Funct. Anal. 224 (2005), 243 -- 280. [2] T. Banica, S.T. Belinschi, M. Capitaine and B. Collins, Free Bessel laws, Canad. J. Math. 63 (2011), 3 -- 37. [3] T. Banica, J. Bichon and B. Collins, The hyperoctahedral quantum group, J. Ramanujan Math. Soc. 22 (2007), 345 -- 384. [4] T. Banica and S. Curran, Decomposition results for Gram matrix determinants, J. Math. Phys. 51 (2010), 1 -- 14. [5] T. Banica, S. Curran and R. Speicher, Classification results for easy quantum groups, Pacific J. Math. 247 (2010), 1 -- 26. [6] T. Banica, S. Curran and R. Speicher, Stochastic aspects of easy quantum groups, Probab. Theory Related Fields, to appear. [7] T. Banica and D. Goswami, Quantum isometries and noncommutative spheres, Comm. Math. Phys. 298 (2010), 343 -- 356. [8] T. Banica and A. Skalski, Two-parameter families of quantum symmetry groups, J. Funct. Anal. 260 (2011), 3252 -- 3282. [9] T. Banica and R. Speicher, Liberation of orthogonal Lie groups, Adv. Math. 222 (2009), 1461 -- 1501. 22 TEODOR BANICA AND ADAM SKALSKI [10] J. Bhowmick, F. D'Andrea and L. Dabrowski, Quantum isometries of the finite noncommutative geometry of the standard model, arxiv:1009.2850. [11] J. Bhowmick and D. Goswami, Quantum isometry groups: examples and computations, Comm. Math. Phys. 285 (2009), 421 -- 444. [12] J. Bhowmick and D. Goswami, Quantum group of orientation-preserving Riemannian isometries, J. Funct. Anal. 257 (2009), 2530 -- 2572. [13] J. Bhowmick and D. Goswami, Quantum isometry groups of the Podles spheres, J. Funct. Anal. 258 (2010), 2937 -- 2960. [14] J. Bhowmick, D. Goswami and A. Skalski, Quantum isometry groups of 0-dimensional manifolds, Trans. Amer. Math. Soc. 363 (2011), 901 -- 921. [15] J. Bhowmick and A. Skalski, Quantum isometry groups of noncommutative manifolds associated to group C∗-algebras, J. Geom. Phys. 60 (2010), 1474 -- 1489. [16] J. Bichon, Quantum automorphism groups of finite graphs, Proc. Amer. Math. Soc. 131 (2003), 665 -- 673. [17] D. Bisch and V.F.R. Jones, Algebras associated to intermediate subfactors, Invent. Math. 128 (1997), 89 -- 157. [18] E. Christensen and C. Ivan, Spectral triples for AF C∗-algebras and metrics on the Cantor set, J. Operator Theory 56 (2006), 17 -- 46. [19] B. Collins and P. ´Sniady, Integration with respect to the Haar measure on unitary, orthogonal and symplectic groups, Comm. Math. Phys. 264 (2006), 773 -- 795. [20] A. Connes, Noncommutative geometry, Academic Press (1994). [21] A. Connes and M. Marcolli, Noncommutative geometry, quantum fields and motives, AMS (2008). [22] P. Di Francesco, O. Golinelli and E. Guitter, Meanders and the Temperley-Lieb algebra, Comm. Math. Phys. 186 (1997), 1 -- 59. [23] D. Goswami, Quantum group of isometries in classical and noncommutative geometry, Comm. Math. Phys. 285 (2009), 141 -- 160. [24] F. Hiai and D. Petz, The semicircle law, free random variables and entropy, AMS (2000). [25] V.F.R. Jones, Index for subfactors, Invent. Math. 72 (1983), 1 -- 25. [26] V.A. Marchenko and L.A. Pastur, Distribution of eigenvalues in certain sets of random matrices, Mat. Sb. 72 (1967), 507 -- 536. [27] A. Nica and R. Speicher, Lectures on the combinatorics of free probability, Cambridge Univ. Press (2006). [28] R. Speicher, Multiplicative functions on the lattice of noncrossing partitions and free convolution, Math. Ann. 298 (1994), 611 -- 628. [29] R. Speicher, Combinatorial theory of the free product with amalgamation and operator-valued free probability theory, Mem. Amer. Math. Soc. 132 (1998). [30] W.T. Tutte, The matrix of chromatic joins, J. Combin. Theory Ser. B 57 (1993), 269 -- 288. [31] D.V. Voiculescu, Addition of certain noncommuting random variables, J. Funct. Anal. 66 (1986), 323 -- 346. [32] D.V. Voiculescu, K.J. Dykema and A. Nica, Free random variables, AMS (1992). [33] S. Wang, Free products of compact quantum groups, Comm. Math. Phys. 167 (1995), 671 -- 692. [34] S. Wang, Quantum symmetry groups of finite spaces, Comm. Math. Phys. 195 (1998), 195 -- 211. [35] S.L. Woronowicz, Compact matrix pseudogroups, Comm. Math. Phys. 111 (1987), 613 -- 665. [36] S.L. Woronowicz, Tannaka-Krein duality for compact matrix pseudogroups. Twisted SU(N) groups, Invent. Math. 93 (1988), 35 -- 76. QUANTUM ISOMETRY GROUPS 23 T.B.: Department of Mathematics, Cergy-Pontoise University, 95000 Cergy-Pontoise, France. [email protected] A.S.: Institute of Mathematics of the Polish Academy of Sciences, ul. ´Sniadeckich 8, 00-956 Warszawa, Poland. [email protected]
1707.01782	4	1707	2018-10-16T12:14:48	The Calkin algebra is $\aleph_1$-universal	[ "math.OA", "math.LO" ]	We discuss the existence of (injectively) universal C-algebras and prove that all C-algebras of density character $\aleph_1$ embed into the Calkin algebra, $Q(H)$. Together with other results, this shows that each of the following assertions is relatively consistent with ZFC: (i) $Q(H)$ is a $2^{\aleph_0}$-universal C-algebra. (ii) There exists a $2^{\aleph_0}$-universal C-algebra, but $Q(H)$ is not $2^{\aleph_0}$-universal. (iii) A $2^{\aleph_0}$-universal C-algebra does not exist. We also prove that it is relatively consistent with ZFC that (iv) there is no $\aleph_1$-universal nuclear C-algebra, and that (v) there is no $\aleph_1$-universal simple nuclear C*-algebra.	math.OA	math	THE CALKIN ALGEBRA IS ℵ1-UNIVERSAL ILIJAS FARAH, ILAN HIRSHBERG, AND ALESSANDRO VIGNATI Abstract. We discuss the existence of (injectively) universal C∗-algebras and prove that all C∗-algebras of density character ℵ1 embed into the Calkin algebra, Q(H). Together with other results, this shows that each of the following assertions is relatively consistent with ZFC: (i) Q(H) is a 2ℵ0 -universal C∗-algebra. (ii) There exists a 2ℵ0 -universal C∗-algebra, but Q(H) is not 2ℵ0 -universal. (iii) A 2ℵ0 -universal C∗-algebra does not exist. We also prove that it is relatively consistent with ZFC that (iv) there is no ℵ1-universal nuclear C∗-algebra, and that (v) there is no ℵ1-universal simple nuclear C∗-algebra. 1. Introduction Let H denote the separable infinite-dimensional complex Hilbert space. The Calkin algebra Q(H) is the quotient B(H)/K (H) of the algebra B(H) of all bounded linear operators on H over the ideal of all compact operators. Given a category C of metric structures and a cardinal κ, an object A ∈ C is (injectively)1 κ-universal if it has density character κ2 and every object B ∈ C of density character at most κ is isometric to a substructure of A. The question whether the Calkin algebra can be ℵ1-universal for the cat- egory of C∗-algebras answered in Theorem A was raised by Piotr Koszmider (personal correspondence) and in [43, Question E]. Theorem A. All C∗-algebras of density character at most ℵ1 embed into the Calkin algebra. Therefore the Continuum Hypothesis implies that the Calkin algebra is an ℵ1-universal C∗-algebra. Date: October 17, 2018. 2010 Mathematics Subject Classification. 46L05, 03E35, 03E75. Key words and phrases. C∗-algebras, Calkin algebra, ℵ1-universal, extension theory. IH and AV's visit to Toronto were supported by NSERC. IH was supported by the Israel Science Foundation, grant no. 476/16. IF's visit to CRM was supported by the Clay Mathematics Institute. AV is supported by a PRESTIGE co-fund Scholarship and an FWO scholarship. 1The dual notion, surjective universality, is trivialized in the category of unital C∗- algebras. Since every unital C∗-algebra is generated by its unitary group, the full group C∗-algebra associated with the free group Fκ, C∗(Fκ), is surjectively κ-universal for every infinite cardinal κ, and, in the abelian setting, C([0, 1]κ) is surjectively κ-universal. 2The density character of a metric space is the smallest cardinality of a dense subset. 1 2 ILIJAS FARAH, ILAN HIRSHBERG, AND ALESSANDRO VIGNATI One of the ingredients of our proof is the analysis of the Extw-group of simple, separable, and unital C∗-algebras that tensorially absorb the Cuntz algebra O2. By a fundamental result of Kirchberg ([27]), O2 is the universal separable nuclear C∗-algebra, and even the universal separable exact C∗-algebra. The following theorem (and more; see Theorem 4.3) will be proved in §4. Theorem B. It is relatively consistent with ZFC that there is no ℵ1- universal nuclear C∗-algebra, and that there is no ℵ1-universal nuclear, sim- ple, C∗-algebra. Our proof of Theorem B requires basic command of the method of forcing, as exposed e.g., in [30, IV]. Acknowledgments. We would like to thank George Elliott, Jamie Gabe, Piotr Koszmider, N. Christopher Phillips, and Chris Schafhauser for helpful remarks and to Bradd Hart for pointing out to some omissions in the early draft of the present paper. We would also like to thank Bartosz Kwasniewski for bringing [31], used in the proof of Lemma 5.2, to our attention. Part of this work was completed during AV's and IH's visits to Toronto in the sum- mer of 2017 and the authors' visit to the program "IRP Operator Algebras: Dynamics and Interactions" at CRM (Barcelona). The authors would like to thank the organizers, in particular Francesc Perera, for their support and hospitality. 2. The proof of Theorem A This section is entirely devoted to the proof of Theorem A. Familiarity with model theory, in particular axiomatizability and the different layers of saturation, is required (see [15], or [14] for an overview of the concept of saturation). For information on C∗-algebras see [3] and for analytic K- homology see [24]. The main technical difficulty in the proof of Theorem A is posed by the absence of reasonable saturation properties in the Calkin algebra. The sim- plest instance of this is the fact that the image of the unilateral shift in Q(H) is a unitary with full spectrum and no square root. As pointed out in the introduction to [36], this implies that Q(H) is not injective (in a categorical sense) for separable C∗-algebras and complicates construction of outer au- tomorphisms of Q(H). More sophisticated obstructions to saturation, and even homogeneity, in Q(H) were exhibited in [14, §4] and [17], respectively. All of these obstructions are of K-theoretic nature. A unital C∗-algebra A is purely infinite and simple if it is infinite di- mensional and for every nonzero positive a ∈ A there is x ∈ A such that xax∗ = 1. The Cuntz algebra O2 is the universal C∗-algebra generated by two isometries s and t satisfying s∗s = t∗t = 1 = ss∗ + tt∗. THE CALKIN ALGEBRA IS ℵ1-UNIVERSAL 3 Let O = {A : A is unital, purely infinite, simple, and A ⊗ O2 ∼= A}. (Since O2 is nuclear, there is no ambiguity in what tensor product is used. In this case there is a unique C∗-norm on the algebraic tensor product.) Lemma 2.1. Every C∗-algebra A embeds into a C∗-algebra B ∈ O of the same density character as A. If A is unital then the embedding can be chosen to be unital. We provide two proofs of Lemma 2.1, one of model-theoretic and one of operator-algebraic flavour. The first proof of Lemma 2.1. The class O is separably axiomatizable by ([15, Theorem 2.5.1 and Theorem 2.5.2]). We first consider the case when A is separable. Then A is isomorphic to a subalgebra of Q(H), and the embedding can be chosen to be unital if A is unital. By the downward Lowenheim -- Skolem theorem ([15, Theorem 2.6.2]) we can find a separable elementary submodel C of Q(H) into which A embeds. Then C is simple and purely infinite, and C ⊗ O2 is as required. Now suppose A is not separable and let κ be its density character. Again by the downward Lowenheim -- Skolem theorem we can find a separable ele- mentary submodel A0 of A. By the first paragraph, we can find a separable B0 ∈ O into which A0 embeds. By the standard elementary chain argument ([1, Proposition 7.10]) we construct a κ-saturated elementary extension B1 of B0. Writing A as a union of an elementary chain of submodels of density character < κ and using the saturation of B1, we can embed A into B1. Again by the downward Lowenheim -- Skolem theorem we can find B2 of den- sity character κ such that A ⊆ B2 ⊆ B1 and B2 is elementary equivalent to B1, and therefore to B0. Being elementarily equivalent to B0, B2 is purely infinite and simple, B2 might not be O2-stable, essentially by [22], however B = B2 ⊗ O2 satisfies all requirements. (cid:3) The following proof relies on a result proved in §5. The second proof of Lemma 2.1. The Cuntz-Pimsner algebra OE (see Lemma 5.2), constructed in [5, §4], is simple, and has the same density character as A. Moreover A embeds unitally into it. OE ⊗ O2 then provides the necessary object. (cid:3) We shall need the semigroups Extw(A) and Extw cpc(A) associated to a separable and unital C∗-algebra A. An injective unital ∗-homomorphism π : A → Q(H) is the Busby invariant of an extension of A by K (H). By a slight abuse of terminology, we say that such a ∗-homomorphism is an extension (see [24, Proposition 2.6.3]). Two extensions θj : A → Q(H), for j = 1, 2 are weakly equivalent if there is a unitary u ∈ Q(H) such that θ1 = Ad u ◦ θ2. 4 ILIJAS FARAH, ILAN HIRSHBERG, AND ALESSANDRO VIGNATI Since M2(Q(H)) ∼= Q(H), the set of extensions of A is equipped with the direct sum operation. The set of weak equivalence classes of extensions of A forms a semigroup, denoted Extw(A). An extension θ : A → Q(H) is semisplit if there exists a completely positive contraction (c.p.c.) ϕ : A → B(H) such that (denoting the quotient map from B(H) onto Q(H) by π) π ◦ ϕ = θ. If ϕ is a unital ∗-homomorphism then we say that θ is split. A split extension exists when A is separable, and Voiculescu's theorem ([24, Theorem 3.4.7]) implies that it acts as the unit in Extw(A). Let Extw cpc(A) := {θ ∈ Extw(A) : θ is semisplit}. Stinespring's theorem ([3, Theorem II.6.9.7]) easily implies that Extw Extw(A)−1, the group of all invertible elements of Extw(A). cpc(A) = The following is a standard application of quasicentral approximate units. Lemma 2.2. Suppose that a separable C∗-algebra A is an inductive limit of subalgebras An, for n ∈ N. If θ : A → Q(H) is an extension such that its restriction to An is semisplit for every n, then θ is semisplit. Proof. Let δn > 0 be small enough so that for all operators e and a satisfying 0 ≤ e ≤ 1, kak ≤ 1, and k[e, a]k < δn we have k[e1/2, a]k < 2−n. Let lift for θ ↾ An. By the Arveson Extension Theorem ([3, ψn be a c.p.c. Theorem II.6.9.12]) we can extend ψn to a c.p.c. map ψn : A → B(H). Let E = π−1(θ(A)) and let an, for n ∈ N, be an enumeration of a dense subset of the unit ball of A whose intersection with the unit ball of An is dense for all n. By [24, Proposition 3.2.8] we can find a sequence fn, for n ∈ N, which is an approximate identity for K (H) that is quasicentral for E. By refining this sequence, we may assume that the following conditions hold for all i, j, k, and n with i, j, k ≤ n. 1. k[fn, ψi(aj)]k < δn. 2. k(1 − fn)(ψi(aj) − ψk(aj))k < 2−n, if aj ∈ Ai ∩ Ak. The second condition can be assured because the assumptions imply ψi(aj)− ψk(aj) is compact, and the first condition can be assured by the quasicen- trality of the sequence. Given these conditions we have k[(fn+1 − fn)1/2, ψn(aj)]k < 2−n for all j ≤ n. Therefore ψ(a) = X (fn+1 − fn)1/2 ψn(a)(fn+1 − fn)1/2 n is well-defined since the finite partial sums converge in the strong operator topology. Since every partial sum is c.p.c., so is ψ. For all aj ∈ Ai and all n we also have that ψ(aj) − ψn(aj) is compact and therefore ψ is a c.p.c. lift of θ as required. (cid:3) Proposition 2.3. Suppose A ∈ O is separable. Then Extw Proof. Recall that an endomorphism ϕ of a C∗-algebra A is asymptotically inner if there exists a continuous path of unitaries ut, for 0 ≤ t < ∞, such cpc(A) = 0 THE CALKIN ALGEBRA IS ℵ1-UNIVERSAL 5 Since Extw that u0 = 1 and ϕ(a) = limt→∞(Ad ut)a for all a ∈ A. It follows from [35, Lemma 2.2.1] that any unital endomorphism of O2 is asymptotically inner. Suppose A ∼= A ⊗ O2 and let s and t be two standard generators of O2. It follows that the endomorphism ζ(a) = (1 ⊗ s)a(1 ⊗ s∗) + (1 ⊗ t)a(1 ⊗ t∗) of A is asymptotically inner. cpc(A) is a group, it suffices to prove that each of its elements cpc(A) if ζ : A → A and θ : A → Q(A) is an extension, then cpc(A), we have cpc(A) is homotopy invariant.3 (cid:3) is idempotent. The semigroup of endomorphisms of A acts on Extw by composition: ζ.θ = θ◦ ζ is an extension of A. However, as elements of Extw [θ◦ ζ] = [θ] + [θ]. By [2, Corollary 18.5.4], Extw Thus Extw cpc(A) is trivial, as required. We are now ready to prove Theorem A. By Lemma 2.1 it suffices to prove that every A ∈ O of density character ℵ1 embeds into Q(H). By the down- ward Lowenheim -- Skolem theorem there exists an increasing chain Aα, for α < ℵ1, of separable elementary submodels satisfying A = Sα<ℵ1 Aα and for every limit ordinal δ < ℵ1 we have Aδ = Sα<δ Aα. Each Aα is unital, purely infinite and simple, and it absorbs O2, since these properties are elementary for separable C∗-algebras ([15, Theorem 2.5.1 and Theorem 2.5.2]). cpc(Aα) such for all α < β < ℵ1 we We want to find extensions ϕα ∈ Extw have ϕα ∈ Extw cpc(Aα) and ϕβ↾Aα = ϕα. cpc(A0). Suppose ϕα has been defined for all α < β. Choose ϕ0 ∈ Extw If β is a successor ordinal, let α be such that α + 1 = β. Fix ψ ∈ Extw cpc(Aβ). Then Proposition 2.3 implies that both ψ′ := ψ↾Aα and ϕα are split. By Voiculescu's theorem ([24, Theorem 3.4.7]) there exists a unitary u ∈ Q(H) such that ϕα = Ad u ◦ ψ′ and ϕβ = Ad u ◦ ψ is as required. Now suppose β is a limit ordinal. Then ϕβ is already defined on a dense subalgebra Sα<β Aα of Aβ, and Lemma 2.2 implies that its continuous ex- tension to A is semisplit. Aα, for all a ∈ A the ordinal α(a) = min{α : a ∈ Aα} is well-defined. Then Φ(a) = ϕα(a)(a) extends each ϕα, provides the desired embedding of A into Q(H), and completes the proof of Theorem A. This describes the recursive construction. Since A = Sα<ℵ1 3. Corollaries of Theorem A and related results By GCH, we mean the Generalized Continuum Hypothesis, which is the statement 2λ = λ+ for every cardinal λ. Definition 3.1. Following [28] we say that a cardinal κ is far from the GCH there exists a cardinal λ such that the following holds (λ+ denotes the least 3To match the notation in Blackadar's book, what we denote by Extw(A) is denoted w(A, C). By [2, Proposition 15.14.2], for any unital separable C∗-algebra A w(A) ∼= Ext(A, C). Thus Extw cpc(A) ∼= Ext(A, C)−1. there by Extu we have Extu 6 ILIJAS FARAH, ILAN HIRSHBERG, AND ALESSANDRO VIGNATI cardinal greater than λ): (⋆) λ+ < κ < 2λ. Corollary 3.2. Each of the following assertions is relatively consistent with ZFC : 1. Q(H) is a 2ℵ0-universal C∗-algebra. 2. There exists a 2ℵ0 -universal C∗-algebra, but Q(H) is not 2ℵ0 -universal. 3. A 2ℵ0-universal C∗-algebra does not exist. Proof. (1) Assume the Continuum Hypothesis. Theorem A implies that Q(H) is 2ℵ0 -universal. (2) We shall prove that the Proper Forcing Axiom, PFA, implies the conclusion.4 If 2κ = 2ℵ0 for all κ < 2ℵ0, then [1, Proposition 7.10] implies that the theory of Q(H) has a saturated model B of density character 2ℵ0. By Lemma 2.1 (and its proof), such B is a 2ℵ0 -universal C∗-algebra. It is well-known that PFA implies 2ℵ1 = 2ℵ0 = ℵ2 (e.g. see [30]). By [43, Corollary 5.3.14 and Theorem 5.3.15] (also [33]) PFA implies that there exist a closed subset X of βN\ N such that C(X) does not embed into the Calkin algebra. Such an X can be chosen as follows. Let Z0 = {S ⊆ N : limn→∞ S∩n/n = 0}, the ideal of asymptotic density zero sets. (Any other dense analytic P-ideal would do in place of Z0; see [43].) Identifying βN with the set of all ultrafilters on N, we may let X = {U ∈ βN : U ∩ Z0 = {∅}}. (3) By standard forcing techniques ([30]) the assertion ℵ2 < 2ℵ0 < 2ℵ1 is relatively consistent with ZFC . (For example, start from a model of GCH, add ℵ4 Cohen subsets of ℵ1, then add ℵ3 Cohen reals.) Therefore κ = 2ℵ0 and λ = ℵ1 satisfy the inequality (⋆) and we conclude that 2ℵ0 is far from the GCH in this model. We shall prove that (⋆) and κℵ0 = κ together imply that there is no κ- universal C∗-algebra.5 By [9, §3.3] every Banach space embeds isometrically into an operator space (and therefore into a C∗-algebra) of the same den- sity character. Hence a κ-universal C∗-algebra would also be a κ-universal Banach space. But, by [42, Corollary 2.4], (⋆) implies that there is no (iso- metrically) κ-universal Banach space, and this concludes the proof. (cid:3) A sketch of an alternative, self-contained (and, we believe, more informa- tive) proof that the condition (⋆) in Corollary 3.2 (3) together with κℵ0 = κ implies there is no κ-universal C∗-algebra is in order. It uses the following definition adapted to the continuous context from [28, Definition 5.1] (see [42, Definition 1.1]). Definition 3.3. A theory T has the Strict Order Property (SOP) if there exists a formula ψ(¯x, ¯y) of the language of T in 2n variables for some n ≥ 1 4Readers concerned with the consistency strength issues may rest assured that only a small fragment of PFA with zero large cardinal strength is required in [43]. 5See Remark 3.4 for a sketch of a self-contained proof THE CALKIN ALGEBRA IS ℵ1-UNIVERSAL 7 with the following two properties. First, in every model of T the relation a ≺φ b given by a ≺φ b ⇐⇒ φ(a, b) = 0 and φ(b, a) = 1 defines a partial ordering. Second, in every model of T there are arbitrarily long finite ≺φ-chains. If the theory T is complete, then by a compactness argument (i.e., taking an ultraproduct) the second requirement can be replaced by the requirement that there exists an infinite ≺φ chain in some model of T . Since the theory of C∗-algebras is not complete, we opt for the current formulation. Remark 3.4. Here is the promised alternative proof of Corollary 3.2 (3). Instead of using [42, Corollary 2.4], we follow the lines of its proof. By [28, Theorem 3.10], (⋆) implies that there is no κ-universal linear order, and moreover that any theory T with SOP does not have a 2ℵ0-universal model. As in [16, Lemma 5.3], consider the following condition in the language of C∗-algebras ϕ(x, y) = max(1 − kx∗xk,1 − ky∗yk,kx∗xy∗y − y∗yk). Two elements x and y of a C∗-algebra A satisfy ϕ(x, y) = 0 if and only if kxk = kyk = 1 and in the second dual A∗∗ of A the support projection of y∗y is below the spectral projection of x∗x corresponding to 1. Therefore ϕ(x, y) defines a partial order, (cid:22)ϕ, on A. Every infinite-dimensional C∗-algebra contains an infinite (cid:22)ϕ chain (consider any, necessarily infinite-dimensional, masa or see [16, Lemma 5.3]). Thus every infinite-dimensional C∗-algebra has the Strict Order Property. Since ϕ is quantifier-free, an embedding of A into B is an embedding of the poset (A,(cid:22)ϕ) into (B,(cid:22)ϕ). Hence if C is a κ-universal C∗-algebra for some cardinal κ, then every linear ordering of cardinality κ embeds into the linearization of (C,(cid:22)ϕ), which has size κℵ0 = κ. The latter is a κ-universal linear ordering, a contradiction. The following is a poor man's version of Theorem B. Corollary 3.5. If 2ℵ0 is far from the GCH then no C∗-algebra of density character 2ℵ0 is universal for all abelian C∗-algebras of density character 2ℵ0 . In particular no 2ℵ0 -universal exact C∗-algebra. Proof. By Remark 3.4 the theory of C([0, 1]) has the Strict Order Property witnessed by a quantifier-free formula and the first claim follows by the argument of the latter part of Remark 3.4. Since every abelian C∗-algebra is exact (and being abelian is axiomatizable, [15, Theorem 2.5.1]), the second claim follows. (cid:3) The following lemma is a special case of [38, Proposition 2.53]. Lemma 3.6. Suppose A is a C∗-algebra and B is a C∗-subalgebra of A that contains an approximate unit for A. Then the inclusion from B into A extends to an injection from M (B) into M (A), and M (B)/B is isomorphic to a subalgebra of M (A)/A. 8 ILIJAS FARAH, ILAN HIRSHBERG, AND ALESSANDRO VIGNATI From Theorem A and Lemma 3.6 we immediately have the following. Corollary 3.7. Let A be a unital separable C∗-algebra. If Q(H) is 2ℵ0- universal, then so is the corona of A ⊗ K (H). (cid:3) We record an easy consequence of a trick first used in [35, Theorem 4.3.11]. Proposition 3.8. If κ < 2ℵ0, there is no κ-universal C∗-algebra. Proof. This follows from [25, Theorem 2.3 and Remark 2.10]. The space OS3 of three-dimensional operator spaces can be equipped by a metric δcb such that the space of all operator spaces that embed into a C∗-algebra A has density at most equal to the density character of A ([25, Proposition 2.6(a)]), and the space (OS3, δcb) has density character 2ℵ0.6 (cid:3) 4. The proof of Theorem B A C∗-algebra is approximately matricial, or AM, if it is a unital inductive limit of full matrix algebras Mn(C) for n ∈ N. All AM algebras are nuclear, simple, and have a unique trace ([20]) and every separable AM algebra is UHF. However, nonseparable AM algebras are not necessarily UHF, and can be quite pathological ([18]). Our proof of Theorem B will use a class of AM algebras associated with graphs introduced in [12]. Let P denote the poset for adding ℵ2 Cohen reals (denoted Fn(ℵ2, 2,ℵ0) in [30]). If G ⊆ P is generic over a model M of a sufficiently large fragment of ZFC then every C∗-algebra A in M is identified with its completion in M [G]. Lemma 4.1. Suppose that M is a transitive model of a sufficiently large fragment of ZFC and that G is generic over M for P. Then there exists a C∗-algebra D(G) in M [G] with the following properties. 1. The algebra D(G) is unital, nuclear, and simple. 2. Its density character is ℵ1. 3. It is not isomorphic to a subalgebra of any C∗-algebra in M . In addition, we can choose D(G) to be stably finite with a unique trace and faithfully representable on a separable Hilbert space. The reader will notice that Theorem A implies that the C∗-algebra D(G) provided by Lemma 4.1 embeds into the Calkin algebra, which apparently leads to a contradiction. A reassurance that we do not have a proof that ZFC is inconsistent may therefore be appreciated. The Calkin algebra as computed in M [G] is much larger than the completion of the Calkin algebra as computed in M , and while D(G) embeds into the former by Theorem A it does not embed into the latter by the following proof. Proof of Lemma 4.1. This proof is based on [28, Fact on p. 889] and a construction from [12]. We define a simple bipartite graph Γ(G) (isomorphic to the graph denoted by G and/or G∗ in [28]) as follows. The vertex set of 6This is analogous to the fact that the space D(T ) of quantifier-free types in models of theory T has density 2ℵ0 whenever it is nonseparable; see [28]. THE CALKIN ALGEBRA IS ℵ1-UNIVERSAL 9 Γ(G) is N⊔ℵ1. Let cξ, for ξ < ℵ1, enumerate the Cohen generic reals coded by G. They are functions from N to {0, 1}, and a pair {m, ξ} forms an edge if and only if cξ(m) = 1. Let D(G) be the graph CCR C∗-algebra defined in [12, §1]. It is the universal C∗-algebra given by the generators um, for m ∈ N, and vξ, for ξ < ℵ1 and the following relations for all m and n in N and all ξ and η in ℵ1. m = 1, m = u∗ ξ vξ = vξv∗ m, u2 ξ , v2 mum = umu∗ ξ = 1, 1. um = u∗ ξ = v∗ 2. vξ = v∗ 3. umun = unum, 4. vξvη = vηvξ, 5. umvξ = vξum, if m is not adjacent to ξ, and 6. umvξ = −vξum, if m is adjacent to ξ. The set P(N) is considered with the Cantor set topology and the compatible metric d(a, b) = (min(a∆b) + 1)−1, Claim 4.2. For ξ < ℵ1 let aξ = {m ∈ N : cξ(m) = 1}. The family A = {aξ : ξ < ℵ1} is dense in P(N). Also, if K and L are disjoint finite subsets of ℵ1 and K is nonempty, the set aξ \ [ \ aη ξ∈K η∈L is infinite. (A family with this property is called independent.) Proof. Both properties are easy consequences of the fact that hcξ+m : m < ωi does not belong to any closed nowhere dense subset of P(N)N coded in M [hcη : η < ξi], for all ξ < ℵ1. (cid:3) By the claim, the family A is dense and independent. By [12, Lemma 1.4 and the proof of Lemma 1.6], the C∗-algebra D(G) is AM and it has a faithful representation on a separable Hilbert space. It remains to prove that D(G) is not isomorphic to a subalgebra of any C∗-algebra A in M . Suppose otherwise, and let Φ : D(G) → A be a unital (and therefore necessarily injective) -homomorphism. For ξ < ℵ1 let Gξ denote the intersection of G with the poset for adding the first ξ Cohen reals. Therefore M [Gξ] = M [hcη : η < ξi]. We denote this model by Mξ. By the ccc-ness of P, the values of Φ(um) for m ∈ N are decided by countable maximal antichains and therefore there exists ζ < ℵ1 such that in Mζ for every m ∈ N there is wm ∈ A7 such that Φ(um) = wm. The model M [G] is a forcing extension of Mζ by the quotient of P with its regular subordering that added hcξ : ξ < ζi. This quotient ordering Pζ is isomorphic to the poset for adding ℵ1 Cohen reals, and every cη, for η ≥ ζ, is Cohen over Mζ ([30]). Fix η ≥ ζ and find a condition p in G8 and w ∈ A 7Note that A stands for the metric completion of A as computed in Mζ. 8Formally, p is in the quotient G/Gζ but since in the case of Cohen reals the iteration and the product coincide we can think of p being in G. 10 ILIJAS FARAH, ILAN HIRSHBERG, AND ALESSANDRO VIGNATI such that p (cid:13) kw − Φ(vη)k < 1/4. Fix m ∈ N. Then in M [G] we have (writing [x, y] = xy − yx) k[wm, w] − [wm, Φ(vη)]k ≤ 2kwmkkw − Φ(vη)k < 1/2. Since wm = Φ(um) and Φ is an injective -homomorphism, we have [wm, Φ(vη)] = 0 if cη(m) = 0 and [wm, Φ(vη)] = 2wmΦ(vη) if cη(m) = 1. Therefore cη is decided by p and it belongs to Mζ; contradiction. (cid:3) Proof of Theorem B. To prove that it is relatively consistent with ZFC that there is no ℵ1-universal nuclear C∗-algebra, and that there is no ℵ1-universal nuclear, simple, C∗-algebra, we start from a model of the Continuum Hy- pothesis and add ℵ2 Cohen reals. Suppose that A is a C∗-algebra of den- sity ℵ1 universal for nuclear, simple, C∗-algebras of density character ℵ1. By the ccc-ness of this poset, every C∗-algebra of density character ℵ1 is added by a poset for adding ℵ1 Cohen reals, and by Lemma 4.1 any further batch of ℵ1 Cohen reals adds a unital, nuclear, and simple C∗-algebra of density character ℵ1 that is not isomorphic to a subalgebra of A; contradic- tion. (cid:3) We can do better; here is a sample (note that the C∗-algebra A is not even assumed to be nuclear, and see also Remark 4.4 below). Theorem 4.3. It is relatively consistent with ZFC that no C∗-algebra A of density character ℵ2017 is universal for all unital, simple, nuclear, and stably finite C∗-algebras that have density character ℵ1 and a faithful representation on a separable Hilbert space. Proof. Start from a model of the Continuum Hypothesis and add ℵ2018 Co- hen reals. By the ccc-ness of the forcing, if A is any C∗-algebra of density character < ℵ2018, then it belongs to an intermediate model obtained by adding ℵ2017 Cohen reals. By Lemma 4.1 any further batch of ℵ1 Cohen reals adds an AM C∗-algebra of density character ℵ1 that has a faithful representation on a separable Hilbert space but is not isomorphic to a sub- algebra of A. Since AM algebras are nuclear, unital, simple, and stably finite, this concludes the proof. (cid:3) Remark 4.4. The proof of Lemma 4.1 shows that analogous statements can be proved for models of a first-order theory (possibly in the logic of metric structures) with the Independence Property. This was known to the au- thors of [28]. The novelty is that the nuclear C∗-algebras do not form an axiomatizable class (see [15]). An observant reader familiar with model the- ory will have noticed that the proof of Lemma 4.1 relies on the fact that the category of nuclear, simple, and unital C∗-algebras includes a class of EM-models generated by the indiscernibles. This fact was first used in [21] where it was proved that there are 2ℵ1 nonisomorphic AM algebras of den- sity character ℵ1. Note however that the methods of [21] do not seem to imply Theorem B. THE CALKIN ALGEBRA IS ℵ1-UNIVERSAL 11 5. More on the absence of universal nuclear C∗-algebras We want to use the Strict Order Property (Definition 3.3) to prove the following: Theorem 5.1. Suppose κ is a cardinal far from the GCH. Then the follow- ing hold. 1. There exists no universal abelian C∗-algebra of density character κ. 2. There exists no universal nuclear C∗-algebra of density character κ. 3. There exists no universal nuclear, simple, C∗-algebra of density char- acter κ. In case when κ = κℵ0 we can follow the same strategy as in Remark 3.4. In case this equality fails, the proof of the theorem uses [41, Theorem 2.12] in a manner similar to, but easier than, that in the proof of [42, Corol- lary 2.6]. While the theory of Banach spaces only has a technical weakening of SOP known as SOP4, the theory of any infinite-dimensional C∗-algebra has the full SOP (the relation used in the proof of [16, Lemma 5.3] is clearly transitive). Lemma 5.2 below is a consequence of [29, Theorem 3.1] and standard results (Remark 5.3), but we provide a self-contained proof. It uses the analysis of Cuntz -- Pimsner algebras associated to A− B C∗-correspondences and we recall the definition (see [5, §4.6] for more details). If A and B are C∗-algebras, an A− B C∗-correspondence is a B-Hilbert module E together with a faithful -representation of A in the algebra of adjointable linear operators on E. Additional properties of C∗-correspondences used in the proof will be introduced as needed. Lemma 5.2. Every nuclear C∗-algebra is isomorphic to a subalgebra of a simple, nuclear C∗-algebra of the same density character. Proof. Let A be a C∗-algebra. Let π : A → B(H) be a faithful representation of A on a Hilbert space H such that π(A) ∩ K (H) = {0}. Such an H can always be chosen to have the same density character as A. With π defining a left action, we can view H as an A− C C∗-correspondence. Consider A as a Hilbert module over itself, viewed as a C − A C∗-correspondence (with C acting on the left as scalars). Let E = H ⊗C A. Let OE be the Cuntz-Pimsner algebra associated to E. Then A is embed- ded in OE, by construction OE has the same density character, and by [5, Theorem 4.6.25] and its corollary, if A is nuclear then so is OE. (Likewise, if A is exact then so is OE.). It remains to show that OE is simple, which we do by verifying the conditions of [40, Theorem 3.9]. For that, we need to show that the C∗- correspondence E is full, nonperiodic and minimal. To say that E is full means that hE, Ei is dense in A. let a ∈ A be any positive element, and choose a unit vector ξ ∈ E, then hξ ⊗ √a, ξ ⊗ √ai = a. As all positive elements can be obtained, we have hE, Ei = A. Indeed, 12 ILIJAS FARAH, ILAN HIRSHBERG, AND ALESSANDRO VIGNATI To say that E is nonperiodic means that no tensor power (over A) of E, E⊗n, is unitarily equivalent to the trivial C∗-correspondence A. By our construction, the left action of A on E⊗n has trivial intersection with the compacts for any n > 0, whereas the left action on the trivial C∗- correspondence is via compact operators, and thus they are not unitarily equivalent. (cid:3) To say that E is minimal means that there is no nontrivial ideal J in A such that hE, JEi ⊆ J. Indeed, suppose J is a nontrivial ideal. Let a ∈ J be a non-zero element. As π is faithful, we can choose vectors ξ, η ∈ H such that hξ, π(a)ηi = 1. If b ∈ A r J is positive then b = Dξ ⊗ √b, a · η ⊗ √bE belongs to hE, JEi. Thus E is minimal, as required. Remark 5.3. Lemma 5.2 is also a consequence of the results of [29, The- In it Kumjian proved that every separable nuclear unital C∗- orem 3.1]. algebra A is isomorphic to a unital subalgebra of a separable nuclear unital simple C∗-algebra OE (the UCT is not needed for this; see the second sen- tence of the proof). The separability assumption on A can be removed as follows. If A is nonseparable, then A can be written as an inductive limit of a σ-closed system of its separable elementary submodels Aλ, for λ ∈ Λ. The Cuntz -- Pimsner algebra OEλ associated to Aλ as in [29, §1] is nuclear, sim- ple, purely infinite, and unital and OE is the inductive limit of the system OEλ, for λ ∈ Λ. In addition, this system is σ-directed complete in the sense of [20]: if Aλ = Sn Aλn then OEλ = Sn OEλn . By a closing up argument, this implies that OE has a separable elementary submodel isomorphic to OEλ for some λ. Since being simple and purely infinite is axiomatizable ([15, Theorem 2.15]), the conclusion follows. The following lemma is proved by mimicking the proof of [41, Theo- rem 2.12]; the details are worked out in [39]. Lemma 5.4. Suppose that T is a theory in a continuous language with the SOP and that κ is an infinite cardinal far from the GCH. Then T has no universal model of density character κ. (cid:3) Remark 5.5. 1. We note that the results [40] depend on results from an unpublished manuscript (number 15 in the list of references of [40]). However, a more general result was shown in [31, Theorem 9.15], and thus there is no gap in the literature. 2. As the left action in our construction has trivial intersection with the compact adjointable operators, it follows that the Cuntz-Pimsner algebra OE coincides with the Toeplitz-Pimsner algebra TE. Thus, in the separable setting, it follows from the results in Section 4 of [37] that OE is KK-equivalent to A. In the non-separable setting one cannot talk about KK-equivalence, however it follows from the results in Section 8 of [26] that K∗(OE) ∼= K∗(A) (as unordered groups). THE CALKIN ALGEBRA IS ℵ1-UNIVERSAL 13 3. Lemma 5.2 implies that the existence of a κ-universal nuclear C∗- algebra is equivalent to the existence of a κ-universal simple nuclear C∗-algebra for every infinite cardinal κ. Proof of Theorem 5.1. (1) and (2) are immediate consequences of the fact that each infinite-dimensional C∗-algebra has SOP and Lemma 5.4. (3) follows from Lemma 5.2. (cid:3) 6. Concluding Remarks A positive answer to the following question would imply that the conclu- sion of Theorem B is independent from ZFC. Question 6.1. Is it relatively consistent with ZFC that there exists a uni- versal nuclear C∗-algebra of density character ℵ1? What about a universal exact C∗-algebra of density character ℵ1? While every separable exact C∗-algebra is nuclearly embeddable by [27], it is not known whether every exact C∗-algebra is nuclearly embeddable. It is therefore not impossible that two parts of Question 6.1 have different answers. Theorem A does not answer Question 6.1 because the Calkin algebra is neither nuclear nor exact. In the standard model-theoretic terminology, the following question asks whether the category of separable nuclear C∗-algebras has amalgamation. This is not to be confused with any of the standard amalgamation construc- tions used in operator algebras. Question 6.2. Suppose Φ : A → B, Ψ : A → C are injective ∗-homomorphisms between nuclear C∗-algebras. Is there a nuclear C∗-algebra D and injective ∗-homomorphisms Φ1 : B → D, Ψ1 : C → D such that the diagram com- mutes? We discuss briefly the connection with amalgamation in the C∗-algebraic context. The following idea comes from Jamie Gabe. Viewing A as included in B and in C as above, we can form the amalgamated free product B ∗A C. The amalgamated free product can in general fail to be exact, let alone nu- clear. If however we know that there exist conditional expectations from B onto Φ(A) and from C onto Ψ(A), then one can form the reduced amal- gamated free product (see [5, Section 4.7] for a definition and discussion). Assuming that those conditional expectations have faithful GNS represen- tations (which is automatic if, for example, all C∗-algebras in questions are simple), it follows from [8, Corollary 5.7] that the reduced amalgamated free product of B and C over A is exact. By Kirchberg's embedding theorem, the reduced amalgamated free product then embeds in O2. This gives a partial positive answer to Question 6.2. By using standard techniques, a positive answer to Question 6.2 would imply that under the Continuum Hypothesis there exists an ℵ1-universal nuclear C∗-algebra. The following lemma shows that it suffices to answer an 'easier' version of Question 6.2. 14 ILIJAS FARAH, ILAN HIRSHBERG, AND ALESSANDRO VIGNATI Lemma 6.3. Every nuclear C∗-algebra A is isomorphic to a subalgebra of a C∗-algebra that is an inductive limit of a net of C∗-algebras each of which is isomorphic to O2. Proof. By Lemma 5.2, A is isomorphic to a subalgebra of a simple, nuclear C∗-algebra B of the same density character. We claim that C = B ⊗ O2 is as required. Since B is simple, it is equal to the inductive limit of simple and separable C∗-algebras. We can moreover assure that these algebras are nuclear, either by closing up or by taking an elementary submodel and using [15, Theorem 5.7.3]. But if B = limλ Bλ, then C = limλ Bλ ⊗ O2, and by Kirchberg's O2-absorption theorem Bλ ⊗ O2 ∼= O2 for all λ. (cid:3) In relation to the conclusion of Lemma 6.3 it should be noted that in- ductive limit of C∗-algebras isomorphic to O2 can be quite unruly. By the main result of [18], Jensen's ♦ ([30, III.7]) implies the existence of a C∗- algebra that is an inductive limit of C∗-algebras isomorphic to O2 but it is not isomorphic to its opposite algebra, all of its irreducible representations are unitarily equivalent, and all of its automorphisms are inner. Question 6.4. Suppose Φ : O2 → O2, Ψ : O2 → O2 are unital ∗-homomorphisms. Are there unital ∗-homomorphisms Φ1 : O2 → O2, Ψ1 : O2 → O2 such that Φ1 ◦ Φ = Ψ1 ◦ Ψ? ℵ1 As in the discussion following Question 6.2, if the images of Φ and Ψ admit conditional expectations onto them then the answer is positive. We do not know whether there are subalgebras of O2 isomorphic to O2 which do not admit conditional expectations onto them. By using Lemma 6.3, standard techniques show that a positive answer to Question 6.4 would imply a positive answer to Question 6.1. A stan- dard descriptive set-theoretic argument shows that a a positive answer to Question 6.4 is equivalent to a Σ1 2 statement and therefore absolute between transitive models of ZFC containing all countable ordinals. Because of this the answer to this question is unlikely to be independent from ZFC. For ex- ample, if this question can be resolved by using the Continuum Hypothesis (or ♦, or Martin's Axiom. . . ) then it can be resolved in ZFC alone. This however still leaves a possibility that Question 6.4 cannot be resolved in ZFC. See [13, §3 and §A.4] for a discussion of the absoluteness phenomenon. We can also search for potential target algebras to replace O2. A unital purely infinite and simple C∗-algebra A is in Cuntz standard form if [1A] = 0 in K0(A). It can be shown that A is in standard form if and only if O2 embeds unitally in A. Every Kirchberg algebra is stably isomorphic to one in standard form, denoted Ast. This association is unique up to isomorphism. Given this, as O2 embeds unitally into Ost ∞ serves as a universal algebra which admits unital embeddings of any separable nuclear C∗-algebra. Denote by Oℵ1 the analogue of O∞ corresponding to ℵ1-many isometries with pairwise orthogonal ranges. Question 6.5. Is Ost a universal nuclear C∗-algebra of density character ℵ1? ∞, Ost THE CALKIN ALGEBRA IS ℵ1-UNIVERSAL 15 6.1. Remarks on universality in related categories. Isomorphic embeddings of Banach spaces. Theorem A was inspired by [4, Theorem 1.4], where the analogous statement for 2ℵ0-univeral Banach spaces was proved. Brech and Koszmider constructed a forcing extension in which an isometrically 2ℵ0-Banach space exists, but ℓ∞/c0 is not isometri- cally, or even isomorphically, 2ℵ0-universal Banach space. The result of [42, Corollary 2.4] used in the proof of Corollary 3.2 was improved in [4, The- orem 1.3], where it was proved that consistently there is no isomorphically 2ℵ0-universal Banach space. Linear orders. The existence of universal linear orders is a well-studied subject ([28]). Much attention has been devoted to the question of 2ℵ0- universality of P(N)/ Fin. Since the Calkin algebra is its noncommutative analogue (see e.g. [44]), we shall concentrate on the role of P(N)/ Fin. While it is consistent that the Continuum Hypothesis fails and P(N)/ Fin is 2ℵ0- universal ([32]), it is not clear whether the assertion 'Q(H) is a 2ℵ0 -universal C∗-algebra' is relatively consistent with the failure of the Continuum Hy- pothesis. The question on whether for a given C∗-algebra A there exists a ccc forcing notion that forces an embedding of A into Q(H) was given a positive answer in [19]. Notably, the structure of the small category of linear orders that embed into P(N)/ Fin is remarkably malleable in ZFC (see [10, §1]) and very rigid if a fragment of PFA holds ([11]). Surjective universality for compact Hausdorff spaces. A compact Hausdorff space X is said to be κ-universal if it is surjectively universal among compact Hausdorff spaces of weight κ. By Gelfand -- Naimark duality, this is equivalent to C(X) being an injectively universal unital abelian C∗- algebra. The Continuum Hypothesis implies that βN \ N is an ℵ1-universal compact Hausdorff space (Parovicenko's theorem) and that βR+ \ R+ is an ℵ1-universal connected compact Hausdorff space ([7]). As in Corollary 3.2, PFA implies that βN \ N is not 2ℵ0-universal because it does not map onto the Stone space of the Lebesgue measure algebra ([6]). II1-factors. In [34] it was proved that there is no κ-universal II1-factor for any κ < 2ℵ0. As in Corollary 3.2, κ<κ = κ implies there is a κ-universal II1- factor. The theory of II1-factors has the Order Property ([16, Lemma 3.2]) but it is not known whether it has the Strict Order Property. If it does, the argument from Remark 3.4 would imply that the existence of a cardinal λ such that λ+ < 2ℵ0 < 2λ implies there is no 2ℵ0 -universal II1-factor. As a curiosity, we note that Connes' Embedding Problem has the positive so- lution if and only if the Continuum Hypothesis implies that an ultrapower of the hyperfinite II1-factor is a 2ℵ0-universal II1-factor. Similarly, Kirch- berg's Embedding Problem ([23]) has the positive solution if and only if the Continuum Hypothesis implies that an ultrapower of O2 is a 2ℵ0 -universal C∗-algebra. 16 ILIJAS FARAH, ILAN HIRSHBERG, AND ALESSANDRO VIGNATI References [1] I. Ben Yaacov, A. Berenstein, C.W. Henson, and A. Usvyatsov. Model theory for metric structures. In Z. Chatzidakis et al., editors, Model Theory with Applications to Algebra and Analysis, Vol. II, number 350 in London Math. Soc. Lecture Notes Series, pages 315 -- 427. 2008. [2] B. Blackadar. K-theory for operator algebras, volume 5 of Mathematical Sciences Re- search Institute Publications. Cambridge University Press, Cambridge, second edition, 1998. [3] B. Blackadar. Operator algebras, volume 122 of Encyclopaedia of Mathematical Sci- ences. Springer-Verlag, Berlin, 2006. Theory of C ∗-algebras and von Neumann alge- bras, Operator Algebras and Non-commutative Geometry, III. [4] C. Brech and P. Koszmider. On universal Banach spaces of density continuum. Israel J. Math., 190:93 -- 110, 2012. [5] N. P. Brown and N. Ozawa. C ∗-algebras and finite-dimensional approximations, vol- ume 88 of Graduate Studies in Mathematics. American Mathematical Society, Prov- idence, RI, 2008. [6] A. Dow and K.P. Hart. The measure algebra does not always embed. Fundamenta Mathematicae, 163:163 -- 176, 2000. [7] A. Dow and K.P. Hart. A universal continuum of weight ℵ. Trans. Amer. Math. Soc., 353(5):1819 -- 1838, 2001. [8] K. J. Dykema and D. Shlyakhtenko. Exactness of Cuntz-Pimsner C ∗-algebras. Proc. Edinb. Math. Soc. (2), 44(2):425 -- 444, 2001. [9] E. G. Effros and Z-J. Ruan. Operator spaces, volume 23 of London Math. Soc. Mono- graphs. New Series. The Clarendon Press, Oxford University Press, New York, 2000. [10] I. Farah. Embedding partially ordered sets into ωω. Fundamenta Mathematicae, 151:53 -- 95, 1996. [11] I. Farah. Analytic quotients: theory of liftings for quotients over analytic ideals on the integers. Mem. Amer. Math. Soc., 148(702):xvi+177, 2000. [12] I. Farah. Graphs and CCR algebras. Indiana Univ. Math. Journal, 59:1041 -- 1056, 2010. [13] I. Farah. Absoluteness, truth, and quotients. In C.T. Chong et al., editors, Infinity and Truth, volume 25 of Lecture Notes Series, Institute for Mathematical Sciences, National University of Singapore, pages 1 -- 24. World Scientific, 2013. [14] I. Farah and B. Hart. Countable saturation of corona algebras. C.R. Math. Rep. Acad. Sci. Canada, 35:35 -- 56, 2013. [15] I. Farah, B. Hart, M. Lupini, L. Robert, A. Tikuisis, A. Vignati, and W. Winter. Model theory of C∗-algebras. to appear in Mem. Amer. Math. Soc. arXiv:1602.08072. [16] I. Farah, B. Hart, and D. Sherman. Model theory of operator algebras I: Stability. Bull. London Math. Soc., 45:825 -- 838, 2013. [17] I. Farah and I. Hirshberg. The Calkin algebra is not countably homogeneous. Proc. Amer. Math. Soc., 144(12):5351 -- 5357, 2016. [18] I. Farah and I. Hirshberg. Simple nuclear C∗-algebras not isomorphic to their oppo- sites. Proc. Natl. Acad. Sci. USA, 114(24):6244 -- 6249, 2017. [19] I. Farah, G. Katsimpas, and A. Vaccaro. Embedding C∗-algebras into the Calkin algebra. arXiv:1810.00255, 2018. [20] I. Farah and T. Katsura. Nonseparable UHF algebras I: Dixmier's problem. Adv. Math., 225(3):1399 -- 1430, 2010. [21] I. Farah and T. Katsura. Nonseparable UHF algebras II: Classification. Math. Scand., 117(1):105 -- 125, 2015. [22] S. Ghasemi. SAW algebras are essentially non-factorizable. Glasg. Math. J., 57(1):1 -- 5, 2015. THE CALKIN ALGEBRA IS ℵ1-UNIVERSAL 17 [23] I. Goldbring and T. Sinclair. On Kirchberg's embedding problem. J. Funct. Anal., 269:155 -- 198, 2015. [24] N. Higson and J. Roe. Analytic K-homology. Oxford Mathematical Monographs. Ox- ford University Press, Oxford, 2000. [25] M. Junge and G. Pisier. Bilinear forms on exact operator spaces and B(H) ⊗ B(H). Geom. Funct. Anal., 5(2):329 -- 363, 1995. [26] T. Katsura. On C ∗-algebras associated with C ∗-correspondences. J. Funct. Anal., 217(2):366 -- 401, 2004. [27] E. Kirchberg and N. C. Phillips. Embedding of exact C ∗-algebras in the Cuntz algebra O2. J. Reine Angew. Math., 525:17 -- 53, 2000. [28] M. Kojman and S. Shelah. Nonexistence of universal orders in many cardinals. The Journal of Symbolic Logic, 57(3):875 -- 891, 1992. [29] A. Kumjian. On certain Cuntz -- Pimsner algebras. Pacific J. Math., 217(2):275 -- 289, 2004. [30] K. Kunen. Set theory, volume 34 of Studies in Logic (London). College Publications, London, 2011. [31] B.K. Kwasniewski and R. Meyer. Aperiodicity, topological freeness and more: from group actions to Fell bundles. to appear in Studia Math. arXiv:1611.06954, 2016. [32] R. Laver. Linear orders in ωω under eventual dominance. Studies in Logic and the Foundations of Mathematics, 97:299 -- 302, 1979. [33] P. McKenney and A. Vignati. Forcing axioms and coronas of nuclear C∗-algebras. arXiv:1806.09676, 2018. [34] N. Ozawa. There is no separable universal II1-factor. Proc. Amer. Math. Soc., 132(2):487 -- 490, 2004. [35] N. C. Phillips. A classification theorem for nuclear purely infinite simple C∗-algebras. Doc. Math., 5:46 -- 114, 2000. [36] N.C. Phillips and N. Weaver. The Calkin algebra has outer automorphisms. Duke Math. Journal, 139:185 -- 202, 2007. [37] M. V. Pimsner. A class of C ∗-algebras generalizing both Cuntz-Krieger algebras and crossed products by Z. In Free probability theory (Waterloo, ON, 1995), volume 12 of Fields Inst. Commun., pages 189 -- 212. Amer. Math. Soc., Providence, RI, 1997. [38] I. Raeburn and D.P. Williams. Morita equivalence and continuous-trace C∗-algebras. Number 60 in Math. Surveys and Monographs. Amer. Math. Soc., 1998. [39] S. Reiss and A. Usvyatsov. Sufficient conditions for non-existence of universal metric structures. Unpublished. [40] J. Schweizer. Dilations of C∗-correspondences and the simplicity of Cuntz -- Pimsner algebras. Journal of Functional Analysis, 180(2):404 -- 425, 2001. [41] S. Shelah. Toward classifying unstable theories. Annals of Pure and Applied Logic, 80(3):229 -- 255, 1996. [42] S. Shelah and A. Usvyatsov. Banach spaces and groups: order properties and universal models. Israel Journal of Mathematics, 152(1):245 -- 270, 2006. [43] A. Vignati. Logic and C∗-algebras: Set Theoretical dichotomies in the theory of con- tinuous quotients. PhD thesis, York University, Toronto, 2017. [44] N. Weaver. Set theory and C ∗-algebras. Bull. Symb. Logic, 13:1 -- 20, 2007. 18 ILIJAS FARAH, ILAN HIRSHBERG, AND ALESSANDRO VIGNATI (I. Farah) Department of Mathematics and Statistics, York University, 4700 Keele Street, North York, Ontario, Canada, M3J 1P3 E-mail address: [email protected] URL: http://www.math.yorku.ca/∼ifarah (Ilan Hirshberg) Department of Mathematics, Ben Gurion University of the Negev, P.O.B. 653, Be'er, Sheva 84105, Israel E-mail address: [email protected] URL: http://www.math.bgu.ac.il/~ilan/ (A. Vignati) Institut de Math´ematiques de Jussieu - Paris Rive Gauche (IMJ- PRG), UP7D - Campus des Grands Moulins, Batiment Sophie Germain, 8 Place Aur´elie Nemours, Paris, 75013, France. Currently at: Department of Math- ematics, KU Leuven, Celestijnenlaan 200B, B-3001 Leuven, Belgium E-mail address: [email protected] URL: http://www.automorph.net/avignati
1803.04828	2	1803	2018-03-19T00:09:31	Amenable actions of discrete quantum groups on von Neumann algebras	[ "math.OA" ]	We introduce the notion of Zimmer amenability for actions of discrete quantum groups on von Neumann algebras. We prove generalizations of several fundamental results of the theory in the noncommutative case. In particular, we give a characterization of Zimmer amenability of an action $\alpha:\Bbb G\curvearrowright N$ in terms of $\hat{\Bbb{G}}$-injectivity of the von Neumann algebra crossed product $N\ltimes_\alpha\Bbb G$. As an application we show that the actions of any discrete quantum group on its Poisson boundaries are always amenable.	math.OA	math	AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS ON VON NEUMANN ALGEBRAS MOHAMMAD S. M. MOAKHAR Abstract. We introduce the notion of Zimmer amenability for actions of discrete quantum groups on von Neumann algebras. We prove gen- eralizations of several fundamental results of the theory in the non- commutative case. In particular, we give a characterization of Zimmer amenability of an action α : G y N in terms of G-injectivity of the von Neumann algebra crossed product N ⋉α G. As an application we show that the actions of any discrete quantum group on its Poisson boundaries are always amenable. Contents Introduction 1. 2. Preliminaires 3. von Neumann algebra braided tensor products 4. Amenable actions 5. Examples 6. Amenable actions and crossed products: Kac algebra case 7. Amenable actions and crossed products: general case References 1 4 6 8 15 16 20 24 1. Introduction There are many different equivalent conditions that characterize amenabil- ity of a locally compact group G. One such characterization is in terms of a fixed point property of affine actions of G. In [22], Zimmer introduced the notion of amenable actions as a natural generalization of this fixed point property. In subsequent work, Adams, Elliott and Giordano characterized Zimmer amenability in terms of the existence of an equivariant conditional expectation [1]. In [2], Delaroche extended Zimmer's definition to the set- ting of group actions on von Neumann algebras. In this paper, we introduce the notion of Zimmer amenability for actions of discrete quantum groups on von Neumann algebras. 2010 Mathematics Subject Classification. Primary 46L89, 46L55; Secondary 46L07, 22D25. 1 2 M. S. M. MOAKHAR Definition 4.1. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N . Then α is called amenable if there exists a conditional expectation Eα : ℓ∞(G) ⊗ N → α(N ) such that (id ⊗ Eα)(∆ ⊗ id) = (∆ ⊗ id)Eα. This definition coincides with Delaroche's definition [2, D´efinition. 3.4] when G is a discrete group. Also observe that a discrete quantum group is amenable if and only if its action on the trivial space is amenable in the above sense. We prove, similarly to the classical result, the action of every discrete quantum group on itself by its co-multiplication is amenable (Proposition 5.1). Moreover, we show a connection between amenability of discrete quantum groups and amenability of their actions on von Neumann algebras which is in fact a noncommutative version of [2, Proposition 3.6]: Theoem 4.7. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N . The following are equivalent: 1. The quantum group G is amenable. 2. The action α is amenable and there exists an invariant state on N . In the case of Kac algebras, this theorem provides a new characterization for amenability of G in terms of amenability of the canonical action of G on its dual Kac algebra. Theorem 5.2. Let G be a discrete Kac algebra. Then G is amenable if and only if the canonical action of G on L∞( G) is amenable. One of Zimmer's main motivations to introduce and study the notion of amenable actions was the applications in the theory of random walks on G and their associated to Poisson boundaries of G. He proved that for any G, the action of G on its Poisson boundaries is always amenable [22, Theorem 5.2]. We establish the noncommutative analogue of this result in the case of discrete quantum group actions. Theorem 5.3. Let G be a discrete quantum group and let µ ∈ ℓ1(G) be a state. The canonical action of G on the Poisson boundary Hµ is amenable. In [23], Zimmer studied more properties of the amenable action and he characterized amenability of the action in terms of injectivity of the corre- sponding crossed product [23, Theorem 2.1]. In [2], Delaroche generalized this result to the case of actions on an arbitrary von Neumann algebra. In fact she proved that an action α : G y N is amenable if and only if there exists a conditional expectation from B(L2(G)) ⊗ N onto N ⋉α G [2, Proposition 4.1]. She used this result to show that amenability of the ac- tion on an injective von Neumann algebra is equivalent to injectivity of the corresponding crossed product. For discrete quantum group actions, we will characterize Zimmer amenability in terms of the existence of a conditional expectation that satisfies an equivariant condition coming from induced G action. More precisely we have AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS 3 Theorem 7.5. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N . The following are equivalent: 1. The action α is amenable. 2. There is an equivariant conditional expectation E : (cid:0)B(ℓ2(G)) ⊗ N, ∆op ⊗ id) → (cid:0)N ⋉α G, α(cid:1). As a direct consequence we will prove a noncommutative analogue of [2, Corollaire 4.2] for the general discrete quantum group actions. Corollary 7.7. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N . The following are equivalent: 1. The von Neumann algebra N is injective and the action α is amenable. 2. The crossed product N ⋉α G is G-injective. In the case of the trivial action of G on the trivial space, Theorem 7.5 provides a duality between amenability of G and injectivity of the dual von Neumann algebra L∞( G) in the category of T (ℓ2(G))-modules where T (L2(G)) is the predual of B(L2(G)). This perfect duality was initially investigated by Crann and Neufang in [6], (see also [5, 7]). Moreover in the case of discrete Kac algebra actions, we will show that the equivariant condition in Theorem 7.5 can be eliminated. In fact we have Theorem 6.3. Let α : G y N be an action of a discrete Kac algebra G on a von Neumann algebra N . The following are equivalent: 1. The action α is amenable. 2. There is a conditional expectation from B(ℓ2(G)) ⊗ N onto N ⋉α G. Beside this introduction, this paper includes six other sections. In section 2, we recall some notions about discrete quantum groups and their actions on von Neumann algebras. In section 3, we construct the von Neumann algebra braided tensor product and we use this notion to obtain a version of diagonal action in the setting of quantum groups. In section 4, we intro- duce the notion of amenable actions and we study some of its properties. In section 5, we give some examples of amenable actions. In particular we prove that the action of any discrete quantum groups on any of its Poisson boundaries is amenable. In section 6, we study actions of discrete Kac alge- bras. The main result of this section generalize the well-known fact about the equivalence of amenability of discrete Kac algebra G and injectivity of L∞( G). In section 7, we consider the latter result in the case of discrete quantum group actions. Acknowledgement. We are grateful to Massoud Amini for his contin- uous encouragement throughout this project. We would also like to thank Mehrdad Kalantar and Jason Crann for their helpful comments. 4 M. S. M. MOAKHAR 2. Preliminaires In this section we review some basic notions about discrete quantum groups and their actions on von Neumann algebras. A discrete quantum group G is a quadruple (ℓ∞(G), ∆, ϕ, ψ), where ℓ∞(G) = Li∈I Mni(C) is a von Neumann algebra direct sum of matrix algebras, ∆ : ℓ∞(G) → ℓ∞(G) ⊗ ℓ∞(G) is a co-associative co-multiplication, and ϕ and ψ are nor- mal faithful semi-finite left, respectively right, invariant weights on ℓ∞(G), that is, ϕ((ω ⊗ id)∆(x)) = ω(1)ϕ(x), ψ((id ⊗ ω)∆(x)) = ω(1)ψ(x), x ∈ Mϕ, ω ∈ ℓ1(G), x ∈ Mψ, ω ∈ ℓ1(G). A discrete quantum group G = (ℓ∞(G), ∆, ϕ, ψ) is a Kac algebra, if ϕ equals ψ and is a trace. The pre-adjoint of ∆ induces an associative completely contractive mul- tiplication ∗ : f ⊗ g ∈ ℓ1(G) b⊗ ℓ1(G) → f ∗ g = (f ⊗ g) ∆ ∈ ℓ1(G) on ℓ1(G). Moreover, this maps induces left and right actions of ℓ1(G) on ℓ∞(G) given by: (2.1) µ ∗ x := (id ⊗ µ)∆(x), x ∗ µ := (µ ⊗ id)∆(x). For a fixed µ ∈ ℓ1(G), the map x 7→ x ∗ µ is normal, completely bounded on ℓ∞(G). This map is called the Markov operator, if µ is moreover a state. A discrete quantum group G is said to be amenable if there exists a state m ∈ ℓ∞(G)∗ satisfying h m, x ∗ f i = h f, 1 ih m, x i, x ∈ ℓ∞(G), f ∈ ℓ1(G). The corresponding GNS Hilbert spaces ℓ2(G, ϕ) and ℓ2(G, ψ) are isomor- phic and are denoted by the same notation ℓ2(G). The (left) fundamental unitary W of G is a unitary operator on ℓ2(G) ⊗ ℓ2(G), satisfying the pen- tagonal relation W12W13W23 = W23W12, in which we used the leg notation W12 = W ⊗ 1, W23 = 1 ⊗ W and W13 = (1 ⊗ σ)W12, where σ(x ⊗ y) = y ⊗ x is the flip map on B(H ⊗ K). The right fundamental unitary V with the same properties is defined in a similar way on B(ℓ2(G) ⊗ ℓ2(G)). Let T (ℓ2(G)) be the predual of B(ℓ2(G)). Define the von Neumann alge- bra L∞( G) to be the weak-closure of {(ρ ⊗ id)W : ρ ∈ T (ℓ2(G))}. Consider the map ∆ : L∞( G) → L∞( G) ⊗ L∞( G) given by ∆(x) = W ∗(1 ⊗ x) W , where W = σW ∗σ. There exists a normal state ϕ on L∞( G) which is both invariant of left and right such that the triple G = (L∞( G), ∆, ϕ) is a compact quantum group called the dual quantum group of G. The opposite co-multiplication ∆op is given by ∆op = σ ◦ ∆. The fun- damental unitary W op associated to ∆op is defined by W op = σ V σ, and therefore W op ∈ L∞( G) ⊗ ℓ∞(G)′. AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS 5 The fundamental unitary W of G induces a co-associative co-multiplication on B(ℓ2(G)) defined by ∆ℓ : T ∈ B(ℓ2(G)) 7→ W ∗(1 ⊗ T )W ∈ B(ℓ2(G)) ⊗ B(ℓ2(G)). It is clear that the restriction of ∆ℓ to ℓ∞(G) is the original co-multiplication ∆ on ℓ∞(G). The pre-adjoint of ∆ℓ induces associative completely contrac- tive multiplication on the predual T (ℓ2(G)). ∗ : ω ⊗ τ ∈ T (ℓ2(G)) ⊗T (ℓ2(G)) 7→ ω ∗ τ = ∆ℓ ∗(ω ⊗ τ ) ∈ T (ℓ2(G)). If h T (ℓ2(G))∗T (ℓ2(G))i denotes the linear span of ω∗τ with ω, τ ∈ T (ℓ2(G)) we have (2.2) h T (ℓ2(G)) ∗ T (ℓ2(G))i = T (ℓ2(G)). Similarly to the equations (2.1), there are left and right actions of T (ℓ2(G)) on B(ℓ2(G)). There is also a co-associative co-multiplication on B(ℓ2(G)) induced by the right fundamental unitary V which is defined by ∆r : T ∈ B(ℓ2(G)) 7→ V (T ⊗ 1)V ∗ ∈ B(ℓ2(G)) ⊗ B(ℓ2(G)). In a same way, the pre-adjoint of ∆r induces associative completely contrac- tive multiplication on the predual T (ℓ2(G)) with the property (2.2). Let G be a discrete quantum group. By [5, Proposition 4.2.18], there is a conditional expectation E0 from B(ℓ2(G)) onto ℓ∞(G) such that for any x ∈ B(ℓ2(G)) and f ∈ T (ℓ2(G)), we have (2.3) E0(cid:0)(f ⊗ id)∆ℓ(x)(cid:1) = (f ⊗ id)∆ℓ(E0(x)). In particular, for any x ∈ L∞( G), E0(x) = ϕ(x)1 where ϕ is the normal invariant state of the compact quantum group G. A (left) action α : G y N of a discrete quantum group G on a von Neumann algebra N is an injective ∗-homomorphism α : N → ℓ∞(G) ⊗ N satisfying (∆ ⊗ id) α = (id ⊗ α) α . The action of dual quantum group G is defined similarly. Let α : G y N be an action of the discrete quantum group G on the von Neumann algebra N . A state ω on N is said to be invariant if (f ⊗ ω)α = h f, 1 iω. We denote by N α = {x ∈ N : α(x) = 1 ⊗ x} the fixed point algebra of the action α : G y N . Let θ be a normal semi-finite faithful weight on N , and let Hθ be the GN S Hilbert space of θ. It is proved in [18, Theorem 4.4] that α is implemented by a unitary Uα ∈ ℓ∞(G) ⊗ B(Hθ), that is, (2.4) α(x) = Uα (1 ⊗ x) U ∗ α (x ∈ N ) . 6 M. S. M. MOAKHAR Definition 2.1. Let α : G y N and β : G y M be two actions of the discrete quantum group G on von Neumann algebras N and M . Then a map Φ : N → M is equivariant if (id ⊗ Φ)α = β ◦ Φ. To indicate the actions, we say that the map Φ : (N, α) → (M, β) is equi- variant, or that Φ is (α, β)-equivariant. In the case α = β, we say that Φ is α-equivariant. The (von Neumann algebra) crossed product of the action α : G y N is defined by G ⋉α N := {α(N ) ∪ (L∞( G) ⊗ 1)}′′ ⊆ B(ℓ2(G)) ⊗ N. Analogously to the classical setting, there is a characterization of the crossed product G ⋉α N as the fixed point algebra of a certain action of G on B(ℓ2(G)) ⊗ N as follows: Theorem 2.2 ([9], Theorem 11.6). Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N and let χ be the flip map defined by χ(a⊗b) = b⊗a. Then there is a left action β on the von Neumann algebra B(ℓ2(G)) ⊗ N defined by β : x ∈ B(ℓ2(G)) ⊗ N 7→ (σV ∗σ ⊗ 1)(cid:0)(χ ⊗ 1)(id ⊗ α)(x)(cid:1)(σV σ ⊗ 1), such that G ⋉α N = (B(ℓ2(G)) ⊗ N )β. If α : G y N is an action of a discrete quantum group G on a von Neumann algebra N , there is is also a natural action α of ( G, ∆op) on G⋉αN which is called the dual action of α and is defined by α(α(x)) = 1 ⊗ α(x), α(x ⊗ 1) = ∆op(x) ⊗ 1, for all x ∈ N for all x ∈ L∞( G). In fact, we have α(N ) = (N ⋉α G) α [18, Theorem 2.7]. 3. von Neumann algebra braided tensor products In order to some technical obstacles we need to use a version of diagonal action for discrete quantum group actions. This section is devoted to a brief introduction to Yetter -- Drinfeld actions and braided tensor products in von Neumann algebra setting. For an overview of these notions, we refer to [3] and [15]. Let G = (ℓ∞(G), ∆, ϕ, ψ) be a discrete quantum group. Consider the triple (M, β, γ), where M is a von Neumann algebra on which β and γ of the discrete quantum group G and the dual quantum group G act. We say M is the G-YD-algebra if the actions β and γ satisfy the following Yetter -- Drinfeld condition: (3.1) (ad(W ) ⊗ id)(id ⊗ γ)β = (σ ⊗ id)(id ⊗ β)γ, AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS 7 where ad(W ) = W · W ∗. In this case, if α is any action of G on a von Neumann algebra N , then similarly to [19, Proposition 8.3], we have span{γ(M )12α(N )13} weak = span{α(N )13γ(M )12} weak* . Hence the weak-closed linear span of {γ(a)12α(b)13 : a ∈ M, b ∈ N } is a von Neumann subalgebra of B(ℓ2(G)) ⊗ M ⊗ N , which is called the braided ten- sor product of von Neumann algebras M and N , and is denoted by M ⊠ N . There is a ∗-homomorphism β ⊠ α : M ⊠ N → ℓ∞(G) ⊗ (M ⊠ N ) given by β ⊠ α(X) = W ∗ 12Uβ 13(1 ⊗ X)Uβ ∗ 13W12, where the unitary operator Uβ implements the action β by (2.4). In particular, on the set of generators {γ(M )12α(N )13} we have (β ⊠ α)(γ(a)12α(b)13) = W ∗ = W ∗ = W ∗ = W ∗ = W ∗ = W ∗ = W ∗ 12Uβ 13γ(a)23α(b)24U ∗ 12Uβ 13γ(a)23U ∗ 12(σ ⊗ id)(Uβ 23γ(a)13U ∗ 12(cid:0)(σ ⊗ id)(id ⊗ β)γ(a)(cid:1)123α(b)24W12 12(cid:0)(σ ⊗ id)(id ⊗ β)γ(a)(cid:1)123W12W ∗ 12(cid:0)(σ ⊗ id)(id ⊗ β)γ(a)(cid:1)123W12(cid:0)(∆ ⊗ id)α(b)(cid:1)124 12W12(cid:0)(∆ ⊗ id)α(b)(cid:1)124 12W12(cid:0)(id ⊗ γ)β(a)(cid:1)123W ∗ 12α(b)24W12 β 13 W12 α(b)24W12 )α(b)24W12 β 23 β 13 = (cid:0)(id ⊗ γ)β(a)(cid:1)123(cid:0)(∆ ⊗ id)α(b)(cid:1)124 = (cid:0)(id ⊗ γ)β(a)(cid:1)123(cid:0)(id ⊗ α)α(b)(cid:1)124. Therefore (β ⊠ α)(γ(a)12α(b)13) = (cid:0)(id ⊗ γ)β(a)(cid:1)123(cid:0)(id ⊗ α)α(b)(cid:1)124. Now it is straightforward to check that the normal ∗-homomorphism β ⊠ α is in fact an action of the discrete quantum group G on the von Neumann algebra M ⊠ N . If L and M are G-YD-algebras and N is a von Neumann algebra on which G acts, then similarly to [15], we can construct the braided tensor products (L ⊠ M ) ⊠ N and L ⊠ (M ⊠ N ) and there is a natural identification (3.2) (L ⊠ M ) ⊠ N ∼= L ⊠ (M ⊠ N ). For any discrete quantum group G, there is an action γ : G y ℓ∞(G) given by Observe that γ(x) = W ∗(1 ⊗ x) W . (ad(W ) ⊗ id)(id ⊗ γ)∆(x) = (σ ⊗ id)(id ⊗ ∆)γ(x). It implies that the pair (∆, γ) satisfies the compatibility condition (3.1) and therefore ℓ∞(G) is a G-YD-algebra. In this paper, we always consider braided tensor products whose first legs are ℓ∞(G). 8 M. S. M. MOAKHAR The following is the von Neumann algebraic version of [3, Lemma 1.24]. We included the proof for the convenience of the reader. Lemma 3.1. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N . There exists an equivarinat ∗-isomorphism Tα : (ℓ∞(G) ⊠ N, ∆ ⊠ α) → (ℓ∞(G) ⊗ N, ∆ ⊗ id) such that Tα(1 ⊠ a) = α(a) for all a ∈ N and Tα(x ⊠ 1) = x ⊗ 1 for all x ∈ ℓ∞(G). Proof. It is sufficient to define Tα on the set of generators {γ(a)12α(b)13}. For all a ∈ ℓ∞(G) and b ∈ N Tα(γ(a)12α(b)13) := (id⊗α−1)(cid:16)(σ⊗id)(cid:0)(σW ∗σ⊗1)(γ(a)12α(b)13)(σW σ⊗1)(cid:1)(cid:17). Then the map Tα : ℓ∞(G) ⊠ N → ℓ∞(G) ⊗ N is well-defined. Indeed, for a ∈ ℓ∞(G) and b ∈ N we have (σW ∗σ ⊗ 1)(γ(a)12α(b)13)(σW σ ⊗ 1) = (σW ∗σ ⊗ 1)(cid:0)( W ∗ ⊗ 1)(1 ⊗ a ⊗ 1)( W ⊗ 1)α(b)13(cid:1)(σW σ ⊗ 1) = ( W ⊗ 1)(cid:0)( W ∗ ⊗ 1)(1 ⊗ a ⊗ 1)(σW ∗σ ⊗ 1)α(b)13(cid:1)(σW σ ⊗ 1) = (1 ⊗ a ⊗ 1)(σW ∗σ ⊗ 1)α(b)13(σW σ ⊗ 1) 12α(b)23W12(cid:1) = (1 ⊗ a ⊗ 1)(σ ⊗ id)(cid:0)W ∗ 12α(b)23W12(cid:1) = (σ ⊗ id)(cid:0)(a ⊗ 1 ⊗ 1)W ∗ = (σ ⊗ id)(cid:0)(a ⊗ 1 ⊗ 1)(∆ ⊗ id)(α(b))(cid:1). Therefore by definition of Tα we have Tα(γ(a)12α(b)13) = (id ⊗ α−1)(cid:0)(a ⊗ 1 ⊗ 1)(id ⊗ α)(α(b))(cid:1) = (a ⊗ 1)α(b). Since the linear span of {(ℓ∞(G) ⊗ 1)α(N )} is weak dense in ℓ∞(G) ⊗ N , Tα is a ∗-isomorphism from ℓ∞(G) ⊠ N onto ℓ∞(G) ⊗ N and it is clear that for all a ∈ N and x ∈ ℓ∞(G), Tα(1 ⊠ a) = α(a) and Tα(x ⊠ 1) = x ⊗ 1. (cid:3) 4. Amenable actions In this section, we introduce the notion of amenable action of discrete quantum groups on von Neumann algebras. This definition is a gener- alization of the amenable action of discrete groups on von Neumann al- gebras introduced in [2, D´efinition 3.4]. Recall that the homomorphism α : G → Aut(M ) is called an action of a discrete group G on a von Neu- mann algebra M . If τ denotes the left translation action of G on ℓ∞(G), then the action α is called amenable if there exists an equivariant conditional expectation P : (ℓ∞(G) ⊗ M, τ ⊗ α) → (1 ⊗ M, α), i.e., P (τg ⊗ αg) = αg ◦ P, g ∈ G. AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS 9 There exists an automorphism Tα on ℓ∞(G) ⊗ M defined by Tα(cid:0)X (δg ⊗ xg)(cid:1) = X (δg ⊗ α−1 g (xg)). g∈G g∈G Since α(x) = Pg∈G(δg ⊗ αg(x)) for all x ∈ M , we have Tα(1 ⊗ M ) = α(M ). In some sense this means that the automorphism Tα make it possible to get away with the "twisting" effect of α. It is straightforward to check that (τg ⊗ id) ◦ Tα = Tα ◦ (τg ⊗ αg), for all g ∈ G. So Tα is an equivariant isomorphism from (ℓ∞(G) ⊗ M, τ ⊗ α) onto (ℓ∞(G) ⊗ M, τ ⊗ id). In summary, we have the following commutative diagram for the amenable action α of a discrete group G on a von Neumann algebra N : (ℓ∞(G) ⊗ M, τ ⊗ α) P (1 ⊗ M, τ ⊗ α) Tα Tα (4.1) (ℓ∞(G) ⊗ M, τ ⊗ id) P (α(M ), τ ⊗ id) This diagram allows us to define an equivalent definition for the amenable action of discrete groups on von Neumann algebras. Let α : G → Aut(M ) be an action of a discrete group G on a von Neumann algebra M and let τ be the left translation action on ℓ∞(G). Then the action α is called amenable if there exists an equivariant conditional expectation P : (ℓ∞(G) ⊗ M, τ ⊗ id) → (α(M ), τ ⊗ id). Motivated by this definition, we introduce the notion of the amenable action of discrete quantum groups on von Neumann algebras. Definition 4.1. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N . Then α is called amenable if there exists a conditional expectation Eα : ℓ∞(G) ⊗ N → α(N ) such that (4.2) (id ⊗ Eα)(∆ ⊗ id) = (∆ ⊗ id)Eα. Remark 4.2. The diagram (4.1) shows that the Definition 4.1 coincides with the classical definition of amenable actions introduced in [2]. Remark 4.3. The trivial action tr : G y C of a discrete quantum group G on the trivial space is amenable if and only if G is amenable. Indeed, if the trivial action tr is amenable, then there is an equivariant conditional expectation Etr : (ℓ∞(G) ⊗ C, ∆ ⊗ id) → (C ⊗ 1, ∆ ⊗ id). Define a state m on ℓ∞(G) by Etr(x ⊗ 1) = m(x)1 ⊗ 1. Then m(x ∗ f )1 ⊗ 1 = Etr(cid:0)(x ∗ f ) ⊗ 1(cid:1) = Etr(cid:0)(f ⊗ id ⊗ id)(∆ ⊗ id)(x ⊗ 1)(cid:1) 10 M. S. M. MOAKHAR = (f ⊗ id ⊗ id)(id ⊗ Etr)(∆ ⊗ id)(x ⊗ 1) = (f ⊗ id ⊗ id)(∆ ⊗ id)Etr(x ⊗ 1) = m(x)(f ⊗ id ⊗ id)(∆ ⊗ id)(1 ⊗ 1) = m(x)f (1)1 ⊗ 1. Therefore m is an invariant mean on ℓ∞(G). For the converse, if G is amenable, there exists an invariant state m on ℓ∞(G). Define the condi- tional expectation P : ℓ∞(G) ⊗ C → C ⊗ C by P = m ⊗ id. It is easy to check that P is (∆ ⊗ id)-equivariant. (See also Theorem 4.7.) Definition 4.4. Let α : G y M and β : G y N be actions of a discrete quantum group G on von Neumann algebras M and N , respectively, where M is a von Neumann subalgebra of N . Then 1. The triple (N, G, β) is an extension of (M, G, α) if α is the restriction of β to M and there is a conditional expectation from M onto N . 2. For the extension (N, G, β) of (M, G, α), the pair (N, M ) is called amenable, if there is an equivariant conditional expectation P from (N, β) onto (M, α). Proposition 4.5. Let (N, G, β) be an extension of (M, G, α). 1. If the action α is amenable, then the pair (N, M ) is amenable. 2. If the action β is amenable and the pair (N, M ) is amenable, then the action α is amenable. Proof. (1): Assume that (N, G, β) is an extension of (M, G, α) and therefore there is a conditional expectation Q from N onto M . Since α is amenable we have an equivariant conditional expectation Eα from (ℓ∞(G) ⊗ M, ∆ ⊗ id) onto (α(M ), ∆ ⊗ id). Define the conditional expectation P : N → M by P = α−1 ◦ Eα ◦ (id ⊗ Q) ◦ β. Then we have (id ⊗ P )β = (id ⊗ α−1 ◦ Eα)(id ⊗ Q)(id ⊗ β)β = (id ⊗ α−1)(id ⊗ Eα)(id ⊗ Q)(∆ ⊗ id)β = (id ⊗ α−1)(id ⊗ Eα)(∆ ⊗ id)(id ⊗ Q)β = (id ⊗ α−1)(∆ ⊗ id)Eα(id ⊗ Q)β = (id ⊗ α−1)(id ⊗ α)Eα(id ⊗ Q)β = α ◦ P. It shows that P : (N, β) → (M, α) is equivariant. (2): Now suppose that β is amenable, then there is an equivariant condi- tional expectation Eβ from (ℓ∞(G) ⊗ N, ∆ ⊗ id) onto (β(N ), ∆ ⊗ id). Since the pair (N, M ) is amenable, there is also a conditional expectation P from N onto M such that (id ⊗ P )β = α ◦ P . Hence the composition AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS 11 (id ⊗ P )E : ℓ∞(G) ⊗ N → α(M ) is a conditional expectation such that (id ⊗ id ⊗ P )(id ⊗ E)(∆ ⊗ id) = (id ⊗ id ⊗ P )(∆ ⊗ id)E = (∆ ⊗ id)(id ⊗ P )E. Since α(M ) ⊆ ℓ∞(G) ⊗ M , by restricting of (id ⊗ P )E to ℓ∞(G) ⊗ M we obtain an equivariant conditional expectation from (ℓ∞(G) ⊗ M, ∆ ⊗ id) onto (α(M ), ∆ ⊗ id), and therefore the action α is amenable. (cid:3) Let α : G y N be an action of a discrete quantum group G on a von Neu- mann algebra N . Consider the conditional expectation E0 as the equation (2.3) and fix an arbitrary state f ∈ ℓ1(G). Then E0 ⊗ f ⊗ id is a condi- tional expectation from B(ℓ2(G)) ⊗ ℓ∞(G) ⊗ N onto B(ℓ2(G)) ⊗ 1 ⊗ N . By restricting we obtain a conditional expectation from ℓ∞(G) ⊠ N onto 1 ⊠ N . So the triple (ℓ∞(G) ⊠ N, ∆ ⊠ α, G) is an extension of (1 ⊠ N, ∆ ⊠ α, G). Proposition 4.6. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N . Then the action α is amenable if and only if for the extension (ℓ∞(G) ⊠ N, ∆ ⊠ α, G) of (1 ⊠ N, ∆ ⊠ α, G), the pair (ℓ∞(G) ⊠ N, 1 ⊠ N ) is amenable. Proof. By Lemma 3.1, there exists an equivariant ∗-isomorphism between (ℓ∞(G) ⊠ N, ∆ ⊠ α) and (ℓ∞(G) ⊗ N, ∆ ⊗ id). Since (ℓ∞(G) ⊠ N, ∆ ⊠ α, G) is an extension of (1 ⊠ N, ∆ ⊠ α, G), the action α is amenable if and only if the pair (ℓ∞(G) ⊠ N, 1 ⊠ N ) is amenable. (cid:3) The following result is a noncommutative version of [2, Proposition 3.6]. Theorem 4.7. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N . The following are equivalent: 1. The quantum group G is amenable. 2. The action α is amenable and there exists an invariant state on N . Proof. (2) ⇒ (1): suppose that ω is an invariant state on N and Eα is an equivariant conditional expectation from ℓ∞(G) ⊗ N onto α(N ) coming from amenability of the action α. Define a state m on ℓ∞(G) by h m, x i = h ω, α−1 ◦ Eα(x ⊗ 1) i. Then m is a left invariant state on ℓ∞(G). Indeed, for any x ∈ ℓ∞(G) and f ∈ ℓ1(G) we have h m, x ∗ f i = h m, (f ⊗ id)∆(x) i = h ω, α−1 ◦ Eα(cid:0)((f ⊗ id)∆(x)) ⊗ 1(cid:1) i = h ω, α−1(cid:0)(f ⊗ id ⊗ id)(id ⊗ Eα)(∆ ⊗ id)(x ⊗ 1)(cid:1) i = h ω, α−1(cid:0)(f ⊗ id ⊗ id)(∆ ⊗ id)Eα(x ⊗ 1)(cid:1) i = h ω, α−1(cid:0)(f ⊗ id ⊗ id)(id ⊗ α)Eα(x ⊗ 1)(cid:1) i = h ω, (f ⊗ id)Eα(x ⊗ 1) i 12 M. S. M. MOAKHAR = f (1)h ω, α−1 ◦ Eα(x ⊗ 1) i = f (1)h m, x i, where we use the fact that ω is invariant in the penultimate step. (1) ⇒ (2): suppose that m is a left invariant mean on ℓ∞(G). Fix η ∈ N∗ and define ω := (m ⊗ η)α. Then for any f ∈ ℓ1(G) and x ∈ N , we have h f ⊗ ω, α(x) i = h ω, (f ⊗ id)α(x) i = h m ⊗ η, (f ⊗ id ⊗ id)(id ⊗ α)α(x) i = h m ⊗ η, (f ⊗ id ⊗ id)(∆ ⊗ id)α(x) i = h m, (f ⊗ id)∆(cid:0)(id ⊗ η)α(x)(cid:1) i = f (1)h m, (id ⊗ η)α(x) i = f (1)h m ⊗ η, α(x) i = f (1)h ω, x i, it shows ω is an invariant state on N . Now we prove that the action α is amenable. First, we claim that for any x ∈ ℓ∞(G) ⊠ N , we have (∆ ⊠ α)(x) = (ad(W ∗) ⊗ id ⊗ id)(σ ⊗ id ⊗ id)(id ⊗ ∆ ⊗ id)(x). (4.3) Since the linear span of {γ(ai)12α(bi)13} is weak* dense in ℓ∞(G) ⊠ N , we need to show (4.3) for the set of generators. For any a ∈ ℓ∞(G) and b ∈ N we have (∆ ⊠ α)(cid:0)γ(a)12α(b)13(cid:1) = (cid:0)(id ⊗ γ)∆(a)(cid:1)123(cid:0)(id ⊗ α)α(b)(cid:1)124 12(σ ⊗ id)(cid:0)(id ⊗ ∆)γ(a)(cid:1)W12(cid:1)123(cid:0)(id ⊗ α)α(b)(cid:1)124 12(σ ⊗ id)(cid:0)(id ⊗ ∆)γ(a)(cid:1)W12(cid:1)123(cid:0)(∆ ⊗ id)α(b)(cid:1)124 12(σ ⊗ id)(cid:0)(id ⊗ ∆)γ(a)(cid:1)W12(cid:1)123W ∗ 12(cid:0)(σ ⊗ id)(id ⊗ ∆)γ(a)(cid:1)123α(b)24W12 = (cid:0)(ad(W ∗) ⊗ id)(σ ⊗ id)(id ⊗ ∆)γ(a)(cid:1)123(cid:0)(id ⊗ α)α(b)(cid:1)124 = (cid:0)W ∗ = (cid:0)W ∗ = (cid:0)W ∗ = (ad(W ∗) ⊗ id ⊗ id)(σ ⊗ id ⊗ id)(cid:2)(cid:0)(id ⊗ ∆)γ(a)(cid:1)123 α(b)14(cid:3) = (ad(W ∗) ⊗ id ⊗ id)(σ ⊗ id ⊗ id)(id ⊗ ∆ ⊗ id)(γ(a)12 α(b)13). 12α(b)24W12 = W ∗ Hence we conclude the equality (4.3). Let E0 : B(ℓ2(G)) → ℓ∞(G) be the normal conditional expectation given by (2.3). Therefore for any x ∈ L∞( G) and f, ω ∈ ℓ1(G), we have h ω ⊗ f, (id ⊗ E0)∆ℓ(x)i = h f, E0(cid:0)(ω ⊗ id)∆ℓ(x)(cid:1)i = h f, (ω ⊗ id)∆ℓ(E0(x))i = h ω ⊗ f, ϕ(x)1 ⊗ 1i. Hence (4.4) (id ⊗ E0)∆ℓ(x) = ϕ(x)1 ⊗ 1, AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS 13 for all x ∈ L∞( G). Consider the conditional expectation E0⊗m⊗id from the von Neumann algebra B(ℓ2(G)) ⊗ ℓ∞(G) ⊗ N onto ℓ∞(G) ⊗ 1 ⊗ N . Then by restricting, there is a conditional expectation E from ℓ∞(G) ⊠ N onto 1 ⊠ N . We show that the conditional expectation E is (∆ ⊠ α)-equivariant. For any a ∈ ℓ∞(G) and b ∈ N , by the equality (4.4) we have (id ⊗ E0 ⊗ id ⊗ id)hW ∗ 12γ(a)23 α(b)24W12i = (id ⊗ E0 ⊗ id ⊗ id)hW ∗ = (id ⊗ E0 ⊗ id ⊗ id)h(cid:0)(∆ℓ ⊗ id)γ(a)(cid:1)123(cid:0)(id ⊗ α)α(b)(cid:1)124i = (cid:0)(id ⊗ E0 ⊗ id)(∆ℓ ⊗ id)γ(a)(cid:1)123(cid:0)(id ⊗ α)α(b)(cid:1)124 = (id ⊗ ϕ ⊗ id ⊗ id)(cid:0)γ(a)23(cid:1)(cid:0)(id ⊗ α)α(b)(cid:1)124. 12α(b)24W12i 12γ(a)23W12W ∗ Consider x ∈ ℓ∞(G) ⊠ N asX i∈I γ(ai)12α(bi)13. Since m is an invariant mean, the equality (4.3) and the above calculation yield that (id ⊗ E)(∆ ⊠ α)(x) 12(cid:0)(σ ⊗ id ⊗ id)(id ⊗ ∆ ⊗ id)(x)(cid:1)W12i 12(cid:0)(σ ⊗ id ⊗ id)(id ⊗ ∆ ⊗ id)(x)(cid:1)W12i = (id ⊗ E)h(ad(W ∗) ⊗ id ⊗ id)(σ ⊗ id ⊗ id)(id ⊗ ∆ ⊗ id)(x)i = (id ⊗ E)hW ∗ = (id ⊗ E0 ⊗ m ⊗ id)hW ∗ = (id ⊗ E0 ⊗ id)hW ∗ = (id ⊗ E0 ⊗ id)hW ∗ = (id ⊗ E0 ⊗ m ⊗ id)hW ∗ = (id ⊗ E0 ⊗ m ⊗ id)hW ∗ = (id ⊗ id ⊗ m ⊗ id)hX 12(1 ⊗ x)W12i 12X (id ⊗ ϕ ⊗ id ⊗ id)(cid:0)γ(ai)23(cid:1)(cid:0)(id ⊗ α)α(bi)(cid:1)124i, 12(cid:0)(σ ⊗ id)(id ⊗ id ⊗ m ⊗ id)(id ⊗ ∆ ⊗ id)(x)(cid:1)W12i 12(cid:0)(σ ⊗ id)(id ⊗ id ⊗ m ⊗ id)(x134)(cid:1)W12i γ(ai)23α(bi)24 W12i i∈I i∈I where we use the normality of the conditional expectation E0 in the last equality. On the other hand, repeating the calculation show that (∆⊠α)E(x) = (∆ ⊠ α)(E0 ⊗ m ⊗ id)(x) = (∆ ⊠ α)(E0 ⊗ m ⊗ id)(cid:2)X = (∆ ⊠ α)(id ⊗ m ⊗ id)hX i∈I γ(ai)12 α(bi)13(cid:3) (E0 ⊗ id ⊗ id)(cid:2)γ(ai)12 α(bi)13(cid:3)i i∈I 14 M. S. M. MOAKHAR = (∆ ⊠ α)(id ⊗ m ⊗ id)hX = (∆ ⊠ α)(id ⊗ m ⊗ id)hX = (id ⊗ id ⊗ m ⊗ id)hX i∈I i∈I i∈I (E0 ⊗ id ⊗ id)(cid:2)γ(ai)12(cid:3)α(bi)13i ( ϕ ⊗ id ⊗ id)(cid:2)γ(ai)12(cid:3) α(bi)13i (id ⊗ ϕ ⊗ id ⊗ id)(cid:0)γ(ai)23(cid:1)(id ⊗ α)α(bi)(cid:1)124i From these two calculations, it follows that the pair (ℓ∞(G) ⊠ N, 1 ⊠ N ) is amenable and by Proposition 4.6 the action α is amenable. (cid:3) Theorem 4.8. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N . Then there is an equivariant isomorphism Φ from (cid:0)(ℓ∞(G) ⊠ N ) ⋉∆⊠α G, \∆ ⊠ α(cid:1) onto (cid:0)B(ℓ2(G)) ⊗ N, \∆ ⊠ α(cid:1) such that Φ maps (1 ⊠ N ) ⋉∆⊠α G onto N ⋉α G. Proof. Consider the equivariant ∗-isomorphism Tα, given by Lemma 3.1, from (ℓ∞(G) ⊠ N, ∆ ⊠ α) onto (ℓ∞(G) ⊗ N, ∆ ⊗ id). Then the isomorphism Φ is obtained from the identification: (ℓ∞(G) ⊠ N ) ⋉∆⊠α G = [ (∆ ⊠ α)(ℓ∞(G) ⊠ N ) ∪ (L∞( G) ⊗ 1ℓ∞(G) ⊠ N ) ]′′ ∼= [ (id ⊗ Tα)(∆ ⊠ α)(ℓ∞(G) ⊠ N ) ∪ (L∞( G) ⊗ 1ℓ∞(G) ⊗ N ) ]′′ = [ (∆ ⊗ id)Tα(ℓ∞(G) ⊠ N ) ∪ (L∞( G) ⊗ 1 ⊗ 1)V12V ∗ = [ (∆ ⊗ id)Tα(ℓ∞(G) ⊠ N ) ∪ V12(L∞( G) ⊗ 1 ⊗ 1)V ∗ = [ (∆ ⊗ id)Tα(ℓ∞(G) ⊠ N ) ∪ (∆ ⊗ id)(L∞( G) ⊗ 1) ]′′ ∼= [ Tα(ℓ∞(G) ⊠ N ) ∪ (L∞( G) ⊗ 1) ]′′ ∼= [(ℓ∞(G) ⊗ N ) ∪ (L∞( G) ⊗ 1) ]′′ = B(ℓ2(G)) ⊗ N 12 ]′′ 12 ]′′ ′ where in the fourth equality, we used the fact that V ∈ L∞( G) particular, for any x ∈ ℓ∞(G) ⊠ N and any x ∈ L∞( G) we have (4.5) Φ(cid:0)(∆ ⊠ α)(x)(cid:1) = Tα(x), Φ(x ⊗ 1ℓ∞(G) ⊠ N ) = x ⊗ 1ℓ∞(G) ⊗ N . ⊗ ℓ∞(G). In From Lemma 3.1, we know that Tα(1 ⊠ N ) = α(N ), and therefore by the same calculations we have (1 ⊠ N ) ⋉∆⊠α G ∼= N ⋉α G. In order to show the equivariant condition, it is sufficient to check the equality on the set of generators of (ℓ∞(G) ⊠ N ) ⋉∆⊠α G. Suppose that x ∈ ℓ∞(G) ⊠ N , then since W op ∈ L∞( G) ⊗ ℓ∞(G)′, by (4.5) we have (cid:0)( ∆op ⊗ id) ◦ Φ(cid:1)(∆ ⊠ α)(x) = ( ∆op ⊗ id)(Φ(∆ ⊠ α)(x)) = ( ∆op ⊗ id)Tα(x) = W op∗ 12 (1 ⊗ Tα(x)) W op 12 AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS 15 = 1 ⊗ Tα(x) = 1 ⊗ Φ((∆ ⊠ α)(x)) = (id ⊗ Φ)(cid:0)1 ⊗ (∆ ⊠ α)(x)(cid:1) = (id ⊗ Φ)(cid:0)(\∆ ⊠ α)(∆ ⊠ α)(x)(cid:1). On the other hand, by (4.5), for any x ∈ L∞( G) we have (cid:0)( ∆op ⊗ id) ◦ Φ(cid:1)(x ⊗ 1ℓ∞(G) ⊠ N ) = ∆op(x) ⊗ 1ℓ∞(G) ⊗ N = (id ⊗ Φ)(\∆ ⊠ α)(x ⊗ 1ℓ∞(G) ⊠ N ). (cid:3) 5. Examples In this section, we give some examples of amenable actions of discrete quantum groups on von Neumann algebras. In Theorem 4.7 we showed that the amenable quantum group G acts amenably on any von Neumann algebras. Also, Proposition 4.5 shows that it is possible to get new amenable actions by appropriate restrictions. Below we give more concrete examples of amenable actions. As an application of amenable actions, the action of any discrete group G on ℓ∞(G) is always amenable [2, Remarques 3.7.(b)]. The next result is the noncommutative analogue of that. Proposition 5.1. Every discrete quantum group acts amenably on itself. Proof. Define the map Φ : ℓ∞(G) ⊗ ℓ∞(G) → ℓ∞(G) by Φ(A) = (id ⊗ ε)(A), in which ε is the co-unit in ℓ1(G). Then Φ ◦ ∆ = (id ⊗ ε)∆ = id. So the map Φ is a left inverse of the co-multiplication ∆ and therefore the map E∆ := ∆ ◦ Φ is a conditional expectation from ℓ∞(G) ⊗ ℓ∞(G) onto ∆(ℓ∞(G)). Moreover, for any A ∈ ℓ∞(G) ⊗ ℓ∞(G) we have (id ⊗ E∆)(∆ ⊗ id)(A) = (id ⊗ ∆ ◦ Φ)(∆ ⊗ id)(A) = (id ⊗ ∆)(id ⊗ id ⊗ ε)(∆ ⊗ id)(A) = (id ⊗ ∆)∆(cid:0)(id ⊗ ε)(A)(cid:1) = (∆ ⊗ id)(cid:0)(∆ ◦ Φ)(A)(cid:1) = (∆ ⊗ id)E∆(A). Hence E∆ : (ℓ∞(G) ⊗ ℓ∞(G), ∆⊗id) → (∆(ℓ∞(G)), ∆⊗id) is an equivariant conditional expectation. (cid:3) For a discrete quantum group G the restriction of the extended comulti- plication ∆ℓ to L∞( G) provides an action of G on the von Neumann algebra L∞( G). we next prove in the case of Kac algebras, amenability of the latter action is equivalent to amenability of the Kac algebra G. Theorem 5.2. Let G be a discrete Kac algebra. Then G is amenable if and only if the canonical action ∆ℓ : G y L∞( G) is amenable. L∞(G) 16 M. S. M. MOAKHAR Proof. Since G is a Kac algebra, by [10, Corollary 3.9], the tracial Haar state ϕ of the dual quantum group G is invariant with respect to the ac- tion ∆ℓ is equivalent to amenability of G. . Hence by Theorem 4.7, amenability of the action ∆ℓ L∞(G) L∞(G) (cid:3) Remark 5.3. In [8], Crann defined a notion of inner amenability for quan- tum groups as the existence of an invariant state for the canonical action : G y L∞( G). The same proof shows that Theorem 5.2 holds for ∆ℓ the inner amenable quantum group G in the sense of Crann. L∞(G) In the next result, we state the noncommutative version of Zimmer's classical result [22, Theorem 5.2] that all Poisson boundaries are amenable G-space. Let us first recall the definition of noncommutative Poisson bound- aries in the sense of Izumi [11]. Let µ ∈ ℓ1(G) be a state. Recall in this case Φµ(x) = (µ ⊗ id)∆(x) is a unital, normal completely positive map on ℓ∞(G). The space of fixed point Hµ = {x ∈ ℓ∞(G) : Φµ(x) = x} is a w∗-closed operator system in ℓ∞(G). There is a conditional expectation Eµ from ℓ∞(G) onto Hµ. Then the corresponding Choi -- Effros product induces the von Neumann algebraic structure on Hµ [4]. This von Neumann algebra is called noncommutative Poisson boundary with respect to µ. For more details on noncommutative Poisson boundaries we refer the reader to [13] and [14]. By [14, Proposition 2.1], the restriction of ∆ to Hµ induces a left action ∆µ of G on the von Neumann algebra Hµ. We prove this action is amenable. Theorem 5.4. Let G be a discrete quantum group and let µ ∈ ℓ1(G) be a state. The left action ∆µ of G on the Poisson boundary Hµ is amenable. Proof. The conditional expectation Eµ : ℓ∞(G) → Hµ is equivariant, see e.g. the proof of [14, Proposition 2.1], and therefore the pair (ℓ∞(G), Hµ) is amenable. Since by Proposition 5.1 the action of discrete quantum group on itself is amenable, it follows from Proposition 4.5 that the left action ∆µ is amenable. (cid:3) Remark 5.5. In [21], Vaes and Vergnioux introduced the amenable action of a discrete quantum group on a unital C-algebra. They proved that the canonical C-algebraic action of a universal discrete quantum group on its boundary is always amenable. Therefore the related crossed product becomes nuclear. 6. Amenable actions and crossed products: Kac algebra case In this section we characterize amenability of actions in term of von Neu- mann algebra crossed products. Classically, the action α : G y X of a discrete group G on a standard probability space (X, ν) is amenable if and only if the crossed product L∞(X, ν) ⋉α G is injective. Delaroche extended this result to the action of locally compact groups on arbitrary von Neumann algebras. We prove a noncommutative version of this result in the case of AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS 17 discrete Kac algebra actions on von Neumann algebras. This in particular generalizes a part of Theorem 4.5 in [16] and also [17, Corollary 3.17] which establish the equivalence between amenability of a discrete Kac algebra G and injectivity of L∞( G). The case of actions of general discrete quantum groups on von Neumann algebras is discussed in next section. Lemma 6.1. Let α : G y N be an action of a discrete Kac algbera G on a von Neumann algebra N and let M be a von Neumann subalgebra of N which is invariant under α. The following are equivalent: 1. There is an equivariant conditional expectation P : (N, α) → (M, α). 2. There is a conditional expectation E : N ⋉α G → M ⋉α G. Proof. (1) ⇒ (2): since P : (N, α) → (M, α) is an equivariant condi- tional expectation, it follows E = id ⊗ P is a conditional expectation from B(ℓ2(G)) ⊗ N onto B(ℓ2(G)) ⊗ M such that (χ ⊗ id)(id ⊗ α)E = (id ⊗ E)(χ ⊗ id)(id ⊗ α), where χ is the flip map. Then id ⊗ E is a conditional expectation from ℓ∞(G) ⊗ B(ℓ2(G)) ⊗ N onto ℓ∞(G) ⊗ B(ℓ2(G)) ⊗ M . Recall the left action β of G on B(ℓ2(G)) ⊗ N , defined in Theorem 2.2. One can see that for any y ∈ B(ℓ2(G)) ⊗ N we have β ◦ E(y) = (id ⊗ E)β(y). In particular, if y ∈ (B(ℓ2(G)) ⊗ N )β we have β ◦ E(y) = (id ⊗ E)β(y) = (id ⊗ E)(1 ⊗ y) = 1 ⊗ E(y), which implies E(y) ∈ (B(ℓ2(G)) ⊗ M )β. Thus in view of Theorem 2.2, the restriction of E is a conditional expectation from N ⋉α G onto M ⋉α G. (2) ⇒ (1): suppose E : N ⋉α G → M ⋉α G is the conditional expectation and ϕ is the tracial Haar state of the dual Kac algebra G. There is a canonical conditional expectation E ϕ from M ⋉α G onto α(M ) defined by (6.1) E ϕ(x) = ( ϕ ⊗ id)α(x), for all x ∈ M ⋉α G. We claim that E ϕ ◦ E : (N ⋉α G, ∆ ⊗ id) → (α(M ), ∆ ⊗ id) is an equivari- ant conditional expectation. Since the left fundamental unitary W lies in ℓ∞(G) ⊗ L∞( G), for all z ∈ N ⋉α G we have (id ⊗ E)(∆ ⊗ id)(z) = (id ⊗ E)(W ∗ 12z23W12) 12(1 ⊗ E(z))W12 = W ∗ = (∆ ⊗ id)(E(z)). Therefore in order to conclude the claim, it is sufficient to show that the canonical conditional expectation E ϕ : (M ⋉α G, ∆ ⊗ id) → (α(M ), ∆ ⊗ id) is equivariant. First, consider x ∈ L∞( G). Then (id ⊗ E ϕ)(∆ ⊗ id)(x ⊗ 1) = (id ⊗ E ϕ)(W ∗ 12(1 ⊗ x ⊗ 1)W12) = (id ⊗ ϕ ⊗ id)(id ⊗ α)(W ∗ = (id ⊗ ϕ ⊗ id ⊗ id)(id ⊗ ∆op ⊗ id)(W ∗ 12(1 ⊗ x ⊗ 1)W12) 12(1 ⊗ x ⊗ 1)W12) 18 M. S. M. MOAKHAR = (id ⊗ id ⊗ ϕ ⊗ id)(id ⊗ ∆ ⊗ id)(W ∗ = (id ⊗ id ⊗ ϕ ⊗ id)(W ∗ 13(1 ⊗ 1 ⊗ x ⊗ 1)W13). 12(1 ⊗ x ⊗ 1)W12) Now consider a complete orthonormal system {ej}j∈J . Then Similarly to the proof of [17, Corollary 3.17], for any normal states f ∈ ℓ1(G) and ω ∈ M∗, and any vector state ωξ ∈ T (ℓ2(G)) we get h ωξ ⊗ f ⊗ ω, (id ⊗ E ϕ)(∆ ⊗ id)(x ⊗ 1)i = f (1)ω(1)h ωξ ⊗ ϕ, W ∗(1 ⊗ x)W i (ωξ,ej ⊗ id)(W )∗ x(ωξ,ej ⊗ id)(W )i j∈J h ϕ, (ωξ,ej ⊗ id)(W )∗ x(ωξ,ej ⊗ id)(W )i = h ϕ, (ωξ ⊗ id)(W ∗(1 ⊗ x)W )i = h ϕ,X = X = X = X h ϕ, (ω J ξ, Jej j∈J j∈J j∈J h ϕ, (ωξ,ej ⊗ id)(W )(ωξ,ej ⊗ id)(W )∗ xi ⊗ id)(W )∗(ω Jξ, Jep ⊗ id)(W )xi = ω Jξ(1) ϕ(x) = ωξ(1) ϕ(x), where J is the modular conjugation for the tracial Haar state ϕ. So for any x ∈ L∞( G) we have (id ⊗ E ϕ)(∆ ⊗ id)(x ⊗ 1) = ϕ(x)1 ⊗ 1 ⊗ 1 = (∆ ⊗ id)( ϕ ⊗ id ⊗ id)( ∆op(x) ⊗ 1) = (∆ ⊗ id)( ϕ ⊗ id)(cid:0) α(x ⊗ 1)(cid:1) = (∆ ⊗ id)E ϕ(x ⊗ 1). Hence for any x ∈ L∞( G) and x ∈ M we get (id ⊗ E ϕ)(∆ ⊗ id)(cid:0)(x ⊗ 1)α(x)(cid:1) = (id ⊗ E ϕ)(cid:0)(∆ ⊗ id)(x ⊗ 1)(∆ ⊗ id)α(x)(cid:1) = (id ⊗ E ϕ)(cid:0)(∆ ⊗ id)(x ⊗ 1)(id ⊗ α)α(x)(cid:1) = (id ⊗ E ϕ)(cid:0)(∆ ⊗ id)(x ⊗ 1)(cid:1)(id ⊗ α)α(x) = (∆ ⊗ id)(cid:0)E ϕ(x ⊗ 1)(cid:1)(∆ ⊗ id)α(x) = (∆ ⊗ id)E ϕ(cid:0)(x ⊗ 1)α(x)(cid:1). Since the crossed product M ⋉α G is generated by {ℓ∞( G) ⊗ 1, α(M )}, it follows that the conditional expectation E ϕ is (∆ ⊗ id)-equivariant which completes the proof of claim. Define the conditional expectation P := α−1 ◦ E ϕ ◦ E ◦ α from N onto M . We show that (id ⊗ P )α = α ◦ P . Since E ϕ ◦ E is equivariant with respect to the action ∆ ⊗ id, for all a ∈ N we have (id ⊗ E ϕ ◦ E)(id ⊗ α)α(a) = (id ⊗ E ϕ ◦ E)(∆ ⊗ id)α(a) AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS 19 = (∆ ⊗ id)(E ϕ ◦ E)α(a) = (id ⊗ α) ◦ (E ϕ ◦ E)(α(a)), where in the last equality we use that (E ϕ ◦ E)α(a) ∈ α(M ). Now it follows (id ⊗ P )α = (id ⊗ α−1)(id ⊗ E ϕ ◦ E)(id ⊗ α)α = E ϕ ◦ E ◦ α = α ◦ P. (cid:3) The following is the noncommutative analogue of the main result of [2] for discrete Kac algebra actions: (See [2, Theorem 4.2].) Theorem 6.2. Let α : G y N be an action of a discrete Kac algebra G on a von Neumann algebra N . The following are equivalent: 1. The action α is amenable. 2. There is a conditional expectation from (ℓ∞(G) ⊠ N ) ⋉∆⊠α G onto (1 ⊠ N ) ⋉∆⊠α G. 3. For any extension (M, G, β) of (1 ⊠ N, G, ∆⊠α), the pair (M, 1 ⊠ N ) is amenable. Proof. (1) ⇒ (2): suppose that α is amenable, then by Proposition 4.6 the pair (ℓ∞(G) ⊠ N, 1 ⊠ N ) is amenable which means there is an equivariant conditional expectation from (ℓ∞(G) ⊠ N, ∆⊠α) onto (1 ⊠ N, ∆⊠α). Hence (2) follows by Lemma 6.1. (2) ⇒ (3): suppose that (M, G, β) is an extension of (1 ⊠ N, G, ∆ ⊠ α), let q be a conditional expectation from M onto 1 ⊠ N . Then id ⊗ q is a conditional expectation from B(ℓ2(G)) ⊗ M onto B(ℓ2(G)) ⊗ (1 ⊠ N ), and thus Theorem 4.8 yields a conditional expectation E : (ℓ∞(G) ⊠ M ) ⋉∆⊠β G → (ℓ∞(G) ⊠ (1 ⊠ N )) ⋉∆⊠β G. Moreover by the assumption there is a conditional expectation from the crossed product (ℓ∞(G) ⊠ N ) ⋉∆⊠α G onto (1 ⊠ N ) ⋉∆⊠α G. By the equality (3.2), it is equivalent to the existence of a conditional expectation E0 from (cid:0)ℓ∞(G) ⊠ (1 ⊠ N )(cid:1) ⋉∆⊠β G onto (cid:0)1 ⊠ (1 ⊠ N )(cid:1) ⋉∆⊠β G. By composing, we obtain the conditional expectation E0 ◦ E : (ℓ∞(G) ⊠ M ) ⋉∆⊠β G → (cid:0)1 ⊠ (1 ⊠ N )(cid:1) ⋉∆⊠β G. Hence from Lemma 6.1, we have an equivariant conditional expectation Q from (ℓ∞(G) ⊠ M, ∆ ⊠ β) onto (cid:0)1 ⊠ (1 ⊠ N ), ∆ ⊠ β(cid:1). Since (M, G, β) is an extension of (1 ⊠ N, G, ∆ ⊠ α), the restriction of Q to 1 ⊠ M yields a conditional expectation Q0 : β(M ) ∼= 1 ⊠ M → 1 ⊠ (1 ⊠ N ) ∼= β(1 ⊠ N ) such that (id ⊗ Q0)(∆ ⊠ β) = (∆ ⊠ β) ◦ Q0. Hence for any a ∈ M we have (id ⊗ Q0)(id ⊗ β)β(a) = (id ⊗ β)Q0(cid:0)β(a)(cid:1). Now define a conditional expectation P := β−1 ◦ Q0 ◦ β from M onto 1 ⊠ N . Then for all a ∈ M we have (id ⊗ P )β(a) = (id ⊗ β−1)(id ⊗ Q0)(id ⊗ β)β(a) = (id ⊗ β−1)(id ⊗ β)Q0(cid:0)β(a)(cid:1) 20 M. S. M. MOAKHAR = β ◦ P (a), which implies P : (M, β) → (1 ⊠ N, ∆ ⊠ α) is an equivariant conditional expectation. Hence the pair (M, 1 ⊠ N ) is amenable. (3) ⇒ (1): consider the canonical extension (ℓ∞(G) ⊠ N, G, ∆ ⊠ α) of the triple (1 ⊠ N, G, ∆⊠α). Then by the assumption the pair (ℓ∞(G) ⊠ N, 1 ⊠ N ) must be amenable. Hence the action α is amenable by Proposition 4.6. (cid:3) Theorem 6.3. Let α : G y N be an action of a discrete Kac algebra G on a von Neuamnn algebra N . Then the following are equivalent: 1. The action α is amenable. 2. There is a conditional expectation from B(ℓ2(G)) ⊗ N onto N ⋉α G. Proof. By Theorem 4.8, there is an isomorphism from (ℓ∞(G) ⊠ N ) ⋉∆⊠α G onto B(ℓ2(G)) ⊗ N which maps (1 ⊠ N ) ⋉∆⊠α G onto N ⋉α G. So the Theorem follows by the equivalence of (1) and (2) in Theorem 6.2. (cid:3) Corollary 6.4. Let α : G y N be an action of a discrete Kac algebra G on a von Neuamnn algebra N . Then the following are equivalent: 1. The von Neumann algebra N is injective and the action α is amenable. 2. The crossed product N ⋉α G is injective. Proof. (1) ⇒ (2): if N is injective then so is B(ℓ2(G)) ⊗ N . If α is amenable, Theorem 6.3 yields a conditional expectation from B(ℓ2(G)) ⊗ N onto the crossed product N ⋉α G. Since B(ℓ2(G)) ⊗ N is injective, N ⋉α G is also injective. (2) ⇒ (1): since the crossed product N ⋉α G is injective, there is a condi- tional expectation from B(ℓ2(G)) ⊗ N onto N ⋉α G. Therefore by Theorem 6.3, the action α is amenable. Moreover since there is always the canonical conditional expectation from N ⋉α G on α(N ), it follows that α(N ) and equivalently N , is injective. (cid:3) 7. Amenable actions and crossed products: general case In this section, we generalize the duality of Corollary 6.3 to the setting of discrete quantum group actions. For this end, we basically need to show Lemma 6.1 for general discrete quantum groups. Recall that in the proof of the implication (2) to (1) of Lemma 6.1, we construct an equivariant condi- tional expectation from (α(N ), ∆ ⊗ id) onto (α(M ), ∆ ⊗ id) by composing the restriction Eα(N) with the canonical conditional expectation E ϕ. In the case of discrete Kac algebras, E ϕ is automatically equivariant with respect to the action ∆ ⊗ id. But this is no longer the case in the general setting of discrete quantum group actions, since the Haar state ϕ is not a trace. To overcome this issue, we impose an extra assumption on the conditional ex- pectation E : N ⋉α G → M ⋉α G to be equivariant with respect to the dual action α. This would imply that E maps α(N ) onto α(M ), hence use of the canonical conditional expectation E ϕ is no longer necessary. This inspired AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS 21 by the work of Crann and Neufang in [6], where they proved a character- ization of amenability of the general locally compact quantum group G in terms of covariant injectivity of the dual von Neumann algebra L∞( G). Lemma 7.1. Let G be a discrete quantum group. Then for any y ∈ B(ℓ2(G)) we have (id ⊗ ∆r) ∆op(y) = ( ∆op ⊗ id)∆r(y). Proof. Let x ∈ ℓ∞(G) and x ∈ L∞( G). Since the fundamental unitaries ′ W op and V lie in L∞( G) ⊗ ℓ∞(G)′ and L∞( G) ⊗ ℓ∞(G), respectively, we have ∆op(x) = 1 ⊗ x and ∆r(x) = x ⊗ 1. Therefore (id ⊗ ∆r) ∆op(x) = (id ⊗ ∆r)(1 ⊗ x) = 1 ⊗ ∆r(x) = ( ∆op ⊗ id)∆r(x), and (id ⊗ ∆r) ∆op(x) = ∆op(x) ⊗ 1 = ( ∆op ⊗ id)(x ⊗ 1) = ( ∆op ⊗ id)∆r(x). Since the co-multiplications ∆r and ∆op are homomorphisms, and the linear span of {xx : x ∈ ℓ∞(G), x ∈ L∞( G)} is weak* dense in B(ℓ2(G)) [20, Proposition 2.5], we obtain the desired equality on B(ℓ2(G)). (cid:3) In the following we use the same idea as [7, Proposition 4.2] to show an automatic equivariant property with respect to the dual action. Proposition 7.2. Let G be a discrete quantum group and let N be a von Neumann algebra. Then any (∆r ⊗ id)-equivariant map on B(ℓ2(G)) ⊗ N is automatically ( ∆op ⊗ id)-equivariant. Proof. Let Φ be an equivariant map on (B(ℓ2(G)) ⊗ N, ∆r ⊗ id). Consider normal states τ, ω ∈ T (ℓ2(G)), f ∈ ℓ1(G) and g ∈ N∗. Then for any x ∈ B(ℓ2(G)) ⊗ N we have h f ⊗ τ ⊗ω ⊗ g, (id ⊗ ∆r ⊗ id)(id ⊗ Φ)( ∆op ⊗ id)(x)i = h τ ⊗ ω ⊗ g, (∆r ⊗ id)Φ(cid:0)(f ⊗ id ⊗ id)( ∆op ⊗ id)(x)(cid:1)i = h τ ⊗ ω ⊗ g, (id ⊗ Φ)(cid:2)(∆r ⊗ id)(cid:0)(f ⊗ id ⊗ id)( ∆op ⊗ id)(x)(cid:1)(cid:3)i = h ω ⊗ g, Φ(cid:2)(τ ⊗ id ⊗ id)(∆r ⊗ id)(cid:0)(f ⊗ id ⊗ id)( ∆op ⊗ id)(x)(cid:1)(cid:3)i = h ω ⊗ g, Φ(cid:2)(f ⊗ τ ⊗ id ⊗ id)(id ⊗ ∆r ⊗ id)(cid:0)( ∆op ⊗ id)(x)(cid:1)(cid:3)i = h ω ⊗ g, Φ(cid:2)(f ⊗ τ ⊗ id ⊗ id)( ∆op ⊗ id ⊗ id)(cid:0)(∆r ⊗ id)(x)(cid:1)(cid:3)i = h ω ⊗ g, (f ⊗ τ ⊗ id ⊗ id)( ∆op ⊗ id ⊗ id)Φ(cid:0)(∆r ⊗ id)(x)(cid:1)i = h f ⊗ τ ⊗ ω ⊗ g, ( ∆op ⊗ id ⊗ id)(∆r ⊗ id)(cid:0)Φ(x)(cid:1)i = h f ⊗ τ ⊗ ω ⊗ g, (id ⊗ ∆r ⊗ id)( ∆op ⊗ id)(cid:0)Φ(x)(cid:1)i. 22 M. S. M. MOAKHAR Since {(τ ⊗ ω)∆r : τ, ω ∈ T (ℓ2(G))} spans a dense subset of T (ℓ2(G)), see (2.2), it follows (id ⊗ Φ)(cid:0)( ∆op ⊗ id)(x)(cid:1) = ( ∆op ⊗ id)Φ(x). (cid:3) Corollary 7.3. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N and let M be a von Neumann subalgebra of N which is invariant under α. If E : (N ⋉α G, ∆r ⊗ id) → (M ⋉α G, ∆r ⊗ id) is an equivariant conditional expectation, then E is equivariant with respect to the dual action α. Proof. Note that the dual action α is the restriction of ∆op⊗id to the crossed product N ⋉α G ⊆ B(ℓ2(G)) ⊗ N . Hence Proposition 7.2 implies that the conditional expectation E is equivariant with respect to α. (cid:3) Lemma 7.4. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N and let M be a von Neumann subalgebra of N which is invariant under α. The following are equivalent: 1. There is an equivariant conditional expectation P : (N, α) → (M, α). 2. There is an equivariant conditional expectation E : (N ⋉α G, α) → (M ⋉α G, α). Proof. (1) ⇒ (2): similarly as in the proof of Lemma 6.1, we see that the restriction of id ⊗ P : B(ℓ2(G)) ⊗ N → B(ℓ2(G)) ⊗ M to the crossed product N ⋉α G yields a conditional expectation E from N ⋉α G onto M ⋉α G. It is easy to see that E is (∆r ⊗ id)-equivariant. Thanks to Corollary 7.3 the conditional expectation E is equivariant with respect to the dual action α. (2) ⇒ (1): suppose that E : (N ⋉α G, α) → (M ⋉α G, α) is an equivairant conditional expectation. Then for all x ∈ N we have 1L∞( G) ⊗ E(α(x)) = (id ⊗ E)(cid:0)1L∞( G) ⊗ α(x)(cid:1) = (id ⊗ E) ◦ α(α(x)) = α ◦ E(α(x)). It follows that E(α(x)) is in the fixed point algebra of the dual action α on M ⋉α G. Hence E(α(N )) ⊆ α(M ). Now define the conditional expectation P := α−1 ◦ E ◦ α from N onto M . Similarly to the proof of Lemma 6.1, we show that (id ⊗ P )α = α ◦ P . Since the fundamental unitary W lies in ℓ∞(G) ⊗ L∞( G), it follows that the condtional expectation E is (∆ ⊗ id)- equivariant. Now for any x ∈ N we have (id ⊗ E)(id ⊗ α)α(x) = (id ⊗ E)(∆ ⊗ id)α(x) = (∆ ⊗ id)E(α(x)) = (id ⊗ α)E(α(x)), where in the last equality we use that E(α(x)) ∈ α(M ). Now it follows (id ⊗ P )α = (id ⊗ α−1)(id ⊗ E)(id ⊗ α)α = E ◦ α = α ◦ P. (cid:3) AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS 23 The equivalence of (1) and (3) in the following result is a noncommutative analogue of Zimmer's classical result [23, Theorem. 2.1]. Theorem 7.5. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N . The following are equivalent: 1. The action α is amenable. 2. There is an equivariant conditional expectation E : (cid:0)(ℓ∞(G) ⊠ N ) ⋉∆⊠α G, \∆ ⊠ α(cid:1) → (cid:0)(1 ⊠ N ) ⋉∆⊠α G, \∆ ⊠ α(cid:1). 3. There is an equivariant conditional expectation E : (cid:0)B(ℓ2(G)) ⊗ N, ∆op ⊗ id) → (cid:0)N ⋉α G, α(cid:1). Proof. By Theorem 4.8, the statements (2) and (3) are equivalent. To con- clude (1) and (2), thanks to Proposition 4.6 amenability of α is equivalent to amenability of the pair (ℓ∞(G) ⊠ N, 1 ⊠ N ) which by Lemma 7.4 is equiva- lent to the existence of a \∆ ⊠ α-equivariant conditional expectation E from (ℓ∞(G) ⊠ N ) ⋉∆⊠α G onto (1 ⊠ N ) ⋉∆⊠α G (cid:3) Remark 7.6. Since the trivial action tr : G y C is amenable if and only if G is amenable, and C ⋉tr G = L∞( G), the equivalence of (1) and (3) in Theorem 7.5 in fact gives a generalization of the main result of [6]. Suppose that β : G y K is an action of a discrete quantum group G on a von Neumann algebra K. We say that K is G-injective if for every unital completely isometric equivariant map ι : (M, α1) → (N, α2) and every unital completely positive equivariant map Ψ : (M, α1) → (K, β) there is a unital completely positive equivariant map Ψ : (N, α2) → (K, β) such that Ψ ◦ ι = Ψ. Corollary 7.7. Let α : G y N be an action of a discrete quantum group G on a von Neumann algebra N . The following are equivalent: 1. The von Neumann algebra N is injective and the action α is amenable. 2. The crossed product N ⋉α G is G-injective. Proof. (1)⇒(2): the proof is similar to the proof of Corollary 6.4, only that we use Theorem 7.5 instead of Theorem 6.3. (2)⇒(1): since the crossed product N ⋉α G is G-injective, the identity map on N ⋉α G can be extended to an equivariant conditional expectation from (B(ℓ2(G)) ⊗ N, ∆op ⊗ id) onto (N ⋉α G, α). Hence by Theorem 7.5, the action α is amenable. Moreover there is always the canonical conditional expectation from N ⋉α G onto α(N ), it follows that N is injective. (cid:3) Corollary 7.8 ([14], Corollary 2.5). Let G be a discrete quantum group and let µ ∈ ℓ1(G) be a state. The von Neumann algebra crossed product Hµ ⋉∆µ G is injective. Proof. By Theorem 5.4 the action of G on its Poisson boundaries is always amenable, and therefore the result follows by Corollary 7.7. (cid:3) 24 M. S. M. MOAKHAR Remark 7.9. Crann and Kalantar informed us in a recent unpublished paper they have independently defined a notion of Zimmer amenability in the setting of actions of locally compact quantum groups on von Neumann algebras, where they used a homological approach. But their definition is equivalent to Definition 4.1 in the case of discrete quantum groups. They have obtained a similar result as Corollary 7.7 in that general context. References [1] S. Adams, G. Elliott and T. Giordano, Amenable actions of groups, Trans. Amer. Math. Soc. 344 (1994),no. 2, 803 -- 822. [2] C. Anantharaman-Delaroche, Action moyennable d'un groupe localement compact sur une alg`ebre de von Neumann, Math. Scand. 45 (1979), 289 -- 304. [3] S. Barlak, G. Szab´o and C. Voigt, The spatial Rokhlin property for actions of compact quantum groups, J. Funct. Anal. 272 (2017), 2308 -- 2360. [4] M.-D. Choi and E. Effros, Injectivity and operator spaces, J. Funct. Anal. 24 (1977), no. 2, 156 -- 209. [5] J. Crann, Homological manifestations of quantum group duality, Ph.D. thesis, Carleton University Press, Ottawa, 2015. [6] J. Crann and M. Neufang, Amenability and covariant injectivity of locally compact quantum groups, Trans. Amer. Math. Soc. 368 (2016), no. 1, 495 -- 513. [7] J. Crann, Amenability and covariant injectivity of locally compact quantum groups II, Canadian J. Math. 69 (2017), no. 5, 1064 -- 1086. [8] J. Crann, Inner amenability and approximation properties of locally compact quantum groups, Indiana Univ. Math. J. to appear. arXiv:1709.017770. [9] M. Enock, Measured quantum groupoids in action, M´em. Soc. Math. Fr. (N.S.), 114, 2008. [10] M. Izumi, Non-commutative Poisson boundaries and compact quantum group actions, Adv. Math. 169 (2002), no. 1, 1 -- 57. [11] M. Izumi, Non-commutative Poisson boundaries, in: Discrete geometric analysis, Contemp. Math. 347, Amer. Math. Soc., Providence, RI, 2004, pp. 69 -- 81. [12] M. Junge, M. Neufang and Z.-J. Ruan, A representation theorem for locally compact quantum groups, Internat. J. Math. 20 (2009), no. 3, 377 -- 400. [13] M. Kalantar, M. Neufang and Z.-J. Ruan, Poisson boundaries over locally compact quantum groups, Internat. J. Math. 24 (2013), no. 3, 1350023, 21 pp. [14] M. Kalantar, M. Neufang and Z.-J. Ruan, Realization of quantum group Poisson boundaries as crossed products, Bull. London Math. Soc 46 (2014), 1267 -- 1275. [15] R. Nest and C. Voigt, Equivariant Pointcar´e duality for quantum group actoins, J. Funct. Anal. 258 (2010), 1466 -- 1503. [16] Z.-J. Ruan, Amenability of Hopf von Neumann algebras and Kac algebras, J. Funct. Anal. 139 (1996), no. 2, 466 -- 499. [17] R. Tomatsu, Amenable discrete quantum groups, J. Math. Soc. Japan 58 (2006) no. 4, 949 -- 964. [18] S. Vaes, The unitary implementation of a locally compact quantum group action, J. Funct. Anal. 180 (2001), 426 -- 480. [19] S. Vaes, A new approach to induction and imprimitivity results, J. Funct. Anal. 229 (2005), 317 -- 374. [20] S. Vaes and A. van Daele The Heisenberg commutation relations, commuting squares and the Haar measure on locally compact quantum groups, Operator algebras and Mathematical Physics 379 -- 400 (Constanta, 2001) (Theta, Bucharest, 2003) [21] S. Vaes and R. Vergnioux The boundary of universal discrete quantum groups, exact- ness and factoriality, Duke. Math. J. 140 (2007), 35 -- 84. AMENABLE ACTIONS OF DISCRETE QUANTUM GROUPS 25 [22] R. J. Zimmer, Amenable ergodic group actions and an application to Poisson bound- aries of random walks, J. Funct. Anal. 27 (1978), 350 -- 372. [23] R. J. Zimmer, Hyperfinite factors and amenable ergodic group actions, Inven. Math. 41 (1977), 23 -- 31. Department of Mathematics, Tarbiat Modares University, Tehran 14115- 134, Iran E-mail address: [email protected]
1610.08945	4	1610	2017-03-08T08:56:23	Constructing MASAs with prescribed properties	[ "math.OA" ]	We consider an iterative procedure for constructing maximal abelian $^*$-subalgebras (MASAs) satisfying prescribed properties in II$_1$ factors. This method pairs well with the intertwining by bimodules technique and with properties of the MASA and of the ambient factor that can be described locally. We obtain such a local characterization for II$_1$ factors $M$ that have an {\it s-MASA}, $A\subset M$ (i.e., for which $A \vee JAJ$ is maximal abelian in $\Cal B(L^2M)$), and use this strategy to prove that any factor in this class has uncountably many non-intertwinable singular (respectively semiregular) s-MASAs.	math.OA	math	CONSTRUCTING MASAS WITH PRESCRIBED PROPERTIES Sorin Popa University of California, Los Angeles Abstract. We consider an iterative procedure for constructing maximal abelian ∗-subalgebras (MASAs) satisfying prescribed properties in II1 factors. This method pairs well with the intertwining by bimodules technique and with properties of the MASA and of the ambient factor that can be described locally. We obtain such a local characterization for II1 factors M that have an s-MASA, A ⊂ M (i.e., for which A ∨ J AJ is maximal abelian in B(L2M)), and use this strategy to prove that any factor in this class has uncountably many non-intertwinable singular (respectively semiregular) s-MASAs. 7 1 0 2 r a M 8 ] . A O h t a m [ 4 v 5 4 9 8 0 . 0 1 6 1 : v i X r a 0. Introduction Given a separable II1 factor M , one can construct a maximal abelian ∗-subalgebra (abreviated hereafter as MASA) A in M as an inductive limit of finite partitions. This iterative procedure pairs well with properties of MASAs that can be charac- terized locally, allowing the construction of A in a manner that makes "more and more" of the desired properties be satisfied. This technique has been initiated in [P81a], [P81d] where it was used to prove that any separable II1 factor M contains a MASA A ⊂ M whose normalizer NM (A) := {u ∈ U(M ) uAu∗ = A} generates a factor (A is semiregular in M ; see [P81a]), as well as a MASA A ⊂ M whose normalizer is trivial, i.e. NM (A) = U(A) (A is singular in M ; see [P81d]). In this paper we obtain more refined applications of this method, by combin- ing it with two additional ingredients: the intertwining by bimodule technique ([P01], [P03]) and local properties of the ambient II1 factor M , such as existence Supported by NSF Grant DMS-1400208, a Simons Fellowship, and Chaire d'Excellence de la FSMP 2016 1 Typeset by AMS-TEX 2 SORIN POPA of non-trivial central sequences (i.e., property Gamma of [MvN43]) and s-thin ap- proximation, a property that we introduce here and which will be defined shortly. Recall in this respect that if Q, P are von Neumann subalgebras in a II1 factor M , then we write Q ≺M P if there exists a Hilbert Q − P sub-bimodule H ⊂ L2M such that dimHP < ∞. In certain cases (notably if Q, P are MASAs) this condition is equivalent to the existence of a non-zero partial isometry v ∈ M such that v∗v ∈ Q and vQv∗ ⊂ P . Our first result shows that any separable II1 factor M contains an uncountable family of singular (respectively semiregular) MASAs {Ai}i such that Ai 6≺M Aj, ∀i 6= j, with A containing non-trivial central sequences of M whenever M does. This will in fact follow from the following stronger result. 0.1. Theorem. Let M be a separable II1 factor M and N ⊂ M a subfactor with trivial relative commutant, N′ ∩ M = C. Let Pn ⊂ M be a sequence of von Neumann subalgebras such that N 6≺M Pn, ∀n. Then N contains a singular (respectively semiregular) maximal abelian ∗-subalgebra A of M such that A 6≺M Pn, ∀n. Moreover, if N ≃ R, then one can take A so that to satisfy NM (A)′′ = N , and if N contains non-trivial central sequences of M , then A can be taken so that to contain non-trivial central sequences of M as well. We then consider the class of II1 factors M which have an s-MASA, i.e., a MASA A ⊂ M such that the von Neumann algebra A ∨ JAJ ⊂ B(L2M ), generated by left and right multiplication by elements in A on the Hilbert space L2M , is a MASA in B(L2M ). We obtain a local characterization of factors in this class, by proving that M has an s-MASA if and only if it satisfies the following approximation property, that we call s-thin: for any finite partition {pi}i ⊂ M , any finite set F ⊂ M and any ε > 0, there exist a partition {qj}j ⊂ M refining {pi}i and an element ξ ∈ M such that any x ∈ F can be ε-approximated in the norm-k k2 by linear combinations of elements of the form qjξqk. We show that factors with s-MASAs are closed under amplifications and inductive limits and combine their local characterization with the iterative procedure to prove the following: 0.2. Theorem. If M has an s-MASA, then there exist uncountably many non- intertwinable s-MASAs in M , which in addition can be chosen singular (resp. semiregular). The typical example of s-MASAs in II1 factors are the Cartan (or regular) MASAs, i.e., MASAs A ⊂ M for which NM (A)′′ = M (cf [FM77]). Any group measure space II1 factor M = L∞(X) ⋊ Γ, obtained from a free ergodic measure preserving action Γ y X of a countable group Γ on a probability measure space (X, µ), has A = L∞(X) as a Cartan subalgebra, which is thus also an s-MASA. CONSTRUCTING MASAS 3 The above result shows that such factors necessarily have singular s-MASAs as well. Note that when M is hyperfinite, this fact was already known since ([D54], [Pu61]), where the first concrete exemples of singular s-MASAs were given. By [OP07], [PV11], [PV12], there are large classes of group measure space II1 factors that have unique (up to unitary conjugacy) Cartan subalgebras (= regu- lar MASAs), while by Theorem 0.2 above, such a factor always has "many" non conjugate semiregular s-MASAs. There are by now several classes of II1 factors known to have no Cartan sub- algebras, obtained first by using free probability theory ([Vo96], then by using deformation-rigidity theory ([OP07], [CS11], [CSU11], [PV11], [PV12], [I12]). It is interesting to note that in each case when one could prove absence of Cartan MASAs by using free probability, the same techniques could be used to show ab- sence of s-MASAs as well (notably for the free group factors L(Fn), cf. [G98]). While there is much evidence that II1 factors with s-MASAs but no Cartan subalgebras do exist, the problem of constructing such examples remains open. Another open problem is to find new proofs for the non-existence of s-MASAs in certain II1 factors, such as the free group factors L(Fn). But perhaps the most "urgent" open problem in this direction is to find an intrinsic, local characterization of II1 factors having Cartan subalgebras. Such an intrinsic characterization may lead to interesting applications in deformation-rigidity theory. It may also allow to prove that the class of factors with Cartan MASAs is close to inductive limits, a permanence property that, as we mentioned above, factors with s-MASAs do have. We discuss these open problems and other related questions in the last section of the paper. This work has been finalized while I was visiting RIMS and Kyoto University in September 2016. I am very grateful to Masaki Izumi and Narutaka Ozawa for the warm hospitality extended to me during my stay. Added in the proof. In their very recent paper Thin II1 factors with no Cartan subalgebras (math.OA/1611.02138), Anna Krogager and Stefaan Vaes were able to construct a large class of II1 factors that have s-MASAs but no Cartan subalgebras (in fact are even strongly solid) thus solving a problem stated above and in 5.1.2. 1. Preliminaries All finite von Neumann algebras that we consider in this paper will come with a fixed normal faithful trace state, denoted τ , and they will always be assumed separable with respect to the Hilbert norm k k2 implemented by τ . If M is a finite von Neumann algebra and B ⊂ M is a von Neumann subalgebra, then EB denotes the unique τ -preserving conditional expectation of M onto B. If B ⊂ M is merely 4 SORIN POPA a weakly closed ∗-subalgebra of M (so 1B not necessarily equal to 1M ), then we use the same notation EB for the unique trace preserving expectation of 1BM 1B onto B that preserves the trace state τ (·)/τ (p) on 1BM 1B. We denote by U(M ) the unitary group of a von Neumann algebra M . If X is a Banach space and S ⊂ X is a subset, then we denote by (S)1 the set of elements in S that have norm at most 1. For all notations that are not specified in the text, we send the reader to the expository notes ([P06]), and for basics on von Neumann algebras to the classic book [D57] or the recent [AP17]. 1.1. Perturbation of projections. The following result is well known, but we state it here in the specific form needed in this paper. 1.1.1. Lemma. Let M be a finite von Neumann algebra, B ⊂ M a diffuse von Neumann subalgebra and e ∈ P(M ). Then there exists a projection f ∈ B of trace equal to τ (e) such that kf − ek2 ≤ 14ke − EB(e)k2 + p13ke − EB(e)k2. Proof. By (Lemma 1.1 in [P81c]), if p is the spectral projection of EB(e) corre- sponding to the interval [1/2, 1], then kp − EB(e)k2 ≤ 13ke − EB(e)k2. By using the Cauchy-Schwartz inequality, this implies τ (p) − τ (e) = τ (p) − τ (EB(e)) ≤ 13ke − EB(e)k2. Thus, if we take f ∈ P(B) so that τ (f ) = τ (e) and satisfying f ≤ p if τ (p) ≥ τ (e) and f ≥ p if τ (p) ≤ τ (e), then τ (f ) − τ (p) ≤ 13ke − EB(e)k2. Altogether, kf − ek2 ≤ kf − pk2 + kp − EB(e)k2 + kEB(e) − ek2 ≤ 14ke − EB(e)k2 + p13ke − EB(e)k2. (cid:3) 1.2. Embedding L∞([0, 1]) in II1 factors. We will view a diffuse abelian von Neumann subalgebra A of a separable II1 factor M as an embedding of L∞([0, 1]) ≃ A into M . Thus, if L∞([0, 1]) is represented as an inductive limit of finer and finer partitions (e.g., dyadic) generating the σ-algebra of Lebesgue measurable subsets of [0, 1], then such an embedding is determined by the corresponding increasing sequence of finite dimensional subalgebras An ր A. As pointed out in [P13], any embedding L∞([0, 1]) ≃ A ⊂ M acts weak mixingly on M ⊖ A′ ∩ M , and this entails the following 2-independence property: 1.2.1. Theorem [P13]. Let M be a finite von Neumann algebra and B ⊂ M a diffuse von Neumann subalgebra. Given any finite set F ⊂ M ⊖ B ∨ (B′ ∩ M ), any n ≥ 2 and any ε > 0, there exists a partition of 1 with projections in B, p1, ..., pn ∈ P(B) of trace 1/n, such that kpixpik2 2 − τ (pi)2τ (x∗x) ≤ ε, ∀x ∈ F . The fact that L∞([0, 1]) ≃ A ⊂ M is a MASA is an extremality condition for the embedding, which can be described locally as follows (see e.g. [P81a]): CONSTRUCTING MASAS 5 1.2.2. Lemma. Let M be a separable finite von Neumann algebra and B ⊂ M a von Neumann subalgebra. Let An ⊂ B be an increasing sequence of finite di- . Let {xj}j ⊂ M be mensional von Neumann subalgebras and denote A = ∪nAn a countable set, k k2-dense in the unit ball of M . Then A is maximal abelian in n∩M (EB(xj)) − EAn(xj)k2 = 0, ∀j ≥ 1. Moreover, if A B if and only if limn kEA′ is maximal abelian in B, then limn kEA′ n∩(B∨B ′∩M )(xj) − EAn∨B ′∩M (xj)k2 = 0, ∀j ≥ 1. w Proof. Since {xj}j dense in (M )1 in the norm k k2 implies {EB(jj)}j k k2-dense in (B)1, the first part amounts to A being maximal abelian in B iff A′ ∩ B = A. The last part follows from by combining the first part with the fact that ∩nA′n ∩ (B ∨ B′ ∩ M ) = (A′ ∩ B) ∨ (B′ ∩ M ) while An ∨ (B′ ∩ M ) ր A ∨ (B′ ∩ M ). (cid:3) Let us also mention a result that's essentially contained in (Section A.1 of [P92]), but which we derive here from 1.2.1 and 1.2.2 above: 1.2.3. Corollary. Let M be a separable finite von Neumann algebra and B ⊂ M a diffuse von Neumann subalgebra. There exists a MASA A in B such that A′ ∩ M = A ∨ (B′ ∩ M ). Moreover, if B = N is a II1 factor, then one can take A to be contained in a hyperfinite II1 subfactor A ⊂ R ⊂ N satisfying R′ ∩ M = N′ ∩ M . Proof. Let {xj}j ⊂ (M )1 be k k2-dense sequence in the unit ball of M . We construct recursively an increasing sequence of finite dimensional abelian von Neu- m∩M (xj) − EAm∨B ′∩M (xj)k2 ≤ 2−m for all mann algebras Am ⊂ B such that kEA′ 1 ≤ j ≤ m. Assuming we have constructed these algebras up to m = n, we con- struct An+1 as follows. By Theorem 1.2.1, given any α > 0, there exists an abelian finite dimensional ∗-subalgebra A0 ′∩M (xj) − ′∩B)∨(B ′∩M )(xj)k2 < α, 1 ≤ j ≤ n + 1. Then by taking fist a MASA A0 in B E(A0 that contains A0 n+1 and then using Lemma 1.2.2, we find a finite dimensional abelian subalgebra Aα ′∩B)∨(B ′∩M )(xj) − n+1 (xj)k2 ≤ 2−n−1, for all EAα 1 ≤ j ≤ n + 1. Taking α sufficiently small and letting An+1 = Aα n+1∨B ′∩M (xj)k2 < α and kEAα n+1 containing An such that kEA0 ′∩M (EB(xj)) − EAα n+1, we get n+1 n+1 n+1 ⊂ A0 that contains A0 n+1, such that kE(Aα n+1 n+1 kEA′ n+1∩M (xj) − EAn+1∨B ′∩M (xj)k2 < 2α ≤ 2−n−1, while we still have kEA′ But then A = ∪nAn w n+1∩M (EB(xj))−EAn+1 (xj)k2 ≤ 2−n−1, for all 1 ≤ j ≤ n+1. ⊂ B clearly satisfies the required condition. In the case B = N is a II1 factor, then we construct the increasing sequence of abelian subalgebras An above to also be dyadic (i.e., all its minimal projections be have the same trace, equal to some 2−kn ) and so that each An be the diagonal of a 6 SORIN POPA ij}i,j, such that each en 11 = M0 and the finite set F0 = {en 1ixken . Assuming we have constructed Am ⊂ Rm = sp{em matrix algebra Rn with matrix units {en ij is a sum of some en+1 ij }i,j for 1 ≤ m ≤ n, we kl construct it for m = n+1 by applying the first part to the inclusion B0 = en 11 ⊂ en 11M en j1 1 ≤ k ≤ n + 1, 1 ≤ i, j ≤ n}, with appropriate α, to get a finite dimensional subalgebra A1 n+1 ⊂ B0 such that kEA1 n+1 to be dyadic. We then choose matrix units {ekl}k,l ⊂ B0 such that ekk are the minimal projections of A1 1j k, l, 1 ≤ i, j ≤ n}, Rn+1 = sp{en+1 0∩M0(x)k2 < α, ∀x ∈ F0. Moreover, we can take A1 n+1 and then define {en+1 st }s,t = {en ′∩M0 (x)−EB0∨B ′ n+1 i1eklen 11N en st }s,t. w w Thus, if we let A = ∪nAn as above and put R = ∪nRn , then we still get the condition A′ ∩ M = A ∨ N′ ∩ M , but also R′ ∩ M = R′ ∩ (A′ ∩ M ) = N′ ∩ M . (cid:3) 1.3. Intertwining subalgebras in factors. We recall here some basic facts about the "intertwining" subordination relation between subalgebras in II1 factors, from [P01], [P03]. We will follow the presentation [P05a] of this topic, which emphasized the "intertwining space" between subalgebras. Thus, if M is a finite von Neumann algebra and Q, P ⊂ M are weakly closed ∗- subalgebras of M , then IM (Q, P ) denotes the set of vectors ξ ∈ L2(1QM 1P with the property that the Hilbert Q − P bimodule spQξP ⊂ L2M has finite dimension as a right P -module. This space is clearly invariant to taking sums and to multiplication by Q from the left and P from the right. We call it the intertwining Q − P sub- bimodule of M . The space IM (Q, P ) has left support ≤ 1Q and right support ≤ 1P , it is invariant to multiplication from the left by Q′ ∩ M and from the right by P ′ ∩ M and it is increasing in P and decreasing in Q. Also, IM (Q, P ) = IM (Q1, P1) whenever Q1 ⊂ Q, P ⊂ P1 have finite index in the sense of [PP84] (either the "probabilistic" definition, or the existence of a finite orthonormal basis; see 1.2 in [P94] for the equivalence between these alternative definitions). Moreover, if q ∈ Q, p ∈ P are projections that have central trace of support 1 in Q, respectively P , then IM (qQq, pP p) = qIM (Q, P )p. We'll denote as usual by hM, P i the basic construction algebra, defined as the commutant in B(L2(M 1P )) of the algebra of right multiplication by elements in P . It is also equal to the von Neumann algebra generated by operators of the form xeP y∗, with x, y ∈ M 1P , acting on ξ ∈ M 1P ⊂ L2(M 1P ) by xeP y∗(ξ) = xEP (y∗ξ). Then the projection sQ,P := ∨{s(ξeP ξ∗) ξ ∈ IM (Q, P )} is equal to the support of the direct summand of Q′ ∩ 1QhM, P i1Q generated by projections that are finite in 1QhM, P i1Q (where we have used the notation s(T ) for the support projection of a positive operator T ). Thus, if ξ ∈ L2M , then ξ ⊥ IM (Q, P ) iff ξeP ξ∗sQ,P = 0 CONSTRUCTING MASAS 7 and iff ξeP ξ∗ is orthogonal on any projection q′ ∈ Q′ ∩ hM, P i with q′hM, P iq′ finite. If IM (Q, P ) 6= 0, then we say that Q can be intertwined into P inside M , and write Q ≺M P . Theorem 2.1 in [P03] shows that this condition is equivalent to the following: there exist projections p ∈ P , q ∈ Q, a unital isomorphism ψ : qQq → pP p (not necessarily onto) and a partial isometry v ∈ M such that vv∗ ∈ (qQq)′ ∩ qM q, v∗v ∈ ψ(qQq)′ ∩ pM p, xv = vψ(x), ∀x ∈ qQq, and x ∈ qQq, xvv∗ = 0, implies x = 0. Justified by this 2nd characterization, one also uses the terminology a corner of Q can be embedded into P inside M (cf. 2.4 in [P03]). By (2.1 in [P03]), the relation Q ≺M P is also equivalent to the fact that the action AdU(Q) has a non-zero part that's "compact relative to P ". This means by definition that the commutant of Q in the semifinite von Neumann algebra 1QhM, P i1Q contains non-zero finite projections or, equivalently, that the action AdU(Q) y L2(1QhM, P i1Q, T r) has non-zero fixed points. By (1.3 in [P01]), Q ≺M P is also equivalent to the fact that Q′ ∩ 1QhM, P i1Q contains non-zero elements from the ideal J (hM, P i) of elements in hM, P i that are "compact relative to P ". We will use the notation Q 6≺M P when the above conditions are not satis- fied, i.e., when IM (Q, P ) = 0. This means that the action Ad-action of U(Q) on L2(1QhM, P i1Q, T r) is ergodic. With the terminology (2.9 in [P05b]), in the case 1Q = 1M this amounts to AdU(Q) y M being weak mixing relative to P . We recall from (2.3 in [P03]) some useful necessary and sufficient criteria for the condition Q 6≺M P to be satisfied. 1.3.1. Theorem. Let M be a finite von Neumann algebra and P, Q ⊂ M be weakly closed ∗-subalgebras. For each q ∈ P(Q), fix Uq ⊂ U(qQq) a subgroup generating qQq as a von Neumann algebra. The following conditions are equivalent: (1) Q 6≺M P (2) There exists a total subset X ⊂ M and a sequence un ∈ U1 such that limn kEP (xuny)k2 = 0, ∀x, y ∈ X. (3) Given any q ∈ P(Q) there exists a sequence of unitary elements un ∈ Uq such that limn kEP (xuny)k2 = 0, ∀x, y ∈ M . Moreover, if P is regular in M , then the above are also equivalent to: (4) There exists a total subset X ⊂ M and a sequence un ∈ U1 such that limn kEP (xun)k2 = 0, ∀x ∈ X. The proof of the above theorem in ([P03]) actually shows the following more general result, involving the intertwining space (cf. [P05a]): 8 SORIN POPA 1.3.2. Theorem. With the same assumptions as in 1.3.1, if X ⊂ L2(1QM 1P ), then the following conditions are equivalent: (1) X ⊥ IM (Q, P ). (2) There exist un ∈ U1 such that limn kEP (ξ∗unξ)k1 = 0, ∀ξ ∈ X. (3) There exist un ∈ U1 such that limn kEP (η∗unξ)k1 = 0, ∀ξ ∈ X, η ∈ L2M . 1.3.3. Remarks. (a) Property 1.3.2(2) above, for characterizing the orthogonal in L2M of the intertwining space IM (Q, P ), can be traced back to ([P81b]), where this type of condition appears in the case of subalgebras Q = L(G1), P = L(G2) of M = L(G), arising from subgroups G1, G2 ⊂ G, as well as for general M and Q = P (as in 1.4 below). (b) The relation Q ≺M P is a "virtual" subordination relation, in the sense that it is "insensitive to finite index perturbations": if Q or P are replaced by subalgebras Q1 ⊂ Q, P1 ⊂ P of finite index (in the sense of one of the definitions in ([PP84]), then we still have Q1 ≺M P1. In particular, if Q has a finite dimensional direct summand, then Q ≺M P for any P ⊂ M , and if there exist projections p ∈ P , p′ ∈ P ′ ∩ M such that pP pp′ = pp′M pp′, then Q ≺M P for any Q ⊂ M . The relation Q ≺M P is also insensitive to localization to "corners" of the algebras involved, i.e., it is sufficient to be satisfied under non-zero projections of Q, P (or of their commutants in M ). (c) Related to (b) above, let us underline here that the notions of finite index (up to taking "corners") for an inclusion of finite von Neumann algebras, P ⊂ M considered in [PP84], generalizing the Jones index in the case of inclusions of factors [J83], translates into the relation M ≺M P . More precisely, this last condition means that there exist projections p ∈ P , p′ ∈ P ′∩M such that pP pp′ = P0 ⊂ M0 = pp′M pp′ has finite index, either in the sense that there exists a finite orthonormal basis of M0 over P0 ([PP84]) or that EP0(x) ≥ cx, ∀x ∈ (M0)+, for some c > 0 (see A.1 in [V07]). In turn, the opposite relation M 6≺M P translates into the fact that P has uniform infinite index in M and it amounts to U(pp′M pp′) containing sequences of elements that are "more and more" perpendicular to pP pp′, for any p ∈ P(P ), p′ ∈ P(P ′ ∩ M ). This type of condition characterizing infinite index can be traced back to (2.2 in [PP84]). (d) Since it is determined by its behavior on corners, the subordination relation ≺M is not transitive in general. For instance, if we take Q = pM p + C(1 − p), P = Cp + C(1 − p) then we have M ≺M Q, Q ≺M P , but M 6≺M P . For this same reason, requiring Q ≺M P and P ≺M Q, does not define a "reasonable" equivalence CONSTRUCTING MASAS 9 relation ∼M between subalgebras of M (e.g., the previous example would show that M ∼ C). However, for MASAs of M , we have the following (cf. A.1 in [P01]): 1.3.4. Theorem. Let M be a finite von Neumann algebra and A, B ⊂ M be MASAs in M . Then A ≺M B if and only if B ≺M A and if and only if there exists a non-zero partial isometry v ∈ M such that vv∗ ∈ B, v∗v ∈ A and vAv∗ = Bvv∗. 1.4. Normalizing subalgebra. If M is a finite von Neumann algebra and B ⊂ M is a von Neumann subalgebra, then we denote NM (B) := {u ∈ U(M ) uBu∗ = B}, the normalizer of B in M . The von Neumann algebra it generates, NM (B)′′, is called the normalizing von Neumann algebra of B in M . A von Neumann subalgebra B is singular in M if any automorphism Ad(u) implemented by some u ∈ NM (B) is inner, i.e., it is of the form Ad(v) for some v ∈ U(B). This is the same as requiring that NM (B) = U(B)U(B′ ∩ M ). If in turn NM (B)′′ = M , then we say that B is regular in M . This terminology has been introduced in [D54], in the case B = A ⊂ M is a maximal abelian ∗-subalgebra (MASA) in M . Note that for a MASA A ⊂ M , being singular means that NM (A)′′ = A, or equivalently NM (A) = U(A), i.e., the normalizer of A in M acts trivially on A. A regular MASA will be called a Cartan subalgebra (or Cartan MASA) in M . We will also consider MASAs A ⊂ M for which NM (A)′′ is a factor (equivalently, NM (A) acts ergodically on A), which will be called semi-regular (cf. [D54]). More generally, recall from ([P97] and 1.4 in [P01]) that if B ⊂ M is a von Neumann subalgebra then qNM (B) denotes the set of all x ∈ M with the property that there exists x1, ..., xn ∈ M such that Bx ⊂ ΣixiB and xB ⊂ ΣiBxi. The space qNM (B) is a ∗-subalgebra and we see that, by (Lemma 1.4.2 in [P01]), one has qNM (B) = IM (B, B) ∩ IM (B, B)∗ ∩ M , with the weak closure being a von Neumann subalgebra of M . Note that if B = A is a MASA in M , then by (1.3 in [P01]) we have qN (A) = spNM (A) = IM (A, A) ∩ M and this space is k k2-dense in IM (A, A). Thus, the normalizing von Neumann algebra of A satisfies NM (A)′′ = qNM (A) and the orthogonal of this space in L2M coincides with IM (A, A)⊥. So Theorem 1.3.2 entails the following criterion for estimating the size of the normalizer of A in M : w 1.4.1. Corollary. Let M be a finite von Neumann algebra, A ⊂ M a MASA and N = NM (A)′′ its normalizing von Neumann algebra. The following conditions are equivalent for an element ξ ∈ L2M : (1) ξ ⊥ N . (2) ∃{un}n ⊂ U(A) such that limn kEA(ξ∗unξ)k1 = 0. 10 SORIN POPA (3) For any n ≥ 2 and any ε > 0, there exists an abelian von Neumann subalgebra A0 ⊂ A generated by n projections of equal trace such that kEA(ξ∗yξ)k1 ≤ ε, ∀y ∈ {A0 ⊖ C1 kyk ≤ 1}. Proof. By applying 1.3.2 to the case Q = P = A, and taking into account that for a MASA A ⊂ M one has IM (A, A)⊥ = NM (A)⊥, it follows that (1) ⇔ (2). Then by taking U1 ⊂ U(A) to be a subgroup satisfying U′′1 = A and τ (u) = 0, u2 = 1, ∀u ∈ U1 \ {1}, we get (2) ⇔ (3). (cid:3) Finally, let us note that a singular MASA A ⊂ M means an embedding of the diffuse abelian von Neumann algebra L∞([0, 1]) ≃ A ⊂ M so that the Ad-action of its unitary group on M is weak mixing relative to A, a Cartan MASA is an embedding so that this action is compact relative to A, while a semi-regular MASA is an embedding having a large relative compact part. 2. Constructing MASAs with control of intertwiners Results in [P81a], [P81d] show that any separable II1 factor M has semi-regular and singular MASAs. The proof consists in constructing an embedding of L∞([0, 1]) ≃ A ⊂ M as an inductive limit of dyadic partitions An ր A that become "more and more extremal in M " (resulting into A being a MASA), while also controlling the normalizer of A, making it become singular (in [P81d]), respectively semi-regular (in [P81a]). For A to become singular, one needs An to "become more and more relative weak mixing". For it to become semi-regular, it is sufficient to build An so that fixed matrix units having An as diagonal are in the normalizer (i.e., in the relative compact part), at each step n. We will show below how one can use much more of the intertwining by bimod- ules criteria within such iterative procedure, allowing us to construct embeddings L∞([0, 1]) ≃ A ⊂ M so that to be weak mixing relative to a given countable family of subalgebras of M . One can in fact even control such relative weak mixingness when M is embedded into larger II1 factors, thus leading to super-rigidity type properties for A. Moreover, we will do the construction so that to also take into account local properties of the ambient factor M , such as existence of central se- quences (in this section), and s-thin approximation (in the next section). 2.1. Theorem. Let N be a separable II1 factor and N ֒→ Mn be embeddings of N into separable II1 factors such that N′ ∩ Mn is of type I, ∀n. Let also Pn ⊂ Mn be von Neumann subalgebras. 1◦ There exists a MASA A ⊂ N such that for each n one has: (a) NMn(A) = U(A ∨ N′ ∩ Mn); CONSTRUCTING MASAS 11 (b) IMn (A, Pn)⊥ = IMn (N, Pn)⊥; (c) M′n ∩ Aω is non-trivial whenever M′n ∩ N ω is non-trivial. In particular, A is singular in N and if N′ ∩ Mn = C1, then A is a singular MASA in Mn which contains non-trivial central sequences of Mn whenever N does. 2◦ There exists a semiregular MASA A ⊂ N such that for each n one has: (a) A′ ∩ Mn = A ∨ N′ ∩ Mn; (b) NMn(A)′′ ⊂ N ∨ N′ ∩ Mn; (c) IMn (A, Pn)⊥ = IMn (N, Pn)⊥. (d) M′n ∩ Aω is non-trivial whenever M′n ∩ N ω is non-trivial. Moreover, if N ≃ R then one can take A ⊂ N such that NMn (A)′′ = N ∨ N′ ∩ Mn. Proof. For each Mn choose a sequence {xn and a sequence {ξn IMn (N, Pn). k }k ⊂ (Mn)1 that's k k2-dense in (Mn)1 k }k ⊂ L2(Mn) ⊖ IMn (N, Pn) that's k k2-dense in L2(Mn) ⊖ Let also {em}m ⊂ {e ∈ P(N ) τ (e) ≤ 1/2} be a k k2-dense sequence. To prove 1◦, we construct recursively a sequence of finite dimensional abelian von Neumann subalgebras Am ⊂ N together with projections fm ∈ P(Am), with τ (fm) = τ (em), and unitary elements vm ∈ U(Amfm), wm, um ∈ U(Am), satisfying the following properties for all 1 ≤ i, j, k ≤ m: (2.1.1) (2.1.2) (2.1.3) (2.1.4) (2.1.5) kfm − emk2 ≤ 13kem − EA′ m−1∩N (em)k2 kEA′ m∩Mk (xk i ∗vmxk j )(1 − fm)k2 ≤ 2−m, kEA′ m∩Mk (xk j ) − EAm∨N ′∩Mk (xk j )k2 ≤ 2−m, kEPk(xk i ∗wmξk j )k2 ≤ 2−m, k[xk i , um]k2 ≤ 2−m, kEAm−1(um)k2 ≤ 2−m. Assume we have constructed (Am, fm, vm, wm, um) satisfying these properties for m = 1, 2, ..., n. By applying Lemma 1.1.1 to B = A′n ∩ N and e = en+1, it follows that there exists fn+1 ∈ A′n ∩ N such that kfn+1 − en+1k2 ≤ 13ken+1 − n∩N (en+1)k2 and τ (fn+1 = τ (en+1). By Corollary 1.2.3, there exists a MASA EA′ B0 ⊂ (1 − fn+1)N (1 − fn+1) satisfying the property 12 SORIN POPA B′0 ∩ (1 − fn+1)Mk(1 − fn+1) = B0 ∨ (N′ ∩ Mk)(1 − fn+1), ∀1 ≤ k ≤ n + 1. Since (Anfn+1)′ ∩ fn+1N fn+1 is type II1 and B0 ∨ (N′ ∩ Mk)(1 − fn+1) are of type I, for each k we have (Anfn+1)′ ∩ fn+1N fn+1 6≺Mk B ∨ (N′ ∩ Mk)(1 − fn+1). Thus, there exists vn+1 ∈ U((Anfn+1)′ ∩ fn+1N fn+1) with the property that (2.1.6) kEB ′ 0∩Mk ((1 − fn+1)xk i ∗vn+1xk j (1 − fn+1))k2 < 2−n−1, 1 ≤ i, j, k ≤ n + 1. Moreover, we may clearly assume vn+1 has finite spectrum. We then take a re- finement A0 n+1fn+1 contains vn+1, while A0 n+1(1−fn+1) "approximates" B0 well enough (in the sense of Lemma 1.2.1) so that, due to (2.1.6) and its strict inequality, we still have n+1 of An in N that contains fn+1, such that A0 (2.1.7) kEA0 n+1 ′∩Mk ((1−fn+1)xk i ∗vn+1xk j (1−fn+1))k2 < 2−n−1, 1 ≤ i, j, k ≤ n+1. On the other hand, by Corollary 1.2.3, there exists a finite dimensional abelian von Neumann subalgebra A1 n+1 in N that contains A0 n+1 and satisfies (2.1.8) kEA1 n+1 ′∩Mk (xk j ) − EA1 n+1∨N ′∩Mk (xk j )k2 ≤ 2−n−1, 1 ≤ i, j, k ≤ n + 1. Now, since A1 n+1′ ∩ N has finite index in N , by Section 1.3 we have IMk (A1 N, Pk) = IMk (N, Pk), and thus ξk Thus, by Thereom 1.3.2, there exists a unitary element wn+1 ∈ (A1 that j ⊥ IMk ((A1 n+1′ ∩ n+1′ ∩ N, Pk), for all 1 ≤ j, k ≤ n + 1. n+1′ ∩ N ) such (2.1.9) kEPk(xk i ∗wn+1ξk j )k2 < 2−n−1, 1 ≤ i, j, k ≤ n + 1. Also, we may clearly assume wn+1 has finite spectrum. We take A2 the finite dimensional abelian von Neumann algebra generated by A1 Due to (2.1.7), (2.1.8) and (2.1.9), A2 n+1 satisfies conditions (2.1.1) − (2.1.4). n+1 ⊂ N to be n+1 and wn+1. Finally, by using the fact that M′n ∩ N ω 6= C implies M′n ∩ N ω diffuse (because Mn is a factor), it follows that for any α > 0 there exists a projection p ∈ N of trace 1/2 such that k[x, p]k2 < α and τ (px) − τ (p)τ (x) < α for all x ∈ {xk i 1 ≤ i, k ≤ n + 1} ∪ (A2 n+1)1. By taking α sufficiently small and using Lemma 1.1.1, n+1′ ∩ N ) sufficiently close to 1 − 2p so it follows that there exists un+1 ∈ U(A2 that we have k[xk i , un+1]k2 ≤ 2−n−1, for all 1 ≤ i, k ≤ n + 1. Thus, if we define n+1 ∨{un+1}′′, then conditions (2.1.1)−(2.1.5) are satisfied for m = n+1. An+1 = A2 CONSTRUCTING MASAS 13 Define A = ∪nAn . Condition (2.1.3) clearly implies A′ ∩ Mk = A ∨ N′ ∩ Mk, ∀k, while condition (2.1.4) implies IMn (A, Pn)⊥ = IMn (N, Pn)⊥, ∀k. w Let u ∈ NMm(A). If Ad(u) acts non-trivially on A, then there exists a non-zero projection e ∈ A of trace ≤ 1/2 such that u∗eu ≤ 1 − e. Let n0 be large enough so that 2−n0 < kek2/60. Since {en}n is k k2-dense in the set of projections of N of trace ≤ 1/2, there exists n ≥ n0 such that ken − ek2 < kek2/60 and such that there exists j, k ≤ n with kxk j − euk2 < kek2/60. With fn, vn ∈ An ⊂ A as given by the construction, (2.1.1) implies kfn − enk2 ≤ 13ken − EA(en)k2 = 13ken − e + EA(e) − EA(en)k2 ≤ 26ken − ek2 ≤ 26kek2/60. Since u∗evnu ∈ A ⊂ A′n ∩ Mm, we would then get the estimates (2.1.10) kefnk2 = kevnk2 = kEA′ n∩Mm(u∗evnu)(1 − e)k2 ≤ kEA′ ≤ kEA′ n∩Mm(u∗vnu)(1 − fn)k2 + kfn − enk2 + ken − ek2 n∩Mm(xk j )(1 − fn)k2 + 29kek2/60 ≤ 30kek2/60, ∗vnxk j where for the very last inequality we used (2.1.1). But since ken − ek2 < kek2/60, we also have kfne − ek2 ≤ kfn − ek2 ≤ kfn − enk2 + ken − ek2 ≤ 27kek2/60, which together with (2.1.10) implies that 30kek2/60 ≥ kefnk2 ≥ kek2 − 27kek2/60 = 33kek2/60, a contradiction. This shows that NMk (A) = U(A′ ∩ Mk) = U(A ∨ N′ ∩ Mk), ∀k, finishing the proof that A satisfies all the conditions in part 1◦ of the theorem. To prove 2◦, note first that by Corollary 1.2.3 there exists a hyperfine II1 sub- factor R ⊂ N such that R′ ∩ Mk = N′ ∩ Mk, ∀k. It is then sufficient to construct a Cartan subalgebra A of R such that A′∩Mk = A∨N′∩Mk, NMk (A)′′ = R∨N′∩Mk, IMn (A, Pn)⊥ = IMn (N, Pn)⊥, ∀k. In other words, it is sufficient to prove the last part of 2◦, where one assumes N ≃ R and want to construct a Cartan MASA A ⊂ N whose normalizing algebra is exactly N ∨ N′ ∩ Mk. To this end, we construct recursively a sequence of commuting dyadic matrix subalgebras Rm ⊂ N (i.e., Rm ≃ M2km×2km (C), for some km ≥ 1), with diagonal subalgebras Dm ⊂ Rm, such that if we denote Nm = ∨m k=1Dk, there k=1Rk, Am = ∨m 14 SORIN POPA exist a projection fm of trace 1/2 in Dm and unitary elements vm ∈ U(Dmfm), wm ∈ U(Dm), so that if we denote yk i ), then the following properties are satisfied for 1 ≤ i, j, k ≤ m: i − EN∨N ′∩Mk (xk i = xk (2.1.11) (2.1.12) (2.1.13) (2.1.14) (2.1.15) kEA′ m∩Mk (yk i ∗vmyk j )(1 − fm)k2 ≤ 1/10; kfm(yk i )(1 − fm)k2 ≥ 2kyk i k2/5; kEA′ m∩Mk (xk j ) − EAm∨N ′∩Mk (xk j )k2 ≤ 2−m; kEPk(xk i ∗wmξk j )k2 ≤ 2−m; kENm(xk i ) − EN (xk i )k2 ≤ 2−m. Assuming we have constructed these objects up to m = n, we construct them for m = n + 1 as follows. Noticing that the finite set F = {yk j 1 ≤ j, k ≤ n + 1} is perpendicular to Nn ∨ N′ ∩ Mk (which is equal to the commutant in Mk of N′n ∩ N ), by Lemma 1.2.1 we can first pick a projection f of trace 1/2 in N′n ∩ N such that f is almost 2-independent to F . In particular, we can choose f so that for all 1 ≤ j, k ≤ n + 1 we have (2.1.16) kf yk j (1 − f )k2 ≥ 2kyk j k2/5. By Corollary 1.2.3, there exists a MASA B0 in (1 − f )(N′n ∩ N )(1 − f ) such that B′0 ∩ (1 − f )Mk(1 − f ) = B0 ∨ (Nn ∨ N′ ∩ Mk)(1 − f ), ∀1 ≤ k ≤ n + 1. Since f (N′n ∩ N )f is type II1 and B0 ∨ (Nn ∨ N′ ∩ Mk)(1 − f ) is type I, the former cannot be intertwined into the latter inside Mk, so by Theorem 1.3.2 there exists a unitary element v ∈ f (N′n ∩ N )f such that for all 1 ≤ i, j, k ≤ n + 1 we have (2.1.17) kEB ′ 0∩(1−f )Mk(1−f )((1 − f )xk i ∗vxk j (1 − f ))k2 < 1/10. Moreover, we may choose v so that to belong to a dyadic finite dimensional 1 ⊂ f (N′n ∩ N )f . Also, by approximating B0 sufficiently well abelian subalgebra B1 CONSTRUCTING MASAS 15 with a dyadic finite dimensional subalgebra B0 1 ≤ i, j, k ≤ n + 1 the estimates 1 ⊂ B0, we will still have for all (2.1.18) kEB0 1 ′∩(1−f )Mk(1−f )((1 − f )xk i ∗vxk j (1 − f ))k2 < 1/10. We now take B1 ⊂ (B1 ∨ Nn)′ ∩ N to be a dyadic finite dimensional abelian j ⊥ IMk (N, Pk) = IMk ((B1 ∨ subalgebra containing B1 Nn)′ ∩ N, Pk), there exists a unitary element w ∈ B′1 ∩ N such that 1 (1 − f ). Since ξk 1f + B0 (2.1.19) kEPk (ξk i ∗wξk j )k2 ≤ 2−n−1, 1 ≤ i, j, k ≤ n + 1, We may clearly also assume w lies in a dyadic finite dimensional abelian subal- gebra B2 ⊂ N′n ∩ N that contains B1. Moreover, by using Corollary 1.2.3 again, we may also assume B2 is so that A0 n+1 = An ∨ B2 satisfies (2.1.20) kEA0 n+1 ′∩Mk (xk j ) − EA0 n+1∨N ′∩Mk (xk j )k2 ≤ 2−n−1. Take now R0 n+1 ⊂ N′n ∩ N to be a (dyadic) finite dimensional factor having B2 as a diagonal algebra. Finally, since N ≃ R, there exists a dyadic finite dimensional factor Rn+1 ⊂ N′n ∩N that contains R0 n+1, such that if we define Nn+1 = Nn ∨Rn+1 then (2.1.21) kENn+1(xk i ) − EN (xk i )k2 ≤ 2−n−1, 1 ≤ i, k ≤ n + 1. Thus, if we take Dn+1 to be a diagonal of Rn+1 that contains B2 and denote An+1 = An ∨ Dn+1, Nn+1 = Nn ∨ Rn+1, fn+1 = f , vn+1 = v, wn+1 = w, then (2.1.16) − (2.1.21) insure that conditions (2.1.11) − (2.1.15) are satisfied for n + 1. . Condition (2.1.15) clearly implies that R = N , while condition (2.1.13) implies A′ ∩ Mk = A ∨ N′ ∩ Mk and (2.1.14) implies IMn (A, Pn)⊥ = IMn (N, Pn)⊥, ∀k. Let now R = ∨kRk = ∪nNn , A = ∨kDk = ∪nAn w w By construction, we have that A is Cartan in N , so that NMk (A)′′ contains N ∨ N′ ∩ Mk. If this inclusion is strict for some k, then by the factoriality of N there must exist an automorphism θ of A implemented by a unitary u ∈ NMk (A) such that θ ◦ Ad(v) acts freely on A, ∀v ∈ NN (A). Thus, EN∨N ′∩Mk (u) = 0. j − uk2 ≤ 1/20. This implies that kyk j − uk2 ≤ 1/20 j − u∗vnuk2 ≤ 1/10, while by (2.1.12) we ∗vnyk Let xk j ∈ Mk be so that kxk and that for each n ≥ k we have kyk j also have kfnu(1 − fn)k2 ≥ kfnyk j (1 − fn)k2 − 1/10 ≥ 2kyk j k2/5 − 1/10 16 SORIN POPA Thus, since u∗vnu ∈ A, by (2.1.11) we get 1/10 ≥ kEA∨N ′∩Mk (yk i ∗vnyk j )(1 − fn)k2 ≥ k(1 − fn)u∗vnu(1 − fn)k2 − 1/10 = kfnu(1 − fn)k2 − 1/10 ≥ 2kyk j k2/5 − 1/5 ≥ 2/5 − 1/20 − 1/5 = 3/20, which is a contradiction. (cid:3) Recall that in [P81d] one proves existence of singular MASAs not only in II1 factors, but also in II∞ and IIIλ factors, for 0 < λ < 1. This result is obtained as a consequence of stronger statement about MASAs in a (separable) II1 factor M , showing that given any group of automorphisms G of the associated II∞ factor M∞ = M ⊗B(ℓ2N) such that G/Int(M∞) is countable, there exists a G-singular MASA A ⊂ M , i.e., a maximal abelian subalgebra of M with the property that if θ ∈ G satisfies θ(a) ⊂ A, for all a in a "corner" Ap of A, then θ acts as the identity on Ap. Let us note here that this type of result is in fact covered by the above general theorem: 2.2. Corollary. Let M be a II1 factor with a sequence of von Neumann subalgebras of uniform infinite index Pn ⊂ M (in the sense of 1.3.3.(c)). Let also G ⊂ Aut(M∞) be a subgroup of automorphisms of M∞ that contains Int(M∞) and is so that G/Int(M ) is countable. Then there exists a G-singular MASA A ⊂ M such that A 6≺M Pn, ∀n. In particular, M contains uncountably many non-intertwinable G-singular MASAs, which in addition can be taken to contain non-trivial central sequences of M whenever M has property Gamma. Proof. Let θn ∈ G be a sequence of automorphisms of M∞ such that G = ∪nθn ◦ Int(M∞). For each n, let sn be so that T r ◦ θn = snT r and denote tn = 1 + sn. Let Mn = M tn and fn ∈ Mn a projection of trace τ (fn) = 1/(1 +sn). Thus, fnMnfn ≃ M and we can embed M into Mn as the subfactor {x ⊕ θn(x) x ∈ M ≃ fnMnfn}. By part 1◦ of Theorem 2.1, there exists a singular MASA A ⊂ M such that A′ ∩ Mn = Afn + A(1 − fn) and NMn(A) = U(Afn + A(1 − fn)). It is immediate to see that this means A is G-singular in M . Moreover, by 2.1.1◦ we can take A so that to also satisfy A 6≺Mn Pn (when Pn is viewed as a subalgebra of Mn). This of course implies A 6≺M Pn as well. To prove the last part, let F be a maximal family of G-singular MASAs of M such that A, B are not intertwinable for any A 6= B in F . Assume F is countable and note that M 6≺Mn A, ∀A ∈ F . By applying Theorem 2.1 to N = M ⊂ Mn (with Mn as above), ∀n, and {Pn}n = F , it follows that there exists a singular MASA C ⊂ M such that C 6≺M A, ∀A ∈ F , contradicting the maximality of F . (cid:3) CONSTRUCTING MASAS 17 2.3. Corollary. Any separable II1 factor contains an uncountable family of mutu- ally non-conjugate semi-regular MASAs, which in addition can be chosen to contain non-trivial central sequences if the ambient factor has property Gamma. Proof. The same argument as in the above proof applies, using 2.1.2◦ instead of 2.1.1◦. (cid:3) It has been shown in (2.5 of [P81b]) that if P ⊂ M is a von Neumann subalgebra and u ∈ U(M ) satisfies the property that for all n ≥ 1 and all ε > 0 there exists a subalgebra A0 ≃ L(Z/nZ) ⊂ Q with u∗A0u ⊥ε P , then u ⊥ NM (P ). Along these lines, one can now deduce the following stronger result (generalizing 1.4.1 as well): 2.4. Corollary. Let M be a finite von Neumann algebra. Let Q, P ⊂ M be diffuse von Neumann subalgebras and ξ ∈ L2M . The following conditions are equivalent: 1◦ ξ ⊥ IM (Q, P ). 2◦ For any n ≥ 1 and any ε > 0 there exists u ∈ U(Q) such that un = 1, τ (uk) = 0, 1 ≤ k < n and kEP (ξ∗ukξ)k1 ≤ ε, 1 ≤ k < n. 3◦ For any n ≥ 1 and any ε > 0, there exists an n-dimensional abelian von Neumann subalgebra A0 ⊂ Q such that τ (p) = 1/n for any minimal projection in A0 and kEP (ξ∗xξ)k1 ≤ ε for all x ∈ (A0)1 with τ (x) = 0. Proof. We clearly have 3◦ ⇔ 2◦ and by Theorem 1.3.2 we have 2◦ ⇒ 1◦. To prove 1◦ ⇒ 3◦, let ξ ⊥ IM (Q, P )⊥. Apply first Theorem 2.1 to get a MASA A ⊂ Q with the property that IM (A, P )⊥ = IM (Q, P )⊥. Representing A as the von Neumann algebra of the countable group Z/2Z⊕∞, we get a unitary group U1 ⊂ U(A) such that U′′1 = A, u2 = 1, τ (u) = 0, ∀u ∈ U1\{1}. Applying Theorem 1.3.2 to ξ ⊥ IM (A, P ), we get a sequence {um}m ⊂ U1 such that limm kEP (ξ∗umξ)k1 = 0. Taking α > 0 appropriately small and A0 to be an n-dimensional subalgebra of A generated by projections of trace 1/n that's α-contained in {um m0 ≤ m ≤ m1}′′, for some m0 ≤ m1 sufficiently large so that kEP (ξ∗umξ)k1 < α, ∀m ≥ m0, one gets condition 3◦ satisfied. (cid:3) 3. Thin factors and MASAs with bounded multiplicity In the 1950s, W. Ambrose and I.M. Singer have considered MASAs in II1 factors A ⊂ M with the property that the von Neumann algebra A ∨ JAJ ⊂ B(L2M ), generated by the left-right multiplication on L2M by elements in A, is maximal abelian in B(L2M ). Noticing that this is equivalent to A ∨ JAJ having a cyclic vector (i.e., ∃ξ ∈ L2M with [AξA] = L2M ), this property is analogue to an inclusion of groups H ⊂ G with just one (non-trivial) double co-set over H. (Note however that the algebra framework makes it so that regular MASAs do satisfy this property (cf. [FM77]), while for a normal subgroup H ⊂ G, H \G/H = G/H is always large.) 18 SORIN POPA In [Pu61] L. Pukanszky took this idea further, by noticing that the type of the algebra (A ∨ JAJ)′ ⊂ B(L2M ) is an invariant for the isomorphism class of a MASA inclusion A ⊂ M , and that if an inclusion of groups H ⊂ G with G ICC and H abelian is so that H \ G/H − H has n identical classes, then the MASA inclusion A = L(H) ⊂ L(G) = M has the property that A ∨ JAJ has multiplicity n on L2M ⊖ L2A. In other words, the commutant algebra (A ∨ JAJ)′, which is always equal to AeA ≃ A on the reducing space L2A, is homogeneous of type In on L2(M ⊖A). Taking appropriate examples H ⊂ G with G locally finite ICC (inspired by a construction in [D54]), he was able to give examples of singular MASAs in the hyperfinite II1 factor R that have "multiplicity n", for each 1 ≤ n < ∞, and are thus mutually non-conjugate by automorphisms of R. The type of the von Neumann algebra (A ∨ JAJ)′(1 − eA) ⊂ B(L2(M ⊖ A)) (i.e., the list of multiplicities appearing in its decomposition as a direct sum of homogeneous type Ini algebras, 1 ≤ ni ≤ ∞) is what one generically calls the Pukanszky invariant of A ⊂ M . Of this, we will retain here only the supremum over all the multiplicities 1 ≤ ni ≤ ∞ in the decomposition (A ∨ JAJ)′ = ⊕iL∞(Xi) ⊗ Mni×ni (C). 3.1. Definitions. 1◦ Let M be a II1 factor and A ⊂ M an abelian von Neumann subalgebra. We denote by m(A ⊂ M ) the supremum over all 1 ≤ m ≤ ∞ with the property that (A ∨ JAJ)′ has a type Im direct summand, and call it the multiplicity of A ⊂ M . Notice that m(A ⊂ M ) = 1 if and only if A ∨ JAJ is maximal abelian in B(L2M ), and that this implies A is maximal abelian in M (see 3.3 below). An abelian von Neumann subalgebra A in M with the property that A ∨ JAJ is maximal abelian in B(L2M ) is called an s-MASA of M . 2◦ If M is a II1 factor then we denote ma(M ) = min{m(A ⊂ M ) A a MASA in M }. Thus, ma(M ) = 1 if and only if M has an s-MASA. A II1 factor with this property is called an s-thin factor. Let us note right away that the multiplicity of MASAs behaves well to taking tensor products and intermediate subfactors: if Ai ⊂ Mi, i = 1, 2, are inclusions of MASAs, then m(A1⊗A2 ⊂ M1⊗M2) = m(A1 ⊂ M1)m(A2 ⊂ M2); also, if B ⊂ Q ⊂ P is a MASA in P with Q an intermediate factor, then m(B ⊂ Q) ≤ m(B ⊂ P ). 3.2. Examples. By (2.9 in [FM77]), any Cartan MASA in a II1 factor is an s-MASA. But by [Pu61], there do exist singular s-MASAs as well. For instance, when M ≃ R is the hyperfine II1 factor, then besides its (unique by [CFW81]) Cartan subalgebra D, R contains an s-MASA A that is singular. More precisely, the following example of a singular MASA A ⊂ R from ([D54]) has been shown in ([Pu61]) to be an s-MASA: represent R as the group factor associated with the amenable ICC group G of affine transformation on Q, with its abelian subgroups CONSTRUCTING MASAS 19 T = Q (translations), H = Q∗ (homotheties); since G is ICC relative to both T and H, D = L(T ), A = L(H) are MASAs in R = L(G); since T is normal in G, D is a Cartan subalgebra in R, while since H acts transitively on T \ {0}, A follows singular in R, with the vector ξ = ξ0 + ξ1 ∈ ℓ2(G) = L2R cyclic for A ∨ JAJ, where ξ0 is the vector corresponding to the trivial element in T ⊂ G (so translation by 0) and ξ1 is the element in T ⊂ G corresponding to translation by 1. 3.3. Remark. As we mentioned in 3.1.1◦ above, if A is a von Neumann subalgebra of M , then the condition "A ∨ JM AJM maximal abelian in B(L2M )" implies that A is a MASA in M . To see this, note first that this condition implies B = A′ ∩ M abelian. Indeed, since eB ∈ (A ∨ JM AJM )′ = A ∨ JM AJM , it follows that eB(A ∨ JM AJM )eB is maximal abelian in B(L2B). Since by the commutativity of B we have A = JBAJB, this forces A = B. The terminology "s-MASA in M " can thus be viewed as emphasizing a strength- ening of the property of being a MASA in M . The prefix "s" can also be viewed as hinting to the terminology simple MASA, which has been sometimes used for abelian von Neumann subalgebras satisfying this property (N.B.: this has been the original Ambrose-Singer terminology, carried on in [K67], [JP82], [Ge97]). The usage of the adjective "simple" for a MASA can however be misleading, as (non- trivial) abelian von Neumann algebras do have non-trivial ideals and are thus not simple as rings... The terminology "simple MASA" may trigger additional confu- sion as it has also been used by Takesaki in [T63], but for a different class of MASAs A ⊂ M , via a characterization which has in fact been later shown equivalent to A being singular (cf. [H79]). From this point on, if S is a non-empty subset of a Hilbert space H, we will use the notation [S] for the Hilbert subspace generated by S, i.e., [S] = sp(S), and also for the orthogonal projection of H onto this space, the difference being always clear from the context. Also, if ξ ∈ H and H0 ⊂ H is a vector subspace, then the notation ξ ∈δ H0, for some δ > 0, means that there exists η ∈ H0 such that kξ − ηk2 < δ. While if X ⊂ H, then X ⊂δ H0 stands for ξ ∈δ H0, ∀ξ ∈ X. The following result provides alternative characterizations of MASA-multiplicity. 3.4. Proposition. Let M be a separable II1 factor, A ⊂ M an abelian von Neu- mann subalgebra and n0 ≥ 1 an integer. The following conditions are equivalent: 1◦ Any type Im direct summand of (A ∨ JAJ)′ satisfies m ≤ n0. 2◦ There exists X ⊂ L2M with X ≤ n0 such that [AXA] = L2M . 3◦ ∀F ⊂ M finite, ∀δ > 0, there exists X ⊂ L2M such that X ≤ n0 and F ⊂δ spAXA. Moreover, if n0 = 1, then in 2◦ above one can take X = {b} with b = b∗ ∈ M . 20 SORIN POPA If in addition A is a MASA in M , then 1◦ − 3◦ are also equivalent to: 4◦ The representation Ad(U(A)) y L2(M ⊖A)) admits a cyclic set of n0 vectors. Before proving this result, let us notice the following: 3.5. Lemma. Let A ⊂ B(H) be an abelian von Neumann algebra acting on the Hilbert space H. 1◦ The supremum over all n ≤ ∞ such that A′ has a In direct summand is equal to the minimum over all m ≤ ∞ for which there exists X ⊂ H with X = m and [AX] = H. 2◦ For any η1, η2 ∈ H, the set L = {t ∈ C \ {0} [A′(η1 + tη2)] 6= [A′(η1)] ∨ [A′(η2)]} is at most countable. Proof. 1◦ This part of the statement is the case "B abelian" of the Murray-von Neumann coupling constant theorem (see [vN43]) relating a finite von Neumann algebra B ⊂ B(H) with its commutant B′ ⊂ B(H), by the "factor of multiplicity" dimBH. 2◦ This part is just (Lemma 3.5 in [P82]). (cid:3) Proof of 3.4. Denote A = A ∨ JAJ and B = (A ∨ JAJ)′ ∩ B(L2M ). By 3.5.1◦ in the above Lemma, we have 1◦ ⇔ 2◦, and we clearly have 2◦ ⇒ 3◦. Condition 3◦ shows that there exists a sequence of subsets Xn ⊂ L2M of car- dinality at most n0 such that if we denote pn = [AXnA] ∈ B then pn → 1 in the so-topology. By Lemma 3.5.1◦, each pnBpn is of finite type with all homoge- neous type Im summands satisfying m ≤ n0. By ([PP84], or 1.2 in [P94]), this is equivalent to having the (probabilistic) Pimsner-Popa index of Apn ⊂ pnBpn at most equal to n0. Since the definition of this index behaves well to limits (see e.g., [PP84]), it follows that the index of A ⊂ B is ≤ n0 as well, which in turn means that any type Im direct summand of B must satisfy m ≤ n0. Thus, 3◦ ⇒ 1◦. Let us now prove that if there exists ξ ∈ L2M such that [AξA] = L2M , then there exists b = b∗ ∈ M such that [AbA] = L2M . Let us first notice that there exists η0 = η∗0 ∈ L2M such that [Aη0A] = [AξA] = L2M . Indeed, this follows by noticing that A = A ∨ JAJ satisfies A′ = A and so one can apply Lemma 3.5.2◦ to η1 = ℜξ, η2 = ℑξ, A ⊂ B(L2M ), to get some t ∈ R such that η0 = η1 + tη2 satisfies [A(η0)] = [A(η1] ∨ [A(η2)] = [A(ξ)]. Let us also note that if for some η ∈ L2M and F ⊂ L2M a finite set we have F ⊂δ spAηA, then there exists c = c(δ; η, F ) ≤ 1 such that if η′ ∈ L2M satisfies kη′ − ηk2 < c, then we still have F ⊂δ spAη′A. Denote yn = e[−n,−n+1)(η0)η0 + e[n−1,n)(η0)η0, n ≥ 1, and note that yn = y∗n ∈ k=1yk M are mutually orthogonal in L2M with kynk ≤ 2n, and that η0 = limm Σm CONSTRUCTING MASAS 21 in L2M . By letting b1 = y1 and then applying 3.5.2◦ and the above observation for m ≥ 2, we obtain recursively some scalars tm ≤ 2−mc(2−m−1, {y1, ..., ym}, bm−1)/2m such that bm = bm−1 + tmym satisfy {y1, ..., ym} ⊂ [AbmA]. Thus, if we denote b = limm bm ∈ (Mh)1 (note that this limit exists in operator norm, because kbm − bm−1k ≤ 2−m), then kb − bm−1k2 ≤ kb − bm−1k ≤ c(2−m−1, {y1, ..., ym}, bm−1), and thus {y1, ..., ym} ⊂2−m [AbA]. Thus, all yi belong to [AbA], and hence so does η0 = Σiyi, implying that [AbA] ⊃ [Aη0A] = L2M . To prove 4◦ ⇔ 2◦ under the condition that A is a MASA in M , we need to show that that there exists a set X ⊂ L2(M ⊖ A) with X ≤ n0, such that sp{uξu∗ u ∈ U(A), ξ ∈ X} is dense in L2(M ⊖ A). But by the "linearization principle" (5.1 in [P89]), since A′ ∩ M = A, a set X ⊂ L2(M ⊖ A) is cyclic for Ad(U(A)) y L2(M ⊖ A)) if and only [AXA] = L2(M ⊖ A). (cid:3) 3.6. Theorem. Let M be a separable II1 factor and n0 ≥ 1 an integer. The following conditions are equivalent: 1◦ ma(M ) ≤ n0. 2◦ For any finite dimensional abelian von Neumann subalgebra A0 ⊂ M , any finite subset F ⊂ M and any δ > 0, there exists a finite dimensional abelian von Neumann algebra A1 ⊂ M containing A0 and X1 ⊂ L2M with X1 ≤ n0 such that F ⊂δ spA1X1A1. 3◦ There exists a sequence of positive numbers tn ց 0 such that each II1 factor N = M tn satisfies the the following property: (3.6.3) For any F ⊂ N finite and δ > 0, there exist A1 ⊂ N finite dimensional abelian and X1 ⊂ L2N , with X1 ≤ n0, such that F ⊂δ spA1X1A1. Proof. We clearly have 1◦ ⇒ 2◦. By applying property 2◦ to A0 = Ce + C(1 − e), for projections e in M of trace τ (e) = tn, we see that 2◦ ⇒ 3◦. To prove 2◦ ⇒ 1◦, let {xn}n ⊂ (M )1 be a sequence of elements that's k k2-dense in (M )1. We first construct recursively an increasing sequence of finite dimensional abelian von Neumann sbalgebras Am ⊂ M together with a sequence of subsets Xm ⊂ L2M , with Xm ≤ n0, such that {x1, ..., xm} ⊂2−m spAmXmAm. Assume (Am, Xm) have been constructed for 1 ≤ m ≤ n. Apply 2◦ to F = {x1, ..., xn+1}, B0 = An and δ = 2−n−1 to get a larger finite dimensional algebra B1 ⊃ B0, with a subset Xn+1 ⊂ L2M having at most n0 elements, such that if we let An+1 = B, then {x1, ..., xn+1} ⊂2−n−1 spAn+1Xn+1An+1. Let now A = ∪nAn . By construction, it follows that ∀F ⊂ M finite and ε > 0, there exists X ⊂ L2M , with X ≤ n0, such that F ⊂ε spAXA. But then 3.4.3◦ w 22 SORIN POPA above implies that m(A ⊂ M ) ≤ n0. Let us finally prove that 3◦ ⇒ 2◦. Let F ⊂ (M )1 be a finite set, B0 ⊂ M a finite dimensional abelian von Neumann subalgebra and δ > 0. We need to prove that there exists a larger finite dimensional abelian algebra B1 ⊃ B0 with a set X1 ∈ L2M of at most n0 vectors such that F ⊂δ spB1X1B1. It is clearly sufficient to prove this for B0 generated by minimal projections {qi 0 ≤ i ≤ m} with τ (qi) = tn, 1 ≤ i ≤ m and τ (q0) < tn, where n is sufficiently large so that tn < δ/100. Let v1, ..., vm be partial isometries in M such that v∗i vi = q1 and viv∗i = qi, 1 ≤ i ≤ m. Denote F ′ = {v∗i yvj y ∈ F, 1 ≤ i, j ≤ n} ⊂ q1M q1. By property 3◦, there exists a finite dimensional abelian von Neumann algebra A1 ⊂ q1M q1 and a subset X1 ∈ q1L2(M )q1 of at most n0 elements, such that F ′ ⊂δ/√m spA1X1A1. For each ξ1 ∈ X1, denote η1 = Σm i,j=1viξ1v∗j . Let Y1 be the set of all such η1 and denote B1 = Cq0 + Σm i=1viA1v∗i . Fix y ∈ F . Since v∗i yvj ∈ F ′, there exist aij ∈ spA1X1A1 such that kv∗i yvj − aijk2 ≤ δ2/m. 2,q1M q1 Note that, by the definitions of B1 and η1, the element b = Σm i,j=1viaijv∗j belongs to spB1Y1B1. Moreover, by using Pythagoras theorem in M , we have ky − bk2 2 = ky − Σm i,j=1viaijv∗j k2 2 = Σm i,j=1kv∗i yvj − aijk2 2 = Σm i,j=1kv∗i yvj − aijk2 2,q1M q1 τ (q1) ≤ m2τ (q1)δ2/m = δ2τ (1 − q0). This shows that F ⊂δ spB1Y1B1, thus finishing the proof. (cid:3) Motivated by property (3.6.3) above, we will also consider the following: 3.7. Definition. Let M be a II1 factor. We denote by wma(M ) the minimum over all cardinalities 1 ≤ m ≤ ∞ with the property that given any finite set F ⊂ M and any ε > 0, there exist a subset X1 ⊂ L2M with X1 ≤ m and a finite dimensional abelian ∗-subalgebra B1 ⊂ M such that F ⊂ε [B1X1B1]. The II1 factor M is weak s-thin if wma(M ) = 1, i.e., if for any finite set F ⊂ M and any ε > 0, there exists a finite dimensional abelian von Neumann subalgebra B1 ⊂ M and a vector η1 ⊂ L2M such that F ⊂ε spB1η1B1. Note that ma(M ), wma(M ) are both isomorphism invariants for M , that mea- sure the "thinness" of M relative to its abelian subalgebras and satisfy ma(M ) ≥ wma(M ). Notice also that the invariant ma(M ) is very much in the spirit of what was called n-weak thinness in [GP99], which denoted the minimal cardinality 1 ≤ n ≤ ∞ with the property that there exist hyperfinite von Neumann subalgebras R0, R1 ⊂ M and a set X ⊂ L2M with X ≤ n such that [R0XR1] = L2M . More precisely, the minimal such n obviously satisfies n ≤ ma(M ). CONSTRUCTING MASAS 23 3.8. Corollary. 1◦ We have ma(M ) = ma(M t), for any t > 0. In particular, if M is an s-thin factor, then its amplifications M t are s-thin factors, ∀t > 0. Also, if M is weak s-thin, then Mn×n(M ) is weak s-thin, ∀n ≥ 1. 2◦ If a II1 factor M has non-trivial fundamental group (e.g., if M is a McDuff factor), then wma(M ) = ma(M ). In particular, such a II1 factor is s-thin iff it is weak s-thin. 3◦ If a II1 factor M is generated by an increasing sequence of subfactors Mn ⊂ M , then ma(M ) ≤ lim supn ma(Mn) and wma(M ) ≤ lim supn wma(Mn). In par- ticular, if all Mn are s-thin (resp. weak s-thin), then M is s-thin (resp, weak s-thin). Proof. To prove 1◦, note first that if we assume ma(M ) ≤ n0, then by 1◦ ⇒ 3◦ in Proposition 3.6 there exists tn ց 0 such that wma(M tn) ≤ n0. But then wma((M t)tn/t) ≤ n0, with tn/t ց 0, and thus by applying 3◦ ⇒ 1◦ in 3.6, it follows that ma(M t) ≤ n0. Part 2◦ is immediate from the equivalence 1◦ ⇔ 3◦ in Proposition 3.6. Part 3◦ can be easily deduced from the characterizations in Proposition 3.6. To see this, assume first that ma(Mn) ≤ n0, ∀n. We will directly construct from this a MASA A in M such that ma(M ) ≤ n0. Let {xk}k be a sequence of elements in the unit ball of ∪nMn which is k k2- dense in (M )1. Assume we have constructed finite dimensional abelian von Neu- mann subalgebras A1 ⊂ A2... ⊂ Am in the ∗-algebra ∪nMn together with sub- sets X1, ..., Xm ⊂ ∪nMn, with Xi ≤ n0, such that {x1, ..., xk} ⊂2−k spAkXkAk, 1 ≤ k ≤ m. Let K ≥ 1 be large enough such that {x1, ..., xm+1} ⊂ MK and Am ⊂ MK. By applying 3.6.2◦ to B0 = Am, F = {x1, ..., xm+1} and δ = 2−m−1, as well as the fact that ma(MK) ≤ n0, we get a larger finite dimensional abelian von Neumann subalgebra Am+1 ⊃ Am in MK (which can thus be viewed as a sub- algebra in M ⊃ MK), with a set of ≤ n0 vectors Xm+1 ⊂ L2(MK) ⊂ L2M with at most n0 elements, such that {x1, ..., xm+1} ⊂2−m−1 spAm+1Xm+1Am+1. w Letting now A = ∪mAm , it is trivial to see that A ⊂ M satisfies condition 5◦ in Proposition 3.4, and thus satisfies m(A ⊂ M ) ≤ n0, implying that ma(M ) ≤ n0. Assume now that wma(Mn) ≤ n0, ∀n. To prove that wma(M ) ≤ n0, it is clearly sufficient to show that 3.7 holds true for any finite subset F in a prescribed k k2-dense subset X ⊂ (M )1. But then the inequality is trivial for F ⊂ (∪nMn)1. (cid:3) 4. Singular and semiregular s-MASAs in s-thin factors We prove in this section that if M has an s-MASA (i.e. M is s-thin), then it has singular and semi-regular s-MASAs (in fact, many of them). In other words, if M 24 SORIN POPA admits cyclic MASA actions of L∞([0, 1]), then it has cyclic MASA actions that are relative weak mixing, respectively have a "large" relative compact part. 4.1. Theorem. Let N be a separable s-thin factor and N ֒→ Mn be embeddings of N into separable II1 factors such that N′ ∩ Mn is of type I, ∀n. For each n, let Pn ⊂ Mn be a von Neumann subalgebra such that N 6≺Mn Pn. 1◦ There exists a singular s-MASA A ⊂ N such that NMn (A)′′ = A ∨ N′ ∩ Mn and A 6≺Mn Pn, ∀n. 2◦ There exists a semiregular s-MASA A ⊂ N such that NMn (A)′′ ⊂ N ∨N′∩Mn and A 6≺Mn Pn, ∀n. Moreover, in both 1◦ and 2◦, if M′n ∩ N ω 6= C then one can choose A so that M′n ∩ Aω 6= C as well. Proof. We proceed exactly as in the proof of Theorem 2.1, constructing A iteratively, but with an additional "local requirement" which will insure that in the end, besides being singular (resp. semiregular) in N , A is an s-MASA in N , with its Ad-action on Mn being weak mixing relative to Pn, ∀n. Thus, we take {em}m ⊂ {e ∈ P(N ) τ (e) ≤ 1/2} to be a k k2-dense sequence. Also, we let {xk}k ⊂ (N )1 be k k2-dense in (N )1 and for each Mn we choose a sequence {xn k }k ⊂ (Mn)1 that's k k2-dense in (Mn)1. To prove Part 1◦, we construct recursively an increasing sequence of finite di- mensional abelian von Neumann subalgebras Am ⊂ N together with projections fm ∈ P(Am) and unitary elements vm ∈ U(Amfm), wm ∈ U(Am), as well as a vector ξm ∈ L2N , such that for each 1 ≤ i, j, k ≤ m we have: (4.1.1) (4.1.2) (4.1.3) (4.1.4) (4.1.5) kfm − emk2 ≤ 13kem − EA′ m−1∩N (em)k2 kEA′ m∩Mk (xk i ∗vmxk j )(1 − fm)k2 ≤ 2−m, 1 ≤ i, j, k ≤ m kEA′ m∩Mk (xk j ) − EAm∨N ′∩Mk (xk j )k2 ≤ 2−m, kEPk (xk i ∗wmxk j )k2 ≤ 2−m, 1 ≤ i, j, k ≤ m. {x1, ..., xm} ⊂2−m spAmξmAm. Assume we have constructed (Am, fm, vm, wm, ξm) satisfying these properties for m = 1, 2, ..., n. By the proof of Theorem 2.1, we can first construct a finite CONSTRUCTING MASAS 25 n+1 and unitary elements vn+1 ∈ A1 dimensional abelian algebra A1 A1 (4.1.1) − (4.1.4) are satisfied for m = n + 1 and with A1 An+1. Finally, since A1 a refinement An+1 ⊂ N of A1 (4.1.5) is satisfied for m = n + 1. Note that since An+1 contains A1 (4.1.1) − (4.1.4) will be satisfied for m = n + 1. n+1 ⊂ M that contains An, with a projection fn+1 ∈ n+1, such that conditions n+1 playing the role of n+1 is contained in the s-thin factor N , by 3.6.2◦ there exists n+1 and a vector ξn+1 ∈ L2N such that condition n+1, conditions n+1fn+1, wn+1 ∈ A1 If we now denote A = ∪nAn , then the same argument as in the proof of Theorem 2.1 shows that due to conditions (4.1.1) − (4.1.4) we have A′ ∩ Mk = A ∨ N′ ∩ Mk, A 6≺Mk Pk and NMk (A) = U(A ∨ N′ ∩ Mk), while condition (4.1.5) implies A ⊂ N satisfies condition 3.4.4◦ with n0 = 1 and is thus an s-MASA in N . In turn, to prove part 2◦ we construct recursively a sequence of commuting dyadic matrix subalgebras Rm ⊂ N (i.e., Rm ≃ M2km×2km (C), for some km ≥ 1), with diagonal subalgebras Dm ⊂ Rm, such that if we denote Nm = ∨m k=1Rk, Am = ∨m k=1Dk, there exist a projection fm of trace 1/2 in Dm, unitary elements vm ∈ U(Dmfm), wm ∈ U(Dm), and a vector ξm ∈ L2N , such that if we denote i = xk yk i ), then the following properties are satified for 1 ≤ i, j, k ≤ m: i − EN∨N ′∩Mk (xk w (4.1.6) (4.1.7) (4.1.8) (4.1.9) kEA′ m∩Mk (yk i ∗vmyk j )(1 − fm)k2 ≤ 1/10; kfnyk i (1 − fm)k2 ≥ 2kyk i k2/5; kEA′ m∩Mk (xk j ) − EAm∨N ′∩Mk (xk j )k2 ≤ 2−m; kEPk (xk i ∗wmxk j )k2 ≤ 2−m; (4.1.10) {x1, ..., xm} ⊂2−m spAmξmAm. Assuming we have constructed these objects up to m = n, we construct them for m = n + 1 as follows. From the proof of 2.1.2◦, we can first construct a dyadic matrix algebra R1 n+1 that commutes with Nn, together with a diagonal subalgebra D1 n+1 of trace 1/2 and unitaries vn+1 ∈ D1 n+1, N 1 n+1, then conditions (4.1.6) − (4.1.9) are satisfied for m = n + 1, with A1 n+1 in the role of An+1 ⊂ Nn+1. n+1, such that if we denote A1 n+1 = Nn ∨ R1 n+1 ⊂ N 1 n+1, a projection fn+1 ∈ D1 n+1fn+1, wn+1 ∈ D1 n+1 = An ∨ D1 n+1 ⊂ R1 26 SORIN POPA Let {eij}i,j∈J be matrix units for N 1 n+1, with eii generating An, and denote F = {e1ixkej1 1 ≤ k ≤ n + 1, i, j ∈ J}. Since N0 = (N 1 n+1)′ ∩ N is s-thin (as an amplification of N ), it follows that there exists an abelian finite dimensional von Neumann subalgebra B0 ⊂ N0 and a vector η0 ∈ L2(N0), such that F ⊂α spB0η0B0, where α = 2−n−1/J. Moreover, we may clearly also assume that B0 is dyadic. If we now denote B1 = B0 ∨ An ∈ N and η1 = Σi,jei1η0e1j ∈ L2N , then Pythagoras Theorem implies that {x1, ..., xn+1} ⊂2−n−1 spB1η1B1. Finally, we take a dyadic matrix algebra R0 n+1 ⊂ N0 having B0 as a diagonal subalgebra and denote Rn+1 = R1 n+1 ∨ B0, Nn+1 = Nn ∨ Rn+1, An+1 = B1 = An ∨ Dn+1, ξn+1 = η1. It is then immediate to see that all conditions (4.1.6) − (4.1.10) are satisfied for m = n + 1. n+1, Dn+1 = D1 n+1 ∨ R0 Let A = ∪nAn = ∨nDn, R = ∪nNn = ∨nRn. Like in the proof of 2.1.2◦, condition (4.1.8) insures that A′ ∩ Mk = A ∨ N′ ∩ Mk while (4.1.9) implies that A 6≺Mk Pk, ∀k. Also, by the definitions of A and R we see that NMk (A)′′ ⊃ R ∨ N′ ∩ Mk, and thus A is semiregular in N . On the other hand, condition (4.1.10) shows that NMk (A)′′ ⊂ N ∨ N′ ∩ Mk. Finally, condition (4.1.10) combined with the case n0 = 1 of 3.4.4◦ show that A is s-MASA in N . Finally, let us note that in the proof of the existence of singular s-MASAs in 1◦ (resp. of semi-regular s-MASAs in 2◦), if we assume M′n ∩ N ω 6= C, then exactly as in the proof of Theorem 2.1.1◦ (respectively 2.1.2◦), one can complement the list of conditions in the recursive construction with a condition insuring that A contains non-trivial central sequences of Mn. (cid:3) w w 4.2. Corollary. Let M be a separable s-thin II1 factor. Then M has uncount- ably many mutually non-intertwinable singular (respectively semiregular) s-MASAs. Moreover, if M has the property Gamma, then all these MASAs can be taken to contain non-trivial central sequences of M . Proof. The argument in the proofs of Corollaries 2.2 and 2.3 works exactly the same way, by using Theorem 4.1 in lieu of Theorem 2.1. (cid:3) 5. Final remarks and open problems 5.1. Absence of Cartan MASAs versus s-MASAs. The first examples of (separable) II1 factors without Cartan subalgebras were obtained by Voiculescu in [Vo96], who used free probability methods to prove that the free group factors L(Fn) do not have Cartan MASAs. It was then realized that a suitable adaptation of the argument in [Vo96] shows that L(Fn) doesn't have s-MASAs ([Ge98]), nor even MASAs with finite multiplicity ([GP99]), in fact ma(L(Fn)) = ∞. Similar arguments can be used to show that any II1 factor of the form M = N1 ∗ N2, with N1, N2 finitely generated diffuse von Neumann subalgebras of Rω, satisfies CONSTRUCTING MASAS 27 ma(M ) = ∞. Indeed1, this follows by combining (4.1 in [GP99]) with the lower estimates on free entropy dimension in [Ju01] and the additivity of Voiculescu's free entropy dimension ([Vo96]). On the other hand, during the last ten years, a large number of results about absence of Cartan MASAs have been obtained through deformation rigidity theory ([OP07], [CS11], [CSU11], [PV11], [PV12], [I12]). For instance, it was shown in [PV11] that L(Fn)⊗N has no Cartan subalgebras for any finite factor N . In many "absence of Cartan MASA" results that are obtained through deformation-rigidity theory, one actually obtains classes of II1 factors M that are strongly solid in the sense of [OP07], i.e., the normalizing algebra of any diffuse amenable B ⊂ M (in particular of any MASA A ⊂ M ) is amenable. This is notably the case for the free group factors M = L(Fn) ([OP07]) and more generally for all factors L(Γ) arising from non-elementary hyperbolic groups Γ ([CS11]). Notice that no result about automatic amenability of normalizing algebras of MASAs could be obtained using free probability, while absence of s-MASAs could not be shown by using deformation rigidity theory! Finally, let us point out that absence of Cartan MASAs in a II1 factor M amounts to having no relative compact actions by MASAs on M , while strong solidity means the relative compact part of any such action is amenable. Also, absence of s-MASAs in M means there are no cyclic actions by MASAs on M . 5.1.1. Problem. It would be interesting to find new proofs of absence of s-MASAs in certain factors. This is particularly the case for the II1 factor L(Fn), where a direct, "elementary" proof seems possible. 5.1.2. Problem. We have no examples of II1 factors with s-MASAs but without Cartan subalgebras. One class of factors that may provide such examples are the crossed product factors of the form M = R ⋊ Γ, with Γ = Γ1 × Γ2 where Γ1, Γ2 are groups in one of the classes in [PV11], [PV12], for which one knows that any regular MASA of M is necessarily contained in R (after conjugacy by a unitary). Thus, if the Γ-action on R "mixes well" the Cartan MASAs of R, then one should be able to show that R cannot contain regular MASAs of M . 5.1.3. Problem. Is the weak s-thin property equivalent to s-thin ? More generally, do we always have wma(M ) = ma(M )? We saw that once a factor M has non- trivial fundamental group, then the two properties are equivalent, but it is not clear wether this is the case for any II1 factor. 5.1.4. Problem. Another question we leave open is whether ma(M ) < ∞ implies ma(M ) = 1 and whether there are permanence properties relating ma(M ) with 1I am grateful to Dima Shlyakhtenko for pointing out to me this line of arguments. 28 SORIN POPA the multiplicity invariant ma(N ) of its subfactors of finite Jones index N ⊂ M . In particular, whether M is s-thin if and only if N is s-thin. 5.2. Local characterization of factors with Cartan MASAs. As we have seen above, existence of Cartan MASAs is a property of II1 factors, that many II1 factors, such as L(Fn), do not have. Factors with Cartan MASAs are precisely the ones that admit relative compact actions of L∞([0, 1]). Let us call CF the class of such II1 factors. If A is a MASA in a II1 factor M , then there are ways to characterize the regularity property NM (A)′′ = M which does not specifically mention the normalizer of A. Thus, it is shown in [PS01] that A is Cartan in M iff there exists U0 = U∗0 ⊂ U(M ) such that spU0 = M and A ∋ a 7→ EA(uau∗) ∈ A are c.p. maps with discrete (countable) Fubini decomposition, ∀u ∈ U0, and also iff this is true for U0 = U. It is also shown in [PV14] that A is Cartan iff M = hM, Ai has the Kadison-Singer norm-paving property relative to A = A∨JAJ and iff there exists a normal conditional expectation of M onto A. However, there exists no local, intrinsic characterization of factors M in the class CF , that does not specifically use the Cartan MASA of M . Such a characterization would certainly be very interesting. It may be useful in deformation-rigidity theory, but also for studying permanence properties of CF , such as stability to inductive limits, to finite index extension/restriction, or to crossed products by amenable groups. The criterion 1.4.1(2) may be of help in this direction. A related question is to find an intrinsic, local characterization of factors with unique (up to unitary conjugacy) Cartan subalgebra. There are reasons to believe that any irreducible subfactor N ⊂ M of a factor M in the class CF has Jones index equal to the square norm of a (finite or infinite) bipartite graph. This may even be true for the (possibly larger) class of all s-thin factors. We will discuss the motivations behind this conjecture in a future paper. 5.3. Strengthened singularity. As shown in ([P81d]), any II1 factor M with the property (T) has a MASA A ⊂ M with the property that the only automorphisms of M that normalize A are the automorphism of M implemented by unitaries in A. With the terminology in the remark before Corollary 2.2, this amounts to A being Aut(M )-singular. Equivalently, if θ : M ≃ N is an isomorphism of M onto another II1 factor, then θ is in some sense uniquely determined by its restriction to A, θA. For this reasons, a MASA with this property in a II1 factor M is called super-singular in M (see 5.1 in [P13]). It is shown in [P13] that, besides property (T) factors, the hyperfinite II1 factor has super-singular MASAs as well. It is an open problem whether any separable II1 factor has super-singular MASAs. The proof of Corollary 2.2 shows that the following property for a MASA A ⊂ N implies super-singularity: given any embedding of N into a II1 factor M0 such that CONSTRUCTING MASAS 29 [M0 : N ] < ∞, one has NM0 (A) = U(A)U(N′ ∩ M0). By ([P86]), any property (T) II1 factor N has a MASA A ⊂ N satisfying this strengthened super-singularity. Indeed, by [J83] any embedding with finite index N ֒→ M0 arises from a basic construction M0 = hN, P i, for some subfactor P ⊂ N with [N : P ] = [M0 : N ], while by (Theorem 4.5.1 in [P86]) there are only countably many subfactors of finite index of N up to unitary conjugacy, so Theorem 2.1 applies. On the other hand, Theorem 2.1 suggests the following question: does there exist a separable II1 factor N with a MASA A ⊂ N having the property that for any embedding of N into a separable II1 factor M0 with N′ ∩ M0 atomic, one has NM0 (A) = U(A)U(N′ ∩ M0)? The answer to this question is however negative: if A ⊂ N is a MASA that one also identifies with a Cartan MASA in the hyperfinite II1 factor, A ≃ L∞([0, 1]) ֒→ R, then M0 = N ∗A R has the property that N′ ∩M0 = C1 but NM0(A)′′ ⊃ R. Another strengthening of the singularity property for a MASA A ⊂ M is ob- tained by requiring A to be maximal amenable (or equivalently, maximal AFD, by [C75]) in M , i.e., to be so that there exists no intermediate amenable subalgebra A ⊂ B ⊂ M with A 6= B (so M must be non-amenable). The existence of such MASAs was discovered in [P81c], where it was shown that A = L∞([0, 1]) is maxi- mal amenable in M = A ∗ P . In particular, the MASA Au generated by one of the generators of the free group u ∈ Fn is maximal amenable in M = L(Fn). We have conjectured in the early 1980s that any non-amenable II1 factor contains maximal amenable MASAs. We will discuss this problem in details in a forthcoming paper. References [AP17] C. Anantharaman, S. Popa: "An introduction to II1 factors", www.math.ucla.edu/∼popa/Books/IIun-v13.pdf [CS11] I. Chifan, T. Sinclair: On the structural theory of II1 factors of negatively curved groups, Ann. Sci. de l'´Ecole Norm. Sup. 46 (2013), 1-34. [CSU12] I. Chifan, T. Sinclair, B. Udrea: On the structural theory of II1 factors of nega- tively curved groups II, Adv. in Math. 245 (2013), 208-236. [C75] A. Connes: Classification of injective factors, Ann. of Math., 104 (1976), 73-115. [CFW81] A. Connes, J. Feldman, B. Weiss: An amenable equivalence relation is generated by a single transformation, Erg. Theory Dyn. Sys. 1 (1981), 431-450. [D54] J. Dixmier: Sousanneaux abeliens maximaux dans les facteurs de type fini, Ann. of Math. 59 (1954), 279-286. [D57] J. Dixmier: "Les algebres d'operateurs dans l'espace hilbertien", Gauthier-Vill- ars, Paris 1957, 1969. [FM77] J. Feldman, C.C. Moore: Ergodic equivalence relations, cohomology, and von Neumann algebras II, Trans. AMS 234 (1977), 325-359. 30 SORIN POPA [G98] L.M. Ge: Applications of free entropy to finite yon Neumann algebras, Amer. J. Math. 119 (1997), 467-485. [GP98] L.M. Ge, S. Popa: On some decomposition properties for factors of type II1, Duke Math. J., 94 (1998), 79-101. [H79] P. Hahn: Reconstruction of a factor from measures on Takesakis unitary equiv- alence relation, J. Funct. Anal. 31 (1979) 263-271. [J83] V.F.R. Jones: Index for subfactors, 72 (1983), 1-25. [JP81] V.F.R. Jones, S. Popa: Some properties of MASAs in factors, in Proceedings of VI-th Conference in Operator Theory, Herculane-Timisoara 1981, I. Gohberg (ed.), Birkhauser Verlag, 1982, pp 210-220. [Ju01] K. Jung: The free entropy dimension of hyperfinite von Neumann algebras, Trans. AMS 355 (2003), 5053-5089. [K67] R.V. Kadison: Problems on von Neumann algebras, Baton Rouge Conference 1967, unpublished notes. [MvN36] F. Murray, J. von Neumann: On rings of operators, Ann. Math. 37 (1936), 116-229. [MvN43] F. Murray, J. von Neumann: Rings of operators IV, Ann. Math. 44 (1943), 716-808. [OP07] N. Ozawa, S. Popa: On a class of II1 factors with at most one Cartan subalgebra, Annals of Mathematics 172 (2010), 101-137 (math.OA/0706.3623) [PP84] M. Pimsner, S. Popa, Entropy and index for subfactors, Annales Scient. Ecole Norm. Sup., 19 (1986), 57-106. [P81a] S. Popa: On a problem of R.V. Kadison on maximal abelian -subalgebras in factors, Invent. Math., 65 (1981), 269-281. [P81b] S. Popa: Orthogonal pairs of -subalgebras in finite von Neumann algebras, J. Operator Theory, 9 (1983), 253-268. [P81c] S. Popa: Maximal injective subalgebras in factors associated with free groups, Advances in Math., 50 (1983), 27-48. [P81d] S. Popa: Singular maximal abelian *-subalgebras in continuous von Neumann algebras, J. Funct. Analysis, 50 (1983), 151-166. [P82] S. Popa: Notes on Cartan subalgebras in type II1 factors. Mathematica Scandi- navica, 57 (1985), 171-188. [P86] S. Popa: Correspondences, INCREST Preprint 56/1986, www.math.ucla.edu/ popa/preprints.html [P89] S. Popa: Classification of subfactors: the reduction to commuting squares, Invent. Math., 101 (1990), 19-43. [P92] S. Popa: Classification of amenable subfactors of type II, Acta Mathematica, 172 (1994), 163-255. CONSTRUCTING MASAS 31 [P94] S. Popa: Classification of subfactors and of their endomorphisms, CBMS Lecture Notes, 86, Amer. Math. Soc. 1995. [P97] S. Popa: Some properties of the symmetric enveloping algebras with applications to amenability and property T, Documenta Mathematica, 4 (1999), 665-744. [P01] S. Popa: On a class of type II1 factors with Betti numbers invariants, Ann. of Math. 163 (2006), 809-899. [P03] S. Popa: Strong Rigidity of II1 Factors Arising from Malleable Actions of w-Rigid Groups I, Invent. Math., 165 (2006), 369-408. (math.OA/0305306). [P05a] S. Popa: "Deformation-rigidity theory", NCGOA mini-course, Vanderbilt Uni- versity, May 2005. [P05b] S. Popa: Cocycle and orbit equivalence superrigidity for malleable actions of w-rigid groups, Invent. Math. 170 (2007), 243-295 (math.GR/0512646). [P06] S. Popa: Deformation and rigidity for group actions and von Neumann alge- bras, in "Proceedings of the International Congress of Mathematicians" (Madrid 2006), Volume I, EMS Publishing House, Zurich 2006/2007, pp. 445-479. [P13] S. Popa: A II1 factor approach to the Kadison-Singer problem, Comm. Math. Physics. 332 (2014), 379-414 (math.OA/1303.1424). [PS01] S. Popa, D. Shlyakhtenko: Cartan subalgebras and bimodule decomposition of type II1 factors, Math. Scandinavica 92 (2003), 93-102. [PV11] S. Popa, S. Vaes: Unique Cartan decomposition for II1 factors arising from arbitrary actions of free groups, Acta Mathematica, 194 (2014), 237-284 [PV12] S. Popa, S. Vaes: Unique Cartan decomposition for II1 factors arising from arbitrary actions of hyperbolic groups, Journal fur die reine und angewandte Mathematik, 690 (2014), 433-458. [PV14] S. Popa, S. Vaes: Paving over arbitary MASAs in von Neumann algebras, Anal- ysis and PDE (2015) 101-123 (math.OA/1412.0631) [Pu61] L. Pukanszky: On maximal abelian subrings in factors of type II1, Canad. J. Math. 12 (1960), 289-296. [T63] M. Takesaki: On the unitary equivalence among the components of decomposi- tions of representations of involutive Banach algebras and the associated diagonal algebras, Tohoku Math. J. bf 15 (1963), 365-393. [V07] S. Vaes: Explicit computations of all finite index bimodules for a family of II1 factors, Ann. Sci. de l'´Ecole Norm. Sup. 41 (2008), 743-788. [Vo96] D. Voiculescu: The analogues of entropy and Fisher information measure in free probability III, GAFA 6 (1996), 172-199. Math.Dept., UCLA, Los Angeles, CA 90095-1555 E-mail address: [email protected]
1608.01092	1	1608	2016-08-03T07:13:22	Herz-Schur multipliers of dynamical systems	[ "math.OA", "math.FA" ]	We extend the notion of Herz-Schur multipliers to the setting of non-commutative dynamical systems: given a C-algebra $A$, a locally compact group $G$, and an action $\alpha$ of $G$ on $A$, we define transformations on the (reduced) crossed product $A\rtimes_{r,\alpha} G$ of $A$ by $G$, which, in the case $A = \mathbb{C}$, reduce to the classical Herz-Schur multipliers. We also introduce a class of Schur $A$-multipliers, establish its characterisation which generalise the classical descriptions of Schur multipliers and present a transference theorem in the new setting, identifying isometrically the Herz-Schur multipliers of the dynamical system $(A,G,\alpha)$ with the invariant part of the Schur $A$-multipliers. We discuss special classes of Herz-Schur multipliers, in particular, those which are associated to a locally compact abelian group $G$ and its canonical action on the $C^$-algebra $C^*(\Gamma)$ of the dual group $\Gamma$.	math.OA	math	HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS A. MCKEE, I. G. TODOROV, AND L. TUROWSKA Contents Introduction 1. 2. Schur A-multipliers 3. Herz-Schur multipliers and transference 4. Multipliers of the weak* crossed product 5. Two classes of multipliers 5.1. Multipliers from the Haagerup tensor product 5.2. Groupoid multipliers 6. Convolution multipliers References 1. Introduction 1 4 13 28 32 32 35 38 47 The notion of a Schur multiplier has its origins in the work of I. Schur in the early 20th century, and is based on entry-wise (or Hadamard) product of matrices. More specifically, a bounded function ϕ : N × N → C is called a Schur multiplier if (ϕ(i, j)ai,j ) is the matrix of a bounded linear operator on ℓ2 whenever (ai,j) is such. A concrete description of Schur multipliers, which found numerous applications thereafter, was given by A. Grothendieck in his R´esum´e [12] (see also [30]). A measurable version of Schur multipliers was developed by M. S. Birman and M. Z. Solomyak (see [3] and the references therein) and V. V. Peller [28]. More concretely, given standard measure spaces (X, µ) and (Y, ν) and a function ϕ : X × Y → C, one defines a linear transformation Sϕ on the space of all Hilbert-Schmidt operators from H1 = L2(X, µ) to H2 = L2(Y, ν) by multiplying their integral kernels by ϕ; if Sϕ is bounded in the operator norm (in which case ϕ is called a measurable Schur multiplier ), it is extended to the space K(H1, H2) of all compact operators from H1 into H2 by continuity. The map Sϕ is defined on the space B(H1, H2) of all bounded linear operators from H1 into H2 by taking the second dual of the constructed map on K(H1, H2). A characterisation of measurable Schur multipliers, extending Grothendieck's result, was obtained in [13] and [28] (see also [17] and [38]). Namely, a function ϕ ∈ L∞(X × Y ) was shown to be a Schur multiplier if and only if ϕ coincides almost k=1 ak(x)bk(y), where (ak)k∈N everywhere with a function of the form P∞ 1 2 A. MCKEE, I. TODOROV, AND L. TUROWSKA and (bk)k∈N are families of essentially bounded measurable functions such k=1 bk(y)2 < ∞. that esssupx∈XP∞ k=1 ak(x)2 < ∞ and esssupy∈Y P∞ Among the large number of applications of Schur multipliers is the de- scription of the space M cbA(G) of completely bounded multipliers (also known as Herz-Schur multipliers) of the Fourier algebra A(G) of a locally compact group G, introduced by J. de Canni`ere and U. Haagerup in [7]. Namely, as shown by M. Bozejko and G. Fendler [5] , M cbA(G) can be isometrically identified with the space of all Schur multipliers on G × G of Toeplitz type. An alternative proof of this result was given by P. Jolissaint [15]. Herz-Schur multipliers have been highly instrumental in operator algebra theory, providing the route to defining and studying a number of approxi- mation properties of group C-algebras and group von Neumann algebras (we refer the reader to [20], [6] and [18]). Here one uses the fact that every Herz-Schur multiplier on a locally compact group G gives rise to a (com- pletely bounded) map on the von Neumann algebra VN(G) of G, leaving invariant the reduced C-algebra C ∗ r (G) of G. In view of the large number of applications of Herz-Schur multipliers in operator algebra theory, it is natural to seek generalisations going beyond the context of group algebras. The main goal of this paper is to extend the notion of Herz-Schur multipliers to the setting of non-commutative dy- namical systems. Given a C-algebra A, a locally compact group G, and an action α of G on A, we define transformations on the (reduced) crossed product A⋊r,α G of A by G, which, in the case A = C, reduce to the classical Herz-Schur multipliers and, in the case of a discrete group G, to multipliers defined recently by E. Bedos and R. Conti in [2]. More generally, we intro- duce Schur A-multipliers which, in the case A = C, reduce to the classical measurable Schur multipliers. In Section 2, we establish a characterisation of Schur A-multipliers that generalises the classical description of Schur multipliers (Theorem 2.6). We exhibit a large class of Schur A-multipliers defined in terms of Hilbert A-bimodules, and show that it exhausts all Schur A-multipliers in the case A is finite-dimensional (Theorem 2.7). In Section 3, we prove a transference theorem in the new setting, identifying isometri- cally the Hezr-Schur multipliers of the dynamical system (A, G, α) with the invariant part of the Schur A-multipliers (see Theorems 3.8 and 3.18). We introduce bounded multipliers of (A, G, α) and relate them to Herz-Schur (A, G, α)-multipliers, extending a corresponding result from [7]. In Section 4, we provide a description of a more general and closely related class of mul- tipliers, namely, Herz-Schur multipliers associated with weak closed crossed products, as the commutant of the scalar valued Herz-Schur multipliers as- sociated with elements of M cbA(G) (Theorem 4.3). While in the case A = C this description is straightforward, here we need to use structure theory of crossed products and some recent results from [1]. The rest of the paper is devoted to special classes of Herz-Schur multi- pliers. Namely, in Section 5, we consider multipliers naturally associated HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 3 with the Haagerup tensor product of two copies of A, and multipliers de- fined on groupoids. In the former case, we relate our notion to examples of Herz-Schur multipliers exhibited in [2, Theorem 4.5] in the case of a discrete group. In the latter case, we show that completely bounded multipliers of the Fourier algebra of a groupoid, defined in [33], form a subclass of the class of Herz-Schur multipliers introduced in the present work. The results in Section 6 were our original motivation for the present pa- per. Here, we consider the case of a locally compact abelian group G and its canonical action α on the C-algebra C ∗(Γ) of the dual group Γ. We focus on a special class F(G) of Herz-Schur multipliers, which we call convolution multipliers, and its natural subclass Fθ(G) of weak extendible convolution multipliers. We show that the Fourier-Stieltjes algebras B(G) and B(Γ) can both be viewed as subspaces of Fθ(G), while Fθ(G) is a subspace of their Fubini product. When the crossed product of C ∗(Γ) ⋊α G is canoni- cally identified with the space K(L2(G)) of all compact operators on L2(G), the elements u of B(G) give rise to the measurable Schur multipliers corre- sponding to u via the aforementioned Bozejko-Fendler classical transference theorem, while the elements of B(Γ) correspond to a well-known class of completely bounded maps, arising from a representation of the measure al- gebra M (G) of G on K(L2(G)), studied in a variety of contexts in both operator algebra theory and quantum information theory, and by a number of authors including F. Ghahramani [11], M. Neufang and V. Runde [24], M. Neufang, Zh.-J. Ruan and N. Spronk [25] and E. Størmer [39]. The main result of the section are Theorem 6.7 and 6.10, where we identify the set Fθ(G), and an associated subset, of convolution multipliers of G as subsets of the joint commutant of the two families described above. The paper uses various notions from Operator Space Theory; we refer the reader to [4], [9], [26] or [31] for the basics. For background and notation on crossed products, which will be needed in Sections 3, 5 and 6, we refer the reader to [43]. We finish this section with setting some notation. If E and F are vector spaces, we let E ⊙ F be their algebraic tensor product. For a Banach space X , we let B(X ) (resp. K(X )) be the algebra of all bounded linear (resp. compact) operators on X , and denote by IX the identity operator on X . If H and K are Hilbert spaces, we denote by H ⊗ K their Hilbertian tensor product; for operators S ∈ B(H) and T ∈ B(K), we let S⊗T be the bounded operator on H ⊗ K given by (S ⊗ T )(ξ ⊗ η) = Sξ ⊗ T η. The (norm closed) spacial tensor product of two (norm closed) operator spaces U ⊆ B(H) and V ⊆ B(K) will be denoted by U ⊗ V. If U and V are weak* closed, their weak* spacial tensor product will be denoted by U ¯⊗V. 4 A. MCKEE, I. TODOROV, AND L. TUROWSKA 2. Schur A-multipliers Let (X, µ) be a standard measure space; this means that µ is a Radon measure with respect to some complete metrisable separable locally com- pact topology (called an admissible topology) on X. For p = 1, 2 and a Banach space E, we write Lp(X, E) for the corresponding Lebesgue space of all (equivalence classes of) weakly measurable p-summable E-valued func- tions on X (see e.g. [43, Appendix B]). If H and K are separable Hilbert spaces and E ⊆ B(H, K) is a weak* closed subspace, let L∞(X, E) be the space of all (equivalence classes of) bounded E-valued functions T on X such that, for every ξ ∈ H and every η ∈ K, the functions x → T (x)ξ and x → T (x)∗η are weakly measurable. Note that L∞(X, E) contains all bounded weakly measurable functions from X into E. Analogously to [40, Chapter IV, Section 7], we often identify an element g of L∞(X, E) with the operator Dg from L2(X, H) into L2(X, K) given by (Dgξ)(x) = g(x)(ξ(x)), x ∈ X. We write k · kp for the norm on Lp(X, E), p = 1, 2, ∞. In the case E coincides with the complex field C, we simply write Lp(X). If f ∈ Lp(X) and a ∈ E, we let f ⊗ a ∈ Lp(X, E) be the function given by (f ⊗ a)(x) = f (x)a, x ∈ X. We fix throughout the section a separable Hilbert space H. For a ∈ L∞(X), let Ma ∈ B(L2(X)) be the operator given by Maξ = aξ; set DX = {Ma : a ∈ L∞(X)}. Note that the identification L2(X) ⊗ H ≡ L2(X, H) yields a unitary equiv- alence between L∞(X, B(H)) and DX ¯⊗B(H) [40, Theorem 7.10]. Let (Y, ν) be a(nother) standard measure space. We equip the direct products X × Y and Y × X with the corresponding product measures. It is easy to see that, if k ∈ L2(Y × X, B(H)) and ξ ∈ L2(X, H) then, for almost all y ∈ Y , the function x → k(y, x)ξ(x) is weakly measurable; moreover, ZX (1) kk(y, x)ξ(x)kdµ(x) ≤ ZX kk(y, x)kkξ(x)kdµ(x) ≤ kξk2(cid:18)ZX kk(y, x)k2dµ(x)(cid:19)1/2 . It follows that the formula (2) (Tkξ)(y) =ZX k(y, x)ξ(x)dµ(x), y ∈ Y, defines a (weakly measurable) function Tkξ : Y → H. Lemma 2.1. Let k ∈ L2(Y × X, B(H)). Equation (2) defines a bounded operator Tk : L2(X, H) → L2(Y, H) with kTkk ≤ kkk2. Moreover, Tk = 0 if and only if k = 0 almost everywhere. HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 5 Proof. Let ξ ∈ L2(X, H). Then, by (1), kTkξk2 = ZY kTkξ(y)k2dν(y) ≤ kξk2 kk(y, x)k2dµ(x)dν(y) = kkk2 2kξk2 2. 2ZY ZX Thus, Tk is bounded and its norm does not exceed kkk2. It is clear that if k = 0 almost everywhere then Tk = 0. Conversely, If suppose that Tk = 0. Choose a countable dense subset {ei}i∈N of H. ξ ∈ L2(X) and η ∈ L2(Y ) then ZX×Y hk(y, x)ei, ejiξ(x)η(y)dxdy = hTk(ξ ⊗ ei), η ⊗ eji = 0, and it follows that hk(y, x)ei, eji = 0 almost everywhere, for all i, j. Since k(y, x) is a bounded operator, k(y, x) = 0 for almost all (x, y). (cid:3) If M ⊆ B(H) is a C-subalgebra, let S2(Y × X, M) = {Tk : k ∈ L2(Y × X, M)}. Note that, if w ∈ L2(Y × X) and a ∈ M then (3) Letting Tw⊗a = Tw ⊗ a. def = K(L2(X), L2(Y )) K be the space of all compact operators from L2(X) into L2(Y ), we have that S2 ⊙ M is norm dense in K ⊗ M (here S2 denotes the space of all Hilbert- Schmidt opertors from L2(X) into L2(Y )). We conclude that S2(Y × X, M) is norm dense in K ⊗ M and equip it with the operator space structure arising from its inclusion into K ⊗ M. We fix throughout the section a non-degenerate separable C-algebra A ⊆ B(H). If B is a(nother) C-algebra, we denote by CB(A, B) the space of all completely bounded linear maps from A into B and write CB(A) = CB(A, A). A function ϕ : X × Y → CB(A, B(H)) will be called pointwise measurable if, for every a ∈ A, the function (x, y) → ϕ(x, y)(a) from X × Y into B(H) is weakly measurable. Let ϕ : X × Y → CB(A, B(H)) be a bounded pointwise measurable function. For k ∈ L2(Y × X, A), let ϕ · k : Y × X → B(H) be the function given by (ϕ · k)(y, x) = ϕ(x, y)(k(y, x)), (y, x) ∈ Y × X. It is easy to show that ϕ · k is weakly measurable; since ϕ is bounded, ϕ · k ∈ L2(Y × X, B(H)) and kϕ · kk2 ≤ kϕk∞kkk2. Let Sϕ : S2(Y × X, A) → S2(Y × X, B(H)) 6 A. MCKEE, I. TODOROV, AND L. TUROWSKA be the linear map given by Sϕ(Tk) = Tϕ·k, By Lemma 2.1, Sϕ is well-defined. k ∈ L2(Y × X, A). Definition 2.2. A bounded pointwise measurable map ϕ : X × Y → CB(A, B(H)) will be called a Schur A-multiplier if the map Sϕ is completely bounded. It follows from the discussion after Lemma 2.1 that a bounded pointwise measurable function ϕ : X × Y → CB(A, B(H)) is a Schur A-multiplier if and only if the map Sϕ possesses a completely bounded extension to a map from K ⊗ A into K ⊗ B(H) (which will still be denoted by Sϕ). We let S(X, Y ; A) be the space of all Schur A-multipliers and endow it with the norm (4) kϕkS = kSϕkcb; it follows from Lemma 2.1 that if Sϕ = 0 then ϕ = 0 almost everywhere, and hence (4) indeed defines a norm on S(X, Y ; A). Note that Schur C-multipliers coincide with the classical (measurable) Schur multipliers [28], [17]. A special role in our considerations will be played by Schur A-multipliers ϕ for which ϕ(x, y) ∈ CB(A) for all (x, y) ∈ X × Y , that is, ones for which the range of ϕ(x, y) is in A. In this case, Sϕ is a map on S2(Y × X, A). We let S0(X, Y ; A) be the space of all such Schur A-multipliers. The next proposition shows that S0(X, Y ; A) does not depend on the faithful - representation of A. Proposition 2.3. Let θ : A → B(K) be a faithful -representation of A on a separable Hilbert space K. A bounded pointwise measurable map ϕ : X×Y → CB(A) is a Schur A-multiplier if and only if the (bounded pointwise measur- able) map ϕθ : X × Y → CB(θ(A)), given by ϕθ(x, y)(θ(a)) = θ(ϕ(x, y)(a)), a ∈ A, is a Schur θ(A)-multiplier. Moreover, kϕkS = kϕθkS. Proof. Let B = θ(A). Note that the map id ⊗θ : K ⊙ A → K ⊙ B given by (id ⊗θ)(T ⊗ a) = T ⊗ θ(a) extends to a complete isometry from K ⊗ A onto K ⊗ B [4]. Let k ∈ L2(Y × X, A); then the map kθ = θ ◦ k belongs to L2(Y × X, B). We claim that Tkθ = (id ⊗θ)(Tk). (5) To see (5), note first that, by (3), it holds when k = k′ ⊗ a for some k′ ∈ L2(Y × X) and a ∈ A; hence, by linearity, it holds if k ∈ L2(Y × X) ⊙ A. By [40, Proposition 7.4], there exists a sequence (ki)i∈N ⊆ L2(Y × X) ⊙ A such that kki − kk2 →i→∞ 0. It follows that kθ ◦ ki − θ ◦ kk2 →i→∞ 0. By Lemma 2.1, Tki → Tk and Tθ◦ki → Tθ◦k in the operator norm. Thus, (id ⊗θ)(Tki) → (id ⊗θ)(Tk) in the operator norm, and we conclude that Tkθ = (id ⊗θ)(Tk). HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 7 By (5), in other words, (id ⊗θ)(Sϕ(Tk)) = Sϕθ (Tkθ ); Sϕθ = (id ⊗θ) ◦ Sϕ ◦ (id ⊗θ−1). It follows that Sϕ is completely bounded if and only if Sϕθ is so and that, in this case, kϕkS = kϕθkS. (cid:3) Proposition 2.3 allows us to view the elements of S0(X, Y ; A) indepen- dently of the particular faithful representation of A on a separable Hilbert space; we will thus in the sequel refer to A-valued Schur A-multipliers with- out the need to specify a particular representation. Lemma 2.4. Let ϕ ∈ S(X, Y ; A). If C ∈ DX, D ∈ DY and T ∈ K ⊗ A then (6) Sϕ((D ⊗ IH)T (C ⊗ IH)) = (D ⊗ IH)Sϕ(T )(C ⊗ IH). Proof. By continuity and linearity, it suffices to establish (6) in the case T = Tk ⊗ a, where k ∈ L2(Y × X) and a ∈ A. Assuming that C = Mc and D = Md, where c ∈ L∞(X) and d ∈ L∞(Y ), we have that both Sϕ((D⊗IH )T (C ⊗IH )) and (D⊗IH )Sϕ(T )(C ⊗IH ) coincide with Tk′, where k′ is the (weakly measurable) function from Y × X into B(H) given by k′(y, x) = c(x)d(y)k(x, y)ϕ(x, y)(a). (cid:3) Lemma 2.5. Let E and L be separable Hilbert spaces and θ : K(E) ⊗ A → B(L) be a non-degenerate -representation. Then there exists a separable Hilbert space K, a unitary operator U : L → E ⊗ K and a non-degenerate -representation ρ : A → B(K) such that U θ(b ⊗ a)U ∗ = b ⊗ ρ(a), b ∈ K(E), a ∈ A. Proof. Let M (K(E) ⊗ A) be the multiplier algebra of K(E) ⊗ A. There exists a unital -homomorphism θ : M (K(E)⊗A) → B(L) extending θ (see e.g. [27, Proposition 3.12.10]). The map x → θ(x ⊗ IH) is clearly a non-degenerate -representation of K(E) on L. Thus, there exists a separable Hilbert space K and a unitary operator U : L → E ⊗ K such that U θ(b ⊗ IH )U ∗ = b ⊗ IK, b ∈ K(E). Let θ : M (K(E) ⊗ A) → B(E ⊗ K) be given by θ(T ) = U θ(T )U ∗, T ∈ M (K(E) ⊗ A). For a ∈ A and b ∈ K(E), the operators θ(b ⊗ IH) and θ(IE ⊗ a) commute. It follows that θ(IE ⊗ a) = IE ⊗ ρ(a), for some operator ρ(a) ∈ B(K). Since 8 A. MCKEE, I. TODOROV, AND L. TUROWSKA θ is a unital -homomorphism, the map ρ : A → B(K) is easily seen to be a non-degenerate -homomorphism. Moreover, if b ∈ K(E) and a ∈ A then U θ(b ⊗ a)U ∗ = U θ(b ⊗ IH)θ(IE ⊗ a)U ∗ = U θ(b ⊗ IH)U ∗U θ(IE ⊗ a)U ∗ = θ(b ⊗ IH)θ(IE ⊗ a) = (b ⊗ IK)(IE ⊗ ρ(a)) = b ⊗ ρ(a). (cid:3) Theorem 2.6. Let ϕ : X × Y → CB(A, B(H)) be a bounded pointwise measurable function. The following are equivalent: (i) ϕ is a Schur A-multiplier; (ii) there exist a separable Hilbert space K, a non-degenerate -representa- tion ρ : A → B(K), V ∈ L∞(X, B(H, K)) and W ∈ L∞(Y, B(H, K)) such that, for all a ∈ A, ϕ(x, y)(a) = W ∗(y)ρ(a)V (x), for almost all (x, y) ∈ X × Y . Proof. (i)⇒(ii) Suppose that ϕ ∈ S(X, Y ; A). Let E = L2(Y ) ⊕ L2(X) and Φ : K(E) ⊗ A → K(E) ⊗ B(H) be given by Φ(cid:18)(cid:18)x1,1 x1,2 x2,1 x2,2(cid:19) ⊗ a(cid:19) =(cid:18)0 Sϕ(x1,2 ⊗ a) (cid:19) . 0 0 (7) It is clear that Φ is a completely bounded map with kΦkcb = kSϕkcb. By the Haagerup-Paulsen-Wittstock Theorem, there exist a Hilbert space L, a non-degenerate -homomorphism θ : K(E) ⊗ A → B(L) and operators V0, W0 ∈ B(E ⊗ H, L) such that Φ(T ) = W ∗ 0 θ(T )V0, T ∈ K(E) ⊗ A. As K(E)⊗A is separable, we may assume that L is separable. By Lemma 2.5, there exist a separable Hilbert space K, a unitary operator U : L → E ⊗ K and a -representation ρ : A → B(K) such that U θ(b ⊗ a)U ∗ = b ⊗ ρ(a), b ∈ K(E), a ∈ A. Let W = U W0 and V = U V0. Then Φ(b ⊗ a) = W ∗(b ⊗ ρ(a)) V , b ∈ K(E), a ∈ A. Writing V and W in two by two matrix form and recalling (7), we conclude that there exist bounded operators V : L2(X, H) → L2(X, K) and W : L2(Y, H) → L2(Y, K) such that (8) Let Sϕ(b ⊗ a) = W ∗(b ⊗ ρ(a)) V , b ∈ K, a ∈ A. S def= [{T V L2(X, H) : T ∈ K(L2(X)) ⊗ ρ(A)}]. Clearly, S is invariant under K(L2(X)) ⊗ ρ(A). Thus, the projection onto S has the form IL2(X) ⊗ E, for some projection E ∈ ρ(A)′. Moreover, (9) V = (IL2(X) ⊗ E) V . HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 9 Setting ρ = id ⊗ρ (so that ρ is a map from K ⊗ A into K ⊗ B(K)), by (8) and (9), we now have (10) Sϕ(T ) = W ∗ ρ(T )(IL2(X) ⊗ E) V , T ∈ K ⊗ A. Note, further, that if c ∈ L∞(X) and d ∈ L∞(Y ) then (11) ρ((M ∗ d ⊗ IH)T (Mc ⊗ IH)) = (M ∗ d ⊗ IK)ρ(T )(Mc ⊗ IK ). Let W = (IL2(Y ) ⊗ E) W . Since ρ(T )(IL2(X) ⊗ E) = (IL2(Y ) ⊗ E)ρ(T ), we conclude from (10) that (12) Sϕ(T ) = W ∗(IL2(Y ) ⊗ E)ρ(T ) V = W ∗ ρ(T ) V , for every T ∈ K ⊗ A. Identities (11) and (12) and Lemma 2.4 imply that W ∗(M ∗ d ⊗ IK)ρ(T ) V = (M ∗ d ⊗ IH )W ∗ ρ(T ) V , T ∈ K ⊗ A. (13) Thus, Dρ(T ) V ξ, (Md ⊗ IK )W ηE =Dρ(T ) V ξ, W (Md ⊗ IH)ηE, for all ξ ∈ L2(X, H) and all η ∈ L2(Y, H). We conclude that (IL2(Y ) ⊗ E)(Md ⊗ IK)W = (IL2(Y ) ⊗ E)W (Md ⊗ IH) and hence (Md ⊗ IK)W = W (Md ⊗ IH) for all d ∈ L∞(Y ). It follows easily that W ∈ L∞(Y, B(H, K)) (see [40, Theorem 7.10]). Let now T def = [{T W L2(Y, H) : T ∈ K(L2(Y )) ⊗ ρ(A)}]. The projection onto T has the form IL2(Y )⊗F for some projection F ∈ ρ(A)′. Letting V = (IL2(X) ⊗ F ) V , and using similar arguments to the ones above, one shows that (Mc ⊗ IK )V = V (Mc ⊗ IH) for all c ∈ L∞(X) and hence that V ∈ L∞(X, B(H, K)). Note that W = (IL2(Y ) ⊗ F )W and hence, by (12), (14) Sϕ(T ) = W ∗(IL2(Y ) ⊗ F )ρ(T ) V = W ∗ ρ(T )(IL2(Y ) ⊗ F ) V = W ∗ ρ(T )V, for every T ∈ K ⊗ A. Let k ∈ L2(Y × X) and a ∈ A. For ξ ∈ L2(X, H) and η ∈ L2(Y, H) we have (15) hSϕ(Tk ⊗ a)ξ, ηi =ZY ZX k(y, x)hϕ(x, y)(a)ξ(x), η(y)idµ(x)dν(y). 10 A. MCKEE, I. TODOROV, AND L. TUROWSKA On the other hand, by (14), hSϕ(Tk ⊗ a)ξ, ηi = hW ∗(Tk ⊗ ρ(a))V ξ, ηi = h(Tk ⊗ ρ(a))V ξ, W ηi = ZY ZX = ZY ZX k(y, x)hρ(a)(V (x)ξ(x)), W (y)η(y)idµ(x)dν(y) k(y, x)hW (y)∗ρ(a)V (x)ξ(x), η(y)idµ(x)dν(y). Comparing the last identity with (15) and taking into account that these identities hold for all k ∈ L2(Y × X), we conclude that (16) hϕ(x, y)(a)ξ(x), η(y)i = hW (y)∗ρ(a)V (x)ξ(x), η(y)i almost everywhere, for all ξ ∈ L2(X, H) and all η ∈ L2(Y, H). If the measures µ and ν are finite, take ξ = χX ⊗ ξ0 and η = χY ⊗ η0, where ξ0, η0 ∈ H. The separability of H and (16) imply that ϕ(x, y)(a) = W (y)∗ρ(a)V (x), for almost all (x, y) ∈ X × Y. If the measures µ and ν are not finite, the proof is completed by choosing increasing sequences (Xn)n∈N and (Yn)n∈N, each of whose terms has finite measure, and letting ξ = χXn ⊗ ξ0 and η = χYn ⊗ η0, with ξ0, η0 ∈ H. (ii)⇒(i) The assumption shows that the mapping Sϕ : S2(Y × X, A) → S2(Y × X, B(H)) satisfies Sϕ(Th ⊗ a) = W ∗(Th ⊗ ρ(a))V, h ∈ L2(Y × X), a ∈ A. By linearity, (17) Sϕ(Tk) = W ∗Tρ◦kV, whenever k ∈ L2(Y × X) ⊙ A. Let k ∈ L2(Y × X, A) be arbitrary. By [40, Proposition 7.4], there exists a sequence (ki)i∈N ⊆ L2(Y × X) ⊙ A with kki − kk2 →i→∞ 0. Using (5), (17), Lemma 2.1 and the fact that ϕ is bounded, we obtain Sϕ(Tk) = lim i→∞ Sϕ(Tki) = lim i→∞ W ∗Tρ◦kiV = W ∗( lim i→∞ Tρ◦ki)V = W ∗( lim i→∞ ρ(Tki)V = W ∗ ρ(Tk)V. Thus, Sϕ has a completely bounded extension to K ⊗ A (namely, the map T → W ∗ ρ(T ))V ) and hence ϕ is a Schur A-multiplier. (cid:3) Remarks (i) The proof of Theorem 2.6 shows that if ϕ ∈ S(X, Y ; A) then the operator valued functions V and W can be chosen so that kϕkS = esssup kW (x)k esssup kV (y)k. x∈X y∈Y (ii) In the case A = C, Theorem 2.6 reduces to the well-known charac- terisation of measurable Schur multipliers due to U. Haagerup [13] and V. HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 11 V. Peller [28] (see also [17]). Indeed, in this case, ρ is equal to the identity representation of C and hence ϕ has a representation of the form (18) ϕ(x, y) = hw(y), v(x)i, where v : X → K and w : Y → K are weakly measurable essentially bounded functions, for some separable Hilbert space K. If, in Theorem 2.6, the operator valued functions V and W can be chosen to be weakly measurable, then we will say that the Schur A-multiplier ϕ has a weakly measurable representation. In the next theorem we exhibit a class of A-valued Schur A-multipliers possessing a weakly measurable representation which exhausts all such multipliers in the case A is finite dimensional. Recall that a Hilbert A-bimodule is a right Hilbert A-module N , equipped with a def = θ(a)(ξ), a ∈ A, ξ ∈ N , for some - left A-module action given by a · ξ representation θ of A into the C-algebra of all adjointable operators on N . As is customary in the literature on Hilbert modules, we assume linearity on the second variable of the A-valued inner product, denoted here by h··iA. Theorem 2.7. Let A ⊆ B(H) be a separable C-algebra and ϕ : X × Y → CB(A) be a bounded pointwise measurable function. Consider the conditions: (i) there exists a countably generated Hilbert A-bimodule N and bounded weakly measurable functions v : X → N and w : Y → N such that (19) ϕ(x, y)(a) = hw(y)a · v(x)iA , for almost all (x, y) ∈ X × Y for every a ∈ A. (ii) ϕ is a Schur A-multiplier possessing a weakly measurable representa- tion. Then (i) ⇒ (ii). If A is finite-dimensional then (i) ⇔ (ii). Proof. (i)⇒(ii) It follows for instance from [8, Example 2.8] that there exist a separable Hilbert space K, an isometry τ : N → B(K) and a faithful - representation π : A → B(K) such that τ (a·z) = π(a)τ (z), τ (z·b) = τ (z)π(b) and π(hz1, z2iA) = τ (z1)∗τ (z2), z, z1, z2 ∈ N , a, b ∈ A. For all a ∈ A, we have that π(ϕ(x, y)(a)) = π(hw(x)a · v(y)iA) = τ (w(x))∗π(a)τ (v(y)) a.e. Moreover, the maps τ ◦ v and τ ◦ w are weakly measurable. By Theorem 2.6, the map ϕπ : X × Y → CB(π(A)), given by ϕπ(x, y)(π(a)) = π(ϕ(x, y)(a)), is a Schur π(A)-multiplier. By Proposition 2.3, ϕ is a Schur A-multiplier. Assume now that A is finite dimensional. By Proposition 2.3, we may k=1Mnk ⊆ B(H), where Mn denotes, as Cnk, nk ∈ N, k = identify A with the C-algebra ⊕m customary, the n by n matrix algebra and H = ⊕m 1, . . . , m. k=1 Suppose that ϕ is a Schur A-multiplier, K is a separable Hilbert space, V : X → B(H, K), W : Y → B(H, K) weakly measurable functions, and ρ : A → B(K) a non-degenerate -representation, such that, for every a ∈ A, we 12 A. MCKEE, I. TODOROV, AND L. TUROWSKA have ϕ(x, y)(a) = W (y)∗ρ(a)V (x) for almost all (x, y) ∈ X × Y . The space def B(H, K) is an operator A-bimodule with respect to the actions a·T = ρ(a)T def = T a, a ∈ A, T ∈ B(H, K). Let Pk be the projection in B(H) and T · a k=1 PkT Pk, T ∈ B(H). Clearly, Ψ is a completely positive projection from B(H) onto A. We equip B(H, K) with the A-valued inner product given by hS, T iA = Ψ(S∗T ). As the projections k=1 Pk = I, we have onto the summand Cnk and Ψ(T ) =Pm Pk, k = 1, . . . , m, are mutually orthogonal and Pm hS, SiA = 0 if and only if S = 0. Moreover, hS, T · aiA = Ψ(S∗T a) = Ψ(S∗T )a = hS, T iAa, and hence N def = B(H, K) is a right Hilbert A-module. In addition, ha · S, T iA = Ψ(S∗ρ(a)∗T ) = hS, a∗ · T iA, showing that the map θa : S → a · S is adjointable and that the map a → θa is a -representation; thus, N is a Hilbert A-bimodule and ϕ(x, y)(a) = hW (x), a · V (x)iA for almost all (x, y) ∈ X × Y . As H is finite dimensional and K is separable, N is countably generated. (cid:3) Let B = B(L2(X), L2(Y )) for brevity. Proposition 2.8. If ϕ ∈ S(X, Y ; A) then the map Sϕ has a unique ex- tension to a completely bounded weak continuous map from B ¯⊗A∗∗ into B ¯⊗B(H). Proof. Let P : B(H)∗∗ → B(H) be the canonical projection (that is, the adjoint of the inclusion map of the trace class on H into B(H)∗). Then the map id ⊗P : B ¯⊗B(H)∗∗ → B ¯⊗B(H) is weak* continuous and completely contractive (see e.g. [4] and [9, Proposition 7.1.6]). Let K = K(L2(Y ) ⊕ L2(X)) and B = B(L2(Y ) ⊕ L2(X)). Write PX (resp. PY ) for the projection from L2(Y ) ⊕ L2(X) onto L2(X) (resp. L2(Y )). Let Φ : K ⊗ A → K ⊗ B(H) be given by (20) Φ(cid:18)(cid:18)x1,1 x1,2 x2,1 x2,2(cid:19) ⊗ a(cid:19) =(cid:18)0 Sϕ(x1,2 ⊗ a) (cid:19) . 0 0 By [14, Example 1], given a C ∗-algebra B, there is a canonical normal - isomorphism ( K ⊗ B)∗∗ ∼= B ¯⊗B∗∗. (21) Hence we may view the second dual Φ∗∗ as a completely bounded map from B ¯⊗A∗∗ to B ¯⊗B(H)∗∗, extending Φ. As K ⊗ A is weak dense in B ¯⊗A∗∗ we have that for any T ∈ B ¯⊗A∗∗ there exists Ψ(T ) ∈ B ¯⊗B(H)∗∗ such that In particular, Φ∗∗(T ) =(cid:18)0 Ψ(T ) 0 (cid:19) . 0 Φ∗∗((PY ⊗ id)T (PX ⊗ id)) =(cid:18)0 Ψ((PY ⊗ id)T (PX ⊗ id)) 0 0 (cid:19) , HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 13 and the mapping Ψ = ΨB ¯⊗A∗∗ : B ¯⊗A∗∗ → B ¯⊗B(H)∗∗ is completely bounded and weak* continuous. Hence the composition (id ⊗P ) ◦ Ψ : B ¯⊗A∗∗ → B ¯⊗B(H) is a completely bounded weak* continuous map, extending Sϕ. The fact that this extension is unique follows by weak* density. (cid:3) We will use the same symbol, Sϕ, to denote the map obtained in Propo- sition 2.8. We note that if Sϕ satisfies equation (14), that is, if Sϕ(S) = W ∗(id ⊗ρ)(S)V for all S ∈ K ⊗ A, then Sϕ(T ⊗ a) = W ∗(T ⊗ ρ(a))V for all T ∈ B and all a ∈ A∗∗, where ρ has been canonically extended to A∗∗. While Proposition 2.8 implies that, if ϕ ∈ S0(X, Y ; A), then the map Sϕ on K ⊗ A has a weak* continuous extension to B ⊗ A∗∗, an analogous extension is not guaranteed to exist in representations of A different from the universal one. This motivates the following definition. Definition 2.9. Let A be a separable C-algebra and θ be a faithful - representation of A on a separable Hilbert space. An element ϕ ∈ S0(X, Y ; A) will be called a Schur θ-multiplier if the map Sϕθ : K ⊗ θ(A) → K ⊗ θ(A) can be extended to a weak* continuous map on B ¯⊗θ(A)′′. The notion of a Schur θ-multiplier will be used in the subsequent sections. 3. Herz-Schur multipliers and transference In this section, we introduce and study Herz-Schur multipliers of crossed products. We assume throughout that G is a locally compact group. Left Haar measure on G will be denoted by mG or m and integration with respect to mG along the variable s will be denoted by ds. Let λG : G → B(L2(G)) t ξ(s) = ξ(t−1s), ξ ∈ L2(G), be the left regular representation of G; thus, λG s, t ∈ G. We write C ∗ r (G) (resp. VN(G)) for the reduced group C-algebra (resp. the von Neumann algebra) of G, that is, for the closure in the norm topology (resp. in the weak topology) of λG(L1(G)). As customary, we let A(G) (resp. B(G), Bλ(G)) be the Fourier (resp. the Fourier-Stieltjes, the reduced Fourier-Stieltjes) algebra of G. We note the canonical identifications A(G)∗ = VN(G), C ∗(G)∗ = B(G) and C ∗ r (G)∗ = Bλ(G) [10]. Let A be a separable C-algebra. In this section, unless otherwise stated, H will denote the Hilbert space of the universal representation of A; we consider A as a C-subalgebra of B(H). Let α : G → Aut(A) be a continuous (with respect to point-norm topology) group homomorphism; thus, (A, G, α) is a C-dynamical system. The space L1(G, A) is a -algebra with respect to the product × given by (f × g)(t) =RG f (s)αs(g(s−1t))ds and the involution ∗ given by f ∗(s) = ∆(s)−1αs(f (s−1)∗). Let π : A → B(L2(G, H)) be the -representation defined by (π(a)ξ)(t) = αt−1(a)(ξ(t)), t ∈ G, and λ : G → B(L2(G, H)) be the (continuous) unitary representation given by (λtξ)(s) = ξ(t−1s), s, t ∈ G. Note that π(αt(a)) = λtπ(a)λ∗ t , t ∈ G; 14 A. MCKEE, I. TODOROV, AND L. TUROWSKA thus, the pair (π, λ) is a covariant representation of (A, G, α) and hence gives rise to a -representation π ⋊ λ : L1(G, A) → B(L2(G, H)) given by (π ⋊ λ)(f ) =ZG π(f (s))λsds, f ∈ L1(G, A). The reduced crossed product A ⋊α,r G of A by α is, by definition, the closure of (π ⋊ λ)(L1(G, A)) in the operator norm of B(L2(G, H)) [27, 7.7.4]. We let A ⋊w∗ α,r G be the weak* closure of A ⋊α,r G. A bounded function F : G → B(A) will be called pointwise measurable if, for every a ∈ A, the map s → F (s)(a) is a weakly measurable function from G into A. Suppose that (ρ, τ ) is a covariant representation of the dynamical system (A, G, α) on the Hilbert space K. We say that F is (ρ, τ )-fiber continuous, if the map G → B(K), s → ρ(F (s)(a))τs, is weak* continuous for every a ∈ A. We will say that F is fiber continuous if F is (π, λ)-fiber continuous. Note that if F is bounded and point norm continuous then it is pointwise measurable and fiber continuous. We further say that F is almost (ρ, τ )-fiber continuous if, for every ω ∈ B(K)∗ and every a ∈ A, the function s → hρ(F (s)(a))τs, ωi coincides, up to a null set, with a continuous function. Almost (π, λ)-fiber continuous functions will be referred to simply as almost fiber continuous. For each f ∈ L1(G, A), let F · f ∈ L1(G, A) be the function given by (F · f )(s) = F (s)(f (s)), s ∈ G. It is easy to see that if F is pointwise measurable then F · f is weakly measurable and hence F · f ∈ L1(G, A) for every f ∈ L1(G, A); in fact, kF · f k1 ≤ kF k∞kf k1, where kF k∞ = sups∈G kF (s)k. Definition 3.1. A pointwise measurable function F : G → CB(A) will be called a Herz-Schur (A, G, α)-multiplier if the map SF : (π ⋊ λ)(L1(G, A)) → (π ⋊ λ)(L1(G, A)) given by is completely bounded. SF ((π ⋊ λ)(f )) = (π ⋊ λ)(F · f ) We denote by S(A, G, α) the set of all Herz-Schur (A, G, α)-multipliers. If F ∈ S(A, G, α) then the map SF extends to a completely bounded map on A ⋊r,α G. This (unique) extension will be denoted again by SF . We let kF km = kSF kcb. Remark 3.2. (i) If F1, F2 ∈ S(A, G, α), letting F1 + F2 : G → CB(A) and F1F2 : G → CB(A) be given by (F1 + F2)(s) = F1(s) + F2(s) and (F1F2)(s) = F1(s) ◦ F2(s), we see that SF1+F2 = SF1 + SF2 and SF1F2 = HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 15 SF1SF2. Thus, S(A, G, α) is an algebra with respect to the operations just defined. (ii) Recall that a bounded continuous function u : G → C is called a completely bounded (or Herz-Schur ) multiplier [7] of the Fourier algebra A(G) of G if uv ∈ A(G) for every v ∈ A(G), and the map mu : v → uv on A(G) is completely bounded. The space of all Herz-Schur multipliers of A(G) will be denoted as usual by M cbA(G). If u ∈ M cbA(G) then the dual Su of mu is a completely bounded (and weak* continuous) linear map on the von Neumann algebra VN(G) of G, such that Su(λG t , t ∈ G. Moreover, Su leaves the reduced C-algebra C ∗ r (G) of G invariant, and t ) = u(t)λG Su(cid:18)ZG f (s)λsds(cid:19) =ZG u(s)f (s)λsds, f ∈ L1(G). The reduced crossed product of C by the (unique) action α of a locally compact group G on C coincides with C ∗ r (G). Identifying B(C) with C in the natural way, we have that a bounded continuous function u : G → C is a Herz-Schur (C, G, α)-multiplier if and only if u is a Herz-Schur multiplier. (iii) Suppose that θ : A → B(K) is a faithful -representation of A on a Hilbert space K. Let πθ : A → B(L2(G, K)) be given by (πθ(a)ξ)(t) = θ(αt−1(a))(ξ(t)), t ∈ G, while λθ : G → B(L2(G, K)) be given by (λθ t ξ)(s) = ξ(t−1s), s, t ∈ G. Then the pair (πθ, λθ) is a covariant representation of (A, G, α). Since A is assumed to be universally represented, up to a - isomorphism, K is a closed subspace of H that reduces A, πθ(a) is the restriction of π(a) to L2(G, K), while λθ s is the restriction of λs to L2(G, K). w∗ In the sequel, we let A⋊α,θG = (πθ ⋊λθ)(A⋊αG) and A⋊w∗ . By [27, Theorem 7.7.5], the closure of (πθ ⋊ λθ)(L1(G, A)) is -isomorphic to A ⋊r,α G and a pointwise measurable function F : G → CB(A) is a Herz-Schur (A, G, α)-multiplier if and only if the map α,θG = A ⋊α,θ G (22) Sθ F : (πθ ⋊ λθ)(f ) 7→ (πθ ⋊ λθ)(F · f ) is completely bounded. Thus, Herz-Schur (A, G, α)-multipliers can be de- fined starting with any faithful representation of A instead of its universal representation. In the case A = C, the maps on C ∗ r (G) associated with Herz-Schur mul- tipliers automatically have a weak* continuous extension to (completely bounded) maps on the weak* closure VN(G) of C ∗ r (G). Such extension is not ensured to exist in the general case -- this motivates the following definition. Definition 3.3. Let A be a separable C-algebra, K be a Hilbert space and θ : A → B(K) be a faithful -representation. A function F : G → CB(A) will be called a θ-multiplier if the map Φθ F : πθ(a)λθ t 7→ πθ(F (t)(a))λθ t , t ∈ G, a ∈ A, has an extension to a bounded weak* continuous map on A ⋊w∗ α,θ G. 16 A. MCKEE, I. TODOROV, AND L. TUROWSKA A θ-multiplier F will be called a Herz-Schur θ-multiplier if the extension of Φθ F to A ⋊w∗ α,θ G is completely bounded. We note that, in Definition 3.3, we do not require the pointwise measur- ability of the function F . The weak* continuous extension of the map Φθ F therein will still be denoted by the same symbol. Remark 3.4. Let A be a separable C-algebra, K be a Hilbert space and θ : A → B(K) be a faithful -representation. Suppose that F : G → B(A) is a bounded map and Φ : A ⋊w∗ α,θ G is a bounded weak* continuous map such that, for almost all t ∈ G, α,θ G → A ⋊w∗ Φ(πθ(a)λθ t ) = πθ(F (t)(a))λθ t , a ∈ A. Then, for any ω ∈ B(L2(G, K))∗ and f ∈ L1(G, A), the function s 7→ hπθ(F (s)(f (s)))λθ s, ωi is measurable, and Φ((πθ ⋊ λθ)(f )) = (πθ ⋊ λθ)(F · f ), f ∈ L1(G, A). Proof. Let ω ∈ B(L2(G, K))∗ and f ∈ L1(G, A). Since the function s 7→ hπθ(f (s))λθ s, ωi. We have s, Φ∗(ω)i is measurable so is s 7→ hπθ(F (s)(f (s)))λθ (cid:28)Φ(cid:18)Z πθ(f (s))λθ sds, Φ∗(ω)(cid:29) s, Φ∗(ω)ids sds(cid:19) , ω(cid:29) = (cid:28)Z πθ(f (s))λθ = Z hπθ(f (s))λθ = Z hΦ(πθ(f (s))λθ = Z hπθ(F (s)(f (s)))λθ = (cid:28)Z πθ(F (s)(f (s)))λθ = h(πθ ⋊ λθ)(F · f ), ωi. s), ωids s, ωids sds, ω(cid:29) The claim follows. (cid:3) Lemma 3.5. Let θ be a faithful -representation of A on a Hilbert space K. Let F : G → CB(A) be a pointwise measurable map for which there exists C > 0 such that f ∈ L1(G, A). k(πθ ⋊ λθ)(F · f )k ≤ Ck(πθ ⋊ λθ)(f )k, (23) For ω ∈ B(L2(G, K))∗, let gω(s) = hπθ(F (s)(a))λθ coincides with an element of Bλ(G) up to a null set. Proof. Let πθ be the -representation of A on L2(G, K) ⊗ L2(G) given by πθ(a) = πθ(a) ⊗ IL2(G), a ∈ A. It is easy to see that the pair (πθ, λθ ⊗ λG) is a covariant representation. We first establish the following: Claim. The representation πθ ⋊(λθ ⊗λG) is unitarily equivalent to a direct sum of copies of the representation πθ ⋊ λθ. s, ωi, s ∈ G. Then gω HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 17 Proof of the Claim. Let U1 : L2(G)⊗L2(G) → L2(G, L2(G)) be the unitary operator given by U1(ξ ⊗ η)(s) = ξ(s)λG s−1η, ξ, η ∈ L2(G). Let (ηi)i∈I be an orthonormal basis for L2(G) and let L2(G)I be the direct sum of I-copies of L2(G). Let U2 : L2(G, L2(G)) → L2(G)I be the operator given by U2f = (gi)i∈I , where gi(s) = hf (s), ηii, s ∈ G. It is easy to see that U2 is a unitary operator. Set U = U2U1; thus, U is a unitary operator from L2(G) ⊗ L2(G) onto L2(G)I . For an operator T on L2(G), we shall denote by T ∞ its ampliation on L2(G)I , given by T ∞(ξi)i∈I = (T ξi)i∈I . In what follows we will use some natural identifications: of K⊗L2(G)⊗L2(G) with L2(G, K) ⊗ L2(G), of L2(G, K)I with L2(G, K I ), and of K ⊗ L2(G)I with L2(G, K)I , the latter being given by ξ ⊗ (ξi)∈I → (ξi)i∈I , where ξi(s) = ξi(s)ξ Let ξ1 ∈ K, ξ2, ξ3 ∈ L2(G), a ∈ A and t ∈ G. Let ζi be the function given by ζi(s) = ξ2(s)hξ3, λG s ηii, s ∈ G. For almost all s ∈ G we have t ξ2 ⊗ λG t ξ3)(s) (IK ⊗ U )πθ(a)(ξ1 ⊗ λG = (θ(αs−1(a))ξ1(λG = (θ(αs−1(a))ξ1(λG = (θ(αs−1(a))ξ1λG = πθ(a)∞(λθ = πθ(a)∞((λθ s−1 λG t ξ2)(s)hλG t ξ2)(s)hξ3, λG t ζi(s))i∈I t (ξ1 ⊗ ζi)(s))i∈I t ⊗ IL2(G))(IK ⊗ U )ξ1 ⊗ ξ2 ⊗ ξ3))(s). t ξ3, ηii)i∈I t−1sηii)i∈I The calculations imply (IK ⊗ U )(πθ ⋊ (λθ ⊗ λG))(f ) = (πθ ⋊ λθ(f ))∞(IK ⊗ U ), f ∈ L1(G, A), and the Claim is proved. Fix ω ∈ B(L2(G, H))∗ and write g = gω. Let v ∈ Bλ(G) and w be the linear functional on {λG(f ) : f ∈ L1(G)} given by w(λG(f )) =ZG f (s)g(s)v(s)ds. Let f ∈ L1(G), a ∈ A and f (s) = f (s)a, s ∈ G; clearly, f ∈ L1(G, A). Fix ξ, η ∈ L2(G) and let ωξ,η ∈ B(L2(G))∗ be the vector functional given by 18 A. MCKEE, I. TODOROV, AND L. TUROWSKA ωξ,η(T ) = hT ξ, ηi, T ∈ B(L2(G)). Then hλG(f g)ξ, ηi = ZG = (cid:28)ZG = (cid:28)ZG s )ds, ω ⊗ ωξ,η(cid:29) s )ds, ω ⊗ ωξ,η(cid:29) f (s)hπθ(F (s)(a))λθ s, ωihλG s ξ, ηids (πθ(F (s)( f (s))) ⊗ IL2(G))(λθ s ⊗ λG (πθ(F (s)( f (s))))(λθ s ⊗ λG = hπθ ⋊ (λθ ⊗ λG)(F · f ), ω ⊗ ωξ,ηi. Using the Claim, we have kλG(f g)k ≤ kωkkπθ ⋊ (λθ ⊗ λG)(F · f )k = kωkk(πθ ⋊ λθ)(F · f )k. sds, using (23), we have kw(λG(f )k ≤ kvkBλ(G)kλG(f g)k ≤ CkωkkvkBλ(G)k(πθ ⋊ λθ)( f )k As (πθ ⋊ λθ)(F · f ) =R f (s)πθ(F (s)(a))λθ ≤ CkωkkvkBλ(G)(cid:13)(cid:13)(cid:13)(cid:13) ≤ CkωkkvkBλ(G)kπθ(a)k(cid:13)(cid:13)(cid:13)(cid:13) sds(cid:13)(cid:13)(cid:13)(cid:13) πθ(a)Z f (s)λθ sds(cid:13)(cid:13)(cid:13)(cid:13) Z f (s)λθ ≤ CkωkkvkBλ(G)kπθ(a)kkλθ(f )k = CkωkkvkBλ(G)kπθ(a)kkλG(f )k. Hence, there exists z ∈ Bλ(G) such that ω(λG(f )) = R f (s)z(s)ds, f ∈ L1(G). It follows that z = vg almost everywhere. As this holds for any v ∈ Bλ(G), we have that g is almost everywhere equal to a function from Bλ(G) (see [7, Proposition 1.2]). (cid:3) Remark 3.6. For ω ∈ B(L2(G, K))∗, let gω(s) = hπθ(F (s)(a))λθ s, ωi and let bω ∈ Bλ(G) be such that bω = gω almost everywhere. If G is second countable and K is separable, then under the assumption of the previous lemma there exists a null subset N ⊆ G such that gω(t) = bω(t) for any t ∈ G\N and ω ∈ B(L2(G, K))∗. Indeed, in this case we have that B(L2(G, K)∗ is separable. Let {ωn : n ∈ N} be a dense subset of B(L2(G, K)∗ and a ∈ A. Let Nn ⊆ G be a null set such that gωn(s) = bωn(s) for all s ∈ G \ Nn, and N = ∪n∈NNn. Clearly, N is a null set. If ω ∈ B(L2(G, K)∗, let {ωn(k)}k be a subsequence converging to ω in norm. Then {bωn(k)}k is a Cauchy sequence of bounded continuous functions: letting C = kωk sups∈G kF (s)k, given ε > 0 there exists L ∈ N such that for any l, k > L, we have bωn(k)(t) − bωn(l)(t) = gωn(k)(t) − gωn(l)(t) ≤ Ckωn(k) − ωn(l)k ≤ Cε, whenever t ∈ G \ N . As bωn(k) is continuous, bωn(k)(t) − bωn(l)(t) < Cε for every t ∈ G. Thus, the sequence {bωn(k)}k converges to a continuous function, say b. On the other hand, bωn(k)(t) → gω(t) whenever t ∈ G \ N . HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 19 Therefore gω(t) = b(t) for t ∈ G \ N . As bω and b are continuous, and bω = gω almost everywhere, we have b = bω, giving the statement. For the rest of the section we will assume that G is a second countable locally compact group. In this case, the measure space (G, m) is standard. If t ∈ G, let us call a Dirac family at t a net (fU )U ⊆ L1(G) consisting of non-negative functions, indexed by the directed set of all open neighbour- hoods of t with compact closure, with supp fU ⊆ U and kfU k1 = 1. Lemma 3.7. Let (A, G, α) be a C-dynamical system, F be a Herz-Schur (A, G, α)-multiplier, and θ be a faithful -representation of A on a separable Hilbert space K. Then there exists a null set N ⊆ G such that if t ∈ G \ N and (fU )U is a Dirac family at t then F ((πθ ⋊ λθ)(fU ⊗ a)) →U πθ(F (t)(a))λθ Sθ t in the weak* topology, for every a ∈ A. Proof. Let a ∈ A. By Lemma 3.5 and Remark 3.6 there exists a null set N ⊆ G such that for any ξ, η ∈ L2(G, H) there exists a continuous function bξ,η : G → C such that hπθ(F (s)(a))λθ sξ, ηi = bξ,η(s) for all s ∈ G \ N. Fix ξ, η ∈ L2(G, H). For t ∈ G \ N , set CU = sup s∈U bξ,η(s) − bξ,η(t). Since bξ,η is continuous, CU →U 0. We have hSθ F ((πθ ⋊ λθ)(fU ⊗ a))ξ, ηi − hπθ(F (t)(a))λθ t ξ, ηi(cid:12)(cid:12)(cid:12) sξ, ηids −Z hπθ(F (t)(a))fU (s)λθ (cid:12)(cid:12)(cid:12) t ξ, ηids(cid:12)(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)(cid:12) Z hπθ(F (s)(a))fU (s)λθ ≤ Z fU (s)hπθ(F (s)(a))λθ = Z fU (s)bξ,η(s) − bξ,η(t)ds ≤ CUZ fU (s)ds = CU −→ U 0. sξ, ηi − hπθ(F (t)(a))λθ t ξ, ηids The statement follows from the fact that the weak operator topology and the weak* topology coincide on bounded sets. (cid:3) If ϕ : G × G → CB(A) is a bounded pointwise measurable function, let T (ϕ) : G × G → CB(A) be the function given by T (ϕ)(s, t)(a) = αt(ϕ(s, t)(αt−1 (a))), a ∈ A. It is easy to see that, for each a ∈ A, the function (s, t) → T (ϕ)(s, t)(a) from G × G into A is bounded and weakly measurable. The inverse T −1 of T is given by T −1(ϕ)(s, t)(a) = αt−1(ϕ(s, t)(αt(a))), a ∈ A. For a map F : G → CB(A), let N (F ) : G × G → CB(A) be the function given by N (F )(s, t) = F (ts−1), 20 and A. MCKEE, I. TODOROV, AND L. TUROWSKA N (F ) = T −1(N (F )); thus, N (F ) : G × G → CB(A) is the function given by N (F )(s, t)(a) = αt−1(F (ts−1)(αt(a))), a ∈ A, s, t ∈ G. Note that if F is pointwise measurable then so is N (F ). The next theorem is a dynamical system version of the well-known de- scription of Herz-Schur multipliers in terms of Schur multipliers [5]. Recall that, given a map ϕ : X × Y → CB(A) and a faithful -representation θ of A, we let ϕθ : X × Y → CB(θ(A)) be the map given by ϕθ(x, y)(θ(a)) = θ(ϕ(x, y)(a)), a ∈ A. Note that, if ϕ is pointwise measurable then so is ϕθ. Theorem 3.8. Let (A, G, α) be a C-dynamical system and F : G → CB(A) be a pointwise measurable map. The following are equivalent: (i) F is a Herz-Schur (A, G, α)-multiplier; (ii) N (F ) is a Schur A-multiplier. Moreover, if (i) holds then kF km = kN (F )kS. Proof. (i)⇒(ii) Suppose that F is a Herz-Schur (A, G, α)-multiplier and let θ be a faithful -representation of A on a separable Hilbert space H ′. By the Haagerup-Paulsen-Wittstock Theorem, there exist a separable Hilbert space K, a -representation ρ : A ⋊α,θ G → B(K) and operators V, W : L2(G, H ′) → K such that (24) F (T ) = W ∗ρ(T )V, Sθ T ∈ A ⋊α,θ G, and kSF kcb = kV kkW k. Let A ⋊α G be the full crossed product of A by α, q : A ⋊α G → A ⋊α,θ G be the quotient map and ρ = ρ ◦ q; thus, ρ : A ⋊α G → B(K) is a -representation. Let ρA : A → B(K) be a -representation and ρG : G → B(K) be a strongly continuous unitary representation such that ρ = ρA ⋊ ρG. For f ∈ L1(G, A), we have ρ(f ) =Z ρA(f (s))ρG(s)ds. Setting T = (πθ ⋊ λθ)(f ) in equation (24), we have (25) Z πθ(F (s)(f (s)))λθ sds = W ∗(cid:18)Z ρA(f (s))ρG(s)ds(cid:19) V. Standard arguments show that, if a ∈ A and (fU )U is a Dirac family at the point t ∈ G, then (26) Z ρA((fU ⊗ a)(s))ρG(s)ds −→ U ρA(a)ρG(t) in the weak operator topology. Taking f = fU ⊗ a in (25) and using Lemma 3.7, we obtain a null set N such that (27) πθ(F (t)(a))λθ t = W ∗ρA(a)ρG(t)V, for all t ∈ G \ N. HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 21 For s and t in G, let α(s), β(t) ∈ B(L2(G, H ′), K) be given by α(s) = ρG(s−1)V λθ s, β(t) = ρG(t−1)W λθ t ; then for every ξ ∈ L2(G, H ′), the functions s → α(s)ξ and s → β(s)ξ are weakly continuous and hence α and β belong to L∞(G, B(L2(G, H ′), K)). Using (27), for all (s, t) ∈ G × G with ts−1 ∈ G \ N , we obtain β(t)∗ρA(a)α(s) = λθ = λθ = λθ = πθ(αt−1(F (ts−1)(αt(a))) = πθ(N (F )(s, t)(a)). t−1 W ∗ρG(t)ρA(a)ρG(s−1)V λθ t−1 W ∗ρA(αt(a))ρG(ts−1)V λθ s t−1 πθ(F (ts−1)(αt(a)))λθ ts−1 λθ s s As {(s, t) : ts−1 ∈ N } is a null set for the product measure m × m, by Theorem 2.6, N (F ) is a Schur A-multiplier. Moreover, esssup kα(s)k = kV k and esssup kβ(t)k = kW k s∈G t∈G and hence (28) kN (F )kS ≤ kV kkW k = kF km. (ii)⇒(i) Let θ : A → B(H ′) be a faithful -representation, where H ′ is a def = N (F )θ is a Schur A-multiplier. separable Hilbert space. Suppose that ϕ Fix f ∈ Cc(G, A). For ξ ∈ L2(G, H ′), for almost all t ∈ G we have (πθ ⋊ λθ)(f )ξ(t) = Z πθ(f (s))λθ sξ(t)ds sξ(t))ds = Z θ(αt−1(f (s)))(λθ = Z θ(αt−1(f (s)))(ξ(s−1t))ds = Z ∆(r)−1θ(αt−1(f (tr−1)))(ξ(r))dr. Fix a compact set K ⊆ G. Then the function (29) hK : (t, r) → χK×K(t, r)∆(r)−1θ(αt−1(f (tr−1))) belongs to L2(G × G, θ(A)). Note that (30) We have ThK = (MχK ⊗ IH ′)(πθ ⋊ λθ)(f )(MχK ⊗ IH ′). ϕ · hK(t, r) = χK×K(t, r)∆(r)−1θ(αt−1(F (tr−1)(αt(αt−1(f (tr−1)))) = χK×K(t, r)∆(r)−1θ(αt−1(F (tr−1)(f (tr−1)))). Let ξ, η ∈ L2(G, H ′) have compact support and (Kn)∞ sequence of compact sets such that G = ∪∞ n=1 be an increasing n=1Kn (such a sequence exists 22 A. MCKEE, I. TODOROV, AND L. TUROWSKA since G is second countable and hence σ-compact). Then (31) hSϕ(ThKn )ξ, ηi =ZKn×Kn On the other hand, ∆(r)−1hθ(αt−1 (F (tr−1)(f (tr−1))))ξ(r), η(t)idrdt. ∆(r)−1hθ(αt−1(F (tr−1)(f (tr−1))))ξ(r), η(t)i ≤ ∆(r)−1kF k∞kf (tr−1)kkξ(r)kkη(t)k, while ∆(r)−1kf (tr−1)kkξ(r)kkη(t)kdrdt = hf ′ ∗ ξ′, η′i, ZG×G where f ′, ξ′, η′ : G → R are the functions given by f ′(s) = kf (s)k, ξ′(s) = kξ(s)k and η′(s) = kη(s)k, s ∈ G. Thus, an application of the Lebesgue Dominated Convergence Theorem shows that the right hand side of (31) converges to Z ∆(r)−1hθ(αt−1(F (tr−1)(f (tr−1))))ξ(r), η(t)idrdt = h(πθ ⋊ λθ)(F ·f ))ξ, ηi; thus, (32) lim n→∞ hSϕ(ThKn )ξ, ηi = h(πθ ⋊ λθ)(F · f )ξ, ηi. By (30), ThKn →n→∞ (πθ ⋊ λθ)(f ) in the weak topology. It follows that h(πθ ⋊ λθ)(F · f ))ξ, ηi ≤ lim sup n∈N hSϕ(ThKn )ξ, ηi ≤ kϕkSkξkkηk lim sup n∈N kThKn k ≤ kϕkSkξkkηkk(πθ ⋊ λθ)(f )k. Thus, k(πθ ⋊ λθ)(F · f ))k ≤ kϕkSk(πθ ⋊ λθ)(f )k, and hence the map Sθ Sθ F is completely bounded and kSθ Herz-Schur multiplier and F is bounded. Similar arguments show that, in fact, F kcb ≤ kϕkS. By Remark 3.2 (iii), F is a kF km ≤ kN (F )kS. The last inequality and (28) show that kF km = kN (F )kS and the proof is complete. (cid:3) Remark 3.9. In the case A = C, Theorem 3.8 reduces to the classical transference theorem for Herz-Schur multipliers [5]: a continuous function u : G → C is a Herz-Schur multiplier of A(G) if and only if N (u) is a Schur multiplier on G × G. HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 23 Corollary 3.10. Let θ be a faithful -representation of A on a separable Hilbert space and F : G → CB(A) be a pointwise measurable function. The following are equivalent: (i) F is a Herz-Schur multiplier such that Sθ F can be extended to a weak continuous map on A ⋊w∗ α,θ G; (ii) there exists a weak* continuous completely bounded linear map Φ on A ⋊w∗ α,θ G such that for almost all t ∈ G (33) Φ(πθ(a)λθ t ) = πθ(F (t)(a))λθ t , a ∈ A. In particular, (i) holds true if N (F ) is a Schur θ-multiplier. Proof. (i)⇒(ii) Suppose that F : G → CB(A) is a Herz-Schur multiplier and let Φ be the (unique) weak* continuous extension of Sθ α,θ G. By Lemma 3.7, there exists a null set N ⊆ G such that, if (fU )U is a Dirac family at t ∈ G \ N then, for every a ∈ A, F to A ⋊w∗ Φ((πθ ⋊ λθ)(fU ⊗ a)) →U πθ(F (t)(a))λθ t , while, as can be easily checked, (πθ ⋊ λθ)(fU ⊗ a) →U πθ(a)λθ t (both limits are in the weak operator topology). It follows that (33) holds for all t ∈ G \ N . (ii)⇒(i) By Remark 3.4, if f ∈ L1(G, A), then Φ(cid:18)ZG πθ(f (t))λθ t dt(cid:19) =ZG πθ(F (t)(f (t)))λθ t dt, i.e. Φ is a weak* continuous extension of Sθ Remark 3.2 (iii) implies that F ∈ S(G, A, α). F . Since Φ is completely bounded, Suppose that the map SN (F )θ has a weak* continuous extension to a map on B(L2(G)) ¯⊗θ(A)′′. By the proof of Theorem 3.8, if G = ∪∞ n=1Kn, where Kn is an increasing sequence of compact subsets of G, f ∈ Cc(G, A) and ThK is the operator with the kernel hK given by (29), then ThKn → (πθ ⋊ λθ)(f ) in the weak* topology. As SN (F )θ has a weak* continuous extension, we have hSN (F )θ (ThKn )ξ, ηi → hSN (F )θ ((πθ ⋊ λθ)(f ))ξ, ηi, ξ, η ∈ L2(G, H ′). On the other hand, by (32), hSN (F )θ (ThKn )ξ, ηi → hSθ F (πθ ⋊ λθ(f ))ξ, ηi. Thus, Sθ weak* continuous extension to A ⋊w∗ F is the restriction of SN (F )θ to A ⋊α,θ G, and hence Sθ F possesses a (cid:3) α,θ G. Remark 3.11. We remark that if in Corollary 3.10 we assume also that F : G → CB(A) is (πθ, λθ)-fiber continuous then condition (ii) is equivalent 24 A. MCKEE, I. TODOROV, AND L. TUROWSKA to F being a Herz-Schur θ-multiplier. Therefore, in this case, F is a Herz- Schur θ-multiplier if and only if F is a Hez-Schur multiplier such that Sθ F possesses a weak-continuous extension to A ⋊w∗ α,θ G. Corollary 3.12. Let θ be a faithful -representation of A on a separable Hilbert space and F : G → CB(A) be a Herz-Schur θ-multiplier. Then supt∈G kF (t)kcb ≤ kF km. Proof. Immediate from Corollary 3.10 and the fact that λθ tries. s and πθ are isome- (cid:3) An equivariant representation of the dynamical system (A, G, α) on a Hilbert A-module N is a pair (ρ, τ ), where ρ : A → L(N ) is a -representation of A on N and τ is a homomorphism from G into the group I(N ) of all C-linear, invertible, bounded maps on N , which satisfy: (1) ρ(αs(a)) = τ (s)ρ(a)τ (s)−1 for all s ∈ G and a ∈ A; (2) αs(hξ, ηiA) = hτ (s)ξ, τ (s)ηiA, for all s ∈ G and ξ, η ∈ N ; (3) τ (s)(ξ · a) = (τ (s)ξ) · a, for all s ∈ G, ξ ∈ N and a ∈ A; (4) the map s 7→ τ (s)ξ is continuous for every ξ ∈ N . This definition was given in [2] for discrete twisted dynamical systems. An example of such an equivariant representation can be obtained as follows. Let H G A = {ξ : G → A : ξ(·)∗ξ(·) ∈ L1(G, A)}. We have that H G A is a Hilbert A-module with respect to the right action given by (ξ·a)(s) := ξ(s)a and the inner product given by hξ, ηiA =RG ξ(s)∗η(s)ds. The regular equivariant representation of (A, G, α) on H G given by A is the pair (ρ, τ ) (τ (t)ξ)(s) = αt(ξ(t−1s)). It is easy to check that (ρ, τ ) satisfies the conditions (1)-(4). ρ(a)ξ(h) = aξ(h), The following corollary was proved in [2, Theorem 4.8] for discrete dy- namical systems using different arguments. Corollary 3.13. Let (ρ, τ ) be an equivariant representation of (A, G, α) on a countably generated Hilbert A-module N , and let ξ, η ∈ N . Define F (t)(a) = hξ, ρ(a)τ (t)ηiA, t ∈ G, a ∈ A. Then F is a Herz-Schur (A, G, α)-multiplier. Proof. By Theorem 3.8, it suffices to show that N (F ) is a Schur A-multiplier. We have N (F )(s, t)(a) = αt−1 (F (ts−1)(αt(a))) = αt−1 (hξ, ρ(αt(a))τ (ts−1)ηiA) = hτ (t−1)ξ, τ (t)−1ρ(αt(a))τ (t)τ (s−1)ηiA = hτ (t−1)ξ, ρ(a)τ (s−1)ηiA As kτ (t)ξk2 = khτ (t)ξ, τ (t)ξiAk = kαt(hξ, ξiA)k = khξ, ξiAk = kξk2 HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS for all t ∈ G, the statement follows from Theorem 2.7. 25 (cid:3) We next identify the Schur multipliers of the form N (F ) as the "invariant" part of S0(G, G; A). Let ρ be the right regular representation of G on L2(G), i.e. (ρrξ)(s) = ∆(r)1/2ξ(sr), ξ ∈ L2(G), s, r ∈ G. Let αr = (Adρr) ⊗ αr, where, as usual, AdU is the map given by AdU (T ) = U T U ∗; we have that αr is a -automorphism of K(L2(G)) ⊗ A [16, Theorem 11.2.9]. Definition 3.14. A Schur A-multiplier ϕ : G × G → CB(A) will be called invariant if Sϕ commutes with αr for every r ∈ G. We denote by Sinv(G, G; A) the set of all invariant Schur A-multipliers. If w is a (possibly vector-valued) function defined on G × G, for r ∈ G, we let wr be the function given by wr(s, t) = w(sr, tr). Lemma 3.15. If k ∈ L2(G × G, A) then αr(Tk) = Tk, where k ∈ L2(G × G, A) is given by k(t, s) = ∆(r)αr(kr(t, s)), s, t ∈ G. Proof. First note that if k ∈ L2(G × G, A) then k ∈ L2(G × G, A) and (34) indeed, Let kkk2 = kkk2; kkk2 =Z ∆2(r)kαr(k(tr, sr))k2dtds =Z kk(t′, s′)k2dt′ds′. Θ, Θ′ : L2(G × G, A) → B(L2(G, H)) be the maps defined by Θ(k) = Tk, Θ′(k) = αr(Tk). By Lemma 2.1 and (34), Θ and Θ′ are continuous. Suppose that k ∈ L2(G×G, A) is given by k = h⊗a, where h ∈ L2(G×G) and a ∈ A. As ρ∗ rξ(s) = ∆(r)−1/2ξ(sr−1), ξ ∈ L2(G), s ∈ G, we have (ρrThρ∗ rξ)(s) = ∆(r)1/2(Thρ∗ r)ξ(sr) = ∆(r)1/2Z h(sr, x)(ρ∗ = Z h(sr, x)ξ(xr−1)dx =Z h(sr, yr)ξ(y)∆(r)dy, , where hr(t, s) = ∆(r)h(tr, sr), s, t ∈ G. Thus, r ξ)(x)dx that is, ρrThρ∗ r = Thr αr(Tk) = αr(Th ⊗ a) = Thr ⊗ αr(a), and so Θ′(k) = Θ(k). As h and a vary, the functions k span a dense subspace of L2(G × G, A), and hence Θ = Θ′ by continuity. (cid:3) In the proof of Theorem 3.18 below, we will need the following improve- ment of [41, Lemma 3.9]. 26 A. MCKEE, I. TODOROV, AND L. TUROWSKA Lemma 3.16. Let X be a separable Banach space and w : G × G → B(X ) be a bounded function, such that, for every a ∈ X , the function (s, t) → w(s, t)(a) is weakly measurable, and wr = w almost everywhere, for every r ∈ G. Then there exists a bounded function u : G → B(X ) such that, for every a ∈ X , the function s → u(s)(a) is measurable and, up to a null set, w = N (u). Proof. The map φ : G×G → G×G, given by φ(y, x) = (y, xy), is continuous (and hence measurable) and bijective, and Fubini's Theorem shows that it preserves null sets in both directions. By assumption, for all r ∈ G, we have that wr(s, x) = w(s, x) for almost all (s, x) ∈ G × G. Thus, wr(φ(s, x)) = w(φ(s, x)) for almost all (s, x) ∈ G × G, that is, w(sr, xsr) = w(s, xs) for almost all (x, s) ∈ G × G. We claim that w(sr, xsr) = w(s, xs) for almost all (x, s, r) ∈ G × G × G. In fact, let S ⊆ X be a countable dense subset. For every a ∈ S, we have kw(sr, xsr)(a) − w(s, xs)(a)kdxdsdr kw(sr, xsr)(a) − w(s, xs)(a)kdxds(cid:19) dr = 0. ZG×G×G = ZG(cid:18)ZG×G Thus, there exists a null set Na ⊆ G × G × G such that w(sr, xsr)(a) = w(s, xs)(a) for all (x, s, r) 6∈ Na. Let N = ∪a∈S Na. Then w(sr, xsr)(a) = w(s, xs)(a) for all (x, s, r) 6∈ N and all a ∈ S. Since w(sr, xsr) and w(s, xs) are bounded operators on X , we conclude that w(sr, xsr) = w(s, xs) for all (x, s, r) 6∈ N . Thus, there exists s0 ∈ G such that (35) w(s0r, xs0r) = w(s0, xs0), for almost all (x, r) ∈ G × G. For each x ∈ G, let u(x) = w(s0, xs0). Clearly, u : G → B(X ) is a bounded function such that, for every a ∈ X , the function x → u(x)(a) is weakly measurable. Now (35) implies that w(y, xy) = u(x) for almost all (x, y) ∈ G × G. Letting u : G × G → X be the map given by u(s, t) = u(t), we thus have that w(y, xy) = u(y, x) for almost all (x, y) ∈ G × G. It follows that w(φ−1(y, xy)) = u(φ−1(y, x)) for almost all (x, y) ∈ G × G, that is, w(y, x) = u(xy−1) for almost all (x, y) ∈ G × G. The proof is complete. (cid:3) Lemma 3.17. Let ϕ ∈ S0(G, G; A). The following are equivalent: (i) ϕ is an invariant Schur A-multiplier; (ii) T (ϕ)r = T (ϕ) almost everywhere, for every r ∈ G. Proof. Assume that A is faithfully represented on a separable Hilbert space. (i)⇒(ii) Let r ∈ G, a ∈ A and k ∈ L2(G × G). By Lemma 3.15, (Sϕ ◦ αr)(Tk ⊗ a) = Tk1, where k1 : G × G → A is the function given by k1(t, s) = ∆(r)k(tr, sr)ϕ(s, t)(αr(a)), s, t ∈ G, and (αr ◦ Sϕ)(Tk ⊗ a) = Tk2, where k2 : G × G → A is the function given by k2(t, s) = ∆(r)k(tr, sr)αr(ϕ(sr, tr)(a)), s, t ∈ G. HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 27 By Lemma 2.1, k1 = k2 almost everywhere, and hence ϕ(sr, tr)(a) = αr−1(ϕ(s, t)(αr(a))), for almost all (s, t) ∈ G × G. Thus, for every a ∈ A, T (ϕ)(sr, tr)(a) = αtr(ϕ(sr, tr)(αr−1t−1 (a))) = αtr(αr−1(ϕ(s, t)(αr(αr−1t−1(a)))) = αt(ϕ(s, t)(αt−1 (a))) = T (ϕ)(s, t)(a) for almost all (s, t) ∈ G × G. Since A is separable, we conclude that T (ϕ)(sr, tr) = T (ϕ)(s, t) for almost all (s, t) ∈ G × G. (ii)⇒(i) follows by reversing the steps in the previous paragraph and using the density in K(L2(G)) ⊗ A of the linear span of the operators of the form Tk ⊗ a, with k ∈ L2(G × G) and a ∈ A. (cid:3) Theorem 3.18. The map N is a linear isometry from S(A, G, α) onto Sinv(G, G; A). Proof. By Theorem 3.8, the map N is a linear isometry from S(A, G, α) into S0(G, G; A). By the definition of N , we have that T (N (F ))r = T (N (F )) almost everywhere for every r ∈ G and every F ∈ S(A, G, α). By Lemma 3.17, the image of N is in Sinv(G, G; A). It remains to show that N is surjective. To this end, let θ be a faithful -representation of A on a separable Hilbert space K; we identify A with its image θ(A) under θ and let ϕ ∈ Sinv(G, G; A). By Lemmas 3.17 and 3.16, there exists a bounded function F : G → B(A) such that N (F ) = T (ϕ) almost everywhere and such that, for every a ∈ A, the function s → F (s)(a), is weakly measurable. It follows that N (F ) = ϕ almost everywhere. Since ϕ(x, y) is completely bounded for all (x, y), we have that F (s) ∈ CB(A) for all s ∈ G. As ϕ is a Schur A-multiplier, it follows from the proof of Theorem 3.8 that the map (πθ ⋊ λθ)(f ) → (πθ ⋊ λθ)(F · f ) on (πθ ⋊ λθ)(L1(G, A)) is completely bounded. By Remark 3.2 (iii), F ∈ S(A, G, α). (cid:3) We now consider bounded, as opposed to completely bounded, multipliers, and use them to characterise Herz-Schur θ-multipliers in the spirit of [7] in Proposition 3.19 below. Let Γ be a a locally compact group. For a function F : G → CB(A), let F Γ : Γ × G → CB(A) be given by F Γ(x, s) = F (s), x ∈ Γ, s ∈ G. If α is an action of G on A, we let αΓ : Γ × G → Aut(A) be given by αΓ (x,s)(a) = αs(a), a ∈ A, x ∈ Γ, s ∈ G. We have the following characterisation of Herz-Schur θ-multipliers, similar to [7, Theorem 1.6]. Proposition 3.19. Let F : G → CB(A) and θ : A → B(K) be a faithful -representation. The following are equivalent: (i) F is a Herz-Schur θ-multiplier; (ii) for each locally compact group Γ, the function F Γ is a θ-multiplier; (iii) for Γ = SU (2), the function F Γ is a θ-multiplier. 28 A. MCKEE, I. TODOROV, AND L. TUROWSKA Proof. Let πΓ,θ : A → B(L2(Γ × G, K)) be the -representation given by (x−1,t−1)(a)(ξ(x, t)). Identifying the Hilbert space L2(Γ × πΓ,θ(a)ξ(x, t) = αΓ G, K) with L2(Γ) ⊗ L2(G, K) in the natural way, we see that On the other hand, πΓ,θ(a) = IL2(Γ) ⊗ πθ(a), a ∈ A. λθ (x,t) = λΓ x ⊗ λθ t , x ∈ Γ, t ∈ G. Suppose that f ∈ L1(Γ × G, A) has the form f (x, s) = g(x)h(s), where f ∈ L1(Γ) and h ∈ L1(G, A). We have ZΓ×G (36) Thus, (37) πΓ,θ(f (x, s))λθ (g(x)IL2(Γ) ⊗ πθ(f (s)))(λΓ x ⊗ λθ s)dxds (x,s)dxds = ZΓ×G = (cid:18)ZΓ g(x)λΓ x dx(cid:19) ⊗(cid:18)ZG πθ(f (s))λθ sds(cid:19) . A ⋊w∗ αΓ,θ (Γ × G) = VN(Γ) ¯⊗(A ⋊w∗ α,θ G). F on A ⋊w∗ (i)⇒(ii) Suppose that F : G → CB(A) is a Herz-Schur θ-multiplier. Since t ) = πθ(F (t)(a))λθ t , is com- F is a (completely) bounded the map Φθ F (πθ(a)λθ pletely bounded and weak continuous, id ⊗Φθ map on VN(Γ) ¯⊗(A ⋊w∗ α,θ G) (see e.g. [7, Lemma 1.5]). Moreover, α,θ G, given by Φθ (38) πθ,Γ(F Γ(a))λθ (x,s) = (id ⊗Φθ It follows that the map Φθ bounded weak* continuous map on A ⋊w∗ a θ-multiplier. F Γ : πθ,Γ(a)λθ F )(πθ,Γ(a)λθ (x,s)). (x,s) → πθ,Γ(F Γ(a))λθ (x,s) extends to a αΓ,θ (Γ × G); in other words, F Γ is (ii)⇒(iii) is trivial. (iii)⇒(i) We have that VN(SU (2)) ≡ ⊕n∈NMn, where Mn is the n by n matrix algebra. Hence VN(SU (2)) ¯⊗(A ⋊w∗ α,θ G) ≡ ⊕∞ n=1(cid:16)Mn ⊗ (A ⋊w∗ α,θ G)(cid:17) . F Γ is a bounded weak* continuous map on VN(SU (2)) ¯⊗(A ⋊w∗ As Φθ equations (37) and (38) now imply that k idMn ⊗Φθ hence Φθ F is completely bounded. α,θ G), F Γk for all n and (cid:3) F k ≤ kΦθ 4. Multipliers of the weak* crossed product In this section, we consider the weak* extendable Herz-Schur multipliers introduced in Definition 3.3, and characterise them as the commutator of the "scalar valued" multipliers described in Proposition 4.1 below. We fix, throughout the section, a C ∗-dynamical system (A, G, α). As before, A is assumed to be separable, while G is assumed to be second countable. If θ : A → B(K) is a faithful -representation, where K is a separable Hilbert HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 29 space, we let αθ : G → Aut(θ(A)) be given by αθ t (θ(a)) = θ(αt(a)), t ∈ G, a ∈ A. We call α a θ-action, if αθ t can be extended to a weak continuous automorphism (which we will denote in the same fashion) of θ(A)′′, such s(x) from G into θ(A)′′ is weak-continuous for each that the map s → αθ x ∈ θ(A)′′. Proposition 4.1. Let u : G → C be a bounded continuous function, and let Fu : G → CB(A) be given by Fu(t)(a) = u(t)a, a ∈ A, t ∈ G. The following are equivalent: (i) Fu is a Herz-Schur (A, G, α)-multiplier; (ii) u ∈ M cbA(G). Moreover, if (i) holds then Fu is a Herz-Schur θ-multiplier for every faith- ful representation θ of A on a separable Hilbert space. Proof. Set F = Fu. We have (39) N (F )(s, t)(a) = u(ts−1)a, a ∈ A, s, t ∈ G. We assume, without loss of generality, that A is a non-degenerate C- subalgebra of B(H), for a separable Hilbert space H. (i)⇒(ii) By Theorems 2.6 and 3.8, there exist a separable Hilbert space K, a non-degenerate -representation ρ : A → B(K) and elements V, W of L∞(G, B(H, K)) such that u(ts−1)a = W (t)∗ρ(a)V (s), for almost all s, t ∈ G and all a ∈ A. Let (ai)∞ ξ ∈ H and every i ∈ N, we have i=1 be a bounded approximate identity for A. Then, for a unit vector hu(ts−1)aiξ, ξi = hρ(ai)V (s)ξ, W (t)ξi, for almost all s, t ∈ G. Since A ⊆ B(H) is non-degenerate and ρ is a non-degenerate representation, passing to a limit along i, we obtain u(ts−1) = hV (s)ξ, W (t)ξi, for almost all s, t ∈ G. By [5], u ∈ M cbA(G). (ii)⇒(i) As G is second countable, by [5], there exist weakly measurable functions ξ, η : G → ℓ2 such that u(ts−1) = hξ(s), η(t)i, for almost all s, t ∈ G. Let ρ : A → B(H ∞) be the countable ampliation of the identity represen- tation of A. Write ξ(s) = (ξi(s))i∈N and η(t) = (ηi(t))i∈N, s, t ∈ G. Let V (s) : H → H ∞ (resp. W (t) : H → H ∞) be given by V (s) = (ξi(s)IH )i∈N (resp. W (t) = (ηi(t)IH )i∈N). Then V, W ∈ L∞(G, B(H, H ∞)) and ∞ W (t)∗ρ(a)V (s) = ξi(s)ηi(t)a = u(ts−1)a, Xi=1 for almost all s, t ∈ G and all a ∈ A. It follows by (39) and Theorems 2.6 and 3.8 that Fu is a Herz-Schur (A, G, α)-multiplier. 30 A. MCKEE, I. TODOROV, AND L. TUROWSKA Now suppose that u ∈ M cbA(G) and denote by Ψu the weak continuous completely bounded map on B(L2(G)) corresponding to the function u via classical transference [5] (see Remark 3.9). Let θ : A → B(K) be a faithful - representation of A, for some separable Hilbert space K. Note that N (F ) is a Schur θ-multiplier; indeed, we have that N (F )θ(s, t)(θ(a)) = u(ts−1)θ(a), a ∈ A, and hence SN (F )θ = ΨuK(L2(G)) ⊗idθ(A). It follows that SN (F )θ is the restriction to K(L2(G)) ⊗ θ(A) of the weak continuous map Ψu ⊗ idθ(A)′′. By Corollary 3.10 and Remark 3.11, Fu is a Herz-Schur θ-multiplier. (cid:3) In what follows we denote by Sθ arising from the previous proposition. u the weak* continuous map on A ⋊w∗ α,θ G It is well-known that an essentially bounded function on G that is in- variant under right translations agrees almost everywhere with a constant function. The next lemma is a dynamical system version of this fact. For a -representation θ : A → B(K) such that α is a θ-action, let ¯πθ be the -representation of θ(A)′′ on L2(G, K) given by (¯πθ(a)ξ)(s) = α−1 s (a)(ξ(s)), a ∈ θ(A)′′, ξ ∈ L2(G, K). Recall that if ρ : G → B(L2(G)) is the right regu- lar representation of G we write αr for the map Adρr ⊗ αr on K(L2(G)) ⊗ θ(A); we have that the map αr can be extended to a weak* continuous map B(L2(G) ¯⊗θ(A)′′, denoted in the same fashion. Lemma 4.2. Let (A, G, α) be a C ∗-dynamical system and θ : A → B(K) be a faithful -representation, where K is a separable Hilbert space, such that α is a θ-action. Then (40) (DG ¯⊗θ(A)′′) ∩ (A ⋊w∗ α,θ G) = ¯πθ(θ(A)′′). Proof. Suppose that D belongs to both L∞(G, θ(A)′′) and A ⋊w∗ α,θ G. By [23, Theorem 1.2, Chapter II], there exists a null set M ⊆ G such that αr(D) = D whenever r ∈ M c. However, αr(D) = Dr, where Dr ∈ L∞(G, θ(A)′′) is given by Dr(s) = αr(D(sr)), s ∈ G. Let D ∈ L∞(G, θ(A)′′) be defined by D(s) = αs(D(s)), s ∈ G. For every r ∈ M c, D(sr) = αsr(D(sr)) = αs(αr(D(sr)) = αs(D(s)) = D(s), for almost all s. As in the proof of Lemma 3.16, there exists s0 ∈ G such that D(s0r) = D(s0) for almost all r ∈ G. Thus, there exists a ∈ θ(A)′′ such that D(t) = a for almost all t ∈ G, and hence D(t) = αt−1(a) for almost all t ∈ G; in other words, D = ¯πθ(a). We thus showed that the intersection on the left hand side of (40) is contained in ¯πθ(θ(A)′′). The converse inclusion is trivial. (cid:3) Let K be a Hilbert space. If ω ∈ B(H)∗, we let Lω be the (unique) weak continuous linear map from B(K⊗H) into B(K) such that Lω(b⊗a) = ω(a)b, a ∈ B(H), b ∈ B(K). Recall that a weak* closed subspace U ⊆ B(H) is said to have property Sσ [19] if V ¯⊗U = {T ∈ B(K) ¯⊗U : Lω(T ) ∈ V for all ω ∈ B(H)∗}, for every weak* closed subspace V ⊆ B(K). HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 31 For the proof of the next theorem, we recall that an operator T ∈ B(L2(G)) is said to be supported on a measurable subset E ⊆ G × G if Mχβ T Mχα = 0 whenever α, β ⊆ G are measurable sets with (α × β) ∩ E = ∅. It is easy to see that the space of operators supported on the set {(s, ts) : s ∈ G} coincides with DGλG t . Theorem 4.3. Let (A, G, α) be a C-dynamical system and θ be a faithful -representation of A on a separable Hilbert space K such that α is a θ- action and θ(A)′′ possesses property Sσ. Let Φ be a completely bounded weak* continuous map on A ⋊w∗ α,θ G. The following are equivalent: u = Sθ (i) ΦSθ (ii) For each t ∈ G, there exists completely bounded map Ft : θ(A)′′ → uΦ for all u ∈ M cbA(G); θ(A)′′ such that Φ(¯πθ(a)λθ t ) = ¯πθ(Ft(a))λθ t , a ∈ θ(A)′′. Proof. (i)⇒(ii) Given u ∈ M cbA(G), let Ψu be the weak* continuous com- pletely bounded map on B(L2(G)) corresponding to the Schur multiplier N (u) (see Remark 3.9). We claim that (41) (Lω ◦ Sθ u)(T ) = (Ψu ◦ Lω)(T ), T ∈ A ⋊w∗ α,θ G, ω ∈ B(K)∗. For S ∈ B(K) and T ∈ B(L2(G)), we have Lω((S ⊗ T )(IK ⊗ λG by weak* continuity, we obtain t )) = ω(S)T λG t = Lω(S ⊗ T )λG t ; Lω(R(IK ⊗ λG (42) On the other hand, if t ∈ G then λθ (43) t )) = Lω(R)λG t , R ∈ B(L2(G, K)). t = IK ⊗ λG t , and it follows that Lω(Sθ u(πθ(a)λθ t )) = u(t)Lω(πθ(a)λθ t ) = u(t)Lω(πθ(a))λG t , a ∈ A, t ∈ G. Since πθ(a) ∈ L∞(G, θ(A)′′), we have that Lω(πθ(a)) ∈ DG, for every ω ∈ B(K)∗. Since Ψu is a DG-bimodule map, using equation (42) we obtain (44) Ψu(Lω(πθ(a)λθ t )) = Ψu(Lω(πθ(a))λG t ) = Lω(πθ(a))Ψu(λG t ) = u(t)Lω(πθ(a))λG t . Equation (41) follows from the weak* continuity of Lω and Sθ sition 4.1), after comparing (43) and (44). u (see Propo- Let a ∈ θ(A)′′, t ∈ G and T = ¯πθ(a)λθ t . Set J = {u ∈ M cbA(G) : u(t) = 1}. If u ∈ J then Sθ with Sθ u, we have u(T ) = T and hence, by (41) and the fact that Φ commutes Ψu(Lω(Φ(T ))) = Lω(Φ(T )). Thus, for every u ∈ J, the operator Lω(Φ(T )) is u-harmonic in the sense of [24]. It follows from [1, Corollary 3.7] that Lω(Φ(T )) is supported on the set {(x, y) ∈ G × G : yx−1 ∈ Z}, where Z = {s ∈ G : u(s) = 1, for all u ∈ J}. By the regularity of A(G) and the fact that A(G) ⊆ M cbA(G), we have that 32 A. MCKEE, I. TODOROV, AND L. TUROWSKA Z = {t} and hence, by (42) and the paragraph before the statement of the theorem, Lω(Φ(T )λθ t−1) = Lω(Φ(T ))λG t−1 ∈ DG. Since this holds for every ω ∈ B(K)∗ and θ(A)′′ is assumed to possess property Sσ, we conclude that Φ(T )λθ t−1 ∈ DG ¯⊗θ(A)′′. On the other hand, Φ(T )λθ ¯πθ(at) for some at ∈ θ(A)′′, and hence Φ(T ) = ¯πθ(at))λθ at, we have Φ(T ) = ¯πθ(Ft(a)))λθ bounded since Φ is so. t−1 ∈ A ⋊w∗ α,θ G. By Lemma 4.2, Φ(T )λθ t−1 = t . Writing Ft(a) = t . The map Ft is linear and completely (ii)⇒(i) For t ∈ G and a ∈ θ(A)′′, we have Φ(Sθ u(¯πθ(a)λθ t )) = u(t)Φ(¯πθ(a)λθ t ) = u(t)¯πθ(Ft(a))λθ t = Sθ u(Φ(¯πθ(a)λθ t )). The commutation relations now follow by linearity and weak* continuity. (cid:3) 5. Two classes of multipliers In this section, we describe two special classes of Herz-Schur multipliers and relate them to maps that have been studied previously. 5.1. Multipliers from the Haagerup tensor product. Multipliers of the type studied in this subsection have been considered in the case of a discrete group in [2]. Let A be a separable non-degenerate C-subalgebra of B(H), where H is a separable Hilbert space, and C∞(A) be the col- umn operator space over A; thus, the elements of C∞(A) are the sequences i ai converges in norm. Recall [4] that the Haagerup tensor product A ⊗h A consists, by definition, of all sums i )i∈N ∈ C∞(A). Let β : X → C∞(A) and γ : Y → C∞(A) be bounded weakly measurable functions. Write β(x) = (βi(x))i∈N, x ∈ X, and γ(y) = (γi(y))i∈N, y ∈ Y . It is clear that, in particular, βi ∈ L∞(X, A) and γi ∈ L∞(Y, A) for each i ∈ N. (ai)i∈N ⊆ A such that the series P∞ u = P∞ i=1 bi ⊗ ai, where (ai)i∈N, (b∗ i=1 a∗ Let ϕβ,γ : X × Y → A ⊗h A be given by (45) ϕβ,γ(x, y) = ∞ Xi=1 γi(y)∗ ⊗ βi(x), (x, y) ∈ X × Y. P∞ by Φu(a) = P∞ Note that A ⊗h A embeds canonically into CB(A); for an element u = i=1 bi ⊗ ai of A ⊗h A, the corresponding map Φu : A → A is given i=1 biaai, a ∈ A. We thus view ϕβ,γ(x, y) as a completely bounded map on A. It is easy to see that the partial sums of (45) define weakly measurable functions, and since the convergence of the series is in norm, [43, Lemma B.17] shows that the function ϕβ,γ is weakly measurable. In particular, ϕβ,γ is pointwise measurable. HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 33 Proposition 5.1. Let β : X → C∞(A) and γ : Y → C∞(A) be bounded weakly measurable functions. Then ϕβ,γ is a Schur id-multiplier. Moreover, ∞ (46) Sϕβ,γ (T ) = γ∗ i T βi, T ∈ K ⊗ A, where the series converges in norm. Xi=1 Proof. First note that β∗ i (x)βi(x)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) and that a similar estimate holds for P∞ i . It follows that the series on the right hand side of (46) converges in norm. Let k ∈ L2(Y × X, A), ξ ∈ L2(X, H) and η ∈ L2(Y, H). Then ∞ = essupx∈Xkβ(x)k2, i=1 γiγ∗ ∞ Xi=1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ∞ Xi=1 γ∗ ∞ β∗ i βi(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) = essupx∈X(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xi=1 i Tkβiξ, η+ = ZX×Y = ZX×Y hϕβ,γ(x, y)(k(y, x))ξ(x), η(y)idxdy hk(y, x)βi(x)ξ(x), γi(y)η(y)idxdy Xi=1 = hTϕβ,γ ·kξ, ηi. It follows that ϕβ,γ is a Schur A-multiplier. Identity (46) now follows by boundedness. Since the map expressed by the right hand side of (46) is weak* extendible, we conclude that ϕβ,γ is in fact a Schur id-multiplier. (cid:3) Recall that a subset E ⊆ G × G is called marginally null if there exists a null set M ⊆ G such that E ⊆ (M × G) ∪ (G × M ). Proposition 5.2. Let β : X → C∞(A) and γ : Y → C∞(A) be bounded weakly measurable functions. The following are equivalent: (i) there exists F ∈ S(A, G, α) such that Sid F coincides with the restriction of Sϕβ,γ to A ⋊α,id G; (ii) for every a ∈ A, the function ϕa : G × G → A given by ∞ ϕa(s, t) = αt(γi(t))∗aαt(βi(s)), s, t ∈ G, Xi=1 has the property that, for every r ∈ G, ϕa(sr, tr) = ϕa(s, t) for almost all (s, t). Moreover, if (i) holds then the map Sid F has an extension to a bounded weak* continuous map on A ⋊w∗ α,id G. Proof. (i)⇒(ii) By Proposition 5.1, the map Sϕβ,γ has a weak* continuous extension to a completely bounded map on B(L2(G)) ¯⊗A′′. Since Sid F is the restriction of Sϕβ,γ , it possesses a weak* continuous extension to a completely bounded map on A ⋊w∗ α,id G. 34 A. MCKEE, I. TODOROV, AND L. TUROWSKA Let a ∈ A and s ∈ G. Note that, if βs i (t) = s . By Corollary 3.10, for almost all s ∈ G, we i ∈ L∞(G, A) is given by βs βi(s−1t), then λid have s βi = βs i λid πid(F (s)(a)) = ∞ Xi=1 γ∗ i πid(a)λid s βi! (λid s )∗ = γ∗ i πid(a)βs i , a ∈ A. ∞ Xi=1 Therefore, if ξ ∈ L2(G, H) then, for almost all s, t ∈ G, we have αt−1(F (s)(a))(ξ(t)) = ∞ Xi=1 Xi=1 = ∞ γ∗ i πid(a)βs i ξ! (t) γi(t)∗αt−1(a)βi(s−1t)(ξ(t)). A standard argument using the separability of H now shows that, for almost all s, t ∈ G, we have αt−1(F (s)(a)) = ∞ Xi=1 γi(t)∗αt−1(a)βi(s−1t), and, since the series on the right hand side converges is norm, (47) ∞ Xi=1 αt(γ∗ i (t))aαt(βi(s−1t)) = F (s)(a), i.e. ϕa(t, s−1t) = F (s)(a) for almost all s, t ∈ G. As the map (s, t) 7→ (t, s−1t) is continuous, bijective and preserves null sets in both directions, we obtain ϕa(s, t) = F (st−1)(a) for almost all (s, t) ∈ G × G. Hence, for each r ∈ G, ϕa(sr, tr) = ϕa(s, t) almost everywhere on G × G. (ii)⇒(i) As γ(t), β(s) ∈ C∞(A) for all (s, t), we have that ϕβ,γ(s, t)(a) = lim n→∞ γi(t)∗aβi(s) n Xi=1 in norm and, in particular, ϕβ,γ(s, t)(a) ∈ A for all a ∈ A. Hence αt(ϕβ,γ(s, t)(αt−1 (a)) = αt lim Xi=1 = lim n→∞ n n→∞ ∞ γi(t)∗αt−1(a)βi(s)! n Xi=1 αt(γi(t)∗)aαt(βi(s)) αt(γi(t))∗aαt(βi(s)) = ϕa(s, t). = Xi=1 HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 35 Thus T (ϕβ,γ)(s, t)(a) = ϕa(s, t) for all s, t ∈ G. Fix r ∈ G and let S ⊆ A be a countable dense subset. Then, for every a ∈ S we have T (ϕβ,γ)r(s, t)(a) = T (ϕβ,γ)(sr, tr)(a) = ϕa(sr, tr) = ϕa(s, t) = T (ϕβ,γ)(s, t)(a), for almost all (s, t) ∈ G × G. It follows that there exists a set E ⊆ G × G whose complement is null, such that T (ϕβ,γ)r(s, t)(a) = T (ϕβ,γ)(s, t)(a) for all (s, t) and all a ∈ S. Fix (s, t) ∈ E. By the boundedness of the maps T (ϕβ,γ)r(s, t) and T (ϕβ,γ)(s, t), we have that T (ϕβ,γ)r(s, t)(a) = T (ϕβ,γ)(s, t)(a), for all a ∈ A. Thus, T (ϕβ,γ)r = T (ϕβ,γ) almost everywhere, for all r ∈ G. By Lemma 3.17, Theorem 3.18 and Proposition 5.1, there exists F ∈ S(A, G, α) such that N (F ) = ϕβ,γ almost everywhere. (cid:3) 5.2. Groupoid multipliers. In this subsection, we relate Herz-Schur mul- tipliers to the multipliers of the Fourier algebra of a groupoid. We refer the reader to [21] and [32] for more details on the background, which we now recall. Let G be a locally compact group acting on a locally compact Hausdorff space X; thus, we are given a map X × G → X, (x, s) → xs, jointly continuous and such that x(st) = (xs)t for all x ∈ X and all s, t ∈ G. The set G = X × G is a groupoid, where the set G2 of composable pairs is given by G2 = {[(x1, t1), (x2, t2)] : x2 = x1t1}, and if [(x1, t1), (x2, t2)] ∈ G2, the product (x1, t1) · (x2, t2) is defined to be (x1, t1t2), while the inverse (x, t)−1 of (x, t) is defined to be (xt, t−1). The domain and range maps are given by d((x, t)) := (x, t)−1 · (x, t) = (xt, e), r((x, t)) := (x, t) · (x, t)−1 = (x, e). The unit space G0 of the groupoid, which is by defnition equal to the common image of the maps d and r, can therefore be canonically identified with X. Let λ be the left Haar measure on G. The groupoid G can be equipped with the Haar system {λx : x ∈ X}, where λx = δx × λ and δx is the point mass at x. The space Cc(G) of compactly supported continuous functions on G is a ∗-algebra with respect to the convolution product given by (f ∗ g)(x, t) =Z f (x, s)g(xs, s−1t)ds, and the involution given by f ∗(x, s) = f (xs, s−1). We equip Cc(G) with the norm kf kI = max(cid:26)sup x∈XZ f dλx, sup x∈XZ f ∗dλx(cid:27) . The completion of Cc(G) with respect to this norm is denoted by LI(G), and its enveloping C ∗-algebra C ∗(G) is called the groupoid C ∗-algebra of G. Let A = C0(X) and αt(a)(x) = a(xt), t ∈ G, x ∈ X. Then α : t 7→ αt is a continuous homomorphism from G to Aut(A). Identifying Cc(X × G) = 36 A. MCKEE, I. TODOROV, AND L. TUROWSKA Cc(G) with a subspace of Cc(G, A), we see that the ∗-algebra structure on Cc(G, A), associated with the action α (see the beginning of Section 3), coincides with the one on Cc(G) except for the absence of the modular function in the definition of the involution. However, the C ∗-algebras C ∗(G) and the full crossed product A ⋊α G are isomorphic via the map φ given by φ(f )(x, s) = ∆−1/2(s)f (x, s), f ∈ Cc(X × G). In fact, for f, g ∈ Cc(X × G), we have φ(f ∗ g)(x, s) = ∆−1/2(s)(f ∗ g)(x, s) = ∆−1/2(s)Z f (x, t)g(xt, t−1s)dt, while φ(f ) × φ(g)(x, s) = Z φ(f )(x, t)φ(g)(xt, t−1s)dt = Z ∆−1/2(t)f (x, t)∆−1/2(t−1s)g(xt, t−1s)dt = ∆−1/2(s)Z f (x, t)g(xt, t−1s)dt; hence, φ(f ∗ g) = φ(f ) × φ(g). In addition, φ(f ∗)(x, s) = ∆−1/2(s)f (xs, s−1), while φ(f )∗(x, s) = ∆−1(s)φ(f )(xs, s−1) = ∆−1(s)∆−1/2(s−1)f (xs, s−1) = ∆−1/2(s)f (xs, s−1), giving φ(f )∗ = φ(f ∗). By [21, p. 9], the map φ extends to a -isomorphism from C ∗(G) onto A ⋊α G. Let µ be a measure on X and ν = µ × λ; thus, for a measurable subset E of X × G, we have ν(E) =R λx(Ex)dµ(x) (for x ∈ X, we have set Ex = E∩({x}×G)). For a measurable subset E, set ν −1(E) =R λx((E−1)x)dµ(x). Let Ind(µ) be the -representation of Cc(G) on L2(G, ν −1) given by (Ind(µ)(f )ξ)(x, t) =Z f (x, s)ξ(xs, s−1t)ds, f ∈ Cc(G). One can check that kInd(µ)(f )k ≤ kf kI [21]; hence Ind(µ) can be extended to C ∗(G). If supp µ = X then the map f 7→ Mf , where Mf is the operator of multiplication by f on L2(X, µ), is a faithful -representation θ of C0(X). The corresponding regular representation πθ ⋊ λθ of the crossed product A ⋊α G on L2(G, L2(X, µ)) = L2(X × G, µ × λ) is given by (πθ⋊λθ)(f )ξ(x, t) =Z αt−1 (f (x, s))ξ(x, s−1t)ds =Z f (xt−1, s)ξ(x, s−1t)ds, for f ∈ Cc(X × G) and ξ ∈ L2(X × G, µ × λ). Let Jξ(x, t) = ξ(xt, t−1); then J is a unitary operator from L2(G, ν) to L2(G, ν −1) with J −1η(x, t) = HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 37 η(xt, t−1). Let also U ξ(x, t) = ∆−1/2(t)ξ(x, t−1); thus, U is a unitary oper- ator on L2(G, ν). We have (U −1J −1Ind(µ)(f )JU ξ)(x, t) =Z f (xt−1, s)ξ(x, s−1t)∆−1/2(s)ds; we thus see that (πθ ⋊ λθ) ◦ φ is unitarily equivalent to Ind(µ). Let I be the intersection of the kernels of Ind(µ) as µ varies over the measures of X. The quotient C ∗(G)/I is called the reduced C ∗-algebra of G and denoted by C ∗ red(G). It follows from [21, Proposition 2.17] that Ind(µ) is a faithful representation of C ∗ red(G) is isomorphic to the reduced crossed product C ∗-algebra of the C ∗-dynamical system (C0(X), G, α). red(G) if supp µ = X. Therefore C ∗ A measure µ on X is called quasi-invariant if the measures ν and ν −1 are equivalent. It is known that µ is quasi-invariant if and only if the measures µ and µ · s are equivalent for any s ∈ G (here µ · s(E) = µ(Es−1)). If δ(·, s) is the Radon-Nikodym derivative d(µ · s−1)/dµ and D is the Radon- Nikodym derivative dν/dν −1 then D(x, s) = ∆(s)/δ(x, s), x ∈ X, s ∈ G (see [32, Chapter I, 3.21]). In what follows we will asume that X possesses a quasi-invariant measure µ such that supp µ = X. The groupoid G equipped with such a measure µ is called a measured groupoid [32]. Next we would like to point out a connection between its multipliers, studied in [33], and Herz-Schur (C0(X), G, α)-multipliers. The Hilbert space L2(G, ν) carries a representation Reg of Cc(G) defined by (Reg(f )ξ)(x, s) =Z f (x, t)ξ(xt, t−1s)D−1/2(x, t)dt, and unitarily equivalent to Ind(µ) via the unitary operator V from L2(G, ν) to L2(G, ν −1) given by V ξ = D1/2ξ. The von Neumann algebra VN(G) of G is defined to be the bicommutant Reg(Cc(G))′′ [33, 2.1]. The Fourier algebra A(G) of the measured groupoid G was defined in [33] and is, similarly to the case where G is a group, a Banach algebra of complex- valued continuous functions on G. By [33, Propsition 3.1], the operator Mϕ of multiplication by ϕ ∈ L∞(G) is a bounded linear map on A(G) if and only if the map Reg(f ) → Reg(ϕf ), f ∈ Cc(G), is bounded. The function ϕ is in this case called a multiplier of A(G). If the map Mϕ is moreover completely bounded then ϕ is called a completely bounded multiplier of A(G). Following [33], we denote by M A(G) (resp. M0A(G)) the set of all multipliers (resp. completely bounded multipliers) of A(G). For a bounded continuous function ϕ : X × G → C and t ∈ G, let Fϕ(t) be the linear map on C0(X) given by Fϕ(t)(a)(x) = ϕ(x, t)a(x), a ∈ C0(X), x ∈ X. Proposition 5.3. Let ϕ : X × G → C be a bounded continuous function. Then (i) the map Fϕ is a θ-multiplier if and only if ϕ is a multiplier of A(G); 38 A. MCKEE, I. TODOROV, AND L. TUROWSKA (ii) the map Fϕ is a Herz-Schur (C0(X), G, α)-multiplier if and only if ϕ is a completely bounded multiplier of A(G). Proof. Both statements follow from the previous paragraphs, Remark 3.2 (iii), the definition of (Herz-Schur) θ-multipliers and the fact that kReg(f )k = k(πθ ⋊ λθ)(φ(f ))k, f ∈ Cc(G). (cid:3) The following statement gives the result of [33, Proposition 3.8] in case G is a locally compact second countable group. Corollary 5.4. Let θ : G → C be a bounded continuous function and ϕ : X × G → C be the function given by ϕ(x, t) = θ(t). Then ϕ is a completely bounded multiplier of A(G) if and only if θ ∈ M cbA(G). Proof. The statement follows from Proposition 5.3 and Proposition 4.1. (cid:3) The next corollary provides a new description of the completely bounded multipliers of A(G). We write H = L2(X, µ). Corollary 5.5. Let ϕ : X × G → C be a bounded continuous function. Assume that ϕ is a completely bounded multiplier of A(G). Then there exist a separable Hilbert space K and functions V, W ∈ L∞(G, B(H, K)) such that, for almost all s, t ∈ G, we have that W ∗(t)V (s) ∈ DX and ϕ(xt−1, ts−1) = (W ∗(t)V (s))(x), for almost all x ∈ X. Proof. By Proposition 5.3, F = Fϕ is a Herz-Schur (C0(X), G, α)-multiplier. By Theorem 3.8, N (F ) is a Schur C0(X)-multiplier. We have (48) N (F )(s, t)(a)(x) = ϕ(xt−1, ts−1)a(x). Hence there exist a separable Hilbert space K, a non-degenerate ∗-represen- tation ρ : C0(X) → B(K) and functions V, W ∈ L∞(G, B(H, K)) such that ϕ(xt−1, ts−1)a(x)ξ(x) = W ∗(t)ρ(a)V (s)ξ(x), a ∈ C0(X), ξ ∈ L2(X, µ). Taking an approximate unit (an)n∈N of C0(X) in (48) and letting n → ∞, we obtain the statement. (cid:3) 6. Convolution multipliers Throughout this section, we will assume that G is an abelian locally com- pact group, and will write the group operations additively. Let A be a sep- arable C-algebra and (A, G, α) be a C-dynamical system. For a measure µ ∈ M (G), let αµ : A → A be the completely bounded map given by αµ(a) =RG αr(a)dµ(r) (see [37],[39]). Definition 6.1. A family Λ = (µt)t∈G, where µt ∈ M (G), t ∈ G, will be called a convolution (A, G, α)-multiplier (or simply a convolution mul- tiplier), if the map FΛ : G → CB(A) given by FΛ(t) = αµt , t ∈ G, is a Herz-Schur (A, G, α)-multiplier. HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 39 For a convolution multiplier Λ = (µt)t∈G, we let kΛkm = kFΛkm. Since G is assumed to be abelian, we have that αµ ◦ αr = αr ◦ αµ for every r ∈ G and every µ ∈ M (G). It is well-known that in this case the map αµ : A → A lifts to a completely bounded map on th the crossed product; the following proposition provides a concrete route to this fact. Proposition 6.2. Let µ ∈ M (G), µt = µ for every t ∈ G, and Λ = (µt)t∈G. Then Λ is a convolution multiplier and kΛkm ≤ kµk. Proof. Note that π(αr(a)) = λrπ(a)λ∗ r, r ∈ G. Set F = FΛ. If a ∈ A then π(F (s)(a)) = π(αµ(a)) =ZG Hence, if f ∈ L1(G, A) then π(αr(a))dµ(r) =ZG λrπ(a)λ∗ rdµ(r). SF ((π ⋊ λ)(f )) = ZG π(F (s)(f (s))λsds = ZG(cid:18)ZG = ZGZG = ZG λr(cid:18)ZG λrπ(f (s))λ∗ rdµ(r)(cid:19) λsds λrπ(f (s))λsλ∗ rdsdµ(r) π(f (s))λsds(cid:19) λ∗ rdµ(r). The claims follow from the fact that the mapping T 7→ R λrT λ∗ rdµ(r) is a completely bounded map on B(L2(G, H)) with completely bounded norm dominated by kµk (see [37]). (cid:3) In this section, we will be concerned with a special class of convolution multipliers, which we now define. Let Γ be the dual group of G. The C- algebra C ∗(Γ) of Γ is canonically -isomorphic to its reduced C-algebra C ∗ r (Γ) (see e.g. [27, Theorem 7.3.9]). We let θ : C ∗(Γ) → B(L2(Γ)) be the associated (faithful) -representation. An element s ∈ G will be viewed as a character (and, in particular, a unimodular function) on Γ. For s ∈ G, let αs : λΓ(L1(Γ)) → λΓ(L1(Γ)) be the map given by αs(λΓ(f )) = λΓ(sf ), f ∈ L1(Γ). Note that, if f ∈ L1(Γ), s ∈ G and x ∈ Γ, then αs(λΓ(f )) = MsλΓ(f )M−s. (49) Thus, αs extends canonically to an automorphism of C ∗ r (Γ). By abuse of notation, we consider αs as an automorphism of C ∗(Γ); thus, (C ∗(Γ), G, α) is a C-dynamical system. By (49), α is a θ-action. Moreover, by [27, Theorem 7.7.7], C ∗(Γ) ⋊α G is -isomorphic to the C-subalgebra C ∗(Γ) ⋊α,θ G of B(L2(G × Γ)). Given a bounded measurable function ψ : G × Γ → C and t ∈ G (resp. x ∈ Γ), let the function ψt : Γ → C (resp. ψx : G → C) given by ψt(y) = 40 A. MCKEE, I. TODOROV, AND L. TUROWSKA ψ(t, y) (resp. ψx(s) = ψ(s, x)). We call ψ admissible if ψt ∈ B(Γ) for every t ∈ G and supt kψtkB(Γ) < ∞. Assuming that ψ is addmissible, let Fψ(t) : C ∗ r (Γ) → C ∗ r (Γ) be the map given by Fψ(t)(λΓ(g)) = λΓ(ψtg), g ∈ L1(Γ). By abuse of notation, we consider Fψ(t) as a map on C ∗(Γ). Set F(G) = {ψ : G × Γ → C : ψ is admissible and Fψ is a Herz-Schur (C ∗(Γ), G, α)-multiplier} and Fθ(G) = {ψ : G × Γ → C : ψ is admissible and Fψ is a Herz-Schur θ-multiplier}. Clearly, the space F(G) is an algebra with respect to the operations of point- wise addition and multiplication, and Fθ(G) is a subalgebra of F(G). For ψ ∈ F(G), let kψkm = kFψkm, and use Sψ to denote the map SFψ . For µ ∈ M (G), set µ(x) =RGhx, sidµ(s), x ∈ Γ. Proposition 6.3. Let ψ : G × Γ → C be an addmissible function. The following are equivalent: (i) ψ ∈ F(G); (ii) for each t ∈ G, there exists µt ∈ M (G) such that ψ(t, x) = µt(x), t ∈ G, x ∈ Γ, and the family (µt)t∈G is a convolution (C ∗(Γ), G, α)-multiplier. Proof. Note that, if µ ∈ M (G) and g ∈ L1(Γ) then (50) αµ(λΓ(g)) = Z αr(λΓ(g))dµ(r) =ZG(cid:18)ZG hs, rig(s)λΓ s ds(cid:19) dµ(r) = ZG µ(s)g(s)λΓ s ds = λΓ(µg). (i)⇒(ii) If ψ is admissible then, for every t ∈ G, ψt ∈ B(Γ) and hence, by Bochner's theorem, there exists µt ∈ M (G) such that ψt = µt (see e.g. [35, Section I]). It follows from (50) that the family (µt)t∈G is a convolution multiplier. (ii)⇒ (i) By (50), Fψ(t) = αµt . The claim now follows from the definition (cid:3) of a convolution multiplier. For a family Λ = (µt)t∈G of measures in M (G), let ψΛ : G × Γ → C be the function given by ψΛ(t, x) = µt(x). Call Λ admissible if the function ψΛ is admissible. By Proposition 6.3, an admissibe family of measures Λ is a Herz-Schur (C ∗(Γ), G, α)-multiplier if and only if ψΛ ∈ F(G). By abuse of terminology, we will hence call the elements of F(G) convolution multipliers. Corollary 6.4. Let g ∈ L∞(Γ) and let ψ : G × Γ → C be given by ψ(s, x) = g(x), s ∈ G, x ∈ Γ. The following are equivalent: (i) ψ ∈ F(G); (ii) g ∈ B(Γ). HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 41 Moreover, if (i) holds then ψ ∈ Fθ(G). Proof. The equivalence of (i) and (ii) follows from Propositions 6.2 and 6.3. Suppose that g ∈ B(Γ). Then the map on C ∗ r (Γ) corresponding to g via classical transference has a (completely bounded) weak* continuous extension Φg : VN(Γ) → VN(Γ). Thus, the restriction of the map Φg ⊗ id to C ∗(Γ) ⋊w∗ . By Remark 3.11, Fψ is a Herz-Schur θ-multiplier. (cid:3) α,θ G is a weak* continuous extension of Sθ Fψ It will be convenient, in the sequel, to denote by Sg the map Sψ, where ψ and g are as in Corollary 6.4. The rest of the paper will be devoted to properties of the spaces F(G) and Fθ(G). In the next theorem, we identify an elementary tensor u ⊗ h, where u ∈ B(G) and h ∈ B(Γ), with the function (s, x) → u(s)h(x), s ∈ G, x ∈ Γ. Let F(B(G), B(Γ)) be the complex vector space of all separately continuous functions ψ : G × Γ → C such that, for every s ∈ G (resp. x ∈ Γ), the function ψs : Γ → C (resp. ψx : G → C) belongs to B(Γ) (resp. B(G)). Theorem 6.5. (i) The inclusions B(G) ⊙ B(Γ) ⊆ Fθ(G) ⊆ F(B(G), B(Γ)) hold. (ii) Suppose that ψ ∈ Fθ(G). Then kψxkB(G) ≤ kψkm for every x ∈ Γ and kψskB(Γ) ≤ kψkm for every s ∈ G. (iii) Let ψ : G × Γ → C be an admissible function, such that the function G → B(Γ), sending s to ψs, is continuous. Suppose that (ψk)k∈N ⊆ F(G), supk∈N kψkk∞ < ∞ and ψk → ψ pointwise. Then ψ ∈ F(G). Proof. (i) The first inclusion follows from Proposition 4.1 and Corollary 6.4. Let ψ ∈ Fθ(G) and fix x ∈ Γ. The map Ψψs corresponding to ψs via x , x ∈ Γ. classical transference satisfies the identities Ψψs(λΓ Thus x) = ψs(x)λΓ Φθ ψ(¯πθ(λΓ x)λθ s) = ¯πθ(ψs(x)λΓ = ψs(x)¯πθ(λΓ x )λθ s s = ¯πθ(λΓ x)λθ x )(ψx(s)λG s ⊗ I). On the other hand, It follows that the map ¯πθ(λΓ x )λθ s = ¯πθ(λΓ x)(λG s ⊗ I). s → ψx(s)λG λG s is bounded, and hence ψx is a Herz-Schur multiplier giving ψx ∈ B(G). The fact that ψs ∈ B(Γ) for every s ∈ G is implicit in the definition of the space F(G). (ii) The inequalities kψxkB(Γ) ≤ kψkm, x ∈ Γ, follow from the proof of (i). The inequalities kψskB(Γ) ≤ kψkm, s ∈ G, follow from Corollary 3.12 and the fact that kψskB(Γ) = kFψ(s)kcb, s ∈ G. 42 A. MCKEE, I. TODOROV, AND L. TUROWSKA (iii) Suppose that kψkk∞ ≤ C for every k ∈ N. Let ξ, η ∈ L2(G × Γ) and f ∈ L1(G, Cc(Γ)). A direct verification shows that hSθ ψk ((πθ ⋊ λθ)(f ))ξ, ηi = h(πθ ⋊ λθ)(ψkf )ξ, ηi = Z ht, xiψk(t, y)f (t, y)ξ(s − t, x − y)η(s, x)dsdtdxdy. On the other hand, ht, xiψk(t, y)f (t, y)ξ(s − t, x − y)η(s, x) ≤ Cf (t, y)ξ(s − t, x − y)η(s, x), and the L1-norm of the latter function is equal to Chf ∗ ξ, ηi. By the Lebesgue Dominated Convergence Theorem, (51) hSθ ψk ((πθ ⋊ λθ)(f ))ξ, ηi → h(πθ ⋊ λθ)(ψf )ξ, ηi. Now let fi,j ∈ L1(G, Cc(Γ)), i, j = 1, . . . , m. By (51), the operator matrix ((πθ ⋊ λθ)(f )))i,j converges weakly to ((πθ ⋊ λθ)(ψf ))i,j . The fact that (Sθ ψk the map s → ψs is continuous implies that ψ is weakly measurable. Since kSθ kcb ≤ C for all k, we conclude that ψ ∈ F(G) and that kψkm ≤ C. (cid:3) ψk In view of Theorem 6.5, it is natural to ask the following question. Question 6.6. Can Fθ(G) be characterised as a topological closure of B(G)⊙ B(Γ)? Let CBw∗(C ∗(Γ) ⋊α,θ G) be the space of all completely bounded maps on α,θ G. C ∗(Γ) ⋊α,θ G which admit a weak* continuous extension to C ∗(Γ) ⋊w∗ Set S = {Sθ ψ : ψ ∈ Fθ(G)}. As usual, if J is a family of linear transformations acting on a vector space, we denote by J ′ its commutant. We recall that, for g ∈ B(Γ), we let g denote the map on C ∗ Sθ Theorem 6.7. We have g (λΓ(f )) = λΓ(gf ), f ∈ L1(Γ). r (Γ) given by Sθ u, Sθ v : u ∈ B(G), v ∈ B(Γ)}′. S = CBw∗(C ∗(Γ) ⋊α,θ G) ∩ {Sθ (52) In particular, S is a maximal abelian subalgebra of CBw∗(C ∗(Γ) ⋊α,θ G). Proof. Note that S is a commutative subalgebra of CBw∗(C ∗(Γ) ⋊α,θ G) and, by Theorem 6.5 (i), contains the maps of the form Sθ v , where u ∈ B(G) and v ∈ B(Γ). It follows that it is contained in the commutant on the right hand side of (52). u and Sθ Let A = C ∗(Γ). To establish the reverse inclusion, suppose that Φ ∈ CBw∗(C ∗(Γ) ⋊α,θ G) commutes with the operators of the form Sθ u and Sθ v , where u ∈ B(G) and v ∈ B(Γ). Since G is abelian, VN(G) possesses prop- erty Sσ (see e.g. [19, Theorem 1.9]). By Theorem 4.3, for each t ∈ G, there exists a completely bounded map Ft : VN(Γ) → VN(Γ) such that Φ(¯πθ(a)λθ t ) = ¯πθ(Ft(a))λθ t , a ∈ VN(Γ). HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 43 t = Φ(Sθ v (¯πθ(a)λθ If v ∈ B(Γ), f ∈ L1(Γ) and a = λΓ(f ) then ¯πθ(Ft(λΓ(vf )))λθ t )) = ¯πθ(vFt(λΓ(f ))λθ t )) = Sθ t . Thus, Ft(λΓ(vf )) = vFt(λΓ(f )) for each v ∈ B(Γ). Let Ft be the restriction of Ft to C ∗ t (A(Γ)) ⊆ B(Γ) and F ∗ It follows that F ∗ t (u) = ψtu for some function ψt : Γ → C. As ψtu ∈ B(Γ) for all u ∈ A(Γ), by [35, Theorem 3.8.1], ψt ∈ B(Γ). Hence, for u ∈ A(Γ), we have t (u) for any u ∈ A(Γ) and v ∈ B(Γ). is a map on B(Γ). In particular, F ∗ r (Γ). Then F ∗ t t (vu) = v F ∗ v (Φ(¯πθ(a)λθ hFt(λΓ(f )), ui = hλΓ(f ), F ∗ t (u)i = hλΓ(f ), ψtui = hλΓ(ψtf ), ui and Ft(λΓ(f )) = λΓ(ψtf ). The proof is complete. (cid:3) Our next aim is to identify the joint commutant of two families of com- pletely bounded maps on K(L2(G)), in terms of multipliers of Herz-Schur type. Recall that, for a ∈ L∞(G), we denote by Ma the operator on L2(G) given by Maf = af , f ∈ L2(G), and set C = {Ma : a ∈ C0(G)}. We let id be the identity representation of C. For t ∈ G, let βt : C0(G) → C0(G) be given by βt(h)(s) = h(s − t), h ∈ C0(G). By abuse of notation, we denote by βt the corresponding map on the C-algebra C. Note that (53) βt(T ) = λG t T λG −t, T ∈ C, and that (C, G, β) is a C-dynamical system. Note also that βµ(Ma) = Mµ∗a, µ ∈ M (G), a ∈ L∞(G). (54) Note that C ⋊β,id G is a C-subalgebra of B(L2(G × G)). Let F : L2(G) → L2(Γ) be the Fourier transform, so that Fξ(x) = (55) F ∗MtF = λG t RGhx, siξ(s)ds, ξ ∈ L1(G) ∩ L2(G), x ∈ Γ. Then where f : G → C is the function given by f (t) =RΓ ht, xif (x)dx. In partic- r (Γ)F = C. Moreover, if f ∈ L1(Γ) then and F ∗λΓ(f )F = M f , t ∈ G, f ∈ L1(Γ), ular, F ∗C ∗ F ∗αt(λΓ(f ))F = βt(F ∗λΓ(f )F), (56) giving F ∗αt(T )F = βt(F ∗T F), T ∈ C ∗ r (Γ), t ∈ G. Let F = I ⊗F ∗; thus, F is a unitary operator from L2(G×Γ) onto L2(G×G). It is well-known that F (C ∗(Γ) ⋊α,θ G) F ∗ = C ⋊β,id G. Let Λ = (µt)t∈G be a family of measures in M (G), such that the function ψΛ is in F(G). Then ψΛ gives rise to a Herz-Schur (C, G, β)-multiplier given by π(Mg)λt = π(F ∗λΓ(g)F)λt 7→ π(F ∗Fψt(λΓ(g))F)λt = π(Mµt∗g)λt. 44 A. MCKEE, I. TODOROV, AND L. TUROWSKA This observation was our motivation for the chosen terminology for con- volution multipliers. Note that the convolution multipliers are of different nature than Herz-Schur (C, G, β)-multipliers considered in Section 5.2. The pair (id, λG) is a covariant representation of (C, G, β) (see (53)); in addition, id ⋊λG is a faithful representation of C ⋊β G on L2(G) and its image coincides with the algebra K(L2(G)) of all compact operators on L2(G) (see [34] and [43]). For ψ ∈ F(G), let Eψ : K(L2(G)) → K(L2(G)) be the (completely bounded) map given by Eψ((id ⋊λG)( F T F ∗)) = (id ⋊λG)( F Sθ ψ(T ) F ∗), T ∈ C∗(Γ) ⋊α,θ G. We extend Eψ to a weak continuous map on B(L2(G)), denoted in the same r ∈ B(L2(G)) be the corresponding right regular fashion. For r ∈ G, let ρG unitary on L2(G), that is, ρG r f (s) = f (s + r), s, r ∈ G, f ∈ L2(G). For a measure µ ∈ M (G), consider the map Θ(µ) ∈ CB(B(L2(G)), given by (57) Θ(µ)(T ) =ZG It is easy to see that ρG r T ρG −rdµ(r), T ∈ B(L2(G)). (58) Θ(µ)(Ma) = Mµ·a = Mµ∗a, a ∈ L∞(G), where (µ · a)(s) = RG a(s + r)dµ(r) and µ is the measure on G given by µ(E) = µ(−E). Note that Θ(µ) is a VN(G)-bimodule map and leaves DG invariant. Recall that, for every u ∈ B(G), the function N (u) given by N (u)(s, t) = u(t−s), is a Schur multiplier [5] (see also Remrak 3.9). Thus, the correspond- ing map Ψu : B(L2(G)) → B(L2(G)) is a completely bounded DG-bimodule map that leaves VN(G) invariant. Proposition 6.8. Suppose that µ ∈ M (G) and u ∈ B(G). Let ψµ and ψu be the elements of F(G) given by ψµ(s, x) = µ(x) and ψu(s, x) = u(s), s ∈ G, x ∈ Γ. Then (i) Eψµ = Θ(µ), and (ii) Eψu = Ψu. Proof. Let f ∈ Cc(G, Cc(Γ)) be given by f (s) = f0(s)g, for a certain g ∈ Cc(Γ) and a certain f0 ∈ Cc(G). Let i : Cc(G, Cc(Γ)) → C ∗(Γ) ⋊α,θ G be the embedding map given by i(h)ξ(t) =ZG We have α−t(λΓ(h(s)))λθ sξ(t)ds, ξ ∈ L2(G × Γ), h ∈ Cc(G × Γ). i(f )ξ(t) =ZG α−t(λΓ(g))f0(s)λθ sξ(t)ds HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 45 and, by (55) and (56), (59) Fi(f ) F ∗ξ(t) = ZG = ZG = ZG F ∗α−t(λΓ(g))Ff0(s)λθ sξ(t)ds β−t(F ∗(λΓ(g))F)f0(s)λθ sξ(t)ds β−t(Mg)f0(s)λθ sξ(t)ds. Let T, Tµ : G → C be the maps given by T (s) = Mf0(s)g and Tµ(s) = Mf0(s)µ∗g, s ∈ G; clearly, T, Tµ ∈ L1(G, C). Let j : L1(G, C) → C ⋊β,id G be the canonical injection. By (59), Fi(f ) F ∗ = j(T ) and F i(ψµf ) F ∗ = j(Tµ). (i) Note that Sθ ψµ(i(f )) = i(ψµf ) = i(f0 ⊗ (µg)). We have Eψµ(MgλG(f0)) = Eψµ((id ⋊λG)(j(T ))) = Eψµ((id ⋊λG)( F i(f ) F ∗)) = (id ⋊λG)( F Sθ = (id ⋊λG)(j(Tµ)) = Mµ∗gλG(f0). ψµ(i(f )) F ∗) = (id ⋊λG)( F i(f0 ⊗ µg) F ∗) By (58) and the modularity of Θ(µ) over VN(G) we now have Eψµ(MgλG(f0)) = Mµ∗gλG(f0) = Mµ·gλG(f0) = Θ(µ)(MgλG(f0)). Since the operators of the form MgλG(f0) span a dense subspace of K(L2(G)), it follows that Eψµ = Θ(µ). (ii) Similarly to (i), we have Eψu(MgλG(f0)) = (id ⋊λG)( F i(uf0 ⊗ g) F ∗) = MgλG(uf0). Since, by [15], λG(uf0) = Ψu(λG(f0)), and Ψu is a C-bimodule map, we obtain that Eψu(MgλG(f0)) = Ψu((MgλG(f0)). The statement now follows by the density of the linear span of the operators of the form MgλG(f0) in K(L2(G)). (cid:3) Definition 6.9. Let (θ, τ ) be a covariant representation of the dynamical system (A, G, α). We say that F : G → CB(A) is a Herz-Schur (θ, τ )- multiplier if the map θ(a)τs → θ(F (s)(a)τs, s ∈ G, a ∈ A can be extended to a weak-continuous completely bounded map on the weak closed hull of (θ ⋊ τ )(A ⋊α G). Let Fid,λG(G) be the set of admissible functions ψ : G × Γ → C such that the corresponding Fψ : G → CB(C) is a Herz-Schur (id, λG)-multiplier for the dynamical system (C, G, β). 46 A. MCKEE, I. TODOROV, AND L. TUROWSKA Theorem 6.10. Let E = {Eψ : ψ ∈ Fid,λG(G)}. Then E = CB(K(L2(G))) ∩ {Θ(µ), Ψu : µ ∈ M (G), u ∈ B(G)}′. In particular, E is a maximal abelian subalgebra of CB(K(L2(G)). Proof. The fact that E is contained in the right hand side follows from the fact that E is a commutative subalgebra of CB(K(L2(G))) and, by Proposi- tion 6.8, contains the maps Θ(µ) and Ψu, where u ∈ B(G) and µ ∈ M (G). To prove the reverse inclusion, we modify the arguments in the proof of Theorem 4.3. Let Φ ∈ CB(K) commute with Ψu and Θ(µ), for all µ ∈ M (G) and all u ∈ B(G). The map Φ has a unique extension to a weak* continuous completely bounded map on B(L2(G)), which will be denoted by the same symbol. Let t ∈ G, a ∈ L∞(G) and T = MaλG t . Set J = {u ∈ B(G) : u(t) = 1}. Then for u ∈ J we have Ψu(MaλG t ) = u(t)MaλG t = MaλG t . As ΦΨu = ΨuΦ, we obtain Ψu(Φ(T )) = Φ(T ). By [1] and the remark before Theorem 4.3, we conclude that Φ(T )λG −t ∈ DG. Therefore, there exists at ∈ L∞(G) such that Φ(MaλG t ) = MatλG t . Let Ft(a) = at, t ∈ G. Then Ft is a linear map on DG. Since Φ is completely bounded and weak* continuous, Ft is so, too. Since Φ commutes with Θ(µ), µ ∈ M (G), we have Φ(Mµ·aλG t ) = Mµ·Ft(a)λG t , giving Ft(µ·a) = µ·Ft(a). The map Ft is the adjoint of a bounded linear map Ψt : L1(G) → L1(G) such that Ψt(µ ∗ f ) = µ ∗ Ψt(f ), µ ∈ M (G), f ∈ L1(G). By [35, Theorem 3.8], there exists νt ∈ M (G) such that Ft(a) = νt ∗ a, a ∈ L∞(G). Let ψ(t, x) = νt(x), t ∈ G, x ∈ Γ. Then ψ ∈ Fid,λG(G) and Φ is the weak* extension of Eψ. (cid:3) Remark 6.11. Assume G is arbitrary and let βt : C → C be given by t−1. Then (C, G, β) is a C ∗-dynamical system. For a βt(Ma) = λG measure µ we consider the map Θ(µ) ∈ CB(L2(G)) given by (57). We have Θ(µ)(C) ⊆ C. Moreover, for each r ∈ G we have βr ◦ Θ(µ) = Θ(µ) ◦ βr: t MaλG βr(Θ(µ)(Ma)) = λG s MaρG ρG s−1dµ(s)(cid:19) λG r−1 r (cid:18)ZG = ZG ρG s λG r MaλG r−1ρs−1dµ(s) = Θ(µ)(βr(Ma)). Hence Θ(µ) gives rise to a completely bounded map on C ⋊β,r G, i.e. the function F , given by F (t)(Ma) := Θ(µ)(Ma), is a Herz-Schur (C, G, β)- multiplier. For Λ = {µt}t∈G we let FΛ(t)(Ma) = Θ(µt)(Ma). The class of HERZ-SCHUR MULTIPLIERS OF DYNAMICAL SYSTEMS 47 Herz-Schur multipliers FΛ includes the convolution multipliers examined in the present section, whose study will be pursued elsewhere. Acknowledgement. We would like to thank Sergey Neshveyev and Adam Skalski for many helpful conversations during the preparation of this paper. References [1] M. Anoussis, A. Katavolos and I. G. Todorov, Ideals of the Fourier algebra, supports and harmonic operators, to appear in the Math. Proc. Cambridge Plilos. Soc. [2] E. Bedos and R. Conti, Fourier series and twisted C-crossed products, J. Fourier Anal. Appl 21 (2015), 32-75. [3] M. S. Birman and M. Z. Solomyak, Double operator integrals in a Hilbert space, Int. Eq. Oper. Th. 47 (2003), no. 2, 131-168. [4] D. P. Blecher and C. Le Merdy, Operator algebras and their modules -- an oper- ator space approach, Oxford University Press, 2004. [5] M. Bozejko and G. Fendler, Herz-Schur multipliers and completely bounded mul- tipliers of the Fourier algebra of a locally compact group, Boll. Un. Mat. Ital. A (6) 2 (1984), no. 2, 297-302. [6] N. P. Brown and N. Ozawa, C-algebras and finite dimensional approximations, American Mathematical Society, 2008. [7] J. de Canniere and U. Haagerup, Multipliers of the Fourier algebras of some simple Lie groups and their discrete subgroups, Amer. J. Math. 107 (1985), no. 2, 455-500. [8] S. Echterhoff, S. Kaliszewski, J. Quigg, I. Raeburn Naturality and induced representations, Bull. Austral. Math. Soc. 61 (2000), no. 3, 415-438. [9] E. Effros and Z.-J. Ruan, Operator Spaces, Oxford University Press, 2000. [10] P. Eymard, L'alg`ebre de Fourier d'un groupe localement compact, Bull. Soc. Math. France 92 (1964), 181-236. [11] F. Ghahramani, Isometric representations of M (G) on B(H), Glasgow Math. J. 23 (1982), 119-122. [12] A. Grothendieck, R´esum´e de la th´eorie m´etrique des produits tensoriels topologiques, Boll. Soc. Mat. Sao-Paulo 8 (1956), 1-79. [13] U. Haagerup, Decomposition of completely bounded maps on operator algebras, un- published manuscript. [14] T. Huruya, The second dual of a tensor product of C* -algebras, II, Sci. Rep. Niigata Univ. Ser. A 11 (1974), 21-23. [15] P. Jolissaint, A characterisation of completely bounded multipliers of Fourier alge- bras, Colloquium Math. 63 (1992) 311-313. [16] R. V. Kadison and J. R. Ringrose, Fundamentals of the theory of von Neumann algebras II, Academic Press, 1986. [17] A. Katavolos and V. I. Paulsen, On the ranges of bimodule projections, Canad. Math. Bull. 48 (2005) no. 1, 97-111. [18] S. Knudby, The weak Haagerup property, preprint, arXiv 1401.7541. [19] J. Kraus, The slice map problem for σ-weakly closed subspaces of von Neumann algebras, Trans. Amer. Math. Soc. 279 (1983), no. 1, 357-376. [20] U. Haagerup and J. Kraus, Approximation properties for group C-algebras and group von Neumann algebras, Trans. Amer. Math. Soc. 344 (1994), no. 2, 667-699. [21] P. S. Muhly and J. N. Renault, C-algebras of multivariable Wienner-Hopf oper- ators, Trans. Amer. Math. Soc. 274 (1982), no. 1, 1-44. [22] P. S. Muhly and B. Solel, Tensor algebras over C-correspondences: representa- tions, dilations, and C-envelopes, J. Funct. Anal. 158 (1998), no. 2, 389-457. 48 A. MCKEE, I. TODOROV, AND L. TUROWSKA [23] Y. Nakagami and M. Takesaki, Duality for crossed products of von Neumann algebras, Springer-Verlag, 1979. [24] M. Neufang and V. Runde, Harmonic operators: the dual perspective, Math. Z. 255 (2007), no. 3, 669-690. [25] M. Neufang, Z-J. Ruan and N. Spronk, Completely isometric representations of McbA(G) and U CB( bG)∗, Trans. Am. Math. Soc. 360 (2008), no. 3, 1133-1161. [26] V. I. Paulsen, Completely bounded maps and operator algebras, Cambridge Univer- sity Press, 2002. [27] G. K. Pedersen, C-algebras and their automorphism groups, Academic Press, 1979. [28] V. V. Peller, Hankel operators in the theory of perturbations of unitary and selfad- joint operators, Funct. Anal. Appl. 19 (1985), no. 2, 111-123. [29] M. Pimsner, A class of C-algebras generalizing both Cuntz-Krieger algebras and crossed products by Z, "Free Probability Theory" (D. Voiculescu, Ed.), Fields Instit. Comm. (1997) 12, 189-212. [30] G. Pisier, Similarity problems and completely bounded maps, Springer-Verlag, 2001. [31] G. Pisier, Introduction to operator space theory, Cambridge University Press, 2003. [32] J. Renault, A groupoid approach to C-algebras, Springer-Verlag, 1980. [33] J. Renault, The Fourier algebra of a measured groupoid and its multipliers, J. Funct. Anal. 145 (1997), no. 2, 455-490. [34] M. A. Rieffel, On the uniqueness of the Heisenberg commutation relations, Duke Math. J. 39 (1972), 745-752. [35] W. Rudin, Fourier analysis on groups, John Wiley & Sons, 1990. [36] V. S. Shulman, I. G. Todorov and L. Turowska, Closable multipliers, Int. Eq. Oper. Th. 69 (2011), no. 1, 29-62. [37] R. R. Smith and N. Spronk, Representations of group algebras in spaces of com- pletely bounded maps, Indiana Univ. Math. J. 54 (2005), no. 3, 873-896. [38] N. Spronk, Measurable Schur multipliers and completely bounded multipliers of the Fourier algebras, Proc. London Math. Soc (3) 89 (2004), 161-192. [39] E. Størmer, Regular abelian Banach algebras of linear maps of operator algebras, J. Funct. Anal. 37 (1980), no. 3, 331-373. [40] M. Takesaki, Theory of operator algebras I, Springer-Verlag, 2003. [41] I. G. Todorov and L. Turowska, Sets of p-multiplicity in locally compact groups, Studia Math. 226 (2015), 75-93. [42] J. Tomiyama, On the projection of norm one in W-algebras, Proc. Japan Acad. 33 (1957), no. 10, 608-612. [43] D. P. Williams, Crossed products of C*-algebras, American Mathematical Society, 2007. Pure Mathematics Research Centre, Queen's University Belfast, Belfast BT7 1NN, United Kingdom E-mail address: [email protected] Pure Mathematics Research Centre, Queen's University Belfast, Belfast BT7 1NN, United Kingdom E-mail address: [email protected] Department of Mathematical Sciences, Chalmers University of Technol- ogy and the University of Gothenburg, Gothenburg SE-412 96, Sweden E-mail address: [email protected]
1905.10332	2	1905	2019-05-27T03:16:49	The hyperrigidity of tensor algebras of C$^*$-correspondences	[ "math.OA", "math.FA" ]	Given a C$^*$-correspondence $X$, we give necessary and sufficient conditions for the tensor algebra $\mathcal T_X^+$ to be hyperrigid. In the case where $X$ is coming from a topological graph we obtain a complete characterization.	math.OA	math	THE HYPERRIGIDITY OF TENSOR ALGEBRAS OF C∗-CORRESPONDENCES ELIAS G. KATSOULIS AND CHRISTOPHER RAMSEY Abstract. Given a C∗-correspondence X, we give necessary and suf- ficient conditions for the tensor algebra T + In the case where X is coming from a topological graph we obtain a complete characterization. X to be hyperrigid. 1. Introduction A not necessarily unital operator algebra A is said to be hyperrigid if given any non-degenerate ∗-homomorphism τ : C∗ env(A) −→ B(H) then τ is the only completely positive, completely contractive extension of the restricted map τA. Arveson coined the term hyperrigid in [1] but he was not the only one considering properties similar to this at the time, e.g. [4]. There are many examples of hyperrigid operator algebras such as those which are Dirichlet but the situation was not very clear in the case of ten- sor algebras of C∗-correspondences. It was known that the tensor algebra of a row-finite graph is hyperrigid [4], [5] and Dor-On and Salmomon [3] showed that row-finiteness completely characterizes hyperrigidity for such graph correspondences. These approaches, while successful, did not lend themselves to a more general characterization. The authors, in a previous work [11], developed a sufficient condition for hyperrigidity in tensor algebras. In particular, if Katsura's ideal acts non-degenerately on the left then the tensor algebra is hyperrigid. The motivation was to provide a large class of hyperrigid C∗-correspondence examples as crossed products of operator algebras behave in a very nice manner when the operator algebra is hyperrigid. This theory was in turn leveraged to provide a positive confirmation to the Hao-Ng isomorphism problem in the case of graph correspondences and arbitrary groups. For further reading on the subject please see [9, 10, 11]. In this paper, we provide a necessary condition for the hyperrigidity of a tensor algebra, that a C∗-correspondence cannot be σ-degenerate, and show 2010 Mathematics Subject Classification. 46L07, 46L08, 46L55, 47B49, 47L40. Key words and phrases: C∗-correspondence, tensor algebra, hyperrigid, topological graph, operator algebra. 1 2 E.G. KATSOULIS AND C. RAMSEY that this completely characterizes the situation where the C∗-correspondence is coming from a topological graph, which generalizes both the graph cor- respondence case and the semicrossed product arising from a multivariable dynamical system. 1.1. Regarding hyperrigidity. The reader familiar with the literature recognizes that in our definition of hyperrigidity, we are essentially asking that the restriction on A of any non-degenerate representation of C∗ env(A) possesses the unique extension property (abbr. UEP). According to [3, Proposition 2.4] a representation ρ : A → B(H), degenerate or not, has the UEP if and only if ρ is a maximal representation of A, i.e., whenever π is a representation of A dilating ρ, then π = ρ ⊕ π′ for some representation π′. Our definition of hyperrigidity is in accordance with Arveson's nomenclature [1], our earlier work [7, 11] and the works of Dor-On and Salomon [3] and Salomon [16], who systematized quite nicely the non-unital theory. An alternative definition of hyperrigidity for A may ask that any repre- sentation of C∗ env(A), not just the non-degenerate ones, possesses the UEP when restricted on A. It turns out that for operator algebras with a positive contractive approximate unit1, such a definition would be equivalent to ours [16, Proposition 3.6 and Theorem 3.9] . However when one moves beyond operator algebras with an approximate unit, there are examples to show that the two definitions differ. One such example is the non-unital operator algebra AV generated by the unilateral forward shift V . It is easy to see that AV is hyperrigid according to our definition and yet the zero map, as a representation on H = C, does not have the UEP. (See for instance [16, Example 3.4].) 2. Main results A C∗-correspondence (X, C, ϕX ) (often just (X, C)) consists of a C∗-algebra C, a Hilbert C-module (X, h , i) and a (non-degenerate) ∗-homomorphism ϕX : C → L(X) into the C∗-algebra of adjointable operators on X. An isometric (Toeplitz) representation (ρ, t, H) of a C∗-correspondence (X, C) consists of a non-degenerate ∗-homomorphism ρ : C → B(H) and a linear map t : X → B(H), such that ρ(c)t(x) = t(ϕX (c)(x)), and for all c ∈ C and x, x′ ∈ X. These relations imply that the C∗-algebra generated by this isometric representation equals the closed linear span of t(x)∗t(x′) = ρ((cid:10)x, x′(cid:11)), t(x1) · · · t(xn)t(y1)∗ · · · t(ym)∗, xi, yj ∈ X. Moreover, there exists a ∗-homomorphism ψt : K(X) → B, such that ψt(θx,y) = t(x)t(y)∗, 1which includes all operator algebras appearing in this paper HYPERRIGIDITY OF TENSOR ALGEBRAS 3 where K(X) ⊂ L(X) is the subalgebra generated by the operators θx,y(z) = xhy, zi, x, y, x ∈ X, which are called by analogy the compact operators. The Cuntz-Pimsner-Toeplitz C∗-algebra TX is defined as the C∗-algebra generated by the image of (ρ∞, t∞), the universal isometric representation. This is universal in the sense that for any other isometric representation there is a ∗-homomorphism of TX onto the C∗-algebra generated by this representation in the most natural way. The tensor algebra T + X of a C∗-correspondence (X, C) is the norm-closed subalgebra of TX generated by ρ∞(C) and t∞(X). See [14] for more on these constructions. Consider Katsura's ideal JX ≡ ker ϕ⊥ X ∩ ϕ−1 X (K(X)). An isometric representation (ρ, t) of (X, C, ϕX ) is said to be covariant (or Cuntz-Pimsner) if and only if ψt(ϕX (c)) = ρ(c), for all c ∈ JX. The Cuntz-Pimsner algebra OX is the universal C∗-algebra for all isometric covariant representations of (X, C), see [13] for further de- tails. Furthermore, the first author and Kribs [8, Lemma 3.5] showed that OX contains a completely isometric copy of T + X ) ≃ OX . In [11] a suffi- cient condition for hyperrigidity was developed, Katsura's ideal acting non- degenerately on the left of X. To be clear, non-degeneracy here means that ϕX(JX )X = X which by Cohen's factorization theorem implies that we actually have ϕX (JX)X = X. We turn now to the hyperrigidity of tensor algebras. X and C∗ env(T + Theorem 2.1 (Theorem 3.1, [11]). Let (X, C) be a C∗-correspondence with X countably generated as a right Hilbert C-module. If ϕX (JX) acts non- degenerately on X, then T + X is a hyperrigid operator algebra. The proof shows that if τ ′ : OX −→ B(H) is a completely contractive and completely positive map that agrees with a ∗-homomorphism of OX on T + X then the multiplicative domain of τ ′ must be everything. This is ac- complished through the multiplicative domain arguments of [2, Proposition 1.5.7] and the fact that by X being countably generated, Kasparov's Stabi- lization Theorem implies the existence of a sequence {xn}∞ n=1 in X so that Pk n=1 θxn,xn, k = 1, 2, . . . , is an approximate unit for K(X). After quite a lot of inequality calculations one arrives at the fact that all of T + X is in the multiplicative domain and thus so is OX . A C∗-correspondence (X, C) is called regular if and only if C acts faithfully on X by compact operators, i.e., JX = C. We thus obtain the following which also appeared in [11]. Corollary 2.2. The tensor algebra of a regular, countably generated C∗- correspondence is necessarily hyperrigid. 4 E.G. KATSOULIS AND C. RAMSEY We seek a converse to Theorem 2.1. Definition 2.3. Let (X, C) be a C∗-correspondence and let JX be Kat- sura's ideal. We say that ϕX (JX) acts σ-degenerately on X if there exists a representation σ : C → B(H) so that ϕX (JX )X ⊗σ H 6= X ⊗σ H. Remark 2.4. In particular, if there exists n ∈ N so that (ϕX (JX) ⊗ id)X ⊗n ⊗σ H 6= X ⊗n ⊗σ H. then by considering the Hilbert space K := X ⊗n−1 ⊗σ H, we see that ϕX (JX )X ⊗σ K 6= X ⊗σ K. and so ϕX (JX) acts σ-degenerately on X. The following gives a quick example of a σ-degenerate action. Note that this is possibly stronger than having a not non-degenerate action. Proposition 2.5. Let (X, C) be a C∗-correspondence. If (ϕX (JX )X)⊥ 6= {0}, then ϕX (JX ) acts σ-degenerately on X. Proof. Let 0 6= f ∈ (ϕX (JX )X)⊥. Let σ : C → B(H) be a ∗-representation and h ∈ H so that σ(cid:0)hf, f i1/2(cid:1)h 6= 0. Then, hf ⊗σ h, f ⊗σ hi = hh, σ((hf, f i)hi = kσ(cid:0)hf, f i1/2(cid:1)hk 6= 0. A similar calculation shows that 0 6= f ⊗σ h ∈ (ϕX (JX )X ⊗σ H)⊥ and we are done. We need the following Lemma 2.6. Let (X, C) be a C∗-correspondence and (ρ, t) an isometric representation of (X, C) on H. (i) (ii) If M ⊆ H is an invariant subspace for (ρ ⋊ t)(T + X ), then the re- striction (ρM, tM) of (ρ, t) on M is an isometric representation. If ρ(c)h = ψt(ϕX (c))h, for all c ∈ JX and h ∈ [t(X)H]⊥, then (ρ, t) is a Cuntz-Pimsner representation. Proof. (i) If p is the orthogonal projection on M, then p commutes with ρ(C) and so ρM(·) = pρ(·)p is a ∗-representation of C. Furthermore, for x, y ∈ X, we have tM(x)∗tM(y) = pt(x)∗pt(y)p = pt(x)∗t(y)p = pρ(hx, yi)p = ρM(hx, yi) and the conclusion follows. HYPERRIGIDITY OF TENSOR ALGEBRAS 5 (ii) It is easy to see on rank-one operators and therefore by linearity and continuity on all compact operators K ∈ K(X) that t(Kx) = ψt(K)t(x), x ∈ X. Now if c ∈ JX, then for any x ∈ X and h ∈ H we have ρ(c)t(x)h = t(ϕX (c)x)h = ψt(ϕX (c))t(x)h. By assumption ρ(c)h = ψt(ϕX (c))h, for any h ∈ [t(X)H]⊥ and the conclu- sion follows. Theorem 2.7. Let (X, C) be a C∗-correspondence. If Katsura's ideal JX acts σ-degenerately on X then the tensor algebra T + X is not hyperrigid. Proof. Let σ : C → B(H) so that ϕX (JX)X ⊗σ H 6= X ⊗σ H and let M0 := (ϕX (JX)X ⊗σ H)⊥. We claim that (1) (ϕX (JX) ⊗ I)M0 = {0}. Indeed for any f ∈ M0 and j ∈ JX we have (cid:10)(ϕX (j) ⊗ I)f , (ϕX (j) ⊗ I)f(cid:11) = hf, (ϕX (j∗j) ⊗ I)f i = 0 since f ∈ (ϕX (JX)X ⊗σ H)⊥. This proves the claim. We also claim that (2) (ϕX (C) ⊗ I)M0 = M0. Indeed this follows from the fact that (ϕX (C) ⊗ I)(ϕX (JX)X ⊗σ H) = ϕX (JX)X ⊗σ H, which is easily verified. Using the subspace M0 we produce a Cuntz-Pimsner representation (ρ, t) of (X, C) as follows. Let (ρ∞, t∞) be the universal representation of (X, C) on the Fock space F(X) = ⊕∞ n=0X ⊗n, X ⊗0 := C. Let ρ0 : C −→ B(F(X) ⊗σ H); c 7−→ ρ∞(c) ⊗ I t0 : X −→ B(F(X) ⊗σ H); x 7−→ t∞(x) ⊗ I. Define M : = 0 ⊕ M0 ⊕ (X ⊗ M0) ⊕ (X ⊗2 ⊗ M0) ⊕ . . . = (ρ0 ⋊ t0)(T + X )(0 ⊕ M0 ⊕ 0 ⊕ 0 ⊕ . . . ) ⊆ F(X) ⊗σ H, with the second equality following from (2). Clearly, M is an invariant subspace for (ρ0 ⋊ t0)(T + X ). Let ρ := ρ0M and t := t0M. By Lemma 2.6(i), (ρ, t) is a representation of (X, C). We claim that (ρ, t) is actually Cuntz-Pimsner. 6 E.G. KATSOULIS AND C. RAMSEY Indeed by Lemma 2.6(ii) it suffices to examine whether ψt(ϕX (j))h = ρ(j)h, for any h ∈ M ⊖ t(X)M. Note that since t(X)M = 0 ⊕ 0 ⊕ (X ⊗ M0) ⊕ (X ⊗2 ⊗ M0) ⊕ ..., we have that M ⊖ t(X)M = 0 ⊕ M0 ⊕ 0 ⊕ 0 ⊕ . . . . From this it follows that for any h ∈ M ⊖ t(X)M we have t0(x)∗h ∈ (C ⊗σ H) ⊕ 0 ⊕ 0 ⊕ ..., x ∈ X and so in particular for any j ∈ JX we obtain ψt(ϕX (j))h ∈ t0M(X)(t0M)(X)∗h = {0}. On the other hand, ρ(j)h ∈ 0 ⊕ (ϕX (JX) ⊗ I)M0 ⊕ 0 ⊕ 0 ⊕ · · · = {0}, because of (2). Hence (ρ, t) is Cuntz-Pimsner. At this point by restricting on T + X , we produce the representation ρ⋊t T + X coming from a ∗-representation of its C∗-envelope OX , which admits a is a non-trivial X ), then ρ⋊ t T + X . Proposition 2.4 [3] shows ρ ⋊ t T + of T + dilation, namely ρ0 ⋊ t0 T + dilation of ρ⋊ t T + is not a maximal representation of T + does not have the UEP and so T + . If we show now that ρ0 ⋊ t0 T + , i.e. M0 is not reducing for (ρ0 ⋊ t0)(T + X is not hyperrigid, as desired. X X X X X X Towards this end, note that M⊥ = C ⊕ (ϕX (JX)X ⊗σ H) ⊕ (X ⊗ M0)⊥ ⊕ . . . and so t0(X)M⊥ = 0 ⊕ (XC ⊗σ H) ⊕ 0 ⊕ 0 ⊕ · · · * M⊥ Therefore M⊥ is not an invariant subspace for (ρ0 ⋊ t0)(T + not a reducing subspace for (ρ0 ⋊ t0)(T + X ). This completes the proof. X ) and so M is 3. Topological graphs A broad class of C∗-correspondences arises naturally from the concept of a topological graph. For us, a topological graph G = (G0, G1, r, s) consists of two σ-locally compact spaces G0, G1, a continuous proper map r : G1 → G0 and a local homeomorphism s : G1 → G0. The set G0 is called the base (vertex) space and G1 the edge space. When G0 and G1 are both equipped with the discrete topology, we have a discrete countable graph. With a given topological graph G = (G0, G1, r, s) we associate a C∗- correspondence XG over C0(G0). The right and left actions of C0(G0) on Cc(G1) are given by (f F g)(e) = f (r(e))F (e)g(s(e)) HYPERRIGIDITY OF TENSOR ALGEBRAS 7 for F ∈ Cc(G1), f, g ∈ C0(G0) and e ∈ G1. The inner product is defined for F, H ∈ Cc(G1) by hF Hi (v) = X e∈s−1(v) F (e)H(e) for v ∈ G0. Finally, XG denotes the completion of Cc(G1) with respect to the norm (3) kF k = sup v∈G0 hF F i (v)1/2. When G0 and G1 are both equipped with the discrete topology, then associated with G coincides with the quiver the tensor algebra T + algebra of Muhly and Solel [14]. See [15] for further reading. G ≡ T + XG Given a topological graph G = (G0, G1, r, s), we can describe the ideal JXG as follows. Let sce = {v ∈ G0 v has a neighborhood V such that r−1(V ) = ∅} G0 G0 fin = {v ∈ G0 v has a neighborhood V such that r−1(V ) is compact} Both sets are easily seen to be open and in [12, Proposition 1.24] Katsura shows that ker ϕXG = C0(G0 sce) and ϕ−1 XG (K(XG)) = C0(G0 fin). From the above it is easy to see that JXG = C0(G0 reg), where G0 reg := G0 fin\G0 sce. We need the following Lemma 3.1. Let G = (G0, G1, r, s) be a topological graph. Then r−1(cid:0)G0 reg(cid:1) = G1 if and only if r : G1 → G0 is a proper map satisfying r(G1) ⊆ (cid:0)r(G1)(cid:1)◦ . Proof. Notice that r−1(G0 reg) = r−1(G0 fin) ∩ r−1(G0 sce)c fin) = r−1(G0 First we claim that r−1(G0 reg(cid:1) = G1 is equivalent to r−1(G0 and so r−1(cid:0)G0 fin) = G1 if and only if r is a proper map. Indeed, assume that r−1(G0 fin) = G1 and let K ⊆ r(G1) compact in the relative topology. For every x ∈ K, let Vx be a compact neighborhood of x such that r−1(Vx) is compact and so r−1(Vx ∩ K)) is also compact. By compactness, there exist x1, x2, . . . , xn ∈ K so that K = ∪n i=1(Vxi ∩ K) and so sce) = G1 and so r−1(K) is compact. r−1(K) = ∪n i=1r−1(Vxi ∩ K) Conversely, if r is proper then any compact neighborhood V of any point in G0 is inverted by r−1 to a compact set and so r−1(G0 fin) = G1. sce) = ∅ if and only r(G1) ⊆ (cid:0)r(G1)(cid:1)◦. We now claim that r−1(G0 8 E.G. KATSOULIS AND C. RAMSEY Indeed, e ∈ r−1(G0 ∅ is equivalent to as desired. sce) is equivalent to r(e) ∈ (r(G1)c)◦ and so r−1(G0 sce) = r(G1) ⊆ (cid:16)(r(G1)c)◦(cid:17)c = (cid:0)r(G1)(cid:1)◦ , If G = (G0, G1, r, s) is a topological graph and S ⊆ G1, then N (S) denotes the collection of continuous functions F ∈ XG with FS = 0, i.e., vanishing at S. The following appears as Lemma 4.3(ii) in [6]. Lemma 3.2. Let G = (G0, G1, r, s) be a topological graph. S2 ⊆ G1 closed, then If S1 ⊆ G0, N (r−1(S1) ∪ S2) = span{(f ◦ r)F fS1 = 0, FS2 = 0} Theorem 3.3. Let G = (G0, G1, r, s) be a topological graph and let XG the C∗-correspondence associated with G. Then the following are equivalent (i) the tensor algebra T + XG (ii) ϕ(JXG ) acts non-degenerately on XG (iii) r : G1 → G0 is a proper map satisfying r(G1) ⊆ (cid:0)r(G1)(cid:1)◦ is hyperrigid Proof. If ϕ(JXG) acts non-degenerately on XG, then Theorem 2.1 shows that T + XG is hyperrigid. Thus (ii) implies (i). For the converse, assume that ϕ(JXG ) acts degenerately on XG. If we verify that ϕ(JXG) acts σ-degenerately on XG, then Theorem 2.7 shows that T + XG is not hyperrigid and so (i) implies (ii). Towards this end note that JXG = C0(U ) for some proper open set U ⊆ reg but this is not really needed for this G0. (Actually we know that U = G0 part of the proof!) Hence (4) ϕ(JXG )XG = span{(f ◦ r)F fU c = 0} = N (r−1(U )c), according to Lemma 3.2. Since ϕ(JXG) acts degenerately on XG, (4) shows that r−1(U )c 6= ∅. Let e ∈ r−1(U )c and let F ∈ Cc(G1) ⊆ XG with F (e) = 1 and F (e′) = 0, for any other e′ ∈ G1 with s(e′) = s(e). Consider the one dimensional representation σ : C0(G0) → C coming from evaluation at s(e). We claim that ϕXG(JXG)XG ⊗σ C 6= XG ⊗σ C. Indeed for any G ∈ ϕ(JXG)XG = N (r−1(U )c) we have hF ⊗σ 1, G ⊗σ 1i = h1, σ(hF, Gi1) = hF, Gis(e) = X s(e′)=s(e) F (e′)G(e′) = F (e)G(e) = 0. HYPERRIGIDITY OF TENSOR ALGEBRAS 9 Furthermore, hF ⊗σ 1, F ⊗σ 1is(e) = F (e)2 = 1 and so 0 6= F ⊗σ 1 ∈ (ϕXG (JXG)XG ⊗σ C)⊥. This establishes the claim and finishes the proof of (i) implies (ii). Finally we need to show that (ii) is equivalent to (iii). Notice that (4) implies that ϕ(JXG) acts degenerately on XG if and only if The conclusion now follows from Lemma 3.1. r−1(U )c = r−1(G0 reg)c = ∅. The statement of the previous Theorem takes its most pleasing form when X is hyperrigid if and only is G1 is G0 is a compact space. In that case T + compact and r(G1) ⊆ G0 is clopen. Acknowledgement. Both authors would like to thank MacEwan University for providing project funding to bring the first author out to Edmonton for a research visit. The first author received support for this project in the form of a Summer Research Award from the Thomas Harriott College of Arts Sciences at ECU. He also expresses his gratitude to Adam Dor-On and Guy Salomon for several discussions regarding their work on hyperrigidity. The second author was partially supported by an NSERC grant. References [1] W. Arveson, The noncommutative Choquet boundary II: hyperrigidity, Israel J. Math. 184 (2011), 349 -- 385. [2] N. Brown and N. Ozawa, C∗ -- algebras and finite-dimensional approximations, Grad- uate Studies in Mathematics 88, American Mathematical Society, Providence, RI, 2008. xvi+509 pp. [3] A. Dor-On and G. Salomon, Full Cuntz-Krieger dilations via non-commutative bound- aries, J. London Math. Soc. 98 (2018), 416 -- 438. [4] B. Duncan, Certain free products of graph operator algebras, J. Math. Anal. Appl. 364 (2010), 534 -- 543. [5] E.T.A. Kakariadis, The Dirichlet property for tensor algebras, Bull. Lond. Math. Soc. 45 (2013), 1119 -- 1130. [6] E. Katsoulis, Local maps and the representation theory of operator algebras, Trans. Amer. Math. Soc. 368 (2016), 5377 -- 5397. [7] E. Katsoulis, C∗-envelopes and the Hao-Ng isomorphism for discrete groups, Inter. Math. Res. Not. (2017), 5751 -- 5768. [8] E. Katsoulis and D. Kribs, Tensor algebras of C ∗-correspondences and their C∗- envelopes, J. Funct. Anal. 234 (2006), 226 -- 233. [9] E. Katsoulis and C. Ramsey, Crossed products of operator algebras, Mem. Amer. Math. Soc 258 (2019), no. 1240, vii+85 pp. [10] E. Katsoulis and C. Ramsey, Crossed products of operator algebras: applications of Takai duality, J. Funct. Anal. 275 (2018), 1173 -- 1207. [11] E. Katsoulis and C. Ramsey, The non-selfadjoint approach to the Hao-Ng isomor- phism problem, preprint arXiv:1807.11425. [12] T. Katsura, A class of C∗-algebras generalizing both graph algebras and homeomor- phism C∗-algebras. I. Fundamental results, Trans. Amer. Math. Soc. 356 (2004), 4287 -- 4322. 10 E.G. KATSOULIS AND C. RAMSEY [13] T. Katsura, On C∗-algebras associated with C∗-correspondences, J. Funct. Anal. 217 (2004), 366 -- 401. [14] P. Muhly and B. Solel, Tensor algebras over C ∗-correspondences: representations, dilations, and C ∗-envelopes, J. Funct. Anal. 158 (1998), 389 -- 457. [15] I. Raeburn, Graph algebras, CBMS Regional Conference Series in Mathematics 103, American Mathematical Society, Providence, RI, 2005. [16] G. Salomon, Hyperrigid subsets of Cuntz-Krieger algebras and the property of rigidity at zero, J. Operator Theory 81 (2019), 61 -- 79. Department of Mathematics, East Carolina University, Greenville, NC 27858, USA E-mail address: [email protected] Department of Mathematics and Statistics, MacEwan University, Edmon- ton, AB, Canada E-mail address: [email protected]
1804.04558	1	1804	2018-04-12T15:09:41	Some rigidity results for II$_1$ factors arising from wreath products of property (T) groups	[ "math.OA", "math.GR" ]	We show that any infinite collection $(\Gamma_n)_{n\in \mathbb N}$ of icc, hyperbolic, property (T) groups satisfies the following von Neumann algebraic \emph{infinite product rigidity} phenomenon. If $\Lambda$ is an arbitrary group such that $L(\oplus_{n\in \mathbb N} \Gamma_n)\cong L(\Lambda)$ then there exists an infinite direct sum decomposition $\Lambda=(\oplus_{n \in \mathbb N} \Lambda_n )\oplus A$ with $A$ icc amenable such that, for all $n\in \mathbb N$, up to amplifications, we have $L(\Gamma_n) \cong L(\Lambda_n)$ and $L(\oplus_{k\geq n} \Gamma_k )\cong L((\oplus_{k\geq n} \Lambda_k) \oplus A)$. The result is sharp and complements the previous finite product rigidity property found in [CdSS16]. Using this we provide an uncountable family of restricted wreath products $\Gamma\cong\Sigma\wr \Delta$ of icc, property (T) groups $\Sigma$, $\Delta$ whose wreath product structure is recognizable, up to a normal amenable subgroup, from their von Neumann algebras $L(\Gamma)$. Along the way we highlight several applications of these results to the study of rigidity in the $\mathbb C^*$-algebra setting.	math.OA	math	SOME RIGIDITY RESULTS FOR II1 FACTORS ARISING FROM WREATH PRODUCTS OF PROPERTY (T) GROUPS IONUT CHIFAN AND BOGDAN TEODOR UDREA ABSTRACT. We show that any infinite collection (Γn)n∈N of icc, hyperbolic, property (T) groups satisfies the following von Neumann algebraic infinite product rigidity phenomenon. If Λ is an arbitrary group such that L(⊕n∈N Γn) ∼= L(Λ) then there exists an infinite direct sum decomposition Λ = (⊕n∈NΛn) ⊕ A with A icc amenable such that, for all n ∈ N, up to amplifications, we have L(Γn) ∼= L(Λn) and L(⊕k≥nΓk) ∼= L((⊕k≥nΛk) ⊕ A). The result is sharp and complements the previous finite product rigidity property found in [CdSS16]. Using this we provide an uncountable family of restricted wreath products Γ ∼= Σ ≀ ∆ of icc, property (T) groups Σ, ∆ whose wreath product structure is recognizable, up to a normal amenable subgroup, from their von Neumann algebras L(Γ). Along the way we highlight several applications of these results to the study of rigidity in the C∗-algebra setting. 1. INTRODUCTION For a countable infinite group Γ we denote by ℓ2Γ the Hilbert space of all square sum- mable complex functions on Γ. Each element γ ∈ Γ gives rise to a unitary operator uγ : ℓ2Γ → ℓ2Γ by group translation uγ(ξ)(λ) = ξ(γ−1λ), where λ ∈ Γ and ξ ∈ ℓ2Γ. The bicommutant {uγγ ∈ Γ} ′′ inside the algebra of all bounded linear operators B(ℓ2 Γ), is denoted by L(Γ) and it is called the group von Neumann algebra of Γ. The algebra L(Γ) is a II1 factor (has trivial center) precisely when all nontrivial conjugacy classes of Γ are infinite (icc), this being the most interesting for study [MvN43]. Ever since their introduction, the classification of these factors is a core direction of research driven by the following fundamental question: What aspects of the group Γ are remembered by L(Γ)? This emerged as an interesting yet intriguing theme since these al- gebras tend to have little memory of the initial group. This is best illustrated by Connes' celebrated result asserting that all amenable icc groups give isomorphic factors, [Co76]. Hence very different groups like the group of all finite permutations of the positive in- tegers, the lamplighter group, or the wreath product of the integers with itself give rise to isomorphic factors. Consequently, the von Neumann algebric structure has no mem- ory of the typical discrete algebraic group invariants like torsion, rank, or generators and relations. In this case the only information the von Neumann algebra retains is the amenability-an approximation property-of the group. In the non-amenable case the situation is radically different and an unprecedented progress has been achieved through the emergence of Popa's deformation/rigidity the- ory [Po06]. Using this completely new conceptual framework it was shown that various properties of groups, such as their representation theory or their approximations, can be completely recovered from their von Neumann algebras. As a result, for large classes of I.C. was supported by NSF Grant DMS #1600688. 1 group factors, many remarkable structural properties such as primeness, (strong) solid- ity, classification of normalizers of algebras, etc could be successfully established [Po03, IPP05, Po06, Po08, OP07, OP08, CH08, CI08, Pe09, PV09, FV10, Io10, IPV10, HPV10, CP10, Si10, Va10, CS11, CSU11, Io11, PV11, HV12, PV12, Io12, Bo12, BHR12, Is12, BV13, Va13, Is14, CIK13, VV14, BC14, CKP14, CdSS16, DHI16, CdSS17]. For additional information we refer the reader to the following survey papers [Po06, V10, Io12, Io17]. One of the most impressive milestone in this study is Ioana-Popa-Vaes' discovery of the first examples of groups that can be completely recovered from their von Neumann algebras (W∗-superrigid1 groups), [IPV10]. See also the subsequent result [BV13] and the more recent work [CI17]. These results pushed the classification problem of group factors to new boundaries and exciting possibilities. In this direction an interesting and wide open theme is to identify a comprehensive list of canonical constructions in group theory (direct sum, free product, HNN-extension, wreath product, etc) that are recoverable from their von Neumann algebras. 1.1. Statements of main results. Over a decade ago Ozawa and Popa discovered the first unique prime factorization results for tensor product of II1 factors, [OP03]. This work has had deep consequences to the classification of II1 factors and has generated significant subsequent developments. Some of Ozawa-Popa's results have been strengthened con- siderably in [CdSS16] where was unveiled a large class of product groups Γ1 × Γ2 whose product structure is a feature completely recognizable at the level of their von Neumann algebras L(Γ1 × Γ2). Precisely, whenever Γ1, Γ2 are hyperbolic icc groups (e.g. non-abelian free groups) and Λ is an arbitrary group such that L(Γ1 × Γ2) = L(Λ) then Λ admits a non- trivial product decomposition Λ = Λ1 × Λ2 and there exists a scalar t > 0 such that, up to unitary conjugacy, we have L(Γ1) = L(Λ1)t and L(Γ2) = L(Λ2)1/t. The result still holds if one assumes, more generally, that Γ1, Γ2 are just icc biexact groups, [Oz03]. Isono studied unique prime factorization aspects for infinite tensor product of factors and several interesting results have emerged in [Is16]. Motivated in part by these re- sults it is natural to investigate whether "product rigidity" properties, similar with ones in [CdSS16], would hold in the context of infinite direct sums groups. Specifically, if one considers Γ = ⊕n∈NΓn with Γn's icc non-amenable groups, it would be interest- ing to understand how much of the infinite direct sum structure of Γ is retained by its von Neumann algebra L(Γ). Right away one may notice a sharp contrast point with the aforementioned finite product situation. Since L(Γ) canonically decomposes as an infinite tensor product L(Γ) = ¯⊗n∈NL(Γn) it follows that L(Γ) is a McDuff factor and hence L(Γ) = L(Γ) ¯⊗R, where R is the hyperfinite factor; consequently, we have that L(⊕n∈N Γn) = L((⊕n∈N Γn) ⊕ A), for any icc amenable group A. This observation shows that, in the best case scenario, L(Γ) could remember the direct sum feature of the under- lying group only up to an amenable subgroup which typically lies in the tail of the infinite tensor product. It is therefore natural to investigate under which circumstances it is pos- sible to completely reconstruct the infinite direct sum feature only up to this obstruction. Building upon previous techniques from [IPV10, Io11, CdSS16, DHI16, CdSS17] and us- ing the classification of normalizers from [PV12] we found infinitely many classes of Γn's for which this problem has a positive answer. 1Γ is W∗-superrigid if whenever Λ is an arbitrary group so that L(Γ) = L(Λ) then it follows that Λ = Γ. 2 Theorem A. Let (Γn)n∈N an infinite collection of property (T), biexact, weakly amenable, icc groups. Assume that Λ is an arbitrary group satisfying L(⊕n∈N Γn) = L(Λ). Then Λ admits an infinite direct sum decomposition Λ = (⊕n∈NΛn) ⊕ A, where Λn is icc, weakly amenable, property (T) group for all n and A is a icc amenable group. Moreover, for each k ∈ N there exist scalars t1, t2, ..., tk+1 > 0 satisfying t1t2...tk+1 = 1 and a unitary u ∈ L(Λ) so that uL(Γn)tnu∗ = L(Λn) for all k ≥ n ≥ 1; and uL(⊕n≥k+1Γn)tk+1u∗ = L((⊕n≥k+1Λn) ⊕ A). The result applies to several concrete classes of groups a such as: (1) the uniform lattices Γn in Sp(kn, 1) with kn ≥ 2 or any icc groups in their measure equivalence class; and (2) Gromov's random groups with density satisfying 3−1 < d < 2−1. While the conclusion of the previous theorem shows a strong identification of the von Neumann algebras of the group factors Γn's, in general one cannot recover these groups. To see this note that Voiculescu's compression formula for free group factors gives L(F2) = L(F5) ⊗ M2(C). This implies that L(⊕n∈NF2) = ¯⊗n∈NL(F2) = ¯⊗n∈N(L(F5) ⊗ M2(C)) = ( ¯⊗n∈N(L(F5)) ¯⊗R = L((⊕n∈N F5) ⊕ A), for every icc amenable group A. Theorem A can be successfully used to shed light towards rigidity aspects in the C∗- algebraic setting. Precisely, when it is combined with [BKKO14, Theorem 1.3] one gets the following version of infinite product rigidity for reduced group C∗-algebras. Corollary B. Let (Γn)n∈N an infinite collection of property (T), biexact, weakly amenable, icc groups. Assume that Λ is an arbitrary group satisfying C∗ r (Λ). Then Λ admits an infinite direct sum decomposition Λ = ⊕n∈NΛn, where the Λn's are icc, weakly amenable, property (T) groups. r (⊕n∈NΓn) = C∗ When compared side by side, Theorem A and Corollary B highlight again the funda- mental difference between C∗-algebras and von Neumann algebras; the absence of the infinite amenable direct summand of Λ in the conclusion of Corollary B exemplifies once more the fact that the WOT closure is considerably larger than the normic closure, thus triggering significant loss of algebraic information in the von Neumann algebraic setting. Restricted wreath product groups manifest a remarkable rigid behavior in the von Neu- mann algebraic setting. In fact a large portion of the groups/actions known to be recon- structible of from their von Neumann algebras arise from constructions involving wreath product groups or Bernoulli shifts [Po03, Po05, PV09, Io10, IPV10, PV11, BV13]. A com- mon feature of these examples is that the core or the wreath product groups involved are amenable and in many cases even abelian. For example, with the exception of [CI17], all known examples of W∗-superrigid2 groups are of the form H ≀I Γ where H is finite [IPV10, BV13, B14]. However significantly fewer rigidity results are known in general for wreath product factors L(H ≀I Γ) when H is nonamenable. In this direction we men- tion in passing Ioana's strong rigidity results which asserts that L(Fn ≀ A) 6= L(Fm ≀ B) whenever n, m ≥ 2 and A, B are nonisomorphic icc amenable groups, [Io06]. Theorem A can be successfully used to provide new insight towards this problem as well. For in- stance, using it in combination with various technical outgrowths of previous methods 2a group K is W∗-superrigid if whenever T is an arbitrary group such that L(K) = L(T) then K = T. 3 from [Po03, Io06, IPV10, CdSS16, CI17] we obtain the following wreath product rigidity result up to an amenable subgroup for group factors. Theorem C. Let H be icc, weakly amenable, biexact property (T) group. Let Γ = Γ1 × Γ2, where Γi are icc, biexact, property (T) group. Let H ≀ Γ be the corresponding (plain) wreath product. Let Λ be an arbitrary group and let θ : L(H ≀ Γ) → L(Λ) be a ∗-isomorphism. Then one can find non-amenable icc groups Σ, Ψ, an amenable icc group A, and an action Ψ yα A such that we can decompose Λ as semidirect product Λ = (Σ(Ψ) ⊕ A) ⋊β⊕α Ψ, where Ψ yβ Σ(Ψ) is the Bernoulli shift. In addition, there exist a group isomorphism δ : Γ → Ψ, a character η : Γ → T, a ∗-isomorphism θ : L(H(Γ)) → L(Σ(Ψ) ⊕ A) and a unitary u ∈ L(Λ) so that for all x ∈ L(H(Γ)), γ ∈ Γ we have θ(xuγ) = η(γ)uθ(x)vδ(γ) u∗. Here {uγ γ ∈ Γ} and {vλ λ ∈ Ψ} are the canonical group unitaries of L(Γ) and L(Ψ), respec- tively. We notice the theorem still holds for slightly more general situations, e.g. generalized wreath products with finite or even possible just amenable stabilizers (see Theorem 5.4). However at this time we do not know the full extent of these cases as it seems to rely on heavy constructions on group theory (see Remarks 5.5). As before, Theorem C in combination with [BKKO14, Theorem 1.3] lead to the fol- lowing stronger version of wreath product rigidity in the context of reduced group C∗- algebras. Corollary D. Let H be icc, weakly amenable, biexact property (T) group. Let Γ = Γ1 × Γ2, where Γi are icc, biexact, property (T) group. Let H ≀ Γ be the corresponding wreath product. Let Λ be an arbitrary group so that C∗ r (Λ). Then Λ = Σ ≀ Γ, where Σ is an icc, weakly amenable, property (T) group. r (H ≀ Γ) = C∗ In connection with Theorem C it is natural to investigate whether similar statements hold if one relaxes the biexactness assumptions on H or the product assumption Γ. In this situation, building upon the previous techniques from [IPV10, KV15] one can show the following strong rigidity statement holds just for property (T) groups H and Γ. Theorem E. Let H, Γ be icc, torsion free groups. Also assume H has property (T) and Γ admits an infinite, almost normal subgroup with relative property (T). Let Γ y I be a transitive action on a countable set satisfying the following conditions: a) The stabilizer StabΓ(i) has infinite index in Γ for each i ∈ I; b) There is k ∈ N such that for each J ⊆ I satisfying J ≥ k we have StabΓ(J) < ∞; c) The orbit StabΓ(i) · j is infinite for all i 6= j. Denote by G = H ≀I Γ the corresponding generalized wreath product. Let Λ be any torsion free group and let θ : L(G) → L(Λ) be a ∗-isomorphism. Then Λ admits a generalized wreath product decomposition Λ = Σ ≀I Ψ satisfying all the properties enumerated in a) − c). In addition, there exist a group isomorphism δ : Γ → Ψ, a character η : Γ → T, a ∗-isomorphism θ : L(H) → L(Σ) and a unitary u ∈ L(Λ) such that for every x ∈ L(H(I)) and γ ∈ Γ we have θ(xuγ) = η(γ)uθ⊗I (x)vδ(γ)u∗. Here {uγγ ∈ Γ} and {vλλ ∈ Ψ} are the canonical unitaries of L(Γ) and L(Ψ), respectively. 4 The previous theorem applies to many natural families of generalized wreath groups H ≀I Γ, including: a) any icc torsion free prop (T ) group H and any torsion free, hyperbolic, property (T) group Γ together with a maximal amenable subgroup Σ < Γ and the action Γ y I = Γ/Ω by translation on right cosets Γ/Ω; b) (see [IPV10]) Let H be any icc property (T) group and Γ = Z2 ⋊ SL2(Z). Take a matrix A in SL2(Z) having the modulus of the eigenvalues larger than one, and the subgroup S = {B ∈ SL2(Z) BAB−1 = ±A}. The action Γ y I = Γ/S is given by left translations. 1.2. Acknowledgments. The first author is grateful to the American Institute of Mathe- matics for their kind hospitality during the workshop "Classification of group von Neu- mann algebras" where part of this work was carried out. The first author is also indebted to Professor Denis Osin for many discussions and suggestions regarding this project. 2. PRELIMINARIES 2.1. Notations. Given a von Neumann algebra M we will denote by U (M) its unitary group and by P(M) the set of all its nonzero projections. All algebras inclusions N ⊆ M are assumed unital unless otherwise specified. For any von Neumann subalgebras P, Q ⊆ M we denote by P ∨ Q the von Neumann algebra they generate in M. All von Neumann algebras M considered in this article will be tracial, i.e., endowed with a unital, faithful, normal functional τ : M → C satisfying τ(xy) = τ(yx) for all x, y ∈ M. This induces a norm on M by the formula kxk2 = τ(x∗x)1/2 for all x ∈ M. The k · k2-completion of M will be denoted by L2(M). For a countable group Γ we denote by {uγγ ∈ Γ} ∈ U(ℓ2 Γ) its left regular represen- tation given by uγ(δλ) = δγλ, where δλ : Γ → C is the Dirac mass at {λ}. The weak operatorial closure of the linear span of {uγγ ∈ Γ} in B(ℓ2 Γ) is the so called group von Neumann algebra and will be denoted by L(Γ). L(Γ) is a II1 factor precisely when Γ has infinite non-trivial conjugacy classes (icc). Given a group Γ and a subset F ⊆ Γ we will be denoting by hFi the subgroup of Γ generated by F. Given a group action Γ y I on a countable set I, for any subset F ⊂ I we denote Stab(F ) = {γ ∈ Γ g · i = i, ∀i ∈ F } and Norm(F ) = {γ ∈ Γ γ · F = F }. Given a subgroup Λ 6 Γ we will often consider the virtual centralizer of Λ in Γ, i.e. vCΓ(Λ) = {γ ∈ Γ : γΛ < ∞}. Notice vCΓ(Λ) is a subgroup normalized by Λ. When Λ = Γ, the virtual centralizer is denoted by vZ(Γ) := vCΓ(Γ) and called the virtual center of Γ; this is nothing else but the FC-radical of Γ. Hence Γ is icc precisely when vZ(Γ) = 1. 2.2. Popa's intertwining techniques. Over a decade ago, Popa introduced in [Po03, The- orem 2.1 and Corollary 2.3] a powerful analytic criterion for identifying intertwiners be- tween arbitrary subalgebras of tracial von Neumann algebras. This is now termed Popa's intertwining-by-bimodules technique. Theorem 2.1. [Po03] Let (M, τ) be a separable tracial von Neumann algebra and let P, Q ⊆ M be (not necessarily unital) von Neumann subalgebras. Then the following are equivalent: (1) There exist p ∈ P(P), q ∈ P(Q), a ∗-homomorphism θ : pPp → qQq and a partial (2) For any group G ⊂ U (P) such that G ′′ = P there is no sequence (un)n ⊂ G satisfying isometry 0 6= v ∈ qMp such that θ(x)v = vx, for all x ∈ pPp. kEQ(xuny)k2 → 0, for all x, y ∈ M. 5 If one of the two equivalent conditions from Theorem 3.8 holds then we say that a corner of P embeds into Q inside M, and write P ≺M Q. If we moreover have that Pp ′ ≺M Q, for any projection 0 6= p ′ ∈ P ′ ∩ 1P M1P (equivalently, for any projection 0 6= p ′ ∈ Z(P ′ ∩ 1P M1P)), then we write P ≺s 2.3. Quasinormalizers of groups and algebras. Given groups Ω 6 Γ, the one-side quasi- (1) Γ (Ω) ⊆ Γ is the set of all γ ∈ Γ for which there is a finite normalizer semigroup QN (1) Γ (Ω) iff [Ω : set F ⊆ Γ so that Ωγ ⊆ FΩ, [JGS10, Section 5]; equivalently, γ ∈ QN (1) Γ (Ω) coincides with the one-side commensurator of Ω in Γ. Similarly, the quasinormalizer (also called the commensurator) QNΓ(Ω) is the set of all γ ∈ Γ for which there exists a finite set F ⊆ Γ such that Ωγ ⊆ FΩ and γΩ ⊆ ΩF; equivalently, γΩγ−1 ∩ Ω] < ∞. Thus QN M Q. (1) Γ (Ω). γ ∈ QNΓ(Ω) iff [Ω : γΩγ−1 ∩ Ω] < ∞ and [γΩγ−1 : γΩγ−1 ∩ Ω] < ∞. From definitions one checks that QNΓ(Ω) 6 Γ is a subgroup satisfying Ω ⊆ QNΓ(Ω) ⊆ QN The von Neumann algebraic counterparts of (one-sided) quasinormalizers have played a major role in the recent classification results in this area [Po99, Po01, Po03, IPP05]. Given an inclusion Q ⊆ M, the quasi-normalizer QNM(Q) is the ∗-algebra of all elements x ∈ M such that there exist x1, x2, ..., xk ∈ M so that Qx ⊆ Pi xiQ and xQ ⊆ Pi Qxi, [Po99]. The von Neumann algebra QNM(Q) ′′ is called the quasi-normalizing algebra of Q inside (1) M (Q) is the set of all x ∈ M such that there M. Similarly, the intertwiner space QN exist x1, x2, ..., xk ∈ M so that Qx ⊆ Pi xiQ, [Po99, JGS10]. The von Neumann algebra (1) M (Q) ′′ is called the one-sided quasinormalizer algebra of Q inside M. QN As usual NM(Q) = {u ∈ U (M) uQu∗ = Q} denotes the normalizing group and NM(Q) ′′ denotes the normalizing algebra of Q in M. Notice that Q ⊆ NM(Q) ⊆ QNM(Q) ⊆ QN The following computations of (one-sided) quasinormalizer algebras for inclusions of group von Neumann algebras will be essential for our arguments. Theorem 2.2 (Corollary 5.2 [JGS10]). If Ω 6 Γ is an inclusion of groups, the following hold: (1) M (Q) ⊆ M. (1) QNL(Γ)(L(Ω)) ′′ = L(H1), where H1 = QNΓ(Ω) = QN (1) Γ (Ω)i 6 Γ. (2) QN (1) L(Γ)(L(Ω)) ′′ = L(H2), where H2 = hQN (1) Γ (Ω) ∩ QN (1) Γ (Ω)−1; Theorem 2.3 (Theorem [Po03]; Proposition 6.2 [JGS10]). Let Q ⊆ M be an inclusion of tracial von Neumann algebras. Then for any p ∈ P(Q) we have (1) QNpMp(pQp) ′′ = pQNM(Q) ′′ p, and (2) QN (1) pMp(pQp) ′′ = pQN (1) M (Q) ′′ p. 2.4. Height of elements in group von Neumann algebras. Following [Io10, Section 4] and [IPV10, Section 3] the height hΓ(x) of an element x ∈ L(Γ) is absolute value of the largest Fourier coefficient, i.e., hΓ(x) = supγ∈Γ τ(xuγ), where {uγγ ∈ Γ} are the canonical unitaries of M implemented by Γ. For a subset S ⊆ L(Γ), we denote by hΓ(S) = infx∈S hΓ(x). In this section we prove two elementary lemmas on height that will be used in the proof Theorem C. 6 Lemma 2.4. Assume that L(Γ) = M and consider the subsets Σ ⊆ Γ and S ⊆ (M)1. If there exist x, y ∈ M such that hΣ(xS y) > 0 then one can find a finite subset F ⊂ Λ such that hFΣF(S) > 0. In particular, we have hΛ(S) > 0 iff hΛ(uS u∗) > 0 for some u ∈ U (M). Proof. Using Kaplansky's Theorem for every ε > 0 there is a finite subset Fε ⊂ Λ and xε, yε ∈ (M)1 supported on Fε so that kx − xεk2, ky − yεk2 ≤ ε. Using these estimates together with triangle inequality, for every s ∈ S and γ ∈ Σ we have τ(xsyuγ) ≤ 4ε + τ(xεsyεuγ) ≤ 4ε + X λ,µ∈Fε τ(xεuλ−1 )τ(yεuµ−1)τ(uλsuµuγ) ≤ 4ε + Fε2 max ν∈FεγFε τ(suν). This implies that hΣ(xS y) ≤ 4ε + Fε2hFεΣFε(S) and hence hFεΣFε(S) ≥ (hΣ(xS y) − 4ε)Fε −2. Letting ε > 0 small enough we get the first part conclusion. For the remaining part, the reverse implication follows from above. The direct implica- (cid:3) tion follows from the reverse implication by replacing S with uS u∗ and u with u∗. Lemma 2.5. Assume that L(Γ) = L(Λ) = M. Consider the comultiplication along Λ i.e. the embedding ∆ : M → M ¯⊗M given by ∆(vλ) = vλ ⊗ vλ, where {vλ λ ∈ Λ} are the canonical group unitaries of M implemented by Λ. If S ⊆ (M)1 and there are x, y ∈ (M ¯⊗M)1 so that hΓ×Γ(x∆(S)y) > 0 then hΛ(S) > 0. Proof. Since hΓ×Γ(x∆(S)y) > 0 then Lemma 2.4 implies that (2.1) Fix s ∈ S. Let s = Pλ∈Λ τ(svλ−1 )vλ and note that ∆(s) = Pλ∈Λ τ(svλ−1 )vλ ⊗ vλ. Using these formulas and Cauchy-Schwarz inequality, for any γ1, γ2 ∈ Γ, we have hΓ×Γ(∆(S)) > 0. τ(vλuγ1)τ(vλuγ2) ≤ hΛ(s)kuγ1 k2kuγ2 k2 = hΛ(s). λ∈Λ This further implies hΓ×Γ(∆(S)) ≤ hΛ(S) and using (2.1) we get the conclusion. (cid:3) 2.5. Relative amenability. A tracial von Neumann algebra (M, τ) is called amenable if there exists a state φ : B(L2(M)) → C such that φM = τ and φ is M-central (i.e. φ(xT) = φ(Tx) for all x ∈ M, T ∈ B(L2(M))), [Co76]. Making use of the basic construc- tion for inclusions of algebras [Ch79, Jo81] this concept was further generalized in [OP07] to subalgebras. Let (M, τ) be a tracial von Neumann algebra, p ∈ M be a projection, and P ⊆ pMp, Q ⊆ M be von Neumann subalgebras. Following [OP07, Section 2.2] we say that P is amenable relative to Q inside M if there exists a P-central state φ : phM, eQip → C such that φ(x) = τ(x), for all x ∈ pMp. Here hM, eQi denotes the basic construction for the inclusion Q ⊆ M, i.e. the commutant of the Q-right action on B(L2(M)) [Jo81]. In this section we prove a relative amenability result for subalgebras that "cluster at infinity" in an infinite tensor product of factors. The result will be essentially used to derive our infinite product rigidity result for group factors. Our proof is an adaptation of an argument due to Ioana. See also [Is16, Lemma 4.4] and [HU15, Proposition 4.2] for similar results. 7 τ ⊗ τ(∆(s)(uγ1 ⊗ uγ2)) ≤ X ≤ hΛ(s) X λ∈Λ τ(svλ−1 )τ(vλuγ1)τ(vλuγ2) Proposition 2.6. Let M ⊆ ( M, τ) be finite von Neumann algebras satisfying the following: (1) there are subalgebras B, C ⊆ M and C ⊆ C ⊆ M so that M = B ∨ C and M = B ∨ C; (2) there are descending family Bn ⊆ M of subalgebras and an ascending family Cn ⊆ M of subalgebras such that ∪nCn = C, ∩nBn = B and M = Bj ∨ Cj for all j. (3) there exist θ, θn ∈ Aut( M) such that θnBn = idBn for all n, θn → θ pointwise, Let p ∈ B be a nonzero projection and let A ⊆ pMp be a von Neumann subalgebra so that for each n ∈ N there is un ∈ U (pMp) satisfying un Au∗ n ⊆ Bn. Consider the M-M bimodule H = L2( M) given by the actions x · ξ · y = xξθ(y) for all x, y ∈ M and ξ ∈ L2( M). Then there exists a sequence of vectors (ξn)n ⊂ L2(p Mp) satisfying lim n lim n kx · ξn − ξn · xkH = 0, for all x ∈ A, and hx · ξn, ξniH = τ(x), for all x ∈ pMp. (2.2) (2.3) n ⊆ Bn and θnBn = idBn then for every x ∈ A we have θn(unxu∗ nθn(un)θn(x) = xu∗ nθn(un) and letting ξn = u∗ n) = nθn(un) ∈ U (p M p) n. This implies that u∗ Proof. Since un Au∗ unxu∗ we conclude that for all x ∈ A and n ∈ N we have ξnθn(x) = xξn. Since un ∈ U (pMp) and θnBn = idBn we have that kξnkH = ku∗ n. Since kunk ∞ x ∈ A we have (2.4) nθn(un)k2 = kpk2 for all ≤ 1 then using (2.4) and θn → θ pointwise one can check that for every lim n kx · ξn − ξn · xkH = lim n ≤ lim n kxξn − ξnθ(x)k2 = lim n kξn(θn(x) − θ(x))k2 kξnk kθn(x) − θ(x)k2 = lim n ∞ kθn(x) − θ(x)k2 = 0. n xξn) = τ(xξnξ∗ Finally, since ξn ∈ U (p M p) we have hx · ξn, ξniH = τ(ξ∗ n) = τ(x) for all x ∈ pMp. Altogether, the above relations give the desired conclusion. (cid:3) Proposition 2.7. Let M = ¯⊗i∈N Mi ¯⊗B. Let A ⊆ M be a von Neumann algebra for which there n ⊆ ¯⊗i≥kn Mi ¯⊗B exist sequences (kn)n ⊆ N and (un)n ⊂ U (M) such that kn ր ∞ and un Au∗ for all n. Then A is amenable relative to B inside M. Proof. Denote by ¯⊗i∈N Mi = C and notice that M = C ¯⊗B. Let C = C ¯⊗C and M = C ¯⊗B and notice that M ⊂ M. For every n ∈ N denote by Cn = ¯⊗kn−1 i=1 Mi, by Dn = ¯⊗i≥kn Mi and by Bn = Dn ¯⊗B and notice that ∪nCn = C and ∩nBn = B. Next let θn ∈ Aut( M) satisfying θn(x ⊗ y) = y ⊗ x for all x, y ∈ Cn and θn = id on Dn ¯⊗Dn ¯⊗B. Notice that θn → θ pointwise, where θ ∈ Aut( M) satisfies θ(x ⊗ y) = y ⊗ x for all x, y ∈ C and θ = id on B. One can check all the conditions in the statement of Proposition 2.6 are satisfied. Thus if we consider the M-M bimodule H := L2( M) = L2(C) ¯⊗L2(C) ¯⊗L2(B) with the actions given by x · ξ · y = xξθ(y) for all x, y ∈ M and ξ ∈ H there exist a sequence of unit vectors (ξn)n ∈ H such that lim n lim n kx · ξn − ξn · xkH = 0, for all x ∈ A hx · ξn, ξniH = lim n hξn · x, ξniH = τ(x), for all x ∈ M. (2.5) Let hM, e1⊗Bi be the basic construction for 1 ¯⊗B ⊂ M and let Tr be the semifinite trace on hM, e1⊗Bi. Next we notice that, as M-M-bimodules, L2(hM, e1⊗Bi, Tr) is isomorphic to H 8 via the map (x ⊗ y)e1⊗B(z ⊗ 1) → (x ⊗ y) · (1 ⊗ 1 ⊗ 1) · (z ⊗ 1), for x, z ∈ C and y ∈ B. Indeed it is clear this is M-M-bimodular and also for all xi, zi ∈ C and yi ∈ B we have 2 ⊗ 1)E1⊗B(x∗ h(x1 ⊗ y1)e1⊗B(z1 ⊗ 1), (x2 ⊗ y2)e1⊗B(z2 ⊗ 1)iTr = = Tr((z2 ⊗ 1)∗e1⊗B(x2 ⊗ y2)∗(x1 ⊗ y1)e1⊗B(z1 ⊗ 1)) = Tr((z1z∗ =τC(x∗ =τC(x∗ =h(x1 ⊗ 1 ⊗ y1)(1 ⊗ 1 ⊗ 1)θ(z1 ⊗ 1), (x2 ⊗ 1 ⊗ y2)(1 ⊗ 1 ⊗ 1)θ(z2 ⊗ 1))iH =h(x1 ⊗ y1) · (1 ⊗ 1 ⊗ 1) · (z1 ⊗ 1), (x2 ⊗ y2) · (1 ⊗ 1 ⊗ 1) · (z2 ⊗ 1)iH. 2 x1)τC⊗B(z1z∗ 2 x1)τC(z∗ 2 ⊗ y∗ 2z1)τB(y∗ 2y1) 2y1) 2 x1 ⊗ y∗ 2y1)e1⊗B) This combined with (2.5) and [OP07, Theorem 2.1] show that A is amenable relative to B inside M. (cid:3) 3. PROOF OF THEOREM A This section is devoted to the proof of Theorem A. In essence this result is an infinite analog of the "product rigidity" phenomenon for group factors found in [CdSS16]. In fact our methods build upon the general strategy developed in [CdSS16] and still use in a cru- cial way the ultrapower techniques from [Io11] as well as the intertwining/combinatorial aspects developed in [OP03, IPV10, CdSS16, DHI16, CdSS17] and the classification of nor- malizers from [PV12]. Since our exposition will focus primarily on the novel aspects of these techniques we recommend the reader to consult the aforementioned works as some of these results will be heavily used throughout the section. To ease our exposition we first introduce the following notation: Notation 3.1. Let {Γi}i∈I be a collection of icc, weakly amenable, biexact groups and de- note by Γ = ⊕i∈I Γi. For any subset S ⊆ I, we denote ΓS = ⊕i∈SΓi. Denote by M = L(Γ), let t > 0 be a scalar, and assume that Mt = L(Λ) for an arbitrary group Λ. Following [IPV10], let ∆ : Mt → Mt ¯⊗Mt be the commultiplication along Λ, i.e. ∆(vλ) = vλ ⊗ vλ, where {vλ}λ∈Λ are the canonical unitaries generating L(Λ). Proposition 3.2. Assume Notation 3.1. Then for every i ∈ I there exists j ∈ I such that ∆(L(Γ I\{i})t) ≺Mt ¯⊗Mt Mt ¯⊗L(Γ I\{j})t. Proof. Using that Γi's are weakly amenable and biexact we show next the following Claim 3.3. For every i, j ∈ I one of the following holds: I\{i})t) ≺Mt ¯⊗Mt Mt ¯⊗L(Γ a) ∆(L(Γ b) ∆(L(Γi )) ≺Mt ¯⊗Mt Mt ¯⊗L(Γ I\{j})t. I\{j})t, or Proof of Claim 3.3. One has the following decomposition Mt ¯⊗Mt = Mt ¯⊗L(Γ I\{j})t ¯⊗L(Γj). Fix A ⊂ ∆(L(Γi)) a diffuse amenable subalgebra. Using [PV12, Theorem 1.6], we have either c) A ≺Mt ¯⊗Mt Mt ¯⊗L(Γ d) NMt ¯⊗Mt(A) ′′ is amenable relative to Mt ¯⊗L(Γ I\{j})t, or 9 I\{j})t. Suppose d) holds. As ∆(L(Γ either I\{i})t) ⊆ NMt ¯⊗Mt(A) ′′, then [PV12, Theorem 1.6] implies I\{i})t) ≺ Mt ¯⊗L(Γ I\{j})t, or e) ∆(L(Γ f) NMt ¯⊗Mt ∆(L(Γ I\{i})t) ′′ is amenable relative to Mt ¯⊗L(Γ I\{j})t. I\{i}))t) ′′, then [IPV10, However, f) cannot hold. Indeed, since ∆(Mt) ⊆ NMt ¯⊗Mt(∆(L(Γ Theorem 7.2(2)] would imply Γi is finite, a contradiction. Hence, for every diffuse subal- gebra A ⊂ L(Γi), either c) or e) must occur. Using [BO08, Appendix], we get the claim. (cid:4) Now assume by contradiction the conclusion does not hold. By Claim 3.3, for every j ∈ I we have ∆(L(Γi )) ≺Mt ¯⊗Mt Mt ¯⊗L(Γ (3.1) Next we observe that Z(∆(L(Γi )) ′ ∩ Mt ¯⊗Mt) = C1. To see this, let z ∈ Z(∆(L(Γi )) ′ ∩ Mt ¯⊗Mt). Since ∆(L(Γ I\{i})t) ⊂ ∆(L(Γi )) ′ ∩ Mt ¯⊗Mt, one can check that z ∈ ∆(Mt) ′ ∩ Mt ¯⊗Mt. However, since Λ is icc we have ∆(Mt) ′ ∩ Mt ¯⊗Mt = C1 and our claim follows. I\{j}). Hence, applying [DHI16, Thus (3.1) further implies that ∆(L(Γi)) ≺s Mt ¯⊗Mt M ¯⊗L(Γ I\{j})t. Lemma 2.8 (2)], for every finite subset F ⊂ I we have ∆(L(Γi )) ≺Mt ¯⊗Mt Mt ¯⊗L(Γ I\F)t. (3.2) Next we show that (3.2) implies the following Claim 3.4. ∆(L(Γi)) is amenable relative to Mt ⊗ 1. Proof of Claim 3.4. Let In = {n, n + 1, n + 2, ...}. Since ∆(L(Γi)) ′ ∩ Mt ¯⊗Mt is a factor, then using [OP03, Proposition 12], for every n ∈ N there is tn > 0 and un ∈ U (Mt ¯⊗Mt) so that un∆(L(Γi ))u∗ n ⊂ (Mt ¯⊗L(ΓIn )t)tn. Naturally, we have the following inclusions Mt ¯⊗L(ΓIn )ttn ⊂ Mt ¯⊗L(ΓIn ) ¯⊗L(Γn−1) = Mt ¯⊗L(ΓIn−1 ). Thus, for every n ∈ I, there is un ∈ U (Mt ¯⊗Mt) so that un∆(L(Γi ))u∗ n ⊂ Mt ¯⊗L(ΓIn−1 ). (3.3) Thus the claim follows from (3.3) and Proposition 2.7. (cid:4) Finally, Claim 3.4 and [IPV10, Proposition 7.2] imply L(Γi) amenable, a contradiction. (cid:3) Proposition 3.5. Assume Notation 3.1. Then for all i ∈ I, there exists a non-amenable subgroup Λi 6 Λ with non-amenable centralizer CΛ(Λi) such that L(Γ Proof. This follows directly from Proposition 3.2 and [DHI16, Theorem 4.1], (see also the proof of [CdSS16, Theorem 3.3]). (cid:3) I\{i})t ≺Mt L(Λi). Theorem 3.6. Assume Notation 3.1. In addition, assume that Γi has property (T), for all i ∈ I. For each i ∈ I there is a decomposition Λ = Ψi ⊕ Θi, a scalar ti > 0 and ui ∈ U (M) satisfying (3.4) I\{i})t ¯⊗L(Γi) = A ¯⊗B. By Proposition 3.5, Proof. Fix i ∈ I and write Mt = L(ΓI )t = L(Γ we have A ≺Mt L(Λi) for some non-amenable group Λi 6 Λ with non-amenable CΛ(Λi). By [CKP14, Proposition 2.4], there exist nonzero projections a ∈ A, q ∈ L(Λi), a partial i = L(Ψi) and ui L(Γ ui L(Γi)ti u∗ i = L(Θi). I\{i})t/tiu∗ 10 isometry v ∈ Mt, a subalgebra D ⊆ qL(Λi)q, and a ∗-isomorphism φ : aAa → D such that (3.5) (3.6) Notice that vv∗ ∈ D ′ ∩ qMtq and v∗v ∈ (aAa) ′ ∩ aMt a = a ⊗ B. Hence there is a projection b ∈ B satisfying v∗v = a ⊗ b. Picking u ∈ U (Mt) so that v = u(a ⊗ b) then (3.6) gives D ∨ (D ′ ∩ qL(Λi)q) ⊆ qL(Λi)q has finite index, and φ(x)v = vx ∀ x ∈ aAa. (3.7) Passing to the relative commutants, we obtain vv∗(D ′ ∩ qMtq)vv∗ = u(a ⊗ bBb)u∗. This further implies that there exist s1, s2 > 0 Dvv∗ = vaAav∗ = u(aAa ⊗ b)u∗. (D ′ ∩ qMtq)z = u(a ⊗ bBb)s1 u∗ = L(Γi)s2, (3.8) where z is the central support projection of vv∗ in D ′ ∩ qMq. Now notice D ′ ∩ qMtq ⊇ (qL(Λi)q) ′ ∩ qMtq = (L(Λi) ′ ∩ Mt)q ⊇ L(CΛ(Λi))q, where L(CΛ(Λi)) has no amenable direct summand since CΛ(Λi) is a non-amenable group. Moreover we also have D ′ ∩ qMtq ⊇ D ′ ∩ qL(Λi)q. Thus (L(Λi) ′ ∩ Mt)z and (D ′ ∩ qL(Λi)q)z are commuting subalgebras of (D ′ ∩ qMtq)z where (L(Λi) ′ ∩ Mt)z has no amenable direct summand. Since Γi was assumed to be bi-exact, then using (3.8) and [Oz03, Theorem 1] it follows that (D ′ ∩ qL(Λi)q)z is purely atomic. Thus, cutting by a central projection r ′ ∈ D ′ ∩ q(Λi)q and using (3.5) we may assume that D ⊆ qL(Λi)q is a finite index inclusion of algebras. Processing as in the second part of [CdSS16, Claim 4.4], we may assume that D ⊆ qL(Λi)q is a finite index of II1 factors. Moreover one can check that if one replaces v by the partial isometry of the polar decomposition of r ′v 6= 0 then all relations (3.6),(3.7) and (3.8) are still satisfied. In addition, we can assume without any loss of generality that the support projection satisfies s(EL(Λi)(vv∗) = q. Thus, following the terminology introduced in [CdSS17, Definition 4.1] we actually have that a corner of A is spatially commensurable to a corner of L(Λi), i.e. A ∼=com Mt L(Λi). (3.9) Performing the downward basic construction [Jo81, Lemma 3.1.8], there exists e ∈ P(qL(Λi)q) and a II1 subfactor R ⊆ D ⊆ qL(Λi)q = hD, ei such that [D : R] = [qL(Λi)q : D] and Re = eL(Λi)e. Keeping with the same notation, by relation (3.6) the restriction φ−1 : R → aAa is an injective ∗-homomorphism such that T = φ−1(R) ⊆ aAa is a finite Jones index subfactor and φ−1(y)v∗ = v∗y, for all y ∈ R. (3.10) Let θ ′ : Re → R be the ∗-isomorphism given by θ(xe) = x. Since e has full central support in hD, ei one can see that ev 6= 0. Letting w0 be a partial isometry so that w∗ 0v∗e = v∗e, then Re = eL(Λi)e together with (3.10) imply that θ = φ−1 ◦ θ ′ : eL(Λi)e → aAa is an injective ∗-homomorphism satisfying θ(eL(Λi)e) = T and Notice that w∗ 12] one can further assume that w∗ (3.11) 0w0 ∈ (T ′ ∩ aAa) ¯⊗B and proceeding as in the proof of [OP03, Proposition 0w0 ∈ Z (T ′ ∩ aAa) ¯⊗B. Since [aAa : T] < ∞ then 0y, for all y ∈ eL(Λi)e. 11 θ(y)w∗ 0 = w∗ T ′ ∩ aAa is finite dimensional and so is Z (T ′ ∩ aAa). Thus, replacing the partial isometry w0 by w := w0r0, for some minimal projection r0 ∈ Z (T ′ ∩ aAa) satisfying r0w∗ 0v∗e 6= 0, we see that all relations above still hold including relation (3.11). Moreover, we can assume that w∗w = z1 ⊗ z2, for some nonzero projections z1 ∈ Z (T ′ ∩ aAa) and z2 ∈ B. Using relation (3.11) we get (3.12) Since T ⊆ aAa is finite index inclusion of II1 factors then by the local index formula [Jo81] it follows Tz1 ⊆ z1 Az1 is a finite index inclusion of II1 factors as well. Also, we have w∗L(Λi)w = θ(eL(Λi)e)w∗w = Tz1 ⊗ z2. (w∗ L(Λi)w) ′ ∩ (z1 ⊗ z2)Mt(z1 ⊗ z2) = ((Tz1) ′ ∩ z1 Az1) ¯⊗z2Bz2. (3.13) Altogether, the previous relations imply that Tz1 ¯⊗z2Bz2 ⊆ Tz1 ∨ (Tz ′ 1 ∩ z1 Az1) ¯⊗z2Bz2 = w∗L(Λi)w ∨ w∗(L(Λi) ′ ∩ Mt)w = w∗L(Λi)w ∨ (cid:0)(w∗ L(Λi)w) ′ ∩ (z1 ⊗ z2)Mt(z1 ⊗ z2)(cid:1) ⊆ z1 Az1 ¯⊗z2Bz2. (3.14) Since Tz1 ⊆ z1 Az1 if a finite index inclusion of II1 factors then so is Tz1 ¯⊗z2Bz2 ⊆ z1 Az1 ¯⊗z2Bz2. Let f := ww∗ and notice f = re, for some projection r ∈ L(Λi) ′ ∩ Mt. Letting u ∈ U (Mt) such that w∗ = uww∗ = u f , then relation (3.14) further implies that f (L(Λi) ∨ (L(Λi) ′ ∩ Mt)) f = L(Λi) f ∨ f (L(Λi) ′ ∩ Mt) f ⊆ f Mt f (3.15) is an inclusion of finite index II1 factors. In addition, (3.14) gives that dimC(Z ( f (L(Λi ) ∨ (L(Λi) ′ ∩ Mt)) f )) ≤ [z1 Az1 ¯⊗z2Bz2 : Tz1 ¯⊗z2Bz2] < ∞. Since the central support of e in qL(Λi)q equals q then (3.15) implies that q(L(Λi)qr ∨ r(L(Λi) ′ ∩ Mt))rq = qr(L(Λi) ∨ (L(Λi) ′ ∩ Mt))qr ⊆ qrMtqr, (3.16) is a finite index inclusion of II1 factors. In particular, qL(Λi)qr and r(L(Λi) ′ ∩ Mt)rq are commuting II1 factors. To this end we notice that since 0 6= r0w∗ 0v∗e = w∗v∗e then 0 6= w∗v∗e1/2. Thus 0 6= w∗v∗ew = v∗ew and since v, w are partial isometries we conclude that 0 6= vv∗eww∗. However since ww∗ = r0w0w∗ 0 ≤ s(v∗e) then ww∗ ≤ e. Combining with the above it follows that 0 6= vv∗ww∗ and hence z f = zww∗ 6= 0. Thus further implies that zr 6= 0. (3.17) Next we show the following Claim 3.7. r(L(Λi) ′ ∩ Mt)rq has property (T). Proof of Claim 3.7. Since D ⊆ qL(Λi)q is a finite index inclusion of II1 factors then so is Dr ∨ r(L(Λi) ′ ∩ Mt)rq ⊆ qL(Λi)qr ∨ r(L(Λi) ′ ∩ Mt)rq. Using (3.16) it follows that Dr ∨ r(L(Λi) ′ ∩ Mt)rq ⊆ rqMtrq is finite index as well. Hence Dr ∨ r(L(Λi) ′ ∩ Mt)rq ⊆ Dr ∨ r(D ′ ∩ qMtq)r is also a finite index inclusion. Since D is a factor one can check that EDr∨r(L(Λi) ′∩Mt)rq(x) = E(L(Λi) ′∩Mt)q(x), for all x ∈ r(D ′ ∩ qMtq)r. This combined with the above entail that r(L(Λi) ′ ∩ Mt)rq ⊆ r(D ′ ∩ qMtq)r is finite index. By [Po94, Theorem 1.1.2 (ii)] r(L(Λi) ′ ∩ Mt)rz ⊆ r(D ′ ∩ qMtq)rz is finite index and since property (T) passes 12 to amplifications and finite index subalgebras, then (3.8) implies that r(L(Λi) ′ ∩ Mt)rz has property (T). As r(L(Λi) ′ ∩ Mt)rq is a factor we conclude r(L(Λi) ′ ∩ Mt)rq has property (T). (cid:4) Now consider Ω := vCΛ(Λi) = {λ ∈ Λ λΛi < ∞}, the virtual centralizer of Λi in Λ. Using [CdSS16, Claim 4.7] we have [Λ : ΛiΩ] < ∞ and hence ΛiΩ 6 Λ is an icc subgroup; in particular, vZ(Λi Ω) = 1. Consider vZ(Ω) = {ω ∈ Ω ωΩ < ∞}, the virtual center of Ω. Since Λi normalizes Ω one can check that vZ(Ω) 6 vZ(Λi Ω). Since the latter is trivial we get vZ(Ω) = 1 and hence Ω is icc. Let (On)n∈N be a countable enumeration of all the orbits under conjugation by Λi. Denote by Ωk = hO1, ..., Oki 6 Ω, the subgroup generated by On, n = 1, k. Ω k's form an ascending sequence of subgroups Ωk. Thus ΛiΩk is an ascending sequence satisfying normalized by Λi such that Ω = ∪∞k=1 ΛiΩk. Since r(L(Λi) ′ ∩ Mt)rq ⊂ L(ΛiΩ) has property (T) there is k0 ∈ N ΛiΩ = ∪∞k=1 such that r(L(Λi) ′ ∩ Mt)rq ≺L(ΛiΩ) L(Λi Ωk0). (3.18) Next we show the following Claim 3.8. There exists k ≥ k0 such that qL(Λi)qr ∨ r(L(Λi) ′ ∩ Mt)rq ≺L(ΛiΩ) L(ΛiΩk). Proof of the Claim 3.8. Using Popa's intertwining techniques, (3.18) implies the existence of xℓ, yℓ ∈ L(ΛiΩ), ℓ = 1, j and c > 0 satisfying jX ℓ=1 kEL(ΛiΩ k0 )(xℓuyℓ)k2 ≥ c, (3.19) for all u ∈ U (r(L(Λi ) ′ ∩ Mt)rq). Since ΛiΩk ր ΛiΩ for every ε > 0 there is k ≥ k0 so that (3.20) k)(xℓ) − xℓk2 < ε, kEL(ΛiΩ jX jX kEL(ΛiΩ k)(yℓ) − yℓk < ε. ℓ=1 ℓ=1 Using (3.19) together with inequalities kmznk2 ≤ kmk then for all u ∈ U (r(L(Λi ) ′ ∩ Mt)rq) we have ∞ kzk2knk ∞ for all m, n, z ∈ Mt c ≤ ≤ jX ℓ=1 jX ℓ=1 kEL(ΛiΩ k)(xℓuyl)k2 kEL(ΛiΩ k)((xℓ − EL(ΛiΩ k)(xℓ))uyℓ)k2 + jX ℓ=1 kEL(ΛiΩ k)(EL(ΛiΩ k)(xℓ)u(yℓ − EL(ΛiΩ k)(yℓ))k2 + jX ℓ=1 kEL(Λi Ωk)(EL(Λi Ωk)(xℓ))uEL(Λi Ωk)(yℓ)k2 ≤ε max 1≤ℓ≤j (kyℓk ∞ + kxℓk ) + ∞ ≤2dε + jd2kEL(ΛiΩ k)(u)k2, jX ℓ=1 kxℓk kyℓk ∞ ∞ kEL(ΛiΩk)(u)k2 13 for all u ∈ U (r(L(Λi) ′ ∩ Mt)rq). Letting ε = c ∞ }. This shows there is k ≥ k0 so that kEL(ΛiΩ k0 )(u)k2 ≥ 4d then for all u ∈ U (r(L(Λi ) ′ ∩ Mt)rq) where d := max1≤ℓ≤j{kxℓk c−2εd ∞ , kyℓk jd2 we have kEL(Λi Ω k)(u)k2 ≥ c 2jd2 > 0. This implies for all a ∈ U (qL(Λi)qr) and u ∈ U (r(L(Λi) ′ ∩ Mt)rq) we have c 2jd2 . k)(au)k2 = kaEL(Λi Ω k)(u)k2 = kEL(ΛiΩ k)(u)k2 ≥ kEL(ΛiΩ (3.21) As U (qL(Λi)qr)U (r(L(Λi ) ′ ∩ Mt)rq) generates qL(Λi)qr ∨ r(L(Λi) ′ ∩ M)rq, (3.21) gives the claim. (cid:4) Now, since q(L(Λi)qr ∨ r(L(Λi) ′ ∩ M)rq ⊆ rqL(Λi Ω)rq is a finite index inclusion, then rqL(Λi Ω)rq ≺L(ΛiΩ) (L(Λi Ωk) and hence L(Λi Ω) ≺L(ΛiΩ) L(ΛiΩk). By [CI17, Lemma 2.2] it follows that ΛiΩk 6 ΛiΩ has finite index and by increasing k we can assume that ΛiΩk = ΛiΩ. Let Λ ′ := CΛi(Ωk) 6 Λi and notice [Λi : Λ ′] < ∞. Thus [ΛiΩ : Λ ′Ωk] < ∞ and since ΛiΩ is icc then Λ ′Ωk is also icc. In particular, we also have Λ ′ ∩ Ωk = 1. As Λ ′Ωk 6 ΛiΩ is finite index the Λ ′Ωk ∩ Ω 6 Ω is also finite index. In particular since Ω is icc it follows that Λ ′Ωk ∩ Ω is also icc. Letting Λ ′′ := Λ ′ ∩ Ω the above considerations imply that Λ ′′Ω k ∩ Ω. This forces Λ ′′ to be either trivial or icc. However, since by construction Λ ′′ = vZ(Λ ′′) then Λ ′′ = 1. Since Λ ′ 6 Λi finite index it follows that Λi is icc-by-finite and hence finite-by-icc. k = Λ ′Ω This together with (3.9) and [CdSS17, Theorem 4.6] show there exists Σ 6 CΛ(Λi) such that [Λ : ΛiΣ] < ∞ and B ∼=com M L(Σ). Also since Λ is icc then so are Λi and Σ. Finally, using [CdSS17, Theorem 4.7] there is a decomposition Λ = Ψi ⊕ Θi, ui ∈ U (Mt) and t > 0 I\{i})tu∗ such that ui Atu∗ (cid:3) i = L(Ψi) and ui B1/tu∗ i = ui L(Γi)tu∗ i = L(Θi). i = ui L(Γ Theorem 3.9. Let (Γn)n∈N a countable infinite collection of property (T), biexact, weakly amenable, icc groups. Assume that Λ is an arbitrary group satisfying L(⊕n Γn) = L(Λ). Then there exists an infinite direct sum decomposition Λ = (⊕nΛn) ⊕ A where A is either trivial or icc amenable group. Moreover, for each k ∈ N there exist scalars t1, ..., tk+1 > 0 satisfying t1t2 · · · tk+1 = 1 and a unitary u ∈ L(Λ) so that uL(Γn)tn u∗ = L(Λn) for all n = 1, k; and uL(⊕n≥k+1 Γn)tk+1u∗ = L(⊕n≥k+1Λn ⊕ A). (3.22) 1 = L(Λ1) and v1L(ΓN\{1})t/t1v∗ Proof. Using Theorem 3.6 there exist a product decomposition Λ = Λ1 ⊕ Θ1, v1 ∈ U (M), and t1 > 0 such that v1L(Γ1)t1v∗ 1 = L(Θ1). Applying The- orem 3.6 again in the last relation for the group ΓN\{1} there exist a product decomposition Θ1 = Λ2 ⊕ Θ2, v2 ∈ U (v1 L(ΓN\{1})t/t1v∗ 2 = L(Λ2) and v2L(ΓN\{1,2})t/(t1t2)v∗ 2 = L(Θ2). Proceeding inductively one has Θn−1 = Λn ⊕ Θn, a uni- N\1,n−1)t/(t1t2···tn−1)v∗ tary vn ∈ U (vn−1L(Γ n = L(Λn) n = L(Θn). Altogether these relations show that Θn > Θn+1 and vnL(Γ for all n and also Λ = ⊕N Λn ⊕ Σ, where A = ∩nΘn. In addition, for every k ∈ N letting n−1) and tn > 0 such that vnL(Γn)tn v∗ 1) , and t2 > 0 such that v2L(Γ2)t2v∗ N\1,n)t/(t1t2···tn)v∗ 14 uk := v1v2 · · · vk we see that k = L(Λi) for all i = 1, k and uk L(Γi)tiu∗ uk L(Γ N\1,k)t/(t1t2···tk)u∗ k = L(⊕i≥k+1Λi ⊕ A). (3.23) k L(A)uk ⊆ L(Γ Since L(Γk) is a II1 factor the second relation in (3.23) show that for each k ∈ N one can N\1,k)t/(t1t2···tk). Using Proposition 2.7 and the find uk ∈ U (M) such that u∗ same argument as in the proof of Claim 3.4 if follows that A is icc amenable as desired. (cid:3) Remarks 3.10. We conjecture that Theorem 3.9 still holds true without the property (T) assumption on the Γi's. We point out that property (T) was used in the proof of Theorem 3.6 only to derive relation (3.18); in other words the (increasing) sequence of subgroups Ωk becomes stationary. We believe this conclusion can still be achieved without the property (T) assumption. However at this time we are unable to prove this. Proof of Corollary B. First we argue that the group Γ = ⊕nΓn has trivial amenable radi- cal. So let B ⊳ Γ be a normal amenable subgroup. Thus the von Neumann subalgebra L(B) ⊆ L(Γ) = L(⊕n6=kΓk) ¯⊗L(Γk) is regular and amenable. Applying [PV12, Theo- rem 1.4] it follows that L(B) ≺ L(⊕n6=kΓn). Since B is normal Γ we further deduce from [CI17, Lemma 2.2] that [B : Bk] < ∞ where Bk := B ∩ (⊕n6=kΓn) < ⊕n6=kΓn. Since Bk ⊳ Γ is normal it follows that B/Bk ⊳ Γ/Bk is a finite normal subgroup. As Γ/Bk = (⊕n6=kΓn/Bk) ⊕ Γk, if πk : Γ/Bk → Γk is the canonical projection map it fol- lows that πk(B/Bk) ⊳ Γk is a finite normal subgroup. As Γk is icc we have πk(B/Bk) = 1 and hence B/Bk < ⊕n6=kΓn/Bk; in particular, B = Bk < ⊕n6=kΓn. Since this holds for every positive integer k then B < ∩k(⊕n6=kΓn) = 1, thus giving the desired claim. r (Γ) → C∗ [BKKO14, Theorem 1.3] implies that the reduced C∗-algebra C∗ r (Γ) has the unique trace property. Letting φ : C∗ r (Λ) be a ∗-isomorphism of C∗-algebras it follows that φ lifts to a ∗-isomorphism φ : L(Γ) → L(Λ) of von Neumann algebras. By Theorem 3.9 we have that Λ = (⊕nΛn) ⊕ A with A icc amenable; moreover, the corresponding relations (3.22) also hold. Since C∗ r (Λ) has the unique trace property then [BKKO14, Theorem 1.3] implies that A = 1 and the first part of the conclusion is proved. The remaining part of the conclusion follows directly from relations 3.22. (cid:3) To ease our exposition we first introduce the following notation: 4. PROOF OF THEOREM E Notation 4.1. Let H0, Γ be icc groups such that H0 has property (T) and Γ admits an infi- nite, almost normal subgroup Γ0 6 Γ with relative property (T). Let Γ y I be an action on a countable set I satisfying the following conditions: a) For each i ∈ I we have [Γ : StabΓ(i)] < ∞; b) There is k ∈ N such that for each J ⊆ I with J ≥ k we have StabΓ(J) < ∞. Denote by G = H0 ≀I Γ the corresponding generalized wreath product. Denote by M = L(G) and assume that M = L(Λ) for an arbitrary group Λ. Let ∆ : M → M ¯⊗M be the commultiplication along Λ, i.e. ∆(vλ) = vλ ⊗ vλ, where {vλ}λ∈Λ are the canonical unitaries generating L(Λ). Proposition 4.2. Assume Notation 4.1. Then the following hold: 15 c) ∆(L(H d) There exists u ∈ U (M ¯⊗M) such that u∆(L(Γ))u∗ ⊆ L(Γ × Γ). M⊗M L(H (I) 0 )) ≺s (I) 0 ) ¯⊗L(H (I) 0 ); Proof. We denote A0 = L(H0) and A = A(I) 0 . Note that M = A ⋊ Γ, the action being given by generalized Bernoulli shifts. Write M = L(Λ) and denote by ∆ : M → M ¯⊗M the associated co-multiplication. Note that M ¯⊗M = (A ¯⊗A) ⋊ (Γ × Γ). The inclusion ∆(A0) ⊂ M ¯⊗M = M ¯⊗(A ⋊ Γ) is rigid. Denote by P ⊂ M ¯⊗M the quasi- normalizer of ∆(A0). Note that ∆(A) ⊂ P. By applying [IPV10, Theorem 4.2], we see that one of the following has to hold: (1) ∆(A0) ≺M ¯⊗M M ⊗ 1; (2) P ≺M ¯⊗M M ¯⊗(A ⋊ StabΓ(i)), for some i ∈ I; (3) v∗Pv ⊂ M ¯⊗L(Γ) for some partial isometry 0 6= v ∈ M ¯⊗M. (1) is impossible since A0 is diffuse. Suppose (3) holds. Then by the remark above we have that v∗∆(A)v ⊂ M ¯⊗L(Γ). There are two possibilities: either ∆(A) ≺M ¯⊗M M ¯⊗L(StabΓ(i)) for some i, or ∆(A) ⊀M ¯⊗M M ¯⊗L(StabΓ), for all i ∈ I. In the first case we again have two possibilities: either there exists a maximal finite subset G ∋ i such that ∆(A) ≺M ¯⊗M M ¯⊗L(StabΓ(G)), or there is no such subset. If the first sub-case holds then [IPV10, Lemma 4.1.3] gives that QNM ¯⊗M(∆(A)) ′′ ≺M ¯⊗M M ¯⊗L(Norm(G)). Since the quasi-normalizer of ∆(A) contains ∆(M), this implies ∆(M) ≺M ¯⊗M M ¯⊗L(Norm(G)). As StabΓ(G) is a finite index subgroup of Norm(G), it follows that ∆(M) ≺M ¯⊗M M ¯⊗L(StabΓ(G)), and hence ∆(M) ≺M ¯⊗M M ¯⊗L(StabΓ(i)) for some i, which implies by [IPV10, Lemma 7.2.2] that L(StabΓ(i)) ⊂ M has finite index, which is a contradiction. If the second sub-case holds, by taking G with G ≥ κ we get that ∆(A) ≺M ¯⊗M M ⊗ 1, a contradiction. It follows that (2) must hold, hence P ≺M ¯⊗M M ¯⊗(A ⋊ StabΓ(i)), which further implies ∆(A) ≺M ¯⊗M M ¯⊗(A ⋊ StabΓ(i)), for some i ∈ I. Again we have two possibilities: either there exists a finite maximal subset G ⊂ I such that ∆(A) ≺M ¯⊗M M ¯⊗(A ⋊ StabΓ(G)), or it doesn't. In the first sub-case we get, by [IPV10, Lemma 4.1.3], that QNM ¯⊗M(∆(A)) ′′ ≺M ¯⊗M M ¯⊗(A ⋊ Norm(G)) and again as above, that ∆(M) ≺M ¯⊗M M ¯⊗(A ⋊ StabΓ(i)), for some i ∈ G, which by [IPV10, Lemma 7.2.2] implies that [M : A ⋊ StabΓ(i)] is finite, a contradiction. In the second sub-case, by taking a G with G ≥ κ, we obtain that ∆(A) ≺M ¯⊗M M ¯⊗A, which is what we wanted. The maximal projection q ∈ ∆(A) ′ ∩ M ¯⊗M such that ∆(A)q ≺s M ¯⊗M M ¯⊗A is non-zero and belongs to the center of the normalizer of ∆(A) in M ¯⊗M. This center is contained in ∆(M) ′ ∩ M ¯⊗M = C1. It follows that q = 1, hence ∆(A) ≺s M ¯⊗M M ¯⊗A. By symmetry we obtain that also ∆(A) ≺s M ¯⊗M A ¯⊗M and finally that ∆(A) ≺s M ¯⊗M A ¯⊗A, showing part c). Next we prove part d). First notice from the assumptions that the inclusion ∆(L(Γ0 )) ⊂ M ¯⊗(A ⋊ Γ) is rigid. Denote by P the quasi-normalizer of ∆(L(Γ0)) inside M ¯⊗M. Note that P contains ∆(L(Γ)). We apply again [IPV10, Theorem 4.2] and we see that one of the following has to hold: (1) ∆(L(Γ0 )) ≺M ¯⊗M M ¯⊗1; (2) P ≺M ¯⊗M M ¯⊗(A ⋊ StabΓ(i)), for some i ∈ I; (3) vPv∗ ⊂ M ¯⊗L(Γ), for some v ∈ U (M ¯⊗M). Note (1) cannot be true because ∆(L(Γ0 )) is diffuse. Suppose (2) is true. This implies in particular that ∆(L(Γ)) ≺M ¯⊗M M ¯⊗(A ⋊ StabΓ(i)). But since ∆(A) ≺s M ¯⊗M A ¯⊗A, by 16 the same argument as in the beginning of the proof of [Io10, Theorem 8.2], we would get ∆(A ⋊ Γ) = ∆(M) ≺M ¯⊗M M ¯⊗(A ⋊ StabΓ(i)), which by [IPV10, Lemma 7.2.2] implies that A ⋊ StabΓ(i) ⊂ M has finite index, a contradiction. So (3) must be true, hence a fortiori v∆(L(Γ))v∗ ⊂ M ¯⊗L(Γ). Repeating the argument for the inclusion v∆(L(Γ))v∗ ⊂ M ¯⊗L(Γ) = (A ⋊ Γ) ¯⊗L(Γ), we obtain an unitary u ∈ M ¯⊗M such that u∆(L(Γ))u∗ ⊂ L(Γ) ¯⊗L(Γ), as desired. (cid:3) Proposition 4.3. Assume Notation 4.1. In addition assume that H0, Γ and Λ are torsion free groups. Then the following hold: e) ∆(L(H f) There exists w ∈ U (M ¯⊗M) such that w∆(Γ)w∗ ⊆ T(Γ × Γ). M⊗M L(H (I) 0 )) ≺s (I) 0 ) ¯⊗L(H (I) 0 ); Proof. Since e) follows directly from Proposition 4.2 we will only argue for f). From Propo- sition 4.2 there exists u ∈ U (M ¯⊗M) such that u∆(L(Γ))u∗ ⊆ L(Γ × Γ). Denote by G = {u∆(uγ)u∗ γ ∈ Γ} ⊂ U (L(Γ × Γ)). Since G normalizes u∆(A)u∗, which satisfies u∆(A)u∗ ≺s A ¯⊗A, the argument in Step 5 of the proof of [IPV10, Theorem 5.1] implies that hΓ×Γ(G) > 0. We have that G ′′ = u∆(L(Γ))u∗ ⊀M ¯⊗M L(CΓ×Γ(γ1, γ2)), Indeed, suppose this is not true, and u∆(L(Γ))u∗ ≺ for any (γ1, γ2) ∈ Γ × Γ − {e}. L(CΓ×Γ((γ1, γ2))) = L(CΓ(γ1)) ¯⊗L(CΓ(γ2)), with γ2 6= e. Again using the fact that ∆(A) ≺M ¯⊗M A ¯⊗A, we infer that ∆(M) = ∆(A ⋊ Γ) ≺ M ¯⊗(A ⋊ CΓ(γ2)). [IPV10, Lemma 7.2.2] then implies that A ⋊ CΓ(γ2) has finite index in M, which is a contradiction, since Γ is icc. Also, the representation {Ad(v)}v∈G on L2(L(Γ × Γ)) ⊖ C1 is weakly mixing, be- cause it is in fact weakly mixing on L2(M ¯⊗M) ⊖ C1. Indeed, let H ⊂ L2(M ¯⊗M) be a finite dimensional {Ad(v)}v∈G invariant subspace. Then H0 = uHu∗ is a finite dimen- sional {Ad∆(uγ)}γ∈Γ-invariant subspace of L2(M ¯⊗M). Denote by K the closed linear span of H0∆(M). Then K is a ∆(LΓ) − ∆(M) bi-module, which is finitely generated as a right module. Since L(Γ) ⊀M L(CΛ(s)), for any s ∈ Λ − {e}, [IPV10, Proposition 7.2.3] implies that K ⊂ ∆(L2 M), so in particular H0 ⊂ ∆(L2 M). Hence ∆−1(H0) ⊂ L2 M is a (I) finite dimensional {Ad(uγ)}γ∈Γ-invariant subspace. As the inclusion Γ 6 H 0 ⋊ Γ is icc, the representation {Ad(uγ)}γ∈Γ on L2(M) ⊖ C1 is weakly mixing, which further implies that H = C1, as claimed. Now we apply [KV15, Theorem 4.1] to deduce that there exists a unitary w ∈ L(Γ × Γ) such that wGw∗ ⊂ T(Γ × Γ). By replacing w with wu, we may assume that w∆(uγ)w∗ ∈ T(Γ × Γ) for all γ ∈ Γ. (cid:3) Theorem 4.4. Let H0, Γ be icc torsion free groups such that H0 has property (T) and Γ admits an infinite, almost normal subgroup Γ0 6 Γ with relative property (T). Let Γ y I be a transitive action on a countable set I satisfying the following conditions: a) For each i ∈ I we have [Γ : StabΓ(i)] < ∞; b) There is k ∈ N such that for each J ⊆ I satisfying J ≥ k we have StabΓ(J) < ∞; c) For every i 6= j we have that StabΓ(i) · j = ∞. Denote by G = H0 ≀I Γ the corresponding generalized wreath product. Let Λ be any torsion free group and let θ : L(G) → L(Λ) be a ∗-isomorphism. Then Λ admits a wreath product decomposition Λ = Σ0 ≀I Ψ satisfying the following properties: there exist a group isomorphism ρ : Γ → Ψ, a character η : Γ → T, a ∗-isomorphism θ0 : L(H0) → L(Σ0) and a unitary 17 v ∈ L(Λ) such that for every x ∈ L(H (I) 0 ) and γ ∈ Γ we have θ(xuγ) = η(γ)v∗θ ¯⊗I 0 (x)vδ(γ)v. Here {uγ γ ∈ Γ} and {vλ λ ∈ Ψ} are the canonical unitaries of L(Γ) and L(Ψ), respectively. (I) 0 ), and notice from assumptions we have that θ(L(G)) = L(Λ) = M. Proof. Let A = L(H Using Proposition 4.3 one can find w ∈ U (M ¯⊗M), group homomorphisms δi : Γ → Γ, and a character ω : Γ → T such that w∆(θ(uγ))w∗ = ω(γ)θ(uδ1 (γ)) ⊗ θ(uδ2 (γ)) for all γ ∈ Γ. Then aplying verbatim Steps 4 and 5 in the proof of [IPV10, Theorem 8.2] one can find an injective group homomorphism ρ : Γ → Λ and a character η : Γ → T satisfying (4.1) θ(uγ) = η(γ)vρ(γ), for all γ ∈ Γ. Denote by Ψ = ρ(Γ). In addition, these proofs also show there is v ∈ U (M) such that w = (v∗ ⊗ v∗)∆(v). Henceforth the canonical unitaries θ(uγ), γ ∈ Γ will be replaced by vθ(uγ)v∗ and A will be replaced by vAv∗. Under these conventions we prove that Claim 4.5. ∆(θ(A)) ⊂ θ(A) ¯⊗θ(A). Proof of Claim 4.5. By Proposition 4.2, ∆(θ(A)) ≺s θ(A) ¯⊗θ(A). This means that for ev- ery ǫ > 0, there exists a finite subset e ∈ S ⊂ θ(Γ) such that kd − PS×S(d)k2 ≤ ǫ, for all d ∈ U (∆(θ(A))). But since, according to Proposition 4.3, ∆(θ(A)) is invariant to Ad(∆(θ(uγ ))) = Ad(θ(uγ) ⊗ θ(uγ)) for all γ ∈ Γ, we see that kd − PµSµ−1×µSµ−1(d)k2 ≤ ǫ, for all d ∈ U (∆(θ(A))) and µ ∈ θ(Γ). As Γ is icc we can find µ ∈ θ(Γ) such that µSµ−1 ∩ S = {e} (see for instance [CSU13, Proposition 3.4]). By the triangle inequality this further implies that kd − Eθ(A) ¯⊗θ(A)(d)k2 = kd − P(µSµ−1∩S)×(µSµ−1∩S)(d)k2 ≤ 2ǫ, for all d ∈ U (∆(θ(A))). As ǫ is arbitrary, this implies ∆(θ(A)) ⊂ θ(A) ¯⊗θ(A). (cid:4) From Claim 4.5 and [IPV10, Lemma 7.1.2] it follows that θ(A) = L(Σ), for some Σ < Λ. Since the uγ's normalize A, it follows that Ψ ∋ ρ(γ) normalizes Σ, for all γ. Consider the action of Ψ → Aut(Σ) given by Ψ ∋ λ → Ad(λ) ∈ Aut(Σ) and observe Λ splits as a semidirect product Λ = Σ ⋊ Ψ, because L(Λ) = θ(A) ⋊ θ(Γ). For the remaining part consider A0 = L(H0) and denote by Ai 0 the copy of A0 in posi- tion i ∈ I. Next we show that 0)) ⊂ θ(Ai Claim 4.6. ∆(θ(Ai 0) ¯⊗θ(Ai 0), for all i ∈ I. 0) ¯⊗θ(Ai Proof of Claim 4.6. Using (4.1) we note that ∆(θ(Ai 0)) is fixed by Ad(∆(θ(uγ ))) = Ad(θ(uγ) ⊗ θ(uγ)), for all γ ∈ StabΓ(i). Due to the assumption that StabΓ(i) · j is infinite for all i 6= j, the representation Ad{θ(uγ) ⊗ θ(uγ)}γ∈StabΓ(i) is weakly mixing on L2(M ¯⊗M) ⊖ L2(θ(Ai (cid:4) Hence from Claim 4.6 and [IPV10, Lemma 7.1.2] for every i ∈ I there exists a subgroup 0) = L(Σi). Since the action Γ y I is transitive, it follows that Σ0. Moreover, this entails that the action Ψ → Σi) is induced by the generalized Bernoulli action of Ψ y I and hence (cid:3) Σi < Λ such that θ(Ai Σi Aut(Σ) = Aut(L I Λ = Σ0 ≀I Γ. The rest of the statement follows from the previous observations. ∼= Σ0 for all i, and then that Σ = LI 0)), so it follows that ∆(θ(Ai 0)) ⊂ θ(Ai 0) ¯⊗θ(Ai 0). 18 5. PROOF OF THEOREM C Theorem 5.1. Let H0, Γ be icc, property (T) group. Also assume that Γ = Γ1 × Γ2, where Γi are nonamenable biexact groups for all i = 1, 2. Let Γ y I be an action on a countable set I satisfying the following conditions: a) The stabilizer StabΓ(i) is amenable for each i ∈ I; b) There is k ∈ N such that for each J ⊆ I satisfying J ≥ k we have StabΓ(J) < ∞. decomposition Λ = Σ ⋊ Φ satisfying the following properties: there exist a group isomorphism Denote by G = H0 ≀I Γ the corresponding generalized wreath product. Let Λ be an arbitrary group and let θ : L(G) → L(Λ) be a ∗-isomorphism. Then Λ admits a semidirect product δ : Γ → Φ, a character ζ : Γ → T, a ∗-isomorphism θ0 : L(H 0 ) → L(Σ) and a unitary (I) t ∈ L(Λ) such that for every x ∈ L(H (I) 0 ) and γ ∈ Γ we have θ(xuγ) = ζ(γ)tθ0 (x)vδ(γ)t∗. Here {ug γ ∈ Γ} and {vλ λ ∈ Φ} are the canonical unitaries of L(Γ) and L(Φ), respectively. Proof. From assumptions we have that θ(L(G)) = L(Λ) = M. Denote by A0 = θ(L(H0)) and A = θ(L(H I 0). Also to simplify the writing, throughout the proof we will identify Γ with θ(Γ), etc. Thus note that M = A ⋊ Γ, the action being given by generalized Bernoulli shifts. Consider ∆ : M → M ¯⊗M the comultiplication along Λ. Note that M ¯⊗M = (A ¯⊗A) ⋊ (Γ × Γ). Theorem 4.2 implies that (1) ∆(A) ≺s (2) there is u ∈ U (M ¯⊗M) such that u∆(L(Γ))u∗ ⊆ L(Γ × Γ). M ¯⊗M A ¯⊗A, and Next we show the following (1) Λ (Φ) = Φ, d ∈ P(L(Φ)) and µ ∈ Claim 5.2. There exist a subgroup Φ < Λ with QN U (M) satisfying h = µdµ∗ ∈ L(Γ) and µdL(Φ)dµ∗ = hL(Γ)h. Proof of Claim 5.3. Let K := {Γ × Γ1, Γ × Γ2, Γ1 × Γ, Γ2 × Γ}. Since by [BO08, Lemma 15.3.3] Γ × Γ is biexact relatively to K and ∆(L(Γ1 )) and ∆(L(Γ2)) are commuting non-amenable factors then [BO08, Theorem 15.1.5] implies that there are Ψ ∈ K and i = 1, 2 so that u∆(L(Γi ))u∗ ≺L(Γ×Γ) L(Ψ). Since the flip automorphism of M ¯⊗M acts identically on ∆(L(Γi )) we can assume without any loss of generality Ψ = Γ × Γ1 and i = 1. Hence u∆(L(Γ1 ))u∗ ≺L(Γ×Γ) L(Γ × Γ1). The using [DHI16, Theorem 4.1] (see also [CdSS16, Theorem 3.3])) this further implies there exists a subgroup Σ < Λ with non-amenable centralizer Υ := CΛ(Σ) and L(Γ1) ≺M L(Σ). Passing to the intertwining of the relative commutants we have that L(Υ) ⊆ L(Σ) ′ ∩ M ≺M L(Γ1) ′ ∩ M = L(Γ2). Thus there are projections e ∈ L(Υ), f ∈ L(Γ2), a partial isometry v ∈ M, and an injective unital ∗-homomorphism φ : eL(Υ)e → f L(Γ2) f such that (5.1) Denote by T := φ(eL(Υ)e) and notice that q ′ := vv∗ ∈ T ′ ∩ f M f and p := v∗v ∈ eL(Υ)e ′ ∩ eMe = (L(Υ) ′ ∩ M)e. Since T is a non-amenable factor then a) implies that T ⊀ L(StabΓ(i)) for all i and using [Po03, Theorem 3.1] we have QNf M f (T) ′′ ⊆ L(Γ). In particular, q ′ ∈ L(Γ) and by (5.1) there is u ∈ U (M ¯⊗M) such that ueL(Υ)epu∗ ⊆ L(Γ). φ(x)v = vx, for all x ∈ eL(Υ)e. 19 Since L(Γ) is a factor, the same argument from [IPP05, Theorem 5.1, page 26] shows that one can perturb u to a new unitary such that we further have uL(Υ)pu∗ ⊆ L(Γ). (5.2) Since Υ is non-amenable then uL(Υ)pu∗ ⊀ L(StabΓ(i)) for all i and (5.2) combined with [Po03, Theorem 3.1] and the quasinormalizer formula show that upL(QNΛ(Υ)) ′′ pu∗ ⊆ QNupMpu∗(uL(Υ)pu∗ ) ′′ ⊆ L(Γ). Since L(Γ) is a factor, the same argument as before further implies that uL(QNΛ(Υ))z ′ u∗ ⊆ L(Γ), where z ′ is the central support of p in L(QNΛ(Υ)). Notice Σ 6 vCΛ(Υ) < QNΛ(Υ) and hence uL(Σ)z ′ u∗ ⊆ L(Γ). Letting (1) Ω := vCΛ(Σ), Θ := QN Λ (ΣΩ) and using the same arguments as before, we can further find η ∈ U (M ¯⊗M) and a projection z ∈ Z(L(Θ)) such that ηL(Θ)zη∗ ⊆ L(Γ). (5.3) Since Υ, Σ < Θ are commuting non-amenable groups and Γ is biexact relatively to {Γ1, Γ2}, [BO08, Theorem 15.1.5] implies that ηL(Σ)zη∗ ≺L(Γ) L(Γk), for some k = 1, 2. Again, wlog we can assume k = 1. Passing to the relative commutants intertwining we get L(Γ2) = L(Γ1) ′ ∩ L(Γ) ≺L(Γ) (ηL(Σ)zη∗ ) ′ ∩ ηzη∗ L(Γ)ηzη∗ ⊆ ηL(Ω)zη∗. (5.4) Now let {Ok}k be a countable enumeration of all the finite orbits under conjugation by Σ and notice that ∪kOk = Ω. Consider Ωk := hO1, ..., Oki 6 Λ and note that Ωk ր Ω. Since L(Γ2) has property (T) then (5.4) implies that (5.5) By [CKP14, Proposition 2.4], there exist nonzero projections a ∈ L(Γ2), q ∈ L(Ωk), a partial isometry w ∈ L(Γ), a subalgebra D ⊆ ηqL(Ωk)qzη∗, and a ∗-isomorphism ψ : aL(Γ2)a → D such that D ∨ (D ′ ∩ ηqL(Ωk)qzη∗) ⊆ ηqL(Ωk)qzη∗ has finite index, and ψ(x)w = wx ∀ x ∈ aL(Γ2)a. (5.6) (5.7) Let r = ηqzη∗ and notice that ww∗ ∈ D ′ ∩ rL(Γ)r and w∗w ∈ (aL(Γ2 )a) ′ ∩ aL(Γ)a = L(Γ1) ¯⊗Ca. Hence there is a projection b ∈ L(Γ1) satisfying w∗w = b ⊗ a. Picking c ∈ U (L(Γ)) so that w = c(b ⊗ a) then (5.7) gives L(Γ2) ≺L(Γ) ηL(Ωk)zη∗ for some k. (5.8) Passing to the relative commutants, we obtain ww∗(D ′ ∩ rL(Γ)r)ww∗ = c(bL(Γ1)b ⊗ Ca)c∗. Hence there exist s1, s2 > 0 satisfying Dww∗ = wL(Γ2)w∗ = c(Cb ⊗ aL(Γ2)a)c∗. (D ′ ∩ rL(Γ)r)y = c(bL(Γ1 )b ⊗ Ca)s2 c∗ ∼= L(Γ1)s1, (5.9) where y is the central support projection of ww∗ in D ′ ∩ rL(Γ)r. Notice D ′ ∩ rL(Γ)r ⊇ (ηqL(Ωk)qzη∗) ′ ∩ rL(Γ)r = η(L(Ωk) ′ ∩ L(Θ))qzη∗ ⊇ ηL(CΣ(Ωk))qzη∗. From the definition of Ω k)] < ∞. Since Σ is non-amenable it follows that CΣ(Ωk) is also non-amenable and hence ηL(CΣ(Ωk))qzη∗ has no amenable direct summand. Moreover we also have D ′ ∩ rL(Γ)r ⊇ D ′ ∩ ηqL(Ωk)qzη∗. In conclu- sion (ηL(CΣ(Ωk))qzη∗)y and (D ′ ∩ ηqL(Ωk)qzη∗)y are commuting subalgebras of (D ′ ∩ rL(Γ)r)y where (ηL(CΣ(Ωk))qzη∗)y has no amenable direct summand. Since Γi was k it follows that [Σ : CΣ(Ω 20 assumed to be bi-exact, then using (5.9) and [Oz03, Theorem 1] it follows that (D ′ ∩ ηqL(Ωk)qzη∗)y is purely atomic. Thus, cutting by a central projection r ′ ∈ D ′ ∩ ηqL(Ωk)qzη∗ k)qzη∗ is a finite index inclusion of al- and using (5.6) we may assume that D ⊆ ηqL(Ω gebras. Proceeding as in the second part of [CdSS16, Claim 4.4], we may assume that D ⊆ ηqL(Ωk)qzη∗ is a finite index inclusion of II1 factors. Moreover one can check that if one replaces w by the partial isometry of the polar decomposition of r ′w 6= 0 then all relations (5.7),(5.8) and (5.9) are still satisfied. Using relation (5.8), the quasinormalizer compresion formula, and the fact that D ⊆ ηqL(Ωk)qzη∗ is a finite index inclusion of II1 factors we can see that c(b ⊗ a)L(Γ)(b ⊗ a)c∗ = QNc(b⊗a)M(b⊗a)c∗(c(Cb ⊗ (aL(Γ2 )a))c∗) ′′ = QNww∗ Mww∗(Dww∗) ′′ = ww∗QNrMr(D) ′′ww∗ = ww∗QNηqzMqzη∗(ηqL(Ωk )qzη∗) ′′ww∗ = ww∗ηqzQNL(Λ)(L(Ωk)) ′′qzηww∗. Letting Ξ = QNΛ(Ωk), then the previous relation and formula [JGS10, Corollary 5.2] (1) imply that c(b ⊗ a)L(Γ)(b ⊗ a)c∗ = ww∗ηL(Ξ)η∗ww∗. Since QN G (Γ) = Γ, this formula and [JGS10, Corollary 5.2] further imply that c(b ⊗ a)L(Γ)(b ⊗ a)c∗ = QN (1) c(b⊗a)M(b⊗a)c∗(c(b ⊗ a)L(Γ)(b ⊗ a)c∗) ′′ (1) ww∗ηMη∗ww∗(ww∗ηL(Ξ)η∗ww∗) ′′ = QN = ww∗ηL(Φ)η∗ww∗, (5.10) (1) Λ (Ξ) = QN where Φ = hQN(1) Λ (Ξ)i. Hence in particular we have ww∗ηL(Ξ)η∗ww∗ = ww∗ηL(Φ)η∗ ww∗ and by [CdSS16, Proposition 2.6] it follows that [Φ : Ξ] < ∞. This entails that Φ = (1) Λ (Φ). Note the above relations imply that ww∗ ∈ ηL(Ξ)η∗ ⊆ ηL(Φ)η∗. QN Consider d ∈ P(L(Φ)) such that ww∗ = ηdη∗ and letting µ := c∗η and h := b ⊗ a then relation (5.10) gives the desired conclusion. (cid:4) Claim 5.3. There exists a unitary w ∈ M such that wL(Φ)w∗ = L(Γ). Proof of Claim 5.3. From Claim 5.2, there exists Φ 6 Λ with Φ = QN and µ ∈ U (M) satisfying h = µdµ∗ ∈ L(Γ) and (1) Λ (Φ), d ∈ P(L(Φ)) µdL(Φ)dµ∗ = hL(Γ)h. (5.11) As Γ has property (T), (5.11) implies that dL(Φ)d is a property (T) von Neumann algebra. By [CI17, Lemma 2.13] it follows that Φ is a property (T) group. Fix r ∈ P((µL(Φ)µ∗ ) ′ ∩ M) and note that µL(Φ)µ∗r is a property (T) von Neumann algebra. Thus, using [, Theo- rem] we have that either (1) µL(Φ)µ∗r ≺M L(Γ), or (2) µL(Φ)µ∗r ≺M L(HF) for some finite F ⊂ I. If (2) would hold then we would have that L(H I\F) = L(HF) ′ ∩ M ≺M (µL(Φ)µ∗ r) ′ ∩ (1) rMr = rµ(L(Φ) ′ ∩ M)µ∗r. On the other hand since QN Λ (Φ) = Φ we have µ(L(Φ) ′ ∩ 21 M)µ∗ = Z(µL(Φ)µ∗ ) . Altogether these would show that H I\F is amenable and hence H is amenable, a contradiction. So (1) must hold for every r ∈ P((µL(Φ)µ∗ ) ′ ∩ M). This entails that µL(Φ)µ∗ ≺s M L(Γ) and using (5.11) and [CI17, Lemma 2.6] one can find w ∈ U (M) such that wL(Φ)w∗ = L(Γ). (cid:4) Next consider the subgroup G = {u∆(uγ)u∗ γ ∈ Γ} 6 U (L(Γ × Γ)). Since G nor- malizes u∆(A)u∗, which by (1) satisfies u∆(A)u∗ ≺s A ¯⊗A, the argument in Step 5 in the proof of [IPV10, Theorem 5.1] implies that hΓ×Γ(G) > 0. Then using Lemmas 2.4-2.5 we further have that hΛ(Γ) > 0. Using Lemma 2.4 we get hΛ(w∗Γw) > 0 and by Claim 5.3 we further conclude that hΦ(w∗Γw) > 0. Thus by [IPV10, Theorem 3.1] one can find t ∈ U (M), a character ζ : Γ → T and a group isomorphism δ : Γ → Φ such that tuγt∗ = ζ(γ)vδ(γ), for all γ ∈ Γ. (5.12) Letting Ω := (t∗ ⊗ t∗)∆(t) ∈ U (M ¯⊗M) we then have Ω∆(uγ)Ω∗ = ζ(γ)(uγ ⊗ uγ) for all γ ∈ Γ. Next, we replace the canonical unitaries uγ, γ ∈ Γ by tuγt∗ and A by tAt∗. Then (2) combined with the argument from the proof of Claim 4.5 in Theorem 4.4 further shows that ∆(A) ⊂ A ¯⊗A. Hence using [IPV10, Lemma 7.1.2] there exists a subgroup Σ < Λ such that A = L(Σ). Since the uγ's normalize A, it follows that vδ(γ) normalizes Σ, for all γ. Moreover, since L(Λ) = A ⋊ Γ, then Λ admits a semidirect product decomposition Λ = Σ ⋊ Φ. Altogether the previous considerations give the conclusion of the theorem. (cid:3) Theorem 5.4. Let H be icc, weakly amenable, biexact property (T) group. Let Γ = Γ1 × Γ2, where Γi are icc, biexact, property (T) group. Assume that Γ y I is an action on a countable infinite set I that satisfies the following properties: a) The stabilizer StabΓ(i) is amenable for each i ∈ I; b) There is k ∈ N such that for each J ⊆ I satisfying J ≥ k we have StabΓ(J) < ∞. Let G = H ≀I Γ be the corresponding generalized wreath product. Let Λ be an arbitrary group and let θ : L(G) → L(Λ) be a ∗-isomorphism. Then one can find non-amenable icc groups Σ0, Ψ, an amenable icc group A, and an action Ψ yα A such that we can decompose Λ as semidirect product Λ = (Σ is the generalized Bernoulli action. In Ψ, where Ψ yβ Σ (I) 0 ⊕ A) ⋊β⊕α (I) 0 addition, there exist a group isomorphism δ : Γ → Ψ, a character η : Γ → T, a ∗-isomorphism (I) 0 ⊕ A) and u ∈ U (L(Λ)) so that for every x ∈ L(H(I)) and γ ∈ Γ we θ0 : L(H(I)) → L(Σ have θ(xuγ) = η(γ)uθ0(x)vδ(γ)u∗. Here {ug γ ∈ Γ} and {vλ λ ∈ Ψ} are the canonical unitaries of L(Γ) and L(Ψ), respectively. Proof. Let G = H ≀I Γ satisfies the conditions stated in the Theorem 5.1. Let Λ be an arbi- trary group and assume that θ : L(G) → L(Λ) is an ∗-isomorphism. Using Theorem 5.1, after composing θ with an inner automorphism of M one can find a semidirect product decomposition of Λ = Σ ⋊β Ψ, a group isomorphism δ : Γ → Ψ, and a character η : Γ → T such that (5.13) θ(uγ) = η(γ)vδ(γ) for all γ ∈ Γ. Moreover, we have θ(L(Σ)) = L(H(I)). Since H are icc, biexact, weakly amenable, property (T) groups then Theorem A implies that one can decompose Σ = ⊕i∈I Σi ⊕ A, 22 where A is trivial or amenable icc. In addition, for every finite subset F ⊂ I there exist u ∈ U (L(Λ)) and scalars ti > 0 for i ∈ F such that uL(Σi)ti u∗ = θ(L(Hi)) for all i ∈ F, and uL(⊕i∈I\F Σi ⊕ A)Qi∈F t−1 i u∗ = θ(L(HI\F)). (5.14) Next we show that Σi ∼= Σ0 for all i and there exists an action Ψ yα A such that Λ = (Σ(I) ⊕ A) ⋊b⊕α Ψ, where Ψ yb Σ(I) is the generalized Bernoulli action induced by Γ y I. Fix i, j ∈ I and γ ∈ Γ such that γi = j. Let F ⊂ I be a finite set such that {i, j} ⊆ F. Using the first relation of (5.14) for i and j in combination with (5.13) we get uL(Σj)tju∗ = θ(L(Hj)) = θ(uγ)uL(Σi )ti u∗θ(u∗ δ(γ). In particular this relation implies that L(Σj) ≺M L(βδ(γ)(Σi)) and L(βδ(γ)(Σi)) ≺M L(Σj). Since Σj, βδ(γ)(Σi) are normal subgroups of Λ these intertwinings combined with [CI17, Lemma 2.2] imply that Σj is commensurable with βδ(γ)(Σi); in other words [Σj : Σj ∩ βδ(γ)(Σi)] < ∞ and [βδ(γ)(Σi) : Σj ∩ βδ(γ)(Σi)] < ∞. (5.15) Since βδ(γ)(Σi) 6 ⊕i∈I Σi ⊕ A using the second relation in (5.15) there exists a finite subset j ∈ J ⊂ I so that βδ(γ)(Σi) 6 ΣJ ⊕ A. Thus we have the following normal subgroups Σj ∩ βδ(γ)(Σi) ⊳ βδ(γ)(Σi) ⊳ ΣJ ⊕ A. Taking the quotient we get a finite normal subgroup δ(γ)L(βδ(γ)(Σi))ti vδ(γ)u∗v∗ γ) = vδ(γ)uv∗ J\{j} ⊕ A. J\{j} ⊕ A). However, since Σ J\{j} ⊕ βδ(γ)(Σi)/Σj ∩ βδ(γ)(Σi) ⊳ (ΣJ ⊕ A)/Σj ∩ βδ(γ)(Σi) = Σj/Σj ∩ βδ(γ)(Σi) ⊕ Σ Hence βδ(γ)(Σi)/Σj ∩ βδ(γ)(Σi) ⊳ vZ(Σj/Σj ∩ βδ(γ)(Σi) ⊕ Σ A) is icc by (5.15) we have vZ(Σj/Σj ∩ βδ(γ)(Σi) ⊕ Σ J\{j} ⊕ A) = Σj/Σj ∩ βδ(γ)(Σi). Alto- gether these relations show that βδ(γ)(Σi)/Σj ∩ βδ(γ)(Σi) ⊳ Σj/Σj ∩ βδ(γ)(Σi) and hence βδ(γ)(Σi) 6 Σj. Similarly one can show that βδ(γ)(Σi) > Σj and hence βδ(γ)(Σi) = Σj. ∼= Σ0 for all i. Moreover there is an action Ψ y I which induces a This shows that Σi generalized Bernoulli action Ψ yb ⊕i∈I Σi. Also since the action of Ψ yβ Σ leaves the subgroup ⊕I Σi invariant then Ψ will also leave invariant CΣ(⊕I Σi) = A. Hence there exists an action Ψ yα A such that Λ = (Σ(I) Ψ. The remaining part of the 0 ⊕ A) ⋊b⊕α statement follows directly from the above considerations. (cid:3) Proof of Corollary D. This follows proceeding in the same manner as in the proof of Corol- lary B and using Theorem 5.4. (cid:3) Remarks 5.5. When considering generalized Bernoulli actions it is clear the conditions presented in the statements of Theorems 5.1, 5.4 are satisfied when all the stabilizers of action Γ y I are finite. On the other hand, if one wants to tackle the infinite amenable stabilizers situation, pro- ducing examples seems far more challenging. In this direction we would like to present a possible approach for this which was suggested to us by Professor Denis Osin dur- ing the AIM workshop "Classification of group von Neumann algebras". Consider Σ0 an icc finitely generated amenable group. By [AMO06, Theorem 1.2] there exists an icc supragroup Σ0 < Γ0 that has property (T) and is hyperbolic relatively to Σ0. Hence by [Oz06] it follows that Γ0 is biexact. Let Γ = Γ0 × Γ0 and consider the diagonal subgroup Σ = diag(Σ0 ) < Γ. Also let Γ y I = Γ/Σ be the action by left multiplication on the 23 right cosets Γ/Σ. Since Σ0 < Γ0 is icc and almost malnormal it follows that the one-sided (1) Γ (Σ) = Σ. In turn this is equivalent with condition c) in quasinormalizer satisfies QN Theorem 4.4. Finally one can check that condition b) in Theorems 5.1, 5.4 is equivalent with the property that the group Σ has finite height in Γ (or it is almost malnormal). This is equivalent to the following property: there exists k ∈ N such that any subset F < Σ0 with F ≥ k has finie centralizer CΓ0(F). While f.g. groups like this exist in general (e.g. monster groups) it is unclear if one can construct amenable examples. In any case a possible positive answer to this last group theoretic question would lead to a class of generalized wreath products constructions with non-amenable core that are recognizable from the von Neumann algebraic setting. Indeed, Theorem 4.2 together with the argument from the proof of Claim 4.6 in Theorem 4.4 give the following Corollary 5.6. Let H0, Γ be icc, property (T) groups. Also assume that Γ = Γ1 × Γ2, where Γi are nonamenable biexact groups for all i = 1, 2. Let Γ y I be an action on a countable set I satisfying the following conditions: a) The stabilizer StabΓ(i) is amenable for each i ∈ I; b) There is k ∈ N such that for each J ⊆ I satisfying J ≥ k we have StabΓ(J) < ∞. c) The orbit StabΓ(i) · j is infinite for all i 6= j. Denote by G = H0 ≀I Γ the corresponding generalized wreath product. Let Λ be any torsion free group and let θ : L(G) → L(Λ) be a ∗-isomorphism. Then Λ admits a wreath product decomposition Λ = Σ0 ≀I Ψ satisfying all the properties enumerated in a)-c). In addition there exist a group isomorphism ρ : Γ → Ψ, a character η : Γ → T, a ∗-isomorphism θ0 : L(H0) → L(Σ0) (I) 0 ) and γ ∈ Γ we have and a unitary v ∈ L(Λ) such that for every x ∈ L(H ¯⊗I θ(xuγ) = η(γ)v∗θ 0 (x)vδ(γ)v. Here {uγ γ ∈ Γ} and {vλ λ ∈ Ψ} are the canonical group unitaries of L(Γ) and L(Ψ), respec- tively. [AMO06] G. Arzhantseva, A. Minasyan, D. Osin, The SQ-universality and residual properties of relatively hy- REFERENCES [BO06] [B14] perbolic groups, J. of Algebra, 315 (2007), 165-177. I. Belegradek, D. Osin, Rips construction and Kazhdan property (T), Groups, Geom., Dynam. 2 (2008), 1–12. M. Berbec, W∗-superrigidity for wreath products with groups having positive first ℓ2-Betti number, In- ternat. J. Math. 26 (2015), 1550003, 27 pp. [BV13] M. Berbec, S. Vaes, W∗-superrigidity for group von Neumann algebras of left-right wreath products, Proc. Lond. Math. Soc. (3) 108 (2014), 1116–1152. [BKKO14] E. Breuillard, M. Kalantar, M. Kennedy, N. Ozawa C∗-simplicity and the unique trace property for discrete groups, Publ. Math. Inst. Hautes ÃL'tudes Sci. 126 (2017), 35–71. [BO08] N. P. Brown and N. Ozawa, C∗-algebras and finite-dimensional approximations, Graduate Studies in [Bo12] [BC14] Mathematics, vol. 88, AMS, Providence, RI. R, Boutonnet, W∗-superrigidity of mixing Gaussian actions of rigid groups, Adv. Math., 244 (2013), 69-9-0. R, Boutonnet and A. Carderi, Maximal amenable von Neumann subalgebras arising from maximal amenable subgroups, to appear in Geom. Funct. Anal., ArXiv:1411.4093. [BHR12] R. Boutonnet, C. Houdayer, and S. Raum, Amalgamated free product type III factors with at most one Cartan subalgebra. Compos. Math. 150 (2014), 143–174. 24 [CdSS16] I. Chifan, R. de Santiago, and T. Sinclair, W∗-rigidity for the von Neumann algebras of products of hyperbolic groups, Geom. Funct. Anal. 26 (2016), 136-159. [CI08] [CH08] [CK14] [CI17] [CS11] [CKP14] [CdSS17] I. Chifan, R. de Santiago, and W. Sucpikarnon, Tensor product decompositions of II1 fastors arising from extensions of amalgamated free product groups, arXiv:1710.05019. I. Chifan and A. Ioana, Ergodic subequivalence relations induced by a Bernoulli action, Geom. Funct. Anal. 20 (2010), 53–67. I. Chifan and C. Houdayer, Bass-Serre rigidity results in von Neumann algebras, Duke Math. J. 153 (2010), 23–54. I. Chifan and Y. Kida, OE and W∗ superrigidity results for actions by surface braid groups, Preprint February 2014, ArXiv:1502.02391. I. Chifan, Y. Kida, and S. Pant, Structural results for the von Neumann algebras associated with surface braid groups, preprint, Int. Math. Res. Not. 2016 (2016), 4807–4848. I. Chifan, A. Ioana, Amalgamated free product rigidity for group von Neumann algebras, to appear in Adv. Math. (arXiv:1705.07350). I. Chifan and T. Sinclair, On the structural theory of II1factors of negatively curved groups, Ann. Sci. Éc. Norm. Sup. 46 (2013), 1–33. I. Chifan, T. Sinclair, and B. Udrea, On the structural theory of II1 factors of negatively curved groups, II. Actions by product groups, Adv. Math. 245 (2013), 208–236. I. Chifan and J. Peterson, Some unique group measure space decomposition results, Duke Math. J. 162 (2013), 1923–1966. I. Chifan, A. Ioana, and Y. Kida, W∗-superrigidity for arbitrary actions of central quotients of braid groups, Math. Ann. 361 (2015), 563–582. I. Chifan, T. Sinclair and B. Udrea, Inner amenability for groups and central sequences in factors, Er- godic Theory Dynam. Systems 36 (2016), 1106–1029. [Ch79] E. Christensen, Subalgebras of a finite algebra, Math. Ann. 243 (1979) 17-29. [Co76] A. Connes, Classification of injective factors, Ann. Math. 104 (1976) 73-115. [CH89] M. Cowling and U. Haagerup, Completely bounded multipliers of the Fourier algebra of a simple Lie [CSU13] [CSU11] [CIK13] [CP10] group of real rank one, Invent. Math. 96 (1989) 507-549. [DHI16] D. Drimbe, D. Hoff, A. Ioana, Prime II1 factors arising from irreducible lattices in products of rank one [JGS10] [FV10] simple Lie groups, arXiv:1611.02209. J. Fang, S. Gao, and R. Smith, The Relative Weak Asymptotic Homomorphism Property for Inclusions of Finite von Neumann Algebras, Intern. J. Math. 22 (2011) 991–1011. P. Fima and S. Vaes, HNN extensions and unique group measure space decomposition of II1 factors, Trans. Amer. Math. Soc. 364 (2012), 2601–2617. L. Ge, On maximal injective subalgebras of factors, Adv. Math. 118 (1996), 34–70. [G96] [HPV10] C. Houdayer, S. Popa, and S. Vaes, A class of groups for which every action is W∗-superrigid, Groups Geom. Dyn. 7 (2013), 577–590. [HV12] C. Houdayer and S. Vaes, Type III factors with unique Cartan decomposition, J. Math. Pures Appl. 100 (2013), 564–590. [HU15] C. Houdayer, Y. Ueda, Rigidity of free product von Neumann algebras, to appear in Compos. Math. [Io10] A. Ioana, W∗-superrigidity for Bernoulli actions of property (T) groups, J. Amer. Math. Soc. 24 (2011), 1175–1226. A. Ioana, Rigidity results for wreath product II1 factors, J. Funct. Anal. 252 (2007), 763–791. [Io06] [IPP05] A. Ioana, J. Peterson, and S. Popa, Amalgamated free products of weakly rigid factors and calculation of their symmetry groups. Acta Math. 200 (2008), 85–153. [Io12a] A. Ioana, Cartan subalgebras of amalgamated free product II1 factors, With an appendix by Ioana and [Io11] [Io12] [Io17] Stefaan Vaes. Ann. Sci. Ec. Norm. Super. (4) 48 (2015), 71–130. A. Ioana, Uniqueness of the group measure space decomposition for Popa's HT factors, Geom. Funct. Anal. 22 (2012), 699–732. A. Ioana, Classification and rigidity for von Neumann algebras, European Congress of Mathematics, 601-625, Eur. Math. Soc., Zurich, 2013, 46–02. A. Ioana, Rigidity for von Neumann algebras, submitted to Proceedings of the ICM 2018. 25 [IPV10] A. Ioana, S. Popa, and S. Vaes, A Class of superrigid group von Neumann algebras, Ann. of Math. (2) [Is12] [Is14] 178 (2013), 231–286. Y. Isono, Examples of factors which have no Cartan subalgebras, to appear in Trans. Amer. Math. Soc., ArXiv1209.1728. Y. Isono, Some prime factorization results for free quantum group factors, to appear in J. Reine Angew. Math., ArXiv:1401.6923. Y. Isono, On fundamental groups of tensor product II1 factors, Preprint 2016, arXiv:1608.06426. V.F.R. Jones, Index for subfactors, Invent. Math. 72 (1983), 1–25. [Is16] [Jo81] [KV15] A. S. Krogager, S. Vaes, A class of II1 factors with exactly two group measure space decompositions, J. Math. Pures Appl. (9) 108 (2017), 88–110. [Ma79] G.A. Margulis, Finiteness of quotients of discrete groups, Func. Anal. Appl. 13 (1979), 178-187. [MvN43] F. J. Murray and J. von Neumann, On rings of operators, IV, Ann. of Math. 44 (1943), 716-808. [Oz03] N. Ozawa, Solid von Neumann algebras, Acta Math., 192 (2004), 111–117. [Oz05] N. Ozawa, A Kurosh type theorem for type II1 factors, Int. Math. Res. Not., Volume 2006, Article ID97560 (21 pages). [Oz06] N. Ozawa, Boundary amenability of relatively hyperbolic groups, Topology Appl. 53 (2006) 2624-2630. [OP03] N. Ozawa and S. Popa, Some prime factorization results for type II1 factors, Invent. Math., 156 (2004), 223–234. [OP07] N. Ozawa and S. Popa, On a class of II1 factors with at most one Cartan subalgebra, Ann. of Math. 172 (2010), 713-749. [OP08] N. Ozawa and S. Popa, On a class of II1 factors with at most one Cartan subalgebra. II, Amer. J. Math. [Pe06] [Pe09] [Po94] [Po99] [Po01] [Po03] [Po04] [Po08] [Po06] [Po05] [Po06] [PV09] [PV11] [PV12] [Si10] 132 (2010), 841–866. J. Peterson, L2-rigidity in von Neumann algebras, Invent. Math. 175 (2009), no. 2, 417–433. J. Peterson, Examples of group actions which are virtually W∗-superrigid, Preprint February 2009, ArXiv:1002.1745. S. Popa, Classification of subfactors and their endomorphisms. CBMS Regional Conference Series in Mathematics, 86. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 1995. x+110 pp. S. Popa, Some properties of the symmetric enveloping algebra of a factor, with applications to amenability and property (T), Doc. Math. 4 (1999), 665–744. S. Popa, On a class of type II1 factors with Betti numbers invariants, Ann. of Math. 163 (2006), 809-899. S. Popa, Strong Rigidity of II1 Factors Arising from Malleable Actions of w-Rigid Groups I, Invent. Math. 165 (2006), 369–408. S. Popa, Strong Rigidity of II1 Factors Arising from Malleable Actions of w-Rigid Groups II, Invent. Math. 165 (2006), 409–453. S. Popa, On the superrigidity of malleable action with spectal gap, J. Amer. Math. Soc. 21 (2008), 981– 1000. S. Popa, On Ozawa's property for free group factors, Int. Math. Res. Not. Vol. 2007 : article ID rnm036, 10 pages. S. Popa, Cocycle and orbit equivalence superrigidity for malleable actions of w-rigid groups, Invent. Math. 170 (2007), 243–295. S. Popa, Deformation and rigidity for group actions and von Neumann algebras, International Congress of Mathematicians. Vol. I, 445–477, Eur. Math. Soc., Zürich, 2007. S. Popa and S. Vaes, Group measure space decomposition of II1 factors and W∗-superrigidity, Invent. Math. 182 (2010), 371–417. S. Popa and S. Vaes, Unique Cartan decomposition for II1 factors arising from arbitrary actions of free groups, Acta Math. 212 (2014), 141–198. S. Popa and S. Vaes, Unique Cartan decomposition for II1 factors arising from arbitrary actions of hyper- bolic groups, J. Reine Angew. Math. 694 (2014), 215–239. T. Sinclair, Strong solidity of group factors from lattices in SO(n, 1) and SU(n, 1), J. Funct. Anal. 260 (2011), no. 11, 3209–3221. [SW12] O. Sizemore and A. Winchester, A unique prime decomposition result for wreath product factors, Pacific J. Math. 265 (2013), no. 1, 221–232. 26 [Va07] [V10] [Va10] [Va13] [VV14] [V96] S. Vaes, Explicit computations of all finite index bimodules for a family of II1 factors, Ann. Sci. Éc. Norm. Sup. 41 (2008), 743-788. S. Vaes, Rigidity for von Neumann algebras and their invariants, Proceedings of the International Con- gress of Mathematicians (Hyderabad, India, 2010), Vol. III, 1624–1650, Hindustan Book Agency, New Delhi, 2010. S Vaes, One-cohomology and the uniqueness of the group measure space decomposition of a II1 factor, Math. Ann. 355 (2013), 661-696. S. Vaes, Normalizers inside amalgamated free products von Neumann algebras, Publ. Res. Inst. Math. Sci. 50 (2014), 695–721. S. Vaes and P. Verraedt, Classification of type III Bernoulli crossed products, Adv. Math. 281 (2015), 296–332. D-V. Voiculescu,The analogues of entropy and of Fisher's information measure in free probability theory: the absence of Cartan subalgebras, Geom. Funct. Anal. 6 (1996), no. 1,172–199. DEPARTMENT OF MATHEMATICS, THE UNIVERSITY OF IOWA, 14 MACLEAN HALL, IA 52242, USA E-mail address: [email protected] DEPARTMENT OF MATHEMATICS, THE UNIVERSITY OF IOWA, 14 MACLEAN HALL, IA 52242, USA E-mail address: [email protected] 27
1008.0903	1	1008	2010-08-05T02:40:54	Dilations of interaction groups that extend actions of Ore semigroups	[ "math.OA" ]	We show that every interaction group extending an action of an Ore semigroup by injective unital endomorphisms of a C-algebra, admits a dilation to an action of the corresponding enveloping group on another unital C-algebra, of which the former is a C*-subalgebra: the interaction group is obtained by composing the action with a conditional expectation. The dilation is essentially unique if a certain natural condition of minimality is imposed. If the action is induced by covering maps on the spectrum, then the expectation is faithful.	math.OA	math	DILATIONS OF INTERACTION GROUPS THAT EXTEND ACTIONS OF ORE SEMIGROUPS FERNANDO ABADIE Abstract. We show that every interaction group extending an action of an Ore semigroup by injective unital endomorphisms of a C ∗-algebra, admits a dilation to an action of the corresponding enveloping group on another unital C ∗-algebra, of which the former is a C ∗-subalgebra: the interaction group is obtained by composing the action with a condi- tional expectation. The dilation is essentially unique if a certain natural condition of minimality is imposed. If the action is induced by covering maps on the spectrum, then the expectation is faithful. 1. Introduction and preliminaries The notions of interaction groups and their crossed products have been introduced and studied by Exel in [9], with the aim of dealing with irre- versible dynamical systems. The mentioned paper emerges as a culmination of previous work in the subject appeared in [6], [8], [7], [11]. Related work may be found as well in [13], [3], [14], [5]. Recently, Exel and Renault studied in [10] a family of interaction groups that extend actions of some semigroups on unital commutative C ∗-algebras. Suppose that X is a compact Hausdorff space and θ : X → X is a covering map. For A = C(X), let α : A → A be the dual map of θ, i.e.: α(a) = a ◦ θ, which is a unital injective endomorphism of A. In case there exists a transfer operator ([6]) for α, that is, a positive linear map L : A → A such that L(α(a)b) = aL(b), ∀a, b ∈ A, then V : Z → B(A) (here B(A) is the algebra of bounded operators from A into itself) given by Vn =(αn n ≥ 0 L−n n < 0 is called an interaction group (Definition 1.1). This interaction group is clearly an extension of the action ¯α : N × A → A given by (n, a) 7→ αn(a). Conversely, it can be shown that if W : Z → B(A) is an interaction group that extends ¯α, then W−1 is a transfer operator for α, and W is retrieved from α and W−1 from the construction above. That is: interaction groups that extend ¯α are in a natural bijection with transfer operators for α. In the same way, interaction groups that extend the action of an Ore semigroup correspond to semigroups of transfer operators corresponding to the endomorphisms of the 2000 Mathematics Subject Classification. Primary 46L05. Key words and phrases. Interaction groups. Partially supported by CNPq, Brazil, Processo 151654/2006-9. 1 2 FERNANDO ABADIE action. In the case of actions on commutative algebras the work in [10] shows that one can replace transfer operators by cocycles (see Definition 3.9). We will show later that an interaction group as the one above can be written as the composition of an action β with a conditional expectation F : Vn = F βn, ∀n ∈ Z, a decomposition that reflects the combination of the deterministic and probabilistic elements included in the concept of interaction group. On the other hand, it seems that interaction groups are closely related with partial actions. Propositions 1.3 and 1.7 below are instances of this relation. Moreover, under certain conditions one may construct interaction groups from actions of groups and conditional expectations, in a way that resembles the construction of partial actions by the restriction of global ones. In fact, suppose that A is a C ∗-subalgebra of the unital C ∗-algebra B, F : B → B is a conditional expectation with range A, and β : G × B → B is an action of a group G on B. Let Ft : B → B be given by Ft := βtF βt−1 . Then Ft is a conditional expectation onto βt(A), ∀t ∈ G. It is not hard to prove that if F Ft = FtF , ∀t ∈ G, then V : G → B(A) such that Vt(a) = F (βt(a)), ∀a ∈ A, t ∈ G is an interaction group (provided F (βt(1A)) = 1A, ∀t ∈ G, see 1.2 below). With the same spirit of the work done in [1], although with different meth- ods, we show in the present paper that any interaction group that extends an action of an Ore semigroup by unital injective endomorphisms (for instance those studied in [10]) is of this form, that is, it can be obtained by compos- ing an action with a conditional expectation. The existence of the action is due to Laca's Theorem (see [12] and Theorem 2.2 below) on the dilation of actions of Ore semigroups. The conditional expectation is constructed as the limit of the directed system of transfer operators corresponding to the endomorphisms of the Ore semigroup action. The structure of the present paper is the following. In the rest of this section we study some relations between interaction groups and partial ac- tions and we introduce the notion of dilation of an interaction group. In the next section we prove our main result, Theorem 2.4, and in the final one we see how this theorem applies, with a refinement, to the interaction groups studied by Exel and Renault in [10]. 1.1. Interaction groups. We show here how to get interaction groups from suitable pairs of actions and conditional expectations. Recall that a par- tial representation of a group G on a Banach space A is a map V : G → B(A), the Banach algebra of bounded linear operators on A, such that Ve = Id Vs−1VsVt = Vs−1Vst ∀s, t ∈ G ∀s, t ∈ G VsVtVt−1 = VstVt−1 (e the unit of G)   Definition 1.1. An interaction group is a triple (A, G, V ) where A is a unital C ∗-algebra, G is a group, and V is a map from G into B(A), which satisfies: DILATIONS OF INTERACTION GROUPS 3 (1) Vt is a positive unital map, ∀t ∈ G. (2) V is a partial representation. (3) Vt(ab) = Vt(a)Vt(b) if either a or b belongs to Vt−1(A). If the group G is understood we will put just (A, V ) (or even V if A is un- ψ → (A′, G, V ′) t ψ, derstood as well) instead of (A, G, V ). A morphism (A, G, V ) is a unital homomorphism of C ∗-algebras ψ : A → A′ such that ψVt = V ′ ∀t ∈ G. It will be useful for our purposes to consider the following couple of cat- egories, TG and DG associated to a group G. The objects of TG are triples T = (B, β, F ), where β is an action of the group G on the unital C ∗- algebra B, and F : B → B is a conditional expectation, that is, a norm one idempotent whose range is a C ∗-subalgebra of B. Recall that a condi- tional expectation F is a positive F (B)-bimodule map. If T = (B, β, F ), T ′ = (B′, β′, F ′) ∈ TG, by a morphism φ : T → T ′ we mean a unital homo- morphism of C ∗-algebras φ : B → B′ such that φF = F ′φ and φβt = β′ tφ, ∀t ∈ G. The category DG is the full subcategory of TG whose objects (B, β, F ) satisfy the following two conditions: a) F βtF (1) = F (1), ∀t ∈ G, and b) FrFs = FsFr, ∀r, s ∈ G, where Fr = βrF βr−1, ∀r ∈ G. Note that F (1) is the unit of F (B), and that Fr is a conditional expectation with range βr(F (B)). Proposition 1.2. Let T = (B, β, F ) ∈ TG, and A := F (B). If F Ft = FtF , ∀t ∈ G, then: (1) FrFs = FsFr, ∀r, s ∈ G, and FrFs is a conditional expectation with range βr(A) ∩ βs(A). (2) If Vt := F βtA, ∀t ∈ G, then the map V : G → B(A) given by t 7→ Vt is a partial representation and satisfies condition 3. of 1.1. Moreover the range of Vt is A ∩ βt(A), ∀t ∈ G. (3) If V is the map defined in 2, then V is an interaction group if and only if F (βt(1A)) = 1A for every t ∈ G. That is: V is an interaction group if and only if T ∈ DG. Proof. We have βrF βr−1βsF βs−1 = βsFs−1rF βs−1 = βsF Fs−1rβs−1 = βsF βs−1βrF βr−1βsβs−1. That is, FrFs = FsFr, and therefore FrFs is a conditional expectation with range Fr(B) ∩ Fs(B). On the other hand Fr(B) = βrF βr−1(B) = βrF (B) = βr(A). Hence FrFs(B) = βr(A) ∩ βs(A). As for 2., V is a partial representation: Vs−1VsVt = F Fs−1F βt = F Fs−1βt = Vs−1Vst, VsVtVt−1 = F βstFt−1 F βt−1 A = VstF βt−1 IdA = VstVt−1. 4 FERNANDO ABADIE If x ∈ A and a = Vt−1(x), b ∈ B, then: Vt(ab) = Vt(Vt−1 (x)b) = F (Ft(x)βt(b)) = F (FtF (x)βt(b)) = F (F Ft(x)βt(b)) = F Ft(x)F (βt(b)) = VtVt−1(x)Vt(b) = Vt(a)Vt(b). Since Vt(ba) = Vt(a∗b∗)∗, we have shown that V satisfies condition 3. of 1.1. On the other hand, Vt(A) = F βt(A) = F Ft(βt(A)) = A ∩ βt(A), because F Ft is a conditional expectation with range A ∩ βt(A) ⊆ βt(A). Now, if V is an interaction group, then F βt(1A) = Vt(1A) = 1A, ∀t ∈ G. Conversely, if F βt(1A) = 1A ∀t ∈ G, then Vt is a positive unital map. In addition Ve = F βeA = F IdA = IdA; hence V is an interaction group. (cid:3) 1.2. The partial action of an interaction group. We will see now that every interaction group has naturally associated a partial action of the group on the same algebra. Recall that a partial action of a discrete group G on a set X is a pair ({Xt}t∈G, {γt}t∈G) where, for every t ∈ G, Xt is a subset of X, γt : Xt−1 → Xt is a bijection, and γst extends γsγt, ∀s, t ∈ G. It is also assumed that γe = idX . When X is a C ∗-algebra, it is usually supposed that Xt is an ideal and that γt is an isomorphism of C ∗-algebras. So we warn the reader that for the partial actions we consider in this paper the sets Xt will be unital C ∗-subalgebras. Proposition 1.3. Suppose that V : G → B(A) is an interaction group. For t ∈ G let At := Vt(A), and γt : At−1 → At be such that γt(a) = Vt(a). Then: (1) Every At is a unital C ∗-subalgebra of A (with the same unit), and γt is an isomorphism between At−1 and At, ∀t ∈ G. (2) The map Et : A → A given by Et := VtVt−1 is a conditional expecta- tion onto At, ∀t ∈ G, and ErEs = EsEr, ∀r, s ∈ G. (3) The pair γ := ({At}t∈G, {γt}t∈G) is a partial action of G on A. Proof. We already know by [9, 3.2] that At is a unital C ∗-subalgebra of A with unit Vt(1A) = 1A, and that γt is an isomorphism, ∀t ∈ G. Since V is a partial representation we have that γe = Ve = Id. Suppose now that c belongs to the domain of γsγt, that is, c ∈ At−1 is such that γt(c) ∈ As−1. Then γsγt(c) ∈ As and γsγt(c) = VsVt(Vt−1(γt(c))) = Vst(Vt−1 (γt(c))) = Vst(c) ∈ Ast. Then γsγt(c) ∈ As ∩ Ast, and we may apply γt−1s−1 to γsγt(c). Since V is a partial action we obtain: γt−1s−1γsγt(c) = Vt−1s−1Vsγt(c) = Vt−1Vs−1Vsγt(c) = γt−1γs−1γsγt(c) = c, whence γst(c) = γsγt(c). This shows that γst extends γsγt, ∀s, t ∈ G, and therefore γ is a partial action. (cid:3) Observe that if V is an interaction group of the type considered in 1.2, then Er = F FrA, and ErEs = F FrFsA (with the notations of 1.2 and 1.3). DILATIONS OF INTERACTION GROUPS 5 The usual notion of partial actions of groups on C ∗-algebras requires that the domains of the partial automorphisms are ideals. In the commu- tative case, partial actions on a C ∗-algebra correspond exactly with partial actions on the spectrum of the algebra, where the domains of the partial homeomorphisms are open subsets of the spectrum ([2, Proposition 1.5]). Instead, partial actions on unital commutative C ∗-algebras as the ones con- sidered in 1.3 lead to a different notion of partial action on a topological space. In fact, let A = C(X) be a unital commutative C ∗-algebra, and let γ = ({At}, {γt}) be a partial action of G on A, where each At is a unital subalgebra of A, with the same unit. Then the dual notion of the partial action γ should be expressed in terms of the spectra of the subalgebras At and the maps induced by γ between them. Although we will not give here the exact conditions that such a collection of spaces and maps must satisfy, it is clear that the result is not a partial action in the usual sense, as the spectrum of At is not a subspace but a quotient of X. 1.3. Dilations of interaction groups. We introduce next the notion of dilation of an interaction group V , and we study its relation with the partial action associated with V . Definition 1.4. Let V : G → B(A) be an interaction group. A dilation of V is a pair (i, T ), where T = (B, β, F ) ∈ TG and i : A → B is a homomorphism of C ∗-algebras such that iVt = F βti, ∀t ∈ G. If B = span{βti(a) : a ∈ A, t ∈ G}, we say that the dilation is minimal. The dilation is called faithful if so is F , and it is called admissible if T ∈ DG (recall that a positive map F is called faithful when b 6= 0 implies F (b∗b) 6= 0). Proposition 1.5. Let V : G → B(A) be an interaction group, and suppose that (i, T ) is a minimal dilation of V , where T = (B, β, F ). Then we have F ((F Ft − FtF )(b)∗(F Ft − FtF )(b)) = 0, ∀b ∈ B. Proof. We must show that F βtF βt−1(b) − βtF βt−1 F (b) belongs to the left ideal LF := {b ∈ B : F (b∗b) = 0} of B, ∀b ∈ B. Since B is the closed linear span of the set ∪s∈Gβsi(A), it is enough to prove that F βtF βt−1 (βsi(a)) − βtF βt−1F (βsi(a)) ∈ LF , that is, VtVt−1s(a) − βtVt−1Vs(a) ∈ LF , ∀s ∈ G, a ∈ A. Since F is an A-bimodule map which is the identity operator on A, and since F βtA = Vt, we have that the expression F(cid:0)(VtVt−1s(a) − βtVt−1Vs(a))∗(VtVt−1s(a)−βtVt−1 Vs(a))(cid:1) is equal to VtVt−1s(a∗)(cid:0)VtVt−1s(a)− VtVt−1Vs(a)(cid:1) −VtVt−1Vs(a∗)(cid:0)VtVt−1s(a)−VtVt−1Vs(a)(cid:1), which is zero because V is a partial representation. (cid:3) Corollary 1.6. Any minimal and faithful dilation of an interaction group is admissible. Suppose that β is an action of G on the C ∗-algebra B, and that A is a C ∗-subalgebra of A. The restriction of β to A is the partial action βA := ({A′ t−1, t(a) := βt(a), ∀a ∈ A′ t := A ∩ βt(A) and γ′ t}t∈G, {γ′ t}t∈G), where A′ 6 FERNANDO ABADIE t ∈ G. In case that the C ∗-algebra generated by {βt(a) : a ∈ A, t ∈ G} is all of B, we say that β is an enveloping action for γ′. Proposition 1.7. Suppose that V : G → B(A) is an interaction group with dilation (i, (B, β, F )), where A is a C ∗-subalgebra of B and i : A → B is the natural inclusion. Let γ be the partial action of G on A given by Proposition 1.3, and let γ′ := βA. Then At ⊇ A′ t := A ∩ βt(A) and γt(a) = t(a), ∀t ∈ G, a ∈ A′ γ′ t−1 . If the dilation is admissible then γ = βA. In particular if the dilation is faithful then γ is the restriction of β to A. Proof. If a ∈ A, then a ∈ A′ βt(a) = Vt(a). Then if a ∈ A′ t ⊆ At and γ′ shows that A′ F Ft = FtF , then if a ∈ At we have: t−1 ⇐⇒ βt(a) ∈ A ⇐⇒ βt(a) = F βt(a) ⇐⇒ t−1 we have γ′ t(a) = βt(a) = Vt(a) ∈ At, which t = VtA′ . On the other hand, if = γtA′ t−1 t−1 a = VtVt−1(a) = F Ft(a) = FtF (a) = βt(F βt−1 F (a)) ∈ A ∩ βt(A) = A′ t, whence At = A′ tion 1.5 and Corollary 1.6. t, and γt = γ′ t. The last two assertions follow from Proposi- (cid:3) Corollary 1.8. Suppose that V : G → B(A) is an interaction group with admissible dilation (i, (B, β, F )), where i : A → B is an embedding (i.e.: i is injective). Then the restriction of β to C := span{βti(a) : t ∈ G, a ∈ A} is an enveloping action for the partial action γ of G on A given by Propo- sition 1.3. In particular, if the dilation is minimal then β is an enveloping action for γ. 2. The dilation A cancelative monoid P is called an Ore semigroup if P r ∩ P s 6= ∅, ∀r, s ∈ P . It follows by induction that P is an Ore semigroup if and only if Pt1 ∩ . . . ∩ Ptn 6= ∅, ∀t1, . . . , tn ∈ P . Then P is partially ordered by the relation r ≤ s ⇐⇒ s ∈ P r (equivalently: r ≤ s ⇐⇒ P r ⊇ P s), and it is even directed by that relation. Any cancelative abelian monoid P is an Ore semigroup. In fact, such a monoid embeds in its Grothendieck group G, and every element t ∈ G can be written as t = v−1u, with u, v ∈ P . Therefore, if r, s ∈ P , writing rs−1 = u−1v, with u, v ∈ P , gives t := ur = vs ∈ P r ∩ P s, so P is an Ore semigroup (and P ∋ t ≥ r, s). More generally, we have the following theorem [12, Theorem 1.1.2], which shows that there is a functor from the category of Ore semigroups into the category of groups: Theorem 2.1 (Ore, Dubreil). A semigroup P can be embedded in a group G with P −1P = G if and only if it is an Ore semigroup. In this case the group G is determined up to canonical isomorphism and every semigroup homomorphism φ from P into a group H extends uniquely to a group ho- momorphism ϕ : G → H. DILATIONS OF INTERACTION GROUPS 7 In case P is an Ore semigroup we say that the group G in 2.1 is the enveloping group of P . A key ingredient in our process of dilating the interaction groups under consideration is Laca's theorem [12, 2.1.1]. For the convenience of the reader we recall it below: Theorem 2.2 (M. Laca, [12]). Assume P is an Ore semigroup with en- veloping group G = P −1P and let α be an action of P by unital injective endomorphisms of a unital C ∗-algebra A. Then there exists a C ∗-dynamical system (B, G, β), unique up to isomorphism, consisting of an action β of G by automorphisms of a C ∗-algebra B and an embedding i : A → B such that: (1) β dilates α, that is, βt ◦ i = i ◦ αt, for t in P , and (2) (B, G, β) is minimal, that is, St∈P β−1 Note that i is unital: (i(A)) is dense in B. t βt−1 i(a)i(1A) = βt−1 (i(a)βt(i(1A))) = βt−1 (i(aαt(1A))) = i(a), ∀t ∈ P, so taking adjoints and recalling that {βt−1 (i(a)) : t ∈ P, a ∈ A} is dense in B, we see that i(1A) = 1B. From now on G will denote the enveloping group of the Ore semigroup P . Lemma 2.3. Let α be an action of the Ore semigroup P by unital injective endomorphisms of the unital C ∗-algebra A, and suppose that V : G → B(A) is an interaction group such that V P = α. If (i, (B, β, F )) is an admissible dilation of V , then βti = iVt = iαt, ∀t ∈ P . Proof. Note first that for t ∈ G: F Fti = F βt(F βt−1 i) = F βtiVt−1 = iVtVt−1. If now t ∈ P we have Vt−1αt = idA, and therefore βti = βtiVt−1αt = βtF βt−1 iαt = FtF iαt = F Ftiαt = iVtVt−1 Vt = iVt. (cid:3) Theorem 2.4. Let α be an action of the Ore semigroup P by unital injective endomorphisms of the unital C ∗-algebra A, and suppose that V : G → B(A) is an interaction group such that V P = α. Then V has a minimal admis- sible dilation (i, T ), where T = (B, β, F ) and i : A → B is an embedding, which has the following universal property. If (i′, (B′, β′, F ′)) is another ad- missible dilation of V , then there exists a unique morphism φ : (B, β, F ) → (B′, β′, F ′) such that φi = i′. Therefore the dilation (i, T ) is unique up to isomorphism in the class of minimal and admissible dilations. Proof. Let i : (A, α) → (B, β) be the minimal dilation of (A, α) provided by Laca's theorem. We suppose, as we can do, that i is the natural inclusion, so A ⊆ B. We proceed next to define a conditional expectation F : B → A. To this end note first that if r, s ∈ P , with r ≤ s, and ar, as ∈ A are such 8 FERNANDO ABADIE that βr−1(ar) = βs−1(as), then βsr−1(ar) = as, so αsr−1(ar) = as by 2.3. Therefore Ls(as) = Lsαsr−1(ar) = Vs−1Vsr−1(ar) = LsαsLr(ar) = Lr(ar). Thus we may define F0 : St∈P βt−1 (A) → B such that F0(b) = Lt(βt(b)), ∀b ∈ βt−1 (A). Since kF0(b)k = kLt(βt(b))k ≤ kbk, F0 extends uniquely to a bounded operator F : B → A, which is easily seen to be positive and to satisfy F 2 = F and F (B) = A. Then F is a conditional expectation with range A. We claim that (B, β, F ) is a minimal admissible dilation of V . In fact, if t ∈ G and r, s ∈ P are such that t = r−1s, then F βtA = F βr−1βrtA = F βr−1αs = Lrαs = Vr−1VrVr−1s = Vr−1s = Vt. Since St∈P βt−1(A) is dense in B we have that (B, β, F ) is minimal, and to see that it is also admissible, it is enough to check that F Ftβr−1A = FtF βr−1A, ∀t ∈ G, r ∈ P . On the one hand we have (2.1) On the other hand, let t ∈ G, t = u−1v, u, v ∈ P . Using Lemma 2.3 and recalling that EuEv = EvEu, we have F Ftβr−1A = F βtF βt−1r−1A = VtF Vt−1r−1 = EtVr−1 (2.2) FtF βr−1A = βu−1βvF βv−1uVr−1 = βu−1VvVv−1uVr−1 = βu−1EvEuVuVr−1 = βu−1EuEvVuVr−1 = βu−1VuVu−1VvVv−1VuVr−1 = βu−1βuVtVt−1Vr−1 = EtVr−1 From (2.1) and (2.2) we conclude that (B, β, F ) is admissible. We see next that (B, β, F ) has the claimed universal property. Then suppose that (i′, (B′, β′, F ′)) is another admissible dilation of V . By Lemma 2.3 we have that β′P = i′α, and then by the universal property of the pair (B, β) there exists a unique homomorphism φ : B → B′ such that φi = i′ and β′ tφ = φβt ∀t ∈ G. In particular φβr−1i = β′ r−1i′, ∀r ∈ P . Thus r−1φi = β′ F ′φβr−1i = F ′β′ r−1i′ = i′Vr−1 = φiVr−1 = φF βr−1i, ∀r ∈ P. The equality φF = F ′φ follows now from the density of Sr∈P βr−1i(A) in B and the continuity of the involved maps. Remark 2.5. Suppose V and V ′ are interaction groups that extend actions by injective unital endomorphisms of the Ore semigroup P . Suppose as well that ψ : (A, V ) → (A′, V ′) is a morphism of interaction groups, and let (i, T ) and (i′, T ′) be the corresponding minimal admissible dilations of V and V ′. Then (i′ψ, T ′) is an admissible dilation of V , so there exists a unique morphism φ : T → T ′ such that φi = i′ψ. In this way we obtain a functor from the category of interaction groups that extend actions by injective unital endomorphisms of the Ore semigroup P into the category DG, where G is the enveloping group of P . (cid:3) We end the section with a result concerning enveloping actions. DILATIONS OF INTERACTION GROUPS 9 Proposition 2.6. Let V be an interaction group like in 2.4, and let γ be the partial action associated to V via 1.3. Then γ has an enveloping action, which is unique up to isomorphism. Proof. It follows from 2.4 and 1.8 that the action β provided by Theorem 2.4 is an enveloping action for γ. Suppose now that β′ : G × B′ → B′ is another enveloping action for γ, where B′ is a C ∗-algebra which contains A. To show that β and β′ are isomorphic, it is enough to show that β′ satisfies properties 1. and 2. of Theorem 2.2. It is clear that β′ satisfies the first property, so let us see that it also verifies the second one. Note that if t = r−1s ∈ G, with r, s ∈ P , then β′ r−1(A). On the other hand, suppose r, s ∈ P , with r ≤ s. Then, since sr−1 ∈ P , we have A ⊇ αsr−1(A) = β′ t(A) ⊆ r−1(A), ∀t ∈ G, because P is directed by its partial order. This r−1(A), as we wanted to prove. (cid:3) Sr∈P β′ implies that B′ is the closure of Sr∈P β′ r−1(A). Thus β′ r−1αs(A) ⊆ β′ 3. Dilations of Exel-Renault interaction groups t(A) = β′ sβ′ r−1(A), so β′ s−1(A) ⊇ β′ In this section we specialize to certain interaction groups occuring on commutative C ∗-algebras. More precisely, we are interested in the inter- action groups studied in [10]. In that work, the authors considered right actions θ : P × X → X, where P is an Ore semigroup with enveloping group G, and θt is an onto local homeomorphism of the compact Hausdorff space X, that is, θt : X → X is a covering map. Dualizing, θ induces a left action α of P by injective unital endomorphisms of A = C(X). It is shown in [10] that for α to be extended to an interaction group V : G → B(A) it is enough that there exists a certain map ω : P × X → [0, 1], asso- ciated to θ. This map is called a cocycle and is determined by the fact that Et(a)(x) = Pθt(y)=θt(x) ω(t, y)a(y), ∀t ∈ P , a ∈ A and x ∈ X, where Et = VtVt−1. In this case Theorem 2.4 can be applied, so one concludes that the interaction groups considered by Exel and Renault in [10] have mini- mal admissible dilations. We mention in passing that for these interaction groups Theorem 2.4 could be proved by using exclusively measure-theoretic arguments, but we will not do it here. The aim of this section is to show that the minimal admissible dilations of the Exel-Renault interaction groups are also faithful. 3.1. Conditional expectations on commutative C ∗-algebras. We be- gin by giving a characterization of conditional expectations from a commuta- tive unital C ∗-algebra onto a unital C ∗-subalgebra, suitable for our purposes. We also describe the transfer operators for an endomorphism induced by a covering map. For more information about conditional expectations we refer the reader to [15] and [4]. We fix a notation we will use until the end of the present paragraph. Let B = C(Z) be a unital C ∗-algebra and A = C(X) a unital C ∗-subalgebra of B. Note that X is homeomorphic to the quotient space of Z with respect 10 FERNANDO ABADIE to the relation z ∼ z′ ⇐⇒ a(z) = a(z′), ∀a ∈ A. Let π : Z → X be the corresponding quotient map. Observe that an element a ∈ C(X), when seen as an element of B, sends z ∈ Z into a(π(z)). Denote by P (Z) the set of regular Borel probability measures on Z. Proposition 3.1. With the above notation, let F : B → A be a unital linear map. Then F is positive if and only if there exists a map µ : X → P (Z) that is w∗-continuous and such that (3.1) F (b)(x) =ZZ b(z)dµx(z), ∀b ∈ B, x ∈ X. Equation (3.1) establishes a bijective correspondence between unital positive linear maps F : B → A and w∗-continuous maps µ : X → P (Z). Proof. Let ǫx : A → C be evaluation in x ∈ X. Then if F is positive ǫx ◦ F is a state of B. Let µx be the probability measure provided by the ∀b ∈ B. Since F (b) is a continuous function defined on X, it follows that x 7→ µx is w∗-continuous. Conversely, it is clear that if a w∗-continuous Riesz-Markov representation theorem, such that ǫx ◦ F (b) = RZ b(z)dµx(z), map µ : X → P (Z) is such that F (b)(x) = RZ b(z)dµx(z), ∀b ∈ B, x ∈ X, then F (b) is positive whenever b is positive. Finally, it is obvious that the correspondence F 7→ µ is one to one and onto. (cid:3) It is clear that in Proposition 3.1 above A does not need to be a subalgebra of B. Example 3.2. Suppose α : B → B is an injective unital endomorphism and let A := α(B). Then there exists a homeomorphism ¯ξ : X → Z such that α(b) = b ◦ ¯ξ ∈ A. Therefore α(b)(x) = b(¯ξ(x)) =RZ b(z)dδ¯ξ(x), where δz denotes the Dirac measure concentrated at z. Thus the map µ provided by 3.1 for α is given by: µx = δ¯ξ(x). Proposition 3.3. A linear map F : B → A is an onto conditional ex- pectation if and only if there exists a map µ : X → P (Z) such that µ is w∗-continuous, supp(µx) ⊆ π−1(x), ∀x ∈ X, and F (b)(x) =Z b(z)dµx(z), ∀b ∈ B, x ∈ X. If there exists such a map µ, then it is unique, and F is faithful if and only if the interior of the set Zµ := {z ∈ Z : z /∈ supp(µπ(z))} is empty. Consequently, if supp(µx) = π−1(x), ∀x ∈ X, then F is faithful. Proof. Suppose first that there exists such a map µ. If a ∈ A, b ∈ B and x ∈ X: F (ab)(x) =Zπ−1(x) a(π(z))b(z)dµx(z) = a(x)Zπ−1(x) = (aF (b))(x). b(z)dµx(z) DILATIONS OF INTERACTION GROUPS 11 Then F (ab) = aF (b). A similar computation shows that F (a) = a, ∀a ∈ A, whence F is a conditional expectation. Conversely, suppose that F : B → A is a conditional expectation, and let µ : X → P (Z) be the map provided by Proposition 3.1 for the unital positive map F . Let us see that supp(µx) ⊆ π−1(x). Suppose z /∈ π−1(x). Then π(z) 6= x, so there exist open disjoint sets Vz and Vx in X such that π(z) ∈ Vz and x ∈ Vx. Let a ∈ A be such that a(X) = [0, 1], with supp(a) ⊆ Vx and a(x) = 1. Then, since a = F (a), a(x) = 1, and supp(a ◦ π) ⊆ π−1(Vx): 1 = F (a)(x) =ZZ a ◦ π dµx =Zπ−1(Vx) a ◦ π dµx ≤ µx(π−1(Vx)) ≤ 1. It follows that µx(Z \ π−1(Vx)) = 0. Then π−1(Vz) ∩ π−1(Vx) = ∅, hence µx(π−1(Vz)) = 0. Since π−1(Vz) is open we have that π−1(Vz) ∩ supp(µx) = ∅. This shows that z /∈ supp(µx) and therefore supp(µx) ⊆ π−1(x). Suppose now that there exists a non -- empty open subset V of Z such that z /∈ supp(µπ(z)), ∀z ∈ V . Let b ∈ B+ be a non -- zero element such that supp(b) ⊆ V . Then for all x ∈ X we have F (b)(x) =Rπ−1(x)∩V b(z)dµx(z) = 0 since π−1(x) ∩ V ∩ supp(µx) = ∅. Thus F is not faithful. Conversely, if F is not faithful, let 0 6= b ∈ B+ ∩ ker F , and V ⊆ supp(b) such that b(z) ≥ δ, for some positive δ and for all z ∈ V . Then, if z0 ∈ V and x = π(z0) we have b dµx ≥ δµx(π−1(x) ∩ supp(b)) ≥ δµx(V ∩ π−1(x)). 0 = F (b)(x) =Zπ−1(x) This shows that z0 /∈ supp(µπ(z)), ∀z0 ∈ V . (cid:3) Corollary 3.4. Let ξ : Z → Z be an onto continuous map and α : B → B its dual map. Let A = C(X) be the range of α and π : Z → X the canonical projection. Then a map L : B → B is a transfer operator for α if and only if there exists a w∗-continuous map ν : Z → P (Z) such that (3.2) L(b)(z) =Zξ−1(z) b(u)dνz(u) with supp(νz) ⊆ ξ−1(z), ∀z ∈ Z. In this case the map ν is unique. More precisely, if L is a transfer operator, then νz = µπ(z ′), where µ is the map associated by 3.3 to the conditional expectation αL, and z′ is any element of ξ−1(z). Proof. Suppose that L : B → B is a transfer operator for α. Then F := αL is a conditional expectation onto A. By Proposition 3.3 we have F (b)(x) = there exists z′ such that z = ξ(z′). Then, as F = αL, we get: L(b)(z) = Rπ−1(x) b(u)dµx(u), for a unique w∗-continuous map µ : X → P (Z). Conse- quently we have F (b)(z) =Rπ−1(π(z)) b(u)dµπ(z)(u). Since ξ is onto, for z ∈ Z L(b)(ξ(z′)) = F (b)(z′) = Rπ−1(π(z ′)) b(u)dµπ(z ′)(u) = Rξ−1(z) b(u)dµπ(z ′)(u), ∀b ∈ B, z′ ∈ ξ−1(z). So if ν : Z → P (Z) is the map associated to the unital positive map L on B by 3.1, we have, for z, z′ ∈ Z, with ξ(z′) = z: 12 FERNANDO ABADIE L(b)(z) = Rξ−1(z) b(u)dµπ(z ′)(u) = RZ b(u)dνz(u), and therefore νz = µπ(z ′), ∀z′ ∈ ξ−1(z). In particular, if ξ(z′) = z, then: supp(νz) = supp µπ(z ′) ⊆ π−1(π(z′)) = ξ−1(ξ(z′)) = ξ−1(z). Conversely, it is readily checked that a map L given by (3.2) for such a map ν is a transfer operator for α. (cid:3) As an immediate consequence of the above result we have Corollary 3.5. Let α be as in Corollary 3.4. Then the map L 7→ αL is a bijective correspondence between the sets of transfer operators for α and conditional expectations onto α(B). Moreover αL is faithful if and only if so is L. Proof. The last assertion, which is not implied by 3.4, follows from the injectivity of α. (cid:3) When the quotient map π : Z → X is a covering map we can be more precise: Corollary 3.6. Suppose that the quotient map π : Z → X is a covering map. Then a linear map F : B → A is an onto conditional expectation if and only if there exists a continuous map ω : Z → [0, 1] such that Pz∈π−1(x) ω(z) = 1, ∀x ∈ X, and F (b)(x) =Pz∈π−1(x) ω(z)b(z), ∀b ∈ B, x ∈ X. In this case the map ω is unique, and F is faithful if and only if the set ω−1(0) is nowhere dense. Proof. If x ∈ X, then π−1(x) is finite, because π is a local homeomorphism and Z is compact. Thus a map µ : X → P (Z) such that supp(µx) ⊆ π−1(x) is nothing but a map ω : Z → [0, 1] such that Pz∈π−1(x) ω(z) = 1 and µx = Pz∈π−1(x) ω(z)δz, ∀x ∈ X. Then, if ω is continuous, µ is w∗-continuous. Suppose conversely that µ is w∗-continuous. Fix z0 ∈ Z, and let V be an open neighborhood of z0 on which the restriction of π is a homeomorphism onto its image. Let b ∈ C(Z) be such that supp(b) ⊆ V , and b = 1 on a neighborhood U of z0. If z ∈ U : ω(z) = ω(z)b(z) = Xz ′∈π−1(π(z)) ω(z′)ZZ b dδz ′ =ZZ b dµπ(z) Since µ is w∗-continuous and π is continuous, it follows that ω also is con- tinuous because: lim z→z0 ω(z) = lim z→z0ZZ b dµπ(z) =ZZ b dµπ(z0) = ω(z0) Note finally that if z ∈ π−1(x), then z ∈ supp(µx) if and only if ω(z) 6= 0. The proof now follows by combining the considerations above with Propo- sition 3.3. (cid:3) If, in the situation of Corollary 3.6, the map ω exists and is positive, it follows from [15, Proposition 2.8.9] that the associated conditional expecta- tion F is of index-finite type (in the sense of Watatani, [15]), and moreover DILATIONS OF INTERACTION GROUPS 13 Index F (z) = 1/ω(z), ∀z ∈ Z. In particular F is faithful, a fact that also fol- lows from 3.3 and 3.5. Conversely, if F is of index-finite type, then Index F is positive ([15, Lemma 2.3.1]) and ω(z) = 1/Index F (z). Example 3.7. Consider the covering map θ : S1 → S1 given by θ(z) = z2, and let α : B → B be its dual map, where B = C(S1). Then A := α(B) = {b ∈ B : b(z) = b(−z), ∀z ∈ S1}. Consider any continuous function ω′ : [0, π] → [0, 1] such that ω′(π) = 1 − ω′(0), and let ω : [0, 2π] → [0, 1] be the extension of ω′ such that ω(t) = 1 − ω′(t − π), ∀t ∈ (π, 2π]. Since ω is continuous and ω(0) = ω(2π), we can look at ω as a continuous map by 3.6 ω defines a conditional expectation Fω. Consider the construction above for the following three cases: ω′ 3(t) = from S1 into [0, 1]. It is clear that Pz2=z0 ω(z) = 1, ∀z0 ∈ S1. Therefore (0 . Then Fω1 is of index-finite type, Fω2 is faithful but if t ∈ [0, π 2 ] if t ∈ [ π 2 , π] 1(t) = 1 2 , ω′ 2(t) = t π and ω′ 2t π − 1 not of index-finite type, and Fω3 is not faithful. Corollary 3.8. Let ξ : Z → Z be a covering map, and let α : B → B be its dual map, with range A = C(X). Then a linear map L : B → B is a transfer operator for α if and only if there exists a continuous map ω : Z → [0, 1] such that Pz∈π−1(x) ω(z) = 1, ∀x ∈ X, and (3.3) ω(z′)b(z′), ∀z ∈ Z. L(b)(z) = Xz ′∈ξ−1(z) In this case the map ω is unique. Proof. Since ξ is a covering map if and only if so is π, and ξ−1(z) = π−1(π(z)), ∀z ∈ Z, our claims follow from Corollary 3.6, Corollary 3.4, and their proofs. (cid:3) 3.2. Exel -- Renault interaction groups. Suppose again that P is an Ore semigroup with enveloping group G, so G = P −1P , and that θ : P × X → X is a right action, where X is a compact Hausdorff space, and each θt is a covering map. Then θ induces a left action α : P × A → A, where A = C(X) and αt(a) = a ◦ θt, ∀a ∈ A. Suppose in addition that V : G → B(A) is an interaction group that extends α, that is, Vt = αt, ∀t ∈ P . For each t ∈ P let Xt be the spectrum of the C ∗-subalgebra At := Vt(A), so At = C(Xt), and let πt : X → Xt be the corresponding canonical projection. Note that πt(x) = πt(x′) ⇐⇒ θt(x) = θt(x′) and πt is a local homeomorphism because θt is. Since Et := VtVt−1 : A → At is a conditional expectation, by Corollary 3.6 there exists a unique map ωt : X → [0, 1] such that Et(a)(x) = ωt(y)a(y), ∀a ∈ A, x ∈ X. Thus associated with the family of Pθt(y)=θt(x) conditional expectations {Et}t∈P there is a unique map ω : P × X → [0, 1] such that Et(a)(x) = Xθt(y)=θt(x) ω(t, y)a(y) 14 FERNANDO ABADIE ∀t ∈ P, a ∈ A, x ∈ X. This map ω is continuous and satisfies (3.4) ω(t, y) = 1, Xy∈θ−1 t (x) ∀t ∈ P , x ∈ X. The map ω also satisfies the cocycle property: (3.5) ω(rs, x) = ω(r, x)ω(s, θs(x)), ∀r, s ∈ P , x ∈ X, which reflects the fact that Vs−1r−1 = Vs−1Vr−1, ∀r, s ∈ P . Moreover, due to the commutativity of the conditional expectations Es and Er, ω also satifies the coherence property: (3.6) ω(s, x)Wr(C s,r x,y) = ω(r, x)Ws(C r,s x,y), C s,r x ∩ C r x = θ−1 y , with C s ∀r, s ∈ P , x, y ∈ X, where, for S ⊆ X, we put Wr(S) := Px∈S ω(r, x), and that Vr−1(a)(x) =Py∈θ−1 x,y = C s Since Vr−1 is a transfer operator for αr, r ∈ P , it follows by Corollary 3.8 r (x) ω(r, y)a(y), ∀a ∈ A, x ∈ X. Then, as Vs−1Vs = IdA, ∀s ∈ P , if t = r−1s ∈ G, r, s ∈ P , we have: Vt = Vr−1s = Vr−1sVs−1Vs = Vr−1VsVs−1Vs = Vr−1Vs = Vr−1αs. Therefore: s (θs(x)). (3.7) Vt(a)(x) = Vr−1αs(a)(x) = Xy∈θ−1 r (x) ω(r, y)a(θs(y)), ∀a ∈ A, x ∈ X. We recall from [10] the following definition: Definition 3.9. A continuous map ω : P × X → [0, 1], such that w(t, y) > 0 ∀(t, y) ∈ P × X will be called a normalized coherent cocycle or just cocycle for the action θ if it satisfies (3.4), (3.5) and (3.6). It is proved in [10, Theorem 2.8] that every normalized coherent cocycle associated with θ defines, by means of formula (3.7), an interaction group V ω that extends α. Such a V ω will be called an Exel -- Renault interaction group. Since each Et is an index-finite type conditional expectation, as observed after Corollary 3.6, we must have ω(t, x) = 1/Index Et(x), ∀t ∈ P, x ∈ X. Therefore, if r, s ∈ P and t = r−1s, from (3.7) we have: (3.8) Vt(a)(x) = Vr−1αs(a)(x) = Xy∈θ−1 r (x) a(θs(y)) Index Er(y) If V and V ′ are Exel-Renault interaction groups which extend α, for- mula (3.8) gives us a simple relation between them: in fact, if t = r−1s ∈ G, with r, s ∈ P , then with obvious notation we have (3.9) Vt(a) = Vr−1αs(a) = V ′ r−1(cid:16) IndexE′ r IndexEr αs(a)(cid:17), ∀a ∈ A. DILATIONS OF INTERACTION GROUPS 15 3.3. Dilations of Exel -- Renault interaction groups. In this final para- graph we will show that every Exel-Renault interaction group has a minimal faithful dilation. Let V = V ω : G → B(A) be an Exel-Renault interaction group extending α : P → B(A) as in the previous paragraph. Let (B, β, F ) be the minimal admissible dilation of V . From properties 1. and 2. of Theorem 2.2 we have that B is the direct limit of copies of A with connecting maps αr (alternatively see the proof of this result in [12]). Then B = C(Z), where (Z, {qr}r∈P ) is the inverse limit of the system ({Xr = X}r∈P , {θsr−1}e≤r≤s}. Concretely, we have Z = {z : P → X/ z(r) = θsr−1(z(s)), ∀r, s ∈ P, r ≤ s}, and qr : Z → X given by qr(z) = z(r). The dual (right) action β of β is described by the following formulae. For t ∈ P and z ∈ Z, to compute βtz(r) we choose s ∈ P such that s ≥ r, t. Then we have βtz(r) = θsr−1(z(st−1)) In other words, if t ∈ P , then qr βt = θsr−1qst−1, and β−1 t z(r) = z(rt). ∀s ≥ r, t, and qr β−1 In particular qe βt = θtqe. The inclusion of A into B is given by a 7→ aqe, ∀a ∈ A. Note that Zx := q−1 e (x) = {z ∈ Z : z(e) = x} is the inverse limit of the system ({Zx(r)}r∈P , {θsr−1}e≤r≤s}, where Zx(r) = θ−1 Lemma 3.10. Let ({Yr}r∈P , {σs spaces with inverse limit (Y, {pr}). Then the family V := {p−1 P, V ⊆ Yr open subset} is a basis for the topology of Y . r}r≤s) be an inverse limit of topological r (V ) : r ∈ t = qrt. r (x). y ∈ Tn j=1 p−1 rj ps(y) = prj (y), ∀j = 1, . . . , n. Since every σs Proof. Let V1, . . . , Vn be open subsets of Yr and r1, . . . , rn ∈ P . Suppose rj (Vj). Pick any element s ∈ P such that s ≥ rj, ∀j = 1, . . . , n. rj is continuous, there rj (V ) ⊆ Vj, ∀j = 1, . . . , n. rj (Vj), which shows that V is a basis for the (cid:3) Then σs exists an open neighborhood V of ps(y) such that σs Thus y ∈ p−1 topology of Y , since it is already a sub -- basis for it. j=1 p−1 s (V ) ⊆ Tn We have next the main result of this section. Theorem 3.11. Every Exel-Renault interaction group has a minimal faith- ful dilation, unique up to isomorphism. a unique map µ : X → P (Z) such that F (b)(x) = RZx Proof. We will use the above just introduced notation. Let (i, (B, β, F )) be the minimal admissible dilation of the Exel-Renault interaction group V : G → B(A), with i : A → B the natural inclusion. By 3.3 there exists b(z)dµx(z), ∀b ∈ B and x ∈ X, and to see that F is faithful is enough to show that the support of µx is exactly Zx. Observe that, since qe βr = θrqe, ∀r ∈ P , we have βr(Zy). Then if a ∈ A, x ∈ X, it Zx = q−1 r (z))dµx(z). Therefore e (Zx(r)) = Uy∈Zx(r) e (x) = βrq−1 aqe( β−1 (3.10) follows that Vr−1(a)(x) = F βr−1(a)(x) =RZx Vr−1(a)(x) = Xy∈Zx(r)Z βr(Zy) aqe( β−1 r (z))dµx(z) = Xy∈Zx(r) µx( βr(Zy))a(y) 16 FERNANDO ABADIE On the other hand, if y0 ∈ Zx(r), then V −1 Er(a)(y0). Thus (3.11) r (a)(x) = V −1 r (a)(θr(y0)) = V −1 r (a)(x) = Er(a)(y0) = Xy∈θ−1 r (θr(y0)) ω(r, y)a(y) = Xy∈Zx(r) ω(r, y)a(y) Comparing (3.10) and (3.11) we see that, by the uniqueness of ω(r, y), we must have µx( βr(Zy)) = ω(r, y) > 0. Since by Lemma 3.10 the family r (y) = βr(Zy) : r ∈ P, y ∈ Zx(r)} is a basis for the topology of Zx, we {q−1 conclude that the support of µx is Zx, and hence F is faithful. (cid:3) 2P{y:y2=x} a(y) is a transfer operator for α. Example 3.12. Consider the local homeomorphism θ : S1 → S1 given by θ(x) = x2, and let α : A → A be its dual map, where A = C(S1). Then L : A → A given by L(a)(x) = 1 It is easy to see that the cocycle ω : N × S1 → [0, 1] associated to the interaction group V induced by L is given by: ω(n, y) = 1 2n , ∀n ∈ N, y ∈ S1. Let (i, (B, β, F )) be the minimal dilation of V . Then B = C(Z), where the space Z is the solenoid: Z = {z : N → S1/ z(n) = z(n + 1)2, ∀n ∈ N}. The inclusion i : A → B is the dual map of q0 ( we use the notation of Theorem 3.11: thus q0 : Z → S1 is given by q0(z) = z(0)). The action β is the one determined by the shift β(b)(z)(n) = b(z(n + 1)). Thus β(z)(n) = z(n + 1), and q0 β = θq0. To find the corresponding conditional expectation F : B → A we need to describe the measures µx on Zx = q−1 0 (x) = {z ∈ Z : z(0) = x}. Note that, since every y ∈ S1 has exactly two roots, then for each n ∈ N the set Zx(n) (= qn(Zx)) has 2n elements, namely the roots of X 2n −x n (y) = βn(Zy), and therefore µx is determined in C. If y ∈ Zx(n), then q−1 by the fact that we must have µx(q−1 n (y)) = µx( βn(Zy)) = ω(n, y) = 1 2n . We will finish our work by explicitely computing the conditional expec- tation F on elements of a dense subalgebra of B. Consider the additive group M := {m : N → Z/ m(k) = 0 for all but a finite set of natural num- bers k}. Given m ∈ M let bm ∈ B be given by bm = 1 if m = 0 and m = b−m. Therefore the Stone-Weierstrass theorem implies that the selfadjoint subal- gebra B0 := span{bm : m ∈ M} of B is dense in B. For m ∈ M let ¯m be defined as ¯m = 0, if m = 0, and ¯m = max{k ∈ N : m(k) 6= 0}. Then k=0 m(k)/2k . Note that the previous formula also works for m = 0. Therefore, if m ∈ M bm(z) = Qk∈N z(k)m(k). Then bm ∈ B, bm1+m2 = bm1bm2 and b∗ k=0 z( ¯m)m(k)2 ¯m−k = z( ¯m)2 ¯m P ¯m it is clear that we have bm(z) = Q ¯m and x ∈ S1: F (bm)(x) = RZx and therefore we have (3.12) F (bm)(x) = 1 bm(z) dµx(z) = Py∈Zx( ¯m)R β ¯m(Zy) bm(z) dµx(z) 2 ¯m X{y: y2 ¯m =x} y2 ¯m P ¯m k=0 m(k)/2k Note that F (bm) ∈ C(S1), that is, the right hand side of equation (3.12) is a continuous function of x ∈ S1. DILATIONS OF INTERACTION GROUPS 17 The main portion of the present research was done during a visit of the au- thor to the Universidade Federal de Santa Catarina (Florian´opolis, Brazil). The author wishes to thank Ruy Exel and the people of the Departamento de Matem´atica there for their warm hospitatility. References [1] Fernando Abadie, Enveloping actions and Takai duality for partial actions, J. Funct. Anal. 197 (2003), 14 -- 67. [2] Fernando Abadie, On partial actions and groupoids, Proc. Amer. Math. Soc. 132 (2004), no. 4, 1037 -- 1047. [3] N. Brownlowe and Iain Raeburn, Exel's crossed products and relative Cuntz -- Pimsner algebras, Math. Proc. Cambridge Philos. Soc. 141 (2006), no. 3, 497 -- 508. [arXiv:math.OA/0408324]. [4] E. Blanchard, D´eformations de C ∗-alg`ebres de Hopf, Bull. Soc. Math. France 124 (1996), 141 -- 215. [5] Valentin Deaconu, C ∗-algebras of commuting endomorphisms, Advances in opera- tor algebras and mathematical physics, 47 -- 55, Theta Ser. Adv. Math., 5, Theta, Bucharest, 2005. [arXiv:math.OA/0406624]. [6] Ruy Exel, A new look at the crossed-product of a C ∗-algebra by an en- domorphism, Ergodic Theory Dynam. Systems 23 (2003), no. 6, 1733 -- 1750. [arXiv:math.OA/0012084]. [7] Ruy Exel, Crossed -- Products by Finite Index Endomorphisms and KMS states, J. Funct. Anal. 199 (2003), no. 1, 153 -- 188. [arXiv:math.OA/00105195]. no J. Funct. Anal. (2007), [8] R. Exel, Interactions, [arXiv:math.OA/0409267]. 244 1., 26 -- 67. [9] Ruy Exel, A new look at the crossed-product of a C ∗-algebra by a semigroup of endomorphisms, Ergodic Theory Dynam. Systems 28 (2008), no 3, 749 -- 789. [arXiv:math.OA/0511061]. [10] Ruy Exel, Jean Renault, Semigroups of local homeomorphisms and interac- tion groups, Ergodic Theory Dynam. Systems 27 (2007), no. 6, 1737 -- 1771. [arXiv:math.OA/0608589]. [11] R. Exel and A. Vershik, C ∗-algebras of Irreversible Dynamical Systems, Canad. J. Math. 58 (2006), no. 1, 39 -- 63. [arXiv:math.OA/0203185]. [12] Marcelo Laca, From endomorphisms to automorphisms and back: dilations and full corners, J. London Math. Soc. (2) 61 (2000), no. 3, 893 -- 904. [13] Nadia S. Larsen, Crossed products by abelian semigroups via transfer operators, preprint, [arXiv:math.OA/0502307]. [14] Danilo Royer, The crossed product by a partial endomorphism and the covariance algebra, J. Math. Anal. Appl. 323 (2006), 33 -- 41. [15] Yasuo Watatani, Index for C ∗-subalgebras, Mem. Amer. Math. Soc. 83 (1990), no. 424. Centro de Matem´atica-FC , Universidad de la Rep´ublica. 11 400 Igu´a 4225, Montevideo, URUGUAY. E-mail address: [email protected]
1604.03900	3	1604	2017-03-08T19:02:17	On finite free Fisher information for eigenvectors of a modular operator	[ "math.OA" ]	Suppose $M$ is a von Neumann algebra equipped with a faithful normal state $\varphi$ and generated by a finite set $G=G^*$, $\|G\|\geq 2$. We show that if $G$ consists of eigenvectors of the modular operator $\Delta_\varphi$ with finite free Fisher information, then the centralizer $M^\varphi$ is a $\mathrm{II}_1$ factor and $M$ is either a type $\mathrm{II}_1$ factor or a type $\mathrm{III}_\lambda$ factor, $0<\lambda\leq 1$, depending on the eigenvalues of $G$. Furthermore, $(M^\varphi)'\cap M=\mathbb{C}$, $M^\varphi$ does not have property $\Gamma$, and $M$ is full provided it is type $\mathrm{III}_\lambda$, $0<\lambda<1$.	math.OA	math	ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR BRENT NELSON Abstract. Suppose M is a von Neumann algebra equipped with a faithful normal state ϕ and generated by a finite set G = G∗, G ≥ 2. We show that if G consists of eigenvectors of the modular operator ∆ϕ with finite free Fisher information, then the centralizer M ϕ is a II1 factor and M is either a type II1 factor or a type IIIλ factor, 0 < λ ≤ 1, depending on the eigenvalues of G. Furthermore, (M ϕ)′ ∩ M = C, M ϕ does not have property Γ, and M is full provided it is type IIIλ, 0 < λ < 1. Introduction Given random variables x1, . . . , xn in a non-commutative probability space (M, ϕ), it is natural to ask what information about the distribution of a polynomial p ∈ Chx1, . . . , xni can be gleaned from the distributions of x1, . . . , xn. If p = x1 + x2 or p = x1x2 with x1 freely independent from x2, the theory of free additive and multiplicative convolutions tells us everything about the distribution of p, but (until recently) without the strict regularity condition of free independence little could be deduced about the distribution of a general polynomial. Shlyakhtenko and Skoufranis studied the distributions of matrices of polynomials in freely independent random variables x1, . . . , xn and their adjoints, and in particular showed that if x1, . . . , xn were semicircular random variables, then any self-adjoint polynomial has diffuse spectrum [24]. Mai, Speicher, and Weber later improved upon this result by showing that if x1, . . . , xn are self-adjoint random variables, not necessarily freely independent or having semicircular distributions but instead having finite free Fisher information, then x1, . . . , xn are algebraically free, any non-constant self-adjoint polynomial p ∈ Chx1, . . . , xni has diffuse spec- trum, and W ∗(x1, . . . , xn) contains no zero divisors for Chx1, . . . , xni [17]. Charlesworth and Shlyakhtenko further improved on this result by weakening the assumption of finite free Fisher information to having full free entropy dimension, and showed that under stronger assumptions on x1, . . . , xn one can assert that the spectral measure of p ∈ Chx1, . . . , xni is non-singular [3]. These techniques have since been applied by Hart- glass to show that certain elements in C∗-algebras associated to weighted graphs have diffuse spectrum [15]. In this paper, these techniques are brought to bear on non-tracial von Neumann algebras. We consider a von Neumann algebra M with a faithful normal state ϕ, and a finite generating set G. We will further assume that G has finite free Fisher information with respect to the state ϕ, and that each y ∈ G is an "eigenoperator"; that is, scaled by the modular automorphism group: σϕ y y for some λy > 0. Under these assumptions, we obtain a criterion for when polynomials ChGi in the centralizer M ϕ have diffuse spectrum (cf. Corollary 5.10). Our context is inspired by Shlyakhtenko's free Araki-Woods factors, which are non-tracial von Neumann algebras generated by generalized circular elements (operators scaled by the action of the modular automorphism group, cf. [21, Section 4]). t (y) = λit Regularity conditions on x1, . . . , xn can also have consequences on the von Neumann algebra generated by these operators. Indeed, Dabrowski [12] showed that if x1, . . . , xn in a tracial non-commutative probability space have finite free Fisher information, then these operators generate a factor without property Γ. The non-tracial analogue of this result, which considers the centralizer M ϕ as well as M , is the content of the two main results of this paper. The first is concerned with factoriality: Theorem A. Let M be a von Neumann algebra with a faithful normal state ϕ. Suppose M is generated by a finite set G = G∗, G ≥ 2, of eigenoperators of σϕ with finite free Fisher information. Then (M ϕ)′∩M = C. In particular, M ϕ is a II1 factor and if H < R×+ is the closed subgroup generated by the eigenvalues of G Research supported by NSF grants DMS-1502822 and DMS-0838680. 1 2 then BRENT NELSON M is a factor of type  III1 IIIλ II1 if H = R×+ if H = λZ, 0 < λ < 1 if H = {1}. Lacking property Γ is, for tracial von Neumann algebras, equivalent to the more general property of a von Neumann algebra being "full" (cf. Subsection 1.2). Consequently, the following theorem is the other half of the non-tracial analogue to Dabrowski's result: Theorem B. Let M be a von Neumann algebra with a faithful normal state ϕ. Suppose M generated by a finite set G = G∗, G ≥ 2, of eigenoperators of σϕ with finite free Fisher information. Then M ϕ does not have property Γ. Furthermore, if M is a type IIIλ factor, 0 < λ < 1, then M is full. The structure of the paper is as follows. In Section 1 we recall various notions relevant to the study of non-tracial von Neumann algebras. We also recall the definition of Dirichlet and completely Dirichlet forms. The context of our results is established in Section 2, wherein eigenoperators are defined and studied. In Section 3 we analyze derivations on the non-tracial von Neumann algebra M , conjugate variables, and free Fisher information. Of particular interest are "µ-modular" derivations, namely those derivations that interact nicely with the modular automorphism group σϕ. Section 4 is dedicated to the study of closable µ-modular derivations. In order to show that these closures are still derivations when restricted to the centralizer M ϕ, we study Dirichlet forms arising from µ-modular derivations. This is also used to establish a type of Kaplansky's density theorem for operators in the domain of a µ-modular derivation. Contraction resolvents associated to these derivations are also considered here. In Section 5 we produce a criterion for when polynomials ChGi in M ϕ are diffuse, and deduce when monomials in M ϕ have an atom at zero and of what size. Section 6 combines the analysis of the previous two sections to show that derivations associated to y ∈ G give rise to derivations (enjoying many of the same properties) that are related to the polar decomposition of y. Furthermore, the derivations associated to y for y ∈ G are in some sense tracial derivations, which we exploit in the proofs of our main theorems in Section 7. Theorem A is proven using a contraction resolvent argument similar to the one used in [12]. The type classifications of these von Neumann algebras are deduced using the well-known invariants recalled in Section 1. Theorem B is proven by using the derivations associated to y for y ∈ G to appeal to a tracial result of Curran, Dabrowski, and Shlyakhtenko from [11]. Acknowledgments. I would like to thank Dimitri Shlyakhtenko for the initial idea of this paper and his guidance in the early stages. I am indebted to Yoann Dabrowski, who greatly accelerated the progress of this paper with discussions we had while at Mathematisches Forschungsinstitut Oberwolfach, and suggested proofs that significantly improved Theorem B. Fabio Cipriani also helped immensely by guiding me through the literature on Dirichlet forms. Thanks to Ian Charlesworth, Michael Hartglass, and Benjamin Hayes who all provided an abundance of support and advice. Finally, I would also like to thank the anonymous referee for their helpful comments and suggestions. Throughout the paper, M will denote a von Neumann algebra with a faithful normal state ϕ. 1. Preliminaries 1.1. Arveson spectrum and Connes's S(M ) invariant. Suppose M is a factor. We recall some invariants for later use in establishing the type classification of M in Theorem A. The following exposition can be found in greater generality in [26, Chapter XI]. Identify R×+ as the dual group of R via the pairing so that the Fourier transform F on L1(R) is defined R ⊗ R×+ ∋ (t, λ) 7→ λit, (F f )(λ) =ZR λ−itf (t) dt, f ∈ L1(R). ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 3 Denote by A(R×+) := F (L1(R)), and for f ∈ A(R×+) let f denote its inverse image under the Fourier transform. When f is integrable this can be computed by the inverse Fourier transform: f (t) =ZR× + λitf (λ) dλ f ∈ A(R×+) ∩ L1(R×+). The following definitions are due to Arveson (cf. [1]), but we use the notation from [26, Chapter XI]. For all x ∈ M and f ∈ A(R×+), denote Define for each x ∈ M The σϕ-spectrum of x is defined σf (x) =ZR f (−t)σϕ t (x) dt. I(x) = {f ∈ A(R×+) : σϕ f (x) = 0}. and the Arveson spectrum of σϕ is defined Spσϕ (x) := {λ ∈ R×+ : f (λ) = 0, f ∈ I(x)}, Sp(σϕ) :=(λ ∈ R×+ : f (λ) = 0, f ∈ \x∈M I(x)) . Lemma 1.1 ( [26, Lemma XI.1.3.(v)]). For an open subset U ⊂ R×+ define M σϕ subspace of M spanned by 0 (U ) as the weak∗ closed Then λ ∈ R×+ belongs to Sp(σϕ) if and only if M σϕ 0 (U ) 6= {0} for every open neighborhood U of λ. f (x) : x ∈ M, f ∈ A(R×+) with supp(f ) ⊂ Uo . nσϕ Given a projection p ∈ M ϕ, the restriction of σϕ to pM p is denoted (σϕ)p. The Connes spectrum of σϕ [8, Definition 2.2.1] is defined Γ(σϕ) :=\p Sp((σϕ)p), where the intersection is over non-zero projections p ∈ M ϕ. Lemma 1.2 ( [8, Proposition 2.2.2.(c)]). If M is a factor with faithful normal state ϕ such that M ϕ is a factor, then Γ(σϕ) = Sp(σϕ). For a faithful semi-finite normal weight ψ on M , let Sψ denote the closure of the map M ∋ x 7→ x∗ when viewed as a densely defined operator on L2(M, ψ). The polar decomposition Sψ = Jψ∆1/2 yields an anti- ψ linear isometry Jψ and the modular operator ∆ψ, which is a positive non-singular (unbounded) operator densely defined on L2(M, ψ). Suppose M is a factor. Denoting the set of all faithful semi-finite normal weights on M by W0, the modular spectrum of M [8, Definition 3.1.1] is defined S(M ) := \ψ∈W0 spectrum(∆ψ). If S(M ) = {1}, then M is semi-finite. Otherwise M is a type III factor and S(M ) determines its type classification: M is of type  III0 IIIλ III1 if S(M ) = {0, 1} if S(M ) = {0} ∪ λZ for 0 < λ < 1 if S(M ) = [0, +∞). Lemma 1.3 ( [8, Theorem 3.2.1, Corollary 3.2.7]). If M is a factor with faithful normal state ϕ then S(M ) ∩ R×+ = Γ(σϕ), and if M ϕ is a factor then S(M ) = spectrum(∆ϕ). 4 BRENT NELSON 1.2. Full von Neumann algebras. Let Aut(M ) denote the group of automorphisms on M , and let Int(M ) denote the group of inner automorphisms (i.e. those automorphisms implemented via conjugation by a unitary u ∈ M ). On Aut(M ) we consider the topology of point-wise norm convergence in M∗: a net {αι} ⊂ Aut(M ) converges to α ∈ Aut(M ) if for every φ ∈ M∗ we have 0 = lim ι kφ ◦ (αι − α)kM∗ = lim ι sup x∈M φ(αι(x) − α(x)) kxk . Definition 1.4 ( [9, Definition 3.5]). A von Neumann algebra M with separable predual M∗ is full when Int(M ) is closed in Aut(M ) with respect to the above topology. Given φ ∈ M∗ and x ∈ M we define [x, φ] ∈ M∗ by [x, φ](y) := φ([y, x]). Definition 1.5. For φ ∈ M∗, an operator-norm-bounded sequence (zj)j∈N in M is said to be in the asymp- totic centralizer with respect to φ if k[zj, φ]kM∗ → 0. Theorem 3.1 of [9] tells us the following two conditions are equivalent for a von Neumann algebra M with separable predual: (i) M is full. (ii) Whenever a bounded sequence (zj)j∈N in M is in the asymptotic centralizer with respect to φ for all φ ∈ M∗, there exists a bounded sequence of scalars (cj)j∈N so that zj − cj → 0 ∗-strongly. For any faithful normal state ψ, we define a norm on M by kxk♯ ψ :=qkxk2 ψ + kx∗k2 ψ x ∈ M, and recall that for uniformly bounded sequences in M convergence with respect to this norm coincides with ∗-strong convergence. It is an easy exercise to see that the sequence of scalars in condition (ii) can be replaced with (ψ(zj))j∈N for any faithful normal state ψ on M . Moreover, zj can be replaced with zj − ψ(zj) so that we need only consider ψ-centered sequences. Finally, when M is a II1 factor this condition is equivalent to not having property Γ. 1.3. Entire elements and a bimodule structure for L2(M, ϕ). While L2(M, ϕ) is a left M -module, unlike in the tracial case, the right action of M on itself does not in general extend to a bounded action on L2(M, ϕ). However, there is a ∗-subalgebra of M containing the centralizer M ϕ for which the right action extends to a bounded action on L2(M, ϕ). Recall that a map from a complex domain valued in a Banach space is analytic if it can locally be expressed as a norm-convergent power series with coefficients in the Banach space. Such a map is entire if the domain is C. Definition 1.6 ( [26, Definition VIII.2.2]). An element x ∈ M is said to be entire if the M -valued map R ∋ t 7→ σϕ t (x) has an extension to an entire function, denoted by fx : C → M . The set of entire elements will be denoted M∞. Given x ∈ M∞ and a ∈ M , we compute kaxkϕ = kJϕaxkϕ = k∆1/2 ϕ x∗a∗kϕ ≤ kfx∗(−i/2)kk∆1/2 ϕ a∗kϕ = kfx∗(−i/2)kkakϕ. Thus this right action of M∞ extends to L2(M, ϕ) and L2(M, ϕ) is an M -M∞ bimodule. Remark 1.7. We note that for ξ ∈ L2(M, ϕ) and x ∈ M∞ the right action we are considering is equivalent to ξ · x = (Jϕσϕ i/2(x)∗Jϕ)ξ = Sϕx∗Sϕξ, whereas the usual right action of M on L2(M, ϕ) is given by Jϕx∗Jϕξ. We make this adjustment for the sake of the derivations; that is, so that when ξ ∈ M , the action of x is simply right-multiplication. In Section 5.1 we will make use of the usual right action. ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 5 1.4. Dirichlet forms. We refer the reader to [4, Section 4] and [5, Section 4] for further details. We begin with a discussion of standard forms of a von Neumann algebra. Dirichlet forms are defined in terms of a certain monotonicity condition on Hilbert spaces, and hence require a choice of positive cone. In our context, the Hilbert space will be L2(M, ϕ) and a standard form offers a convenient choice of positive cone. Since M ⊂ dom (Sϕ) = dom (∆1/2 s.a.(M, ϕ) denote the closed subspace of Jϕ-real vectors. We note that ϕ M+, then +(M, ϕ), Jϕ) is a standard form of M . A vector ξ ∈ L2(M, ϕ) is Jϕ-real if Jϕξ = ξ; we ϕ ), we have M ⊂ dom (∆1/4 (M, L2(M, ϕ), L2 let L2 ϕ ). Let L2 +(M, ϕ) = ∆1/4 We will consider complex-valued, sesquilinear forms (linear in the right entry) s.a.(M, ϕ) = ∆1/4 L2 ϕ Ms.a.. E : dom (E ) × dom (E ) → C, defined on a dense subspace dom (E ) ⊂ L2(M, ϕ), along with the associated quadratic form E [ξ] := E (ξ, ξ), ξ ∈ dom (E ). The quadratic form is Jϕ-real if Jϕdom (E ) ⊂ dom (E ) and E [Jϕξ] = E [ξ] for all ξ ∈ dom (E ). We will restrict our attention to the case when E [·] is valued in [0,∞). Such forms are closed if dom (E ) is a Hilbert space with the inner product hξ, ηiE := hξ, ηiϕ + E (ξ, η), and are closable if the identity map from dom (E ) to L2(M, ϕ) extends injectively to the Hilbert space completion of dom (E ) with respect to k · kE . Equivalently, E is closed if whenever {ξn}n∈N ⊂ dom (E ) is a Cauchy sequence with respect to k ·kE and ξn → ξ with respect to k ·kϕ, then ξ ∈ dom (E ) and ξn → ξ with respect to k · kE . E is closable if whenever {ξn}n∈N ⊂ dom (E ) is a Cauchy sequence with respect to k · kE and ξn → 0 with respect to k · kϕ, then limn E [ξn] = 0. In particular, if E [ξ] = kT ξk2 ϕ for some operator T with dom (T ) = dom (E ), then E is closed (resp. closable) if and only if T is closed (resp. closable). If one extends a closed form E to all of L2(M, ϕ) by letting E ≡ +∞ outside of dom (E ), then E is lower semicontinuous: whenever {ξn}n∈N ⊂ dom (E ) converges to ξ ∈ L2(M, ϕ) we have In this case, to show ξ ∈ dom (E ) (for the unextended form) it suffices to show E [ξ] < +∞. Consider the following closed, convex set: E [ξ] ≤ lim inf n→∞ E [ξn]. ϕ x : x ∈ Ms.a., x ≤ 1} s.a.(M, ϕ), we let ξ ∧ 1 denote the projection of ξ onto C. C = {∆1/4 For ξ ∈ L2 k·kϕ A quadratic form (E , dom (E )) is Markovian if for every Jϕ-real ξ ∈ dom (E ), ξ ∧ 1 ∈ dom (E ) with A Dirichlet form is a closed Markovian form. E [ξ ∧ 1] ≤ E [ξ]. For any n ∈ N, one has a canonical extension of E to L2(Mn(M ), 1 n ϕ ◦ Tr) in terms of the quadratic form E (n) defined on the subspace {[ξi,j ]n i,j=1 : ξi,j ∈ dom (E ), 1 ≤ i, j ≤ n} by E (n)[[ξi,j ]n i,j=1] = n Xi,j=1 E [ξi,j]. A form is completely Markovian (resp. completely Dirichlet ) if E (n) is Markovian (resp. Dirichlet) for every n ≥ 1. 2. Eigenoperators We will assume that M is reasonably well-behaved under the action of the the modular automorphism group σϕ; that is, M is generated by operators for whom the action of σϕ is merely multiplication by a scalar. Such operators will be known as "eigenoperators" of σϕ. We begin with some equivalent conditions for having such generators and the analytic implications of their existence. 6 BRENT NELSON 2.1. Regarding the generators. We begin with a proposition that, given a finitely generated von Neumann algebra M with faithful normal state ϕ and minimal assumptions on the generators, will allow us to freely switch to generators with more convenient behavior under the modular operator ∆ϕ. Proposition 2.1. Let M be a von Neumann algebra with a faithful normal state ϕ. Then the following are equivalent: k=1[A]jkxk for each j = 1, . . . , n where A ∈ Mn(C) is a positive definite matrix of the form A = diag(A1, . . . , Ak, 1, . . . , 1) where for each j = 1, . . . , k (i) M is generated by {e1, . . . , en} ⊂ Ms.a. such that ∆ϕej ∈ span{e1, . . . , en} for each j = 1, . . . , n. (ii) M is generated by {x1, . . . , xn} ⊂ Ms.a. such that ∆ϕxj =Pn −i(λj − λ−1 j ) λj + λ1 j Aj = 1 2(cid:18) λj + λ−1 j i(λj − λ−1 j ) (cid:19) for some λj ∈ (0, 1). Furthermore, the covariance of these generators is given by ϕ(xkxj) =(cid:20) 2 1 + A(cid:21)jk . (iii) M is generated by {c1, c∗1, . . . , ck, c∗k, z2k+1, . . . , zn} where ∆ϕcj = λjcj for some λj ∈ (0, 1), j = 1, . . . , k, and zj are self-adjoint elements satisfying ∆ϕzj = zj, j = 2k + 1, . . . , n. Moreover, elements of GL(n, C) linearly relate the three sets of generators in the following sense. If V = (v1, . . . , vn) and W = (w1, . . . , wn) are n-tuples formed by two sets of the above generators, then there is a Q ∈ GL(n, C) such that V = QW . Consequently the modular automorphism group {σϕ t }t∈R of ϕ can be extended to σϕ Che1, . . . , eni = Chx1, . . . , xni = Chc1, c∗1, . . . , ck, c∗k, z2k+1, . . . , zni . z , z ∈ C, on In particular, for any z ∈ C we have k=1[Aiz]jkxk ∀j ∈ {1, . . . , n}, j cj z (xj ) =Pn σϕ σϕ z (cj) = λiz σϕ z (zj) = zj Proof. We first note that (iii)⇒(i) is clear. (i)⇒(ii). Since ϕ is faithful, we assume without loss of generality that {e1, . . . , en} ⊂ L2(M, ϕ) is a linearly independent set. Furthermore, Re h·,·iϕ is a positive definite symmetric bilinear form and so by a Gram- Schmidt process we may assume hej, ekiϕ ∈ iR when j 6= k and kejkϕ = 1 while still maintaining ej = e∗j for each j = 1, . . . , n. Now, the modular operator ∆ϕ is a positive, non-singular operator on L2(M, ϕ). The condition on ∀j ∈ {1, . . . , k}, and ∀j ∈ {2k + 1, . . . , n}. {e1, . . . , en} implies that span{e1, . . . , en} is a ∆ϕ-invariant subspace. Define A ∈ Mn(C) by ∆ϕej = n Xk=1 [A]jkek ∀j = 1, . . . , n. We first claim that A = A∗. Let Λ ∈ Mn(C) be the covariance matrix: [Λ]jk := hek, ejiϕ for all j, k = 1, . . . , n. Note that Λ is positive definite since {e1, . . . , en} is a linearly independent set. Also, [ΛT Λ]jk = n Xl=1 [Λ]lj[Λ]lk = = = n n Xl=1 Xl=1 Xl=1 n hej, eliϕ hek, eliϕ hel, ejiϕhel, ekiϕ − hel, ejiϕ · − hel, ekiϕ = [ΛΛT ]jk, ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 7 ϕ = ∆−1/2 ϕ Jϕ, where Jϕ is an anti-linear isometry. which implies ΛT Λ−1 = Λ−1ΛT . Now, recall Sϕ = Jϕ∆1/2 We have for each j, k ∈ {1, . . . , n} [A]jl hek, eliϕ =ek, Xl=1 Xl=1 = hJϕ∆ϕej, Jϕekiϕ =D∆−1/2 = hej, ekiϕ = [ΛT ]jk, [AΛ]jk = ϕ n n [A]jlel+ϕ = hek, ∆ϕejiϕ Jϕ∆1/2 ϕ ej, JϕekEϕ = hSϕej, Sϕekiϕ so that A = ΛT Λ−1 = Λ−1ΛT . By a similar computation we have ΛA∗ = ΛT or A∗ = Λ−1ΛT = A. We also note that this computation also shows A = Λ−1/2ΛT Λ−1/2 AT = (Λ−1)T Λ = (Λ−1ΛT )−1 = A−1, hence A is positive definite with inverse A−1 = AT . Thus for each t ∈ R, Ut := Ait is a unitary matrix because A > 0 and an orthogonal matrix because t = (Ait)T = A−it = U−1 U T t . Hence, the entries of Ut are real and t 7→ Ut is an orthogonal representation of R on Rn. It follows that up to conjugating by some orthogonal matrix with real entries, ∀t ∈ R we have Ut = diag(R1(t), . . . , Rk(t), 1, . . . , 1) where for each j = 1, . . . , k for some λj ∈ (0, 1]. Using the formula Rj(t) =(cid:18) cos(t log λj) − sin(t log λj) cos(t log λj ) (cid:19) sin(t log λj ) i log(A)tv = lim t→0 1 t (Ut − 1)v ∀v ∈ Cn, we see that A has the desired form up to conjugation by an orthogonal matrix with real entries. Redefine A to be the desired conjugated form and let x1, . . . , xn ∈ Ms.a. be the image of {e1, . . . , en} under this orthogonal change of basis. Then clearly {x1, . . . , xn} generate M and we have ∆ϕxj = Pn k=1[A]jkxk for each j = 1, . . . , n. Redefine Λ to be the covariance matrix of {x1, . . . , xn}. Note that {x1, . . . , xn} is orthonormal with respect to Re h·,·iϕ since these elements are obtained from {e1, . . . , en} by an orthogonal change of basis. Hence so that 1 + ΛT Λ−1 = 2Λ−1. By the same computation as above we have that ΛT Λ−1 = A. It follows that (cid:2)Λ + ΛT(cid:3)jk = 2Re hxk, xjiϕ = [2]jk, Λ = 2 1 + A . (ii)⇒(iii). Define Q ∈ Mn(C) by Q = diag(Q1, . . . , Qk, 1, . . . , 1) where for each j = 1, . . . , n Qj = 1 √2(cid:18) 1 −i i (cid:19) . 1 Then it is easy to check that QAQ∗ = diag(λ1, λ−1 the new generators as the entries of the n-tuple QX. If we write 1 , . . . , λk, λ−1 k , 1, . . . , 1). Write X = (x1, . . . , xn) and define (c1, b1, . . . , ck, bk, z2k+1, . . . , zn) = QX, then because x1, . . . , xn are self-adjoint it is clear from the definition of Q that bj = c∗j , j = 1, . . . , k, and zj = z∗j , j = 2k + 1, . . . , n. Furthermore, n n n ∆ϕcj = ∆ϕ[Q]jkxk = [Q]jk[A]klxl = λj[Q]jlxl = λjcj, Xk=1 Xk,l=1 Xl=1 for each j = 1, . . . , k. A similar computation yields ∆ϕzj = zj for each j = 2k + 1, . . . , n. 8 BRENT NELSON It is clear from their constructions that the various sets of generators are linearly related by invertible matrices. Finally, the extension of the modular automorphism group is given by σϕ z (x) := ∆iz ϕ x∆−iz ϕ . The action of ∆ϕ on the vectors xj , cj, and zj then implies the claimed formulas. (cid:3) Remark 2.2. The relationship between the sets of generators in (ii) and (iii) is precisely the relationship between quasi-free semicircular random variables and generalized circular elements (cf. [21, Section 4] and [19, Section 3]). Remark 2.3. If k = 0 in either condition (ii) or (iii), then all of the generators are fixed points of ∆ϕ and hence ϕ is a trace. Remark 2.4. Suppose y ∈ M ⊂ L2(M, ϕ) is an eigenvector of ∆ϕ with eigenvalue λ > 0. Then y∗ is also an eigenvector with eigenvalue λ−1 because ∆ϕSϕ = Sϕ∆−1 ϕ . Therefore if y = y∗ then λ = 1. Hence if M is generated by a finite set G = G∗ of eigenvectors of ∆ϕ, then G is of the form in condition (iii). Definition 2.5. If y ∈ M ⊂ L2(M, ϕ) is an eigenvector of ∆ϕ with eigenvalue λ > 0, we say that y is an eigenoperator of the modular automorphism group σϕ = {σϕ t }t∈R with eigenvalue λ. Observe that every element of M ϕ is an eigenoperator with eigenvalue 1. In particular, if ϕ = τ is a trace, then every element of M is an eigenoperator. Proposition 2.6. Let M be a von Neumann algebra with faithful normal state ϕ. Suppose y ∈ M is a eigenoperator of σϕ with eigenvalue λ. If y = vy is the polar decomposition, then y ∈ M ϕ and v is an eigenoperator with eigenvalue λ. Proof. We note that Thus y∗y ∈ M ϕ, and consequently y = √y∗y ∈ M ϕ. It follows that y = λ−itσϕ of the polar decomposition we have σϕ t (v) = λitv. σϕ t (y∗y) = σϕ t (y)∗σϕ t (y) = λ−ity∗λity = y∗y. t (v)y, and so by uniqueness (cid:3) Henceforth M will be a von Neumann algebra with a faithful normal state ϕ, which is generated by a finite set G = G∗ of eigenoperators of σϕ. By Remark 2.4 we may use Proposition 2.1 to call on generators {x1, . . . , xn} of the form in (ii). We will often switch between these two generating sets depending on convenience, but will always denote generators of the form in (ii) by x1, . . . , xn whereas elements of G will generally be denoted by y. Denote P := ChGi = Chx1, . . . , xni . We also note that M has a separable predual since it can be faithfully represented on L2(M, ϕ) and is finitely generated. Remark 2.7. The formulas at the end of Proposition 2.1 imply P ⊂ M∞ with fp(z) = ∆iz for p ∈ P. Furthermore, using the density of P in L2(M, ϕ) and its invariance under ∆iz one can deduce from the proof of [26, Lemma VI.2.3] that fx(z) = ∆iz for every x ∈ M∞. 2.2. Implications of eigenoperators as generators. Observe that since σϕ −i is a homomorphism, any product of eigenoperators is again an eigenoperator. In particular, any monomial in ChGi is an eigenvector of ∆ϕ and so the modular operator ∆ϕ can be written z (p) ϕ for every z ∈ C, ϕ = σϕ ϕ x∆−iz ϕ p∆−iz ϕ ∆ϕ = Xλ∈R× + λπλ, on L2(M, ϕ); that is, ϕ is almost periodic (cf. [7]). Moreover, πλ converges strongly to the identity. For each λ ∈ R×+, let Eλ denote πλL2(M, ϕ), the eigenspace for pairwise orthogonal projections {πλ}λ∈R× Pλ∈R× of ∆ϕ corresponding to the eigenvalue λ. + + ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 9 For each λ ∈ R×+, define a map Eλ on B(L2(M, ϕ)) by Eλ(T ) = Xµ∈R× + πλµT πµ. Using the orthogonality of the projections πλ, it is easy to see that Eλ(T ) is bounded with kEλ(T )k ≤ kTk and that Eλ ◦ Eλ = Eλ. Moreover, since each πλ ∈ (M ϕ)′ ∩ B(L2(M, ϕ)), Eλ is left and right M ϕ-linear. Lemma 2.8. Let λ ∈ R×+. If p ∈ M is an eigenoperator of σϕ with eigenvalue µ ∈ R×+, then Eλ(p) = δλ=µp. Furthermore, Eλ(M ) ⊂ M . Proof. If ξ ∈ Eν, then clearly pξ ∈ Eµν and therefore πλν pξ = δλ=µpξ, establishing the first claim. Now, given x ∈ M , we can use Kaplansky's density theorem to find xn ∈ ChGi converging strongly to x and satisfying kxnk ≤ kxk; in particular, xn converges σ-strongly to x. We claim that Eλ(xn) converges strongly to Eλ(x). Indeed, we fix ξ ∈ L2(M, ϕ) and compute kEλ(x − xn)ξk2 ϕ = Xµ∈R× + kπλµ(x − xn)πµξk2 k(x − xn)πµξk2 ϕ. ϕ ≤ Xµ∈R× + This tends to zero by the σ-strong convergence of xn to x since kπµξk2 ϕ = kξk2 ϕ < ∞. Xµ∈R× + Since each xn is a sum of eigenoperators of σϕ, Eλ(xn) ∈ ChGi by the above discussion. Hence Eλ(x) ∈ M as the strong limit of polynomials. (cid:3) From this lemma we can deduce that Eλ(M ) consists of all eigenoperators in M with eigenvalue λ and that ϕ ◦ Eλ = δλ=1ϕ on M . We think of Eλ as a "conditional expectation" onto Eλ(M ), but recognize that Eλ(M ) is not an algebra for λ 6= 1. However, E1 = Eϕ, which is the actual conditional expectation onto the centralizer M ϕ. Corollary 2.9. (M ϕ)′ ∩ M = (M ϕ)′ ∩ M ϕ. Proof. Let x ∈ (M ϕ)′ ∩ M . Without loss of generality, we may assume x is self-adjoint. Since kx − Eϕ(x)k2 ϕ = Xλ6=1 kEλ(x)k2 ϕ, it suffices to show Eλ(x) = 0 for all λ 6= 1. Moreover, because Eλ(x)∗ = Eλ−1(x), we need only consider λ > 1. Fix such a λ, and note that Eλ(x) ∈ (M ϕ)′ ∩ M since Eλ is left and right M ϕ-linear. Let Eλ(x) = vEλ(x) be the polar decomposition, so that v is an eigenoperator with eigenvalue λ by Proposition 2.6. Observe that for a ∈ M ϕ, v∗avEλ(x) = v∗aEλ(x) = v∗Eλ(x)a = Eλ(x)a. Thus, if we define θ : M ϕ → M ϕ by θ(a) := v∗av, then ϕ(Eλ(x)a) = ϕ(θ(a)Eλ(x)) = ϕ(Eλ(x)θ(a)) = ··· = ϕ(Eλ(x)θk(a)) for any k ∈ N. Consequently, kEλ(z)k2 ϕ = ϕ(Eλ(x)Eλ(x)) = ϕ(cid:0)Eλ(x)θk(Eλ(x))(cid:1) = ϕ(cid:0)σϕ i (vk)Eλ(x)(v∗)kEλ(x)(cid:1) = λ−kϕ(vkEλ(x)(v∗)kEλ(x)) = λ−k(cid:10)vkEλ(x)(v∗)k,Eλ(x)(cid:11)ϕ ≤ λ−kkvkEλ(x)(v∗)kkkEλ(x)kϕ ≤ λ−kkEλ(x)kkEλ(x)kϕ. (cid:3) Letting k → ∞ we see kEλ(z)kϕ = 0. Thus x = Eϕ(x) ∈ M ϕ. For a subset I ⊂ R×+, let πλ πI :=Xλ∈I 10 and We define BRENT NELSON EI = span{Eλ : λ ∈ I} = πI L2(M, ϕ). M0 := {x ∈ M : x ∈ E[1/n,n] for some n ∈ N}, which is easily seen to be a ∗-algebra containing M ϕ and P, and hence is strongly dense in M . Clearly M0 ⊂ M∞. We will use the following lemma frequently to simplify the analysis of the modular operator. Lemma 2.10. Let D ⊂ M be an algebra generated by eigenoperators, and assume D is dense in L2(M, ϕ). Then for every z ∈ C, D is a core of ∆z ϕ. ϕ is closed, its graph G(∆z Proof. Since ∆z G(∆z ϕ) has the following orthogonal decomposition: ϕ) = G(∆z ϕ D)⊥. Thus, it suffices to show G(∆z ϕ D) ⊕ G(∆z ϕ) satisfies ϕ D)⊥ = 0. If ξ ∈ dom (∆z ϕx(cid:11)ϕ hξ, xiϕ +(cid:10)∆z ϕξ, ∆z then Dξ,(cid:16)1 + ∆2Re (z) with eigenvalue λ is simply scaled by 1 + λ2Re (z)), the density of D in L2(M, ϕ) implies ξ = 0. Remark 2.11. An easy consequence of this lemma when z = 1 ∀x ∈ D, = 0 for all x ∈ D. Noting that (cid:16)1 + ∆2Re (z) 4 is that Dk·k♯ (cid:17) D = D (since an eigenoperator ϕ ⊂ dom (∆1/4 (cid:17) xEϕ ϕ ). In fact, = 0, (cid:3) ϕ ϕ implies that ∆1/4 ϕ xn → ∆1/4 ϕ x so long as xn → x ∈ L2(M, ϕ) and {xn}n∈N is uniformly bounded. k∆1/4 ϕ xk2 ϕ =Dx, ∆1/2 ϕ xEϕ ≤ kxkϕkxk From this remark we can somewhat simplify our later analysis of Dirichlet forms. Recall that for ξ ∈ L2 s.a.(M, ϕ), ξ ∧ 1 is the projection of ξ onto the closed, convex set ϕ x : x ∈ Ms.a., x ≤ 1} C = {∆1/4 k·kϕ . ϕ x] ∧ 1 = ∆1/4 Lemma 2.12. For x ∈ M ϕ be self-adjoint, [∆1/4 Proof. Since x ∈ M ϕ, we have ∆1/4 By [2, Theorem 5.2], the projection x ∧ 1 is characterized by hx − x ∧ 1, x ∧ 1 − ξi ≥ 0 Note that by the functional calculus we have x − f (x) ∈ L2 ϕ x = x and ∆1/4 ∀ξ ∈ C. +(M, ϕ) and that ϕ f (x), where f (t) = min{t, 1} for t ∈ R. ϕ f (x) = f (x) in L2 s.a.(M, ϕ). Note that f (x) ∈ C. hx − f (x), f (x)iϕ = hx − f (x), 1iϕ . +(M, ϕ) and hence So for a ∈ Ms.a. with a ≤ 1, we have ∆1/4 ϕ (1 − a) ∈ L2 ϕ aEϕ Dx − f (x), f (x) − ∆1/4 Since such elements a are dense in C, this proves that f (x) satisfies (1). = hx − f (x), 1 − aiϕ ≥ 0. (1) (cid:3) 3. Non-tracial Differential Calculus In the tracial case, derivations on a von Neumann algebra have proven to be powerful tools in both Popa's deformation/rigidity theory and free probability. Provided the modular automorphism group interacts with the derivation in a nice way, one should expect similar results in the non-tracial case. Here we study, in particular, the notion of "µ-modularity," and give several examples of derivations exhibiting this behavior. We will also examine derivations previously considered in [18], which will be used to prove an L2-homology type result in Subsection 5.1. ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 11 3.1. Non-tracial derivatives, µ-modularity, conjugate variables, and free Fisher information. All of the derivations we consider will be valued in M ⊗ M op or L2(M ¯⊗M op, ϕ ⊗ ϕop), so we begin with some conventions on these spaces. First, we shall usually denote the latter space simply by L2(M ¯⊗M op), as the only GNS representation of M ¯⊗M op we will consider is the one with respect to ϕ ⊗ ϕop. For elements x◦ ∈ M op, we will also usually suppress the "◦" notation, and use the notation # for the natural multiplication on this space: On M ⊗ M op we consider three involutions: (a ⊗ b)#(c ⊗ d) = (ac) ⊗ (db). (a ⊗ b)∗ := a∗ ⊗ b∗ (a ⊗ b)† := b∗ ⊗ a∗ (a ⊗ b)⋄ := b ⊗ a. The first involution is used in the definition of h · ,· iHS: ha ⊗ b, c ⊗ diHS := ϕ ⊗ ϕop((a ⊗ b)∗#c ⊗ d). The second involution corresponds to the adjoint on HS(L2(M, ϕ)) when we identify it with L2(M ¯⊗M op) via the map where P1 ∈ B(L2(M, ϕ)) is the projection onto the cyclic vector 1 ∈ L2(M, ϕ). The third involution arises as the composition of the first two. Ψ(a ⊗ b) := aP1b, Note that x · (a ⊗ b) := (xa) ⊗ b (a ⊗ b) · y := a ⊗ (by) extend to operator-norm-bounded left and right actions of M on L2(M ¯⊗M op). Definition 3.1. Let B ⊂ M be a ∗-subalgebra. A derivation is a C-linear map δ : B → M ⊗ M op satisfying the Leibniz rule: We call B the domain of δ and write dom (δ) := B. The conjugate derivation to δ, denoted by δ, is a derivation with dom (δ) = dom (δ) defined by δ(x) = δ(x∗)† for x ∈ dom (δ). δ(ab) = δ(a) · b + a · δ(b). Given the required codomain, this definition is more restrictive than the usual definition of a derivation, but is sufficiently general for our present work. Note that if C ⊂ dom (δ), the Leibniz rule implies δ(z) = 0 for all z ∈ C. Most of the derivations we consider will, by virtue of the regularity conditions imposed on the generators of M , interact nicely with the modular automorphism group. We give this property the following name. Definition 3.2. Suppose B ⊂ M∞, and that δ : B → M∞ ⊗ M op is µ-modular if it satisfies ∞ is a derivation. For µ > 0, we say that δ δ ◦ σϕ z (x) = µiz(σϕ z ⊗ σϕ z ) ◦ δ(x) ∀z ∈ C, x ∈ B. (2) Remark 3.3. Note that if δ is µ-modular, then its conjugate derivation δ is µ−1-modular. We will assume initially that G is an algebraically free set: no non-trivial polynomial is zero. However, we will see that this assumption is in fact redundant in the context of our main results (cf. Remark 3.8). Note that the linear relation between G and {x1, . . . , xn} implies that x1, . . . , xn are algebraically free as well. For each y ∈ G, we define then extend δy to P by the Leibniz rule and linearity. Then δy : P → P ⊗P op is a derivation. If λy > 0 is the eigenvalue of y, then using the formulas at the end of Proposition 2.1 it is easy to see that δy is λy-modular. Also, the conjugate derivation δy is simply δy∗. δy(y′) = δy=y′1 ⊗ 1, When ϕ is a trace (and consequently y = y∗ for each y ∈ G), the derivations δy are examples of Voiculescu's free difference quotients (cf. [28]). When ϕ is not a trace, we still consider them to be free difference quotients, 12 BRENT NELSON but under the following slightly more general definition (the main difference being that the defining variable a need not be self-adjoint). Definition 3.4. Given a ∗-subalgebra B ⊂ M , let a ∈ M be algebraically free from B and denote by B[a] the ∗-algebra generated by B and a. If a is not self-adjoint further assume a∗ is algebraically free from a. The free difference quotient with respect to a is a derivation defined by δa : B[a] → B[a] ⊗ B[a]op and the Leibniz rule. If a is not self-adjoint, we also set δa(a∗) = 0. δa(a) = 1 ⊗ 1, δa(b) = 0 ∀b ∈ B, When a = a∗ we have δa = δa, and when a is not self-adjoint we have δa = δa∗, provided the latter derivation is well-defined. If B ⊂ M∞, then δa is µ-modular if a is an eigenoperator with eigenvalue µ. In the spirit of [29, Definitions 3.1 and 6.1], we make the following definitions. Definition 3.5. For a derivation δ, the conjugate variable to δ with respect to ϕ is defined to be a vector ξ ∈ L2(W ∗(dom (δ)), ϕ) such that hξ, xiϕ = h1 ⊗ 1, δ(x)iϕ⊗ϕop ∀x ∈ dom (δ), (3) provided such a vector exists. Note that the conjugate variable is uniquely determined by (3). Also, observe that δ∗(1 ⊗ 1) is the conjugate variable to δ provided 1 ⊗ 1 ∈ dom (δ∗) when δ : L2(W ∗(dom (δ)), ϕ) → L2(M ¯⊗M op) is viewed as a densely defined operator. For a free difference quotient δa, a ∈ M , with dom (δa) = B[a], we denote the conjugate variable by Jϕ(a : B). Definition 3.6. For a ∈ M algebraically free from a subalgebra B ⊂ M , the free Fisher information for a over B with respect to ϕ is defined as Φ∗ϕ(a : B) := kJϕ(a : B)k2 ϕ when Jϕ(a : B) exists, and as +∞ otherwise. For multiple elements a1, . . . , an ∈ M that are algebraically free over B, let B[ak : k 6= j] denote the ∗-algebra generated by B and the elements {ak}k6=j. Then the free Fisher information for a1, . . . , an over B is defined as n Φ∗ϕ(a1, . . . , an : B) := Φ∗ϕ(aj : B[ak : k 6= j]) Xj=1 when Jϕ(aj : B[ak : k 6= j]) exists for each j = 1, . . . , n, and as +∞ otherwise. When a1, . . . , an are just algebraically free (over C), then the free Fisher information for a1, . . . , an is defined Φ∗ϕ(a1, . . . , an) := Φ∗ϕ(a1, . . . , an : C). Remark 3.7. Notions of conjugate variables and free Fisher information for non-tracial von Neumann algebras were previously considered by Shlyakhtenko in [23]. We will see in Section 5.1 that our current definitions are strongly related (cf. Remark 5.1). Remark 3.8. When ϕ = τ is a trace, [17, Theorem 2.5] implies that finite free Fisher information for a1, . . . , an (or equivalently the existence of the conjugate variables to their free difference quotients) guaran- tees that a1, . . . , an are algebraically free. In fact, the proof of this theorem never invokes the trace condition on τ and therefore holds when ϕ is merely a faithful normal state, as in our context. Thus we will not require in our results that G be algebraically free, as this will be a consequence of Φ∗ϕ(G) < ∞. For each y ∈ G, we let ξy denote the conjugate variable to δy with respect to ϕ, provided it exists. One immediate consequence of the existence of conjugate variables is that the derivation is closable, as we shall see in the following lemma. In fact, when we assume that the free Fisher information is finite, it is usually to make use of this property. ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 13 Lemma 3.9. Let δ : dom (δ) → M∞ ⊗ M op ∞ be a derivation with dom (δ) ⊂ M∞ dense in L2(M, ϕ). If η ∈ dom (δ∗) when viewed as a densely defined map δ : L2(M, ϕ) → L2(M ¯⊗M op), then for a, b ∈ dom (δ) we have a · η, η · b ∈ dom (δ∗) with In particular, if the conjugate variable ξ to δ exists then dom (δ) ⊗ dom (δ)op ⊂ dom (δ∗) with δ∗(a · η) = a · δ∗(η) − (ϕ ⊗ σϕ δ∗(η · b) = δ∗(η) · σϕ −i ⊗ σϕ −i)hη#(σϕ −i(b) − (1 ⊗ ϕ)hη#(σϕ −i(b) − m(1 ⊗ ϕ ⊗ σϕ i )(cid:16)δ(a)⋄(cid:17)i i )(cid:16)δ(b)⋄(cid:17)i . −i)(1 ⊗ δ + δ ⊗ 1)(a ⊗ b) −i ⊗ σϕ δ∗(a ⊗ b) = a · ξ · σϕ (where m(a ⊗ b) = ab), and δ is closable. Proof. Let η ∈ dom (δ∗), b ∈ dom (δ), and x ∈ dom (δ). We compute hη · b, δ(x)iHS = hη, (1 ⊗ b∗)#δ(x)iHS = hη, δ(xb∗) − x · δ(b∗)iHS = hδ∗(η), xb∗iϕ −(cid:10)η#(σϕ =Dδ∗(η) · σϕ −i ⊗ σϕ −i(b) − (1 ⊗ ϕ)hη#(σϕ i )(δ(b∗)∗), x ⊗ 1(cid:11)HS −i ⊗ σϕ i )(δ(b)⋄)i , xEϕ . A similar computation yields the formula for δ∗(a · η). formulas we have for a, b ∈ dom (δ): Now, if the conjugate variable ξ to δ exists, recall that ξ = δ∗(1 ⊗ 1). So applying the previous two δ∗(a ⊗ b) = aδ∗(1 ⊗ b) − (σϕ −i ⊗ ϕ)h(σϕ −i(b) − (ϕ ⊗ 1)h(σϕ −i(b) − (ϕ ⊗ σϕ −i(b) − m(1 ⊗ ϕ ⊗ σϕ i ⊗ σϕ i ⊗ σϕ −i) ◦ δ(a)#b ⊗ 1i −i) ◦ δ(b)ii − (1 ⊗ ϕ)hδ(a)#σϕ −i)ha ⊗ 1#δ(b)i − (1 ⊗ ϕ)hδ(a)#σϕ −i(b) ⊗ 1i −i(b) ⊗ 1i = ahξ · σϕ = a · ξ · σϕ = a · ξ · σϕ −i)(1 ⊗ δ + δ ⊗ 1)(a ⊗ b). Thus δ is closable since dom (δ∗) contains the dense set dom (δ) ⊗ dom (δ)op. We conclude this subsection by noting that µ-modularity forces the conjugate variable to a derivation to be well-behaved under the modular operator. In particular, this will apply to ξy, y ∈ G, when they exist. Lemma 3.10. Let δ be a µ-modular derivation for some µ > 0, with dom (δ) a ∗-algebra generated by eigenoperators that is dense in L2(M, ϕ). If the conjugate variable ξ to δ exists, then ξ ∈ dom (∆z ϕ) for all z ∈ C with ∆z ϕξ = µzξ. (cid:3) (4) (5) (6) Furthermore, ξ ∈ dom (Sϕ) with Sϕξ = µ ξ, where ξ is the conjugate variable to δ, which exists if and only if ξ does. In particular, ∆ϕSϕξ = ξ. Proof. For x ∈ dom (δ) we compute (cid:10)ξ, ∆¯z =D1 ⊗ 1, δ ◦ σϕ =D1 ⊗ 1, µ¯z(σϕ = hµz1 ⊗ 1, δ(x)iHS = hµzξ, xiϕ . This computation suffices since dom (δ) is a core of ∆¯z ϕx(cid:11)ϕ ¯−iz(x)EHS ¯−iz ⊗ σϕ ϕ by Lemma 2.10. ¯−iz) ◦ δ(x)EHS 14 BRENT NELSON Since dom (Sϕ) = dom (∆1/2 dom (δ): ϕ ), the previous argument implies ξ ∈ dom (Sϕ). So we compute for x ∈ hSϕξ, xiϕ =DJϕ∆1/2 ϕ x, ∆ϕξEϕ = hx∗, µξiϕ = µhδ(x∗), 1 ⊗ 1iHS = µϕ ⊗ ϕop(δ(x)⋄) = µϕ ⊗ ϕop(δ(x)) =Dµ1 ⊗ 1, δ(x)EHS , which shows ξ exists and equals µ−1Sϕξ. Finally, (6) follows from combining (4) for z = 1 and (5). (cid:3) 3.2. Quasi-free difference quotients. Let A be the matrix from Proposition 2.1. For each j = 1, . . . , n we define ∂j(xk) =(cid:20) 2 1 + A(cid:21)kj 1 ⊗ 1, and then extend ∂j to P by the Leibniz rule and linearity. Then ∂j is a derivation with conjugate derivation ∂j determined by Furthermore, it follows from (cid:16) 2 If we let δj denote the free difference quotient with respect to xj, j = 1, . . . , n, then the {∂j}n are linearly related as follows: −i(p) = ∂j(p). = 2 1+A(cid:17)T (7) j=1 and {δj}n j=1 1 ⊗ 1. 2 1 + A(cid:21)jk ∂j(xk) =(cid:20) 1+A−1 that for p ∈ P (σϕ i ⊗ σϕ i ) ◦ ∂j ◦ σϕ ∂j = n Xk=1(cid:20) 2 1 + A(cid:21)kj δk and ∂j = n Xk=1(cid:20) 2 1 + A(cid:21)jk δk. Such derivations have been previously considered in [18], and we formally define them here. Definition 3.11. Suppose a1, . . . , an ∈ M∞ are self-adjoint and that there is an n × n matrix A > 0 which determines the covariance and action of the modular operator: 2 1 + A(cid:21)jk ϕ(akaj) =(cid:20) σϕ Xk=1 −i(aj) = [A]jkak. n ∂aj := n Xk=1(cid:20) 2 1 + A(cid:21)kj δak . If a1, . . . , an are algebraically free then the quasi-free difference quotients are defined on Cha1, . . . , ani as The conjugate variables to ∂aj with respect to ϕ are defined as in Definition 3.5 and denoted by J A j]), provided they exist. ϕ (aj : C[ak : k 6= The conjugate variables to ∂1, . . . , ∂n will be denoted ξ1, . . . , ξn, respectively. All quasi-free difference quotients satisfy the corresponding version of (7). One consequence of this is the following lemma, analogous to part of Lemma 3.10. Lemma 3.12. If ξ ∈ L2(M, ϕ) is the conjugate variable to a quasi-free difference quotient ∂, dom (∂) = P, then ξ ∈ dom (Sϕ) with Sϕξ = ξ. ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 15 Proof. Recall that Sϕ = Jϕ∆1/2 ϕ = ∆−1/2 ϕ D∆1/2 ϕ Jϕp, ξEϕ Jϕ. We compute for p ∈ P =D∆ϕJϕ∆1/2 ϕ p, ξEϕ =(cid:10)σϕ −i(p∗), ξ(cid:11)ϕ =(cid:10)∂ ◦ σϕ −i(p∗), 1 ⊗ 1(cid:11)HS =D(σϕ −i ⊗ σϕ =D ∂(p∗), 1 ⊗ 1EHS = ϕ ⊗ ϕop(∂(p)⋄) = ϕ ⊗ ϕop(∂(p)) = hξ, piϕ . −i) ◦ ∂(p∗), 1 ⊗ 1EHS ϕ , has domain dom (∆−1/2 ϕ (and hence of Jϕ∆−1/2 Now, the adjoint of Sϕ, Jϕ∆−1/2 ∆−1/2 formula. Remark 3.13. Recall from Proposition 2.1 that there exists Q ∈ GL(n, C) such that ). Recall from Lemma 2.10 that P is a core of ). Therefore the above computation shows ξ ∈ dom (Sϕ) with the claimed (cid:3) ϕ ϕ It follows that the free difference quotients are related by (Q−1)T : (c1, c∗1, . . . , ck, c∗k, z2k+1, . . . , zn) = Q · (x1, . . . , xn). (δc1 , δc∗ 1 , . . . , δck , δc∗ k , δz2k+1 , . . . , δzn ) = (Q−1)T · (δx1, . . . , δxn). j=1 exist if and only if the conjugate variables Moreover, the conjugate variables {Jϕ(xj : C[xk : k 6= j])}n {ξy}y∈G exist, and are related by (Q−1)∗: (ξc1 , ξc∗ 1 , . . . , ξck , ξc∗ k , ξz2k+1, . . . , ξzn ) = (Q−1)∗ · (Jϕ(x1 : C[xk : k 6= 1]), . . . , Jϕ(xn : C[xk : k 6= n])) . Consequently Φ∗ϕ(x1, . . . , xn) is finite if and only if Φ∗ϕ(G) is finite. Similarly, the linear relation between the quasi-free difference quotients ∂j and the free difference quotients δj implies {ξj}n j=1 exist if and only if {Jϕ(xj : C[xk : k 6= j])}n j=1 exist, and are related by ξj = n Xk=1(cid:20) 2 1 + A(cid:21)jk Jϕ(xk : C[xℓ : ℓ 6= k]). (8) 4. Closable µ-Modular Derivations In order to gain insights into the von Nuemann algebra M (as opposed to just the ∗-algebra P) we must necessarily consider µ-modular derivations δ which are closable, such as when the conjugate variable exists. By employing the theory of Dirichlet forms, we will see that the restriction of ¯δ to M ϕ satisfies the Leibniz rule. This result will allow us to establish a type of Kaplansky's density theorem (cf. Theorem 4.8), as well as some bounds for ¯δ (when the conjugate variable exists) that imply the domain of δ∗ is in fact quite large. 4.1. Some preliminary observations. Any closable operator is assumed to have dense domain. In order to simplify the exposition, for any (unbounded) closed operator T : H1 → H2, we denote kξkT =qkξk2 H1 + kT ξk2 H2 ξ ∈ dom (T ). We first observe that µ-modularity for a closable derivation has implications for how the closure restricts to the eigenspaces of ∆ϕ. Lemma 4.1. Let δ : L2(M, ϕ) → L2(M ¯⊗M op) be a closable µ-modular derivation for some µ > 0, with dom (δ) generated by eigenoperators of σϕ. If we denote the closure by ¯δ, then for I, J ⊂ R×+ disjoint subsets (i) πI dom (¯δ) ⊂ dom (¯δ); (ii) dom (δ) ∩ EI is a core of ¯δ πI dom (¯δ); and (iii) ¯δ(πI dom (¯δ)) ⊥ ¯δ(πJ dom (¯δ)). Furthermore, if the conjugate derivation δ is closable with closure ¯δ ◦ Sϕ ◦ πI (·) = [¯δ ◦ πI (·)]† for any subset I ⊂ R×+ bounded above. ¯δ, then Sϕ(πI dom (¯δ)) ⊂ dom ( ¯δ) with 16 BRENT NELSON Proof. Since dom (δ) is generated by eigenoperators, πI dom (δ) = dom (δ) ∩ EI for any I ⊂ R×+. The µ- modularity implies that if p ∈ dom (δ)∩EI and q ∈ dom (δ)∩EJ for disjoint sets I and J, then hδ(p), δ(q)iHS = 0. Now, given ξ ∈ dom (¯δ), let {pn}n∈N ⊂ dom (δ) approximate ξ in the k · k¯δ-norm. Then kξ − pnk2 ϕ = kπI (ξ − pn)k2 ϕ + k(1 − πI )(ξ − pn)k2 ϕ implies πI pn converges to πI ξ in L2(M, ϕ). Also, since 1 − πI = πJ for J = R×+ \ I, µ-modularity implies kδ(pn − pm)k2 ϕ = kδ(πI (pn − pm))k2 ϕ + kδ((1 − πI )(pn − pm))k2 ϕ. Finally, let I ⊂ R×+ be bounded above by λ > 0. Then clearly EI ⊂ dom (∆1/2 So δ(πI pn) is a Cauchy sequence and must converges to some η ∈ L2(M ¯⊗M op); that is, πI ξ ∈ dom (¯δ). This establishes (i), and our proof establishes (ii) and (iii). ϕ ) = dom (Sϕ). Given ξ ∈ πI dom (¯δ), let {pn}n∈N ∈ πI dom (δ) approximate ξ in the k · k¯δ-norm. Since ∆ϕ is bounded on EI (and hence so is Sϕ = Jϕ∆1/2 ϕ ), Sϕpn converges to Sϕξ. Since † is an isometry we have that δ(p∗n) = δ(pn)† converges to ¯δ(ξ)†. (cid:3) From the above lemma, we obtain an immediate (albeit partial) extension of the Leibniz rule, which we shall use later to obtain a more robust result. Recall that Mϕ acts boundedly on L2(M, ϕ) as SϕM∞Sϕ. Lemma 4.2. Let δ : L2(M, ϕ) → L2(M ¯⊗M op) be a closable µ-modular derivation, for some µ > 0, with dom (δ) generated by eigenoperators of σϕ. Then ¯δ is defined on the products dom (¯δ) · dom (δ) and dom (δ) · (cid:0)πI dom (¯δ)(cid:1), I ⊂ R×+ bounded above, and satisfies the Leibniz rule on these products. Moreover, ¯δ is defined on ker(δ) · dom (¯δ) · ker(δ) and satisfies the Leibniz rule on this product. Proof. For ξ ∈ dom (¯δ), let {xn}n∈N ⊂ dom (δ) approximate ξ in the k · k¯δ-norm. Then for p ∈ dom (δ) we have kξ · p − xnpkϕ ≤ kσϕ i/2(p)kkξ − xnkϕ → 0. Similarly, say δ(p) =Pj aj ⊗ bj ∈ M∞ ⊗ M op ∞ (a finite sum) then k[(ξ − xn) · aj] ⊗ bjkHS ≤Xj k(ξ − xn) · δ(p)kHS ≤Xj k¯δ(ξ) · p + ξ · δ(p) − δ(xnp)kHS = k¯δ(ξ) · p + ξ · δ(p) − δ(xn) · p − xn · δ(p)kHS → 0, i/2(aj )kkξ − xnkϕkb∗jkϕ → 0. kσϕ Therefore it is clear that so that ξ · p ∈ dom (¯δ) with ¯δ(ξ · p) = ¯δ(ξ) · p + ξ · δ(p). For ξ ∈ πI dom (¯δ), I ⊂ R×+ bounded above, the proof of the final assertion in Lemma 4.1 shows that if {xn}n∈N ⊂ dom (δ) ∩ EI approximates ξ in the k · k¯δ-norm, then in fact kξ − xnk# ϕ → 0. Thus if p ∈ dom (δ) and δ(p) =Pj aj ⊗ bj, then kajkϕkσϕ −i/2(bj)kkSϕ(ξ − xn)kϕ → 0. kδ(p) · (ξ − xn)kHS ≤Xj kajkϕkSϕbj · (ξ − xn)kϕ ≤Xj We of course also have kp · ξ − pxnkϕ → 0, and kp · ¯δ(ξ) − p · δ(xn)kHS → 0, so that p · ξ ∈ dom (¯δ) with ¯δ(p · ξ) = δ(p) · ξ + p · ¯δ(ξ). observe that pxnq k · k¯δ-approximates p · ξ · q since δ(pxnq) = p · δ(xn) · q. Finally, for p, q ∈ ker(δ) ⊂ dom (δ) and ξ ∈ dom (¯δ) k · k¯δ-approximated by {xn}n∈N ⊂ dom (δ), we easily (cid:3) We will also frequently make use of the following lemma, which is a variation of [6, Lemma 7.2]. We first establish some notation. Let a ∈ M be self-adjoint with spectrum contained in a compact interval I ⊂ R. We consider a representation LRa of C(I) ⊗ C(I) = C(I × I) in C∗(a) ⊗ C∗(a)op defined by LRa(f ⊗ g) = f (a) ⊗ g(a) for f, g ∈ C(I). In particular, if h ∈ C(I × I) factors as h(s, t) = f (s)g(t), then LRa(h) = f (a) ⊗ g(a). For f ∈ C1(I), denote f (s, t) :=(cid:26) f′(t) f (t)−f (s) t−s if t = s otherwise ∈ C(I × I). ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 17 Lemma 4.3. Let δ : L2(M, ϕ) → L2(M ¯⊗M op) be a closable derivation with closure ¯δ and dom (δ) a unital ∗-algebra. If a = a∗ ∈ dom (δ) has spectrum contained in a compact interval I ⊂ R, then for every f ∈ C1(I) we have f (a) ∈ dom (¯δ) with ¯δ(f (a)) = LRa( f )#δ(a). Moreover, if g ∈ C(I) is Lipschitz with constant C, then g(a) ∈ dom (¯δ) with k¯δ(g(a))kHS ≤ Ckδ(a)kHS. Proof. The proof of the first part is identical to that in [6, Lemma 7.2], but we note that δ(1) = 0 allows us to consider f ∈ C1(I) with f (0) 6= 0. For Lipschitz functions, approximate by functions in C1(I). (cid:3) 4.2. Dirichlet forms arising from non-tracial derivations. Fix a µ-modular derivation δ with dom (δ) a ∗-algebra generated by eigenoperators of σϕ. We consider the following quadratic form on L2(M ϕ, ϕ): E[ · ] := kδ( · )k2 HS + kδ( · )k2 HS, with dom (E) = M ϕ ∩ dom (δ). If δ = δ, E is defined by the above formula divided by a factor of two. While (M ϕ, ϕ) is a tracial von Neumann algebra, the results of this section are not entirely subsumed by the tracial case since δ may be valued outside of M ϕ ⊗ (M ϕ)op. for p ∈ dom (E ) we have We claim that E is Jϕ-real. Indeed, recall that Jϕ M ϕ = Sϕ M ϕ and that † is an isometry for k · kHS. So, E [Jϕp] = kδ(p∗)k2 = kδ(p)†k2 HS + kδ(p∗)k2 HS + kδ(p)†k2 HS HS = E [p]. We also note that if the conjugate variables to either δ or δ exist (and hence both exist by Lemma 3.10), then clearly E is closable, say with closure ¯E . One has that ξ ∈ dom ( ¯E ) if and only if ξ ∈ L2(M ϕ, ϕ) ∩ ¯δ) and there exists a sequence (pn)n∈N ⊂ M ϕ ∩ dom (δ) such that pn → ξ in L2(M ϕ, ϕ) and dom (¯δ) ∩ dom ( simultaneously δ(pn) → ¯δ(ξ) and δ(pn) → Proposition 4.4. Let δ : dom (δ) → M∞ ⊗ M op ∞ be a µ-modular derivation for some µ > 0, with dom (δ) generated by eigenoperators of σϕ. Assume the conjugate variable to δ exists so that E is closable with closure ¯E . Then ¯E is a completely Dirichlet form on L2(M ϕ, ϕ). Proof. Let η = Jϕη ∈ dom ( ¯E ), and let (pn)n∈N ⊂ dom (E ) be a sequence converging to η with respect to k·k ¯E . By replacing pn with pn + Jϕpn ∈ dom (E ), we may assume each pn is Jϕ-real. Since Jϕ M ϕ = S M ϕ , each pn is self-adjoint, and so by Lemma 2.12 pn ∧ 1 = f (pn) where f (t) = min{t, 1}. We first claim pn ∧ 1 ∈ dom ( ¯E ) with ¯δ(ξ) in L2(M ¯⊗M op). ¯E [pn ∧ 1] ≤ E [pn]. Indeed, let I ⊂ R be an interval containing 1 and the spectrum of pn. Then there exists a sequence {gk}k∈N of polynomials with real coefficients such that gk approximate f uniformly on I, and g′k are uniformly bounded by say 1 + 1 k on I. Then {gk(pn)}k∈N is a sequence of self-adjoint operators in dom (E ) that by Lemma 4.3 satisfy E [gk(pn)] ≤(cid:18)1 + 1 k(cid:19)2 E [pn]. Moreover, gk(pn) → f (pn) in L2(M ϕ, ϕ). Extending ¯E to all of L2(M ϕ, ϕ) by letting ¯E ≡ +∞ outside of dom ( ¯E ), we have ¯E [pn ∧ 1] = ¯E [f (pn)] ≤ lim inf k→∞ E [gk(pn)] ≤ E [pn] from the lower semicontinuity guaranteed by ¯E being closed. In particular, ¯E [pn ∧ 1] < +∞ so that pn ∧ 1 ∈ dom ( ¯E ). Now, since (·) ∧ 1 is a projection onto a closed convex set, we know (pn ∧ 1)n∈N converges to η ∧ 1 with respect to k · kϕ. Thus, using lower semicontinuity again we have ¯E [pn ∧ 1] ≤ lim inf n→∞ ¯E [η ∧ 1] ≤ lim inf n→∞ Thus, ¯E is Markovian and hence Dirichlet. E [pn] = ¯E [η]. Given n ∈ N, we note that the canonical extension ¯E (n) is defined for T ∈ Mn(dom ( ¯E )) by ¯E (n) [T ] =(cid:13)(cid:13)(cid:0)¯δ ⊗ In(cid:1) (T )(cid:13)(cid:13) 2 1 n (ϕ⊗ϕop)◦Tr +(cid:13)(cid:13)(cid:13)(cid:16)¯δ ⊗ In(cid:17) (T )(cid:13)(cid:13)(cid:13) 2 . 1 n (ϕ⊗ϕop)◦Tr 18 BRENT NELSON If ξ is the conjugate variable to δ, let ξ denote the conjugate variable to δ, which exists by Lemma 3.10. It is easy to see that δ ⊗ In and δ ⊗ In are derivations on Mn(dom (E )) with conjugate variables ξ ⊗ In and ξ ⊗ In, respectively. Consequently, by the same argument preceding the proposition, we see that ¯E (n) is closed. The argument showing that ¯E is a Dirichlet form relied only on the functional calculus, hence we repeat the argument to see that ¯E (n) is also a Dirichlet form. Thus ¯E is completely Dirichlet. (cid:3) Using the proof of [5, Proposition 4.7], we obtain the following as an immediate corollary. Corollary 4.5. Let δ : dom (δ) → M∞ ⊗ M op ∞ be a µ-modular derivation for some µ > 0, with dom (δ) generated by eigenoperators of σϕ. Assume the conjugate variable to δ exists so that E is closable with closure ¯E . Then the set M ϕ ∩ dom ( ¯E ) is a ∗-algebra. For the remainder of this section, we will assume that the conjugate variable to δ exists, so that the ¯δ, respectively. We ¯δ) that can be approximated by elements conjugate variable to δ also exists and both δ and δ are closable with closures ¯δ and define dom (δ⊕ δ) to be the subspace of elements in dom (¯δ)∩ dom ( of dom (δ) simultaneously in the k · k¯δ and k · k¯δ norms. When δ = δ, this set is simply dom (¯δ). Proposition 4.6. Let δ : dom (δ) → M∞ ⊗ M op ∞ be a µ-modular derivation for some µ > 0, with dom (δ) generated by eigenoperators of σϕ. Assume the conjugate variable to δ exists. Then the set M ϕ ∩ dom (δ ⊕ δ) is a ∗-algebra on which both ¯δ and Proof. Since M ϕ ∩ dom (δ ⊕ δ) = M ϕ ∩ dom ( ¯E ) by the remarks preceding Proposition 4.4, we see this set is a ∗-algebra by Corollary 4.5. ¯δ being similar). Let (pn)n∈N ⊂ M ϕ ∩ dom (δ) approximate a in the k·k¯δ-norm. Since L2(M, ϕ) and L2(M ¯⊗M op) admit bounded right actions of M∞ we have Let a, b ∈ M ϕ ∩ dom (δ ⊕ δ). We will show ¯δ(ab) = ¯δ(a) · b + a · ¯δ(b) (the proof for ¯δ satisfy the Leibniz rule. Also, since kp∗n−a∗kϕ = kpn−akϕ → 0, pn·¯δ(b) converges to a·¯δ(b) weakly against M∞⊗M op ∞ ⊂ L2(M ¯⊗M op). Thus δ(pn)·b+pn·¯δ(b) converges weakly to ¯δ(a)·b+a·¯δ(b) against M∞⊗M op ∞ . On the other hand, from Lemma 4.2 we know pnb ∈ dom (¯δ) with ¯δ(pnb) = δ(pn)· b + pn · ¯δ(b). Consequently, for any η ∈ dom (δ) ⊗ dom (δ)op (which also lies in dom (δ∗) by Lemma 3.9) we have kpnb − abkϕ → 0 and kδ(pn) · b − ¯δ(a) · bkHS → 0. (cid:3) (cid:10)¯δ(a) · b + a · ¯δ(b), η(cid:11)HS = lim n→∞(cid:10)δ(pn) · b + pn · ¯δ(b), η(cid:11)HS n→∞(cid:10)¯δ(pnb), η(cid:11)HS = lim = hab, δ∗(η)iϕ =(cid:10)¯δ(ab), η(cid:11)HS . Since dom (δ) ⊗ dom (δ)op is dense, we have ¯δ(ab) = ¯δ(a) · b + a · ¯δ(b). = lim n→∞hpnb, δ∗(η)iϕ One particular consequence of this proposition is that for any x ∈ M ϕ ∩ dom (δ ⊕ δ), we have p(x) ∈ M ϕ ∩ dom (δ ⊕ δ) for any polynomial p. By the same argument as in Lemma 4.3 we obtain the following corollary. Corollary 4.7. Let δ : dom (δ) → M∞ ⊗ M op ∞ be a µ-modular derivation for some µ > 0, with dom (δ) generated by eigenoperators of σϕ. Assume the conjugate variable to δ exists. Let a = a∗ ∈ M ϕ ∩ dom (δ ⊕ δ) with spectrum contained in a compact interval I ⊂ R. Then for any f ∈ C1(I), f (a) ∈ M ϕ ∩ dom (δ ⊕ δ) ¯δ). Moreover, if g ∈ C(I) is Lipschitz with constant C, then with ¯δ(f (a)) = LRa( f )#¯δ(a) (and similarly for g(a) ∈ M ϕ ∩ dom (δ ⊕ δ) with k¯δ(g(a))kHS ≤ Ck¯δ(a)kHS (and similarly for ¯δ). We conclude with the following analogue of [12, Proposition 6], which we obtain via the same proof. This result is a version of Kaplansky's density theorem for elements in the domain of a closed derivation. Theorem 4.8. Let δ : dom (δ) → M∞ ⊗ M op ∞ be a µ-modular derivation for some µ > 0, with dom (δ) generated by eigenoperators of σϕ. Assume the conjugate variable to δ exists so that δ and δ are closable ¯δ, respectively. For any x ∈ M ϕ ∩ dom (δ ⊕ δ), λ > 0, there exists a sequence (pn)n∈N ⊂ with closures ¯δ and M ϕ ∩ dom (δ) converging ∗-strongly to x and approximating x simultaneously in the k · k¯δ and k · k¯δ norms such that kpnk ≤ kxk for all n ∈ N. ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 19 4.3. Norm boundedness of δ∗. By exhibiting some boundedness conditions for a derivation δ, we will be able to see that the domain of δ∗ is in fact quite large. The following proposition is the non-tracial analogue of [12, Lemma 12]. Proposition 4.9. Let δ : dom (δ) → M∞ ⊗ M op ∞ be a µ-modular derivation for some µ > 0, with dom (δ) generated by eigenoperators of σϕ. Assume the conjugate variable ξ to δ exists so that δ and δ are closable with closures ¯δ and ¯δ, respectively. Then for u ∈ M ϕ ∩ dom (δ ⊕ δ) unitary ¯δ(u))kϕ = kξkϕ, and ¯δ(u))kϕ = kξkϕ. ku · ξ − (1 ⊗ ϕ)( −i)( kξ · u − (ϕ ⊗ σϕ For a ∈ M ϕ ∩ dom (δ ⊕ δ) self-adjoint we have ka · ξ − (1 ⊗ ϕ)( −i)( kξ · a − (ϕ ⊗ σϕ ¯δ(a))kϕ ≤ kakkξkϕ, and ¯δ(a))kϕ ≤ kakkξkϕ. For any x ∈ M ϕ ∩ dom (δ ⊕ δ) we have kx · ξ − (1 ⊗ ϕ)( kξ · x − (ϕ ⊗ σϕ −i)( Consequently, for x ∈ M ϕ ∩ dom (δ ⊕ δ) we have k(1 ⊗ ϕ)( k(ϕ ⊗ σϕ −i( ¯δ(x))kϕ ≤ 2kxkkξkϕ, and ¯δ(x))kϕ ≤ 2kxkkξkϕ. ¯δ(x)kϕ ≤ 3kxkkξkϕ, and ¯δ(x))kϕ ≤ 3kxkkξkϕ. (9) Moreover, for any λ ∈ R×+ and any p ∈ Eλ(dom (δ)) we have k(1 ⊗ ϕ)( k(ϕ ⊗ σϕ −i( ¯δ(p)kϕ ≤ 3kpkkξkϕ, and ¯δ(p))kϕ ≤ 3λ1/2kpkkξkϕ. kb(p)k2 −i/2(x∗)kkwkϕ), which we now prove. Proof. It is clear that the estimates in (9) follow immediately from the previous ones (along with the well known inequality kwxkϕ ≤ kσϕ For notational simplicity, we write b(x) := (1 ⊗ ϕ)h¯δ(x)i for x ∈ M ϕ ∩ dom (δ ⊕ δ). We first establish the inequalities which involve b. We let ξ denote the conjugate variable to δ, which exists by Lemma 3.10. First, we consider p ∈ dom (δ). Using Lemma 3.9 we have ϕ =Db(p) ⊗ 1, δ(p)EHS =Db(p) ξ − (1 ⊗ ϕ)(δ(b(p))), pEϕ =Db(p) ξ − (1 ⊗ ϕ ⊗ ϕ)(δ ⊗ 1)(δ(p)), pEϕ =Db(p) ξ − (1 ⊗ ϕ ⊗ ϕ)(1 ⊗ δ)(δ(p)), pEϕ =Db(p) ξ − (1 ⊗ ϕ)(δ(p)#1 ⊗ Sϕ( ξ)), pEϕ =Db(p) ξ − (1 ⊗ ϕ)(δ(p) · ξ), pEϕ = hb(p), pξiϕ − hδ(p), p ⊗ (Sϕξ)iHS , where we have used (6) in the final two equalities. We can obtain the equality of the first and last expressions for any x ∈ M ϕ ∩ dom (δ ⊕ δ) by applying Theorem 4.8 to approximate x by polynomials p with kpk ≤ kxk. 20 BRENT NELSON We next compute We focus on the third term above: kb(x)k2 ϕ = hb(x), xξiϕ −(cid:10)¯δ(x), x ⊗ (Sϕξ)(cid:11)HS = hb(x), xξiϕ −(cid:10)x∗ · ¯δ(x), 1 ⊗ (Sϕξ)(cid:11)HS = hb(x), xξiϕ −(cid:10)¯δ(x∗x), 1 ⊗ (Sϕξ)(cid:11)HS +(cid:10)¯δ(x∗) · x, 1 ⊗ (Sϕξ)(cid:11)HS . (cid:10)¯δ(x∗) · x, 1 ⊗ (Sϕξ)(cid:11)HS =D¯δ(x)†, 1 ⊗ Sϕ(xξ)EHS ¯δ(x)EHS =D(xξ) ⊗ 1, = hxξ, b(x)iϕ . Thus we have shown and consequently kb(x)k2 kxξ − b(x)k2 ϕ = 2Re hb(x), xξiϕ −(cid:10)¯δ(x∗x), 1 ⊗ Sϕ(ξ)(cid:11)HS , ϕ −(cid:10)¯δ(x∗x), 1 ⊗ Sϕ(ξ)(cid:11)HS . ϕ = kuξk2 = u1+u2 kuξ − b(u)k2 ϕ = kxξk2 a ϕ = kξk2 ϕ. as a sum of two unitaries u1, u2. In particular, (10) Now, if x = u is a unitary then the above reduces to For self-adjoint x = a, let α > 1 and write 2 αkak a αkak a + is1 − αkak − is1 − u1 = u2 = a2 α2kak2 α2kak2 . a2 Then u1, u2 ∈ M ϕ ∩ dom (δ ⊕ δ) by Corollary 4.7 and by the first part of the proof we have kaξ − b(a)kϕ ≤ αkak 2 (ku1ξ − b(u1)kϕ + ku2ξ − b(u2)kϕ) = αkakkξk2 ϕ. Letting α → 1 yields the desired inequality. Finally, for generic x simply write it as the sum of its real and imaginary parts, use the triangle inequality, and apply the previous bound. Towards proving the inequalities without b, we recall that the Tomita operator Sϕ has the polar decom- ϕ , where Jϕ is an anti-linear isometry and ∆ϕ is the modular operator. In particular, ϕ Sϕ. Hence for x ∈ M ϕ ∩ dom (δ ⊕ δ) we have ξx − (ϕ ⊗ σϕ position Sϕ = Jϕ∆1/2 1 = SϕSϕ = Jϕ∆1/2 −i)( ϕ ϕ hx∗µ ξ − (σϕ −i/2(x∗)µ∆1/2 i ⊗ ϕ)(¯δ(x∗))i ξ − (1 ⊗ ϕ)(σϕ −i/2(x∗) ξ − (1 ⊗ ϕ)(¯δ(σϕ ¯δ(x)) = Jϕ∆1/2 = Jϕhσϕ = Jϕµ1/2hσϕ = Jϕµ1/2hx∗ ξ − (1 ⊗ ϕ)(¯δ(x∗))i −i)( −i/2)(¯δ(x∗))i i/2 ⊗ σϕ i/2(x∗)))i where we have used (4) with z = −i/2 in the second-to-last equality. Consequently ¯δ(x))kϕ = µ1/2kx∗ ξ − (1 ⊗ ϕ)(¯δ(x∗))kϕ kξx − (ϕ ⊗ σϕ (11) If x = u is unitary, then this and the previously established equality for unitaries yields kξu − (ϕ ⊗ σϕ −i)( ¯δ(u))kϕ = µ1/2k ξkϕ = kξkϕ where the last equality follows from a simple computation using (5) and (4). The inequalities for a, x ∈ M ϕ ∩ dom (δ ⊕ δ) self-adjoint and generic, respectively, also follow from (11) by using the previously established inequalities and µ1/2k ξkϕ = kξkϕ. ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 21 Now, fix λ ∈ R×+, and let p ∈ Eλ(dom (δ)). Then (10) holds for p: kpξ − b(p)k2 ϕ = kpξk2 = kpξk2 = kpξk2 ϕ − hδ(p∗p), 1 ⊗ Sϕ(ξ)iHS ϕ −(cid:10)ξ ⊗ 1, δ(p∗p)†(cid:11)HS ϕ − hξ, b(p∗p)iHS . Then, since p∗p ∈ M ϕ, we can use (9) to obtain kpξ − b(p)k2 ϕ + kξkϕ3kp∗pkkξkϕ ϕ ≤ kpξk2 ≤ 4kpk2kξk2 ϕ. Hence kb(p)kϕ ≤ kpξkϕ + kpξ − b(p)kϕ ≤ 3kpkkξkϕ The final estimate then follows from (ϕ ⊗ σϕ −i)( ¯δ(p)) = Jϕ∆1/2 ϕ (σϕ i ⊗ ϕ)(¯δ(p∗)) i/2)(¯δ(p∗)) = Jϕ(1 ⊗ ϕ)(σϕ i/2 ⊗ σϕ = Jϕµ1/2(1 ⊗ ϕ)(¯δ(σϕ i/2(p∗))) = Jϕµ1/2λ1/2(1 ⊗ ϕ)(¯δ(p∗)), and the fact that µ1/2k ξkϕ = kξkϕ. Corollary 4.10. Let δ : dom (δ) → M∞ ⊗ M op ∞ be a µ-modular derivation for some µ > 0, with dom (δ) generated by eigenoperators of σϕ. Assume that the conjugate variable ξ to δ exists. Then the closures of the densely defined maps (1 ⊗ ϕ) ◦ δ and (ϕ ⊗ σϕ −i) ◦ δ on L2(M, ϕ) (both with domain dom (δ)) satisfy for any λ ∈ R×+ and x ∈ Eλ(M ) (cid:3) (cid:13)(cid:13)(cid:13) (1 ⊗ ϕ) ◦ δ(x)(cid:13)(cid:13)(cid:13)ϕ ≤ 3kxkkξkϕ, −i) ◦ δ(x)kϕ ≤ 3λ1/2kxkkξkϕ. k(ϕ ⊗ σϕ and Furthermore, for every λ, γ ∈ R×+ we have Eλ(M ) ⊗ Eγ(M )op ⊂ dom (δ∗). Proof. Denote the two above maps by B and C, respectively. We first show B and C are closable by showing the L2-dense set dom (δ) lies in the domain of their adjoints. Given p, x ∈ dom (δ) we have hp, B(x)iϕ = ϕ(p∗(1 ⊗ ϕ)(δ(x))) = ϕ ⊗ ϕop((p∗ ⊗ 1)#δ(x)) = Dp ⊗ 1, δ(x)Eϕ ≤ kδ∗(p ⊗ 1)kϕkxkϕ, and hence p ∈ dom (B∗). Similarly, we have hp, C(x)iϕ = ϕ(p∗(ϕ ⊗ σϕ −i)(δ(x))) = ϕ((ϕ ⊗ 1)(δ(x))p∗) = ϕ ⊗ ϕop((1 ⊗ p∗)#δ(x)) ≤ kδ∗(1 ⊗ p)kϕkxkϕ, so that p ∈ dom (C∗). Thus B and C are closable, and we let ¯B and ¯C denote their closures. Let x ∈ Eλ(M ) for some λ ∈ R×+, then by Kaplansky's density theorem we can find a sequence (pn)n∈N ⊂ dom (δ) which converges to x in L2(M, ϕ) and satisfies kpnk ≤ kxk for each n. By replacing pn with Eλ(pn) 22 BRENT NELSON for each n ∈ N, we may assume pn ∈ Eλ(dom (δ)). For w ∈ dom (B∗), using the final bounds in Proposition 4.9 we have hx, B∗(w)iϕ = lim = lim n→∞hpn, B∗(w)iϕ n→∞D(1 ⊗ ϕ)(δ(pn), wEϕ ≤ lim sup n→∞ ≤ 3kxkkξkϕkwkϕ, 3kpnkkξkϕkwkϕ which shows that x is in the domain of (B∗)∗ = ¯B with k ¯B(x)kϕ ≤ 3kxkkξkϕ. Similarly, we have x ∈ dom ( ¯C) with k ¯C(x)kϕ ≤ 3λ1/2kxkkξkϕ. Now, for λ, γ ∈ R×+ let a ∈ Eλ(M ) and b ∈ Eγ(M ). As above, we let (pn)n∈N ⊂ Eλ(dom (δ)) and (qn)n∈N ⊂ Eγ(dom (δ)) be sequences converging to a and b in L2(M, ϕ), respectively, and satisfying kpnk ≤ kak and kqnk ≤ kbk for all n ∈ N. Since Sϕ is bounded on Eγ, we note that (q∗n)n∈N converges to b∗ in L2(M, ϕ). Consequently, (pn ⊗ qn)n∈N converges to a ⊗ b in L2(M ¯⊗M op). Using the formula in Lemma 3.9, we have for any w ∈ dom (¯δ) = lim (cid:10)a ⊗ b, ¯δ(w)(cid:11)HS = lim n→∞(cid:10)pn ⊗ qn, ¯δ(w)(cid:11)HS n→∞(cid:10)pnξσϕ n→∞ hkσϕ ≤ lim sup ≤ lim sup n→∞ −i(qn) − B(pn)σϕ −i/2(qn)kkpnξkϕ + kσϕ −i(qn) − pnC(qn), w(cid:11)ϕ −i/2(qn)kkB(pn)kϕ + kpnkkC(qn)kϕikwkϕ 7γ1/2kqnkkpnkkξkϕkwkϕ ≤ 7γ1/2kbkkakkξkϕkwkϕ, Thus a ⊗ b ∈ dom (δ∗). 4.4. Contraction resolvent arising as deformations of δ∗¯δ. Let δ : L2(M, ϕ) → L2(M ¯⊗M op) be a closable µ-modular derivation, for some µ > 0, with dom (δ) generated by eigenoperators of σϕ. One of the key steps in the proofs of our main theorems will be to show that certain central elements z lie in ker(¯δ), and so it will be useful to have a way to approximate such z with elements from dom (¯δ). Towards this end, in this subsection we replicate in M ϕ the analysis of contraction resolvents given in [12, Section 1]. The lemma we prove shows that we can in fact approximate such z with similarly central elements in dom (¯δ). We refer the reader to [16, Chapter I] for a more general treatment of contraction resolvents. We consider L := δ∗¯δ, a self-adjoint operator with dense domain. For each t > 0 define Tt := e−tL. Then {Tt}t>0 is a strongly continuous contraction semigroup with infinitesimal generator −L. For each α > 0 define ηα = α(α + L)−1. Proposition 1.10 of [16] implies (cid:3) α ηα}α>0 is a strongly continuous contraction resolvent (cf. [16, Definition 1.4]). In particular: e−αsTs ds, and that { 1 ηα = αZ ∞ (1) For all α > 0, ηα is a k · kϕ-contraction; and (2) ηα converges strongly to the identity as α → ∞ . Define ζα := (ηα)1/2. Then we have by [20, Lemma 3.2] 1 0 ηα(t + ηα)−1 dt = ζα = 1 π Z ∞ 0 1 √t Moreover, Range(ηα) = dom (L) ⊂ dom (¯δ) (cf. the proof of [16, Proposition 1.5]). π Z ∞ 0 1 √t(1 + t) ηα(1+t)/t dt. (12) Furthermore, Range(ζα) = dom (L1/2) = dom (¯δ), where the latter equality follows from kL1/2(x)kϕ = k¯δ(x)kϕ for all x ∈ dom (L1/2). Consequently, ¯δ ◦ ζα defines a bounded operator. For x ∈ L2(M, ϕ) we have lim α→∞ kx − ζα(x)kϕ = lim 0 1 π Z ∞ α→∞(cid:13)(cid:13)(cid:13)(cid:13) π Z ∞ 1 0 ≤ lim α→∞ 1 √t(1 + t)(cid:0)x − ηα(1+t)/t(x)(cid:1) dt(cid:13)(cid:13)(cid:13)(cid:13)ϕ √t(1 + t)(cid:13)(cid:13)x − ηα(1+t)/t(x)(cid:13)(cid:13)ϕ dt. 1 ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 23 Using property (1) above, we see that the integrand is dominated by convergence theorem implies 1√t(1+t) 2kxkϕ and so the dominated Recalling that L2(M, ϕ) admits bounded left and right actions of M ϕ, we have the following lemma. lim α→∞kx − ζα(x)kϕ = 0. (13) Lemma 4.11. For x ∈ ker(δ) ∩ ker(δ) ∩ M ϕ and ξ ∈ L2(M, ϕ) ζα(ξ · x) = ζα(ξ) · x and ζα(x · ξ) = x · ζα(ξ). Proof. First note that δ(x) = 0 implies L(x) = 0, and so using δ(x) = 0, Lemma 3.9, and Lemma 4.2 we have L(ξ · x) = L(ξ) · x and L(x · ξ) = x · L(ξ) for all ξ ∈ dom (L). We next claim ηα(ξ · x) = ηα(ξ) · x and ηα(x · ξ) = x · ηα(ξ) for all ξ ∈ L2(M, ϕ). Suppose η = ηα(ξ · x) and η′ = ηα(ξ), which we note are contained in dom (L) ⊂ dom (¯δ). Then (α + L)(η − η′ · x) = ξ · x − η′ · x − L(η′) · x = ξ · x − (α + L)(η′) · x = ξ · x − ξ · x = 0, 1 α 1 α 1 α hence ηα(ξ · x) = η = η′ · x = ηα(ξ) · x. Similarly, ηα(x · ξ) = x · ηα(ξ). Finally, using (12) we have ζα(ξ · x) = ζα(ξ) · x and ζα(x · ξ) = x · ζα(ξ). (cid:3) 5. Diffuse Elements in the Centralizer In this section we will show that there is an abundance of diffuse elements in the centralizer M ϕ. Recall our notation from Section 2: where G = G∗ consists of eigenoperators of σϕ and x1, . . . , xn are generators of the form in Proposition 2.1.(ii). Then, more precisely, we will give a condition for when elements in ChGi ∩ M ϕ are diffuse. We begin by replicating [10, Theorem 4.4] in our non-tracial context. P = ChGi = Chx1, . . . , xni , 5.1. An L2-homology estimate. Let P1 ∈ B(L2(M, ϕ)) denote the projection onto the cyclic vector. We let Ψ : M ⊗ M op → FR(L2(M, ϕ)) be the isometry into the finite-rank operators defined by Ψ(a ⊗ b)ξ = aP1bξ, a, b ∈ M, ξ ∈ L2(M, ϕ). For the matrix A ∈ Mn(C) as in Proposition 2.1, let Γ(Rn, Ait)′′ be the free Araki-Woods factor cor- responding to the orthogonal group {Ait}t∈R (cf. [21]). It is generated by quasi-free semicircular elements s1, . . . , sn and admits a free quasi-free state ϕA satisfying for each j = 1, . . . , n n [Aiz]jksk. σϕ z (sj) = Xk=1 The covariance of the system is given by ϕA(sjsk) =h 2 1+Aikj Let HA = L2(Γ(Rn, Ait)′′, ϕA), then HA can be identified with a Fock space on which each sj = ℓ(ej) + ℓ(ej)∗ is a sum of left creation and left annihilation operators for an orthonormal basis {e1, . . . , en} of Rn obliquely embedded in Cn. Letting rj := r(ej ), j = 1, . . . , n, be the corresponding right creation operators, we have that . P1 = hsj, rkiϕA [sj, rk] = hej, ekiHA 1 + A(cid:21)kj We say that {r1, . . . , rn} is a quasi-dual system to {s1, . . . , sn} with covariance We consider the free product (M, θ) = (M, ϕ)∗ (Γ(Rn, Ait)′′, ϕA). Then, by using the right regular repre- sentation for r1, . . . , rn on L2(M, θ) = (L2(M, ϕ), 1)∗(HA, Ω), we can realize these operators in B(L2(M, θ)), where they satisfy [x, rk] = 0 for all x ∈ M . Remark 5.1. Observe that Γ(Rn, Ait)′′ ∼= Γ(Ms.a. ⊂ M )′′ since s1, . . . , sn have the same covariance as the generators x1, . . . , xn and also vary under the modular operator in the same way. Consequently, the maps P1 =(cid:20) 2 1+A . P1. 2 24 BRENT NELSON P ∋ p 7→ δj(p)#sj, j = 1, . . . , n, are exactly the derivations considered in [23], in which Shlyakhtenko defined the conjugate variable to this derivation as an element ξ ∈ L2(M, ϕ) satisfying hξ, piϕ = hsj, δj(p)#sjiθ ∀p ∈ P, provided it exists. The freeness condition implies that for a, b ∈ M we have hsj, asjbiθ = θ(sj asjb) = ϕ(a)ϕ(b)ϕ(s2 j ) = h1 ⊗ 1, a ⊗ biHS . Thus if Jϕ(xj : C[xk : k 6= j]) exists then hJϕ(xj : C[xk : k 6= j]), piϕ = h1 ⊗ 1, δj(p)iHS = hsj, δj(p)#sjiθ . Hence Jϕ(xj : C[xk : k 6= j]) is also a conjugate variable in the sense of [23]. Moreover, since kxjkϕ = 1 for each j = 1, . . . , n, this also implies Φ∗ϕ(x1, . . . , xn) equals the free Fisher information from [23, Definition 2.5]. For each ǫ > 0 and j = 1, . . . , n define xj(ǫ) = xj + √ǫsj and let Mǫ = W ∗(xj (ǫ) : j = 1, . . . , n) ⊂ M. Since xj and sj vary linearly in the same way under the action of σθ, it is easy to see that ∀ǫ > 0 Mǫ is globally invariant under σθ. Therefore, by [26, Theorem IX.4.2], there is a conditional expectation Eǫ : M → Mǫ. For each ǫ > 0 let pǫ denote the orthogonal projection from L2(M, θ) to L2(Mǫ, θ) so that for x ∈ M we have pǫxpǫ = Eǫ(x)pǫ. Define for each j = 1, . . . , n rj(ǫ) = pǫ 1√ǫ rjpǫ so that [xj(ǫ), rk(ǫ)] = pǫ 1 √ǫ [xj + √ǫsj, rk]pǫ = pǫ[sj, rk]pǫ =(cid:20) 2 1 + A(cid:21)kj P1, where we have used the fact that Mǫ is unital to conclude P1 = pǫP1pǫ. Hence {r1(ǫ), . . . , rn(ǫ)} is a quasi-dual system to {x1(ǫ), . . . , xn(ǫ)} with covariance We will associate the following quantity to x1, . . . , xn, which can be thought of as a type of free entropy dimension: 2 1+A . d⋆ A(x1, . . . , xn) := n − lim inf ǫ→0 ǫ n Xj=1(cid:13)(cid:13)J A θ (xj (ǫ) : C[xk(ǫ) : k 6= j])(cid:13)(cid:13) 2 θ . This is the non-tracial analogue of the quantity δ⋆ that appears in the tracial case as the result of formally applying L'Hopital's rule to the free entropy dimension δ∗ (cf. [10, Section 4.1]). We note that we present this quantity merely as a convenient notation. Since our only examples are those where the lim inf in the definition of d⋆ A is zero, and since we do not have a corresponding non-tracial non-microstates free entropy to compare this quantity with, it is unclear if this is really the correct definition for a non-tracial free entropy dimension. In any case, the following lemma, which is proved using arguments from [29, Propositions 3.6 and 3.7] adapted to the present context, tells us that d⋆ A(x1, . . . , xn) = n in the situations we will consider. Lemma 5.2. For each ǫ > 0 and j ∈ {1, . . . , n}, J A θ (xj (ǫ) : C[xk(ǫ) : k 6= j]) = 1 √ǫEǫ(sj), and if J A ϕ (xj : C[xk : k 6= j]) exists then In particular, when J A ϕ (xj : C[xk : k 6= j]). ϕ (xj : C[xk : k 6= j]) exists for each j = 1, . . . , n, we have d⋆ Proof. We first note that for any p ∈ Chx1(ǫ), . . . , xn(ǫ)i and any j ∈ {1, . . . , n} θ (xj (ǫ) : C[xk(ǫ) : k 6= j]) = pǫJ A J A A(x1, . . . , xn) = n. ∂xj (ǫ)(p) = ∂sj (p) 1 √ǫ Consequently, Next, we claim that J A θ (xj (ǫ) : C[xk(ǫ) : k 6= j]) = 1 √ǫ pǫJ A θ (sj : C[xk(ǫ) : k 6= j]). J A θ (sj : C[xk(ǫ) : k 6= j]) = J A ϕA(sj : C[sk : k 6= j]) = sj. ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 25 The last equality follows by [18, Proposition 2.2]. Thus to show the first equality we must demonstrate By expanding each xk(ǫ) = xk + √ǫsk, it suffices to show hsj, piθ =(cid:10)1 ⊗ 1, ∂sj p(cid:11)θ⊗θop ∀p ∈ Chx1(ǫ), . . . , xk(ǫ)i . m hsj, c0p1c1 ··· pmcmiθ = θ ⊗ θop(c0p1c1 ··· ck−1∂sj (pk)ck ··· pmcm), (14) where c0, . . . , cm ∈ P and p1, . . . , pm ∈ Chs1, . . . , sni. We proceed by induction on m. Recalling that θ(sj) = ϕA(sj) = 0 and using free independence (twice) we have hsj, c0p1c1iθ = ϕ(c0)ϕ(c1)ϕA(sjp1) = ϕ(c0)ϕ(c1)ϕA ⊗ ϕop = θ ⊗ θop(c0∂sj (p1)c1). A (∂sj (p1)) Let m ≥ 2 and suppose (14) holds for elements of the form c0p1c1 ··· pℓcℓ with ℓ < m. For a ∈ M, write a = a−θ(a). Upon writing pk = pk +θ(pk) for each k = 1, . . . , m and ck = ck +θ(ck) for each k = 1, . . . , m−1 and expanding, it suffices by the induction hypothesis to consider the case when p1, . . . , pm and c1, . . . , cm−1 are centered with respect to θ. We then have by free independence Xk=1 hsj, c0p1 ··· pmcmiθ = θ(sjc0p1 ··· pmcm) + θ(c0)θ(sj p1 ··· pmcm) = 0. On the other hand, if we write ∂sj (pk) =Pℓ ak m ℓ ⊗ bk ℓ for each k = 1, . . . , m then θ(c0p1c1 ··· ck−1ak Xk=1Xℓ ℓ ) vanish if k ≥ 2, while the factors θ(bk ℓ )θ(bk ℓ ck ··· pmcm). The factors θ(c0p1c1 ··· ck−1ak m ≥ 2, at least one of these conditions always holds and we have proved the claim. ℓ ck ··· pmcm) vanish if k ≤ m− 1. Since That J A θ (xj (ǫ) : C[xk(ǫ) : k 6= j]) = pǫJ A ϕ (xj : C[xk : k 6= j]), holds via a similar argument. Hence, when J A ϕ (xj : C[xk : k 6= j]) exists for each j = 1, . . . , n, we have n n ǫ lim inf ǫ→0 Xj=1(cid:13)(cid:13)J A θ (xj(ǫ) : C[xk(ǫ) : k 6= j])(cid:13)(cid:13) 2 θ ≤ lim ǫ→0 ǫ Xj=1(cid:13)(cid:13)J A ϕ (xj : C[xk : k 6= j])(cid:13)(cid:13) 2 ϕ = 0. (cid:3) Thus d⋆ A(x1, . . . , xn) = n. We let 1 denote the cyclic vector in L2(M, θ). Define Sθ to be the closure of the operator y1 7→ y∗1, . The restrictions of Sθ to the densely defined on M1 ⊂ L2(M, θ), with polar decomposition Sθ = Jθ∆1/2 subspaces L2(M, ϕ) and HA yield Sϕ = Jϕ∆1/2 and SϕA = JϕA∆1/2 ϕA , respectively. ϕ θ For T ∈ B(L2(M, θ)) we define ρ(T ) := JθT ∗Jθ, which we recall from Remark 1.7 is the usual right action of M on L2(M, θ), but differs from the one we have been considering thus far. An easy computation shows that ρ is an isometry on Ψ(M ⊗ M op) with respect to the Hilbert-Schmidt norm. We also note that for a, b ∈ M , f, g ∈ M∞, and ξ ∈ L2(M, ϕ) ρ(f )aP1bρ(g)ξ = h1, bρ(g)ξiϕ ρ(f )a1 = hg∗Jϕbξ, 1iϕ aJϕf∗1 = hJϕg1, bξiϕ aσϕ −i/2(g∗)1, bξEϕ =Dσϕ −i/2(f )P1σϕ = aσϕ −i/2(f )1 aσϕ i/2(g)bξ, −i/2(f )1 thus we define the actions ρ(f ) · (a ⊗ b) · ρ(g) := aσϕ −i/2(f ) ⊗ σϕ i/2(g)b. (15) 26 BRENT NELSON Lemma 5.3. For x ∈ M and rj, sj as above, we have ρ(rj )x1 = 0. If a, b ∈ M∞ we have ρ(aρ(rj )b)1 = σϕ −i/2(a)sjσϕ −i/2(b)1. Proof. For x ∈ M , we have Jθx1 = Jϕx1 = ∆1/2 ϕ x∗1 ∈ L2(M, ϕ). Therefore ρ(rj )x1 = Jθr∗j (∆1/2 ϕ x∗1) = 0. Next, we note the following identity for c ∈ M∞: Jθc1 = ∆1/2 θ Sθc1 = ∆1/2 ϕ c∗1 = σϕ −i/2(c∗)1. We compute ρ(aρ(rj )b)1 = Jθb∗JθrjJθa∗Jθ1 = Jθb∗Jθrjσϕ = Jθb∗Jθhσϕ = Jθhb∗ ⊗ ∆1/2 −i/2(a)sjσϕ = σϕ −i/2(a)1 −i/2(a) ⊗ eji = Jθb∗h∆1/2 ϕA ej ⊗ a∗i = σϕ −i/2(b)1, ϕA ej ⊗ a∗i −i/2(a) ⊗ ej ⊗ σϕ −i/2(b) as claimed. Lemma 5.4. For all x ∈ M and T ∈ B(L2(M, θ)) we have Proof. We simply compute Tr(P1[T1, ρ(x)]) =Dρ(x∗)1,(cid:16)∆−1/2 Tr(P1[T, ρ(x)]) = h1, T ρ(x)1 − ρ(x)T 1iθ θ ρ(T ) − T(cid:17) 1Eθ = hT ∗1, ρ(x)1iθ − hρ(x)∗1, T 1iθ = hx∗1, JθT ∗1iθ − hρ(x∗)1, T 1iθ =D∆−1/2 =Dρ(x∗)1, (∆−1/2 ρ(T ) − T )1Eθ Jθx1, JθT ∗1Eθ − hρ(x∗)1, T 1iθ . θ θ (cid:3) (cid:3) Lemma 5.5. Let δ > 0. Given T =Pℓ aℓP1bℓ ∈ Ψ(M ⊗ M op), there exists ǫ0 > 0 so that for each ǫ ∈ (0, ǫ0) we can find xℓ, yℓ ∈ Mǫ such that if T (ǫ) :=Xℓ xℓP1yℓ, then Proof. This follows from exactly the same argument as in [10, Lemma 4.2], except we must note that kT − T (ǫ)kHS ≤ kT − T (ǫ)k1 < δ. kaP1bk1 = sup = sup kTk∞=1hT, aP1biT r kTk∞=1T r(P1bT ∗aP1) kTk∞=1hT b∗1, a1iϕ = kakϕkb∗kϕ. So we must approximate b∗ (rather than b) by polynomials in the k · kϕ-norm. = sup (cid:3) Note that for a, b ∈ Mǫ and ξ ∈ L2(M, θ) we have so that [T (ǫ), pǫ] = 0 for T (ǫ) ∈ Ψ(Mǫ ⊗ M op ǫ ) as in the previous lemma. aP1bpǫξ = hb∗1, pǫξiθ a1 = hb∗1, ξiθ pǫa1 = pǫaP1bξ, ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 27 Proposition 5.6. With the notation as above, suppose that d⋆ Ψ(M∞ ⊗ M op j=1[Tjk, ρ(xj )] = 0 for each k = 1, . . . , n satisfies A(x1, . . . , xn) = n. Then any set {Tjk}n j,k=1 ⊂ ∞ ) such that Pn n Xj,k=1P1, Tjk(cid:20) 2 1 + A(cid:21)kj+HS = 0. Proof. For each 1 ≤ j, k ≤ n we can write ℓ P1bjk ajk ℓ , Tjk =Xℓ for some aij ℓ , bij ℓ ∈ M∞. We compute n Xj,k=1P1,(cid:20) 2 1 + A(cid:21)kj Tjk+HS = = = = = Tr Tjkρ (cid:20) 2 1 + A(cid:21)kj P1!! Tr(Tjkρ([xj(ǫ), rk(ǫ)])) Tr([ρ(xj (ǫ)), Tjk]ρ(rk(ǫ))) Tr([ρ(√ǫsj), Tjk]ρ(rk(ǫ))) Tr(Tjk[√ǫρ(rk(ǫ)), ρ(sj )]). n n n Xj,k=1 Xj,k=1 Xj,k=1 Xj,k=1 Xj,k=1 n n Let δ > 0, then by Lemma 5.5 there exists ǫ0 > 0 such that if ǫ < ǫ0 then for each Tij we can find Tij(ǫ) ∈ Ψ(Mǫ ⊗ M op ǫ ) satisfying kTij − Tij(ǫ)k1 ≤ δ. Then since we have k[√ǫρ(rk(ǫ)), ρ(sj )]k =(cid:13)(cid:13)(cid:13)(cid:13) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Tjk+HS Xj,k=1P1,(cid:20) 1 + A(cid:21)kj 2 n (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) √ǫJθ[pǫ 1 √ǫ ≤ 4n2δ +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 2kr∗jkksjk ≤ 4, r∗j pǫ, sj]Jθ(cid:13)(cid:13)(cid:13)(cid:13) Tr(Tjk(ǫ)[√ǫρ(rk(ǫ)), ρ(sj)])(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xj,k=1 n . We focus on the latter term. Note that [pǫ, Jθ] = 0 and denote ξj(ǫ) = 1√ǫ pǫsjpǫ. We have Tr(Tjk(ǫ)[√ǫρ(rk(ǫ)), ρ(sj )]) = n Xj,k=1 Tr(Tjk(ǫ)[ρ(rk), ρ(√ǫξj(ǫ))]). n Xj,k=1 Using k√ǫξj(ǫ)k ≤ ksjk ≤ 2 and switching back to Tij from Tij(ǫ) we have n Xj,k=1P1,(cid:20) 2 1 + A(cid:21)kj Tjk+HS (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ 8n2δ +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n Xj,k=1 Tr(Tjk[ρ(rk), ρ(√ǫξj (ǫ))])(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) . 28 BRENT NELSON Now, we have that JθxJθ ∈ M′ ∩ B(L2(M, θ)) for x ∈ M so ajk considering the last term in the above inequality we have ℓ and bjk ℓ commute with ρ(√ǫξj(ǫ)). Hence ℓ , ρ(√ǫξj(ǫ))]) Tr(Tjk[ρ(rk), ρ(√ǫξj (ǫ))]) = n Xj,k=1 = = = = n θ n n Tr(P1[bjk ℓ ρ(rk)ajk Xj,k=1Xℓ Xj,k=1Xℓ Dρ(√ǫξj (ǫ)∗)1,(cid:16)∆−1/2 Xj,k Xℓ Dρ(√ǫξj (ǫ)∗)1, ∆−1/2 Jθ√ǫξj (ǫ)1, σϕ Xj,k Xℓ D∆−1/2 Xj,k Xℓ D√ǫξj (ǫ)1, σϕ −i/2(bjk n n θ θ σϕ −i/2(bjk ℓ )skσϕ −i/2(bjk ℓ )skσϕ ℓ )1Eθ −i/2(ajk ℓ )1Eθ −i/2(ajk ℓ )1Eθ ℓ )skσϕ −i/2(ajk ρ(bjk ℓ ρ(rk)ajk ℓ ) − bjk ℓ ρ(rk)ajk ℓ (cid:17) 1Eθ where we have used Lemmas 5.4 and 5.3, and that ∆−1/2 So by applying the Cauchy -- Schwarz inequality to our previous computation we obtain Jθξj(ǫ)1 = Sθξj(ǫ)1 = ξj(ǫ)1 since sj is self-adjoint. θ n (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xj,k=1 Tr(Tjk[ρ(rk), ρ(√ǫξj(ǫ))])(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Thus we have shown so far that n Xj,k=1P1,(cid:20) 2 1 + A(cid:21)kj Tjk+HS (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ℓ )skσϕ −i/2(ajk ℓ )1+θ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) σϕ −i/2(bjk ℓ )skσϕ −i/2(ajk n n n n n −i/2(bjk σϕ ≤(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xj=1√ǫξj(ǫ)1, Xk=1Xℓ 1/2 θ ≤ Xj=1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kξj(ǫ)k2 Xk=1Xℓ Xj=1 ǫ   1/2 ≤ 8n2δ + Xj=1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Xk=1Xℓ ǫ  θ kξj(ǫ)k2  Xj=1 n n n (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) −i/2(bjk σϕ ℓ )skσϕ −i/2(ajk 1/2 θ 2  ℓ )1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  ℓ )1(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) 1/2 . 2 θ   Since d⋆ the desired equality after first letting ǫ tend to zero and then δ. A(x1, . . . , xn) = n, the first factor in the second term above tends to zero as ǫ → 0. Thus we have (cid:3) Corollary 5.7. Let M be a von Neumann algebra with faithful normal state ϕ. Suppose M is generated by a finite set G = G∗ of eigenoperators of σϕ with finite free Fisher information. If T1, . . . , Tn ∈ Ψ(M∞ ⊗ M op ∞ ) satisfy Pn Proof. Let M, θ, and Mǫ be as above. Lemma 5.2 implies d⋆ j=1[Tj, ρ(xj )] = 0, then T1 = ··· = Tn = 0. A(x1, . . . , xn) = n. We compute n Xj=1 kTjk2 HS = = = n n Xj,m=1 [1]mj hTm, TjiHS Xj,k,m=1(cid:20) 1 + A (cid:21)mk(cid:20) Xj,k,m=1(cid:20) 1 + A (cid:21)km 2 2 n 2 1 + A(cid:21)kj hTm, TjiHS Tj+HS Tm,(cid:20) 1 + A(cid:21)kj 2 We can write for each k and m (cid:20) 1 + A 2 (cid:21)km Tm =Xℓ akm ℓ P1bkm ℓ , ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 29 for some akm ℓ , bkm ℓ ∈ M , so that kTjk2 n Xj=1 HS = n Xj,k P1,(cid:20) 2 1 + A(cid:21)kj n Xm=1Xℓ (akm ℓ )∗Tj(bkm ℓ )∗+HS . We note that since ρ(xj) ∈ M′ ∩ B(L2(M, ϕ)), we have for each k = 1, . . . , n Xm=1Xℓ Xj=1" n 0 = n )∗! # (akm n n ℓ )∗Tj(bkm ℓ )∗, ρ(xj)# . ℓ Xm=1Xℓ )∗ ⊗ (bkm Thus if we set Tjk :=Pn m=1Pℓ(akm Xj=1 kTjk2 of Proposition 5.6 and hence HS = n n ℓ )∗Tj(bkm ℓ Xj,k P1,(cid:20) ℓ (akm [Tj, ρ(xj )] = Xj=1 )∗ for each k = 1, . . . , n then {Tjk}n )∗+HS 1 + A(cid:21)kj )∗Tj(bkm Xm=1Xℓ (akm 2 n ℓ ℓ = 0, j,k=1 satisfies the hypothesis implying Tj = 0 for each j = 1, . . . , n. (cid:3) 5.2. The existence of diffuse elements in the centralizer. Whenever we refer to a monomial in this section, we mean a monomial in ChGi. To avoid confusion, we will temporarily return to using the notation ChGi in place of P. We also remind the reader that the notations Eλ, EI , and Eλ were defined in Subsection 2.2. We define E∞ ≤1 = {x ∈ ChGi : [(ϕ ⊗ 1) ◦ δy1]··· [(ϕ ⊗ 1) ◦ δym](x) ∈ E(0,1], ∀n ≥ 0 and y1, . . . , ym ∈ G}. Observe that for x = y1 ··· ym ∈ ChGi a monomial, x ∈ E∞ ≤1 is equivalent to yj ··· ym ∈ Eλj (M ) with λj ≤ 1 for each j = 1, . . . , m. If wj ∈ C(cid:10)cj, c∗j(cid:11) ∩ E∞ ≤1 are monomials for each j = 1, . . . , n then a monomial formed by any interleaving of the factors in w1, . . . , wn is also in E∞ ≤1. Lemma 5.8. For x ∈ Eλ(M ), λ ≤ 1, if xp = 0 for some non-zero projection p ∈ M , then there exists a non-zero projection q ∈ M ϕ such that qx = 0. Proof. If x = vx is the polar decomposition, then v ∈ Eλ(M ) by Proposition 2.6. Now, xp = 0 implies pker(x), the projection onto the kernel of x, is non-zero. Moreover, if pran(x) and pran(x∗) are the projections onto the closures of the ranges of x and x∗, respectively, then we have pker(x) = 1 − pran(x∗) = 1 − v∗v pker(x∗) = 1 − pran(x) = 1 − vv∗. Hence ϕ(pker(x∗)) = 1 − ϕ(vv∗) = 1 − ϕ(v∗σϕ −i(v)) = 1 − λϕ(v∗v) = (1 − λ) + λϕ(pker(x)) > 0. So letting q = pker(x∗) 6= 0, we have x∗q = 0 or qx = 0, and q = 1 − vv∗ ∈ M ϕ. Theorem 5.9. Assume Φ∗ϕ(G) < ∞. Suppose for x ∈ ChGi there is a monomial x0 of highest degree with non-zero coefficient such that x0 ∈ E∞ Proof. Suppose, towards a contradiction, p ∈ M ϕ is a non-zero projection such that xp = 0. Let λ1 ∈ (0, 1] be such that x0 ∈ Eλ1 (M ). Then ≤1. Then there is no non-zero projection p ∈ M ϕ such that xp = 0. (cid:3) 0 = kxpk2 ϕ = kπλ1 xpk2 ϕ + k(1 − πλ1 )xpk2 ϕ, and hence 0 = πλ1 xp = Eλ1 (xp) = Eλ1 (x)p. Lemma 5.8 implies there is a non-zero projection q1 ∈ M ϕ such that q1Eλ1 (x) = 0. Thus 0 = (q1 ⊗ p)#(Eλ1 (x) ⊗ 1 − 1 ⊗ Eλ1 (x)) = (q1 ⊗ p)#Xy∈G δy(Eλ1 (x))#(y ⊗ 1 − 1 ⊗ y). 30 BRENT NELSON Setting Ty := (σϕ the above equality yields −i/2 ⊗ σϕ i/2) [(q1 ⊗ p)#δy(Eλ1 (x))] for each y ∈ G and applying σϕ 0 = Xy∈G −i/2(y) ⊗ 1 − 1 ⊗ σϕ i/2(y)) = Xy∈G Ty#(σϕ [ρ(y), Ty], −i/2 ⊗ σϕ i/2 to each side of by (15). So by Corollary 5.7, Ty = 0 for each y ∈ G and hence (q1 ⊗ p)#δy(Eλ1 (x)) = 0 for each y ∈ G. Let x0 = y1 ··· ym for y1, . . . , ym ∈ G. Then in particular, (ϕ ⊗ 1)(q1 · δy1(Eλ1 (x)))p = 0. The condition on x0 implies that yℓ ··· ym ∈ E(0,1] for each ℓ ∈ {2, . . . , m}, and so iterating the above argument we can find λ2, . . . , λm ∈ (0, 1] and non-zero projections q2, . . . , qm so that for each ℓ ∈ {2, . . . , m} if (ϕq)(·) = ϕ(q·) then [(ϕqm) ⊗ 1] ◦ δym ◦ Eλm ◦ ··· ◦ [(ϕq1) ⊗ 1] ◦ δy1 ◦ Eλ1 (x)p = 0. But since x0 is of highest degree in x, say with coefficient α 6= 0, we have 0 = [(ϕqm) ⊗ 1] ◦ δym ◦ Eλm ◦ ··· ◦ [(ϕq1) ⊗ 1] ◦ δy1 ◦ Eλ1 (x)p = αϕ(qm)··· ϕ(q1)p, a contradiction. Corollary 5.10. Assume Φ∗ϕ(G) < ∞. Suppose for x ∈ ChGi there is a monomial x0 of highest degree with non-zero coefficient such that x0 ∈ E∞ ≤1. (cid:3) (1) If x ∈ Eλ(M ) for λ ≤ 1, then x has no atoms at zero. (2) If x ∈ M ϕ, then x is diffuse. Proof. If x ∈ Eλ(M ), λ ≤ 1, had an atom at zero, then there would be a non-zero projection p such that xp = 0. Hence we may apply Theorem 5.9 with pker(x) = 1 − v∗v ∈ M ϕ, v coming from the polar decomposition of x, to obtain a contradiction. For x ∈ M ϕ, we simply apply (1) to each translation x − α ∈ E1(M ) = M ϕ, α ∈ C. Corollary 5.11. Assume Φ∗ϕ(G) < ∞. For y1, . . . , ym ∈ G, let λj be the eigenvalue of yj ··· ym for each j = 1, . . . , m, and suppose (cid:3) λ1 = max 1≤j≤m λj . . Then the element x = y∗m ··· y∗1y1 ··· ym ∈ (M ϕ)+ is diffuse when λ1 ≤ 1 and otherwise has exactly one atom, which is at zero and of size 1 − 1 Proof. Define λ0 = λm+1 = 1. First suppose λ1 ≤ 1. For each j = 1, . . . , m, the eigenvalues of yj ··· ym and y∗j ··· y∗1y1 ··· ym are easily seen to be λj and λj+1, respectively. Each of these is less than one by assumption, so x ∈ E∞ ≤1 and is diffuse by Corollary 5.10. Now suppose λ1 > 1 and consider x := y1 ··· ymy∗m ··· y∗1 ∈ (M ϕ)+. For each j = 1, . . . , m, the eigenvalues of y∗j ··· y∗1 and yj ··· ymy∗m ··· y∗1 are λj+1 , respectively. By assumption, each of these is less than and λj λ1 λ1 λ1 λ1 λ1 = 1, so x ∈ E∞ ≤1 and hence is diffuse. Now, for any d ≥ 1 we have i (y1 ··· ym)xk−1y∗m ··· y∗1) = ϕ(xk) = ϕ(σϕ 1 λ1 ϕ(xk). λ1 at zero. Consequently, for any polynomial p with p(0) = 0, we have ϕ(p(x)) = λ−1 1 ϕ(p(x)). By approximating a Dirac mass δt for t ∈ (0,∞) by such polynomials, we see that x has no atoms away from zero since x is diffuse. On the other hand, by approximating 1 − δ0 by such polynomials we see that x must have an atom of size 1 − 1 Example 5.12. Let y ∈ G have eigenvalue λ ≤ 1. Then by Corollary 5.11, y∗y is diffuse and yy∗ has an atom of size 1 − λ at zero. Remark 5.13. One application of Corollary 5.10 is to operators arising from loops on a weighted graph which begin and end on a vertex of minimal weight. Following [14, Section 4], given an oriented bipartite graph Γ = (V, E), one can associate to each edge e ∈ E a generalized circular operator c(e) acting on a (cid:3) ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 31 Fock space FΓ associated to the graph. Let ϕΓ denote the vacuum state on B(FΓ), then the action of its associated modular automorphism group σϕΓ on the operators c(e) is known: σϕΓ µ(s(e))(cid:19)iz z (c(e)) =(cid:18) µ(t(e)) , z ∈ C, where s(e), t(e) ∈ V are the source and target vertices for e, respectively, and µ is the weighting given by the Perron -- Frobenius eigenvector of the adjacency matrix of Γ. It then follows that c(e1)c(e2)··· c(em) is always an eigenoperator of σϕΓ , and in particular is in the centralizer with respect to ϕΓ whenever e1e2 ··· em is a loop in Γ. Fix a loop e1 ··· em in Γ and observe that the eigenvalue of c(ej)··· c(em), 1 ≤ j ≤ m, is µ(t(ej))) µ(s(ej))) µ(t(ej+1)) µ(s(ej+1)) ··· µ(t(em)) µ(s(em)) = µ(t(em)) µ(s(ej)) , since t(ek) = s(ek+1). Thus if t(em) = s(e1) has minimal µ-weight amongst all vertices traversed by the loop, then c(e1)··· c(em) is diffuse by Corollary 5.10. A more direct proof of this result as well as a broader consideration of graphs and weightings µ can be found in [15]. 6. Derivations From The Polar Decomposition Recall that for y ∈ G, δy is the free difference quotient with respect to y, which is λy-modular. When it exists, we denote by ξy the conjugate variable to δy. From these derivations and the polar decomposition of y we will construct new derivations. In particular, we will construct a derivation associated to y which can be restricted to the tracial context of (M ϕ, ϕ). 6.1. The derivations δy and δv. For y ∈ G with y 6= y∗ and polar decomposition y = vy, we now consider two new derivations δy and δv. Set dom (δy) = dom (δv) = P and for p ∈ P define these derivations by δy(p) := δy(p)#(v ⊗ 1) + δy∗(p)#(1 ⊗ v∗), δv(p) := δy(p)#(v ⊗ y) − δy∗(p)#(y ⊗ v∗). Observe that δy and δv are 1-modular, since v ∈ Eλy (M ). Also δy = δy and δv = −δv. Lemma 6.1. Fix y ∈ G with polar decomposition y = vy, and suppose y 6= y∗ and that ξy exists. Then 1 ⊗ 1 ∈ dom (δ∗ y ) ∩ dom (δ∗v ) so that δy and δv are closable as densely defined maps δy, δv : L2(M, ϕ) → L2(M ¯⊗M op). ¯δv(y) = 0, ¯δv(y∗) = [v ⊗ v∗,y∗], ¯δv(v) = v ⊗ v∗v, and ¯δv(v∗) = −v∗v ⊗ v. Proof. We compute for p ∈ P (cid:10)1 ⊗ 1, δy(p)(cid:11)HS = h1 ⊗ 1, δy(p)#v ⊗ 1iHS + h1 ⊗ 1, δy∗(p)#1 ⊗ v∗iHS + hλy1 ⊗ v, δy∗ (p)iHS . =(cid:10)λ−1 y v∗ ⊗ 1, δy(p)(cid:11)HS Now, Corollary 4.10 implies v∗ ⊗ 1 ∈ dom (δ∗y) and 1 ⊗ v ∈ dom (δ∗y∗ ). Thus which implies 1 ⊗ 1 ∈ dom (δ∗ y We note that (cid:10)1 ⊗ 1, δy(p)(cid:11)HS =(cid:10)λ−1 y δ∗y(v∗ ⊗ 1) + λyδ∗y∗(1 ⊗ v), p(cid:11)ϕ ). Hence δy is closable by Lemma 3.9. The proof for δv is similar. , δy(y∗y) = (1 ⊗ y)#(1 ⊗ v∗) + (y∗ ⊗ 1)#(v ⊗ 1) = 1 ⊗ (v∗y) + (y∗v) ⊗ 1 = 1 ⊗ y + y ⊗ 1, Furthermore, y,y∗ ∈ dom (¯δy) with and y,y∗, v, v∗ ∈ dom (¯δv) with ¯δy(y) = v∗v ⊗ 1 + 1 ⊗ v∗v − v∗v ⊗ v∗v and ¯δy(y∗) = v ⊗ v∗, 32 BRENT NELSON since v∗v = 1 − pker(y) = 1 − pker(y). Let I ⊂ [0,∞) be a compact interval containing the spectrum of y∗y, and let µ be the spectral measure of y∗y. For each α > 0, consider the function By Lemma 4.3, fα(y∗y) ∈ dom (¯δy) and ¯δy(fα(y∗y)) is identified (via the functional calculus) with fα(t) =pt + α2 ∈ C1(I). (√t + √s) = √t + √s √t + α2 + √s + α2 , fα(t) − fα(s) t − s which by Lebesgue's dominated convergence theorem (DCT) converges in L2(I × I, µ × µ) to the function 1 − χ{(0,0)}(t, s), where χE denotes the characteristic function on a set E. Since v∗v = 1 − pker(y∗y) = χ(0,∞)(y∗y) and 1 − χ{(0,0)}(t, s) = χ(0,∞)(t) + χ(0,∞)(s) − χ(0,∞)(t)χ(0,∞)(s), this limit is identified via the Borel functional calculus for y∗y with v∗v ⊗ 1 + 1 ⊗ v∗v − v∗v ⊗ v∗v. Thus ¯δy(fα(y∗y)) → v∗v⊗1+1⊗v∗v−v∗v⊗v∗v in L2(M ¯⊗M op). Since fα(t) converges to √t uniformly on I, it follows that fα(y∗y) → y in L2(M, ϕ). Hence, y ∈ dom (¯δy) with ¯δy(y) = v∗v⊗1+1⊗v∗v−v∗v⊗v∗v. Now let I ⊂ [0,∞) be a compact interval containing the spectrum of y, and let µ be the spectral measure of y. For each α > 0, consider the function gα(t) = t (t + α)2 ∈ C1(I). Since y∗ = vyv∗ and v is an eigenoperator with eigenvalue λy, we have kygα(y)y∗ − y∗kϕ = λykygα(y)y − ykϕ. So ygα(y)y∗ → y∗ in L2(M, ϕ) since t2gα(t) → t in L2(I, µ) by Lebesgue's DCT. By Corollary 4.7 we have gα(y) ∈ dom (¯δy). Then, since gα(y) ∈ M ϕ and y, y∗ ∈ M0 we have by Proposition 4.6 that ygα(y)y∗ ∈ dom (¯δy) with ¯δy(ygα(y)y∗) = v ⊗ gα(y)y∗ + y · LRy(gα) · y∗ + ygα(y) ⊗ v∗. We claim that this converges to v ⊗ v∗ in L2(M ¯⊗M op). Let p = v∗v = 1 − pker(y), then a straightforward computation shows kv ⊗ gα(y)y∗+y · LRy(gα) · y∗ + ygα(y) ⊗ v∗ − v ⊗ v∗kHS = kp ⊗ gα(y)y + y · LRy(gα) · y + ygα(y) ⊗ p − p ⊗ pkHS. It is readily seen (after computing gα) that p⊗gα(y)y+y·LRy(gα)·y+ygα(y)⊗p−p⊗p corresponds via the Borel functional calculus for y to χ(0,∞)(t)sgα(s)+(α2 − ts)gα(t)gα(s) + tgα(t)χ(0,∞)(s) − χ(0,∞)(t)χ(0,∞)(s) tα s2 t2 (s + α)2 − χ(0,∞)(s)(cid:21) + sα (s + α)2 . (t + α)2 This converges to zero in L2(I × I, µ× µ) by Lebesgue's DCT. Hence y∗ ∈ dom(¯δy) with ¯δy(y∗) = v⊗ v∗. (t + α)2(cid:21)(cid:20) Seeing that y ∈ dom (¯δv) is much easier. Indeed, =(cid:20)χ(0,∞)(t) − δv(y∗y) = −(y ⊗ v∗) · y + y∗ · (v ⊗ y) = −y ⊗ (v∗vy) + (yv∗v) ⊗ y = 0, so by Lemma 4.3 fα(y∗y) ∈ dom (¯δv) with ¯δv(fα(y∗y)) = 0 for all α > 0. This, in turn, implies y ∈ dom (¯δv) with ¯δv(y) = 0. Towards showing v ∈ dom (¯δv), we consider the elements y(α + y)−1, α > 0, which we claim converge to v in L2(M, ϕ) as α → 0. Once again let I ⊂ [0,∞) be a compact interval containing the spectrum of y, and let µ be the spectral measure of y. Noting that ky(α + y)−1 − vkϕ = ky(α + y)−1 − v∗vkϕ, ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 33 we see that the convergence follows because y(α +y)−1− v∗v corresponds via the Borel functional calculus of y to the function t(α + t)−1 − χ(0,∞)(t), which converges to zero on L2(I, µ) by Lebesgue's DCT. Furthermore, (α + y)−1 ∈ dom (¯δv) with ¯δv((α + y)−1) = 0 by Lemma 4.3. So by the first part of Lemma 4.2 we have ¯δv(y(α + y)−1) = (v ⊗ y) · (α + y)−1, which converges to v ⊗ v∗v as α → 0. Thus v ∈ dom (¯δv) with ¯δv(v) = v ⊗ v∗v. The proof for v∗ is similar using (α + y)−1y∗ to approximate v∗. Finally, observe that since δv = −δv we have dom (δv ⊕ δv) = dom (¯δv). Thus Proposition 4.6 implies y∗ = vyv∗ ∈ dom (¯δv) with ¯δv(y∗) = v ⊗ v∗vyv∗ − vyv∗v ⊗ v∗ = v ⊗ v∗y∗ − y∗v ⊗ v∗ = [v ⊗ v∗,y∗] as claimed. Remark 6.2. Consider the derivation In light of Lemma 6.1, we can see δv := δy#(vv∗ ⊗ vy) − δy∗ #(yv∗ ⊗ vv∗) = δv#(v∗ ⊗ v). (cid:3) which is exactly the derivation defined in [22, Section 2.3] -- the inspiration for δv. δv(v) = vv∗ ⊗ v δv(v∗) = −v∗ ⊗ vv∗, The primary benefit of the derivation ¯δy, is that it gives us a derivation on the tracial von Neumann algebra M ϕ. By composing this with Eϕ⊗E op ϕ , we can consider an exclusively tracial context. In the following lemma, we check a few details for these derivations that will be relevant for the proofs of Theorems A and B. Lemma 6.3. For y ∈ G satisfying y 6= y∗ and λy ≤ 1, assume ξy exists. Let Y1 = y, Y2 = y∗, and D = P[Y1, Y2] ∩ M ϕ (a dense subset of L2(M ϕ, ϕ)). Then ϕ ) ◦ ¯δYj D δj := (Eϕ ⊗ E op j = 1, 2 are 1-modular derivations valued in M ϕ ⊗ (M ϕ)op that satisfy for j = 1, 2: (i) 1 ⊗ 1 ∈ dom (δ∗j ) with δ∗j (1 ⊗ 1) ∈ L2(M ϕ, ϕ); (ii) δj is closable; (iii) (1 ⊗ ϕ) ◦ δj extends to a bounded map from M ϕ to L2(M ϕ, ϕ). Moreover, δj(Yk) = δj=kUj for j, k = 1, 2, where U1 := 1 ⊗ 1 and U2 := vv∗ ⊗ 1 + 1 ⊗ vv∗ − vv∗ ⊗ vv∗. Also, δj#Uj = δj for j = 1, 2. Proof. Note that P ∩ M ϕ = Eϕ(P) is dense in L2(M ϕ, ϕ): P is dense in L2(M, ϕ) and kξ − pk2 Consequently, D is dense. ϕ = kξ − Eϕ(p)k2 ϕ + k(1 − Eϕ)(p)k2 ϕ ∀ξ ∈ L2(M ϕ, ϕ), p ∈ P. Lemma 6.1 and Proposition 4.6 imply that ¯δy and ¯δy∗ are 1-modular derivations on D. Since D ⊂ M ϕ we see that δ1 and δ2 satisfy the Leibniz rule and are therefore 1-modular derivations valued in M ϕ⊗(M ϕ)op. Lemma 6.1 implies that 1 ⊗ 1 ∈ dom (δ∗ (1 ⊗ 1) ∈ L2(M, ϕ). Then Lemma 3.10 implies ) with ξy := δ∗ y y ξy is invariant under ∆ϕ and therefore contained in L2(M ϕ, ϕ). Since Eϕ ⊗ E op ϕ (1 ⊗ 1) = 1 ⊗ 1, it follows that 1⊗ 1 ∈ dom (δ∗1) with δ∗1(1⊗ 1) = δ∗ (1⊗ 1) = ξy. Similarly for y∗ and δ2. This establishes (i). Since y the conjugate variables exist, Lemma 3.9 implies δ1 and δ2 are closable. Thus (ii) is established, and (iii) follows from Corollary 4.10. Note that λy ≤ 1 implies y∗y is diffuse by Corollary 5.10. Hence v∗v = 1 and the formula for δj(Uk), j, k = 1, 2, follows from Lemma 6.1. Now, δ1#U1 = δ1#(1⊗1) = δ1 is clear, and towards showing δ2#U2 = δ2 observe that δ2#U2 = (Eϕ ⊗ E op ϕ )(cid:0)¯δY2 D #U2(cid:1) . 34 BRENT NELSON So it suffices to show ¯δy∗#U2 = U2. Recall that δy∗ = δy∗ #(v∗ ⊗ 1) + δy#(1 ⊗ v). Since (v∗ ⊗ 1)#U2 = v∗ ⊗ 1, and similarly for 1 ⊗ v, we have δy∗#U2 = U2. Let ξ ∈ dom (¯δy∗) be approximated by {pn}n∈N ⊂ P in the k · k¯δy∗ -norm. Then for any η ∈ L2(M ¯⊗M op) we have (cid:10)¯δy∗(ξ)#U2, η(cid:11)HS =(cid:10)¯δy∗(ξ), η#U∗2(cid:11)HS = lim = lim n→∞(cid:10)δy∗(pn), η#U∗2(cid:11)HS n→∞(cid:10)δy∗(pn)#U2, η(cid:11)HS n→∞(cid:10)δy∗(pn), η(cid:11)HS =(cid:10)¯δy∗(ξ), η(cid:11)HS , = lim where we have used U2 ∈ M ϕ ⊗ (M ϕ)op. Thus ¯δy∗#U2 = ¯δy∗ and the desired formula holds. (cid:3) We are very grateful to Yoann Dabrowski who suggested to us the following lemma, which is a modified version of [11, Proposition 3.21]. Lemma 6.4. For y ∈ G satisfying y 6= y∗ and λy ≤ 1, assume ξy exists. With the same notations as in the previous lemma, consider δ0 : D → M ϕ ⊗ (M ϕ)op defined for p ∈ D by 2 Then 1 ⊗ 1 ∈ dom (δ∗0) with δ0(p) := Xj=1 δj(p)#(Yj ⊗ 1 − 1 ⊗ Yj) − [p, 1 ⊗ 1]. 2 In particular, δ0 is a closable operator, with closure ¯δ0 satisfying L2(C[Y1, Y2], ϕ) ⊂ ker(¯δ0). Moreover, for any x ∈ dom (¯δ0) ∩ M ϕ we have δ∗0(1 ⊗ 1) = [Yj, δ∗j (1 ⊗ 1)]. Xj=1 k[x, U2]k2 HS = −(cid:10)¯δ0(x), [x, 1 ⊗ 1](cid:11)HS 2 + Xj=1(cid:16)(cid:10)[x, Yj], [x, δ∗j (Uj)](cid:11)ϕ + 2Re (cid:10)(ϕ ⊗ 1 − 1 ⊗ ϕop) ◦ ¯δj(x), [x, Yj ](cid:11)ϕ(cid:17) . (16) Proof. For p ∈ D note that ϕ ⊗ ϕop([p, Uj]) = 0 for j = 1, 2. Thus, we have * 2 Xj=1 [Yj, δ∗j (1 ⊗ 1)], p+ ϕ h1 ⊗ 1, δj([Yj , p])iHS 2 2 = = Xj=1 Xj=1 Xj=1 = h1 ⊗ 1, δ0(p)iHS . = 2 h1 ⊗ 1, [Uj, p]iHS + hYj ⊗ 1 − 1 ⊗ Yj, δj(p)iHS h1 ⊗ 1, δj(p)#(Yj ⊗ 1 − 1 ⊗ Yj)iHS So 1 ⊗ 1 ∈ dom (δ∗0) with the claimed image. Hence δ0 is closable by Lemma 3.9. dom (¯δ0) with δ0 identically zero on this subspace. For p ∈ C[Y1, Y2], δ0(p) = 0 since U2#(Y2 ⊗ 1− 1⊗ Y2) = Y2⊗ 1− 1⊗ Y2. It follows that L2(C[Y1, Y2], ϕ) ⊂ We now establish (16). First note that Uj ∈ dom (δ∗k) for j, k = 1, 2 by Corollary 4.10, and so the right- hand side is well-defined. Also, by Theorem 4.8 it suffices to establish this formula for p ∈ D. We have δj([p, Yj]) = [δj(p), Yj] + [p, Uj]. So k[p, Uj]k2 HS = hδj([p, Yj]), [p, Uj]iHS − h[δj(p), Yj], [p, Uj]iHS =: Ij − IIj. ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 35 We compute Ij = h[p∗, δj([p, Yj])], UjiHS = hδj([p∗, [p, Yj]]), UjiHS − h[δj(p∗), [p, Yj]], UjiHS =(cid:10)[p, Yj], [p, δ∗j (Uj)](cid:11)HS + hδj(p∗), [Uj, [p∗, Yj]]iHS Focusing on the second term, we use δj = δj and U†j = Uj to observe that hδj(p∗), [Uj, [p∗, Yj]]iHS = h[Uj, [p∗, Yj]], δj(p)†iHS = ϕ ⊗ ϕop([Uj, [p∗, Yj]]∗#δj (p)†) = ϕ ⊗ ϕop(([Uj, [p∗, Yj]]†)∗#δj(p)) =(cid:10)[Uj, [p∗, Yj]]†, δj(p)(cid:11)HS = h[[Yj , p], Uj], δj(p)iHS = h[Uj, [p, Yj]], δj(p)iHS . Thus we have shown Next, using that ϕ is a trace on M ϕ, we have Ij =(cid:10)[p, Yj], [p, δ∗j (Uj)](cid:11)HS + h[Uj, [p, Yj]], δj(p)iHS . IIj = hδj(p), [[p, Uj], Yj]iHS = hδj(p), [p, [Uj, Yj]]iHS − hδj(p), [Uj, [p, Yj]]iHS . Note that [Uj, Yj] = 1 ⊗ Yj − Yj ⊗ 1. So considering [p, [Uj, Yj]] appearing in the first term above, we have [p, [Uj, Yj]] = (p ⊗ 1 − 1 ⊗ p)#[Uj, Yj] = [p, 1 ⊗ 1]#[Uj, Yj] = [p, 1 ⊗ 1]#(1 ⊗ Yj − Yj ⊗ 1). 2 Since ϕ ⊗ ϕop is tracial on M ϕ ⊗ (M ϕ)op we have Xj=1 = h−δ0(p) − [p, 1 ⊗ 1], [p, 1 ⊗ 1]iHS = − hδ0(p), [p, 1 ⊗ 1]iHS − k[p, 1 ⊗ 1]k2 HS. hδj(p), [p, [Uj, Yj]]iHS = hδj(p)#(1 ⊗ Yj − Yj ⊗ 1), [p, 1 ⊗ 1]iHS Xj=1 2 Thus we have shown 2 Xj=1 −IIj = hδ0(p), [p, 1 ⊗ 1]iHS + k[p, 1 ⊗ 1]k2 HS + 2 Xj=1 hδj(p), [Uj, [p, Yj]]iHS Combining this with out previous observations about Ij, we have 2 2 Xj=1 k[p, Uj]k2 HS = Xj=1(cid:16)(cid:10)[p, Yj], [p, δ∗j (Uj)](cid:11)HS + 2Re hδj(p), [Uj, [p, Yj]]iHS(cid:17) + hδ0(p), [p, 1 ⊗ 1]iHS + k[p, 1 ⊗ 1]k2 HS from each side gives the desired formula, provided we show HS. Subtracting k[p, U1]k2 HS = k[p, 1 ⊗ 1]k2 hδj(p), [Uj, [p, Yj]]iHS = h(ϕ ⊗ 1 − 1 ⊗ ϕop) ◦ δj(p), [p, Yj]iϕ . We compute hδj(p), [Uj, [p, Yj]]iHS = (ϕ ⊗ ϕop)(δj(p)∗#(1 ⊗ [p, Yj] − [p, Yj] ⊗ 1)#Uj) = (ϕ ⊗ ϕop)(Uj#δj(p)∗#(1 ⊗ [p, Yj] − [p, Yj] ⊗ 1)) = (ϕ ⊗ ϕop)(δj(p)∗#(1 ⊗ [p, Yj] − [p, Yj] ⊗ 1)), 36 BRENT NELSON where we have used Lemma 6.3 to assert δj#Uj = δj in the last equality. Continuing, we obtain hδj(p), [Uj, [p, Yj]]iHS = (ϕ ⊗ ϕop)(δj (p)∗#(1 ⊗ [p, Yj] − [p, Yj] ⊗ 1)) = ϕ(cid:16)[p, Yj](ϕ ⊗ 1)(δj(p)∗) − (1 ⊗ ϕop)(δj (p)∗)[p, Yj ](cid:17) = ϕ(cid:16)(ϕ ⊗ 1)(δj(p))∗[p, Yj] − (1 ⊗ ϕop)(δj (p))∗[p, Yj](cid:17) = h(ϕ ⊗ 1 − 1 ⊗ ϕop) ◦ δj(p), [p, Yj]iϕ , as claimed. 6.2. Extending ¯δy via the predual (M ¯⊗M op)∗. Fix y ∈ G with λy ≤ 1, so that y is diffuse by Corollary 5.11. Though we won't use it to prove either of our main theorems, there does exist a closed extension of ¯δy that is defined on v and v∗ and sends both to zero. To see this extension, though, one must first expand the codomain of the derivations to the predual (M ¯⊗M op)∗. Let M be a von Neumann algebra with faithful normal state ψ. Recall that L2(M, ψ) can be embedded into M∗ the predual of M via (cid:3) L2(M, ψ) ∋ ξ 7→ ωξ ∈ M∗, where ωξ(x) = h1, xξiψ for x ∈ M. Lemma 6.5. Suppose a ∈ M satisfies a ∈ Mψ. Then kωak ≤ ψ(a). Proof. If ψ is a trace, then this follows by [25, Equation V.2.(2)]. The proof presented here is identical modulo the additional non-tracial hypothesis. Suppose a has polar decomposition a = va, and let x ∈ M have polar decomposition x = wx. Then ωa(x)2 = ψ(xa)2 = ψ(a ≤ ψ(a 1 2x 2 wx 2 wxw∗a 1 1 2 )2 2 va 2 )ψ(a 1 1 1 1 2 v∗xva But wxw∗ = x∗ ≤ kxk1 v∗xv ≤ kxkv∗v ≤ kxk1, , and so that continuing our previous estimate we have ωa(x)2 ≤ kxk2ψ(a)2. Using the above embedding, we think of δy as a densely defined map 1 2 ). (cid:3) with adjoint δy : L2(M, ϕ) → (M ¯⊗M op)∗ δ⋆ y : M ¯⊗M op → L2(M, ϕ). Proposition 6.6. Fix y ∈ G with λy ≤ 1 and polar decomposition y = vy, and suppose y 6= y∗ and that ξy exists. Then δy is closable as a densely defined map say with closure δ 1 y, and v, v∗ ∈ dom (δ δy : L2(M, ϕ) → (M ¯⊗M op)∗, y(v∗) = 0. y) with δ y(v) = δ 1 1 1 Proof. The same proof as in Lemma 6.1 shows that δ⋆ (1⊗1) ∈ L2(M, ϕ) exists and that P⊗P op ⊂ dom (δ⋆ ). y y We claim that the weak density of P ⊗P op in M ¯⊗M op suffices to establish closability of δy. Indeed, suppose dom (δy) ∋ xn → 0 in L2(M, ϕ) and δy(xn) → ω ∈ (M ¯⊗M op)∗. Given ǫ > 0, let ζ ∈ P ⊗ P op be such that Then, we have kωk ≤ ω(ζ) + ǫ. kωk ≤ lim = lim n→∞(cid:10)1 ⊗ 1, ζ#δy(xn)(cid:11)HS + ǫ (ζ∗), xnEϕ + ǫ = ǫ. n→∞Dδ⋆ y ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 37 y and note that it is an Recall from the proof of Lemma 6.1, that y(α +y)−1 converges in L2(M, ϕ) to v as α → 0. Furthermore, Thus ω = 0 and δy is closable as a map into (M ¯⊗M op)∗. Denote this closure by δ extension of ¯δy by the aforementioned embedding. (α + y)−1 ∈ dom (¯δy) ⊂ dom (δ y) by Corollary 4.7. The computation 1 1 along with the formulas in Lemma 6.1 imply α+s 1 α+t − 1 t − s = −1 (α + t)(α + s) 1 δ y((α + y)−1) = −(cid:2)(α + y)−1 ⊗ (α + y)−1(cid:3) #¯δy(y) = −(α + y)−1 ⊗ (α + y)−1 1 (note v∗v = 1 since λy ≤ 1). By Lemma 4.2, y(α + y)−1 ∈ dom (¯δy) ⊂ dom (δ y) with 1 δ y(y(α + y)−1) = v ⊗ (α + y)−1 − y(α + y)−1 ⊗ (α + y)−1 = v(cid:2)1 − y(α + y)−1(cid:3) ⊗ (α + y)−1 = vα(α + y)−1 ⊗ (α + y)−1 = v√α(α + y)−1 ⊗ √α(α + y)−1, It is easy to see that v√α(α + y)−1 ⊗ √α(α + y)−1 = √α(α + y)−1 ⊗ √α(α + y)−1 ∈ (M ¯⊗M op)ϕ⊗ϕop , which can be identified with the function gα(t, s) = √α α + t √α α + s . Now, let I ⊂ [0,∞) be a compact interval containing the spectrum of y, let µ be the spectral measure of y, and let m be the Lebesgue measure. Since ¯δy(y) = 1 ⊗ 1, ¯δy is the free difference quotient for y, and hence 1 ⊗ 1 ∈ dom (δ∗ ) implies that y has finite free Fisher information. Consequently, [27] and [29] y imply that p := dµ/dm ∈ L3(R, m). Thus, by Holder's inequality we have kgαkL1(I×I,µ×µ) ≤ kgαkL 3 2 (I×I,m×m)kpk2 L3(I,m) An easy computation shows kgαkL3/2(I×I,m×m) → 0 as α → 0. Lemma 6.5 then implies δ in (M ¯⊗M op)∗ as α → 0, and hence v ∈ dom (δ that y and v belong to the same eigenspace Eλ. y(y(α+y)−1) → 0 y(v) = 0. The proof for v∗ is similar, but does use y) with δ (cid:3) 1 1 1 3 Remark 6.7. Note that unless the spectrum of y is bounded away from zero, gα(s, t)2 does not converge to 2 (I × I, m× m). This means the above argument cannot show that gα(s, t) → 0 in L2(I × I, µ× µ). zero in L Let ¯δy denote the closure of δy as a map into L2(M ¯⊗M op). By the above proposition, we can consider an extension dy of ¯δy defined on span{Chv, v∗i ∪ dom (¯δy)} so that dy(v) = dy(v∗) = 0, the rest of its definition being determined by the Leibniz rule and ¯δy. As a map into (M ¯⊗M op)∗, it is clear that dy is y; however, this also implies that dy is closable as a map into L2(M ¯⊗M op). Indeed, closable with closure δ if (xn)n∈N ⊂ dom (dy) converges to zero in L2(M, ϕ) and (dy(xn))n∈N converges to some η ∈ L2(M ¯⊗M op), then (since kωζk ≤ kζkHS for ζ ∈ L2(M ¯⊗M op)) it follows that (dy(xn))n∈N converges to ωη as elements of (M ¯⊗M op)∗. Hence ωη = 0, since δ 1 1 y is closed, which then implies η = 0 because ha, ηiHS = ωη(a∗) = 0 ∀a ∈ M ¯⊗M op. Thus, as maps into L2(M ¯⊗M op), the closure of dy is a closed extension of ¯δy with the claimed properties regarding v, v∗. 38 BRENT NELSON 7. Main Results We conclude with the proofs of our main results. We also include a few easy corollaries that minimize the hypotheses. Theorem A. Let M be a von Neumann algebra with a faithful normal state ϕ. Suppose M is generated by a finite set G = G∗, G ≥ 2, of eigenoperators of σϕ with finite free Fisher information. Then (M ϕ)′∩M = C. In particular, M ϕ is a II1 factor and if H < R×+ is the closed subgroup generated by the eigenvalues of G then M is a factor of type  III1 IIIλ II1 if H = R×+ if H = λZ, 0 < λ < 1 if H = {1}. Proof. By Corollary 2.9, (M ϕ)′ ∩ M = (M ϕ)′ ∩ M ϕ. Also M′ ∩ M ⊂ M′ ∩ M ϕ ⊂ (M ϕ)′ ∩ M ϕ, so it suffices to show M ϕ is a factor. Fix z ∈ (M ϕ)′∩ M ϕ. Pick y ∈ G such that y 6= y∗ (if no such y exists then M ϕ = M and we simply appeal to the tracial result [12, Theorem 1]). By replacing y with y∗ if necessary, we may assume λy ≤ 1 so that y∗y is diffuse by Corollary 5.10. Let δ := δ2 be as in Lemma 6.3 with Y2 = y∗. Then δ(y∗y) = Eϕ ⊗ E op ϕ (v∗ ⊗ y + y∗ ⊗ v) = 0 where y∗ = v∗y∗ is the polar decomposition. Let {ζα}α>0 be the maps derived from the contraction resolvent associated to L = δ∗¯δ. Then by Lemma 4.11 we have Then Proposition 4.6 implies 0 = ζα([z, y∗y]) = [ζα(z), y∗y], 0 = ¯δ([ζα(z), y∗y]) = [¯δ ◦ ζα(z), y∗y]. Hence ¯δ ◦ ζα(z) is a Hilbert -- Schmidt operator commuting with a diffuse operator, implying ¯δ ◦ ζα(z) = 0 for all α > 0. Since limα→∞ ζα(z) = z by (13) we see that z ∈ dom (¯δ) with ¯δ(z) = 0. But then, invoking Proposition 4.6 again as well as the computations in Lemma 6.1, we have 0 = ¯δ([z,y∗]) = [z, p ⊗ 1 + 1 ⊗ p − p ⊗ p], where p = vv∗. Applying 1 ⊗ ϕ yields zp + z(1 − p)ϕ(p) = pϕ(z) + (1 − p)ϕ(pz). Multiplying both sides by p and applying ϕ reveals that ϕ(zp) = ϕ(z)ϕ(p). So we can rewrite the above equation as: Then multiplying by p + 1 ϕ(p) (1 − p) yields zp + z(1 − p)ϕ(p) = pϕ(z) + (1 − p)ϕ(p)ϕ(z). zp + z(1 − p) = ϕ(z)p + ϕ(z)(1 − p), or z = ϕ(z) ∈ C as claimed. Now, M ϕ is a factor containing the diffuse element y∗y; that is, M ϕ is a II1 factor. As for the type classification of M , first note that if H = {1} then G ⊂ M ϕ and hence M = M ϕ is a II1 factor. Otherwise, we appeal to the modular spectrum S(M ). Recall the notation established in Subsection 1.1. Since M and M ϕ are factors, S(M ) ∩ R×+ = Γ(σϕ) by Lemma 1.3 and Γ(σϕ) = Sp(σϕ) by Lemma 1.2. Thus, by Connes's classification it suffices to show H = Sp(σϕ). Let λ be an eigenvalue for some non-zero monomial p ∈ ChGi (hence λ ∈ H). Then for any f ∈Tx∈M I(x) we have Hence f (λ) = 0 and therefore λ ∈ Sp(σϕ). Since the Arveson spectrum is closed, this implies H ⊂ Sp(σϕ). If H = R×+, then this forces equality. Otherwise, given µ ∈ R×+ \ H there is an open set U ⊂ R×+ \ H containing µ. Suppose f ∈ A(R×+) satisfies supp(f ) ⊂ U . Then any monomial p ∈ ChGi has an eigenvalue λ 6∈ U and hence f (−t)λit dt · p = f (λ)p. t (p) dt =ZR f (−t)σϕ 0 = σϕ f (p) =ZR Consequently M σϕ σϕ f (p) = f (λ)p = 0. 0 (U ) = {0}, and so µ 6∈ Sp(σϕ) by Lemma 1.1. (cid:3) ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 39 Other than in the type classification of M , the above proof only required that the other generators G \ {y, y∗} were annihilated by δy and δy∗ . Consequently, we have the following corollary. Corollary 7.1. Suppose M is a von Neumann algebra with a faithful normal state ϕ. Let y ∈ M be an eigenoperator of σϕ and B ⊂ M a unital ∗-subalgebra which is globally invariant under σϕ. Letting N denote the von Neumann algebra generated by B and y, if Φ∗ϕ(y, y∗ : B) < ∞ then N ϕN is a II1 factor and N is a factor. We note that Φ∗ϕ(y, y∗ : B) < ∞ in particular implies that y and y∗ are distinct elements which are algebraically free over B. Theorem B. Let M be a von Neumann algebra with a faithful normal state ϕ. Suppose M generated by a finite set G = G∗, G ≥ 2, of eigenoperators of σϕ with finite free Fisher information. Then M ϕ does not have property Γ. Furthermore, if M is a type IIIλ factor, 0 < λ < 1, then M is full. Proof. Pick y ∈ G such that y 6= y∗ (if no such y exists, then M ϕ = M and we simply appeal to the tracial result [12, Theorem 13]). By replacing y with y∗ if necessary, we may assume λy ≤ 1 so that y∗y is diffuse by Corollary 5.10. Letting y = vy be the polar decomposition, it follows that v∗v = 1. Let Y1 = y, Y2 = y∗, δ1 and δ2 as in Lemma 6.3, and δ0 as in Lemma 6.4. We claim that W ∗(Y1, Y2) does not have property Γ and consequently is non-amenable. Suppose (zn)n∈N ⊂ W ∗(Y1, Y2) is a uniformly bounded, asymptotically central sequence. Without loss of generality, we may assume ϕ(zn) = 0 for all n ∈ N. By Lemma 6.4, ¯δ0(zn) = 0 for all n ∈ N and so (16) implies 2 k[zn, U2]k2 HS = Xj=1(cid:16)(cid:10)[zn, Yj], [zn, δ∗j (Uj)](cid:11)ϕ + 2Re (cid:10)(ϕ ⊗ 1 − 1 ⊗ ϕ) ◦ ¯δj(zn), [zn, Yj](cid:11)ϕ(cid:17) . HS → 0. Let p = vv∗ then ϕ + 2hp, zniϕ 2. (17) k[ζα(zn), U2]k2 HS ≤ 2 k[ζα(zn), Yj ]kϕkznk(cid:0)2kδ∗j (Uj)kϕ + 12kδ∗j (1 ⊗ 1)kϕ(cid:1) Xj=1 + Kk[ζα(zn), 1 ⊗ 1]kHSXx∈S k[x, ¯δ0 ◦ ζα(zn)]kHS. Lemma 6.4 implies that W ∗(Y1, Y2) ⊂ ker(¯δ0). So by Lemma 4.11 we have [ζα(zn), Yj] = ζα([zn, Yj]) for j = 1, 2, and [x, ¯δ0◦ζα(zn)] = ¯δ0◦ζα([x, zn]) for all x ∈ S. Thus the above estimate implies limn k[ζα(zn), U2]k2 HS = 0. By [20, Theorem 4.3], {ηα}α>0 converges uniformly on (W ∗(Y1, Y2)′ ∩ (M ϕ)ω)1. By (12), {ζα}α>0 also converges uniformly. Using (17) we therefore have lim n→∞k[zn, U2]k2 HS = lim n→∞ lim α→∞k[ζα(zn), U2]k2 HS = lim α→∞ lim n→∞k[ζα(zn), U2]k2 HS = 0. Therefore HS = kznp +pϕ(p)zn(1 − p)k2 kznk2 Thus by the uniform boundedness of (zn) and Corollary 4.10, we have k[zn, U2]k2 one easily checks that ϕ + kpzn +pϕ(p)(1 − p)znk2 k[zn, U2]k2 (1 − p)!(cid:16)pzn +pϕ(p)(1 − p)zn(cid:17)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) pϕ(p) (1 − p)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) pϕ(p) k[zn, U2]k2 p + HS, p + 1 1 2 2 ϕ there exists S ⊂ W ∗(Y1, Y2) and constant K > 0 such that ϕ =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) kηkHS ≤ KXx∈S k[x, η]kHS ∀η ∈ L2(W ∗(Y1, Y2) ¯⊗W ∗(Y1, Y2)). (18) ¯δ0. For ω a free Let {ζα}α>0 be the maps derived from the contraction resolvent associated to L0 := δ∗0 ultra-filter, let (zn)n∈N ∈ W ∗(Y1, Y2)′ ∩ (M ϕ)ω, which we can assume satisfies ϕ(zn) = 0 for all n ∈ N. Fix α > 0, then for each n ∈ N (16), Corollary 4.10, and (18) imply which tends to zero. Thus W ∗(Y1, Y2) does not have property Γ, and consequently is non-amenable. Now, as W ∗(Y1, Y2) is non-amenable, it admits a non-amenability set (see [13, Definition 2.4]). That is, 40 BRENT NELSON From this and the same computation which showed W ∗(Y1, Y2) does not have property Γ, we have kznk2 ϕ → 0. Thus W ∗(Y1, Y2)′ ∩ (M ϕ)ω = C, and in particular M ϕ does not have property Γ. As discussed in Subsection 1.2, M is full if and only if any uniformly bounded sequence (xn)n∈N of ϕ- centered elements of M in the asymptotic centralizer with respect to φ for all φ ∈ M∗ converges ∗-strongly to zero. Recall that on uniformly bounded subsets of M the ∗-strong topology coincides with the topology defined by the norm kxk# Let (xn)n∈N ⊂ M be a sequence satisfying the above hypothesis. From the proof of Theorem A, we know that if M is of type IIIλ, 0 < λ < 1, then S(M ) = {0} ∪ λZ. So by Lemma 1.3, spectrum(∆ϕ) = {0} ∪ λZ and hence 1 is isolated in the spectrum of ∆ϕ. Thus we may apply [9, Proposition 2.3.(2)] to assert that ϕ = √2kEϕ(xn)kϕ → 0. We will show that (Eϕ(xn))n∈N kxn−Eϕ(xn)k# is asymptotically central and therefore converges to zero in L2(M ϕ, ϕ) since M ϕ does not have property Γ. ϕ → 0. So, it suffices to show kEϕ(xn)k# ϕ := ϕ(x∗x + xx∗)1/2, x ∈ M . Fix z ∈ M ϕ and set an := [Eϕ(xn), z]. Then k[Eϕ(xn), z]k2 ϕ = ϕ(a∗nEϕ(xn)z − a∗nzEϕ(xn)) = ϕ([a∗n,Eϕ(xn)]z) = ϕ([a∗n, xn]z) + ϕ([a∗n,Eϕ(xn) − xn]z). Letting φ(·) := ϕ(·z) ∈ M∗, we see that the first term in the last expression above is bounded by ka∗nkk[xn, φ]k, which tends to zero since (xn)n∈N is uniformly bounded and in the asymptotic centralizer with respect to φ. Thus it suffices to bound the second term: ϕ([a∗n,Eϕ(xn) − xn]z) = ϕ((Eϕ(xn) − xn)[z, a∗n]) ≤ kEϕ(xn) − xnk# ≤ kEϕ(xn) − xnk# ϕ k[z, a∗n]kϕ ϕ 4kEϕ(xn)kkzk2. This tends to zero and completes the proof. (cid:3) We are again grateful to Yoann Dabrowski who suggested the above proof showing that M ϕ does not have property Γ. In particular, this allowed us to replace the hypothesis G ≥ 3 with G ≥ 2. following corollary. As with Corollary 7.1, we used only that G \ {y,y∗} is annihilated by δy and δy∗. Hence we have the Corollary 7.2. Suppose M is a von Neumann algebra with a faithful normal state ϕ. Let y ∈ M be an eigenoperator of σϕ and B ⊂ M a unital ∗-subalgebra which is globally invariant under σϕ. Letting N denote the von Neumann algebra generated by B and y, then N ϕN does not have property Γ, and if 1 is isolated in the spectrum of ∆ϕN then N is full. References [1] William Arveson, On groups of automorphisms of operator algebras, J. Functional Analysis 15 (1974), 217 -- 243. MR0348518 (50 #1016) [2] Haim Brezis, Functional analysis, Sobolev spaces and partial differential equations, Universitext, Springer, New York, 2011. MR2759829 [3] Ian Charlesworth and Dimitri Shlyakhtenko, Free entropy dimension and regularity of non-commutative polynomials, J. Funct. Anal. 271 (2016), no. 8, 2274 -- 2292. MR3539353 [4] Fabio Cipriani, Dirichlet forms and Markovian semigroups on standard forms of von Neumann algebras, J. Funct. Anal. 147 (1997), no. 2, 259 -- 300. MR1454483 (98i:46067) [5] , Dirichlet forms on noncommutative spaces, Quantum potential theory, 2008, pp. 161 -- 276. MR2463708 (2010f:46112) [6] Fabio Cipriani and Jean-Luc Sauvageot, Derivations as square roots of Dirichlet forms, J. Funct. Anal. 201 (2003), no. 1, 78 -- 120. MR1986156 (2004e:46080) [7] Alain Connes, ´Etats presque p´eriodiques sur une alg`ebre de von Neumann, C. R. Acad. Sci. Paris S´er. A-B 274 (1972), A1402 -- A1405. MR0295092 , Une classification des facteurs de type III, Ann. Sci. ´Ecole Norm. Sup. (4) 6 (1973), 133 -- 252. MR0341115 (49 #5865) , Almost periodic states and factors of type III1, J. Functional Analysis 16 (1974), 415 -- 445. MR0358374 (50 [8] [9] #10840) [10] Alain Connes and Dimitri Shlyakhtenko, L2-homology for von Neumann algebras, J. Reine Angew. Math. 586 (2005), 125 -- 168. MR2180603 (2007b:46104) ON FINITE FREE FISHER INFORMATION FOR EIGENVECTORS OF A MODULAR OPERATOR 41 [11] Stephen Curran, Yoann Dabrowski, and Dimitri Shlyakhtenko, Free analysis and planar algebras, ArXiv e-prints (November 2014), available at 1411.0268. [12] Yoann Dabrowski, A note about proving non-Γ under a finite non-microstates free Fisher information assumption, J. Funct. Anal. 258 (2010), no. 11, 3662 -- 3674. MR2606868 (2011d:46135) [13] Yoann Dabrowski and Adrian Ioana, Unbounded derivations, free dilations, and indecomposability results for II1 factors, Trans. Amer. Math. Soc. 368 (2016), no. 7, 4525 -- 4560. MR3456153 [14] Alice Guionnet, Vaughan F. R. Jones, and Dimitri Shlyakhtenko, Random matrices, free probability, planar algebras and subfactors, Quanta of maths, 2010, pp. 201 -- 239. MR2732052 (2012g:46094) [15] Michael Hartglass, Free product C*-algebras associated to graphs, free differentials, and laws of loops, Canad. J. Math. (2016). [16] Zhi Ming Ma and Michael Rockner, Introduction to the theory of (nonsymmetric) Dirichlet forms, Universitext, Springer- Verlag, Berlin, 1992. MR1214375 (94d:60119) [17] Tobias Mai, Roland Speicher, and Moritz Weber, Absence of algebraic relations and of zero divisors under the assumption of full non-microstates free entropy dimension, ArXiv e-prints (February 2015), available at 1502.06357. [18] Brent Nelson, Free monotone transport without a trace, Comm. Math. Phys. 334 (2015), no. 3, 1245 -- 1298. MR3312436 [19] , Free transport for finite depth subfactor planar algebras, J. Funct. Anal. 268 (2015), no. 9, 2586 -- 2620. MR3325530 [20] Jesse Peterson, L2-rigidity in von Neumann algebras, Invent. Math. 175 (2009), no. 2, 417 -- 433. MR2470111 (2010b:46128) [21] Dimitri Shlyakhtenko, Free quasi-free states, Pacific J. Math. 177 (1997), no. 2, 329 -- 368. MR1444786 (98b:46086) [22] , Free Fisher information with respect to a completely positive map and cost of equivalence relations, Comm. Math. Phys. 218 (2001), no. 1, 133 -- 152. MR1824202 (2002e:46077) [23] , Free Fisher information for non-tracial states, Pacific J. Math. 211 (2003), no. 2, 375 -- 390. MR2015742 (2005a:46139) [24] Dimitri Shlyakhtenko and Paul Skoufranis, Freely independent random variables with non-atomic distributions, Trans. Amer. Math. Soc. 367 (2015), no. 9, 6267 -- 6291. MR3356937 [25] M. Takesaki, Theory of operator algebras. I, Encyclopaedia of Mathematical Sciences, vol. 124, Springer-Verlag, Berlin, 2002. Reprint of the first (1979) edition, Operator Algebras and Non-commutative Geometry, 5. MR1873025 (2002m:46083) [26] Masamichi Takesaki, Theory of operator algebras. II, Encyclopaedia of Mathematical Sciences, vol. 125, Springer-Verlag, Berlin, 2003. Operator Algebras and Non-commutative Geometry, 6. MR1943006 (2004g:46079) [27] Dan-Virgil Voiculescu, The analogues of entropy and of Fisher's information measure in free probability theory. I, Comm. Math. Phys. 155 (1993), no. 1, 71 -- 92. MR1228526 [28] [29] , The analogues of entropy and of Fisher's information measure in free probability theory. II, Invent. Math. 118 (1994), no. 3, 411 -- 440. MR1296352 (96a:46117) , The analogues of entropy and of Fisher's information measure in free probability theory. V. Noncommutative Hilbert transforms, Invent. Math. 132 (1998), no. 1, 189 -- 227. MR1618636 (99d:46087) Department of Mathematics, University of California, Berkeley E-mail address: [email protected]
1110.6858	1	1110	2011-10-31T16:45:25	The second local multiplier algebra of a separable C*-algebra	[ "math.OA" ]	Several examples of (separable) C-algebras with the property that their second (iterated) local multiplier algebra is strictly larger than the first have been found by various groups of authors over the past few years, thus answering a question originally posed by G. K. Pedersen in 1978. This survey discusses a systematic approach by P. Ara and the author to produce such examples on the one hand; on the other hand, we present new criteria guaranteeing that the second and the first local multiplier algebra of a separable C-algebra agree. For this class of C*-algebras, each derivation of the local multiplier algebra is inner.	math.OA	math	THE SECOND LOCAL MULTIPLIER ALGEBRA OF A SEPARABLE C-ALGEBRA MARTIN MATHIEU This paper is dedicated to Victor Shulman on his 65th birthday. Abstract. Several examples of (separable) C-algebras with the property that their second (iterated) local multiplier algebra is strictly larger than the first have been found by various groups of authors over the past few years, thus answering a question originally posed by G. K. Pedersen in 1978. This survey discusses a systematic ap- proach by P. Ara and the author to produce such examples on the one hand; on the other hand, we present new criteria guaranteeing that the second and the first local multiplier algebra of a separable C-algebra agree. For this class of C-algebras, each derivation of the local multiplier algebra is inner. 1 1 0 2 t c O 1 3 ] . A O h t a m [ 1 v 8 5 8 6 . 0 1 1 1 : v i X r a 1. Introduction In 1978 the late G. K. Pedersen introduced an algebra extension of a general C-algebra A which he called the "C-algebra of essential multipliers". His interest in this C- algebra was stipulated by possible applications in operator theory on C-algebras as was the interest of the authors of the few papers that appeared on this topic around the same time; see [15] and [16]. Having lain dormant for a good while, this theme was taken up again by P. Ara and the present author who re-discovered the same C- algebra independently of Pedersen's work in the late 1980's and coined the terminology "local multiplier algebra", introducing, accordingly, the notation Mloc(A). An important connection with the symmetric algebra of quotients in noncommutative ring theory was subsequently made, which led to a fuller understanding of the structure of Mloc(A). For a discussion of this interplay, see [19]. A comprehensive account can be found in our monograph [2], which also contains a wealth of applications to a number of classes of operators between C-algebras, thus continuing Pedersen's ideas. In his seminal paper [21], Pedersen asked two questions which spurred quite some research in the past decades. Iterating the construction of the local multiplier algebra (see the definition in the subsequent section) one obtains the following tower of C- algebras which, a priori, does not have a largest element. A ⊆ Mloc(A) ⊆ Mloc(Mloc(A)) ⊆ . . . (1.1) This led Pedersen to ask Question 1. Is Mloc(Mloc(A)) = Mloc(A) for every C-algebra A? The main result in [21] is the following. 2000 Mathematics Subject Classification. Primary 46L05. Secondary 46L06, 46M20. Key words and phrases. Local multiplier algebra, injective envelope, C-algebra, sheaf theory. 1 2 MARTIN MATHIEU Theorem 1.1. Let A be a separable C-algebra. Every derivation d : A → A extends uniquely to a derivation d : Mloc(A) → Mloc(A) and there is y ∈ Mloc(A) such that d = ad y (that is, dx = [x, y] = xy − yx for all x ∈ Mloc(A)). A derivation of Mloc(A) may or may not leave A invariant; thus Pedersen's second question reads: Question 2. Is every derivation d : Mloc(A) → Mloc(A) inner in Mloc(A), provided A is a separable C-algebra? It appears that the answer to Question 2 is still not known in full generality. However, a negative answer to Question 1 was provided in [3]. Since then, several other examples of C-algebras with the property that the second local multiplier algebra is strictly larger than the first have surfaced [4], [9]. Some of them are separable C-algebras, others are not. On the other hand, Somerset, in [24], gave fairly general conditions on a separable C-algebra A implying that both Questions 1 and 2 have positive answers. Until recently, it was not understood how these two different directions could fit into a more comprehensive framework. This is now achieved in our paper [6], and the aim of the present survey is to explain the historical development that led to this more detailed analysis of the differences between the first and the second local multiplier algebra of a separable C-algebra. 2. The History Let I be a closed, two-sided ideal of a C-algebra A which is essential (that is, aI = 0 for some a ∈ A implies that a = 0). If J is another such essential ideal of A which is contained in I, then the multiplier algebra M(I) is canonically embedded as a C- subalgebra into M(J) by restriction of multipliers to the smaller ideal. In this way, we obtain a directed system of C-algebras with isometric connecting morphisms, where I runs through the directed set Ice(A) of all closed, two-sided, essential ideals of A. The −→ Ice(A)M(I), the local multiplier algebra of A. direct limit of this system is Mloc(A) = lim There are several other very useful descriptions of the local multiplier algebra, which are all discussed in detail in our monograph [2]. These lead, e.g., to a representation of −→ Ice(A)Z(M(I)) [2, Section 3.1] which the centre Z = Z(Mloc(A)) of Mloc(A) as Z = lim cannot be deduced directly from the defining formula for Mloc(A) above. Another important characterisation of Mloc(A), which we will rely on heavily in the following, was first obtained by Frank and Paulsen in [17]; see also [4, Section 4.3]. For a C-algebra A, let us denote by I(A) its injective envelope as introduced by Hamana in [18]; see also [20]. We emphasise that I(A) is not an injective object in the category of C-algebras and -homomorphisms but in the category of operator spaces and complete contractions. However, it turns out that, nevertheless, I(A) is a C-algebra canonically containing A as a C-subalgebra. For a concise discussion of these facts suited for our purposes, see [4]. Under this embedding of A into I(A), the local multiplier algebra Mloc(A) is nothing but the completion of the set of all y ∈ I(A) which act as a multiplier on some I ∈ Ice(A). Since I(Mloc(A)) = I(A) [4, Proposition 2.14], we see that all higher local THE SECOND LOCAL MULTIPLIER ALGEBRA OF A SEPARABLE C -ALGEBRA 3 multiplier algebras of A are contained in I(A) wherefore (1.1) improves to A ⊆ Mloc(A) ⊆ Mloc(Mloc(A)) ⊆ . . . ⊆ I(A). (2.1) As a result, we can use the injective envelope to study Question 1. For instance, if A is commutative, then Mloc(A) is a commutative AW-algebra, hence injective. It follows that Mloc(A) = I(A) and, since the local multiplier algebra of every AW-algebra is the algebra itself [2, Theorem 2.3.8], we find that Mloc(Mloc(A)) = Mloc(I(A)) = I(A) = Mloc(A), one of the possible ways to affirm Pedersen's question in the commutative case. Evidently, Mloc(A) = M(A) for each simple C-algebra A. Since Mloc(M(A)) = Mloc(A), it follows that Mloc(Mloc(A)) = Mloc(A) in this case too. (We note in passing that Mloc(A) itself can be simple without A being simple unital, see [1], in which case we also have Mloc(Mloc(A)) = Mloc(A).) On the basis of these two positive answers, it appears to be close at hand to investigate a C-algebra A which is the tensor product of a commutative and a simple one: A = C ⊗ B with C commutative (and, without loss of generality, unital) and B simple. The surprise comes in Section 3 below where we shall pin down the properties of C and B determining whether the answer to Question 1 is negative or positive. Another important contribution is due to Somerset who proved in [24] that the answer to both Question 1 and Question 2 is positive for every unital, separable C-algebra A which contains sufficiently many maximal ideals; to be precise, he assumed that the primitive ideal space Prim(A) contains a dense Gδ subset consisting of closed points. This topological condition will feature again in Section 3 below. The first class of examples of C-algebras for which Question 1 has a negative answer was found in [3]. These are certain unital separable approximately finite-dimensional C-algebras which are primitive and (necessarily) anti-liminal. We employed non-stable K-theory to describe these C-algebras. A very different method, the theory of Hilbert modules over commutative AW-algebras, was applied in [4] to prove that algebras of the form C(X) ⊗ B(H), where H is a separable Hilbert space and X is a Stonean space with additional properties, also provide a negative answer to Question 1. From this it is easy to see that some separable, liminal C-algebras such as C[0, 1] ⊗ K(H) have the same property. The latter example was independently found by Argerami, Farenick and Massey in [9]. Both approaches make use of the injective envelope as well as of formulas for Mloc(A) and I(A) which are (fortunately) available in these special cases. The same three authors recently studied the local multiplier algebra of certain continuous trace C-algebras with similar techniques [10], [11] leading to the general result that the second local multiplier algebra of those C-algebras is injective [11, Theorem 6.7]. In this section, we shall explain how the following dichotomy arises. Let X be a compact metric space which is perfect (that is, contains no isolated points). Let B be any 3. The Results 4 MARTIN MATHIEU separable simple C-algebra. Then we have the following alternative A = C(X) ⊗ B B unital rrrrrrrrrrrrrrrrrrrrrr HHHHHHHHHHHHHHHHHHHH B non-unital Mloc(Mloc(A)) = Mloc(A) Mloc(Mloc(A)) 6= Mloc(A) Consequently, there is a plethora of examples of C-algebras for which Question 1 has a negative answer! The first of the two main results below explains the non-unital case. Theorem 3.1 ([6, Corollary 3.8]). Let B and C be separable C-algebras and suppose that at least one of them is nuclear. Suppose further that B is simple and non-unital and that Prim(C) contains a dense Gδ subset consisting of closed points. Let A = C ⊗ B. Then Mloc(A) = Mloc(Mloc(A)) if and only if Prim(A) contains a dense subset of isolated points. The sufficient condition in this theorem applies to any C-algebra A: Let X = Prim(A), X1 the set of isolated points in X and X2 = X \ X1. Then X1 and X2 are disjoint open subsets of X with corresponding orthogonal closed ideals I1 = A(X1) and I2 = A(X2) of A. If X1 is dense, I1 is the minimal essential closed ideal of A so Mloc(A) = M(I1). It follows that Mloc(Mloc(A)) = Mloc(M(I1)) = Mloc(I1) = Mloc(A). It is thus surprising that this very general condition is indeed necessary for C-algebras of the above type -- and therefore provides us with an easy, systematic way of producing counterexamples. Note that, since B or C is nuclear, the tensor product is unambiguous and Prim(C ⊗ B) = Prim(C) × Prim(B) [23, Theorem B.45]. We shall indicate in Section 4 how to obtain Theorem 3.1. The unital case in the above alternative fits into the following more general result which applies to C(X) ⊗ B since this C-algebra is quasicentral if and only if B is unital (the following formula for the centres applies: Z(C ⊗ B) = Z(C) ⊗ Z(B) [8, Theorem 3]). Theorem 3.2 ([6, Theorem 4.7]). Let A be a quasicentral separable C-algebra such that Prim(A) contains a dense Gδ subset consisting of closed points. If B is a C-subalgebra of Mloc(A) containing A then Mloc(B) ⊆ Mloc(A). In particular, Mloc(Mloc(A)) = Mloc(A). Therefore, for every unital separable C-algebra with Hausdorff primitive spectrum, the first and the second local multiplier algebras agree; this was already observed in [24]. Recall that A is said to be quasicentral if no primitive ideal of A contains the centre Z(A) of A. This class of C-algebras was introduced and studied initially by Delaroche in the late 1960's [12], [13] and has turned out to be useful on various occasions. Another "classical" notion that arises when dealing with topological spaces associated to a C- algebra is the one of a separated point. A point t in a topological space X is called THE SECOND LOCAL MULTIPLIER ALGEBRA OF A SEPARABLE C -ALGEBRA 5 separated if t and every point t′ outside the closure of {t} can be separated by disjoint open neighbourhoods. The set Sep(A) of all separated points in a separable C-algebra A is a dense Gδ subset of Prim(A) and consists precisely of those t ∈ Prim(A) for which the norm function t 7→ ka + tk is continuous for every a ∈ A [14]. We shall soon make good use of Sep(A) when we outline the arguments for Theorem 3.2 in Section 4 below. But first let us draw an immediate consequence of Theorem 3.2 for Question 2. Corollary 3.3 ([6, Corollary 4.9]). Let A be a quasicentral separable C-algebra such that Prim(A) contains a dense Gδ subset consisting of closed points. Then every deriva- tion of Mloc(A) is inner. The argument is in fact a reduction to Pedersen's theorem, Theorem 1.1 above. Starting with a derivation d : Mloc(A) → Mloc(A) one can construct a separable C-subalgebra B of Mloc(A) containing A which is d-invariant. By Theorem 1.1, d is inner in Mloc(B) which, however, is contained in Mloc(A) by Theorem 3.2. For the details of the proof, see [6]. Note that the same argument applies to every separable C-algebra A with the property that any separable C-subalgebra B of Mloc(A) that contains A has its local multiplier algebra Mloc(B) contained in Mloc(A). Corollary 3.3 was obtained in the unital case in [24]. 4. Proofs We will first attend to the proof of Theorem 3.2. The key tool are two formulas, valid for an arbitrary C-algebra A, that describe both Mloc(A) and I(A) in a compatible way. They rest on our sheaf theory for C-algebras which was developed in [5]. Rather than giving the formal definition of a sheaf of a C-algebra [5, Definition 3.1], It is straightforward that both we shall content ourselves here with two examples. constitute contravariant functors from the category OPrim(A) of open subsets of the primitive ideal space of the C-algebra A with inclusions as arrows into the category C ∗ 1 of unital C-algebras with unital -homomorphisms as arrows, that is, presheaves of unital C-algebras over the base space OPrim(A). It takes a bit more work to verify the unique gluing property for the multiplier sheaf [5, Proposition 3.4] whereas it is fairly easy to establish this in the case of the injective envelope sheaf. Example 4.1 (The multiplier sheaf). Let A be a C-algebra with primitive ideal space Prim(A). We define MA : OPrim(A) → C ∗ 1 , MA(U) = M(A(U)), where M(A(U)) denotes the multiplier algebra of the closed ideal A(U) of A associated to the open subset U ⊆ Prim(A) and M(A(U)) → M(A(V )), V ⊆ U, are the restriction homomorphisms. This is the multiplier sheaf of A over OPrim(A). Example 4.2 (The injective envelope sheaf). Let I(B) denote the injective envelope of a C-algebra B. We define IA : OPrim(A) → C ∗ 1 , IA(U) = pU I(A) = I(A(U)), 6 MARTIN MATHIEU where pU denotes the unique central open projection in I(A) such that pU I(A) is the injective envelope of A(U). The mappings I(A(U)) → I(A(V )), V ⊆ U, are given by multiplication by pV . This is the injective envelope sheaf of A over OPrim(A). To every presheaf we can associate in a canonical way an upper semicontinuous C-bundle [5, Theorem 5.6]. Let A be a presheaf of C-algebras over the topologi- A(U) as the direct cal space X. For t ∈ X, we define the stalk at t by At := lim limit of C-algebras of the directed family {A(U)}, where U ranges over the family of At and let π(a) = t if a ∈ At. For s ∈ A(U) and t ∈ U, we have a canonical -homomorphism A(U) → At and we denote by s(t) the image of s under this mapping. There is a canonical topology on the total space A (which is uniquely determined in case X is Hausdorff) turning (A, π, X) into an upper semicontinuous C-bundle over X. all open neighbourhoods of t in X. Take A := Ft∈X −→t∈U Let us denote by AMA and AIA, respectively, the C-bundles associated in this way to MA and IA, respectively. For any upper semicontinuous C-bundle (A, π, X) and a subset Y ⊆ X, we shall write Γb(Y, A) for the C-algebra of bounded continuous local sections on Y [5, Lemma 5.2]. Moreover, T stands for the downwards directed family of dense Gδ subsets of X. The following two formulas are the main results, Theorem 7.6 and Theorem 7.7, in [5]. They require the additional concept of the derived sheaf of a sheaf of C-algebras [5, Proposition 7.4]: Mloc(A) = alg lim −→ T ∈T Γb(T, AMA) I(A) = alg lim −→ T ∈T Γb(T, AIA). (4.1) The algebraic direct limit, that is, the limit before completion to a C-algebra, is already complete since Prim(A) is a Baire space and hence countable intersections of dense Gδ subsets remain within T . As AMA is a sub-bundle of AIA, the above formulas (4.1) open up a new way to compare the second local multiplier algebra with the first. Indeed, if we take y ∈ Mloc(Mloc(A)) ⊆ I(A), by (4.1), y is contained in some C-subalgebra Γb(T, AIA) and will belong to Mloc(A) once we find T ′ ⊆ T , T ′ ∈ T such that y ∈ Γb(T ′, AMA). We shall now outline how to use this approach to prove Theorem 3.2 for the case B = Mloc(A) (the general case only requires small modifications). Take y ∈ M(J) for some closed essential ideal J of Mloc(A). There is T ∈ T such that y ∈ Γb(T, AIA). By our assumption on Prim(A) and the remarks before Corollary 3.3 we can assume, without restricting the generality, that T consists of closed separated points of Prim(A). By the results in [6], there is an element h ∈ J such that h(t) 6= 0 for all t ∈ T when viewed as a section on T , and there is a separable C-subalgebra B ⊆ J with the properties AhA ⊆ B and y ∈ M(B). Now take a countable dense subset {bn n ∈ N} in B and Tn ∈ T such that bn ∈ Γb(Tn, AMA) for each n ∈ N. In order to simplify the notation, let us put A = AMA for the rest of the proof. Setting T ′ = Tn Tn ∩ T ∈ T , we have B ⊆ Γb(T ′, A) and hence Bt = {b(t) b ∈ B} ⊆ At (t ∈ T ′). THE SECOND LOCAL MULTIPLIER ALGEBRA OF A SEPARABLE C* -ALGEBRA 7 In general, there is In a next step, we aim to describe the fibres At in more detail. a homomorphism ϕt : At → Mloc(A/t) [5, Section 6]; however, this need neither be injective nor surjective. From our hypotheses, we can conclude for each t ∈ T ′: A quasicentral ⇒ A/t unital t closed ⇒ A/t simple (cid:27) ⇒ Mloc(A/t) = A/t and, moreover, that ϕt is an isomorphism. The surjectivity of ϕt rests on the existence of local identities in quasicentral C-algebras: ∀ t ∈ Prim(A) ∃ U1 ⊆ Prim(A) open, t ∈ U1, ∃ z ∈ Z(A)+, kzk = 1 : z + A(U2) = 1A/A(U2), where U2 = Prim(A) \ U1 [6, Lemma 4.3]. As a result, At is a simple, unital C-algebra for each t ∈ T ′ and thus, At = At h(t) At = (A/t)h(t)(A/t) = (AhA)t ⊆ Bt ⊆ At (t ∈ T ′). Taking bt ∈ B with bt(t) = 1At we obtain y(t) = y(t) 1At = (ybt)(t) ∈ At for all t ∈ T ′. This entails that y ∈ Γb(T ′, AMA) with T ′ ⊆ T , proving that y ∈ Mloc(A), as desired. In the proof of the "only if"-part in Theorem 3.1, the sheaf theoretic concepts can be pushed to the background, and we will merely roughly indicate the main argument in the following. It is easy to reduce the general case to the case when Prim(A) is perfect (i.e., has no isolated points) using the direct sum decomposition Mloc(A) = Mloc(I1) ⊕ Mloc(I2) (in the notation of the argument for the "if"-part directly after the statement of Theorem 3.1 above). The strategy now is the following: to show that Mloc(Mloc(A)) 6= Mloc(A) under the hypothesis that Prim(A) is perfect (and all the other assumptions in Theorem 3.1) we aim to identify a closed essential ideal K of Mloc(A) with the property that Mloc(A) M(K). The following C-subalgebra KA of Mloc(A) was already introduced by Somerset in [24]: KA is the closure of the set of all elements of the form Pn∈N anzn, where {an} ⊆ A is a bounded family and {zn} ⊆ Z(Mloc(A)) consists of mutually orthogonal projections. (These infinite sums exist in Mloc(A) by [2, Lemma 3.3.6], for example. Note also that Z(Mloc(A)) is countably decomposable since A is separable.) For a separable C-algebra A with the property that Prim(A) contains a dense Gδ subset of closed points, the following statements hold (see [6, Section 2]). (i) KA is an essential ideal in Mloc(A). (ii) If KI = KA for all I ∈ Ice(A) then Mloc(KA) = M(KA). (iii) Let y ∈ I(A). If ya ∈ KA for all a ∈ A then y ∈ M(KA). Combining these we obtain the following result. Proposition 4.3 ([6, Theorem 3.2]). M (3) loc (A) = M (2) loc (A) = M(KA), where M (n) algebra of A. loc (A) = Mloc(cid:0)M (n−1) loc (A)(cid:1), n ≥ 2 denotes the n-fold iterated local multiplier The main work in the proof of Theorem 3.1 consists in constructing an element q ∈ M(KA) \ Mloc(A). This uses the special tensor product structure A = C ⊗ B. The topological assumptions on X = Prim(A) = Prim(C) lead us to a dense Gδ subset S of X consisting of closed separated points which is a Polish space (using [14]). Since S itself 8 MARTIN MATHIEU is perfect, every non-empty open subset of S contains an open subset which has non- empty boundary. This allows us to choose a suitable sequence (zn)n∈N of projections in Z(M(Kn ⊗ B)) ⊆ Mloc(A), where Kn ∈ Ice(C). As B is simple and non-unital, there is a strictly increasing approximate identity (en)n∈N of B with enen+1 = en and j=1 zj ⊗p2j, n ∈ N. We thus obtain an increasing sequence (qn)n∈N in Mloc(A)+ bounded by 1. Since n=1 zn⊗p2n exists in I(A)+ and has norm 1. ken+1−enk = 1 for all n. Put p1 = e1, pn = en−en−1 for n ≥ 2 and set qn = Pn I(A) is monotone complete, q = supn qn = P∞ It remains to show (a) q ∈ M(KA); (b) q /∈ Mloc(A). The first assertion is established by using that the bounded central closure cA = AZ is σ-unital (as A is separable) together with approximation in the strict topology of M(KA). For (b), we use special properties of the sequence (zn)n∈N chosen above relying on the topological properties of S indicated above. It is shown that, if q were in Mloc(A), this would lead to a contradiction. This completes the proof of Theorem 3.1. It is clear from Proposition 4.3, which holds for every separable C-algebra A with the property that Prim(A) contains a dense Gδ subset of closed points, that the methods discussed above cannot create an example of a C-algebra A with the property that M (2) loc (A); in fact, it appears that no such concrete example is known at present. loc (A) 6= M (3) References [1] P. Ara and M. Mathieu, A simple local multiplier algebra, Math. Proc. Cambridge Phil. Soc. 126 (1999), 555 -- 564. [2] P. Ara and M. Mathieu, Local multipliers of C-algebras, Springer-Verlag, London, 2003. [3] P. Ara and M. Mathieu, A not so simple local multiplier algebra, J. Funct. Anal. 237 (2006), 721 -- 737. [4] P. Ara and M. Mathieu, Maximal C-algebras of quotients and injective envelopes of C- algebras, Houston J. Math. 34 (2008), 827 -- 872. [5] P. Ara and M. Mathieu, Sheaves of C-algebras, Math. Nachr. 283 (2010), 21 -- 39. [6] P. Ara and M. Mathieu, When is the second local multiplier algebra of a C-algebra equal to the first?, Bull. London Math. Soc. (2011), in press. [7] R. J. Archbold, Density theorems for the centre of a C-algebra, J. London Math. Soc. (2) 10 (1975), 189 -- 197. [8] R. J. Archbold, On the centre of a tensor product of C-algebras, J. London Math. Soc. (2) 10 (1975), 257 -- 262. [9] M. Argerami, D. R. Farenick and P. Massey, The gap between local multiplier algebras of C-algebras, Quart. J. Math. 60 (2009), 273 -- 281. [10] M. Argerami, D. R. Farenick and P. Massey, Injective envelopes and local multiplier algebras of some spatial continuous trace C-algebras, Quart. J. Math., to appear. [11] M. Argerami, D. R. Farenick and P. Massey, Injective envelopes and local multiplier algebras of some spatial continuous trace C-algebras, II, preprint, February 2011; arXiv:1102.4869. [12] C. Delaroche, Sur les centres des C-alg`ebres, Bull. Sc. math. 91 (1967), 105 -- 112. [13] C. Delaroche, Sur les centres des C-alg`ebres, II, Bull. Sc. math. 92 (1968), 111 -- 128. [14] J. Dixmier, Sur les espaces localement quasi-compacts, Canad. J. Math. 20 (1968), 1093 -- 1100. [15] G. A. Elliott, Derivations determined by multipliers on ideals of a C-algebra, Publ. Res. Inst. Math. Sci. 10 (1974/75), 721 -- 728. THE SECOND LOCAL MULTIPLIER ALGEBRA OF A SEPARABLE C -ALGEBRA 9 [16] G. A. Elliott, Automorphisms determined by multipliers on ideals of a C-algebra, J. Funct. Anal. 23 (1976), 1 -- 10. [17] M. Frank and V. I. Paulsen, Injective envelopes of C-algebras as operator modules, Pacific J. Math. 212 (2003), 57 -- 69. [18] M. Hamana, Injective envelopes of C-algebras, J. Math. Soc. Japan 31 (1979), 181 -- 197. [19] M. Mathieu, The local multiplier algebra: blending noncommutative ring theory and functional analysis, in: Proc. Conf. in honour of Robert Wisbauer's 65th birthday, (Porto, 8 -- 10 September 2006), Trends in Math., Birkhauser, Basel, 2008, 301 -- 312. [20] V. I. Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in Adv. Math. 78, Cambridge Univ. Press, Cambridge, 2002. [21] G. K. Pedersen, Approximating derivations on ideals of C-algebras, Invent. Math. 45 (1978), 299 -- 305. [22] G. K. Pedersen, C-algebras and their automorphism groups, Academic Press, London, 1979. [23] I. Raeburn and D. P. Williams, Morita equivalence and continuous-trace C-algebras, Math. Surveys and Monographs 60, Amer. Math. Soc., Providence, RI, 1998. [24] D. W. B. Somerset, The local multiplier algebra of a C*-algebra. II, J. Funct. Anal. 171 (2000), 308 -- 330. Department of Pure Mathematics, Queen's University Belfast, Belfast BT7 1NN, Northern Ireland E-mail address: [email protected]
1008.1003	3	1008	2012-05-02T10:24:11	The Nakayama automorphism of the almost Calabi-Yau algebras associated to SU(3) modular invariants	[ "math.OA", "hep-th", "math-ph", "math.AG", "math-ph" ]	We determine the Nakayama automorphism of the almost Calabi-Yau algebra A associated to the braided subfactors or nimrep graphs associated to each SU(3) modular invariant. We use this to determine a resolution of A as an A-A bimodule, which will yield a projective resolution of A.	math.OA	math	The Nakayama automorphism of the almost Calabi-Yau algebras associated to SU (3) modular invariants David E. Evans and Mathew Pugh School of Mathematics, Cardiff University, Senghennydd Road, Cardiff, CF24 4AG, Wales, U.K. November 11, 2018 Abstract We determine the Nakayama automorphism of the almost Calabi-Yau algebra A associated to the braided subfactors or nimrep graphs associated to each SU (3) modular invariant. We use this to determine a resolution of A as an A-A bimodule, which will yield a projective resolution of A. 1 Introduction The SU(2) and SU(3) modular invariant partition functions were classified by Cappelli, Itzykson and Zuber [15] and Gannon [35] respectively. The SU(2) theory is closely related to the preprojective algebras of Coxeter-Dynkin quivers. The object of this paper is to study the analogous finite dimensional superpotential algebras associated to the SU(3) invariants. The classical McKay correspondence relates finite subgroups Γ of SU(2) with the algebraic geometry of the quotient Kleinian singularities C2/Γ [63] but also with the classification of SU(2) modular invariants [15, 73], classification of subfactors of index less than 4 [57, 42, 50, 2, 43], and quantum subgroups of SU(2) [59, 60, 70, 5, 6, 10, 11]. The study of quotient singularities and their resolution has been assisted with the study of the structure of certain noncommutative algebras. Minimal resolutions of Kleinian singularities can be described via the moduli space of representations of the preprojective algebra associated to the action of Γ. This leads to general programme to understand singularities via a noncommutative algebra A, often called a noncommutative resolution, whose centre corresponds to the coordinate ring of the singularity [66]. The algebra should be finitely generated over its centre, and the desired favourable resolution is the moduli 1 space of representations of A, whose category of finitely generated modules is derived equivalent to the category of coherent sheaves of the resolution. In the case of a quotient singularity C3/Γ for a finite subgroup Γ of SU(3), the corresponding noncommutative algebra A is a Calabi-Yau algebra of dimension 3. Calabi-Yau algebras arise naturally in the study of Calabi-Yau manifolds, providing a noncommutative version of conventional Calabi-Yau geometry. An algebra A is Calabi- Yau of dimension n if the bounded derived category of the abelian category of finite dimensional A-modules is a Calabi-Yau category of dimension n. In this case the global dimension of A is n [12]. The derived category of coherent sheaves over an n-dimensional Calabi-Yau manifold is a Calabi-Yau category of dimension n and they appear naturally in the study of boundary conditions of the B-model in superstring theory over the manifold. For more on Calabi-Yau algebras, see e.g. [12, 38]. In [38, Remark 4.5.7] Ginzburg introduced, in his terminology, q-deformed Calabi- Yau algebras. In the case where q is not a root of unity, these algebras are Calabi-Yau algebras of dimension 3. In this paper we will study these algebras in the case where q is a root of unity, which are the SU(3) generalizations of preprojective algebras [37] for the Coxeter-Dynkin diagrams ADE. We will call these algebras almost Calabi-Yau algebras. In Section 2 we bring together the strands needed from subfactor theory, modular ten- sor categories and their modules, planar algebras and categorical approaches. We begin in Section 2.1 by describing our notation for the representation theory of SU(n) at level k. In Section 2.2 we recall the generalized Temperley-Lieb algebras for SU(n), which are representations of the Hecke algebra. Then in Section 2.3 we review the description of the Verlinde algebra and fusion rules for quantum SU(n) in terms of endomorphisms of a hyperfinite type III1 factor N. This system of endomorphisms has the structure of a modular tensor category. For a braided inclusion N ⊂ M we obtain a module category which produces a nimrep, or non-negative matrix integer representation of the fusion rules, which is associated to the SU(n) modular invariants. In Section 2.4 we give a subfactor description of the Verlinde algebra, where the space of intertwiners is identified with the algebra of paths on the nimrep graph. In Section 2.5 we move to the categorical picture, and give a diagrammatic and categorical description of the Verlinde algebra and fusion rules for quantum SU(2). This is based on the diagrammatic representation of the Temperley-Lieb algebra of Kauffman [49]. A categorical approach to Temperley-Lieb algebras was given in [65, 71, 16]. In Section 2.6 we describe SU(2) module categories in terms of preprojective algebras, following the ideas of Cooper [16]. We relate this in Section 2.7 to braided subfactors using the Goodman-de la Harpe-Jones construction [39] and its manifestation in the bipartite graph planar algebra construction of Jones [47]. We give the analogous description for SU(3), where the diagrammatic and categorical descrip- tion of Section 2.8 is based on the SU(3)-Temperley-Lieb algebra and its diagrammatic representation constructed in [27] using the A2-webs of Kuperberg [53]. In Section 2.9 we describe the module categories in terms of (almost) Calabi-Yau algebras, following the ideas of Cooper [16]. The corresponding nimreps, the SU(3) ADE graphs, are the graphs associated to the SU(3) modular invariants. In Section 2.10 we relate the construction of these almost Calabi-Yau algebras to braided subfactors using the SU(3) Goodman-de la Harpe-Jones construction [26] and its manifestation in the SU(3)-graph planar algebra construction [28]. Then in Section 3 we compute the Hilbert series of dimensions associated to these 2 almost Calabi-Yau algebras. The McKay graph of SU(3) is built out of closed paths of length 3, which corresponds to the fact that the fundamental representation ρ of SU(3) satisfies ρ ⊗ ρ ⊗ ρ ∋ 1. One can build an Ocneanu cell system W on the McKay graph of a subgroup of SU(3) or an ADE graph G, which attaches a complex number to each closed path of length three on the edges of G [59]. These yield relations on paths of equal length, and one obtains a superpotential algebra A = A(G, W ) by taking the quotient by the ideal generated by these relations. For the ADE graphs G, we take potentials built on the cell system W computed in [25], and study the Hilbert series of dimensions of the corresponding quotient algebras A(G, W ), which are almost Calabi-Yau algebras. The preprojective algebras for SU(2) braided subfactors, and the almost Calabi-Yau algebras for SU(3) braided subfactors, are given by the image under a functor F of the quotient S of the tensor algebra generated by the fundamental sector by a tensor ideal which makes S symmetric, that is, F (S) = A(G, W ), (1) where the functor F is essentially the module category arising from the braided subfactor. If Hn is the matrix of dimensions of paths of length n in some graph G in the quotient algebra A = A(G, W ), with the indices of the matrix labeled by the vertices, then the the adjacency matrix of G. Then if G is the McKay graph of a subgroup of SU(3) then A = A(G, W ) is a Calabi-Yau algebra of dimension 3 [38, Theorem 4.4.6], and by [12, Theorem 4.6] its Hilbert series is given by: matrix Hilbert series HA of the algebra A is defined as HA(t) = P Hntn. Let ∆G be HA(t) = 1 1 − ∆Gt + ∆T G t2 − t3 . (2) We prove in Theorem 3.1 that if G is a finite SU(3) ADE graph which carries a cell system W , and thus yields a braided subfactor [26], then HA(t) = 1 − P th 1 − ∆Gt + ∆T G t2 − t3 , (3) where P is a permutation matrix corresponding to a Z3 symmetry of the graph, and h = k + 3 is the Coxeter number of G, where k is the level of SU(3). The permutation matrix P corresponds precisely to the Nakayama permutation for A. This result was mentioned without proof in [29]. In [13, Proposition 3.14] the Hilbert series was given for a (p, q)-Koszul algebra (or almost Koszul algebra), of which (3) is a particular case, where A = A(G, W ) is a (h − 3, 3)-Koszul algebra, see Section 2.9. The dual A∗ = Hom(A, C) is an A-A bimodule, not usually identified with AAA or 1A1 with standard right and left actions but with 1Aβ with standard left action and the standard right action twisted by an automorphism β, the Nakayama automorphism [72]. The Nakayama automorphism measures how far away A is from being symmetric. In the case of a preprojective algebra Π of a Dynkin quiver, this Nakayama automorphism is identified (up to a sign) with an involution on the underlying Dynkin diagram, which is trivial in all cases, except for the Dynkin diagrams An, D2n+1, E6 where it is the unique non-trivial involution [19, 20]. In the case of SU(3), the Nakayama automorphism is identified in Theorem 4.6 with the Z3 symmetry of the underlying ADE graph given by the permutation P in (3). This answers a question we posed in [29, p.411]. The Hilbert 3 series (3) for A is a key ingredient in our determination of the Nakayama automorphism for the SU(3) ADE graphs. We also use the Ocneanu cells W (△) computed in [25], and exploit their Z3-invariance and in most cases the non-vanishing of (certain linear combinations of) cells which appear in the determination of a basis for A. It does not appear that the non-vanishing of these linear combinations can be deduced merely from the existence of cells in [59]. In Section 4 we obtain the first part of a resolution of A as an A-A bimodule A ⊗R A[3] → A ⊗R V ∗ ⊗R A[1] → A ⊗R V ⊗R A → A ⊗R A → A → 0, where R, V are the A-A bimodules generated by the vertices, edges of G respectively, the A-A bimodule V ∗ is the dual space of V , and B[m] denotes the graded space B but with grading shifted by m. The algebra A is a Calabi-Yau algebra if and only if the kernel of the leftmost map is zero, as in [38, 12, 14]. In our case, this kernel Ω4(A) is non-zero and is determined in Theorem 5.1. We show that Ω4(A) is isomorphic as an A-A bimodule to 1Aβ−1, and thus obtain a finite resolution of A as an A-A bimodule 0 → 1Aβ−1[h] → A ⊗R A[3] → A ⊗R V ∗ ⊗R A[1] → A ⊗R V ⊗R A → A ⊗R A → A → 0. This almost Calabi-Yau condition should be compared with the Calabi-Yau condition expressed above. This resolution of A will yield a projective resolution of A as an A-A bimodule. The objective in deriving this resolution is to provide a basis for the computation of the Hochschild (co)homology and cyclic homology of the algebras A(G, W ) for the SU(3) ADE graphs. Beginning with a pair (G, W ) given by a cell system W on an SU(3) ADE graph G, we construct a subfactor N ⊂ M which yields a nimrep which recovers the graph G as described in Section 2.3. Then we can construct the algebra A(G, W ) whose Hochschild (co)homology and cyclic homology only depends on the original pair (G, W ), or equivalently, on the subfactor N ⊂ M. Thus the Hochschild (co)homology and cyclic homology of A should be regarded as invariants for the subfactor N ⊂ M. 2 Preliminaries In this section we bring together the strands needed from subfactor theory, modular tensor categories and their modules, planar algebras and categorical approaches as outlined in the Introduction. 2.1 Representations of SU (n) and SU (n)k Here we describe our notation for the representation theory of SU(n) at level k ≤ ∞. Every irreducible representation λm of SU(n) is classified by a signature, or highest weight, m = (m1, m2, . . . , mn−1), where mi are integers such that m1 ≥ m2 ≥ · · · ≥ mn−1 ≥ 0, for i = 1, . . . , n − 1. A signature m can be represented by a Young diagram with at most n − 1 rows, and mi boxes in the ith row, i = 1, . . . , n − 1. The irreducible positive energy representations of the loop group of SU(n) at level k, or SU(n)k, are described by the irreducible representations of SU(n) whose signature has at most k columns. 4 For SU(2) the signatures are just the integers k ≥ 0. The fundamental representation is ρ = λ1, and the irreducible representations of SU(2) satisfy the fusion rules λm ⊗ ρ = λm−1⊕ λm+1 for m ≥ 1, and λ0⊗ ρ = ρ. The fusion graph for SU(2) is the infinite Dynkin diagram A∞. It is well known that the kth symmetric product of C2 gives the irreducible level k representation. The irreducible representations of SU(2)k satisfy the same fusion rules as those for SU(2), only now λm is also understood to be zero if m > k. For SU(3) the signatures are pairs (m1, m2) with m1 ≥ m2 ≥ 0. We will replace the pair (m1, m2) by the Dynkin labels (p, l) = (m2 − m1, m1), where now p, l ≥ 0. The conjugate representation of λ(p,l) is λ(p,l) = λ(l,p). The fundamental representation ρ = λ(1,0) generates every irreducible representation of SU(3) with its conjugate ρ, and the irreducible representations of SU(3) satisfy the fusion rules λ(p,l) ⊗ ρ = λ(p,l−1) ⊕ λ(p−1,l+1) ⊕ λ(p+1,l), λ(p,l) ⊗ ρ = λ(p−1,l) ⊕ λ(p+1,l−1) ⊕ λ(p,l+1), where λ(r,s) is understood to be zero if r < 0 or s < 0. The fusion graph is the infinite graph A(∞) (see [25, Figure 4]). The irreducible representations of SU(3)k satisfy the same fusion rules as those for SU(3), only now λ(p,l) is also understood to be zero if p + l > k, and the fusion graph is the truncated graph A(k+3). (4) 2.2 The generalized Temperley-Lieb algebras Let Mn = End(Cn). By Weyl duality, the fixed point algebra of ⊗mMn under the product adjoint action of SU(n) is generated by a representation σ → gσ on ⊗mCn of the group ring of the symmetric, or permutation, group Sm. This algebra is generated by unitary operators gj, j = 1, . . . , m − 1, which represent transpositions (j, j + 1), satisfying the relations g2 j = 1, gigj = gjgi, gigi+1gi = gi+1gigi+1, i − j > 1, and the vanishing of the antisymmetrizer sgn(σ)gσ = 0. Xσ∈Sm (5) (6) (7) (8) Writing gj = 1 − Uj, these unitary generators and relations are equivalent to the self- adjoint generators 1, Uj, j = 1, . . . , m − 1, and relations H1: H2: H3: U 2 i = δUi, UiUj = UjUi, i − j > 1, UiUi+1Ui − Ui = Ui+1UiUi+1 − Ui+1, where δ = 2, and the analogue of (8). There is a q-version of this algebra, which is a representation of a Hecke algebra. This is the centralizer of a representation of the quantum group SU(n)q (or the universal enveloping algebra), with a deformation of (5) to (q−1 − gj)(q + gj) = 0. 5 (9) The invertible generators gj, j = 1, . . . , m − 1, satisfy the relations (6), (7), (9) and the vanishing of the q-antisymmetrizer [18] Xσ∈Sm (−q)Iσgσ = 0, (10) where gσ =Qi∈Iσ gi if σ =Qi∈Iσ (i, i + 1). Then writing gj = q−1 − Uj, we are interested in the generalized Temperley-Lieb algebra generated by self-adjoint operators 1, Uj, j = 1, . . . , m − 1, satisfying H1-H3 and the analogue of (10), where now δ = q + q−1. In the cases n = 2, 3, which we are interested in, (10) reduces for SU(2) to the Temperley-Lieb condition and for SU(3) it is UiUi±1Ui − Ui = 0, (Ui − Ui+2Ui+1Ui + Ui+1) (Ui+1Ui+2Ui+1 − Ui+1) = 0. (11) (12) There are minor errors in a parallel discussion in Section 2 of the published version of [27] which have been corrected in the arXiv version. 2.3 Braided subfactors and modular invariants Let A and B be type III von Neumann factors. A unital ∗-homomorphism ρ : A → B is called a B-A morphism. The positive number dρ = [B : ρ(A)]1/2 is called the statistical dimension of ρ; here [B : ρ(A)] is the Jones-Kosaki index [46, 52] of the subfactor ρ(A) ⊂ B. Some B-A morphism ρ′ is called equivalent to ρ if ρ′ = Ad(u) ◦ ρ for some unitary u ∈ B. The equivalence class [ρ] of ρ is called the B-A sector of ρ. If ρ and σ are B-A morphisms with finite statistical dimensions, then the vector space of intertwiners Hom(ρ, σ) = {t ∈ B : tρ(a) = σ(a)t , a ∈ A} is finite-dimensional, and we denote its dimension by hρ, σi. A B-A morphism is called irreducible if hρ, ρi = 1, i.e. if Hom(ρ, ρ) = C1B. Then, if hρ, τi 6= 0 for some (pos- sibly reducible) B-A morphism τ , then [ρ] is called an irreducible subsector of [τ ] with multiplicity hρ, τi. An irreducible A-B morphism ρ is a conjugate morphism of the ir- reducible ρ if and only if [ρρ] contains the trivial sector [idA] as a subsector, and then hρρ, idBi = 1 = hρρ, idAi automatically [42]. The Verlinde algebra is realised in the subfactor models by systems of endomorphisms NXN of the hyperfinite type III1 factor N. That is, NXN denotes a finite system of finite index irreducible endomorphisms of a factor N in the sense that different elements of NXN are not unitary equivalent, for any λ ∈ NXN there is a representative λ ∈ NXN of the conjugate sector [λ], and NXN is closed under composition and subsequent irre- ducible decomposition. In the case of WZW models associated to SU(n) at level k, the Verlinde algebra is a non-degenerately braided system of endomorphisms NXN , labelled by the positive energy representations of the loop group of SU(n)k on a type III1 factor λνν which exactly match those of the positive energy representations [67]. The fusion matrices Nλ = [N σ ρλ]ρ,σ are a family of commuting normal matrices which give a representation themselves of the fusion rules of the positive energy N, with fusion rules λµ = Lν N µ 6 representations of the loop group of SU(n)k, NλNµ = Pν N µ fusion matrices can be simultaneously diagonalised: λνNν. This family {Nλ} of Nλ =Xσ Sσ,λ Sσ,1 SσS∗ σ, (13) where 1 is the trivial representation, and the eigenvalues Sσ,λ/Sσ,1 and eigenvectors Sσ = [Sσ,µ]µ are described by the statistics S matrix. Moreover, there is equality between the statistics S- and T - matrices and the Kac-Peterson modular S- and T - matrices which perform the conformal character transformations [48], thanks to [33, 32, 67]. The key structure in the conformal field theory is the modular invariant partition function Z. In the subfactor setting this is realised by a braided subfactor N ⊂ M where trivial (or permutation) invariants in the ambient factor M when restricted to N yield Z. This would mean that the dual canonical endomorphism is in Σ(NXN ), i.e. decomposes as a finite linear combination of endomorphisms in NXN . Indeed if this is the case for the inclusion N ⊂ M, then the process of α-induction allows us to analyse the modular invariant, providing two extensions of λ on N to endomorphisms α± λ of M, such that the matrix Zλ,µ = hα+ Let NXM , MXM denote a system of endomorphisms consisting of a choice of rep- resentative endomorphism of each irreducible subsector of sectors of the form [λι], [ιλι] respectively, for each λ ∈ NXN , where ι : N ֒→ M is the inclusion map which we may con- sider as an M-N morphism, and ι is a representative of its conjugate N-M sector. The action of the system NXN on the N-M sectors NXM produces a nimrep (non-negative λνGν, whose spectrum reproduces exactly the diagonal part of the modular invariant, i.e. matrix integer representation of the fusion rules) GλGµ = Pν N µ µi is a modular invariant [10, 7, 22]. λ , α− Gλ =Xσ Sσ,λ Sσ,1 ψσψ∗ σ, (14) with the spectrum of Gλ = {Sµ,λ/Sµ,1 with multiplicity Zµ,µ} [11, Theorem 4.16]. The labels µ of the non-zero diagonal elements are called the exponents of Z, counting multi- plicity. A modular invariant for which there exists a nimrep whose spectrum is described by the diagonal part of the invariant is said to be nimble. The systems NXN , NXM , MXM are (the irreducible objects of) tensor categories of endomorphisms with the Hom-spaces as their morphisms. Thus NXN gives a braided modular tensor category, and NXM a module category. The structure of the module cat- egory NXM is the same as a tensor functor F from NXN to the category Fun(NXM , NXM ) of additive functors from NXM to itself, see [62]. That is, F is essentially the module category NXM . The classification of SU(2) modular invariants is due to Cappelli, Itzykson and Zuber [15]. They label the modular invariant with an ADE graph G such that the diagonal part Zµ,µ of the invariant is exactly the multiplicity of the eigenvalue Sµ,ρ/Sµ,1 of G, where 1, ρ denote the trivial, fundamental representations respectively. Since these ADE graphs can be matched to the affine Dynkin diagrams -- the McKay graphs of the representation theory of the finite subgroups of SU(2) -- di Francesco and Zuber [18] were guided to find candidates for classifying graphs for SU(3) modular invariants by first considering the McKay graphs of the finite subgroups of SU(3) to produce a candidate list of ADE 7 graphs whose spectra described the diagonal part of the modular invariant. They proposed candidates for most of the modular invariants, except for the conjugate invariants A∗ In as they restricted themselves to only look for graphs which are three-colourable. the subfactor theory, this is understood in the following way. Suppose N ⊂ M is a braided subfactor which realises the modular invariant ZG. Evaluating the nimrep G at the fundamental representation ρ, we obtain for the inclusion N ⊂ M a matrix Gρ, which is the adjacency matrix for the ADE graph G which labels the modular invariant. Every SU(2) modular invariant is realised, and all nimreps are realised by subfactors [59, 60, 70, 5, 6, 10, 11], apart from the tadpole nimreps of the orbifolds of the even A's (see e.g. [8] for an explanation of the failure of the tadpole nimreps). Behrend, Pearce, Petkova and Zuber [1] (see also [73]) systematically proposed nimreps as a framework for boundary conformal field theory. The N-M system NXM corresponds to boundary fields in their language, and the M-M system MXM to defect lines. Bockenhauer and Evans [3] understood that nimrep graphs for the SU(3) conjugate invariants were not three-colourable. This was also realised simultaneously by Behrend, Pearce, Petkova and Zuber [1] and Ocneanu [60]. The figures for the complete list of the ADE graphs are given in [1, 60, 25]. The classification of SU(3) modular invariants was shown to be complete by Gannon [35], and the complete list is given in [26]. Ocneanu claimed [59, 60] that all SU(3) modular invariants were realised by subfactors and this was shown in [70, 5, 6, 10, 8, 9, 25, 26]. However, the classification of nimreps is incomplete if one relaxes the condition that the nimrep be compatible with a modular invariant [36, 41]. Ostrik [51, 62] took up a categorical description of subfactor α-induction, see [62, Remark 14], and this was taken further by Fjelstad, Frohlich, Fuchs, Schweigert and Runkel as a categorical framework for conformal field theory. See [34] for a review. 2.4 Subfactors Suppose we have a system of endomorphisms NXN of a type III1 factor N for SU(n)k, k ≤ ∞, where ρ denotes the endomorphism in NXN corresponding to the fundamental generator. We can form the tunnel · · · ⊂ ρρρ(N) ⊂ ρρ(N) ⊂ ρ(N) ⊂ N. (15) [λ(0)][ρ] = Ls By decomposing the sectors of 1, ρ, ρρ, ρρρ, . . . into irreducible sectors we can obtain the Bratteli diagram of the higher relative commutants of ρ(N) ⊂ N. If [λ(0)] is an irreducible at an even level of the Bratteli diagram and [λ(0)][ρ] decomposes into irreducibles as i=1[λ(i)], for irreducible sectors [λ(i)], i = 1, . . . , s, then there is an edge from the vertex [λ(0)] in the Bratteli diagram to the vertices [λ(i)], whilst if [λ(0)] is an irreducible at an odd level of the Bratteli diagram, we consider instead the decomposition of [λ(0)][ρ] into irreducibles. The Bratteli diagram obtained in this way is identical to that obtained for the Jones-Wenzl type II1 SU(n) subfactors [69]. The principal graph is the bipartite graph constructed by deleting at each level the vertices belonging to the old sectors (that is, any vertex at a given level which appeared at a previous level of the Bratteli diagram) and the edges emanating from them [39, Definition 4.6.5]. The decomposition of the sectors of 1, ρ, ρρ, ρρρ, . . . into irreducibles yields the dual principal graph in a similar way. 8 The decomposition of the sectors of the form (ρρ)m and (ρρ)mρ will not usually produce all the irreducible sectors in NXN , so to obtain all the irreducible sectors we also consider the decomposition of more general products ρmρl into irreducibles, m, l ≥ 0. In this way we recover the graph A when n = 2, 3, with vertices labelling the irreducible sectors, and the edges representing multiplication by the fundamental generator ρ. This corresponds to idempotent completion in the categorical language of Section 2.5. The principal graph, respectively dual principal graph is however only the 0-1, 0-(n − 1) part of the full graph A, where the edges are now undirected, and where either would be the entire graph only in the case when n = 2 [23]. There is an identification between intertwiners and explicit paths on the intertwining graph A [44], [24, Section 3.5]. For a sector [λi] at an even level in the Bratteli diagram, the intertwiners Hom(ρλi, λj) are identified with the edges from [λi] to [λj] on A, whilst for a sector [λ′ i, λj) are identified with the edges from [λ′ i] to [λj]. Let T (aj) denote an intertwiner labeled by an edge aj of A, and for a path x = a1a2 · · · as on A, define T (x) := T (a1)T (a2) · · · T (as). Then the T (x) are an orthogonal basis of the intertwiners between some endomorphisms. The spaces of intertwiners are the Hilbert spaces on which the system NXN acts. In this way the spaces Hom(ρm1ρl1, ρm2ρl2) of morphisms are identified with the span of all pairs (x1, x2) of paths x1, x2 on A and its opposite graph Aop where all the edges of A are reversed, where xj has mj, lj edges on A, Aop respectively, j = 1, 2. In particular, the algebras Hom(ρmρl, ρmρl) are identified with the path algebra (in the usual operator algebraic sense [24]) on A, Aop. Jones projections ej [46] for the tunnel (15) are identified with those in the path algebra on A, Aop, where edges are alternately on A and Aop [23]. The Jones projections in the path algebra are given by the product c∗c, where the annihilation and creation operators c, c∗ are defined in Section 2.6. i] at an odd level in the Bratteli diagram, the intertwiners Hom(ρλ′ We now focus on the cases n = 2, 3, where for k < ∞, q is a (k + n)th root of unity. For n = 2, let ρ = ρ denote the endomorphism corresponding to the fundamental generator of SU(2)k. The tunnel (15) defines Jones projections ej which generate the Temperley-Lieb algebra. The intertwiner space is generated by the Jones projections ej, so that Hom(ρm, ρm) ∼= T Lm := alg(1, e1, e2, . . . em−1). Jones-Wenzl projections fm = 1 − e1 ∨ · · · ∨ em−1 are given by [68]: fm+1 = fm − [2]q[m]q [m + 1]q fmemfm, (16) where the quantum integer [m]q is defined by [m]q = (qm − q−m)/(q − q−1). Here, a ∨ b denotes the projection such that Ran(a∨ b) = Ran(a) + Ran(b) where a, b are projections on a Hilbert space. For m < k − 1 ≤ ∞, the Jones-Wenzl projection fm is the minimal central projection corresponding to the new sector that appears at level m in the Bratteli diagram. For the fixed point algebra (⊗mM2)SU (2) which is equal to the Temperley-Lieb algebra with q = 1, the Jones-Wenzl projection fm is the projection on the (m + 1)- dimensional representation indexed by m in the intertwiner space Hom(ρm, ρm), where ρ is the fundamental representation of SU(2) on C2. For n = 3, we take the fundamental generator ρ of SU(3)k. The Jones projections ej, j = 1, . . . , 2m − 1, for the tunnel (15) are identified with projections in the algebras Hom(ρmρm, ρmρm), whilst the algebras Hom(ρm, ρm) are generated by the A2-Temperley- Lieb operators Uj, j = 1, . . . , m − 1. For the fixed point algebra (⊗mM3)SU (3), there is 9 a generalized Jones-Wenzl projection f(m,0) which is the projection on the representation indexed by λ(m,0) in the intertwiner space Hom(ρm, ρm), where ρ is the fundamental representation of SU(3) on C3. More generally, if we have an A2-Temperley-Lieb algebra generated by self-adjoint operators Ui with parameter δ = q + q−1, generalized Jones- Wenzl projections f(m,0) (called projectors in [64], also called clasps [53], magic elements [61]) are defined by [64]: f(0,0) = 1, f(m+1,0) = f(m,0) − f(1,0) = 1, [m]q f(m,0)Umf(m,0). [m + 1]q Note that the nimrep graph given by the module category NXM is not the principal graph for the braided inclusion N ⊂ M. The principal graph of a subfactor of index < 4 can only be A, Deven, E6 or E8, whilst the nimrep graph of a subfactor is any Coxeter- Dynkin diagram including Dodd and E7. Indeed, the principal graphs of the braided inclusions N ⊂ M usually have index which exceeds 4, and are those of the Goodman-de la Harpe-Jones construction and their generalizations discussed in [39, 24, 26]. The even, odd vertices of the ADE graphs are the A-A, B-B systems respectively for a subfactor A ⊂ B. For the braided SU(2) subfactors, all An vertices (both even and odd) are represented as N-N sectors, and all the vertices of the classifying graph G appear as N-M sectors. When we consider modular invariants, their module categories or nimreps, other graphs G will appear. The above intertwining discussion already leads us to the path Hilbert space CA, which is the vector space of paths on A, identified with the intertwin- ers T (x) where x is a path on A. The module category NXM from a braided inclusion N ⊂ M yields a nimrep G and we obtain the path Hilbert space CG, the vector space of paths on G = Gρ, identified with the intertwiners T (x) where x is now a path on G. De- note by (CG)j the space of paths of length j on G. The path Hilbert space CG is a graded algebra where multiplication (CG)i× (CG)j → (CG)i+j of two paths x ∈ (CG)i, y ∈ (CG)j is given by concatenation of paths xy, and is defined to be zero if r(x) 6= s(y), where s(x), r(x) denote the source, range vertices of the path x respectively. The endomorphisms End ((CG)j) on (CG)j are the (CG)j × (CG)j-matrices, with rows and columns labeled by the paths of length j on G. The End ((CG)j) have an algebra structure given by matrix multiplication. Thus there are two different notions of path algebra of G. In the theory of however, we will work with the graded algebra CG, as in e.g. [12]. operator algebras, the path algebra of G is usuallyLj≥0 End ((CG)j) [24]. In this paper 2.5 SU (2) Categorical Approach In this section we will describe the Verlinde algebra and fusion rules for SU(2) in the diagrammatic and categorical language of the Temperley-Lieb algebra [49, 65, 71, 16]. Let q be real or a root of unity, so that δ = [2]q is real. Denote by Tm,n the set of all planar diagrams consisting of a rectangle with m, n vertices along the top, bottom edge respectively, and with (m + n)/2 curves, called strings, inside the rectangle so that each vertex is the endpoint of exactly one string, and the strings do not cross each other. Let Vm,n denote the free vector space over C with basis Tm,n. Composition RS of diagrams R ∈ Tm,n, S ∈ Tn,p is given by gluing S vertically below R such that the vertices at 10 Figure 1: diagram Ei ∈ Tn the bottom of R and the top of S coincide, removing these vertices, and isotoping the glued strings if necessary to make them smooth. Any closed loops which may appear are removed, contributing a factor of δ. The resulting diagram is in Tm,p. This composition is clearly associative, and composition in V =Sm,n≥0 Vm,n is defined as its linear extension. The adjoint R∗ ∈ Tn,m of a diagram in R ∈ Tm,n is given by reflecting R about a horizontal line halfway between the top and bottom vertices of the diagram. This action is extended conjugate linearly to V. Let Ei denote the diagram in Tn := Tn,n illustrated in Figure 1. For δ ≥ 2 there is an isomorphism Vn ∼= T Ln given by δ−1Ei → ei. We will now define the Temperley-Lieb category T L as a matrix category T L = Mat(C). We begin by defining C to be the tensor category whose objects are projections in Vn := Vn,n, and whose morphisms Hom(p1, p2) between projections pi ∈ Vni, i = 1, 2, are given by the space p2Vn2,n1p1. We will use fraktur script to denote morphisms. The tensor product is defined on the objects and morphisms by horizontal juxtaposition. The trivial object id0 is the empty diagram which is a projection in V0. (The category C is the idempotent completion, or Karoubi envelope, of the category whose objects are non-negative integers, and whose morphisms are given by Vm,n.) In order to be able to take direct sums, we define the matrix category T L = Mat(C) to be the category with objects given by formal direct sums of objects in C, and morphisms Hom(p1 ⊕ · · · ⊕ pn1, q1 ⊕ · · · ⊕ qn2) given by n2 × n1 matrices, where the i, j-th entry is in Hom(pj, qi). The tensor product on T L is given on objects by (p1 ⊕ · · · ⊕ pn1) ⊗ (q1 ⊕ · · · ⊕ qn2) = (p1 ⊗ q1) ⊕ (p1 ⊗ q2) ⊕ · · · ⊕ (pn1 ⊗ qn2), and on morphisms by the usual tensor product on matrices with the tensor product for C on matrix entries. A projection p ∈ T L is called simple if hp, pi = 1. We write T Ln := Vn, and ρ for the identity object in T L1 consisting of a single vertical string. Then the identity diagram in T Ln, given by n vertical strings, is expressed by ρn := ⊗nρ. We have dim(T L0) = dim(T L1) = 1 and T L0, T L1 have simple projections f0 (the empty diagram), f1 = ρ respectively. Moving to T L2, the identity diagram ρ2 is a projection but is not simple, since hρ2, ρ2i = 2. One of these morphisms is the 1 = δE1, e1 = δ−1E1 is a projection. In identity diagram, the other is E1 = idE1. Since E2 fact, e1 is isomorphic to f0, as can be seen from the following isomorphisms ψ : f0 → e1, , and ψ∗ is defined by reflecting ψ about its horizontal axis. Then we have ψ∗ψ = f0 = idf0 and ψψ∗ = e1 = ide1, where the morphism e1 = δ−1E1. , and hf1, ρ2i = 0 (by parity), we Since hf0, ρ2i = 1, where the morphism is given by have the decomposition ρ2 = f0 ⊕ f2, where f2 is a simple projection in T L2. In the same way, we obtain at each level n that ρn is a linear combination of f0, . . . , fn−1 plus a new projection fn, which turns out to be simple. The morphisms fp = idfp are Jones-Wenzl projections, and satisfy a similar recursion relation to (16), with em replaced by em. These satisfy the properties [69]: ψ∗ : e1 → f0, where ψ =p[2]q −1 11 fp ⊗ ρ ∼= fp−1 ⊕ fp+1. ψ = fp+1  , ρp = ⌊p/2⌋Mj=0 (cid:20) p j (cid:21) fp−j, (17) (18) tr(fp) = [p + 1]q, where tr(fp) is given by connecting the ith string from the left along the top is connected to the ith string from the left along the bottom for each i = 1, . . . , p. For p ≥ p′ the Jones-Wenzl projections also satisfy the property fp(idρi ⊗ fp′ ⊗ idρp−p′−i) = fp = (idρi ⊗ fp′ ⊗ idρp−p′ −i)fp, for any 0 ≤ i ≤ p − p′. The morphisms fi and objects fi are denoted by ei, Xi respectively in [16]. In [56] there is some abuse of notation with both the objects and the morphisms given by the Jones-Wenzl projections denoted by f (i). We have the relation This is seen from the isomorphisms ψ : fp ⊗ ρ → fp−1 ⊕ fp+1, and ψ−1 = ψ∗, where and ψ∗ is defined as the transpose of ψ where we replace each entry a in ψ by the reflection a∗ of a about its horizontal axis. Then it easy to check using the above properties and (16) that ψψ∗ = fp−1 ⊕ fp+1 = idfp−1 ⊕ idfp+1 and ψ∗ψ = fp ⊗ idρ = idfp⊗ρ. Suppose that we have obtained simple projections fp for 0 ≤ p ≤ n, and we obtain a new projection fn+1 at level n + 1 as above. From the relation (17) we obtain the decomposition j , 0 ≤ p ≤ n. Then from the relation (17) with p = n we obtain the decomposition (18) with p = n + 1. Since fp is simple for j−1 for binomial coefficients C p j − C p j (cid:21) = C p where(cid:20) p 0 ≤ p ≤ n, hρn+1, ρn+1i = hfn+1, fn+1i +P⌊(n+1)/2⌋ j=0 (cid:20) n + 1 j (cid:21) − 1 = hfn+1, fn+1i + cn − 1, n /(n + 1) is the nth Catalan number, which gives the dimension of T Ln in where cn = C 2n the generic case. Thus we see that hfn+1, fn+1i = 1, so that fn+1 is indeed simple. In the generic case, δ ≥ 2, the Temperley-Lieb category T L is semisimple, that is, every projection is a direct sum of simple projections, and for any pair of non-isomorphic simple projections p1, p2 we have hp1, p2i = 0. We recover the infinite Dynkin diagram A∞, where vertices are labeled by the projections fi and edges represent tensoring by ρ. 12 In the non-generic case, δ = [2]q < 2, where q is a k + 2th root of unity, we have [k + 2]q = 0. Then tr(fk+1) = [k + 2]q = 0. Thus the negligible morphisms are those in the unique proper tensor ideal in the Temperley-Lieb category generated by fk+1 [40]. The quotient T L(k) := T L/hfk+1i is semisimple with simple objects f0 = id0, f1 = ρ, f2, . . . , fk which satisfy the fusion rules (17) for p < k, and fk ⊗ ρ ∼= fk−1. Thus we recover the Dynkin diagram Ak+1, where the vertices are labeled by the projections fi and edges represent tensoring by ρ. 2.6 SU (2) module categories In this section we describe SU(2) module categories in terms of preprojective alge- bras. Then in the subsequent Section 2.7 we relate this to braided subfactors using the Goodman-de la Harpe-Jones construction [39, 24, 11] and its manifestation in the bipartite graph planar algebra construction [47]. reverse edgeea ∈ G1 from j to i. As described in Section 2.4, the irreducible sectors in As usual NXN will be a braided system of endomorphisms on a factor N, and N ⊂ M will be a braided inclusion with classifying graph G = Gρ, of ADE type, arising from the nimrep G of NXN acting on NXM . Denote by G0, G1 the vertices, edges of G respectively. The graph G is directed, and for every edge a ∈ G1 from vertex i to j, there is a unique NXN label the vertices of the Dynkin diagram Ak+1, and the edges of Ak+1 represent multiplication by the fundamental generator, and the irreducible endomorphisms satisfy the same fusion rules as the projections fi in the category T L(k). We will use λi to denote endomorphisms in NXN , whilst fi will denote the abstract object in the category T L(k). Semisimple module categories over Cq (where Cq = T L for q = ±1 or q not a root of unity, and Cq = T L(k) when q is an k + 2th root of unity) where classified in [21]: A semisimple Cq-module category D is abelian, and is equivalent as an abelian category to the category MI of I-graded vector spaces, where I are the (isomorphism classes of) simple objects of D. The structure of a Cq-module category on MI is the same as a tensor functor F from Cq to Fun(MI,MI) ∼= MI×I, the category of additive functors from MI to itself. Thus the module category D = NXM gives rise to a monoidal functor F from the Temperley-Lieb category T L(k) to Fun(NXM , NXM ), given by Gλ(i, j) Ci,j, (19) F (f ) = Mi,j∈G0 where λ = λp is an irreducible endomorphism in NXN identified naturally with the Jones- Wenzl projection f = fp. The Ci,j are 1-dimensional R-R bimodules, where R = (CG)0. The category of R-R bimodules has a natural monoidal structure given by ⊗R, or more i,j ⊗ E(2) i,j , r = 1, 2. Then we have R-R bimodules (E(1) ⊗R E(2))i,k, where (E(1) ⊗R E(2))i,k =Lj∈G0 explicitly, E(1) ⊗R E(2) =Li,k∈G0 j,k ) for all R-R bimodules E(r) =Li,j∈G0 F (ρ) =Li,j∈G0 ∆G(i, j) Ci,j = (CG)1 and F (ρm) = ⊗m(CG)1 = (CG)m. The set of all edges a form a basis for (CG)1. The functor F is defined on the morphisms of T L by specifying annihilation and creation operators c, c∗ respectively [58, Section 4]: E(r) (E(1) √µr(a) √µs(a) s(a), c(ab) = δb,ea 13 (20) c∗(i) = Xa∈G1:r(a)=i √µs(a) √µi eaa, ) = c∗, F ( ) = c. (21) p=0 F (fp), where the pth graded part is Σp = Let Σ be the graded algebra Σ = L∞ where (µj)j is the Perron-Frobenius eigenvector for the Perron-Frobenius eigenvalue δ of G. Then we set F ( F (fp) = F (λp). The multiplication µ is defined by µp,l = F (fp+l) : Σp ⊗R Σl → Σp+l. Preprojective algebras associated to graphs were introduced in [37], and it was shown that they are finite dimensional if and only if the graphs are of ADE type. They have since found many other applications, including to Kleinian singularities [17] and to Nakajima's quiver varieties [55]. The preprojective algebra of G is the graded al- ))i ⊂ CG is the two-sided gebra defined by Π = CG/hIm (F ( in CG. Its pth graded part is ideal generated by the image of the creation operators Πp = (CG)p/hIm (F ( ))i to i=1 Im(F (ei)), the linear span of the images in CGp of the morphisms ei = idei. i=1 Im(F (ei)) = (CG)p/Im(F (e1) ∨ · · · ∨ F (ep−1)) is isomorphic to ker(F (e1) ∨ · · · ∨ F (ep−1)) = Im(1 − F (e1) ∨ · · · ∨ F (ep−1)) = Im(F (fp)). These are the essential paths EssPathp = ker(F (e1) ∨ · · · ∨ F (ep−1)) of Ocneanu [58], and we have (CG)p, which is equal to Pp−1 Now (CG)p/Pp−1 ))ip, where hIm (F ( ))ip is the restriction of hIm (F ( ))i, where hIm (F ( Πp = (CG)p/ p−1Xi=1 Im(F (ei)) ∼= Im(F (fp)) = F (fp) = Σp. The isomorphism ϕ : Σp → Πp is given by the natural inclusion of Σp in (CG)p, then passing to the quotient (CG)p/hIm (F ( ))ip = Πp. That ϕ is an isomorphism as algebras is seen as follows, see [16, Proposition 5.5.6]. Since the quotient map π : CG → Π is an algebra homomorphism, the multiplication of the images of Σr and Σs in Π is equal to the multiplication of F (fr) and F (fs) in CG and then taking the quotient. Now the image of µr,s(Σr ⊗ Σs) in CG is F (fr+s), which is equal to the image of the multiple of F (fr) and F (fs) in CG under F (fr+s), since fp(ρi ⊗ fp′ ⊗ ρp−p′−i) = fp for any 0 ≤ i ≤ p − p′, p′ ≤ p. Since fr+s = idρr+s + φ, where φ is a linear combination of ei, we see that ))i so that π ◦ F (fr+s) = π. Thus ϕ is an algebra homomorphism. Im(F (φ)) ⊂ hIm (F ( Applying the functor F to the construction in Section 2.5 we obtain the identification F (fp) ⊗R F (ρ) ∼= F (fp−1) ⊕ F (fp+1), which yields Σp ⊗R (CG)1 ∼= Σp−1 ⊕ Σp+1. (22) Let Λ be the graded coalgebra Λ = F (f0)⊕ F (f1)⊕ F (f0) = (CG)0 ⊕ (CG)1 ⊕ (CG)0 with comultiplication ∆, where ∆1,1 : Λ2 → Λ1 ⊗S Λ1 is given by ∆1,1 = F ( ) and the other comultiplications are trivial [16]. Let δ = [2]q and suppose [m]q 6= 0 for all m ≤ n for some n ∈ N. Then for all p ≤ n − 2 we obtain the following exact sequence: 0 −→ Σp−1 ⊗R Λ2 −→ Σp ⊗R Λ1 −→ Σp+1 ⊗R Λ0 −→ 0, (23) where the connecting maps are given by the Koszul differential, the composite map dp,l = (µp,1⊗1)◦(1⊗∆1,l−1) : Σp⊗R Λl → Σp⊗R Λ1⊗R Λl−1 = Σp⊗R Σ1⊗R Λl−1 → Σp+1⊗R Λl−1. 14 Thus for δ = [2]q < 2, where q is a k + 2th root of unity, we have Σ =Lk p=0 F (fp), since fp = 0 in T L(k) for p ≥ k + 1. This means that Im(F (e1) ∨ · · · ∨ F (ek)) = (CG)k+1. The short exact sequence (23) degenerates for p = k to give 0 −→ Σk−1 ⊗R Λ2 −→ Σk⊗RΛ1 −→ 0, and we see that the pair (Π, Λ) is almost Koszul, in the sense of [13], where the preprojective algebra Π is a (k, 2)-Koszul algebra [13, Corollary 4.3] [16, Corollary 5.6.16]. In the generic case, δ ≥ 2, there is an analogous pair (Π, Λ) which is Koszul [13], where Koszul duality is a generalization of the duality between symmetric and antisymmetric algebras. 2.7 Bipartite graph planar algebras and the GHJ construction Jones [47] introduced the graph planar algebra construction for a bipartite graph G. We will show that the functor F defined in (19) recovers this bipartite graph planar algebra construction. The planar algebra P G of a finite bipartite graph G, introduced in [47], is the path algebra on G where paths may start at any of the even vertices of G, and where the mth graded part P G m is given by all pairs of paths of length m on G which start at the same even vertex and have the same end vertex. Let P be the set of tangles in a disc with an even number of vertices on its outer disc, and a finite number of internal discs, each with an even number of vertices, such that each vertex is an endpoint of a string. Internal discs with 2m vertices on their boundary are labeled by elements of P G m. The presenting map Z : P → P G is defined uniquely [47, Theorem 3.1], up to isotopy, by first isotoping the strings of a tangle T with internal discs in such a way that T may be divided into horizontal strips where in each strip only cups, caps, internal discs or through strings appear. Then each cup, cap is given by the local operators (21), (20) respectively, which operate on the elements of P G inserted in the internal discs, and the outer boundary of the tangle T yields an element of P G. The planar algebra P G is a planar ∗-algebra, with ∗-operation defined on matrix units by (x1, x2)∗ = (x2, x1). The ∗-structure on a tangle T is given by reflecting about a horizontal line which bisects T , and replacing every label of T by its adjoint. The tangles Ei in Figure 1 are thus self-adjoint. m+1 We have a tower of algebras P G is given by the graph G. There is a positive definite inner product defined from the trace on P G. We have the inclusion P ∅ m ⊂ P G 2 ⊂ · · · , where the inclusion P G m for each m, and a double sequence 0 ⊂ P G 1 ⊂ P G m := Z(Vm) ⊂ P G 1 ⊂ P ∅ 0 ⊂ P ∅ P ∅ 2 ⊂ · · · ∩ ∩ ∩ P G 0 ⊂ P G 1 ⊂ P G 2 ⊂ · · · Then P ∅ := Z(V) is the embedding of the Temperley-Lieb algebra into the path algebra of G, which is used to construct the Goodman-de la Harpe-Jones (GHJ) subfactors [39]. Let P G denote the von Neumann algebra GNS-completion of P G with respect to the trace. Then for q = Z(∗G) the minimal projection in P G 0 corresponding to the distinguished vertex ∗G of G with lowest Perron-Frobenius weight, we have an inclusion qP ∅ ⊂ qP Gq which gives the Goodman-de la Harpe-Jones subfactor NA ⊂ NG, where N ′ A ∩ NG = qP G 0 q = C. 15 Figure 2: A2 webs The limit of the sequence of inclusions 1 ⊂ P G 0 ⊂ P G P G 2 ⊂ · · · ∩ ∩ ∩ P G 1 ⊂ P G 2 ⊂ P G 3 ⊂ · · · gives a subfactor NG ⊂ MG in a similar way. Thus we obtain a commuting square of inclusions [39]: NA ⊂ NG ∩ ∩ MA ⊂ MG (24) This commuting square allows us to compute the dual canonical endomorphism θ of the GHJ subfactor, from which we construct the nimrep graph G = Gρ [11]. Since the functor F in Section 2.6 is defined by the annihilation and creation operators in (20), (21), we see that F is equivalent to the presenting map Z above. The embedding P A ⊂ P G is given by the image under F of the morphisms in the Temperley-Lieb category. 2.8 SU (3) Categorical Approach In this section we describe the Verlinde algebra and fusion rules for SU(3) in the dia- grammatic and categorical language of Kuperberg spiders [53, 27, 16], namely the A2- Temperley-Lieb category. Let q be real or a root of unity, so that δ = [2]q is real. The A2-Temperley-Lieb alge- bra is the generalized Temperley-Lieb algebra generated by a family {Uj} of self-adjoint operators which satisfy the relations H1-H3 and the vanishing of the q-antisymmetrizer for SU(3) which gives (12). We call a vertex a source vertex if the string attached to it has orientation away from the vertex. Similarly, a sink vertex will be a vertex where the string attached has orientation towards the vertex. A string s is a sequence of signs +, −. For two (possibly empty) strings s1, s2, an A2-s1, s2-tangle T is a tangle on a rectangle with strings s1, s2 along the top, bottom edges respectively, generated by A2 webs (see Figure 2) such that every free end of T is attached to a vertex along the top or bottom of the rectangle in a way that respects the orientation of the strings, every vertex has a string attached to it, and the tangle contains no closed loops or elliptic faces. Along the top edge the points + are source vertices and − are sink vertices, while along the bottom edge the roles are reversed. We define the vector space V A2 s1,s2 to be the free vector space over C with basis T A2 s1,s2. Kuperberg relations K1-K3 below. That is, composition in V A2 s1,s2 by the Kuperberg ideal generated by the s1,s2 is defined as follows. We define V A2 s1,s2 to be the quotient of V A2 16 Figure 3: The tangle Wi ∈ V A2 m . The composition RS ∈ V A2 s1,s3 of an A2-s1, s2-tangle R and an A2-s2, s3-tangle S is given by gluing S vertically below R such that the vertices at the bottom of R and the top of S coincide, removing these vertices, and isotoping the glued strings if necessary to make them smooth. Any closed loops which may appear are removed, contributing a factor of α = [3]q, as in relation K1 below. Any elliptic faces that appear are removed using relations K2, K3 below. The composition is associative and is extended linearly to elements in V A2 s1,s2. K1: K2: K3: There is a braiding on V A2 local diagrams in V A2 s1,s2 (see [53, 64]), for any q ∈ C: s1,s2, defined locally by the following linear combinations of (25) The braiding satisfies type II and type III Reidemeister moves, and a braiding fusion relation [27, Equations (8), (9)], provided δ = [2]q and α = [3]q. (m,n),(m′,n′) := V A2 Thus it is sufficient to work over V A2 +m−n,+m′ −n′ , where +k−l is the string of k signs + followed by l signs −, since, for any arbitrary string s with m signs + and n signs − and string s′ with m′ signs + and n′ signs −, there is an isomorphism ι between V A2 s,s′ and V A2 (m,n),(m′,n′) given by using the braiding to permute the order of the signs in s to +m−n, and the inverse braiding to permute the order of the signs in s′ to +m′ A diagrammatic representation of the Hecke algebra for SU(3) is as follows: Let Wi ∈ V A2 (m,0),(m,0) be the tangle illustrated in Figure 3. A ∗-operation can be m , T ∗ is the m-tangle obtained by reflecting defined on V A2 T about a horizontal line halfway between the top and bottom vertices of the tangle, and reversing the orientations on every string. Then ∗ on V A2 m is the conjugate linear extension of ∗ on T A2 m . For δ ∈ R (so q ∈ R or q a root of unity), the ∗-operation m , where for an m-tangle T ∈ T A2 m := V A2 −n′ . 17 leaves the Kuperberg ideal invariant due to the symmetry of the relations K1-K3. For m ∈ N∪{0} we define the algebra A2-T Lm to be alg(idm, wii = 1, . . . , m− 1), where idm is the identity diagram given by m vertical strings, and wi is the image of Wi ∈ V A2 m in the quotient space V A2 (m,0),(m,0). The wi's in A2-T Lm are clearly self-adjoint, and satisfy the relations H1-H3 and (12) [27]. m := V A2 Diagrammatically, the generalized Jones-Wenzl projections f(m,0) of Section 2.4 are given as follows: f(1,0) is given by a single vertical string in T A2 is defined inductively by [64, (2.1.0)-(2.1.2)]: +,−, whilst f(m,0) ∈ V A2 (m,0),(m,0) (26) The generalized Jones-Wenzl projections f(m,n) ∈ V A2 (m,n) := V A2 (m,n),(m,n) are defined inductively by [64, (2.1.7)]: (27) We will define the A2-Temperley-Lieb category by A2-T L = Mat(C A2), where C A2 is the tensor category whose objects are projections in V A2 (m,n), and whose morphisms are Hom(p1, p2) = p2V A2 (m2,n2),(m1,n1)p1, for projections pi ∈ V A2 (mi,ni), i = 1, 2. We write A2-T L(m,n) = V A2 (m,n), and ρ, ρ for the identity projections in A2-T L(1,0), A2-T L(0,1) re- spectively consisting of a single string with orientation downwards, upwards respectively. Then the identity diagram in A2-T L(m,n), given by m + n vertical strings where the first m strings have downwards orientation and the next n have upwards orientation, is ex- pressed as ρmρn. We have dim(A2-T L(0,0)) = dim(A2-T L(1,0)) = dim(A2-T L(0,1)) = 1 and A2-T L(0,0), A2-T L(1,0) and A2-T L(0,1) have simple projections f(0,0) (the empty diagram), f(1,0) = ρ and f(0,1) = ρ respectively. Moving to level 2, that is, A2-T L(k,l) such that k + l = 2, consider first A2-T L(2,0). The identity diagram ρ2 is a projection but is not simple, since hρ2, ρ2i = 2. One of these morphisms is the identity diagram, the other is w1 = idw1. Now u1 = δ−1w1 is a projection, isomorphic to f(0,1), as can be seen from the isomorphisms ψ : f(0,1) → u1 and ψ∗ : u1 → f(0,1), where ψ = √δ . Then we have ψψ∗ = u1 = idu1, ψ∗ψ = f(0,1) = idf(0,1). Since hf(0,1), ρ2i = 1, where the morphism , and hf ′, ρ2i = 0 for f ′ = f(0,0), f(1,0) (by parity), we have the decom- is given by position ρ2 = f(0,1) ⊕ f(2,0), where f(2,0) is a simple projection in A2-T L(2,0). Similarly, ρ2 = f(1,0) ⊕ f(0,2), where f(0,2) is a simple projection in A2-T L(0,2). Finally at level 2 consider A2-T L(1,1). The simple projection e1 := α−1E1 ∈ Hom(ρ ⊗ ρ, ρ ⊗ ρ) is isomor- phic to f(0,0), as in the SU(2) case (see Section 2.5). Since hf(0,0), ρ ⊗ ρi = 1, where the −1 18 , and hf ′, ρ ⊗ ρi = 0 for f ′ = f(1,0), f(0,1) (by parity), we have morphism is given by the decomposition ρ ⊗ ρ = f(0,0) ⊕ f(1,1), where f(1,1) is a simple projection in A2-T L(1,1). In the same way, we obtain we obtain at each level that ρmρn is a linear combination of f(i,j) for i, j ≥ 0, 0 ≤ i+j < m+n such that i−j ≡ m−n mod 3, plus a new projection f(m,n), which turns out to be simple. The morphisms f(p,l) = idf(p,l) are generalized Jones- Wenzl projections which satisfy the recursion relations (26) and (27). These satisfy the properties [64]: tr(f(p,l)) = [p + 1]q[l + 1]q[p + l + 2]q [2]q , where tr is defined as in Section 2.5. For p ≥ p′ and l ≥ l′, these generalized Jones- Wenzl projections also satisfy the property f(p,l)(idρiρj ⊗ f(p′,l′) ⊗ idρp−p′ −iρl−l′ −j ) = f(p,l) = (idρiρj ⊗ f(p′,l′) ⊗ idρp−p′−iρl−l′−j )f(p,l), for any 0 ≤ i ≤ p − p′, 0 ≤ j ≤ l − l′. This property also holds if we conjugate either f(p,l) or idρiρj ⊗ f(p′,l′) ⊗ idρp−p′−iρl−l′−j by any braiding. We have the relations (28) given by the isomorphisms ψ : f(p,l) ⊗ ρ → f(p,l−1) ⊕ f(p−1,l+1) ⊕ f(p+1,l), and ψ−1 = ψ∗, where f(p,l) ⊗ ρ ∼= f(p,l−1) ⊕ f(p−1,l+1) ⊕ f(p+1,l), ψ = . (29)  f(p+1,l) 19  Then it easy to check using the above properties and (27) that ψψ∗ = f(p,l−1) ⊕ f(p−1,l+1) ⊕ f(p+1,l) and ψ∗ψ = f(p,l) ⊗ idρ. Similarly we have f(p,l) ⊗ ρ ∼= f(p−1,l) ⊕ f(p+1,l−1) ⊕ f(p,l+1), (30) so the f(p,l) satisfy the fusion rules for SU(3), given in (4). Suppose that we have obtained simple projections f(p,l) for 0 ≤ p + l ≤ n, and we obtain a new projection f(j,n−j+1) at level n + 1 as above, for some 0 ≤ j ≤ n + 1. From the relation (28) with p = j, l = n− j and from dimension considerations we see that hf(j,n−j+1), f(j,n−j+1)i = 1, so that f(j,n−j+1) is indeed simple. In the generic case, δ ≥ 2, the A2-Temperley-Lieb category A2-T L is semisimple and for any pair of non-isomorphic simple projections p1, p2 we have hp1, p2i = 0. We recover the infinite graph A(∞), where the vertices are labeled by the projections f(p,l) and the edges represent tensoring by ρ. In the non-generic case where q is a k + 3th root of unity, we have tr(f(p,l)) = 0 for p + l = k + 1. By (27), f(p′,l′) = 0 for all p′, l′ ≥ k + 2 if f(p,l) = 0 for p + l = k + 1. Thus the negligible morphisms are the ideal hf(p,l)p + l = k + 1i generated by f(p,l) such that p + l = k + 1. The quotient A2-T L(k) := A2-T L/hf(p,l)p + l = k + 1i is semisimple with simple objects f(p,l), p, l ≥ 0 such that p + l ≤ k which satisfy the fusion rules (28) and (30), where f(p′,l′) is understood to be zero if p′ < 0, l′ < 0 or p′ + l′ ≥ k + 1. Thus we recover the graph A(k+3), where the vertices are labeled by the projections f(p,l) and the edges represent tensoring by ρ. 2.9 SU (3) module categories In this section we describe SU(3) module categories in terms of certain algebras of paths. Then in the subsequent Section 2.10 we relate this to braided subfactors using the SU(3) Goodman-de la Harpe-Jones construction [26] and its manifestation in the SU(3)-graph planar algebra construction [28]. As usual NXN = {λ(p,l) 0 ≤ p, l, p + l ≤ k < ∞} will be braided system of endomor- phisms of SU(3)k on a factor N, and N ⊂ M will be a braided inclusion with classifying graph G = Gρ, of ADE type, arising from the nimrep G of NXN acting on NXM . Then the module category gives rise to a monoidal functor F from the A2-Temperley-Lieb category A2-T L(k) to Fun(NXM , NXM ), where F is given by (19), where now λ = λ(p,l) is an ir- reducible endomorphism in NXN identified with the generalized Jones-Wenzl projections f = f(p,l). We denote by Gop the opposite graph of G obtained by reversing the orientation of every edge of G. Then we have that F (ρmρn) is the R-R bimodule with basis given by all paths of length m + n on G, Gop, where the first m edges are on G and the last n edges are on Gop, where R = (CG)0. In particular F (ρm) = (CG)m. the corresponding edge with opposite orientation on Gop. We define annihilation operators cl, cr by: If a ∈ G1 is an edge on G, we denote byea ∈ Gop 1 and creation operators c∗ Gop, and (µj)j is the Perron-Frobenius eigenvector for the Perron-Frobenius eigenvalue α r as their adjoints, where a is an edge on G andeb an edge on l , c∗ 20 √µr(a) √µs(a) s(a), cl(aeb) = δs(a),s(b) √µs(a) √µr(a) cr(eba) = δr(a),r(b) r(a), (31) Figure 4: left and right cups; left and right caps Figure 5: incoming and outgoing Y-forks; incoming and outgoing inverted Y-forks of G. Define the following fork operators g, g by: g (ea) = g(a) = 1 √µs(a)µr(a)Xb1,b2 √µs(a)µr(a)Xb1,b2 1 W (△(a,b1,b2) s(a),r(a),r(b1))b1b2, W (△(a,b1,b2) s(a),r(a),r(b1))eb1eb2, (32) (33) where △(a,b1,b2) s(a),r(a),r(b1) denotes the closed loop of length 3 on G along the edges a, b1 and b2, and W (△) are the Ocneanu cells on G constructed in [25]. We also define f = g∗ and f = g∗. Then the functor F is defined on the morphisms of A2-T L by assigning the following operators to the morphisms given in Figures 4 and 5: to the left, right caps the annihilation operators cl, cr respectively, given by (31), to the left, right cups the creation operators c∗ r respectively, to the incoming, outgoing Y-forks the operators g, g respectively given in (32), (33), and to the incoming, outgoing inverted Y-forks the operators f, f respectively. Let Σ be the graded algebra Σ =L∞ p=0 F (f(p,0)), where the pth graded part is Σp = F (f(p,0)) = F (λ(p,0)). The multiplication µ is defined by µp,l = F (f(p+l,0)) : Σp ⊗R Σl → Σp+l, where f(p,l) = idf(p,l). l , c∗ ))i, where hIm(F ( ))ip, where hIm(F ( We define a graded algebra Π by Π = CG/hIm(F ( to (CG)p, which is equal to Pp−1 The quotient (CG)p/Pp−1 is the two-sided ideal generated by the image of the operators part is Πp = (CG)p/hIm(F ( morphisms Ui = idUi. ))i ⊂ CG in CG. Its pth graded ))i i=1 Im(F (Ui)), the union of the images on CGp of the i=1 Im(F (Ui)) = (CG)p/Im(F (U1)∨· · ·∨F (Up−1)) is isomorphic to ker(F (U1) ∨ · · · ∨ F (Up−1)). Clearly ker(F (U1) ∨ · · · ∨ F (Up−1)) ⊃ Im(F (f(p,0))) since Uif(p,0) = 0 for i = 1, . . . , p − 1. For a ∈ ker(F (U1) ∨ · · · ∨ F (Up−1)), a = a · F (f(p,0)) since the only term in f(p,0) which does not contain a Ui is the identity, which has coefficient 1. Thus a ∈ Im(F (f(p,0))) so ker(F (U1) ∨ · · · ∨ F (Up−1)) = Im(F (f(p,0))). ))ip is the restriction of hIm(F ( Then we have Πp = (CG)p/ p−1Xi=1 Im(F (Ui)) ∼= Im(F (f(p,0))) = F (f(p,0)) = Σp. (34) The isomorphism is given by the natural inclusion of Σp in (CG)p, then passing to the quotient (CG)p/hIm(F ( ))ip = Πp. That this map is an isomorphism as algebras follows by an analogous argument to that in the SU(2) case [16, Theorem 7.3.5]. 21 Applying the functor F to the construction in Section 2.8 we obtain the identifications F (f(k,l)) ⊗R F (ρ) ∼= F (f(p,l−1)) ⊕ F (f(p−1,l+1)) ⊕ F (f(p+1,l)) and F (f(p−2,l)) ⊗R F (ρ) ∼= F (f(p−2,l)) ⊕ F (f(p,l−1)) ⊕ F (f(p−1,l+1)), which yield when l = 0: Σp ⊗R (CG)1 ∼= F (f(p−1,1)) ⊕ Σp+1, Σp−1 ⊗R (CGop)1 ∼= Σp−2 ⊕ F (f(p−1,1)). (35) (36) ) : Λ3 → Λ1 ⊗R Λ2, ∆2,1 = F ( Let Λ be the graded coalgebra Λ = F (f(0,0)) ⊕ F (f(1,0)) ⊕ F (f(0,1)) ⊕ F (f(0,0)) = (CG)0 ⊕ (CG)1 ⊕ (CGop)1 ⊕ (CG)0. The comultiplication is ∆ is given by ∆1,1 = F ( ) : Λ2 → Λ1 ⊗R Λ1, ∆1,2 = F ( ) : Λ3 → Λ2 ⊗R Λ1, and the other comultiplications are trivial. Let δ = [2]q and suppose [m]q 6= 0 for all m ≤ n for some n ∈ N. Then for all p ≤ n − 3 we obtain the following exact sequences: d−→ Σp+1 ⊗R Λ0 −→ 0, −→ F (f(p−1,1)) −→ 0, d−→ Σp−1 ⊗R Λ2 0 −→ F (f(p−1,1)) −→ Σp ⊗R Λ1 0 −→ Σp−2 ⊗R Λ3 F (ψ2) F (f(p−1,1)) (37) (38) where ψ = (ψi) is the isomorphism given in (29) and d is the Koszul differential defined in Section 2.6. These sequences can be combined to give another exact sequence: 0 −→ Σp−2 ⊗R Λ3 −→ Σp−1 ⊗R Λ2 −→ Σp ⊗R Λ1 −→ Σp+1 ⊗R Λ0 −→ 0, (39) where now all the connecting maps are given by the Koszul differential d. Thus when q is a k + 3th root of unity, we have Σ =Lk p=0 F (f(p,0)), since f(p,0) = 0 in A2-T L(k) for p ≥ k + 1. This means that Im(F (U1) ∨ · · · ∨ F (Uk)) = (CG)k+1. The exact sequence (39) degenerates for p = k to give 0 −→ Σk−2 ⊗R Λ3 −→ Σk−1 ⊗R Λ2 −→ Σk ⊗R Λ1 −→ 0, and for p = k + 1 it degenerates to give 0 −→ Σk−1 ⊗R Λ3 −→ Σk ⊗R Λ2 −→ 0. Then we see that the pair (Π, Λ) is almost Koszul, where the algebra Π is a (k, 3)-Koszul algebra [16, Corollary 7.4.19]. Corollary 7.3.9]. In the generic case, δ ≥ 2, there is an analogous pair (Π, Λ) which is Koszul [16, The isomorphism between the two algebras Π and Σ is a key ingredient in the deter- mination of the Hilbert series in Section 3. 2.10 SU (3)-graph planar algebras and the SU (3)-GHJ construc- tion In [28] we introduced the A2-graph planar algebra construction for an SU(3) ADE graph G. The A2-graph planar algebra P G of an SU(3) ADE graph G is the path algebra on G and Gop. We will show that the functor F defined in Section 2.9 recovers this A2-graph planar algebra construction. The presenting map Z : P → P G is defined uniquely [28, Theorem 5.1], up to isotopy, by first isotoping the strings of T in such a way that the diagram T may be divided into horizontal strips so that each horizontal strip only contains the following elements: a (left or right) cup, a (left or right) cap, an (incoming or outgoing) Y-fork, or an (incoming or outgoing) inverted Y-fork, see Figures 4 and 5. Then Z assigns to the left, right caps the annihilation operators cl, cr respectively, given by (31), to the left, right cups the 22 creation operators c∗ r respectively, to the incoming, outgoing Y-forks the operators g, g respectively given in (32), (33), and to the incoming, outgoing inverted Y-forks the operators f, f respectively. l , c∗ We have a tower of algebras P G 0,0 ⊂ P G 0,1 ⊂ P G is given by the m-(m + 1) part of the graph G. There is a positive definite inner product defined from the trace on P G. We have the inclusion P ∅ 0,m for each m, and we have a double sequence 0,m := Z(V A2 m ) ⊂ P G 0,2 ⊂ · · · , where the inclusion P G m ⊂ P G m+1 0,1 ⊂ P ∅ 0,0 ⊂ P ∅ P ∅ 0,2 ⊂ · · · ∩ ∩ ∩ 0,0 ⊂ P G 0,1 ⊂ P G P G 0,2 ⊂ · · · A∩NG = qP G Then P ∅ := Z(V A2) is the embedding of the A2-Temperley-Lieb algebra into the path algebra of G, which is used to construct the A2-Goodman-de la Harpe-Jones subfactors [26]. Let P G denote the GNS-completion of P G with respect to the trace. Then for q = Z(∗G) the minimal projection in P G 0 corresponding to the vertex ∗G of G, we have an inclusion qP ∅ ⊂ qP Gq which gives the A2-Goodman-de la Harpe-Jones subfactor NA ⊂ 0,m ⊂ P G NG, where N ′ 0,m}m is a periodic sequence of commuting squares of period 3, in the sense of Wenzl in [69]. Thus we obtain a commuting square of inclusions as in (24), which allows us to compute the dual canonical endomorphism θ of the A2-Goodman-de la Harpe-Jones subfactor, from which we construct the nimrep graph G = Gρ [26]. Since the functor F in Section 2.9 is defined by the annihilation and creation operators given by (31), and the incoming, outgoing (inverted) Y-fork operators given in (32), (33), we see that F is equivalent to the presenting map Z above. The embedding P A ⊂ P G is given by the image under F of the morphisms in the A2-Temperley-Lieb category. 0,0q = C and the sequence {qP ∅ 3 Hilbert series of the almost Calabi-Yau algebras In Section 3.1 we introduce an algebra A(G, W ) associated to a finite graph G (an SU(3) ADE graph or the McKay graph GΓ of finite subgroup Γ ⊂ SU(3)) which carries a cell system W . In the case where G is an SU(3) ADE graph, these algebras are called almost Calabi-Yau algebras and are shown to be isomorphic to the almost Koszul algebras Π of Section 2.9. We will determine a formula for the Hilbert series which counts the dimensions of these algebras in Section 3.2. 3.1 The algebras A(G, W ) In this section we introduce the algebra A(G, W ) associated to a finite graph G which carries a cell system W . For any finite directed graph G, let [CG, CG] denote the subspace of CG spanned by all commutators of the form xy − yx, for x, y ∈ CG. If x, y are paths in CG such that r(x) = s(y) but r(y) 6= s(x), then xy − yx = xy, so in the quotient CG/[CG, CG] the path xy will be zero. Then any non-cyclic path, i.e. any path x such that r(x) 6= s(x), will be zero in CG/[CG, CG]. If x = a1a2 · · · ak is a cyclic path in CG, then a1a2 · · · ak − aka1 · · · ak−1 = 0 in CG/[CG, CG], so a1a2 · · · ak is identified with aka1 · · · ak−1. Similarly, 23 x = a1a2 · · · ak is identified with every cyclic permutation of the edges aj, j = 1, . . . , k. So the commutator quotient CG/[CG, CG] may be identified, up to cyclic permutation of the arrows, with the vector space spanned by cyclic paths in G. One defines a derivation is over all indices j such that aj = a. Then for a potential Φ ∈ CG/[CG, CG], which is some linear combination of cyclic paths in G, we define the algebra ∂a : CG/[CG, CG] → CG by ∂a(a1 · · · an) =Pj aj+1 · · · ana1 · · · aj−1, where the summation A(CG, Φ) = CG/hρai, p=0 H p jitp, where the H p HA for A(CG, Φ) as HA(t) = P∞ which is the quotient of the path algebra by the two-sided ideal generated by the relations ρa = ∂aΦ ∈ CG, for all edges a of G. If Φ is homogeneous, we define the Hilbert series ji are matrices which count the dimension of the subspace {ixj x ∈ A(CG, Φ)p}, where A(CG, Φ)p is the subspace of A(CG, Φ) of all paths of length p, and i, j ∈ A(CG, Φ)0. Suppose A(CG, Φ) is a Calabi-Yau algebra of dimension d = 3 and that deg Φ = 3, that is, Φ is a linear combination of cyclic paths of length 3 on G. Then HA(t) is given by (2) [12, Theorem 4.6]. For G an SU(3) ADE graph or the McKay graph GΓ of a finite subgroup Γ ⊂ SU(3), we define a homogeneous potential Φ by [38, Remark 4.5.7]: Φ = Xa,b,c∈G1 W (△(a,b,c) s(a),s(b),s(c)) · △(a,b,c) s(a),s(b),s(c) ∈ CG/[CG, CG], (40) Now let G be an SU(3) ADE graph and let (x1, x2) ∈ End(cid:0)(CG)0(CG)p(CG)0(cid:1) be matrix for a cell system W [59], and we will denote by A(G, W ) the algebra A(CG, Φ). units, where x1, x2 ∈ (CG)p for p = 0, 1, . . . , as in Section 2.2, which act on CG by (x1, x2)y = δy,x2x1. Applying the functor F to the morphisms Up in A2−T L we obtain a representation of the Hecke algebra on CG given by (c.f. (32) and Figure 3): s(b),r(b),r(a1)) (xa1a2, xa3a4), φ−1 s(a1)φ−1 r(a2)W (△(b,a3,a4) s(b),r(b),r(a3))W (△(b,a1,a2) (41) F (Up) = Xx,ai,b where the summation is over all paths x of length p − 1 and edges ai, b of G such that the paths xa1a2, xa3a4 make sense. These operators were shown to satisfy the relations H1-H3 in [26]. Let hρaip denote the restriction of the ideal hρai in (CG)p, which is isomorphic to i=1 Im(F (Ui)) = Πp ∼= Σp, where Π, Σ are the graded algebras defined in Section 2.9. i=1 Im(F (Ui)). Then A(G, W )p ∼= (CG)p/Pp−1 Pp−1 3.2 Hilbert Series of A(G, W ) In this section we give the Hilbert series of the algebra A(G, W ) where G is an SU(3) ADE graph G with cell system W . In this case we will call A(G, W ) an almost Calabi- Yau algebra. When G = GΓ is the McKay graph of a finite subgroup Γ ⊂ SU(3), A(G, W ) is a Calabi-Yau algebra of dimension 3 [38, Theorem 4.4.6] and its Hilbert series is thus given by (2). Let NXN = {λ(p,l) 0 ≤ p, q, p + l ≤ k < ∞} denote a non-degenerately braided system of endomorphisms on a type III1 factor N, which is generated by ρ = λ(1,0) and its 24 conjugate ρ = λ(0,1), where the irreducible endomorphisms λ(p,l) satisfy the fusion rules of SU(3)k given in (4). Let N ⊂ M be a braided subfactor with nimrep G which realises the modular invariant ZG at level k, where G = Gρ, and let I = NXM . The nimrep G sends λ ∈ NXN to the graph Gλ. Let S = Lk p=0 F (λ(p,0)) = Σ. Then since A(G, W ) = Π ∼= Σ (see Section 3.1), p=0 λ(p,0), so that F (S) = Lk F (S) ∼= A(G, W ), (42) proof of Theorem 4.4.6 in [38]. The tensor algebra T ρ =L∞ that is, the algebra A(G, W ) is given by the module category associated to the inclusion N ⊂ M. The algebra S has another description, based on [55] in the context of SU(2), and the j=0 ρj is the free algebra in the category A2-T L at level k generated by ρ. Under the functor F defined by (19), T ρ maps to F (T ρ) = CG, c.f. Section 2.9. Let T ′ be the quotient of T ρ by the two-sided ideal hρi generated by ρ ⊂ ρ2. It can be shown inductively for each grade that T ′ j = Sj. For j = 0, 1, the result is trivial. Since ρ2 = ρ + λ(2,0) by (4), we have T ′ 2 = λ(2,0) = S2. Now consider j = 3. From (4) we obtain ρ3 = ρρ + λ(1,1) + λ(3,0). Since ρρ = id + λ(1,1), we see that the ideal hid + λ(1,1)i ⊂ hρi. Then λ(1,1) ∈ hρi and T ′ 3 = λ(3,0) = S3. The situation for general j is similar. Thus we see that S is the symmetric algebra given by the quotient of the tensor algebra T ρ by hρi, see also [38, (4.5.1)]. Let π, γ denote the composite morphisms π : λ(0,0) ֒→ ρ3 ֒→ T ρ and γ : ρ ֒→ ρ2 ֒→ T ρ. Then π maps under F to F (π) : CI → CG, which sends 1 ∈ CI to the potential Φ of (40), and F (γ) : F (ρ) → CG sends the reverse edge ea of a to the relation ρa = ∂aΦ. Then under F , S = T ρ/hρi is mapped to F (T ρ)/hF (ρ)i ∼= CG/hF (γ(ρ))i = CG/hρai = A(G, W ). Reversing the argument, exactness of F yields F (S) = F (T ρ)/hF (ρ)i which by the above discussion is isomorphic to CG/hF (γ(ρ))i = CG/hρai = A(G, W ). The fusion rules (4) yield the recursion relation λ(j+1,0) ⊕ (ρ⊗ λ(j−1,0)) = (ρ⊗ λ(j,0)) ⊕ λ(j−2,0), j = 2, 3, . . . , k − 1, and so each λ(j,0) can be written recursively in terms of the three irreducible endomorphisms ρ, ρ and λ(0,0) = id. Summing over all j, and using t to keep track of the grading, we find that S =Lk p=0 λ(p,0)tp satisfies (cid:0)S ⊕ (t2ρ ⊗ S) ⊕ tk+3λ(k,0)(cid:1) = λ(0,0) ⊕ (tρ ⊗ S) ⊕ t3S. G HA(t) − t3HA(t) = 1 − tk+3P, HA(t) − t∆GHA(t) + t2∆T Then applying F to (43) we obtain the following equation for the Hilbert series of A: (43) where P = Gλ(k,0). From P 3 = Gλ(k,0) ⊗ Gλ(k,0) ⊗ Gλ(k,0) = Gλ(0,0) = 1, we see that the matrix P is an automorphism of the graph G of order 3. Since each Gλ(j,0) can be written recursively in terms of Gρ = ∆G, Gρ = ∆T G and Gλ(0,0) = 1, the permutation P can be determined for each graph G separately using standard Mathematica computations, and we obtain the following result: Theorem 3.1 Let HA(t) denote the Hilbert series of A(G, W ), for an SU(3) ADE graph G with adjacency matrix ∆G, Coxeter number h = k + 3 and cell system W . Then HA(t) = 1 − P th 1 − ∆Gt + ∆T G t2 − t3 , 25 (44) where P is the permutation matrix corresponding to a Z3 symmetry of the graph. It is the identity for D(n), A(n)∗, n ≥ 5, E (8)∗, E (12) , l = 1, 2, 4, 5, and E (24). For the remaining graphs A(n), D(n)∗ and E (8), let V be the permutation matrix corresponding to the clockwise rotation of the graph by 2π/3. Then l P = V 2 V V 2n for for A(n), n ≥ 4, for D(n)∗, n ≥ 5. E (8), The numerator and denominator in (44) commute, since any permutation matrix which We warn that we have not yet realised the graph E (12) corresponds to a symmetry of the graph G commutes with ∆G and ∆T G . subfactor, as we have not been able to construct a cell system on E (12) has claimed that the graph E (12) above proof would hold in this case also. as the nimrep produced by a . However, Ocneanu does have a cell system built on it [59], and hence the 4 4 4 In [13, Proposition 3.14] the Hilbert series was given for a (p, q)-Koszul algebra (or almost Koszul algebra), where the permutation matrix P is equal to the product of the permutation matrices given by the Nakayama permutations for A and its Koszul dual Λ. It was shown in Section 2.9 that A = A(G, W ) is a (h − 3, 3)-Koszul algebra. Then the Nakayama permutation for the coalgebra Λ being trivial is equivalent to Nakayama permutation of A being given by the permutation matrix P . 4 Nakayama automorphism for SU (3) ADE graphs When G is an SU(3) ADE graph, the dual A∗ = Hom(A, C) of the algebra A = A(G, W ) is identified as an A-A bimodule with 1Aβ, with standard left action and the right ac- tion twisted by an automorphism β, the Nakayama automorphism. In this section we determine the Nakayama automorphism β in Theorem 4.6. 4 The explicit cell systems W computed in [25] and knowledge of the Hilbert series (44) for A are key ingredients in both these results. Thus these results are not proven for the SU(3) ADE graph E (12) , since we were not able to compute an explicit cell system W for this graph in [25]. In the remainder of the paper, any reference to an SU(3) ADE graph will not include the graph E (12) We begin with some preliminary results, including Proposition 4.4 whose lengthy proof will be the content of Section 4.1. This section is based closely on [13, Section 4.2], which is in the context of preprojective algebras of the ADE graphs in SU(2). . 4 Let A = A(G, W ) for an SU(3) ADE graph G and let Ap denote its pth graded part. The edges a ∈ G1 are a basis for A1. With the potential Φ defined in (40), A has a relation ρa for each edge a ∈ G1 given by W (a,b,b′) s(a),r(a),r(b)bb′, (45) ρa = Xb,b′∈G1 where W (a,b,b′) s(a),r(a),r(b) := W (△(a,b,b′) s(a),r(a),r(b)). Let h = k + 3 denote the Coxeter number of G. The image of the endomorphism λ(k,0) under the functor F defines a unique permutation ν of the graph G, which is described 26 as follows. If the permutation matrix P in Theorem 3.1 is the identity matrix, then the permutation ν on the graph G is just the identity. For the other graphs, the permutation ν is given on the vertices of G by the permutation matrix P and on G1 by the unique permutation on the edges of G such that s(ν(a)) = ν(s(a)) and r(ν(a)) = ν(r(a)) (note that there are no double edges on the graphs G for which P is non-trivial). For any vertex i of G, it can be seen from the Hilbert series of A(G, W ) that the space i·Ah−3·ν(i) is one-dimensional. For k a vertex for which there is an edge from k to i, it can be seen from the Hilbert series of A(G, W ) that the space i· Ah−4· ν(k) is one-dimensional, except where there is a double edge from k to i on G, in which case dim(i·Ah−4·ν(k)) = 2. Let uiν(i) denote a generator of i · Ah−3 · ν(i), and let {vm iν(k), m ∈ {1, pki}} denote a basis a−→ j for i· Ah−4· ν(k), where pki denotes the number of edges from k to i. For each edge i and each edge k b−→ i on G there are non-zero scalars λ(m) such that , µ(m′) a b jν(i) = uiν(i) = µ(m′) a avm λ(m) for any m ∈ {1, pij}, m′ ∈ {1, pki}. Definition 4.1 We call a cell system W on G ν-invariant if W (a,b,c) all triangles i vm′ iν(k)ν(b), b a−→ j b−→ k c−→ i on G. (46) i,j,k = W (ν(a),ν(b),ν(c)) ν(i)ν(j)ν(k) for Proposition 4.2 Let W be a ν-invariant cell system on G. There is a constant C such that a λ(m′) λ(m) a µ(m′) µ(m) λ(m′′) b µ(m′′) c c b = C (47) a−→ j b−→ k c−→ i on G, and all m ∈ {1, pij}, m′ ∈ {1, pjk}, m′′ ∈ for all triangles i {1, pki}. Proof : The dual A∗ of A is an A-A bimodule with the products ϕx, xϕ defined by (ϕx)(y) = ϕ(xy), (xϕ)(y) = ϕ(yx), for ϕ ∈ A∗ and x, y ∈ A. The element dual to vm jν(i) is (vm ν(i)i is the element dual to uiν(i), since ν(i)ia, where s(a) = i, r(a) = j, and u∗ ν(i)j)∗ = λ(m) a u∗ (vm ν(i)j)∗(vm jν(i)) = λ(m) a (u∗ ν(i)ia)(vm jν(i)) = λ(m) a u∗ ν(i)i(avm jν(i)) = u∗ ν(i)i(uiν(i)) = 1. Similarly, the element µ(m) Then (46) dualises to give a′ ν(a′)u∗ ν(j)j, where s(a′) = i, r(a′) = j, is also dual to vm jν(i). a u∗ λ(m) ν(i)ia = (vm ν(i)j)∗ = µ(m) a′ ν(a′)u∗ ν(j)j. Let a = a′ and m ∈ {1, pij}. Then multiplying on the right by (λ(m) for a vertex k such that there is a triangle i a a−→ j (48) )−1W (a,b,c) ijk b in (48) b−→ k −→ i on G, we have a µ(m′) µ(m) a λ(m′) λ(m) ν(a)ν(b)u∗ W (a,b,c) ijk b b ν(k)k, W (a,b,c) ijk u∗ ν(i)iab = µ(m) a λ(m) a W (a,b,c) ijk ν(a)u∗ ν(j)jb = 27 where the second equality follows from (48) for a choice of m′ ∈ {1, pjk}. Summing over c−→ i on G, and all vertices j and edges a, b such that there is a triangle i making a choice of m = mj, m′ = m′ j for each j, the L.H.S. is zero by the relation ρc in (45), and so we obtain σ(m)u∗ b−→ k a−→ j ν(k)k = 0 where σ(m) =Xj,a,b µ(mj ) a µ λ(mj ) a λ (m′ b (m′ b j ) j ) W (a,b,c) ijk ν(a)ν(b), 1, m2, m′ and m = (m1, m′ r) where r is the number of vertices j in the sum- mation. Suppose σ(m) 6= 0. Then there exists a non-zero v ∈ k · Ah−5 · ν(i) such that ν(k)k = V ∗ 6= 0 (since v 6= 0) which is a vσ(m) = ukν(k) ∈ k · Ah−3 · ν(k), and σ(m)u∗ contradiction. Thus σ(m) = 0, which implies 2, . . . , mr, m′ a µ(m′) µ(m) a λ(m′) λ(m) b b W (a,b,c) ijk = ξikW (ν(a),ν(b),ν(c)) ν(i)ν(j)ν(k) , (49) where ξik ∈ C does not depend on j, a or b. For j = j1, the left hand side of (49) only depends on m1 and m′ 1, whilst for j = j2, the left hand side of (49) now only depends on m2 and m′ ξik, which only depends on i, k and m′′. Then since W is ν-invariant, we have 2. Thus ξik does not depend on m, but only on i, k. Define ξ′ ′′ ) ik := µ(m λ(m′′) c c a µ(m′) µ(m) a λ(m′) λ(m) b µ(m′′) c λ(m′′) c b = ξ′ ik, (50) kj (which only depends on k, j and m′) and ξ′ which only depends on i, k and m′′. However, by a similar argument, the left hand side is also equal to ξ′ ji (which only depends on j, i and m). Hence the left hand side does not depend on i, j, k or the choices of m, m′ b−→ k c−→ i on G. (cid:3) and m′′. So we have ξ′ ik in (50) does not depend on the choice of m ∈ {1, pij}, we obtain as an ji =: 1/C for all triangles i a−→ j Since ξ′ kj = ξ′ ik = ξ′ immediate corollary: Corollary 4.3 For any ν-invariant cell system W on G, λ(m) a µ(m) a = λ(m′) a µ(m′) a a,a′ for any edge i a−→ j on G, and m, m′ ∈ {1, pij}. We will now define an alternative basis {va i −→ j and each edge k where λa, µa are non-zero scalars. This basis will be used in Section 5. For any double edge (a, a′) on G with s(a) = i, r(a) = j, if we set va λ(2) jν(i) and va′ a′ v2 jν(i) = uiν(i), b ∈ {a, a′}, where λb = εbλ(1) λbbvb −→ i on G we have λa′a′va jν(i), then we have a′va jν(i) − λ(2) a v2 a − λ(1) b /(λ(1) := λ(1) b λ(2) a′ λ(2) a v1 jν(i) b,b′ jν(i)} for i·Ah−4·ν(k) such that for any edges iν(k)ν(b′) = δb,b′uiν(i), jν(i) = δa,a′uiν(i), µb′vb jν(i) := λ(1) jν(i) = 0 = ava′ a′ v1 jν(i) − jν(i) and a′ ) and εa = 1, a λ(2) 28 εa′ = −1. Then from (46) and Corollary 4.3, we see that va a /(λ(1) and µbvb µ(1) a′ µ(2) jν(i)ν(b) = uiν(i), b ∈ {a, a′}, where µa = µ(1) a µ(2) a′ /(λ(1) Let va a µ(2) jν(i) = v1 a′ λ(2) a ). a′ − µ(1) jν(i), λa = λ(1) jν(i)ν(a′) = 0 = va′ a − µ(1) a′ µ(2) jν(i)ν(a) a′ ) and µa′ = a λ(2) double edge (a, a′) from i to j let vb b ∈ {a, a′}. Thus we have defined an alternative basis {va there for each edge i a when there is only one edge a from i to j, and if there is a jν(i), λb be as defined in the previous paragraph, where jν(i)} for i · Ah−4 · ν(k) such that a−→ j and each edge k b−→ i on G we have λaava jν(i) = uiν(i) = µbvb iν(k)ν(b), (51) where λa, µa are non-zero scalars. Since dim(i · Ah−3 · l) = 0 for all l 6= ν(i), we have bva jν(i) = 0 = va jν(i)ν(b) (52) when a 6= b. We will usually write vjν(i) for va j. Dualising (51) we thus get: jν(i) where there is only one edge a from i to λau∗ ν(i)ia = (va ν(i)j)∗ = µa′ν(a′)u∗ ν(j)j. (53) 4.1 Computation of the constant C In this section we compute the value of the constant C in Proposition 4.2. Proposition 4.4 Let G be an SU(3) ADE graph, which is not E (12) mutation ν of G defined in Section 4, the constant C in Proposition 4.2 is 1. 4 . Then for the per- The proof of Proposition 4.4 is done in a case-by-case method, where we will use the , we did not Ocneanu cells W (△) computed in [25]. For the graphs D(n), D(n)∗ and E (12) claim to have computed all cell systems up to equivalence in [25]. 1 · · · We will begin by describing the general strategy. We choose a vertex i of G which is the source of only one edge a, and similarly the range of only one edge b. We denote by j, k the range, source vertices of the edges a, b respectively. Note that there must necessarily be (at least one) edge c from j to k on G. We choose a non-zero path uiν(i) = [i j l1 l2 lh−6 ν(k) ν(i)] ∈ i · Ah−3 · ν(i) of length h − 3 from i to ν(i), where l1, . . . , lh−6 are vertices of G. We let the elements vjν(i) and viν(k) be the paths vjν(i) = [j l1 l2 · · · lh−6 ν(k) ν(i)] ∈ j·Ah−4·ν(i) and viν(k) = [i j l1 l2 · · · lh−6 ν(k)] ∈ i·Ah−4·ν(k). Then uiν(i) = avjν(i) = viν(k)ν(b) in A, and we have λa = µb = 1. Note that since j · Ah−4 · ν(i) and i · Ah−4 · ν(k) are now one-dimensional, we omit the notation m = 1 from λ(m) . We now form the paths vjν(i)ν(a) ∈ j · Ah−3· ν(j), bviν(k) ∈ k · Ah−3· ν(k), a′ and transform these using (46) so that we have , µ(m) a′ vjν(i)ν(a) = bviν(k) = λc µa µc λb cvkν(j), vkν(j)ν(c), 29 (54) (55) where the same path vkν(j) appears in the right hand side of both equalities. By Propo- sition 4.2, λc/µa = Cµcµb/λaλb = Cµc/λb, since λa = µb = 1, so that (54) becomes vjν(i)ν(a) = µc λb Ccvkν(j). (56) Then we obtain the value of µc/λb from (55), and substituting into (56) we can determine the value of C. For the graphs where ν is the identity, that the constant C which appears in (56) is 1 follows from the following considerations. There is a conjugation on the A graphs given by the conjugation on the representations of SU(3) given in Section 2.1. There is also a conjugation on the other SU(3) ADE graphs: The conjugation τ : NXN → NXN on the braided system of endomorphisms of SU(3)k on a factor N, given by the conjugation on the representations of SU(3), induces a conjugation τ : NXM → NXM such that Gλ = τ Gλτ , where Gλa = λa for λ ∈ NXN , a ∈ NXM . We call a path symmetric if it is invariant under reversal of the path and taking the conjugate of the graph, i.e. a path x = [a1 a2 · · · al−1 al] is symmetric if xc := [al al−1 · · · a2 a1] = x, where v denotes the image of the vertex v under conjugation of the graph. We note that for a path to be symmetric, it must start at a vertex of colour p and end at a vertex of colour −p mod 3. We choose a non-zero path uiν(i) which is symmetric. We then form the path vjν(i)ν(a) as above, and transform this to a scalar multiple d1 of cvkν(j), where vkν(j) is a basis path in k·Ah−4·ν(j). We also find a non-zero path c′v′ ∈ j·Ah−3·ν(j), where v′ ∈ k·Ah−4·ν(j) is a symmetric path of length h− 4, and transform this to a scalar multiple d2 of cvkν(j). Then we obtain that vjν(i)ν(a) = d1d−1 2 c′v′. Now, since uiν(i) is symmetric, (vjν(i)ν(a))c = bviν(k). Then bviν(k) = (vjν(i)ν(a))c = (d1d−1 2 v′ν(c′) = d1vkν(j)ν(c). So we see that the same coefficient d1 appears in both (55) and (56), so that C = 1. However, this method will not work for the graph E (12) even though it has ν = id, since for all the possible choices for the vertices j, k, one of these must be of colour 0, hence the conjugation of the graph does not interchange j ↔ k. , E (12) and E (24) it is not clear that chosen paths are non-zero in l·Ah−3·ν(l), for l = i, j, k, and we will need to construct a basis for the space of all paths which start at the vertices i, j (and also k in the case of E (12) ). The computations of these basis paths are lengthy and are contained in the Appendix to [31]. We will summarize the results for the computation of C for these graphs here, with the detailed computations given in the Appendix to [31]. In what follows we will use the notation ars to denote the edge from r to s where this edge is unique. For the graphs E (12) 2 c′v′)c = d1d−1 , E (12) 5 5 5 1 2 The identity A graphs: We will first compute the value of the constant C for the graphs A(n). The infinite graph A(∞), illustrated in [25, Figure 4], is the fusion graph for the irreducible represen- tations of SU(3), whilst for finite n, the graph A(n) is the fusion graph for the irreducible representations of SU(3)n, as described in Section 2.1. We will write (p, q) for the irre- ducible representation λ(p,q). For A(n) the automorphism ν is the clockwise rotation of the graph by 2π/3. Choosing the vertices i = (0, 0), j = (1, 0) and k = (0, 1), we have ν(i) = (n − 3, 0), ν(j) = (n − 4, 1) and ν(k) = (n − 4, 0). The unique cell system W (up to equivalence) was computed in [25, Theorem 5.1], and we use the same notation for the cells here. The cell system W is ν-invariant. 30 Figure 6: Two paths in j · Ah−3 · ν(j) for A(n): vjν(i)ν(aij) on the left and aijvkν(j) on the right. Figure 7: Two paths in k · Ah−3 · ν(k) for A(n): akiviν(k) on the left and vkν(j)ν(aij) on the right. There is only one possible non-zero path of length h − 3 from i = (0, 0) to ν(i) = (n − 3, 0), which is given by uiν(i) = [(0, 0) (1, 0) (2, 0) · · · (n − 4, 0) (n − 5, 0)]. Any path of length > h − 3 will be zero in A since using the relations on A it can be transformed to a path which begins [i j k · · · ] = 0 since [i j k] = ρaki = 0 in A, where ars is the edge from r to s. Then vjν(i) = [(1, 0) (2, 0) · · · (n − 4, 0) (n − 5, 0)] and viν(k) = [(0, 0) (1, 0) (2, 0) · · · (n − 4, 0)]. We form vjν(i)ν(aij) and using the relation ρν(ajk ) we obtain the path [(1, 0) (2, 0) · · · (n − 4, 0) (n − 5, 1) (n − 4, 1)], as shown on the left hand side of Figure 6, with coefficient −W∇(n−5,0)/W△(n−4,0), where we use the notation W△(k,m) = W(k,m)(k+1,m)(k,m+1) and W∇(k,m) = W(k+1,m)(k,m+1)(k+1,m+1). Continuing in this way we obtain the path ajkvkν(j) = [(1, 0) (0, 1) (1, 1) (2, 1) · · · (n − 4, 1)], as shown on the right hand side of Figure 6, with coefficient ξ = (−1)n−4 W∇(0,0)W∇(1,0) . . . W∇(n−5,0) W△(1,0)W△(2,0) . . . W△(n−4,0) . Similarly, we form akiviν(k) and transform using the relations to obtain akiviν(k) = (−1)n−4 W∇(0,0)W∇(1,0) . . . W∇(n−5,0) W△(0,0)W△(1,0) . . . W△(n−5,0) ajkvkν(j) = ξ W△(n−4,0) W△(0,0) ajkvkν(j) = ξ ajkvkν(j), where the last equality follows from the ν-invariance of the cell system W . Thus we see that C = 1 for the graphs A(n). The only properties of the cells W (△) that are used here are their ν-invariance and the fact that they are non-zero. The orbifold D graphs: We will now consider the graphs D(n), which are Z3 orbifolds of the graphs A(n). The graph D(9) is illustrated in Figure 8. The weights W (△) for A(n) are invariant under the Z3 symmetry of the graph given by rotation by 2π/3. Thus there is an orbifold solution 31 Figure 8: Graph D(9) 2 1 1 1 ,i(1) 2 ,i(2) 3 1 ,i(0) 2 ,i(1) 3 1 ,i(2) 2 ,i(0) 3 k , i(1) k , i(2) ) = W (△i(2) ) = W (△i(1) 1 → i(j2) 2 → i(j3) ) for A(n), where i(0) on A(n) which lie on a closed loop of length three i(j1) for the cell system W on D(n) where the weights W (△) are given by the corresponding weights for A(n) [25, Theorems 6.1 & 6.2]. More precisely, excluding triangles △ which contain one of the triplicated vertices (k, k)l in the case where n = 3k + 3, the weight W (△i1,i2,i3) for the triangle △i1,i2,i3 = i1 → i2 → i3 → i1 on D(n) is given by the weight W (△i(0) k are the three vertices of A(n) which are identified under the Z3 action to give the vertex ik of D(n), k = 1, 2, 3. If for a triangle △i1,i2,i3 on D(n) there is no choice of vertices i(j1) , i(j2) , i(j3) 3 → i(j1) , then 3 we have W (△i1,i2,i3) = 0. When n = 3k + 3, the weight W (△) for a triangle △ which contain one of the triplicated vertices (k, k)l is just given by one third of the weight for the corresponding triangle on A(3k+3). Thus the relations (45) for D(n) are given precisely by the relations for A(n), except for the relations ργ, ργ′ in the case where n = 3k + 3, which involve the triplicated vertices (k, k)l. However, these last two relations are not used to show C = 1, and thus the result for D(n) follows from the result for A(n) under the orbifold procedure. The vertex (n− 3, 0) of A(n) is identified with the distinguished vertex (0, 0) of A(n). Thus with i the distinguished vertex of D(n) with lowest Perron-Frobenius weight, the element uiν(i) is a closed loop of length n − 3 starting and ending at i. We see that the permutation ν must be the identity for D(n). We will illustrate the general D case by giving the computations for the graph D(9). We label the vertices as in Figure 8, and denote the two edges in the double edge by γ and γ′. We write Ws(a)s(b)s(c) for W (△(abc) s(a)s(b)s(c) when one of a, b or c is the edge η ∈ {γ, γ′}. We note that W γ′ 369 = 0 [25, Theorem 6.2]. We choose the vertices i = 1, j = 4 and k = 7. Then the path uiν(i) is obtained from the corresponding element for A(9), so uiν(i) = u11 = [1483571] which is symmetric. We have vjν(i) = v41 = [483571], viν(k) = v17 = [148357]. Since ν = id, we only need to transform vjν(i)ν(a) = [4835714], where a is the edge from i to j, to a scalar multiple of a path γ′v′ where v′ is some symmetric path. We will underline at each stage the subpath of length 2 which we transform using the relations. We have s(a)s(b)s(c) where a, b, c 6∈ {γ, γ′}, and W η 269 = W γ [4835714] = = [4835724] = W247 W147 W247W259W γ′ 369W268 W147W257W359W368 [4835924] = W247W259W γ′ 369 W147W257W359 W247W259W γ′ 369W268W247 W147W257W359W368W248 [4836′924] [4726′924], W247W259 W147W257 [4826′924] = 32 Figure 9: Graphs A(7)∗, A(8)∗ Figure 10: Graph D(7)∗ where [69], [6′9] denote the edge from 6 to 9 along γ, γ′ respectively. The path [4726′924] is non-zero, with [726′924] symmetric. Here it is important that the cells W (△) for some of the triangles which contain one of the double edges are zero, otherwise the path vjν(i) = v41 = [483571] might be zero in A. The only other property of the cells W (△) that is used here is the fact that those that appear in the coefficient of [4726′924] in the above equality are non-zero. If there was another inequivalent cell system on D such that this coefficient is non-zero and the Hilbert series of A was given by (44), then the result C = 1 would hold for this cell system also. The conjugate A∗ and conjugate orbifold D∗ graphs: The proof for these graphs is slightly different to that for all the other graphs in that we do not choose i to be a vertex which is the source of only one edge. Here the vertex i is chosen to be the source of exactly two edges. First consider the graphs A(n)∗, where ν = id. The unique cell system W (up to equivalence) was computed in [25, Theorems 7.1, 7.3 & 7.4], and we use the same notation for the cells here. The A(n)∗ graphs are illustrated in [25, Figure 11]. We illustrate the cases n = 7, 8 in Figure 9. The labelling we use here for the vertices of A(2m+1)∗ is the reverse of the labelling used in [25]. The relations in A(A(n)∗, W ) are W112[121] + W111[111] = 0, Wa−1,a,a[a(a − 1)a] + Wa,a,a[aaa] + Wa,a,a+1[a(a + 1)a] = 0, Wa,a,a+1[aa(a + 1)] + Wa,a+1,a+1[a(a + 1)(a + 1)] = 0, Wa,a,a+1[(a + 1)aa] + Wa,a+1,a+1[(a + 1)(a + 1)a] = 0, (57) (58) (59) where a = 2, . . . , p − 1 in (57), and a = 1, . . . , a′ in (58), (59), where p = ⌊(n − 1)/2⌋, a′ = p − 1 for even n, and a′ = p − 2 for odd n. For even n we have the extra relation Wp−1,p,p[p(p− 1)p] + Wp,p,p[ppp] = 0, and for odd n we have the extra relation [p(p− 1)(p− 1)] = [(p − 1)(p − 1)p] = 0. We first consider the even case n = 2m + 2. We choose the vertices i = j = k = 1 as illustrated in Figure 9. The element uiν(i) = u1ν(1) of length 2m + 1 is (up to some scalar) [111 · · · 1]. We show this is non-zero by induction. When m = 1, we see using the relations that [1111] = −(W112/W111)[1121] = (W 2 111)[1221] = −(W112/W111)[1211], so that all paths of length 3 are equal, up to scalar multiple, in A. We assume that all non-zero paths in A(A(2m+2)∗, W ) are proportional to the path u(k) 1ν(1) = [111 · · · 1] of length 2m− 1 for m = k. For m = k + 1, any path in A(A(2k+4)∗, W ) of length 2k + 1 must have one of the following forms, where a{j} = [v0v1 · · · v2k−1] is a path of length 2k − 1 with 112/W 2 33 (I) [a{1}11] = −W112/W111[a{1}21], (II) [a{2}11] = −W122/W112[a{2}21], v2k−1 = j: (III) [a{3}21]. Any path of form (I) is clearly proportional to u(k+1) 1ν(1) = [111 · · · 1] of length 2k−1, since any path a{1} in A(A(2k+4)∗, W ) of length 2k−3 must be proportional to u(k) 1ν(1) (we note that a non-zero coefficient when m = k may become zero when m = k+1 since we replace the quantum integer [s]q(k) by [s]q(k+1) in the weights W , where q(m) = e2πi/m). We now consider paths of form (II). There are three cases: (a) Suppose v2k−4 = 1 in a{2}. Then using the relations we see that [a{2}11] = −W111/W112[a{1}11], where a{1} is obtained from a{2} by replacing its last edge by a closed loop from 1 to 1. This path is now of form (I). (b) If v2k−4 = 2, then we have [a{2}11] = −W112/W122[a{1}11], where now a{1} is obtained from a{2} by replacing its last edge by an edge from 2 to 1, and this new path is of form (I). (c) If v2k−4 = 3, then we move to consider v2k−5. Then we will have three cases similar to (a)-(c). If we are in case (c) we consider v2k−6, and continuing in this way get v2k−l−1 ≤ v2k−l for some l, and we are in case (a) or (b). Finally, consider paths of form (III). Suppose v2k−4 = 2. Then using the relations we obtain [1 · · · v2k−52321] = − W222 W223 [1 · · · v2k−52221] − W122 W223 [1 · · · v2k−52121] = W112W222 W122W223 [1 · · · v2k−52211] + W111W122 W112W223 [1 · · · v2k−52111] = (cid:18)W111W122 W112W223 − W 2 W 2 112W222 122W223(cid:19) [1 · · · v2k−52111], and we are in case (I). If v2k−4 = 3, 4, we proceed as in (II). Thus any path in A(A(2k+4)∗, W ) of length 2k + 1 is equal to ξu(k+1) 1ν(1) = ξ[111 · · · 1], for some ξ ∈ R (possibly zero). Thus for A(2m+2)∗ we choose uiν(i) = u1ν(1) = [111 · · · 1], and we have vjν(i) = viν(k) = [111 · · · 1] of length 2m − 2. Then vjν(i)ν(aij) = [111 · · · 1] = u1ν(1) = ajkvkν(j), where vkν(j) = v1ν(1) = [111 · · · 1]. Similarly akiviν(k) = vkν(j)ν(ajk). Thus we obtain C = 1. The situation for A(2m+1)∗ follows similarly. The graphs D(n)∗ are (three-colourable) unfolded versions of the graphs A(n)∗, where we replace every vertex v of A(n)∗ by three vertices v0, v1, v2, where va is of colour a, such that there are edges v0 → w1, v1 → w2 and v2 → w0 if and only if there is an edge from v to w on A(n)∗. The graph D(7)∗ is illustrated in Figure 10. Then the proof for D(n)∗ follows exactly as the proof for A(n)∗ where now we write a suffix for each vertex indicating the colour of that vertex. Due to the three-coloured nature of the graphs D(n)∗, we see that the maximum path uiν(i) = u10ν(10) of length n − 3 must end at the vertex ν(10) = 1r, where r ≡ n mod 3. Thus the permutation ν should be the identity, clock- wise rotation of the graph by 2π/3, or anticlockwise rotation of the graph by 2π/3, for n ≡ 0, 1, 2 mod 3 respectively. Since ν is not always the identity for the graphs D(n)∗ we require ν-invariance of the cells here. The orbifold cell systems computed in [25, Theo- rems 8.1 & 8.2] are ν-invariant. If another inequivalent ν-invariant cell system could be found such that uiν(i) = u10ν(10) is non-zero and the Hilbert series of A was given by (44), then the result C = 1 would hold for this cell system also. 34 Figure 11: Graph E (8) Figure 12: Graph E (8)∗ The graphs E (8), E (8)∗ for the conformal embeddings SU(3)5 ⊂ SU(6)1 and SU(3)5 ⊂ SU(6)1 ⋊ Z3 [70, 5, 26]: We will first consider the graph E (8)∗, illustrated in Figure 12. The unique cell system W (up to equivalence) was computed in [25, Theorem 10.1], and we use the same notation for the cells here. The quotient algebra A has the relations [123] = [231] = [324] = [432] = 0, [232], [333] = −W233 W333 [312] = [322] + [332], [323], [222] = −W223 W222 −W123 W223 −W234 W223 [243] = [223] + [233], (60) (61) (62) (63) where W222 = −W333 and W223 = W233. We choose the vertices i = 1, j = 2 and k = 3. We choose the path [122331] of length 5 to be the path uiν(i), which is nonzero since [122331] = −(W243/W223)[124331] using (63) and (60), and no relations can be used on [124331] except for the one which transforms it back to [122331]. The only other relation which can be used on [122331] is [122331] = −(W243/W233)[122431]. Then vjν(i)ν(aij) = v2ν(1)ν(a12) = [223312], which we transform to the (non-zero) path [233122]: [223312] = − = −2 W223 W123 W222 W123 [223322] − [222222], W233 W123 [223332] = W223 W123 [223222] + W 2 233 W123W333 [223232] where in the penultimate equality we also used (60) and in the last equality we used (61) once for [223222] and twice for [223232]. Similarly, [233122] = − W223 W123 [233222] − W233 W123 [233322] = −2 W222 W123 [222222], so that vjν(i)ν(aij) = [223312] = [233122] = ajkv′, where v′ = [33122] is symmetric. Here we need the fact that the coefficient of [222222] in both the above equalities is not zero for the cell system W , as well as the fact that they are non-zero. Since the graph E (8), illustrated in Figure 11, is the (three-colourable) unfolded version of E (8)∗, the result for E (8) follows in the same way as the result for D(2n+1)∗ follows from A(2n+1)∗. The unique cell system W (up to equivalence) was computed in [25, Theorem 35 Figure 13: Graphs E (12) 1 and E (12) 2 9.1]. Due to the three-coloured nature of E (8), we see that since uiν(i) = u10ν(10) has length 5, ν must be the non-trivial permutation which sends 10 → 12. Since ν is not the identity for the graph E (8) we require ν-invariance of the cells here. The cell system W for E (8), computed in [25, Theorem 9.1] is ν-invariant. 1 1 1 1 1 ijk ϕ(i)ϕ(j)ϕ(k) = W +(ϕ(a)ϕ(b)ϕ(c)) for the conformal embedding SU(3)9 ⊂ (E6)1 [70, 6, 26]: Two inequivalent solutions for the cell system W ± for the graph E (12) The graph E (12) , illustrated in Figure 13, were computed in [25, Theorem 12.1]. We will use the solution W +. The solution W − is obtained from W + by taking the conjugation of the graph E (12) , that is, W −(abc) where ϕ is the map which reflects the graph E (12) about the plane which passes through the vertices i1, i2, i3 and p, and reverses the direction of each edge. For E (12) the automorphism ν is the identity. We choose the vertices i = is, j = js and k = ks, for some s ∈ {1, 2, 3}. We first computed a basis for the space of paths which start from the vertices is, js. The explicit details of these computations are given in the Appendix to [31]. We will denote by [rp], [r′p] the path which goes along the edge α, α′ respectively, and similarly by [pq], [p′q] the path which goes along the edge β, β′ respectively. Let uiν(i) = uisν(is) be the (non-zero) path uiν(i) = [isjsrpqrpqksis] of length 9. Using the computations contained in the Appendix to [31], we obtain vjν(i)ν(aij) = [jsrpqrpqksisjs] = d[jskspqksisjsrpjs] = dajkv′, where v′ = [kspqksisjsrpjs] is symmetric and d is a non-zero scalar. Hence C = 1. In the computations for the graph E (12) we have used the orbifold cell system which was constructed explicitly in [25]. If another inequivalent cell system could be found which satisfied the non-vanishing of many coefficients which appear in the computations of a basis for A, and such that the Hilbert series of A was given by (44), then the result C = 1 would hold for this cell system also. 1 2 2 The graph E (12) for the conformal embedding SU(3)9 ⊂ (E6)1 ⋊ Z3 [9, 26]: The graph E (12) . Every cell system W for E (12) is equivalent to either a cell system W + or the inequivalent cell system W − [25, Theorem 11.1]. We will use the solution W +. The solution W − is obtained from W + by taking the conjugation of the graph E (12) where , illustrated in Figure 13, is a Z3 orbifold of E (12) , that is, W −(abc) = W +(ϕ(a)ϕ(b)ϕ(c)) ϕ(i)ϕ(j)ϕ(k) ijk 2 1 1 36 Figure 14: Labelled graph E (12) 5 2 2 2 ϕ is the map which reflects the graph E (12) about the plane which passes through the vertices i, p1, p2 and p3, and reverses the direction of each edge. For the graph E (12) the automorphism ν is the identity. We choose i, j, k as labelled on the graph E (12) in Figure 13. We first computed a basis for the space of paths which start from the vertices i, k. The explicit details of these computations are given in the Appendix to [31]. Here q = e2πi/12, and we will write [m] for the quantum integer [m]q. We let a±, b± denote the scalars which is symmetric. This path is non-zero in A since a± = q[2][4] ±p[2][4], b± = q[2]2 ±p[2][4]. Let uiν(i) be the path [ijr1p1jkp1q1ki], p[3][4] µ! [ijr1p1jkp1q1ki] = −a+[ijr1p1q1kp1q1ki] = −b−[ijr1p1q1r2p1q1ki]. p[2] Then vjν(i)ν(aij) is given by [kijr1p1jkp1q1k] = d[kp1q1kp1jr1p1jk] = dajkv′, where v′ is symmetric and d is a non-zero scalar, and we obtain C = 1. + a− 1 5 5 M S The Moore-Seiberg invariant ZE (12) is realised by a braided subfactor which produces for the twisted orbifold D(12) invariant: The graph E (12) the nimrep E (12) [26, Section 5.4], illustrated in Figure 14. The unique cell system W (up to equivalence) was computed in [25, Theorem 13.1]. For the graph E (12) the automorphism ν is the identity. There are four possible choices for the vertices j, k: these are (3,8), (5,9), (14,3) and (15, 4). We see that it will not be possible to use the quicker method described above for the case where ν = id, since conjugating the graph will not interchange the vertices j ↔ k for any of the choices of j, k. We choose the vertices i = 10, j = 15 and k = 4. We first computed a basis for the space of paths which start from the vertices 10, 15, 4. The explicit details of these computations are given in the Appendix to [31]. Here q = e2πi/12, and we will write [m] for the quantum integer [m]q. 5 We choose uiν(i) = u10ν(10) = [10, 15, 2, 9, 16, 5, 8, 14, 4, 10]. Then v15ν(10)ν(a10,15) = [15, 2, 9, 16, 5, 8, 14, 4, 10, 15] = c[15, 4, 7, 12, 1, 7, 13, 1, 7, 15] = ca15,4v4ν(15), 37 Figure 15: Labelled graph E (24) where c = −[2]p[3]3(cid:16)p[6]3 + [3][4](cid:17) /[4]2[6], and C = 1 follows from a4,10v10ν(4) = [4, 10, 15, 2, 9, 16, 5, 8, 14, 4] = c[4, 7, 12, 1, 7, 13, 1, 7, 15, 4] = cv4ν(15)ν(a15,4). The graph E (24) for the conformal embedding SU(3)21 ⊂ (E7)1 [70, 6, 26]: For the graph E (24), illustrated in Figure 15, the automorphism ν is the identity. The unique cell system W (up to equivalence) was computed in [25, Theorem 14.1]. We choose the vertices i = 1, j = 9 and k = 17. We first computed a basis for the space of paths which start from the vertices 1, 9. The explicit details of these computations are given in the Appendix to [31]. Here q = e2πi/24, and we will write [m] for the quantum integer [m]q. Let u be the path [1, 9, 18, 6, 13, 22, 4, 14, 19, 2, 11, 19, 2, 11, 22, 5, 14, 21, 3, 10, 17, 1], which is symmetric, and is non-zero as shown in the Appendix to [31]. We choose uiν(i) = u1ν(1) := u. Then vjν(i)ν(aij) = v9ν(1)ν(a1,9) is given by [9, 18, 6, 13, 22, 4, 14, 19, 2, 11, 19, 2, 11, 22, 5, 14, 21, 3, 10, 17, 1, 9] [9, 17, 2, 11, 20, 5, 14, 21, 3, 10, 17, 2, 9, 18, 6, 136, 22, 4, 12, 19, 2, 9] ajkv′, = − = − [4][9][10]p[2][3]5 [7]2p[5]3 [4][9][10]p[2][3]5 [7]2p[5]3 where v′ is symmetric. Then C = 1. (cid:3) 4.2 Determining the Nakayama automorphism for A(G, W ) We will now determine the Nakayama automorphism for the algebras A = A(G, W ), where G is an SU(3) ADE graph which is not E (12) The algebra A = A(G, W ) is a Frobenius algebra, that is, there is a linear function f : A → C such that (x, y) := f (xy) is a non-degenerate bilinear form (this is equivalent . 4 38 to the statement that A is isomorphic to its dual A∗ = Hom(A, C) as left (or right) A-modules). There is an automorphism β of A, called the Nakayama automorphism of A (associated to f ), such that (x, y) = (y, β(x)). Then there is an A-A bimodule isomorphism A∗ → 1Aβ [72]. Using the notation of Section 3, we will define a non- degenerate form on A by setting f to be the function which is 0 on every element of A of length < h− 3, and 1 on uiν(i), for some i ∈ G1. Then using the relation (x, y) = (y, β(x)) this determines the value of f on ujν(j), for all other j ∈ G1. We will normalize the ujν(j) such that f (ujν(j)) = 1 for all j ∈ G1. From equation (51) we see that a∗ = λava jν(i), jν(i))∗ = µaν(a), where a is an edge from i to j on G. (va Definition 4.5 Let β denote the automorphism of A defined on G by β = ν, where ν is the permutation defined on G in Section 4. Let W be a ν-invariant cell system on G. With β defined above we have β (ρa) = Xb,b′ W (a,b,b′) i,r(b),k ν(b)ν(b′) = Xb,b′ W (ν(a),ν(b),ν(b′ )) ν(i)ν(r(b))ν(k) ν(b)ν(b′) = ρν(a), for a an edge from k to i. The following lemma shows that β is the Nakayama automor- phism for A: Theorem 4.6 The A-A bimodule A∗ contains an element u∗ such that A∗ = Au∗ = u∗A, and u∗a = β(a)u∗ for all a ∈ A. Thus A∗ is isomorphic to 1Aβ as an A-A bimodule. Proof : As in [13, Corollary 4.7], any u∗ =Pi ζiu∗ β(a)u∗ for each a ∈ A1. By Definition 4.5 this becomesPl ζlu∗ ν(i)i with non-zero ζi ∈ C will generate A∗ as both a left and right module. We will show that the ζi can be chosen such that u∗a = ν(l)l, so ν(i)ia by (53). Thus we need to choose the ζi ν(l)la = ν(a)Pl ζlu∗ ν(j)j = ζj(λa/µa)u∗ ν(i)ia = ζjν(a)u∗ that ζiu∗ such that 1 = ζj ζi λa µa . (64) Similarly, for any triangle i Then (64) for a and b gives a−→ j b−→ k c−→ i on G, we need (64) with a replaced by b. 1 = ζi ζk µaµb λaλb = ζi ζk λc µc , by Propositions 4.2 and 4.4. Thus (64) will also be satisfied with i, j replaced by k, i respectively. We may then consider a maximal connected subgraph G′ of G such that for all vertices i, j on G′, with an edge i → j, there is a choice of non-zero ζi, ζj such that (64) is satisfied. Suppose G′ 6= G so that there is a vertex k of G′ which is the source of an edge γ to a vertex l which is not on G′. If we choose ζl = (µc/λc)ζk then (64) is now satisfied for all vertices i, j on the graph G′′, which is the graph obtained from G′ by adjoining the vertex l and the edge c. This contradicts the assumption that G′ is the maximal subgraph such that (64) is satisfied, hence G′ = G. Then there is an A-A bimodule isomorphism φ : 1Aβ → A∗ given by φ(a) = au∗, such that φ(xaβ(y)) = xaβ(y)u∗ = xau∗y = xφ(a)y, for all a ∈ 1Aβ, x, y ∈ A. Corollary 4.7 For every a ∈ G1 we have λa = µa. Proof : Using equation (51), 1 = λa(a, va jν(i)) = λa(va jν(i), β(a)) = λa/µa. (cid:3) (cid:3) 39 5 A finite resolution of A as an A-A bimodule In this section we determine a finite resolution of A = A(G, W ) as an A-A bimodule, where G is an SU(3) ADE graph. Let A = A(G, W ), where G is either an SU(3) ADE graph or the McKay graph GΓ of a finite subgroup Γ ⊂ SU(3), and W is a cell system on G. Consider the following complex: (65) µ2−→ A ⊗R V ⊗R A µ1−→ A ⊗R A µ0−→ A −→ 0, A ⊗R A[3] where the A-A bimodules R, V and eV are given by R = (CG)0, V = (CG)1 and eV = La∈G1 Cea. We denote by B[m] the space B but with grading shifted by m. The A-A bimodule eV is isomorphic to the dual space of V . The connecting A-A bimodule maps are given by µ3−→ A ⊗R eV ⊗R A[1] µ0(x ⊗ y) = xy, µ1(x ⊗ a ⊗ y) = xa ⊗ y − x ⊗ ay, µ2(x ⊗ea ⊗ y) = Xb,b′∈G1 µ3(x ⊗ y) = Xa∈G1 x ⊗ea ⊗ ay, where x, y ∈ A, a ∈ G1 and we denote by Wabb′ the cell W (△(a,b,b′) xa ⊗ea ⊗ y −Xa∈G1 s(a),s(b),s(b′)). Wabb′(xb ⊗ b′ ⊗ y + x ⊗ b ⊗ b′y), The sequence (65) is exact and the kernel of µ3 is zero if and only if A is a Calabi- Yau algebra of dimension 3 [12, Theorem 4.3], e.g. if G is the McKay graph of a finite subgroup of SU(3) -- see [30] for these subgroups and their graphs. When G is an SU(3) ADE graph, we will see that the kernel of µ3 is non-zero. As in [14], applying the functor F = −⊗A R to the two-sided complex (65), we obtain the following one-sided complex of right A modules: A[3] F (µ3) −→ A ⊗R eV [1] F (µ2) −→ A ⊗R V F (µ1) −→ A F (µ0) −→ R −→ 0, (66) where the connecting A module maps are given by F (µ0) the projection of A onto its zero- Wabb′xb ⊗ b′ and F (µ3)(x) = setting to show that the full complex (65) is exact if and only if the one-sided complex (66) is exact. graded part R, F (µ1)(x ⊗ a) = xa, F (µ2)(x ⊗ea) =Pb,b′∈G1 Pa∈G1 xa⊗ea, where y ∈ A, a ∈ G1. The proof of [14, Proposition 7.2.1] carries over to our It was shown in Section 2.9 that the one-sided sequence (39) is exact, for k ≤ h − 2. Thus the bottom row of diagram (67) is shown to be exact everywhere except at the first term A in degree h − 3, where the Koszul differentials di are given by d1(x ⊗ a) = xa, There are isomorphisms ψi such that the following diagram commutes: Wabb′xb ⊗ b′, and d3(x) =Pa∈G1 d2(x ⊗ea) = √µs(a)µr(a) (√µr(a)/√µs(a))xa ⊗ea. F (µ3) −1Pb,b′∈G1 −→ A ⊗R eV d3−→ A ⊗R eV ↓ψ2 A ↓ψ3 A F (µ1) −→ A ↓ψ0 d1−→ A F (µ2) −→ A ⊗R V ↓ψ1 d2−→ A ⊗R V 40 (67) where ψ0 = ψ1 = id, ψ2(x ⊗ea) = √µs(a)µr(a) x ⊗ea and ψ3(x) = µr(x)x. Then the top row of diagram (67) is exact everywhere except at the first term A in degree h − 3, thus the complex (65) is shown to be exact everywhere except at the first term A⊗R A[3]. We will now compute the kernel of the map µ3, that is, the fourth syzygy Ω4(A) of A. The proof of the following theorem is based on the proof of Theorem 4.9 in [13]. Theorem 5.1 Let G be an SU(3) ADE graph which is not E (12) , and W be a ν-invariant cell system on G. The fourth syzygy Ω4(A) = Ker(µ3) of A = A(G, W ) is isomorphic to 1Aβ−1 as an A-A bimodule, where β is the Nakayama automorphism defined in Definition 4.5. Thus 4 0 → 1Aβ−1[h] µ4−→ A⊗R A[3] µ1−→ A⊗R A is a finite resolution of A as an A-A bimodule, where µ4(x) = xPj xj ⊗ x∗ µ3−→ A⊗ReV ⊗R A[1] homogeneous basis for A and x∗ µ2−→ A⊗R V ⊗R A j is its dual basis under the non-degenerate form on A. µ0−→ A → 0 (68) j , the xj are a Proof : Since A is a (h − 3, 3)-Koszul ring (see Section 2.9), by [13, Theorem 3.15] we see that Ω4(A) is generated both as a left A-module and as a right A-module by its component Z with total degree h, where Z ⊆ (A ⊗R A[3])h = Ar ⊗R Ah−3−r, h−3Xr=0 and that Z → A0 ⊗R Ah−3[3] is an R-R bimodule isomorphism. Since A0 ∼= R and Ah−3 ∼= 1Rν as R-R bimodules, we see that Z ∼= 1Rν as an R-R bimodule, and has a basis given by elements wiν(i) which project onto i ⊗ uiν(i) in the one-dimensional subspaces i · A0 ⊗R Ah−3 · ν(i) of A0 ⊗R Ah−3. Since µ3(wiν(i)) = 0 we must have wiν(i) = iXj xj ⊗ x∗ j , (69) where the xj are a homogeneous basis for A and x∗ j degenerate form on A, i.e. xjx∗ Ah−2−r in µ3(wiν(i)), for r = 1, . . . , h − 3: is its dual basis under the non- j = us(xj)ν(s(xj )). To see this, consider the terms in Ar ⊗R r−1 y∈Bi ′∈Bi r Xa∈G1 where Bi x ∈ C. Then for each y′ ∈ Bi λya r, ya ⊗ea ⊗ y∗ − Xa∈G1 k denotes a basis of i· Ak. Now ya =Px∈Bi y, a(y′)∗) = ( Xy∈Bi ( Xy∈Bi so that a(y′)∗ =Py∈Bi ya ⊗ea ⊗ y∗ − Xy′∈Bi Xy∈Bi y′ ⊗ea ⊗ a(y′)∗ = Xy∈Bi ya, (y′)∗) = ( Xy∈Bi λya λya x∈Bi r y ′∈Bi r r−1 r−1 r−1 r−1 r−1 r−1 r 41 λya y′ y∗. Then for fixed a ∈ G1 we have y′ ⊗ea ⊗ a(y′)∗, y x x for any y ∈ Bi λya r−1, a ∈ G1, where r x x, (y′)∗) = Xy∈Bi λya x δy′,x = Xy∈Bi r−1 λya y′ , r−1 x∈Bi r y′ y′ ⊗ea ⊗ y∗ − Xy∈Bi ′∈Bi r y r−1 y′ y∗ = 0, y′ ⊗ea ⊗ λya as required. Let w = Pi∈G0 wiν(i). Then we have Z = Rw = wR and by [13, Theorem 3.15] Ω4(A) = Aw = wA. Then there is an automorphism γ of A such that xw = wγ(x) If for all x ∈ A, which we will assume to have degree zero since w is homogeneous. we take x = i ∈ A0 we see from (69) that γ(i) = ν(i) = β(i), and so we must have γ(a) = γaν(a) = γaβ(a) for some γa ∈ C \ {0}. If we now let x = ν(a) ∈ A1, we obtain wiν(i)ν(a) = γν(a)awjν(j). Then by (69) and comparing the terms in A1 ⊗R Ah−3, we have Xb∈G1 λbb ⊗ vb lν(i)ν(a) = γν(a)a ⊗ ujν(j), using the identification b∗ = λbvb lν(i). Using (51), (52) on the left hand side we obtain (λa/µa)a ⊗ ujν(j) = γν(a)a ⊗ ujν(j). Thus γν(a) = λa/µa = 1 for all a ∈ G1, using Corollary 4.7, and we have γ = β as required. Then there is an A-A bimodule isomorphism φ : 1Aβ−1 → Ω4(A) given by φ(a) = aw, such that φ(xaβ−1(y)) = xaβ−1yw = xawy = xφ(a)y, for all a ∈ 1Aβ−1, x, y ∈ A. The isomorphism φ thus defines an A-A bimodule map µ4 : 1Aβ−1[h] −→ A ⊗R A[3] given by µ4(x) = xPj xj ⊗ x∗ Acknowledgements This work was supported by the Marie Curie Research Training Network MRTN-CT-2006-031962 EU-NCG. The authors would like to thanks Alastair King and the referees for their careful reading of earlier versions of this paper and their comments which have greatly improved the exposition. j . (cid:3) References [1] R. E. Behrend, P. A. Pearce, V. B. Petkova and J.-B. Zuber, Boundary conditions in rational conformal field theories, Nuclear Phys. B 579 (2000), 707 -- 773. [2] J. Bion-Nadal, An example of a subfactor of the hyperfinite II1 factor whose principal graph invariant is the Coxeter graph E6, in Current topics in operator algebras (Nara, 1990), 104 -- 113, World Sci. Publ., River Edge, NJ, 1991. [3] J. Bockenhauer, Lecture at Warwick Workshop on Modular Invariants, Operator Algebras and Quotient Singularities. September 1999. [4] J. Bockenhauer and D. E. Evans, Modular invariants, graphs and α-induction for nets of subfactors. I, Comm. Math. Phys. 197 (1998), 361 -- 386. [5] J. Bockenhauer and D. E. Evans, Modular invariants, graphs and α-induction for nets of subfactors. II, Comm. Math. Phys. 200 (1999), 57 -- 103. [6] J. Bockenhauer and D. E. Evans, Modular invariants, graphs and α-induction for nets of subfactors. III, Comm. Math. Phys. 205 (1999), 183 -- 228. [7] J. Bockenhauer and D. E. Evans, Modular invariants from subfactors: Type I coupling matrices and intermediate subfactors, Comm. Math. Phys. 213 (2000), 267 -- 289. [8] J. Bockenhauer and D. E. Evans, Modular invariants and subfactors, in Mathematical physics in mathematics and physics (Siena, 2000), Fields Inst. Commun. 30, 11 -- 37, Amer. Math. Soc., Providence, RI, 2001. 42 [9] J. Bockenhauer and D. E. Evans, Modular invariants from subfactors, in Quantum sym- metries in theoretical physics and mathematics (Bariloche, 2000), Contemp. Math. 294, 95 -- 131, Amer. Math. Soc., Providence, RI, 2002. [10] J. Bockenhauer, D. E. Evans and Y. Kawahigashi, On α-induction, chiral generators and modular invariants for subfactors, Comm. Math. Phys. 208 (1999), 429 -- 487. [11] J. Bockenhauer, D. E. Evans and Y. Kawahigashi, Chiral structure of modular invariants for subfactors, Comm. Math. Phys. 210 (2000), 733 -- 784. [12] R. Bocklandt, Graded Calabi Yau algebras of dimension 3, J. Pure Appl. Algebra 212 (2008), 14 -- 32. [13] S. Brenner, M. C. R. Butler and A. D. King, Periodic algebras which are almost Koszul, Algebr. Represent. Theory 5 (2002), 331 -- 367. [14] N. Broomhead, Dimer models and Calabi-Yau algebras, PhD thesis, University of Bath, 2008. [15] A. Cappelli, C. Itzykson, C. and J.-B. Zuber, The A-D-E classification of minimal and A(1) 1 conformal invariant theories, Comm. Math. Phys. 113 (1987), 1 -- 26. [16] B. Cooper, Almost Koszul Duality and Rational Conformal Field Theory, PhD thesis, University of Bath, 2007. [17] W. Crawley-Boevey and M. P. Holland, Noncommutative deformations of Kleinian singu- larities, Duke Math. J. 92 (1998), 605 -- 635. [18] P. Di Francesco and J.-B. Zuber, SU(N ) lattice integrable models associated with graphs, Nuclear Phys. B 338 (1990), 602 -- 646. [19] K. Erdmann and N. Snashall, On Hochschild cohomology of preprojective algebras. I, II, J. Algebra 205 (1998), 391 -- 412, 413 -- 434. [20] K. Erdmann and N. Snashall, Preprojective algebras of Dynkin type, periodicity and the second Hochschild cohomology, in Algebras and modules, II (Geiranger, 1996), CMS Conf. Proc. 24, 183 -- 193, Amer. Math. Soc., Providence, RI, 1998. [21] P. Etingof and V. Ostrik, Module categories over representations of SLq(2) and graphs, Math. Res. Lett. 11 (2004), 103 -- 114. [22] D. E. Evans, Critical phenomena, modular invariants and operator algebras, in Operator algebras and mathematical physics (Constant¸a, 2001), 89 -- 113, Theta, Bucharest, 2003. [23] D. E. Evans and Y. Kawahigashi, Orbifold subfactors from Hecke algebras, Comm. Math. Phys. 165 (1994), 445 -- 484. [24] D. E. Evans and Y. Kawahigashi, Quantum symmetries on operator algebras, Oxford Math- ematical Monographs. The Clarendon Press Oxford University Press, New York, 1998. Oxford Science Publications. [25] D. E. Evans and M. Pugh, Ocneanu Cells and Boltzmann Weights for the SU (3) ADE Graphs, Munster J. Math. 2 (2009), 95 -- 142. arXiv:0906.4307 [math.OA]. 43 [26] D. E. Evans and M. Pugh, SU (3)-Goodman-de la Harpe-Jones subfactors and the realisa- tion of SU (3) modular invariants, Rev. Math. Phys. 21 (2009), 877 -- 928. arXiv:0906.4252 [math.OA]. [27] D. E. Evans and M. Pugh, A2-Planar Algebras I, Quantum Topol. 1 (2010), 321-377. arXiv:0906.4225 [math.OA]. [28] D. E. Evans and M. Pugh, A2-Planar Algebras II: Planar Modules, J. Funct. Anal. 261 (2011), 1923 -- 1954. arXiv:0906.4311 [math.OA]. [29] D. E. Evans and M. Pugh, Spectral Measures and Generating Series for Nimrep Graphs in Subfactor Theory, Comm. Math. Phys. 295 (2010), 363 -- 413. arXiv:0906.4314 [math.OA]. [30] D. E. Evans and M. Pugh, Spectral Measures and Generating Series for Nimrep Graphs in Subfactor Theory II: SU (3), Comm. Math. Phys. 301 (2011), 771-809. arXiv:1002.2348 [math.OA]. [31] D. E. Evans and M. Pugh, The Nakayama automorphism of the almost Calabi-Yau algebras associated to SU (3) modular invariants. arXiv:1008.1003 (math.OA). [32] K. Fredenhagen, K.-H. Rehren and B. Schroer, Superselection sectors with braid group statistics and exchange algebras. II. Geometric aspects and conformal covariance, Rev. Math. Phys. Special issue (1992), 113 -- 157. [33] J. Frohlich and F. Gabbiani, Braid statistics in local quantum theory, Rev. Math. Phys. 2 (1990), 251 -- 353. [34] J. Fuchs, I. Runkel and C. Schweigert, Twenty five years of two-dimensional rational con- formal field theory, J. Math. Phys. 51 (2010), 015210, 19 pages. [35] T. Gannon, The classification of affine SU(3) modular invariant partition functions, Comm. Math. Phys. 161 (1994), 233 -- 263. [36] T. Gannon, Private communication, 2001. [37] I. M. Gel'fand and V. A. Ponomarev, Model algebras and representations of graphs, Funk- tsional. Anal. i Prilozhen. 13 (1979), 1 -- 12. [38] V. Ginzburg, Calabi-Yau algebras, 2006. arXiv:math/0612139 [math.AG]. [39] F. M. Goodman, P. de la Harpe and V. F. R. Jones, Coxeter graphs and towers of algebras MSRI Publications, 14, Springer-Verlag, New York, 1989. [40] F. M. Goodman and H. Wenzl, Ideals in the Temperley-Lieb Category. Appendix to Freed- man, Michael H., A magnetic model with a possible Chern-Simons phase, Comm. Math. Phys. 234 (2003), 129 -- 183. [41] T. Graves, Representations of affine truncations of representation involutive-semirings of Lie algebras and root systems of higher type, MSc thesis, University of Alberta, 2010. [42] M. Izumi, Application of fusion rules to classification of subfactors, Publ. Res. Inst. Math. Sci. 27 (1991), 953 -- 994. [43] M. Izumi, On flatness of the Coxeter graph E8, Pacific J. Math. 166 (1994), 305 -- 327. 44 [44] M. Izumi, Subalgebras of infinite C ∗-algebras with finite Watatani indices. II. Cuntz- Krieger algebras, Duke Math. J. 91 (1998), 409 -- 461. [45] M. Jimbo, A q-analogue of U (gl(N + 1)), Hecke algebra, and the Yang-Baxter equation, Lett. Math. Phys. 11 (1986), 247 -- 252. [46] V. F. R. Jones, Index for subfactors. Invent. Math. 72 (1983), 1 -- 25. [47] V. F. R. Jones, The planar algebra of a bipartite graph, in Knots in Hellas '98 (Delphi), Ser. Knots Everything 24, 94 -- 117, World Sci. Publ., River Edge, NJ, 2000. [48] V. G. Kac, Infinite-dimensional Lie algebras, Third Edition, Cambridge University Press, Cambridge, 1990. [49] L. H. Kauffman, State models and the Jones polynomial, Topology 26 (1987), 395 -- 407. [50] Y. Kawahigashi, On flatness of Ocneanu's connections on the Dynkin diagrams and classi- fication of subfactors, J. Funct. Anal. 127 (1995), 63 -- 107. [51] A. Kirillov, Jr. and V. Ostrik, On a q-analogue of the McKay correspondence and the ADE classification of sl2 conformal field theories, Adv. Math. 171 (2002), 183 -- 227. [52] H. Kosaki, Extension of Jones' theory on index to arbitrary factors, J. Funct. Anal. 66 (1986), 123 -- 140. [53] G. Kuperberg, Spiders for rank 2 Lie algebras, Comm. Math. Phys. 180 (1996), 109 -- 151. [54] R. Longo, Index of subfactors and statistics of quantum fields. II. Correspondences, braid group statistics and Jones polynomial, Comm. Math. Phys. 130 (1990), 285 -- 309. [55] A. Malkin, V. Ostrik and M. Vybornov, Quiver varieties and Lusztig's algebra, Adv. Math. 203 (2006), 514 -- 536. [56] S. Morrison, E. Peters and N. Snyder, Skein Theory for the D2n Planar Algebras, J. Pure Appl. Algebra 214 (2010) 117 -- 139. arXiv:0808.0764 [math.OA]. [57] A. Ocneanu, Quantized groups, string algebras and Galois theory for algebras, in Operator algebras and applications, Vol. 2, London Math. Soc. Lecture Note Ser. 136, 119 -- 172, Cambridge Univ. Press, Cambridge, 1988. [58] A. Ocneanu, Paths on Coxeter diagrams: from Platonic solids and singularities to minimal models and subfactors. (Notes recorded by S. Goto), in Lectures on operator theory, (ed. B. V. Rajarama Bhat et al.), The Fields Institute Monographs, 243 -- 323, Amer. Math. Soc., Providence, R.I., 2000. [59] A. Ocneanu, Higher Coxeter Systems (2000). Talk given at MSRI. http://www.msri.org/publications/ln/msri/2000/subfactors/ocneanu. [60] A. Ocneanu, The classification of subgroups of quantum SU(N ), in Quantum symmetries in theoretical physics and mathematics (Bariloche, 2000), Contemp. Math. 294, 133 -- 159, Amer. Math. Soc., Providence, RI, 2002. [61] T. Ohtsuki and S. Yamada, Quantum SU(3) invariant of 3-manifolds via linear skein theory, J. Knot Theory Ramifications 6 (1997), 373 -- 404. 45 [62] V. Ostrik, Module categories, weak Hopf algebras and modular invariants, Transform. Groups 8 (2003), 177 -- 206. [63] M. Reid, La correspondance de McKay, Ast´erisque (2002), 53 -- 72. S´eminaire Bourbaki, Vol. 1999/2000. [64] L. C. Suciu, The SU (3) Wire Model, PhD thesis, The Pennsylvania State University, 1997. [65] V. G. Turaev, Quantum invariants of knots and 3-manifolds, de Gruyter Studies in Math- ematics, 18, Walter de Gruyter & Co., Berlin, 1994. [66] M. van den Bergh, Non-commutative crepant resolutions, in The legacy of Niels Henrik Abel, 749 -- 770, Springer, Berlin, 2004. [67] A. Wassermann, Operator algebras and conformal field theory. III. Fusion of positive energy representations of LSU(N ) using bounded operators, Invent. Math. 133 (1998), 467 -- 538. [68] H. Wenzl, On sequences of projections, C. R. Math. Rep. Acad. Sci. Canada 9 (1987), 5 -- 9. [69] H. Wenzl, Hecke algebras of type An and subfactors, Invent. Math. 92 (1988), 349 -- 383. [70] F. Xu, New braided endomorphisms from conformal inclusions, Comm. Math. Phys. 192 (1998), 349 -- 403. [71] S. Yamagami, A categorical and diagrammatical approach to Temperley-Lieb algebras. arXiv:math/0405267 [math.QA]. [72] K. Yamagata, Frobenius algebras, in Handbook of algebra, Vol. 1, 841 -- 887, Elsevier, Ams- terdam, 1996. [73] J.-B. Zuber, CFT, BCFT, ADE and all that, in Quantum symmetries in theoretical physics and mathematics (Bariloche, 2000), Contemp. Math. 294, 233 -- 266, Amer. Math. Soc., Providence, RI, 2002. 46 Appendix to: The Nakayama automorphism of the almost Calabi-Yau algebras associated to SU (3) modular invariants David E. Evans and Mathew Pugh School of Mathematics, Cardiff University, Senghennydd Road, Cardiff, CF24 4AG, Wales, U.K. A Computation of certain basis paths for A(G, W ) for the graphs E (12) 1 , E (12) 2 , E (12) 5 and E (24) 1 2 , E (12) Here we provide the computations for a basis for (a subspace of) the quotient path space A(G, W ) for the exceptional SU(3) ADE graphs E (12) and E (24). These computa- tions are needed for the proof of Proposition 4.4 in The Nakayama automorphism of the almost Calabi-Yau algebras associated to SU(3) modular invariants. Due to the length of these computations, it was not practical to include them in the original paper, thus for completeness they are reproduced here instead. We do not write down a basis for the en- tire quotient space A(G, W ), but rather only for the subspaces i· A(G, W ) and j · A(G, W ) for two vertices i, j of G, where i is a vertex which is the source of only one edge a of G, and j is the range of this vertex. For the graph E (12) 5 we also write a basis for the subspace k · A(G, W ), where k is the source of the unique edge b whose range is the vertex i. We also include details of the explicit computation of the constant C = 1 for each of these graphs. These computations were only summarised in [31]. For each graph G we set q = e2πi/h, where h is the Coxeter number of G, and we will write [m] for the quantum integer [m]q = (qm − q−m)/(q − q−1). E (12) 1 graph for the conformal embedding SU (3)9 ⊂ (E6)1 A.1 Two inequivalent solutions for the cell system W ± for the graph E (12) , illustrated in Figure 13, were computed in [25, Theorem 12.1]. We will use the solution W +. The solution W − is obtained from W + by taking the conjugation of the graph E (12) , that is, W −(abc) where ϕ is the map which reflects the graph E (12) about the plane which passes through the vertices i1, i2, i3 and p, and reverses the direction of each edge. For E (12) the automorphism ν is the identity. We choose i = is, j = js and k = ks, for some s ∈ {1, 2, 3}. We will write out a basis for the space of paths which start from the vertices is, js. We will denote by l the length of the paths. Suppose we have a basis for the paths of length k. We will consider every path of length k + 1 obtainable = W +(ϕ(a)ϕ(b)ϕ(c)) ϕ(i)ϕ(j)ϕ(k) ijk 1 1 1 1 47 by adding an extra edge to the end of each of the basis paths of length k. We will first list the basis paths of length k + 1. These will be followed by computations (contained within parentheses) showing how all the other paths obtained can be written in terms of these basis paths. We will mark basis paths with an asterisk, e.g. [isjsrpq]∗. We will underline the subpaths of length 2 on which we have used a relation at each stage, and if a subpath of length k > 2 is underlined, this will indicate that we are using a relation on that path found when considering l = k. Often when a relation is used on a path b of length k, which gives b as a linear combination of basis paths bi of length k, some of these paths bi are easily shown to be zero because the subpath given by the first l < k edges of bi was shown to be zero when considering paths of length l. When this is the case, we will usually omit the paths bi which are known to be zero in this way. We will denote by [rp], [r′p] the path which goes along the edge α, α′ respectively, and similarly by [pq], [p′q] the path which goes along the edge β, β′ respectively. Let a±, b± denote the scalars ∓b∓), and ǫs the third root a± =q[2][4] ±p[2][4], b± =q[2]2 ±p[2][4], c± = (a2 of unity ǫs = e2πi(s−1)/3 for s = 1, 2, 3. ±b±)/(a2 Paths starting at vertex is: Paths [isjsr] ([isjsks] = 0) [isjsrp] = −ǫs a− a+ [isjsrpjs] = −ǫs [isjsr′p] − ǫlp[3][4] p[2] [isjsr′pjs], a− a+ (cid:18)[isjsrpjs±1] = −ǫs±1 (cid:18)[isjsrp′q] = b+ b− a− a+ [isjsr′pq] = −ǫs 1 a+ [isjsksp] = −ǫs a− a+ [isjsr′p] [isjsrpq] [isjsr′pjs±1] = ǫsǫs±1[isjsrpjs±1] ⇒ [isjsrpjs±1] = 0(cid:19) a+b+ a−b− [isjsrpq]∗(cid:19) [4] [2] b+ a− 5 [isjsrpqks], [isjsrpjsr] = ǫs [isjsrpqr] [isjsrpjsks] = −ǫs p[2] p[3][4] (cid:18)[isjsrpqks±1] = −ǫs a+ a− a−[isjsrpqks] − ǫs p[2] p[3][4] [isjsrp′qks±1] = c+[isjsrpqks±1] a+[isjsrp′qks] = −ǫs p[2] p[3][4] ⇒ [isjsrpqks±1] = 0(cid:19) a−(1 + c+)[isjsrpqks]∗! [isjsrpqrp] = −ǫs [isjsrpqksp] [2] [4] a+ b− b+ b− [isjsrpqrpq] (cid:16)[isjsrpqksis] = 0(cid:17) (cid:18)[isjsrp′q] = [isjsrpqrpjs±1] = −ǫs±1 ⇒ [isjsrpqrpjs±1] = 0! (cid:18)[isjsrpqr′p] = ǫs [2] [4] a− b+ [isjsrpqksp] = −ǫs a−b− a+b+ [isjsrpqrp]∗(cid:19) [isjsrpqr′pq] = ǫs [2] [4] a− b− [isjsrpqkspq] = −ǫs [isjsrpqrpq]∗(cid:19) a− a+ a2 − a− a+ [isjsrpqr′pjs±1] = −ǫsǫs±1 [2] [4] a+b+ [isjsrpkspjs±1] = −ǫsǫs±1c−[isjsrpqrpjs±1] (cid:18)[isjsrpqrpjs] = −ǫs [2] [4] a+ b− [isjsrpqkspjs] = 0(cid:19) [isjsrpqrp′qks±1] = ǫsǫs±1[isjsrpqrpqks±1] ⇒ [isjsrpqrpqks±1] = 0(cid:19) l 2 3 4 6 7 8 [isjsrpqrpqks] (cid:16)[isjsrpqrpqr] = 0(cid:17) (cid:16)[isjsrpqrpqksp] = 0(cid:17) a+ a− (cid:18)[isjsrpqrpqks±1] = −ǫs±1 (cid:16)[isjsrpqrpqksisjs] = 0(cid:17) [isjsrpqrpqksis] 9 10 48 Figure 16: Graphs E (12) 1 and E (12) 2 Paths starting at vertex js: l 2 3 4 Paths [jsrp], [jsr′p] ([jsksis] = 0) [jsrpjs′ ] = −ǫs′ [jsr′pjs′ ] (s′ ∈ {s, s ± 1}), a− a+ [jsksp′q] − ǫs [jsr′p ′ q] = −ǫsp[3][4] p[2] 1 a− b− a− a+ [jsrpqks′ ] [jsrp′qks], (s′ ∈ {s, s ± 1}), = −ǫs[jsrpq]∗ − ǫs (1 + c+)[jsr′pq](cid:19) [jsrpjsks] = −ǫs p[2] a−[jsrpqks]∗ − ǫs p[2] p[3][4] p[3][4] [jsrpjs±1ks±1] = −ǫs±1 p[2] p[3][4] a−[jsrpqks±1] − ǫs±1 p[2] p[3][4] a−[jsrpqks±1] + p[2] p[3][4] = −ǫs±1 p[2] p[3][4] = −ǫs±1 p[2] p[3][4] a− (ǫsc+ − ǫs±1) [jsrpqks±1] − ǫs p[2] = p[2] p[3][4] p[3][4] = p[2] a− (ǫsc+ − ǫs±1) [jsrpqks±1] − ǫs p[2] p[3][4] p[3][4] = p[2] p[3][4] ⇒ [jsrpjs±1ks±1] = p[2] p[3][4] [jsrpqks±1] − ǫs±1p[3][4] p[2] [jsrp′qks±1] = −ǫs±1 Now [jsrp′qr] = −ǫs [jsrpjsr] − ǫs+1 a2 +b+ a−b− a− a+ a+ b− [2] [4] a+ b− [2] [4] 1 a+ [jsksp] = −ǫs p[2] p[3][4] [jsrp′q] = p[3][4] p[2] [jsrpq], a+ a− a+[jsrp]∗ − ǫs p[2] p[3][4] [jsrp′q] = [jsr′pq] b+ b− a−[jsr′p]∗! 1 a+ [jskspq] − ǫs a+b+ a−b− [jsr′pq] [jsrpqr], [jsrp′qr] a+[jsrp′qks]∗! a−[jsrpqks±1] − ǫs±1 p[2] p[3][4] a+b+ [jsr′pqks±1] a+[jsrp′qks±1] [jsr′p′qks±1] + ǫs±1 a+b+ a−b− [jsr′pjs±1ks±1] a+b+ b− (1 + c+)[jsr′pqks±1] − c+[jsrpjs±1ks±1] a+(1 + c+)[jsrp′qks±1] − c+[jsrpjs±1ks±1] a− (ǫsc+ − ǫs±1 + ǫs±1ǫs(1 + c+)) [jsrpqks±1] + (ǫs±1ǫs + (ǫs±1ǫs − 1)c+) [jsrpjs±1ks±1] a−λ(4) s±1[jsrpqks±1]∗ where λ(4) s±1 = ǫsc+ − ǫs±1 + ǫs±1ǫs(1 + c+) (1 + c+)(1 − ǫs±1ǫs) ! a− a+ (cid:16)ǫs±1 − ǫs±1λ(4) s±1(cid:17) [jsrpqks±1]∗! [jsrpjs±1ks±1] = − [jsrpjs+1r] − ǫs−1 [2] [4] a+ b− [jsrpjs−1r] = −ǫs a+b+ a−b− [jsrpqr] + (ǫsǫs+1 − ǫs+1) [2] [4] a+ b− 49 [jsrpjs+1r] + (ǫsǫs−1 − ǫs−1) [2] [4] a+ b− [jsrpjs−1r], (†) but also [jsrp′qr] = = −ǫs b+ b− a− a+ [jsr′pqr] = ǫs [2] [4] a− b− [jsr′pjsr] + ǫs+1 [2] [4] a− b− [jsr′pjs+1r] + ǫs−1 [2] [4] a− b− [jsr′pjs−1r] [jsr′p′qr] + (ǫs+1 − ǫsǫs+1) [2] [4] a− b− [jsr′pjs+1r] + (ǫs−1 − ǫsǫs−1) [2] [4] a− b− [jsr′pjs−1r] = ǫs = ǫs a− a+ a− a+ [jsrpqr] + a2 − a2 + (1 + c+)[jsr′pqr] + (ǫs − ǫs+1) [2] [4] a− b− [jsrpjs+1r] + (ǫs − ǫs−1) [2] [4] a− b− [jsrpjs−1r] [jsrpqr] + (c− + 1) [jsrp′qr] + (ǫs − ǫs+1) [2] [4] a− b− [jsrpjs+1r] + (ǫs − ǫs−1) [2] [4] a− b− [jsrpjs−1r] ⇒ [jsrp′qr] = −ǫs a+b+ a−b− [jsrpqr] + (ǫs+1 − ǫs) [2] [4] a+ b− c+[jsrpjs+1r] + (ǫs−1 − ǫs) [2] [4] a+ b− c+[jsrpjs−1r]. (††) Then from (†), (††) we obtain [jsrpjs−1r] = µ[jsrpjs+1r] where µ = ǫs+1 − ǫsǫs+1 + (ǫs+1 − ǫs)c+ ǫsǫs−1 − ǫs−1 + (ǫs − ǫs−1)c+ . So [4] [2] b+ a− [jsrpqr] = ǫs[jsrpjsr] + ǫs+1[jsrpjs+1r] + ǫs−1[jsrpjs−1r] = ǫs[jsrpjsr] + (ǫs+1 + ǫs−1µ) [jsrpjs+1r], (†††) and [4] [2] b− a+ [jsrp′qr] = −ǫs[jsrpjsr] − ǫs+1[jsrpjs+1r] + −ǫs−1[jsrpjs−1r] = −ǫs[jsrpjsr] − (ǫs+1 + ǫs−1µ) [jsrpjs+1r], giving [jsrpjs+1r] = ǫs i 1 − µ b+ a− [jsrpqr]∗ + ǫs i 1 − µ b− a+ [jsrp′qr]∗! (cid:18)[jsrpjs−1r] = µ[jsrpjs+1r] = ǫs (cid:18)By (†††), [jsrpjsr] = ǫs b+ a− [4] [2] iµ 1 − µ b+ a− [jsrpqr]∗ + ǫs iµ 1 − µ b− a+ [jsrp′qr]∗(cid:19) [jsrpqr] − ǫs (ǫs+1 + ǫs−1µ) [jsrpjs+1r] = ǫs b+ a− (cid:18) [4] [2] + a+ a− 5 [jsrpqksis] = −ǫs i (ǫs+1 + ǫs−1µ) 1 − µ [jsrp′qksis], (cid:19) [jsrpqr]∗ + ǫs b− a+ i (ǫs+1 + ǫs−1µ) 1 − µ [jsrp′qr]∗(cid:19) [jsrpqksp], [jsrpqks+1p], [jsrpqks−1p] (cid:18)[jsrpqks±1is±1] = −ǫs±1 Now [jsrpqrp] = −ǫs a− b+ a+b+ a− = ǫsp[3][4] p[2] a+b+ (cid:16)ǫsλ(4) +ǫs+1ǫs [2] [4] a2 − = a+ a− [jsrp′qks±1is±1] =(cid:16)1 − ǫs ± 1λ(4) s±1(cid:17) [jsrpqks±1is±1] ⇒ [jsrpqks±1is±1] = 0(cid:19) i(1 − µ)[jsrpjs+1rp] − ǫs [jsrp′qrp] a−b− a+b+ i(1 − µ)[jsrpjs+1ks+1p] + 1 b+ i(1 − µ)[jsrpjs+1r′p] + [2] [4] a− b+ [jsrp′qksp] a− b+ [jsrp′qks+1p] + ǫs−1ǫs [2] [4] a− b+ [jsrp′qks−1p] [jsrp′qksp] − ǫs−1ǫs s+1(1 + i(1 − µ)) − ǫsǫs+1(cid:17) [jsrpqks+1p] − ǫs a+b+ (cid:16)ǫs−1 − ǫs−1λ(4) s+1(1 + i(1 − µ)) − ǫsǫs+1(cid:17) [jsrpqks+1p] − ǫs a2 − + [2] [4] a− b+ a2 − a+b+ (cid:16)ǫsλ(4) 1 a− +ǫs−1ǫs [jsrp′qks−1p] + [jsrp′qksp] − ǫs−1ǫs [2] [4] a− b+ [2] [4] 1 b+ s−1(cid:17) [jsrpqks−1p] = = 1 a− [jsrpqrp] − ǫs b− a+b+ [jsrp′qrp] [jsrpqrp] + ǫs+1ǫs [2] [4] 1 b+ [jsrp′qks+1p] a2 − a+b+ (cid:16)ǫs−1 − ǫs−1λ(4) s−1(cid:17) [jsrpqks−1p] a− a+b+ (cid:18) [2] [4](cid:16)ǫs+1ǫs − ǫsλ(4) s+1(cid:17) + a−ǫsλ(4) s+1(1 + i(1 − µ)) − a−ǫsǫs+1(cid:19) [jsrpqks+1p] − ǫs 1 a− −ǫs+1ǫs a− [2] [4] (a− + ǫs)(cid:16)ǫs−1 − ǫs−1λ(4) s−1(cid:17) [jsrpqks−1p] + [2] [4] [jsrp′qksp] [jsrpqrp] but also [jsrpqrp] = −ǫs [jsrpqksp] − ǫs+1 [jsrpqks+1p] − ǫs−1 [jsrpqks−1p] a− b+ [2] [4] a+ b− so that [jsrp′qksp] = −ǫs [jsrpqksp]∗ + λ(5) 1 [jsrpqks+1p]∗ + ǫsǫs−1 a+b+ [2] a+ [4] b− a+b+ a−b− a− a+ (cid:16)ǫs−1 − ǫs−1λ(4) s−1(cid:17) [jsrpqks−1p]∗ where λ(5) 1 = ǫsǫs+1 (cid:18)[jsrpqrp] = −ǫs [2] [4] a+ b− a− a+ (cid:16)ǫs+1 − ǫs+1λ(4) s+1(cid:17) − ǫs [jsrpqksp]∗ − ǫs+1 [2] [4] a+ b− [2] [4] a+ b− [4] [2] a2 − a+ i(1 − µ) ǫs + a− s+1.! λ(4) [jsrpqks+1p]∗ − ǫs−1 [2] [4] a+ b− 50 [jsrpqks−1p]∗(cid:19) [2] [4] (cid:18)[jsrpqr′p] = ǫs (cid:18)[jsrp′qrp] = −ǫs = ǫs [2] [4] a2 +b+ a−b2 − [jsrp′qr′p] = ǫs a− b+ [jsrpqksp]∗ + ǫs+1 [2] [4] a− b+ [jsrpqks+1p]∗ + ǫs−1 [2] [4] [2] [4] a+ b− [jsrp′qksp] − ǫs+1 [2] [4] a+ b− [jsrp′qks+1p] − ǫs−1 [jsrpqksp]∗ + λ(5) 2 [jsrpqks+1p]∗, where λ(5) 2 = [jsrpqks−1p]∗(cid:19) [jsrp′qks−1p] a− b+ a+ b− i(1 − µ) ǫs + a− s+1.! λ(4) [2] [4] a2 − b− [2] [4] a− b+ [jsrp′qksp] + ǫs+1 [2] [4] a− b+ [jsrp′qks+1p] + ǫs−1 [2] [4] a− b+ [jsrp′qks−1p] = − [2] [4] a+ b− [jsrpqksp]∗ + λ(5) 3 [jsrpqks+1p]∗ + λ(5) 4 [jsrpqks−1p]∗, where λ(5) 3 = (ǫsǫs+1 − ǫs+1) and λ(5) 4 = (ǫsǫs−1 − ǫs−1) [2] [4] [2] [4] a2 − s+1(cid:17) − ǫs a+b+ (cid:16)ǫs+1 − ǫs+1λ(4) s−1(cid:17) .! a+b+ (cid:16)ǫs−1 − ǫs−1λ(4) a2 − a3 − a+b+ i(1 − µ) ǫs + a− λ(4) s+1 6 [jsrpqksisjs] = −p[4] p[2] [jsrpqks+1pq] = −ǫs+1 (cid:18)From [jsrpqkspjs+1] = −ǫs [4] [2] b− a+ [jsrpqkspjs], [jsrpqkspjs±1], [jsrpqkspq] = −ǫs a+ a− [jsrpqks+1p′q] (cid:16)[jsrpqks±1pjs±1] = 0(cid:17) [jsrpqrpjs+1] − ǫsǫs−1[jsrpqks−1pjs+1] a+ a− [jsrpqksp′q], = ǫsǫs+1 [jsrpqr′pjs+1] − ǫsǫs−1[jsrpqks−1pjs+1] [4] [2] a−b− a2 + = ǫsǫs+1c−[jsrpqkspjs+1] + ǫsǫs−1 (c− − 1) [jsrpqks−1pjs+1] giving (c− − 1) [jsrpqks−1pjs+1] = (ǫsǫs−1 − c−) [jsrpqkspjs+1]∗! (cid:18)Now [jsrpqks+1pjs] = −ǫs+1 (Similarly (c− − 1) [jsrpqks+1pjs−1] = (ǫsǫs+1 − c−) [jsrpqkspjs−1]∗) b− a+ [4] [2] = ǫsǫs+1 [jsrpqr′pjs] − ǫsǫs+1[jsrpqkspjs] − ǫs−1ǫs+1[jsrpqks−1pjs] [4] [2] a−b− a2 + [jsrpqrpjs] − ǫsǫs+1[jsrpqkspjs] − ǫs−1ǫs+1[jsrpqks−1pjs] = ǫsǫs+1 (c− − 1) [jsrpqkspjs] + ǫsǫs+1c−[jsrpqks+1pjs] − (ǫs−1ǫs+1 − c−) [jsrpqks−1pjs], so that (1 − ǫsǫs+1c−) [jsrpqks+1pjs] = ǫsǫs+1 (c− − 1) [jsrpqkspjs] − (ǫs−1ǫs+1 − c−) [jsrpqks−1pjs]. (‡) But also [jsrpqks+1pjs] = − [jsrp′qrpjs] + ǫs 1 λ(5) 2 [2] [4] a2 +b+ a−b2 − 1 λ(5) 2 [jsrpqkspjs] [jsrp′qr′pjs] + ǫs [2] [4] a2 +b+ a−b2 − 1 λ(5) 2 [jsrpqkspjs] = ǫs a− a+ 1 λ(5) 2 = −ǫs [2] [4] a− b− (1 − c+) 1 λ(5) 2 [jsrpqkspjs] + ǫs a− a+ λ(5) 3 λ(5) 2 [jsrpqks+1pjs] + ǫs a− a+ λ(5) 4 λ(5) 2 [jsrpqks−1pjs], so that 1 − ǫs a− a+ 2 ! [jsrpqks+1pjs] = −ǫs λ(5) 3 λ(5) [2] [4] a− b− (1 − c+) 1 λ(5) 2 [jsrpqkspjs] + ǫs a− a+ λ(5) 4 λ(5) 2 [jsrpqks−1pjs]. (‡‡) Then from (‡), (‡‡) we obtain λ(5) 3 λ(5) 1 = 1 − ǫs where λ(6) a− a+ and λ(6) 2 = ǫs [2] [4] a− b− 1 λ(5) 2 λ(6) 1 [jsrpqks−1pjs] = λ(6) 2 [jsrpqkspjs]∗ a− a+ 2 ! (ǫs+1ǫs−1 − c−) + ǫs (1 − ǫsǫs+1c−) (1 − c+) − ǫsǫs+1 1 − ǫs λ(5) 4 λ(5) 2 (1 − ǫsǫs+1c−) a− a+ 2 ! (1 − c−).! λ(5) 3 λ(5) a+b+ a−b− [jsrpqkspq] − ǫsǫs−1 51 a− a+ (cid:16)ǫs−1 − ǫs−1λ(4) 1 (cid:17) [jsrpqks−1pq] From (‡), where λ(6) λ(5) [jsrpqks+1pjs] = λ(6) 3 [jsrpqkspjs]∗, 3 = −ǫsǫs+1(1 − c−) − (ǫs+1ǫs−1 − c−) 1 ! (1 − ǫsǫs+1c−)−1 .! λ(6) 2 λ(6) 1 [jsrpqks+1pq] = [jsrp′qkspq] + ǫs = −ǫsp[3][4] p[2] a− (cid:18) [4] b+ + = [2] a−b− +ǫs a− a+ 1 a+ [jsrpjskspq] − ǫs a− a+ (1 − c+)[jsrpqkspq] − ǫsǫs−1 = ǫs[jsrpjsrpq] + [jsrpjsr′pq] − ǫs (1 − c+)[jsrpqkspq] − ǫsǫs−1 iµ 1 − µ(cid:19) [jsrpqrpq] + ǫs iµ 1 − µ [jsrp′qrpq] + ǫs b+ a+ (cid:18) [4] [jsrp′qr′pq] − ǫs (1 − c+)[jsrpqkspq] − ǫsǫs−1 iµ a2 + 1 − µ a− a+ b− a+ a− a+ 1 (cid:17) [jsrpqks−1pq] 1 (cid:17) [jsrpqks−1pq] a− a− a+ (cid:16)ǫs−1 − ǫs−1λ(4) a+ (cid:16)ǫs−1 − ǫs−1λ(4) 1 − µ(cid:19) [jsrpqr′pq] a+ (cid:16)ǫs−1 − ǫs−1λ(4) a− [2] iµ + 1 (cid:17) [jsrpqks−1pq] 2 ! , λ(5) 3 − ǫs b− a+ λ(5) = λ(6) 4 [jsrpqks+1pq] + λ(6) where λ(6) 4 = ǫs+1 and λ(6) 5 = ǫs−1 a− a+ −ǫsǫs−1 a− a+ 5 [jsrpqks−1pq] [2] [4] (ǫsǫs+1 − c+)(cid:18)1 + (ǫsǫs−1 − c+)(cid:18)1 + a+ (cid:18)ǫs−1 − ǫs−1 [2] [4] a− iµ 1 − µ(cid:19) . Thus λ(6) 5 [jsrpqks−1pq] = (λ(5) 1 − λ(6) a−(ǫsǫs±1 − ǫs±1)[jsrpqkspqks±1] 7 [jsrpqksisjsr], [jsrpqkspqks], [jsrpqkspjs±1ks±1] = −ǫs±1 p[2] p[3][4] = p[2] p[3][4] [jsrpqks+1pqks] = −ǫsp[3][4] p[2] ⇒ [jsrpqks+1pqks] = 0! Now [jsrpqks−1pjs+1r] = ǫs+1 1 a− i (ǫs+1 + ǫs−1µ) 1 − µ i (ǫs+1 + ǫs−1µ) 1 − µ (cid:19) + (cid:19) + ǫs i (ǫs+1 + ǫs−1µ) 1 − µ ǫs a−b− a2 + a−b− i (ǫs+1 + ǫs−1µ) a2 + 1 − µ λ(5) 4 4 )[jsrpqks+1pq]∗! a−[jsrpqkspqks±1] − ǫs±1 p[2] p[3][4] a+[jsrpqksp′qks±1] [jsrpqksisjsks] = 0! [jsrpqkspjsks] = −p[2] p[4] a+ a− [jsrpqks+1pjsks] − ǫs [jsrpqks+1p′qks] = ǫsǫs+1[jsrpqks+1pqks] [4] [2] b+ a− [jsrpqks−1pqr] − ǫsǫs+1[jsrpqks−1pjsr] = −ǫs+1ǫs−1 [4] [2] a+b+ a2 − [jsrpqks−1p′qr] − ǫsǫs+1 λ(6) 2 λ(6) 1 [jsrpqkspjsr] = c+[jsrpqks−1pjsr] + ǫs+1ǫs−1c+[jsrpqks−1pjs+1r] − ǫsǫs+1 λ(6) 2 λ(6) 1 [jsrpqkspjsr] = ǫs+1ǫs−1c+[jsrpqks−1pjs+1r] + (c+ − ǫsǫs+1) λ(6) 2 λ(6) 1 [jsrpqkspjsr] So (1 − ǫs+1ǫs−1c+) [jsrpqks−1pjs+1r] = (c+ − ǫsǫs+1) λ(6) 2 λ(6) 1 [jsrpqkspjsr]. Then [jsrpqkspjs+1r] = c− − 1 ǫsǫs−1 − c− [jsrpqks−1pjs+1r] = λ(7) 1 [jsrpqkspjsr]∗, where λ(7) 1 = λ(6) 2 λ(6) 1 (c− − 1)(c+ − ǫsǫs+1) (ǫsǫs−1 − c−)(1 − ǫs+1ǫs−1c+) Similarly [jsrpqkspjs−1r] = λ(7) [jsrpqkspqr] = ǫs a− b+ [2] [4] 2 [jsrpqkspjsr]∗, = [2] [4] a− b+ (cid:16)ǫs + ǫs+1λ(7) 1 + ǫs−1λ(7) 2 (cid:17) [jsrpqkspjsr]∗! [jsrpqks+1pqr] = ǫs = [2] [4] a− b+ λ(6) 3 (cid:18)ǫs + ǫs−1 [2] [4] a− b+ [jsrpqks+1pjsr] + ǫs−1 [2] [4] a− b+ [jsrpqks+1pjs−1r] c+ − ǫsǫs−1 1 − ǫs+1ǫs−1c+(cid:19) [jsrpqkspjsr]∗! 52 .! where λ(7) 2 = λ(6) 3 (c− − 1)(c+ − ǫsǫs−1) (ǫsǫs+1 − c−)(1 − ǫs+1ǫs−1c+) .! [jsrpqkspjsr] + ǫs+1 [2] [4] a− b+ [jsrpqkspjs+1r] + ǫs−1 [2] [4] a− b+ [jsrpqkspjs−1r] 1 a− [jsrpqks+1pjs−1ks−1] − ǫs−1 a+ a− [jsrpqks+1p′qks−1] [jsrpqkspjs−1ks−1] + ǫs+1ǫs−1[jsrpqks+1pqks−1] [jsrpqks+1pqks−1] = −ǫs−1p[3][4] p[2] c− − ǫsǫs+1 1 − c− 1 a− c− − ǫsǫs+1 = −ǫs−1p[3][4] p[2] ⇒ (ǫs+1ǫs−1 − 1)[jsrpqks+1pqks−1] = ǫs−1p[3][4] p[2] Similarly (ǫs−1ǫs+1 − 1)[jsrpqks−1pqks+1] = ǫs+1p[3][4] p[2] [jsrpqkspjs+1ks+1]∗! (cid:16)[jsrpqkspjs+1ks+1is+1] = 0(cid:17) = ǫs+1p[3][4] p[2] 4 )[jsrpqks+1pqks+1] = λ(6) 5 [jsrpqks−1pqks+1] (ǫs−1ǫs+1 − 1)(1 − c−) [jsrpqkspqksis], giving (λ(5) [jsrpqkspqksp] 1 − λ(6) c− − ǫsǫs−1 1 a− 1 a− 1 a− 1 − c− λ(6) 5 1 − c− c− − ǫsǫs+1 [jsrpqkspjs−1ks−1]∗! [jsrpqkspjs+1ks+1], (cid:16)[jsrpqkspjs−1ks−1is−1] = 0(cid:17) a−b− [4] [2] = −ǫs b− a+ Now [jsrpqkspqksp] = −ǫs a+b+ (cid:16)ǫs + ǫs+1λ(7) ǫsǫs−1 − ǫs−1p[3][4] p[2] = p[2] a− (cid:16)c− + ǫs(1 + c−)(cid:16)ǫs+1λ(7) p[4] 1 + ǫs−1λ(7) ǫsǫs−1 1 a− a+ − so [jsrpqksisjsrp] = 1 λ(8) 1 a+ a− [jsrpqksisjsr′p] = −ǫs [jsrpqkspjs+1ks+1p] = −ǫs+1 p[2] p[3][4] [jsrpqkspqrp] − ǫsǫs+1[jsrpqkspqks+1p] − ǫsǫs−1[jsrpqkspqks−1p] 2 (cid:17) [jsrpqkspjsrp] − [jsrpqkspjs−1ks−1p] ǫsǫs+1 ǫsǫs+1 − ǫs+1p[3][4] p[2] 1 a− [jsrpqkspjs+1ks+1p] [jsrpqkspqksp]∗ where λ(8) [jsrpqksisjsrp] = −ǫs a+ 1 + ǫs−1λ(7) 2 (cid:17)(cid:17) [jsrpqksisjsrp], 1 = p[2] a− (cid:16)c− + ǫs(1 + c−)(cid:16)ǫs+1λ(7) p[4] [jsrpqkspqksp]∗! a+[jsrpqkspjs+1rp] − ǫs+1 p[2] p[3][4] 1 λ(8) 1 a+ a− a−[jsrpqkspjs+1r′p] a−λ(7) 1 [jsrpqkspjsr′p] 1 + ǫs−1λ(7) 2 (cid:17)(cid:17) .! = −ǫs+1 p[2] p[3][4] [4]p[3] = ǫs+1 [2] [2] = [4]p[3] a+λ(7) a+λ(7) [2] 1 [jsrpqksisjsrp] + ǫs+1 1 [jsrpqkspjsrp] − ǫs+1 p[2] p[3][4] [4]p[3] [jsrpqkspqksp]∗! [4]p[3] λ(7) 1 λ(8) 1 [2] [jsrpqkspqkspjs] Similarly [jsrpqkspjs−1ks−1p] = [jsrpqkspqksisjs] = −p[4] p[2] (ǫs+1 − ǫsǫs+1)a+ a−λ(7) 1 [jsrpqksisjsr′p] (ǫs−1 − ǫsǫs−1)a+ [jsrpqkspqksp]∗! λ(7) 2 λ(8) 1 8 9 (We know from the Hilbert series of [31, Theorem 3.1] that all paths of length 9 ending at any other vertex v 6= js are zero) Let uiν(i) = uisν(is) be the (non-zero) path uiν(i) = [isjsrpqrpqksis] of length 9. We have [jskspqksisjsrpjs] = −ǫs p[2] p[3][4] = −ǫs p[2] p[3][4] = −ǫs p[2] p[3][4] a+[jsrpqksisjsrpjs] − ǫs p[2] p[3][4] a+[jsrpqksisjsrpjs] − ǫs p[2] p[3][4] a+(1 − c−)[jsrpqksisjsrpjs] 53 a−[jsr′pqksisjsrpjs] a−b− b+ [jsrp′qksisjsrpjs] = −ǫs p[2] p[3][4] [4]p[3] = ǫs [2] a+(1 − c−)λ(8) a+(1 − c−)λ(8) 1 [jsrpqkspqkspjs] 1 [jsrpqkspqksisjs]. Then vjν(i)ν(aij) = [jsrpqrpqksisjs] = −ǫs [2] [4] a+ b− [jsrpqkspqksisjs] − ǫs+1 [2] [4] [jsrpqks−1pqksisjs] a+ b− [2] [4] a+ b− [jsrpqks+1pqksisjs] = −ǫs [jsrpqkspqksisjs] − ǫs−1 [2] [4] a+ b− λ(5) 1 − λ(6) λ(6) 5 4 [jsrpqks+1pqksisjs] −ǫs−1 [2] a+ [4] b− a+ b− [2] [4] [jsrpqkspqksisjs] = −ǫs = d[jskspqksisjsrpjs] = dajkv′, where v′ = [kspqksisjsrpjs] is symmetric and d = −ǫsp[3]/b−λ(8) Hence C = 1. 1 (1 − c−) is non-zero. 2 2 2 graph for the conformal embedding SU (3)9 ⊂ (E6)1 ⋊ Z3 , illustrated in Figure 16, is a Z3 orbifold of E (12) E (12) A.2 The graph E (12) . Every cell system W for E (12) is equivalent to either a cell system W + or the inequivalent cell system W − [25, Theorem 11.1]. We will use the solution W +. The solution W − is obtained from W + by taking the conjugation of the graph E (12) where ϕ is the map which reflects the graph E (12) about the plane which passes through the vertices i, p1, p2 and p3, and reverses the direction of each edge. For the graph E (12) the automorphism 2 ν is the identity. We choose i, j, k as labelled on the graph E (12) in Figure 16. We will write out a basis for the space of paths which start from the vertices i, k. Let a±, b± ∓b∓). denote the scalars a± =q[2][4] ±p[2][4], b± =q[2]2 ±p[2][4] and c± = (a2 = W +(ϕ(a)ϕ(b)ϕ(c)) , that is, W −(abc) Paths starting at vertex i: ±b±)/(a2 ϕ(i)ϕ(j)ϕ(k) ijk 1 2 2 1 l 2 3 4 Paths [ijr1], [ijr2], [ijr3] ([ijk] = 0) [ijr1p1] = − a+ a− [ijr2p1], [ijr2p2] = − a+ a− [ijr3p2], [ijr3p3] = − a+ a− [ijr1p3] [ijr1p1q1] = − [ijr2p2q2] = − [ijr3p3q3] = − [ijr2p1q1] = − [ijr3p2q2] = − [ijr1p3q3] = − a+b+ a−b− a+b+ a−b− a+b+ a−b− [ijr2p2q1] = c+[ijr3p2q1], [ijr3p3q2] = c+[ijr1p3q2], [ijr1p1q3] = c+[ijr2p1q3], [ijr1p1j] = − [ijr1p3j] = [ijr3p3j] = − a+ a− a+ a− a+ a− a+ a− a+ a− [ijr3p2j] = [ijr2p2j] = − a+ a− [ijr2p1j] 54 [ijr1p1jk] − a− a+ [ijr1p1q3k] . 1 a+ − a− a+ 1 µ! [ijr1p1jk] 1 µ!−1 1 a+ − [ijr1p1q1k]∗! µ!−1 a− a+ 1 [ijr1p1q1k]∗, (1 + c+)[ijr1p1q3k] 1 = = a− 1 a+ 1 a+ a− a+ a− a+ a− a+ = −c2 (1 + c+)[ijr1p1q3k] c−[ijr2p2jk] − c−[ijr1p1jk] − c−[ijr2p2q1k] + p[3][4] p[2] −[ijr1p1q1k] + p[3][4] p[2] − − c+ − 1(cid:1) [ijr1p1q3k] + p[3][4] a+ (cid:0)c2 p[2] where µ = p[2] p[3][4] [ijr1p1q3k] = −p[3][4] p[2] and comparing this with the first equality [ijr1p1q1k] = −p[3][4] p[2] Thus [ijr1p1q1k] = −p[3][4] p[2] Similarly [ijr2p2jk] = µ[ijr2p2q1k] and [ijr3p3jk] = µ[ijr3p3q2k]. − + c−(cid:1) [ijr1p1jk], we obtain [ijr1p1jk] = µ[ijr1p1q3k], a+ (cid:0)c2 [ijr1p1jk] − c+ − c2 − 1 + c− + c2 − a− a+ 1 a+ 1 a+ a− 1 a+ − a− a+ 1 µ! [ijr2p2jk] = [ijr2p2q2k], = −p[3][4] p[2] and similarly [ijr3p3q3k] = [ijr1p1q1k]∗! so [ijr2p2q2k] = [ijr1p1q1k]∗, From the previous computation we have [ijr1p1jk] = −p[3][4] p[2] Similarly [ijr2p2jk] = −p[3][4] p[2] and [ijr3p3jk] = −p[3][4] p[2] µ!−1 [ijr1p1q1k]∗! µ!−1 a− a+ 1 a+ 1 a+ [ijr1p1q1r2p1] = [ijr1p1q1kp1] = − − 1 1 [ijr2p2q2k] = −p[3][4] p[2] [ijr3p3q3kp1] = [ijr3p3q3r1p1], 1 a+ − a− a+ [ijr2p2q2r3p2] = [ijr3p3q3r1p3] = [ijr2p2q2kp2] = [ijr3p3q3kp3] = [ijr1p1q1kp2] = [ijr1p1q1r2p2], [ijr2p2q2kp3] = [ijr2p2q2r3p3] [ijr1p1q1ki] = −p[3][4] p[2] [ijr1p1q1r2p1q1] = [ijr1p1q1r2p2q1] = 1 a+ − a− a+ 1 µ! [ijr1p1jki] = 0! a− a+ a+ b− a+ b− a+ b− 6 7 [ijr2p2q2r3p2q2] = [ijr3p3q3r1p3q3] = [ijr2p2q2r3p3q2] = [ijr3p3q3r1p1q3] = b+ b− b+ b− b+ b− [ijr2p2q2r3p2q1], [ijr3p3q3r1p3q2], [ijr1p1q1r2p1q3], 5 [ijr1p1q1r2] = [ijr1p1jr2], [ijr2p2q2r3] = a+ b− a+ b− [ijr2p2jr3], [ijr3p3q3r1] = a+ b− [ijr3p3jr1], [ijr1p1q1k] Now [ijr1p1q1k] = −p[3][4] p[2] = −[ijr1p3q2k] − a+ a− 1 a+ [ijr1p1jk] − a− a+ [ijr1p1q3k] = p[3][4] p[2] (1 + c−)[ijr1p3q3k] = −c−[ijr2p2q2k] + (1 + c+)[ijr2p1q3k] 1 a− [ijr1p3jk] − a−b− a+b+ [ijr1p3q3k] a+ b− a+ b− a+ b− b+ b− b+ b− b+ b− Now [ijr1p1q1r2p1j] = and [ijr1p1q1r2p1j] = − a+ b− a− a+ [ijr1p1q1kp1j], but also [ijr1p1q1r2p1j] = − a− a+ [ijr1p1q1r2p2j] = − a− b− [ijr1p1q1kp2j], [ijr1p1q1r2p2j] = − a− a+ [ijr2p2q2r3p2j] = a2 − a2 + [ijr2p2q2r3p3j] = a2 − a+b− [ijr2p2q2kp3j] a2 − a+b− [ijr1p1q1kp3j]. Now since [ijr1p1q1kij] = 0, [ijr1p1q1kij] = [ijr1p1q1kp1j] + [ijr1p1q1kp2j] + [ijr1p1q1kp3j] = = −p[4] p[2] 55 b− a− (cid:18) a− a+ + a+ a− − 1(cid:19) [ijr1p1q1r2p1j] = 0, so [ijr1p1q1r2p1j] = [ijr2p2q2r3p2j] = [ijr3p3q3r1p3j] = 0.! 8 [ijr1p1q1r2p1q1k] = − a− a+ a−b− a+b+ [ijr1p1q1r2p1q3k] = − [ijr3p3q3r1p3q3k] = c−[ijr3p3q3r1p3q2k] = c−[ijr2p2q2r3p2q2k] b− b+ (cid:18)[ijr2p2q2r3p2q2r3] = (cid:18)[ijr1p1q1r2p1q1r2] = a+ b− [ijr1p1q1r2p1jr2] = 0(cid:19) a+ b− [ijr2p2q2r3p2jr3] = 0(cid:19) (cid:18)[ijr1p1q1r2p1q1kp1] = b− a+ (cid:18)[ijr3p3q3r1p3q3r1] = [ijr1p1q1r2p1q1r2p1] = 0(cid:19) 9 [ijr1p1q1r2p1q1ki] a+ b− [ijr3p3q3r1p3jr1] = 0(cid:19) b− b+ [ijr1p1q1r2p1q1kp2] = (cid:18)[ijr1p1q1r2p1q1kp3] = (cid:16)[ijr1p1q1r2p1q1kij] = 0(cid:17) Paths starting at vertex k: c−[ijr2p2q2r3p2q2kp2] = a−b− a+b+ [ijr3p3q3r1p3q3kp3] = b− a− b2 − a+b+ c−[ijr2p2q2r3p2q2r3p2] = 0! c−[ijr3p3q3r1p3q3r1p3] = 0(cid:19) Paths [kp1q1] = − a+ a− [kp2q1], [kp2q2] = − [kp3q2], [kp3q3] = − a+ a− a+ a− [kp1q3], [kp1j], [kp2j], [kp3j] 10 l 2 3 4 [kp2q2r3] = − a− a+ [kp3q2r3] = a2 − a+b+ [kp3jr3], [kij] = −p[2] p[4] [kp1q1r2] = − [kp1j]∗ − p[2] p[4] [kp2q1r2] = [kp2j]∗ − p[2] p[4] a2 − [kp2jr2], [kp3j]∗! [kp3q3r1] = − [kp1q3r1] = [kp1jr1], a− a+ a− a+ [kp2jr1], b− a+ a+b+ a2 − a+b+ [kp3jr2], [kp1jr3], [kp2q2k], [kp3q3k] [kp1q1k], b− a+ [kp1q1r2]∗(cid:19) (cid:18)[kp2jr3] = (cid:18)[kp1jr2] = [kp1jk] = − p[2] a−[kp1q3k] − p[2] p[3][4] p[3][4] Similarly [kp2jk] = p[2] p[3][4] (cid:18)[kp3jr1] = [kp3q3r1]∗(cid:19) a+ ([kp3q3k]∗ − [kp1q1k]∗)! [kp2q2r3]∗(cid:19) a+[kp1q1k] = p[2] p[3][4] and [kp3jk] = p[2] p[3][4] a+ ([kp1q1k]∗ − [kp2q2k]∗) , b− a+ [kp2q2kp2] = [kp2q2r3p2], [kp3q3kp3] = [kp3q3r1p3], [kp1q1kp1] = [kp1q1r2p1], b− a+ b− a+ b− a+ a+ ([kp2q2k]∗ − [kp3q3k]∗) .! [kp1q1kp2] = − b+ a− [kp1q1r2p2], [kp2q2kp3] = − b+ a− [kp2q2r3p3], [kp3q3kp1] = − b+ a− [kp3q3r1p1], [kp1q1kp3], [kp2q2kp1], [kp3q3kp2], [kp1q1ki] = − a− a+ [kp2q1ki] = [kp2q2ki] = − a− a+ [kp3q2ki] = [kp3q3ki] 1 a+ [kp1jkp2] − a− a+ [kp1jr2p2] = [kp1q1kp2] − [kp3q3kp2] − a−b− a2 + [kp1q1r2p2] and [kp3jr2p1] = (1 + c−)[kp3q3kp1]∗ − [kp2q2kp1]∗! a+ a− [kp3q3kp3] + a+ a− [kp1q1kp3] − a2 +b+ a3 − [kp3q3r1p3] [kp1jr3p2] = −p[3][4] p[2] = (1 + c−)[kp1q1kp2]∗ − [kp3q3kp2]∗! Similarly [kp2jr1p3] = (1 + c−)[kp2q2kp3]∗ − [kp1q1kp3]∗ [kp1jr3p3] = −p[3][4] p[2] Similarly [kp2jr1p1] = (1 + c+)[kp3q3kp3]∗! [kp1q1kp3]∗ − [kp2q2kp1]∗ − [kp1jr1p3] = − [kp1jkp3] − a+ a− a+ a− a+ a− a+ a− a+ a− 1 a− = and [kp3jr2p2] = a+ a− [kp3q3kp2]∗ − a+ a− (1 + c+)[kp1q1kp1]∗ (1 + c+)[kp2q2kp2]∗! 56 5 [kp1q1kp1q1] = − [kp1q1kp3q3] = − a− a+ a− a+ [kp1q1kp2q1], [kp2q2kp2q2] = − [kp1q1kp1q3], [kp2q2kp1q1] = − a− a+ a− a+ [kp2q2kp3q2], [kp3q3kp3q3] = − [kp2q2kp2q1], [kp3q3kp2q2] = − a− a+ a− a+ [kp3q3kp1q3], [kp3q3kp3q2], [kp1q1kp1j] = [kp2q2kp2j] = [kp3q3kp3j] = b− a+ b− a+ b− a+ [kp1q1r2p1j] = − [kp2q2r3p2j] = − [kp3q3r1p3j] = − [kp3q3kp1q1] = − a− a+ [kp3q3kp2q1] = a−b− a2 + a−b− a2 + a−b− a2 + a2 − a2 + [kp1q1r2p2j] = c−[kp1q1kp2j], [kp2q2r3p3j] = c−[kp2q2kp3j], [kp3q3r1p1j] = c−[kp3q3kp1j], [kp1q1kij] a− a2 + [kp3jkp2q1] a−c−[kp3jkp1q1] [kp3q2kp2q1] + p[3][4] p[2] [kp3jr3p2q1] a− a+ b+ a− c−[kp3jr1p1q1] + p[3][4] p[2] −(cid:1) [kp3q3kp1q1] a− a+ 1 + c− + c2 − 1 + c+ + c− where µ1 = .! [kp2q2kp2q1] − a2 − a2 + [kp3jr2p2q1] − [kp2q2kp2q1] − c−[kp3jr2p1q1] − [kp2q2r3p2q1] = − = − = − a− a+ a− a+ a− a+ (1 + c+)[kp2q2kp2p1] + a− a+ a−b− a2 + = (1 + c+ + c−) [kp2q2kp1p1] −(cid:0)c− + c2 ⇒ µ1[kp3q3kp1q1] = [kp2q2kp1p1]∗ = (1 + c+)[kp2q2kp1p1] + c−[kp3q3r1p1q1] − c−[kp3q2kp1q1] − c−[kp3q3kp1q1] (Similarly µ1[kp1q1kp2q2] = [kp3q3kp2p2]∗ and µ1[kp2q2kp3q3] = [kp1q1kp3p3]∗) [kp1q1kp3j] = −[kp1q1kp1j] − [kp1q1kp2j] − p[4] [kp1q1kij] = −(1 + c+)[kp1q1kp1j]∗ − p[4] p[2] p[2] [kp2q2kp1j] = −[kp2q2kp2j] − [kp2q2kp3j] − p[4] [kp2q2kij] = −(1 + c+)[kp2q2kp2j]∗ − p[4] p[2] p[2] Similarly [kp3q3kp2j] = −(1 + c+)[kp3q3kp3j]∗ − p[4] p[2] [kp1q1kij]∗! [kp2q2kp2q2r3] = [kp1q1kp1q1r2] = [kp1q1kp1jr2], [kp2q2kp2jr3], [kp3q3kp3q3r1] = [kp1q1kij]∗! [kp1q1kij]∗! a+ b− 6 a+ b− a+ b− [kp3q3kp3jr1], [kp2q2kp1q1r2], [kp3q3kp2q2r3], [kp1q1kp3q3r1], [kp1q1kp1jk], [kp1q1kp1q3r1] = a+b+ a2 − [kp1q1kp3q3r1]∗! a+b+ a2 − [kp2q2kp1q1r2]∗ and [kp3q3kp3jr3] = [kp2q2kp1jk], [kp3q3kp1jk] a+b+ a2 − [kp3q3kp2q2r3]∗! b+ a− [kp1q1kp1jr1] = − Similarly [kp2q2kp2jr2] = [kp1q1kijr1] = −p[2] p[4] [kp1q1kp3jr1] = −p[2] p[4] [kp1q1kp2jr1] − p[2] [kp1q1kp1jr1] − p[2] p[4] p[4] (1 + c−)[kp1q1kp1jr1] − p[2] [kp1q1kp3q3r1] = −p[2] p[4] p[4] Similarly [kp1q1kijr2] = [kp2q2kijr2] = −p[2] p[4] and [kp1q1kijr3] = [kp3q3kijr3] = −p[2] p[4] b− a+ a+b+ a+b+ a2 − a2 − c−[kp1q1kp2q2r3] = (1 + 2c−) [kp2q2kp1q1r2]∗ (1 + 2c−) [kp3q3kp2q2r3]∗! (cid:18)[kp1q1kp1jr3] = c−[kp1q1kp2jr3] = (cid:18)Similarly [kp2q2kp2jr1] = Now [kplqlkplqlk] = − a− a+ b− a+ 1 µ1 b− a+ 1 µ1 b− a+ c−[kp3q3kp2q2r3]∗(cid:19) 1 µ1 b− a+ c−[kp2q2kp1q1r2]∗(cid:19) c−[kp1q1kp3q3r1]∗ and [kp3q3kp3jr2] = [kplqlkplql−1k] − p[3][4] p[2] 1 a+ [kplqlkpljk] = [kplqlkpl−1ql−1k] − p[3][4] p[2] 1 a+ [kplqlkpljk] a+b+ a2 − (1 + 2c−) [kp1q1kp3q3r1]∗! 57 [kpl+1ql+1kpl−1jk] − p[3][4] p[2] [kp2q2kp3jk] − p[3][4] p[2] 1 a+ 1 a+ µ1 µ1 [kp2q2kp2jk] 1 a+ [kplqlkpljk]. [kp1q1kp1jk] [kp3q3kp3jk] − p[3][4] p[2] µ2 1 1 a+ [kp3q3kp1jk] µ1 1 a+ [kp2q2kp1jk] 1c+(cid:1) [kp1q1kp1jk]∗ + µ1[kp2q2kp1jk]∗ − µ2 1(cid:1) a−[kp1q1kp3q2k] −(cid:0)1 − µ3 1(cid:1) [kp1q1kp3jk] 1(1 + c−)[kp3q3kp1jk]∗! µ1 1 µ1 µ2 1 1 µ1 µ2 1 = µ2 = µ3 1 + c− 1 a+ 1 a+ 1 a+ 1 a+ 1 a+ 1 a+ 1 a+ 1 a+ 1 a+ [kplqlkpljk] [kp1q1kp1jk] [kp2q2kp1jk] − p[3][4] p[2] = µ1[kpl+1ql+1kpl−1ql−1k] − p[3][4] p[2] = µ1[kpl+1ql+1kpl+1ql+1k] − p[3][4] p[2] So [kp1q1kp1q1k] = µ1[kp2q2kp2q2k] − p[3][4] p[2] 1[kp3q3kp3q3k] − p[3][4] [kp3q3kp1jk] − p[3][4] p[2] p[2] −p[3][4] p[2] [kp1q1kp2jk] − p[3][4] 1[kp1q1kp1q1k] − p[3][4] p[2] p[2] [kp2q2kp1jk] − p[3][4] +p[3][4] p[2] p[2] 1c+(cid:1) [kp1q1kp1jk] + p[3][4] 1[kp1q1kp1q1k] − p[3][4] a+ (cid:0)1 + µ3 p[2] p[2] −p[3][4] p[2] Thus p[2] p[3][4](cid:0)1 − µ3 1(cid:1) a+[kp1q1kp1q1k] = −(cid:0)1 + µ3 p[2] 1(cid:1) a+[kp1q1kp3q3k] = − p[2] p[3][4](cid:0)1 − µ3 p[3][4](cid:0)1 − µ3 = p[2] p[3][4](cid:0)1 − µ3 = − p[2] p[3][4](cid:0)1 − µ3 = p[2] p[3][4](cid:0)1 − µ3 1(cid:1) a+[kp1q1kp2q2k] + (1 + c+)[kp1q1kp1jk] 1(cid:1) a−[kp1q1kp2q1k] −(cid:0)1 − µ3 1(cid:1) a−[kp1q1kp1q1k] −(cid:0)1 − µ3 (1 + c−)[kp3q3kp1jk]. [kp1q1kp1jk] 1 a+ 1 a+ = µ3 = −µ3 1(1 + c+)[kp1q1kp1jk]∗ + µ1[kp2q2kp1jk]∗ − µ2 µ2 1 1 µ1 = −µ2 p[2] 1(cid:1) a+[kp2q2kp2q2k] = − p[2] p[3][4](cid:0)1 − µ3 p[3][4](cid:0)1 − µ3 = p[2] p[3][4](cid:0)1 − µ3 1(cid:1) a+[kp2q2kp3q3k] +(cid:0)1 − µ3 1(cid:1) = p[2] p[3][4](cid:0)1 − µ3 1(cid:1) 1 a+[kp1q1kp3q3k] +(cid:0)1 − µ3 1(cid:1) 1(1 + c+)[kp1q1kp1jk]∗ +(cid:0)2 − µ3 1 + c+(cid:1) Now, from the computation of [kp1q1kp1q1k] we obtain 1(cid:1) a+[kp3q3kp3q3k] = p[2] p[2] 1(cid:1) 1 p[3][4](cid:0)1 − µ3 p[3][4](cid:0)1 − µ3 [kp2q2kp2jk] −(cid:0)1 − µ3 1(cid:1) 1 −(cid:0)1 − µ3 1(cid:1) 1 =(cid:18) 1 − 1 − µ1(1 + c+)(cid:19) [kp1q1kp1jk]∗ + µ2 µ2 1 1 µ2 1 µ1 µ1 1 + c+ 1 1 + c− Now [kp3q3kp3q3k] = − 1[kp2q2kp1jk]∗ −(cid:0)µ3 [kp3q3kp1q3k] = [kp3q3kp1q1k] + p[3][4] p[2] [kp2q2kp1q1k] + p[3][4] p[2] [kp3q3kp1jk]. a− a+ 1 a+ = 1 µ1 58 1(cid:1) [kp1q1kp2jk] + (1 + c+)[kp1q1kp1jk] 1(cid:1) [kp1q1kp1jk] 1(1 + c−)[kp3q3kp1jk]∗! 1(cid:1) a−[kp2q2kp2q3k] −(cid:0)1 − µ3 1(cid:1) [kp2q2kp2jk] [kp2q2kp1jk] 1 + c+ 1 1 + c+ [kp2q2kp1jk] [kp2q2kp1jk]∗ − µ1(1 + c−)[kp3q3kp1jk]∗! 1(cid:1) [kp3q3kp1jk] a+[kp1q1kp1q1k] +(cid:0)1 − µ3 [kp2q2kp1jk] +(cid:0)1 − µ3 1(cid:1) 1 1 + c−(cid:1) [kp3q3kp1jk]∗! [kp1q1kp1jk] µ2 1 1 a+ [kp3q3kp1jk] 1(cid:1) µ1a+[kp3q3kp3q3k] −(cid:0)1 − µ3 1(cid:1) µ1[kp3q3kp1jk] 1 + c−(cid:1) [kp3q3kp1jk]∗! 1[kp2q2kp1jk]∗ − µ1(cid:0)1 + µ3 1(cid:1) a−[kp3q3kp3q2k] 1(cid:1) [kp3q3kp3jk] 1[kp2q2kp1jk]∗ − µ3 1[kp3q3kp1jk]∗! − µ1 − µ2 p[3][4](cid:0)1 − µ3 1(cid:1) a+[kp2q2kp1q1k] = p[2] 1(1 + c+)(cid:19) [kp1q1kp1jk]∗ + µ3 p[3][4](cid:0)1 − µ3 1(cid:1) a+[kp3q3kp2q2k] = − p[2] 1(cid:1) a+[kp3q3kp3q3k] +(cid:0)1 − µ3 − 1 − µ1(1 + c+)(cid:19) [kp1q1kp1jk]∗ + µ2 µ1 Thus p[2] p[3][4](cid:0)1 − µ3 =(cid:18) 1 p[2] p[3][4](cid:0)1 − µ3 = p[2] p[3][4](cid:0)1 − µ3 =(cid:18) 1 (cid:16)[kp1q1kijk] = 0(cid:17) (cid:16)[kp1q1kp1jki] = 0(cid:17) Now [kpljkp3q3r1p1] = − [kp1q1kp1jkp1], µ2 1 (cid:16)[kp2q2kp1jki] = 0(cid:17) a2 − a+b+ a− a+ = − a− b+ [kpljkp1jr2p1] − p[3][4] p[2] a+ a− a− a+b+ [kpljkp1jkp1], a2 + a−b− 7 [kp2q2kp1jkp1], [kp2q2kp1jkp2], [kp3q3kp1jkp2], (cid:16)[kp3q3kp1jki] = 0(cid:17) [kpljkp1q3r1p1] = [kpljkp1jr1p1] [kp3q3kp1jkp3], [kp1q1kp1jkp3] a− a+b+ [kpljkp1jkp1] [kpljkp1q3kp1] + a+b+ a− b+ [kpljkp1q1kp1] [kplql−1kp1q3kp1] [kplqlkp1q1kp1] [kpl−1ql−1kp1q3kp1] a+a− b+ [kplqlkp1q1kp1]. (∗) and [kpljkp2q1r2p1] = − [kpljkp1q1r2p1] = − [kpljkp1jr2p1], thus [kpljkp3q3r2p1] + [kpljkp2q1r2p1] = − = (1 + c+) a2 − a+b+ [kpljkp1jr1p1] + p[3][4] p[2] a3 − a− b+ (1 + c+)[kpljkp1jr2p1] − p[3][4] p[2] a− a2 − [kpljkp1jkp1] + (1 + c+) a+b+ Also, [kp1jr3p2jr2p1] = − [kp1jkp3jr2p1] (1 + c+)a−[kplqlkp1q3r1p1] a3 − a+b+ a+a− b+ (1 + c+) [kplqlkp1jr1p1] a2 − a+b+ a− b+ a− b+ a2 − b+ a2 − b+ a2 − b+ (1 + c+) a+a− b+ +(1 + c+) −(1 + c+) a− b+ a2 − b+ a+b+ a2 − b+ [kplql−1kp1jkp1] − (1 + c+) [kpl−1ql−1kp1jkp1] − (1 + c+) [kplqlkp1q3kp1] − p[2] p[3][4] [kplqlkp1q3kp1] − p[2] p[3][4] [kplql−1kp1jr1p1] − p[2] p[3][4] [kplqlkp1jkp1] − p[2] p[3][4] [kplql−1kp1q1kp1] − p[2] p[3][4] (1 + c+)a−[kpl−1ql−1kp1q3r1p1] + p[2] p[3][4] [kplqlkp1jkp1] + p[2] p[3][4] [kpl−1ql−1kp1q1kp1] − p[2] p[3][4] = − p[2] p[3][4] − p[2] p[3][4] = − p[2] p[3][4] − p[2] p[3][4] [kp1jr3p3jr2p1] = [kp1jr1p3jr2p1] + p[3][4] p[2] [kp1q3r1p3jr2p1] − p[3][4] p[2] [kp1jkp3q3r1p1] + p[3][4] [kp1q3kp3jr2p1] − p[3][4] p[2] p[2] +p[3][4] p[2] [kp3q3kp3jr1p1] − p[3][4] p[2] [kp1q1kp3q3r1p1] + p[3][4] p[2] (1 + c+)[kp3q3kp3q3r1p1] − p[3][4] p[2] (1 + c−)[kp3q3kp1jkp1] + (1 + c+)[kp3q3kp3jkp1] + [kp1jkp3jr1p1] − a2 −b− a3 + [kp1q1kp3jkp1] [kp1q1kp3jkp1] [kp1jkp3jkp1] [3][4] [2] 1 a2 + a+b+ a−b− a+b+ a−b− a−b− a3 + a−b− a2 + + a−b− a2 + a− a2 + a− a2 + = − b+ a− a− a+ 1 a+ 1 a+ 1 a+ 1 a+ 1 a+ = − = − = − [kp1q3kp3jkp1] [kp1q3kp3q3r1p1] 59 a−b− a2 + [kp1q1kp3q3r1p1] [kp1q1kp2q1r2p1] a− a+ a−b− a+b+ 1 a+ b−[kp1q1kp1q1r2p1] b+[kp1q3kp3q3r1p1] + (1 + c+)[kp1q1kp1jkp1]. We have [kp1q1kp1jkp1] = − p[2] p[3][4] +p[3][4] p[2] a−[kp1q1kp1jr1p1] − p[2] p[3][4] = p[2] b+[kp1q1kp1q3r1p1] − p[2] p[3][4] p[3][4] [kp1q1kp3q3r1p1] + p[2] = − p[2] p[3][4] p[3][4] [kp1jkp3q3r1p1] − p[2] = p[2] p[3][4] p[3][4] [kp1jkp3q3r1p1] − p[2] = p[2] p[3][4] p[3][4] = − p[2] p[3][4] + p[2] p[3][4] = p[2] p[3][4] [kp3jkp2q1r2p1] − p[2] p[3][4] [kp3q3r1p3q3r1p1] − [kp1q3r1p3q3r1p1] + [kp3q3kp3q3r1p1] − a+[kp1jr3p2jr2p1]. [kp2jkp3q3r1p1] − [kp2jkp3q3r1p1] − [kp1jr3p2q1r2p1] b+ a− b+ a− b+ a− b+ a− b+ a− b+ a− b+ a− a+b+ a+b+ b+b− b+b− a+ a− − a− a− (∗∗) a+[kp1q1kp1jr2p1] b+b− a+ [kp1q1r2p2q1r2p1] b+[kp1jr2p2q1r2p1] [kp3jkp3q3r1p1] + b+ a− [kp1jkp2q1r2p1] [kp2jkp2q1r2p1] − b+ a− [kp3jkp3q3r1p1] b+(1 + c+)[kp1q1kp1q3r1p1] b+(1 + c+)[kp3q3kp1q3r1p1] a+[kp3q3kp1q1kp1] a+ a+ a−b− a−b− Then by (∗), (∗∗), [kp1q1kp3q3r1p1] a−(1 + c+)[kp1q1kp1q3kp1] a+[kp3q3kp3q3kp1] + p[2] p[3][4] a+[kp1q1kp1q1kp1] − p[2] p[3][4] a−[kp3q3kp1q3kp1] + p[2] p[3][4] (1 + c+)[kp3q3kp3q3r1p1] − p[2] p[3][4] [kp1q1kp1jkp1] = p[2] p[3][4] a−[kp1q1kp1q3kp1] − p[2] − p[2] p[3][4] p[3][4] + (2 + c+ + c−) [kp3q3kp1jkp1] − p[2] p[3][4] + p[2] p[3][4] a+(1 + c−)[kp3q3kp3q3kp1] − p[2] = p[2] p[3][4] p[3][4] + p[2] a+(1 + c−)[kp1q1kp3q3kp1] − p[2] p[3][4] p[3][4] a−(1 + c+)[kp3q3kp1q3kp1] + (2 + c+ + c−) [kp3q3kp1jkp1] + p[2] + p[2] p[3][4] p[3][4] a+ (2 + c+ + c−) [kp1q1kp3q3kp1] + p[2] a+[kp1q1kp1q1kp1] + p[2] = − p[2] p[3][4] p[3][4] p[3][4] − p[2] p[3][4] =(cid:18)1 + λ1 1(cid:19) [kp1q1kp1jkp1] + λ2 +(cid:1) , 1(cid:1)(cid:0)2 + 2c+ + c− + c2 where λ1 = µ1(cid:0)1 − µ2 and λ3 =(cid:0)1 − µ2 1(cid:1)(cid:0)2 + c+ + 2c− + c2 −(cid:1) . Similarly [kp1q1kp2jkp2] = − a+ (2 + c+ + c−) [kp3q3kp3q3kp1] + (2 + c+ + c−) [kp3q3kp1jkp1] [kp2q2kp1jkp1]∗.! Thus we obtain [kp3q3kp1jkp1] = − λ2 = µ1 (1 − µ1) (1 + c+ + c−) , [kp3q3kp2jkp2] which gives [kp1q1kp1jkp1]∗ − [kp2q2kp1jkp1] + λ3 a+[kp1q1kp1q1kp1] [kp2q2kp2jkp2] − 1 − µ3 1 1 − µ3 1 1 − µ3 λ1 λ3 λ2 λ3 λ1 λ3 λ2 λ3 1 1 1 [kp1q1kp1jkp2] = λ1 λ3 and [kp2q2kp3jkp3] = − c− 1 + c+ λ1 λ3 [kp3q3kp3jkp3] − [kp2q2kp1jkp2]∗ + λ2 λ3 c−(1 + c−)[kp3q3kp1jkp2]∗, λ2 λ3 [kp1q1kp3jkp3] which gives 60 [kp3q3kp1jkp1] a+[kp3q3kp1q1kp1] a+[kp3q3kp1q1kp1] c−(1 + c−)[kp3q3kp1jkp3]∗ − (1 + c+)(1 + c−)[kp1q1kp1jkp3]∗.! λ2 λ3 [kp1q1kp1q1kp1] λ1 λ3 a+ b− µ1 1 − µ3 1 1 b− [kp2q2kp1jkp1] 1 b− (cid:0)1 + µ3 1a−c+(cid:1) [kp1q1kp1jkp1] + p[3][4] p[2] (1 + c−)[kp3q3kp1jkp1] 1 − µ3 1 µ2 1 1 − µ3 1 1 b− 1 − µ3 1 µ1 1 − µ3 1 a+ b− 1 1 µ2 λ1 λ3 1a−c+ + 1(1 + c−)(cid:19) [kp1q1kp1jkp1]∗ b− (cid:18)1 + µ3 b− (cid:18)1 + [kp2q2kp2q2kp2] + p[3][4] p[2] µ1(1 + c−)(cid:19) [kp2q2kp1jkp1]∗! b− (cid:0)2 − µ3 1 + c+(cid:1) 1 − µ3 1 λ2 λ3 1 1 (1 + c−)[kp3q3kp1jkp2] µ1 1 − µ3 1 1 b− 1 b− (cid:18)(cid:0)2 − µ3 (1 + c−) 1 + c+(cid:1) b− (cid:18)1 + 1 1 + c+ 1 λ2 λ3 − µ2 λ1 λ3 1c−(cid:19) [kp2q2kp1jkp2]∗ µ1(1 + c−)(cid:19) [kp3q3kp1jkp2]∗! µ1 1 − µ3 1 a+ b− [kp3q3kp3q3kp3] 1 1 + c+ [kp2q2kp1jkp2] [kp2q2kp1jkp3] = 1 − µ3 1 1 1 1 1 [kp1q1kp1q1r2p1] = = −p[3][4] p[2] −p[3][4] p[2] = −p[3][4] p[2] +p[3][4] p[2] [kp2q2kp2q2r3p2] = −p[3][4] p[2] = p[3][4] p[2] −p[3][4] p[2] [kp3q3kp3q3r1p3] = = p[3][4] p[2] +p[3][4] p[2] = p[3][4] p[2] −p[3][4] p[2] = −p[3][4] p[2] −p[3][4] p[2] = −p[3][4] p[2] −p[3][4] p[2] = p[3][4] p[2] +p[3][4] p[2] = p[3][4] p[2] 1 1 1 1 1 1 − µ3 1 1 − µ3 1 1 − µ3 1 µ2 1 1 − µ3 1 1 b− 1 b− (cid:18) 1 µ2 1 1 1 − 1 − µ1(1 + c+)(cid:19) [kp1q1kp1jkp3] [kp2q2kp1jp3] − p[3][4] b− (cid:0)µ3 1 + c−(cid:1) [kp3q3kp1jkp3] p[2] µ1(1 + c−)(cid:19)(cid:19) [kp1q1kp1jkp3]∗ − 1 − µ1(1 + c+)(cid:18)1 − 1c−(1 + c−)(cid:19) [kp3q3kp1jkp3]∗! 1 + c−(cid:1) − 1 − µ3 1 λ1 λ3 λ2 λ3 µ2 1 µ2 1 b− (cid:18) 1 b− (cid:18)(cid:0)µ3 1 1 1 − µ3 1 [kp3q3kp3q3r1p1] = − a− b+ [kp3q3kp3q3kp1] [kp1q1kp1q1r2p2] = − a− b+ [kp1q1kp1q1kp2] a− a+b+ (cid:18) 1 a− a+b+ 1 − µ3 1 µ2 1 1 − µ3 1 a− µ2 1 a+b+ (cid:18)(cid:18) 1 a+b+ (cid:18)µ2 a− 1 − µ3 1 1 1 − µ3 1 1 a− µ2 1 − 1 − µ1(1 + c+)(cid:19) [kp1q1kp1jkp1] [kp2q2kp1jp1] + p[3][4] a+b+ (cid:0)µ3 p[2] 1 + c−(cid:1)(cid:19) [kp1q1kp1jkp1]∗ − 1 − µ1(1 + c+)(cid:19) + 1 + c−(cid:1)(cid:19) [kp2q2kp1jp1]∗! λ3 (cid:0)µ3 λ3 (cid:0)µ3 1 + c−(cid:1) [kp3q3kp1jkp1] 1 − µ3 1 1 + λ2 λ1 a− a+b+ (cid:0)1 + µ3 a− 1a−c+(cid:1) [kp1q1kp1jkp2] − p[3][4] p[2] (1 + c−)[kp3q3kp1jkp2] 1 − µ3 1 µ2 1 µ1 a− 1 − µ3 1 a+b+ [kp2q2kp1jkp2] 1 − µ3 1 a+b+ a− a+b+ (cid:18)−µ1 + λ1 λ3 (cid:0)1 + µ3 1a−c+(cid:1) c− 1 + c+(cid:19) [kp2q2kp1jkp2]∗ 61 8 [kp1q1kp1jkp1q1], [kp2q2kp1jkp2q2], − λ2 λ3 = λ1 λ3 [kp1q1kp1jkp1q1] − λ3 λ2 a+b+ a−b− 1 1 + c− =(cid:18) λ1 λ3 + λ3 λ2 c+ 1 + c−(cid:19) [kp1q1kp1jkp1q1] − ⇒ [kp2q2kp1jkp1q1] = µ2[kp1q1kp1jkp1q1]∗, From the previous computation [kp3q3kp1jkp2q1] = a+ a− (cid:18)[kp2q2kp1jkp2q1] = − λ1 λ3 λ2 λ3 a+ a− [kp1q1kp1jkp2q1] + λ1 λ2 a− a+ 1 (1 + c+)(1 + c−) [kp2q2kp1jkp2q1] λ1 λ2 1 (1 + c+)(1 + c−) [kp2q2kp1jkp1q1] where (cid:18)− λ2 λ3 + λ1 λ2 1 (1 + c+)(1 + c−)(cid:19) µ2 = λ1 λ3 + λ3 λ2 c+ 1 + c− .! [kp2q2kp1jkp1q1] + a+ a− λ1 λ3 [kp2q2kp1jkp1q1] = − a+ a− [kp1q1kp1jkp1q1] =(cid:18) a+ µ2[kp1q1kp1jkp1q1]∗(cid:19) a− λ2 λ3 µ2 + a+ a− λ1 λ3(cid:19) [kp1q1kp1jkp1q1]∗! +p[3][4] p[2] = p[3][4] p[2] +p[3][4] p[2] = p[3][4] p[2] +p[3][4] p[2] 1 a− 1 − µ3 1 a+b+ (1 + c−)(cid:18)µ2 1 + λ2 λ3 (cid:0)1 + µ3 1a−c+(cid:1) c−(cid:19) [kp3q3kp1jkp2]∗! [kp2q2kp2q2r3p3] = − a− b+ [kp2q2kp2q2kp3] 1 1 − µ3 1 a− a+b+ (cid:0)2 − µ3 1 + c+(cid:1) 1 1 + c+ [kp2q2kp1jkp3] µ2 1 1 − µ3 1 a− a+b+ µ1 (1 + c+)[kp1q1kp1jkp3] − p[3][4] p[2] (1 + c−)[kp3q3kp1jkp3] a− 1 − µ3 1 a+b+ 1 1 − µ3 1 a− a+b+ (cid:18)µ2 1(1 + c+) − 1 a− 1 − µ3 1 a+b+ (1 + c−)(cid:18)µ1 + λ1 λ3 (cid:0)2 − µ3 λ3 (cid:0)2 − µ3 −(cid:19) [kp1q1kp1jkp3]∗ 1 + c+(cid:1) c2 1 + c+(cid:1)(cid:19) [kp3q3kp1jkp3]∗! λ2 [kp3q3kp1jkp3q3], [kp1q1kp1jkp1j] [kp2q2kp1jkp1q1] = λ1 λ3 [kp1q1kp1jkp1q1] + [kp3q3kp1jkp1q1] = λ1 λ3 [kp1q1kp1jkp1q1] − a− a+ [kp3q3kp1jkp2q1] [kp2q2kp1jkp2q2] = c+(1 + c+)[kp1q1kp1jkp2q2] − λ2 λ3 (1 + c+)(1 + c−)[kp3q3kp1jkp2q2] = − a+b+ a−b− (1 + c+)[kp1q1kp1jkp3q2] − (1 + c+)(1 + c−)[kp3q3kp1jkp2q2] 1 1 + c− = λ3 λ2 = − a+b+ a−b− c+ λ3 λ2 1 + c− [kp2q2kp1jkp3q2] − [kp3q3kp1jkp3q2] − (1 + c+)(1 + c−)[kp3q3kp1jkp2q2] λ2 λ3 [kp2q2kp1jkp2q2] +(cid:18) λ1 λ2 (1 + c+)(1 + c−)(cid:19) [kp3q3kp1jkp2q2] λ2 λ3 λ1 λ2 − a− a+ λ2 λ3 [kp2q2kp1jkp2q2]∗! ⇒ [kp3q3kp1jkp2q2] = µ2 1 (1 + c+)(1 + c−) From the previous computation [kp1q1kp1jkp3q2] = − a−b− a+b+ λ2 λ3 (1 + c−)[kp3q3kp1jkp2q2] − a−b− a+b+ λ1 λ3 [kp2q2kp1jkp2q2] = − a−b− a+b+ (cid:18) λ1 λ3 + µ2 λ2 λ3 1 1 + c+(cid:19) [kp2q2kp1jkp2q2]∗! (cid:18)[kp3q3kp1jkp3q2] = − λ1 λ3 c−(1 + c−)[kp3q3kp1jkp3q3] = a+ a− [kp3q3kp1jkp2q2] = −µ2 a+ a− 1 (1 + c+)(1 + c−) [kp2q2kp1jkp2q2]∗(cid:19) λ2 λ3 (1 + c+)(1 + c−)[kp1q1kp1jkp3q3] + [kp2q2kp1jkp3q3] = = λ2 λ3 λ2 λ3 (1 + c+)(1 + c−)[kp1q1kp1jkp3q3] − (1 + c+)(1 + c−)[kp1q1kp1jkp3q3] + a− a+ a− a+ [kp2q2kp1jkp1q3] λ3 λ2 [kp3q3kp1jkp1q3] + a− a+ λ1 λ2 [kp1q1kp1jkp1q3] 62 =(cid:18) λ2 λ3 (1 + c+)(1 + c−) − λ1 λ2(cid:19) [kp1q1kp1jkp3q3] − [kp3q3kp1jkp3q3]∗! c− 1 + c+ λ3 λ2 [kp3q3kp1jkp3q3] ⇒ [kp1q1kp1jkp3q3] = −µ2 From the previous computation [kp2q2kp1jkp1q3] = λ1 λ3 c−(1 + c−)[kp3q3kp1jkp3q3] − (1 + c+)(1 + c−)[kp1q1kp1jkp3q3] λ2 λ3 λ3 λ2 λ3 c− + µ2 c−(cid:19) [kp3q3kp1jkp3q3]∗! = (1 + c−)(cid:18) λ1 (cid:18)[kp1q1kp1jkp1q3] = − We have [kp3q3kp1jkp3j] = −[kp3q3kp1jkp1j] − [kp3q3kp1jkp2j] [kp1q1kp1jkp3q3] = µ2 a−b− a+b+ 1 + c+ a+ a− 1 [kp3q3kp1jkp3q3]∗(cid:19) = λ1 λ3 [kp1q1kp1jkp1j] + λ2 λ3 [kp2q2kp1jkp1j] − [kp3q3kp1jkp2j], [kp1q1kp1jkp1j] = −[kp1q1kp1jkp2j] − [kp1q1kp1jkp3j] = − λ1 λ3 c− 1 + c+ [kp2q2kp1jkp2j] − λ2 λ3 c−(1 + c−)[kp3q3kp1jkp2j] − [kp1q1kp1jkp3j], and [kp2q2kp1jkp2j] = −[kp2q2kp1jkp1j] − [kp2q2kp1jkp3j] = −[kp2q2kp1jkp1j] + λ2 λ3 (1 + c+)(1 + c−)[kp1q1kp1jkp3j] − λ1 λ3 c−(1 + c−)[kp3q3kp1jkp3j]. Then from (I)-(III) we obtain [kp3q3kp1jkp2j] (I) = λ1 λ3 λ2 λ3 [kp1q1kp1jkp1j] + [kp2q2kp1jkp1j] − [kp3q3kp1jkp3j] (III) = λ1 λ3 [kp1q1kp1jkp1j] −(cid:18)1 + λ1λ2 λ2 3 c−(1 + c−)(cid:19) [kp3q3kp1jkp3j] + λ2 2 λ2 3 (1 + c+)(1 + c−)[kp1q1kp1jkp3j] − λ2 λ3 [kp2q2kp1jkp2j] (1 + c−)3[kp3q3kp1jkp2j] (II) + − λ3 λ3 λ2 2 λ2 3 λ3 2 λ3 3 λ1λ2 2 λ3 3 c−(1 + c−)(cid:19) [kp2q2kp1jkp2j] −(cid:18)1 + (1 + c+)(1 + c−)(cid:19) [kp1q1kp1jkp1j] − = (cid:18) λ1 −(cid:18) λ2 (1 + c−)3(cid:19) [kp3q3kp1jkp2j] =(cid:18) λ1 −(cid:18) λ2 c−(1 + c−)(cid:19) [kp2q2kp1jkp2j] −(cid:18)1 + λ1λ2 2 λ3 3 λ3 2 λ3 3 λ2 2 λ2 3 λ3 λ3 − + ⇒(cid:18)1 + λ1λ2 λ2 3 c−(1 + c−)(cid:19) [kp3q3kp1jkp3j] (1 + c+)(1 + c−)(cid:19) [kp1q1kp1jkp1j] λ1λ2 λ2 3 c−(1 + c−)(cid:19) [kp3q3kp1jkp3j]. − λ2 2 λ2 3 (III) = c−(cid:18) λ2 2 λ2 3 Similarly from (I)-(III) we obtain [kp1q1kp1jkp3j] (II) = − λ1 λ3 c− 1 + c+ (I) = − λ1 λ3 c− 1 + c+ [kp2q2kp1jkp2j] −(cid:18)1 + λ1λ2 λ2 3 c−(1 + c−)(cid:19) [kp1q1kp1jkp1j] c−(1 + c−)[kp2q2kp1jkp1j] + c−(1 + c−)[kp3q3kp1jkp3j] λ2 λ3 λ2 λ3 [kp2q2kp1jkp2j] − c−(1 + c−)[kp3q3kp1jkp2j] − [kp1q1kp1jkp1j] (1 + c−) − λ1 λ3 1 1 + c+(cid:19) [kp2q2kp1jkp2j] −(cid:18)1 + − λ3 2 λ3 3 (1 + c−)3[kp1q1kp1jkp3j] + c−(1 + c−)(cid:18) λ2 (1 + c−)3(cid:19) [kp1q1kp1jkp3j] = c−(cid:18) λ2 −(cid:18)1 + c−(1 + c−)(cid:19) [kp1q1kp1jkp1j] + c−(1 + c−)(cid:18) λ2 λ1λ2 2 λ3 3 λ1λ2 λ2 3 (1 + c−) − λ1 λ3 2 λ2 3 λ3 2 λ3 3 λ3 λ3 + λ1λ2 λ2 3 c−(1 + c−)(cid:19) [kp1q1kp1jkp1j] c−(1 + c−)(cid:19) [kp3q3kp1jkp3j] 1 + c+(cid:19) [kp2q2kp1jkp2j] 1 + λ1λ2 2 λ3 3 c−(1 + c−)(cid:19) [kp3q3kp1jkp3j]. ⇒(cid:18)1 + 63 (I) (II) (III) (IV) (V) = −(1 + c−)(cid:18) λ2 +(1 + c−)(cid:18) λ2 2 λ2 3 λ3 (1 + c+) + (1 + c−)2 − λ1λ2 2 λ3 3 λ1 λ3 (1 + c−)2(cid:19) [kp1q1kp1jkp1j] −(cid:18)1 + c−(cid:19) [kp3q3kp1jkp3j]. λ1λ2 λ2 3 c−(1 + c−)(cid:19) [kp2q2kp1jkp2j] (VI) Thus we have expressed [kp3q3kp1jkp2j], [kp1q1kp1jkp3j] and [kp2q2kp1jkp1j] in terms of [kp1q1kp1jkp1j], [kp2q2kp1jkp2j] and [kp3q3kp1jkp3j]. We now want to express [kp2q2kp1jkp2j] and [kp3q3kp1jkp3j] in terms of [kp1q1kp1jkp1j]. a−b− [kp1q1kp1q1r2p2j] = c−[kp1q1kp1q1kp2j]. b− a+ Consider [kp1q1kp1q1kp1j] = a2 + Thus we obtain 0 = [kp1q1kp1q1kp1j] − c−[kp1q1kp1q1kp2j] [kp1q1kp1q1r2p1j] = − µ2 λ1 λ3 1c+ + 1(cid:1) [kp1q1kp1jkp1j] + µ1[kp2q2kp1jkp1j] − µ2 = −(cid:0)µ3 +(cid:0)µ3 1 + c−(cid:1) [kp1q1kp1jkp2j] − µ1c−[kp2q2kp1jkp2j] + µ2 1(1 + c−)(cid:19) [kp1q1kp1jkp1j] +(cid:18)µ1 + =(cid:18)−(cid:0)µ3 1c+ + 1(cid:1) + +c−(cid:18) λ1 − µ1(cid:19) [kp2q2kp1jkp2j] +(cid:18)µ2 λ3 (cid:0)µ3 1 + c−(cid:1) Multiply through by (cid:18)1 + 0 =(cid:26)(cid:18)1 + (1 + c−)3(cid:19)(cid:18) λ1 (1 + c−)3(cid:19) : 1c+ − 1(cid:19) µ2 1(1 + c−) − µ3 1 + c+ λ3 2 λ3 3 λ3 2 λ3 3 λ3 1 1(1 + c−)[kp3q3kp1jkp1j] 1c−(1 + c−)[kp3q3kp1jkp2j] λ2 λ3 1(1 + c−)(cid:19) [kp2q2kp1jkp1j] µ2 1c−(1 + c−) + λ2 λ3 c−(1 + c−)(cid:19) [kp3q3kp1jkp2j]. From (I)-(III) we also obtain [kp2q2kp1jkp1j] (1 + c+)(1 + c−)[kp1q1kp1jkp3j] − λ2 λ3 (III) = λ1λ2 λ2 3 (II) = −(cid:18)1 + − λ2 λ3 c−(1 + c−)(cid:19) [kp2q2kp1jkp2j] − λ2 2 λ2 3 (1 + c−)3[kp3q3kp1jkp2j] (1 + c+)(1 + c−)[kp1q1kp1jkp1j] − c−(1 + c−)[kp3q3kp1jkp3j] λ1 λ3 λ1 λ3 c−(1 + c−)[kp3q3kp1jkp3j] − [kp2q2kp1jkp2j] (I) λ3 (1 + c+) + = −(1 + c−)(cid:18) λ2 +(1 + c−)(cid:18) λ2 (1 + c−)3(cid:19) [kp2q2kp1jkp1j] λ1λ2 2 λ3 3 λ1 λ3 (1 + c−)2 − λ3 2 λ3 3 2 λ2 3 ⇒(cid:18)1 + (1 + c−)2(cid:19) [kp1q1kp1jkp1j] −(cid:18)1 + c−(cid:19) [kp3q3kp1jkp3j] − λ3 2 λ3 3 (1 + c−)3[kp2q2kp1jkp1j] λ1λ2 λ2 3 c−(1 + c−)(cid:19) [kp2q2kp1jkp2j] (1 + c+)(1 + c−) + λ1λ2 2 λ3 3 (1 + c−)3(cid:19) + µ1(1 + c−)(cid:19)(cid:18) λ2 λ3 λ3 −µ1(cid:18) λ2 +c−(1 + c−)(cid:18) λ2 λ3 (cid:0)µ3 (1 + c−)3(cid:19)(cid:18) λ1 +(cid:26)(cid:18)1 + λ3 2 λ3 3 −µ2 1(1 + c−)2(cid:18) λ2 λ3 + +(cid:26)µ1(cid:18) λ2 λ3 + µ1(1 + c−)(cid:19)(cid:18) λ2 −c−(1 + c−)(cid:18) λ2 λ3 (cid:0)µ3 1 + c−(cid:1) + µ2 − λ3 c− λ2 2 λ2 3 1(cid:19)(cid:18) λ1 (1 + c+)(1 + c−)(cid:19)(cid:27) [kp1q1kp1jkp1j] 1 + c−(cid:1) + µ2 − µ1c−(cid:19) −(cid:18)µ1 + 1 + c−(cid:1) λ3 (cid:0)µ3 c−(1 + c−)(cid:19)(cid:27) [kp2q2kp1jkp2j] c−(1 + c−)(cid:19) 1 + c−(cid:1) 3 (cid:0)µ3 (1 + c−)3 − λ1λ2 2 λ3 3 1 + c+ λ2 2 λ2 λ1 λ3 2 λ2 3 1(cid:19)(cid:18)1 + λ1λ2 λ2 3 c−(1 + c−)(cid:19)(cid:27) [kp3q3kp1jkp3j] 1 1 + c+(cid:19)(cid:18)1 + = λ(1) 1 [kp1q1kp1jkp1j] + λ(1) 2 [kp2q2kp1jkp2j] + λ(1) 3 [kp3q3kp1jkp3j]. λ1λ2 λ2 3 c−(1 + c−)(cid:19) (VII) Now consider [kp1q1kp3q3kp3j] = b− a+ [kp1q1p1kq3r1p3j] = − a−b− a2 + Thus we obtain 0 = [kp1q1kp3q3kp3j] − c−[kp1q1kp3q3kp1j] [kp1q1kp3q3r1p1j] = c−[kp1q1kp3q3kp1j]. = −µ3 1(1 + c+)[kp1q1kp1jkp3j] + µ1[kp2q2kp1jkp3j] − µ2 +µ3 1(1 + c−)[kp1q1kp1jkp1j] − µ1c−[kp2q2kp1jkp1j] + µ2 = −(cid:18)µ3 1(1 + c+) + λ2 λ3 µ1(1 + c+)(1 + c−)(cid:19) [kp1q1kp1jkp3j] +(cid:18) λ1 λ3 1(1 + c−)[kp3q3kp1jkp3j] 1c−(1 + c−)[kp3q3kp1jkp1j] µ1c−(1 + c−) − µ2 1(1 + c−)(cid:19) [kp3q3kp1jkp3j] 64 +(cid:18)µ3 1(1 + c−) − Multiply through by (cid:18)1 + 0 =(cid:26)µ2 1(1 + c−)(cid:18)1 + λ3 2 λ3 3 µ2 λ1 λ3 λ3 2 λ3 3 1c−(1 + c−)(cid:19) [kp1q1kp1jkp1j] −(cid:18)µ1c− + (1 + c−)3(cid:19) : (1 + c−)3(cid:19)(cid:18)µ1 − c−(cid:19) + µ1(1 + c+)(cid:18)µ2 λ1 λ3 + λ2 λ3 µ1c−(1 + c−)(cid:18)1 + λ2 λ3 +(cid:26)−µ1(cid:18)µ2 1 + λ2 λ3 (1 + c−)(cid:19)(cid:18) λ2 2 λ2 3 (1 + c−)(cid:19)(cid:18)(1 + c+) + c−(cid:19) (1 + c−)2 − λ1 λ3 (1 + c−)(cid:19)(cid:18)1 + λ1λ2 λ2 3 c−(1 + c−)(cid:19)(cid:27) [kp2q2kp1jkp2j] λ2 λ3 µ2 1c−(1 + c−)(cid:19) [kp2q2kp1jkp1j]. 1 + λ2 λ3 (1 + c−)(cid:19)(cid:18)1 + (1 + c−)2(cid:19)(cid:27) [kp1q1kp1jkp1j] λ1λ2 λ2 3 λ1λ2 λ2 3 c−(1 + c−)(cid:19) +µ1c−(cid:18)1 + µ1 λ2 λ3 + {0} [kp3q3kp1jkp3j] = λ(2) 1 [kp1q1kp1jkp1j] + λ(2) λ(2) 1 λ(2) 2 Thus [kp2q2kp1jkp2j] = − Then from (VII) [kp3q3kp1jkp3j] = − 2 [kp2q2kp1jkp2j]. [kp1q1kp1jkp1j]∗! λ(1) 1 λ(1) 3 [kp1q1kp1jkp1j] − λ(1) 2 λ(1) 3 [kp2q2kp1jkp2j] = 1 − λ(1) 2 λ(2) λ(1) 3 λ(2) λ(1) 2 1 λ(2) 2 [kp1q1kp1jkp1j]∗! 1 a+ µ2 c− 1 + c+ [kp3q3kp1jkp3jk] + µ2 a3 −b− a3 +b+ 1 1 + c+ [kp3q3kp1jkp3q2k] c− 1 + c+! [kp1q1kp1jkp1jk] − µ2 c− 1 + c+ [kp3q3kp1jkp2q2k] 65 3 !!(cid:18)1 + λ(1) 2 λ(1) λ1λ2 λ2 3 c−(1 + c−)(cid:19)) [kp1q1kp1jkp1j]∗ λ1λ2 λ2 3 c−(1 + c−)(cid:19)) [kp1q1kp1jkp1j]∗ (1 + c−)3(cid:19) [kp3q3kp1jkp2j] − + λ3 λ3 2 λ3 3 1 λ(1) 3 λ(2) 1 λ(2) 2 λ2 (1 + c+)(1 + c−) + λ(1) (1 + c−)3(cid:19) [kp1q1kp1jkp3j] 1 + c+(cid:19) c−(1 + c−) − 1!(cid:18)1 + (1 + c−) − 1 λ(2) λ1 λ3 λ2 λ3 2 λ2 3 1 2 1 − λ(1) 2 λ(2) λ(1) 3 λ(2) 2 − λ3 λ3 2 λ3 3 λ2 2 λ2 3 From (IV) (cid:18)1 + =( λ1 From (V) (cid:18)1 + =(− c−(cid:18) λ2 + λ(1) From (VI) (cid:18)1 + =( λ(2) λ(2) 1 λ(2) 2 λ3 2 λ3 3 λ2 λ3 1 λ(2) 2 − (1 + c−)3(cid:19) [kp2q2kp1jkp1j] (1 + c+)(1 + c−)!(cid:18)1 + (1 + c−)(cid:18) λ2 1 λ(2) 2 λ2 3 λ1λ2 λ2 3 2 (1 + c−)2 − c−(1 + c−)(cid:19) + 1 − λ(1) 2 λ(2) λ(1) λ(1) 3 λ(2) 2 9 [kp1q1kp1jkp1q1k] 1 a+ Now [kp1q1kp1jkp1q1k] = −p[3][4] p[2] = −p[3][4] p[2] = −p[3][4] p[2] = −p[3][4] p[2] = −p[3][4] p[2] [kp1q1kp1jkp1jk] + p[3][4] p[2] a+ 1 − µ2 1 − λ(1) 2 λ(2) λ(1) λ(1) 3 λ(2) [kp1q1kp1jkp1jk] − µ2 1 λ(2) 1 a+ 1 a+ c− 1 2 2 [kp1q1kp1jkp1jk] + [kp1q1kp1jkp3q3k] [kp3q3kp1jkp3q3k] 1 + c+ λ1 λ3 c−(cid:19)) [kp1q1kp1jkp1j]∗ 1 a+ [kp1q1kp1jkp1jk] − a− a+ [kp1q1kp1jkp1q3k] 2(1 + c+)−3[kp2q2kp1jkp2q2k] c− 1 + c+! [kp1q1kp1jkp1jk] − µ2 1 + c+! [kp1q1kp1jkp1jk] c− 2 −µ2 −µ3 c− c− 1 1 2 2 2 2 1 1 + µ2 2 1 a+ 1 + c+ a− a+ 1 λ(2) 2 1 λ(2) 2 1 λ(2) 1 λ(2) 2 λ(2) 1 λ(2) 2 µ2 2(1 + c+)−3[kp2q2kp1jkp2jk] + µ2 2 a+ 1 − µ2 a+ 1 − µ2 2 λ(2) 1 − λ(1) λ(1) 3 λ(2) λ(1) λ(1) 2 λ(2) 1 − λ(1) λ(1) 3 λ(2) a+ 1 − µ2 a+ 1 − µ2 1 − λ(1) 2 λ(2) λ(1) λ(1) 3 λ(2) 2(1 + c+)−3[kp2q2kp1jkp1q1k] 1 − λ(1) 2 λ(2) λ(1) λ(1) 3 λ(2) 2(1 + c+)−3[kp1q1kp1jkp1q1k] 1 λ(2) 1 = −p[3][4] p[2] = −p[3][4] p[2] +p[3][4] p[2] = −p[3][4] p[2] = −p[3][4] p[2] a+ 1 − µ2 ⇒ −p[3][4] p[2] 2(1 + c+)−3(cid:1) [kp1q1kp1jkp1q1k]∗ =(cid:0)1 + µ3 From the above computation [kp2q2kp1jkp2q2k] (1 + c+)3[kp1q1kp1jkp1q1k] + p[3][4] p[2] (1 + c+)−3! µ3 = µ2 − µ2 1 − λ(1) 2 λ(2) λ(1) λ(1) 3 λ(2) a+ 1 − µ2 = µ3[kp1q1kp1jkp1q1k]∗, where 1 − λ(1) 2 λ(2) λ(1) λ(1) 3 λ(2) λ(2) 1 λ(2) 2 λ(2) 1 λ(2) 2 λ(2) 1 λ(2) 2 1 λ(2) 2 1 + c+ 1 + c+ + µ2 2 + µ2 2 + µ2 2 1 µ2 2 2 c− = − 1 2 2 (1 + c+)−3[kp2q2kp1jkp2q1k] (1 + c+)−3! [kp1q1kp1jkp1jk] (1 + c+)−3! [kp1q1kp1jkp1jk] (1 + c+)−3! [kp1q1kp1jkp1jk] λ(1) 2 λ(2) 1 − λ(1) λ(1) 3 λ(2) 2 1 λ(2) 2 c− 1 + c+! [kp1q1kp1jkp1jk] . 1 − λ(1) 2 λ(2) λ(1) λ(1) 3 λ(2) λ(2) 1 λ(2) 2 1 λ(2) − 2 2 c− 1 + c+ 2 c− 1 + c+ 1 − µ2 We also obtain [kp3q3kp1jkp3q3k] = − 1 µ2 c+(1 + c+)[kp1q1kp1jkp1q1k] − p[3][4] p[2] = µ4[kp1q1kp1jkp1q1k]∗, where 1 − µ2 c− 1 + c+ λ(1) 2 λ(2) 1 − λ(1) λ(1) 3 λ(2) 2 1 λ(2) 2 + µ2 2 λ(2) 1 λ(2) 2 (1 + c+)−3! µ4 = µ2 2 1 (1 + c+)(1 + c−) + 1 − λ(1) 2 λ(2) λ(1) λ(1) 3 λ(2) 2 1 λ(2) 2 − µ2 λ(2) 1 λ(2) 2 1 (1 + c+)(1 + c−) . 1 a+ 1 µ2 c+(1 + c+)[kp1q1kp1jkp1jk] (We know from the Hilbert series of [31, Theorem 3.1] that all paths of length 9 that end at q1, q2, q3 are zero.) Let u be the path [ijr1p1jkp1q1ki], which is symmetric. This path is non-zero in A since 1 µ! [ijr1p1jkp1q1ki] = −a+[ijr1p1q1kp1q1ki] = −b−[ijr1p1q1r2p1q1ki]. + a− p[3][4] p[2] We choose uiν(i) := u. Then vjν(i)ν(aij) is given by [kijr1p1jkp1q1k] = −p[2] p[4] [kp1jr1p1jkp1q1k] −p[2] p[4] [kp2jr1p1jkp1q1k] −p[2] p[4] 66 [kp3jr1p1jkp1q1k] [kp2jr1p1jkp1q1k] [kp2q2kp1jkp1q1k] λ2 λ3 c−(1 + c+)(cid:19)(cid:19) 6= 0. a+ a+ λ2 λ3 λ1 λ3 a+ a− b− a+ a+ a− a+ a− a−b− a+b+ = d1[kp1q1kp1jkp1q1k], (1 + c+)[kp1q1kp1jkp1q1k] (1 + c+)[kp3q3r1p1jkp1q1k] −p[2] p[4] (1 + c+)[kp3q3kp1jkp1q1k] −p[2] p[4] c−(cid:19) [kp1q1kp1jkp1q1k] c−(1 + c+)(cid:19) [kp2q2kp1jkp1q1k] = −p[2] p[4] = −p[2] p[4] +p[2] p[4] = p[2] (1 + c+)(cid:18)1 − p[4] −p[2] a−(cid:18)1 − p[4] where d1 = p[2] a−(cid:18)(1 + c+)(cid:18)1 − p[4] Now p[2] (1 − µ2 p[3][4] = p[2] p[3][4] = − p[2] (1 − µ2 p[3][4] 1(1 + c+)[kp1q1kp1jkp1jk] − µ1[kp2q2kp1jkp1jk] + µ2 1(cid:18)µ1(1 + c+) − −µ1(cid:18)1 + (1 + c−)(cid:19) [kp1q1kp1jkp1jk] µ1(1 + c−)(cid:19) [kp2q2kp1jkp1jk] µ1(1 + c−) − λ2 c−(cid:19) − µ2(cid:18)1 − λ3 + λ1µ1(1 + c−) λ3 3 + λ3 1)a−[kp1q1kp1jr1p1jk] 1)a+[kp1q1kp3q3kp1jk] [kp1q1kp3q3r1p1jk] (1 − µ2 1) λ1 λ3 = d4[kp1q1kp1jkp1q1k] 1(1 + c+) − a+b+ a−b− = µ3 = µ2 λ1 λ3 λ1 λ3 λ1 λ3 3 1(1 + c−)[kp3q3kp1jkp1jk] = µ1(cid:18)µ2 where d4 = µ1(cid:18)µ2 with d2 = λ(2) 1 λ(2) 2 − and d3 = 1 − µ2 λ3 + λ1µ1(1 + c−) λ3 3 + λ3 3 6= 0, 2(1 + c−)3 d2(cid:19) [kp1q1kp1jkp1jk] 2(1 + c−)3 d2(cid:19) d−1 c−(1 + c−)(cid:19) (1 + c−)2 − λ(2) 1 λ(2) 2 c−(cid:19) λ1 λ3 . λ1 λ3 µ1(1 + c−) − λ2 3 1(1 + c+) − λ2 λ3 (1 + c+)(1 + c−)!(cid:18)1 + λ1λ2 λ2 3 1 λ(2) 2 + λ(1) 2 λ(2) 1 − λ(1) λ(1) 3 λ(2) 1 − λ(1) 2 λ(2) λ(1) 1 λ(2) λ(1) 3 λ(2) 2 2 2 (1 + c+)(cid:18)λ2 2 λ2 3 c− + µ2 2 1 + c+ (1 + c+)3 Then vjν(i)ν(aij) = d1d−1 obtain C = 1. 4 [kp1q1kp1jr1p1jk] = d1d−1 4 ajkv′, where v′ is symmetric, and we 67 Figure 17: Labelled graph E (12) 5 5 5 M S is realised by a braided subfactor which produces the for the twisted orbifold D(12) invariant A.3 The graph E (12) The Moore-Seiberg invariant ZE (12) nimrep E (12) [26, Section 5.4], illustrated in Figure 17. The unique cell system W (up to equivalence) was computed in [25, Theorem 13.1]. For the graph E (12) the automorphism ν is the identity. There are four possible choices for the vertices j, k: these are (3,8), (5,9), (14,3) and (15, 4). We see that it will not be possible to use the quicker method described above for the case where ν = id, since conjugating the graph will not interchange j ↔ k for any of the choices of j, k. We choose i = 10, j = 15 and k = 4. We will write out a basis for the space of paths which start from the vertices 10, 15, 4. Here q = e2πi/12, and we will write [m] for the quantum integer [m]q. Paths starting at vertex 10: 5 l 2 3 4 5 6 7 Paths [10, 15, 2], ([10, 15, 4] = 0) [10, 15, 2, 9], [10, 15, 2, 6] [10, 15, 2, 9, 16], [10, 15, 2, 9, 13] = [10, 15, 2, 6, 13] (cid:18)[10, 15, 2, 7] = − [2] [4] [10, 15, 4, 7] = 0(cid:19) [10, 15, 2, 7, 13] − p[2]3 p[3][4] [2] p[3] [10, 15, 2, 6, 13] = − p[2]3 p[3][4] [2] [10, 15, 2, 9, 13, 3] [10, 15, 2, 9, 13, 5], [10, 15, 2, 6, 13, 1] = [10, 15, 2, 7, 12] = 0! [10, 15, 2, 6, 12] = −p[2] p[4] [10, 15, 2, 9, 16, 5] = −p[4] p[2] [10, 15, 2, 9, 13, 1] = − p[2]3 [4]p[3] p[3][4] [10, 15, 2, 9, 13, 5, 8] = p[4] [10, 15, 2, 9, 16, 5, 8] = −p[4] p[2] p[2] [10, 15, 2, 9, 13, 3, 7] = p[3] [10, 15, 2, 9, 13, 1, 7] − p[2] (cid:16)[10, 15, 2, 9, 16, 5, 9] = 0(cid:17) p[2][6] p[6] [10, 15, 2, 9, 13, 3, 8, 14] = −p[3][4] [10, 15, 2, 9, 16, 5, 8, 14] = p[4] p[6] p[2] [10, 15, 2, 9, 16, 5, 9, 13] = 0! [10, 15, 2, 9, 16, 5, 8, 13] = −p[2] p[4] [10, 15, 2, 6, 12, 1] = 0! [10, 15, 2, 9, 13, 3, 11, 14] [10, 15, 2, 9, 13, 3, 8], 68 (cid:16)[10, 15, 2, 9, 13, 2] = 0(cid:17) [10, 15, 2, 9, 13, 3, 11] [10, 15, 2, 9, 13, 2, 7] = 0! [10, 15, 2, 9, 16, 5, 8, 14, 4] [10, 15, 2, 9, 16, 5, 8, 17] = p[4] p[2] (cid:16)[10, 15, 2, 9, 16, 5, 8, 14, 3] = 0(cid:17) [10, 15, 2, 9, 16, 5, 8, 14, 4, 10, 15] = −p[4] p[2] [10, 15, 2, 9, 16, 5, 8, 14, 4, 10] [10, 15, 2, 9, 13, 3, 8, 17] = 0! (cid:16)[10, 15, 2, 9, 16, 5, 8, 14, 1] = −p[4][10, 15, 2, 9, 16, 5, 18, 13, 1] = 0(cid:17) (cid:16)[10, 15, 2, 9, 16, 5, 8, 14, 4, 7] = 0(cid:17) [10, 15, 2, 9, 16, 5, 8, 14, 4, 7, 15] = 0! Paths starting at vertex 15: 8 9 10 l 2 3 4 5 6 Paths [15, 2, 6], 1 [15, 2, 7, 13, 1], [2] 1 [15, 4, 7], [15, 2, 7, 12], [15, 2, 6, 13], [15, 2, 7, 13], [15, 2, 9, 16] [15, 2, 6, 12, 1] = [15, 2, 9] ([10, 15, 4] = 0) [15, 2, 6, 13, 1] = − [15, 2, 7, 12, 1] = − (cid:16)[15, 2, 7, 15] = 0(cid:17) [15, 2, 6, 13]∗! [15, 2, 7] = −p[2] p[4] [15, 2, 6, 12] = −p[2] p[4] [15, 4, 7, 14] = 0! [15, 2, 7, 14] = −p[2] p[4] [15, 2, 9, 13] = [15, 2, 7, 13]∗ − p[2]3 p[3] p[3][4] p[2][4] [15, 2, 6, 13, 2] = −p[4] p[2] [15, 2, 9, 16, 5] = −p[4] p[2] [15, 2, 6, 13, 2, 7] = −p[4] p[2] = −[2][4][15, 2, 6, 13, 2, 7] − [4]p[2][6][15, 2, 6, 13, 3, 7], ⇒ [15, 2, 6, 13, 3, 7] = − [15, 2, 6, 12, 2, 9] = −p[2] p[4] [15, 2, 7, 13, 1, 6] = [2][15, 2, 6, 12, 1, 6] = − p[3] [15, 2, 6, 13, 5]∗ − p[2][4] p[3] p[3] [15, 2, 6, 12, 2, 7] = p[3][4] p[2] [15, 2, 6, 12, 2] = [15, 2, 7, 12, 2] = −p[4] p[2] (cid:16)[15, 2, 7, 13, 3] = 0(cid:17) [15, 2, 6, 12, 2, 6] = p[2]3[4] p[3] p[2]3[4] [15, 2, 6, 13, 2, 9] = [15, 2, 6, 13, 3, 11] [15, 2, 6, 13, 3, 8], [15, 2, 6, 13, 3], [15, 2, 6, 13, 5], [15, 2, 7, 13, 5] [15, 2, 6, 12, 1, 8], [15, 2, 9, 13, 5] = [15, 2, 6, 13, 5, 9] [15, 2, 7, 13, 2], [2] [4] [2][4] 1 [4] [2] [15, 2, 7, 13, 5]∗! = [15, 2, 7, 13, 2, 6] = −[2][4][15, 2, 7, 13, 1, 6] [15, 2, 6, 12, 1, 7] = −[4]p[3][15, 2, 6, 13, 1, 7] [4]p[2][6] [3]2 [15, 2, 6, 13, 2, 7]∗! [2][4] p[3] [15, 2, 6, 13, 5, 8] = −[15, 2, 6, 13, 3, 8] − p[6] p[4] [15, 2, 7, 13, 5, 8] = −p[6] p[4] [15, 2, 6, 12, 1, 7, 13] = −p[6] p[4] [15, 2, 6, 12, 1, 8, 17], [15, 2, 6, 13, 3, 8, 14] [15, 2, 7, 13, 1, 8] = − [15, 2, 6, 12, 1, 8, 13], (cid:16)[15, 2, 6, 12, 1, 7, 12] = 0(cid:17) (cid:18)[15, 2, 6, 12, 2, 9, 16] = [2] [4] [15, 2, 6, 12, 2, 9, 13] = [15, 2, 6, 13, 5, 9, 16] = 0(cid:19) [15, 2, 6, 13, 2, 6] [15, 2, 6, 12, 1, 8]∗! [15, 2, 6, 13, 1, 8] = −[15, 2, 6, 13, 3, 8]∗ + p[6] [4]p[2] [2]p[6] p[4] [15, 2, 6, 12, 1, 7, 15] = − ⇒ [15, 2, 7, 13, 1, 6] = 0 [15, 2, 6, 12, 1, 8]∗! [15, 2, 6, 12, 1, 7, 14] = −p[4][15, 2, 6, 12, 1, 8, 14], [15, 2, 6, 12, 2, 7, 15] = 0! [15, 2, 6, 12, 2, 7, 13] = −[2][15, 2, 6, 12, 1, 7, 13]∗! p[3] (cid:16)[15, 2, 6, 13, 3, 8, 17] = 0(cid:17) p[3] [2] 1 69 [15, 2, 6, 12, 1, 7, 13]∗! [15, 2, 6, 13, 3, 8, 14] [3]2 [15, 2, 6, 12, 1, 7, 14, 1] [15, 2, 6, 13, 3, 8, 14, 4] [15, 2, 6, 13, 3, 7, 13] = − p[2][4] [15, 2, 6, 13, 3, 8, 13] = p[2][6] p[3] [15, 2, 6, 13, 3, 11, 14] = −[15, 2, 6, 13, 3, 7, 14] − p[6] p[2][3] [15, 2, 6, 13, 3, 8, 14]∗! [15, 2, 6, 13, 1, 7, 14]∗ − p[6] p[2][3] = p[3]5 p[2]3[4][6] [15, 2, 6, 12, 1, 7, 14, 3] = p[4] [15, 2, 6, 12, 1, 7, 13, 1] = −p[6] p[4] (cid:16)[15, 2, 6, 12, 1, 7, 13, 2] = 0(cid:17) [15, 2, 6, 12, 1, 8, 13, 5] = 0! [15, 2, 6, 12, 1, 7, 13, 5] = −p[6] p[4] [15, 2, 6, 12, 1, 8, 17, 3] = −[15, 2, 6, 12, 1, 8, 13, 3] − p[6] p[2][3] = p[4] p[6] [15, 2, 6, 13, 3, 8, 14, 3] = −p[2][3] p[6] [15, 2, 6, 12, 1, 7, 13, 1, 7] = p[2][6] p[3] (cid:16)[15, 2, 6, 12, 1, 7, 13, 1, 6] = 0(cid:17) [15, 2, 6, 12, 1, 7, 13, 1, 8] = −p[4] p[6] [15, 2, 6, 13, 3, 8, 14, 4, 7] = −p[6] p[4] [15, 2, 6, 12, 1, 7, 13, 3] + p[4][6] p[2][3] [15, 2, 6, 13, 3, 8, 13, 3] = p[3]5 p[4][6] [15, 2, 6, 12, 1, 7, 13, 3, 11] = p[3] p[2][6] [15, 2, 6, 12, 1, 7, 13, 3, 8] = −p[3][4] [6]p[2] p[6]3 [15, 2, 6, 12, 1, 7, 13, 1, 7] − p[3]5 p[4][6] [3]2p[6] p[2][4]3 [6]p[2][4]3 (cid:18)q[6]3 + [3][4](cid:19) [15, 2, 6, 12, 1, 7, 13, 1, 7]∗! [15, 2, 6, 12, 1, 7, 14, 1, 8] = − [15, 2, 6, 12, 1, 7, 13, 1, 7, 15] [3]p[4] [15, 2, 6, 12, 1, 7, 13, 3, 7], [3][4] = − = − = [6] [3]2 [15, 2, 6, 12, 1, 7, 13, 1, 8] [15, 2, 6, 12, 1, 7, 13, 3, 7] (cid:16)[15, 2, 6, 12, 1, 7, 13, 1, 7, 12] = 0(cid:17) (cid:16)[15, 2, 6, 12, 1, 7, 13, 1, 7, 13] = 0(cid:17) [15, 2, 6, 13, 3, 8, 14, 4, 10, 15] = −p[4] p[2] (cid:16)[15, 2, 6, 12, 1, 7, 13, 1, 7, 15, 2] = 0(cid:17) [2][4][6](cid:18)q[6]3 + [3][4](cid:19) [15, 2, 6, 12, 1, 7, 13, 1, 7, 15]∗(cid:19) (cid:16)[15, 2, 6, 12, 1, 7, 13, 1, 7, 15, 4] = 0(cid:17) [15, 2, 6, 13, 3, 8, 14, 4, 7, 15] [3]2 = [15, 2, 6, 13, 3, 8, 14, 1, 7] − [15, 2, 6, 13, 3, 8, 14, 3, 7] [15, 2, 6, 12, 1, 7, 13, 3] = p[3] p[2][6] (cid:16)[15, 2, 6, 12, 1, 7, 14, 4] = 0(cid:17) [15, 2, 6, 12, 1, 7, 14, 3], [15, 2, 6, 12, 1, 8, 14, 3] [3]p[6](cid:18)[3] +q[6]3(cid:19) [15, 2, 6, 12, 1, 7, 13, 3]∗! [15, 2, 6, 12, 1, 7, 13, 3]∗! [15, 2, 6, 13, 3, 8, 14, 4, 10] [15, 2, 6, 12, 1, 7, 14, 3, 11] = 0! [15, 2, 6, 12, 1, 7, 14, 3, 8] ⇒ [15, 2, 6, 12, 1, 7, 13, 1, 8] = 0! (cid:16)[15, 2, 6, 12, 1, 7, 13, 1, 7, 14] = 0(cid:17) 7 8 9 10 l 2 3 4 Paths starting at vertex 4: Paths [4, 7, 12], [4, 7, 13], [4, 7, 13, 1], [4, 7, 15] = −p[2] p[4] [4, 7, 12, 1] = − p[4] p[2]3 [4, 7, 15, 2] = p[2]3 (cid:16)[4, 7, 13, 3] = 0(cid:17) p[3][4] [4, 7, 13, 2, 6] = −p[3][4, 7, 13, 1, 6] = p[2]3[3] p[4] [4, 10, 15] ([4, 7, 14] = 0) [4, 7, 12, 2], [4, 7, 13, 2], [4, 7, 13, 5] [2] [4, 7, 12, 2]∗ + [4, 7, 13, 2]∗! [4, 7, 12, 1, 6] = −q[2]3[4, 7, 12, 2, 6] p[3] (cid:16)[4, 7, 15,4] = 0(cid:17) 70 = −[2][4][4, 7, 13, 2, 6] +p[2][3][4][4, 7, 15, 2, 6] ⇒ [4, 7, 15, 2, 6] = p[3] p[2][4] [4, 7, 13, 2, 7] = p[3] [4, 7, 12, 2, 7] [2] [4, 7, 12, 1, 7] = p[2] p[4] (⇒ [4, 7, 15, 2, 7] = 0) [4, 7, 13, 2, 6] = [2][3] [4] [4, 7, 12, 1, 6]∗! [4, 7, 13, 5, 9] [2] [4, 7, 15, 2, 7] [4, 7, 12, 2, 9], [4, 7, 12, 1, 8], [4, 7, 13, 1, 7] = −p[2][3] p[4] = [4, 7, 13, 2, 7] − p[3] p[2][4] [4, 7, 13, 2, 9] = −p[2] p[4] [4, 7, 13, 2, 7] = p[3] [4, 7, 12, 1, 7]∗! [4, 7, 13, 1, 7] = −p[2][3] p[4] [4, 7, 13, 5, 8] = −p[6] [4, 7, 13, 1, 8] =q[2]3[6][4, 7, 12, 1, 8]∗! p[4] [4, 7, 12, 1, 6, 12] = −p[2][4][4, 7, 12, 1, 7, 12], [4, 7, 12, 1, 7, 14] = −p[4][4, 7, 12, 1, 8, 14], [4, 7, 12, 1, 7, 15] = − p[3] [4, 7, 12, 1, 8, 13] = −p[2]3 p[6] [4, 7, 12, 2, 9, 13] = − p[2]3 p[3][4] [4, 7, 13, 2, 9, 13] = − p[2]3 p[3][4] [4, 7, 12, 2, 7, 15] = 0! [4, 7, 12, 1, 6, 13]∗ − p[4] p[6] p[3] [4, 7, 12, 1, 6, 13]∗ − [2][4, 7, 12, 1, 7, 13]∗! p[3] = p[2]3 p[4]3 [4, 7, 12, 2, 6, 13] + [4, 7, 13, 2, 6, 13] + [4, 7, 12, 1, 8, 17], [2] [2] 1 [4, 7, 12, 1, 6, 13]∗ − p[2]3 p[4] [4, 7, 12, 1, 6, 13, 3], [4, 7, 12, 1, 7, 13]∗! = − [2]3 [4] [4, 7, 12, 1, 7, 13, 1], [4, 7, 13, 2, 7, 13] [4, 7, 12, 1, 6, 13], [4, 7, 12, 1, 7, 13], [4, 7, 12, 2, 9, 16] [4, 7, 13, 2, 9, 16] = −p[2] p[4] [4, 7, 12, 1, 7, 13]∗! [4, 7, 12, 2, 7, 13] [4, 7, 13, 5, 9, 16] = 0! 5 6 7 (We know from the Hilbert series of Theorem 3.1 that all paths of length 6 which start at 4 and end at 2 are zero.) [4, 7, 12, 1, 6, 13, 5], [4, 7, 12, 1, 7, 13, 3] [4, 7, 15, 4, 7, 12, 1] = 0! 1 [3] [4, 7, 12, 1, 7, 13, 3]∗! [4, 7, 12, 1, 7, 13, 1]! [4, 7, 12, 1, 7, 14, 3] = p[2][6] p[3] [4, 7, 15, 2, 7, 12, 1] = = [2] [4] [2][3] [4, 7, 12, 1, 7, 13, 1] p[4][6] [4, 7, 12, 1, 6, 12, 1] − p[4] p[6] [4, 7, 12, 1, 7, 13, 1] = −p[4] p[6] [4, 7, 12, 1, 6, 12, 1] = [4, 7, 15, 2, 6, 12, 1] = − p[4] [3]p[2] [4, 7, 12, 1, 7, 14, 1] = −p[2]3 [4, 7, 12, 1, 7, 12, 1] − p[4] p[6] p[6] [4, 7, 12, 1, 7, 13, 5] = −p[2]3 [4, 7, 12, 1, 6, 13, 5]∗! p[4] [4, 7, 12, 1, 8, 17, 3] = −[4, 7, 12, 1, 8, 13, 3] − p[6] p[2][3] [3]p[4]! [4, 7, 12, 1, 7, 13, 3]∗! (cid:16)[4, 7, 12, 1, 7, 14, 4] = 0(cid:17) [4, 7, 12, 1, 7, 13, 3] + p[6] [4, 7, 12, 1, 6, 13, 3] + p[4] p[6] p[2][3][4] [4, 7, 12, 1, 6, 13, 3]∗ + p[4] p[6] [4, 7, 12, 2, 9, 16, 5] = −p[4] p[2] = p[2]3 p[6] = p[2]3 p[6] [4, 7, 12, 1, 6, 13, 5]∗(cid:19) [4, 7, 12, 2, 9, 13, 5] = − [4, 7, 12, 1, 8, 14, 3] We will let ξ denote the path [4, 7, 12, 1] of length 3 [2][3]2 [2] [4] = − [6] [4] + [4, 7, 12, 1, 7, 14, 3] [ξ, 7, 13, 1, 7], [ξ, 7, 13, 1, 8], [ξ, 6, 13, 3, 11] [4, 7, 12, 1, 6, 13, 5] +p[2][4][4, 7, 12, 1, 7, 13, 5] [ξ, 6, 13, 3, 7] = p[3] p[2][6] [ξ, 6, 13, 1, 7] = 0! 71 [ξ, 7, 13, 1, 8]∗! [ξ, 7, 13, 3, 8] [6] [ξ, 7, 14, 1, 8] = [3]p[4] [ξ, 7, 13, 1, 8] − p[4] p[2]3 [ξ, 7, 14, 3, 11] = 0! [ξ, 6, 13, 5, 8, 14] (We know from the Hilbert series of Theorem 3.1 that all paths of length 8 which start at 4 and end at 12, 13, 17 are zero.) [3] [ξ, 7, 14, 3, 8] = − p[6] [ξ, 7, 13, 5, 8] = − p[6] p[2]3 (cid:16)[ξ, 6, 13, 5, 9] = 0(cid:17) [ξ, 7, 13, 3, 11] = p[3] p[2][6] [ξ, 6, 13, 3, 8, 14] = p[6] p[2][3] [ξ, 7, 13, 1, 7, 15] [ξ, 7, 13, 3, 8] = p[3] (cid:16)[ξ, 7, 13, 1, 6] = 0(cid:17) p[2][6] [ξ, 6, 13, 3, 8] = −[4, 7, 12, 1, 6, 13, 5, 8] = p[4] p[2]3 [6]p[2]3 (cid:18)q[6]3 + [3][4](cid:19) [ξ, 7, 13, 1, 8]∗! = − 1 [ξ, 7, 13, 3, 7] = − p[3] [ξ, 7, 13, 1, 7]∗! p[2][6] [ξ, 7, 13, 1, 7, 14] = −p[4][ξ, 7, 13, 1, 8, 14], [ξ, 6, 13, 3, 11, 14] = −[ξ, 6, 13, 3, 7, 14] − p[6] p[2][3] [2]2p[3][6](cid:18)q[6]3 + [3][4](cid:19) [ξ, 7, 13, 1, 8, 14]∗! [ξ, 7, 13, 1, 7, 14, 4] = −p[2] p[4] (cid:16)[ξ, 7, 13, 1, 7, 14, 1] = 0(cid:17) (cid:16)[ξ, 7, 13, 1, 7, 14, 3] = 0(cid:17) [ξ, 7, 13, 1, 7, 14, 4, 10] = −p[2] [ξ, 7, 13, 1, 7, 15, 4, 10] = 0! p[4] [ξ, 7, 13, 1, 7, 15, 4] = 1 8 9 10 (cid:16)[ξ, 7, 13, 1, 7, 15, 2] = 0(cid:17) (cid:16)[ξ, 7, 13, 1, 7, 14, 4, 7] = 0(cid:17) We choose uiν(i) = u10ν(10) = [10, 15, 2, 9, 16, 5, 8, 14, 4, 10]. Then v15ν(10)ν(a10,15) = [15, 2, 9, 16, 5, 8, 14, 4, 10, 15] [2] [2] [15, 2, 6, 13, 5, 8, 14, 4, 10, 15] [15, 2, 6, 13, 3, 8, 14, 4, 10, 15] [15, 2, 6, 12, 1, 8, 14, 4, 10, 15] [15, 2, 7, 13, 5, 8, 14, 4, 10, 15] + ([2] + [4]) [15, 2, 6, 12, 1, 8, 14, 4, 10, 15] p[3] p[3] [15, 2, 6, 12, 1, 8, 14, 4, 10, 15]− = −p[2][4] p[3] = p[2]3[6] p[3] +p[2][6] [4]p[3] = p[2][6] [4]p[3] −p[3]3 [4][6] (cid:16)p[6]3 + [3][4](cid:17) [15, 2, 6, 12, 1, 7, 13, 1, 7, 15] = −p[2]3[3][6] p[4]3 +p[2][3]3 [6]p[4]3(cid:16)p[6]3 + [3][4](cid:17) [15, 2, 7, 12, 1, 7, 13, 1, 7, 15] [2]p[3]3 [4]2[6] (cid:16)p[6]3 + [3][4](cid:17) [15, 4, 7, 12, 1, 7, 13, 1, 7, 15] = ca15,4v4ν(15), [15, 2, 6, 12, 1, 7, 14, 4, 10, 15] = − a4,10v10ν(4) = [4, 10, 15, 2, 9, 16, 5, 8, 14, 4] = −p[4] p[2] 72 [4, 7, 15, 2, 9, 16, 5, 8, 14, 4] and Figure 18: Labelled graph E (24) = [2] [4] [2] = − [4, 7, 13, 2, 9, 16, 5, 8, 14, 4] [4, 7, 12, 1, 6, 13, 5, 8, 14, 4] + [4, 7, 12, 2, 9, 16, 5, 8, 14, 4] −p[2][4] p[3] p[3] [2]2p[3]3 p[3] = −p[2][3]3 [4][6] (cid:16)p[6]3 + [3][4](cid:17) [4, 7, 12, 1, 7, 13, 1, 8, 14, 4] = p[2][3]3 [6]p[4]3(cid:16)p[6]3 + [3][4](cid:17) [4, 7, 12, 1, 7, 13, 1, 7, 14, 4] [2]p[3]3 [4]2[6] (cid:16)p[6]3 + [3][4](cid:17) [4, 7, 12, 1, 7, 13, 1, 7, 15, 4] = cv4ν(15)ν(a15,4). [4, 7, 13, 5, 9, 16, 5, 8, 14, 4] = − Then C = 1. E (24) graph for the conformal embedding SU (3)21 ⊂ (E7)1 A.4 For the graph E (24), illustrated in Figure 18, the automorphism ν is the identity. The unique cell system W (up to equivalence) was computed in [25, Theorem 14.1]. We choose i = 1, j = 9 and k = 17. We will write out a basis for the space of paths which start from the vertices 1, 9. Paths starting at vertex 1: Paths [1, 9, 18] ([1, 9, 17] = 0) [1, 9, 18, 6] =: ξ1 [1, 9, 18, 6, 13], (cid:16)[1, 9, 18, 2] = 0(cid:17) [1, 9, 18, 6, 15] (cid:16)[1, 9, 18, 6, 11] = 0(cid:17) l 2 3 4 5 6 [1, 9, 18, 6, 13, 22] = − [1, 9, 18, 6, 15, 24] [4] [4] [1, 9, 18, 6, 15, 22], p[9] p[9] (cid:16)[ξ1, 13, 22, 6] = 0(cid:17) (cid:16)[ξ1, 13, 22, 5] = 0(cid:17) [4]p[3] p[5][9] 73 [ξ1, 13, 22, 4], [ξ1, 13, 22, 7] = − [ξ1, 15, 22, 7] = [ξ1, 15, 24, 7], [ξ1, 15, 24, 8] (cid:16)[1, 9, 18, 6, 13, 20] = 0(cid:17) 7 8 9 10 11 12 13 [3] [ξ1, 13, 22, 4, 12], [ξ1, 13, 22, 7, 14], [ξ1, 13, 22, 7, 16] = [ξ1, 13, 22, 4, 14, 19], [ξ1, 13, 22, 4, 14] = −p[7] [ξ1, 15, 24, 7, 16] = −p[2][3][4] [4]p[3] p[5][9] p[5][9] [ξ1, 13, 22, 7, 15] = [4]p[3] (cid:16)[ξ1, 13, 22, 4, 11] = 0(cid:17) p[5][9] [ξ1, 13, 22, 4, 12, 19] = −p[5][7] [3]p[9] [ξ1, 13, 22, 7, 14, 23] = p[7] [ξ1, 13, 22, 4, 14, 23] = −p[7] p[3][5] (cid:16)[ξ1, 13, 22, 4, 14, 22] = 0(cid:17) (cid:16)[ξ1, 13, 22, 7, 16, 24] = 0(cid:17) [ξ1, 13, 22, 4, 12, 19, 4] = − p[7] [3]p[5] [ξ1, 13, 22, 4, 12, 19, 5] = [ξ1, 13, 22, 4, 14, 23, 3] = −p[3][5] [7]p[5] [3]2p[9] [ξ1, 13, 22, 4, 14, 19, 3] − p[9] [ξ1, 13, 22, 4, 12, 21, 4], [ξ1, 13, 22, 4, 12, 19, 2] [3] [4] [4] [ξ1, 15, 24, 8, 16] [ξ1, 15, 24, 7, 15] = 0! [ξ1, 13, 22, 4, 12, 21] = − [ξ1, 13, 22, 7, 16, 23] [3]p[5] p[7] [ξ1, 13, 22, 4, 14, 21], [ξ1, 13, 22, 4, 12, 21, 3], [ξ1, 13, 22, 4, 12, 19, 3] = p[5] p[3] [ξ1, 13, 22, 7, 14, 19, 5] = 0! [ξ1, 13, 22, 4, 14, 21, 3] [ξ1, 13, 22, 4, 12, 19, 3] + p[7][9] = p[3]3[9] [4]p[7] [4][5]p[3][7](cid:0)[3]2[5] + [7](cid:1) [ξ1, 13, 22, 4, 12, 19, 3]∗! p[9] [3][4] = [ξ1, 13, 22, 7, 14, 21, 3] We will let ξ2 denote the path [ξ1, 13, 22, 4, 12] = [1, 9, 18, 6, 13, 22, 4, 12] of length 7. (cid:16)[ξ1, 13, 22, 4, 14, 23, 7] = 0(cid:17) [ξ2, 19, 2, 11] = −p[7] [3] [ξ2, 19, 4, 11], [ξ2, 19, 2, 9], [ξ2, 19, 3, 10], [ξ2, 19, 2, 10] = −p[4] p[2] [ξ2, 19, 3, 12] = −p[3][ξ2, 19, 4, 12] [ξ2, 19, 3, 14] = −p[7] [3] [7] [3]2p[5] ⇒ [ξ2, 19, 4, 14] = 0! ⇒ [ξ2, 19, 3, 14] = 0 [ξ2, 19, 4, 14] = [ξ2, 21, 4, 14] = − [ξ2, 21, 3, 14] = − [7] [3]2[5] [ξ2, 19, 3, 14] [7] p[3]5[5] [ξ2, 19, 2, 11, 20], (cid:16)[ξ2, 19, 3, 12, 21] = 0(cid:17) [3] [ξ2, 19, 4, 12, 19] [ξ2, 19, 3, 12, 19] [ξ2, 19, 2, 11, 18], [ξ2, 19, 2, 11, 22] = −p[7] [ξ2, 19, 2, 9, 18] = −p[5] [ξ2, 19, 2, 9, 17] = −p[3][ξ2, 19, 2, 10, 17], p[3] [ξ2, 19, 2, 10, 19] = −p[3][ξ2, 19, 2, 11, 19] = p[7] [ξ2, 19, 4, 11, 19] = −p[3][9] p[3] p[2][4] [ξ2, 19, 4, 11, 22] = 0! = p[9] p[2][4] [ξ2, 19, 2, 9, 17, 2] = −p[3][ξ2, 19, 2, 9, 18, 2] =p[5][ξ2, 19, 2, 11, 18, 2] = −p[3][5][ξ2, 19, 2, 11, 19, 2] =p[5][ξ2, 19, 2, 10, 19, 2], [ξ2, 19, 2, 11, 18, 6] = −p[9] [3]p[9] p[2][4][7] (cid:16)[ξ2, 19, 2, 9, 17, 1] = 0(cid:17) (cid:16)[ξ2, 19, 2, 10, 19, 3] = 0(cid:17) (cid:16)[ξ2, 19, 2, 11, 19, 4] = 0(cid:17) [ξ2, 19, 2, 11, 18, 2, 11] = p[2][3] [ξ2, 19, 2, 11, 18, 6, 11] = −p[2] p[4] p[4] [ξ2, 19, 2, 11, 18, 6, 13] = −p[9] [ξ2, 19, 2, 11, 20, 6, 13] = −p[7][9] [ξ2, 19, 2, 11, 19, 5] = − [ξ2, 19, 2, 11, 19, 5, 11], [ξ2, 19, 2, 11, 19, 2, 11] [ξ2, 19, 2, 11, 20, 5, 13] [ξ2, 19, 2, 11, 20, 6], [4][5] [4] [4] [ξ2, 19, 2, 11, 20, 5] = −p[2][7] p[3][4] [7]p[2] [3][5]p[4] = [ξ2, 19, 2, 11, 19, 5, 13], [ξ2, 19, 2, 11, 18, 6, 15] 74 14 15 16 17 18 19 20 21 22 l 2 3 4 5 [ξ3, 6, 11, 22, 7, 16, 23] (cid:16)[ξ3, 6, 11, 22, 7, 15] = 0(cid:17) [ξ3, 6, 11, 22, 7, 16]∗! [ξ3, 6, 11, 22, 7, 14, 23] = p[7] p[3][5] [ξ3, 6, 11, 22, 7, 14, 22] = 0! [3] (cid:16)[ξ3, 6, 11, 22, 5, 14, 21, 4] = 0(cid:17) (cid:16)[ξ3, 6, 11, 22, 5, 14, 21, 3, 14] = 0(cid:17) [4] [ξ3, 6, 11, 22, 5, 14, 21, 3] = − [ξ3, 6, 11, 22, 5, 14, 22] = −p[7] (cid:16)[ξ3, 6, 11, 22, 5, 14, 19] = 0(cid:17) (cid:16)[ξ3, 6, 11, 22, 7, 16, 24] = 0(cid:17) p[9] (cid:16)[ξ3, 6, 11, 22, 5, 14, 23, 7] = 0(cid:17) (cid:16)[ξ3, 6, 11, 22, 5, 14, 21, 3, 12] = 0(cid:17) (cid:16)[ξ3, 6, 11, 22, 5, 14, 21, 3, 10, 19] = 0(cid:17) (cid:16)[ξ3, 6, 11, 22, 5, 14, 21, 3, 10, 17, 2] = 0(cid:17) [ξ3, 6, 11, 22, 5, 14, 21, 3, 10, 17, 1] [ξ3, 6, 11, 22, 5, 14, 21, 3, 10, 17] [ξ3, 6, 11, 22, 5, 14, 21, 3, 10] [ξ3, 6, 11, 22, 5, 14, 23, 3] (cid:16)[ξ3, 6, 11, 22, 5, 14, 21, 3, 10, 17, 1, 9] = 0(cid:17) Paths starting at vertex 9: We will let ξ3 denote the path [ξ2, 19, 2, 11, 18] = [1, 9, 18, 6, 13, 22, 4, 12, 19, 2, 11, 18] of length 11. (cid:16)[ξ2, 19, 2, 11, 19, 5, 14] = 0(cid:17) [ξ2, 19, 2, 11, 19, 5, 11, 22] = p[2][3][9] p[4][5] [ξ3, 6, 15, 22] = −p[3][5] [ξ3, 6, 11, 22] − p[9] [4] [4] [ξ2, 19, 2, 11, 19, 5, 13, 22] = p[3]2[5][9] [7] [ξ3, 6, 13, 22] = − [ξ3, 6, 13, 22] [ξ3, 6, 11, 22]∗! [2][7] p[3]3[5] [ξ3, 6, 11, 22, 7], [ξ3, 6, 15, 24, 8] [ξ3, 6, 13, 22, 6] = −p[3]3[5]3 ⇒ [ξ3, 6, 11, 22, 6] = 0(cid:19) [7] [ξ3, 6, 13, 20, 6] (cid:16)[ξ3, 6, 11, 22, 4] = 0(cid:17) [ξ3, 6, 15, 24] [ξ3, 6, 13, 20], [ξ3, 6, 11, 22, 5], (cid:16)[ξ2, 19, 2, 11, 19, 2, 10] = 0(cid:17) (cid:16)[ξ2, 19, 2, 11, 18, 2, 9] = 0(cid:17) [ξ3, 6, 11, 20] = p[5] p[3] [ξ3, 6, 11, 22] = −p[2][7] p[3][4] (cid:16)[ξ3, 6, 11, 18] = 0(cid:17) [ξ3, 6, 11, 20, 5] = −p[5][7] [3]p[9] [ξ3, 6, 11, 20, 6] = −p[3][5] p[9] [ξ3, 6, 15, 24, 7] = −p[5] p[3] [ξ3, 6, 11, 22, 5, 14] = −p[7] (cid:16)[ξ3, 6, 11, 20, 5, 11] = 0(cid:17) [ξ3, 6, 15, 24, 8, 16] = −p[4] [7]p[2][4] p[2] [ξ3, 6, 11, 22, 5, 14, 23] = −p[7] (cid:16)[ξ3, 6, 11, 20, 5, 13] = 0(cid:17) ⇒ [ξ3, 6, 11, 20, 6] = 0 [ξ3, 6, 11, 22, 5, 14, 21], [ξ3, 6, 15, 24, 7, 16] = [ξ3, 6, 11, 22, 6] = − [ξ3, 6, 11, 22, 7, 14], [ξ3, 6, 15, 22, 7] = [ξ3, 6, 11, 20, 6] [2][7] [3]2 [3]2[5] [3]2[5] = − [3]2 [7] [3] [7] [3] [ξ3, 6, 11, 22, 7]∗! [ξ3, 6, 11, 22, 7, 16] ([9, 17, 1] = 0) [9, 17, 2, 11], [9, 18, 6, 13], [9, 18, 6, 15] [9, 18, 6] [9, 18, 2, 11] = p[2] p[3][4] (cid:16)[9, 17, 2, 10] = 0(cid:17) Paths [9, 17, 2] = −p[3][9, 18, 2], [9, 18, 6, 11] = −p[2] p[4] (cid:16)[9, 17, 2, 9] = 0(cid:17) [9, 18, 6, 13, 20] = p[3] p[5] [9, 18, 6, 11, 18] = p[2] p[3][4] [9, 18, 6, 15, 22] = −p[3][5] [9, 18, 6, 13, 22, 5] = − p[7] [3]p[5] [4] [9, 18, 6, 11, 20], [9, 18, 6, 11, 22], [9, 18, 6, 13, 22], [9, 18, 6, 15, 24] [9, 18, 6, 11, 19] = p[2] [9, 17, 2, 11, 18] = 0! p[3][4] [9, 18, 6, 13, 22]∗! [9, 18, 6, 11, 22]∗ − p[9] [9, 18, 6, 13, 20, 5] = − p[7] [5]p[3] [9, 18, 6, 11, 20, 5] = [4] 75 [9, 17, 2, 11, 19] = 0! [9, 18, 6, 11, 22, 5], [7] p[3]3[5][9] 6 7 [9, 18, 6, 11, 22, 7], [9, 18, 6, 13, 22, 6] = p[5] p[3] [9, 18, 6, 15, 24, 7] = −p[5] p[3] [9, 18, 6, 15, 22, 6] = −p[3][5] [9, 18, 6, 15, 22, 5] = −p[3][5] p[3]3[5][9] = p[7] ⇒ [9, 18, 6, 13, 22, 4, 11] = [9, 18, 6, 13, 22, 5, 11] = [3][5] [4] [4] [7] [9, 18, 6, 13, 22, 6, 11] = [9, 18, 6, 13, 22, 5, 14], [9, 18, 6, 13, 20, 6] = [9, 18, 6, 11, 20, 6] = −p[3][5] p[9] [9, 18, 6, 13, 22, 7], [9, 18, 6, 13, 22, 4], [9, 18, 6, 15, 24, 8] [9, 18, 6, 11, 22, 6], (cid:16)[9, 18, 6, 11, 22, 4] = 0(cid:17) [9, 18, 6, 13, 22, 7]∗! [4] ! [9, 18, 6, 13, 22, 6] = 0! − p[9] [9, 18, 6, 15, 22, 7] = [4] [4] [5] [4] [9, 18, 6, 11, 22, 7]∗ + p[5][9] [4]p[3] [9, 18, 6, 13, 22, 6] = p[9] [9, 18, 6, 11, 22, 6] − p[9] [9, 18, 6, 13, 22, 5] = −[2]p[9][9, 18, 6, 13, 22, 5]∗! [9, 18, 6, 11, 22, 5] − p[9] [9, 18, 6, 11, 22, 5, 11] = − p[7] p[3][5][9] [7]p[2][10] [3]2p[5]3 [4]p[2][5] p[10] ⇒ [9, 18, 6, 13, 22, 6, 11] = [9, 18, 6, 13, 22, 5, 11]∗ [9, 18, 6, 13, 22, 4, 11] − [9, 18, 6, 13, 22, 5, 11] [9, 18, 6, 11, 22, 6, 11] p[7] [3]2[5] [3][5] [4] [7] [9, 18, 6, 13, 22, 7, 15], [9, 18, 6, 11, 22, 7, 16], [9, 18, 6, 13, 22, 5, 11]∗! [9, 18, 6, 11, 22, 5, 13] [9, 18, 6, 13, 22, 7, 16]∗! [9, 18, 6, 13, 22, 7, 16] [9, 18, 6, 13, 22, 6, 13] = p[5] p[9] [9, 18, 6, 11, 22, 6, 13] = −p[3][5] p[9] ⇒ [9, 18, 6, 13, 22, 5, 13] = 0(cid:19) [9, 18, 6, 13, 22, 5, 14]∗! [9, 18, 6, 11, 22, 5, 14] = −p[3][5][9] p[7] [9, 18, 6, 13, 22, 6, 15]∗! [9, 18, 6, 11, 22, 6, 15] = p[4][9] p[2][3][5] [9, 18, 6, 13, 22, 5, 14]∗! [9, 18, 6, 13, 22, 4, 14]∗ − p[7] [9, 18, 6, 11, 22, 7, 16]∗ − p[5][9] p[2][4] p[2][3][4] [9, 18, 6, 15, 24, 7, 16] = − [3] [5] [9, 18, 6, 13, 22, 5, 14, 19], [9, 18, 6, 13, 22, 5, 14, 21], [9, 18, 6, 13, 22, 6, 15] = −p[2] p[4] [9, 18, 6, 13, 22, 4, 14], [9, 18, 6, 13, 22, 4, 12], 1 [9, 18, 6, 13, 22, 5, 13] = − [7] [3] [3] = − [3]2[5] [9, 18, 6, 13, 22, 5, 13] p[3] [9, 18, 6, 11, 22, 7, 14] = −p[7] [9, 18, 6, 13, 22, 7, 15] = −p[4] p[2] [9, 18, 6, 13, 22, 7, 14] = −p[7] [9, 18, 6, 15, 24, 8, 16] = −p[4] p[2] [9, 18, 6, 13, 22, 5, 11, 19] = −p[10] p[4] [9, 18, 6, 13, 22, 5, 11, 22] = −p[2][4] p[5] p[3] [3]p[5] p[7] (cid:16)[9, 18, 6, 13, 22, 5, 11, 18] = 0(cid:17) (cid:16)[9, 18, 6, 13, 22, 7, 15, 22] = −p[3][9, 18, 6, 13, 22, 7, 14, 22] [9, 18, 6, 13, 22, 7, 15, 24] = − [9, 18, 6, 13, 22, 4, 12, 21] = − [9, 18, 6, 13, 22, 7, 16, 24], = p[7] p[3] = p[7][10] p[3][4] [9, 18, 6, 13, 22, 4, 14, 22] + p[7] p[3] [9, 18, 6, 13, 22, 4, 11, 22] − p[5][7] p[2][3][4] 1 [9, 18, 6, 11, 22, 7, 16, 23] = −p[5] p[3] [9, 18, 6, 13, 22, 4, 14, 19] = −p[2][4] p[5] [9, 18, 6, 13, 22, 5, 14, 22], [9, 18, 6, 13, 22, 4, 14, 21], (cid:16)[9, 18, 6, 13, 22, 5, 11, 20] = 0(cid:17) [9, 18, 6, 13, 22, 5, 14, 22] [9, 18, 6, 13, 22, 5, 14, 23], [9, 18, 6, 13, 22, 4, 12, 19], [9, 18, 6, 13, 22, 4, 14, 23] [9, 18, 6, 13, 22, 5, 11, 22] = p[3]3[5]3 p[2][4][7] [9, 18, 6, 13, 22, 5, 14, 23]∗! [9, 18, 6, 13, 22, 5, 11, 22]∗! [9, 18, 6, 11, 22, 7, 14, 23] = [9, 18, 6, 13, 22, 4, 11, 19] − [9, 18, 6, 13, 22, 4, 12, 19] [5]p[9] p[7] [3]p[9] p[5][7] 76 8 9 We will let ξ denote the path [9, 18, 6, 13, 22] of length 4. [3]p[9] p[5][7] [ξ, 7, 15, 22, 7] = − [3][5]2 [4] [4] [9, 18, 6, 13, 22, 5, 11, 19]∗ − [9, 18, 6, 13, 22, 7, 14, 23] = [ξ, 5, 14, 21, 4] − [ξ, 5, 11, 22, 4] [3]p[9] p[5][7] [12]p[5][7] [3][4]p[9] [ξ, 5, 14, 21, 3], [ξ, 4, 12, 19, 2], [ξ, 5, 11, 19, 4] = −[ξ, 5, 14, 19, 4] = [ξ, 5, 14, 22, 4] [ξ, 5, 11, 19, 3], [ξ, 5, 11, 22, 5], [ξ, 4, 12, 21, 3], [ξ, 4, 12, 21, 4] [ξ, 5, 14, 21, 3]∗! ⇒ [ξ, 5, 14, 21, 4] = [9, 18, 6, 13, 22, 4, 12, 19]∗! [9, 18, 6, 13, 22, 5, 14, 23]∗! (cid:16)[ξ, 5, 11, 22, 6] = 0(cid:17) [ξ, 5, 14, 19, 3] − p[9] [4] = − [2]p[4]3 p[10] = p[5][7] p[3]3 [ξ, 5, 14, 19, 3]∗ − p[9] (cid:16)[ξ, 7, 15, 24, 8] = 0(cid:17) [4] [ξ, 5, 14, 21, 3] = p[3][5] p[4][10] [ξ, 5, 11, 22, 7]∗! [9, 18, 6, 13, 22, 7, 16, 23] = −p[5] p[3] [ξ, 5, 14, 23, 7], [ξ, 5, 14, 22, 7] = p[2][4] p[3][5] [ξ, 4, 12, 19, 4] = − p[7] [3]p[5] [ξ, 5, 14, 21, 4] + p[2][4] p[5] [ξ, 5, 11, 22, 4]∗! [3]p[9] p[5][7] [9, 18, 6, 13, 22, 4, 14, 23]∗ + p[5][7] p[3]3 [ξ, 5, 11, 19, 5] = − p[5] p[2][4] [ξ, 5, 11, 22, 4] = p[4] p[10] [ξ, 5, 11, 22, 7] = −p[2][4] p[5] [ξ, 4, 12, 19, 3] = p[5] p[3] (cid:16)[ξ, 5, 11, 19, 2] = 0(cid:17) [ξ, 5, 14, 23, 3] = −p[3][5] [ξ, 7, 15, 24, 7] = −p[5] p[3] p[2][4][7] [ξ, 4, 14, 23, 3] = −p[3][5] [ξ, 5, 11, 19, 3] + p[3]3[9] [ξ, 4, 12, 19, 3] + p[7][9] [4]p[7] [3][4]p[5] [ξ, 5, 11, 19, 3]∗! [2]p[3][4][5] p[10] [ξ, 4, 12, 19, 5] = −p[2][4][7] [ξ, 4, 11, 19, 5] − p[5][7] [3]p[9] [3]p[9] [ξ, 4, 11, 22, 5] + p[4][5][7] [3]p[9][10] [ξ, 4, 12, 19, 3]∗! [ξ, 4, 12, 21, 3] = p[3] (cid:16)[ξ, 4, 14, 23, 7] = −p[3][ξ, 4, 14, 22, 7] = −p[2][3][4][5][ξ, 5, 11, 22, 7]∗(cid:17) p[5] [ξ, 5, 11, 19, 4, 12] = − p[4] [ξ, 5, 11, 22, 4, 12] = p[4] p[10] p[3][10] [ξ, 5, 11, 22, 4, 11] = p[5] p[2][10] [ξ, 4, 12, 19, 2, 10] = −p[4] p[2] [ξ, 4, 12, 19, 3, 12] = −p[3][ξ, 4, 12, 19, 4, 12] [ξ, 5, 14, 21, 3, 12] = p[7] [ξ, 5, 14, 21, 3, 14] = −p[3][ξ, 5, 14, 21, 4, 14] = − [ξ, 5, 11, 22, 7, 14] = p[2][4] p[3][5] [5]p[3][4] p[2]3[7][9] (cid:16)[ξ, 5, 11, 19, 3, 10] = 0(cid:17) [3][4]p[5][9] [12]p[5][7] [4]p[3][9] p[3][5] [2]p[3][4][5] p[10] [2]p[7][9] [5]p[3] = p[5][7] [3]p[9] [ξ, 5, 11, 22, 4, 12]∗! [ξ, 5, 11, 22, 4, 14]∗! (cid:16)[ξ, 5, 11, 22, 5, 13] = 0(cid:17) [ξ, 5, 11, 22, 5, 11], [ξ, 5, 11, 22, 4, 14], [ξ, 5, 14, 21, 3, 10], [ξ, 4, 12, 19, 2, 9], [ξ, 4, 14, 19, 3] − p[9] [4] [ξ, 4, 12, 19, 3, 10] = − [ξ, 4, 14, 23, 3, 10], [ξ, 4, 12, 19, 2, 11], [5][12]p[2][7] [3]p[9][10] [ξ, 5, 11, 22, 5]∗! [ξ, 5, 14, 21, 4, 12] = [3] [7][12] [ξ, 5, 11, 19, 3, 14], [ξ, 5, 11, 19, 3, 12], [ξ, 4, 14, 19, 5] [ξ, 4, 14, 22, 5] = = = [ξ, 4, 12, 19, 3]∗ + [ξ, 4, 14, 21, 3] [ξ, 4, 12, 21, 3] [ξ, 5, 14, 23, 7, 14] = − [ξ, 5, 14, 23, 3, 14] [4] 77 = − p[4] p[10] [ξ, 5, 11, 19, 3, 14] + p[9] p[3][5] [ξ, 5, 14, 21, 3, 14] = − p[4] p[10] (cid:16)[ξ, 5, 11, 22, 7, 15] = 0(cid:17) [ξ, 4, 12, 19, 3, 14] = −p[7] [3] (cid:16)[ξ, 5, 11, 22, 7, 16] = 0(cid:17) [ξ, 4, 12, 19, 4, 14] − p[2][7][10] [3]p[5] = [ξ, 4, 12, 21, 4, 14] − [ξ, 5, 11, 22, 5, 14] [ξ, 4, 12, 19, 5, 14] [ξ, 5, 11, 19, 3, 14]∗ − [12]p[7] [3][4] [ξ, 5, 14, 22, 4, 14]∗! = − [ξ, 4, 12, 21, 3, 14] − [ξ, 5, 11, 22, 4, 14] − [ξ, 5, 11, 19, 3, 14] = − [ξ, 4, 12, 19, 3, 14] − [ξ, 5, 11, 22, 4, 14] − [ξ, 5, 11, 19, 3, 14] 10 [2][12]p[4][5][7] [3]p[9][10] [2][12]p[4][5][7] [3]p[9][10] [ξ, 5, 11, 22, 4, 14]∗! [12]p[5]3 [4]p[9] [7] [7] [3]2p[5] p[3]5[5] [7] [3]2[5] [2][7][12]p[5] [3]2p[9] [2][7][12]p[5] [3]2p[9] [2][7][12]p[5] [3]2p[9] [3][12]p[5]3 p[4][7][9][10] ⇒ [ξ, 4, 12, 19, 3, 14] = − [ξ, 5, 11, 19, 3, 14]∗ − [ξ, 4, 12, 19, 4, 11] = − [3] p[7] [ξ, 4, 12, 19, 2, 11] − [ξ, 4, 12, 19, 5, 11] [3] [3] = − [ξ, 5, 11, 22, 5, 11] [ξ, 4, 12, 19, 2, 11] − = − = − [ξ, 5, 11, 19, 3, 14, 19], [ξ, 5, 11, 22, 4, 14, 22], [ξ, 5, 11, 22, 4, 11, 19], [ξ, 5, 11, 19, 3, 14, 22], [ξ, 5, 11, 22, 4, 11, 18], [ξ, 4, 12, 19, 2, 11]∗ − [ξ, 5, 11, 19, 3, 14]∗ − p[7] p[7] [12]p[5] p[4][9][10] [5][12]p[2][7] [3]p[9][10] [2][12]p[5][7] [3]p[9] [12]p[5][7] [3][4]p[9] [ξ, 5, 11, 22, 4, 11]∗! [ξ, 4, 12, 21, 4, 14] = p[7] p[3]3[5] [ξ, 5, 11, 22, 4, 14]∗! [ξ, 4, 12, 19, 4, 14] = − p[7] [3]p[5] [ξ, 5, 11, 19, 3, 12, 19] = −p[3][5] p[9] [ξ, 5, 11, 22, 4, 11, 22] = p[4] p[10] [ξ, 4, 12, 19, 2, 9, 17] = −p[3][ξ, 4, 12, 19, 2, 10, 17], [ξ, 4, 12, 19, 2, 10, 19] = −p[3][ξ, 4, 12, 19, 2, 11, 19], [ξ, 5, 11, 22, 4, 11, 20] = p[5] (cid:16)[ξ, 5, 11, 19, 3, 14, 23] = 0(cid:17) p[2][10] [ξ, 5, 11, 22, 4, 14, 19] = −p[2][4] [3]p[9] p[5][7] p[5] [ξ, 5, 11, 22, 4, 11, 19]∗ + p[3][4][9] p[5][7][10] [ξ, 5, 11, 22, 4, 14, 21] = − p[7] [ξ, 5, 11, 22, 4, 12, 21] = p[4][7] p[3]3[5][10] [3]p[5] [ξ, 5, 11, 22, 4, 14, 23] = −[ξ, 5, 11, 22, 5, 14, 23] − ⇒ [ξ, 5, 11, 22, 4, 14, 23] = 0! [ξ, 5, 14, 21, 3, 10, 19] = −p[9] [ξ, 5, 14, 21, 3, 12, 19] = p[3][5] = −p[2][4] p[5] [ξ, 5, 11, 22, 4, 11, 19] − [3] p[7] [4] [4] [ξ, 5, 11, 19, 3, 12, 19]∗! [ξ, 4, 12, 21, 3, 14] = p[7] [3][5] [ξ, 4, 12, 19, 3, 14] [ξ, 5, 11, 19, 3, 12, 21] = −p[3] p[5] [ξ, 5, 11, 19, 3, 14, 21], [ξ, 5, 14, 21, 3, 10, 17], [ξ, 4, 12, 19, 2, 9, 18] = −p[5] p[3] [ξ, 4, 12, 19, 2, 11, 20] [ξ, 4, 12, 19, 2, 11, 18], [ξ, 5, 11, 22, 5, 11, 20] = 0! [ξ, 5, 11, 22, 4, 12, 19] [ξ, 5, 11, 19, 3, 12, 21]∗! [ξ, 5, 11, 22, 7, 14, 23] = −[ξ, 5, 11, 22, 4, 14, 23] [ξ, 5, 11, 19, 3, 12, 19] + [ξ, 5, 11, 22, 4, 14, 19] = = [7][12] [2][7][12] p[3]3[4]3[5][10] p[3]3[4][5][10] [ξ, 5, 11, 19, 3, 12, 19]∗ − [ξ, 5, 14, 21, 3, 14, 19] [ξ, 5, 11, 22, 4, 11, 19]∗! [5][12]p[7] [4]2p[9] [12]p[2][5][7] p[4]3[9] 78 [ξ, 4, 12, 19, 3, 12, 19] = − [ξ, 4, 12, 19, 3, 14, 19] [4] p[9] [ξ, 4, 12, 19, 3, 10, 19] − p[3][5] p[9] [ξ, 5, 11, 22, 4, 11, 19]∗! [ξ, 5, 11, 19, 3, 14, 19] + [5]2[12]p[3] [4][9] [ξ, 4, 12, 19, 2, 10, 19]∗ − [ξ, 4, 12, 19, 2, 10, 19] + [ξ, 5, 11, 22, 4, 14, 19] [ξ, 4, 12, 19, 2, 11, 22] = −p[7] [3] [ξ, 4, 12, 19, 4, 11, 22] − p[7] [3] [ξ, 4, 12, 19, 5, 11, 22] [ξ, 4, 12, 19, 4, 14, 22] − [ξ, 5, 11, 22, 5, 11, 22] [5]2[12]p[3]3 [9]p[4][7][10] [12]p[2][3][5]3 [9]p[4] [5][7][12]p[2] [3]2p[9][10] [7][12]p[5] [3]2p[4][9][10] [6][7][12]p[5][9] [2][3]3[4] = p[2][4] p[9] = p[2][4] p[9] = − p[4][7] [3]p[10] [12]p[5][7] [3][10]p[9] [12]p[5][7] [3][10]p[9] [5][12]p[3]3 p[4][7][9][10] = − = = [ξ, 5, 11, 19, 3, 14, 22] + [ξ, 5, 11, 22, 4, 14, 22] − [ξ, 5, 11, 22, 4, 11, 22] [2][7][12]p[5] [3]2p[9] [ξ, 5, 11, 22, 4, 11, 22]∗! 11 [ξ, 5, 11, 19, 3, 14, 22]∗ − [ξ, 5, 11, 19, 3, 12, 21]∗! [ξ, 4, 12, 19, 3, 14, 21] [ξ, 5, 11, 19, 3, 14, 21] − [ξ, 5, 11, 19, 3, 12, 19, 2], [ξ, 5, 11, 19, 3, 12, 21, 3], [ξ, 5, 11, 19, 3, 12, 21, 4], [ξ, 5, 11, 22, 4, 14, 21] = − [2][12]p[4][5][7] [3]p[9][10] [ξ, 4, 12, 19, 3, 12, 21] = p[3] p[5] [5][12]p[3] [4]p[9] [ξ, 5, 11, 19, 3, 12, 19, 3] = p[5] p[3] [ξ, 5, 11, 19, 3, 12, 19, 4] = − p[7] [3]p[5] [ξ, 5, 11, 19, 3, 12, 19, 5] = −p[3][5] p[9] [ξ, 5, 11, 22, 4, 11, 18, 2] = −p[3][ξ, 5, 11, 22, 4, 11, 19, 2], [ξ, 4, 12, 19, 2, 9, 17, 2] = −p[3][ξ, 4, 12, 19, 2, 9, 18, 2] =p[5][ξ, 4, 12, 19, 2, 11, 18, 2] = −p[3][5][ξ, 4, 12, 19, 2, 11, 19, 2] =p[5][ξ, 4, 12, 19, 2, 10, 19, 2], [ξ, 5, 11, 19, 3, 14, 22, 4] = − p[5] p[2][4] = p[9] p[2][3][4] [ξ, 5, 11, 19, 3, 14, 19, 5] = p[3][4][5] p[9][10] [ξ, 5, 11, 19, 3, 12, 21, 4] = p[2][4][9] p[3] [ξ, 5, 11, 19, 3, 12, 19, 4] − p[3][5][9] p[2][4][7] [ξ, 5, 11, 19, 3, 12, 19, 4]∗! [3]p[9] p[2][4][7] [ξ, 5, 11, 19, 3, 14, 19, 4] − [ξ, 5, 11, 19, 3, 14, 22, 6], [ξ, 5, 11, 19, 3, 14, 22, 5], [ξ, 5, 14, 21, 3, 10, 17, 1], [ξ, 5, 11, 19, 3, 14, 21, 4] [ξ, 4, 12, 19, 2, 10, 19, 5] [ξ, 4, 12, 19, 2, 9, 18, 6], [ξ, 5, 11, 22, 4, 11, 22, 6] = − [ξ, 5, 11, 22, 5, 11, 22, 6] [5]p[3] [4]p[2][10] [ξ, 5, 11, 22, 4, 14, 22, 6] + p[3]3[5] [10]p[7] [ξ, 5, 11, 19, 3, 14, 22, 6] (cid:16)[ξ, 5, 11, 19, 3, 14, 22, 7] = 0(cid:17) [ξ, 5, 11, 22, 4, 11, 18, 6] = −p[3][5] [4] [4] [ξ, 5, 11, 19, 3, 14, 22, 6] = p[3][5] [ξ, 5, 11, 22, 5, 14, 22, 6] = p[3][5] p[4][10] p[4][10] [ξ, 5, 11, 22, 4, 11, 22, 6] + p[3]3[5] = p[3][5] [10]p[7] = −[ξ, 5, 11, 22, 4, 11, 18, 6] + p[3]3[5] [10]p[7] [4]p[3]3[5] [10][12]p[7] [3]p[9] p[2][4][7] ⇒ [ξ, 5, 11, 22, 4, 11, 18, 6] = [ξ, 5, 11, 19, 3, 14, 22, 6]∗! [ξ, 5, 11, 22, 4, 12, 19, 3] − p[5] p[2][4] [ξ, 5, 11, 22, 4, 12, 21, 3] + p[9] p[2][3][4] = p[3][5][9] p[2][4][7] [ξ, 5, 11, 19, 3, 14, 22, 6] [ξ, 5, 11, 22, 4, 11, 19, 3] = − [ξ, 5, 11, 22, 4, 11, 19, 4] = p[5] p[2][10] [ξ, 5, 11, 22, 5, 11, 19, 4] = p[5] p[2][4] 79 [ξ, 5, 11, 22, 4, 14, 21, 3] = [ξ, 5, 11, 22, 4, 14, 19, 3] [4]p[2][7][9] [3]2p[5][10] [ξ, 5, 11, 19, 3, 12, 21, 3]∗! [ξ, 5, 11, 22, 5, 11, 22, 4] [ξ, 5, 11, 22, 4, 14, 22, 5] [ξ, 5, 11, 19, 3, 12, 19, 5] [3]p[4] p[7][10] = −[ξ, 5, 11, 22, 5, 14, 22, 4] = −[ξ, 5, 11, 22, 4, 14, 22, 4] − [ξ, 5, 11, 19, 3, 14, 22, 4] = −[ξ, 5, 11, 22, 4, 11, 19, 4] + = −p[10] p[4] [ξ, 5, 11, 19, 3, 12, 21, 4] [ξ, 5, 11, 19, 3, 12, 21, 4] [ξ, 5, 11, 22, 4, 11, 22, 4] + [4]p[2][3][9] p[7][10] [4]p[2][3][9] p[7][10] [ξ, 5, 11, 19, 3, 12, 19, 4]∗! [4]2p[2][3][9] [12]p[7][10] [ξ, 5, 11, 22, 4, 11, 22, 5] = − p[5] p[2][10] [ξ, 5, 11, 22, 4, 14, 19, 5] = −[ξ, 5, 11, 22, 4, 11, 19, 5] + p[3][9] p[2][7][10] [ξ, 5, 11, 19, 3, 12, 19, 5]∗! [4]p[3][9] [12]p[2][7][10] [ξ, 5, 11, 22, 4, 14, 22, 7] = 0! ⇒ [ξ, 5, 11, 22, 4, 11, 19, 4] = − [ξ, 5, 11, 22, 4, 11, 19, 5] = − p[5] p[2][4] = p[5] p[2][4] ⇒ [ξ, 5, 11, 22, 4, 11, 19, 5] = [ξ, 5, 11, 22, 4, 11, 22, 7] = p[4] p[10] [ξ, 5, 14, 21, 3, 10, 17, 2] = −p[5] p[3] [2][7][12] [ξ, 5, 11, 19, 3, 12, 19, 2] + [ξ, 5, 11, 22, 4, 11, 19, 2] [ξ, 5, 14, 21, 3, 10, 19, 2] [5][12]p[2][7] p[3][4]3[9] [5][12]p[2][7] [3]p[4]3[9] (cid:16)[ξ, 4, 12, 19, 2, 10, 19, 3] = 0(cid:17) [ξ, 5, 11, 19, 3, 12, 19, 2]∗ − (cid:16)[ξ, 4, 12, 19, 2, 9, 17, 1] = 0(cid:17) [ξ, 4, 12, 19, 2, 10, 19, 4] = −p[3][ξ, 4, 12, 19, 2, 11, 19, 4] = −p[3][10] p[4] [6][7][12]p[5][9][10] [2]p[3]5[4]3 [6][7][12]p[5][9] [2][4]p[3]5 [ξ, 5, 11, 19, 3, 14, 22, 4] + [ξ, 5, 11, 19, 3, 12, 21, 4] + = [ξ, 5, 11, 19, 3, 12, 19, 4]∗! [ξ, 4, 12, 19, 2, 11, 20, 5] = −p[2][4][7] [3]p[9] [ξ, 4, 12, 19, 2, 11, 19, 5] − p[5][7] [3]p[9] [5][7][12] [3]2[9][10] [ξ, 5, 11, 22, 4, 11, 18, 2]∗! [ξ, 4, 12, 19, 2, 11, 22, 4] [ξ, 5, 11, 22, 4, 11, 22, 4] [ξ, 5, 11, 22, 4, 11, 19, 4] [ξ, 4, 12, 19, 2, 11, 22, 5] [5][6][12]p[7]3 [2][3]4[4] = − = − [2][7][12] [3]2p[4][10] [3]2p[4][10] = = − [12]p[5][7] p[3][4][9][10] [12]p[2][5][7] [3]p[10] [4][6]p[2][7][10] [3]p[5] = p[2][4][7] p[3]3[9] = p[2][4][7] p[3]3[9] [4]p[3] p[5][9] [4]p[3] p[5][9] = = [ξ, 4, 12, 19, 2, 10, 19, 5] − [ξ, 5, 11, 19, 3, 14, 22, 5] + [ξ, 5, 11, 22, 4, 11, 22, 5] [ξ, 4, 12, 19, 2, 11, 20, 6] = − [ξ, 4, 12, 19, 2, 10, 19, 5]∗ − [6][7]p[4][10] p[3]5[5][9] [ξ, 5, 11, 19, 3, 12, 19, 5]∗! [ξ, 4, 12, 19, 2, 11, 18, 6] − p[3][5] p[9] [ξ, 5, 11, 19, 3, 14, 22, 6] + [4] p[9] [ξ, 4, 12, 19, 2, 11, 22, 6] [ξ, 4, 12, 19, 2, 9, 18, 6] − [ξ, 4, 12, 19, 2, 9, 18, 6]∗ − [ξ, 5, 11, 19, 3, 14, 22, 6]∗ + [5][12]p[7] [9][10]p[3] [5][12]p[7] [9][10]p[3] [ξ, 5, 11, 19, 3, 12, 19, 2, 10] = −p[4] p[2] [ξ, 5, 11, 19, 3, 12, 19, 3, 14], 12 [ξ, 5, 11, 19, 3, 12, 19, 2, 9], [ξ, 5, 11, 19, 3, 12, 19, 3, 10], [ξ, 5, 11, 19, 3, 12, 19, 2, 11], [ξ, 5, 11, 19, 3, 12, 19, 4, 11], [ξ, 5, 11, 19, 3, 12, 19, 5, 13] = p[3][4][5] p[9][10] [ξ, 4, 12, 19, 2, 9, 17, 2, 11], [ξ, 5, 11, 19, 3, 14, 22, 5, 13] = − p[4][5] p[9][10] [ξ, 4, 12, 19, 2, 9, 18, 6, 13], [ξ, 4, 12, 19, 2, 9, 18, 6, 15] [ξ, 5, 11, 19, 3, 14, 22, 6, 13], [5][6][7][12] [2][4]p[3]5 [4][5][6]p[7] [2][10]p[3]3 [ξ, 5, 11, 22, 4, 11, 22, 6] [ξ, 5, 11, 19, 3, 14, 22, 6]∗! 80 [ξ, 5, 11, 19, 3, 12, 19, 3, 12] = −p[3][ξ, 5, 11, 19, 3, 12, 19, 4, 12] = [12]p[7][10] [4]2p[2][9] [ξ, 5, 11, 22, 4, 11, 19, 3, 12] = − [ξ, 5, 11, 19, 3, 12, 19, 3, 12] [7][12] [3]2[4][5] [ξ, 5, 11, 22, 4, 11, 19, 4, 12] = − [12]p[7][10] [4]2p[2][3][9] ⇒ [ξ, 5, 11, 19, 3, 12, 19, 3, 12] = 0) [3] [ξ, 5, 11, 19, 3, 12, 19, 5, 11] = − p[7] [ξ, 5, 11, 19, 3, 14, 22, 6, 11] = p[2][7][10] [3]p[5] [ξ, 5, 11, 19, 3, 14, 22, 4, 14] [ξ, 5, 11, 19, 3, 12, 19, 3, 14] + [2][4][ξ, 5, 11, 19, 3, 12, 19, 5, 14] [ξ, 5, 11, 19, 3, 12, 19, 2, 11]∗ − [ξ, 5, 11, 19, 3, 12, 19, 4, 11]∗! [ξ, 5, 11, 19, 3, 14, 22, 4, 11] − p[7] [3] [ξ, 5, 11, 19, 3, 14, 22, 5, 11] = [2]p[4][7][9][10] p[3]3[5] = p[9][10] p[3][4][5] [ξ, 5, 11, 19, 3, 12, 19, 4, 11] − p[7][9][10] p[3]3[4][5] [ξ, 5, 11, 19, 3, 12, 19, 2, 11]∗ + p[3][7][9][10] p[4][5] [ξ, 5, 11, 19, 3, 12, 19, 5, 11] [ξ, 5, 11, 19, 3, 12, 19, 4, 11]∗! (cid:16)[ξ, 5, 11, 19, 3, 12, 19, 4, 12] = 0(cid:17) [ξ, 5, 11, 19, 3, 12, 19, 5, 14] = p[3][4][5] p[9][10] [ξ, 5, 11, 19, 3, 14, 22, 5, 14] = −p[3][4][5] p[9][10] = − [ξ, 5, 11, 19, 3, 12, 19, 4, 14] = ⇒ [ξ, 5, 11, 19, 3, 12, 19, 5, 14] = [4]p[2][5] p[10] [3][4]p[2][5] p[7][10] [3]2[4]p[2]3[5] [6][9]p[7][10] [ξ, 5, 11, 19, 3, 12, 19, 3, 14]∗! [6][9]p[7] (cid:16)[ξ, 5, 11, 19, 3, 14, 22, 6, 15] = 0(cid:17) (cid:16)[ξ, 5, 11, 22, 4, 11, 18, 2, 9] = 0(cid:17) [ξ, 5, 11, 22, 4, 11, 18, 2, 10] = −p[3][ξ, 5, 11, 22, 4, 11, 19, 2, 10] = p[3][4] p[2] ⇒ [ξ, 5, 11, 19, 3, 12, 19, 4, 14] = [2][3]2 [ξ, 5, 11, 19, 3, 12, 19, 3, 14]∗ [ξ, 5, 11, 19, 3, 12, 19, 3, 10]∗! = − [ξ, 5, 11, 22, 4, 11, 19, 3, 10] [4]p[2][7][9] [3][5]p[10] [3][4]p[9] [12]p[2][7][10] [ξ, 5, 11, 19, 3, 14, 22, 6, 11] = − [ξ, 5, 11, 19, 3, 12, 19, 2, 11]∗ − [ξ, 5, 11, 22, 4, 11, 18, 2, 11] = −p[4] [ξ, 5, 11, 22, 4, 11, 18, 6, 11] = − p[4]3[3]3[5] [10][12]p[2][7] p[2] [ξ, 5, 11, 19, 3, 12, 19, 4, 11]∗! [3]2[4]p[9] [12]p[2][10] [ξ, 5, 11, 19, 3, 12, 19, 2, 9]∗! [ξ, 5, 14, 21, 3, 10, 17, 1, 9] = −p[4] [7][12]p[2] [3]2p[10] p[2] (cid:16)[ξ, 4, 12, 19, 2, 9, 17, 2, 9] = −p[3][ξ, 4, 12, 19, 2, 9, 18, 2, 9] = 0(cid:17) (cid:16)[ξ, 4, 12, 19, 2, 9, 17, 2, 10] = 0(cid:17) [ξ, 4, 12, 19, 2, 9, 17, 2, 11]∗! [ξ, 4, 12, 19, 2, 9, 18, 6, 11] = −p[2] [ξ, 4, 12, 19, 2, 9, 18, 2, 11] = p[2] p[4] p[3][4] [ξ, 4, 12, 19, 2, 10, 19, 5, 11] = − p[7] [ξ, 4, 12, 19, 2, 10, 19, 2, 11] − [ξ, 4, 12, 19, 2, 10, 19, 4, 11] [ξ, 5, 14, 21, 3, 10, 17, 2, 9] = [3] [3] = − [ξ, 4, 12, 19, 2, 11, 19, 2, 11] + [ξ, 4, 12, 19, 2, 9, 17, 2, 11]∗ + [4][6]p[2][7][10] [3]p[5] [4][6]p[2][7][10] [3]p[5] (cid:16)[ξ, 4, 12, 19, 2, 10, 19, 5, 13] = −p[3][ξ, 4, 12, 19, 2, 11, 19, 5, 13] [ξ, 4, 12, 19, 2, 11, 20, 5, 13] + p[3][5] p[2][4] [ξ, 4, 12, 19, 2, 11, 20, 6, 13] − p[5] p[2][4] = p[3]3 p[7] p[5][7] = p[3]3[9] p[2][4][7] [5]p[3]3[9] [7]p[2][4] = [ξ, 5, 11, 19, 3, 12, 19, 4, 11] [ξ, 5, 11, 19, 3, 12, 19, 4, 11]∗! [ξ, 4, 12, 19, 2, 11, 22, 5, 13] [ξ, 4, 12, 19, 2, 11, 22, 6, 13] 81 [ξ, 4, 12, 19, 2, 9, 18, 6, 13] − [ξ, 5, 11, 19, 3, 14, 22, 6, 13] [ξ, 5, 11, 22, 4, 11, 18, 6, 13] + [ξ, 5, 11, 22, 4, 11, 22, 6, 13] [ξ, 4, 12, 19, 2, 9, 18, 6, 13] − [ξ, 5, 11, 19, 3, 14, 22, 6, 13] [5][6][7][12]p[9] [3]3p[2]3[4]3 [5][12]p[2][4][7] [3][10]p[9] [6]p[2][4]3[7] [3]p[9] [4][6]p[2][7][10] [3]p[5] [3]2p[5] [ξ, 5, 11, 19, 3, 12, 19, 3, 14]∗! [4][6][7] [4][6][7] [3]2[5] [ξ, 5, 11, 19, 3, 12, 19, 5, 13]∗! [ξ, 4, 12, 19, 2, 10, 19, 4, 14] = − [ξ, 5, 11, 19, 3, 12, 21, 4, 14] 13 We let ξ2, ξ3 denote the paths ξ2 = [ξ, 5, 11, 19, 3] = [9, 18, 6, 13, 22, 5, 11, 19, 3], and ξ3 = [ξ, 4, 12, 19, 2] = [9, 18, 6, 13, 22, 4, 12, 19, 2] of length 8. [ξ2, 12, 19, 2, 11, 18], [ξ2, 12, 19, 2, 10, 19] = −p[3][ξ2, 12, 19, 2, 11, 19], [ξ2, 12, 19, 3, 14, 22], [ξ3, 9, 18, 6, 11, 20] = p[4][5] p[2] [ξ3, 9, 18, 6, 13, 20], [ξ3, 9, 18, 6, 13, 22], (We know from the Hilbert series of [31, Theorem 3.1] that all paths of length 13 which start at 9 and end at 17 are zero.) [ξ3, 9, 18, 6, 15, 24] = = = − [3]2p[4][5] [7]p[2] [6][12]p[5]3[9] p[2]3[3]3[4] [3]2p[4][5] [7]p[2] [3]2p[4][5] [7]p[2] [ξ, 4, 12, 19, 2, 9, 18, 6, 13]∗ + [ξ, 4, 12, 19, 2, 10, 19, 5, 14] = − p[5] p[2][4] [4][6][7] [ξ, 5, 11, 19, 3, 12, 21, 3, 14] = = p[3]5[5] [4] [ξ2, 12, 19, 2, 11, 22], [ξ2, 12, 19, 2, 11, 20], [ξ2, 12, 19, 2, 9, 18] = −p[5] p[3] [ξ3, 9, 17, 2, 11, 20] = p[3][4] p[2] [ξ2, 12, 19, 3, 14, 19] = − p[3][5] (cid:16)[ξ2, 12, 19, 3, 14, 23] = 0(cid:17) [ξ2, 12, 19, 4, 11, 19] = − p[5] p[2][4] Now [ξ2, 14, 22, 6, 11, 20] = p[5] p[3] [ξ2, 12, 19, 2, 10, 19]∗! [2]p[3]3 [6][9]p[7] = p[3][5][9][10] p[4][7] Then [ξ2, 12, 19, 4, 11, 20] = p[4][5] p[3][7][9][10] [4]p[7] ⇒ [ξ2, 12, 19, 4, 11, 20] = − [3]p[7] [ξ2, 12, 19, 2, 11, 20] − [5] [7] [3][6] [2][4] = − = − [ξ2, 12, 19, 4, 11, 22] = p[4] p[10] [ξ2, 12, 19, 5, 13, 20] = − [3]p[5] p[7] [ξ2, 12, 19, 5, 13, 22] = −p[5][7] [3]p[9] = p[5] [6][7] [ξ2, 12, 19, 2, 11, 22]∗ + = p[5] p[9] ([3]2[6] − [4][7])[ξ2, 12, 19, 2, 11, 20]∗! (cid:16)[ξ2, 12, 19, 3, 14, 21] = 0(cid:17) [ξ2, 12, 19, 4, 11, 20]. 1 [ξ2, 12, 19, 4, 14, 19] = − [ξ2, 12, 19, 2, 10, 19]∗! [ξ2, 12, 19, 2, 11, 18]∗! [ξ2, 12, 19, 3, 10, 19] = p[2][4] p[3][5] [ξ2, 12, 19, 4, 11, 18] = − [3]p[7] [3]2p[2][5] [6][9]p[4][7] [ξ2, 14, 22, 6, 13, 20] = −p[9][10] p[3][4] [ξ2, 12, 19, 2, 11, 20] − p[3][5][9][10] p[4][7] [ξ2, 12, 19, 5, 13, 20] [ξ2, 12, 19, 3, 14, 19] [ξ2, 14, 22, 6, 11, 20] − [ξ2, 12, 19, 2, 11, 20] [ξ2, 12, 19, 5, 11, 20] = −p[3]3[5][9][10] [7]p[4] 1 [3]p[7] [ξ2, 12, 19, 4, 11, 20] [ξ2, 12, 19, 2, 11, 20]∗! [2][3]2p[4] [6][9]p[7][10] [3]2p[5] [7] [ξ2, 12, 19, 4, 14, 22] = [ξ2, 12, 19, 3, 14, 22]∗! [3]p[5] p[7] [ξ2, 12, 19, 5, 11, 20] = [ξ2, 12, 19, 2, 11, 20] + [ξ2, 12, 19, 4, 11, 20] [ξ2, 12, 19, 5, 11, 22] − p[2][4][7] [3]p[9] [2]2[3]p[4]3[5] [6]p[9]3[10] [ξ2, 12, 19, 3, 14, 22]∗! [ξ2, 12, 19, 5, 14, 22] 82 (cid:16)[ξ3, 9, 17, 2, 11, 19] = 0(cid:17) [ξ3, 10, 19, 5, 11, 22] + [3] (cid:16)[ξ3, 9, 17, 2, 11, 18] = 0(cid:17) [ξ3, 9, 17, 2, 11, 22] = −p[5][7] =p[9][ξ3, 10, 19, 5, 13, 22] + p[2][4][7] [3]2p[4][5][9] [7]p[2] [3]2p[4][5][9] [7]p[2] [ξ3, 9, 18, 6, 15, 22] = −p[3][5] [ξ3, 9, 18, 6, 13, 22]∗ − [ξ3, 9, 18, 6, 13, 22] − [3] [4] = = [4][6]p[2][7][9][10] [3]p[5] [4][6]p[2][7][10] [ξ3, 9, 18, 6, 11, 22]∗ − p[9] [4] [3] [4][6][7]p[2][10] [3]2 [ξ2, 12, 19, 4, 11, 22] [6]p[2][4]3[7]3 [3]3[5] [ξ3, 10, 19, 5, 14, 22] + [ξ2, 12, 19, 3, 14, 22] [ξ2, 12, 19, 5, 13, 22] + [ξ2, 12, 19, 3, 14, 22] [6][12]p[2][4][7]3 [3]3[5] [ξ2, 12, 19, 2, 11, 22]∗ + p[2]3[4]3[7] [ξ3, 9, 18, 6, 13, 22]∗! [9] ([3]2 + 1)[ξ2, 12, 19, 3, 14, 22]∗! 14 [ξ2, 12, 19, 2, 9, 18, 6], [ξ2, 12, 19, 2, 11, 22, 4], [ξ2, 12, 19, 2, 10, 19, 4] = −p[3][ξ2, 12, 19, 2, 11, 19, 4] = −p[3][10] p[4] (cid:16)[ξ2, 12, 19, 2, 10, 19, 3] = 0(cid:17) (cid:16)[ξ2, 12, 19, 2, 10, 19, 2] = 0(cid:17) [ξ2, 12, 19, 2, 11, 20, 6], [ξ3, 9, 18, 6, 15, 24, 8], [ξ2, 12, 19, 2, 11, 22, 7], [ξ3, 9, 18, 6, 13, 22, 7] [6][7] p[5]([3]2[6] − [4][7]) [ξ2, 12, 19, 2, 11, 22, 5] − [ξ2, 12, 19, 5, 13, 20, 5] = [ξ2, 12, 19, 5, 13, 22, 5] [3][6]p[7] ([3]2[6] − [4][7]) [ξ2, 12, 19, 3, 14, 22, 5] [2]2[3]2p[4]3[5][7] p[9]3[10]([3]2[6] − [4][7]) [3]2[6] ([3]2[6] − [4][7]) [ξ2, 12, 19, 2, 11, 19, 5] + [ξ2, 12, 19, 2, 11, 20, 5] [ξ2, 12, 19, 2, 10, 19, 5], [ξ3, 9, 18, 6, 13, 22, 5], = (cid:16)[ξ2, 12, 19, 2, 9, 18, 2] = 0(cid:17) [ξ2, 12, 19, 2, 11, 20, 5] = [3][6]p[5][7] p[9]([3]2[6] − [4][7]) [3][6]p[2][4][7] p[9]([3]2[6] − [4][7]) [2]2[3]2[4]p[5][7] p[9]3([3]2[6] − [4][7]) [3]2[6] = + [ξ2, 12, 19, 2, 10, 19, 5] [ξ2, 12, 19, 2, 11, 20, 5] [ξ2, 12, 19, 2, 11, 20, 6]∗! [ξ2, 12, 19, 2, 10, 19, 4]∗! [ξ2, 12, 19, 2, 10, 19, 5]∗! [ξ2, 12, 19, 3, 14, 19, 5] = = [4] ([3]2[6] − [4][7]) [6]2[7]p[2] [12]p[3][4][5]3 ⇒ [ξ2, 12, 19, 2, 11, 20, 5] = − p[2]3[3]3 p[4][7][9]3 [ξ2, 12, 19, 2, 10, 19, 5]∗! [ξ2, 12, 19, 2, 11, 22, 5] = −p[2][4] p[5] [ξ2, 12, 19, 2, 11, 19, 5] − [ξ2, 12, 19, 2, 11, 20, 5] + p[2]3[3]3[4][7] p[9]3([3]2[6] − [4][7]) [ξ2, 12, 19, 2, 10, 19, 5]∗! [3]p[9] p[5][7] [ξ2, 12, 19, 2, 11, 18, 6]∗ − p[9] p[3][5] p[3] [ξ2, 12, 19, 3, 14, 19, 5] = −p[2][10] p[3][5] [ξ2, 12, 19, 3, 14, 19, 4] = − [ξ2, 12, 19, 2, 11, 22, 6] = − p[3][5] [ξ2, 12, 19, 3, 14, 22, 4] = − p[5] p[2][4] [ξ2, 12, 19, 3, 14, 22, 5] = −p[10] p[4] [ξ2, 12, 19, 3, 14, 22, 6] = [3][5]p[10] p[4][7] = −p[3][4][5][10] [ξ2, 12, 19, 4, 11, 18, 6] − p[3][5][9][10] p[4][7] p[7] [ξ2, 12, 19, 2, 11, 18, 6] + p[4][5][9][10] = p[4][5][10] [7]p[3] [6]p[3] [ξ2, 12, 19, 2, 9, 18, 6]∗ + p[4][5][9][10] = −p[4][10] [6]p[3] [ξ2, 12, 19, 4, 11, 22, 6] [7] 1 (cid:16)[ξ2, 12, 19, 3, 14, 22, 7] = 0(cid:17) [ξ2, 12, 19, 4, 11, 20, 6] [ξ2, 12, 19, 2, 11, 20, 6] [ξ2, 12, 19, 2, 11, 20, 6]∗! 83 − = − [ξ2, 12, 19, 2, 11, 22, 5] − [ξ3, 9, 17, 2, 11, 22, 5] = − [ξ3, 9, 18, 6, 13, 22, 5]∗ + [3]2p[9] [ξ3, 9, 17, 2, 11, 20, 5] = −p[5][7] [3]p[9] [3][5]p[4] p[2][7] [7]p[2]3[4]3[5] [4][6][7]p[2][5][10] [3]p[9]3 [ξ2, 12, 19, 2, 10, 19, 5]∗! [3][5]p[4] p[2][7] [ξ3, 9, 17, 2, 11, 20, 6] = −p[3][5] [5]p[3]5[4] p[9] [7]p[2] [4][6]p[2][5][7][10] [ξ2, 12, 19, 2, 11, 22, 6] − p[2]3[3][4]3[5][7] p[9]3 p[3][9] [4]2[6]p[2][7][10] [3]2p[4][5]3 [7]p[2] [4]3[5]p[2]5[7][10] [2]2[7]p[4]3[10] p[3]3[9]3 [4]2p[2]3[3][5][10] [ξ3, 9, 17, 2, 11, 22, 6] = − [3]p[9] [ξ2, 12, 19, 2, 11, 20, 6] + [ξ3, 9, 18, 6, 13, 20, 6] + [ξ2, 12, 19, 2, 11, 18, 6] [6][9] = − − − = − [3]2[5] [7] [ξ3, 9, 17, 2, 11, 20, 6] − [ξ2, 12, 19, 2, 9, 18, 6] − [4]2p[2]3[3][5][10] p[7][9]3 [4]p[2][3][5][10] p[7][9]3 p[7][9]3 ([3]2 + 1)[ξ2, 12, 19, 2, 9, 18, 6] [6][9] [4]3[5]p[2]5[7][10] [4]2[5]p[2]3[7][10] [6][9] [ξ2, 12, 19, 2, 11, 20, 6] [ξ2, 12, 19, 2, 11, 20, 6]∗! ⇒ [ξ3, 9, 17, 2, 11, 20, 6] = − [ξ2, 12, 19, 2, 9, 18, 6]∗ − [ξ3, 9, 17, 2, 11, 22, 4] − [ξ2, 12, 19, 2, 11, 22, 4] [2][6]p[4][7]3[10] [3]3p[5][9] [2]3[4][12]p[7]3 p[3]5[5][9]3 [ξ3, 9, 17, 11, 20, 6] [ξ2, 12, 19, 2, 10, 19, 4]∗! [7]p[2] [3]2p[4][5][9] − [ξ3, 9, 18, 6, 13, 22, 4] = [2]2[4]p[7]3 [3]2p[5][9]3 [2]p[4][5][10] p[7][9]3 [2][5][6]p[7][10] p[3]3[4] [ξ3, 9, 18, 6, 13, 22, 6] = p[5] p[3] [ξ3, 9, 18, 6, 15, 24, 7] = −p[5] p[3] [ξ2, 12, 19, 2, 9, 18, 6, 13], = − = ([3]2 + 1)[ξ2, 12, 19, 3, 14, 22, 4] = [ξ2, 12, 19, 2, 9, 18, 6]∗ − [2]2[5]p[4]3[7][10] [ξ3, 9, 18, 6, 13, 20, 6] = p[2] p[3][4] [6][9]p[3] [2]p[5][9] p[3] [ξ3, 9, 18, 6, 15, 22, 7] = [5] [4] [ξ2, 12, 19, 2, 9, 18, 6, 15], [ξ2, 12, 19, 2, 11, 20, 6]∗! [ξ3, 9, 18, 6, 11, 22, 7] + p[5][9] [4]p[3] [ξ3, 9, 18, 6, 13, 22, 7]∗! [ξ2, 12, 19, 2, 10, 19, 4, 11], [ξ2, 12, 19, 2, 11, 22, 7]∗ + [ξ3, 9, 18, 6, 13, 22, 7] [ξ3, 9, 18, 6, 13, 22, 5] ([3]2 + 1)[ξ2, 12, 19, 3, 14, 22, 5] [ξ3, 9, 18, 6, 13, 22, 6] ([3]2 + 1)[ξ2, 12, 19, 3, 14, 22, 6] 15 [ξ2, 12, 19, 2, 10, 19, 4, 14] = −p[2][10] p[5] [ξ3, 9, 18, 6, 13, 22, 5, 14], [ξ3, 9, 18, 6, 13, 22, 7, 16] [ξ2, 12, 19, 2, 10, 19, 5, 14], [ξ2, 12, 19, 2, 11, 22, 7, 16], (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 11] = 0(cid:17) (cid:16)[ξ2, 12, 19, 2, 10, 19, 4, 12] = 0(cid:17) [ξ2, 12, 19, 2, 10, 19, 5, 13] = − p[3][5] p[2][10] (cid:16)[ξ2, 12, 19, 2, 10, 19, 5, 11] = −[ξ2, 12, 19, 2, 10, 19, 4, 11]∗(cid:17) [ξ2, 12, 119, 3, 14, 22, 5, 13] = p[5] p[2][10] [5]p[4][9] [6]p[2][3] [ξ2, 12, 19, 2, 9, 18, 6, 13] + p[4][7][9] [6]p[2][3] [ξ2, 12, 19, 2, 9, 18, 6, 13] + [ξ2, 12, 19, 2, 11, 20, 5, 13] [ξ2, 12, 19, 2, 11, 20, 6, 13] [ξ2, 12, 19, 2, 9, 18, 6, 13] − [ξ2, 12, 19, 2, 10, 19, 5, 13] [2][3] [6][9] = −p[4][5] [7]p[2] = −p[4][5] [7]p[2] = −p[4][5] [7]p[2] ⇒ [ξ2, 12, 19, 2, 10, 19, 5, 13] = − [ξ2, 12, 19, 2, 11, 20, 6, 11] = −p[3][ξ2, 12, 19, 2, 11, 20, 5, 11] = [3]2p[5]3 [7]2p[2]3[4] [ξ2, 12, 19, 2, 10, 19, 4, 11]∗! [ξ2, 12, 19, 2, 9, 18, 6, 13]∗! [3]2p[2]3 p[4][7][9]3 = − [3]2p[2]3 p[4][7][9]3 [ξ2, 12, 19, 2, 10, 19, 5, 11] [ξ2, 12, 19, 3, 14, 22, 6, 13] 84 [ξ2, 12, 19, 2, 11, 20, 6, 13] = p[7] [5] [ξ2, 12, 19, 2, 11, 20, 5, 13] = − p[2]3[3]3 [5]p[4][9]3 [ξ2, 12, 19, 2, 9, 18, 6, 13]∗! Now [ξ3, 9, 18, 6, 13, 22, 6, 15] = −p[2] p[4] [ξ2, 12, 19, 2, 11, 22, 6, 15] [ξ3, 9, 18, 6, 13, 22, 7, 15] = − [ξ2, 12, 19, 2, 10, 19, 5, 13] [6]p[5][7][10] [3]p[4][9] [ξ2, 12, 19, 2, 11, 22, 7, 15] [ξ2, 12, 19, 2, 11, 18, 6, 15] − [ξ2, 12, 19, 2, 11, 20, 6, 15] [ξ2, 12, 19, 2, 9, 18, 6, 15] − [ξ2, 12, 19, 2, 11, 20, 6, 15. Then [ξ2, 12, 19, 2, 11, 20, 6, 15] [ξ3, 9, 18, 6, 13, 22, 6, 15] − [ξ3, 12, 19, 2, 9, 18, 6, 15] ([3]2 + 1)[ξ2, 12, 19, 2, 9, 18, 6, 15] − [ξ2, 12, 19, 2, 11, 20, 6, 15] ⇒ ([2][4]2[7] + [3]2[6])[ξ2, 12, 19, 2, 11, 20, 6, 15] = − [ξ2, 12, 19, 2, 11, 22, 7, 14] = −p[7] [3] ([3]2 + 1)[ξ2, 12, 19, 2, 9, 18, 6, 15]∗! [ξ2, 12, 19, 2, 11, 22, 5, 14] [ξ2, 12, 19, 2, 10, 19, 4, 14] − [ξ2, 12, 19, 2, 10, 19, 5, 14] [6]p[7][10] p[2][3]3 [6]p[7][10] p[2][3]3 [6]p[3] [2][4][7]p[5][9] [4][6]p[3] p[5][9] [ξ2, 12, 19, 2, 11, 22, 4, 14] − p[7] [6][3]2 [2][4]2[7] [3] [6]2p[2][7]3 [12]p[3]3[4][5]3 = = − = p[3]7[5] [4][7]2p[9]3 [6]p[5][7][10] [3]p[2][9] [4][6]p[7][10] p[2][3]3[9] [4][6]p[7][10] [3]p[2][5][9] [6][9]p[3] [2]2[5]p[4]3[7][10] [6]p[3] [2][4][7]p[5][9] = − = − = − = = p[4][7] p[3]3[10] [2][6]p[4][7] [5]p[3]3[10] = p[4]3 p[2][3][5] = − p[4]3 [5]p[2] [4][12]p[2]7[7]3[10] [5]p[3]5[9]3 [4]p[2]7[10] [6]p[3][7][9]5 [2]p[4][5][10] p[3][7][9]3 [2][12]p[5][10] p[3][4][7][9]3 = = = = [3] [ξ3, 9, 18, 6, 13, 22, 6,11] [ξ2, 12, 19, 2, 11, 20, 6, 11] [ξ2, 12, 19, 2, 10, 19, 4, 14]∗! [ξ2, 12, 19, 2, 11, 20, 6, 15] [ξ2, 12, 19, 2, 11, 22, 7, 15] = −p[4] p[2] [ξ2, 12, 19, 2, 11, 22, 6, 15] [ξ2, 12, 19, 2, 11, 18, 6, 15] + p[4][9] p[2][3][5] ([2][4]2[7] − [6]) [ξ2, 12, 19, 2, 9, 18, 6, 15]∗! ([2][4]2[7] + [3]2[6]) [ξ3, 9, 18, 6, 13, 22, 5, 11] = p[2][10] p[5] [ξ3, 9, 18, 6, 13, 22, 4, 11] − [ξ2, 12, 19, 2, 10, 19, 4, 11] + p[7] [2]2[5]p[3][4]3[10] ([2][7][12] + [3]7[5])[ξ2, 12, 19, 2, 10, 19, 4, 11]∗! [6][9] [ξ3, 9, 18, 6, 13, 22, 5, 13] = − 1 p[3] [ξ3, 9, 18, 6, 13, 22, 6, 13] [2]2[5]p[4]3[7][10] [3][6][9] [ξ2, 12, 19, 2, 9, 18, 6, 13]∗! [ξ2, 12, 19, 2, 9, 18, 6, 13] + [ξ2, 12, 19, 2, 11, 20, 6, 13] [ξ3, 9, 18, 6, 13, 22, 7, 14] = −p[7] [3] [ξ3, 9, 18, 6, 13, 22, 4, 14] − p[7] [3] [ξ3, 9, 18, 6, 13, 22, 5, 14] [ξ2, 12, 19, 2, 10, 19, 4, 14]∗ − p[7] [3] [ξ3, 9, 18, 6, 13, 22, 5, 14]∗! = [2]3[4][7]2[12] p[3]7[5][9]3 [ξ3, 9, 18, 6, 13, 22, 7, 15] = −p[4] p[2] [ξ3, 9, 18, 6, 13, 22, 6, 15] 85 = = [4]p[2][5][10] p[7][9]3 [3]2[4]p[2][5][10] p[7][9]3 = −p[2][4][5][9] p[3] [ξ3, 9, 18, 6, 15, 24, 8, 16] = −p[4] p[2] [4] [ξ2, 12, 19, 2, 9, 18, 6, 15] + [ξ2, 12, 19, 2, 11, 20, 6, 15] [4]2[5]p[2]3[7][10] [ξ2, 12, 19, 2, 9, 18, 6, 15]∗! [6][9]p[3] ([6] − [2][4]2[7]) ([2][4]2[7] + [3]2[6]) [ξ3, 9, 18, 6, 15, 24, 7, 16] [ξ3, 9, 18, 6, 13, 22, 7, 16]∗ − [5][6]p[2][7][10] p[3]3 [ξ2, 12, 19, 2, 11, 22, 7, 16]∗! [ξ2, 12, 19, 2, 9, 18, 6, 15, 24], 16 [ξ2, 12, 19, 2, 9, 18, 6, 15, 22], [ξ3, 9, 18, 6, 13, 22, 5, 14, 21], [ξ2, 12, 19, 2, 9, 18, 6, 13, 22] = − [ξ2, 12, 19, 2, 10, 19, 4, 14, 19], p[9] [ξ2, 12, 19, 2, 10, 19, 4, 11, 19] = − p[5] p[2][4] (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 20] = 0(cid:17) [ξ2, 12, 19, 2, 10, 19, 4, 11, 20] = −[ξ2, 12, 19, 2, 10, 19, 5, 11, 20] = p[7] [3]p[5] [ξ2, 12, 19, 2, 10, 19, 4, 11, 18] = −p[4][7][9]3 [3]2p[2]3 [ξ3, 9, 18, 6, 13, 22, 5, 14, 23] [ξ2, 12, 19, 2, 10, 19, 4, 14, 23], [ξ2, 12, 19, 2, 11, 20, 6, 11, 18] = 0! [ξ2, 12, 19, 2, 10, 19, 5, 13, 20] = [ξ2, 12, 19, 2, 9, 18, 6, 13, 20] = 0(cid:17) [ξ2, 12, 19, 2, 10, 19, 4, 11, 22] = p[4] p[10] = [ξ2, 12, 19, 2, 10, 19, 5, 11, 22] + = −[ξ2, 12, 19, 2, 10, 19, 4, 11, 22] − ⇒ [ξ2, 12, 19, 2, 10, 19, 4, 11, 22] = − [ξ2, 12, 19, 2, 10, 19, 5, 13, 22] [ξ2, 12, 19, 2, 10, 19, 4, 14, 22] = −p[2][4] p[5] [3]p[9] p[5][7] [3]2[5]p[9] p[2]3[4][7]5 [3]2[5]p[4][9] [12]p[2]3[7]5 [ξ2, 12, 19, 2, 9, 18, 6, 13, 22]∗! [ξ2, 12, 19, 2, 9, 18, 6, 13, 22] [ξ2, 12, 19, 2, 10, 19, 5, 14, 22] [2][6]p[4][7] [3]2p[5][10] [ξ2, 12, 19, 2, 10, 19, 4, 14, 23]∗! ([2][4]2[7] − [6]) ([2][4]2[7] + [3]2[6]) [ξ2, 12, 19, 2, 11, 22, 7, 14, 23] = (cid:16)[ξ2, 12, 19, 2, 10, 19, 4, 14, 21] = 0(cid:17) [ξ2, 12, 19, 2, 11, 22, 7, 16, 23] = −p[5] p[3] (cid:16)[ξ2, 12, 19, 2, 11, 22, 7, 16, 24] = −p[3][ξ2, 12, 19, 2, 11, 22, 7, 15, 24] [ξ2, 12, 19, 2, 9, 18, 6, 15, 24]∗! [ξ3, 9, 18, 6, 13, 22, 5, 14, 19] = − p[4] p[10] [ξ3, 9, 18, 6, 13, 22, 5, 14, 22] = − p[5] p[2][4] = p[3][4]3 [5]p[2] = −p[2]7[3][4]3 [6]p[7][9]5 [2]3p[3][4][5][10] [6]p[7][9]5 = p[2][3][5][10] ([2][7][12] + [3]7[5])[ξ2, 12, 19, 2, 10, 19, 4, 11, 19]∗! [ξ3, 9, 18, 6, 13, 22, 5, 11, 22] − [ξ3, 9, 18, 6, 13, 22, 5, 11, 19] [4][6][7]3[9]2[12] = − [ξ3, 9, 18, 6, 13, 22, 7, 14, 23] [ξ3, 9, 18, 6, 13, 22, 7, 16, 23] = −p[5] p[3] = − [2]3[4][7]2[12] [ξ2, 12, 19, 2, 10, 19, 4, 14, 23]∗ + p[5][7] p[3]3 (cid:16)[ξ3, 9, 18, 6, 13, 22, 7, 16, 24] = −p[3][ξ3, 9, 18, 6, 13, 22, 7, 15, 24] [3]4p[9]3 86 [ξ3, 9, 18, 6, 13, 22, 5, 13, 22] [3]p[9] p[2][4][7] [12]p[2][3][5][10] [4][7][9] ([2][7][12] + [3]7[5])[ξ2, 12, 19, 2, 10, 19, 4, 11, 22] − [ξ2, 12, 19, 2, 9, 18, 6, 13, 22] ([2]2[3]2[4]2[5][7][12] + [2][3]9[4]2[5]2 − [6][7]2[9][12]2)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22]∗! [ξ3, 9, 18, 6, 13, 22, 5, 14, 23]∗! 17 18 = [4]p[2][3]5[5][10] p[7][9]3 [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4], [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3] ([2][4]2[7] − [6]) ([2][4]2[7] + [3]2[6]) [ξ2, 12, 19, 2, 9, 18, 6, 15, 24]∗! [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 7], [ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 3], [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 7]∗! = ([2][4]2[7] − [6]) ([2][4]2[7] + [3]2[6]) [ξ2, 12, 19, 2, 10, 19, 4, 14, 22, 4] (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 6] = 0(cid:17) (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 5] = 0(cid:17) [ξ2, 12, 19, 2, 9, 18, 6, 15, 24, 7] = −p[5] [ξ2, 12, 19, 2, 9, 18, 6, 15, 22, 7] = p[5][9] p[3] [4]p[3] [ξ2, 12, 19, 2, 9, 18, 6, 15, 24, 8] = [ξ3, 9, 18, 6, 13, 22, 7, 16, 24, 8] = 0! p[7][9]3 [4]p[2][3]5[5][10] [ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 4] = p[10] p[4] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4]∗! [ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 5] = − p[5] p[2][4] [ξ2, 12, 19, 2, 10, 19, 4, 14, 23, 3] = −p[3][5] [ξ2, 12, 19, 2, 10, 19, 4, 14, 23, 7] = −p[3][ξ2, 12, 19, 2, 10, 19, 4, 14, 22, 7] = [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 4] = −p[5][7] [3]p[9] [3]2p[5]3[9] [2]2[12]p[7]5 [ξ2, 12, 19, 2, 10, 19, 4, 14, 19, 3] = p[2][3] p[4] [5]p[3]5[9][10] [12]p[2]3[7]5 [ξ3, 9, 18, 6, 13, 22, 5, 14, 19, 4] − p[2][4][7] [3]p[9] (cid:16)[ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 2] = 0(cid:17) [3]2[5]p[9][10] [12]p[2]3[7]5 [ξ2, 12, 19, 2, 10, 19, 4, 11, 22, 5] = [4] − [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4]∗! [ξ3, 9, 18, 6, 13, 22, 5, 14, 23, 3] = −p[3][5] = p[2]7[4]3[5] [6][9]3p[3] (cid:0)[2][7][12] + [3]7[5](cid:1) [ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 4] [2]p[5][10] [6][12]p[3][4][7]5[9]5 (cid:0)[2]2[3]2[4]2[5][7][12] + [2][3]9[4]2[5]2 − [6][7]2[9][12]2(cid:1) [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4] [2][12]p[5][10] p[3][4][7][9]3 [ξ3, 9, 18, 6, 13, 22, 5, 14, 19, 3] − p[9] [6]p[7][9]5 (cid:0)[2][7][12] + [3]7[5](cid:1) [ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 3]∗ − p[9] [3]p[2]7[4][5] [4][6][7]3[9]2[12](cid:0)[2]2[3]2[4]2[5][7][12] + [2][3]9[4]2[5]2 − [6][7]2[9][12]2(cid:1) [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 7]∗! [3]p[2][5][10] [ξ3, 9, 18, 6, 13, 22, 5, 14, 23, 7] = −p[3][ξ3, 9, 18, 6, 13, 22, 5, 14, 22, 7] [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3]∗! [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3] = − [4] [4] [4] = = [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12], [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 10], [3] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 5] = 0! [ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 3]∗! [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 7]∗! [ξ3, 9, 18, 6, 13, 22, 5, 14, 22, 4] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 7, 14] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 14] = − (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 7, 15] = 0(cid:17) p[7] (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 11] = 0(cid:17) [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 7, 16] = [12]p[2]3[7]5 [5]p[3]5[9][10] (cid:16)[ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 3, 10] = 0(cid:17) [ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 3, 12] = −p[3][ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 4, 12] [ξ2, 12, 19, 2, 10, 19, 4, 14, 23, 7, 16] = 0! [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12]∗! = [5]p[3]5[9][10] [12]p[2]3[7]5 [ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 3, 14] = −p[7] [3] [ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 4, 14] 87 [2][12]p[10] p[3][4][5][9] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12]∗! = [3][5]p[9][10] [12][7]2p[2]3 [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 14]∗! [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 12] = p[7] [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 14] = −p[3][ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 4, 14] [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 4, 12] = [5] = − [2][12]p[5][10] p[4][7][9] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 14]∗! [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 10, 17] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19] = −p[5][7] [3]p[9] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 21] = [12]p[2]3[7]5 [5]p[3]5[9][10] = [12]p[2]3[7]5 [3]2p[5]3[9][10] [7] [ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 3, 14, 21] = p[7] [3]p[5] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 21] = − [3]2[5] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 14, 19], (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 14, 21] = 0(cid:17) [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 14, 23] = − [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 10, 19] [3] p[7] [ξ2, 12, 19, 2, 10, 19, 4, 11, 19, 3, 12, 21] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 14, 21] ⇒ [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 21] = 0(cid:19) (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 14, 22] = 0(cid:17) [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 7, 14, 23] = 0! = −p[9] [4] [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 12, 19] − p[3][5] [4] [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 14, 19] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19] + [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 14, 19] [ξ2, 9, 18, 6, 13, 22, 5, 14, 21, 3, 10, 17, 1] (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 4] = 0(cid:17) [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 14, 19, 5] = 0! [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 10, 19, 2] [2][5][12]p[3][10] p[4]3[7][9] = − = − [2][12]p[10] p[3][4]3[5] [2]2[12]p[10] p[3][5] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19]∗! [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2], (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 3] = 0(cid:17) [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 5] = −p[5][7] [3]p[9] [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 10, 17, 2] = −p[5] p[3] = [2]2[12]p[10] [3] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2]∗! [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2, 9] (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2, 10] = 0(cid:17) [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 10, 17, 1, 9] = −p[4] p[2] (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2, 9, 17] = 0(cid:17) [12]p[2]3[4][10] = − [3] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2, 9]∗! (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2, 11] = 0(cid:17) [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 10, 17, 2, 9] (cid:16)[ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2, 9, 18] = 0(cid:17) 19 20 21 22 Let u be the path [1, 9, 18, 6, 13, 22, 4, 14, 19, 2, 11, 19, 2, 11, 22, 5, 14, 21, 3, 10, 17, 1], which is symmetric. This path is non-zero in A since [1, 9, 18, 6, 13, 22, 4, 14, 19, 2, 11, 19, 2, 11, 22, 5, 14, 21, 3, 10, 17, 1] 88 [1, 9, 18, 6, 13, 22, 4, 14, 19, 2, 11, 18, 2, 11, 22, 5, 14, 21, 3, 10, 17, 1] [1, 9, 18, 6, 13, 22, 4, 14, 19, 2, 11, 18, 6, 11, 22, 5, 14, 21, 3, 10, 17, 1] [1, 9, 18, 6, 13, 22, 4, 12, 19, 2, 11, 18, 6, 11, 22, 5, 14, 21, 3, 10, 17, 1]. 1 = − p[3] = − p[4] p[2][3] = −p[3][4][9] p[2][5][7] We choose uiν(i) = u1ν(1) := u. Then vjν(i)ν(aij) = v9ν(1)ν(a1,9) is given by [9, 18, 6, 13, 22, 4, 14, 19, 2, 11, 19, 2, 11, 22, 5, 14, 21, 3, 10, 17, 1, 9] [9, 18, 6, 13, 22, 4, 12, 19, 2, 11, 19, 2, 11, 22, 5, 14, 21, 3, 10, 17, 1, 9] [ξ3, 10, 19, 2, 11, 22, 5, 14, 21, 3, 10, 17, 1, 9] [ξ3, 9, 17, 2, 11, 22, 5, 14, 21, 3, 10, 17, 1, 9] = [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 10, 17, 1, 9] [ξ2, 12, 19, 2, 11, 22, 5, 14, 21, 3, 10, 17, 1, 9] ([3]2 + 1)[ξ2, 12, 19, 3, 14, 22, 5, 14, 21, 3, 10, 17, 1, 9]. [ξ2, 12, 19, 2, 11, 22, 5, 14, 21, 3, 10, 17, 1, 9] ([3]2 + 1)[ξ2, 12, 19, 3, 14, 22, 5, 14, 21, 3, 10, 17, 1, 9] [ξ2, 12, 19, 2, 10, 19, 5, 14, 21, 3, 10, 17, 1, 9] = 0, + = − [3]p[9] p[5][7] = p[3][9] p[5][7] = p[3][9] [5]p[7] [9]p[3]5[4] p[2][5][7]3 [4][6]p[2][9][10] [5]p[3] +p[2]3[3][4]3 [5]p[9] [4][6]p[2][9][10] [5]p[3] +p[2]3[3][4]3 [5]p[9] [2]2p[4]3[10] p[5]3[9] [9]p[3]5[4] p[2][5][7]3 [2][4][9][12]p[3]3[10] = − = − p[5][7]3 Now Now so we have vjν(i)ν(aij) = [ξ3, 9, 18, 6, 13, 22, 5, 14, 21, 3, 10, 17, 1, 9] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2, 9] 6= 0. [9, 17, 2, 11, 20, 5, 14, 21, 3, 10, 17, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2, 9] 89 = − = p[4][5] p[2] [3][5]p[4] p[2][7] [5][12]p[2][7] [3]p[10] = so that [9, 18, 6, 13, 20, 5, 14, 21, 3, 10, 17, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2, 9] [9, 18, 6, 13, 22, 5, 14, 21, 3, 10, 17, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2, 9] [ξ2, 12, 19, 2, 9, 18, 6, 13, 22, 4, 12, 19, 2, 9] 6= 0, vjν(i)ν(aij) [4][9][10]p[2][3]5 [7]2p[5]3 = − = ajkv′, [9, 17, 2, 11, 20, 5, 14, 21, 3, 10, 17, 2, 9, 18, 6, 136, 22, 4, 12, 19, 2, 9] where v′ is symmetric. Then C = 1. References [31] D. E. Evans and M. Pugh, The Nakayama automorphism of the almost Calabi- Yau algebras associated to SU(3) modular invariants. Preprint, arXiv:1008.1003 (math.OA). 90
1302.2604	3	1302	2016-11-08T17:15:09	Conformal nets I: coordinate-free nets	[ "math.OA", "math-ph", "math-ph" ]	We describe a coordinate-free perspective on conformal nets, as functors from intervals to von Neumann algebras. We discuss an operation of fusion of intervals and observe that a conformal net takes a fused interval to the fiber product of von Neumann algebras. Though coordinate-free nets do not a priori have vacuum sectors, we show that there is a vacuum sector canonically associated to any circle equipped with a conformal structure. This is the first in a series of papers constructing a 3-category of conformal nets, defects, sectors, and intertwiners.	math.OA	math	CONFORMAL NETS I: COORDINATE-FREE NETS ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR´E HENRIQUES Abstract. We describe a coordinate-free perspective on conformal nets, as functors from intervals to von Neumann algebras. We discuss an operation of fusion of intervals and observe that a conformal net takes a fused interval to the fiber product of von Neumann algebras. Though coordinate-free nets do not a priori have vacuum sectors, we show that there is a vacuum sector canonically associated to any circle equipped with a conformal structure. This is the first in a series of papers constructing a 3-category of conformal nets, defects, sectors, and intertwiners. Contents Implementation of diffeomorphisms Introduction 1. Conformal nets 1.a. Definition of conformal nets 1.b. Sectors of conformal nets 1.c. Fusion along intervals 1.d. Central decomposition 1.e. Conformal embeddings 2. Covariance for the vacuum sector 2.a. 2.b. Conformal circles and their vacuum sectors 3. Finite conformal nets and their sectors 3.a. The index of a conformal net 3.b. Finiteness properties of the category of sectors 3.c. The Hilbert space associated to a surface 3.d. Characterization of finite-index conformal nets 4. Comparing conformal and positive-energy nets 4.a. Circle-based nets 4.b. Positive-energy nets 4.c. The loop group conformal nets Appendix References 1 3 3 5 8 13 15 16 16 19 26 26 28 33 35 38 38 42 44 46 50 Introduction In their work on algebraic quantum field theory, Haag and Kastler studied nets of operator algebras. These are covariant functors from the category of open subsets of space-time to that of C∗-algebras or von Neumann algebras [14, 15, 22, 23]. For two-dimensional conformal field theory, one takes space-time to be two- dimensional Minkowski space M2, or a compactification thereof, and requires the net to be covariant with respect to the group of conformal diffeomorphisms of M2. Since that group contains Diff(R) × Diff(R) as a subgroup of finite index, the natural next step is to consider nets of von Neumann algebras on the real line 1 2 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES or the circle, which are covariant with respect to diffeomorphisms of the line or the circle. The latter correspond to chiral conformal field theories, and are called conformal nets. They have been studied intensively-see for example the papers [7, 9, 10, 19, 40]. Classification results and many more references can be found in Kawahigashi–Longo [26]. Our interest in conformal nets was prompted by the following question of Stephan Stolz and Peter Teichner, which arose in connection with their ongoing program to construct elliptic cohomology using local quantum field theories [34, 35]. Recall that von Neumann algebras form the objects of a 2-category, where the morphisms are bimodules, and the 2-morphisms are maps of bimodules. Given an n-category C equipped with a unit object 1 ∈ C, the (n− 1)-category L := HomC(1, 1) is called the loops on C. The n-category C is then said to deloop L. Question. (Stolz-Teichner, 2004) Does there exist an interesting 3-category that deloops the 2-category of von Neumann algebras? Here, by an "interesting" 3-category, they meant a 3-category other than the obvi- ous one-object 3-category defined by Ob(C) = {1} and HomC(1, 1) = {von Neumann algebras}. Actually, given that von Neumann algebras form a symmetric monoidal 2-category, one should ask for a symmetric monoidal 3-category that deloops von Neumann algebras. Some axiomatizations of the notion symmetric monoidal 3- category were presented in [17]. One of them is the notion of an internal bicategory in the 2-category of symmetric monoidal categories. The present paper is the first of a series [3], the goal of which is to provide a posi- tive answer to the above question of Stolz and Teichner. Namely, we will show that conformal nets form a symmetric monoidal 3-category (an internal bicategory in the 2-category of symmetric monoidal categories) that deloops the symmetric monoidal 2-category of von Neumann algebras. This first paper of our series contains our definition of conformal nets. We also treat the notion of sectors of conformal nets, otherwise known as representations. Our definition of conformal nets is different from the standard definition in two important ways. The first difference is that we do not include a positive-energy assumption. This is important for the following reason. We want certain objects in our 3-category to be dualizable. The natural candidate for the dual of a net A is its complex conjugate ¯A, which is almost never positive-energy. We expect our conformal nets include the usual ones (Virasoro, loop groups, free boson, orbifolds, and cosets, among others), but we also expect them to include new examples which do not have positive-energy, such as the spatial slices of nets on M2 studied in Kawahigashi–Longo–Muger [27]. The second difference is that we expand the category of intervals on which a net is defined, and expand the category of von Neumann algebras in which they take values. Traditionally, conformal nets assign to a subinterval of the circle S1 a von Neumann subalgebra of B(H0) for a fixed "vacuum" Hilbert space H0. In our definition, a "coordinate-free" conformal net assigns abstract von Neumann algebras to abstract intervals. For local quantum field theory such a coordinate- free point of view has been introduced in Brunetti–Fredenhagen–Verch [6]. The coordinate-free definition has the advantage that there is a priori no Hilbert space as part of the structure of a net. (In the hierarchy of our 3-category, Hilbert spaces appear later as 2-morphisms; the vacuum Hilbert space of a net A will be the identity 2-morphism of the identity 1-morphism of the net. In particular, we prove that the vacuum Hilbert space with its action of the diffeomorphism group of the circle can be reconstructed from the von Neumann algebras associated to intervals; our notion of conformal nets is therefore closely related to existing notions in the literature [27].) Moreover, we can glue abstract intervals to one another, but not CONFORMAL NETS I: COORDINATE-FREE NETS 3 arbitrary subintervals of the circle: the gluing of two subintervals is typically no longer a subinterval. This will be of crucial importance for the composition of 1-morphisms, in the third paper of our series. Outline. In the first section, we introduce our definition of conformal nets and study some of their basic properties. We discuss the category Sect(A) of sectors of a conformal net A, and define the monoidal structure on it. We also introduce a coordinate-free version of the category of sectors, SectS(A), that depends on the choice of a circle S. In Section 2, we study the vacuum sector H0(S,A) of a conformal net A, which is the unit object in SectS(A). In Theorem 2.13, we show that if the circle S is equipped with a conformal structure, then the vacuum sector is well defined up to unique unitary isomorphism. The vacuum sector is covariant for conformal maps of circles, and projectively covariant for diffeomorphisms. Section 3 concerns finiteness properties of the category of sectors. We start by discussing the notion of finite index for conformal nets in terms of a certain minimal index of the vacuum sector. We describe the characterization of finite index nets in terms of the category of sectors: the conformal net A has finite index if and only if the category of sectors Sect(A) is fusion. This provides alternative proofs of results of Kawahigashi–Longo–Muger [27] and Longo-Xu [31]. Finally, in Section 4, we relate our definition of conformal nets to other definitions in the literature. We start by presenting a circle-based definition that is in principle equivalent to our usual coordinate-free definition. We then discuss the more classical definition of conformal nets, which includes the positive-energy condition and is not equivalent to our notion. We call these classical nets "positive-energy nets", and check that positive-energy nets yield examples of conformal nets in our sense. We review the construction of positive-energy nets from loop groups, and collect the necessary results from the literature to show that they yield coordinate-free nets. The appendix contains a brief summary of definitions and results about von Neu- mann algebras, Connes fusion, dualizability, statistical dimension, and Haagerup's u-topology. With the exception of the last subsection, these topics are discussed in more detail in our paper [4]. Acknowledgements. We would like to thank Stephan Stolz and Peter Teich- ner for their continual support during this project and for having formulated the question that led to it, and to thank Michael Muger for his invaluable guidance regarding conformal nets. We thank Sebastiano Carpi for pointing out a missing argument and the reference [31], and Marcel Bischoff for pointers to the literature. The last author also would like to thank Michael Hopkins for his suggestion to read Wassermann's articles. The first author was supported by the Sonderforschungs- bereich 878, and the second author was partially supported by a Miller Research Fellowship. 1. Conformal nets 1.a. Definition of conformal nets. All 1-manifolds are compact, smooth, and oriented. The standard circle S1 := {z ∈ C : z = 1} is the set of complex numbers of modulus one, equipped with the counter-clockwise orientation. By a circle, we shall mean a smooth manifold S that is diffeomorphic to the standard circle S1. Similarly, by an interval, we shall mean a smooth 1-manifold that is diffeomorphic to the standard interval [0, 1]. Equivalently, an interval is a compact non-empty 1-manifold with boundary that is connected and simply connected. For a 1-manifold I, we denote by ¯I the same manifold equipped with the opposite orien- tation, by Diff(I) the group of diffeomorphisms of I, and by Diff +(I) the subgroup 4 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES of orientation-preserving diffeomorphisms. Let INT be the category whose objects are intervals and whose morphisms are embeddings, not necessarily orientation- preserving. We also let VN be the category whose objects are von Neumann alge- bras with separable preduals, and whose morphisms are C-linear homomorphisms, and C-linear anti-homomorphisms1. The following notion of conformal nets differs from the standard definition in the literature: our nets are "coordinate-free", in the sense that they take values on all abstract intervals, not only on subintervals of the standard circle; our nets also need not satisfy the usual positive-energy condition. A net is a covariant functor A : INT → VN taking orientation-preserving embed- dings to injective homomorphisms and orientation-reversing embeddings to injective antihomomorphisms. A net is said to be continuous if for any intervals I and J, the natural map HomINT(I, J) → HomVN(A(I),A(J)) is continuous for the C∞ topol- ogy on HomINT(I, J) and Haagerup's u-topology on HomVN(A(I),A(J)), reviewed in the appendix. Given a subinterval I ⊆ K, we will often not distinguish between A(I) and its image in A(K). Definition 1.1. A conformal net is a continuous net A subject to the following conditions. Here, I and J are subintervals of an interval K: (i) Locality: If I, J ⊂ K have disjoint interiors, then A(I) and A(J) are commuting subalgebras of A(K). (ii) Strong additivity: If K = I ∪ J, then A(K) is generated as a von Neumann algbera by its two subalgebras: A(K) = A(I) ∨ A(J). (iii) Split property: If I, J ⊂ K are disjoint, then the map from the algebraic tensor product A(I)⊗algA(J) → A(K) extends to a map from the spatial tensor product A(I) ¯⊗ A(J) → A(K). (iv) Inner covariance: If ϕ ∈ Diff +(I) restricts to the identity in a neighbor- hood of ∂I, then A(ϕ) is an inner automorphism of A(I). (A unitary u ∈ A(I) with Ad(u) = A(ϕ) is said to implement ϕ.) (v) Vacuum sector: Suppose that J ( I contains the boundary point p ∈ ∂I, and let ¯J denote J with the reversed orientation; A(J) acts on L2(A(I)) via the left action of A(I), and A( ¯J) ∼= A(J)op acts on L2(A(I)) via the right action of A(I). In that case, we require that the action of A(J)⊗algA( ¯J) on L2(A(I)) extends to an action of A(J ∪p ¯J): (1.2) A(J)⊗algA( ¯J ) B(L2A(I)) A(J ∪p ¯J) Here, J ∪p ¯J is equipped with any smooth structure extending the given smooth structures on J and ¯J, and for which the orientation-reversing involution that exchanges J and ¯J is smooth. Note that A( ¯J) is canonically isomorphic to A(J)op via the antihomomorphism A(IdJ ) : A(J) → A( ¯J ). That fact was used above in the vacuum axiom in order to identify A( ¯J ) with A(J)op. Also, the proper way of visualizing the interval J ∪p ¯J 1An anti-homomorphism is a unital map satisfying f (ab) = f (b)f (a). CONFORMAL NETS I: COORDINATE-FREE NETS 5 is as submanifold of the circle S := I ∪∂I ¯I: S : I ¯I J p ¯J Here, the circle S is equipped with a smooth structure such that the three embed- dings I ֒→ S, ¯I ֒→ S, J ∪p ¯J ֒→ S are smooth, and the involution S → S that exchanges I with ¯I is smooth. Example 1.3. The trivial conformal net C is given by C(I) = C for any interval I, and C(ι) = IdC for any embedding of intervals ι. Some more substantial examples of conformal nets are discussed in Section 4.c. Given conformal nets A and B, their direct sum is given by (A ⊕ B)(I) := A(I) ⊕ B(I), and their tensor product is (A ⊗ B)(I) := A(I) ¯⊗ B(I). Most of the axioms for A⊕B and A⊗B are straightforward. We just check the vacuum axiom for A⊗B: the Hilbert spaces L2(A(I) ¯⊗ B(I)) and L2(A(I))⊗ L2(B(I)) are isomorphic as A(I) ¯⊗ B(I)-bimodules, so the action of A(J) ⊗alg B(J) ⊗alg A( ¯J ) ⊗alg B( ¯J) extends to A(J ∪p ¯J) ¯⊗ B(J ∪p ¯J). This provides the desired extension of the action of (A ⊗ B)(J) ⊗alg (A ⊗ B)( ¯J) on L2((A ⊗ B)(I)). interior of I. ThenWn A(In) = A(I). Then every element a ∈ A(I) can be written as limn A(ϕn)(a) ∈Wn A(In). Lemma 1.4. Let In ⊂ I be an increasing sequence of intervals whose union is the Proof. Let ϕn : I → I, ϕn(I) = In be a sequence of embeddings that tends to idI . We record the following easy result for future use: (cid:3) Remark 1.5. The continuity condition in the definition of conformal nets is equiv- alent to the following slightly weaker condition. It is enough that for any interval I, the natural map Diff +(I) → Aut(A(I)) be continuous for the C∞ topology on Diff +(I) and the u-topology on Aut(A(I)), see Lemma 4.4. Remark 1.6. It is possible that the condition that the algebras A(I) have separa- ble preduals follows from Definition 1.1, more specifically, from the split property axiom-compare with [16, Proposition 1.6]. Definition 1.7. A conformal net A is called irreducible if every algebra A(I) is a factor. A direct sum of finitely many irreducible conformal nets is called semisimple. 1.b. Sectors of conformal nets. In the 3-category that our series of papers [3] constructs, A-sectors correspond to 2-morphisms A 1A ⇓ 1A . A Here, we shall discuss A-sectors without any reference to the 3-category. Definition 1.8. Let S be a circle and let A be a conformal net. An S-sector of A (also called an A-sector on S) consists of a Hilbert space H and a collection of homomorphisms ρI : A(I) → B(H), I ⊂ S subject to the compatibility condition ρIA(J) = ρJ whenever J ⊂ I. The category of S-sectors of A is denoted SectS(A). 6 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES For the standard circle S1 := {z ∈ C : z = 1}, an S1-sector of A is simply called a sector of A, or an A-sector. The category of A-sectors is denoted Sect(A). Given an interval I, let Diff 0(I) denote the group of diffeomorphisms that restrict to the identity near the boundary of I. Lemma 1.9. Let S be a circle, and let Ii ⊂ S be intervals whose interiors cover S. Let ρi : A(Ii) → B(H) be actions subject to the following two conditions: (i) ρiA(Ii∩Ij ) = ρjA(Ii∩Ij ), (ii) if J ⊂ Ii and K ⊂ Ij are disjoint, then ρi(A(J)) commutes with ρj(A(K)). These actions extend uniquely to the structure of an S-sector on H. Proof. For every interval J ⊂ S, we will construct an action ρJ : A(J) → B(H), uniquely determined by the requirement that ρJA(J∩Iℓ) = ρℓA(J∩Iℓ) for every Iℓ in our cover. Fix an element I0 of our cover, with corresponding action ρ0 of A(I0) on H. Pick a diffeomorphism ϕ = ϕn ◦ . . . ◦ ϕ1 with ϕs ∈ Diff 0(Iis ), for Iis in our cover, such that ϕ(J) ⊂ I0. Let us ∈ A(Iis ) be unitaries implementing ϕs. Identifying the elements us with their images in B(H), we set ρJ (a) := u∗ 1 . . . u∗ nρ0(cid:0)A(ϕ)(a)(cid:1)un . . . u1. For every sufficiently small interval K ⊂ J ∩ Iℓ, we will show that (1.10) ρJA(K) = ρℓA(K). Here, 'sufficiently small' means that the intervals Ks := ϕs(. . . (ϕ1(K))) should be contained in some Iks in our cover, and that for every s, s′ ≤ n, either Ks ⊂ Iis′ or Ks ∩ supp(ϕs′ ) = ∅. For every s ≤ n, we claim that (1.11) u∗ 1 . . . u∗ ∀a ∈ A(K). Note that (1.10) is the special case s = n of this equation. We prove (1.11) by induction on s. The base case (s = 0, k0 = ℓ) is trivial. The induction step reduces to the equation sρks(cid:0)A(ϕs ◦ . . . ◦ ϕ1)(a)(cid:1)us . . . u1 = ρℓ(a) ρks(cid:0)A(ϕs)(b)(cid:1) = usρks−1(b)u∗ s, with b = A(ϕs−1 ◦ . . . ◦ ϕ1)(a). Recall that b ∈ A(Ks−1), us ∈ A(Iis ), and that, by assumption, either Ks−1 ⊂ Iis or Ks−1 ∩ supp(ϕs) = ∅. In the first case, we have usρks−1 (b)u∗ s = usρis (b)u∗ s = ρis(usbu∗ s) = ρis(cid:0)A(ϕs)(b)(cid:1) = ρks(cid:0)A(ϕs)(b)(cid:1). In the second case, by condition (ii), the elements ρks−1(b) and us commute in B(H). It follows that This finishes the proof of (1.11) and hence of (1.10). usρks−1(b)u∗ s = ρks−1(b) = ρks (b) = ρks(cid:0)A(ϕs)(b)(cid:1). Finally, by the strong additivity axiom, it follows from (1.10) that ρJ (a) = ρℓ(a) (cid:3) for every a ∈ A(J ∩ Iℓ). Let S be a circle, let j ∈ Diff −(S) be an orientation-reversing involution fixing the boundary ∂I of some interval I ⊂ S, and let I ′ := j(I). The Hilbert space H := L2(A(I)) is equipped with: • for each J ⊂ I, an action ρJ : A(J) ֒→ A(I) −−−−−−−−−−−−−−−−−→ B(H) left action of A(I) on H of the algebra A(J). CONFORMAL NETS I: COORDINATE-FREE NETS 7 • for each J ⊂ I ′, an action ρJ : A(J) ֒→ A(I ′) A(j) −−−→ A(I)op right action of A(I) on H −−−−−−−−−−−−−−−−−→ B(H) of A(J). morphism • If J ⊂ S satisfies j(J) = J, then by the vacuum sector axiom, the homo- ρJ∩I ⊗ ρJ∩I ′ : A(J ∩ I) ⊗alg A(J ∩ I ′) → B(H) extends to an action ρJ of A(J) on H. Applying Lemma 1.9, we see that L2(A(I)) comes naturally equipped with the structure of an S-sector of A. We call it the vacuum sector of A associated to S, I, and j. We record the following subintervals of the standard circle for future usage: S1 ⊤ := {z ∈ S1 ℑm(z) ≥ 0}, S1 ⊥ := {z ∈ S1 ℑm(z) ≤ 0}, S1 ⊣ := {z ∈ S1 ℜe(z) ≥ 0}, S1 ⊢ := {z ∈ S1 ℜe(z) ≤ 0}. Definition 1.12. We let H0(A) denote the vacuum sector of A associated to the standard circle S1, its upper half S1 ⊤, and the involution z 7→ ¯z. It is defined by and has left actions of A(I) for every I ⊂ S1. H0(A) := L2(A(S1 ⊤)), Given two circles S1, S2 and a diffeomorphism ϕ : S1 → S2, there is a corre- sponding functor (1.13) SectS2 (A) −→ SectS1 (A) H 7→ ϕ∗H. For an orientation-preserving diffeomorphism ϕ, that functor sends (H,{ρJ}J⊂S2) to the S1-sector with underlying Hilbert space H, and actions ρϕ(J) ◦ A(ϕJ ). If the diffeomorphism ϕ is orientation-reversing, then ϕ∗H is the complex conjugate Hilbert space H, equipped with the actions A(J) A(ϕJ ) −−−−→ A(ϕ(J))op ρϕ(J)−−−−→ B(H)op ∗−−→ B(H) = B(H) for J ⊂ S1. Proposition 1.14. Let S1 and S2 be two circles, and let ϕ, ψ ∈ Diff +(S1, S2) be diffeomorphisms. Then the functors ϕ∗, ψ∗ : SectS2(A) → SectS1(A) are non-canonically unitarily naturally equivalent (in other words, there exists a unitary natural equivalence ϕ∗ ∼= ψ∗, but there is no canonical way of choosing such a natural equivalence). Proof. Write ψ ◦ ϕ−1 = ϕ1 ◦ . . .◦ ϕn as a product of diffeomorphisms ϕi ∈ Diff 0(Ii) with support in intervals Ii ⊂ S2, and let ui ∈ A(Ii) be unitaries implementing them (one may arrange this with n = 2). For any sector (H,{ρJ}J⊂S2), conju- gation by ρI1(u1) . . . ρIn (un) provides a unitary isomorphism of S2-sectors from H to (ϕ−1)∗ψ∗H, which can be interpreted as a unitary isomorphism of S1-sectors from ϕ∗H to ψ∗H. The collection of all those isomorphisms is a unitary natural transformation from ϕ∗ to ψ∗. (cid:3) Corollary 1.15. Let S be a circle, and ϕ ∈ Diff +(S) a diffeomorphism. Then for any S-sector H we have ϕ∗H ∼= H. Proof. Take ψ = IdS in the above proposition. (cid:3) 8 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES Corollary 1.16. Let S be a circle, I1, I2 ⊂ S intervals, and let j1, j2 ∈ Diff −(S) be involutions that fix ∂I1 and ∂I2, respectively. Let H1 be the vacuum sector of A associated to S, I1, j1, and let H2 be the vacuum sector of A associated to S, I2, j2. Then H1 and H2 are isomorphic as S-sectors. Proof. Let ϕ ∈ Diff +(S) be a diffeomorphism that sends I1 to I2 and that inter- twines the actions of j1 and j2. By the definition of vacuum sector, the diffeomor- phism ϕ induces an isomorphism ϕ∗H2 ∼= H1 and by Corollary 1.15, we also have ϕ∗H2 ∼= H2. (cid:3) Given a sector H ∈ Sect(A) on the standard circle, and given another circle S, we denote by H(S) ∈ SectS(A) the S-sector ϕ∗H, where ϕ ∈ Diff +(S, S1) is some diffeomorphism. Note that by Proposition 1.14, the S-sector H(S) is well defined up to (non-canonical) unitary isomorphism. In the case of the vacuum sector, the above construction specializes to: Definition 1.17. Given a circle S, we let H0(S,A) ∈ SectS(A) stand for any one of the Hilbert spaces H0(A)(S) considered in Corollary 1.16 or equivalently any one of the Hilbert spaces considered in the paragraph following Corollary 1.16, and call it the vacuum sector of A associated to S. We sometimes abbreviate H0(S,A) by H0(S). That Hilbert space is well defined up to non- canonical unitary isomorphism of S-sectors. We will see later, in Section 2, that H0(S,A) can be determined canonically if we fix a conformal structure on S. For S a circle and I ⊂ S an interval, let us denote by I ′ ⊂ S the closure of its complement in S. Proposition 1.18 (Haag duality). Let A be a conformal net, and S be a circle. Then for any I ⊂ S, the algebra A(I ′) is the commutant of A(I) on H0(S,A). Given intervals J ⊂ K such that J c, the closure of K \ J, is itself an interval, the commutant of A(J) in A(K) is A(J c). Proof. Let j ∈ Diff −(S) be an involution that fixes ∂I, and let H = L2(A(I)) be the vacuum sector associated to S, I, and j. The equation A(I ′) = A(I)′ is obvious on H, and follows for any vacuum H0(S,A) since the two are isomorphic. For the second statement, pick an embedding K → S. Using strong additivity, we then have A(J)′ ∩ A(K) = A(J)′ ∩ A(K ′)′ = (A(J) ∨ A(K ′))′ = A(J ∪ K ′)′ = A(J ′ ∩ K) = A(J c). 1.c. Fusion along intervals. In this section, we introduce the operation ⊛ of fusion of von Neumann algebras, and show that it corresponds to the geometric operation of gluing two intervals and then discarding the part along which they were glued: (cid:3) (1.19) ⊛ Later in this section, we will use this operation to define the operation of fusion of sectors. This operation will also be crucial in the third paper of our series [3], in order to define the composition of two defects. Fusion of von Neumann algebras. Definition 1.20. Let A ← C op, C → B be two homomorphisms between von Neumann algerbas, and let AH and BK be faithful modules. Viewing H as a right CONFORMAL NETS I: COORDINATE-FREE NETS 9 C-module, we may form the Connes fusion H ⊠C K (recalled in the Appendix). One then defines A ⊛C B := (A ∩ C op ′ ) ∨ (C′ ∩ B) ⊂ B(H ⊠C K), (1.21) where the commutants of C op and C are taken in H and K, respectively. This algebra is independent of the choices of H and K (see Proposition 1.23). Warning 1.22. The operation ⊛ has rather poor formal properties. For example, given homomorphisms A1 ← Bop 2 , B2 → A3, such that the images of B1 and Bop 2 commute in A2, one might ask for an associator isomorphism (A1 ⊛B1 A2) ⊛B2 A3 ∼= A1 ⊛B1 (A2 ⊛B2 A3). Such an isomorphism does not exist in general: there are algebras for which one of the following two inclusions 1 , B1 → A2, A2 ← Bop (A1 ∩ Bop 1 ′ ) ∨ (B′ 1 ∩ A1 ∩ Bop 2 ′ ) ∨ (B′ 2 ∩ A3) ֒ →֒→ (A1 ⊛B1 A2) ⊛B2 A3 A1 ⊛B1 (A2 ⊛B2 A3) is an isomorphism, but the other isn't. We present a counterexample (but without further justifications as this would take us too far afield): take a conformal net A with µ(A) > 1 (Definition 3.1), and let A1 = A([0, 2])op, B1 = A([1, 2]), A2 = A([−2, 2]) ∩ A([−1, 0])′, B2 = A3 = A([0, 1])op. Proposition 1.23. The algebra A ⊛C B is independent, up to canonical isomor- phism, of the choice of faithful modules AH and BK. B(H1 ⊠C K1) (A ∩ C op ′ ) ⊗alg (C′ ∩ B) Proof. Let H1 and H2 be faithful A-modules, and let K1 and K2 be faithful B- modules. Upon choosing isomorphisms H1 ⊗ ℓ2 ∼= H2 ⊗ ℓ2 and K1 ⊗ ℓ2 ∼= K2 ⊗ ℓ2, we get the following commutative diagram of algebra homomorphisms: B(cid:0)(H1 ⊗ ℓ2) ⊠C (K1 ⊗ ℓ2)(cid:1) B(cid:0)(H2 ⊗ ℓ2) ⊠C (K2 ⊗ ℓ2)(cid:1). The completions of (A ∩ C op ′ ) ⊗alg (C′ ∩ B) in B(H1 ⊠C K1) and B(H2 ⊠C K2) therefore agree since they might as well be taken in B(cid:0)(H1 ⊗ ℓ2) ⊠C (K1 ⊗ ℓ2)(cid:1) and B(cid:0)(H2 ⊗ ℓ2) ⊠C (K2 ⊗ ℓ2)(cid:1), respectively. If the modules H and K are not faithful, then there is still an action, albeit (cid:3) B(H2 ⊠C K2) ∼= non-faithful, of A ⊛C B on H ⊠C K: Lemma 1.24. Let A ← C op, C → B be homomorphisms, and let AH and BK be any modules. Then the natural map (A ∩ C op ′ ) ⊗alg (C′ ∩ B) → B(H ⊠C K) extends to an action of A ⊛C B on H ⊠C K. (cid:3) 2 , C2 → B2, there is a canonical isomorphism The operation ⊛ is compatible with spatial tensor product in the sense that 1 , C1 → B1, given algebras A1, B1, C1, A2, B2, C2 and homomorphisms A1 ← C op A2 ← C op (1.25) Remark 1.26. In [36], a similar operation A ∗C B is defined, under the name fiber product of von Neumann algebras. It is given by (A1 ¯⊗ A2) ⊛C1 ¯⊗ C2 (B1 ¯⊗ B2) ∼= (A1 ⊛C1 B1) ¯⊗ (A2 ⊛C2 B2). A ∗C B := (A′ ⊗alg B′)′, where the commutants A′ and B′ are taken in B(H) and B(K) respectively, while the last one is taken in B(H ⊠C K). Unlike ⊛, the operation ∗ is associative. There is always an inclusion A ⊛C B ֒→ A∗C B and under favorable circumstances, it can happen that those two algebras agree. This will always be true in the cases that we 10 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES consider (see the section on associativity of composition in the third paper of our series [3] for a precise statement). We work with A ⊛C B as opposed to A ∗C B for technical convenience: it is easier to check that the former commutes with other von Neumann algebras. The algebra of a fused interval. We now make precise the heuristic of picture (1.19). Consider three intervals I, Il, and Ir equipped with two maps il : I → Il and ir : I → Ir. The maps il and ir are orientation-reversing and orientation-preserving, respectively. Moreover, we require that the closure Jl of Il \ il(I) and the closure Jr of Ir \ ir(I) be (non-empty) intervals, and that Il ∪I Ir be a "Y-graph"-see (1.27). We can then define the fused interval Il ⊛I Ir := Jl ∪ Jr ⊂ Il ∪I Ir: (1.27) Il Ir I Il ∪I Ir : Il ⊛I Ir : Note that the fused interval Il ⊛I Ir only inherits a canonical C1 structure. Indeed, we have a natural identification of tangent spaces TpJl −Tpil ∼= TpI Tpir ∼= TpJr p := Jl ∩ Jr but no way to compatibly identify the higher germs of Jl and Jr at p. Pick involu- tions α ∈ Diff −(Il) and β ∈ Diff −(Ir) that fix p, and such that the map αJl ∪ βI : Il = Jl ∪ I → I ∪ Jr = Ir is smooth. We equip Il ⊛I Ir with the smooth structure pulled back via αJl ∪ IdJr : Il ⊛I Ir → Ir or, equivalently, the one pulled back via IdJl ∪ βJr : Il ⊛I Ir → Il. The smooth structure on Il ⊛I Ir depends on the involutions α and β. The distinguishing feature of smooth structures arising this way is that there exists an action of the symmetric group S3 on Il ∪I Ir such that all the induced maps between Il, Ir and Il ⊛I Ir are smooth. The next proposition shows that the algebra A(Il ⊛I Ir) associated to the fused interval does not depend on the above choices, up to canonical isomorphism: Proposition 1.28. Let A be a conformal net, and let I, Il, Ir, Jl, Jr be as above. Then there is a canonical isomorphism A(Il ⊛I Ir) ∼= A(Il) ⊛A(I) A(Ir), compatible with the inclusions of A(Jl) and A(Jr). Proof. Let S := I ∪∂I ¯I be the circle obtained by gluing two copies of I along their common boundary, and let j0 : S → S be the orientation-reversing involution that exchanges I and ¯I. Equip S with any smooth structure compatible with those on I and ¯I, and for which the map j0 is smooth. Recall the involutions α ∈ Diff −(Il) and β ∈ Diff −(Ir) described above. Identify I with its image in Il and Ir under the inclusions il and ir respectively, and let p := Jl ∩ Jr be the trivalent vertex of the Y-graph Il ∪I Ir. Pick orientation- preserving embeddings fl : Jl → I, gr : Jr → ¯I that send p to itself, and satisfy the following two conditions: (i ) the maps f : Il → S and g : Ir → S given by f : Il = Jl ∪ ¯I g : Il = I ∪ Jr fl∪Id ¯I −−−−−→ I ∪ ¯I = S −−−−−→ I ∪ ¯I = S IdI ∪gr are injective and smooth, and (ii ) the equations f ◦ α = j0 ◦ f and g ◦ β = j0 ◦ g are satisfied in a neighborhood of p. Note that the map fl ∪ gr : Il ⊛I Ir → S is then also smooth. CONFORMAL NETS I: COORDINATE-FREE NETS 11 Let H0 := L2(A(I)) be the vacuum sector associated to S, I, and j0. We have two faithful actions A(Il) → B(H0) and A(Ir) → B(H0) associated to f and g. We can therefore compute A(Il)⊛A(I)A(Ir) inside B(cid:0)H0⊠A(I) H0(cid:1) = B(cid:0)H0(cid:1). By Haag duality, the relative commutant of A(I)op in A(Il) is A(Jl), and the relative commutant of A(I) in A(Ir) is A(Jr). We therefore have A(Il) ⊛A(I) A(Ir) = A(Jl) ∨ A(Jr) ⊂ B(cid:0)H0(cid:1), which is equal to A(Jl ∪ Jr) = A(Il ⊛I Ir) by the strong additivity axiom. Corollary 1.29. Let A be a conformal net, let Hl be an A(Il)-module, and let Hr be an A(Ir)-module. Then the two actions A(Jl) and A(Jr) extend to an action of A(Il ⊛I Ir) on Hl ⊠A(I) Hr. Fusion of sectors. Consider now a theta-graph Θ with trivalent vertices p and q, and let S1, S2, S3 ⊂ Θ be its three circle subgraphs, with orientations as drawn below: (cid:3) (1.30) Θ : p q , , . S1 S2 S3 Equip S1, S2, S3 with smooth structures for which there exists an action of the symmetric group S3 on Θ that fixes p and q, permutes the three circles, and such that πSa is smooth for every π ∈ S3 and a ∈ {1, 2, 3}. Let I := S1 ∩ S2, K := S1 ∩ S3, L := S2 ∩ S3. We equip K with the orientation inherited from S1, and give I and L the ones coming from S2. Definition 1.31. Given S1, S2, S3, I, K, L as above, the operation (1.32) ⊠I : SectS1(A) × SectS2 (A) → SectS3(A) of fusion of sectors is given by (H1, H2) 7→ H1 ⊠A(I) H2. That space inherits an A(K) action from H1, and an A(L) action from H2 (both are left actions). If J ⊂ S3 is an interval not contained in and not containing K or L, then J = ((J ∩ S1)∪ ¯I) ⊛I ((J ∩ S2)∪ I), and so A(J) acts on H1 ⊠A(I) H2 by Corollary 1.29. The Hilbert space H1 ⊠A(I) H2 is then an S3-sector by Lemma 1.9. Associativity of fusion. The operation 1.32 satisfies a certain version of associativ- ity. Given a graph that looks as follows as indicated below , with circles subgraphs S1, . . . , S6 S1 S2 S3 S4 S5 S6 then there is an associator coming from the associator of Connes fusion that makes the following diagram commute: SectS1(A) × SectS2(A) × SectS3 (A) SectS1(A) × SectS5(A) SectS4(A) × SectS3(A) SectS6(A). 12 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES Note. There should be a similar canonical associator when the circles S1, . . . , S6 are arranged as follows: S1 S2 S3 S4 S5 S6 but we only know how to construct it non-canonically. We will not discuss this construction. Unitality of fusion. The unitality of fusion of sectors can be formulated as follows. Recall that given a sector H ∈ Sect(A) on the standard circle and given another circle S, we denote by H(S) ∈ SectS(A) the corresponding sector on S (well defined up to non-canonical isomorphism). Lemma 1.33. Let S1, S2, S3 and I = S1 ∩ S2 be as in (1.30). Then for any sector H ∈ Sect(A), there exists (non-canonical) unitary isomorphisms of S3-sectors H(S1) ⊠A(I) H0(S2) ∼= H(S3) H0(S1) ⊠A(I) H(S2) ∼= H(S3). and Proof. We only show the first equality. Let K and L be as above, let j2 be the involution of S2 coming from the action of S3 on S1 ∪ S2, and let ϕ := IdK ∪ j2L : S3 → S1. By definition, we can take H0(S2) = L2(A(I)), with S2-sector structure induced by j2. We then have H(S1) ⊠A(I) H0(S2) = H(S1) ⊠A(I) L2(A(I)) ∼= ϕ∗H(S1) ∼= H(S3). For J ⊂ S3 an interval, the above isomorphism is both A(J ∩ K) and A(J ∩ L)- equivariant, and hence A(J)-equivariant by strong additivity. Corollary 1.34. Letting S1, S2, S3, and I be as in (1.30), we have (cid:3) (1.35) H0(S1) ⊠A(I) H0(S2) ∼= H0(S3). Monoidal fusion products. There is another (closely related) notion fusion of sec- tors, for which the source and target categories are the same category. Let S be a circle. Given an interval I ⊂ S, and an involution j ∈ Diff −(S) fixing ∂I, we can turn any S-sector into an A(I)-A(I)-bimodule by equipping it with the right action induced by A(j) : A(I)op → A(I ′). The fusion of A(I)-A(I)-bimodules then equips the category of S-sectors of A with a monoidal structure:  (1.36) product : unit object : SectS(A) × SectS(A) ⊠I−−→ SectS(A) H ⊠I K := H ⊠A(I) K L2(A(I)) ∈ SectS(A) (We usually drop j from the notation, but occasionally write ⊠I,j when we need to be precise.) The unit object is the vacuum sector of A associated to S, I, and j. We now explain why H ⊠I K is an S-sector. The algebras A(I) and A(I ′) act on H ⊠I K by their respective actions on H and K. If J ⊂ S is an interval that crosses ∂I once, then we have A(J) ∼= A(J ∪ I ′) ⊛A(I) A(J ∪ I) (1.37) by Proposition 1.28. Here, the two maps A(J ∪ I ′) ← A(I)op , A(I) → A(J ∪ I) used in the definition of the right-hand side of (1.37) are induced by j : I → J ∪ I ′, and by the inclusion I ֒→ J ∪ I, respectively. The algebra (1.37) then acts on CONFORMAL NETS I: COORDINATE-FREE NETS 13 H ⊠I K by Corollary 1.29. This construction works for any J ⊂ S that crosses ∂I once, and so H ⊠I K is an S-sector by Lemma 1.9. Proposition 1.38. Let S be a circle, I1, I2 ⊂ S intervals, and j1, j2 ∈ Diff −(S) involutions that fix ∂I1 and ∂I2, respectively. Let ⊠1 and ⊠2 be the monoidal structures on SectS(A) associated to (I1, j1) and (I2, j2), respectively, as in (1.36). Then these two monoidal structures are equivalent (but non-canonically). Proof. The first monoidal structure is given by the data of a monoidal product (X, Y 7→ X ⊠1Y ), a unit object 11, an associator a1 : (X ⊠1Y )⊠1Z ∼→ X ⊠1(Y ⊠1Z), ∼→ X; the a left unit map ℓ1 : 11 ⊠1 X ∼→ X, and a right unit map r1 : X ⊠1 11 second monoidal structure is similarly given by ⊠2, 12, a2, ℓ2, and r2. An equivalence between these two monoidal structures consists of a unitary natural ∼→ 12 transformation σ : X ⊠1 Y ∼→ X ⊠2 Y , and a unitary isomorphism µ : 11 making the following three diagrams commute: (X ⊠1 Y ) ⊠1 Z a1 X ⊠1 (Y ⊠1 Z) σ◦(σ⊠1) σ◦(1⊠σ) (X ⊠2 Y ) ⊠2 Z a2 X ⊠2 (Y ⊠2 Z) 11 ⊠1 X σ◦(µ⊠1) 12 ⊠2 X ℓ1 ℓ2 X X ⊠1 11 r1 σ◦(1⊠µ) X X ⊠2 12 r2 The claim is that it is possible to find such a pair (σ, µ), but that there is no canonical choice. Pick a diffeomorphism ϕ ∈ Diff +(S) that maps I1 to I2, and that intertwines the involutions j1 with j2. By Proposition 1.14, the functor ϕ∗ is naturally isomorphic to the identity functor on SectS(A). Pick such a natural isomorphism v. The pair (σ, µ) is then given by σ : X ⊠1 Y v−1 ⊠v−1 −−−−−−→ (ϕ∗X) ⊠1 (ϕ∗Y ) ∼= ϕ∗(X ⊠2 Y ) v−−→ X ⊠2 Y, and µ : L2(A(I1)) L2(A(ϕI1 )) −−−−−−−−→ ϕ∗(cid:0)L2(A(I2))(cid:1) v−−→ L2(A(I2)). (cid:3) Definition 1.39. Given a circle S, we let H, K 7→ H ⊠ K denote any one of the monoidal structures on SectS(A) considered in Proposition 1.38, and call it "the fusion of H and K". It is well defined up to non-canonical unitary isomorphism. In the case when S is the standard circle, there are two important special cases of (1.36): the vertical fusion and the horizontal fusion on Sect(A), given by H ⊠v K := H ⊠S1 ⊤,j K H ⊠hK := H ⊠S1 ⊢,j′ K, ⊤ is the upper half of the standard circle, S1 respectively. Here, S1 j and j′ are the reflections given by j(z) = ¯z and j′(z) = −¯z. 1.d. Central decomposition. Given a conformal net A and an orientation-preserving embedding f : I → J, we will show that A(f ) : Z(A(I)) → Z(A(J)) is always an isomorphism. In fact, there is an algebra Z(A), called the center of A, that only depends on A, and that is canonically isomorphic to Z(A(I)) for every I. See [8, Sec. 3] for a similar discussion. ⊢ its left half, and 14 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES Proposition 1.40. Let A be a conformal net. Then there is an abelian von Neu- mann algebra Z(A) such that for every interval I we have a canonical isomor- phism Z(A(I)) ∼−→ Z(A), and such that for each embedding J ֒→ I, the inclusion A(J) ֒→ A(I) induces a commutative diagram (1.41) Z(A(J)) Z(A) Z(A(I)) Proof. We first show that ζ(I) := Z(A(I)) is a functor from intervals with orientation- preserving embeddings to von Neumann algebras. We define Z(A) := colim ζ. The natural map Z(A(I)) → Z(A) will be an isomorphism if and only if ζ is equivalent to a constant functor. Checking the latter condition involves two things: (a) given an embedding J ֒→ I, we show the induced map A(J) ֒→ A(I) sends Z(A(J)) isomorphically onto Z(A(I)), and (b) given two orientation-preserving embeddings α, β : J ֒→ I, we show that the induced isomorphisms α∗, β∗ : Z(A(J)) → Z(A(I)) are equal to each other. (a) Let J ֒→ I be an embedding. Without loss of generality, we may assume that I and J share a boundary point. Let J c be the closure in I of the complement of J. By Lemma 1.18, A(J) and A(J c) are each other's commutants in A(I). Hence Z(cid:0)A(J)(cid:1) = Z(cid:0)A(J c)(cid:1) = A(J) ∩ A(J c). An element of A(J) ∩ A(J c) commutes with both A(J c) and A(J), and so it also commutes with A(I) = A(J) ∨ A(J c). An element of Z(A(I)) commutes with both A(J) and A(J c), and is therefore in the intersection A(J c) ∩ A(J) of their commutants. It follows that Z(A(I)) = Z(A(J)). (b) Without loss of generality, we may take I = J, and β = IdI . Pick intervals I ⊂ I+ ⊃ K, and a diffeomorphism ϕ : I+ → I+ that restricts to α on I and to the identity on K: ϕI =α I I i i I+ ∼= ϕ I+ j j K ϕK =IdK K As the maps i∗ : Z(A(I)) → Z(A(I+)) and j∗ : Z(A(K)) → Z(A(I+)) are isomor- phisms, it follows that α∗ : Z(A(I)) → Z(A(I)) is the identity map. We have now canonically identified every Z(A(I)) with Z(A). To show that every diagram (1.41) is commutative, we still need to treat the case of orientation- reversing maps. For that purpose, it is enough to analyze the identity map from J to ¯J. Namely, we need to show that the map Z(A) ∼= Z(A(J)) i∗−−→ Z(A( ¯J )) ∼= Z(A) induced by i := IdJ : J → ¯J is the identity. Assuming the contrary, there would be a non-zero central projection p ∈ A(J) that is orthogonal to its image q := i∗(p). Letting I be an interval containing J and sharing one endpoint (as in the statement CONFORMAL NETS I: COORDINATE-FREE NETS 15 of the vacuum axiom), we would then have the following commutative diagram p ⊗ q ∈ A(I)⊗algA(I)op p ⊗ q ∈ A(J)⊗algA(J)op p ⊗ p ∈ A(J)⊗algA( ¯J) p2 = p ∈ A(J ∪ ¯J) B(L2A(I)) The element p ⊗ p goes to p2 under the bottom map because it is the product of p ⊗ 1 with 1 ⊗ p, and both get mapped to p. Since p ⊗ q acts as zero on L2A(I), by the commutativity of the above diagram, so must p, contradicting the fact that p 6= 0. (cid:3) Given an abelian von Neumann algebra A, let Spec(A) refer to any nice measure space X equipped with an isomorphism L∞(X) ∼= A. Any von Neumann algebra can be written as a direct integral of factors [13] indexed over Spec of its center. Applying this to the algebras A(I), one can then write every conformal net A as a direct integral of irreducible conformal nets: (1.42) A =Z ⊕ x∈Spec(Z(A)) Ax Here, we have secretly used that INT is equivalent to a category with countably many objects (actually one object) and has separable hom spaces. Conformal nets form a category: a morphism A → B is a natural transformation τ : A → B that assigns to each interval I a unital homomorphism τI : A(I) → B(I) of von Neumann algebras. Objectwise spatial tensor product defines a symmetric monoidal structure on that category. Lemma 1.43. Let A and B be conformal nets, and let τ : A → B be a natural transformation. Then τ (Z(A)) ⊂ Z(B). Proof. Let I = [0, 1], J = [1, 2], and K = [0, 2]. By Proposition 1.40, the natural maps Z(A(I)) → Z(A(K)) ← Z(A(J)) are isomorphisms. By locality, the algebra τ (Z(A(I))) commutes with B(J), and the algebra τ (Z(A(J))) commutes with B(I). It follows that the image of Z(A(K)) commutes with B(I) ∨ B(J) = B(K). Corollary 1.44. Let τ be a natural transformation between semisimple conformal nets (Definition 1.7). Then τ is a direct sum of maps of the form (cid:3) A −→ A ⊕ιi−−→ nMi=1 nMi=1 Bi, where A and Bi are irreducible conformal nets, n ∈ N (n = 0 allowed), A →Ln is the diagonal map, and ιi : A → Bi are inclusions. i=1 A In view of the above results, the study of arbitrary conformal nets reduces to that of irreducible conformal nets. From now on, we shall therefore assume that our conformal nets are irreducible. 1.e. Conformal embeddings. In this section, we discuss two classes of mor- phisms of special interest, namely finite morphisms and conformal embeddings, and we show that all finite morphisms are conformal embeddings. 16 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES Let A and B be irreducible conformal nets. Recall that given an interval I, we denote by Diff 0(I) the subgroup of diffeomorphisms that restrict to the identity near the boundary of I. Definition 1.45. A natural tranformation τ : A → B is called finite if for every interval I the map τI : A(I) → B(I) is a finite homomorphism (Definition A.16). We borrow the following terminology from affine Lie algebras [1]: Definition 1.46. A natural tranformation τ : A → B is called a conformal embed- ding if for every ϕ ∈ Diff 0(I), and every unitary u ∈ A(I), Ad(u) = A(ϕ) ⇒ Ad(τ (u)) = B(ϕ). Lemma 1.47. Let τ : A → B be a finite natural transformation between irreducible conformal nets. Then the relative commutant of τ (A(I)) inside B(I) is trivial. Proof. Let I = [0, 1], J = [1, 2], K = [0, 2], and let A(I)c, A(J)c, A(K)c be the commutants of A(I), A(J), A(K) inside B(K). By locality, the algebras A(I)c ∩ B(I) and A(J)c ∩ B(J) commute with A(K) = A(I) ∨ A(J). We therefore have inclusions A(I)c ∩ B(I) ֒→ A(K)c ←֓ A(J)c ∩ B(J). Since τ is finite, these algebras are finite-dimensional by Lemma A.17. They are all of the same dimension, and so the above inclusions are actually isomorphisms. Moreover, A(I)c∩B(I) and A(J)c∩B(J) commute with B(J) and B(I), respectively. It follows that A(I)c ∩B(I) = A(J)c ∩B(J) is central in B(I)∨B(J) = B(K). This finishes the argument since, by assumption, Z(B(K)) = C. Proposition 1.48. If τ : A → B is finite, then it is a conformal embedding. Proof. Let ϕ ∈ Diff 0(I) be a diffeomorphism. By the inner covariance axiom, there exists a unitary v ∈ B(I) such that Ad(v) = B(ϕ). Similarly, there exists u ∈ A(I) such that Ad(u) = A(ϕ) = B(ϕ)A(I), where the restriction occurs along the morphism τ . It follows that Ad(τ (u)v∗)A(I) = IdA(I), and hence that τ (u)v∗ ∈ A(I)′ ∩ B(I). By Lemma 1.47, τ (u) is therefore a scalar multiple of v. It follows that Ad(τ (u)) = Ad(v) = B(ϕ). (cid:3) (cid:3) 2. Covariance for the vacuum sector In this section, we study the natural projective actions of Diff(S1) and its various subgroups on the vacuum sector of a conformal net. The main result of this section is that the vacuum sector construction can be upgraded to a functor from the category of conformal circles to the category of Hilbert spaces. From now on, all conformal nets are irreducible, unless stated otherwise. 2.a. Implementation of diffeomorphisms. Given a Hilbert space H, we let U±(H) = U(H) ∪ U−(H) be the group of unitary and anti-unitary operators on H, equipped with the strong operator topology. This is a topological group2. Note that, on U±(H), the strong, weak, and ultraweak topologies all agree. Let S be a circle, let I0 ⊂ S be an interval, and let j : S → S be an orientation- reversing involution that fixes ∂I0. For a conformal net A, let H0 := L2(A(I0)) ∈ SectS(A) be the vacuum sector associated to S, I0 and j, as in Section 1.b. 2This might be surprising since, on B(H), the map a 7→ a∗ is not continuous for the strong topology. CONFORMAL NETS I: COORDINATE-FREE NETS 17 Definition 2.1. Let ϕ ∈ Diff(S) be a diffeomorphism, and let u ∈ U±(H0) be an operator that is complex linear if ϕ ∈ Diff +(S), and complex antilinear otherwise. We say that u implements ϕ if is a morphism of A-sectors. u : H0 → ϕ∗H0 Unpacking the definition, a unitary u ∈ U(H0) implements a diffeomorphism A(ϕ)(a) = Ad(u)(a) = u a u∗ ϕ ∈ Diff +(S) if (2.2) for all I ⊂ S and a ∈ A(I), and that an anti-unitary u ∈ U−(H0) implements an orientation-reversing diffeomorphism ϕ ∈ Diff −(S) if (2.3) Here, the adjoint u∗ of an antilinear operator u is defined by huξ, ηi = hξ, u∗ηi. A(ϕ)(a) = Ad(u)(a∗) = u a∗u∗. Throughout this section, we will adopt the notation I ′ 0 for the closure of S \ I0. Lemma 2.4. Let u be an (anti-)unitary operator on the Hilbert space H0. In order to check that u implements a diffeomorphism ϕ, it is enough to check (2.2) or (2.3) for a ∈ A(I0) and a ∈ A(I ′ 0). Proof. Let ϕ be a diffeomorphism, and let u be an (anti-)unitary on H0 that satisfies (2.2) (or (2.3)) for all a in A(I0) and A(I ′ 0). Let I ⊂ S be an interval. Consider the subalgebra of all elements a ∈ A(I) that satisfy (2.2) (or (2.3)). That subalgebra is closed in the ultraweak topology and contains A(I∩I0) and A(I∩I ′ 0) by assumption. By strong additivity, it is therefore equal to A(I). (cid:3) Recall that given a von Neumann algebra A, the modular conjugation J : L2(A) → L2(A) is an antilinear involution that satisfies J(aξb) = b∗J(ξ)a∗. Lemma 2.5. The modular conjugation J for L2(A(I0)) (see (A.4)) implements j. Proof. By Lemma 2.4, it is enough to verify (2.3) for a ∈ A(I0) and for b ∈ A(I ′ 0). For a ∈ A(I0) and ξ ∈ H0, we have A(j)(a)ξ = ξa = J(a∗J(ξ)), and for b ∈ A(I ′ 0), we have A(j)(b)ξ = J(J(ξ)A(j)(b∗)) = J(b∗J(ξ)). These equation are equivalent to (2.3) because J is self-adjoint. Corollary 2.6. For any diffeomorphism ϕ ∈ Diff(S), the S-sectors H0 and ϕ∗H0 are unitarily isomorphic. Proof. For ϕ ∈ Diff +(S), this is Corollary 1.15. For ϕ ∈ Diff −(S), write ϕ = j ◦ ψ for some ψ ∈ Diff +(S). By the previous lemma, we have j∗H0 ∼= H0. Therefore ϕ∗H0 = ψ∗j∗H0 ∼= ψ∗H0 ∼= H0. Lemma 2.7. Let ϕ ∈ Diff +(S, ∂I0) be a diffeomorphism that commutes with j, and let ϕ0 := ϕI0 . Then L2(A(ϕ0)) implements ϕ. Proof. By Lemma 2.4, it is enough to verify (2.2) for a ∈ A(I0) and for b ∈ A(I ′ 0). Given a von Neumann algebra A, and an automorphism f : A → A, we always have L2(f )(aξ) = f (a)L2(f )(ξ). Substituting A = A(I0), f = A(ϕ0), and ξ = L2(A(ϕ0))∗η for some η ∈ H0, we get (cid:3) (cid:3) L2(A(ϕ0)) a L2(A(ϕ0))∗η =(cid:0)A(ϕ0)a(cid:1)η, which shows that (2.2) holds for a ∈ A(I0). Given an automorphism f of a von Neumann algebra A, we also have L2(f )(ξa) = (cid:0)L2(f )(ξ)(cid:1)f (a). Substituting A = A(I0), f = A(ϕ0), ξ = L2(A(ϕ0))∗η, and a = A(j)b for some b ∈ A(I ′ 0), we get L2(A(ϕ0))(cid:0)(cid:0)L2(A(ϕ0))∗η(cid:1)A(j)b(cid:1) = η(cid:0)A(ϕ0)A(j)b(cid:1). For b ∈ A(I ′ 0), we also have that 0 := ϕI ′ 0 ξ . . Given a ∈ A(I0), then by (2.8), we have A(ϕ0)(a)ξ = L2(A(ϕ0))(cid:16)(cid:0)L2(A(ϕ0))∗ξ(cid:1)a(cid:17) = L2(A(ϕ0))Ja∗JL2(A(ϕ0))∗ξ =(cid:0)L2(A(ϕ0))J(cid:1)a∗(cid:0)L2(A(ϕ0))J(cid:1)∗ 0)(b)(cid:1) = ξ(cid:0)A(ϕ0)A(j)(b)(cid:1) = L2(A(ϕ0))(cid:0)A(j)(b)(cid:0)L2(A(ϕ0))∗ξ(cid:1)(cid:1) = L2(A(ϕ0))J(cid:16)(cid:0)JL2(A(ϕ0))∗ξ(cid:1)A(j)(b∗)(cid:17) =(cid:0)L2(A(ϕ0))J(cid:1)b∗(cid:0)L2(A(ϕ0))J(cid:1)∗ 0)(b)ξ = ξ(cid:0)A(j)A(ϕ′ A(ϕ′ ξ , 18 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES The left hand side is given by and the right-hand side is L2(A(ϕ0))(cid:0)(cid:0)L2(A(ϕ0))∗η(cid:1)A(j)b(cid:1) = L2(A(ϕ0)) b L2(A(ϕ0))∗η )b(cid:1)η, η(cid:0)A(ϕ0)A(j)b(cid:1) =(cid:0)A(j)A(ϕ0)A(j)b(cid:1) η =(cid:0)A(ϕI ′ 0 which shows that (2.2) holds for b ∈ A(I ′ 0). (cid:3) Recall that, by Remark A.5, an anti-isomorphism f : A → B induces a linear isomorphism L2(f ) : L2(A) → L2(B) that exchanges left and right actions, that is, such that L2(f )(a1ξa2) = f (a2)L2(f )(ξ)f (a1). (2.8) Lemma 2.9. Let ϕ ∈ Diff −(S) be a diffeomorphism that commutes with j and exchanges the endpoints of I0. Let ϕ0 := ϕI0 . Then L2(A(ϕ0)) ◦ J implements ϕ. Proof. Let ϕ′ which finishes the proof by Lemma 2.4. (cid:3) Given a Hilbert space H, equip PU±(H) = PU(H) ∪ PU−(H) := U±(H)/S1 with the quotient strong topology3. Recall that H0 := L2(A(I0)) denotes the vacuum sector associated to S, I0 and j, and that A is assumed to be irreducible. Proposition 2.10. Let A be a conformal net, and let H0 be as above. Then there is a unique continuous representation Diff(S) → PU±(H0), ϕ 7→ [uϕ] such that (i) uϕ is complex linear for ϕ ∈ Diff +(S), and complex antilinear otherwise. (ii) uϕ implements ϕ. Here, u denotes any preimage of [u] in U±(H0). Proof. The vacuum H0 is an irreducible sector. By Schur's lemma, the implemen- tation uϕ of a diffeomorphism ϕ is therefore unique up to phase. Moreover, an implementation always exists since, by Corollary 2.6, H0 ∼= ϕ∗H0 for any diffeo- morphism ϕ. It remains to show that the homomorphism Diff(S) → PU±(H0) is continuous. For a subinterval K ⊂ I whose boundary is contained in the interior of I, write Diff 0,K(I) for the diffeomorphisms of I that fix the complement of K pointwise. The restrictions Diff 0,K(I) → PU±(H0) are continuous by Lemma 2.11 below. The result then follows as the C∞ topology on Diff(S) is the finest one for which the inclusions Diff 0,K(I) ֒→ Diff(S) are continuous. (cid:3) 3With this topology, the projection U±(H) → PU±(H) is a locally trivial bundle. CONFORMAL NETS I: COORDINATE-FREE NETS 19 Given an interval I, by the inner covariance axiom, we have a group homomor- phism Diff 0(I) → Inn(A(I)) ∼= PU(A(I)) := U(A(I))/S1. By definition the net A is continuous for the C∞ topology on Diff 0(I) and the u-topology on Inn(A(I)) (note that we do not claim that the u-topology and the quotient strong topology coincide under the identification Inn(A(I)) = PU(A(I))). Lemma 2.11. Let Diff 0,K(I) be as in the previous proof. Then the map Diff 0,K(I) → PU(A(I)) is continuous with respect to the C∞ topology on Diff 0,K(I) and the quo- tient strong topology on PU(A(I)) = U(A(I))/S1. Proof. Pick an enlargement I of I, such that I is contained in the interior of I. By the split property axiom, the subfactor A(I) ⊂ A( I) satisfies the assumption of Proposition A.19. The two vertical maps in the following diagram are therefore homeomorphisms onto their images: Diff 0,K(I) Diff( I) PU(A(I)) Aut(A( I)). The map Diff( I) → Aut(A( I)) is continuous by our definition of conformal nets, and therefore so is the map Diff 0,K(I) → PU(A(I)). 2.b. Conformal circles and their vacuum sectors. (cid:3) Conformal circles. The group Conf(S1) of conformal maps of the standard circle S1 = {z ∈ C : z = 1} consists of all maps of the form z 7→ αz + β ¯βz + ¯α where α, β ∈ C, α2 − β2 = ±1. Those are the maps that extend to conformal transformations of unit disc in C. We let Conf+(S1) ∼= PSU(1, 1) =n(cid:0) a b c d(cid:1)(cid:12)(cid:12)(cid:12) det(cid:0) a b c d(cid:1)−1 c d(cid:1) = 1,(cid:0) a b −¯b ¯d (cid:1)o(cid:14){±1} =(cid:0) ¯a −¯c be the subgroup where α2 − β2 = 1, and Conf −(S1) := Conf(S1) \ Conf+(S1) its complement. Elements in the former are orientation-preserving maps, while ele- ments in the latter are orientation-reversing. The subgroup Conf+(S1) can also be identified with PSL2(R) by congugating it with the Cayley transform. Explicitely, the identification sends the matrix (cid:0) α β PSL2(R). ¯β ¯α(cid:1) ∈ PSU(1, 1) to 1 2(cid:0) −1 1 −i −i(cid:1)(cid:0) α β ¯β ¯α(cid:1)(cid:0) −1 i 1 i(cid:1) ∈ Definition 2.12. Let S be a circle. A conformal structure τ on S is an orbit of the right Conf(S1) action on the set Diff(S, S1). Thus, a conformal structure on S is an identification S → S1 that is only determined up to elements of Conf(S1). A conformal circle is a circle equipped with a conformal structure. If S and S′ are conformal circles, we write Conf(S, S′) for the set of all diffeomor- phisms S → S′ that are compatible with the conformal structures, and abbreviate Conf(S, S) by Conf(S). We also let Conf +(S, S′) := Conf(S, S′) ∩ Diff +(S, S′), Conf+(S) := Conf +(S, S), and similarly for Conf −. The collection of all conformal circles forms a category. The objects of that category are conformal circles (always equipped with an orientation), and the mor- phisms from S to S′ are given by Conf(S, S′) = Conf+(S, S′) ⊔ Conf−(S, S′). 20 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES The vacuum sector functor. The main goal of this section is to prove the following theorem. Recall that our conformal nets are assumed to be irreducible. Theorem 2.13. Let A be a conformal net. There is a functor S 7→ H0(S,A) from the category of conformal circles to the category of complex Hilbert spaces. For ϕ a conformal map, the operator H0(ϕ,A) is unitary when ϕ is orientation- preserving, and anti-unitary when ϕ is orientation-reversing. Moreover, H0(S,A) is naturally equipped with the following structure: (i) The Hilbert space H0(S,A) is an S-sector of A, and it is a vacuum sector of A associated to S, in the sense of Definition 1.17. (ii) The representation ϕ 7→ H0(ϕ,A) of Conf(S) on H0(S,A) extends to a continuous projective representation ϕ 7→ [uϕ] of Diff(S) satisfying the two conditions listed in Proposition 2.10. (iii) For any interval I ⊂ S, there is a unitary isomorphism of S-sectors vI : H0(S,A) → L2(A(I)) between H0(S,A) and the vacuum sector associated to S, I, and j, where j ∈ Conf−(S) is the involution that fixes ∂I (see Lemma 2.21 below). Moreover, given two conformal circles S and S′, an interval I ⊂ S, and a conformal map ϕ ∈ Conf(S, S′), the diagrams H0(S,A) (2.14) H0(ϕ,A) H0(S′,A) H0(S,A) (2.15) H0(ϕ,A) vI vϕ(I) vI L2(A(I)) L2(A(ϕ)) L2(A(ϕ(I))) L2(A(I)) L2(A(ϕj))◦J vϕj(I) H0(S′,A) L2(A(ϕj(I))) commute. if ϕ ∈ Conf +(S, S′), if ϕ ∈ Conf−(S, S′), (iv) If j ∈ Conf−(S) is the involution that fixes the boundary of I ⊂ S, then vI ◦ H0(j,A) ◦ v∗ I is the modular conjugation on L2(A(I)). define H0(A, S) asR ⊕ Remark 2.16. If the conformal net A is not irreducible, then most of Theorem 2.13 remains true. Indeed, given the direct integral decomposition (1.42) of A, we can x∈Spec(Z(A)) H0(Ax, S). The only piece of structure that is no longer present on H0(A, S) is the projective action of Diff(S). The issue is that the direct sum or direct integral of two projective representations is typically no longer a projective representation (except if the 2-cocycles are equal). Before embarking on the proof of Theorem 2.13, we list a few of its consequences. Definition 2.17. Let S be a conformal circle and A a conformal net. The Hilbert space H0(S,A) constructed in Theorem 2.13 is called the vacuum sector of A on S. The vacuum sector H0(S,A) is a unit for Connes fusion along any interval I ⊂ S. Indeed, given a right A(I)-module H, composing the isometry vI : H0(S,A) → L2(A(I)) with the unit map H ⊠A(I) L2(A(I)) ∼= H, we obtain a natural isomor- phism (2.18) H ⊠A(I) H0(S,A) ∼=−→ H. CONFORMAL NETS I: COORDINATE-FREE NETS 21 Recall that ¯I denotes I with the opposite orientation, and that there is a canonical isomorphism i : L2(A(I)) ∼= L2(A( ¯I)) under which the left/right A(I)-actions on L2(A(I)) corresponds to the right/left A( ¯I)-actions on L2(A( ¯I)). For every I ⊂ S, the vacuum sector H0(S,A) is a right A( ¯I)-module via the isomorphism A(IdI ) : A( ¯I) ∼= A(I)op . Composing i and vI , we obtain a right A( ¯I) linear isomorphism H0(S,A) → L2(A( ¯I)). Let K be any left A( ¯I)-module. Using the left unit map L2(A( ¯I)) ⊠A( ¯I) K → K, we obtain a natural isomorphism (2.19) H0(S,A) ⊠A( ¯I) K ∼=−→ K similar to (2.18). Now let S1 and S2 be conformal circles, let I1 ⊂ S1 and I2 ⊂ S2 be intervals, and let ϕ : I2 → I1 be an orientation-reversing diffeomorphism. Let us also assume that ϕ is the restriction of some element in Conf−(S2, S1). Let I ′ i denote the closure of Si \ Ii. Finally, let S3 = I ′ 2 be the circle obtained by gluing S1 and S2 along ϕ, and then removing the interior of I1. 1 ∪∂I2 I ′ The circle S3 is given a conformal structure as follows. Letting j1 ∈ Conf−(S1) be the involution that fixes ∂I1, the conformal structure on S3 is the one mak- ing j1I ′ : S3 → S1 is then also a conformal map, where j2 ∈ Conf−(S2) is the involution that fixes ∂I2. : S3 → S2 into a conformal map. Note that IdI ′ 1 ∪ j2I ′ 2 1 ∪ IdI ′ 2 S1 : , S2 : , S3 : , j1 : . The following is a refinement of Corollary 1.34 in the sense that (1.35) is now replaced by a canonical isomorphism: Corollary 2.20. Let S1, S2 and S3 be as above. View H0(S1,A) as a right A(I2)- module via A(ϕ−1). Then there is a canonical unitary isomoprhism of S3-sectors of A: H0(S1,A) ⊠A(I2) H0(S2,A) ∼= H0(S3,A). Proof. The isomorphism is given by H0(S1,A) ⊠A(I2) H0(S2,A) vI′ 1 ⊠ vI2 −−−−−−→ L2(A(I ′ 1)) ⊠A(I2) L2(A(I2)) ∼=−→ L2(A(I ′ 1−−→ H0(S3,A). 1)) v∗ I′ (cid:3) Vacuum representations of the conformal group. We now discuss a number of results we will need for the proof of Theorem 2.13. Most importantly, we construct an action of the group Conf(S1) on the vacuum sector of a conformal net. We begin with two well known facts about conformal transformations: Lemma 2.21. Let S be a conformal circle and let ζ, ζ′ ∈ S be two distinct points. (i) There is a unique j ∈ Conf−(S1) that fixes ζ and ζ′ and is an involution. (ii) The subgroup Conf+(S1,{ζ, ζ′}) of Conf +(S1) that fixes both ζ and ζ′ is isomorphic to R. The elements of this subgroup commute with the unique involution j ∈ Conf −(S) that fixes ζ and ζ′. Proof. We may assume without loss of generality that S = S1, ζ = 1 and ζ′ = −1. The stabilizer of the pair (1,−1) is given by Conf(cid:0)S1,{1,−1}(cid:1) = (cid:26) az + b bz + a(cid:12)(cid:12)(cid:12)(cid:12) a, b ∈ R, a2 − b2 = ±1(cid:27) . 22 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES cosh(t)z+sinh(t) cosh(t)z+sinh(t) = cosh(t)¯z+sinh(t) There is an isomorphism from R × Z/2 to the above group that sends (t, 0) to sinh(t)z+cosh(t) and (t, 1) to sinh(t)z+cosh(t) sinh(t)¯z+cosh(t) . The result follows since R × Z/2 has a unique element of order two, and this element is central. (cid:3) Lemma 2.22. Let PSL2(R) → PU(H) be a continuous representation for the quotient strong topology. Then it lifts uniquely to a continuous representation ]PSL2(R) → U(H) of the universal cover of PSL2(R). Proof. The composite ]PSL2(R) ։ PSL2(R) → PU(H) is a continuous projective representation of a semisimple simply-connected Lie group, and thus lifts to U(H) by the main theorem of [2]. Moreover, this lift is unique: any two lifts differ by a character, but the abelianization of ]PSL2(R) is trivial, and so it has no characters. (cid:3) Let A be a conformal net and S a conformal circle. We let j ∈ Conf−(S) be an involution, I ⊂ S an interval whose boundary is fixed by j, and we consider the Hilbert space H0 := L2(A(I)). By Proposition 2.10, H0 carries a projective Diff(S) action ϕ 7→ [uϕ] implementing diffeomorphims. On the subgroup Conf(S), this action lifts to an honest representation ϕ 7→ uϕ: Proposition 2.23. Let S be a conformal circle, and let H0 be as above. Then there exists a unique lift (2.24) Conf+(S) ϕ 7→ uϕ ∃! U(H0) Diff +(S) PU(H0) of the projective action Conf +(S) → PU(H0) constructed in Proposition 2.10 to an honest action Conf +(S) → U(H0). There also exists a lift (2.25) Conf(S) U±(H0) ∃ Diff(S) PU±(H0) of the projective action Conf(S) → PU±(H0) to an honest action Conf(S) → U±(H0). The lift 2.25 is unique up to multiplication by the character Conf(S) ։ π0(Conf(S)) ∼= {±1}. Proof. The following proof is based on a trick from [21, Prop. 1.1]. Let ]Conf+(S) be the universal cover of Conf +(S). By Lemma 2.22, the projective action of Conf+(S) lifts uniquely to an action ]Conf +(S) → U(H0). Let us denote by uϕ ∈ U(H0) the image of an element ϕ ∈ ]Conf+(S). By construction, uϕ implements the image ¯ϕ ∈ Conf+(S) of ϕ. The conjugation action of j on Conf+(S) lifts to an action, again denoted ϕ 7→ jϕj, on its universal cover. Note that (2.26) ]Conf+(S) → U(H0) ϕ 7→ JujϕjJ is a group homomorphism. By Lemma 2.5, JujϕjJ also implements ¯ϕ. The uniqueness in Lemma 2.22 then implies that uϕ = JujϕjJ. Equivalently, we have ujϕj = JuϕJ. Let f ∈ Conf −(S) be an involution that exchanges the two boundary points of I, and let F := L2(A(fI )) ◦ J. The operator F squares to one, commutes with J, and implements f by Lemma 2.9. Let r ∈ ]Conf +(S) be a lift of 'rotation by π/2', CONFORMAL NETS I: COORDINATE-FREE NETS 23 so that r4 generates the kernel of the projection map ]Conf +(S) ։ Conf+(S), and its image ¯r ∈ Conf+(S) satisfies ¯rj¯r−1 = f . I : j : f : ¯r : Finally, let R := ur be the unitary implementing r. By the above computation, the inverse of R is given by R−1 = ur−1 = ujrj = JurJ = JRJ. This implies R = JR−1J. The involutions F and RJR−1 both implement f . It follows that RJR−1 = λF for some λ = ±1. Now we compute R4 = (RJR−1J)2 = (λF J)2 = λ2F 2J 2 = 1, which shows that the action (2.26) descends to Conf +(S). (cid:3) To extend it to Conf(S), it is enough to specify where j should go. The only anti-unitary involutions implementing j are J and −J. Both assignments j 7→ J and j 7→ −J produce homomorphisms Conf(S) → U±(H0). Convention 2.27. As shown above, there are two possible actions of Conf(S) on H0. Henceforth, we will always consider the one that sends j to J. Lemma 2.28. Let ϕ ∈ Conf+(S, ∂I) be a conformal map that commutes with j, and let uϕ be the unitary constructed in 2.23 and 2.27. Then L2(A(ϕI )) = uϕ. Proof. By Lemma 2.21 (ii) the group Conf+(S, ∂I) is isomorphic to R≥0, and in particular is commutative. Consequently, given elements ϕ and ψ of that group, uϕ and uψ commute. If v is any other unitary implementing ψ, then v = θuψ for some unit complex number θ ∈ S1, and hence uϕ and v commute. In particular, uϕ commutes with L2(A(ψI )). The map Conf +(S, ∂I) → S1 (2.29) ϕ 7→ cϕ := u∗ ϕ ◦ L2(A(ϕI )) is therefore a homomorphism. Let f ∈ Conf −(S) and F = L2(A(fI ))◦ J be as in the proof of Proposition 2.23. The involution implementing f being unique up to sign, we have F = ±uf . From the equations F uϕF = uf uϕuf = uf ϕf = uϕ−1 and F L2(A(ϕI )) F = L2(A(fI )) L2(A(ϕI )) L2(A(fI )) = L2(A((f ϕf )I )) = L2(A(ϕ−1I )), it follows that cϕ = F cϕF =(cid:0)F u∗ ϕF(cid:1) ◦(cid:0)F L2(A(ϕI ))F(cid:1) = u∗ The homomorphism (2.29) therefore takes its values in {±1}. Since Conf+(S, ∂I) is connected, it must be trivial. (cid:3) ϕ−1 ◦ L2(A(ϕ−1I )) = cϕ−1. Given a conformal circle S and an interval I ⊂ S, we write L2(A(I)) for the vacuum sector associated to S, I, and the involution j ∈ Conf−(S) that fixes ∂I. Lemma 2.30. For ϕ ∈ Conf+(S), and I ⊂ S an interval, consider the map (2.31) ϕ ◦ L2(A(ϕI )) : L2(A(I)) → L2(A(ϕ(I))). T (ϕ, I) := u∗ Then 24 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES (i) T (ϕ, I) is a morphism of S-sectors. (ii) T (ϕ, I) depends only on I and on ϕ(I), but not on ϕ: Conf+(S) are such that ϕ(I) = ψ(I), then T (ϕ, I) = T (ψ, I). (iii) If ϕ, ψ ∈ Conf+(S), then T (ψ ◦ ϕ, I) = T (ψ, ϕ(I)) ◦ T (ϕ, I). if ϕ and ψ ∈ Proof. (i) By naturality of the vacuum sector construction, the map L2(A(ϕI )) is a morphism of sectors from L2(A(I)) to ϕ∗L2(A(ϕ(I))). The map uϕ : L2(A(ϕ(I))) → ϕ∗L2(A(ϕ(I))) is also a morphism of sectors. The composite u∗ ϕ L2(A(ϕI )) is there- fore also a morphism of sectors. (ii) Let K := ϕ(I) = ψ(I). By applying Lemma 2.28 to ϕ ◦ ψ−1, we get T (ϕ, I) = u∗ = u∗ = u∗ ϕ L2(A(ϕI )) ϕψ−1 L2(cid:0)A(ϕψ−1K )(cid:1) L2(cid:0)A(ψI )(cid:1) ψ u∗ ψ L2(A(ψI )) = T (ψ, I). (iii) The operator T (ψ, ϕ(I)) is a isomorphism of S-sectors. Since the S-sector structure on a vacuum sector uniquely determines the action of Conf +(S) on the sector, the operator T (ψ, ϕ(I)) : L2(A(ϕ(I))) → L2(A(ψ ◦ ϕ(I))) intertwines the two actions of Conf +(S). Therefore T (ψ, ϕ(I)) T (ϕ, I) = T (ψ, ϕ(I)) u∗ ϕ L2(A(ϕI )) ϕ T (ψ, ϕ(I)) L2(A(ϕI )) ψ L2(A(ψϕ(I))) L2(A(ϕI )) ϕ u∗ ψ ϕ L2(A(ψϕI )) = T (ψ ◦ ϕ, I). = u∗ = u∗ = u∗ (cid:3) Construction of the vacuum sector. After all the above preparation, we can finally construct the vacuum sector associated to a conformal circle: Proof of Theorem 2.13. Given a conformal circle S, let I be the category whose objects are the intervals of S, and in which every hom-set Hom(I, J) contains exactly one element. Recall that for an interval I ⊂ S, we write L2(A(I)) for the vacuum sector associated to S, I, and the involution j ∈ Conf−(S) that fixes ∂I. The assignment I 7→ L2(A(I)) extends to a functor from I to the category of S-sectors of A in the following way: given two intervals I, J ∈ I, pick a map ϕ ∈ Conf+(S) that sends I to J. The value of the functor on the unique morphism from I to J is then given by u∗ ϕ◦L2(A(ϕI )) : L2(A(I)) → L2(A(ϕ(I))). By Lemma 2.30, this assignment is well defined, independent of the choice of ϕ, and functorial. We can therefore define H0(S,A) := lim I∈I L2(A(I)). We can therefore think of a vector in H0(S,A) as a collection of vectors(cid:8)ξI ∈ L2(A(I))(cid:9)I∈I subject to the condition that (2.32) for every I ⊂ S, and every ϕ ∈ Conf+(S). Let (2.33) vI : H0(S,A) → L2(A(I)) u∗ ϕ L2(A(ϕI )) (ξI ) = ξϕ(I) {ξJ}J∈I 7→ ξI be the maps that exhibit H0(S,A) as the limit of all the L2(A(I)), and note that they are all isomorphisms as I is equivalent to the trivial category (with only one object and one identity arrow). We equip H0(S,A) with the Hilbert space structure that makes the maps vI unitary. CONFORMAL NETS I: COORDINATE-FREE NETS 25 So far, we have only constructed S 7→ H0(S,A) as a functor from conformal circles and orientation-preserving maps. Explicitly, for {ξI}I⊂S1 ∈ H0(S1,A) and ϕ : S1 → S2 an orientation-preserving map, we have H0(ϕ,A)(cid:0){ξI}(cid:1) =(cid:8)L2(A(ϕϕ−1(K)))(ξϕ−1(K))(cid:9) ∈ H0(S2,A) for K ⊂ S2. Equivalently, the map H0(ϕ,A) is characterized by the fact that (2.34) for every I ⊂ S1. ϕ(I) ◦ L2(A(ϕI )) ◦ vI H0(ϕ,A) = v∗ If now ϕ ∈ Conf −(S1, S2) is an orientation-reversing map, define H0(ϕ,A) := v∗ ϕj(I) ◦ L2(A(ϕjI )) ◦ J ◦ vI , (2.35) where I ⊂ S1 is an interval, and j ∈ Conf−(S1) is the involution that fixes ∂I. To see that (2.35) is independent of the choice of interval I, given another interval I ⊂ S1, pick ψ ∈ Conf+(S1) such that ψ(I) = I, and let j = ψjψ−1 be the involution that fixes ∂ I. Letting A1 = A(I), A1 = A( I), A2 = A(ϕj(I)), and A2 = A(ϕj( I)), we have to show that the following diagram is commutative vI L2(A1) H0(S1,A) v I L2(ψ) L2( A1) u∗ ψ L2( A1) J J J L2(ϕj) L2(A1) L2(A2) v∗ ϕj(I) L2(ψ) L2(ϕψϕ−1) L2( A1) u∗ jψj L2( A1) L2(ϕj) L2( A2) H0(S2,A) , u∗ ϕψϕ−1 L2(ϕj) L2( A2) v∗ ϕj( I) where L2(ψ) stands for L2(A(ψI )), and we have used similar abbreviations for L2(A(ϕjI )), L2(A(ϕj I )), and L2(A(ϕψϕ−1ϕj(I))). The commutativity of the left and right parts of the diagram are given by (2.32). The lower left square is commutative since J = uj on L2( A1). The commutativity of the remaining three squares is clear. Given ϕ ∈ Conf(S1, S2) and ψ ∈ Conf(S2, S3), we still need to show that H0(ψ,A) ◦ H0(ϕ,A) = H0(ψ ◦ ϕ,A). The case when both ψ and ϕ are in Conf + and the case when ψ ∈ Conf + and ϕ ∈ Conf − are clear. The case ψ ∈ Conf− and ϕ ∈ Conf+ is checked as follows: H0(ψ,A) ◦ H0(ϕ,A) =(cid:0)v∗ ψϕj(I) L2(A(ψϕjϕ−1ϕ(I))) J vϕ(I)(cid:1)(cid:0)v∗ ψϕj(I) L2(A(ψϕjϕ−1ϕ(I))) J L2(A(ϕI )) vI ψϕj(I) L2(A(ψϕjϕ−1ϕ(I))) L2(A(ϕI )) J vI ψϕj(I) L2(A(ψϕjI )) J vI = v∗ = v∗ = v∗ = H0(ψ ◦ ϕ,A). ϕ(I) L2(A(ϕI )) vI(cid:1) Finally, we summarize the structure on H0(S,A): The last case ϕ, ψ ∈ Conf− follows from the previous ones since H0(ϕ,A)H0(ψ,A) = H0(ϕj,A)H0(j,A)H0(j,A)H0(jψ,A) = H0(ϕj,A)H0(jψ,A) = H0(ϕψ,A). Here, j ∈ Conf− is an arbitrary involution, and we have used (iv) below to know that H0(j,A) is an involution. (i) The vector space H0(S,A) is an S-sector by construction. (ii) By Proposition 2.10, H0(S,A) comes equipped with a projective Diff(S) action ϕ 7→ [uϕ], uniquely determined by the requirement that uϕ implements ϕ. For ϕ ∈ Conf(S), the map H0(ϕ,A) also implements ϕ, and so [H0(ϕ,A)] = [uϕ]. (iii) The maps vI : H0(S,A) → L2(A(I)) are defined in (2.33); diagram (2.14) is 26 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES equation (2.34), and diagram (2.15) is equation (2.35). (iv) Letting ϕ = j in (2.35), we get H0(j,A) = v∗ J = vI H0(j,A) v∗ I . I J vI , which is equivalent to (cid:3) 3. Finite conformal nets and their sectors 3.a. The index of a conformal net. We have defined a conformal net as functor A : INT → VN from the category of intervals to that of von Neumann algebras. In the second paper of this series [3], we will see that a conformal net can also be used to assign von Neumann algebras to arbitrary compact 1-manifolds, i.e., disjoint unions of intervals and circles. For now we focus on disjoint unions of intervals, extending A by setting A(I1 ∪ . . . ∪ In) := A(I1) ¯⊗ . . . ¯⊗A(In). Definition 3.1. Let S be a circle, split into four intervals I1, I2, I3, I4 such that each Ii intersects Ii+1 (cyclic numbering) in a single point: (3.2) I3 I2 I4 I1 Let A be a conformal net (always assumed irreducible). The algebras A(I1 ∪ I3) = A(I1) ¯⊗ A(I3) and A(I2 ∪ I4) = A(I2) ¯⊗A(I4) act on H0(S,A) and commute with each other. The index µ(A) of the conformal net A is the minimal index of the inclusion A(I1 ∪ I3) ⊆ A(I2 ∪ I4)′: µ(A) := [A(I2 ∪ I4)′ : A(I1 ∪ I3)], where the commutant is computed in B(H0(S,A)). (See Definition A.14 and the paper [4] for recollections about the notion of the minimal index of a subfactor.) Note that the index is an invariant of the net (that is, it does not depend on the choice of circle and intervals), and that it satisfies µ(A ⊗ B) = µ(A) · µ(B). Remark 3.3. In [27], nets of finite index were called completely rational. In our context however, that terminology would be quite misleading. Consider any unitary conformal field theory (not necessarily rational, e.g., the free boson compactified on a circle of irrational squared radius) viewed as a net O 7→ A(O) on two-dimensional Minkowski space. We believe that the restriction of such a net to the zero time slice provides an example of a conformal net with index 1. Recall that a bimodule AHB is dualizable (Definition A.10) if and only if the inclusion A ⊂ B′ has finite index [4]. A conformal net A therefore has finite index if and only if the bimodule A(I1∪I3)H0(S)A(I2∪I4)op has a dual. It is useful to identify this dual: Lemma 3.4. Let S be a circle split into intervals I1, I2, I3, I4 as above, and let ¯S, ¯I1, . . . , ¯I4 be the same manifolds with the reverse orientation. Let A be conformal net with finite index. The dual of the bimodule A(I1∪I3)H0(S)A(I2∪I4)op is A( ¯I2∪ ¯I4)H0( ¯S)A( ¯I1∪ ¯I3)op using the canonical identifications A( ¯I1 ∪ ¯I3)op ∼= A(I1 ∪ I3) and A( ¯I2 ∪ ¯I4)op ∼= A(I2 ∪ I4). CONFORMAL NETS I: COORDINATE-FREE NETS 27 Proof. Let j ∈ Diff −(S) be a diffeomorphism that exchanges I1 with I4, and I2 with I3, and let us abbreviate A(j) by j∗. The S-sector H0(S,A) can be taken to be L2(A(I1 ∪ I2)), with actions aξ = aξ bξ = ξj∗(b) for a ∈ A(I1 ∪ I2) for b ∈ A(I3 ∪ I4) for ξ ∈ L2(A(I1 ∪ I2)). The actions of the bimodule A(I1∪I3)H0(S)A(I2∪I4)op are then given by (a1 ⊗ a3) · ξ · (a2 ⊗ a4) = (a1a2) ξ j∗(a3a4) for a1 ∈ A(I1), a2 ∈ A(I2)op, a3 ∈ A(I3), and a4 ∈ A(I4)op. The dual bimodule is the complex conjugate L2(A(I1 ∪ I2)), with actions (3.5) (a2 ⊗ a4) · ¯ξ · (a1 ⊗ a3) = (a1 ⊗ a3)∗· ξ · (a2 ⊗ a4)∗ = (a∗ 1a∗ 2) ξ j∗(a∗ 3a∗ 4) for a1 ∈ A(I1), a2 ∈ A(I2)op, a3 ∈ A(I3), and a4 ∈ A(I4)op. Here ¯ξ denotes the vector ξ ∈ L2 viewed as an element of L2. Note that H0( ¯S,A) = L2(A(I1 ∪ I2)) = L2(A(I1 ∪ I2)op) has actions given by aη = aη bη = ηj∗(b) for a ∈ A(I1 ∪ I2)op for b ∈ A(I3 ∪ I4)op for η ∈ L2(A(I1∪I2)op). Using the canonical identification between L2(A(I1∪I2)op) and L2(A(I1 ∪ I2)) that exchanges the left A(I1 ∪ I2)op-module structure with the right A(I1 ∪ I2)-module structure and the right A(I1 ∪ I2)op-module structure with the left A(I1 ∪ I2)-module structure, this becomes aξ = ξa bξ = j∗(b)ξ for a ∈ A(I1 ∪ I2)op for b ∈ A(I3 ∪ I4)op for ξ ∈ L2(A(I1 ∪ I2)). The bimodule A( ¯I2∪ ¯I4)H0( ¯S)A( ¯I1∪ ¯I3)op is therefore given by (3.6) a1 ∈ A(I1), a2 ∈ A(I2)op, a3 ∈ A(I3), a4 ∈ A(I4)op. The isomoprhism that inter- twines (3.5) and (3.6) is given by the modular conjugation J : L2(A(I1 ∪ I2)) → L2(A(I1 ∪ I2)). (a2 ⊗ a4) · ξ · (a1 ⊗ a3) = j∗(a3a4) ξ (a1a2), (cid:3) Instead of splitting the circle in four as in (3.2), one can also split it into 2n intervals, for arbitrary n ≥ 2. Lemma 3.7 ([27]). Let S be a circle, split into 2n intervals I1, I2, . . . , I2n such that each Ii intersects Ii+1 (cyclic numbering) in a single point: I3 I4 I2 I1 I2n Let A be a conformal net with finite index, and let ν be the square root of the index of A. Then the bimodule (3.8) A(I1∪I3∪...∪I2n−1)H0(S)A(I2∪I4∪...∪I2n)op is dualizable, and its statistical dimension is given by νn−1. 28 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES Proof. Glue intervals J1, . . . , Jn to the circle S as in the following picture . . . J2 Jn J1 and let S1 := I1 ∪ J1, Si := Ii ∪ Ji ∪ I2n+2−i ∪ ¯Ji−1 for 2 ≤ i ≤ n, Sn+1 := In+1 ∪ ¯Jn: S1 : S2 : S3 : ··· Sn+1 : The bimodule (3.8) can then be factored as (cid:18)H0(S1) ⊗ H0(S3) ⊗ . . .(cid:19) A( ¯J1) ¯⊗ A(J2) ¯⊗ A( ¯J3) ¯⊗ A(J4) ¯⊗ ...(cid:18)H0(S2) ⊗ H0(S4) ⊗ . . .(cid:19). ⊠ Its statistical dimension is therefore (see [4, Prop 5.2]) the product of dim(cid:18)A(I1∪I3∪...∪I2n−1)(cid:0)H0(S1) ⊗ H0(S3) ⊗ . . .(cid:1)A( ¯J1∪J2∪ ¯J3∪J4...)(cid:19) = dim(cid:16)A(I1)H0(S1)A( ¯J1)(cid:17) · dim(cid:16)A(I3∪I2n−1)H0(S3)A(J2∪ ¯J3)(cid:17) · dim(cid:16)A(I5∪I2n−3)H0(S5)A(J4∪ ¯J5)(cid:17)··· = 1 · ν · ν · . . . = ν⌊ n−1 2 ⌋ and dim(cid:18)A( ¯J1∪J2∪ ¯J3∪J4...)(cid:0)H0(S2) ⊗ H0(S4) ⊗ . . .(cid:1)A(I2∪I4∪...∪I2n)op(cid:19) = dim(cid:16)A( ¯J1∪J2)H0(S2)A(I2∪I2n)op(cid:17) · dim(cid:16)A( ¯J3∪J4)H0(S4)A(I4∪I2n−2)op(cid:17)··· = ν · ν · . . . = ν⌈ n−1 2 ⌉, namely ν⌊ n−1 2 ⌋ · ν⌈ n−1 2 ⌉ = νn−1. (cid:3) 3.b. Finiteness properties of the category of sectors. The goal of this section is to prove Theorem 3.14, which says that the category of sectors of a conformal net with finite index is a fusion category (Definition 3.32). Let A be an (irreducible) conformal net with finite index. Hitherto, we have been talking about the vacuum sector H0(S) = H0(S,A) as being associated to a circle S. However, it is sometimes convenient to think of it as being associated to a disk D with ∂D = S. More generally, given an oriented topological surface Σ whose boundary ∂Σ is equipped with a smooth structure, we will associate to it a Hilbert space V (Σ), well defined up to canonical-up-to-phase isomorphism. The construction of V (Σ) is rather involved, and we will only sketch it in Section 3.c. A thorough discussion, will appear in the second paper of this series [3]. In this subsection, we will describe the case when Σ is an annulus. The sector of an annulus. Let Sl be a circle, decomposed into four intervals I1, . . . , I4, as in (3.2), and let Sr be another circle, similarly decomposed into four intervals I5, . . . , I8. Let ϕ : I5 → I1 and ψ : I7 → I3 be orientation-reversing diffeo- morphisms. These diffeomorphisms equip H0(Sl) with the structure of a right CONFORMAL NETS I: COORDINATE-FREE NETS 29 A(I5) ¯⊗ A(I7)-module. We are interested in the Hilbert space Hann := H0(Sl) ⊠ A(I5) ¯⊗A(I7) H0(Sr) This is the space V (Σ) associated to the annulus Σ = Dl ∪I5∪I7 Dr, where Dl and Dr are disks bounding Sl and Sr. (3.9) I2 Dl I4 I1 I3 ϕ ψ I5 I7 Dr I6 I8 Σ=Dl∪Dr , Let Sb := I2 ∪ I8 and Sm := I4 ∪ I6 be the two boundary circles of this annulus. Sl Sr Sb Sm We equip Sb and Sm with smooth structures that are compatible with those on Sl and Sr in the sense that, locally around the trivalent points, there exist actions of S3 as in (1.30) whose restriction to each subinterval is smooth. Equivalently, the smooth structures around a trivalent point should be modeled on (1.27). The Hilbert space Hann is equipped with actions of the algebras A(J) associated to the following subintervals of Sm and Sb: • For each J ⊂ I2 or I4, the algebra A(J) acts on H0(Sl), and thus on Hann . • For each J ⊂ I6 or I8, the algebra A(J) acts on H0(Sr), and thus on Hann . • A special case of (1.25) says that whenever D is a factor, we have (A ¯⊗ D) ⊛C ¯⊗ D (B ¯⊗ D) ∼= A ⊛C B. If J ⊂ Sm or Sb intersects ∂I5 ∪ ∂I7 in one point, and that point is in the interior of J then by Proposition 1.28 and the above isomorphism, we have A(J) ∼= A(cid:0)(J ∩ Sl) ∪ I1 ∪ I3(cid:1) ⊛A(I5) ¯⊗A(I7) A(cid:0)(J ∩ Sr) ∪ I5 ∪ I7(cid:1), where the map (A(I5) ¯⊗A(I7))op → A((J ∩ Sl)∪ I1 ∪ I3) is induced by ϕ∪ ψ. That algebra then acts on Hann by Lemma 1.24. By Lemma 1.9, it follows that Hann is both an Sm-sector and an Sb-sector, and that those two structures commute. We abbreviate this by saying that Hann is an Sm-Sb-sector. Lemma 3.10. Let A be a conformal net with finite index, and let Sl, Sr, Sb, Sm, and I1, . . . , I8 be as above. Then the Sm-Sb-sector H0(Sm) ⊗ H0(Sb) is a direct summand of Hann , with multiplicity one. Proof. Let A := A(I2 ∪ I4), B := A(I1 ∪ I3)op ∼= A(I5 ∪ I7), C := A(I6 ∪ I8)op, and Hl := H0(Sl), Hr := H0(Sr), Hb := H0(Sb), Hm := H0(Sm). Since A is irreducible, and since the intervals I2, I4, I6, I8 cover Sm∪Sb, the Hilbert space Hm⊗ Hb is an irreducible A-C-bimodule. In order to show that Hm⊗ Hb is a direct summand of Hann := Hl ⊠B Hr with multiplicity one, it is therefore enough to show that (3.11) is 1-dimensional. HomA,C(Hm ⊗ Hb, Hl ⊠B Hr) 30 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES Since A has finite index, the bimodule A(Hl)B is dualizable, and its dual is H0( ¯Sl) by Lemma 3.4. Letting Hl := H0( ¯Sl), we may rewrite (3.11) as HomB,C(cid:0) Hl ⊠A (Hm ⊗ Hb), Hr(cid:1). By Corollary 1.34, Hl ⊠A (Hm ⊗ Hb) ∼= Hm ⊠A( ¯I4) Hl ⊠A(I2) Hb is isomorphic to H0(Sr). The above expression then reduces to HomB,C (Hr, Hr), which is indeed 1-dimensional. (cid:3) Sectors of nets of finite index. Using Hann as a tool, we now show that given a finite index conformal net A, there are finitely many isomorphism classes of irreducible A-sectors, and that every irreducible sector is finite, in the following sense. Definition 3.12. Let S be a circle, I ⊂ S an interval, and I ′ the closure of its complement. An A-sector on S is called finite if it is dualizable as an A(I)-A(I ′)op- bimodule. Note that the choice of interval I in the above definition is irrelevant: 1)op-bimodule if and only if it is dualizable as A(I2)-A(I ′ Lemma 3.13. For I1 and I2 subintervals of a circle S, an S-sector is dualizable as A(I1)-A(I ′ 2)op-bimodule. Proof. Pick a diffeomorphism ϕ ∈ Diff +(S) that sends I1 to I2. A sector H is dualizable as A(I2)-A(I ′ 2)op-bimodule if and only if ϕ∗H is dualizable as A(I1)- A(I ′ 1)op-bimodule. The sector ϕ∗H is isomorphic to H by Corollary 1.15. Recall from Section 1.b that for a sector K ∈ Sect(A) on the standard circle S1, given another circle S, we denote by K(S) ∈ SectS(A) the corresponding sector on S. It is given by K(S) := ϕ∗K for any ϕ ∈ Diff +(S, S1), and is only well defined up to non-canonical isomorphism. Let us also recall that our conformal nets are irreducible, and that all the von (cid:3) Neumann algebras are assumed to have separable preduals. Theorem 3.14 (Lemma 13 and Corollaries 14 and 39 of [27]). Let A be a con- formal net with finite index. Then all A-sectors are (possibly infinite) direct sums of irreducible sectors, and all irreducible A-sectors are finite. Moreover, there are only finitely many isomorphism classes of irreducible sectors. Proof. Let Sl, Sr, Sb, Sm and I1, I2, . . . , I8 be as before: I2 Sl I4 Sb I1 I5 Sm I3 I7 Sr I6 I8 and again let Hl := H0(Sl), Hr := H0(Sr), Hb := H0(Sb), Hm := H0(Sm), and Hann := H0(Sl) ⊠A(I5) ¯⊗A(I7) H0(Sr). We will also use the abbreviations A := A(I2 ∪ I4), B := A(I1 ∪ I3)op ∼= A(I5 ∪ I7), C := A(I6 ∪ I8)op, Al := A(I2), Am := A(I4)op, Cm := A(I6)op, Cr := A(I8). Since A(Hl)B and B(Hr)C are dualizable bimodules, Hann = Hl ⊠B Hr is dualizable as an A-C-bimodule, and therefore splits into finitely many irreducible summands (see Lemma A.13). Forget the actions of {A(I)}I⊂Sm on Hann , and only view it as an Sb-sector. The von Neumann algebra generated by {A(I)}I⊂Sb on Hann has a finite-dimensional center. (Otherwise, it would contradict the fact that A(Hann )C splits into finitely many irreducible summands: every central projection in that algebra commutes CONFORMAL NETS I: COORDINATE-FREE NETS 31 with A(I2), A(I4), A(I6), and A(I8), and therefore induces a non-trivial direct sum decomposition of Hann as A-C-bimodule.) We can therefore write Hann as a direct sum of finitely many factorial Sb-sectors: Hann ∼= K1(Sb) ⊕ . . . ⊕ Kn(Sb). (3.15) Here K1, . . . , Kn are A-sectors, and a sector is called factorial if its endomorphism algebra is a factor. Given an arbitrary factorial A-sector K, we now show that there exists a Ki in the above list to which K is stably isomorphic, i.e., such that K ⊗ ℓ2 ∼= Ki ⊗ ℓ2. ¯I4, we have the following isomorphisms Letting S2 := I2 ∪∂I2 of Sl-sectors: ¯I2 and S4 := I4 ∪∂I4 K(S2) ⊠Al Hl ∼= K(Sl) ∼= Hl ⊠Am K(S4) Fusing with Hr over B, we thus get K(S2) ⊠Al Hann ∼= Hann ⊠Am K(S4). We also have K(Sb) ∼= K(S2) ⊠Al Hb. By Lemma 3.10, it follows that K(Sb) ⊗ Hm ∼= K(S2) ⊠Al (Hb ⊗ Hm) ⊂ K(S2) ⊠Al Hann ∼= Hann ⊠Am K(S4). Since Am is a factor, it has only one stable isomorphism class of modules. In particular, K(S4) and L2Am are stably isomorphic as Am-modules. We therefore get a (non-canonical) inclusion of Sb-sectors: K(Sb) ⊗ ℓ2 ∼= K(Sb) ⊗ Hm ⊗ ℓ2 ⊂ Hann ⊠Am K(S4) ⊗ ℓ2 ∼= Hann ⊠Am L2Am ⊗ ℓ2 ∼= Hann ⊗ ℓ2, where the first equality uses an arbitrary unitary isomorphism ℓ2 ∼= Hm ⊗ ℓ2 of Hilbert spaces. The sector K(Sb) is factorial. It therefore maps to a single sum- mand Ki ⊗ ℓ2 of Hann ⊗ ℓ2. It follows that K and Ki are stably isomorphic. In particular, this shows that there are at most finitely many stable isomorphism classes of factorial A-sectors. By [27, Appendix C], since the algebras A(I) have separable preduals, any A- sector can be disintegrated into irreducible ones. As a consequence, if there exists a factorial sector of type II or III, then, again by the arguments in [27, Appendix C], there must be uncountably many non-isomorphic irreducible A-sectors. This is impossible, and so all factorial sectors must be of type I. Let us now go back to Hann and analyse it as an Sb-Sm-sector. Since every summand Ki(Sb) in the decomposition (3.15) is a type I factorial Sb-sector of A, we can write it as Pi⊗Qi, where Pi is an irreducible Sb-sector, and Qi = Hom(Pi, Hann ) is some multiplicity space. The multiplicity spaces Qi then carry residual Sm-sector structures, inherited from that of Hann . The decomposition (3.15) then becomes Al ¯⊗ Am (Hann ) Cr ¯⊗ Cm ∼= Mi Al(Pi) Cr ⊗ Am(Qi) Cm. Since Hann is a dualizable A-C-bimodule, the bimodules Al(Pi)Cr are also dualiz- able. To finish the argument, recall that any irreducible A-sector is isomorphic to one of the Pi, and so any irreducible sector is finite. Duals of finite sectors. Given a conformal net A with finite index, let ∆ = ∆A be the finite set of isomorphism classes of irreducible A-sectors. For every λ ∈ ∆, let Hλ be a representative of the isomorphism class. The set ∆ has an involution λ 7→ ¯λ given by sending a Hilbert space Hλ to its pullback H¯λ ∼= j∗Hλ along some element j ∈ Diff −(S1), as defined in (1.13) (note that H¯λ is only well defined up to non-canonical isomorphism). (cid:3) 32 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES Lemma 3.16. Let i : ¯S → S denote the identity map. Then, there is an isomor- phism of ¯S-sectors i∗Hλ(S) ∼= H¯λ( ¯S). Proof. Without loss of generality, we take S to be the standard circle S1. Let j : S1 → S1 be a reflection, and let us denote by ¯j the same map, viewed as an orientation-preserving map from ¯S1 to S1. We have H¯λ( ¯S1) ∼= ¯j∗H¯λ and H¯λ ∼= j∗Hλ. It follows that H¯λ( ¯S1) ∼= ¯j∗j∗Hλ ∼= i∗Hλ. (cid:3) The following two lemmas describe the duals of sectors with respect to Connes fusion of bimodules and the monoidal product on sectors, respectively. Lemma 3.17. Let S be a circle, decomposed into two subintervals I and I ′. Then the dual of the bimodule A(I)Hλ(S)A( ¯I ′) is A( ¯I ′)H¯λ( ¯S)A(I). Proof. Letting i : ¯S → S be the identity map, we have H¯λ( ¯S) ∼= i∗Hλ(S) by the previous lemma. Here, the sector H¯λ( ¯S) is the complex conjugate of Hλ, with action aξ := a∗ξ for a ∈ A( ¯J ), J ⊂ S. The bimodule A(I)Hλ(S)A( ¯I ′) has actions given by (3.18) and the bimodule A( ¯I ′)H¯λ( ¯S)A(I) has actions given by aξb := abξ a ∈ A(I), b ∈ A( ¯I ′), (3.19) Comparing (3.18) and (3.19), we see that b ¯ξa = a∗ξb∗, and so the two bimodules are dual to each other (see the discussion after A.10 or [4, Cor 6.12]). (cid:3) a ∈ A(I), b ∈ A( ¯I ′). b ¯ξa := ab ¯ξ = a∗b∗ξ Lemma 3.20. The sector H¯λ(S) is dual to Hλ(S) with respect to the monoidal structure (1.36) on SectS(A). Proof. Let j ∈ Diff −(S1) be an involution fixing ∂I, for I ⊂ S1. The sector H¯λ ∼= j∗Hλ is the complex conjugate Hλ, with actions aξ := A(j)(a∗)ξ for a ∈ A(J), J ⊂ S1. Following (1.36), we view Hλ as an A(I)-A(I)-bimodule, with actions (3.21) aξb := aA(j)(b)ξ a, b ∈ A(I). The same procedure on j∗Hλ yields the following left and right actions on Hλ: (3.22) Comparing (3.21) and (3.22), we see that a ¯ξb = b∗ξa∗, and so j∗Hλ is the dual of Hλ (again by [4, Cor 6.12]). (cid:3) a, b ∈ A(I). a ¯ξb := b∗A(j)(a∗)ξ Computation of the annular sector. The following result, even though phrased in a rather different language, is essentially equivalent to [27, Theorem 9]. Recall that A is irreducible. Theorem 3.23. Let A be a conformal net with finite index, and let Sm, Sb, and Hann be as in (3.9). We then have a non-canonical isomorphism of Sm-Sb-sectors (3.24) Hann ∼= Mλ∈∆ Hλ(Sm) ⊗ H¯λ(Sb). We draw this isomorphism as ∼= Mλ∈∆ λ ⊗ ¯λ CONFORMAL NETS I: COORDINATE-FREE NETS 33 Proof. Let Hl = H0(Sl), Hr = H0(Sr), Hl = H0( ¯Sl), and A, B, C, Al, Am, Cm, Cr be as in the proofs of Lemma 3.10 and Theorem 3.14. The Hilbert space Hann = Hl ⊠B Hr is a finite A-C-bimodule and therefore splits into finitely many irreducible summands. By the argument in the proof of Theorem 3.14, each ir- reducible summand is the tensor product of an irreducible Sm-sector and an irre- ducible Sb-sector, and so we can write Hann as a direct sum Nλµ Hλ(Sm) ⊗ Hµ(Sb) Hann ∼= Mλ,µ∈∆ with finite multiplicities Nλµ ∈ N. Given λ, µ ∈ ∆, we now compute Nλµ. By slight abuse of notation, we abbreviate Hλ := Hλ(Sm) and Hµ := Hµ(Sb). We also let K := (Hλ ⊠ Hµ)(Sr), where the operation of fusion of sectors is described in Definition 1.39. We then have HomA,C(cid:0)Hλ ⊗ Hµ, Hann(cid:1) ∼= HomA,C(cid:0)Hλ ⊗ Hµ, Hl ⊠B Hr(cid:1) ∼= HomB,C(cid:0) Hl ⊠A (Hλ ⊗ Hµ), Hr(cid:1) Hµ, Hr(cid:1) ∼= HomB,C(cid:0)Hλ ⊠Aop ∼= HomB,C(cid:0)K, Hr(cid:1) ∼= (C if µ = ¯λ Hl ⊠Al 0 m otherwise, where the last equality is given by Lemma A.12. If follows that Nλµ = δµ¯λ. (cid:3) Remark 3.25. The isomorphism (3.24) is non-canonical. It does not even make sense to ask whether or not it is canonical since the right-hand side of the equation is only well defined up to non-canonical isomorphism. Given λ ∈ ∆, consider the statistical dimension dλ := dim(A(S1 Hλ as a bimodule for two complementary intervals. Corollary 3.26. If a conformal net A has finite index, then the index satisfies ⊤)(Hλ)A(S1 ⊥)) of µ(A) = Xλ∈∆ d2 λ. Proof. Let I1, . . . , I8 be as in (3.9). By the multiplicativity of dimension under Connes fusion [4, Prop 5.2], the statistical dimension of Hann as an A(I2 ∪ I4)- A(I6 ∪ I8)-bimodule is the square of the statistical dimension of A( )H0(A)A( ). In other words, dim(cid:0)A(I2∪I4)(Hann )A(I6∪I8)(cid:1) = µ(A). The result follows from (3.24) and the fact (A.15) that dλ = d¯λ. (cid:3) 3.c. The Hilbert space associated to a surface. Given a closed 1-manifold M , we call a Hilbert space H an M -sector of A if it comes equipped with compatible actions ρI : A(I) → B(H) for all the intervals I ⊂ M . Here, compatible means that ρJ = ρIA(J) whenever J is contained in I, and that A(I) and A(J) commute whenever I and J have disjoint interiors. We denote by SectM (A) the category of M -sectors of A. Lemma 3.27. Let M be a closed 1-manifold, and let Ii ⊂ M be intervals whose interiors cover M . Let ρi : A(Ii) → B(H) be actions subject to the conditions: (i) ρiA(Ii∩Ij ) = ρjA(Ii∩Ij ), (ii) if J ⊂ Ii and K ⊂ Ij are disjoint, then ρi(A(J)) commutes with ρj(A(K)). These actions extend in a unique way to the structure of an M -sector of A. A(I) ⊗alg A(J) on H extends to the spatial tensor product A(I) ¯⊗ A(J). Moreover, if I ⊂ M and J ⊂ M are disjoint intervals, then the action of 34 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES Proof. The first statment is an immediate generalization of Lemma 1.9. If the disjoint intervals I and J belong to the same connected component of M , then we may find an interval K that contains both, in which case the map A(I) ⊗alg A(J) → A(K) → B(H) extends to A(I) ¯⊗ A(J) by the split property. Assume now that I and J belong to two different connected components; call those components S1 and S2. Applying Theorem 3.14 to H viewed as an S1-sector, we may decompose it as (3.28) H ∼= Mλ∈∆ Hλ(S1) ⊗ Mλ, where ∆ is the set of isomorphism classes of irreducible S1-sectors of A, the mul- tiplicity spaces Mλ = homSectS1 (A)(Hλ(S1), H) are Hilbert spaces, and the tensor product is the completed tensor product of Hilbert spaces. The multiplicity spaces Mλ have residual actions of A(I ′) for every interval I ′ ⊂ M not contained in S1. In particular, they are equipped with actions of A(J). It now follows from the form of the decomposition (3.28) that H supports an action of A(I) ¯⊗ A(J). (cid:3) In this section, we give a construction of a Hilbert space V (Σ) ∈ Sect∂Σ(A) associated to a topological surface with smooth boundary-we insist that every connected component of the surface have non-empty boundary. The construction depends on the auxiliary choice of particular kind of cell decomposition of Σ. We will show later, in the second paper of this series [3], that it is actually independent of any choice, and that the construction also makes sense for surfaces without boundary (at least when the conformal net has finite index). A cell decomposition of a topological surface is called regular if all the attaching maps are injective. We call a cell decomposition Σ = D1 ∪ . . . ∪ Dn ordered if the set {D1, . . . , Dn} of 2-cells is ordered. Finally, a cell decomposition is collared if the 1-cells are equipped with smooth structures and with germs of (1-dimensional) local coordinates at their two end-points. The construction of V (Σ) that we present here depends on the choice of a regular ordered collared cell decomposition of Σ. The idea of the construction is to associate to each 2-cell Di ⊂ Σ the vacuum sector H0(∂Di), and to then glue them using Connes fusion. Let Σi := D1∪. . .∪Di, and let us assume that the Mi := Di ∩ Σi−1 contain no isolated points. Give the 1-manifolds Mi the orientations coming from ∂Di. Note that these manifolds have natural smooth structures due to the presence of collars (although they are typically not smoothly embedded in Σ, even if the latter is smooth): in any dimension, if two smooth manifolds have collars along their boundary, then glueing them along some boundary components will again produce a smooth manifold. We also make the assumption that the manifolds Mi are disjoint unions of intervals (in the sequel paper, that condition will be removed). Finally, we define inductively V (Σi) by (3.29) V (Σi) :=(H0(∂D1) V (Σi−1) ⊠A(Mi) H0(∂Di) for i = 1 for 1 < i ≤ n where the algebras A(Mi) are described in the beginning of Section 3.a. For the above construction to work, we need to check that the algebras A(Mi) act on V (Σi−1) and on H0(∂Di). The existence of an action of A(Mi) on H0(∂Di) follows from the split property axiom. To see that A(Mi) acts on V (Σi−1), it is enough to show that V (Σi−1) is a ∂Σi−1-sector of A. Assuming by induction that V (Σi−2) is a ∂Σi−2-sector, that V (Σi−1) is a ∂Σi−1-sector is a consequence of the following lemma. CONFORMAL NETS I: COORDINATE-FREE NETS 35 Lemma 3.30. Let N1, N2, N3 be 1-manifolds equipped with collars, and with an identification of their boundaries ∂N1 = ∂N2 = ∂N3. Orient them so that M1 := N1 ∪ ¯N2 M2 := N2 ∪ N3 M3 := N1 ∪ N3 are closed and oriented (though not necessarily connected): N2 N1 N3 Assume furthermore that none of the connected components of N2 are circles. Let A be a conformal net, let H1 be an M1-sector, and let H2 be an M2-sector. Then H3 := H1 ⊠A(N2) H2 is an M3-sector of A. Proof. Let J ⊂ M3 be an interval that intersects N1 ∩ N3 in at most one point. If J ⊂ N1 (respectively if J ⊂ N3), then A(J) acts on H3 via its action on H1 If the point J ∩ N1 ∩ N3 is in the interior of J then, by (respectively on H2). Proposition 1.28, we have A(J) ∼= A(cid:0)(J ∩ N1) ∪ N2(cid:1) ⊛A(N2) A(cid:0)(J ∩ N3) ∪ N2(cid:1), and that algebra acts on H1 ⊠A(N2) H2 by Lemma 1.24. The Hilbert space H3 is therefore an M3-sector by Lemma 3.27. (cid:3) The requirement that N2 does not have closed components is unnecessary. In the second paper of this series, we will define A(N2) for any 1-manifold N2, and in that more general case the proof of the above lemma goes through unchanged. The collars can also be replaced by a slightly weaker piece of structure: one only needs smooth structures on M1, M2, and M3 whose relationship to each other around a trivalent point is modeled on (1.27). Corollary 3.31. The Hilbert space V (Σ) is well defined (at this point, it still depends on a choice of cell decomposition of Σ), and is naturally equipped with the structure of a ∂Σ-sector. Proof. Assuming by induction that V (Σi−1) is a ∂Σi−1-sector, we apply Lemma 3.30 to H1 := V (Σi−1) and H2 := H0(∂Di), with N2 := Mi as in (3.29), N1 the closure of ∂Σi−1 \ Mi, and N3 the closure of ∂Di \ Mi. (cid:3) 3.d. Characterization of finite-index conformal nets. Recall that a dagger category is a category equipped with an involutive, identity-on-objects, contravari- ant endofunctor. Denote by Hilb the dagger category of Hilbert spaces and bounded linear maps. Let us call a dagger category C a Hilb-category if it is a module over (Hilb,⊗), that is, if there is a functor ⊙ : Hilb × C → C along with invertible associator and unitality natural transformations ⊙ ◦ (id × ⊙) ⇒ ⊙ ◦ (⊗ × id) : Hilb × Hilb × C → C, ⊙ ◦ (1 × id) ⇒ id : C → C subject to the obvious compatibility conditions. Let us call a Hilb-category C a tensor category if it is equipped with a monoidal structure that is compatible with the Hilb-module structure. 36 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES Definition 3.32. A tensor category is called a fusion category 4 if its underlying Hilb-category is equivalent to Hilbn (the category whose objects are n-tuples of Hilbert spaces) for some finite n, and if all its irreducible objects are dualizable. In other words, a tensor category is fusion if it is semisimple with finitely many simples, and every simple is dualizable. (Note that the two possible definitions of dualizable, namely the one with and the one without the normalization condition in Definition A.10 are equivalent to each other, as shown in [4, Thm. 4.12].) We know from Theorem 3.14 and Lemma 3.20 that if A is a conformal net with finite index, then the category of S-sectors of A is fusion with respect to the monoidal structure (1.36). By a result of Longo-Xu [31], the converse also holds. In this section, we give an alternative proof of this result using the coordinate-free point of view. Theorem 3.33. If Sect(A) is a fusion category, then the conformal net A has finite index. The proof is based on the following technical lemma: Lemma 3.34. Let A and B be von Neumann algebras, and let AHB and BKA be bimodules. Let BH A be the complex conjugate of AHB. If AHB is irreducible and (1) AH ⊠B KA is a direct sum of dualizable A-A-bimodules, (2) BK ⊠A HB is a direct sum of dualizable B-B-bimodules, (3) AH ⊠B K ⊠A KB is a direct sum of irreducible A-B-bimodules, (4) AK ⊠B K ⊠A HB is a direct sum of irreducible A-B-bimodules, (where those direct sums are possibly infinite) then AHB is a dualizable bimodule. Proof. Pick dualizable bimodules A(Mi)A and maps fi : Mi → H ⊠B K so that F := L fi :L Mi → H ⊠B K is an isomorphism. Similarly, pick dualizable bimodules B(Nj)B and maps gj : K ⊠A H → Nj so that G :=L gj : K ⊠A H →L Nj is an isomorphism. The composite K ⊠A(cid:0)L Mi(cid:1) 1⊗F−−−→ K ⊠A H ⊠B K G⊗1−−−→(cid:0)L Nj(cid:1) ⊠B K being an isomorphism, one can chose indices i and j so that K ⊠A Mi 1⊗fi−−−→ K ⊠A H ⊠B K gj ⊗1 −−−→ Nj ⊠B K is non-zero. Write M for Mi, and N for Nj. By duality, the composite N ⊠B K 1⊗1⊗coevM 1⊗1⊗f ⊗1 −−−−−−−−→ N ⊠B K ⊠A M ⊠A M −−−−−−→ N ⊠B K ⊠A H ⊠B K ⊠A M −−−−−−→ N ⊠B N ⊠B K ⊠A M 1⊗g⊗1⊗1 ev N ⊗1 −−−−→ K ⊠A M is also non-zero. The bimodules N ⊠B K and K ⊠A M are direct summands of H ⊠A K ⊠B K and K ⊠A K ⊠B H, respectively. By assumption, they can therefore be written as direct sums of irreducible bimodules: N ⊠B K ∼=L Pα K ⊠A M ∼=L Qβ . 4Strictly speaking, we should be calling these 'fusion Hilb-categories', as the term 'fusion cat- egory' usually refers to categories where each object is a finite direct sum of simples. CONFORMAL NETS I: COORDINATE-FREE NETS 37 Pick α and β so that the map Pα ֒→ N ⊠B K (3.35) 1⊗1⊗coevM 1⊗1⊗f ⊗1 −−−−−−−−→ N ⊠ K ⊠ M ⊠ M −−−−−−→ N ⊠ K ⊠ H ⊠ K ⊠ M −−−−−−→ N ⊠ N ⊠ K ⊠ M ev N ⊗1 1⊗g⊗1⊗1 −−−−→ K ⊠A M ։ Qβ is non-zero. Since Pα and Qβ are irreducible, that map is actually an isomorphism. Use (3.35) to identify Pα and Qβ, and call it simply P . The maps Ev : P ⊠A H ֒→ N ⊠B K ⊠A H Coev : L2A coevM−−−−→ M ⊠A M f ⊗1 1⊗g −−→ N ⊠B N ev N−−−→ L2B −−−→ H ⊠B K ⊠A M ։ H ⊠B P then satisfy (Ev ⊠ 1P ) ◦ (1P ⊠ Coev ) = 1P . Since H is irreducible, there is some λ ∈ C for which (1H ⊠ Ev ) ◦ (Coev ⊠ 1H ) = λ1H. Moreover, by evaluating (Ev ⊠ Ev ⊠ 1P )◦ (1P ⊠ Coev ⊠ Coev ) in two different ways, one can see that λ = 1, and so H is dualizable. (cid:3) Proof of Theorem 3.33. Let ∆ be the set of isomorphism classes of A-sectors. Since by assumption the category Sect(A) is semisimple, any object H can be decomposed as (3.36) H ∼= Mλ∈∆ Hλ ⊗ HomSect(A)(cid:0)Hλ, H(cid:1), where the multiplicity spaces HomSect(A)(Hλ, H) are Hilbert spaces, and the tensor product is the completed tensor product of Hilbert spaces. We need to show that H := A(I1∪I3)H0(S)A(I2∪I4)op is a dualizable bimodule, where S and I1, . . . , I4 are as in (3.2). For that, we verify the assumptions in Lemma 3.34 for the bimodules H and K := A(I2∪I4)op H0( ¯S)A(I1∪I3). We only check the first and third conditions of the lemma, as the other two are entirely similar. Following the notation of the lemma, we let A := A(I1 ∪ I3) and B := A(I2 ∪ I4)op. (1) The fusion AH ⊠B KA is the Hilbert space Hann studied in Section 3.b. It ¯I3. By applying is an S1-S3-sector for the circles S1 := I1 ∪∂I1 (3.36) to Hann , viewed as an S1-sector, we get ¯I1 and S3 := I3 ∪∂I3 Hλ(S1) ⊗ Hom(cid:0)Hλ(S1), Hann(cid:1). Hann ∼= Mλ∈∆ The multiplicity space Hom(Hλ(S1), Hann ) is itself an S3-sector, and so we can apply (3.36) once more to write Hann ∼= Mλ,µ∈∆ Hλ(S1) ⊗ Hµ(S3) ⊗ Vλµ, where Vλµ is a multiplicity space. The A-A-bimodule Hλ(S1) ⊗ Hµ(S3) is a ten- sor product of A(I1)Hλ(S1)A( ¯I1)op and A(I3)Hµ(S3)A( ¯I3)op , which are dualizable by assumption. We have thus verified the first condition of Lemma 3.34. 38 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES (3) Observe that H ⊠B K ⊠A K is the Hilbert space V (Σ) associated to the following surface (with the indicated cell decomposition) Σ = with boundary ∂Σ = S. By Corollary 3.31, it is not only an A-B-bimodule, but also an S-sector. By strong additivity, the forgetful functor from S-sectors to A-B- bimodules is fully faithful. Therefore, a subspace of H ⊠B K ⊠A K is an irreducible sub-S-sector if and only if it is an irreducible sub-A-B-bimodule. Since SectS(A) is semisimple by assumption, every S-sector can be written as a direct sum of irreducible S-sectors. Decomposing H ⊠B K ⊠A K as a direct sum of irreducible S-sectors then also provides a decomposition into irreducible A-B-bimodules. This verifies the third condition of Lemma 3.34. (cid:3) 4. Comparing conformal and positive-energy nets 4.a. Circle-based nets. In this section, we present an alternative version of the definition of conformal nets that will provide an intermediary between our notion of coordinate-free conformal nets and existing notions of conformal nets in the literature [19, 30]. Recall that S1 := {z ∈ C : z = 1} denotes the standard circle, and S1 ⊤ := {z ∈ S1 : ℑm(z) ≥ 0} its upper half. Let INTS1 be the poset of subintervals of S1. Given a Hilbert space H (always assumed separable), we write VNH for the poset of von Neumann subalgebras of B(H). Recall that given an interval I ⊂ S1, we denote by I ′ the closure of its complement. projective action Diff(S1) → PU±(H), ϕ 7→ [uϕ] and an order preserving map A net on the circle (A, H) is a Hilbert space H equipped with a continuous INTS1 → VNH , I 7→ A(I). It is required that uϕ be complex linear if ϕ is orientation-preserving and com- plex antilinear if ϕ is orientation-reversing, where uϕ is any representative of [uϕ]. Moreover, the continuity condition refers to the C∞ topology on Diff(S1), and the quotient of the strong topology on U±(H). Definition 4.1. A conformal net on the circle is a net on the circle (A, H), and a unitary isomorphism v : H ⊤)), subject to the following conditions. Here, I, K, and L will denote subintervals of S1. ∼=−→ L2(A(S1 (i) Locality: If I and K have disjoint interiors, then A(I) and A(K) are commuting subalgebras of B(H). (ii) Strong additivity: If L = I ∪ K, then A(L) = A(I) ∨ A(K). (iii) Split property: (iv) Covariance: For ϕ ∈ Diff(S1), we have uϕA(I)u∗ (v) Vacuum: The map v intertwines the A(S1 If I and K are disjoint, then the ultraweak closure of the algebraic tensor product A(I)⊗alg A(K) ⊂ B(H) is the spatial tensor product A(I) ¯⊗ A(K). If ϕ ∈ Diff(S1) restricts to the identity on I ′, then uϕ ∈ A(I). ⊤)-module structures on H and ⊤)). Moreover, letting j : S1 → S1 be complex conjugation, on L2(A(S1 and J the modular conjugation on L2(A(S1 ⊤)), then we have uj = v∗J v in PU−(H). ϕ = A(ϕ(I)). We remind the reader that intervals are by definition closed. In particular, there is a gap between the disjoint intervals appearing in the preceding definition. CONFORMAL NETS I: COORDINATE-FREE NETS 39 Construction (coordinate-free to circle-based conformal nets). Let A be an ir- reducible conformal net (Definition 1.1). By Theorem 2.13, there is a canonical vacuum sector H0(S1,A) associated to the standard circle, and it is equipped with a continuous projective action ϕ 7→ [uϕ] of Diff(S1) such that uϕ implements ϕ. We therefore obtain a net on the circle by setting H := H0(S1,A), and defining A(I) to be the image of A(I) under its action on H0(S1,A). Proposition 4.2. Let A be an irreducible conformal net (Definition 1.1). Then the above construction produces a conformal net on the circle (Definition 4.1). Proof. The locality, strong additivity, and split property axioms of Definition 4.1 follow immediately from the corresponding axioms of Definition 1.1. We have uϕA(I)u∗ ϕ = A(ϕ(I)) because uϕ implements ϕ. If ϕ restricts to the identity on I ′, then since uϕ implements ϕ, uϕ commutes with A(I ′), and so uϕ ∈ A(I ′)′ = A(I) by Haag duality, Proposition 1.18. The isomorphism v : H → L2(A(S1 ⊤)) is the ⊤)-modules, and map vS1 we have uj = H0(j,A) = v∗J v, by parts (ii) and (iv) of Theorem 2.13. (cid:3) Construction (circle-based to coordinate-free conformal nets). Let now (A, H) be a conformal net on the circle. Given an abstract interval I ∈ INT, we let EI be the category whose objects are smooth embeddings of I into S1, and in which every hom-set contains exactly one element. We define a functor from part (iii) of Theorem 2.13. It is a morphism of A(S1 ⊤ AI : EI → VN as follows. At the level of objects, it is given by AI (ι) := A(ι(I)). To define AI on morphisms, we need to provide an (anti-)isomorphism A(ι(I)) → A(ι′(I)) for every pair of elements ι, ι′ ∈ EI . Given ι, ι′ ∈ EI , pick an extension ϕ ∈ Diff(S1) of ι′◦ ι−1 : ι(I) → ι′(I). If ϕ is orientation-preserving then a 7→ uϕau∗ defines an isomorphism A(ι(I)) → A(ι′(I)), and if ϕ is orientation-reversing then ϕ defines an anti-isomorphism A(ι(I)) → A(ι′(I)). If ϕ is another exten- a 7→ uϕa∗u∗ sion of ι′◦ι−1, then ϕ−1 ϕ is the identity on ι(I), and so Ad(uϕ−1 ϕ) is the identity on A(ι(I)) by the covariance and locality axioms. Therefore, the (anti-)isomorphism A(ι(I)) → A(ι′(I)) defined above is independent of the choice of extension ϕ. If ι′′ : I → S1 is a third embedding, one checks, by picking compatible extensions ϕ, ϕ′, ϕ′′ of ι′′ ◦ ι−1, ι′′ ◦ ι′−1, and ι′ ◦ ι−1, that Ad(uϕ) = Ad(uϕ′) ◦ Ad(uϕ′′ ). The above prescription therefore defines a functor AI , and it makes sense to set ϕ A(I) := limι∈EI AI (ι) = limι∈EI A(ι(I)). ∗uϕ By definition, an element a ∈ A(I) is a collection of operators {aι ∈ A(ι(I))}ι∈EI ∗ = aϕ◦ι for each embedding ι : I → S1 and subject to the conditions that uϕaιuϕ ϕ ∈ Diff +(S1), and that uϕaι Given a smooth embedding of abstract intervals f : I → J, there is an induced (anti-)homomorphism A(f ) : A(I) → A(J) that sends a = {aι ∈ A(ι(I))}ι∈EI to ∗ (or b = the unique element b = {b ∈ A((J))}∈EJ that satisfies b = uϕaιuϕ ∗) for every ϕ ∈ Diff +(S1) (respectively ϕ ∈ Diff −(S1)) with  ◦ f = ϕ ◦ ι. uϕa∗ This defines a functor A : INT → VN that sends orientation-preserving embeddings to homomorphisms, and orientation-reversing embeddings to anti-homomorphisms. ∗ = aϕ◦ι for each ι and ϕ ∈ Diff −(S1). ι uϕ Proposition 4.3. Let (A, H) be a conformal net on the circle (Definition 4.1), then the functor A : INT → VN given by the above construction is a coordinate-free conformal net (Definition 1.1). Proof. The locality, strong additivity, and split property axioms for (A, H) imply the corresponding axioms for A. For the remaining requirements, we argue as follows. 40 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES Covariance. Let ϕ ∈ Diff +(I) be a diffeomorphism that restricts to the identity on a neighborhood of ∂I. We assume without loss of generality that I is contained in the standard circle, and let ϕ ∈ Diff +(S1) be the extension of ϕ by the identity map on I ′. The map A(ϕ) : A(I) → A(I) is given by {aι}ι∈EI 7→ {b}∈EI , where the b ∗ for every ψ ∈ Diff +(S1) with are determined by the requirement that b = uψ aιuψ  ◦ ϕ = ψ ◦ ι. Letting ι =  = idI and ψ = ϕ, we learn that bidI = u ϕaidI u∗ ϕ. Under the identification {aι} 7→ aidI of A(I) with A(I), the map A(ϕ) : A(I) → A(I) therefore corresponds to Ad(u ϕ) : A(I) → A(I). Since u ϕ ∈ A(I), it follows that A(ϕ) is an inner automorphism. Continuity. Let I be an interval, which we take, without loss of generality, to be a subinterval of S1. We identify A(I) with A(I) by a = {aι} 7→ aidI . By Lemma 4.4 below, it suffices to check the continuity of the map Diff +(I) → Aut(A(I)). Letting Diff +(S1, I) := {ϕ ∈ Diff +(S1) ϕ(I) = I} and PN(A(I)) := {U ∈ U(H) U A(I)U ∗ = A(I)}/S1, we have the following commutative diagram Diff +(S1) Diff +(S1, I) Diff +(I) u PU(H) uDiff+(S1 ,I) PN(A(I)) Ad Aut(A(I)) where the existence of the middle vertical map is guaranteed by the covariance axiom. The two horizontal maps on the left are subgroup inclusions, equipped with subspace topologies. The map Ad is continuous by Lemma A.18, and the map u is continuous by assumption. The middle vertical is continuous by restriction. The vertical map on the right is ϕ 7→ A(ϕ), and we have to show that it is continuous. The C∞ topology on Diff(I) coincides with the quotient topology under the map Diff +(S1, I) ։ Diff(I). The continuity of ϕ 7→ A(ϕ) therefore follows from that of Ad◦ uDiff+(S1,I). Vacuum sector. Let K ( I contain the boundary point p ∈ ∂I, and let K ∪p ¯K be equipped with any smooth structure that extends the ones on K and ¯K and for which the orientation-reversing involution jK that exchanges K and ¯K is smooth. We have to show that the following diagram can be completed: A(K)⊗algA( ¯K) (id,A(jK )) A(K)⊗algA(K)op A(I)⊗algA(I)op A(K ∪p ¯K) B(L2A(I)). Extend the smooth structure on K ∪p ¯K to one on S := I ∪∂I ¯I (make sure that the involution that exchanges I and ¯I is smooth), and pick an orientation- preserving diffeomorphism ϕ : S → S1 that intertwines the above involution with the standard involution j on S1 and that sends I to the upper half S1 ⊤ of the circle. We then have isomorphisms A(K) → A(ϕ(K)) A(K ∪ ¯K) → A(ϕ(K ∪ ¯K)) A( ¯K) → A(ϕ( ¯K)) A(I) → A(S1 ⊤) CONFORMAL NETS I: COORDINATE-FREE NETS 41 (given by a 7→ aι where ι is the appropriate restriction of ϕ) that make the following diagram commute: A(K)⊗alg A( ¯K) (id, A(jK ) A(K)⊗alg A(K)op A(I)⊗alg A(I)op A(K∪p ¯K) A(ϕ(K))⊗alg A(ϕ(K∪p ¯K)) (id, a7→uj a ∗ A(ϕ( ¯K)) B(L2A(I)) uj ) A(ϕ(K))⊗alg A(ϕ(K))op A(S1 ⊤)⊗alg A(S1 ⊤)op B(H) ∼= B(L2A(S1 ⊤)) Indeed, if a = {aα} ∈ A( ¯K) and b = {bβ} ∈ A(K) are such that A(jK )(a) = b, then ψ for every ψ ∈ Diff −(S1) with β ◦ jK = ψ ◦ α. Setting α = ϕ, β = ϕ, bβ = uψa∗ αu∗ and ψ = j, we get the commutativity of the rear left square: b → ϕuj = bϕ. Let us identify K, ¯K, and K∪ ¯K with their images under ϕ in order to simplify the notation. The existence of the top dotted arrow is now equivalent to the existence of an arrow completing the following diagram: a → aϕ 7→ uja∗ A(K)⊗alg A( ¯K) A(K ∪p ¯K) (id, a7→uj a∗uj ) A(K)⊗alg A(K)op A(S1 ⊤)⊗alg A(S1 ⊤)op ∼= B(H) B(L2A(S1 ⊤)) We claim that the natural action A(K ∪p ¯K) → B(H) provided by the data of a conformal net on the circle makes the above diagram commute. On the subalgebra A(K) of A(K)⊗alg A( ¯K), the commutativity of the above diagram is obvious since, by assumption, the isomorphism v : H → L2(A(S1 ⊤)) ⊤). On the subalgebra A( ¯K), we argue is equivariant for the left actions of A(S1 as follows. Pick a ∈ A( ¯K), with image uja∗uj ∈ A(K)op ⊂ A(S1 ⊤)op. That element goes to J(uja∗uj)∗J under the right action map A(S1 ⊤)), where J is the modular conjugation. By assumption, the isomorphism B(H) → ⊤)) sends uj to J. Since J = J ∗, the above expression then simplifies to B(L2A(S1 J(Ja∗J)∗J = a, which proves the commutativity of the diagram. (cid:3) ⊤)op → B(L2A(S1 Recall the definition of the continuity of a net from the beginning of Section 1.a and Haagerup's u-topology from the Appendix. Lemma 4.4. Given a net A, if the natural maps Diff +(I) → Aut(A(I)) are con- tinuous, then so are the maps HomINT(I, J) → HomVN(A(I),A(J)). Proof. It is enough to show continuity on the subset Hom+ INT(I, J) of orientation- preserving embeddings. Given a generalized sequence ϕi ∈ Hom+ INT(I, J), indexed by the poset I, with limit ϕ, and given a vector ξ ∈ A(J)∗ in the predual, we need to show that A(ϕi)∗(ξ) converges to A(ϕ)∗(ξ) in A(I)∗. Pick an interval K, identify I and J with subintervals of K via some fixed embeddings I ֒→ K, J ֒→ K into its interior, and extend ϕ to a diffeomorphism ϕ ∈ Diff +(K). For each natural number n, pick an extension ϕn,i ∈ Diff +(K) of ϕi such that k ϕn,i − ϕkC n < kϕi − ϕkC n, where k kC n is any norm that induces the Cn topology on Diff +(K). It follows that F -lim ϕn,i = ϕ, 7 7 42 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES where F is the filter5 on N × I generated by the sets {(n, i) ∈ N × I n ≥ n0, i ≥ i0(n)}n0∈N,i0:N→I . Given a lift ξ ∈ A(K)∗ of ξ then, by assumption, F -lim A( ϕn,i)∗( ξ) = A( ϕ)∗( ξ). Composing with the projection π : A(K)∗ ։ A(I)∗, it follows that A(ϕi)∗(ξ) = π(A( ϕn,i)∗( ξ)) converges to π(A( ϕ)∗( ξ)) = A(ϕ)∗(ξ). 4.b. Positive-energy nets. The following notion, that we call "positive-energy net", corresponds to what most people would call "conformal net" in the literature. The goal of this section is to show that positive-energy nets subject to the extra conditions of strong additivity, the split property, and diffeomorphism covariance, yield examples of conformal nets in the sense of Definition 1.1. (cid:3) Definition 4.5. A positive-energy net (A, H) is a Hilbert space H, a unit vector Ω ∈ H called the vacuum vector, a continuous action Conf+(S1) → U(H), ϕ 7→ uϕ, and an order preserving map subject to the following conditions. Here, I and J are subintervals of S1: INTS1 → VNH , I 7→ A(I) (i) Locality: If I and J have disjoint interiors, then A(I) and A(J) are com- muting subalgebras of B(H). (ii) Covariance: For ϕ ∈ Conf+(S1), we have uϕA(I)u∗ (iii) Vacuum vector: The subspace of Conf+(S1)-invariant vectors of H is spanned by Ω. Moreover, Ω is a cyclic vector for the action of the al- ϕ = A(ϕ(I)). gebraWI∈INTS1 A(I). (iv) Positive-energy: Let L0 be the conformal Hamiltonian, defined by the equation urα = eiαL0, where rα ∈ Conf+(S1) is the anticlockwise rotation by angle α. Then L0 is a positive operator. There are three further conditions that may be imposed on a positive-energy net: 1. A positive-energy net satisfies strong additivity if A(I ∪ J) = A(I) ∨ A(J). Note that if the interiors of I and J have non-empty intersection, then that condition is automatic [18]. 2. A positive-energy net satisfies the split property if for any pair of disjoint subintervals I, J ⊆ S1, the closure of the algebraic tensor product A(I)⊗alg A(J) ⊂ B(H) is the spatial tensor product A(I) ¯⊗ A(J). 3. A positive-energy net is diffeomorphism covariant if u : Conf+(S1) → U(H) extends to a continuous projective action of the orientation-preserving diffeomor- phisms Diff +(S1) → PU(H), ϕ 7→ [uϕ], such that (i ) uϕA(I)u∗ (ii ) if ϕ has support in I (i.e., is the identity outside I), then uϕ ∈ A(I), ϕ = A(ϕ(I)) for any lift uϕ ∈ U(H) of [uϕ]. If a positive-energy net is diffeomorphism covariant then, by a result of Carpi and Weiner [11, Theorem 5.5], the extension Diff +(S1) → PU(H) is uniquely determined by the above two conditions. We will use this result to further extend the action to orientation-reversing diffeomorphisms. Recall that S1 ⊤ = {z ∈ S1 : ℑm(z) ≥ 0} is the upper half of the standard circle, and that j : S1 → S1 denotes complex conjugation. Proposition 4.6. Let (A, H) be a positive-energy net. Then there is an A(S1 linear unitary isomorphism v : H ⊤)- ⊤)) such that, letting J = v∗J v with ∼=−→ L2(A(S1 5Recall that a filter F on a set I is a collection of subsets S ⊂ I such that if S ∈ F and T ∈ F then S ∩ T ∈ F , and if S ∈ F and S ⊂ S ′ then S ′ ∈ F . The expression F -lim xi = x means that for every neighborhood U of x, the set {i ∈ I xi ∈ U } is an element of F . CONFORMAL NETS I: COORDINATE-FREE NETS 43 J the modular conjugation, we have JA(I)J = A(j(I)) J uϕ J = uj◦ϕ◦j ∀ I ⊂ S1, ∀ ϕ ∈ Conf +(S1). Proof. By the Reeh-Schlieder theorem [19, Thm. 2.8], Ω is cyclic and separating for each algebra A(I), and in particular for A(S1 ⊤). Since Ω is separating, the ⊤)), ω(a) = haΩ, ΩiH , is faithful. The vector ω1/2 corresponding state ω ∈ L1(A(S1 is therefore cyclic in L2(A(S1 ⊤)-linear isometry ⊤)) → H that sends ω1/2 to Ω. That map is then surjective because Ω is L2(A(S1 cyclic for the action of A(S1 The operator J is the modular conjugation of Tomita-Takesaki theory for the action of A(S1 ⊤) on H with respect to the cyclic and separating vector Ω. Using a result of Borchers [5, Thm. II.9], one can then show that JA(I)J = A(j(I)) and JuϕJ = uj◦ϕ◦j, see [19, Thm. 2.19]. (cid:3) ⊤)), and so there is a unique A(S1 ⊤) on H. Let j : S1 → S1 and J : H → H be as in the previous proposition. Proposition 4.7. If a positive-energy net (A, H) is diffeomorphism covariant, then the formula (4.8) defines an extension of the projective action of Diff +(S1) on H to the group Diff(S1) of all diffeomorphisms of S1. uϕ◦j = uϕ ◦ J, ϕ ∈ Diff +(S1) Proof. In order to show that (4.8) defines a representation, we need to verify that ujϕj = J uϕJ holds up to phase for all ϕ ∈ Diff +(S1). Consider the homomoprhism Diff +(S1) → PU(H) : ϕ 7→ [uϕ] given by uϕ = Jujϕj J. We have to show [uϕ] = [uϕ] for all ϕ ∈ Diff +(S1). For ϕ ∈ Conf+(S1), this equation holds by Proposition 4.6. In particular, both u and u are extensions of uConf +(S1). By the uniqueness result of Carpi and Weiner [11, Theorem 5.5], it suffices to check that u satisfies the same two conditions as u: (i ) uϕA(I)u∗ (ii ) if ϕ has support in I, then uϕ ∈ A(I), ϕ = A(ϕ(I)) ϕ = Jujϕj JA(I)Ju∗ jϕj J = JujϕjA(j(I))u∗ For the first condition, we check using Proposition 4.6 that uϕA(I)u∗ jϕj J = JA(j(ϕ(I)))J = A(ϕ(I)). For the second condition, if ϕ has support in I, then jϕj has support in j(I). By the definition of diffeomorphism covariance, it follows that ujϕj ∈ A(j(I)), and so uϕ = Jujϕj J ∈ JA(j(I))J = A(I) by Proposition 4.6. Proposition 4.9. Every positive-energy net (Definition 4.5) that satisfies strong additivity, the split property, and diffeomorphism covariance extends to a conformal net on the circle (Definition 4.1) and therefore to a conformal net (Definition 1.1). Moreover, the resulting conformal net is irreducible. (cid:3) Proof. We first check that a positive-energy net extends to a conformal net on the circle. The isomorphism v : H → L2(A(S1 ⊤)) is given by Proposition 4.6. The projective action of Diff(S1) on H is provided by Proposition 4.7. The equation uj = J holds by definition, and the condition uψA(I)u∗ ψ = A(ψ(I)) for ψ = ϕj ∈ Diff −(S1) follows from Proposition 4.6: uψA(I)u∗ ϕ = A(ϕ(j(I))) = A(ψ(I)). ψ = uϕJA(I)Ju∗ ϕ = uϕA(j(I))u∗ To get a conformal net from a positive-energy net, apply Proposition 4.3 to the conformal net on the circle associated to the positive-energy net. Irreducibility follows, using [30, Proposition 6.2.9], from the uniqueness of the vacuum vector. (cid:3) (4.11) c(f, g) = 1 2πZS1hf, dgi 44 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES We will later need the following result from the literature: Theorem 4.10. Let (A, H) be a positive-energy net with conformal Hamilton- If e−βL0 is of trace class for all β > 0, then (A, H) satisfies the split ian L0. property. Proof. See [30, Thm 7.3.3] or [19, Lem. 2.12]. (cid:3) 4.c. The loop group conformal nets. In this section, we describe, following [19], the construction of positive-energy nets associated to loop groups. We verify that these positive-energy nets satisfy strong additivity, the split property, and diffeomorphism covariance, and therefore extend to coordinate-free conformal nets. There is such a net associated to each compact, simple, simply connected Lie group G equipped with a choice of a positive integer k ∈ N called the level. Let g be the Lie algebra of G, gC its complexification, and h ⊂ gC a Cartan subalgebra. When dealing with loop groups, it is customary [32] to equip g with the negative definite G-invariant inner product h·,·i such that every short coroot θ ∈ h has hθ, θi = 2-this is the so-called basic inner product. With respect to the dual inner product, a long root α ∈ h∗ then satisfies hα, αi = 2. Let Lg := C∞(S1, g) be the Lie algebra of functions on S1 with values in g, under the pointwise Lie bracket operation. Let c be the 2-cocycle on Lg given by and let fLg be the central extension of Lg by R that corresponds to that cocycle. Finally, let ω be the left invariant closed 2-form on the loop group LG := C∞(S1, G) whose value at the origin is given by c. The integral of ω against a generator of H2(LG, Z) ∼= Z is 2π, and so there is a principal bundle with connection R/2πZ → P → LG whose curvature is ω. The group of connection-preserving automorphisms of P that cover left translations is the simply connected Lie group that integrates the Lie algebra fLg. It is a central extension of LG by the abelian group S1 := R/2πZ, and it is the universal central extension of LG inside the category of Fr´echet Lie groups. S1 → gLG → LG To construct the appropriate Hilbert space representations ofgLG, one starts at the Lie algebra level. The Lie algebra Lg has a dense subalgebra Lgpol given by Laurent polynomial functions C× → gC whose restriction to S1 ⊂ C× take values in g. Its complexification Lgpol ⊗R C = gC[z, z−1] is the algebra of all polynomial functions on C× with values in gC. We denote by g the central extension of Lgpol given by the same cocycle (4.11), and by gC = g ⊗R C the corresponding central extension of gC[z, z−1] by C. Finally, we let g+ ⊂ gC be the restriction of that last central extension to gC[z] ⊂ gC[z, z−1]. Note that the cocycle c is trivial on gC[z], and so g+ splits as a direct sum of Lie algebras: g+ ∼= gC[z] ⊕ C. Let C0,k be the one-dimensional g+-module in which the first summand gC[z] acts by zero, and the second summand C acts by x 7→ kix (the derivative of the kth irreducible representation of S1). We then consider the induced gC-module W0,k := U gC ⊗U g+ C0,k, and let L0,k := W0,k/J be the quotient by its unique maximal proper submodule. The module L0,k can be equipped [25, Chapt. 11] with a positive definite g-invariant inner product. We denote by H0,k its Hilbert space completion. The action of g on L0,k extends to an action of fLg on H0,k by unbounded skew-adjoint operators, and the latter can then be integrated [20, 38] to a continuous unitary representation (4.12) π :gLG −→ U(H0,k). CONFORMAL NETS I: COORDINATE-FREE NETS 45 Moreover, by [20, Thm. 6.7] or [38, Thm. 6.1.2], one can use the Segal-Sugawara formulae to extend the induced map LG → PU(H0,k) to a continuous projective representation LG ⋊ Diff +(S1) −→ PU(H0,k). (4.13) Finally, the projective action of Diff +(S1) on H0,k restricts to an honest action of the conformal group Conf+(S1) ⊂ Diff +(S1), and so one gets a homomorphism (4.14) Note that the infinitesimal generator L0 of the rotation group S1 ⊂ Conf+(S1) has positive spectrum, and satisfies the assumption of Theorem 4.10: gLG ⋊ Conf+(S1) −→ U(H0,k). Lemma 4.15. Let L0 be the operator on H0,k defined by urα = eiαL0 (see Def- inition 4.5), where rα ∈ S1 is the anticlockwise rotation by angle α. Then L0 is a positive self-adjoint operator. Moreover, for every β > 0, the operator e−βL0 is trace class. Proof. Let L0,k(n) ⊂ L0,k denote the subspace where S1 acts by its nth represen- tation, and let W0,k(n) be the corresponding subspace of W0,k. We have to show that L0,k(n) = 0 for n < 0, and that trH0,k (e−βL0) =Xn≥0 dim(L0,k(n))e−βn < ∞. Clearly, since L0,k(n) is a quotient of W0,k(n), it is enough to show that W0,k(n) = 0 for n < 0, and thatPn dim(W0,k(n))e−βn is summable. By the Poincar´e-Birkhoff-Witt theorem, there is an S1-equivariant isomorphism between W0,k = U gC⊗U g+ C0,k and the symmetric algebra on gC/g+ = z−1gC[z−1]. The rotation group S1 acts by its nth representation on the span of z−n. We therefore have W0,k(n) = 0 for n < 0, and Xn dim(W0,k(n))e−βn = Yn≥1(cid:0)1 + e−nβ + e−2nβ + e−3nβ + . . .(cid:1)dim(g) , (cid:3) which converges. for the respective Lie algebras. For an interval I ⊂ S1, we denote by LI G ⊂ LG the subgroup of loops with support in I, and by gLIG the preimage of LI G ingLG. We also write LI g and gLI g Consider π(2) : gLG ×gLG → U(H0,k), π(2)(g, h) = π(gh), with π as in (4.12). Then π(2)(cid:0)gLIG × gLJ G(cid:1) is dense in π(cid:0)]LKG(cid:1), where those two groups are given the The following proposition is proven in [37, Chapter IV, Proposition 1.3.2]: Proposition 4.16. Let I, J ⊂ S1 intersect in one point, and let K be their union. subspace topology from (U(H0,k), strong). Definition 4.17. The loop group net for G at level k is the positive-energy net (AG,k, H0,k) given by AG,k : INTS1 → VNH0,k I 7→ π(cid:0)gLI G(cid:1)′′ along with the action (4.14) of Conf+(S1) on H0,k, and the unique (up to scalar) fixed vector Ω ∈ H0,k for the rotation group S1 ⊂ Conf +(S1). The axioms of positive-energy nets are verified as follows. First note that if I and J ⊂ S1 have disjoint interiors, then the cocycle c(f, g) vanishes for f ∈ LI g and g ∈ LJ g. As a consequence, the Lie algebras gLI g and gLJ g commute inside fLg, the subgroups gLI G and ]LJ G commute in gLG, and the subalgebras AG,k(I) 46 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES and AG,k(J) commute in B(H0,k). The covariance axiom holds because the ac- tion of ϕ ∈ Conf+(S1) conjugates gLI G into ^Lϕ(I)G. Positive-energy follows from Lemma 4.15. Finally, by the classification of unitary positive-energy representa- tions of PSL2(R) ∼= Conf +(S1), any vector that is fixed by S1 ⊂ Conf+(S1) is actually fixed by the whole group Conf+(S1). In particular, Ω is a fixed vector for Conf+(S1). Theorem 4.18. The positive-energy net (AG,k, H0,k) satisfies strong additivity, the split property, and diffeomorphism covariance. Moreover, if G = SU(n), then the conformal net associated (by Proposition 4.9) to the positive energy net (ASU(n),k, H0,k) has finite index (Definition 3.1). Proof. Strong additivity follows from Proposition 4.16 and the split property follows from Lemma 4.15 and Theorem 4.10. We now check diffeomorphism covariance. p −→ Diff +(S1) → Aut(LG) gDiff Let p : gDiff → Diff +(S1) be the central extension by S1 pulled back along (4.13), and let q : Aut(LG) → Aut(gLG) be the isomorphism given by the functoriality of universal central extensions. The semidirect productgLG ⋊gDiff for the action conjugates gLIG into ^Lϕ(I)G, and therefore AG,k(I) into AG,k(ϕ(I)). Indeed, at the Lie algebra level, the action of ϕ ∈ Diff +(S1) on fLg = Lg ⊕ R is simply given by then acts on H0,k. One first observes that the action (4.13) of a diffeomorphism ϕ ϕ·(f, a) = (f◦ϕ−1, a). Let now I ⊂ S1 be an interval, and let ϕ be a diffeomorphism with support in I. Denote by I ′ the closure of S1 \ I. By Haag duality for positive- energy nets [19, Thm. 2.19.(ii)], in order to show that uϕ ∈ AG,k(I), it is enough to argue that it commutes with AG,k(I ′). Equivalently, we have to show that the the action of p( ϕ) = ϕ on ]LI ′G is trivial. The last statement can be verified at the Lie algebra level. chosen lift ϕ ∈ gDiff of ϕ commutes with ]LI ′G inside the groupgLG ⋊gDiff, i.e., that q −→ Aut(gLG) Finally, building on work of Wassermann [39], Feng Xu proves in [42] that (cid:3) (ASU(n),k, H0,k) has finite index. Remark 4.19. For other compact, simple, simply-connected Lie groups, it is ex- pected that the conformal nets AG,k have finite index, as is known to be the case for G = SU(n). It is also expected that the category of positive energy representa- tions of LG at level k is equivalent to the category of sectors for the corresponding conformal net.6 However, as far as we know, those problems are still open. The main theorem of Wassermann [39, p.535] (combined with [29, Theorem 4.1]) shows that the category of positive energy representations of L SU(n) at level k is a fusion category. Therefore, assuming that the category of positive energy representations of L SU(n) at level k is equivalent to the category of sectors for the corresponding conformal net, it would follow that the category Sect(ASU(n),k) is fusion. An application of Theorem 3.33 would then yield an alternative proof that ASU(n),k has finite index. Appendix von Neumann algebras. Given a Hilbert space H, let B(H) denote its algebra of bounded operators. The ultraweak topology on B(H) is the topology of pointwise convergence with respect to the pairing with its predual, the trace class operators. 6We have been informed that this will, in fact, be a consequence of ongoing work of Carpi- Weiner; see also [41]. For the case G = SU(N ), this result is a consequence of [42, Thm. 2.2], [27, Thm. 33], and [42, Thm. 3.5]. CONFORMAL NETS I: COORDINATE-FREE NETS 47 Definition A.1. A von Neumann algebra, is a topological *-algebra7 that is em- beddable as closed subalgebra of B(H) with respect to the ultraweak topology. The spatial tensor product A1 ¯⊗A2 of von Neumann algebras Ai ⊂ B(Hi) is the ultraweak closure in B(H1 ⊗ H2) of their algebraic tensor product A1 ⊗alg A2. Definition A.2. Let A be a von Neumann algebra. A left (right) A-module is a Hilbert space H equipped with a continuous homomorphism from A (respectively Aop) to B(H). We will use the notation AH (respectively HA) to denote the fact that H is a left (right) A-module. The main distinguishing feature of the representation theory of von Neumann algebras is expressed in the following lemma. Here, ℓ2 stands for the Hilbert space ℓ2(N) if all the spaces in the statement of the lemma are separable. Otherwise, it stands for ℓ2(X), where X is any set of sufficiently large cardinality. Lemma A.3. Let A be a von Neumann algebra and let H and K be two faithful left A-modules. Then H ⊗ ℓ2 ∼= K ⊗ ℓ2. In particular, any A-module is isomorphic to a direct summand of H ⊗ ℓ2. The Haagerup L2-space. (See [4, §2] for further details.) A faithful left module H for a von Neumann algebra A is called a standard form if it comes equipped with an antilinear isometric involution J and a selfdual cone P ⊂ H subject to the properties (i) JAJ = A′ on H, (ii) JcJ = c∗ for all c ∈ Z(A), (iii) Jξ = ξ for all ξ ∈ P , (iv) aJaJ(P ) ⊆ P for all a ∈ A where A′ denotes the commutant of A. The standard form is unique up to unique unitary isomorphism [24]. It is an A-A-bimodule, with right action ξa := Ja∗Jξ. The space of continuous linear functionals A → C forms a Banach space A∗ = L1(A) called the predual of A. +(A) := {φ ∈ A∗ φ(x) ≥ 0 ∀x ∈ A+} and two commuting A-actions given by (aφb)(x) := φ(bxa). Given a von Neumann algebra A, its Haagerup L2-space is an A-A-bimodule that is canonically associated to A [28]. It is the completion of It comes with a positive cone L1 C√φ Mφ∈L1 +(A) := {√φ φ ∈ L1 +(A) with respect to some pre-inner product, and is denoted L2(A). The positive cone +(A)}. The space L2A is also equipped in L2A is given by L2 with the modular conjugation JA that sends λ√φ to ¯λ√φ for λ ∈ C, and satisfies JA(aξb) = b∗JA(ξ)a∗. +(A)) is a standard form for the von Neumann (A.4) All together, the triple (L2(A), JA, L2 algebra A. Remark A.5. There is an isomorphism L2(A) ∼= L2(Aop) under which the left action of A on L2A corresponds to the right action of Aop on L2(Aop ), and the right action of A on L2A corresponds to the left action of Aop on L2(Aop). Remark A.6. The assignment A 7→ L2(A) defines a functor from the category of factors and isomorphisms, to the category of Hilbert spaces and bounded linear maps. (This still true for the larger category whose morphisms are finite homomor- phisms between factors [4, Thm 6.7], but in the present paper we only need this functor for isomorphisms.) 7Warning: there is no compatibility between the topology and the algebra structure; the mul- tiplication map (B(H), ultraweak) × (B(H), ultraweak) → (B(H), ultraweak) is not continuous. 48 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES Connes fusion. (See [4, §3] for further details.) Definition A.7. Given two modules HA and AK over a von Neumann algebra A, their Connes fusion H ⊠A K is the completion [12, 33, 39] of (A.8) Hom(cid:0)L2(A)A, HA(cid:1) ⊗A L2(A) ⊗A Hom(cid:0)AL2(A), AK(cid:1) with respect to the inner product(cid:10)φ1⊗ξ1⊗ψ1, φ2⊗ξ2⊗ψ2(cid:11) :=(cid:10)(φ∗ 2φ1)ξ1(ψ1ψ∗ Here, we have written the action of ψi on the right, which means that ψ1ψ∗ for the composite L2(A) ψ∗ 2 ), ξ2(cid:11). 2 stands ψ1−−→ K 2−−→ L2(A). The L2 space is a unit for Connes fusion in the sense that there are canonical unitary isomorphisms AL2(A) ⊠A H ∼= AH (A.9) Dualizability. (See [4, §4] for further details.) A von Neumann algebra whose center is C is called a factor. H ⊠A L2(A)A ∼= HA. and Definition A.10. For A and B factors, given an A-B-bimodule H, we say that a B-A-bimodule ¯H is dual to H if it comes equipped with maps (A.11) subject to the duality equations (R∗ ⊗ 1)(1 ⊗ S) = 1, (S∗ ⊗ 1)(1 ⊗ R) = 1, and to the normalization R∗(x ⊗ 1)R = S∗(1 ⊗ x)S for all x ∈ End(AHB). A bimodule whose dual module exists is called dualizable. S : BL2(B)B → B ¯H ⊠A HB R : AL2(A)A → AH ⊠B ¯HA If AHB is a dualizable bimodule, then its dual bimodule is well defined up to canonical unitary isomorphism [4, Thm 4.22]. Moreover, the dual bimodule is canonically isomorphic to the complex conjugate Hilbert space H, with the actions b ¯ξa := a∗ξb∗ [4, Cor 6.12]. Lemma A.12 ([4, Lemma 4.6]). Let AHB and BKA be dualizable irreducible bi- modules. Then HomA,A(cid:0)H ⊠B K, L2(A)(cid:1) =(C if BKA ∼= B ¯HA otherwise 0 Lemma A.13 ([4, Lemma 4.10]). If AHB is a dualizable bimodule, then its algebra of A-B-bilinear endomorphisms is finite-dimensional. Statistical dimension and minimal index. (See [4, §5] for further details.) Definition A.14. The statistical dimension of a dualizable bimodule AHB is given by dim(AHB) := R∗R = S∗S ∈ R≥0 where R and S are as in (A.11). For non-dualizable bimodules, one declares dim(AHB) to be ∞. Note that from the definition, it is obvious that dim(AHB) = dim(B ¯HA). (A.15) The minimal index [B : A] of an inclusion of factors ι : A → B is the square of the statistical dimension of AL2BB. If AHB is a faithful bimodule between factors, then we have [B′ : A] = [A′ : B] = dim(AHB)2. Definition A.16. Let ι : A → B be an inclusion of factors. If the minimal index [B : A] is finite, we say that ι is a finite homomorphism. As a corollary of Lemma A.13, we have: Lemma A.17 ([4, Lemma 5.15]). Let ι : A → B be a finite homomorphism between factors. Then the relative commutant of ι(A) in B is finite-dimensional. CONFORMAL NETS I: COORDINATE-FREE NETS 49 Haagerup's u-topology. Given von Neumann algebras A and B, the u-topology on Hom(A, B) is defined by declaring that a generalized sequence {ϕi} converges to ϕ ∈ Hom(A, B) if for every ξ ∈ L1(B), we have limi(ξ ◦ ϕi) = ξ ◦ ϕ in L1(A). Equivalently, it is the topology generated by the semi-norms a∈A,kak≤1ξ(ϕ(a)) ϕ 7→ kξ ◦ ϕkL1(A) = sup The subgroup N(A) := {u ∈ U(L2(A)) uAu∗ = A} ⊂ U(L2(A)) is closed for the for ξ ∈ L1(B). strong (= weak) topology on U(L2(A)). Lemma A.18. The adjoint map Ad : N(A) → Aut(A), Ad(u)(a) = uau∗, is con- tinuous for the strong topology on N(A) and the u-topology on Aut(A). Proof.8 Given ξ ∈ L1 +(A), we need to show that fξ : N(A) → C u 7→ sup a∈A,kak≤1ξ(uau∗) Therefore, given a in the unit ball of A, we have is continuous for the strong topology on N(A). Let un → u be a convergent sequence in N(A). For every v ∈ N(A), we have ξ(vav∗) =(cid:10)vav∗√ξ,√ξ(cid:11)L2(A) =(cid:10)av∗√ξ, v∗√ξ(cid:11)L2(A). n) − ξ(uau∗) (cid:12)(cid:12)ξ(unau∗ = hau∗ ≤ 2 · k√ξkL2(A) · k(un − u)∗√ξkL2(A). n → u∗ in the strong topology, we have limn k(un − u)∗√ξkL2(A) = 0, and n) converges to ξ(uau∗) uniformly in the unit ball of A. n√ξ, (un − u)∗√ξiL2(A) + ha(un − u)∗√ξ, u∗√ξiL2(A)(cid:12)(cid:12) (cid:3) Since u∗ so ξ(unau∗ The functoriality of L2 yields a map Aut(A) → N(A), ψ 7→ L2(ψ). Haagerup calls this the canonical implementation. He also shows [24, Prop. 3.5] that it exhibits Aut(A) with the u-topology as a closed subgroup of N(A) with the strong topology. Proposition A.19. Let A0 ⊆ A be a subfactor. Assume that the action of the algebraic tensor product A0⊗alg A′ on L2(A) extends to the spatial tensor product A0 ¯⊗ A′. Then Ad : U(A0) → Aut(A) induces a homeomorphism from PU(A0) with the quotient strong topology onto its image in Aut(A) with the u-topology. Proof. Recall that the canonical implementation ψ 7→ L2(ψ) identifies Aut(A) with a closed subgroup of U(L2(A)). For u ∈ U(A), the canonical implementation of Ad(u) ∈ Aut(A) is L2(Ad(u)) = uJu∗J, where J is the modular conjugation on L2(A). Thus it suffices to show that f : U(A0) → U(L2(A)), u 7→ uJu∗J The map ¯f is well-defined, bijective, and continuous by Lemma A.18. descends to a homeomorphism ¯f : PU(A0) → f (U(A0)). It re- mains to show that ¯f −1 is continuous. Let un and u be elements of U(A0) so that limn f (un) = f (u) in U(L2(A)). We would like to know that limn[un] = [u] in PU(A0). For that, it is enough to show that there exist λn ∈ S1 such that limn λnun = u in U(A0). 8Courtesy of http://mathoverflow.net/questions/87324/ 50 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES Since the algebra generated by A0 and A′ on L2(A) is their spatial tensor product, nJ and u ⊗ Ju∗J of we can identify f (un) and f (u) with the elements un ⊗ Ju∗ U(A0 ¯⊗A). The existence of the λn then follows from Lemma A.20. Lemma A.20. Let A and B be factors. Let {un} be a sequence in U(A), and {vn} a sequence in U(B) so that (cid:3) lim n (un ⊗ vn) = u ⊗ v in U(A ¯⊗ B) n vn = v. for given u ∈ U(A) and v ∈ U(B). Then there exist λn ∈ S1 so that limn λnun = u and limn λ−1 Proof. The identity limn(un⊗vn) = u⊗v is equivalent to limn(u∗un⊗v∗vn) = 1⊗1, so we may assume that u = 1 and v = 1. Pick a faithful representation H of A, and a unit vector ξ ∈ H. Replacing un by λnun and vn by λ−1 n vn for appropriate λn ∈ S1, we may also assume that hunξ ξi ≥ 0. Denote by A1 and B1 the unit balls in A and B. These are compact in the weak (= ultraweak) topology, and so it is enough to show that the limit of any weakly convergent subsequence of {un} is equal to 1, and the same for {vn}. We can therefore assume that u := limn un and and v := limn vn exist. The product map A1×B1 → (A ¯⊗ B)1 is continuous for the weak topology.9 It follows that u ⊗ v = 1 ⊗ 1, and so u = λ and v = λ−1 for some λ ∈ S1. Finally, limnhunξ ξi = huξ ξi = λ is positive, and so λ = 1. (cid:3) Given two von Neumann algebras A and B, recall that HomVN(A, B) denotes the space of homomorphisms and antihomomorphisms from A to B. That set is topologized as the disjoint union of Hom(A, B) and Hom(A, Bop), where both of those Hom sets are given the Haagerup u-topology. References [1] F. A. Bais and P. G. Bouwknegt. A classification of subgroup truncations of the bosonic string. Nuclear Phys. B, 279(3-4):561–570, 1987. [2] V. Bargmann. On unitary ray representations of continuous groups. Ann. of Math. (2), 59:1– 46, 1954. [3] A. Bartels, C. L. Douglas, and A. Henriques. Conformal nets II: Conformal blocks; Con- formal nets III: Fusion of defects; Conformal nets IV: The 3-category; Conformal nets V: Dualizability. In preparation. [4] A. Bartels, C. L. Douglas, and A. Henriques. Dualizability and index of subfactors. arXiv:1110.5671, 2011. [5] H.-J. Borchers. The CPT-theorem in two-dimensional theories of local observables. Comm. Math. Phys., 143(2):315–332, 1992. [6] R. Brunetti, K. Fredenhagen, and R. Verch. The generally covariant locality principle-a new paradigm for local quantum field theory. Comm. Math. Phys., 237(1-2):31–68, 2003. Dedicated to Rudolf Haag. [7] R. Brunetti, D. Guido, and R. Longo. Modular structure and duality in conformal quantum field theory. Comm. Math. Phys., 156(1):201–219, 1993. [8] D. Buchholz, C. D'Antoni, and K. Fredenhagen. The universal structure of local algebras. Comm. Math. Phys., 111(1):123–135, 1987. [9] D. Buchholz, G. Mack, and I. Todorov. The current algebra on the circle as a germ of local field theories. Nuclear Phys. B Proc. Suppl., 5B:20–56, 1988. Conformal field theories and related topics (Annecy-le-Vieux, 1988). [10] D. Buchholz and H. Schulz-Mirbach. Haag duality in conformal quantum field theory. Rev. Math. Phys., 2(1):105–125, 1990. [11] S. Carpi and M. Weiner. On the uniqueness of diffeomorphism symmetry in conformal field theory. Comm. Math. Phys., 258(1):203–221, 2005. [12] A. Connes. G´eom´etrie non commutative. InterEditions, Paris, 1990. 9This fails if one replaces A ¯⊗ B by some other completion of A ⊗alg B. CONFORMAL NETS I: COORDINATE-FREE NETS 51 [13] J. Dixmier. von Neumann algebras, volume 27 of North-Holland Mathematical Library. North-Holland Publishing Co., Amsterdam, 1981. With a preface by E. C. Lance, Trans- lated from the second French edition by F. Jellett. [14] S. Doplicher, R. Haag, and J. E. Roberts. Fields, observables and gauge transformations. I. Comm. Math. Phys., 13:1–23, 1969. [15] S. Doplicher, R. Haag, and J. E. Roberts. Fields, observables and gauge transformations. II. Comm. Math. Phys., 15:173–200, 1969. [16] S. Doplicher and R. Longo. Standard and split inclusions of von Neumann algebras. Invent. Math., 75(3):493–536, 1984. [17] C. L. Douglas and A. Henriques. Internal bicategories. arXiv:1206:4284, 2012. [18] K. Fredenhagen and M. Jorss. Conformal Haag-Kastler nets, pointlike localized fields and the existence of operator product expansions. Comm. Math. Phys., 176(3):541–554, 1996. [19] F. Gabbiani and J. Frohlich. Operator algebras and conformal field theory. Comm. Math. Phys., 155(3):569–640, 1993. [20] R. Goodman and N. R. Wallach. Structure and unitary cocycle representations of loop groups and the group of diffeomorphisms of the circle. J. Reine Angew. Math., 347:69–133, 1984. [21] D. Guido and R. Longo. The conformal spin and statistics theorem. Comm. Math. Phys., 181(1):11–35, 1996. [22] R. Haag. Local quantum physics. Texts and Monographs in Physics. Springer-Verlag, Berlin, 1992. Fields, particles, algebras. [23] R. Haag and D. Kastler. An algebraic approach to quantum field theory. J. Mathematical Phys., 5:848–861, 1964. [24] U. Haagerup. The standard form of von Neumann algebras. Math. Scand., 37(2):271–283, 1975. [25] V. G. Kac. Infinite-dimensional Lie algebras. Cambridge University Press, Cambridge, third edition, 1990. [26] Y. Kawahigashi and R. Longo. Classification of local conformal nets. Case c < 1. Ann. of Math. (2), 160(2):493–522, 2004. [27] Y. Kawahigashi, R. Longo, and M. Muger. Multi-interval subfactors and modularity of rep- resentations in conformal field theory. Comm. Math. Phys., 219(3):631–669, 2001. [28] H. Kosaki. Canonical Lp-spaces associated with an arbitrary abstract von Neumann algebra. Ph.D. thesis, UCLA, 1980. [29] R. Longo. Index of subfactors and statistics of quantum fields. II. Correspondences, braid group statistics and Jones polynomial. Comm. Math. Phys., 130(2):285–309, 1990. [30] R. Longo. Lectures on conformal nets II. http://www.mat.uniroma2.it/longo/Lecture%20Notes.html, 2008. [31] R. Longo and F. Xu. Topological sectors and a dichotomy in conformal field theory. Comm. Math. Phys., 251(2):321–364, 2004. [32] A. Pressley and G. Segal. Loop groups. Oxford Mathematical Monographs. The Clarendon Press Oxford University Press, New York, 1986. Oxford Science Publications. [33] J.-L. Sauvageot. Sur le produit tensoriel relatif d'espaces de Hilbert. J. Operator Theory, 9(2):237–252, 1983. [34] S. Stolz and P. Teichner. What is an elliptic object? In Topology, geometry and quantum field theory, volume 308 of London Math. Soc. Lecture Note Ser., pages 247–343. Cambridge Univ. Press, Cambridge, 2004. [35] S. Stolz and P. Teichner. Supersymmetric field theories and generalized cohomology. In Math- ematical foundations of quantum field theory and perturbative string theory, volume 83 of Proc. Sympos. Pure Math., pages 279–340. Amer. Math. Soc., Providence, RI, 2011. [36] T. Timmermann. An invitation to quantum groups and duality. EMS Textbooks in Mathe- matics. European Mathematical Society, Zurich, 2008. [37] V. Toledano Laredo. Fusion of positive energy representations of lspin2n. Ph.D. thesis, St. Johns College, Cambridge, 1997. [38] V. Toledano Laredo. Integrating unitary representations of infinite-dimensional Lie groups. J. Funct. Anal., 161(2):478–508, 1999. [39] A. Wassermann. Operator algebras and conformal field theory. III. Fusion of positive energy representations of LSU(N ) using bounded operators. Invent. Math., 133(3):467–538, 1998. [40] A. J. Wassermann. Operator algebras and conformal field theory. In Proceedings of the Inter- national Congress of Mathematicians, Vol. 1, 2 (Zurich, 1994), pages 966–979, Basel, 1995. Birkhauser. [41] M. Weiner. Conformal covariance and positivity of energy in charged sectors. Comm. Math. Phys., 265(2):493–506, 2006. [42] F. Xu. Jones-Wassermann subfactors for disconnected intervals. Commun. Contemp. Math., 2(3):307–347, 2000. 52 ARTHUR BARTELS, CHRISTOPHER L. DOUGLAS, AND ANDR ´E HENRIQUES WWU Munster, Mathematisches Institut, Einsteinstr. 62, 48149 Munster, Germany E-mail address: [email protected] URL: http://www.math.uni-muenster.de/u/bartelsa Mathematical Institute, 2429 St Giles', Oxford, OX1 3LB, United Kingdom E-mail address: [email protected] URL: http://people.maths.ox.ac.uk/cdouglas Mathematisch Instituut, Universiteit Utrecht, Postbus 80.010, 3508 TA Utrecht, The Netherlands E-mail address, Corresponding author: [email protected] URL: http://www.staff.science.uu.nl/~henri105
1301.6863	1	1301	2013-01-29T08:20:26	A Helson-Szeg\"o theorem for subdiagonal subalgebras with applications to Toeplitz operators	[ "math.OA", "math.FA" ]	We formulate and establish a noncommutative version of the well known Helson-Szego theorem about the angle between past and future for subdiagonal subalgebras. We then proceed to use this theorem to characterise the symbols of invertible Toeplitz operators on the noncommutative Hardy spaces associated to subdiagonal subalgebras.	math.OA	math	A HELSON-SZEG O THEOREM FOR SUBDIAGONAL SUBALGEBRAS WITH APPLICATIONS TO TOEPLITZ OPERATORS LOUIS E LABUSCHAGNE AND QUANHUA XU Abstract. We formulate and establish a noncommutative version of the well known Helson- Szego theorem about the angle between past and future for subdiagonal subalgebras. We then proceed to use this theorem to characterise the symbols of invertible Toeplitz operators on the noncommutative Hardy spaces associated to subdiagonal subalgebras. 1. Introduction Let T be the unit circle of the complex plane equipped with normalised Lebesgue measure dm. We denote by H p(T) the usual Hardy spaces on T. Let P+ be the orthogonal projection from L2(T) onto H 2(T). The classical Helson-Szego theorem [14] (see also [12, section IV.3]) characterises those positive measures µ on T such that P+ is bounded on L2(T, µ). The condition is that µ is absolutely continuous with respect to dm and the corresponding Radon-Nikod´ym derivative w satisfies (1.1) w = eu+ev for two functions u, v ∈ L∞(T) with kevk∞ < π/2, where ev denotes the conjugate function of v. The motivation of this theorem comes from univariate prediction theory. Let P+ denote the space of all polynomials in z, and P− the space of all polynomials in ¯z without constant term. P = P+ + P− is the space of all trigonometric polynomials. Then P+ is bounded on L2(T, µ) if and only if P+ and P− are at positive angle in L2(T, µ). Recall that the angle between P+ and P− is defined as arccos of the following quantity ρ = sup(cid:8)(cid:12)(cid:12)ZT f ¯gdµ(cid:12)(cid:12) : f ∈ P+, g ∈ P−, kf kL2(T,µ) = kgkL2(T,µ) = 1(cid:9). Thus P+ is bounded on L2(T, µ) if and only if ρ < 1. In multivariate prediction theory one needs to consider the matrix-valued extension of the Helson-Szego theorem. Let Mn denote the full algebra of complex n × n-matrices, equipped with the normalised trace tr. Let P+(Mn) denote the space of all polynomials in z with coefficients in Mn. P−(Mn) and P(Mn) have similar meanings. Let w be an Mn-valued weight on T, i.e. w is an integrable function on T with values in the family of semidefinite nonnegative matrices. For any trigonometric polynomials f and g in P(Mn) define hf, giw =ZT tr(g∗f w)dm and kf kw = hf, f i1/2 w , where a∗ denotes the adjoint of a matrix a. Like in the scalar case, define ρ = sup(cid:8)(cid:12)(cid:12)ZT tr(g∗f w)dm(cid:12)(cid:12) : f ∈ P+(Mn), g ∈ P−(Mn), kf kw = kgkw = 1(cid:9). Again, ρ < 1 if and only if P+ ⊗ IdMn is bounded on P(Mn) with respect to k kw. The problem here is, of course, to characterise w such that ρ < 1 in a way similar to the scalar case. This time the task is much harder, and it is impossible to find a characterisation as nice as (1.1). Numerous works have been devoted to this subject, see, for instance [2, 4, 10, 20, 22, 27]. In particular, Pousson's characterisation in [22] is the matrix-valued analogue of a key intermediate step to (1.1). It is strong enough for applications to the invertibility of Toeplitz operators. 2010 Mathematics Subject Classification. Primary: 46L52. Secondary: 47L38, 42C99. Key words and phrases. Helson-Szego theorem, angle between past and future, invertibility of Toeplitz operators, subdiagonal algebras, Hardy spaces, outer operators, Riesz-Szego factorisation. 1 2 LOUIS E LABUSCHAGNE AND QUANHUA XU The preceding two cases can be put into the more general setting of subdiagonal algebras in the sense of [1]. We will provide an extension of the Helson-Szego theorem in this general setting. This is the first objective of the paper. Our second objective is to study the invertibility of Toeplitz operators. It is well known that the Helson-Szego theorem is closely related to the invertibility of Toeplitz operators. This relationship was remarkably exploited by Devinatz [9]. Pousson [21, 22] then subsequently extended Devinatz's work to the matrix-valued case. Using our extension of the Helson-Szego theorem, we will char- acterise the symbols of invertible Toeplitz operators in the very general setting of subdiagonal algebras. We end this introduction by mentioning the link between the Helson-Szego theorem and Muck- enhoupt's A2 weights. Let w be a weight on T. Hunt, Muckenhoupt and Wheeden [15] proved that the Riesz projection P+ is bounded on L2(T, w) if and only if (1.2) sup 1 IZI w 1 IZI w−1 < ∞, where the supremum runs over all arcs of T. Such a w is called an A2-weight. Thus for a weight w the two conditions (1.1) and (1.2) are equivalent via the boundedness of the Riesz projection. It seems that it is still an open problem to find a direct proof of this equivalence. Hunt, Muckenhoupt and Wheeden's theorem was extended to the matrix-valued case by Treil and Volberg [27]. Namely, let w now be an Mn-valued weight on T. Then P+ ⊗ IdMn is bounded on P(Mn) with respect to k kw if and only if sup IZI I (cid:13)(cid:13)(cid:13)(cid:16) 1 IZI w(cid:17)1/2(cid:16) 1 IZI w−1(cid:17)(cid:16) 1 w(cid:17)1/2(cid:13)(cid:13)(cid:13)Mn < ∞. It is not clear for us how to extend Treil and Volberg's theorem to the case of subdiagonal algebras. On the other hand, Hunt, Muckenhoupt and Wheeden also characterised the boundedness of P+ on Lp(T, w) for any 1 < p < ∞ by the so-called Ap weights. A well known open problem in matrix-valued harmonic analysis is to extend this result to the matrix-valued case; even to the very general one of subdiagonal algebras. 2. Preliminaries Throughout the paper M will be a von Neumann algebra possessing a faithful normal tracial state τ . The associated noncommutative Lp-spaces are denoted by Lp(M). We refer to [19] for noncommutative integration. For a subset S of Lp(M), we will write [S]p for the closure of S in the Lp-topology. On the other hand, S∗ will denote the set of all Hilbert-adjoints of elements of S. When an actual Banach dual of some Banach space is in view, we will for the sake of avoiding confusion prefer the superscript ⋆ . For example the dual of M will be denoted by M⋆. Because M is finite, there will for any von Neumann subalgebra N of M, always exist a normal contractive projection Ψ : M → N satisfying τ ◦ Ψ = τ . This is the so-called normal faithful conditional expectation onto N with respect to τ . A finite subdiagonal algebra of M is a weak* closed unital subalgebra A of M satisfying the following conditions • A + A∗ is weak* dense in M; • the trace preserving conditional expectation Φ : M → A ∩ A∗ = D is multiplicative on A: Φ(ab) = Φ(a)Φ(b), a, b ∈ A. In this case, D is called the diagonal of A. We also set A0 = A ∩ Ker(Φ). In the sequel, A will always denote a finite subdiagonal algebra of M. Subdiagonal algebras are our noncommutative H ∞'s. The most important example is, of course, the classical H ∞(T) on the unit circle. Another example important for multivariate prediction theory is the matrix-valued H ∞(T). More precisely, let M = L∞(T) ⊗ Mn = L∞(T; Mn) equipped with the product trace, and let A = H ∞(T; Mn) -- the subalgebra of M consisting of n × n-matrices with entries in H ∞(T). Many classical results about Hardy spaces on T have been transferred to the matrix-valued case. A third example is the upper triangle subalgebra Tn of Mn. This example is closely related to the second one, and is a finite dimensional nest algebra. We refer to [19, §8] for A NONCOMMUTATIVE HELSON-SZEG O THEOREM 3 more information and historical references on subdiagonal algebras, in particular, on matrix-valued analytic functions. For p < ∞ the Hardy space H p(A) associated with a finite subdiagonal algebra A is defined to be [A]p. The closure of A0 in Lp(M) will be denoted by H p 0 (M). By convention, we put H ∞(A) = A and H ∞ 0 (A) = A0. These spaces exhibit many of the properties of classical H p spaces (see [3, 6, 7, 17, 24, 25]). In particular for 1 < p < ∞, Lp(M) appears as the Banach space direct sum of H p(M) and H p 0 (M) and Lp(D). In the case p = 2, these direct sums are even orthogonal direct sums. 0 (M)∗, with H p(M) appearing as the Banach space direct sum of H p Recall that if a weight w on T satisfies (1.1), then necessarily log w ∈ L1(T), or equivalently, (2.1) exp(cid:0)ZT log w(cid:1) > 0. The integrability of log w is also equivalent to the existence of an outer function h ∈ H 1(T) such that w = h. To state the outer-inner factorisation and prove the Helson-Szego analogue for subdiagonal algebras, we need an appropriate substitute of the latter condition. This is achieved by the Fuglede-Kadison determinant. Recall that the Fuglede-Kadison determinant ∆(a) of an operator a ∈ Lp(M) (p > 0) can be defined by ∆(a) = exp(cid:0)τ (log a)(cid:1) = exp(cid:0)Z ∞ 0 log t dνa(t)(cid:1), where dνa denotes the probability measure on R+ which is obtained by composing the spectral measure of a with the trace τ . It is easy to check that ∆(a) = lim p→0 kakp and ∆(a) = inf ǫ>0 exp τ (log(a + ǫ1)) . As the usual determinant of matrices, ∆ is also multiplicative: ∆(ab) = ∆(a)∆(b). We refer the reader for information on determinant to [11, 1] in the case of bounded operators, and to [8, 13] for unbounded operators. Return to our Hardy spaces. An element h of H p(M) with p < ∞ is said to be an outer element if hA is dense in H p(M). If in addition ∆(h) > 0, we call such an h strongly outer. For an analysis of outer elements in the present context, we refer the interested reader to [6] for p ≥ 1 and [3] for p < 1. We will however pause to summarise the essential points of the theory. For any outer element h of H p(M), both h and Φ(h) necessarily have dense range and trivial kernel. Hence their inverses exist as affiliated operators. For such an outer element, we also necessarily have that ∆(h) = ∆(Φ(h)). If indeed ∆(h) > 0, the equality ∆(h) = ∆(Φ(h)) is sufficient for h to be outer. Using this fact it is now an easy exercise to see that if ∆(h) > 0, then h is an outer element of H p(M) if and only if h∗ is an outer element of H p(M)∗ if and only if h is right outer in the sense that Ah will also be dense in H p(M). In this theory one also has a type of noncommutative Riesz-Szego theorem, in that any f ∈ Lp(M) for which ∆(f ) > 0, may be written in the form f = uh where u ∈ M is unitary and h ∈ H p(M) an outer element of H p(M). Given a state ω on M, we write (πω, L2(ω), Ωω) for the cyclic representation associated to ω. The subspaces A∗ and A0 embed canonically into L2(ω) by means of the operation a 7→ πω(a)Ωω. The angle between A∗ and A0 in L2(ω) is defined to be that between the closed subspaces πω(A∗)Ωω and πω(A0)Ωω. The latter is equal to arccos ρ with ρ given by ρ = sup{hπω(a)Ωω, πω(b)Ωωi : a ∈ A0, b ∈ A∗, kπω(a)Ωωk ≤ 1, kπω(b)Ωωk ≤ 1}. In view of the fact that hπω(a)Ωω, πω(b)Ωωi = ω(b∗a), this may be rewritten as ρ = sup{ω(b∗a) : a ∈ A0, b ∈ A∗, ω(a2) ≤ 1, ω(b2) ≤ 1}. In general 0 ≤ ρ ≤ 1. A∗ and A0 are said to be at positive angle in L2(ω) if ρ < 1. Let P+ be the orthogonal projection from L2(M) onto H 2(M). It is then clear that P+ defines a bounded operator on L2(ω) if and only if ρ < 1. 3. A noncommutative Helson-Szego theorem In this section we present our noncommutative Helson-Szego theorem. This theorem will prove to be an important ingredient in our onslaught on Toeplitz operators in the next section. As re- called previously, the classical Helson-Szego theorem contains the information that any finite Borel measure for which the angle between A and A∗ 0 is positive must necessarily be absolutely continuous 4 LOUIS E LABUSCHAGNE AND QUANHUA XU with respect to Lebesgue measure, and moreover that the Radon-Nikod´ym derivative of this mea- sure must have a strictly positive geometric mean (2.1). Before presenting our noncommutative Helson-Szego theorem, we first show that under some mild restrictions the same claims are true in the noncommutative case. Lp +(M) will denote the positive cone of Lp(M). Proposition 3.1. Let D = A ∩ A∗ be finite dimensional, and let ω be a state on M for which ρ < 1. Then ω is of the form ω = τ (g·) for some g ∈ L1 +(M). Proof. We keep the notation introduced at the end of the previous section. Let ωn and ωs respectively be the normal and singular parts of ω. Firstly note that by [26, III.2.14], there exists a central projection e0 in πω(M)′′ such that for any ξ, ψ ∈ L2(ω) the functionals a 7→ hπω(a)e0ξ, ψi and a 7→ hπω(a)e⊥ 0 ξ, ψi on M are respectively the normal and singular parts of the functional a 7→ hπω(a)ξ, ψi, where e⊥ 0 = 1 − e0. In particular, the triples (e0πω, e0L2(ω), e0Ωω) and (e⊥ 0 Ωω) are copies of the GNS representations of ωn and ωs respectively. 0 L2(ω), e⊥ 0 πω, e⊥ Since ρ < 1, we must have that πω(A0)Ωω ∩ πω(A∗)Ωω = {0}. Now suppose that the singular part ωs of ω is nonzero. By Ueda's noncommutative peak-set theorem [28, Theorem 1] there exist an orthogonal projection e in the second dual M⋆⋆ of M and a contractive element a of A so that • an converges to e in the weak-topology on M⋆⋆; • ωs(e) = ωs(1) (here ωs is identified with its canonical extension to M⋆⋆); • an converges to 0 in the weak-topology on M. Since the expectation Φ is weak-continuous on M, Φ(an) is weak convergent to 0. But then the finite dimensionality of D ensures that Φ(an) converges to 0 in norm. Recall that the bidual M⋆⋆ of M may be represented as the double commutant of M in its universal representation. So when this realisation of M⋆⋆ is compressed to the specific representation engendered by ω, it follows that e yields a projectionee in πω(M)′′ to which πω(an) converges in the weak-topology on πω(M)′′. This weak convergence in πω(M)′′ together with the second bullet above, then yield the facts that • πω(an)Ωω converges to ee Ωω in the weak-topology on L2(ω); • hee Ωω, e⊥ 0 Ωωi = ωs(1). From the first bullet and the fact that {Φ(an)} is a norm-null sequence, it follows that πω(an − e in the weak-topology on M⋆⋆, then surely so does (a∗)n. In terms of the GNS representation Φ(an))Ωω is weakly convergent to e Ωω, and hence that ee Ωω ∈ πω(A0)Ωω. But if an converges to for ω, this means that πω((a∗)n)Ωω also converges to ee Ωω in the weak-topology on L2(ω). But then ee Ωω ∈ πω(A∗)Ωω. Then ee Ωω = 0 since ee Ωω ∈ πω(A0)Ωω ∩ πω(A∗)Ωω. But this cannot be, since by the second bullet this would mean that ωs(1) = hee Ωω, e⊥ 0 Ωωi = 0. Thus our supposition that ωs is nonz ero, must be false. The condition that ρ < 1, is therefore sufficient to force ω to be normal. That is ω is of the form ω = τ (g·) for some g ∈ L1 (cid:3) +(M). The following lemmata present two known elementary facts. Lemma 3.2. For any g ∈ L1 +(M) we have that s(Φ(g)) ≥ s(g), where s(g) denotes the support projection of g. Proof. For simplicity of notation we respectively write s and sΦ for s(g) and s(Φ(g)). Since sΦ ∈ D, we have that τ (s⊥ Φ gs⊥ Φ) = τ ◦ Φ(s⊥ Φ gs⊥ Φ) = τ (s⊥ Φ Φ(g)s⊥ Φ ) = 0. Therefore g1/2s⊥ that sΦ ≥ s. Φ = s⊥ Φ g1/2 = 0. This is sufficient to force s⊥ Φ ⊥ s, which in turn suffices to show (cid:3) Lemma 3.3. Let e be a nonzero projection in D. Then eAe is a finite maximal subdiagonal subalgebra of eMe (equipped with the trace τe(·) = 1 τ (e) τ (·)) with diagonal eAe ∩ (eAe)∗ = eDe. A NONCOMMUTATIVE HELSON-SZEG O THEOREM 5 Proof. The expectation Φ is trivially still multiplicative on the compression eAe. Using the fact that e ∈ D, it is an exercise to see that Φ maps eAe onto eDe. It is also straightforward to see that the weak-density of A + A∗ in M forces the weak*-density of eAe + (eAe)∗ in eMe, and that (eAe)0 = eA0e. (cid:3) Definition 3.4. Adopting the notation of the previous two lemmata, given a nonzero element g ∈ L1 +(M), we define ∆Φ(g) to be the determinant of sΦgsΦ regarded as an element of (sΦMsΦ, τsΦ ) Proposition 3.5. Let D = A ∩ A∗ be finite dimensional, and let g ∈ L1 for which the state ω = τ (g·) satisfies ρ < 1. Then ∆Φ(g) > 0. +(M) be a norm-one element Proof. It is clear from the previous lemmata that we may reduce matters to the case where s(Φ(g)) = 1, and hence we will assume this to be the case. Suppose by way of contradiction that ∆(g) = 0. By the Szego formula for subdiagonal algebras [16], we then have that 0 = ∆(g) = inf{τ (ga − d2) : a ∈ A0, d ∈ D, ∆(d) ≥ 1}. Thus there exist sequences {an} ⊂ A0 and {dn} ⊂ D with ∆(dn) ≥ 1 for all n, so that τ (gan − dn2) → 0 as n → ∞. {dn} ⊂ D+. nan − dn(cid:12)(cid:12)2 It is an exercise to see that then {u∗ By Lemma 2.2 of [5] we may assume all the dn's to be invertible. Now let un ∈ D be the unitary in the polar decomposition dn = undn. nan} ⊂ A0 with . Making the required replacements, we may therefore also assume that an − dn2 =(cid:12)(cid:12)u∗ Since 1 ≤ ∆(dn) ≤ kdnk∞ for all n, we will for the sequences fdn = 1 (n ∈ N), still have that τ (gfan −fdn2) → 0 as n → ∞. Now recall that D is finite dimensional. So by passing to a subsequence if necessary, we may assume that {fdn} converges uniformly to some norm one element d0 of D+. But then by what we showed above, kdnk∞ dn and fan = 1 kdnk∞ an kπg(fan) − π(d0)k2 = τ (gfan − d02)1/2 ≤ τ (gfan −fdn2)1/2 + τ (gfdn − d02)1/2 ≤ τ (gfan −fdn2)1/2 + kfdn − d0k∞τ (g)1/2 → 0. Thus πg(d0) ∈ πg(A0) ∩ πg(A∗). Since Φ(g) is of full support, we have that Φ(g)1/2d0Φ(g)1/2 6= 0. So 0 < τ (Φ(g)1/2d0Φ(g)1/2) = τ (Φ(g)d0) = τ (Φ(gd0)) = τ (gd0). Therefore πg(d0) 6= 0. But this proves that the subspaces πg(A0) and πg(A∗) have a nonzero intersection, and hence that ρ = 1. (cid:3) Remark 3.6. Under the assumption of the previous proposition, the support s(Φ(g)) can be strictly less than 1. Indeed, consider the M2-valued case: M = L∞(T; M2) and A = H ∞(T; M2). Let w be a weight satisfying (1.1) and g = w ⊗ e11, where e11 the matrix whose only nonzero entry is the one at the position (1, 1) which is equal to 1. Then the corresponding ρ is less than 1 but s(Φ(g)) = e11. The following technical lemma is a crucial step in the proof of the classical Helson-Szego the- orem. The challenge one faces in the noncommutative world is that the functional calculus at our disposal in that context is simply not strong enough to reproduce so detailed a statement in that framework. However in the lemma following this one, we present what we believe to be a reasonable noncommutative substitute of this interesting lemma. Lemma 3.7. Let u = e−iψ with ψ a real measurable function on T. Then inf g∈H∞(T) ke−iψ−gk∞ < 1 if and only if there exist an ǫ > 0 and a k0 ∈ H ∞(T) so that k0 ≥ ǫ and ψ + arg(k0) ≤ π 2 − ǫ almost everywhere . Lemma 3.8. Let u be a unitary element of M. Then there exists some f ∈ A so that ku − f k∞ < 1 if and only if there exists h ∈ A so that ℜ(u∗h) is strictly positive. 6 LOUIS E LABUSCHAGNE AND QUANHUA XU Proof. Suppose first that there exists f ∈ A with ku − f k∞ < 1. We then equivalently have that k1 − u∗f k = k1 − f ∗uk < 1. On setting α = k1 − u∗f k, it follows that k1 − ℜ(u∗f )k ≤ α < 1, and hence that This in turn ensures that 0 < (1 − α)1 ≤ ℜ(u∗f ). −α1 ≤ ℜ(u∗f ) − 1 ≤ α1. Conversely suppose that there exists h ∈ A ∩ M−1 so that ℜ(u∗h) ≥ α1 for some 0 < α ≤ kℜ(u∗h)k ≤ khk, where M−1 denotes the subset of invertible elements of M. Given ǫ > 0, set λ = ǫ khk . It then follows that −2λℜ(u∗h) + λ2h2 ≤ −(cid:16) 2αǫ khk − ǫ2(cid:17) 1. (Observe that α that 1 > (cid:16) 2αǫ now set δ =(cid:16) 2αǫ khk ≤ 1 in the above inequality.) It is clear that if ǫ is small enough, we would have khk − ǫ2(cid:17) > 0. Thus we may assume this to be the case. For simplicity of notation we khk − ǫ2(cid:17). It therefore follows from the previous centered inequality that 0 ≤ 1 − u∗(λh)2 = 1 − 2ℜ(u∗(λh) + λh2 ≤ (1 − δ)1. Hence as required, k1 − u∗(λh)k2 ≤ (1 − δ) < 1. (cid:3) We are now finally ready to present our noncommutative Helson-Szego theorem. In view of Propositions 3.1 and 3.5, it is not unreasonable to restrict attention to normal states τ (g·) in this theorem for which ∆Φ(g) > 0. The following result is a sharpening of the result of Pousson [22, Theorem 4.3], in that here the conditions imposed on the unitary u are less restrictive. This sharpening is achieved by means of the preceding Lemma. Theorem 3.9. Let g ∈ L1 state ω = τ (g·). Then ρ < 1 and ∆Φ(g) > 0 if and only if g is of the form g = fRufL where +(M) be given with kgk1 = 1, and denote s(Φ(g)) by sΦ. Consider the • u ∈ M is a partial isometry with initial and final projections sΦ for which there exists some k ∈ sΦAsΦ so that ℜ(u∗k) ≥ αsΦ for some α > 0, • and fL and fR are strongly outer elements of H 2(M) commuting with sΦ for which g + (1 − sΦ) = fL2 = f ∗ R2. If in addition dim D < ∞, we may dispense with the restrictions that ω is normal, and that ∆Φ(g) > 0. Proof. Set s = sΦ for simplicity. Suppose that g satisfies the condition ∆Φ(g) > 0. Using the fact that then ∆Φ(g1/2) = ∆Φ(g)1/2 > 0, it follows from the noncommutative Riesz-Szego theorem (see [6]) that there exist strongly outer elements hL, hR ∈ H 2(sMs) and unitaries vL, vR ∈ sMs for which g1/2 = vLhL = hRvR. (Then also g1/2 = hL = h∗ u = vRvL, fL = hL + s⊥, R.) We set fR = hR + s⊥. It is then clear that g = fRufL and g + s⊥ = fL2 = f ∗ R2. We proceed to show that fL and fR are strongly outer. The proofs of the two cases are identical, and hence we do this for fL only. Notice that log(fL) = log(hL + s⊥) = log(hL)s. Since Φ(fL) = Φ(hL) + s⊥, we similarly have that log(Φ(fL)) = log(Φ(hL))s. It then follows that τ (log fL) = τ (s)τs(log hL) and τ (log Φ(fL)) = τ (s)τs(log Φ(hL)). Thus the outerness of hL yields that τ (log fL) = τ (log Φ(fL)) > −∞, so ∆(fL) = ∆(Φ(fL)) > 0. Then an application of [6, Theorem 4.4] now shows that fL is strongly outer. On the other hand, we have hπg(a)Ωg, π(b)Ωgi = τ (gb∗a) = τ (ufLb∗afR), a ∈ A0, b ∈ A∗. A NONCOMMUTATIVE HELSON-SZEG O THEOREM 7 So ρ = sup{τ (gb∗a) : a ∈ A0, b ∈ A∗, τ (ga2) ≤ 1, τ (gb2) ≤ 1} = sup{τ(cid:0)(u(sfLb∗)(afRs)(cid:1) : a ∈ A0, b ∈ A∗, τ (afRs2) ≤ 1, τ (bf ∗ = sup{τ (uF1F2) : F1 ∈ sH 2(M), F2 ∈ H 2 0 (M)s, kF1k2 ≤ 1, kF2k2 ≤ 1}. Ls2) ≤ 1} In the above computation one has used the fact that fL and fR are strongly outer to approximate F1 and F2 with elements of the form sfLb∗ and afRs where a ∈ A0 and b ∈ A∗. However, it is easy to check that for F1 ∈ sH 2(M), F2 ∈ H 2 0 (M)s F1F2 ∈ H 1 0 (sMs) and kF1F2k1 ≤ kF1k2kF2k2. Conversely, by the Noncommutative Riesz Factorisation theorem [17, 25], for any ǫ > 0 and any F ∈ H 1 0 (sMs) there exist F1 ∈ H 2(sMs) ⊂ sH 2(M) and F2 ∈ H 2 0 (sMs) ⊂ H 2(M)s such that From these discussions we conclude that F = F1F2 and kF1k2kF2k2 ≤ kF k1 + ǫ. ρ = sup{τ (uF ) : F ∈ H 1 = sup{τs(uF ) : F ∈ H 1 0 (sMs), kF k1 ≤ 1} 0 (sMs), τs(F ) ≤ 1}. The norm of the restriction of the functional L1(sMs) → C : a 7→ τs(ua) to H 1 precisely the norm of the equivalence class [u] in the quotient space sMs/(H 1 is well known that 0 (sMs) is by duality 0 (sMs))◦. However, it (cf. e.g., [25] ). From this fact it is now an easy exercise to see that the polar (H 1 nothing but sAs. It therefore follows that 0 (sMs))◦ is sAs = {a ∈ sMs : τs(ab) = 0, b ∈ sA0s} The result now follows from an application of the preceding Lemma. (cid:3) ρ = inf{ku − kk∞ : k ∈ sAs}. 4. Invertibility of Toeplitz operators We start by recalling the definition of Toeplitz operators. Given a ∈ M, the Toeplitz operator Ta with symbol a is defined to be the map Ta : H 2(M) → H 2(M) : b 7→ P+(ab), where P+ denotes the orthogonal projection from L2(M) onto H 2(M). Our basic reference for Toeplitz operators in this context is [18] (see also [23]). We will characterise the symbols of invertible Toeplitz operators. We point out that these results are new even for the matrix-valued case. In achieving this characterisation, we will follow the same basic strategy as Devinatz [9] in his remarkable solution of this problem in the classic setting. Our first result essentially reduces the problem to that of characterising invertible Toeplitz operators with unitary symbols. Theorem 4.1. Let a ∈ M be given. A necessary and sufficient condition for Ta to be invertible is that it can be written in the form a = uk where k ∈ A−1, and u ∈ M is a unitary for which Tu is invertible. Suppose that a ∈ M is indeed of the form a = uk where k ∈ A−1, and u ∈ M is a unitary. It is a simple exercise to see that then Tk is invertible with inverse Tk−1 . Since TaTk−1 = Tu and TuTk = Ta, it is now clear that Ta will then be invertible if and only if Tu is invertible. Proof. The sufficiency of the stated condition was noted in the above discussion. To see the necessity, assume Ta to be invertible. There must therefore exist some g ∈ H 2(M) so that Tag = 1. This in turn can only be true if there exists some h ∈ H 2 0 (M) so that ag = 1 + h∗. By the generalised Jensen inequality [6, 3.3] we have that ∆(a)∆(g) = ∆(ag) = ∆(1 + h∗) ≥ ∆(Φ(1 + h∗)) = ∆(1) = 1. Clearly we then have that ∆(a1/2) = ∆(a)1/2 > 0. So by the noncommutative Riesz-Szego theorem [6, 4.14], there must exist an outer element f ∈ H 2(M) and a unitary v so that a1/2 = vf . (Note then that f ∈ M, so f must belong to A too.) Let w be the unitary in the polar decomposition 8 LOUIS E LABUSCHAGNE AND QUANHUA XU a = wa, and consider b = wa1/2v. Notice that by construction bf = a. Thus TbTf = Ta. We will use this formula to show that Tf is invertible, from which the result will then follow. Firstly note that the injectivity of Ta combined with the above equality, ensures that Tf is injective. Next notice that the equality TbTf = Ta ensures that (Ta)−1Tb is a left inverse for Tf . So Tf must have a closed range. However since f is outer, we also have that [f A]2 = H2(M). Since f A ⊂ Tf (H2(M)), these two facts ensure that the range of Tf is all of H2(M). Hence Tf must be invertible. But if Tf is invertible, then so is T ∗ f = Tf ∗. Since Tf ∗ Tf = Tf 2 = Ta, the operator Ta must be invertible. Since σ(a) ⊂ σ(Ta) by Theorem 3.5 of [18], we must have that 0 6∈ σ(a). In other words a must be strictly positive. But if a is strictly positive, then by Arveson's factorisation theorem there exists some k ∈ A−1 with a = k. Finally let w0 be the unitary in the polar form k = w0k. Then a = ww∗ (cid:3) 0 k, which proves the theorem with u = ww∗ 0. Our next step in achieving the desired characterisation, is to present some necessary struc- tural information regarding unitaries u for which Tu is invertible. We then subsequently use this structural information to obtain a characterisation of invertibility in terms of positive angle. 1 )−1dg−1 Lemma 4.2. Let u ∈ M be a unitary. A necessary condition for Tu to be invertible is that it is of the form u = (g∗ 0 where g0, g1 are strongly outer elements of H 2(M) and d a strongly outer element of L2(D) related by the conditions that 0 , d∗g−1 g∗ 0g0 = d∗(g∗ Proof. Let u ∈ M be a unitary for which Tu is invertible. Since T ∗ u = Tu∗ is then also invertible, it follows that there must exist g0, g1 ∈ H 2(M) so that Tug0 = 1 = Tu∗g1. This in turn means that there exist h0, h1 ∈ H 2 1 ∈ H 2(M) and d = Φ(g0) = Φ(g∗ 1g1)−1d. dg−1 1), 0 (M) with ug0 = 1 + h∗ 0, u∗g1 = 1 + h∗ 1. Notice that we may then apply the generalised Jensen inequality [6, 3.3] to conclude that ∆(g0) = ∆(u)∆(g0) = ∆(ug0) ≥ ∆(1) = 1. Similarly ∆(g1) ≥ 1. By [6, 4.2 & 4.15] this means that both g0 and g1 are injective with dense range, and hence that g−1 exist as affiliated operators. On the other hand, we have that 0 1(1 + h∗ 1ug0 = g∗ g∗ 1 0) ∈ H 1(M)∗ 1) ∈ H 1(M)∗. 0u∗g1 = g∗ 0(1 + h∗ and g∗ and g−1 Hence g∗ 1ug0 ∈ H 1(M) ∩ H 1(M)∗ = L1(D). If we denote this element by d, it follows that u is of the form u = (g∗ that d∗(g∗ 1g1)−1d = g∗ 0g0. It remains to show that g0 and g1 are outer and that d = Φ(g0) = Φ(g∗ 1)−1d g−1 0 . It is then clear 1). To see this notice that since g∗ 1 ∈ H 2(M)∗ and ug0 = 1 + h∗ d = Φ(d) = Φ(g∗ 0 ∈ H 2(M)∗, we have that 0)) = Φ(g∗ 1(1 + h∗ 1ug0) = Φ(g∗ 1)Φ(1 + h∗ 0) = Φ(g∗ 1). Similarly, d = Φ(g0). (Since Φ maps H 2(M ) onto L2(D), this equality also shows that d is in fact in L2(D), and not just L1(D).) It now follows from the equality g∗ 0g0 = d∗(g∗ 1g1)−1d, that ∆(g0)2 = ∆(g∗ 0g0) = ∆(d∗(g∗ 1g1)−1d) = ∆(d∗)2∆(g1)−2 = ∆(Φ(g1))2∆(g1)−2. Since as was shown earlier we have that ∆(g0) ≥ 1, it therefore follows that 0 < ∆(g1) ≤ ∆(Φ(g1)). If we combine this with the generalised Jensen inequality [6, 3.3], we obtain 0 < ∆(g1) = ∆(Φ(g1)). Similarly, 0 < ∆(g0) = ∆(Φ(g0)). Thus by [6, Theorem 4.4], both g0 and g1 are strongly outer. (cid:3) When combined with Theorem 4.1, the following lemma characterises the invertibility of Toeplitz operators in terms of positive angle. If we further combine this lemma with the noncommutative Helson-Szego theorem obtained in the previous section, we end up with the promised structural characterisation of invertible Toeplitz operators with unitary symbols. Lemma 4.3. Let u ∈ M be a unitary of the form described in the previous lemma. Then Tu is invertible if and only if A∗ and A0 are at positive angle with respect to the functional τ (w·), where w = g∗ 0g0 = d∗(g∗ 1g1)−1d. A NONCOMMUTATIVE HELSON-SZEG O THEOREM 9 Proof. First suppose that Tu is invertible. For any a ∈ A the element g0a will belong to H 2(M). So the invertibility of Tu ensures that we can find a constant K > 0 so that kg0ak2 ≤ KkTu(g0a)k2, a ∈ A. Recall that by Lemma 4.2 u is of the form u = (g∗ to 1)−1dg−1 0 . Thus the former inequality translates kg0ak2 ≤ KkP+((g∗ 1)−1da)k2, a ∈ A. Now observe that for any b ∈ A0, the element (g∗ Hence 1)−1db∗ will belong to H 2(M)∗A∗ 0 ⊂ H 2 0 (M)∗. If we now write kf kw for τ (wf ∗f )1/2, then for any a ∈ A and b ∈ A0 we have that P+((g∗ 1)−1da) = P+(cid:0)(g∗ 1)−1da + (g∗ 1)−1db∗(cid:1). ka∗kw = τ (a∗wa)1/2 = kg0ak2 1)−1da + (g∗ 1)−1d(a + b∗)k2 ≤ KkP+(cid:0)(g∗ ≤ Kk(g∗ = Kτ ((a∗ + b)w(a + b∗)) = Kka∗ + bkw 1)−1db∗(cid:1)k2 Thus A∗ and A0 are at positive angle with respect to the functional τ (w·). Conversely, suppose that A∗ and A0 are at positive angle with respect to the functional τ (w·). We first show that Tu has dense range, and hence that it will be invertible whenever it is bounded below. Let a0 ∈ H 2(M) be orthogonal to Tu(H 2(M)). We will show that a0 must then be the zero vector. Given a ∈ A, the orthogonality of a0 to Tu(H 2(M)) together with the fact that u = (g∗ 0 , ensures that 1)−1dg−1 0 = hTu(g0a), a0i = τ (a∗ 1)−1da)) = τ (a∗ = τ (a∗ 0P+((g∗ 0(g∗ 1)−1da). 0Tu(g0a)) However, as was noted in the first part of the proof, for any b ∈ A0 we have that a∗ 0(g∗ 1)−1db∗ ∈ H 2 0 (M)∗, which implies that Thus τ (a∗ 0(g∗ 1)−1db∗) = τ (Φ(a∗ 0(g∗ 1)−1db∗)) = 0. Hence d∗g−1 0(g∗ 1 a0 = 0, so a0 = 0. τ (a∗ 1)−1d(a + b∗)) = 0 for all a ∈ A, b ∈ A0. It remains to show that Tu is bounded below whenever A∗ and A0 are at positive angle with respect to the functional τ (w·). Hence assume that there exists a constant B > 0 so that Since by assumption we have that d = Φ(g∗ follows that ka∗kw ≤ Bka∗ + bkw for all a ∈ A, b ∈ A0. 1 and (g∗ 1), and since both g∗ 1)−1d belong to H 2(M)∗, it d = Φ(d) = Φ(g∗ 1[(g∗ 1)−1d]) = Φ(g∗ 1)Φ((g∗ 1)−1d) = dΦ((g∗ 1)−1d). This yields that Φ((g∗ ∆(g1) = ∆(Φ(g1)) > 0 by [6, Theorem 4.4]. Consequently 1))∆((g∗ 1)−1d) = 1. Now since g∗ 1)−1d) = ∆(Φ(g∗ ∆(d) = ∆(g∗ 1 )∆((g∗ 1 )−1d) = ∆(d)∆((g∗ 1 )−1d). 1 is by assumption strongly outer, we have that Thus since ∆(d) > 0 by the strong outerness of d, we must have that 1 )−1d)). 1)−1d) = 1 = ∆(1) = ∆(Φ((g∗ ∆((g∗ Hence by [6, Theorem 4.4] (g∗ [(g∗ 1)−1dA∗ 0] = H 2 1)−1d is a strongly outer element of H 2(M)∗. But this ensures that 0 (M)∗. Hence for any fixed a ∈ A, we may select a sequence {bn} ⊂ A0 so that (g∗ 1)−1db∗ n → (P+ − Id)[(g∗ 1 )−1da] ∈ H 2 0 (M)∗ in L2(M). 10 LOUIS E LABUSCHAGNE AND QUANHUA XU Finally recall that by assumption g0 = (g∗ sequence as constructed above, we have that 1)−1d. So given any a ∈ A, with {bn} ⊂ A0 the kg0ak2 = ka∗kw ≤ Bka∗ + bnkw n)k2 n)k2 = Bkg0(a + b∗ = Bkg0(a + b∗ = Bk(g∗ = Bk(g∗ 1)−1d(a + b∗ 1)−1d(a + b∗ n)k2 n)k2. Letting n → ∞ now yields kg0ak2 ≤ BkP+[(g∗ 1)−1da]k = BkTu(g0a)k2 for any a ∈ A. Finally note that by assumption g0 is an outer element of H 2(M). With g0A therefore being dense in H 2(M), the above inequality extends by continuity to the claim that kak2 ≤ BkTu(a)k2 for any a ∈ H 2(M). Thus Tu is invertible. (cid:3) Definition 4.4. Given f ∈ M we define the Hankel operator with symbol f by means of the prescription where P− is the orthogonal projection from L2(M) onto H 2(M)∗. Hf : H 2(M) → H 2(M)∗ : x 7→ P−(f x), The following lemma is entirely elementary. Lemma 4.5. Let f ∈ M be given. Then kHf H 2 0 k = sup{τ (f F ) : F ∈ H 1 0 (M), τ (F ) ≤ 1}. Proof. Since for every x ∈ H 2(M) we have that (Id − P−)(x) ∈ H 2 (Id − P−)(x) will be orthogonal to any y ∈ H 2(M)∗. Thus hP−(f a), bi = hf a, bi for any a ∈ H 2 and b ∈ H 2(M)∗. Thus 0 (M), it is clear that such an 0 (M) kHf H 2 0 k = sup{kP−(f a)k : a ∈ H 2 0 (M), kak2 ≤ 1} 0 (M), b ∈ H 2(M)∗, kak2 ≤ 1, kbk2 ≤ 1} = sup{hP−(f a), bi : a ∈ H 2 = sup{hf a, bi : a ∈ H 2 = sup{τ (f ab∗) : a ∈ H 2 = sup{τ (f F ) : F ∈ H 1 0 (M), b ∈ H 2(M)∗, kak2 ≤ 1, kbk2 ≤ 1} 0 (M), b ∈ H 2(M)∗, kak2 ≤ 1, kbk2 ≤ 1} 0 (M), τ (F ) ≤ 1}. Here the last equality follows from the Noncommutative Riesz Factorisation theorem from [25] and [17]. (cid:3) We are now ready to present our final result. When taken alongside Theorem 4.1, this result fully characterises invertible Toeplitz operators. Theorem 4.6. Let u ∈ M be a unitary of the form described in Lemma 4.2. Then the following are equivalent: • Tu is invertible; • there exists k ∈ A such that ℜ(u∗k) is strictly positive; • the Hankel operator Hu restricted to H 2 0 (M) has norm less than 1. Proof. Our aim is to apply Theorem 3.9. In this regard we point out that although this theorem is formulated for norm one elements of L1(M)+, that assumption is one of convenience and not necessity. Hence the value of kwk1 is no essential obstruction to applying this theorem. Next observe that the fact that w = g∗ 0g0, not only ensures that ∆(w) = ∆(g0)2 > 0, but also that w is injective. Thus by Lemma 3.2, s(Φ(w)) = 1. We showed in the proof of the preceding Lemma that ∆((g∗ enables us to conclude from [6, Theorem 4.4] that d∗g−1 and hL = g0, it follows that w is of the form is a strongly outer element of H 2(M). On setting hR = d∗g−1 1 )−1d)). Applying this fact to d∗g−1 1 1)−1d) = 1 = ∆(Φ((g∗ 1 1 w = d∗g−1 1 (g∗ 1)−1d = d∗g−1 1 [(g∗ 1)−1dg−1 0 ]g0 = hRuhL A NONCOMMUTATIVE HELSON-SZEG O THEOREM 11 with hR and hL strongly outer elements of H 2(M) for which we have that 1 )−1d = w1/2. hL = g0 = w1/2 and h∗ R = (g∗ With all the other conditions of this theorem being satisfied, we may now conclude from Theorem 3.9 that A and A∗ 0 are at positive angle with respect to the functional τ (w·) if and only if there exists a k ∈ A such that ℜ(u∗k) is strictly positive. From the proof of Theorem 3.9 we also have that A and A∗ 0 (M), τ (F ) ≤ 1} < 1. The result now follows from an application of the preceding two lemmata. (cid:3) 0 are at positive angle if and only if sup{τ (f F ) : F ∈ H 1 Remark 4.7. We point out that for any unitary u of the form described in Lemma 4.2, the condition in the third bullet of the above theorem cannot be improved in the sense that for such a unitary, Hu must necessarily have norm 1. Suppose that u is of the form u = (g∗ 0 where g0, g1 are strongly outer elements of H 2(M) and d a strongly outer element of L2(D), related by the conditions that dg−1 1 )∗ ∈ H2(M)∗ must then be orthogonal to (Id − P−)(ug0). Hence we get that 1g1)−1d. Notice that (g∗ 1 ∈ H 2(M) and g∗ 1)−1d = (d∗g−1 0g0 = d∗(g∗ 0 , d∗g−1 1)−1dg−1 hHu(g0), (g∗ 1)−1di = hug0, (g∗ 1)−1di 1)−1d, (g∗ 1)−1di 1 (g∗ = h(g∗ = τ (d∗g−1 = τ (g∗ = kg0k2.k(g∗ 0 g0)1/2.τ (d∗g−1 1)−1dk2. 1)−1d) 1 (g∗ 1)−1d)1/2 This can clearly only be the case if kHuk ≥ 1. Since we also have that kHuk ≤ kuk∞ = 1, the claim follows. Acknowledgments. The contributions of the first named author is based upon research supported by the National Research Foundation. Any opinion, findings and conclusions or recommendations expressed in this material, are those of the authors, and therefore the NRF do not accept any liability in regard thereto. The second named author is partially supported by ANR-2011-BS01- 008-01 and NSFC grant No. 11271292. References [1] W. B. Arveson. Analyticity in operator algebras. Amer. J. Math., 89 (1967), 578-642. [2] R. Bruzual and M. Dom´ınguez. Operator-valued extension of the theorem of Helson and Szego. Operator Theory: Advances and Applications. 149 (2004), 139-152. [3] T. N. Bekjan and Q. Xu. Riesz and Szego type factorizations for noncommutative Hardy spaces. J. Operator Theory. 62 (2009), 215-231. [4] M. Bekker and A.P. Ugol'nikov. The Helson-Szego theorem for operator-valued weight. Methods of Funct. Anal. and Topology. 10 (2004), 11-16. [5] D. P. Blecher and L. E. Labuschagne. Characterizations of noncommutative H∞. Integr. Equ. Oper. Theory 56 (2006), 301-321. [6] D. P. Blecher and L. E. Labuschagne. Applications of the Fuglede-Kadison determinant: Szego's theorem and outers for noncommutative H p. Trans. Amer. Math. Soc. 360 (2008), 6131-6147. [7] D. P. Blecher and L. E. Labuschagne. Von Neumann algebraic H p theory, Proceedings of the 5th conference on function spaces. Contemporary Math. 435 (2007) 89-114. [8] L. G. Brown. Lidskii's theorem in the type II case. Geometric methods in operator algebras (Kyoto, 1983), 1-35. Pitman Res. Notes Math. Ser., 123, Longman Sci. Tech., Harlow, 1986. [9] A. Devinatz. Toeplitz operators on H 2 spaces. Trans. Amer. Math. Soc. 112 (1964), 304-317. [10] M. A. Mom´ınguez. A matricial extension of the Helson-Sarason theorem and a characterization of some mul- tivariate linearly completely regular processes. J. Multivariate Anal. 31 (1989), 289-310. [11] B. Fuglede and R.V. Kadison. Determinant theory in finite factors. Ann. Math. 55 (1952), 520-530. [12] J.B. Garnett. Bounded analytic functions. Academic Press, 1981. [13] U. Haagerup and H. Schultz. Brown measures of unbounded operators affiliated with a finite von Neumann algebra. Math. Scand. 100 (2007), 209-263. [14] H. Helson and G. Szego. A problem in prediction theory Ann. Mat. Pure Appli. 51 (1960), 107-138. [15] R. Hunt, B. Muckenhoupt and R. Wheeden. Weighted norm inequalities for conjugate function and Hilbert transform. Trans. Amer. Math. Soc. 176 (1973), 227-251. [16] L. E. Labuschagne. A noncommutative Szego theorem for subdiagonal subalgebras of von Neumann algebras. Proc. Amer. Math. Soc. 133 (2005), 3643-3646. [17] M. Marsalli and G. West. Non-commutative H p spaces. J Operator Theory 40 (1998), 339-355. 12 LOUIS E LABUSCHAGNE AND QUANHUA XU [18] M. Marsalli and G. West. Toeplitz operators with noncommuting symbols. Integr. Equ. Oper. Theory 32 (1998), 65-74. [19] G. Pisier and Q. Xu. Non-commutative Lp-spaces. In Handbook of the geometry of Banach spaces, Vol. 2, pages 1459-1517. North-Holland, Amsterdam, 2003. [20] M. Pouraimadi. A Matricial extension of the Helson-Szegd theorem and its application in multivariate predic- tion. J. Multivariate Anal. 16 (1985), 265-275. [21] H. Pousson. Systems of Toeplitz Operators on H 2. Proc. Amer. Math. Soc. 19 (1968), 603-608. [22] H. Pousson. Systems of Toeplitz Operators on H 2: II. Trans. Amer. Math. Soc. 133 (1968), 527-536. [23] B. Prunaru. Toeplitz and Hankel operators associated with subdiagonal algebras. Proc. Amer. Math. Soc. 139 (2010), 1387-1396. [24] N. Randrianantoanina. Hilbert transform associated with finite maximal subdiagonal algebras. J. Austral. Math. Soc. Ser. A. 65 (1998), 388-404. [25] K.-S. Saito. A note on invariant subspaces for finite maximal subdiagonal algebras. Proc. Amer. Math. Soc. 77 (1979), 348-352. [26] M. Takesaki. Theory of Operator Algebras: Vol 1. Springer, New York, 1979. [27] S. Treil and A. Volberg. Wavelets and the Angle between Past and Future. J. Funct. Anal. 143 (1997), 269-308. [28] Y. Ueda. On peak phenomena for non-commutative H∞. Math. Ann. 343 (2009), 421-429 Internal Box 209, School of Comp., Stat. & Math. Sci., NWU, Pvt. Bag X6001, 2520 Potchefstroom, South Africa E-mail address: [email protected] School of Mathematics and Statistics, Wuhan University, Wuhan 430072, China and Laboratoire de Math´ematiques, Universit´e de Franche-Comt´e, 25030 Besancon, cedex-France E-mail address: [email protected]
1908.09121	1	1908	2019-08-24T10:07:40	Minimal index and dimension for inclusions of von Neumann algebras with finite-dimensional centers	[ "math.OA", "math-ph", "math.CT", "math-ph" ]	The notion of index for inclusions of von Neumann algebras goes back to a seminal work of Jones on subfactors of type ${I\!I}_1$. In the absence of a trace, one can still define the index of a conditional expectation associated to a subfactor and look for expectations that minimize the index. This value is called the minimal index of the subfactor. We report on our analysis, contained in [GL19], of the minimal index for inclusions of arbitrary von Neumann algebras (not necessarily finite, nor factorial) with finite-dimensional centers. Our results generalize some aspects of the Jones index for multi-matrix inclusions (finite direct sums of matrix algebras), e.g., the minimal index always equals the squared norm of a matrix, that we call \emph{matrix dimension}, as it is the case for multi-matrices with respect to the Bratteli inclusion matrix. We also mention how the theory of minimal index can be formulated in the purely algebraic context of rigid 2-$C^*$-categories.	math.OA	math	Minimal index and dimension for inclusions of von Neumann algebras with finite-dimensional centers Luca Giorgetti∗ Dipartimento di Matematica "Guido Castelnuovo" Sapienza Universit`a di Roma Piazzale Aldo Moro, 5, I-00185 Roma, Italy [email protected] Abstract The notion of index for inclusions of von Neumann algebras goes back to a seminal work of Jones on subfactors of type II 1. In the absence of a trace, one can still define the index of a conditional expectation associated to a subfactor and look for expectations that minimize the index. This value is called the minimal index of the subfactor. We report on our analysis, contained in [GL19], of the minimal index for inclusions of arbitrary von Neumann algebras (not necessarily finite, nor factorial) with finite- dimensional centers. Our results generalize some aspects of the Jones index for multi- matrix inclusions (finite direct sums of matrix algebras), e.g., the minimal index always equals the squared norm of a matrix, that we call matrix dimension, as it is the case for multi-matrices with respect to the Bratteli inclusion matrix. We also mention how the theory of minimal index can be formulated in the purely algebraic context of rigid 2-C ∗-categories. 1 Motivation One motivation for studying von Neumann algebras with non-trivial centers and inclusions, or better bimodules, between them comes from the theory of Quantum Information. In an operator-algebraic description of quantum systems, observables are described by the self-adjoint part of a non-commutative von Neumann algebra M (with separable predual), while states correspond to normal faithful positive functionals ϕ : M → C normalized such that ϕ(1) = 1, where 1 denotes the identity operator. Keep in mind as an example the most commonly studied case of finite quantum systems [OP93, Part I] where the algebra generated by the observables is finite-dimensional, thus a multi-matrix algebra. Namely, M ∼= Li=1,...,m Mki(C), where m, ki ∈ N and Mki(C) is the algebra of ki × ki matrices over C, realized on the finite dimensional Hilbert space CN , N = k1 +. . .+km. More generally, the center Z(M) = M∩M′ is the classical part of the system, in the previous case Z(M) ∼= Cm, while each factor in the central decomposition of M, in the previous case Mki(C), is a purely ∗Supported by the ERC Advanced Grant n. 669240 QUEST "Quantum Algebraic Structures and Models", OPAL "Consolidate the Foundations" and GNAMPA -- INdAM. 1 quantum part of the system. Recall that a factor is a von Neumann algebra with center equal to C1. In this note we shall always assume, as in [GL19], that the center is finite-dimensional, Z(M) ∼= Cm, thus M ∼= M i=1,...,m Mi where pi ∈ Z(M) are the minimal central projections and Mi = Mpi are factors (of arbitrary type). In this sense, we study possibly infinite quantum systems with a finite-dimensional classical part. The building blocks of information transfer (communication) from a quantum system N to another M are called quantum channels. In the operator-algebraic setting, they are conventionally described by normal completely positive maps α : N → M such that α(1) = 1. Recall that a map α : N → M is called positive if it sends positive elements of N to positive elements of M, while it is called completely positive if α⊗ idk×k : N ⊗ Mk(C) → M⊗ Mk(C) is positive for every k ∈ N. Communication takes place via transferring states from one system to another by pullback, namely αt(ϕ) := ϕ ◦ α is a state on N whenever ϕ is a state on M. Note also that normal states and normal unital ∗-homomorphisms are examples of completely positive maps. In this note, as in the first part of [GL19], we will mostly restrict ourselves to quantum channels given by inclusion morphisms ι : N ֒→ M associated to unital inclusions of von Neumann algebras N ⊆ M. We will furthermore assume that the inclusion morphism has finite Jones index. This assumption is equivalent to the existence of a conjugate morphism ι : M → N (conjugate quantum channel going in the opposite direction). The notion of Jones index will be reviewed in the next section, while we refer to [LR97] and to the second part of [GL19] for the definition of conjugate morphism and its relation to the theory of (minimal) index. For now, we only mention that these notions of conjugation and of minimal index, and the more fundamental notion of (matricial, intrinsic) dimension, are naturally formulated in a tensor C ∗-categorical language. Namely for abstract 1-arrows X : N → M running between 0-objects N ,M of a 2-C ∗-category. Remark 1.1. More generally, one can think of quantum channels as described by N -M bimodules H, also denoted by NHM, see [Lon18, Sec. 2,3]. Recall that a bimodule is a Hilbert space H with a normal left action of N , l : N → B(H), namely l(n1n2) = l(n1)l(n2), and a normal right action of M, r : M → B(H), namely r(m1m2) = r(m2)r(m1), such that l(N ) and r(M) mutually commute in B(H). Thus a bimodule sees in a "balanced way" the inclusion l(N ) ⊂ r(M)′ and the dual inclusion r(M) ⊂ l(N )′. Moreover, it is known that every normal unital completely positive map gives rise to a bimodule. In this algebraic setup of Quantum Information, Longo [Lon18, Thm. 3.2, Cor. 3.4] gave a mathematical derivation of Landauer's bound for possibly infinite quantum systems [Lan61]. See also [Ben03], [PV01] for an introduction to Landauer's principle and bound, and for an explanation of how these settle the famous Maxwell's demon paradox. The bound is a lower estimate on the amount of energy (heat) that is emitted from the system whenever 1 classical bit of information is deleted (or any logically irreversible operation is performed). Namely, Eα ≥ 1 2 kT log(2) 2 where Eα is the variation of the free energy associated to the channel α, k is Boltzmann's constant and T is the temperature of the environment. The bound is calculated in [Lon18] by means of the matrix dimension Dα, it is in general half of the original Landauer's bound E ≥ kT log(2), and it coincides with the latter in the case of finite quantum systems because the lowest non-trivial possible (scalar) dimension of an inclusion of matrix algebras is 2 instead of √2. The most important properties of the matrix dimension Dα of a quantum channel α, in our case of an inclusion morphism α = ι, are its multiplicativity and additivity: Dβ◦α = DβDα, Dα⊕β = Dα + Dβ where β ◦ α and α ⊕ β denote respectively the composition (or "tensor multiplication") and the direct sum of channels. Moreover, Dα determines the minimal index as the square of its l2-norm, and the (unique) minimal conditional expectation via Perron-Frobenius theory, as we shall explain in the next section. 2 Minimal index Let N ⊂ M be a unital inclusion of von Neumann algebras. The Jones index of the inclusion is a number (≥ 1) that measures the "relative size" of M w.r.t. N . The index equals 1 if N = M, it equals +∞ if M is way bigger than N , and it enjoys the exciting property of being quantized between 1 and 4. Actually, there can be more than one notion of "index" for an inclusion, depending on the type of algebras involved. If N ∼= Mk(C) and M ∼= Mh(C) (finite type I factors) then h = km for some m ∈ N, the inclusion morphism is the amplification, namely Mk(C) ⊗ 1m ⊂ Mh(C), and the index is the square of the multiplicity m2, i.e., the ratio of the algebraic dimensions of M over N . If N and M are factors of type II 1, the Jones index, denoted by [M : N ], is defined in terms of tracial states [Jon83]. This is the original definition of index and in this regime one can observe the already mentioned quantization phenomenon of the index values [Jon83, Thm. 4.3.1]. If N ⊂ M is a multi-matrix inclusion (the simplest instance of inclusion with non-trivial centers), namely N ∼= Lj=1,...,n Mkj (C) and M ∼= Li=1,...,m Mhi(C), denote by k = (k1, . . . , kn)t, h = (h1, . . . , hm)t the vectors of dimensions (with positive integer components), and by Λ the Bratteli inclusion matrix (with positive integer entries) describing the inclusion morphism of N in M, [Bra72]. Then Λk = h is the only consistency condition on the Bratteli diagram associated to the inclusion, and the index (there are several equivalent definitions of index in this case) equals kΛk2 l2 , [GdlHJ89, Ch. 2]. For an inclusion N ⊂ M of finite direct sums of type II 1 factors, the index, again denoted by [M : N ], is defined as the spectral radius of a product of matrices constructed from the unique trace on each factor in M (trace matrix ) and from the square roots of the Jones indices of the subfactors obtained by central decomposition (Jones index matrix ), [GdlHJ89, Ch. 3]. If the inclusion is connected, i.e., Z(N ) ∩ Z(M) = C1 (which is equivalent to the actual connectedness of the adjacency graph of N ⊂ M, see [GdlHJ89, Sec. 1.3]) and if [M : N ] ≤ 4, then [M : N ] = kΛM N k2 l2 by [GdlHJ89, Thm. 3.7.13], where ΛM N is the aforementioned Jones index matrix. The theory of index for inclusion of finite von Neumann algebras has been further extended to inclusions with possibly infinite (atomic or diffuse) centers by Jolissaint in [Jol90]. In the absence of a trace, e.g., for subfactors of type III, one can still define the index of a normal faithful conditional expectation E : M → N , denoted by Ind(E), [Kos86]. An 3 inclusion is said to have finite index if it admits some E with k Ind(E)k < +∞. The index Ind(E) is in general an element of Z(M), with Ind(E) ≥ 1, it is of course a scalar in the case of subfactors, and it gives back the Jones index for a finite subfactor by [M : N ]1 = Ind(Etr), where Etr is the trace-preserving expectation. If the inclusion is not irreducible, i.e. N ′ ∩ M 6= C1, there can be several expectations E : M → N , and one can look for those minimizing the number k Ind(E)k. The minimal index of the inclusion N ⊂ M is then defined to be [M : N ]0 := inf E {k Ind(E)k} This analysis has been performed first in the case of subfactors by Hiai [Hia88], Longo [Lon89] and Havet [Hav90, Sec. 1], where it is shown that there is a unique expectation minimizing the index and this expectation is characterized via a certain sphericality condition (which opens the way to a tensor C ∗-categorical formulation of the minimal index, see [LR97]). Remark 2.1. For finite subfactors it can happen that [M : N ]0 (cid:12) [M : N ]. Equality is attained for the so called extremal subfactors, see [PP86, Sec. 4], [PP91] and [Bur03, Sec. 2.3]. In case of arbitrary inclusions, Jolissaint [Jol91, Thm. 1.8] proved that there is always an expectation minimizing the index (thus called a minimal expectation), but this expectation is not unique in general as shown by Fidaleo and Isola [FI96, Prop. 10, Sec. 5]. The starting points of our analysis are the works of Havet [Hav90, Sec. 2] and Teruya [Ter92], where it is shown that in the case of connected inclusions with finite-dimensional centers there is a unique minimal expectation E0, and its index is a scalar operator in Z(M), i.e., Ind(E0) = [M : N ]01. We mention that the connectedness assumption is almost without loss of generality, as every inclusion can be written as a direct sum of connected ones. Theorem 2.2. [GL19]. Let N ⊂ M be a connected inclusion with finite index and finite- dimensional centers. Let p1, . . . , pm and q1, . . . , qn be the minimal projections in Z(M) and Z(N ) respectively and, whenever piqj 6= 0, consider the subfactors Nij := N piqj ⊂ Mij := qjMpiqj and define dij := [Mij : Nij]1/2 , while dij := 0 otherwise. Then 0 [M : N ]0 = kDk2 l2 where D is the m × n matrix with entries dij. We call D the matrix dimension of N ⊂ M and d := kDkl2 its scalar dimension. : Mij → Nij and a matrix of numbers λE ij = 1 (thus called column-stochastic or Markovian). Namely, λE From the previous theorem, together with the quantization of the index for expectations between factors [Kos86], as in [GdlHJ89, Prop. 3.7.12 (c)] one can conclude that either [M : N ]0 ∈ {4 cos2(π/k), k ∈ N, k ≥ 3} or [M : N ]0 ≥ 4. For every normal faithful conditional expectation E : M → N one can consider a matrix ij ≥ 0 with the property of expectations Eij that Pi λE ijqj := E(pi)qj and Eij(qjxpiqj) := (λE ij one can reconstruct the expectation via E(x) = Pi,j λE ijσij(Eij(qjxpiqj)), x ∈ M, where σij : Nij → Nj := N qj is the inverse of the induction isomorphism yqj 7→ ypiqj, y ∈ N . Moreover, every expectation arises uniquely in this way, [Hav90, Prop. 2.2, 2.3], [Ter92, Prop. 2.1]. Note that by connectedness assumption, the matrix dimension D (or equivalently any matrix with the same pattern of zero and non-zero entries) is indecomposable, i.e., DDt and DtD are irreducible square matrices, [GdlHJ89, Lem. 1.3.2, 2.3.1]. ij)−1E(xpi)piqj for every x ∈ M. From the pair Eij, λE 4 Theorem 2.3. [GL19]. Let N ⊂ M be as in the previous theorem. Consider the eigenvalue equations DtD√ν = d2√ν DDt√µ = d2√µ , . . . , ν1/2 where √ν = (ν1/2 m )t are vectors with strictly positive entries and l2-normalized (thus unique by Perron-Frobenius theory). Then the minimal expectation E0 : M → N is determined by n )t, √µ = (µ1/2 , . . . , µ1/2 1 1 (E0)ij = E0 ij, λE 0 ij = dij d µ1/2 i ν1/2 j where E0 ij : Mij → Nij is the unique minimal expectation in each subfactor, if piqj 6= 0. As a consequence, we have the following "weighted" additivity formula for the scalar dimension d (thus for [M : N ]0) d = X i,j dijν1/2 j µ1/2 i . Moreover, by setting ωl(qj) := νj and ωr(pi) := µi we have two faithful states ωl and ωr on Z(N ) and Z(M) respectively, canonically determined by the inclusion. We call them respectively the left and right state of N ⊂ M. These states provide a characterization of minimality of E0 which extends the previously mentioned (but not explained) sphericality condition in the case of subfactors. Namely, E0 is the only expectation from M onto N fulfilling ωl ◦ E0 = ωr ◦ (E0)′ on N ′ ∩ M where (E0)′ : N ′ → M′ is the dual expectation in the sense of Kosaki [Kos86]. Note that the previous equation makes sense because E0(N ′ ∩ M) = Z(N ) and (E0)′(N ′ ∩ M) = Z(M), and it defines a canonical state on N ′ ∩ M for the inclusion, that we call spherical state of N ⊂ M, denoted by ωs. Remark 2.4. In the subfactor case one has that the square root of the minimal index is additive and multiplicative, namely d = d1 + d2 if d is the dimension of N ⊂ M and d1, d2 are obtained by cutting with projections p1, p2 ∈ N ′ ∩ M such that p1 + p2 = 1. Moreover, let N ⊂ M ⊂ L be two consecutive subfactors, then d = d1d2 if d, d1, d2 are respectively the dimensions of N ⊂ L, N ⊂ M, M ⊂ L. These relations do no longer hold for inclusions with non-trivial centers, indeed one has to replace the scalar dimension with the matrix dimension to have D = D1 + D2 and D = D2D1. In particular, the dimension (thus the minimal index) is in general only submultiplicative d ≤ d1d2, while the minimal index itself can be additive d2 = d2 2 (if one of N or M is a factor). 1 + d2 3 Extremality and super-extremality In the case of finite direct sums of finite factors, as for type II 1 subfactors, one can compare the two theories of index (trace/minimal). Given a connected inclusion of such algebras 5 N ⊂ M with finite index, on one hand, we have the matrix dimension D, the minimal conditional expectation E0 with index [M : N ]0, and the spherical state ωs on N ′ ∩ M. On the other hand, we have the Jones index matrix ΛM N , the Jones index [M : N ] and a uniquely determined trace τ on M, called the Markov trace of N ⊂ M, [GdlHJ89, Sec. 2.7,3.7], which extends to the Jones tower. Definition 3.1. In [GL19, Sec. 3], we called extremal an inclusion which fulfills E0 = Eτ , where Eτ is the Markov trace-preserving expectation, and super-extremal an inclusion which fulfills in addition ωs = τ↾N ′∩M. By [Hav90, Prop. 3.2], [GdlHJ89, Cor. 3.7.4] we have Ind(Eτ ) = [M : N ]1, thus for an extremal inclusion we have [M : N ]0 = [M : N ]. Recall from [GL19, Lem. 3.2] that ωs = τ↾N ′∩M is equivalent to ωl = τ↾Z(N ), where ωl is by definition ωs↾Z(N ). Another condition that one might consider is the equality of matrices D = ΛM N , which corresponds to an entrywise extremality for the inclusion once reduced with every piqj. Clearly for a finite subfactor, all these notions, including super-extremality, boil down to the ordinary notion of extremality, see Remark 2.1. The following results completely settle the analysis of (super-)extremal inclusions of multi- matrices (always assumed to have finite-dimensional centers, thus with finite index). As in the first paragraph of the previous section, if N ⊂ M is the inclusion, denote by k = (k1, . . . , kn)t, h = (h1, . . . , hm)t the vectors of dimensions such that N ∼= Lj=1,...,n Mkj (C) and M ∼= Li=1,...,m Mhi(C), and denote by Λ the Bratteli inclusion matrix. Recall that Λk = h is the consistency of the Bratteli diagram. Theorem 3.2. [GL19]. Let N ⊂ M be a connected multi-matrix inclusion. is also super-extremal if and only if Then D = ΛM N = Λ and the inclusion is always extremal, namely E0 = Eτ . The inclusion The index (we need not specify which one) of a super-extremal multi-matrix inclusion is Λth = d2k. easy to compute and it has the following properties: Proposition 3.3. [GL19]. Let N ⊂ M be as in the previous theorem. If the inclusion is super-extremal then [M : N ]0 = khk2 kkk2 (the ratio of the algebraic dimensions of M over N ). In particular the index must be a positive integer (because rational and algebraic integer) and every positive integer (not only squares of integers) is the index of such an inclusion. l2 l2 Moreover, the index of super-extremal multi-matrix inclusions is clearly multiplicative. 4 Open problems Some natural problems (currently under investigation) that arise from the analysis of the minimal index for von Neumann inclusion reviewed here are: Problem 4.1. In the case of inclusions with infinite-dimensional and atomic centers there can be more than one expectation such that k Ind(E0)k = inf E{k Ind(E)k}. Can one find a preferred, canonical one, whose index is scalar and which is related to the (infinite) matrix dimension in some way? 6 Note that in the case of the previous problem the matrix dimension can be defined as for finite-dimensional centers using minimal central projections. Problem 4.2. In the same situation as above, does the theory of minimal index admit a purely 2-C ∗-categorical (or better 2-W ∗-categorical) formulation, cf. [GLR85]? Namely, does the theory of intrinsic tensor-categorical dimension admit an extension beyond tensor C ∗- categories with simple tensor unit or with finitely reducible tensor unit? What is a "standard solution" of the conjugate equations beyond the previously mentioned regimes? We do not want to ask the same question of categorical translation beyond the case of atomic centers, because we cannot immagine by now a good notion of direct integral of objects in a tensor C ∗- (or W ∗-) category. Problem 4.3. In the case of inclusions with infinite-dimensional and possibly diffuse cen- ters (in the absence of minimal central projections) what is a good substitute of the matrix dimension? Problem 4.4. Study the consequences of and characterize super-extremality for inclusion of finite direct sums of type II 1 factors. Find examples of such inclusions that go beyond tensoring a super-extremal multi-matrix inclusion with a type II 1 factor. i=1 R) (Q-systems) or Bim(Ln Problem 4.5. Study C ∗-Frobenius algebra objects in tensor C ∗-categories with non-simple but finitely reducible tensor unit (e.g. for unitary multi-fusion categories), cf. [BKLR15, Ch. 3], [EGNO15, Ch. 4]. Study the relation between the realization of C ∗-Frobenius algebras in End(Ln i=1 R, where R is a factor. It is known that every unitary fusion category, and more generally every rigid tensor C ∗-category with simple unit can be realized as endomorphisms or bimodules of a factor (that can be chosen either of type II 1 or of type III, and in some cases hyperfinite), [HY00], [Yam03], [BHP12], [GY19]. i=1 R) and extensions of Ln Problem 4.6. Study the problem of realizing unitary multi-fusion categories, or more gener- ally rigid 2-C ∗-categories with finite-dimensional centers, as endomorphisms or bimodules of Ln i=1 R, where R is a factor. Is every such abstract category realizable in operator-algebras, if so, is the realization unique in a suitable sense? Acknowledgements. We thank the Institute of Mathematics "Simion Stoilow" of the Romanian Academy and the West University in Timi¸soara for the kind hospitality during the 27th International Conference in Operator Theory (OT27) in July, 2 -- 6, 2018. We thank the organizers for the invitation to present our work at the conference and for financial support. We also thank Maria Stella Adamo and Yoh Tanimoto for discussions. References [Ben03] C. H. Bennett. Notes on Landauer's principle, reversible computation and Maxwell's demon. Studies in History and Philosophy of Modern Physics, 34:501 -- 510, 2003. [BHP12] A. Brothier, M. Hartglass, and D. Penneys. Rigid C ∗-tensor categories of bimodules over interpolated free group factors. J. Math. Phys., 53:123525 (43 pages), 2012. 7 [BKLR15] M. Bischoff, Y. Kawahigashi, R. Longo, and K.-H. Rehren. Tensor categories and endo- morphisms of von Neumann algebras. With applications to quantum field theory. Springer Briefs in Mathematical Physics, Vol. 3. Springer, Cham, 2015. [Bra72] [Bur03] O. Bratteli. Inductive limits of finite dimensional C ∗-algebras. Trans. Amer. Math. Soc., 171:195 -- 234, 1972. M. Burns. Subfactors, planar algebras and rotations. PhD thesis, University of California at Berkeley, 2003. Available at http://arxiv.org/abs/1111.1362v1. [EGNO15] P. Etingof, S. Gelaki, D. Nikshych, and V. Ostrik. Tensor categories. Mathematical Surveys and Monographs, Vol. 205. American Mathematical Society, Providence, RI, 2015. [FI96] F. Fidaleo and T. Isola. Minimal expectations for inclusions with atomic centres. Internat. J. Math., 7:307 -- 327, 1996. [GdlHJ89] F. M. Goodman, P. de la Harpe, and V. F. R. Jones. Coxeter graphs and towers of algebras. Mathematical Sciences Research Institute Publications, Vol. 14. Springer-Verlag, New York, 1989. [GL19] L. Giorgetti and R. Longo. Minimal index and dimension for 2-C ∗-categories with finite- dimensional centers. Comm. Math. Phys., 370:719 -- 757, 2019. [GLR85] P. Ghez, R. Lima, and J. E. Roberts. W ∗-categories. Pacific J. Math., 120:79 -- 109, 1985. [GY19] L. Giorgetti and W. Yuan. Realization of rigid C ∗-tensor categories via Tomita bimodules. J. Operator Theory, 81:433 -- 479, 2019. [Hav90] J.-F. Havet. Esp´erance conditionnelle minimale. J. Operator Theory, 24:33 -- 55, 1990. [Hia88] [HY00] [Jol90] [Jol91] F. Hiai. Minimizing indices of conditional expectations onto a subfactor. Publ. Res. Inst. Math. Sci., 24:673 -- 678, 1988. T. Hayashi and S. Yamagami. Amenable tensor categories and their realizations as AFD bimodules. J. Funct. Anal., 172:19 -- 75, 2000. P. Jolissaint. Index for pairs of finite von Neumann algebras. Pacific J. Math., 146:43 -- 70, 1990. P. Jolissaint. Indice d'esp´erances conditionnelles et alg`ebres de von Neumann finies. Math. Scand., 68:221 -- 246, 1991. [Jon83] V. F. R. Jones. Index for subfactors. Invent. Math., 72:1 -- 25, 1983. [Kos86] [Lan61] [Lon89] [Lon18] H. Kosaki. Extension of Jones' theory on index to arbitrary factors. J. Funct. Anal., 66:123 -- 140, 1986. R. Landauer. Irreversibility and heat generation in the computing process. IBM Journal of Research and Development, 5:183 -- 191, 1961. R. Longo. Index of subfactors and statistics of quantum fields. I. Comm. Math. Phys., 126:217 -- 247, 1989. R. Longo. On Landauer's principle and bound for infinite systems. Comm. Math. Phys., 363:531 -- 560, 2018. [LR97] R. Longo and J. E. Roberts. A theory of dimension. K-Theory, 11:103 -- 159, 1997. [OP93] [PP86] M. Ohya and D. Petz. Quantum entropy and its use. Texts and Monographs in Physics. Springer-Verlag, Berlin, 1993. M. Pimsner and S. Popa. Entropy and index for subfactors. Ann. Sci. Ecole Norm. Sup, 19:57 -- 106, 1986. 8 [PP91] [PV01] [Ter92] M. Pimsner and S. Popa. Finite-dimensional approximation of pairs of algebras and obstructions for the index. J. Funct. Anal., 98:270 -- 291, 1991. M. B. Plenio and V. Vitelli. The physics of forgetting: Landauer's erasure principle and information theory. Contemp. Phys., 42:25 -- 60, 2001. T. Teruya. Index for von Neumann algebras with finite-dimensional centers. Publ. Res. Inst. Math. Sci., 28:437 -- 453, 1992. [Yam03] S. Yamagami. C ∗-tensor categories and free product bimodules. J. Funct. Anal., 197:323 -- 346, 2003. 9
1902.01041	1	1902	2019-02-04T05:49:41	On Bi-R-Diagonal Pairs of Operators	[ "math.OA" ]	We study the properties of the analogue of R-diagonal operators in the setting of bi-free probability. Products of bi-R-diagonal pairs of operators that are $$-bi-free are studied and powers of such pairs are found to also be bi-R-diagonal. It is moreover shown that the joint $$-distribution of a bi-R-diagonal pair of operators remains invariant under the multiplication by a $*$-bi-free bi-Haar unitary pair and equivalent characterizations of the condition of bi-R-diagonality are developed.	math.OA	math	On Bi-R-Diagonal Pairs of Operators Georgios Katsimpas Abstract We study the properties of the analogue of R-diagonal operators in the setting of bi-free probability. Products of bi-R-diagonal pairs of operators that are ∗-bi- free are studied and powers of such pairs are found to also be bi-R-diagonal. It is moreover shown that the joint ∗-distribution of a bi-R-diagonal pair of operators remains invariant under the multiplication by a ∗-bi-free bi-Haar unitary pair and equivalent characterizations of the condition of bi-R-diagonality are developed. 1. Introduction In the theory of free probability, an R-diagonal operator is an element of a non-commutative ∗-probability space (A, ϕ) whose ∗-distribution coincides with the ∗-distribution of a prod- uct of the form u · p, where the sets {u, u∗} and {p, p∗} are freely independent and u is a Haar unitary, i.e. u is a unitary and ϕ(un) = 0, for all n ∈ Z \ {0}. It is due to this free factorization property that the class of R-diagonal operators constitutes a particu- larly well-behaved class of non-normal operators. From a combinatorial point of view, R-diagonal elements are characterized by having all of their free ∗-cumulants that are either of odd order, or have entries that are not alternating in ∗-terms and non-∗-terms equal to zero. This combinatorial approach has proved to be extremely fruitful in the development of the theory of R-diagonal operators (see [NS06] for an exposition of the combinatorics of free probability). In [NS97], R-diagonal operators were found to satisfy a "free absorption" property, namely that for any elements a, b in some non-commutative ∗-probability space such that a is R-diagonal and the sets {a, a∗} and {b, b∗} are freely independent, the element ab is also R-diagonal. In [HL00], Brown's spectral distribution measure was computed for R- diagonal operators in finite von Neumann algebras, while in [Lar02], powers of R-diagonal operators were shown to be R-diagonal and their determining sequences were computed (see also [NS06, Theorem 15.22] for a proof utilising combinatorial arguments). In [NSS01], a number of equivalent characterizations of R-diagonality were formulated, including conditions on ∗-moments, free cumulants and the freeness of certain self-adjoint matrices from the scalar matrices, with amalgamation over the diagonal scalar matrices, while in [BD18] similar results were obtained on B-valued R-diagonal elements in the operator-valued setting. Distributions of R-diagonal operators have found applications in the non-microstate approach to free entropy, answering questions regarding the minimiza- tion of the free Fisher information in the tracial framework (see [NSS99]). Bi-free probability theory originated in [Voi14] as an extension of the free setting and involves the simultaneous study of left and right actions of algebras on reduced free product spaces. The corresponding notion of bi-free independence found its combinatorial characterization in [CNS15b] (see also [CNS15a] for the development of the combinatorics of bi-free probability in the operator-valued setting). This paper is devoted to the study of the analogue of R-diagonal operators in the bi-free setting, namely bi-R-diagonal pairs 1 of operators and, to this end, the combinatorial approach originally proposed in [Sko16, Section 4] shall be adopted, which utilises the bi-free cumulant functions. For the study of products and powers of bi-R-diagonal pairs, similar arguments are used as to those corresponding to the results in the free case, but more care is required due to the dealing with the lattice of bi-non-crossing partitions and the χ-order. Since products of pairs of operators are considered pointwise (i.e. left operators are multiplied by left operators and right operators are multiplied by right operators), caution ought to be exercised when it comes to the order in which the multiplication takes place and, for the most general cases, it is necessary that the order of the multiplication of right operators is reversed (see Theorem 3.2). However, this is found not to play a role in the case when both pairs in question are bi-R-diagonal and ∗-bi-free (Proposition 3.4). These results imply that bi-R-diagonal pairs of operators satisfy a corresponding "bi-free absorption" property and indicate that such pairs of operators exist in abundance. The absence of characterizations of bi-free phenomena with conditions on moments is an unfortunate theme in the theory of bi-free probability (see, however, [Cha16] for an equivalent formulation of bi-free independence in terms of alternating moments). In par- ticular, a characterization of the condition of bi-R-diagonality in terms of ∗-moments was unable to be obtained. In the setting of free probability, one of the most salient features of the ∗-distribution of an R-diagonal operator is that it remains invariant after the multi- plication by a freely independent Haar unitary, a result obtained with the use freeness in terms of its characterization via moments (see [NSS01, Theorem 1.2] and [NS06, Theorem 15.10]). Bi-Haar unitary pairs of operators constitute the bi-free analogue of Haar unitaries and their joint ∗-distribution is modelled by the left and right regular representations of groups on Hilbert spaces. Theorem 4.4 is the generalization of the aforementioned fact to the bi-free setting and displays the invariance of the joint ∗-distribution of any bi-R- diagonal pair of operators under the multiplication of a ∗-bi-free bi-Haar unitary pair. The proof follows the combinatorial approach instead, utilising the bi-free cumulant functions and hence a new proof follows for the free case as well. In the spirit of [NSS01, Theorem 1.2], [BD18, Theorem 3.1] and by combining results from [Sko16], we obtain Theorem 4.6, displaying equivalent formulations of the condition of bi-R-diagonality. The paper is organized as follows: In Section 2 we list all the necessary preliminary notions on bi-free probability theory and fix the appropriate notation. Here, the notion of a bi-R-diagonal pair of operators is defined and a number of lemmas that will be used in subsequent parts of this manuscript will be stated and proved. Section 3 involves the study of the behaviour of bi-R-diagonal pairs under the taking of sums, products and arbitrary powers, while Section 4 is devoted to showing that the joint ∗-distributions of bi-R-diagonal pairs remain invariant under the multiplication by ∗-bi-free bi-Haar unitary pairs. 2. Preliminaries and Notation In this section we will develop the common preliminaries, fix the appropriate notation and state a number of lemmas to be used later in this manuscript. Our main framework will be that of a non-commutative ∗-probability space, i.e. a pair (A, ϕ) where A is a complex, unital ∗-algebra and ϕ : A → C is a unital, linear map such that for all a ∈ A. ϕ(a∗a) ≥ 0, 2 For any S ⊆ A, we will denote by alg(S) the subalgebra of A generated by the set S. If a1, . . . , an are elements of (A, ϕ), then: (a) their joint distribution is given by the linear functional µ : ChX1, . . . , Xni → C defined as µ(P ) = ϕ(P (a1, . . . , an)), (P ∈ ChX1, . . . , Xni) where ChX1, . . . , Xni denotes the unital algebra of polynomials in n-non-commuting indeterminates X1, . . . , Xn, (b) their joint ∗-distribution is given by the joint distribution of the family {a1, . . . , an, a1 ∗, . . . , an ∗}, (c) the family of their joint ∗-moments is given by {ϕ(c1 · · · ck) : k ≥ 1, ci ∈ {a1, . . . , an, a1 ∗, . . . , an ∗} for all 1 ≤ i ≤ k}. It is clear that for a1, . . . , an, b1, . . . , bn ∈ A, in order to verify equality of joint ∗-distributions of the families {a1, . . . , an} and {b1, . . . , bn}, it suffices to prove that all of their joint ∗- moments coincide. For a1, . . . , an ∈ A and ∅ 6= V = {j1 < j2 < . . . < js} ⊆ {1, . . . , n}, the restriction of the sequence (a1, . . . , an) to the set V is given by In this case, we define (a1, . . . , an)V = (aj1, aj2, . . . , ajs). ϕ((a1, . . . , an)V ) = ϕ(aj1 · aj2 · · · ajs). Also, if π is a partition of the set {1, . . . , n}, then we use the following notation: ϕπ(a1, . . . , an) = YV ∈π ϕ((a1, . . . , an)V ). 2.1. The Lattice of Bi-Non-Crossing Partitions Familiarity with the collection of non-crossing partitions NC(n), multiplicative functions on NC(n) and free cumulants is assumed (see [NS06] for an exposition of the combinatorics of free probability). For n ∈ N, we will be using maps χ ∈ {l, r}n to distinguish between left and right operators in a sequence of n-operators. Any such map gives rise to a permutation sχ on {1, . . . , n} as follows: If χ−1({l}) = {i1 < . . . < ip} and χ−1({r}) = {j1 < . . . < jn−p}, then define: sχ(k) =(ik, jn+1−k, if k ≤ p if k > p From a combinatorial standpoint, the only differences between free and bi-free probability arise from dealing with sχ. 3 The permutation sχ naturally induces a total order on {1, . . . , n} (which we will hence- forth be referring to as the χ-order ) as follows: i ≺χ j ⇐⇒ s−1 χ (i) < s−1 χ (j). Instead of reading {1, . . . , n} in the traditional order, this corresponds to first reading the elements of {1, . . . , n} labelled "l" in increasing order, followed by reading the elements labelled "r" in decreasing order. Note that if V is any non-empty subset of {1, . . . , n}, the map χV naturally gives rise to a map sχV , which should be thought of as a permutation on {1, . . . , V }. Before we discuss the lattice of bi-non-crossing partitions, we fix some notation regard- ing general partitions. For n ∈ N, the collection of all partitions on {1, . . . , n} is denoted by P(n), while the collection of non-crossing partitions on {1, . . . , n} is denoted by NC(n). The elements of any π ∈ P(n) are called the blocks of π and for 1 ≤ i, j ≤ n, we write i∼πj to mean that i and j belong to the same block of π, whereas i ≁π j indicates that i and j belong to different blocks of π. For π, σ ∈ P(n), we write π ≤ σ if every block of π is con- tained in a block of σ. This defines the partial order of refinement on P(n). The maximal element of P(n) with respect to this partial order is the partition consisting of one block (denoted by 1n), while the minimal element is the partition consisting of n-blocks (denoted by 0n). This partial order induces a lattice structure on P(n), hence for π, σ ∈ P(n), the join π ∨ σ (i.e. the minimum element of the non-empty set {ρ ∈ P(n) : ρ ≥ π, σ}) of π and σ is well defined. Definition 2.1. Let n ∈ N and χ ∈ {l, r}n. A partition τ ∈ P(n) is called bi-non-crossing with respect to χ if the partition s−1 the partition obtained by applying the permutation s−1 χ to each entry of every block of τ ) is non-crossing. Equivalently, τ is bi- non-crossing with respect to χ if whenever V, W are blocks of τ and v1, v2 ∈ V, w1, w2 ∈ W are such that χ · τ (i.e. v1 ≺χ w1 ≺χ v2 ≺χ w2, then we necessarily have that V = W. The collection of bi-non-crossing partitions with respect to χ is denoted by BNC(χ). It is clear that BNC(χ) = {τ ∈ P(n) : s−1 χ · τ ∈ NC(n)} = {sχ · π : π ∈ NC(n)} We will be referring to a partition τ simply as bi-non-crossing whenever it is clear from the context which map χ is implemented. Note that in the special case when the map χ is constant, one ends up with the collection of all non-crossing partitions on {1, . . . , n}. Example 2.2. If χ ∈ {l, r}6 is such that χ−1({l}) = {1, 2, 3, 6} and χ−1({r}) = {4, 5}, then (sχ(1), . . . , sχ(6)) = (1, 2, 3, 6, 5, 4) and the partition given by τ = {{1, 4}, {2, 5}, {3, 6}} is bi-non-crossing with respect to χ, even though τ /∈ NC(6). This may also be seen via the following diagrams: 1 2 3 6 4 5 −→ 4 1 2 3 6 5 4 The set of bi-non-crossing partitions with respect to a map χ ∈ {l, r}n inherits a lattice structure from P(n) via the partial order of refinement (although the join operation in BNC(χ) need not coincide with the restriction of the join operation in P(n)). The minimal and maximal elements of BNC(χ) will be denoted by 0χ and 1χ respectively (with 0χ = sχ(0n) = 0n and 1χ = sχ(1n) = 1n). For ∅ 6= V ⊆ {1, . . . , n}, we denote by min<V and min≺χV the minimum element of V with respect to the natural order and the χ-order of {1, . . . , n} respectively. Similar notation will be used for such maximum elements. Definition 2.3. The bi-non-crossing Mobius function is the map µBNC : [n∈N [χ∈{l,r}n BNC(χ) × BNC(χ) → C defined recursively by Xρ∈BNC(n) τ ≤ρ≤λ µBNC(τ, ρ) = Xρ∈BNC(n) τ ≤ρ≤λ µNC(ρ, λ) =(1, 0, if τ = λ if τ < λ whenever τ ≤ λ, while taking the zero value otherwise. The connection between the bi-non-crossing Mobius function and the Mobius function on the lattice of non-crossing partitions µNC is given by the formula µBNC(τ, λ) = µNC(s−1 χ · τ, s−1 χ · λ) for all τ ≤ λ ∈ BNC(χ) and hence µBNC inherits many of the multiplicative properties of µNC (see [CNS15b, Section 3]). The Catalan numbers {Cn}n∈N form a sequence of positive integers one of whose many equivalent definitions concerns the equality of the n-th Catalan number with the number of non-crossing partitions on a set of n-elements (and, as a result, with the number of bi- non-crossing partitions with respect to any map χ ∈ {l, r}n). This sequence will come up when we make reference to the joint ∗-distribution of bi-Haar unitary pairs of operators (Corollary 2.18). We state the following lemma tying the values of the bi-non-crossing Mobius function with the Catalan numbers. Lemma 2.4. Let n ∈ N and χ ∈ {l, r}n. Then, for all τ ∈ BNC(χ) we have that In particular, µBNC(0χ, τ ) = YV ∈τ (−1)V −1 · CV −1. µBNC(0χ, 1χ) = (−1)n−1 · Cn−1, where Cn denotes the n-th Catalan number. Due to the connection between µBNC and µNC, the proof of the aforementioned lemma is based on facts regarding the behaviour of multiplicative functions on NC(n). More specifically, it relies on the canonical factorization of intervals in the lattice of non-crossing partitions and on the multiplicative properties of the Mobius function µNC (see [NS06, Theorem 9.29, Proposition 10.14 and 10.15]). 5 Kreweras complementation map KNC : NC(n) → NC(n) defined in [Kre72] is an im- portant example of a lattice anti-isomorphism. For its descripition, we introduce new symbols 1, 2, . . . , n and consider them interlaced with 1, 2, . . . , n in the following manner: For π ∈ NC(n), its Kreweras complement KNC(π) ∈ NC({1, 2, . . . , n}) ∼= NC(n) is defined to be the largest non-crossing partition having the property 1 1 2 2 . . . n n. π ∪ KNC(π) ∈ NC({1, 1, 2, 2 . . . n, n}). The complementation map found its generalization for the lattice of bi-non-crossing par- titions in [CNS15b, Section 5]. Specifically, for any n ∈ N, χ ∈ {l, r}n and τ ∈ BNC(χ), the Kreweras complement of τ in BNC(χ), denoted by KBNC(τ ), is defined as KBNC(τ ) = sχ · KNC(s−1 χ · τ ) is given by applying the permutation sχ to the Kreweras complement of s−1 χ · τ in i.e. NC(n). Note that in the special case when χ ∈ {l, r}n gives the constant value "l", one obtains KNC. In the following lemma, we list properties of KBNC that we will be making use of. Lemma 2.5. Let n ∈ N and χ ∈ {l, r}n. Then: (a) KBNC : BNC(χ) → BNC(χ) is a bijection, (b) For all τ, λ ∈ BNC(χ) we have that τ ≤ λ ⇐⇒ KBNC(λ) ≤ KBNC(τ ) ⇐⇒ K −1 BNC(λ) ≤ K −1 BNC(τ ), (c) KBNC(0χ) = 1χ and KBNC(1χ) = 0χ. All of these properties are easily verified by the definition of KBNC and by the corre- sponding properties which hold for KNC. We shall now state a combinatorial lemma, which may be of independent interest and involves the following cancellation property for the lattice of bi-non-crossing partitions. This lemma will play a key role in the proof of Lemma 4.3. Lemma 2.6. Let n ∈ N, χ ∈ {l, r}n and consider a family {dτ }τ ∈BNC(χ) of free indeter- minates indexed by the bi-non-crossing partitions BNC(χ). Then, the following holds: Xτ ∈BNC(χ)(cid:16)µBNC(0χ, τ ) · Xλ∈BN C(χ) λ≤KBNC(τ ) dλ(cid:17) = d1χ Proof. Re-arragning the left hand-side of the above expression yields: Xτ ∈BNC(n)(cid:16)µBNC(0χ, τ ) · Xλ∈BNC(χ) λ≤KBNC(τ ) dλ(cid:17) = Xλ∈BNC(χ)(cid:16)dλ · Xτ ∈BNC(χ) λ≤KBNC(τ ) µBNC(0χ, τ )(cid:17) With this remark in hand, it is immediate that to prove the conclusion of the lemma, it suffices to show that for all λ ∈ BNC(χ), we have that µBNC(0χ, τ ) =(1, 0, if λ = 1χ if λ < 1χ Xτ ∈BNC(χ) λ≤KBNC(τ ) 6 We simply state that this condition must also be necessary, because the indeterminates {dτ } satisfy no relations. Fix λ ∈ BNC(χ) and let λ′ ∈ BNC(χ) be such that λ = KBNC(λ′). Observe that since λ ≤ KBNC(τ ) ⇐⇒ KBNC(λ′) ≤ KBNC(τ ) ⇐⇒ τ ≤ λ′, we have that {τ ∈ BNC(χ) : λ ≤ KBNC(τ )} = {τ ∈ BNC(χ) : τ ≤ λ′}. Elementary properties of the Mobius function on the lattice of bi-non-crossing partitions imply that Xτ ∈BNC(χ) λ≤KBNC(τ ) µBNC(0χ, τ ) = Xτ ∈BNC(χ) 0χ≤τ ≤λ′ µBNC(0χ, τ ) =(1, 0, if 0χ = λ′ if 0χ < λ′ Then, an application of Lema 2.5 yields: 0χ = λ′ ⇐⇒ K −1 BNC(1χ) = K −1 BNC(λ) ⇐⇒ λ = 1χ and 0χ < λ′ ⇐⇒ K −1 BNC(1χ) < K −1 BNC(λ) ⇐⇒ λ < 1χ. This completes the proof. Of course, when the map χ ∈ {l, r}n gives the constant value "l", one obtains the analogous result for the lattice of non-crossing partitions. 2.2. Bi-Free Independence and Bi-Free Cumulants We begin by recalling the notion of bi-free independence for pairs of faces in some non- commutative ∗-probability space, originally developed in [Voi14]. Definition 2.7. Let (A, ϕ) be a non-commutative ∗-probability space. (i) A pair of faces in (A, ϕ) consists of a pair (C, D) of unital subalgebras of A. (ii) A family {(Ck, Dk)}k∈K of pairs of faces in (A, ϕ) is said to be bi-freely independent (or simply bi-free) if there exists a family of vector spaces with specified vector states {(Xk, ◦ X k, ξk)}k∈K and unital homomorphisms lk : Ck → L(Xk) and rk : Dk → L(Xk), (where L(Xk) denotes the space of all linear maps on Xk) such that the joint distri- bution of the family {(Ck, Dk)}k∈K with respect to ϕ coincided with the joint distri- ◦ X k, ξk). bution with respect to the vacuum state on the representation on ∗k∈K (Xk, (iii) If Sk and Vk are subsets of A for all k ∈ K, then the family {(Sk, Vk)}k∈K will be said to be bi-free if the family of pairs of faces {(alg(1A ∪ Sk), alg(1A ∪ Vk))}k∈K is bi-free. 7 (iv) If Sk and Vk are subsets of A for all k ∈ K, then the family {(Sk, Vk)}k∈K will be said to be ∗-bi-free if the family {(Sk ∪ S∗ k, Vk ∪ V ∗ k )}k∈K is bi-free. The bi-free cumulant function is the main combinatorial tool utilised in bi-free proba- bility theory and its definition is given below. Definition 2.8. Let (A, ϕ) be a non-commutative ∗-probability space. The bi-free cumu- lant function is the map defined by κ : [n∈N [χ∈{l,r}n κχ,τ (a1, . . . , an) = Xλ∈BN C(χ) λ≤τ BNC(χ) × An → C ϕλ(a1, . . . , an)µBNC(λ, τ ) for each n ∈ N, χ ∈ {l, r}n, τ ∈ BNC(χ) and a1, . . . , an ∈ A. The previous formula is called the moment-cumulant formula and an application of Mobius inversion yields that we must also have that ϕ(a1 · · · an) = Xτ ∈BNC(χ) κχ,τ (a1, . . . , an). It is clear that for n ∈ N, χ ∈ {l, r}n and τ ∈ BNC(χ), the bi-free cumulant map κχ,τ : An → C is multilinear. Multiplicative properties of the bi-free cumulant function yield that In the special case when τ = 1χ, we will denote κχ,1χ simply by κχ. κχ,τ (a1, . . . , an) = YV ∈τ κχV ((a1, . . . , an)V ), for all n ∈ N, χ ∈ {l, r}n, τ ∈ BNC(χ) and a1, . . . , an ∈ A. See [CNS15b] for proofs and discussions on all the aforementioned properties. Note that the result of reading the sequence (a1, . . . , an)V with the indices in the induced χV -order coincides with first reading the sequence (a1, . . . , an) with the indices in the χ-order and then restricting the resulting sequence to s−1 χ (V ). Observe that the moment-cumulant formula implies that for elements X, Y, Z, W ∈ A, then the joint ∗-distribution of the pair (X, Y ) coincides with the joint ∗-distribution of (Z, W ) if and only if all bi-free cumulants with entries in the set {X, X ∗, Y, Y ∗} yield equal values to all bi-free cumulants with entries in the set {Z, Z ∗, W, W ∗}. The following theorem displays the equivalent combinatorial characterization of bi-free independence. Theorem 2.9 ([CNS15b], Theorem 4.3.1). Let (A, ϕ) be a non-commutative ∗-probability space and let {(Ck, Dk)}k∈K be family of pairs of faces in A. The following are equivalent: (i) the family {(Ck, Dk)}k∈K is bi-free, 8 (ii) for all n ∈ N, χ ∈ {l, r}n, a1, . . . , an ∈ A and non-constant map ǫ : {1, . . . , n} → K such that we have that ai ∈ Cǫ(i), Dǫ(i), if χ(i) = l if χ(i) = r (i = 1, . . . , n) κχ(a1, . . . , an) = 0. Considering that a number of the central results of this paper involve products of pairs of operators, the following theorem concerning bi-free cumulants having products of operators as arguments will be used numerous times throughout this manuscript. Theorem 2.10 (Scalar case of [CNS15a], Theorem 9.1.5). Let (A, ϕ) be a non-commutative ∗-probability space, m < n ∈ N, χ ∈ {l, r}m and integers Also, let a1, . . . , an ∈ A. Then, by defining χ ∈ {l, r}n by k(0) = 0 < k(1) < . . . < k(m) = n. χ(q) = χ(pq) with pq being the unique number in {1, . . . , m} such that k(pq − 1) < q ≤ k(pq), we have that : κχ(a1 · · · ak(1), ak(1)+1 · · · ak(2), . . . , ak(m−1)+1 · · · ak(m)) = Xτ ∈BNC( χ) where c0χ = {{k(p − 1) + 1, . . . , k(p)} : p = 1, . . . , m} ∈ BNC( χ). i = 1, . . . , m, then c0χ = s χ(c0χ). We find it convenient to state and prove the following proposition, concerning bi-non-crossing partitions whose blocks have to connect consecu- tive indices in the χ-order. In sections 3 and 4, when discussing the behaviour of products of pairs of operators, the forward direction of this proposition will be used frequently in combination with Theorem 2.10. Note that in the case when there exists t ∈ N such that k(i) = k(i − 1) + t for all κ χ,τ (a1, . . . , an), τ ∨c0χ=1 χ Proposition 2.11. Let n ∈ N and χ ∈ {l, r}2n such that χ(2i − 1) = χ(2i) for every partition τ ∈ BNC( χ), the following are equivalent: (ii) s χ(1)∼τ s χ(2n) and s χ(2i)∼τ s χ(2i + 1) for every i = 1, . . . , n − 1. i ∈ {1, . . . , n}. Also, let c0χ = {{2i − 1, 2i} : i = 1, . . . , n}. Then, for a bi-non-crossing (i) τ ∨c0χ = 1 χ and every block of τ contains an even number of elements, Proof. Since c0χ = s χ(c0χ), it is clear that clause (ii) above implies clause (i). Now, let τ ∈ BNC( χ) be such that τ ∨c0χ = 1 χ and every block of τ contains an even number of elements and let V ∈ τ such that s χ(1) ∈ V (equivalently 1 = min< s−1 q ∈ {1, . . . , 2n} such that s χ(q) = max≺ χV (equivalently q = max<s−1 q must be an even number. χ (V )). Also, let χ (V )). We claim that Indeed, by way of contradiction, suppose that q = 2m − 1 for some m ∈ {2, . . . , n}. We remark that V cannot be equal to {s χ(i) : 1 ≤ i ≤ 2m − 1} since V must contain an even number of elements. Notice that if 2 ≤ p ≤ 2m − 2 is such that s χ(p) /∈ V and V ′ ∈ τ is such that s χ(p) ∈ V ′, then we necessarily must have that V ′ ⊆ {s χ(i) : 2 ≤ i ≤ 2m − 2}; for if there exists i ≥ 2m with s χ(i) ∈ V ′ then we obtain that 9 1 = min<s χ −1(V ) , 2m − 1 = max<s χ −1(V ) and p, i ∈ s χ −1(V ′) with 2 ≤ p ≤ 2m − 2 and 2m ≤ i, which contradicts the fact that s−1 χ · τ ∈ NC(2n). This shows that the set {s χ(i) : 1 ≤ i ≤ 2m − 1}, whose cardinality is obviously odd must be written as a union of blocks of τ, thus τ must contain at least one block with an odd number of elements, contradicting our initial assumption. Hence q = 2m for some m ∈ {1, . . . , n}. If m < n, then let eV = {s χ(i) : 1 ≤ i ≤ 2m} and define λ = {eV ,(cid:0)eV(cid:1)c λ ∈ BNC( χ), V ⊆ eV and that τ,c0χ ≤ λ (cid:12) 1 χ, thus the condition τ ∨c0χ = 1 χ cannot be satisfied. Hence, we must have that q = 2n and this implies that s χ(1)∼τ s χ(2n). Now let i ∈ {1, . . . , n − 1} and V ∈ τ such that s χ(2i) ∈ V. Assume that s χ(2i + 1) /∈ V χ · λ = {1, 2, . . . , 2m} ∪ {2m + 1, . . . , 2n}, we have that }. Since s−1 and we will distinguish between two possibilities: First, let us suppose that s χ(2i) = max≺ χV and let q ∈ {1, . . . , 2n} be such that s χ(q) = min≺ χ(V ). Then, arguing as before, we deduce that we must have q = 2p − 1 for some 1 ≤ p ≤ i (otherwise, if q = 2p with 1 ≤ p < i, then the cardinality of the set {s χ(j) : 2p ≤ j ≤ 2i} is odd and thus τ contains at least one block with an odd number of elements which }, since s−1 χ · λ = {1, . . . , 2(p − 1)} ∪ {2p − 1, . . . , 2i} ∪ {2i + 1, . . . , 2n}, of course cannot happen). But then, by setting eV = {s χ(j) : 2p − 1 ≤ j ≤ 2i} and λ = {eV ,(cid:0)eV(cid:1)c we obtain that λ ∈ BNC( χ), V ⊆ eV and τ,c0χ ≤ λ (cid:12) 1 χ, a contradiction. This shows that it cannot be the case that s χ(2i) = max≺ χ V and hence, there must exist q ∈ {1, . . . , 2n} such that s χ(q) ∈ V and s χ(2i)≺ χs χ(q). Without loss of generality, we may assume that for every v ∈ V \ {s χ(2i), s χ(q)}, we either have that v ≺ χ s χ(2i) or s χ(q) ≺ χ v (i.e. we may assume that s χ(q) is the χ-minimum element of V with this property). If q = 2m for some q ≥ i + 1, then the set {s χ(j) : 2i + 1 ≤ j ≤ 2m − 1} is non-empty and contains an odd number of elements. Thus, arguing as before, it must be written as a union of blocks of τ , which implies that at least one block of τ contains an odd number of elements, a contradiction. If q = 2m − 1 for some m > (i + 1), then the set eV = {s χ(j) : 2i + 1 ≤ j ≤ 2(m − 1)} is non-empty and contains an even number of elements. Let λ = eV ∪(cid:0)eV(cid:1)c V ⊆(cid:0)eV(cid:1)c , it follows that λ ∈ BNC( χ) and τ,c0χ ≤ λ (cid:12) 1 χ, a contradiction. This shows that we must have s χ(2i)∼τ s χ(2i + 1) and this completes the proof. . Then, since 10 2.3. Bi-Haar Unitary Pairs of Operators R-diagonal operators are characterized by having all of their free ∗-cumulants that are either of odd order, or have entries that are not alternating in ∗-terms and non-∗-terms equal to zero. Adopting the combinatorial approach in the bi-free setting, we shall now give the definition of bi-R-diagonal pairs of operators, which will be the central focus of this paper. This definition was first proposed as the correct bi-free generalization of R- diagonal elements in [Sko16, Section 4], but was only utilised to yield examples of R-cyclic pairs of matrices (see Proposition 2.21). Definition 2.12. Let (A, ϕ) be a non-commutative ∗-probability space and X, Y ∈ A. We say that the pair (X, Y ) is bi-R-diagonal if for every n ∈ N, χ ∈ {l, r}n and a1, . . . , an ∈ A such that {X, X ∗}, if χ(i) = l {Y, Y ∗}, if χ(i) = r (i = 1, . . . , n) ai ∈ we have that: (i) κχ(a1, . . . , an) = 0, if n is odd (ii) κχ(a1, . . . , an) = 0, if n is even and the sequence (asχ(1), . . . , asχ(n)) is not in one of the following forms: (a) (Z, Z ∗, . . . , Z, Z ∗), with Z ∈ {X, X ∗, Y, Y ∗}, (b) (X, X ∗, . . . , X, X ∗, Y, Y ∗, . . . , Y, Y ∗), (c) (X ∗, X, . . . , X ∗, X, Y ∗, Y, . . . , Y ∗, Y ), (d) (X, X ∗, . . . , X, X ∗, X, Y ∗, Y, . . . , Y ∗, Y, Y ∗), (e) (X ∗, X, . . . , X ∗, X, X ∗, Y, Y ∗, . . . , Y, Y ∗, Y ), i.e. whenever the sequence (a1, . . . , an) is not alternating in ∗-terms and non ∗-terms when read with the indices in the χ-order, with any number of X-terms followed by any number of Y -terms. It is clear from the definition that if the map χ is constant, then bi-free cumulants reduce to free cumulants and all free cumulants with entries in either {X, X ∗} (if the map χ yields the constant value "l") or {Y, Y ∗} (if the map χ yields the constant value "r") that are of odd order or are not alternating in ∗-terms and non-∗-terms are equal to zero. In particular, if (X, Y ) is a bi-R-diagonal pair, then both X and Y are R-diagonal operators. Also, it is immediate from the moment-cumulant formula that all joint ∗-moments of odd order of a bi-R-diagonal pair are equal to zero, i.e. if the pair (X, Y ) is bi-R-diagonal, then for all k ∈ N and a1, . . . , a2k+1 ∈ {X, X ∗, Y, Y ∗}, it follows that ϕ(a1 · · · a2k+1) = 0. In analogy to the case of free probability and free Haar unitaries, we will define the notion of a bi-Haar unitary pair of operators and compute its bi-free ∗-cumulants. Bi-Haar unitary pairs will act as both the prototypical examples and building blocks of bi-R-diagonal pairs (see Theorem 4.4). First, we recall the definition of a free Haar unitary. Definition 2.13. Let (B, ψ) be a non-commutative ∗-probability space. A unitary v ∈ B is called a free Haar unitary if for all n ∈ Z we have that: ψ(vn) =(1, 0, if n = 0, otherwise. 11 The free ∗-cumulants of a Haar unitary are computed as follows: Proposition 2.14 ([NS06], Proposition 15.1). If v ∈ (B, ψ) is a free Haar unitary, then for every n ∈ N, the non-vanishing free ∗-cumulants of v are given by: κ2n(v, v∗, . . . , v, v∗) = κ2n(v∗, v, . . . , v∗, v) = (−1)n−1 · Cn−1, where Cn denotes the n-th Catalan number. All other free cumulants with entries in the set {v, v∗} vanish. The bi-free generalization of the notion of a Haar unitary was first proposed in [CNS15a, Definition 10.1.2] in the operator-valued setting. Definition 2.15. Let (A, ϕ) be a non-commutative ∗-probability space and ul, ur be unitaries in A. The pair (ul, ur) is called a bi-Haar unitary pair if the following hold: (i) the algebras alg({ul, ul ∗}) and alg({ur, ur ∗}) commute, (ii) for all n, m ∈ Z we have that ϕ(un l · um r ) =(1, 0, if n + m = 0, otherwise. In particular, if the pair (ul, ur) is a bi-Haar unitary, then both ul and ur are free Haar unitaries. Example 2.16. Let G be a group with identity e that contains an element of infinite order (i.e. there exists g0 ∈ G such that gn 0 6= e for all n ∈ Z \ {0}). If λ : G → B(ℓ2(G)) and ρ : G → B(ℓ2(G)) denote the left and right regular representations of G respectively, then it is straightforward to verify that the pair (λ(g0), ρ(g−1 0 )) is a bi-Haar unitary pair, with respect to the vector state corresponding to the identity element of G. In particular, if u denotes the bilateral shift on ℓ2(Z), then the pair (u, u) is a bi-Haar unitary. The commutation assumption on the left and right operators of a bi-Haar unitary pair allows one to reduce the computation of its bi-free cumulants to computing free cumulants of a free Haar unitary. In particular, we have the following: Proposition 2.17. Let (A, ϕ), (B, ψ) be non-commutative ∗-probability spaces and let (ul, ur) be a bi-Haar unitary pair in A and v ∈ B a free Haar unitary. For n ∈ N and χ ∈ {l, r}n, let a1, . . . , an ∈ A such that for all i = 1, . . . , n ai ∈({ul, ul {ur, ur ∗}, ∗}, if χ(i) = l if χ(i) = r and define b1, . . . , bn ∈ B by bi =(v, v∗, if asχ(i) ∈ {ul, ur} ∗, ur if asχ(i) ∈ {ul ∗} for all i = 1, . . . , n. Then, we have that κχ(a1, . . . , an) = κn(b1, . . . , bn), with the quantity on the left-hand side of the equation being a bi-free cumulant and the one on the right-hand side being a free cumulant. 12 Proof. For n, m ∈ Z, the following relation between the joint ∗-moments of (ul, ur) and the ∗-moments of v is immediate by Definitions 2.13 and 2.15 ϕ(ul n · ur m) = ψ(vn+m). and, since the algebras alg({ul, u∗ r}) commute, every joint ∗-moment of the pair (ul, ur) factorizes in a moment that has a form similar to the left hand-side of the previous expression. The moment-cumulant formulas yield that l }) and alg({ur, u∗ and κχ(a1, . . . , an) = Xτ ∈BNC(χ) κn(b1, . . . , bn) = Xπ∈NC(n) ϕτ (a1, . . . , an)µBNC(τ, 1χ), ψπ(b1, . . . , bn)µNC(π, 1n). The main observation needed to lead us to the conclusion of the proof is that for all τ ∈ BNC(χ) and for all V ∈ τ , we have that ϕ((a1, . . . , an)V ) = ψ((b1, . . . , bn)s−1 χ (V )). Indeed, let τ ∈ BNC(χ) and V ∈ τ. Define the sets and I1 = {i ∈ V : ai = ul}, I2 = {i ∈ V : ai = u∗ l }, I3 = {i ∈ V : ai = ur}, I4 = {i ∈ V : ai = u∗ r}. Also, let ni ∈ N to be equal to the cardinality of the set Ii, for all i = 1, 2, 3, 4. Then, by the definition of b1, . . . , bn, we have that ϕ((a1, . . . , an)V ) = ϕ(un1−n2 l un3−n4 r ) = ψ(vn1+n3−n2−n4) = ψ((b1, . . . , bn)s−1 χ (V )). Hence, this implies that for any bi-non-crossing partition τ ∈ BNC(χ) we obtain ϕτ (a1, . . . , an)µBNC(τ, 1χ) = ϕτ (a1, . . . , an)µNC(s−1 χ ·τ, 1n) = ψs−1 χ ·τ (b1, . . . , bn)µNC(s−1 χ ·τ, 1χ). This completes the proof. A combination of Propositions 2.14 and 2.17 gives a complete computation of the bi-free cumulants involving a bi-Haar unitary pair. Corollary 2.18. Let (A, ϕ) be a non-commutative ∗-probability space and (ul, ur) a bi- Haar unitary pair in A. Also, let n ∈ N, χ ∈ {l, r}2n and a1, . . . , a2n ∈ A such that (a) for all i = 1, . . . , 2n we have {ul, ul ∗}, {ur, ur ∗}, if χ(i) = l if χ(i) = r ai ∈ (b) the sequence (asχ(1), . . . , asχ(2n)) is alternating in ∗-terms and non-∗-terms. Then, κχ(a1, . . . , a2n) = (−1)n−1 · Cn−1, where Cn denotes the n-th Catalan number. All other bi-free cumulants with entries in the set {ul, ul r} vanish. In particular, the pair (ul, ur) is bi-R-diagonal. ∗, ur, u∗ 13 2.4. Operator-Valued Bi-Free Independence and R-cyclic Pairs of Ma- trices In the spirit of [CNS15a] and [Sko16], we will present the basic definitions regarding operator-valued bi-free independence and a number of results concerning R-cyclic pairs of matrices. The results that are cited will be used in Section 4 to discuss an equivalent characterization of the condition of bi-R-diagonality, which will be formulated in terms of the bi-freeness of certain matrix pairs from scalar matrices with amalgamation over the diagonal scalar matrices (see Theorem 4.6). Definition 2.19. Let B be a unital algebra. (i) A B-B-bimodule with specified B-vector state is a triple (X , ◦ X , p) where X is a direct sum of B-B-bimodules X = B ⊕ ◦ X and p : X → B is the linear map given by p(b ⊕ η) = b for all b ∈ B and η ∈ ◦ X . (ii) A B-B-non commutative probability space is a triple (A, EA, ε), where A is a unital algebra, ε : B ⊗ Bop → A is a unital homomorphism such that both maps εB⊗1B and ε1B ⊗Bop are injective and EA : A → B is a linear map such that and EA(ε(b1 ⊗ b2)Z) = b1EA(Z)b2, EA(Zε(b ⊗ 1B)) = EA(Zε(1B ⊗ b)), for all b, b1, b2 ∈ B and Z ∈ A. The unital subalgebras of A defined as Al = {Z ∈ A : Zε(1B ⊗ b) = ε(1B ⊗ b)Z for all b ∈ B}, and Ar = {Z ∈ A : Zε(b ⊗ 1B) = ε(b ⊗ 1B)Z for all b ∈ B}, are called the left and right algebras of A respectively. (iii) A pair of B-faces in a B-B-non commutative probability space (A, EA, ε) consists of a pair (C, D) of unital subalgebras of A such that ε(B ⊗ 1B) ⊆ C ⊆ Al and ε(1B ⊗ Bop) ⊆ D ⊆ Ar. (iv) A family {(Ck, Dk)}k∈K of pairs of B-faces in a B-B-non commutative probabil- ity space (A, EA, ε) is said to be bi-free with amalgamation over B if there exist ◦ B-B-bimodules with specified B-vector states {(Xk, Xk, pk)}k∈K and unital homo- morphisms lk : Ck → Ll(Xk) and rk : Dk → Lr(Xk) such that the joint distribution of {(Ck, Dk)}k∈K with respect to EA is equal to the joint distribution of the images of {((λk ◦ lk)(Ck), (ρk ◦ rk)(Dk))}k∈K inside L(∗k∈KXk) with respect to EL(∗k∈K Xk), where λk and ρk denote the left and right regular representations onto Xk ⊆ ∗k∈K Xk, respectively. 14 If Sk ⊆ Al and Vk ⊆ Ar for all k ∈ K, we will say that the family {(Sk, Vk)}k∈K is bi-free with amalgamation over B if the family {(alg(ε(B ⊗ 1B) ∪ Sk), alg(ε(1B ⊗ Bop) ∪ Vk))}k∈K of pairs of B-faces is bi-free with amalgamation over B. See [CNS15a, Section 3] for a discussion on why B-B-non-commutative probability spaces are the correct framework to formulate the notions of operator-valued bi-free prob- ability. There, a combinatorial approach was adopted and the bi-multiplicative operator- valued bi-free cumulant maps were defined and used to characterize operator-valued bi-free independence. Let (A, ϕ) be a non-commutative ∗-probability space and let d ∈ N. In the algebra Md(A) of all d×d matrices over A, consider the unital subalgebras Md(C) and D consisting of all scalar matrices and all diagonal scalar matrices respectively. We will recall from [Sko16, Section 4] the process regarding how to turn L(Md(A)), the space of all linear maps on Md(A), into a D-D-non-commutative probability space. We will denote by [ai,j] a matrix whose (i, j)th entry equals ai,j. Let F : Md(C) → D denote the conditional expectation onto the diagonal and define the unital, linear map ϕd : Md(A) → Md(C) by ϕd([Ti,j]) = [ϕ(Ti,j)] for all [Ti,j] ∈ Md(A). Also, for [ai,j] ∈ D, let L[ai,j]([Ti,j]) =" dXk=1 ai,kTk,j# and R[ai,j ]([Ti,j]) =" dXk=1 ak,jTi,k#, for all [Ti,j] ∈ Md(A). Then, if ε : D ⊗ Dop → L(Md(A)) is defined as ε([ai,j] ⊗ [a′ i,j]) = L[ai,j]R[a′ i,j ] ([ai,j], [a′ i,j] ∈ D) and E : L(Md(A)) → Md(C) is defined as E(Z) = ϕd(Z(Id)) (Z ∈ L(Md(A))) where Id denotes the d × d identity matrix, we have that the triple (L(Md(A)), F ◦ E, ε) is a D-D-non-commutative probability space. We will also need the unital homomorphisms L : Md(A) → L(Md(A))l and R : Md(Aop)op → L(Md(A))r given by L([Zi,j])[Ti,j] =" dXk=1 Zi,kTk,j# and R([Zi,j])[Ti,j] =" dXk=1 Zk,jTi,k#, for all [Zi,j], [Ti,j] ∈ Md(A). In the setting of free probability, there is a connection between R-diagonal operators and R-cyclic matrices (see [NS06, Example 20.5]). In the bi-free setting, R-cyclic pairs of matrices were first defined and studied in [Sko16]. Definition 2.20. [Sko16, Definition 4.4] Let (A, ϕ) be a non-commutative ∗-probability space, d ∈ N, I, J be disjoint index sets and let {[Zk;i,j]}k∈I ∪ {[Zk;i,j]}k∈J ⊆ Md(A). The 15 pair ({[Zk;i,j]}k∈I , {[Zk;i,j]}k∈J ) is called R-cyclic if for all n ∈ N, ω : {1, . . . , n} → I ⊔ J and 1 ≤ i1, . . . , in, j1, . . . , jn ≤ d, by defining χ ∈ {l, r}n as χ(i) = l, r, we have that if ω(i) ∈ I if ω(i) ∈ J (i = 1, . . . , n) κχ(Zω(1);i1,j1, Zω(2);i2 ,j2, . . . , Zω(n);in,jn) = 0 whenever at least one of the relations jsχ(1) = isχ(2), jsχ(2) = isχ(3), . . . , jsχ(n−1) = isχ(n), jsχ(n) = isχ(1) is not satisfied. The following result was mentioned (but not proved) in [Sko16, Section 4] and we include the proof for the convenience of the reader. Proposition 2.21. Let (A, ϕ) be a non-commutative ∗-probability space and let X, Y ∈ A. The following are equivalent: (i) the pair (X, Y ) is bi-R-diagonal, (ii) in M2(A), the pair ([Zi,j]1≤i,j≤2, [Wi,j]1≤i,j≤2) defined as [Zi,j]1≤i,j≤2 =(cid:20) 0 X 0(cid:21) and [Wi,j]1≤i,j≤2 =(cid:20) 0 X ∗ Y 0(cid:21) Y ∗ is R-cyclic. Proof. Let n ∈ N, χ ∈ {l, r}n and ai1,j1, . . . , ain,jn ∈ A such that Zim,jm for some 1 ≤ im, jm ≤ 2, Wim,jm for some 1 ≤ im, jm ≤ 2, if χ(i) = l if χ(i) = r (m = 1, . . . , n) aim,jm = If there exists m ∈ {1, . . . , n} such that im = jm, then aim,jm = 0 and this implies that the cumulant κχ(ai1,j1, . . . , ain,jn) vanishes. Hence we can assume that im 6= jm for all m ∈ {1, . . . , n} and, in this case, the bi-free cumulant κχ(ai1,j1, . . . , ain,jn) has entries in the set {X, X ∗, Y, Y ∗}. The main observation that will make the equivalence of the proposition apparent is that the condition that at least one of the relations jsχ(1) = isχ(2), jsχ(2) = isχ(3), . . . , jsχ(n−1) = isχ(n), jsχ(n) = isχ(1) is not satisfied is equivalent to the statement that the sequence (aisχ(1),jsχ(1), . . . , aisχ(n),jsχ(n)) is either not alternating in ∗-terms and non-∗-terms, or is of odd length. Indeed, first suppose that jsχ(m) 6= isχ(m+1) for some m ∈ {1, . . . , n − 1} and notice that this implies that we must have isχ(m) = isχ(m+1) and jsχ(m) = jsχ(m+1). 16 But this is equivalent to stating that the elements aisχ(m),jsχ(m) and aisχ(m+1),jsχ(m+1) both correspond to either ∗-terms or non-∗-terms and hence the sequence (aisχ(1),jsχ(1), . . . , aisχ(n),jsχ(n)) is not alternating in ∗-terms and non-∗-terms. Next, assume that jsχ(n) 6= isχ(1). As before, we must have that isχ(1) = isχ(n) and jsχ(1) = jsχ(n). This is equivalent to stating that the first and last terms of the sequence (aisχ(1),jsχ(1), . . . , aisχ(n),jsχ(n)) both correspond to either ∗-terms or non-∗-terms, which means that this sequence either is not alternating in ∗-terms and non-∗-terms, or is of odd length. The main result we will need for Theorem 4.6 concerns the following equivalent char- acterization of R-cyclic pairs. Theorem 2.22 ([Sko16], Theorem 4.9). Let (A, ϕ) be a non-commutative ∗-probability space, d ∈ N, I, J be disjoint index sets and let {[Zk;i,j]}k∈I ∪ {[Zk;i,j]}k∈J ⊆ Md(A). The following are equivalent: (i) the pair(cid:0){[Zk;i,j]}k∈I , {[Zk;i,j]}k∈J(cid:1) is R-cyclic, (ii) in the D-D-non-commutative probability space (L(Md(A)), F ◦ E, ε), we have that the family is bi-free from with amalgamation over D. (cid:0)({L([Zk;i,j])}k∈I, {R([Zk;i,j])}k∈J(cid:1) (cid:0)L(Md(C)), R(Md(C)op)(cid:1) 3. Operations Involving Bi-R-Diagonal Pairs In this section, we will study the behaviour of bi-R-diagonal pairs of operators under the taking of sums, products and arbitrary powers, where, in most cases, a ∗-bi-free independence condition will be assumed. The proofs obtained will indicate that most of the results that hold for free R-diagonal elements (see [NS97] and [NS06, Lecture 15]) have corresponding generalizations in the bi-free setting. We begin with the following proposition regarding sums of ∗-bi-free bi-R-diagonal pairs. Proposition 3.1. Let (A, ϕ) be a non-commutative ∗-probability space and X, Y, Z, W ∈ A such that: (a) the pairs (X, Y ) and (Z, W ) are both bi-R-diagonal, (b) the pairs (X, Y ) and (Z, W ) are ∗-bi-free. Then, the pair (X + Z, Y + W ) is also bi-R-diagonal. 17 Proof. Let n ∈ N, χ ∈ {l, r}n and a1, . . . , an ∈ A such that {X + Z, X ∗ + Z ∗}, if χ(i) = l {Y + W, Y ∗ + W ∗}, if χ(i) = r (i = 1, . . . , n) Define b1, . . . , bn, c1, . . . , cn ∈ A by ai ∈  X, X ∗, Y, Y ∗, bi = if ai = X + Z if ai = X ∗ + Z ∗ if ai = Y + W if ai = Y ∗ + W ∗ and ci = Z, Z ∗, W, W ∗, if ai = X + Z if ai = X ∗ + Z ∗ if ai = Y + W if ai = Y ∗ + W ∗  for each i = 1, . . . , n. Then, the multi-linearity of the bi-free cumulants maps combined with the ∗-bi-free independence condition yield that κχ(a1, . . . , an) = κχ(b1, . . . , bn) + κχ(c1, . . . , cn). The conclusion of the proposition follows from the observation that the sequence (asχ(1), . . . , asχ(n)) is alternating in ∗-terms and non-∗-terms if and only if both the sequences (bsχ(1), . . . , bsχ(n)) and (csχ(1), . . . , csχ(n)) are also alternating in ∗-terms and non-∗-terms. With the previous proof in mind, it is easy to see that if exactly one of the ∗-bi- free pairs (X, Y ) and (Z, W ) is bi-R-diagonal, then the pair (X + Z, Y + W ) cannot be bi-R-diagonal. We now proceed to study various cases on products involving bi-R-diagonal pairs. The products of pairs will be considered pointwise, with the condition that the order of the right operators is reversed being necessary for the results concerning the more general cases (see Theorem 3.2 below and also Proposition 4.2). The proofs of these results will require more delicate arguments when compared to the cases of sums involving bi-R- diagonal pairs and, for this, the formula for bi-free cumulants with products of operators as arguments will play a key role. The next theorem states that the product of a bi-R- diagonal pair of operators by any ∗-bi-free pair is also bi-R-diagonal and exhibits the fact that bi-R-diagonal pairs exist in abundance. Theorem 3.2. Let (A, ϕ) be a non-commutative ∗-probability space and let X, Y, Z, W ∈ A such that: (a) the pair (X, Y ) is bi-R-diagonal, (b) the pairs (X, Y ) and (Z, W ) are ∗-bi-free. Then, the pair (XZ, W Y ) is also bi-R-diagonal. Proof. Let n ∈ N, χ ∈ {l, r}n and a1, . . . , an ∈ A be such that ai ∈ {XZ, Z ∗X ∗}, if χ(i) = l {W Y, Y ∗W ∗}, if χ(i) = r (i = 1, . . . , n) 18 X, Z ∗, W, Y ∗, Z, X ∗, Y, W ∗, if ai = XZ if ai = Z ∗X ∗ if ai = W Y if ai = Y ∗W ∗ if ai = XZ if ai = Z ∗X ∗ if ai = W Y if ai = Y ∗W ∗ c2i = c2i−1 =   κχ(a1, . . . , an) = Xτ ∈BNC( χ) = Xτ ∈BNC( χ) τ ∨c0χ=1 χ for each i = 1, . . . , n. Then, an application of Theorem 2.10 yields: Define χ ∈ {l, r}2n by χ(2i − 1) = χ(2i) = χ(i) for each i = 1, . . . , n and c1, . . . , c2n ∈ A as follows: κ χ,τ (c1, . . . , c2n) YV ∈τ κ χV(cid:0)(c1, . . . , c2n)V(cid:1) (1) (2) τ ∨c0χ=1 χ wherec0χ = {{2i − 1, 2i} : i = 1, . . . , n} ∈ BNC( χ). χ−1({r}) = {j1 < . . . < jn−p}, the definition of χ implies that To start, we make some remarks. First of all, if χ−1({l}) = {i1 < . . . < ip} and χ−1({l}) = {2i1 − 1 < 2i1 < . . . < 2ip − 1 < 2ip} and χ−1({r}) = {2j1 − 1 < 2j1 < . . . < 2jn−p − 1 < 2jn−p}. Thus, if i ∈ {1, . . . , n} with asχ(i) = XZ, then cs χ(2i−1) = X and cs χ(2i) = Z (a similar situation occurs when asχ(i) = Z ∗X ∗, since this corresponds to a left operator). Now if asχ(i) = W Y , then cs χ(2i−1) = Y and cs χ(2i) = W (and a similar situation occurs when asχ(i) = Y ∗W ∗ since this corresponds to a right operator). Note that in the latter case, the right operators must appear reversed in the χ-order. Secondly, in order for a bi-non-crossing partition τ to contribute to the sum appearing : i ∈ V } ⊆ {X, X ∗, Y, Y ∗} or : i ∈ V } ⊆ {Z, Z ∗, W, W ∗}; for if there exists V ∈ τ and i 6= j ∈ V such that in (2), we must have that for every V ∈ τ , either {ci {ci ci ∈ {X, X ∗, Y, Y ∗} and cj ∈ {Z, Z ∗, W, W ∗}, then κ χV(cid:0)(c1, . . . , c2n)V(cid:1) = 0 due to the ∗- bi-free independence condition and thus κ χ,τ (c1, . . . , c2n) vanishes. Note that this implies that if n is odd, then κχ(a1 . . . , an) = 0, as then the cardinality of the set {j ∈ {1, . . . , 2n} : cj ∈ {X, X ∗, Y, Y ∗}} is odd and hence for any τ ∈ BNC( χ) there exists V ∈ τ with odd cardinality that contains indices corresponding to elements in {X, X ∗, Y, Y ∗}. Since the pair (X, Y ) is bi-R-diagonal, all bi-free cumulants of odd order with entries in {X, X ∗, Y, Y ∗} vanish, thus κ χV ((c1, . . . , c2n)V ) = 0. 19 We may now assume that n is even and that every block of a bi-non-crossing partition contains indices corresponding to elements either from {X, X ∗, Y, Y ∗} or {Z, Z ∗, W, W ∗}. We must show that the cumulant κχ(a1, . . . , an) vanishes if the sequence (asχ(1), . . . , asχ(n)) is not alternating in ∗-terms and non-∗-terms. When this occurs, by analysing individual cases, we will show that a given bi-non-crossing partition τ ∈ BNC( χ) either yields zero contribution to the sum appearing in (1), or that the relation τ ∨c0χ = 1 χ cannot be Suppose the following situation occurs: satisfied. (asχ(1), . . . , asχ(n)) = (. . . , XZ, XZ, . . .), with asχ(m) = asχ(m+1) = XZ for some m ∈ {1, . . . , n − 1}. This implies the following situation for the χ-order: (cs χ(1), . . . , cs χ(2n)) = (. . . , X, Z, X, Z, . . .), with cs χ(2m−1) = cs χ(2m+1) = X and cs χ(2m) = cs χ(2m+2) = Z. For τ ∈ BNC( χ), let V ∈ τ be such that s χ(2m + 1) ∈ V . To start, consider the case when s χ(2m + 1) = min≺ χV (equivalently, 2m + 1 = min<s−1 χ (V )). Let q ∈ {1, . . . , 2n} such that s χ(q) = max≺ χV (equivalently, q = max<s−1 χ (V )) and notice that we must have that cs χ(q) ∈ {X ∗, Y ∗}. Indeed, if cs χ(q) ∈ {X, Y }, then the sequence (c1, . . . , c2n)V when read in the induced χV -order would have either one of the forms (X, . . . . . . , X) or (X, . . . . . . , Y ) and would thus not be alternating in ∗-terms and non-∗-terms. Since the pair (X, Y ) is bi-R-diagonal, this would imply that κ χV ((c1, . . . , c2n)V ) = 0 and hence κ χ,τ (c1, . . . , c2n) = 0. We assume that cs χ(q) = Y ∗ (with the case when cs χ(q) = X ∗ handled similarly). The following situation follows: (cs χ(1), . . . , cs χ(2n)) = (. . . , X, Z, X, Z, . . . , W ∗, Y ∗, . . .) Since }. s−1 χ · λ = {{2m + 1, . . . , 2p}, {1, . . . , 2m} ∪ {2p + 1, . . . , 2n}} ∈ NC(2n), and, as such, q = 2p for some p ∈ {m+2, . . . , n}. We will show that the relation τ ∨c0χ = 1 χ cannot be satisfied. Indeed, define eV = {s χ(i) : 2m + 1 ≤ i ≤ 2p} and let λ = {eV ,(cid:0)eV(cid:1)c it follows that λ ∈ BNC( χ). It is easily seen thatc0χ ≤ λ and, moreover, τ ≤ λ holds. To see this, first note that V ⊆ eV . For V ′ ∈ τ with V 6= V ′, we must have that either V ′ ⊆ eV or V ′ ⊆(cid:0)eV(cid:1)c crossing. Hence, we have that τ,c0χ ≤ λ (cid:12) 1 χ and it follows that we cannot have that i ∈ {1, . . . , 2m} ∪ {2p + 1, . . . , 2n} and j ∈ {2m + 2, . . . , 2p − 1}. So, suppose that there exists q ∈ {1, . . . , 2n} with s χ(q) ∈ V and But this cannot happen, since {2m + 1, 2p} ⊆ s−1 ; for otherwise there would exist i 6= j ∈ s−1 χ (V ′) such that χ (V ) and the partition s−1 χ · τ is non- s χ(2m + 1) = min≺ χV . s χ(q)≺ χ s χ(2m + 1). 20 We may moreover assume that for all v ∈ V \ {s χ(2m + 1), s χ(q)}, we either have that v≺ χ s χ(q) or s χ(2m + 1)≺ χ v (i.e. that s χ(q) is the χ-maximum element of V with this property). Notice that it must necessarily be that cs χ(q) = X ∗. Indeed, if not, we would have that cs χ(q) = X and then the sequence (c1, . . . , c2n)V when read in the induced χV -order would be of the form (. . . . . . , X, X, . . . . . .), with this implying that κ χV ((c1, . . . , c2n)V ) = 0, since the pair (X, Y ) is bi-R-diagonal. Thus, cs χ(q) = X ∗ and this yields the following situation (cs χ(1), . . . , cs χ(2n)) = (. . . , Z ∗, X ∗, . . . , X, Z, X, Z, . . .). From this, one sees that q = 2p, for some p ∈ {1, . . . , m − 1}. We will show that once again the relation τ ∨c0χ = 1 χ cannot be satisfied. By defining eV = {s χ(i) : 2p + 1 ≤ i ≤ 2m} (and noting that this set is non-empty), let λ = {eV ,(cid:0)eV(cid:1)c as before, it follows that λ ∈ BNC( χ) and τ,c0χ ≤ λ (cid:12) 1 χ. Hence, when }. Observe that V ⊆(cid:0)eV(cid:1)c (asχ(1), . . . , asχ(n)) = (. . . , XZ, XZ, . . .), and, we obtain that the bi-free cumulant κχ(a1, . . . , an) vanishes and the use of similar argu- ments shows that this is also the case when the sequence (asχ(1), . . . , asχ(n)) has one of the following forms (. . . , XZ, W Y, . . .) or (. . . , W Y, W Y, . . .). Now, suppose that the following situation occurs: (asχ(1), . . . , asχ(n)) = (. . . , Y ∗W ∗, Y ∗W ∗, . . .), with asχ(m) = asχ(m+1) = Y ∗W ∗ for some m ∈ {1, . . . , n − 1}. This implies the following situation for the χ-order: (cs χ(1), . . . , cs χ(2n)) = (. . . , W ∗, Y ∗, W ∗, Y ∗, . . .), with cs χ(2m−1) = cs χ(2m+1) = W ∗ and cs χ(2m) = cs χ(2m+2) = Y ∗. For τ ∈ BNC( χ), let V ∈ τ be such that s χ(2m) ∈ V and suppose that s χ(2m) = max≺ χV (equivalently, 2m = max<s−1 χ (V )). Let q ∈ {1, . . . , 2n} such that s χ(q) = min≺ χV (equivalently, q = min<s−1 χ (V )) and notice that we must have that cs χ(q) ∈ {X, Y }. Assume that cs χ(q) = Y (with the case when cs χ(q) = X handled similarly). Hence, we have that (cs χ(1), . . . , cs χ(2n)) = (. . . , Y, W, . . . , W ∗, Y ∗, W ∗, Y ∗, . . .) and it follows that q = 2p − 1 for some p ∈ {1, . . . , m − 1}. By defining eV = {s χ(i) : 2p − 1 ≤ i ≤ 2m} and letting λ = {eV ,(cid:0)eV(cid:1)c the relation τ ∨c0χ = 1 χ cannot be satisfied. }, one sees that V ⊆ eV , λ ∈ BNC( χ) and τ,c0χ ≤ λ (cid:12) 1 χ. Thus, This implies that there must exist q ∈ {1, . . . , 2n} with s χ(q) ∈ V and s χ(2m)≺ χ s χ(q) 21 and we may assume that s χ(q) is the χ-minimum element of V with this property. Notice that it must necessarily be that cs χ(q) = Y and this yields the following situation (cs χ(1), . . . , cs χ(2n)) = (. . . , W ∗, Y ∗, W ∗, Y ∗, . . . , Y, W, . . .), from which one sees that q = 2p − 1, for some p ∈ {m + 2, . . . , n}. As before, by defining eV = {s χ(i) : 2m + 1 ≤ i ≤ 2p − 2} }, one sees that V ⊆ (cid:0)eV(cid:1)c and letting λ = {eV ,(cid:0)eV(cid:1)c Thus, the relation τ ∨c0χ = 1 χ cannot be satisfied. Hence, when (asχ(1), . . . , asχ(n)) = (. . . , Y ∗W ∗, Y ∗W ∗, . . .), , λ ∈ BNC( χ) and τ,c0χ ≤ λ (cid:12) 1 χ. we obtain that the bi-free cumulant κχ(a1, . . . , an) vanishes and the use of similar argu- ments shows that this is also the case when the sequence (asχ(1), . . . , asχ(n)) has one of the following forms (. . . , Z ∗X ∗, Z ∗X ∗, . . .) or (. . . , Z ∗X ∗, Y ∗W ∗, . . .). This completes the proof. The main technical difficulty that results in the length of the previous proof is that we cannot only deal with bi-non-crossing partitions whose blocks contain an even number of elements, thus Proposition 2.11 does not apply. This is because the pair (Z, W ) need not be bi-R-diagonal and hence bi-free cumulants of odd order with entries in {Z, Z ∗, W, W ∗} need not necessarily vanish. We remark that for two ∗-bi-free pairs (X, Y ) and (Z, W ) with the first being bi-R- diagonal, it is not in general true that the pair (XZ, Y W ) will also be bi-R-diagonal, as the following example indicates. We will denote by "tr" the normalized trace on any matrix algebra. Example 3.3. Let (A, ϕ) be a non-commutative ∗-probability space and (ul, ur) be a bi-Haar unitary pair in A. Also, consider the pair (Z, W ) in (M2(C), tr) given as follows: Z =(cid:20)1 0 0 0(cid:21) and W =(cid:20)0 0 1 0(cid:21) In the free product space (A ∗ M2(C), ϕ ∗ tr) the pairs (ul, ur) and (Z, W ) are ∗-bi-free, but for χ ∈ {l, r}4 with χ(1) = χ(2) = l and χ(3) = χ(4) = r, the bi-free cumulant κχ(Z ∗ul ∗, urW ) does not vanish, even though it is not alternating in ∗-terms and non-∗-terms in the χ-order. Indeed, the moment-cumulant formula yields ∗, ulZ, W ∗ur κχ(Z ∗ul ∗, ulZ, W ∗ur ∗, urW ) = Xτ ∈BNC(χ) (ϕ ∗ tr)τ (Z ∗ul ∗, ulZ, W ∗ur ∗, urW )µBNC(τ, 1χ). Using the characterization of free independence in terms of moments, it is seen that all operators that appear in the cumulant above are centred, i.e. the following holds (ϕ ∗ tr)(Z ∗u∗ l ) = (ϕ ∗ tr)(W ∗u∗ r) = (ϕ ∗ tr)(ulZ) = (ϕ ∗ tr)(urW ) = 0. Hence, to find bi-non-crossing partitions that are to yield a non-zero contribution to the sum above, we may only consider partitions on {1, 2, 3, 4} that are bi-non-crossing and all of whose blocks are not singletons. These are the following three bi-non-crossing partitions: τ1 = {{1, 2}, {3, 4}}, τ2 = {{1, 2, 3, 4}} and τ3 = {{1, 3}, {2, 4}}. 22 For τ1, we have that (ϕ ∗ tr)τ1(Z ∗u∗ l , ulZ, W ∗u∗ r, urW ) = (ϕ ∗ tr)(Z ∗u∗ l ulZ) · (ϕ ∗ tr)(W ∗u∗ rurW ) while for τ2 we obtain = tr(Z ∗Z) · tr(W ∗W ) = 1 4 , (ϕ ∗ tr)τ2(Z ∗u∗ l , ulZ, W ∗u∗ r, urW ) = (ϕ ∗ tr)(Z ∗u∗ l ulZW ∗u∗ rurW ) = tr(Z ∗ZW ∗W ) = 1 2 . For the case of τ3, it follows that l , ulZ, W ∗u∗ (ϕ ∗ tr)τ3 (Z ∗u∗ r, urW ) = (ϕ ∗ tr)(Z ∗u∗ l W ∗u∗ r) · (ϕ ∗ tr)(ulZurW ), and it is straightforward to show using the moment-cumulant formula that both terms appearing in the product above are equal to zero. Since µBNC(τ1, 1χ) = µBNC(τ3, 1χ) = −1 and µBNC(τ2, 1χ) = 1, the bi-free cumulant is evaluated as follows κχ(Z ∗ul ∗, ulZ, W ∗ur ∗, urW ) = 1 2 − 1 4 − 0 = 1 4 6= 0. However, when the pairs (X, Y ) and (Z, W ) are both bi-R-diagonal and ∗-bi-free, then it is the case that the resulting pair (XZ, Y W ) is also bi-R-diagonal, as the following proposition shows. Proposition 3.4. Let (A, ϕ) be a non-commutative ∗-probability space and X, Y, Z, W ∈ A such that: (a) the pairs (X, Y ) and (Z, W ) are both bi-R-diagonal, (b) the pairs the pairs (X, Y ) and (Z, W ) are ∗-bi-free. Then, the pair (XZ, Y W ) is also bi-R-diagonal. Proof. Let n ∈ N, χ ∈ {l, r}n and a1, . . . , an ∈ A such that {XZ, Z ∗X ∗}, if χ(i) = l {Y W, W ∗Y ∗}, if χ(i) = r (i = 1, . . . , n) ai ∈ Define χ ∈ {l, r}2n by χ(2i − 1) = χ(2i) = χ(i) for each i = 1, . . . , n and c1, . . . , c2n ∈ A as follows: c2i−1 = c2i =   X, Z ∗, Y, W ∗, Z, X ∗, W, Y ∗, if ai = XZ if ai = Z ∗X ∗ if ai = Y W if ai = W ∗Y ∗ if ai = XZ if ai = Z ∗X ∗ if ai = Y W if ai = W ∗Y ∗ 23 for each i = 1, . . . , n. Then, an application of Theorem 2.10 yields: κχ(a1, . . . , an) = Xτ ∈BNC( χ) = Xτ ∈BNC( χ) τ ∨c0χ=1 χ τ ∨c0χ=1 χ κ χ(c1, . . . , c2n) YV ∈τ κ χV(cid:0)(c1, . . . , c2n)V(cid:1) (1) (2) make the following remarks: where c0χ = {{2i − 1, 2i} : i = 1, . . . , n} ∈ BNC( χ). As in the proof of Theorem 3.2, we First of all, if χ−1({l}) = {i1 < . . . < ip} and χ−1({r}) = {j1 < . . . < jn−p}, the definition of χ implies that χ−1({l}) = {2i1 − 1 < 2i1 < . . . < 2ip − 1 < 2ip} and χ−1({r}) = {2j1 − 1 < 2j1 < . . . < 2jn−p − 1 < 2jn−p}. Thus, if i ∈ {1, . . . , n} is such that asχ(i) = XZ, then cs χ(2i−1) = X and cs χ(2i) = Z (a similar situation occurs when asχ(i) = Z ∗X ∗, since this corresponds to a left operator). Now if asχ(i) = Y W , then cs χ(2i−1) = W and cs χ(2i) = Y (and a similar situation occurs when asχ(i) = W ∗Y ∗ since this corresponds to a right operator). Note that in the latter case, the right operators must appear reversed in the χ-order. Secondly, due to the ∗-bi-free independence condition, in order for a bi-non-crossing partition τ to contribute to the above sum, we must have that for every V ∈ τ , either or {ci : i ∈ V } ⊆ {X, X ∗, Y, Y ∗}, {ci : i ∈ V } ⊆ {Z, Z ∗, W, W ∗}. Observe that this implies that if n is odd, then κχ(a1 . . . , an) = 0, as then the cardinality of the set {j ∈ {1, . . . , 2n} : cj ∈ {X, X ∗, Y, Y ∗}} is odd and hence for any τ ∈ BNC( χ) there exists V ∈ τ with odd cardinality that contains indices corresponding to elements in {X, X ∗, Y, Y ∗}. Since the pair (X, Y ) is bi-R-diagonal, all bi-free cumulants of odd order with entries in {X, X ∗, Y, Y ∗} vanish, thus κ χV ((c1, . . . , c2n)V ) = 0. In addition, in order for a bi-non-crossing partition τ ∈ BNC(χ) to contribute to the sum appearing in (1), every block of τ must contain an even number of elements. Indeed, if V ∈ τ contains an odd number of elements, then we deduce that (additionally assuming that all indices in V correspond to elements either from {X, X ∗, Y, Y ∗} or {Z, Z ∗, W, W ∗}) κ χV ((c1, . . . , c2n)V ) is a bi-free cumulant of odd order involving a bi-R-diagonal pair and thus vanishes. Henceforth, when referring to a bi-non-crossing partition τ contributing to the sum appearing in (1), we will always assume that every block of τ contains indices all corre- sponding to elements either from {X, X ∗, Y, Y ∗} or {Z, Z ∗, W, W ∗} and, by Proposition 2.11, that s χ(1)∼τ s χ(2n) and s χ(2i)∼τ s χ(2i + 1) for every i = 1, . . . , n − 1. We will now show that if the sequence (asχ(1), . . . , asχ(n)) is not alternating in ∗-terms and non-∗-terms, then the cumulant κχ(a1, . . . , an) must vanish. Suppose the following situation occurs: (asχ(1), . . . , asχ(n)) = (. . . , Y W, Y W, . . .), 24 with asχ(m) = asχ(m+1) = Y W for some m ∈ {1, . . . , n − 1}. This implies the following situation for the χ-order: (cs χ(1), . . . , cs χ(2n)) = (. . . , W, Y, W, Y, . . .), with cs χ(2m−1) = cs χ(2m+1) = W and cs χ(2m) = cs χ(2m+2) = Y . If τ is a bi-non-crossing partition contributing to the sum appearing in (1), then the block of τ containing s χ(2m) must also contain s χ(2m + 1). But, since this is impossible, due to the ∗-bi-free independence condition. Hence, when cs χ(2m) = Y and cs χ(2m+1) = W, (asχ(1), . . . , asχ(n)) = (. . . , Y W, Y W, . . .), we obtain that the bi-free cumulant κχ(a1, . . . , an) vanishes and the use of similar argu- ments shows that this is also the case when the sequence (asχ(1), . . . , asχ(n)) has either one of the following forms: (. . . , XZ, XZ, . . .), (. . . , Z ∗X ∗, Z ∗X ∗, . . .) or (. . . , W ∗Y ∗, W ∗Y ∗, . . .). Next, suppose the following situation occurs: (asχ(1), . . . , asχ(n)) = (. . . . . . , XZ, Y W, . . . . . .), with asχ(m) = XZ and asχ(m+1) = Y W for some m ∈ {1, . . . , n − 1}. This implies the following situation for the χ-order: (cs χ(1) . . . , cs χ(2n)) = (. . . . . . , X, Z, W, Y, . . . . . .), with cs χ(2m−1) = X, cs χ(2m) = Z, cs χ(2m+1) = W and cs χ(2m+2) = Y . If τ is a bi-non-crossing partition contributing to the sum appearing in (1), then the block V ∈ τ containing s χ(2m) must also contain s χ(2m+1). As discussed in the beginning of the proof, in order for the cumulant κ χ,τ (c1, . . . , c2n) not to vanish we must have that V ⊆ {j ∈ {1, . . . , 2n} : sj ∈ {Z, Z ∗, W, W ∗}}. But then the entries of the cumulant κ χV ((c1, . . . , c2n)V ) in the induced χV -order would be of the form: (. . . . . . , Z, W, . . . . . .) and this implies that the bi-free cumulant κ χV ((c1, . . . , c2n)V ) vanishes, as it is a cumulant involving the bi-R-diagonal pair (Z, W ) that is not alternating in ∗-terms and non-∗-terms in the induced χV -order. Since this is the case for every possible τ ∈ BNC( χ), we deduce that κχ(a1, . . . , an) = 0. Hence, when (asχ(1), . . . , asχ(n)) = (. . . , XZ, Y W, . . .), we obtain that the bi-free cumulant κχ(a1, . . . , an) vanishes and the use of similar ar- guments shows that this is also the case when the sequence (asχ(1), . . . , asχ(n)) has the form: (. . . , Z ∗X ∗, W ∗Y ∗, . . .). This completes the proof. We now proceed to prove that the condition of bi-R-diagonality is preserved under the taking of arbitrary powers. 25 Theorem 3.5. Let (A, ϕ) be a non-commutative ∗-probability space and let (X, Y ) be a bi-R-diagonal pair in A. Then, for every p ≥ 1 the pair (X p, Y p) is also bi-R-diagonal. Proof. Let n ∈ N, p ≥ 1, χ ∈ {l, r}n and a1, . . . , an ∈ {X p, (X ∗)p, Y p, (Y ∗)p} such that {X p, (X ∗)p}, if χ(i) = l {Y p, (Y ∗)p}, if χ(i) = r (i = 1, . . . , n) ai ∈ Define χ ∈ {l, r}np and c1, . . . , cnp ∈ A as follows: χ((i − 1)p + j) = χ(i) and for each i ∈ {1, . . . , n} and j ∈ {1, . . . , p}. Then, an application of Theorem 2.10 yields: c(i−1)p+j =  κχ(a1, . . . , an) = Xτ ∈BNC( χ) = Xτ ∈BNC( χ) τ ∨c0χ=1 χ τ ∨c0χ=1 χ X, X ∗, Y, Y ∗, if ai = X p if ai = (X ∗)p if ai = Y p if ai = (Y ∗)p κ χ,τ (c1, . . . , cnp) YV ∈τ κ χV(cid:0)(c1, . . . , cnp)V(cid:1) (1) (2) wherec0χ = {{(i − 1)p + 1, . . . , ip} : i = 1, . . . , n} ∈ BNC( χ). To start, we remark that since the pair (X, Y ) is bi-R-diagonal, in order for a bi-non- crossing partition τ ∈ BNC( χ) to have non-zero contribution to the sum appearing in (1), every block of τ must contain indices corresponding to an equal number of ∗-terms and non-∗-terms; for otherwise there would exist a block V ∈ τ with indices corresponding to an unequal number of ∗-terms and non-∗-terms. This implies that the sequence (c1, . . . , cnp)V when read in the induced χV -order will not be alternating in ∗-terms and non-∗-terms and hence κ χ,τ (c1, . . . , cnp) = 0. We will first show that if the sequence (asχ(1), . . . , asχ(n)) is not alternating in ∗-terms and non-∗-terms, then κχ(a1, . . . , an) = 0. Suppose the following situation occurs: (asχ(1), . . . , asχ(n)) = (. . . , X p, Y p, . . .), where asχ(m) = X p and asχ(m+1) = Y p for some m ∈ {1, . . . , n}. This implies the following situation for the χ-order: (cs χ(1), . . . , cs χ(np)) = (. . . , X, X . . . , X, Y, Y, . . . , Y, . . .), where cs χ((m−1)p+k) = X and cs χ(mp+k) = Y , for all k = 1, . . . , p. Let τ ∈ BNC( χ) and V ∈ τ such that s χ(mp + 1) ∈ V. Observe that s χ(mp + k) /∈ V for all k = 2, . . . , p ; for otherwise, the sequence (c1, . . . , cnp)V when read in the induced χV -order would be of the form (. . . . . . , Y, Y, . . . . . .) 26 Consider the case when s χ(mp + 1) = min≺ χV (equivalently, mp + 1 = min<s−1 and this would imply that κ χV (c1, . . . , cnp)V = 0, since the pair (X, Y ) is bi-R-diagonal. χ (V )) and let q ∈ {1, . . . , np} be such that s χ(q) = max≺ χV (equivalently, q = max<s−1 χ (V )). It is easy to see that cs χ(q) = Y ∗. We claim that we must necessarily have that q = tp, for some t ∈ {m + 2, . . . , n}. To see this, suppose that q = tp + k with t ∈ {m + 2, . . . , n − 1} and k ∈ {1, . . . , p − 1}. Define A = {s χ(i) : mp + 1 ≤ i ≤ tp + k} and notice that A has to be written as a union of blocks of τ , which means that if V ′ ∈ τ, V 6= V ′ is such that V ′ ∩ A 6= ∅, then V ′ ⊆ A. Indeed, for such a block V ′, if there existed i 6= j ∈ {1, . . . , np} with s χ(i) ∈ V ′ ∩ A and s χ(j) ∈ V ′ \ A, then this would imply that mp + 2 ≤ i ≤ tp + k − 1, j ∈ {1, . . . , mp + 1} ∪ {tp + k, . . . , np} and {mp + 1, tp + k} ⊆ s−1 χ (V ), which contradicts the fact that the partition s−1 χ · τ is non-crossing. Thus, A has to be written as a union of blocks of τ and since A contains indices corresponding to an unequal number of ∗-terms and non-∗-terms, there must exist a block V ′ ∈ τ with this same property. This yields that κ χV ′ (c1, . . . , cnp)V ′ = 0 and, as a result, the cumulant κ χ,τ (c1, . . . , cnp) vanishes. We may now assume that q = tp, for some t ∈ {m + 2, . . . , n}. By defining eV = {s χ(i) : mp + 1 ≤ i ≤ tp} and letting λ = {eV ,(cid:0)eV(cid:1)c }, one sees that V ⊆ eV , λ ∈ BNC( χ) and τ,c0χ ≤ λ (cid:12) 1 χ. Thus, the relation τ ∨c0χ = 1 χ cannot be satisfied. Hence, it cannot be the case that So, suppose that there exists q ∈ {1, . . . , np} with s χ(q) ∈ V and s χ(mp + 1) = min≺ χV . s χ(q)≺ χ s χ(mp + 1). We may moreover assume that for all v ∈ V \ {s χ(mp + 1), s χ(q)}, we either have that v≺ χ s χ(q) or s χ(2m + 1)≺ χ v (i.e. that s χ(q) is the χ-maximum element of V with this property). Notice that it must necessarily be that cs χ(q) = X ∗ and, arguing as before, it must be the case that q = tp for some t ∈ {1, . . . , m − 1}. Then, by defining eV = {s χ(i) : tp + 1 ≤ i ≤ mp} and letting λ = {eV ,(cid:0)eV(cid:1)c }, one sees that V ⊆ (cid:0)eV(cid:1)c Thus, the relation τ ∨c0χ = 1 χ once again cannot be satisfied. This shows that when , λ ∈ BNC( χ) and τ,c0χ ≤ λ (cid:12) 1 χ. (asχ(1), . . . , asχ(n)) = (. . . , X p, Y p, . . .), we obtain that the bi-free cumulant κχ(a1, . . . , an) vanishes and the use of similar argu- ments shows that this is also the case when the sequence (asχ(1), . . . , asχ(n)) has one of the following forms: (a) (. . . , X p, X p, . . .), 27 (b) (. . . , Y p, Y p, . . .), (c) (. . . , (X ∗)p, (X ∗)p, . . .), (d) (. . . , (X ∗)p, (Y ∗)p, . . .), (e) (. . . , (Y ∗)p, (Y ∗)p, . . .). Hence, if the sequence (asχ(1), . . . , asχ(n)) is not alternating in ∗-terms and non-∗-terms, we have that κχ(a1, . . . , an) = 0. It remains to show that if the cumulant κχ(a1, . . . , an) is of odd order, then it must vanish. Assume that n is an odd number. By the aforementioned considerations, we may assume that the sequence (asχ(1), . . . , asχ(n)) does not contain consecutive elements that both correspond to either ∗-terms or non-∗-terms. Suppose the following situation occurs: (asχ(1), . . . , asχ(n)) = ((X ∗)p, . . . . . . , (Y ∗)p), where asχ(1) = (X ∗)p and asχ(n) = (Y ∗)p. This implies the following situation for the χ-order: (cs χ(1), . . . , cs χ(np)) = (X ∗, X ∗, . . . , X ∗, . . . . . . , Y ∗, Y ∗, . . . , Y ∗), where cs χ(k) = X ∗ and cs χ((n−1)p+k) = Y ∗, for all k = 1, . . . , p. Let τ ∈ BNC( χ) and V ∈ τ such that s χ(1) ∈ V . Also, let q ∈ {1, . . . , np} be such that s χ(q) = max≺ χV . First of all, observe that it must be that q ≥ p + 1; for otherwise, since cs χ(k) = X ∗ for all k = 1, . . . , p, the sequence (c1, . . . , cnp)V when read in the induced χV -order would be of the form (X ∗, . . . . . . , X ∗), and hence has either odd length or is not alternating in ∗-terms and non-∗-terms. This implies that κ χV ((c1, . . . , cnp)V ) = 0, since the pair (X, Y ) is bi-R-diagonal. Secondly, note that it must necessarily be that q = tp for some t ∈ {2, . . . , n}. Indeed, if q = tp + k for some t ∈ {2, . . . , n − 1} and k ∈ {1, . . . , p − 1}, then the set {s χ(i) : 1 ≤ i ≤ tp + k} (which contains indices corresponding to an unequal number of ∗-terms and non-∗-terms) must be written as a union of blocks of τ. Thus, there exists a block V ′ of τ containing indices that correspond to an unequal number of ∗-terms and non-∗-terms and it follows that if that is the case, then κ χV ′ ((c1, . . . , cnp)V ′) = 0. We will now show that q = np. If we assumed that q = tp, for some t ∈ {2, . . . , n − 1}, then by defining and letting λ = {eV ,(cid:0)eV(cid:1)c the relation τ ∨c0χ = 1 χ cannot be satisfied. This shows that when eV = {s χ(i) : 1 ≤ i ≤ tp} }, one sees that V ⊆ eV , λ ∈ BNC( χ) and τ,c0χ ≤ λ (cid:12) 1 χ. Thus, (asχ(1), . . . , asχ(n)) = ((X ∗)p, . . . . . . , (Y ∗)p), we obtain that the bi-free cumulant κχ(a1, . . . , an) vanishes and the use of similar argu- ments shows that this is also the case when the sequence (asχ(1), . . . , asχ(n)) has one of the following forms: (a) (X p, . . . . . . , X p), 28 (b) ((X ∗)p, . . . . . . , (X ∗)p), (c) ((Y ∗)p, . . . . . . , (Y ∗)p), (d) (X p, . . . . . . , Y p), (e) (Y p, . . . . . . , Y p). This completes the proof. We close this section by showing that bi-R-diagonal pairs of operators yield examples of bi-free pairs that consist of self-adjoint operators. Proposition 3.6. Let (A, ϕ) be a non-commutative ∗-probability space and (X, Y ) be a bi-R-diagonal pair in A. Then, the pairs (XX ∗, Y ∗Y ) and (X ∗X, Y Y ∗) are bi-free. Proof. Let n ∈ N, n ≥ 2, χ ∈ {l, r}n and a1, . . . , an ∈ A such that {XX ∗, X ∗X}, if χ(i) = l {Y ∗Y, Y Y ∗}, if χ(i) = r (i = 1, . . . , n) ai ∈ Moreover, suppose that there exist i 6= j ∈ {1, . . . , n} such that ai ∈ {XX ∗, Y ∗Y } and aj ∈ {X ∗X, Y Y ∗}. We will show that κχ(a1, . . . , an) = 0, which will imply that the pairs (XX ∗, Y ∗Y ) and (X ∗X, Y Y ∗) are indeed bi-free. Define χ ∈ {l, r}2n by for each i = 1, . . . , n and c1, . . . , c2n ∈ A as follows: χ(2i − 1) = χ(2i) = χ(i), X, Y ∗, X ∗, Y, X ∗, Y, X, Y ∗, if ai = XX ∗ if ai = Y ∗Y if ai = X ∗X if ai = Y Y ∗ if ai = XX ∗ if ai = Y ∗Y if ai = X ∗X if ai = Y Y ∗ c2i = c2i−1 =   κχ(a1, . . . , an) = Xτ ∈BNC( χ) = Xτ ∈BNC( χ) τ ∨c0χ=1 χ for each i = 1, . . . , n. Then, an application of Theorem 2.10 yields: κ χ,τ (c1, . . . , c2n) YV ∈τ κ χV(cid:0)(c1, . . . , c2n)V(cid:1) (1) (2) 29 τ ∨c0χ=1 χ wherec0χ = {{2i − 1, 2i} : i = 1, . . . , n} ∈ BNC( χ). Since the pair (X, Y ) is bi-R-diagonal, for a bi-non-crossing partition τ ∈ BNC( χ) to have a non-zero contribution to the sum appearing in (1), every block of τ must con- tain an even number of elements, as every bi-free cumulant of odd order with entries in {X, X ∗, Y, Y ∗} vanishes. Our initial assumptions imply that there exists i ∈ {1, . . . , n} such that either asχ(i) ∈ {XX ∗, Y ∗Y } and asχ(i+1) ∈ {X ∗X, Y Y ∗}, or asχ(i) ∈ {X ∗X, Y Y ∗} and asχ(i+1) ∈ {XX ∗, Y ∗Y }. Assume that asχ(i) = XX ∗ and asχ(i+1) = Y Y ∗ (with the remaining cases handled simi- larly). Then, the following situation occurs for the χ-order: (cs χ(1), . . . , cs χ(2n)) = (. . . , X, X ∗, Y ∗, Y, . . .), where cs χ(2i−1) = X, cs χ(2i) = X ∗, cs χ(2i+1) = Y ∗ and cs χ(2i+2) = Y . Note that, due to the definition of the permutation s χ, the right operators must appear reversed in the χ-order. By Proposition 2.11, for τ ∈ BNC( χ) such that τ ∨c0χ = 1 χ there exists V ∈ τ with {s χ(2i), s χ(2i + 1)} ⊆ V . But then, the sequence (c1, . . . , c2n)V when read in the induced χV -order would be of the form (. . . . . . , X ∗, Y ∗, . . . . . .) with this implying that κ χV ((c1, . . . , c2n)V ) = 0, since, as the pair (X, Y ) is bi-R-diagonal, bi-free cumulants with entries in {X, X ∗, Y, Y ∗} that are non-∗-alternating in each of the corresponding χ-orders must vanish. Hence, κ χ,τ (c1, . . . , c2n) = 0 and this finishes the proof. We remark that if (X, Y ) is a bi-R-diagonal pair in some non-commutative ∗-probability space, then it is not necessarily true that the pairs (XX ∗, Y Y ∗) and (X ∗X, Y ∗Y ) are bi- free, as the following example indicates. Example 3.7. Let (ul, ur) be a bi-Haar unitary pair in a non-commutative ∗-probability space (A, ϕ) and consider the pair (Z, W ) in the space (M2(C), tr) defined as follows: Z =(cid:20)1 0 0 0(cid:21) and W =(cid:20)0 0 1 0(cid:21) In the free product space (A ∗ M2(C), ϕ ∗ tr), the pairs (ul, ur) and (Z, W ) are ∗-bi-free and hence, by Theorem 3.2, the pair (ulZ, W ur) is bi-R-diagonal. But, the pairs (Z ∗u∗ l ulZ, u∗ rW ∗W ur) = (Z ∗Z, u∗ r W ∗W ur), and (ulZZ ∗u∗ l , W uru∗ rW ∗) = (ulZZ ∗u∗ l , W W ∗) are not bi-free, since the moment-cumulant formula yields κχ(Z ∗Z, W W ∗) = tr(Z ∗ZW W ∗) − tr(Z ∗Z) · tr(W W ∗) = − 1 4 6= 0. 30 4. Joint ∗-Distributions of Bi-R-Diagonal Pairs In this section, we will be concerned with proving that the joint ∗-distribution of a bi- R-diagonal pair of operators remains invariant under the multiplication with a ∗-bi-free bi-Haar unitary pair. We begin by giving the definition of bi-even and ∗-bi-even pairs of operators, as well as display how this class of pairs of operators can yield examples of bi-R-diagonal pairs. Definition 4.1. Let (A, ϕ) be a non-commutative ∗-probability space and Z, W ∈ A. (i) The pair (Z, W ) is called bi-even if for every k ∈ N and a1, . . . , a2k+1 ∈ {Z, W } we have that that is, all of its joint moments of odd order vanish. ϕ(a1 · . . . · a2k+1) = 0, (ii) The pair (Z, W ) is called ∗-bi-even if for every k ∈ N and a1, . . . , a2k+1 ∈ {Z, Z ∗, W, W ∗} we have that that is, all of its joint ∗-moments of odd order vanish. ϕ(a1 · . . . · a2k+1) = 0, The moment-cumulant formula yields that the pair (X, Y ) is ∗-bi-even if and only if all bi-free cumulants of odd order with entries in {X, X ∗, Y, Y ∗} vanish. It clearly follows that every bi-R-diagonal pair is ∗-bi-even. In the setting of free probability, it is observed that products of free, self-adjoint, even self-adjoint elements of non-commutative ∗-probability spaces all whose elements (i.e. moments of odd order vanish) result in R-diagonal elements ([NS06, Theorem 15.17]). Generalizing this to the bi-free setting, we will show that products of ∗-bi-even pairs (where the order of the right operators is reversed in the product) yield bi-R-diagonal pairs. For this, we have the following proposition, the proof of which will be similar to the proofs of Theorem 3.2 and Proposition 3.4. Proposition 4.2. Let (A, ϕ) be a non-commutative probability space and X, Y, Z, W ∈ A such that: (a) the pairs (X, Y ) and (Z, W ) are both ∗-bi-even, (b) the pairs (X, Y ) and (Z, W ) are ∗-bi-free. Then, the pair (XZ, W Y ) is bi-R-diagonal. Proof. Let n ∈ N, χ ∈ {l, r}n and a1, . . . , an ∈ A be such that {XZ, Z ∗X ∗}, if χ(i) = l {W Y, Y ∗W ∗}, if χ(i) = r (i = 1, . . . , n) ai ∈ Define χ ∈ {l, r}2n by χ(2i − 1) = χ(2i) = χ(i) for each i = 1, . . . , n and c1, . . . , c2n ∈ A as follows: c2i−1 = if ai = XZ if ai = Z ∗X ∗ if ai = W Y if ai = Y ∗W ∗ X, Z ∗, W, Y ∗,  31 Z, X ∗, Y, W ∗, if ai = XZ if ai = Z ∗X ∗ if ai = W Y if ai = Y ∗W ∗ c2i =  κχ(a1, . . . , an) = Xτ ∈BNC( χ) = Xτ ∈BNC( χ) τ ∨c0χ=1 χ τ ∨c0χ=1 χ for each i = 1, . . . , n. Then, an application of Theorem 2.10 yields: κ χ,τ (c1, . . . , c2n) YV ∈τ κ χV(cid:0)(c1, . . . , c2n)V(cid:1) (1) (2) make the following remarks: where c0χ = {{2i − 1, 2i} : i = 1, . . . , n} ∈ BNC( χ). As in the proof of Theorem 3.2, we First of all, if for some i ∈ {1, . . . , n} we have that asχ(i) = XZ , then it follows that cs χ(2i−1) = X and cs χ(2i) = Z (with a similar situation occurring when asχ(i) = Z ∗X ∗, since this corresponds to a left operator). Now, if asχ(i) = W Y , then cs χ(2i−1) = Y and cs χ(2i) = W (and a similar situation occurs when asχ(i) = Y ∗W ∗, since this corresponds to a right operator). Note that in the latter case, the right operators must appear reversed in the χ-order. Since the pairs (X, Y ) and (Z, W ) are ∗-bi-free, in order for a bi-non-crossing partition τ ∈ BNC( χ) to contribute to the sum appearing in (1), we must have that for all V ∈ τ , either or {ci : i ∈ V } ⊆ {X, X ∗, Y, Y ∗}, {ci : i ∈ V } ⊆ {Z, Z ∗, W, W ∗}. Observe that this also implies that if n is odd, then κχ(a1 . . . , an) = 0, as then the cardinality of the set {j ∈ {1, . . . , 2n} : cj ∈ {X, X ∗, Y, Y ∗}} is odd and hence for any τ ∈ BNC( χ) there exists V ∈ τ with odd cardinality that contains indices corresponding to elements in {X, X ∗, Y, Y ∗}. Since the pair (X, Y ) is ∗-bi-even, all bi-free cumulants of odd order with entries in {X, X ∗, Y, Y ∗} vanish, thus κ χV(cid:0)(c1, . . . , c2n)V(cid:1) = 0. In addition, in order for τ to contribute to the above sum, every block of τ must contain an even number of elements. Indeed, if V ∈ τ contains an odd number of elements, then we deduce that (additionally assuming that all indices in V correspond to elements either from {X, X ∗, Y, Y ∗} or {Z, Z ∗, W, W ∗}) κ χV ((c1, . . . , c2n)V ) is a bi-free cumulant of odd order involving a ∗-bi-even pair and thus vanishes. Henceforth, when referring to a bi-non-crossing partition τ contributing to the sum appearing in (1), we will assume that every block of τ contains indices all corresponding to elements either from {X, X ∗, Y, Y ∗} or {Z, Z ∗, W, W ∗} and, by Proposition 2.11, that s χ(1)∼τ s χ(2n) and s χ(2i)∼τ s χ(2i + 1) for every i = 1, . . . , n − 1. 32 We will now show that if the sequence (asχ(1), . . . , asχ(n)) is not alternating in ∗-terms and non-∗-terms, then the cumulant κχ(a1, . . . , an) must vanish. Suppose the following situation occurs: (asχ(1), . . . , asχ(n)) = (. . . . . . , Z ∗X ∗, Y ∗W ∗, . . . . . .), with asχ(m) = Z ∗X ∗ and asχ(m+1) = Y ∗W ∗ for some m ∈ {1, . . . , n − 1}. This implies the following situation for the χ-order: (cs χ(1), . . . , cs χ(2n)) = (. . . . . . , Z ∗, X ∗, W ∗, Y ∗, . . . . . .), with cs χ(2m−1) = Z ∗, cs χ(2m) = X ∗, cs χ(2m+1) = W ∗ and cs χ(2m+2) = Y ∗. Now, if τ is a bi-non-crossing partition contributing to the sum appearing in (1), then the block of τ containing s χ(2m) must also contain s χ(2m + 1). But, since cs χ(2m) = X ∗ and cs χ(2m+1) = W ∗, this is impossible, due to the ∗-bi-free independence condition. Hence, when (asχ(1), . . . , asχ(n)) = (. . . , Z ∗X ∗, Y ∗W ∗, . . .), we obtain that the bi-free cumulant κχ(a1, . . . , an) vanishes and the use of similar argu- ments shows that this is also the case when the sequence (asχ(1), . . . , asχ(n)) has either one of the following forms: (a) (. . . , XZ, XZ, . . .), (b) (. . . , W Y, W Y, . . .), (c) (. . . , Z ∗X ∗, Z ∗X ∗, . . .), (d) (. . . , XZ, W Y, . . .), (e) (. . . , Y ∗W ∗, Y ∗W ∗, . . .). This completes the proof. For a non-commutative ∗-probability space (A, ϕ) and X1, X2, Y1, Y2 ∈ A, consider the pair (Z, W ) in the tensor product space (M2(A), ϕ ⊗ tr) defined by Z =(cid:20) 0 X1 0 (cid:21) and W =(cid:20) 0 X2 Y1 0(cid:21) . Y2 Since any product with entries in {Z, Z ∗, W, W ∗} containing an odd number of elements results in a matrix with zeroes across the diagonal, it follows that (Z, W ) is a ∗-bi-even pair. Such pair is not necessary bi-R-diagonal, since for instance κχ(Z, Z) = (ϕ ⊗ tr)(Z · Z) = 1 2 (ϕ(X1X2) + ϕ(X2X1)) which need not be equal to zero. However, the previous proposition implies that the product of two such pairs that are ∗-bi-free will always be bi-R-diagonal. Actually, matrix pairs arising in this manner can be used to characterize the condition of bi-R-diagonality (see Theorem 4.6). We proceed with a lemma that contains the central combinatorial argument required for proving one of the main results of this section (see Theorem 4.4). At a key point, it makes use of the cancellation property observed in Lemma 2.6. 33 Lemma 4.3. Let (A, ϕ) be a non-commutative ∗-probability space and ul, ur, Z, W ∈ A such that: (a) the pair (ul, ur) is a bi-Haar unitary, (b) the pair (Z, W ) is ∗-bi-even, (c) the pairs (ul, ur) and (Z, W ) are ∗-bi-free. Let m ∈ N, χ ∈ {l, r}2m and a1, . . . , a2m ∈ A with {ulZ, Z ∗ul ∗}, if χ(i) = l {W ur, ur ∗W ∗}, if χ(i) = r (i = 1, . . . , 2m) ai ∈ such that the sequence (asχ(1), . . . , asχ(2m)) is alternating in ∗-terms and non-∗-terms. Define b1, . . . , b2m ∈ A as follows: bi = Then, we have that: Z, Z ∗, W, W ∗,  if ai = ulZ if ai = Z ∗ul if ai = W ur if ai = ur ∗W ∗ ∗ (i = 1, . . . , 2m) κχ(a1, . . . , a2m) = κχ(b1, . . . , b2m). Proof. Let m ∈ N, χ ∈ {l, r}2m and a1, . . . , a2m, b1, . . . , b2m be given as in the statement of the lemma. Define χ ∈ {l, r}4m by χ(2i − 1) = χ(2i) = χ(i) for each i = 1, . . . , 2m and c1, . . . , c4m ∈ A as follows: ∗ if ai = ulZ if ai = Z ∗ul if ai = W ur if ai = ur ∗W ∗ ∗ if ai = ulZ if ai = Z ∗ul if ai = W ur if ai = ur ∗W ∗ c2i = c2i−1 = Z, ∗, ul ur, W ∗, ul, Z ∗, W, ∗, ur   κχ(a1, . . . , a2m) = Xτ ∈BNC( χ) = Xτ ∈BNC( χ) τ ∨c0χ=1 χ for each i = 1, . . . , 2m. Then, an application of Theorem 2.10 yields: κ χ,τ (c1, . . . , c4m) (1) (2) YV ∈τ κ χV(cid:0)(c1, . . . , c4m)V(cid:1) 34 τ ∨c0χ=1 χ we make the following observations: wherec0χ = {{2i − 1, 2i} : i = 1, . . . , 2m} ∈ BNC( χ). As in the proof of Proposition 3.4, First of all, if for some i ∈ {1, . . . , 2m} we have that asχ(i) = ulZ , then it follows that cs χ(2i−1) = ul and cs χ(2i) = Z (with a similar situation occurring when asχ(i) = Z ∗X ∗, since this corresponds to a left operator). Now, if asχ(i) = W ur , then cs χ(2i−1) = ur and cs χ(2i) = W (and a similar situation occurs when asχ(i) = Y ∗W ∗, since this corresponds to a right operator). Note that in the latter case, the right operators must appear reversed in the χ-order. This implies that since the sequence (asχ(1), . . . , asχ(2m)) was assumed to be alternating in ∗-terms and non-∗-terms, then both the sequences (cs χ(1), cs χ(4), cs χ(5), . . . , cs χ(4m−4), cs χ(4m−3), cs χ(4m)) and (cs χ(2), cs χ(3), cs χ(6), cs χ(7), . . . , cs χ(4m−2), cs χ(4m−1)) are also alternating in ∗-terms and non-∗-terms (observe that for any i ∈ {1, . . . , 2m}, the element asχ(i) corresponds to a ∗-term if and only if both the elements cs χ(2i−1) and cs χ(2i) correspond to ∗-terms). Since the pairs (ul, ur) and (Z, W ) are ∗-bi-free, in order for a bi-non-crossing partition τ ∈ BNC( χ) to contribute to the sum appearing in (1), we must have that for all V ∈ τ , either or {ci : i ∈ V } ⊆ {ul, ul ∗, ur, ur ∗}, {ci : i ∈ V } ⊆ {Z, Z ∗, W, W ∗}. In addition, in order for τ to contribute to the above sum, every block of τ must contain an even number of elements. Indeed, if V ∈ τ contains an odd number of elements, then we deduce that (additionally assuming that all indices in V correspond to elements either ∗} or {Z, Z ∗, W, W ∗}) κ χV ((c1, . . . , c2n)V ) is a bi-free cumulant of odd from {ul, ul order involving a ∗-bi-even pair and thus vanishes. ∗, ur, ur Henceforth, when referring to a bi-non-crossing partition τ contributing to the sum appearing in (1), we will assume that τ satisfies the following requirements: (A) every block of τ contains indices all corresponding to elements either from {ul, u∗ l , ur, u∗ r} or {Z, Z ∗, W, W ∗}, (B) s χ(1)∼τ s χ(4m) and s χ(2i)∼τ s χ(2i + 1) for every i = 1, . . . , 2m − 1 (this follows from an application of Proposition 2.11). Define the sets E1 = {s χ(1), s χ(4m)} and Ei+1 = {s χ(2i), s χ(2i + 1)}, for all i = 1, . . . , 2m − 1. We introduce new symbols 1, 2, . . . , m and let Fi = E2i−1 and Gi = E2i, for all i = 1, . . . , m. The notation Gi may seem unnatural, but it is being utilized for clarity, once Kreweras complementation map is implemented later in the proof. We claim that it must be the case that either {j ∈ {1, . . . , 4m} : cj ∈ {ul, ul ∗, ur, ur ∗}} = Fi, m[i=1 35 Indeed, begin by assuming that asχ(1) = ulZ. Since the sequence (asχ(1), . . . , asχ(n)) is alternating in ∗-terms and non-∗-terms, we must have that asχ(2) ∈ {Z ∗u∗ If asχ(2) = ulZ, then for the χ-order it is implied that rW ∗}. l , u∗ cs χ(1) = ul, cs χ(2) = Z, cs χ(3) = Z ∗, cs χ(4) = u∗ l , while if asχ(2) = u∗ rW ∗, then for the χ-order it is implied that cs χ(1) = ul, cs χ(2) = Z, cs χ(3) = W ∗, cs χ(4) = u∗ r, hence in both cases we see that {cs χ(2), cs χ(3)} ⊆ {Z, Z ∗, W, W ∗}. A straightforward induction argument then shows that for all i = 1, . . . , m and j ∈ Gi, one must have that cj ∈ {Z, Z ∗, W, W ∗}. Of course, this also implies that the union of {Fi : i = 1, . . . , m} must be equal to the set of all indices that correspond to elements in {ul, u∗ r}. It clearly follows that similar arguments yield an analogous outcome in the case when asχ(1) ∈ {Z ∗u∗ rW ∗}. Hence, we may assume that l , ur, u∗ l , W ur, u∗ {j ∈ {1, . . . , 4m} : cj ∈ {ul, ul ∗, ur, ur ∗}} = Fi, m[i=1 with the remaining case handled similarly. From this, it follows that {j ∈ {1, . . . , 4m} : cj ∈ {Z, Z ∗, W, W ∗}} = Gi. m[i=1 This assumption, along with requirement (A) above imply that for every V ∈ τ , we have that either V ⊆ Fi, or V ⊆ Gi. m[i=1 m[i=1 or {j ∈ {1, . . . , 4m} : cj ∈ {ul, ul ∗, ur, ur ∗}} = Gi. m[i=1 Due to requirement (B) above and the definitions of the sets Fi and Gi, it is easy to see that for any block V ∈ τ and any i ∈ {1, . . . , m}, we have that V ∩ Fi 6= ∅ ⇐⇒ Fi ⊆ V and V ∩ Gi 6= ∅ ⇐⇒ Gi ⊆ V. For all V ∈ τ with V ⊆ ∪m i=1Fi, define IV = {i ∈ {1, . . . , m} : V ∩ Fi 6= ∅} and let πτ =nIV : V ∈ τ, V ⊆ m[i=1 Fio. It is easy to see that πτ ∈ P(m) and we claim that πτ ∈ NC(m). Indeed, if not, there exist blocks V 6= V ′ ∈ τ with V, V ′ ⊆ ∪m i=1Fi, and integers i1, i2, j1, j2 ∈ {1, . . . , m} such that i1, i2 ∈ IV , ji, j2 ∈ IV ′ and i1 < j1 < i2 < j2. 36 Since i1, i2 ∈ IV , it follows that Fi1 , Fi2 ⊆ V and similarly Fj1, Fj2 ⊆ V ′. Initially, assume that i1 = 1. By the definition of the sets {Fi : i = 1, . . . , m}, it is implied that {s χ(1), s χ(4i2 − 3)} ⊆ V and {s χ(4ji − 3), s χ(4j2 − 3)} ⊆ V ′, or, equivalently, {1, 4i2 − 3} ⊆ s−1 χ (V ) and {4j1 − 3, 4j2 − 3} ⊆ s−1 χ (V ′). But, since 1 = i1 < j1 < i2 < j2, it follows that 1 < 4j1 − 3 < 4i2 − 3 < 4j2 − 3, which contradicts the fact that s−1 i1 ≥ 2, then similarly we obtain χ · τ ∈ NC(4m). Now, if we consider the case when {4i1 − 3, 4i2 − 3} ⊆ s−1 χ (V ) and {4j1 − 3, 4j2 − 3} ⊆ s−1 χ (V ′), with the relations i1 < j1 < i2 < j2 implying that 4i1 − 3 < 4j1 − 3 < 4i2 − 3 < 4j2 − 3, which once again contradicts the fact that s−1 χ · τ ∈ NC(4m). Hence, we must have that πτ ∈ NC(m) and the use of similar arguments yields that if for all V ∈ τ with V ⊆ ∪m i=1Gi we define then, by letting JV =ni ∈ {1, . . . , m} : V ∩ Gi 6= ∅o, Gio, στ =nJV : V ∈ τ, V ⊆ m[i=1 it follows that στ ∈ NC({1, 2, . . . , m}). We claim that we must necessarily have that πτ ∪ στ ∈ NC({(1, 1, 2, 2, . . . , m, m)}). Indeed, if not, there exist blocks V 6= V ′ ∈ τ such that V ⊆ Fi and V ′ ⊆ m[i=1 Gi m[i=1 and integers i1, i2, j1, j2 ∈ {1, . . . , m} with i1, i2 ∈ IV , j1, j2 ∈ JV ′ and i1 ≤ j1 < i2 ≤ j2. ⊆ V ′. By the definitions of the sets IV and JV ′, it follows that Fi1, Fi2 ⊆ V and Gj1 Consider the case when i1 ≥ 2 (with the case when i1 = 1 treated analogously). This yields that , Gj2 {4i1 − 3, 4i2 − 3} ⊆ s−1 χ (V ) and {4j1 − 1, 4j2 − 1} ⊆ s−1 χ (V ′). But then, the relations i1 ≤ j1 < i2 ≤ j2 imply that 4i1 − 3 < 4j1 − 1 < 4i2 − 3 < 4j2 − 1, which contradicts the fact that s−1 χ · τ ∈ NC(4m). 37 Hence, πτ ∪ στ ∈ NC({1, 1, 2, 2, . . . , m, m}) and by the definition of Kreweras comple- mentation map and via the canonical identification NC(m) ∼= NC({1, 2, . . . , m}), this is equivalent to στ ≤ KNC(πτ ). The previously described process implies that any τ ∈ BNC( χ) that satisfies the re- quirements (A) and (B) uniquely determines two non-crossing partitions πτ , στ ∈ NC(m) such that στ ≤ KNC(πτ ). Conversely, any two non-crossing partitions π, σ ∈ NC(m) with σ ≤ KNC(π) uniquely determine a bi-non-crossing partition τ(π,σ) ∈ BNC( χ) that satisfies the requirements (A) and (B) by defining τ(π,σ) =n[i∈V Fi : V ∈ πo[n[i∈V Gi : V ∈ σo. This yields a bijection between all bi-non-crossing partitions that satisfy the requirements (A) and (B) with the set of all bi-non-crossing partitions τ(π,σ) obtained in the aforemen- tioned manner. Thus, the sum appearing in (2) becomes: Xτ ∈BNC( χ) τ ∨c0χ=1 χ YV ∈τ hπ · dσ hπ · dσ π,σ∈NC(m) σ≤KNC(π) κ χV(cid:0)(c1, . . . , c4m)V(cid:1) = Xτ(π,σ)∈BNC( χ) = Xπ,σ∈NC(m) = Xπ∈NC(m) σ≤KNC(π) hπ · Xσ∈NC(m) σ≤KNC(π) dσ! (3) where we have used the notation hπ = YV ∈π dσ = YV ∈σ κ χSi∈V Fi(cid:0)(c1, . . . , c4m)Si∈V Fi(cid:1) i(cid:0)(c1, . . . , c4m)Si∈V Gi(cid:1), κ χSi∈V G and for any π, σ ∈ NC(m). For a fixed π ∈ NC(m), we will compute the value of hπ. Recall that we assumed that {j ∈ {1, . . . , 4m} : cj ∈ {ul, u∗ l , ur, u∗ r}} = and {j ∈ {1, . . . , 4m} : cj ∈ {Z, Z ∗, W, W ∗}} = hence for all V ∈ π, the bi-free cumulant κ χSi∈V Fi(cid:0)(c1, . . . , c4m)Si∈V Fi(cid:1) 38 Fi, Gi, m[i=1 m[i=1 has entries in the set {ul, u∗ l , ur, u∗ r} and the sequence (c1, . . . , c4m)Si∈V Fi is alternating in ∗-terms and non-∗-terms when read in the induced χSi∈V Fi-order. More- over, notice that the cardinality of the union ∪i∈V Fi is equal to two times the cardinality of V . Thus, by a combination of Corollary 2.18 and Lemma 2.4 we obtain hπ = YV ∈π (−1)V −1 · CV −1 = µNC(0n, π). Hence, equation (3) yields that κχ(a1, . . . , a2m) = Xπ∈NC(m) µNC(0n, π) · Xσ∈NC(m) σ≤KNC(π) dσ!, with the right-hand side of the previous equation being equal to d1m , by Lemma 2.6. But then κχ(a1, . . . , a2m) = d1m = κ χSm i=1 G i(cid:0)(c1, . . . , c4m)Sm i=1 Gi(cid:1) = κχ(b1, . . . , b2m), where the elements b1, . . . , b2m are as in the statement of the lemma. This concludes the proof. We are now in a position to state the following theorem (which is the generalization of [NS06, Theorem 15.10] to the bi-free setting), regarding the invariance of the joint ∗- distribution of a bi-R-diagonal pair under the multiplication by a ∗-bi-free bi-Haar unitary pair. Theorem 4.4. Let (A, ϕ) be a non-commutative ∗-probability space and ul, ur, X, Y ∈ A such that: (a) the pair (ul, ur) is a bi-Haar unitary, (b) the pairs (ul, ur) and (X, Y ) are ∗-bi-free. Then, the following are equivalent: (i) the pair (X, Y ) is bi-R-diagonal, (ii) the joint ∗-distribution of the pair (X, Y ) coincides with the joint ∗-distribution of (ulX, Y ur). Proof. By Theorem 3.2 the pair (ulX, Y ur) is bi-R-diagonal and, since equality of joint ∗-distributions is equivalent to the equality of ∗-bi-free cumulants, it follows that the pair (X, Y ) is also bi-R-diagonal. This yields the implication (ii) ⇒ (i). For the converse, we will show the equality of all ∗-bi-free cumulants involving the pairs (X, Y ) and (ulX, Y ur). Since (X, Y ) is bi-R-diagonal, all bi-free cumulants with entries in {X, X ∗, Y, Y ∗} that are either of odd order or that are not alternating in ∗-terms and non-∗-terms in the χ-order must vanish. The same applies to the pair (ulX, Y ur) since it is also bi-R-diagonal. This implies that it suffices to show that for all even numbers n ∈ N, χ ∈ {l, r}n and a1, . . . , an ∈ A with ai ∈ {ulX, X ∗ul ∗}, if χ(i) = l {Y ur, ur ∗Y ∗}, if χ(i) = r (i = 1, . . . , n) 39 such that the sequence (asχ(1), . . . , asχ(n)) is alternating in ∗-terms and non-∗-terms, by setting X, X ∗, Y, Y ∗, if ai = ulX if ai = X ∗ul if ai = Y ur if ai = ur ∗Y ∗ ∗ (i = 1, . . . , n) κχ(a1, . . . , an) = κχ(b1, . . . , bn), bi = we have that  which is exactly what an application of Lemma 4.3 yields. We remark that the conclusion of the previous theorem no longer holds if the order of the multiplication of the right operators is not reversed, as the following example indicates. Example 4.5. Let (A, ϕ), (B, ψ) be two non-commutative ∗-probability spaces and let ul, ur ∈ A, vl, vr ∈ B such that both pairs (ul, ur) and (vl, vr) are bi-Haar unitaries. In the free product space (A ∗ B, ϕ ∗ ψ) these pairs are ∗-bi-free and clearly both bi-R- diagonal. But, the joint ∗-distribution of the pair (vl, vr) does not coincide with the joint ∗-distribution of (ulvl, urvr), since while, by an application of Theorem 2.10, it is easily verified that κχ(vl, v∗ r ) = ψ(vl · v∗ r ) = 1, κχ(ulvl, v∗ r u∗ r) = 0. Gathering the results of this section, one can obtain a theorem similar to [NSS01, Theorem 1.2] (and [BD18, Theorem 3.1] for the operator-valued setting). Theorem 4.6. Let (A, ϕ) be a non-commutative ∗-probability space and X, Y ∈ A. The following are equivalent: (i) the pair (X, Y ) is bi-R-diagonal, (ii) there exists an enlargement 1 ( A, ϕ) of (A, ϕ) and ul, ur ∈ A such that (a) the pair (ul, ur) is a bi-Haar unitary, (b) the pairs (ul, ur) and (X, Y ) are ∗-bi-free, (c) the joint ∗-distribution of the pair (ulX, Y ur) coincides with the joint ∗-distribution of (X, Y ), (iii) for any enlargement ( A, ϕ) of (A, ϕ) and any ul, ur ∈ A such that (d) the pair (ul, ur) is a bi-Haar unitary, (e) the pairs (ul, ur) and (X, Y ) are ∗-bi-free, one has that the the joint ∗-distribution of the pair (ulX, Y ur) coincides with the joint ∗-distribution of (X, Y ), 1An enlargement of a non-commutative ∗-probability space (A, ϕ) is a non-commutative ∗-probability space ( A, ϕ) such that A ⊆ A and ϕA = ϕ. 40 (iv) consider the unital subalgebras M2(C) and D of M2(A) consisting of scalar matrices and diagonal scalar matrices respectively and let the maps ε : D ⊗ Dop → L(M2(A)), L, R : M2(A) → L(M2(A)), E : L(M2(A)) → M2(C) and F : M2(C) → D be as in section 2.4. Also, in M2(A) consider the pair (Z, W ) defined as Z =(cid:20) 0 X 0(cid:21) and W =(cid:20) 0 X ∗ Y 0(cid:21) . Y ∗ Then, in the D-D-non-commutative probability space (L(M2(A), F ◦ E, ε), the pair (L(Z), R(W )) is bi-free from (L(M2(C)), R(M2(C)op)) with amalgamation over D. Proof. The equivalence of (i) and (iii), as well as the implication (ii) ⇒ (i) both follow from Theorem 4.4. Also, the equivalence of (i) and (iv) is a result of Proposition 2.21 and Theorem 2.22. To see that (i) implies (ii), simply consider a non-commutative ∗-probability space (B, ψ) containing a bi-Haar unitary pair (ul, ur) and define ( A, ϕ) to be the free product space (A ∗ B, ϕ ∗ ψ). In ( A, ϕ) the pairs (X, Y ) and (ul, ur) are ∗-bi-free and thus, again by Theorem 4.4, the joint ∗-distribution of the pair (ulX, Y ur) must coincide with the joint ∗-distribution of (X, Y ). Acknowledgements The author would like to thank Professor Paul Skoufranis for his valuable guidance and for numerous helpful conversations throughout the development of this manuscript. References [BD18] M. Boedihardjo and K. Dykema. On algebra-valued R-diagonal elements. Hous- ton J. Math., 44(1):209 -- 252, 2018. [Cha16] I. Charlesworth. An alternating moment condition for bi-freeness. preprint, arXiv:1611.01262, 2016. arXiv [CNS15a] I. Charlesworth, B. Nelson, and P. Skoufranis. Combinatorics of bi-freeness with amalgamation. Comm. Math. Phys., 338(2):801 -- 847, 2015. [CNS15b] I. Charlesworth, B. Nelson, and P. Skoufranis. On two-faced families of non- commutative random variables. Canad. J. Math., 67(6):1290 -- 1325, 2015. [HL00] U. Haagerup and F. Larsen. Brown's spectral distribution measure for R- diagonal elements in finite von Neumann algebras. J. Funct. Anal., 176(2):331 -- 367, 2000. [Kre72] G. Kreweras. Sur les partitions non crois´ees d'un cycle. Discrete Math., 1(4):333 -- 350, 1972. [Lar02] F. Larsen. Powers of R-diagonal elements. J. Operator Theory, 47(1):197 -- 212, 2002. 41 [NS97] A Nica and R Speicher. R-diagonal pairs -- a common approach to Haar uni- taries and circular elements. In Free probability theory (Waterloo, ON, 1995), volume 12 of Fields Inst. Commun., pages 149 -- 188. Amer. Math. Soc., Provi- dence, RI, 1997. [NS06] A. Nica and R. Speicher. Lectures on the combinatorics of free probability, volume 335 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 2006. [NSS99] A. Nica, D. Shlyakhtenko, and R. Speicher. Some minimization problems for the free analogue of the Fisher information. Adv. Math., 141(2):282 -- 321, 1999. [NSS01] A. Nica, D. Shlyakhtenko, and R. Speicher. R-diagonal elements and freeness with amalgamation. Canad. J. Math., 53(2):355 -- 381, 2001. [Sko16] [Voi14] P. Skoufranis. On operator-valued bi-free distributions. Adv. Math., 303:638 -- 715, 2016. D. Voiculescu. Free probability for pairs of faces I. Comm. Math. Phys., 332(3):955 -- 980, 2014. G. Katsimpas, Department of Mathematics and Statistics, York University, 4700 Keelee Street, North York, Ontario, Canada, M3J 1P3 E-mail address: [email protected] URL: http://sites.google.com/view/georgioskatsimpas 42
1810.06525	1	1810	2018-10-15T17:18:51	The Fredholm property for groupoids is a local property	[ "math.OA" ]	Fredholm Lie groupoids were introduced by Carvalho, Nistor and Qiao as a tool for the study of partial differential equations on open manifolds. This article extends the definition to the setting of locally compact groupoids and proves that \enquote{the Fredholm property is local}. Let $\mathcal{G} \rightrightarrows X$ be a topological groupoid and $(U_i)_{i\in I}$ be an open cover of $X$. We show that $\mathcal{G}$ is a Fredholm groupoid if, and only if, its reductions $\mathcal{G}^{U_i}_{U_i}$ are Fredholm groupoids for all $i \in I$. We exploit this criterion to show that many groupoids encountered in practical applications are Fredholm. As an important intermediate result, we use an induction argument to show that the primitive spectrum of $C^(\mathcal{G})$ can be written as the union of the primitive spectra of all $C^(\mathcal{G}\|_{U_i})$, for $i \in I$.	math.OA	math	THE FREDHOLM PROPERTY FOR GROUPOIDS IS A LOCAL PROPERTY RÉMI CÔME Abstract. Fredholm Lie groupoids were introduced by Carvalho, Nistor and Qiao as a tool for the study of partial differential equations on open manifolds. This article extends the definition to the setting of locally compact groupoids and proves that "the Fredholm property is local". Let G ⇒ X be a topological groupoid and (Ui)i∈I be an open cover of X. We show that G is a Fredholm groupoid if, and only if, its reductions GUi are Fredholm groupoids for all Ui i ∈ I. We exploit this criterion to show that many groupoids encountered in practical applications are Fredholm. As an important intermediate result, we use an induction argument to show that the primitive spectrum of C ∗(G) can be written as the union of the primitive spectra of all C ∗(GUi ), for i ∈ I. Contents Introduction 1. 1.1. Motivations: differential equations on singular manifolds 1.2. Main results 1.3. Applications and examples 1.4. Outline of the paper 2. Preliminaries 2.1. The primitive spectrum of a C∗-algebra 2.2. Groupoids and their C∗-algebras 3. Primitive spectrum and groupoid reductions 3.1. Representations induced from a reduction 3.2. Decomposition of the spectrum 3.3. Families of representations 4. Fredholm groupoids 4.1. Definitions 4.2. The Fredholm property is local 4.3. Consequences 5. Examples and Applications 5.1. Differential operators 5.2. Examples References 1 1 2 3 3 4 4 5 6 6 9 10 12 12 13 15 18 18 19 22 1. Introduction 1.1. Motivations: differential equations on singular manifolds. This paper deals with the study of locally compact groupoids, and more specifically with the structure of the primitive spectrum of their associated C∗-algebras. Nevertheless, our underlying motivation is the study of linear differential equations on open man- ifolds or on manifolds with singularities. 1 2 RÉMI CÔME Thus, let M0 be a smooth and complete Riemannian manifold without boundary, and P an order-m differential operator on M0. Let (H s(M0))s∈R be the usual scale of Sobolev spaces on M0 [15]. Under some natural conditions, the operator P extends as a bounded operator (1) P : H s(M0) → H s−m(M0). When M0 is a closed manifold, it is well-known that the operator in (1) is Fredholm if, and only if, it is elliptic, meaning that its principal symbol σ(P ) ∈ C∞(T ∗M0) vanishes only on the zero section [17]. This result has deep consequences concerning spectral theory, differential equations and index theory on closed mani- folds [3, 13, 17]. A natural and important question is to seek extensions of this Fredholm charac- terization for open manifolds. In [5], Carvalho, Nistor and Qiao introduced a very large class of manifolds, called manifolds with amenable ends. To any manifold M0 belonging to this class, we can associate a family of manifolds Mα, for some α ∈ A, which are acted upon by Lie groups Gα, and such that the following theorem holds. Theorem 1.1 ( [5, Theorem 1.1] ). Let P be a "compatible" operator on M0. Then one can associate some Gα-invariant differential operators Pα on Mα such that P : H s(M0) → H s−m(M0) is Fredholm if, and only if, (1) P is elliptic, and (2) each operator Pα : H s(Mα) → H s−m(Mα) is invertible. This very vague statement is made more precise in Theorem 5.3 below. The operators Pα should be thought of as "limit operators" giving some control on the behaviour of P at infinity. Note that Theorem 1.1 remains true if we replace P by an operator in a suitable pseudodifferential calculus or if we consider operators acting between vector bundles [5, 35]. Theorem 1.1 recovers in a unified setting many similar results that were pre- viously known in particular cases [9, 10, 12, 21, 22, 26, 27, 43]. This was made possible by noting that, in many cases, the relevant differential operators are gener- ated by the action of a Lie groupoid G on the manifold M0, in a sense made more precise below. Obtaining Fredholm conditions can then be reduced to studying the representations of the reduced C∗-algebra of G [5, 34]. This motivated the definition of Fredholm groupoids in [5]. Roughly speaking, a Fredholm groupoid is a Lie groupoid G whose unit space is a compact manifold with boundary M , and such that a ∈ 1 + C∗ r (G) is Fredholm ⇔ πx(a) is invertible for any x ∈ ∂M . Here πx stands for the regular representation of G at x. This is very similar in spirit to Theorem 1.1: an element is Fredholm if, and only if, some limit operators are invertible. Let us summarize this method. To a manifold M0, seen as the interior of a compact manifold with boundary M , one associates a Lie groupoid G ⇒ M whose action generates an interesting subalgebra of differential operators on M0. To obtain the Fredholm characterization of Theorem 1.1, the main objective is to prove that the groupoid G is Fredholm. The aim of this papers is to construct a large class of Fredholm groupoids. 1.2. Main results. To anticipate some further applications of the theory, we start by generalizing the definition of Fredholm groupoids to the case of locally compact groupoids. Indeed, for many practical applications, it seems that it would be inter- esting to extend the framework of Theorem 1.1 and the tools surrounding it to the setting of continuously family groupoids [24, 36]. THE FREDHOLM PROPERTY FOR GROUPOIDS IS A LOCAL PROPERTY 3 If G ⇒ X is a topological groupoid and U ⊂ X an open subset, we denote by GU = GU U the reduction of G to U , that is the subgroupoid of G made of all elements whose domain and range lie in U . The saturation of U is the set G · U = r(d−1(U )). Our point in this paper is that the Fredholm property for groupoids is a local property, in the following sense. Theorem 1.2. Let G ⇒ X be a locally compact, second-countable and locally Hausdorff groupoid, endowed with a right-invariant Haar system. Assume that (1) there is an open dense G-invariant subset V ⊂ X with GV ≃ V × V , and (2) we have a family (Ui)i∈I of open subsets of X such that the saturations (G · Ui)i∈I provide an open cover of X. Then G is a Fredholm groupoid if, and only if, each reduction GUi is also a Fredholm groupoid, for i ∈ I. Most results giving sufficient conditions for a groupoid G to be Fredholm assume that G is Hausdorff, which sometimes is not so easy to check [5, 34]. This is not a requirement for Theorem 1.2, which states that it is enough to look at the local structure of G (i.e. reductions). Since most groupoids studied in practical cases are locally very simple, this gives a powerful tool to prove the Fredholm property. To study the representation theory of G in terms of its reductions, we will use the induction theory of C∗-algebras [39, 42]. If G ⇒ X is a locally compact groupoid satisfying some extra assumptions, we associate to each open subset U ⊂ X a continuous induction map between the primitive spectra IndU : Prim(C∗(GU )) → Prim(C∗(G)), which is an homeomorphism onto its image. As a important step, we obtain the following result Theorem 1.3. Let G ⇒ X be a locally compact, second countable, and locally Hausdorff groupoid. Assume that we have a family of open subsets (Ui)i∈I such that their saturations (G · Ui)i∈I is an open cover of X. Then IndUi (Prim C∗(GUi )) . Prim C∗(G) = [i∈I Theorem 1.3 gives us a good description of the representation theory of G in terms of that of its reductions. This is the main tool for the proof of Theorem 1.2. 1.3. Applications and examples. A direct consequence of Theorem 1.2 is that gluing Fredholm groupoids together yields another Fredholm groupoid, which gener- alizes a result of [7]. Moreover, most groupoids appearing in practical applications have a very simple local structure: locally they are reductions of action groupoids. To formalize this, we introduce the class of local action groupoids, and prove some related results. In particular, we show that if a local action groupoid G is locally given by the action of an amenable group, then the groupoid G is Fredholm. Concrete example include the groupoids associated to manifolds with cylindrical ends, asymptotically euclidean manifolds and asymptotically hyperbolic manifolds. We also introduce a groupoid which models the analysis on manifolds with cuspidal points and prove that it is Fredholm. 1.4. Outline of the paper. We begin in Section 2 by recalling some known results concerning the primitive spectrum of a C∗-algebra and the induction mechanism. We then turn our attention to locally compact groupoids and introduce some defi- nitions and notations. Section 3 introduce our main tool, which is the induction functor from the C∗- algebra of a reduction to an open subset. We define this functor and establish some 4 RÉMI CÔME important properties, and then use it to prove the decomposition of the primitive spectrum stated in Theorem 1.3. It is in Section 4 that we start dealing with Fredholm groupoids. We introduce our definition of Fredholm groupoids in the locally compact case and prove Theorem 1.2. We then establish a few consequences. Among them, we define the notion of local isomorphisms of two groupoids and the class of local action groupoids. Finally, Section 5 gives some concrete examples of Fredholm groupoids. To motivate their construction, we step back into the setting of Lie groupoids and recall the link with the study of differential operators on manifolds. We then show how Theorem 1.2 may be used to prove that the groupoids under study are Fredholm (and local action groupoids). Aknowledgements: the author would like to thank Victor Nistor for useful dis- cussions and suggestions. 2. Preliminaries 2.1. The primitive spectrum of a C∗-algebra. We recall in this section the definition of the primitive spectrum of a C∗-algebra, as well as the general induction mechanism for representations. The reader interested in more details may refer to the book of Dixmier [11] for more details on the primitive spectrum and to [39, 42] for the induction procedure. Definition 2.1. Let A be a C∗-algebra. An ideal J ⊂ A is called primitive if it is the kernel of a non-zero irreducible representation of A. The primitive spectrum of A, denoted Prim A, is the set of all primitive ideals in A. For any ideal I ⊂ A, let PrimI A = {J ∈ Prim A, I ⊂ J}, and denote by PrimI A the complement of PrimI A in Prim A. The sets PrimI A, where I ranges through the ideals of A, are precisely the open sets in the Jacobson topology of Prim A. The support of a representation π of A is the closed subset supp π := Primker π A = {J ∈ Prim A, ker π ⊂ J}. For any C∗-algebra A, let R(A) denote the category of unitary equivalence classes of non-degenerate representations of A. A well-known result states that, if I is an ideal of A and π a non-degenerate representation of I on a Hilbert space H, then there is a unique representation of A on H extending π [11]. This defines an induction functor IndA Moreover, the representation IndA Thus IndA I descends to a map I : R(I) → R(A). I π is irreducible if, and only if π is irreducible. IndA I : Prim I → Prim A, which is a homeomorphism onto PrimI A [11]. The ideal I is a particular case of an (A, I)-correspondence in the sense of [41, 42]. If A, B are C∗-algebras, a (A, B)-correspondence E is a full right Hilbert B- module endowed with a morphism π : A → LB(E) such that π(A)E is dense in E (it corresponds to the notion of a B-rigged A-module in [42]). To any such correspondence is associated an induction functor E -Ind : R(B) → R(A) given by the tensor product with E. If E is moreover an (A, B)-imprimitivity bi- module, i.e. both a full left A-module and a full right B-module satisfying some THE FREDHOLM PROPERTY FOR GROUPOIDS IS A LOCAL PROPERTY 5 compatibility conditions [39, 42], then A and B are said to be Morita equivalent and E -Ind is an equivalence of categories. In that case, the functor E -Ind descends to an homeomorphism E -Ind : Prim B → Prim A. 2.2. Groupoids and their C∗-algebras. We now turn our attention to topologi- cal groupoids and fix some definitions and notations. Recall that a groupoid consits of two sets: the set of arrows G and the set of units X, together with five structural morphisms: the domain and range d, r : G → X, the inverse ι : G → G, the inclusion of units u : X → G and the product µ from the space G(2) := {(g, h) ∈ G × G, d(g) = r(h)} of compatible arrows to G. Throughout the paper, we denote by G ⇒ X a groupoid If A is a subset of X, we denote by GA = r−1(A) and with set of units X. GA = d−1(A). The groupoid GA := GA ∩ GA, with units A, is the reduction of G to A. Finally, we denote by G · A := {r(g) g ∈ G, d(g) ∈ A} = r(d−1(A)) the saturation of A in X through the action of G. The reader wishing to learn more about groupoids should refer to [25, 40]. Definition 2.2. A topological groupoid is a groupoid G ⇒ X such that (1) the sets G and X are topological spaces, with X Hausdorff, (2) all five structural maps d, r, ι, u and µ are continuous, (3) the domain map d is open. It follows from Definition 2.2 that ι is a homeomorphism and r is open as well. We shall usually require the topological space G to be locally compact, second countable and locally Hausdorff (in the sense that each element of G should have a Hausdorff neighborhood). If G is Hausdorff, let Cc(G) be the space of C-valued continuous functions with compact support in G. If G is only locally Hausdorff, then Cc(G) denotes instead the linear space generated by continuous functions that are compactly supported in a Hausdorff subset of G (see [8] for why this is a better choice). To define a product on Cc(G), we recall below the standard notion of a Haar system from [40]. In the following definition, we denote by Rg : Gr(g) → Gd(g) the homeomorphism induced by right-multiplication by an element g ∈ G. Definition 2.3. Let G ⇒ X be a locally compact groupoid. A continuous, right- invariant Haar system on G is a family of Borel measures (λx)x∈X such that (1) for all x ∈ X, the support of λx is Gx = d−1(x), (2) (right-invariance) for any g ∈ G, we have (Rg)∗λr(g) = λd(g), (3) (continuity) for any f ∈ Cc(G), the function x 7→ZGx f (g)dλx(g) is continuous. Assume that G is endowed with a continuous, right-invariant Haar system. If f, g ∈ Cc(G), we define their convolution product f ∗ g ∈ Cc(G) by f ∗ g(x) =ZGd(x) f (xy−1)g(y)dλd(x)(y). The full C∗-algebra of G, denoted C∗(G), is the completion of Cc(G) for the norm kf kC ∗(G) = sup π kπ(f )k, 6 RÉMI CÔME where π ranges over all continuous, bounded representations of Cc(G) [40]. The reduced C∗-algebra of G, denoted C∗ r (G), is the completion of Cc(G) for the norm kf kC ∗ r (G) = sup x∈X kπx(f )k. Here πx : Cc(G) → B(L2(Gx)) is the regular representation of G at x, defined by πx(f )ξ = f ∗ ξ. Finally, let G ⇒ X and H ⇒ Y be two locally compact, second-countable and locally Hausdorff groupoids. Recall that G and H are Morita equivalent if there exists a locally compact space Z with a left G-action and a right H-action, such that both actions are free and proper, commute with each other and the anchors Z → X and Z → Y induces bijections G\Z ≃ Y and Z/H ≃ X. Morita equivalent groupoids have Morita equivalent C∗-algebras [33, 44]. It follows from Subsection 2.1 that there are homeomorphisms Prim C∗(G) ≃ Prim C∗(H) and Prim C∗ r (G) ≃ Prim C∗ r (H). 3. Primitive spectrum and groupoid reductions In this section, we show that the primitive spectrum of a groupoid C∗-algebra can be investigated locally: this is the content of Theorem 3.12, which will be our main tool for Section 4. Throughout the section, we shall consider a locally compact, second-countable and locally Hausdorff groupoid G ⇒ X that is endowed with a right-invariant continuous Haar system. 3.1. Representations induced from a reduction. We will show that each re- duction of the groupoid to an open subset U ⊂ X defines an induction functor between the categories of representations of GU = d−1(U ) ∩ r−1(U ) and G. The starting point for our construction is Remark 3.1 below. Remark 3.1. If U ⊂ X is an open subset and W := G · U its saturation, then the reduction GU is Morita equivalent to GW . The equivalence is implemented by the (GW , GU )-space GU = d−1(U ), acted upon by left and right multiplication. Both actions are free and proper, and the domain and range maps induce isomorphisms GW \GU ≃ U and GU /GU ≃ W . According to the results recalled in Section 2, it follows that C∗(GU ) and C∗(GW ) are Morita equivalent [33]. The corresponding (C∗(GW ), C∗(GU ))-imprimitivity bimodule EU is the completion of Cc(GU ) for the norm kf kEU = kf ∗ ∗ f k 1 2 C ∗(GU ), with C∗(GW ) and C∗(GU ) acting by right and left multiplication respectively. Sim- ilarly, there is a Morita equivalence between the reduced algebras C∗ r (GU ) and C∗ r (GW ). We choose to stick to the study of the full algebras for now, but all results of this section apply to their reduced counterparts, as pointed by Remark 3.8. It is well-known that for any G-invariant open subset W ⊂ X, the C∗-algebra C∗(GW ) embeds as an ideal in C∗(G) [32]. We saw in Subsection 2.1 that this implies the existence of an induction functor IndW : R(C∗(GW )) → R(C∗(G)), between the respective categories of non-degenerate representations. Definition 3.2. Let G ⇒ X be a locally compact, second countable and locally Hausdorff groupoid. Let U be an open subset of X, and W := G ·U be its saturation. The induced representation functor IndU : R(C∗(GU )) → R(C∗(G)) THE FREDHOLM PROPERTY FOR GROUPOIDS IS A LOCAL PROPERTY 7 is defined as the composition IndU = IndW ◦EU -Ind. Remark 3.3. A possibly more direct way to define the functor IndU is to ob- serve that EU is a (C∗(G), C∗(GU ))-correspondence, as introduced in Subsection 2.1. Correspondingly, the space GU is a (G, GU )-correspondence (or Hilsum-Skandalis morphism) in the sense of [16, 41]. We do not emphasize this approach too much however, because the factorization of IndU through R(C∗(GW )) given by Definition 3.2 will be of use below. Lemma 3.4. Let G ⇒ X be a locally compact, second-countable and locally com- pact groupoid, and let U ⊂ X be open. If π is a non-degenerate representation of C∗(GU ), then for any f ∈ Cc(GU ) we have kπ(f )k ≤ k IndU π(f )k. Proof. Recall that IndU π is a representation of C∗(G) on H := EU ⊗π H. Let HU ⊂ H be the closed subspace generated by linear combinations of tensors f ⊗ ξ, with ξ ∈ H and f ∈ Cc(GU ). Let also ρ be the restriction of IndU π to Cc(GU ). If f, g ∈ Cc(GU ) and ξ, η ∈ H, then ρ(f )(g ⊗ ξ) = (f ∗ g) ⊗ ξ is in HU , so HU is stable by ρ. Moreover hf ⊗ ξ, g ⊗ ηiHU := hπ(g∗ ∗ f )ξ, ηiH = hπ(f )ξ, π(g)ηiH , so the map Φ : f ⊗ ξ 7→ π(f )ξ extends as an isometry from HU to H. Since π is non-degenerate, the map Φ is an isomorphism. Now Φ(ρ(f )g ⊗ ξ) = π(f )Φ(f ⊗ ξ), so ρ and π are unitarily equivalent. This shows that kπ(f )k = kρ(f )k ≤ k IndU π(f )k for any f ∈ Cc(GU ). (cid:3) A corollary is the following not-so-obvious result, for which we could not find a reference in the litterature. Corollary 3.5. Let G ⇒ X be a locally compact, second countable and locally Hausdorff groupoid. Let U be an open subset of X. Then C∗(GU ) is a subalgebra of C∗(G). Proof. Let f ∈ Cc(GU ). Every representation of C∗(G) restricts as a bounded representation of Cc(GU ), so kf kC ∗(G) ≤ kf kC ∗(GU ). Conversely, if π is a non- degenerate representation of C∗(GU ), then IndU π is a representation of C∗(G) such that kπ(f )k ≤ k IndU (f )k by Lemma 3.4. Therefore kf kC ∗(GU ) ≤ kf kC ∗(G). This shows that the inclusion of Cc(GU ) into Cc(G) extends to a continuous and injective ∗-morphism from C∗(GU ) into C∗(G). (cid:3) For any open subset U ⊂ X, set PrimU C∗(G) := {J ∈ Prim C∗(G), C∗(GU ) ⊂ J}, and let PrimU C∗(G) be its complementary subset in Prim C∗(G). Lemma 3.6. Let U ⊂ X be open, and W = G · U be its saturation. Then PrimU C∗(G) = PrimW C∗(G) and PrimU C∗(G) = PrimW C∗(G). Proof. The algebra C∗(GW ) is the closed, two-sided ideal generated by C∗(GU ) in C∗(G). Thus, if J is a primitive ideal of C∗(G) that contains C∗(GU ), then J also contains all of C∗(GW ). On the other hand, it is obvious that if J contains C∗(GW ), then it also contains the subalgebra C∗(GU ). This proves that PrimU C∗(G) = PrimW C∗(G), and therefore that PrimU C∗(G) = PrimW C∗(G). (cid:3) 8 RÉMI CÔME We can now record the main properties of IndU . Proposition 3.7. Let G ⇒ X be a locally compact, second countable and locally Hausdorff groupoid. Let U be an open subset of X. (1) The functor IndU descends to a continuous map IndU : Prim C∗(GU ) → Prim C∗(G), which is an homeomorphism onto PrimU C∗(G). (2) Let π, ρ be two non-degenerate representations of C∗(GU ) such that π is weakly contained in ρ. Then IndU π is weakly contained in IndU ρ. (3) If π is a non-degenerate representation of C∗(GU ), then (4) For x ∈ U , let πU IndU (supp π) ⊂ supp(IndU π). x be the corresponding regular representation of C∗(GU ) and πx the one of C∗(G). Then IndU πU x = πx. Proof. According to Definition 3.2, the functor IndU is defined as the composition IndW ◦EU -Ind. We highlighted in Section 2 that both IndW and EU -Ind induce continuous maps between primitive spectra, that are homeomorphims onto their respective images. Therefore, the map (2) IndU : Prim C∗(GU ) → Prim C∗(G) ker π 7→ ker(IndU π) is well-defined, continuous, and an homeomorphism onto PrimW C∗(G). The latter coincides with PrimU C∗(G) by Lemma 3.6, which proves Assertion (1). Assertion (2) is a well-known property of the Rieffel induction procedure, whose proof can be found in [39]. Assertion (3) is a direct consequence. Indeed, let J = ker ρ be a primitive ideal contained in supp π. By definition, this is equivalent to ρ ≺ π. Then IndU ρ ≺ IndU π, which means that ker(IndU π) ⊂ ker(IndU ρ). Since IndU J = ker(IndU ρ) by Equation (2), we conclude that IndU J ∈ supp(IndU (π)). This proves the inclusion IndU (supp π) ⊂ supp(IndU π). To prove Assertion 4, let H = EU ⊗πU x ). We need to show that the map L2(GU x Φ : H → L2(Gx) defined by Φ : f ⊗ ξ 7→ f ∗ ξ extends to a Hilbert space isomorphism. Let f, g ∈ Cc(GU ) and ξ, η ∈ Cc(GU definition x ). By hf ⊗ ξ, g ⊗ ηiH = h(g∗ ∗ f ) ∗ ξ, ηiL2(GU x ) = h(g∗ ∗ f ) ∗ ξ, ηiL2(Gx) = hf ∗ ξ, g ∗ ηiL2(Gx), so Φ is an isometry. To show that Φ is onto, let f ∈ Cc(GU ), and let (ξn)n∈N be an approximate unit in Cc(GU ). Then (f ∗ ξn)Gx converges to f Gx in L2(Gx). This proves that the image of Φ contains the dense subset Cc(Gx), hence Φ is onto. The map Φ is therefore an isomorphism. Now, let ρ = IndU πU x . If g ∈ Cc(G), then Φ(ρ(g)(f ⊗ ξ)) := Φ((g ∗ f ) ⊗ ξ) = g ∗ (f ∗ ξ) = πx(g)(f ∗ ξ) = πx(g)Φ(f ⊗ ξ). Since Cc(G) is dense in C∗(G) and ρ and πx are continuous, this proves that IndU πU x and πx define the same class in R(C∗(G)). (cid:3) Remark 3.8. All the constructions of this section can be made in the same way by replacing every full groupoid algebras by their reduced counterparts. More ex- plicitely, if G ⇒ X is a groupoid satisfying the assumptions of Definition 3.2 and U ⊂ X an open subset, then there is an induction functor r (GU )) → R(C∗(G)). IndU : R(C∗ THE FREDHOLM PROPERTY FOR GROUPOIDS IS A LOCAL PROPERTY 9 All the properties from Proposition 3.7 follow if we replace each full agebra by its reduced counterpart. Remark 3.9. It should be noted that C∗(GU ) is a hereditary subalgebra [4] of C∗(G) (and so is C∗ r (G)). A hereditary subalgebra B ⊂ A is always Morita equivalent to the closed, two-sided ideal IB it generates: in our case this ideal is IC ∗(GU ) = C∗(GW ), with W = G · U . This yields an induction functor r (GU ) in C∗ IndA B : R(B) → R(A), which factorizes through R(IB), just as in Definition 3.2. This recasts our con- struction in a more general setting, for which most results we have shown in this subsection should hold. 3.2. Decomposition of the spectrum. As in the previous sections, let G ⇒ X be a locally compact, second countable and locally Hausdorff groupoid, endowed with a right-invariant continuous Haar system. If f ∈ Cc(G) and ϕ ∈ C0(X), we follow the notation of [40] and denote by ϕf the function (ϕ ◦ r) · f (the central dots denotes scalar multiplication, and not convolution). Lemma 3.10. Let A be a C∗-algebra and (Iλ)λ∈Λ a family of ideals in A such that Pλ∈Λ Iλ = A. Then Prim A = [λ∈Λ Prim Iλ, where we identify Prim Iλ with its image PrimIλ A through IndA Iλ . The reader should refer to Subsection 2.1 for the definition of PrimIλ A and the induction map IndA Iλ. Proof. For all J ∈ Prim(A), there is a λ ∈ Λ such that Iλ 6⊂ J. Indeed, if that were not the case, then we would have A =Pλ∈Λ Iλ ⊂ J so J = A, which is not a primitive ideal. Therefore there is a λ ∈ Λ such that J ∈ PrimIλ (A), which proves the proposition. (cid:3) Lemma 3.11. Let ϕ ∈ C0(X), and define Mϕ : Cc(G) → Cc(G) by Mϕ(f ) = ϕf . Then Mϕ extends as a continuous linear map from C∗(G) to itself. Moreover, if U ⊂ X is a G-invariant open subset of X such that supp ϕ ⊂ U , then f 7→ ϕf extends as a continuous map from C∗(G) to C∗(GU ). The first part of the Lemma implies that C0(X) embeds into the multiplier algebra of C∗(G). Proof. The first statement was proven in [40, Proposition 1.14], in which it was shown that kϕf kC ∗(G) ≤ kϕk∞kf kC ∗(G). If U ⊂ X is an open subset such that ϕ ∈ Cc(U ), then ϕ ◦ r ∈ Cc(GU ). If U is moreover G-invariant, then GU = GU , so ϕf ∈ Cc(GU ). We know from [40] that C∗(GU ) is an ideal in C∗(G), so kϕf kC ∗(GU ) = kϕf kC ∗(G) ≤ kϕk∞kf kC ∗(G). This proves the continuity as a map to C∗(GU ). (cid:3) We are ready to prove one of the main theorems of this paper. Again, recall that the definition of PrimU C∗(G) and IndU were introduced in the previous subsection. 10 RÉMI CÔME Theorem 3.12. Let G ⇒ X be a locally compact, second countable and locally Hausdorff groupoid. Assume that we have a family of open subsets (Ui)i∈I such that their saturations (G · Ui)i∈I form an open cover of X. Then Prim C∗(G) = [i∈I Prim C∗(GUi ), where we identify Prim C∗(GUi ) with its image PrimUi C∗(G) through IndUi . Theorem 3.12 is a localization result: the primitive spectrum of C∗(G) can be fully described by restricting our attention to sufficiently many reductions of G to open subsets of the unit space. Proof. Put Wi := G · Ui, for each i ∈ I. The assumption is that (Wi)i∈I is an open cover of X, so let (ϕi)i∈I be a partition of unity subordinate to that cover. If a ∈ C∗(G), then it follows from Lemma 3.11 that ϕia is well defined for all i and belongs to C∗(GWi ). Since Pi∈I ϕi = 1, we have a =Pi∈I ϕia. Thus C∗(GWi ), C∗(G) =Xi∈I which is all we need to apply Lemma 3.10. Therefore Prim C∗(G) = [i∈I PrimWi C∗(G), and we conclude with the identification PrimWi C∗(G) = PrimUi C∗(G) established in Proposition 3.7. (cid:3) Remark 3.13. As was already highlighted in Remark 3.8, all results from this paper remain valid if we replace the full groupoid algebras with their reduced couterparts. More explicitely, under the assumptions of Theorem 3.12, there is a decomposition Prim C∗ r (G) = [i∈I Prim C∗ r (GUi ), where we identify Prim C∗ r (G) through IndUi . Note that the technical Lemma 3.11 is much more easy to prove in the reduced case. Indeed there is no need to use Renault's disintegration theorem here, since we only have to deal with the regular representations of G. r (GUi ) with its image PrimUi C∗ Remark 3.14. It should also be noted that a decomposition similar to that of Theorem 3.12 holds for the full spectrum of C∗(G) (i.e. equivalence classes of irre- ducible representations as defined in [11]). Under the assumptions of Theorem 3.12, we may write \C∗(G) = [i∈I \C∗(GUi ). where \C∗(GUi ) is identified with its image through IndUi . 3.3. Families of representations. The main motivation for Theorem 3.12 is to study the representations of C∗(G) from the representations of its reductions. In particular, it was proven in [34, Proposition 2.1] that a family of representations F of a C∗-algebra A is faithful if, and only if Prim A = [π∈F supp π. THE FREDHOLM PROPERTY FOR GROUPOIDS IS A LOCAL PROPERTY 11 Corollary 3.15. We follow the assumptions of Theorem 3.12. For each i ∈ I, let Fi be a faithful family of non-degenerate representations of C∗(GUi ). Then the family F := {IndUi π i ∈ I, π ∈ Fi} is faithful for C∗(G). Proof. By assumption, for all i ∈ I we have Prim C∗(GUi ) = [π∈Fi supp πi. Using Theorem 3.12, we get (3) Prim C∗(G) = [i∈I PrimUi C∗(G) = [i∈I [π∈Fi IndUi (supp π). It was proven in Proposition 3.7 that IndUi(supp π) ⊂ supp(IndUi π) for any non- degenerate representation π of C∗(GUi ). Thus [π∈Fi IndUi (supp π) ⊂ [π∈Fi supp(IndUi π) Together with Equation (3), we obtain Prim C∗(G) ⊂ [i∈I [π∈Fi supp(IndUi π) ⊂ [i∈I [π∈Fi supp(IndUi π) = [π∈F supp π. The converse inclusion is trivial. This shows that F is a faithful family. (cid:3) As a direct application, recall that a groupoid G is called metrically amenable if the canonical morphism C∗(G) → C∗ r (G) is an isomorphism [40]. Corollary 3.16. Under the assumptions of Theorem 3.12, assume that each groupoid GUi is metrically amenable, for all i ∈ I. Then G is metrically amenable. Proof. For each i ∈ I, let Fi = (πUi x )x∈Ui be the family of all regular representations of GUi . The groupoid GUi is metrically amenable if, and only if, the family Fi is faithful for C∗(GUi ). Now recall from Proposition 3.7 that IndUi πUi x = πx, which is the regular representation of G at x. Corollary 3.15 implies that the family F = (πx)x∈X is faithful for C∗(G). This in turn is equivalent to G being metrically amenable. (cid:3) Definition 3.17 ( Nistor and Prudhon [34] ). Let A be a C∗-algebra. A family F of representations of A is called exhaustive if Prim A = [π∈F supp π. Exhaustive families provide a refinment of faithful families and will be used in Section 4. Corollary 3.18. We follow the assumptions of Theorem 3.12. For each i ∈ I, let Fi be a exhaustive family of representations of C∗(GUi ). Then the family {IndUi π i ∈ I, π ∈ Fi} is exhaustive for C∗(G). The proof is the same as that of Corollary 3.15. 12 RÉMI CÔME 4. Fredholm groupoids The class of Fredholm Lie groupoid was introduced by Carvalho, Nistor and Qiao in [5] as an important tool to study differential equations on manifolds with singularities. Our main result (Theorem 4.5) is that the Fredholm property is local: a groupoid is Fredholm if, and only if, all its reductions to open subsets of the unit space are also Fredholm groupoids. This motivates the definition of local action groupoids in Subsection 4.3.3, which occur naturally in many practical cases. 4.1. Definitions. Let G ⇒ X be a locally compact, second countable, locally Haus- dorff groupoid with a continuous right-invariant Haar system. Throughout this subsection, we will assume that there is a G-invariant, open and dense orbit V ⊂ X such that GV ≃ V × V . Such a set V is necessarily unique. Define the vector representation π0 : C∗ r (G) → B(L2(V )), as the equivalence class of any regular representation πx, for any x ∈ V (all those representations are conjugated through the action of G on its fibers Gx = d−1(x)). Fredholm groupoid are tailored to study differential operators on V , so one usually asks V to have a smooth structure: this is the case, for example, when G is a Lie groupoid, or more generally a continuous family groupoid [24, 37]. However, the differential setting is not needed for the results we seek; thus our definition of a Fredholm groupoids is a strict extension of the original one from [5]. Definition 4.1. A Fredholm groupoid is a locally compact, second-countable, lo- cally Hausdorff groupoid G ⇒ X, endowed with a continuous right-invariant Haar system, such that (1) there is an open, dense G-orbit V such that GV ≃ V × V , (2) the vector representation π0 : C∗ (3) for any a ∈ C∗ r (G) → B(L2(V )) is injective, and r (G), the operator 1 + π0(a) is Fredholm in B(L2(V )) if, and only if, each operator 1 + πx(a) is invertible, for every x ∈ X \ V . An equivalent definition of Fredholm groupoids was given in [5]. Recall the concept of an exhaustive family of representations from Definition 3.17. Proposition 4.2. Let G ⇒ X be a locally compact, second-countable and locally Hausdorff groupoid, endowed with a continuous right-invariant Haar system. Then G is a Fredholm groupoid if, and only if, all the following conditions are met: (1) there is an open, dense G-orbit V such that GV ≃ V × V , (2) the vector representation π0 : C∗ r (G) → C∗ (3) the restriction map C∗ C∗ r (G)/C∗ r (G) → B(L2(V )) is injective, r (GF ) induces an isomorphism r (GV ) ≃ C∗ r (GF ), where F = X \ V , and (4) the family of representations (πx)x∈F is exhaustive for C∗ r (GF ). This Proposition was proven in [5] for Lie groupoids, but without making any use of the smooth structure: thus it extends without any modification to our setting. Note that Conditions (3) and (4) may be checked at once by stating that the family (πx)x∈F is exhaustive for the quotient algebra C∗ r (G)/C∗ r (GU ). Recall the definition of a metrically amenable groupoid from Subsection 3.3. Theorem 4.3. Let G ⇒ X be a locally compact and second-countable groupoid endowed with a continuous right-invariant Haar system. Assume that there is an open, dense and G-invariant subset V ⊂ X such that GV ≃ V × V , and put F = X \ V . Assume moreover that G is Hausdorff and GF metrically amenable. Then G is Fredholm. THE FREDHOLM PROPERTY FOR GROUPOIDS IS A LOCAL PROPERTY 13 Theorem 4.3 gives a sufficient condition for Fredholmness which is satisfied by many groupoids encountered in practical cases (see Subsection 5.2 for examples). When G is moreover a Lie groupoid, its dense orbit V is called a manifold with amenable ends [5]. Proof. This result was proven in [5] for Lie groupoids, so we will only give a sketch of the proof here. First, it follows from a lemma of Khoskham and Skandalis [20] (and the density of V in X) that the vector representation is always injective when G is Hausdorff. This proves Condition (2) of Proposition 4.2. The amenability of GF and GV ≃ V × V imply that G is also metrically amenable. r (G)/C∗ r (GF ) [40], which proves Condition (3). Condition (4) is a result of Nistor and if GF is metrically amenable, then its set of regular representations r (GF ) [34, Theorem 3.18]. This follows from the Effros- (cid:3) It is then a standard fact that the restriction map induces an isomorphism C∗ C∗ Prudhon: (πx)x∈F is exhaustive for C∗ Hahn conjecture, which was proven for amenable groupoids [18]. r (GU ) ≃ Many examples of Fredholm groupoids (as well as their relation with the study of differential equations on open manifolds) will be given in Section 5. 4.2. The Fredholm property is local. Our aim in this section is to use the results of Section 3 to prove our main result, Theorem 1.2. In a nutshell, we show that a groupoid G is Fredholm if, and only if, all its reductions to open subsets of the unit are Fredholm. Lemma 4.4. Let G ⇒ X be a Fredholm groupoid. Then, for any open set U ⊂ X, the reduction GU is also a Fredholm groupoid. Proof. Let V ⊂ X be the unique open dense G-orbit such that GV ≃ V × V , and put F = X \ V . Then V ′ := U ∩ V is the unique open dense GU -orbit such that GV ′ ≃ V ′ × V ′. Let a ∈ C∗ r (GU ). Because π0 is injective on C∗ r (G) and π0(C∗ r (V ′ × V ′)) ≃ K(L2(V ′)), there is an induced isomorphism π0 : C∗ r (GU )/C∗ r (GV ′) ≃ π0(C∗ r (GU ))/K(L2(V ′)). Therefore, for any a ∈ C∗ r (GU ), the operator 1 + π0(a) is Fredholm in B(L2(V ′)) r (GV ′). r (GU )/C∗ if, and only if, the class of 1+a is invertible in the unitarization of C∗ But C∗ r (GU ) is a subalgebra of C∗ r (GU )/C∗ r (G), and C∗ r (GV ′ ) ⊂ C∗ r (GV ′) = C∗ r (G)/C∗(GV ). r (GV ) ∩ C∗ C∗ r (GU ). Thus Hence, 1 + a is invertible in the unitarization of C∗ it is invertible as an element of the unitarization of C∗ r (GU )/C∗ r (GV ′ ) if, and only if, r (G)/C∗(GV ). Now, since G is a Fredholm groupoid, we deduce that 1 + π0(a) is Fredholm in B(L2(V ′)) if, and only if, the operator 1 + πx(a) is invertible for each x ∈ F . But πx(a) = 0 for all x /∈ U . Therefore, the operator 1 + π0(a) is Fredholm if, and only if, the operator 1 + πx(a) is invertible for each x ∈ F ∩ U = U \ V ′. This proves that GU is a Fredholm groupoid. (cid:3) We now establish the converse of Lemma 4.4. Theorem 4.5. Let G ⇒ X be a locally compact, second-countable and locally Hausdorff groupoid, endowed with a right-invariant Haar system. Assume that (1) there is an open dense G-invariant subset V ⊂ X with GV ≃ V × V , and (2) we have a family (Ui)i∈I of open subsets of X such that the saturations (G · Ui)i∈I provide an open cover of X. Then G is a Fredholm groupoid if, and only if, each reduction GUi is also a Fredholm groupoid, for every i ∈ I. 14 RÉMI CÔME Theorem 4.5 is the main result of this paper. It emphasizes the fact that the Fredholmness of a groupoid G is determined by its local structure. In particular, what really matters is the local structure in a neighborhood of the closed set F = X \ V , or in other words how the groupoid GF is glued to the pair groupoid GV = V × V . Proof of Theorem 4.5. Assume that each reduction GUi is a Fredholm groupoid, and let Vi ⊂ Ui be the unique open dense GUi -orbit such that GVi ≃ Vi × Vi. We only have to prove that G satisfies the assumptions (2), (3) and (4) of Proposition 4.2. 0 : C∗ First, for any i ∈ I, let πi r (GVi ) → B(L2(Vi)) be the vector representation of GVi . We know from Proposition 3.7 that IndUi πi 0 is the vector representation r (G) on B(L2(V )). Moreover, because GUi is Fredholm, the representation π0 of C∗ 0 is faithful. Corollary 3.15 implies that π0 is a faithful representation of C∗(G), πi which proves Assumption (2). Now, because (G · Ui)i∈I is an open cover of X, Theorem 3.12 implies that Prim C∗ r (G) = [i∈I Prim C∗ r (GUi ). Since Vi is a GUi -invariant open subset of Ui, we may expand this decomposition: Prim C∗ r (G) = [i∈I(cid:0) Prim C∗ (4) r (GVi )G Prim (C∗ r (GVi )![ [i∈I r (GFi ))(cid:1) Prim (C∗ r (GFi ))! , = [i∈I Prim C∗ r (GVi ) ≃ r (GFi ) given by the fact that GUi is a Fredholm groupoid. But the family (G·Vi)i∈I where we have put Fi := Ui \ Vi and used the isomorphism C∗ C∗ is an open cover of V , so another application of Theorem 3.12 yields r (GUi )/C∗ Prim C∗ r (GV ) = [i∈I Prim C∗ r (GVi ). By substituting this last expression in Equation (4), we obtain (5) Prim C∗ r (G) = Prim C∗ r (GV )[ [i∈I Prim C∗ r (GFi )! , On the other hand, because V is a G-invariant open subset, there is also a decomposition (6) Prim C∗ r (G) = Prim C∗ r (G)/C∗ r (GV )) Combining Equations (5) and (6) proves the inclusion Prim(C∗ r (G)/C∗ r (GFi ). r (GV )G Prim(C∗ r (GV )) ⊂ [i∈I Prim C∗ For i ∈ I and x ∈ Ui, let us denote by πi Recall from Proposition 3.7 that IndUi (πi of G at x), so IndUi (supp πi is exhaustive for C∗ family (πi union of the support of every supp(πi x the regular representation of GUi at x. x) = πx (with πx the regular representation x) ⊂ supp πx. Since GUi is a Fredholm groupoid, the r (GFi ) is the r (GFi ); in other words Prim C∗ x), for x ∈ Fi. Therefore x)x∈Fi Prim(C∗ r (G)/C∗ r (GV )) ⊂ [i∈I [x∈Fi IndUi (supp πi x) ⊂ [x∈F supp πx, THE FREDHOLM PROPERTY FOR GROUPOIDS IS A LOCAL PROPERTY 15 with F := X \ V =Si∈I Fi. On the other hand, the representation πx vanishes on r (GV ) for any x ∈ F , so that supp πx is contained in Prim(C∗ r (GV )). This r (G)/C∗ C∗ proves the equality Prim(C∗ r (G)/C∗ r (GV )) = [x∈F supp πx, which by definition indicates that the family (πx)x∈F is exhaustive for the quotient algebra C∗ r (GV ). This proves Assumptions (3) and (4) of Proposition 4.2 and concludes the proof that G is a Fredholm groupoid. Finally, the "only if" part of Theorem 4.5 is a consequence of Lemma 4.4 above. (cid:3) r (G)/C∗ 4.3. Consequences. We give here several corollaries of Theorema 4.5, which may be used as tools to check the Fredholm property for a given groupoid. Some concrete examples of groupoids and applications of these results will be shown in Section 5. 4.3.1. Gluing groupoids. Let (Ui)i∈I be an open cover of a locally compact, Haus- dorff space X. Assume that we are being given a family of locally compact groupoids (Gi ⇒ Ui)i∈I with isomorphisms φji : GiUi∩Uj → Gj Ui∩Uj , for all i, j ∈ I. A natural construction would be to glue this family into a "bigger" groupoid G ⇒ X. This requires some compatibility assumptions on the family (Gi)i∈I , as was studied in [7, 14]. For instance, the family is said to satisfy the weak gluing condition of [7] if (1) the isomorphisms (φij ) satisfy a cocycle condition with φkj φji = φki and φji = (φij )−1; and (2) for any any i, j ∈ I and any composable pair (gi, hj) ∈ Gi × Gj (that is, such that the domain of gi and the range of gj coincide), there is a k ∈ I and a composable pair (gk, hk) ∈ G(2) k with φik(gk) = gi and φjk(hk) = hj. Under those assumptions, the fibered coproduct of the family (Gi)i∈I along the isomorphisms (φij )i,j∈I , which we write G := Si∈I Gi acquires a natural groupoid structure over X (see [7]). The construction of this glued groupoid is such that each reduction GUi is naturally isomorphic to Gi. Theorem 4.6. Let (Ui)i∈I be an open cover of a locally compact Hausdorff space X, and let (Gi ⇒ Ui)i∈I be a family of groupoids satisfying the weak gluing condition. Let G =Si∈I Gi be the glued groupoid, and assume that (1) there is an open dense G-invariant subset V ⊂ X with GV ≃ V × V , and (2) each groupoid Gi is Fredholm, for i ∈ I. Then the groupoid G is Fredholm. Proof. By construction, each reduction GUi is isomorphic to Gi, hence Fredholm. Since the family (Ui)i∈I is an open cover of X, the result is a direct consequence of Theorem 4.5. (cid:3) 4.3.2. Local isomorphisms. Theorems 3.12 and 4.5 state that the primitive spec- trum of a groupoid's C∗-algebra can be studied locally. This suggests the following notion of local isomorphisms between groupoids. Definition 4.7. Let G ⇒ X and H ⇒ Y be two topological groupoid, and let p ∈ X. (1) A local isomorphism between G and H around p is a triplet (U, φ, V ), where U ⊂ X and V ⊂ Y are open subsets, with p ∈ U and is an isomorphism of topological groupoids. φ : GU → HV 16 RÉMI CÔME (2) We say that G is locally isomorphic to H around p, and we write G ∼p H, if there exists an isomorphism between G and H around p. Recall that the direct product of two groupoids G ⇒ X and H ⇒ Y is the groupoid G × H, with units X × Y , whose structural morphisms are the direct products of those of G and H. Lemma 4.8. Let G1 ⇒ X1 and G2 ⇒ X2 be two topological groupoids. Let p1 ∈ X1 and p2 ∈ X2. Assume that there are groupoids H1, H2 such that G1 ∼p1 H1 and G2 ∼p2 H2. Then G1 × G2 is locally isomorphic to H1 × H2 near (p1, p2). Proof. By assumptions, there are isomorphisms φ1 : G1U1 → H1V1 and φ2 : G2U2 → H2V2 , with p1 ∈ U1 and p2 ∈ U2. Then (p1, p2) ∈ U1 × U2 and φ1 × φ2 : (G1 × G2)U1×U2 → (H1 × H2)V1×V2 is an isomorphism, which proves the lemma. (cid:3) Subsection 4.3.1 introduced the gluing construction of a family of groupoids. We show that gluing groupoids preserves their local structure. Lemma 4.9. Let (Ui)i∈I be an open cover of a topological space X, and let (Gi ⇒ Ui)i∈I be a family of topological groupoids satisfying the weak gluing condition. Let i ∈ I and p ∈ Ui, and assume that there is a groupoid H such that Gi ∼p H. Then Gi ∼p H [i∈I Proof. Let G = Si∈I Gi be the glued groupoid. By definition, we have GUi ≃ Gi; hence any local isomorphism φ : GiU → HV around p induces a local isomorphism GUi∩U ≃ HV around p. (cid:3) Theorem 4.10. Let G ⇒ X be a locally compact, second-countable and locally Hausdorff groupoid. Assume that (1) there is an open dense G-invariant subset V ⊂ X with GV ≃ V × V , and (2) for each p ∈ X, there is a Fredholm groupoid Hp such that G ∼p Hp. Then G is a Fredholm groupoid. The point of Theorem 4.10 is to emphasize again that only the local structure is important to characterize Fredholm groupoids. Proof. Following the assumptions, there is for each p ∈ X an open set Up containing p and such that GUp is isomorphic to a reduction HpVp, with Hp a Fredholm groupoid. Lemma 4.4 implies that HpVp is Fredholm, so GUp is also Fredholm. The conclusion follows from Theorem 4.5 applied to the open cover (Up)p∈X . (cid:3) 4.3.3. Local action groupoids. Many Fredholm groupoids occuring in the study of differential equation on singular spaces are very simple on a local scale: they are locally isomorphic to action groupoids. To formalize this, we introduce here the class of local action groupoids. Many examples of such groupoids will be found in Subsection 5.2 below. Remark 4.11. Recall that, if G is a group acting on a space X on the right, then the corresponding action groupoid (or transformation groupoid) is written X ⋊ G and defined as follows. As a set, we put X ⋊ G := X × G. The domain and range maps are given by d(x, g) = x · g−1 and r(x, g) = x, whereas the product is (x, g)(x · g−1, h) = (x, hg) (see [7, 40] for more details). THE FREDHOLM PROPERTY FOR GROUPOIDS IS A LOCAL PROPERTY 17 If G and X are both locally compact, second-countable and Hausdorff, and if moreover the action is continuous, then X ⋊ G is a locally compact, second- countable, Hausdorff groupoid. The groupoid X ⋊ G is endowed with a natural Haar system (induced by the Haar measure on G), and the reduced groupoid C∗- algebra C∗ r (X ⋊ G) is then isomorphic to the crossed-product algebra C0(X) ⋊r G. Theorem 4.12. Let G be a topological group acting continuously on a space X. Assume that G and X are locally compact, Hausdorff and second-countable. Assume moreover that: (1) there is an open, dense G-orbit V ⊂ X such that the action of G on V is free, transitive and proper, (2) the group G is amenable. Then the groupoid X ⋊ G is Fredholm. Action groupoids of the sort occur naturally in the study of the essential spectrum of differential operators on groups, a notable example being that of the quantum N -body problem on Euclidean space [12, 27, 31]. Proof. Let G = X ⋊ G. Firstly, the assumptions on the action of G on V imply that the map α : V ⋊ G → V × V (x, g) 7→ (x, x · g−1) is continuous and bijective. Moreover α is proper with value in a Hausdorff space, hence closed. Therefore α is an homeomorphism, which shows that GV ≃ V × V Secondly, the amenability of G implies that the groupoid GF = F ⋊ G is metrically amenable [46], where F = X \ G. The result follows from Theorem 4.3. (cid:3) Definition 4.13. A locally compact and second countable groupoid G ⇒ X is said to be a local action groupoid if, for each p ∈ X, there is an action groupoid Xp ⋊ Gp such that G is locally isomorphic to Xp ⋊ Gp near p, where Gp and Xp are both locally compact, Hausdorff and second countable. The main point of Definition 4.13 is that the local structure of a such groupoids indeed, the C∗-algebras of action groupoids and their is very well understood: representations have been much studied in the litterature [46]. Several examples of local action groupoids shall be given in Section 5. Proposition 4.14. Let G and H be local action groupoids. Then G × H is also a local action groupoid. Proof. This follows from Lemma 4.8 and the fact that, if G1 = X1 ⋊ G1 and G2 = X2 ⋊ G2 are action groupoids, then G1 × G2 ≃ (X1 × X2) ⋊ (G1 × G2), where G1 × G2 acts on X1 × X2 by the product action. (cid:3) Proposition 4.15. Let (Gi)i∈I be a family of local action groupoids satisfying the weak gluing condition of Subsection 4.3.1. Then the glued groupoid G = Si∈I Gi is also a local action groupoid. Proof. This is a direct consequence of Lemma 4.9. (cid:3) Since "the Fredholm property is local", it is only natural that Theorem 4.12 generalizes to local action groupoids. Theorem 4.16. Let G ⇒ X be a local action groupoid. Assume that 18 RÉMI CÔME (1) there is an open, dense and G-invariant subset V ⊂ X such that GV ≃ V × V ; (2) for each p ∈ X, there is a local isomorphism G ∼p Xp ⋊ Gp, with Xp, Gp locally compact, second countable and Hausdorff, and Gp amenable. Then G is a Fredholm groupoid. In other words, if G is locally given by the action of an amenable group, then G is Fredholm. Section 5 will provide many examples of such groupoids. Proof. Set p ∈ X, and consider the local isomorphism φp : GUp ≃ (Xp ⋊ Gp)U ′ p, where p ∈ Up. Let Vp = V ∩ Up and V ′ p = φp(Vp). Because GV ≃ V × V , we have GVp ≃ Vp × Vp, hence (Xp ⋊ Gp)V ′ p ≃ V ′ p × V ′ p. Now, because Gi is amenable, the groupoid Xp ⋊ Gp is Hausdorff and amenable. It follows that its reduction (Xp ⋊ Gp)U ′ p is also Hausdorff and amenable. Theorem 4.3 therefore implies that each groupoid (Xp ⋊ Gp)Up is Fredholm. We conclude using Theorem 4.10 that G is a Fredholm groupoid. (cid:3) 5. Examples and Applications We conclude this paper with many examples of Fredholm and local action groupoids. All our examples are motivated by the study of differential equations on singular spaces, so we begin in Subsection 5.1 by discussing the notion of differential opera- tors induced by a Lie groupoid. 5.1. Differential operators. Many Fredholm groupoids that are studied in prac- tical cases are Lie groupoids, i.e. groupoids with a smooth structure. We work in the setting of manifolds with corners, in other words manifolds which are modelled on open subsets of the cube [0; 1]n, for n ∈ N (see [19] for more on that matter). A Lie groupoid in that context is defined as follows. Definition 5.1. A groupoid G ⇒ X is a Lie groupoid if (1) both G and X are manifolds with corners, with X Hausdorff, (2) the structural maps d, r, ι and u are smooth, (3) the map d is a tame submersion, and (4) the composition map µ is smooth. A submersion h : X → Y of manifolds with corners is said to be tame if, for all v ∈ T X, the vector h∗(v) ∈ T Y is inward-poiting if and only v is. If G is a Lie groupoid with units X, the tameness condition ensures that the fibers Gx = d−1(x), for x ∈ X, are smooth manifolds without corners [19]. Note that Definition 5.1 does not require G to be Hausdorff; however, because G is a manifold, it is always locally Hausdorff. The Lie algebroid of a Lie groupoid G ⇒ X is the bundle AG → X of all vectors tangent to the d-fibers Gx, for x ∈ X; in other words AG = [x∈X TxGx = (ker d∗)X . The differential of the range r : G → X gives the anchor map r∗ : AG → T X. [25] Assume now that X is compact and that there is an open, dense and G-invariant subset V ⊂ X such that GV ≃ V × V . Then r∗ is an isomorphism from AGV to T V . Thus any metric g on AG induces a Riemannian metric g0 on V , which we call compatible with AG. Such metrics g0 are always complete, and their equivalence class depends on AG only [1]. Associated to g0 is a well-defined scale of Sobolev spaces (H s(V ))s∈R, which all contain C∞ c (V ) as a dense subset [15]. Naturally H 0(V ) = L2(V ). THE FREDHOLM PROPERTY FOR GROUPOIDS IS A LOCAL PROPERTY 19 Definition 5.2. A differential operator P on a Lie groupoid G with compact units X is a family (Px)x∈X such that (1) for any x ∈ X, the element Px is a differential operator on Gx = d−1(x), (2) (right-invariance) for any g ∈ G, the right multiplication Rg : Gr(g) → Gd(g) gives a conjugation R∗ g−1 Pr(g)R∗ g = Pd(g), and (3) (smoothness) for any f ∈ C∞(G), the map x 7→ Pxfx is smooth, where fx = f Gx. We let Diff(G) be the algebra of all differential operators on G, and Diff m(G) the subspace of operators of order lesser or equal to m ∈ N. Operators of order 1 are just right-invariant vector fields on Sx∈X T Gx, which are in one-to-one correspon- dence with Γ(AG). Thus the algebra Diff(G) may be alternatively described as the universal envelopping algebra of the Lie algebroid AG [35]. The anchor map r∗ : Γ(AG) → Γ(T X), whose image can be restricted to U , induces an injective algebra morphism π0 : Diff(G) → Diff(V ), whose image we denote DiffG(V ) (correspondingly, we write Diff m of Diff m(G) through π0). Operators in Diff1 really be thought of as the "infinitesimal action" of G on its dense orbit V . G (V ) for the image G(V ) (that is, sections of AG) should Two important properties of Diff G(V ) where shown in [1]. First, the algebra Diff G(V ) contains every geometric operator associated to the compatible metric g0 (such as the Laplacian and any generalized Dirac operator). Secondly, any differential operator P ∈ Diff m G (V ) induces a bounded operator P : H s(V ) → H s−m(V ), for any s ∈ R. The main motivation for introducting and studying Fredholm groupoids is to obtain Fredholm conditions for the operators in Diff G(V ). This is illustrated by the following theorem. Theorem 5.3 (Carvalho, Nistor, and Qiao [5]). Let G ⇒ X be a Fredholm Lie groupoid with compact unit space X, and set V ⊂ X its unique dense, open G-orbit. Let P ∈ Diff m G (V ) and set s ∈ R. Then P : H s(V ) → H s−m(V ) is Fredholm if, and only if, (1) P is elliptic and (2) for any x ∈ X \ V , the operator Px : H s(Gx) → H s−m(Gx) is invertible. Recall that a differential operator P on V is called elliptic if its principal symbol σ(P ) ∈ C∞(T ∗M ) vanishes only on the zero-section [17]. The operators Px, for x ∈ X \ V , are called limit operators for P : they are invariant under the action of G on Gx, and are of the same type as P (e.g. if P is the Laplacian on V , then Px is the Laplacian on Gx). Note that Theorem 5.3 remains true if we condider pseudo- differential operators on G or operators acting between vector bundles sections [5]. Many similar results were known in particular cases, see [9, 10, 12, 21, 22, 26, 27, 43] and the reference therein. Remark 5.4. For many practical applications, it may be interesting to extend the definition of differential operators (Definition 5.2) and the corresponding Fredholm and boundedness results to the setting of groupoids that are only longitudinally smooth: see [24, 37]. 5.2. Examples. We conclude this paper with many examples of groupoids occur- ing in practical applications, and show that these groupoids are all local action groupoids, and in particular Fredholm. 20 RÉMI CÔME 5.2.1. The pair groupoid. Let M be a closed manifold, i.e. a compact smooth man- ifold without boundary. The pair groupoid G = M × M is then Fredholm: indeed, the vector representation π0 : C∗ r (M × M ) → B(L2(M )) is an isomorphism onto the ideal K of compact operators on L2(M ) [5]. Thus the operators 1 + π0(a), for a ∈ C∗ r (G), are all Fredholm. Assumption (3) of Definition 4.1 is trivially satisfied in that case, because the boundary set M \ M is empty. Here the algebra Diff G(M ) contains all differential operators on M . Theorem 5.3 then recovers the classical Fredolmness result: a differential operator on M is Fredholm if, and only if, it is elliptic. The groupoid G is a local action groupoid. Indeed, any point p ∈ M has a neighborhood U ⊂ M diffeomorphic to an open subset U ′ ⊂ Rn. Then where α is the action of Rn on itself by translation. GU ≃ U ′ × U ′ ≃ (Rn ⋊α Rn)U ′ , 5.2.2. Cylindrical ends. Consider the smooth action of R on R+ given by α : R × R+ → R+ (t, x) 7→ etx. The fundamental vector fields for this action are the image of the sections of the Lie algebroid of R+ ⋊ R through the anchor map; here they are generated by x∂x. Let now M be a manifold with boundary, and M0 its interior. The b-groupoid over M , which will be denoted Gb, can be obtained in several steps. (1) First, let H ⇒ R+ × ∂M be the product groupoid H = (R+ ⋊α R) × (∂M × ∂M ). The action of R on R+ is free and transitive on R∗ morphic to the pair groupoid of ]0, 1[×∂M . +, so H]0,1[×∂M is iso- (2) Let U ≃ [0, 1[×∂M be a tubular neighborhood of ∂M in M . We now have an identification over the interior: H]0,1[×∂M ≃ (]0, 1[×∂M )2 ≃ (M0 × M0)U\∂M . So H and M0 × M0 may be glued into a groupoid Gb ⇒ M (refer to Sub- section 4.3.1 for more details concerning the gluing procedure). We call Gb the b-groupoid of M . Any metric g0 on M0 that is compatible with Gb is bi-Lipschitz equivalent to the product metric dx2 x2 + h∂M on U ≃ [0, 1[×∂M , with h∂M a metric on ∂M . Thus Gb models manifolds with cylindrical ends. The algebra DiffGb (M0) is that of every differential operator P on M0 which can be written as P = Xα≤m aα(x∂x)α1 ∂α2 y2 . . . ∂αn yn , locally near any point of ∂M , with aα ∈ C∞(M ) and (∂yi )n i=2 a local basis of Γ(T ∂M ). It contains in particular any geometric operator associated to the metric g0. The algebra Diff b(M0) has been extensively studied, and is closely related to the b-calculus of Melrose and the Atiyah-Patodi-Singer index theorem of manifolds with boundaries [2, 30]. The groupoid Gb is obtained by gluing Hausdorff groupoids together, hence Gb is Hausdorff [7]. Moreover, the groupoid Gb is obtained by taking direct products THE FREDHOLM PROPERTY FOR GROUPOIDS IS A LOCAL PROPERTY 21 and gluing together several local action groupoids: it follows from Propositions 4.14 and 4.15 that Gb is also a local action groupoid. The local structure is very simple: for any p ∈ M , we have a local isomorphism G ∼p (R+ × Rn−1) ⋊ (R × Rn−1) The action is given by the product action of Rn−1 on itself (by translation) and R on R+ (by α). Since the acting groups is amenable, we conclude from Theorem 4.16 that Gb is a Fredholm groupoid. This is by no mean a new result [5], but should serve as a case in point to emphasizes the relevance of local action groupoids in practical cases. Remark 5.5. Carvalho and Qiao have recently considered a groupoid very close in nature to the b-groupoid and more suited to the study of boundary layer potential operators on manifolds with boundary [6]. Because the construction is very similar, it is clear that their groupoid is also a local action groupoid (with the same local picture as for Gb). 5.2.3. Cusp metrics. Example 5.2.2 can be generalized by replacing the action of R on R+ by a more general one. For example, let ϕ be a function in C0(R+, R+), vanishing only at 0 and such that (1) ϕ ∈ C∞(R∗ (2) ϕ′ is bounded on R∗ +. +), and Let α : R × R+ → R+ be the flow associated to the continuous vector field x 7→ ϕ(x)∂x on R+. The function ϕ is globally Lipschitz, so this flow is well-defined for any time t ∈ R. A typical example is any function ϕ ∈ C0(R+ ∩ C∞(R∗ +) satisfying (cid:26) ϕ(x) = xr ϕ(x) = 1 if x ∈ [0, 1], and if x ≥ 2, for any r ∈ [1; +∞). Under these assumptions, the open set R∗ + is an orbit of α on which the action is free and transitive, so we may construct a groupoid Gϕ ⇒ M by following the same gluing procedure as in Example 5.2.2. The groupoid Gϕ is a local action groupoid that is locally isomorphic to (R+ × Rn−1) ⋊α (R × Rn−1), hence it is Fredholm by Theorem 4.16. The compatible metrics are bi-Lipschitz equivalent to the complete metric g0(x, p) = dx2 (ϕ(x))2 + h∂M when (x, p) is in a tubular neighborhood [0, 1[×∂M of ∂M , with x > 0. This models manifolds with cusps, see e.g. [9, 22, 38]. The differential operators P ∈ Diff Gϕ(M0) can be written P = Xα≤m aα(ϕ(x)∂x)α1 ∂α2 y2 . . . ∂αn yn , locally near any point of ∂M , with aα ∈ C∞(M ) and (∂yi )n i=2 a local basis of Γ(T ∂M ). However, if ϕ is not smooth at x = 0, then the groupoid Gϕ is not a Lie groupoid, but only a continuous family groupoid [24, 36]. Obtaining Fredholm conditions for operators in Diff Gϕ (M0) therefore requires an extension of Theorem 5.3 to the setting of continuous family groupoids. 5.2.4. Scattering manifolds. Let M be a compact manifold with boundary and in- terior M0. A scattering metric (with respect to M ) is a Riemannian metric g0 on M0 such that, for a tubular neighborhood U ≃ [0, 1[×∂M of ∂M in M , we can write g0(x, p) = dx2 x4 + h∂M x2 , 22 RÉMI CÔME for any (x, p) ∈]0, 1[×∂M ≃ U ∩ M0, and with h∂M a metric on ∂M . A typical example is given by the stereographic compactification of Rn into the half-sphere Sn +: the euclidean metric on Rn is then a scattering metric for Sn + [29, 45]. For this reason, manifolds with scattering metrics are also called asymptotically euclidean. The algebra of scattering differential operators, written Diff sc(M0), is the one containing all differential operators P on M0 that can be written aα(x2∂x)α1 (x∂y2 )α2 . . . (x∂yn )αn , P = Xα≤m locally near any point of ∂M , with aα ∈ C∞(M ) and (∂yi )n Γ(T ∂M ). It contains in particular the Laplacian associated to g0. i=2 a local basis of It was shown in [5, 7] that there is a groupoid Gsc ⇒ M such that Diff G(M0) = Diff sc(M0). Moreover, the groupoid Gsc can be constructed by gluing reductions of the action groupoid Sn + is the only one extending the action of Rn on itself by translation. Thus Gsc is a local action groupoid that is locally isomorphic to Sn + ⋊ Rn around any point. It follows that Gsc is Fredholm by Theorem 4.16. This groupoid and closely related ones were studied in [31, 45] for instance, in relation with the study of the spectrum of the N -body problem on Euclidean space. + ⋊ Rn, where the smooth action of Rn on Sn 5.2.5. Asymptotically hyperbolic. As in Example 5.2.4, let M be a compact manifold with boundary and M0 its interior. An asymptotically hyperbolic metric (with respect to M ) is a Riemannian metric g0 on M0 that can be written g0(x, p) = dx2 + h∂M x2 in a tubular neighborhood [0, 1[×∂M , for x > 0 (here h∂M is a metric on ∂M , as before). A typical example is the hyperbolic space Hn, with its usual metric, compactified into the Poincaré ball. The interesting operators are those that can be written aα(x∂x)α1 (x∂y2 )α2 . . . (x∂yn )αn , P = Xα≤m locally near any point of ∂M , with aα ∈ C∞(M ) and (∂yi )n i=2 a local basis of Γ(T ∂M ). These operators form an algebra, which we call Diff0(M0). This algebra and related pseudodifferential calculi were studied in [10, 23, 28, 43] for instance. It was shown in [5, 7] that there is a groupoid G0 ⇒ M such that Diff G0 (M0) = Diff0(M0). Moreover, this groupoid can be obtained by gluing reductions of the action groupoid + ⋉ Rn−1). + acts on Rn−1 by dilation, and the action of Gn := R∗ Here R∗ right multiplication extends smoothly to the boundary by the formula Xn ⋊ Gn := (R+ × Rn−1) ⋊ (R∗ + ⋉ Rn−1 on itself by (x1, . . . , xn) · (t, ξ2, . . . , ξn) = (tx1, x2 + x1ξ2, . . . , xn + x1ξn), for all (x1, . . . , xn) ∈ Rn and (t, ξ2, . . . , ξn) ∈ Gn. This entails that G0 is a (Haus- dorff) local action groupoid, which is locally isomorphic to Xn ⋊ Gn around each point of M . Because Gn is amenable, Theorem 4.16 again implies that G0 is a Fredholm groupoid. References [1] B. Ammann, R. Lauter, and V. Nistor. "On the geometry of Riemannian manifolds with a Lie structure at infinity." In: International Journal of Mathematics and Mathematical Sciences 1-4 (2004), pp. 161 -- 193. REFERENCES 23 [2] M. F. Atiyah, V. K. Patodi, and I. M. Singer. "Spectral asymmetry and Riemannian geom- etry. III." In: Mathematical Proceedings of the Cambridge Philosophical Society 79.1 (1976), pp. 71 -- 99. [3] M. F. Atiyah and I. M. Singer. "The index of elliptic operators. I." In: Annals of Mathematics. Second Series 87 (1968), pp. 484 -- 530. [4] B. Blackadar. Operator algebras. Vol. 122. Encyclopaedia of Mathematical Sciences. The- ory of C ∗-algebras and von Neumann algebras, Operator Algebras and Non-commutative Geometry, III. Springer-Verlag, Berlin, 2006, pp. xx+517. [5] C. Carvalho, V. Nistor, and Y. Qiao. "Fredholm conditions on non-compact manifolds: theory and examples." In: (Mar. 23, 2017). arXiv: 1703.07953v1 [math.OA]. [6] C. Carvalho and Y. Qiao. "Fredholm Groupoids and Layer Potentials on Conical Domains." In: (July 14, 2018). arXiv: 1807.05418v1 [math.OA]. [7] R. Côme. "Gluing action groupoids: differential operators and Fredholm conditions." In: (Aug. 4, 2018). arXiv: 1808.01442v1 [math.DG]. [8] A. Connes. "A survey of foliations and operator algebras." In: Operator algebras and appli- cations, Part I (Kingston, Ont., 1980). Vol. 38. Proc. Sympos. Pure Math. Amer. Math. Soc., Providence, R.I., 1982, pp. 521 -- 628. [9] M. Dauge. Elliptic boundary value problems on corner domains. Vol. 1341. Lecture Notes in Mathematics. Smoothness and asymptotics of solutions. Springer-Verlag, Berlin, 1988, pp. viii+259. [10] C. Debord, J.-M. Lescure, and F. Rochon. "Pseudodifferential operators on manifolds with fi- bred corners." In: Université de Grenoble. Annales de l'Institut Fourier 65.4 (2015), pp. 1799 -- 1880. J. Dixmier. C ∗-algebras. Translated from the French by Francis Jellett, North-Holland Math- ematical Library, Vol. 15. North-Holland Publishing Co., Amsterdam-New York-Oxford, 1977, pp. xiii+492. [11] [12] V. Georgescu and A. Iftimovici. "Localizations at infinity and essential spectrum of quantum Hamiltonians. I. General theory." In: Rev. Math. Phys. 18.4 (2006), pp. 417 -- 483. [13] P. B. Gilkey. Invariance theory, the heat equation, and the Atiyah-Singer index theo- rem. Vol. 11. Mathematics Lecture Series. Publish or Perish, Inc., Wilmington, DE, 1984, pp. viii+349. [14] M. Gualtieri and S. Li. "Symplectic groupoids of log symplectic manifolds." In: Int. Math. Res. Not. IMRN 11 (2014), pp. 3022 -- 3074. [15] E. Hebey and F. Robert. "Sobolev spaces on manifolds." In: Handbook of global analysis. Elsevier Sci. B. V., Amsterdam, 2008, pp. 375 -- 415, 1213. [16] M. Hilsum and G. Skandalis. "Morphismes K-orientés d'espaces de feuilles et fonctorialité en théorie de Kasparov (d'après une conjecture d'A. Connes)." In: Ann. Sci. École Norm. Sup. (4) 20.3 (1987), pp. 325 -- 390. [17] L. Hörmander. The analysis of linear partial differential operators. III. Vol. 274. Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Pseudodifferential operators. Springer-Verlag, Berlin, 1985, pp. viii + 525. [18] M. Ionescu and D. Williams. "The generalized Effros-Hahn conjecture for groupoids." In: Indiana Univ. Math. J. 58.6 (2009), pp. 2489 -- 2508. [19] D. Joyce. "On manifolds with corners." In: Advances in geometric analysis. Vol. 21. Adv. Lect. Math. (ALM). Int. Press, Somerville, MA, 2012, pp. 225 -- 258. [20] M. Khoshkam and G. Skandalis. "Regular representation of groupoid C ∗-algebras and ap- plications to inverse semigroups." In: J. Reine Angew. Math. 546 (2002), pp. 47 -- 72. [21] V. A. Kondrat′ ev. "Boundary value problems for elliptic equations in domains with conical or angular points." In: Trudy Moskov. Mat. Obšč. 16 (1967), pp. 209 -- 292. [22] V. A. Kozlov, V. G. Maz′ ya, and J. Rossmann. Elliptic boundary value problems in do- mains with point singularities. Vol. 52. Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 1997, pp. x+414. [23] R. Lauter. "Pseudodifferential analysis on conformally compact spaces." In: Memoirs of the American Mathematical Society 163.777 (2003), pp. xvi+92. [24] R. Lauter, B. Monthubert, and V. Nistor. "Pseudodifferential analysis on continuous family groupoids." In: Doc. Math. 5 (2000), pp. 625 -- 655. [25] K. Mackenzie. Lie groupoids and Lie algebroids in differential geometry. Vol. 124. London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 1987, pp. xvi+327. [26] M. Măntoiu. "C ∗-algebras, dynamical systems at infinity and the essential spectrum of generalized Schrödinger operators." In: J. Reine Angew. Math. 550 (2002), pp. 211 -- 229. 24 REFERENCES [27] M. Măntoiu, R. Purice, and S. Richard. "Spectral and propagation results for magnetic Schrödinger operators; a C ∗-algebraic framework." In: J. Funct. Anal. 250.1 (2007), pp. 42 -- 67. [28] R. Mazzeo. "Elliptic theory of differential edge operators. I." In: Comm. Partial Differential Equations 16.10 (1991), pp. 1615 -- 1664. [29] R. B. Melrose. Geometric scattering theory. Stanford Lectures. Cambridge University Press, Cambridge, 1995, pp. xii+116. [30] R. B. Melrose. The Atiyah-Patodi-Singer index theorem. Vol. 4. Research Notes in Mathe- [31] matics. A K Peters, Ltd., Wellesley, MA, 1993, pp. xiv+377. J. Mougel, V. Nistor, and N. Prudhon. "A refined HVZ-theorem for asymptotically homo- geneous interactions and finitely many collision planes." In: Rev. Roumaine Math. Pures Appl. 62.1 (2017), pp. 287 -- 308. [32] P. S. Muhly, J. N. Renault, and D. P. Williams. "Continuous-trace groupoid C ∗-algebras. III." In: Trans. Amer. Math. Soc. 348.9 (1996), pp. 3621 -- 3641. [33] P. S. Muhly, J. N. Renault, and D. P. Williams. "Equivalence and isomorphism for groupoid C ∗-algebras." In: J. Operator Theory 17.1 (1987), pp. 3 -- 22. [34] V. Nistor and N. Prudhon. "Exhaustive families of representations and spectra of pseudo- differential operators." In: J. Operator Theory 78.2 (2017), pp. 247 -- 279. [35] V. Nistor, A. Weinstein, and P. Xu. "Pseudodifferential operators on differential groupoids." In: Pacific J. Math. 189.1 (1999), pp. 117 -- 152. [36] A. L. T. Paterson. "Continuous family groupoids." In: Homology Homotopy Appl. 2 (2000), pp. 89 -- 104. [37] A. L. T. Paterson. "The analytic index for proper, Lie groupoid actions." In: Groupoids in analysis, geometry, and physics (Boulder, CO, 1999). Vol. 282. Contemp. Math. Amer. Math. Soc., Providence, RI, 2001, pp. 115 -- 135. [38] V. Rabinovich, B.-W. Schulze, and N. Tarkhanov. "Boundary value problems in oscillating [39] [40] [41] cuspidal wedges." In: Rocky Mountain J. Math. 34.4 (2004), pp. 1399 -- 1471. I. Raeburn and D. P. Williams. Morita equivalence and continuous-trace C ∗-algebras. Vol. 60. Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 1998, pp. xiv+327. J. Renault. A groupoid approach to C ∗-algebras. Vol. 793. Lecture Notes in Mathematics. Springer, Berlin, 1980, pp. ii+160. J. Renault. "Induced representations and hypergroupoids." In: SIGMA Symmetry Integra- bility Geom. Methods Appl. 10 (2014), Paper 057, 18. [42] M. A. Rieffel. "Induced representations of C ∗-algebras." In: Advances in Mathematics 13 (1974), pp. 176 -- 257. [43] B.-W. Schulze. Pseudo-differential operators on manifolds with singularities. Vol. 24. Stud- ies in Mathematics and its Applications. North-Holland Publishing Co., Amsterdam, 1991, pp. vi+410. J.-L. Tu. "Non-Hausdorff groupoids, proper actions and K-theory." In: Doc. Math. 9 (2004), pp. 565 -- 597. [44] [45] A. Vasy. "Propagation of singularities in many-body scattering." In: Ann. Sci. École Norm. Sup. (4) 34.3 (2001), pp. 313 -- 402. [46] D. P. Williams. Crossed products of C ∗-algebras. Vol. 134. Mathematical Surveys and Mono- graphs. American Mathematical Society, Providence, RI, 2007, pp. xvi+528.
1301.4897	2	1301	2013-12-21T09:54:13	Deformation of C*-algebras by cocycles on locally compact quantum groups	[ "math.OA", "math.QA" ]	Given a C-algebra A with a left action of a locally compact quantum group G on it and a unitary 2-cocycle Omega on \hat G, we define a deformation A_Omega of A. The construction behaves well under certain additional technical assumptions on Omega, the most important of which is regularity, meaning that C_0(G)_Omega\rtimes G is isomorphic to the algebra of compact operators on some Hilbert space. In particular, then A_\Omega is stably isomorphic to the iterated twisted crossed product \hat G^{op}\ltimes_\Omega G\ltimes A. Also, in good situations, the C-algebra A_\Omega carries a left action of the deformed quantum group G_\Omega and we have an isomorphism G_\Omega\ltimes A_\Omega\cong G\ltimes A. When G is a genuine locally compact group, we show that the action of G on C_0(G)_Omega=C*_r(\hat G;Omega) is always integrable. Stronger assumptions of properness and saturation of the action imply regularity. As an example, we make a preliminary analysis of the cocycles on the duals of some solvable Lie groups recently constructed by Bieliavsky et al., and discuss the relation of our construction to that of Bieliavsky and Gayral.	math.OA	math	DEFORMATION OF C∗-ALGEBRAS BY COCYCLES ON LOCALLY COMPACT QUANTUM GROUPS SERGEY NESHVEYEV AND LARS TUSET Abstract. Given a C∗-algebra A with a left action of a locally compact quantum group G on it and a unitary 2-cocycle Ω on G, we define a deformation AΩ of A. The construction behaves well under certain additional technical assumptions on Ω, the most important of which is regularity, meaning that C0(G)Ω ⋊ G is isomorphic to the algebra of compact operators on some Hilbert space. In particular, then AΩ is stably isomorphic to the iterated twisted crossed product Gop ⋉Ω G ⋉ A. Also, in good situations, the C∗-algebra AΩ carries a left action of the deformed quantum group GΩ ∼= G ⋉ A. When G is a genuine locally compact group, we and we have an isomorphism GΩ ⋉ AΩ r ( G; Ω) is always integrable. Stronger assumptions of show that the action of G on C0(G)Ω = C ∗ properness and saturation of the action imply regularity. As an example, we make a preliminary analysis of the cocycles on the duals of some solvable Lie groups recently constructed by Bieliavsky et al., and discuss the relation of our construction to that of Bieliavsky and Gayral. Assume (H, ∆) is a Hopf algebra and A is a left H-module algebra, with the action of H denoted by h ⊗ a 7→ h ⊲ a. Assume also that Ω ∈ H ⊗ H is an invertible element satisfying the cocycle identity Introduction (Ω ⊗ 1)( ∆ ⊗ ι)(Ω) = (1 ⊗ Ω)(ι ⊗ ∆)(Ω). In this case we can consider a new Hopf algebra HΩ, defined by Drinfeld [13], such that HΩ = H as an algebra, but the coproduct is given by ∆Ω = Ω ∆(·)Ω−1. We can also define a new product ⋆Ω on AΩ = A by a ⋆Ω b = m(Ω−1 ⊲ (a ⊗ b)), where m(a ⊗ b) = ab. Then AΩ is an HΩ-module algebra and, as was observed by Majid [24] (see also [7]), for the corresponding smash, or crossed, products we have HΩ#AΩ ∼= H#A. (0.1) Our goal in this paper is to develop a similar theory in the context of C∗-algebras and actions of locally compact quantum groups. Thus, given a C∗-algebra A with an action of a locally compact quantum group G on it, and a unitary cocycle Ω ∈ L∞( G) ¯⊗L∞( G), we want to define a deforma- tion AΩ of A. Note that the deformed quantum group GΩ is defined in full generality by the theory developed by De Commer [10]. Particular cases of our construction of AΩ are of course well-known. For example, when G is dual to a discrete group Γ, A = Γ ⋉γ B (all crossed products in this paper are assumed to be reduced) and the action of G is the dual action γ, then AΩ is nothing else than the twisted crossed product Γ ⋉γ,Ω B, as defined already by Zeller-Meier [45] in the 60s. In the case when G is a compact group, a study of cocycles on G was initiated by Landstad [21] r ( G; Ω), which and Wassermann [40, 41] in the early 80s. They defined twisted group C∗-algebras C ∗ should be thought of as deformations C0(G)Ω. Date: January 21, 2013; revised version December 21, 2013. The research leading to these results has received funding from the Research Council of Norway and the Euro- pean Research Council under the European Union's Seventh Framework Programme (FP/2007-2013) / ERC Grant Agreement no. 307663. 1 2 S. NESHVEYEV AND L. TUSET Another milestone is the work of Rieffel [30] for G = R2d. He was motivated by deformation quantization theory and extended Weyl quantization to actions of R2d on C∗-algebras. His theory is beautiful, but quite complicated, based on an extension of oscillatory integrals to C∗-algebras. A much simpler, although less explicit, approach was later proposed by Kasprzak [17]. His idea was to start with isomorphism (0.1). This isomorphism implies that AΩ can be identified with the fixed point subalgebra of H#A with respect to a coaction of HΩ that corresponds to the dual coaction on HΩ#AΩ. It is easy to check that this coaction of HΩ on H#A is simply the dual coaction of H twisted by Ω. This allows one to describe AΩ in terms of H#A and Ω without using expressions like m(Ω−1 ⊲ (a⊗ b)) that are difficult to makes sense of in the analytic setting. Kasprzak developed this idea in the setting of C∗-algebras in the case when G is an abelian locally compact group and Ω is a continuous 2-cocycle on G. But his theory works equally well when G is the dual of a locally compact group. In the previous paper the first author together with Bhowmick and Sangha [3] extended Kasprzak's approach to the case of measurable cocycles. The problem with such cocycles is that the twisted dual action of G on G ⋉ A (since we deal now with group duals the deformed quantum group GΩ is G) is not always well-defined. Nevertheless, a description of AΩ in terms of generators continues to make sense and, as was shown in [3], these algebras still satisfy a number of properties to be considered as the correct deformations of A. A by no means exhaustive list of other relevant papers on cocycle deformations in the operator algebraic setting includes [4], [5], [6], [8], [11], [15], [16], [18], [22], [23], [32], [36], [37], [38], [39], [44]. In this paper we continue the work started in [3] and define the deformations AΩ for general locally compact quantum groups G and arbitrary unitary cocycles Ω on G. In fact, our primary interest is the group case. This is the situation studied by Landstad and Raeburn [22, 23] for C∗-algebras of the form A = C0(G/H) and a particular class of cocycles on G. But since, as follows from the above discussion, the deformed quantum group GΩ should play a role in the theory, trying to work only with groups and their duals is unnecessarily restrictive. Moreover, in the proofs of a significant number of general results there would be almost no simplifications even if we restricted ourselves to the group situation. The paper is organized as follows. In Section 1 we collect some basic facts on locally compact quantum groups. In Section 2 we study various notions related to cocycles. In particular, here we introduce twisted crossed products. They are related, but not in the most straightforward way, to cocycle crossed products studied by Vaes and Vainerman [36]. Another important notion is regularity of a cocycle, r ( G; Ω)⋊G is isomorphic to the algebra of compact operators which means that the crossed product C ∗ on L2(G). For regular quantum groups we show that regularity of a cocycle Ω is equivalent to the inclusion (K ⊗ 1) W Ω∗(1 ⊗ K) ⊂ K ⊗ K, where K is the algebra of compact operators on L2(G) and W is the multiplicative unitary of G. But we leave open the question of finding somewhat more manageable sufficient conditions for regularity. Section 3 contains our main general results. Here we introduce the deformed algebras AΩ and study such questions as the relation of AΩ to twisted crossed products, existence of an action of GΩ on AΩ, deformation in stages, invariance of AΩ under replacing Ω by a cohomologous cocycle. In Section 4 we specialize to the group case. The main goal is to understand when a cocycle on G is regular, but the outcome is far from satisfactory. We observe that Ω is regular if the action r ( G; Ω) is proper and saturated in the sense of Rieffel [29]. What we are able to prove of G on C ∗ in general is that this action always has a weaker property of integrability; note that integrability in an even weaker sense has already been established by Vaes and Vainerman [36]. One outcome of r ( G; Ω) ⊗ A under the this discussion is that AΩ is generated by the image of a dense subspace of C ∗ operator-valued weight C ∗ In Section 5, in order to illustrate some of our general results, as well as the difficulties one might encounter in analyzing concrete examples, we briefly consider the cocycles on the duals of some r ( G; Ω) ⊗ A → M (C ∗ r ( G; Ω) ⊗ A)G. DEFORMATION OF C∗-ALGEBRAS 3 solvable Lie groups recently constructed by Bieliavsky et al. [6, 5]. A detailed study will appear elsewhere. We finish the paper with a list of open problems in Section 6. Acknowledgement. We would like to thank Jyotishman Bhowmick for fruitful discussions and a careful reading of the manuscript. We are also grateful to Kenny De Commer for comments on the first version of the paper. In this section we will fix our notation and recall some facts on locally compact quantum groups that we will use freely throughout the paper. 1. Preliminaries 1.1. Locally compact quantum groups. Recall [19, 20] that a locally compact quantum group, in the von Neumann algebraic setting, is a pair G = (M, ∆) consisting of a von Neumann algebra M and a coassociative normal unital ∗-homomorphism ∆ : M → M ¯⊗M such that there exist a left invariant n.s.f weight ϕ and a right invariant n.s.f. weight ψ on M . We will often use the suggestive notation L∞(G) for M . Denote by L2(G) the space of the GNS-representation of M defined by the left invariant Haar weight ϕ. Write Λ : Nϕ → L2(G) for the corresponding map, where Nϕ = {x ∈ L∞(G) ϕ(x∗x) < ∞}. Then the multiplicative unitary W of G is defined by for x, y ∈ Nϕ. W ∗(Λ(x) ⊗ Λ(y)) = (Λ ⊗ Λ)(∆(y)(x ⊗ 1)), Therefore, identifying L∞(G) with its image under the GNS-representation defined by ϕ, we have ∆(x) = W ∗(1 ⊗ x)W for x ∈ L∞(G). Throughout the whole paper we will denote by K the C∗-algebra of compact operators on L2(G). We identify K ∗ with B(L2(G))∗. For a subset X of a normed space we denote by [X] the norm closure of the linear span of X. Using this notation the C∗-algebra C0(G) of continuous functions on G vanishing at infinity is defined by C0(G) = [(ι ⊗ ω)(W ) ω ∈ K ∗]. The dual quantum group G = ( M , ∆) is defined by M = {(ω ⊗ ι)(W ) ω ∈ K ∗}−σ-strong∗ , ∆(x) = ΣW (x ⊗ 1)W ∗Σ, where Σ is the flip on L2(G)⊗ L2(G). By definition M is represented on L2(G). This representation is identified with the GNS-representation defined by a left invariant Haar weight ϕ on G, with the corresponding map Λ : N ϕ → L2(G) uniquely defined by the identities (Λ((ω ⊗ ι)(W )), Λ(x)) = ω(x∗) for x ∈ Nϕ and suitable ω ∈ K ∗, namely, for ω such that the map Λ(x) 7→ ω(x∗) extends to a bounded linear functional on L2(G). Under this identification the multiplicative unitary W of G is given by W = ΣW ∗Σ. We then have C0( G) = [(ω ⊗ ι)(W ) ω ∈ K ∗] and W ∈ M (C0(G) ⊗ C0( G)). The pentagon relation for W can be written in the following equivalent forms: (∆ ⊗ ι)(W ) = W13W23, (ι ⊗ ∆)(W ) = W13W12. Denote by J, resp. J, the modular involutions on L2(G) defined by ϕ, resp. ϕ. Then J and J commute up to a scalar factor. The unitary antipode on M , resp. M , is given by R(x) = Jx∗ J , resp. R(a) = Ja∗J. We have (R ⊗ R)(W ) = W , that is, ( J ⊗ J)W ∗( J ⊗ J) = W. 4 S. NESHVEYEV AND L. TUSET In addition to W it is convenient to use another multiplicative unitary V corresponding to the GNS-representation defined by the right Haar weight ψ. It is defined by V = ( J ⊗ J) W ( J ⊗ J ) ∈ L∞( G)′ ¯⊗L∞(G), and we have ∆(x) = V (x ⊗ 1)V ∗ for x ∈ L∞(G). A locally compact quantum group G is called regular [1, 2], if (K ⊗ 1)W (1 ⊗ K) ⊂ K ⊗ K, or equivalently, (1 ⊗ K)W (K ⊗ 1) ⊂ K ⊗ K. This is equivalent to several other conditions. particular, if G is regular then [(K ⊗ 1)W (1 ⊗ K)] = K ⊗ K and [(C0(G) ⊗ 1)W (1 ⊗ C0( G))] = C0(G) ⊗ C0( G). Recall that genuine locally compact groups, and therefore also their duals, are always regular. In 1.2. Actions on operator algebras. A left action of a locally compact quantum group G on a von Neumann algebra N is a normal unital injective ∗-homomorphism α : N → L∞(G) ¯⊗N such that (ι ⊗ α)α = (∆ ⊗ ι)α. A continuous left action of G on a C∗-algebra A is a nondegenerate injective ∗-homomorphism α : A → M (C0(G) ⊗ A) such that (ι ⊗ α)α = (∆ ⊗ ι)α and the following cancelation property holds: The following proposition is a small variation of results of Baaj, Skandalis and Vaes, see Propo- C0(G) ⊗ A = [(C0(G) ⊗ 1)α(A)]. sitions 5.7 and 5.8 in [2]. We include a complete proof for convenience. Proposition 1.1. Assume G is a regular locally compact quantum group, N is a von Neumann algebra and α : N → L∞(G) ¯⊗N is an action of G on N . For a subspace X ⊂ N define Xα = [(ω⊗ ι)α(X) ω ∈ K ∗] ⊂ N. Then for any C∗-subalgebra A ⊂ N , if Aα ⊂ A, then Aα is a C∗-algebra and αAα is a continuous action of G on Aα. Proof. Since AAα ⊂ A by assumption, we have Aα ⊃ [(ω ⊗ ι)α(A(ν ⊗ ι)α(A)) ω, ν ∈ K ∗] = [(ν ⊗ ω ⊗ ι)(α(A)23(∆ ⊗ ι)α(A)) ω, ν ∈ K ∗] = [(ν ⊗ ω ⊗ ι)(α(A)23V12α(A)13V ∗ 12) ω, ν ∈ K ∗] = [(ν ⊗ ω ⊗ ι)(α(A)23V12α(A)13) ω, ν ∈ K ∗] = [(ν ⊗ ω ⊗ ι)(α(A)23α(A)13) ω, ν ∈ K ∗], where in the last step we used that [(K ⊗ 1)V (1 ⊗ K)] = K ⊗ K by regularity. We thus see that AαAα ⊂ Aα. It is also clear that Aα is invariant under the ∗-operation. Thus Aα is a C∗-algebra. In order to show that αAα is an action, observe first that (Xα)α = Xα for any subspace X ⊂ N . Therefore replacing A by Aα we may assume that A = Aα. We then have [α(A)(C0(G) ⊗ 1)] = [α((ω ⊗ ι)α(A))(C0(G) ⊗ 1) ω ∈ K ∗] = [(ω ⊗ ι ⊗ ι)(V12α(A)13V ∗ = [(ω ⊗ ι ⊗ ι)(V12α(A)13(1 ⊗ C0(G) ⊗ 1)) ω ∈ K ∗] = [(ω ⊗ ι ⊗ ι)(α(A)13(1 ⊗ C0(G) ⊗ 1)) ω ∈ K ∗], 12(1 ⊗ C0(G) ⊗ 1)) ω ∈ K ∗] where in the last step we used that [(K ⊗ 1)V (1 ⊗ C0(G)] = K ⊗ C0(G) by regularity. Therefore [α(A)(C0(G) ⊗ 1)] = C0(G) ⊗ A. From this we conclude that α(A) ⊂ M (C0(G) ⊗ A) and αA is a continuous action of G. (cid:3) DEFORMATION OF C∗-ALGEBRAS 5 1.3. Crossed products and duality. Given a continuous left action α of a locally compact quan- tum group G on a C∗-algebra A, the reduced C∗-crossed product G ⋉α A (since we are going to consider only reduced crossed products, we omit r in the notation) is defined by G ⋉α A = [(C0( G) ⊗ 1)α(A)] ⊂ M (K ⊗ A). It is equipped with the dual continuous right action of G, or in other words, with a continuous left action α of Gop, which is the quantum group G with the opposite comultiplication ∆op on L∞( Gop) = L∞( G). Namely, we have α(x) = ( W op ⊗ 1)∗(1 ⊗ x)( W op ⊗ 1) for x ∈ G ⋉α A, where is the multiplicative unitary of Gop (see [20, Section 4]), so W op = (J ⊗ J) W (J ⊗ J) ∈ L∞( G) ¯⊗L∞(G)′ α(α(a)) = 1 ⊗ α(a) for a ∈ A and α(x ⊗ 1) = ∆op(x) ⊗ 1 for x ∈ C0( G). Then the double crossed product is Gop ⋉ α G ⋉α A = [(JC0(G)J ⊗ 1 ⊗ 1)( ∆op(C0( G)) ⊗ 1)(1 ⊗ α(A))]. Since the map Ad(W ∗ ⊗ 1) maps Gop ⋉ α G ⋉α A onto ∆op(x) = W (x ⊗ 1)W ∗ for x ∈ C0( G), [(JC0(G)JC0( G) ⊗ 1 ⊗ 1)(ι ⊗ α)α(A)]. In particular, if [JC0(G)JC0( G)] = K, which is another equivalent formulation of regularity of G, we get the Takesaki-Takai duality Gop ⋉ α G ⋉α A ∼= K ⊗ α(A) ∼= K ⊗ A. Assume now that G is regular and β is a continuous left action of Gop on a C∗-algebra B. Assume also that there exists a unitary X in M (C0(G) ⊗ B) such that (∆ ⊗ ι)(X) = X13X23 and (ι ⊗ β)(X) = W12X13. Consider the ∗-homomorphism η : B → M (K ⊗ B), η(x) = X ∗β(x)X. Then by a Landstad-type result of Vaes [35, Theorem 6.7], the space A = [(ω ⊗ ι)η(B) ω ∈ K ∗] ⊂ M (B) is a C∗-algebra, the formula defines a continuous left action of G on A, and η defines an isomorphism B ∼= G ⋉α A intertwining β with α. Note that if we already have (B, β) = (G ⋉α A, α), then we can take X = W ⊗ 1, in which case η becomes the map defining the Takesaki-Takai isomorphism. α(a) = X ∗(1 ⊗ a)X 6 S. NESHVEYEV AND L. TUSET 2. Dual cocycles 2.1. Twisted group algebras. Assume G is a locally compact quantum group. By a measurable unitary dual 2-cocycle on G, or a measurable unitary 2-cocycle on G, we mean a unitary element Ω ∈ L∞( G) ¯⊗L∞( G) such that We say that Ω is continuous if Ω ∈ M (C0( G) ⊗ C0( G)). Given a measurable unitary 2-cocycle Ω, the cocycle condition can be written as (Ω ⊗ 1)( ∆ ⊗ ι)(Ω) = (1 ⊗ Ω)(ι ⊗ ∆)(Ω). Indeed, we have ( ∆ ⊗ ι)( W Ω∗)Ω∗ 12 = ( W Ω∗)13( W Ω∗)23. (2.1) ( ∆ ⊗ ι)( W Ω∗) = W13 W23( ∆ ⊗ ι)(Ω∗) = W13 W23(ι ⊗ ∆)(Ω∗)(1 ⊗ Ω∗)(Ω ⊗ 1) = W13 W23 W ∗ W23Ω∗ 23Ω∗ 13 23Ω12, which is what we claimed. C∗-algebra C ∗ The von Neumann algebra C ∗ Identity (2.1) shows that the space of operators (ω ⊗ ι)( W Ω∗), ω ∈ K ∗, forms an algebra. The r ( G; Ω) generated by this algebra is called the reduced twisted group C∗-algebra of G. r ( G; Ω)′′ ⊂ B(L2(G)) is denoted by W ∗( G; Ω). The following theorem in full generality is quite nontrivial and follows from results of De Com- mer [9, Propositions 11.2.1 and 11.2.2], which, in turn, rely on an analogue of manageability of multiplicative unitaries for measured quantum groupoids established by Enock [14]. Theorem 2.1. We have C ∗ r ( G; Ω) = [(ω ⊗ ι)( W Ω∗) ω ∈ K ∗] and W Ω∗ ∈ M (K ⊗ C ∗ r ( G; Ω)). For regular quantum groups the theorem is, however, not difficult to prove. Indeed, when G r ( G; Ω) = [(ω ⊗ ι)( W Ω∗) ω ∈ K ∗] was proved in is a compact quantum group, the equality C ∗ [4, Lemma 4.9]. The same arguments work for any regular locally compact quantum group. In Section 3.2 we will also give a proof of this equality for arbitrary locally compact quantum groups r ( G; Ω). that is independent of results of De Commer, by constructing a different set of generators of C ∗ r ( G; Ω)) for regular quantum groups we can adapt the proof of [1, Proposition 3.6] of a similar result for the multiplicative unitary. For this, rewrite identity (2.1) as On the other hand, to show that W Ω∗ ∈ M (K ⊗ C ∗ W ∗ 12( W Ω∗)23( W Ω∗)12 = ( W Ω∗)13( W Ω∗)23. (2.2) Multiplying by K ⊗ 1 ⊗ 1 on the right and applying the slice maps to the second leg we get [(ι ⊗ ω ⊗ ι)( W ∗ 12( W Ω∗)23(K ⊗ 1 ⊗ 1)) ω ∈ K ∗] = W Ω∗(K ⊗ C ∗ r ( G; Ω)). Using that [(1 ⊗ K) W ∗(K ⊗ 1)] = K ⊗ K by regularity, we see that the left hand side equals K ⊗ C ∗ r ( G; Ω), so W Ω∗(K ⊗ C ∗ r ( G; Ω)) = K ⊗ C ∗ r ( G; Ω). Similarly, rewriting (2.2) as ( W Ω∗)12( W Ω∗)∗ 23 = ( W Ω∗)∗ 23 W12( W Ω∗)13, multiplying this identity by K ⊗ 1 ⊗ 1 on the left and applying the slice maps to the second leg, we get Therefore W Ω∗ ∈ M (K ⊗ C ∗ Let us also note the following. K ⊗ C ∗ r ( G; Ω)). r ( G; Ω) = (K ⊗ C ∗ r ( G; Ω)) W Ω∗. DEFORMATION OF C∗-ALGEBRAS 7 Proposition 2.2. If Ω is a continuous unitary 2-cocycle on G, then Proof. Using identity (2.2) in the form W Ω∗ ∈ M (C0( G) ⊗ C ∗ r (G; Ω)). ( W Ω∗)13 = W ∗ 12( W Ω∗)23 W12Ω∗ 12( W Ω∗)∗ 23, from W Ω∗ ∈ M (K ⊗ C ∗ slice maps to the second leg we conclude that W Ω∗ ∈ M (C0( G) ⊗ C ∗ r ( G; Ω)) we see that ( W Ω∗)13 ∈ M (C0( G) ⊗ K ⊗ C ∗ r (G; Ω)). r ( G; Ω)). Applying the (cid:3) The von Neumann algebras W ∗( G; Ω) (in fact, more general von Neumann-algebraic cocycle crossed products) were extensively studied by Vaes and Vainerman [36]. In particular, in [36, Propo- sition 1.4] they showed that there exists a right action β of G on W ∗( G; Ω) such that (ι ⊗ β)( W Ω∗) = W13( W Ω∗)12. This action is given by Another useful formula, which follows from (2.2), see [36, Proposition 1.5], is β(x) = V (x ⊗ 1)V ∗ for x ∈ W ∗( G; Ω). β(x) = ( W Ω∗)21(1 ⊗ x)( W Ω∗)∗ 21. (2.3) r ( G; Ω) defines a continuous action of G on the C∗- Proposition 2.3. The restriction of β to C ∗ algebra C ∗ Proof. Since W ∈ M (K ⊗ C0(G)), from the equality (ι ⊗ β)( W Ω∗) = W13( W Ω∗)12 we get r ( G; Ω). (K ⊗ 1 ⊗ C0(G))(ι ⊗ β)( W Ω∗) = (K ⊗ 1 ⊗ C0(G))( W Ω∗)12. Applying the slice maps to the first leg we get [(1 ⊗ C0(G))β(C ∗ r ( G; Ω))] = C ∗ r ( G; Ω) ⊗ C0(G), which proves the proposition. 2.2. Deformed quantum group. Given a unitary 2-cocycle Ω ∈ L∞( G) ¯⊗L∞( G), we can define a new coproduct ∆Ω on L∞( G) by (cid:3) ∆Ω(x) = Ω ∆(x)Ω∗ for x ∈ L∞( G). By a result of De Commer [10], the pair GΩ = (L∞( G), ∆Ω) is again a locally compact quantum group. We will use the subscript Ω to denote the objects related to GΩ, such as the coproduct, the multiplicative unitary, etc. In order to describe the multiplicative unitary WΩ of GΩ we need to recall some results of Vaes and Vainerman [36]. By [36, Lemma 1.12] the action β of G on W ∗( G; Ω) is integrable, meaning that (ι ⊗ ϕ)β is a n.s.f. operator valued weight from W ∗( G; Ω) to W ∗( G; Ω)β = C1. Therefore we have a n.s.f. weight ϕ on W ∗( G; Ω) such that ϕ(x)1 = (ι ⊗ ϕ)β(x) for x ∈ W ∗( G; Ω)+. By construction W ∗( G; Ω) is represented on L2(G). By [36, Proposition 1.15] this representation can be identified with the GNS-representation defined by the weight ϕ, with the corresponding map Λ : N ϕ → L2(G) given by (2.4) for suitable ω ∈ K ∗. Denote by J the modular involution on L2(G) defined by ϕ. The von Neu- mann algebra L∞( GΩ) = L∞( G) ⊂ B(L2(G)) is in the standard form, so JΩ = J , and by [10, Proposition 5.4] we have (2.5) Λ((ω ⊗ ι)( W Ω∗)) = Λ((ω ⊗ ι)( W )) WΩ = ( J ⊗ J)Ω W ∗(J ⊗ J)Ω∗. From this we immediately get the following proposition. 8 S. NESHVEYEV AND L. TUSET Proposition 2.4. For any measurable unitary 2-cocycle Ω ∈ L∞( G) ¯⊗L∞( G) on G, the element Ω∗ is a measurable unitary 2-cocycle on GΩ, and we have r ( GΩ; Ω∗) = J C ∗ r ( G; Ω) J . C ∗ Therefore C ∗ r ( G; Ω) is ∗-anti-isomorphic to C ∗ r ( GΩ; Ω∗). By Proposition 2.3 we have a continuous r ( GΩ; Ω∗), where GΩ is the dual of GΩ. Using the unitary antipode RΩ(x) = right action of GΩ on C ∗ JΩx∗ JΩ = J x∗ J on C0(GΩ) we can transform this action to a continuous left action βΩ of GΩ on C ∗ r ( G; Ω). Lemma 2.5. We have and r ( G; Ω), βΩ(x) = W ∗ Ω(1 ⊗ x)WΩ for x ∈ C ∗ (ι ⊗ βΩ)( W Ω∗) = ( W Ω∗)13(cid:0)( J ⊗ J) W ∗ Ω( J ⊗ J)(cid:1)12. r ( GΩ; Ω∗) is given by x 7→ VΩ(x⊗ 1)V ∗ Proof. The right action of GΩ on C ∗ of GΩ on C ∗ r ( G; Ω) is defined by Ω. Therefore the left action Since βΩ(x) = ( J ⊗ J )(VΩ( Jx J ⊗ 1)V ∗ Ω )21( J ⊗ J ). (VΩ)21 = ( J ⊗ J)( WΩ)21( J ⊗ J) = ( J ⊗ J)W ∗ Ω( J ⊗ J), we get the first formula for βΩ in the formulation. Similarly, since the right action of GΩ on C ∗ r ( GΩ; Ω∗) maps (ω ⊗ ι)( WΩΩ) into (ω ⊗ ι ⊗ ι)(( WΩ)13( WΩΩ)12), we have (ω ⊗ βΩ)((J1 ⊗ J) WΩΩ(J2 ⊗ J)) = (ω ⊗ ι ⊗ ι)((J1 ⊗ J ⊗ J)( WΩ)12( WΩΩ)13(J2 ⊗ J ⊗ J)) for any ω ∈ K ∗ and any bounded antilinear operators J1 and J2. Since taking J1 = J and J2 = J we get WΩΩ = ( J ⊗ J)( W Ω∗)∗(J ⊗ J), which is exactly the second formula in the formulation. (ι ⊗ βΩ)(( W Ω∗)∗) = (cid:0)( J ⊗ J) WΩ( J ⊗ J)(cid:1)12( W Ω∗)∗ 13, (cid:3) Therefore we have a left action βΩ of GΩ and a right action β of G on C ∗ r ( G; Ω). Using that (ι⊗ β)( W Ω∗) = W13( W Ω∗)12 and the second formula in the lemma above, we see that these actions commute: (βΩ ⊗ ι)β = (ι ⊗ β)βΩ. 2.3. Twisted crossed products. Assume Ω is a measurable unitary 2-cocycle on G and α is a continuous left action of Gop on a C∗-algebra A. Definition 2.6. The reduced twisted crossed product Gop ⋉α,Ω A is defined as the C∗-subalgebra of M (K ⊗ A) generated by (J J C ∗ Proposition 2.7. We have Gop ⋉α,Ω A = [(J J C ∗ α(x) = Ad(cid:0)(1 ⊗ J J ⊗ 1)(W ∗ defines a continuous left action of GΩ on Gop ⋉α,Ω A. r ( G; Ω) JJ ⊗ 1)α(A)], and the formula Ω ⊗ 1)(1 ⊗ JJ ⊗ 1)(cid:1)(1 ⊗ x) r ( G; Ω) JJ ⊗ 1)α(A). Proof. The first part is proved in the standard way. Namely, observe first that DEFORMATION OF C∗-ALGEBRAS 9 whence W op = (J ⊗ J) W (J ⊗ J) = (1 ⊗ J J ) W ∗(1 ⊗ JJ), ∆op(x) = Ad(cid:0)(1 ⊗ J J) W Ω∗(1 ⊗ JJ)(cid:1)(1 ⊗ x) for x ∈ L∞( G). It follows that (1 ⊗ J J ⊗ 1)( W Ω∗ ⊗ 1)(1 ⊗ JJ ⊗ 1)(1 ⊗ α(A)) = (ι ⊗ α)α(A)(1 ⊗ J J ⊗ 1)( W Ω∗ ⊗ 1)(1 ⊗ JJ ⊗ 1). Applying the slice maps to the first leg and using that (K ⊗ 1)α(A) ⊂ K ⊗ A we conclude that [(J J C ∗ r ( G; Ω) JJ ⊗ 1)α(A)] ⊂ [α(A)(J J C ∗ r ( G; Ω) J J ⊗ 1)], and therefore Gop ⋉α,Ω A = [α(A)(J J C ∗ r ( G; Ω) JJ ⊗ 1)] = [(J J C ∗ r ( G; Ω) J J ⊗ 1)α(A)]. For the second part, note that the element (1 ⊗ J J ⊗ 1)(W ∗ Ω ⊗ 1)(1 ⊗ JJ ⊗ 1) ∈ L∞(GΩ) ¯⊗L∞( G)′ ⊗ 1 commutes with 1 ⊗ α(A). Therefore it suffices to check that the formula defines a continuous action of GΩ on J JC ∗ J JC ∗ r ( G; Ω) J J ∋ x 7→ Ad(cid:0)(1 ⊗ J J)W ∗ Ω(1 ⊗ JJ)(cid:1)(1 ⊗ x) r ( G; Ω) JJ. But this is true by Lemma 2.5. (cid:3) Twisted crossed products, or cocycle crossed products, in the von Neumann algebraic setting were defined by Vaes and Vainerman [36]. Our definition is of course related to theirs, but not in the most straightforward way. Namely, assume we are given a left action α of Gop on a von Neumann algebra 21) is a cocycle action of ( Gop)Ω21 = ( GΩ)op on N in the sense of [36, Definition 1.1]. N . Then (α, Ω∗ The von Neumann-algebraic cocycle crossed product of N by ( Gop)Ω21 is defined as the von Neumann algebra generated by α(N ) and W ∗(( Gop)Ω21; Ω∗ Lemma 2.8. Letting X = JJ , we have X ∈ L∞( G) and 21) ⊗ 1, see [36, Definition 1.3]. J JW ∗( G; Ω) J J = JX ∗ JW ∗(( Gop)Ω21; Ω∗ 21) J X J. Proof. The claim that X ∈ L∞( G) is in [10, Section 5]. By Proposition 2.4, applied to the cocycle Ω21 on Gop, we have W ∗(( Gop)Ω21; Ω∗ 21) = JW ∗( Gop; Ω21) J. By [10, Proposition 6.3] we also have Ω∗(X ⊗ X) = ∆(X)( R ⊗ R)(Ω21) = ∆(X)(J ⊗ J)Ω∗ 21(J ⊗ J), whence W Ω∗(X ⊗ X) = (1 ⊗ X) W (J ⊗ J)Ω∗ 21(J ⊗ J) = (1 ⊗ X)(J ⊗ J) W opΩ∗ 21(J ⊗ J), and therefore W ∗( G; Ω) = XJW ∗( Gop; Ω21)JX ∗ = JW ∗( Gop; Ω21) J . It follows that J JW ∗( G; Ω) J J = JJ J W ∗( Gop; Ω21) JJ J = JJ J JW ∗(( Gop)Ω21 ; Ω∗ 21) J JJ J, which is what we need. (2.6) (cid:3) Therefore up to conjugation by JX ∗ J ⊗ 1 our definition of the twisted crossed product by Gop is a C∗-algebraic version of the definition of Vaes and Vainerman of the cocycle crossed product by ( Gop)Ω21. 10 S. NESHVEYEV AND L. TUSET 2.4. Regular cocycles. Assume Ω is a measurable unitary cocycle on G. Recall that we have a continuous right action β of G on C ∗ r ( G; Ω), so we can consider the reduced crossed product C ∗ r ( G; Ω) ⋊β G = [β(C ∗ By (2.3) we have β(x) = ( W Ω∗)21(1⊗x)( W Ω∗)∗ Therefore the conjugation by this unitary maps C ∗ Definition 2.9. A cocycle Ω is called regular if [C ∗ r ( G; Ω))(1 ⊗ JC0( G) J )]. 21. The unitary ( W Ω∗)∗ r ( G; Ω) ⋊β G onto 1 ⊗ [C ∗ r ( G; Ω) JC0( G) J] = K. 21 commutes with 1⊗ J C0( G) J . r ( G; Ω) JC0( G) J ]. Note that by a version of the Takesaki duality [36, Proposition 1.20] the von Neumann algebra r ( G; Ω) JC0( G) J coincides with B(L2(G)) (this will also become clear from the proof generated by C ∗ of Proposition 2.11 below). Therefore regularity of Ω is equivalent to the formally weaker condition Since the representation of C ∗ formulation of regularity of Ω is that the C∗-algebra C ∗ compact operators on some Hilbert space. r ( G; Ω)⋊β G on L2(G) is faithful and irreducible, yet another equivalent r ( G; Ω) ⋊β G is isomorphic to the algebra of r ( G; Ω) J C0( G) J ⊂ K. C ∗ By definition, regularity of the trivial cocycle 1 is the same as regularity of G. Therefore regularity of cocycles is definitely not automatic. Even for regular locally compact quantum groups regularity of a cocycle is a very delicate question. The only easy cases seem to be covered by the following proposition. Proposition 2.10. Any measurable unitary 2-cocycle on G is regular in the following cases: (i) G is a genuine locally compact group; (ii) G is a discrete quantum group. Proof. Part (i) is well-known and is proved in the same way as regularity of G, by observing that r ( G; Ω) J C0( G) J contains a lot of integral operators. Part (ii) is obvious, as already the the space C ∗ algebra C0( G) consists of compact operators. (cid:3) In view of various equivalent characterizations of regularity of quantum groups, it is natural to wonder how regularity of a cocycle is related to properties like (K ⊗ 1) W Ω∗(1 ⊗ K) ⊂ K ⊗ K. We have the following result. Proposition 2.11. For a cocycle Ω on G consider the following conditions: (i) Ω is regular; (ii) (K ⊗ 1) W Ω∗(1 ⊗ K) ⊂ K ⊗ K. Then (i) ⇒ (ii). If G is regular, then the two conditions are equivalent. Proof. In this proof it will be convenient to consider the right action β of G on C ∗ action β′ of Gop, so β′(x) = V21(1 ⊗ x)V ∗ to stabilization this action is dual. Namely, by [36, Propositions 1.8 and 1.9] the unitary Y = V ∗ defines a left action γ of G on B(L2(G)) by r ( G; Ω) as the left r ( G; Ω). On the von Neumann algebra level, up 21Ω∗ 21 21 for x ∈ C ∗ and we have an isomorphism γ(x) = Y (1 ⊗ x)Y ∗, W ∗( G; Ω) ¯⊗B(L2(G)) ∼= (L∞(G) ⊗ 1 ∪ γ(B(L2(G))))′′, x 7→ Y xY ∗, intertwining β′ ⊗ ι with the dual action γ, defined by γ(x) = V21(1 ⊗ x)V ∗ 21. Note also that and by (2.6), V ∗ 21 = (J ⊗ J) W (J ⊗ J) = W op W Ω∗( J ⊗ J) = (J ⊗ J) W opΩ∗ 21, so that DEFORMATION OF C∗-ALGEBRAS 11 We are now ready to prove the proposition. Y = V ∗ 21Ω∗ 21 = (J ⊗ J) W Ω∗( J ⊗ J). Assume condition (i) holds. We claim that the restriction of γ to K defines a continuous action r ( G; Ω), it suffices to of G on K. Since K = [C ∗ show that the restriction of γ to JC0( G) J defines a continuous action. But this is clear, since Ω21 commutes with 1 ⊗ JC0( G) J and therefore r ( G; Ω) J C0( G) J] and Y commutes with 1 ⊗ C ∗ γ( Jx J) = W op(1 ⊗ Jx J)( W op)∗ = (J ⊗ J)( W op)∗(1 ⊗ x) W op(J ⊗ J) = (J ⊗ J) ∆op(x)(J ⊗ J). It follows that (K ⊗ 1)Y (1 ⊗ K)Y ∗(K ⊗ 1) = (K ⊗ 1)γ(K)(K ⊗ 1) ⊂ K ⊗ K. This means exactly that (K ⊗ 1)Y (1 ⊗ K) ⊂ K ⊗ K, which is equivalent to (ii). Assume now that G is regular and condition (ii) holds. We claim again that the restriction of γ to K defines a continuous action of G. By Proposition 1.1 it suffices to show that Kγ = K. Condition (ii) implies that (K ⊗ 1)γ(K)(K ⊗ 1) ⊂ K ⊗ K, whence Kγ ⊂ K. By Proposition 1.1 this already shows that Kγ is a C∗-algebra. Since it is σ-strongly∗ dense in [(ω ⊗ ι)γ(B(L2(G))) ω ∈ K ∗]′′ = B(L2(G)), it follows that Kγ = K. Next, we have an isomorphism intertwining β′ ⊗ ι with γ. This is a C∗-algebraic version of [36, Propositions 1.8], and the proof is basically the same. Briefly, we have the identity r ( G; Ω) ⊗ K ∼= G ⋉γ K, x 7→ Y xY ∗, C ∗ Y ∗ 23 W12Y23 = ( W Ω∗)12Y ∗ 13, which is proved similarly to (2.1). Applying the slice maps to the first leg we conclude that Ad Y ∗ maps G ⋉γ K = [(C0(G) ⊗ 1)γ(K)] = [(ω ⊗ ι ⊗ ι)( W ⊗ 1)(1 ⊗ Y )(1 ⊗ 1 ⊗ K)(1 ⊗ Y ∗) ω ∈ K ∗] onto [(ω ⊗ ι ⊗ ι)( W Ω∗ ⊗ 1)Y ∗ 13(1 ⊗ 1 ⊗ K) ω ∈ K ∗] = C ∗ r ( G; Ω) ⊗ K, as claimed. Consider now the double crossed product Gop ⋉γ G ⋉γ K. By the Takesaki-Takai duality it is isomorphic to K ⊗ K. What is however important to us, is only the equality [( J C0( G) J ⊗ 1)( G ⋉γ K)] = K ⊗ K, which is an immediate consequence of regularity of G, since [ JC0( G) J C0(G)] = K and (2.7) [(K ⊗ 1)γ(K)] = [(KC0( G) ⊗ 1)γ(K)] = K ⊗ K. Applying Ad Y ∗ to both sides of (2.7) and using that Y commutes with JC0( G) J ⊗ 1 we conclude that so Ω is regular. [ JC0( G) J C ∗ r ( G; Ω)] ⊗ K = K ⊗ K, (cid:3) 12 S. NESHVEYEV AND L. TUSET 3. Deformation of C∗-algebras 3.1. Quantization maps. Let Ω be a measurable unitary 2-cocycle on a locally compact quantum group G. Proposition 3.1. For every ν ∈ K ∗, the formula Tν (x) = (ι ⊗ ν)( WΩΩ(x ⊗ 1)( WΩΩ)∗) defines a right G-equivariant map C0(G) → C ∗ of G on C0(G) and C ∗ r ( G; Ω), respectively. Furthermore, we have r ( G; Ω), where we consider the right actions ∆ and β [Tν (C0(G)) ν ∈ K ∗] = [(ω ⊗ ι)( W Ω∗) ω ∈ K ∗] = C ∗ r ( G; Ω). Proof. The proof relies on the following identity: ( WΩΩ)23 W12( WΩΩ)∗ 23 = ( W Ω∗)12( WΩΩ)13. To prove it, write identity (2.2) for the cocycle Ω∗ on GΩ as ( WΩΩ)23( WΩΩ)12( WΩΩ)∗ 23 = ( WΩ)12( WΩΩ)13. (3.1) ( WΩΩ)23(cid:0)( J ⊗ J)Ω(J ⊗ J )(cid:1)12 Substituting ( WΩ)12 on both sides of the above identity with (cid:0)( J ⊗ J)Ω(J ⊗ J) W Ω∗(cid:1)12 we get Since ( WΩΩ)23 and (( J ⊗ J )Ω(J ⊗ J))12 commute, this is exactly (3.1). contained in C ∗ 23 = (cid:0)( J ⊗ J )Ω(J ⊗ J)(cid:1)12( W Ω∗)12( WΩΩ)13. Applying the slice maps to the first and the third legs of (3.1) we see that the image of Tν is W12( WΩΩ)∗ r ( G; Ω) and [Tν (C0(G)) ν ∈ K ∗] = [(ω ⊗ ι)( W Ω∗) ω ∈ K ∗]. It remains to check G-equivariance. For x ∈ C0(G) we compute: β(Tν (x)) = (ι ⊗ ι ⊗ ν)(V12( WΩΩ)13(x ⊗ 1 ⊗ 1)( WΩΩ)∗ = (ι ⊗ ι ⊗ ν)(( WΩΩ)13V12(x ⊗ 1 ⊗ 1)V ∗ = (Tν ⊗ ι)∆(x), 13V ∗ 12) 12( WΩΩ)∗ 13) which finishes the proof of the proposition. (cid:3) As a byproduct we get an alternative proof of part of Theorem 2.1, as promised earlier: the space [Tν (C0(G)) ν ∈ K ∗] is clearly self-adjoint, hence the algebra [(ω ⊗ ι)( W Ω∗) ω ∈ K ∗] is a C∗-algebra, so it coincides with C ∗ r ( G; Ω). The map Tν depends only on the restriction of ν to W ∗( GΩ; Ω∗) = JW ∗( G; Ω) J. It extends to a normal map L∞(G) → W ∗( G; Ω), which we continue to denote by Tν . Note also that, since W Ω∗ ∈ M (K ⊗ C ∗ r ( GΩ; Ω∗)) by Theorem 2.1, identity (3.1) implies that WΩΩ(C0(G) ⊗ 1)( WΩΩ)∗ is a nondegenerate C∗-subalgebra of r ( G; Ω)) and WΩΩ ∈ M (K ⊗ C ∗ M (C ∗ r ( G; Ω) ⊗ C ∗ r ( GΩ; Ω∗)) = M (C ∗ r ( G; Ω) ⊗ JC ∗ r ( G; Ω) J ). r ( G; Ω)), and the map Tν : M (C0(G)) → M (C ∗ This implies that Tν maps M (C0(G)) into M (C ∗ is strictly continuous on bounded sets. Example 3.2. Assume G is the dual of a discrete group Γ, so L∞(G) = W ∗(Γ) ⊂ B(ℓ2(Γ)) and ∆(λs) = λs ⊗ λs for s ∈ Γ. Then L∞( G) = ℓ∞(Γ), and a 2-cocycle on G is a 2-cocycle Ω : Γ× Γ → T on Γ in the usual sense. The multiplicative unitary W is defined by W (δs⊗δt) = δs⊗δst. The twisted s λΩ group C∗-algebra C ∗ t . In this case we have GΩ = G, and (3.1) gives us the known identity r ( G; Ω) is generated by the operators λΩ s = λsΩ(s,·) satisfying λΩ st = Ω(s, t)λΩ r ( G; Ω)) W Ω(λs ⊗ 1)( W Ω)∗ = λΩ s ⊗ λ ¯Ω s . DEFORMATION OF C∗-ALGEBRAS 13 r (Γ) → C ∗ r (Γ; Ω) are given by Tν (λs) = ν(λ ¯Ω Therefore the maps Tν : C ∗ We call the maps Tν the quantization maps. We will write T Ω ♦ ν for Tν when we want to stress that we consider the quantization maps defined by Ω. We can also define dequantization maps going in the opposite direction, although we will not need them in this paper. Namely, for ω ∈ W ∗( G; Ω)∗ we can define s )λΩ s . Sω : C ∗ r ( G; Ω) → C0(G), Sω(x) = (ω ⊗ ι)β(x). Since the elements of the form Tν(x), with x ∈ C0(G) and ν ∈ W ∗( GΩ; Ω∗)∗, span a dense subspace of C ∗ r ( G; Ω), the following computation shows that the image of Sω is contained in C0(G): 21) ∈ C0(G). SωTν(x) = (ω ⊗ ι)(Tν ⊗ ι)∆(x) = (ν ⊗ ω ⊗ ι)((WΩΩ)21(1 ⊗ ∆(x))( WΩΩ)∗ The maps Sω are again right G-equivariant. 3.2. Ω-Deformation. Assume now that A is a C∗-algebra and α is a continuous left action of G on A. Since, as we observed in the previous subsection, WΩΩ(C0(G)⊗ 1)( WΩΩ)∗ is a nondegenerate r ( G; Ω) ⊗ A extend C∗-subalgebra of M (C ∗ to maps r ( G; Ω) J), the maps Tν ⊗ ι : C0(G) ⊗ A → C ∗ r ( G; Ω) ⊗ JC ∗ Tν ⊗ ι : M (C0(G) ⊗ A) → M (C ∗ defined by (Tν ⊗ ι)(x) = (ν ⊗ ι ⊗ ι)(( WΩΩ)21(1 ⊗ x)( WΩΩ)∗ Definition 3.3. The Ω-deformation of a A is the C∗-subalgebra r ( G; Ω) ⊗ A), 21). generated by elements of the form (Tν⊗ι)α(a) for all ν ∈ K ∗ and a ∈ A. The maps (Tν⊗ι)α : A → AΩ are called the quantization maps. AΩ ⊂ M (C ∗ r ( G; Ω) ⊗ A) Note that since the maps Tν are right G-equivariant, we immediately see that AΩ ⊂ {x ∈ M (C ∗ r ( G; Ω) ⊗ A) (β ⊗ ι)(x) = (ι ⊗ α)(x)}. As a first example consider A = C0(G) with the action of G on itself by left translations, so α = ∆. In this case, using that (Tν ⊗ ι)∆(x) = β(Tν (x)) for all x ∈ C0(G), we get C0(G)Ω = β(C ∗ (3.2) This provides a different perspective on the action β of G on W ∗( G; Ω). This action was defined in [36] as a dual action on a twisted crossed product. We can now say that β is simply the right action of G on itself that survives under deformation. More precisely, we have the following general result. r ( G; Ω)) ∼= C ∗ r ( G; Ω). Proposition 3.4. Assume A is a C∗-algebra equipped with a continuous left action α of a locally compact quantum group G and a continuous right action γ of a locally compact quantum group H such r ( G; Ω)⊗A⊗C0(H)) that (ι⊗γ)α = (α⊗ι)γ. Then the restriction of ι⊗γ : M (C ∗ to AΩ defines a continuous right action of H on AΩ. Proof. For any ν ∈ K ∗ we have r ( G; Ω)⊗A) → M (C ∗ [(1 ⊗ 1 ⊗ C0(H))(ι ⊗ γ)(Tν ⊗ ι)α(A)] = [(Tν ⊗ ι ⊗ ι)(α ⊗ ι)(cid:0)(1 ⊗ C0(H))γ(A)(cid:1)] = [(Tν ⊗ ι)α(A)] ⊗ C0(H). This implies that From this we conclude that (ι⊗ γ)(AΩ) ⊂ M (AΩ ⊗ C0(H)) and the restriction of ι⊗ γ to AΩ defines a continuous action of H. (cid:3) [(1 ⊗ 1 ⊗ C0(H))(ι ⊗ γ)(AΩ)] = AΩ ⊗ C0(H). 14 S. NESHVEYEV AND L. TUSET 3.3. Relation to twisted crossed products. Assume B is a C∗-algebra and γ is a continuous left action of Gop on B. Consider the crossed product and the dual action α = γ of G on Gop ⋉γ B given by Gop ⋉γ B = [(JC0(G)J ⊗ 1)γ(B)] α(x) = Ad(cid:0)(1 ⊗ J J ⊗ 1)(W ∗ ⊗ 1)(1 ⊗ JJ ⊗ 1)(cid:1)(1 ⊗ x). Generalizing the isomorphism C0(G)Ω ∼= C ∗ Proposition 3.5. We have ( Gop ⋉γ B)Ω ∼= Gop ⋉γ,Ω B. More precisely, r ( G; Ω) we then have the following result. Ad(cid:0)(1 ⊗ J J ⊗ 1)( W Ω∗)∗ 21(1 ⊗ JJ ⊗ 1)(cid:1) maps the Ω-deformation of Gop ⋉γ B onto 1 ⊗ ( Gop ⋉γ,Ω B). Proof. The conjugation by JJ defines a left G-equivariant isomorphism JC0(G)J ∼= C0(G), where on JC0(G)J we consider the action given by Therefore by (3.2) the Ω-deformation of JC0(G)J is equal to x 7→ Ad(cid:0)(1 ⊗ J J)W ∗(1 ⊗ JJ)(cid:1)(1 ⊗ x). From this, by definition of the Ω-deformation, we conclude that ( Gop ⋉γ B)Ω is generated by (1 ⊗ J J)β(C ∗ r ( G; Ω))(1 ⊗ JJ). (1 ⊗ J J ⊗ 1)β(C ∗ r ( G; Ω)) ⊗ 1)(1 ⊗ JJ ⊗ 1)(1 ⊗ γ(B)). Now recall that β(x) = ( W Ω∗)21(1 ⊗ x)( W Ω∗)∗ 21 by (2.3). Observing also that the unitary (1 ⊗ J J ⊗ 1)( W Ω∗)∗ 21(1 ⊗ JJ ⊗ 1) commutes with 1⊗γ(B), we conclude that the conjugation by this unitary maps ( Gop ⋉γ B)Ω onto the r ( G; Ω) JJ ⊗ 1)(1⊗ γ(B)). But this is exactly 1⊗ ( Gop ⋉γ,Ω B). (cid:3) C∗-algebra generated by (1⊗ J JC ∗ Turning to more general actions, recall that for regular quantum groups any action is stably exterior equivalent to a dual action. Therefore it is natural to expect that up to stabilization Ω- deformations can be expressed in terms of twisted crossed products, at least under some regularity assumptions. In order to formulate the result recall that in Sections 2.3 and 2.4 we already used the unitaries Using formula (2.5) for WΩ we can also write X = JJ ∈ L∞( G) and Y = V ∗ 21Ω∗ 21 = (J ⊗ J) W Ω∗( J ⊗ J). We also define a map ηΩ : α(A) → M ( JC ∗ Y = (1 ⊗ J J)( WΩΩ)∗(1 ⊗ J J). r ( G; Ω) J ⊗ C ∗ r ( G; Ω) ⊗ A) by ηΩ(α(a)) = ( WΩΩ)21(1 ⊗ α(a))( WΩΩ)∗ 21, so that (Tν ⊗ ι)α(a) = (ν ⊗ ι ⊗ ι)ηΩ(α(a)). Theorem 3.6. Assume Ω is a regular cocycle on a locally compact quantum group G. Then for any C∗-algebra A equipped with a continuous left action α of G we have Explicitly, the map Gop ⋉ α,Ω G ⋉α A ∼= K ⊗ AΩ. Ad(cid:0)( JJ ⊗ 1 ⊗ 1)Y ∗ 21(J J ⊗ 1 ⊗ 1)(cid:1) = Ad(cid:0)( JX ∗ J ⊗ 1 ⊗ 1)( WΩΩ)21( J X J ⊗ 1 ⊗ 1)(cid:1) defines such an isomorphism. This map is trivial on J J C ∗ ∆op(C0( G)) ⊗ 1 into x ⊗ 1 ⊗ 1 and 1 ⊗ α(a) ∈ 1 ⊗ α(A) into Ad( J X ∗ J ⊗ 1 ⊗ 1)ηΩ(α(a)). r ( G; Ω) JJ ⊗1⊗1 and it maps ∆op(x)⊗1 ∈ (3.3) DEFORMATION OF C∗-ALGEBRAS 15 We will simultaneously prove the following. Theorem 3.7. If Ω is regular, then AΩ = [(Tν ⊗ ι)α(A) ν ∈ K ∗]. Proof of Theorems 3.6 and 3.7. From the equality Y = (J ⊗ J) W Ω∗( J ⊗ J) it is clear that Y com- r ( G; Ω) J J ⊗ mutes with 1⊗C ∗ 1 ⊗ 1. Clearly, it maps 1 ⊗ α(a) ∈ 1 ⊗ α(A) into Ad( JX ∗ J ⊗ 1 ⊗ 1)ηΩ(α(a)). Finally, for x ∈ C0( G) this homomorphism maps ∆op(x) ⊗ 1 into r ( G; Ω). Hence the homomorphism in the formulation is trivial on J J C ∗ ( WΩΩ)21( ∆op(x) ⊗ 1)( WΩΩ∗)21 = ( WΩΩ ∆(x)( WΩΩ)∗)21 = ( WΩ ∆Ω(x) W ∗ Ω)21 = x ⊗ 1 ⊗ 1. r ( G; Ω) J JC0( G)] = K by regularity of Ω, it follows that this homomorphism maps the Since [J J C ∗ C∗-algebra Gop ⋉ α,Ω G ⋉α A onto In particular, the last space is a C∗-algebra, so it coincides with [(K ⊗ 1 ⊗ 1) Ad( J X ∗ J ⊗ 1 ⊗ 1)ηΩ(α(A))]. [(K ⊗ 1 ⊗ 1) Ad( JX ∗ J ⊗ 1 ⊗ 1)ηΩ(α(A))(K ⊗ 1 ⊗ 1)] = K ⊗ [(ν ⊗ ι ⊗ ι)ηΩ(α(A)) ν ∈ K ∗]. This shows that [(ν ⊗ ι ⊗ ι)ηΩ(α(A)) ν ∈ K ∗] is a C∗-algebra and finishes the proof of both theorems. (cid:3) Note that while the regularity of Ω is a necessary condition for the conclusion of Theorem 3.6 to be true, it is not clear whether this is the case for Theorem 3.7. For example, by the proof of Proposition 3.5, for dual actions Theorem 3.7 remains true for any Ω. 3.4. Deformed action. Recall from Section 2.2 that we have a continuous left action βΩ of GΩ r ( G; Ω), which commutes with the right action β of G. This suggests that the action βΩ ⊗ ι on on C ∗ r ( G; Ω) ⊗ A defines a continuous action on AΩ. In other words, we want to define a left αΩ of GΩ C ∗ on AΩ by (3.4) We can prove that this is indeed a continuous action of GΩ under an additional regularity assumption. Theorem 3.8. Assume Ω is a measurable unitary 2-cocycle on a locally compact quantum group G such that the deformed quantum group GΩ is regular. Then for any C∗-algebra A equipped with a continuous left action α of G, the formula (3.4) defines a continuous left action of GΩ on AΩ. Ω ⊗ 1)(1 ⊗ x)(WΩ ⊗ 1) for x ∈ AΩ ⊂ M (C ∗ r ( G; Ω) ⊗ A). αΩ(x) = (W ∗ Proof. This follows from the proof of [35, Theorem 6.7]. We include a complete argument for 23 = ( WΩ)13( WΩ)12, for a ∈ A we the reader's convenience. Using the identity (VΩ)23( WΩ)12(VΩ)∗ compute: (ι ⊗ αΩ)ηΩ(α(a)) = ( WΩ)32( WΩΩ)31α(a)34( WΩΩ)∗ 12Ω31α(a)34Ω∗ = (VΩ)12( WΩ)31(VΩ)∗ 31( WΩ)∗ 32 31(VΩ)12( WΩ)∗ 31(VΩ)∗ 12, where ηΩ is defined by (3.3). Since (VΩ)∗ 12 and Ω31 commute, we thus get (ι ⊗ αΩ)ηΩ(α(a)) = (VΩ)12ηΩ(α(a))134(VΩ)∗ 12. Multiplying this identity on the left by 1 ⊗ C0(GΩ) ⊗ 1 ⊗ 1, applying the slice maps to the first leg and using that VΩ ∈ M (K ⊗ C0(GΩ)) and [(1 ⊗ C0(GΩ))V ∗ Ω(K ⊗ 1)] = K ⊗ C0(GΩ) by regularity of GΩ, we see that [(C0(GΩ) ⊗ 1 ⊗ 1)αΩ(cid:0)(Tν ⊗ ι)α(A)(cid:1) ν ∈ K ∗] = C0(GΩ) ⊗ [(Tν ⊗ ι)α(A) ν ∈ K ∗]. This implies that αΩ(AΩ) ⊂ M (C0(GΩ) ⊗ AΩ) and that the cancellation property holds. Finally, it is clear that (ι ⊗ αΩ)αΩ = (∆Ω ⊗ ι)αΩ. (cid:3) 16 S. NESHVEYEV AND L. TUSET We now want to give a different picture of AΩ and αΩ based on crossed products. As was mentioned in the introduction, it is inspired by work of Kasprzak [17], and it was our original motivation for the definition of AΩ. Consider the crossed product G ⋉α A and the dual action α of Gop on G ⋉α A. We can then try to define a new deformed action αΩ of ( GΩ)op on G ⋉α A by 21. αΩ(x) = Ω21 α(x)Ω∗ If αΩ is well-defined, GΩ is regular and WΩ⊗1 ∈ M (C0(GΩ)⊗(G⋉α A)), then by a result of Vaes [35, Theorem 6.7] discussed in Section 1.3, the action αΩ is dual to an action of GΩ on a C∗-subalgebra B ⊂ M (G ⋉α A), which can be recovered using the homomorphism ηΩ : G ⋉α A → M (K ⊗ K ⊗ A), ηΩ(x) = (WΩ)∗ (3.5) Note that since α(α(a)) = 1 ⊗ α(a), the definition of ηΩ is consistent with (3.3). Note also that for x = y ⊗ 1 ∈ C0( G) ⊗ 1 ∈ M (G ⋉α A) we have ηΩ(x) = ( WΩ)21( ∆op 12 αΩ(x)(WΩ)12. Ω (y) ⊗ 1)( WΩ)∗ 21 = y ⊗ 1 ⊗ 1. It follows that B = [{(ω ⊗ ι ⊗ ι)ηΩ(G ⋉α A) ω ∈ K ∗}] = [{(Tν ⊗ ι)α(A) ν ∈ K ∗}] = AΩ. Summarizing the above discussion, we have the following result. Theorem 3.9. Under the assumptions of Theorem 3.8 suppose that the formula αΩ(x) = Ω21 α(x)Ω∗ 21 defines a continuous left action of ( GΩ)op on G ⋉α A ⊂ M (K ⊗ A) and that WΩ ⊗ 1 ∈ M (C0(GΩ) ⊗ (G ⋉α A)) ⊂ M (C0(GΩ) ⊗ K ⊗ A). Then G ⋉α A = [(C0( GΩ) ⊗ 1)AΩ] and the map ηΩ defines an isomorphism G ⋉α A ∼= GΩ ⋉αΩ AΩ. Under this isomorphism the deformed dual action αΩ on G ⋉α A becomes the action dual to αΩ. Again, it is not clear to us what the optimal assumptions for the above two theorems are. Note, however, that for nonregular quantum groups it is not even obvious what the correct definition of a continuous action should be, see the discussion in [2]. Even if GΩ is regular, it is doubtful that the map ηΩ defines an isomorphism G ⋉α A ∼= GΩ ⋉αΩ AΩ for any A, since this would imply that the deformed dual action αΩ is well-defined on G ⋉α A, which seems to be overly optimistic already when G is a group and Ω is a measurable, but not continuous, cocycle on G. 3.5. Deformation in stages. If Ω is a cocycle on G and Ω1 is a cocycle on GΩ, then it is easy to check that Ω1Ω is a cocycle on G. Therefore if the deformed action αΩ of GΩ on AΩ is well-defined, then we can compare the Ω1-deformation of AΩ with the Ω1Ω-deformation of A. Theorem 3.10. Assume G is a locally compact quantum group, Ω is a measurable unitary 2-cocycle on G, Ω1 is a measurable unitary 2-cocycle on GΩ, and A is a C∗-algebra equipped with a continuous left action α of G. Suppose the following conditions are satisfied: (i) AΩ = [(Tν ⊗ ι)α(A) ν ∈ K ∗]; (ii) the deformed action αΩ of GΩ is well-defined on AΩ. Then the map x 7→ ( WΩΩ∗ 21 defines an isomorphism AΩ1Ω ∼= (AΩ)Ω1 . Fur- 1)21(1 ⊗ x)( WΩΩ∗ 1)∗ thermore, if one of the deformed actions αΩ1Ω and (αΩ)Ω1 is well-defined, then the other is also well-defined and the isomorphism AΩ1Ω ∼= (AΩ)Ω1 is GΩ1Ω-equivariant. Proof. For the proof we need the following identity: ( WΩΩ)23( WΩ1ΩΩ1Ω)12( WΩΩ)∗ 23 = ( WΩ1ΩΩ1)12( WΩΩ)13. (3.6) DEFORMATION OF C∗-ALGEBRAS 17 In order to show this, similarly to the proof of (3.1) we start with the identity ( WΩΩ)23( WΩΩ)12( WΩΩ)∗ 23 = ( WΩ)12( WΩΩ)13. Next, substitute ( WΩ)12 on both sides for its expression involving ( WΩ1Ω)12 obtained from (2.5): where J1 is the modular involution defined by the dual weight on W ∗( GΩ; Ω1) as explained in Section 2.2. We get WΩ1Ω = ( J1 ⊗ J)Ω1(JΩ ⊗ J) WΩΩ∗ 1, 12( WΩ1ΩΩ1Ω)12( WΩΩ)∗ 23 ( WΩΩ)23(cid:0)( J1 ⊗ J)Ω1(JΩ ⊗ J)(cid:1)∗ Since ( WΩΩ)23 and (cid:0)( J1 ⊗ J )Ω1(JΩ ⊗ J)(cid:1)∗ For a ∈ A we now start computing: = (cid:0)( J1 ⊗ J)Ω1(JΩ ⊗ J)(cid:1)∗ 12 commute, this is exactly (3.6). 12(( WΩ1ΩΩ1)12( WΩΩ)13. 43( WΩ1ΩΩ1)∗ 32. By identity (3.1), applied to the quantum group GΩ and the dual cocycle Ω1, we have (ι ⊗ ηΩ1αΩ)ηΩ(α(a)) = ( WΩ1ΩΩ1)32( WΩ)43ηΩ(α(a))145( WΩ)∗ ( WΩ1ΩΩ1)32( WΩ)43( WΩ1ΩΩ1)∗ 32 = ( WΩΩ∗ 1)43( WΩ1ΩΩ1)42. Therefore (ι ⊗ ηΩ1αΩ)ηΩ(α(a)) = ( WΩΩ∗ = ( WΩΩ∗ By (3.6) the last expression equals 1)43( WΩ1ΩΩ1)42ηΩ(α(a))145( WΩ1ΩΩ1)∗ 1)43( WΩ1ΩΩ1)42( WΩΩ)41α(a)45( WΩΩ)∗ 1)∗ 43 42( WΩΩ∗ 41( WΩ1ΩΩ1)∗ 42( WΩΩ∗ 1)∗ 43. ( WΩΩ∗ 1)43( WΩΩ)21( WΩ1ΩΩ1Ω)42( WΩΩ)∗ 21α(a)45( WΩΩ)21( WΩ1ΩΩ1Ω)∗ 42( WΩΩ)∗ 21( WΩΩ∗ 1)∗ 43 = ( WΩΩ∗ = ( WΩΩ∗ 1)43( WΩΩ)21( WΩ1ΩΩ1Ω)42α(a)45( WΩ1ΩΩ1Ω)∗ 42( WΩΩ)∗ 1)43( WΩΩ)21ηΩ1Ω(α(a))245( WΩΩ)∗ 1)∗ 43. 21( WΩΩ∗ 21( WΩΩ∗ 1)∗ 43 Thus 43( WΩΩ)∗ 21. Applying the slice maps to the first two legs we get the first statement of the theorem. (ι ⊗ ηΩ1αΩ)ηΩ(α(a)) = ( WΩΩ)21( WΩΩ∗ 1)43ηΩ1Ω(α(a))245( WΩΩ∗ 1)∗ In order to show that the isomorphism AΩ1Ω ∼= (AΩ)Ω1 is GΩ1Ω-equivariant we need the identity ( WΩ1Ω)23( WΩΩ∗ 1)12 = ( WΩΩ∗ 1)12( WΩ1Ω)13( WΩ1Ω)23. Since WΩΩ∗ the isomorphism in the formulation of the theorem by θ we compute: 1 = (cid:0)( J1⊗ J)Ω1(JΩ⊗ J)(cid:1)∗ WΩ1Ω, this is simply the pentagon relation for WΩ1Ω. Denoting (αΩ)Ω1(θ(x)) = ( WΩ1Ω)21(1 ⊗ θ(x))( WΩ1Ω)∗ 21 32( WΩ1Ω)∗ 21 1)32( WΩ1Ω)31( WΩ1Ω)21(1 ⊗ 1 ⊗ x)( WΩ1Ω)∗ 21( WΩ1Ω)∗ 1)32αΩ1Ω(x)13( WΩΩ∗ 1)∗ 32 = (ι ⊗ θ)αΩ1Ω(x). This proves the second statement of the theorem. = ( WΩ1Ω)21( WΩΩ∗ = ( WΩΩ∗ = ( WΩΩ∗ 1)32(1 ⊗ 1 ⊗ x)( WΩΩ∗ 1)∗ 31( WΩΩ∗ 1)∗ 32 (cid:3) Example 3.11. (i) It is straightforward to check that the deformation A1 of A with respect to the trivial cocycle 1 is α(A). Therefore given a cocycle Ω such that the assumptions (i) and (ii) in the above theorem 21 defines an isomorphism α(A) ∼= are satisfied, it follows that the map x 7→ ( WΩΩ)21(1 ⊗ x)( WΩΩ)∗ (AΩ)Ω∗. In other words, the map ηΩα is an isomorphism A ∼= (AΩ)Ω∗. (ii) Assume A = Gop ⋉γ B and α = γ. Then by Proposition 3.5 we have AΩ ∼= Gop ⋉γ,Ω B. By Proposition 2.7 we have a dual action on Gop ⋉γ,Ω B. It is easy to check that this is exactly the 18 S. NESHVEYEV AND L. TUSET deformed action αΩ. Furthermore, as we have already remarked, the proof of Proposition 3.5 shows that AΩ = [(Tν ⊗ ι)α(A) ν ∈ K ∗]. Therefore for dual actions conditions (i) and (ii) in the above theorem are always satisfied. For any cocycle Ω1 on GΩ we thus get ( Gop ⋉γ,Ω B)Ω1 ∼= ( Gop ⋉γ B)Ω1Ω ∼= Gop ⋉γ,Ω1Ω B. In particular, for the C∗-algebra C ∗ C ∗ (iii)As a particular case of either of the previous two examples we get an isomorphism r ( G; Ω) equipped with the action βΩ of GΩ we get C ∗ r ( G; Ω1Ω). r ( G; Ω)Ω1 ∼= r ( GΩ; Ω∗) is defined using the action βΩ∗ (x) = W ∗(1 ⊗ x)W of G on where the deformation of C ∗ r ( GΩ; Ω∗). Explicitly, by the first example the isomorphism C0(GΩ) ∼= (C0(GΩ)Ω∗)Ω is given by C ∗ C0(GΩ) ∼= C ∗ r ( GΩ; Ω∗)Ω, x 7→ ηΩ∗ ∆Ω(x) = ( W Ω∗)21(VΩ)23(1 ⊗ x ⊗ 1)(VΩ)∗ 23( W Ω∗)∗ 21, Ω we conclude that the isomorphism C0(GΩ) ∼= so using that C0(GΩ)Ω∗ = VΩ(C ∗ r ( GΩ; Ω∗)Ω is given by C ∗ r ( GΩ; Ω∗) ⊗ 1)V ∗ x 7→ ( W Ω∗)21(1 ⊗ x)( W Ω∗)∗ 21. This is also not difficult to check directly from the definition of C ∗ r ( GΩ; Ω∗)Ω. Note that C ∗ r ( GΩ; Ω∗) = JC ∗ r ( G; Ω) J is nothing other than the deformation of C0(G) with respect to the right action ∆ of G on C0(G). More precisely, a right action of G can be considered as a left action of Gop, which is the quantum group (L∞(G), ∆op). The element ( J ⊗ J)Ω( J ⊗ J) is a dual cocycle on Gop. It is not difficult to check then that the deformation of C0(G) with respect to the left action of Gop and the cocycle ( J ⊗ J)Ω( J ⊗ J) is isomorphic to C ∗ r ( GΩ; Ω∗), and under r ( GΩ; Ω∗) corresponds to the action arising from the left this isomorphism the action βΩ∗ of G on C ∗ action ∆ of G on C0(G). Therefore the isomorphism C0(GΩ) ∼= C ∗ r ( GΩ; Ω∗)Ω is consistent with what we should expect from the case of finite quantum groups, when C0(GΩ) ∼= C0(G) as coalgebras, while the new algebra structure is obtained by duality from ∆Ω = Ω ∆(·)Ω−1, which implies that it is obtained by deforming the original product structure on C0(G) twice, with respect to the left and ♦ right actions of G on C0(G). We finish our general discussion of Ω-deformations with the observation that up to isomorphism the C∗-algebra AΩ depends only on the cohomology class of Ω. Recall that given a 2-cocycle Ω on G and a unitary u ∈ L∞( G), the element Ωu = (u ⊗ u)Ω ∆(u)∗ is again a cocycle on G. The cocycles Ω and Ωu are called cohomologous. The set of cohomology classes of unitary 2-cocycles on G is denoted by H 2( G; T). In general it is just a set. Proposition 3.12. Assume G is a locally compact quantum group, Ω is a measurable unitary 2- cocycle on G and u is a unitary in L∞( G). Then for any C∗-algebra A equipped with a continuous left action of G, the map Ad(u ⊗ 1) defines an isomorphism AΩ ∼= AΩu. u = W ∆(u)Ω∗(u∗ ⊗ u∗) = (1 ⊗ u) W Ω∗(u∗ ⊗ u∗), the map Ad u defines a right G- Proof. Since W Ω∗ equivariant isomorphism W ∗( G; Ω) ∼= W ∗( G; Ωu). This isomorphism maps (ω(· u∗) ⊗ ι)( W Ω∗) into u). Consider the GNS-representations Λ : N ϕ → L2(G) and Λu : N ϕu → L2(G) defined (ω ⊗ ι)( W Ω∗ by the dual weights ϕ on W ∗( G; Ω) and ϕu on W ∗( G; Ωu), as described in Section 2.2. Then the isomorphism W ∗( G; Ω) ∼= W ∗( G; Ωu) defines a unitary u on L2(G) such that uΛ(x) = Λu(uxu∗) for x ∈ N ϕ, so for suitable ω ∈ K ∗. Since Λ((ω(· u∗)⊗ ι)( W Ω∗)) = Λ((ω(· u∗)⊗ ι)( W )) and a similar formula holds for Λu, we in other words have uΛ((ω(· u∗) ⊗ ι)( W Ω∗)) = Λu((ω ⊗ ι)( W Ω∗ u)) uΛ((ω(· u∗) ⊗ ι)( W )) = Λ(ω ⊗ ι)( W )). DEFORMATION OF C∗-ALGEBRAS 19 But Λ((ω(· u∗) ⊗ ι)( W )) = u∗Λ((ω ⊗ ι)( W )), as can be easily checked using that (Λ((ν ⊗ ι)( W )), Λ(y)) = ν(y∗). It follows that u = u. Hence the modular involution Ju defined by the weight ϕu on W ∗( G; Ωu) is equal to u Ju∗. Therefore ( Ju ⊗ J)Ωu(J ⊗ J) = (u ⊗ Ju J)(cid:0)( J ⊗ J)Ω(J ⊗ J)(cid:1)(J ⊗ J) ∆(u)∗(J ⊗ J). Next, we have (J ⊗ J) ∆(u)∗(J ⊗ J) W = (J ⊗ J) ∆(u)∗ W ∗(J ⊗ J) = (J ⊗ J) W ∗(1 ⊗ u∗)(J ⊗ J) = W (1 ⊗ Ju∗ J). Hence, by (2.5), WΩuΩu = ( Ju ⊗ J)Ωu(J ⊗ J) W = (u ⊗ Ju J) WΩΩ(1 ⊗ Ju∗ J ). Recalling the definition of ηΩ we get This gives the result. ηΩu = Ad( Ju J ⊗ u ⊗ 1)ηΩ. (cid:3) In this section we assume that G is a genuine locally compact group. 4. Cocycles on group duals 4.1. Dual cocycles and deformations of the Fourier algebra. Denote by λg, resp. ρg, the operators of the left, resp. right, regular representation of G. Then ( W ξ)(s, t) = ξ(ts, t) = (λ−1 t ξ(·, t))(s) and (V ξ)(s, t) = (ρtξ(·, t))(s) for ξ ∈ L2(G × G). The predual of L∞( G) = W ∗(G) can be identified with the Fourier algebra A(G) ⊂ C0(G), so an element ω ∈ W ∗(G)∗ is identified with the function f (g) = ω(λg) on G. Note that under this identification we have where f (g) = f (g−1). (f ⊗ ι)( W ) = f for f ∈ A(G), Assume now that Ω is a measurable unitary 2-cocycle on G. Then we can define a new product ⋆Ω on A(G) by The associativity of this product is equivalent to the cocycle identity for Ω. Identity (2.1) shows that the formula f1 ⋆Ω f2 = (f1 ⊗ f2)( ∆(·)Ω∗). πΩ(f ) = (f ⊗ ι)( W Ω∗) r ( G; Ω) defines a representation of (A(G), ⋆Ω) on L2(G), and then by definition the C∗-algebra C ∗ r ( G; Ω) = πΩ(A(G)). is generated by πΩ(A(G)). Recall that by Theorem 2.1 we in fact have C ∗ Nevertheless we do not claim that (A(G), ⋆Ω) is itself a ∗-algebra, although this is often the case. In r ( G; Ω). Since by (2.4), Section 5 we will give an example where πΩ(A(G)) is not a ∗-subalgebra of C ∗ we have the representation πΩ is given by Λ((f ⊗ ι)( W Ω∗)) = Λ((f ⊗ ι)( W )) = f , πΩ(f1) f2 = (f1 ⋆Ω f2) for f1 ∈ A(G) and f2 ∈ A(G) ∩ Cc(G). f (g)dg = RG f (g)∆G(g)−1dg, the representation πΩ is simply the left regular In other words, since RG representation of (A(G), ⋆Ω) on itself, with A(G), or more precisely A(G) ∩ Cc(G), completed to a Hilbert space using the scalar product defined by the right Haar measure ∆G(g)−1dg. 20 S. NESHVEYEV AND L. TUSET The left translations of G on itself define automorphisms of (A(G), ⋆Ω). On the level of C ∗ r ( G; Ω) the action by left translations is exactly the action β introduced earlier, so βg(πΩ(f )) = πΩ(f (g−1·)) for f ∈ A(G). This is the reason for the appearance of the right Haar measure, since the average of a function on G with respect to the action by left translations is the integral with respect to a right Haar measure. Conversely, assume we have a product ⋆ on A(G) that is invariant under left translations. Assume there exists an element Ω ∈ W ∗(G) ¯⊗W ∗(G) such that (f1 ⊗ f2)(Ω∗) = (f1 ⋆ f2)(e) for all f1, f2 ∈ A(G), which happens exactly when the map f1 ⊗ f2 7→ (f1 ⋆ f2)(e) extends to a bounded linear functional on the projective tensor product A(G) ⊗A(G). Then f1 ⋆ f2 = (f1 ⊗ f2)( ∆(·)Ω∗). For finite groups this was observed by Movshev [25]. As follows from results in [25], if G is finite and (A(G), ⋆) is semisimple, then Ω is invertible and cohomological to a unitary cocycle, that is, there exists an invertible element a ∈ W ∗(G) such that (a ⊗ a)Ω ∆(a)−1 is unitary. For a related discussion see [22]. 4.2. Ω-Deformations and generalized fixed point algebras. As we know, by [36, Lemma 1.12] the action β of G on W ∗( G; Ω) is integrable. We will now show a stronger property: the action r ( G; Ω) is integrable. Since the integrability property that we will establish appears under of G on C ∗ several different names in the literature, let us say precisely what we mean by this. Assume γ is a continuous action of G on a C∗-algebra B. An element b ∈ B+ is called γ-integrable if there exists an element ψγ(b) ∈ M (B) such that for every state ω on B the function g 7→ ω(γg(b)) is integrable and ZG ω(γg(b))dg = ω(ψγ(b)). Clearly, the element ψγ(b) is uniquely determined by b, and ψγ(b) ∈ M (B)γ. There are several other equivalent definitions of integrable elements, see [31] (note that in [31] γ-integrable elements are called γ-proper). For example, by [31, Proposition 4.4] a positive element b is γ-integrable if and only if the functions g 7→ γg(b)c and g 7→ cγg(b) are unconditionally integrable for all c ∈ B, meaning that their integrals over compact subsets of G form Cauchy nets. γ of γ-integrable positive elements is a hereditary cone in B+, see e.g. [31, Lemma 2.7]. Hence the linear span of P + γ is dense in B+, or equivalently, [P + Proposition 4.1. For any measurable unitary 2-cocycle Ω on G the action β of G on C ∗ integrable. γ is a ∗-algebra. We say that γ is integrable if P + r ( G; Ω) is The set P + γ ] = B. Before we turn to the proof, let us discuss the difference between this statement and the integra- bility of the action of G on W ∗( G; Ω). By [36, Theorem 1.11] the action on W ∗( G; Ω) is ergodic. In r ( G; Ω))β = C1. For x ∈ W ∗( G; Ω)+ in the domain of definition dom ϕ of the dual particular, M (C ∗ weight ϕ we have ZG ω(βg(x))dg = ϕ(x) (4.1) for all normal states ω on W ∗( G; Ω). It follows that if x ∈ C ∗ and ψβ(x) = ϕ(x)1. The difference between the cones P + elements of P + of C ∗ of C ∗ r ( G; Ω)+ is β-integrable, then x ∈ dom ϕ r ( G; Ω) ∩ dom ϕ is that for the β and C ∗ r ( G; Ω), while for the elements β identity (4.1) should be satisfied for all states ω on C ∗ r ( G; Ω)∩ dom ϕ we need only to consider normal states on W ∗( G; Ω). Note also that the density r ( G; Ω) ∩ dom ϕ in C ∗ r ( G; Ω)+ follows already from the proof of [36, Lemma 1.12]. DEFORMATION OF C∗-ALGEBRAS 21 Proof of Proposition 4.1. Let ν be a normal state on W ∗( GΩ; Ω∗). As we already observed in Sec- r ( G; Ω)) is strictly continuous on bounded tion 3.1, the quantization map Tν : M (C0(G)) → M (C ∗ r ( G; Ω), the positive linear functional ωTν on C0(G) is sets. It follows that for any state ω on C ∗ again a state. Since Tν : L∞(G) → W ∗( G; Ω) is a G-equivariant normal u.c.p. map, if f ∈ dom ϕ = L∞(G)+ ∩ L1(G), then Tν(f ) ∈ dom ϕ and ϕ(Tν(f )) = ϕ(f ) = ZG f (g)dg. For the action of G on itself by right translations there is no distinction between integrability of an element f ∈ C0(G)+ in the von Neumann algebraic and the C∗-algebraic sense: both conditions are equivalent to f ∈ L1(G). For f ∈ C0(G)+ ∩ L1(G) and any state ω on C ∗ r ( G; Ω) we then get ZG ω(βg(Tν(f )))dg = ZG ωTν(f (· g))dg = (ωTν )(1)ϕ(f ) = ϕ(Tν (f )), so Tν (f ) is β-integrable. Hence [P + β ] = C ∗ r ( G; Ω). By Proposition 3.1 the span of the spaces Tν(C0(G)), ν ∈ W ∗( GΩ; Ω∗)∗, is dense in C ∗ r ( G; Ω). (cid:3) Returning to a general action γ of G on B, it is easy to see that if b ∈ B+ is γ-integrable and x ∈ M (B)γ, then x∗bx is again γ-integrable and ψγ(x∗bx) = x∗ψγ(b)x. This implies that the span of P + γ is an M (B)γ-bimodule, which in turn implies that the span of ψγ(P + γ ) is a ∗-ideal in M (B)γ. The C∗-algebra [ψγ(P + γ )] ⊂ M (B)γ can be considered as a generalized fixed point algebra for the action γ on B. In general, however, it is too big to have good properties and it is not clear what the correct definition of a generalized fixed point algebra should be, see the discussion in [31]. Consider now a continuous action α of G on a C∗-algebra A. By the observation immediately r ( G; Ω) ⊗ A)β⊗α. We can now prove a slightly more precise after Definition 3.3, we have AΩ ⊂ M (C ∗ result. Proposition 4.2. For any measurable unitary 2-cocycle Ω on G and any C∗-algebra A equipped r ( G; Ω) ⊗ A is integrable and with a continuous action α of G, the diagonal action β ⊗ α of G on C ∗ r ( G; Ω) ⊗ A)β⊗α. the C∗-algebra AΩ is contained in [ψβ⊗α(P + Proof. The first statement follows immediately from the integrability of β. β⊗α)] ⊂ M (C ∗ In order to prove the second statement denote by ρ the action of G on C0(G) by right translations. Then it is easy to check that for any f ∈ C0(G)+ ∩ L1(G) and a ∈ A+ the element f ⊗ a is ρ ⊗ α- integrable and where af = RG f (g)αg(a)dg; note that if we identify M (C0(G) ⊗ A) with Cb(G; M (A)) then by definition α(af )(g) = αg−1(af ). ψρ⊗α(f ⊗ a) = α(af ), From this, arguing as in the proof of the previous proposition, we conclude that the element Tν(f ) ⊗ a is β ⊗ α-integrable and ψβ⊗α(Tν (f ) ⊗ a) = (Tν ⊗ ι)α(af ). Since [ψβ⊗α(P + β⊗α)] is a C∗-algebra, it follows that AΩ is contained in [ψβ⊗α(P + β⊗α)]. (cid:3) In the case when G is a compact group all the analytical difficulties disappear and we get the following. Proposition 4.3. If G is compact, then AΩ = (C ∗ r ( G; Ω) ⊗ A)β⊗α. 22 S. NESHVEYEV AND L. TUSET Proof. When G is compact, every positive element of C ∗ r ( G; Ω) ⊗ A → M (C ∗ ψβ⊗α extends by linearity to a bounded map C ∗ (C ∗ is mapped by ψβ⊗α onto a generating subspace of AΩ. Hence AΩ = (C ∗ r ( G; Ω) ⊗ A is integrable and the map r ( G; Ω) ⊗ A)β⊗α with image r ( G; Ω)⊗A r ( G; Ω)⊗A)β⊗α. As follows from the previous proof, a dense subspace of elements of C ∗ r ( G; Ω) ⊗ A)β⊗α. For an overview of what is known about cocycles on duals of compact groups see [26]. (cid:3) 4.3. Regularity of cocycles and proper actions. A stronger notion than integrability was in- troduced by Rieffel in [29]. Namely, an action γ of G on a C∗-algebra B is called proper, if there exists a dense γ-invariant ∗-subalgebra B0 ⊂ B such that for all b, c ∈ B the functions g 7→ bγg(c) and g 7→ ∆G(g)1/2bγg(c) are norm-integrable∗ and there exists an element x ∈ M (B0)γ ⊂ M (B)γ such that RG aγg(bc)dg = ax for all a ∈ B0. By [31, Proposition 4.6] a proper action is integrable; in fact, if b ∈ B0 then b∗b is γ-integrable. The integrable functions g 7→ ∆G(g)1/2bγg(c) define elements of the reduced crossed product G ⋉γ B. The closure I of the space spanned by such elements is a ∗-ideal in G ⋉γ B. As shown 0)] ⊂ M (B)γ. The action γ in [29], this ideal is strongly Morita equivalent to the C∗-subalgebra [ψγ(B2 is called saturated if B0 can be chosen such that I = G ⋉γ B. The relevance of these notions for us is explained by the following. Proposition 4.4. For a measurable unitary 2-cocycle Ω on G, assume the action β of G on C ∗ is proper and saturated. Then Ω is regular. r ( G; Ω) r ( G; Ω))β = C1, by the above discussion the assumptions of the proposition imply Proof. Since M (C ∗ r ( G; Ω)⋊β G is isomorphic to the algebra that C ∗ of compact operators on some Hilbert space. But this is one of the equivalent characterizations of regularity. (cid:3) r ( G)⋊β G is strongly Morita equivalent to C, that is, C ∗ We expect the saturation property to hold more or less automatically. For example, it holds when G is compact, in which case, however, we do need the above proposition to show regularity. We finish this section with a simple result on continuity of dual cocycles. Proposition 4.5. Assume Ω is a measurable unitary 2-cocycle on G such that both Ω and Ω∗ map Cc(G × G) ⊂ L2(G × G) into L1(G × G) ∩ L2(G × G). Then Ω is continuous, that is, Ω ∈ M (C ∗ Proof. A function f ∈ L1(G × G) ∩ L2(G × G) defines both a vector in L2(G × G) and an element (λ ⊗ λ)(f ) ∈ C ∗ r (G). We claim that if f ∈ Cc(G × G), then r (G) ⊗ C ∗ r (G) ⊗ C ∗ r (G)). Indeed, if ξ ∈ Cc(G × G), then, using that Ω commutes with the operator ζ 7→ ζ ∗ ξ on L2(G × G), we have Ω(λ ⊗ λ)(f ) = (λ ⊗ λ)(Ωf ). Since Cc(G × G) is dense in L2(G × G), this proves our claim. r (G)) ⊂ C ∗ all f ∈ Cc(G × G), which implies that Ω∗(C ∗ Ω(λ ⊗ λ)(f )ξ = Ω(f ∗ ξ) = (Ωf ) ∗ ξ = (λ ⊗ λ)(Ωf )ξ. r (G) ⊗ C ∗ It follows that Ω(C ∗ r (G)) ⊂ C ∗ r (G) ⊗ C ∗ r (G) ⊗ C ∗ r (G). r (G) ⊗ C ∗ r (G). Similarly, Ω∗(λ ⊗ λ)(f ) = (λ ⊗ λ)(Ω∗f ) for (cid:3) 5. Dual cocycles for a class of solvable Lie groups In this section we briefly consider dual cocycles recently constructed by Bieliavsky et al. [6, 5]. ∗We define the modular function so that RG f (gh)dg = ∆G(h)−1 RG f (g)dg, which is opposite to the conventions in [29]. DEFORMATION OF C∗-ALGEBRAS 23 5.1. Deformation of negatively curved Kahlerian Lie groups. As explained in [6], by re- sults of Pyatetskii-Shapiro, negatively curved Kahlerian Lie groups can be decomposed into iterated semidirect products of certain elementary groups, called elementary normal j-groups in [6]. To simplify matters we will consider only the latter groups. Thus, throughout the whole Section 5 we assume that G is a simply connected real Lie group of dimension 2d + 2 with a basis H, {Xj}2d j=1, E of the Lie algebra g satisfying the relations [H, E] = 2E, [H, Xj] = Xj, [E, Xj ] = 0, [Xi, Xj] = (δi+d,j − δi,j+d)E. The map R × R2d × R ∋ (a, v, t) 7→ exp(aH) exp  2d Xj=1 vjXj + tE  ∈ G is a diffeomorphism. In the coordinates (a, v, t) the group law on G takes the form (a, v, t)(a′, v′, t′) = (a + a′, e−a′ v + v′, e−2a′ t + t′ + e−a′ ω0(v, v′)), 1 2 where ω0(v, v′) = Pd i) is the standard symplectic form on R2d. From this formula it is clear that the usual Lebesgue measure on R2d+2 defines a left Haar measure on G. Then the modular function ∆G is given by† i+d − vi+dv′ i=1(viv′ ∆G(a, v, t) = e−(2d+2)a. For every θ ∈ R, θ 6= 0, Bieliavsky and Gayral [6, Section 4.1] construct a new product ⋆θ on a space Eθ(G) of smooth functions on G. The precise definition of Eθ(G) is not important to us. What we need to know is that Eθ(G) contains C ∞ c (G) and there exists a bijective linear map Tθ : S(R2d+2) → Eθ(G) that is compatible with complex conjugation and that extends to a unitary operator L2(R2d+2) → L2(G). The new product is then defined by (f2)), f1 ⋆θ f2 = Tθ(T −1 (f1) ⋆0 θ T −1 θ θ where ⋆0 θ denotes the standard Moyal product on S(R2d+2) defined using the symplectic form ωg((a, v, t), (a′, v′, t′)) = 2(at′ − ta′) + ω0(v, v′). One of the reasons to introduce the map Tθ is that the product ⋆θ becomes left G-invariant. Fur- thermore, it is possible to explicitly write down the distribution kernel of the product: (f1 ⋆θ f2)(g) = ZG×G Kθ(x, y)f1(gx)f2(gy)dx dy for f1, f2 ∈ C ∞ c (G), where Kθ(x, y) = 4 (πθ)2d+2 A(x, y) exp(cid:26) 2i θ S(x, y)(cid:27) and, for x = (a, v, t) and x′ = (a′, v′, t′), A(x, x′) = (cid:0) cosh(a) cosh(a′) cosh(a − a′)(cid:1)d(cid:0) cosh(2a) cosh(2a′) cosh(2a − 2a′)(cid:1)1/2, S(x, x′) = sinh(2a)t′ − sinh(2a′)t + cosh(a) cosh(a′)ω0(v, v′). algebra in Section 4.1 it is then natural to try to define a cocycle Ωθ on G by In view of our discussion of the relation between dual cocycles and deformations of the Fourier Ω∗ θ = ZG×G Kθ(x, y)λx ⊗ λy dx dy on C ∞ c (G × G) ⊂ L2(G × G). †Note again that our definition of the modular function is opposite to the one in [6]. 24 S. NESHVEYEV AND L. TUSET The question is whether this defines a unitary operator on L2(G× G). By [5] this is indeed the case. In fact, the proof is essentially contained already in [6]. Namely, as established in the proof of [6, Proposition 8.45], we have ZG×G Kθ(x, y)K−θ(gx, hy)∆G(x)∆G(y)dx dy = δe(g)δe(h), with the integral understood in the distribution sense. It is not difficult to check that this, together with Kθ(x, y) = K−θ(x, y), implies that Ωθ is an isometry. Similarly, we have ZG×G K−θ(x, y)Kθ(xg, yh)dx dy = δe(g)δe(h), which implies that Ω∗ θ is an isometry. Therefore, by [5], we get a family of unitary 2-cocycles Ωθ on G. The corresponding quantum groups GΩθ provide a nonformal deformation of the Poisson-Lie group G, with the Poisson structure defined by the nondegenerate 2-cocycle ωg on g. In the simplest case d = 0 the group G is the ax + b group. A quantization of the same Poisson-Lie structure on it has been defined by Baaj and Skandalis [33], see also [36, Section 5.3] and [34] (note also that there exists only one, up to isomorphism and rescaling, different Poisson-Lie structure; it has been quantized by Pusz, Woronowicz and Zakrzewski [28, 43]). It would be interesting to check whether for d = 0 the quantum groups GΩθ are isomorphic to the one defined by Baaj and Skandalis. 5.2. Involution on the twisted group algebra. Given the cocycle Ωθ, we can now consider the new product ⋆Ωθ on the Fourier algebra A(G) and the representation πΩθ of (A(G), ⋆Ωθ ) on L2(G) given by πΩθ (f1) f2 = (f1 ⋆Ωθ f2) for f1 ∈ A(G) and f2 ∈ A(G) ∩ Cc(G). At a first glance a bit surprisingly, the algebra πΩθ (A(G)) fails to be a ∗-algebra. Namely, consider the modular function ∆G as the unbounded operator of multiplication by ∆G on L2(G), so identify ∆G with the modular operator defined by the Haar weight on L∞( G) = W ∗(G). Consider also the dense subspace A∞(G) of A(G) spanned by the functions of the form ξ ∗ ζ with ξ, ζ ∈ C ∞ c (G). We then have the following. Lemma 5.1. For any f ∈ A∞(G) we have πΩθ (f )∗ = ∆−1 G πΩθ ( ¯f )∆G on Cc(G) ⊂ L2(G). Proof. Since for f ∈ A∞(G) we have we get (f ⊗ ι)( W (λx ⊗ λy)) = f (x−1·)λy, πΩθ (f ) = ZG×G Kθ(x, y) f (x−1·)λy dx dy on Cc(G). It follows that on Cc(G) we have πΩθ (f )∗ = ZG×G = ZG×G Kθ(x, y)λy−1 ¯f (x−1·)dx dy = ZG×G Kθ(yx, y) ¯f (x−1·)λy−1 dx dy = ZG×G Kθ(x, y) ¯f (x−1y ·)λy−1 dx dy Kθ(y−1x, y−1)∆G(y)−1 ¯f (x−1·)λy dx dy = ∆−1 G ZG×G Kθ(y−1x, y−1) ¯f (x−1·)λy dx dy ∆G. Therefore it suffices to check that Kθ(y−1x, y−1) = Kθ(x, y), or equivalently, A(y−1x, y−1) = A(x, y) and S(y−1x, y−1) = −S(x, y). Both identities are checked by a straightforward computation. (cid:3) DEFORMATION OF C∗-ALGEBRAS 25 We will now give a different proof of this lemma using known properties of the Moyal product. As we discussed in Section 4.1, the representation πΩθ can be thought of as the left regular representation of (A(G), ⋆Ωθ ) with respect to the scalar product defined by the right Haar measure. We can also try to use the left Haar measure. In general we see no reason to expect the corresponding representation to be well-defined. But in the present case, where ⋆θ was constructed using the Moyal product, we do have a representation πθ of (Eθ(G), ⋆θ) on L2(G) defined by πθ(f1)f2 = f1 ⋆θ f2 for f1, f2 ∈ Eθ(G). Furthermore, for this representation we have πθ(f )∗ = πθ( ¯f ). Since by construction the products ⋆θ and ⋆Ωθ coincide on A∞(G) ⊂ A(G) ∩ Eθ(G) (but A∞(G) is not closed under these products), for any f ∈ A∞(G) we have where ∇ is the unbounded involutive operator defined by ∇f = f . Since ∇∗ = ∆−1 is consistent with Lemma 5.1. G ∇ = ∇∆G, this In view of the identity πΩθ (f ) = ∇πθ(f )∇ it may seem surprising that both representations πΩθ and πθ are well-defined. The representation πΩθ is well-defined by our general theory. The reason why πθ is well-defined is that ultimately the product ⋆θ was constructed using a dual cocycle on R2d+2 and this group is unimodular. πΩθ (f ) = ∇πθ(f )∇ on A∞(G) ⊂ L2(G), Turning to C∗-algebras, the obvious conclusion is that the identity map on A∞(G) does not extend r ( G; Ωθ) = πΩθ (A(G)) and πθ(Eθ(G)). This, however, does not exclude the to a ∗-isomorphism of C ∗ possibility that these C∗-algebras are isomorphic in a canonical G-equivariant way. In fact, the above considerations suggest that the conjugation by the involutive unitary ∆−1/2 G ∇ = J J gives such an isomorphism. This will be analyzed in a subsequent publication. The C∗-algebra πθ(Eθ(G)) is the θ-deformation of C0(G) as defined by Bieliavsky and Gayral [6]. To be more precise, instead of the representation πθ they use the Weyl quantization map. But it In particular, as an abstract C∗- is well-known that this gives a quasi-equivalent representation. algebra, πθ(Eθ(G)) is isomorphic to the algebra of compact operators on an infinite dimensional separable Hilbert space. 5.3. Two-parameter deformation. The papers [6] and [5] contain a more general class of defor- mations, with a second parameter of deformation being a function on R. These deformations are obtained by inserting an additional factor into the definition of the map Tθ. We do not need the precise definition of this procedure, see [6, Section 4.1] for details, and will only write down the final answer. Given a smooth function τ on R satisfying certain growth conditions, we have a G-invariant product ⋆θ,τ on a function space Eθ,τ (G) defined by the kernel Kθ,τ (x, x′) = Kθ(x, x′) exp(cid:26)τ (cid:18) 2 θ sinh(2a)(cid:19) + τ (cid:18) 2 θ sinh(−2a′)(cid:19) − τ (cid:18) 2 θ sinh(2a − 2a′)(cid:19)(cid:27) . If τ is purely imaginary, this kernel defines a unitary 2-cocycle Ωθ,τ on G such that Ω∗ θ,τ = ZG×G Kθ,τ (x, y)λx ⊗ λy dx dy on C ∞ c (G × G) ⊂ L2(G × G). In order to understand this cocycle, consider the von Neumann algebra W ∗(R) of R. The conju- gation by the inverse of the Fourier transform, defined by (Ff )(ξ) = f (ξ) = 1 √2π ZR f (t)e−iξtdt, gives an isomorphism L∞(R) ∼= W ∗(R), so the unitary e−τ ∈ L∞(R) defines a unitary uτ = F −1e−τF ∈ W ∗(R). Using the embedding R ֒→ G, t 7→ (0, 0, t), we get an embedding W ∗(R) ֒→ W ∗(G), so we can consider the unitary uτ as an element of W ∗(G). = c (G × G) we have: Ω∗ = Z A(a, a′) exp(cid:26) 2i θ × λx ⊗ λx′ dx dx′ 1 (2π)3/2 Z f1(t1) f1(t′ × exp(cid:26) 2i × exp(cid:26) 2i (2π)3/2 Z f1(t1) f1(t′ × exp(cid:26) 2i × λ(a,v,t−t′ = 1 1) f2(t2) A(a, a′) exp(cid:26) 2i θ cosh(a) cosh(a′)ω0(v, v′)(cid:27) 1 sinh(2a′) + t2(sinh(2a) cosh(2a′) − cosh(2a) sinh(2a′))(cid:1)(cid:27) θ (cid:0)t1 sinh(2a) − t′ (sinh(2a)t′ − sinh(2a′)t)(cid:27) λx ⊗ λx′ dt1 dt′ θ 1 dt2 dx dx′ 1) f2(t2) A(a, a′) exp(cid:26) 2i θ cosh(a) cosh(a′)ω0(v, v′)(cid:27) θ (cid:0) sinh(2a)t′ − sinh(2a′)t(cid:1)(cid:27) 1−t2 cosh(2a)) ⊗ λ(a′,v′,t′−t1−t2 cosh(2a′))dt1 dt′ 1 dt2 dx dx′. 26 S. NESHVEYEV AND L. TUSET Proposition 5.2. We have Ωθ,τ = (uτ⊗uτ )Ωθ ∆(uτ )∗, so the cocycles Ωθ,τ and Ωθ are cohomologous. Proof. It is convenient to prove a slightly different statement. Fix functions f1, f2 ∈ S(R). Consider the corresponding elements bi = F −1fiF of W ∗(R) ⊂ W ∗(G). Explicitly, Consider also the function bi = 1 √2π ZR fi(−t)λ(0,0,t)dt. K(x, x′) = Kθ(x, x′)f1(cid:18)2 θ sinh(2a)(cid:19) f1(cid:18) 2 θ sinh(−2a′)(cid:19) f2(cid:18) 2 θ sinh(2a − 2a′)(cid:19) and the operator Ω∗ = RG×G claim that this is a bounded operator on L2(G × G) and K(x, y)λx ⊗ λy dx dy, which is at least defined on C ∞ c (G × G). We (5.1) Denote 4(πθ)−2d−2A(x, x′) by A(a, a′) (recall that A(x, x′) depends only on the coordinates a, a′). Then on C ∞ θ(b1 ⊗ b1). Ω∗ = ∆(b2)Ω∗ S(x, x′)(cid:27) f1(cid:18) 2 θ sinh(2a)(cid:19) f1(cid:18) 2 θ sinh(−2a′)(cid:19) f2(cid:18) 2 θ sinh(2a − 2a′)(cid:19) The group multiplication formula gives Hence Ω∗ = (a, v, t − t′ 1 − t2 cosh(2a)) = (0, 0,−t2)(a, v, t − t2 sinh(2a))(0, 0,−t′ 1). 1 1) f2(t2) A(a, a′) exp(cid:26) 2i cosh(a) cosh(a′)ω0(v, v′)(cid:27) θ (cid:0) sinh(2a)(t′ + t2 sinh(2a′)) − sinh(2a′)(t + t2 sinh(2a))(cid:1)(cid:27) (2π)3/2 Z f1(t1) f1(t′ × exp(cid:26) 2i × (λ(0,0,−t2) ⊗ λ(0,0,−t2))(λx ⊗ λx′)(λ(0,0,−t′ 1) ⊗ λ(0,0,−t1))dt1 dt′ θ 1 dt2 dx dx′ = ∆(b2)Ω∗ θ(b1 ⊗ b1). Now, choosing a bounded sequence of functions gn ∈ S(R) converging to eτ pointwise and passing to the limit in identity (5.1) applied to the pairs (f1, f2) = (gn, ¯gn), we get the result. (cid:3) Therefore, from our perspective, there is no reason to introduce the second deformation param- eter τ , since by Proposition 3.12 this leads to isomorphic deformations. Note also that on the level of the function spaces Eθ,τ (G) the corresponding G-equivariant isomorphism (Eθ(G), ⋆θ) ∼= (Eθ,τ (G), ⋆θ,τ ) is given by the operator F −1e−τF. DEFORMATION OF C∗-ALGEBRAS 27 6. Open problems The results that we have obtained so far lead to a number of questions that have to be resolved in order to bring the theory to a completely satisfactory level. In this section we list some of the most natural ones. 6.1. Regularity of cocycles. The problem is to find simple, verifiable conditions for regularity. In view of our considerations in Sections 4.2 and 4.3 for group duals, it seems unlikely that such conditions exist. At the same time we do not have a single example of a nonregular cocycle on a regular quantum group. 6.2. Regularity of deformed quantum groups. As has been shown by De Commer [11], reg- ularity of a quantum group is not preserved under cocycle deformation: the nonregular quantum group Eq(2) is obtained by deformation by a cocycle on SUq(2). In this respect we want to formulate the following question: if G is a regular quantum group and Ω is a cocycle on G, is regularity of Ω equivalent to regularity of GΩ? If not, do we have an implication in at least one direction? 6.3. Generalized fixed point algebras. By Proposition 4.2, in the case of cocycles on group duals, for the Ω-deformation AΩ of a C∗-algebra A we have AΩ ⊂ [ψβ⊗α(P + β⊗α)]. It is not difficult to see that the inclusion can be strict already for G = Z and Ω = 1. What is the proper characterization of elements of AΩ in terms of the action β ⊗ α of G on C ∗ Another question is what an analogue of this setting for general quantum groups is. When G is a genuine group, what is of course special, is that any action of G can be viewed as a left or a right action and the tensor product action is always well-defined. When G is a group dual, then again we can always pass from a left to a right action. But in order to define the diagonal action we have to replace the usual tensor product by the braided tensor product ⊠. As was shown by Yamashita [44], when in addition G is compact, so C0(G) = C ∗ r (Γ) for a discrete group Γ, then AΩ is isomorphic to r (Γ; Ω)⊠A)G. But for general quantum groups, when (C0(G), ∆) is neither the fixed point algebra (C ∗ commutative nor cocommutative, it is not clear to us what the correct analogue of the description of AΩ as (a subalgebra of) a fixed point algebra is. r ( G; Ω) ⊗ A? 6.4. Generalization of Rieffel's deformation. In the setting of Section 5, Bieliavsky and Gayral defined a θ-deformation Aθ of any C∗-algebra A equipped with an action of G. The question is how the algebras Aθ are related to our algebras AΩθ . As we have seen, this question is not quite trivial already for A = C0(G). Assuming it can be rigorously settled in this case, for general A both r ( G; Ωθ)⊗ A)β⊗α, and then the question is whether algebras Aθ and AΩθ can be embedded into M (C ∗ they coincide. The analogous question for Rieffel's deformation has an affirmative answer [3, 27]. There are several reasons why it will be difficult to give a similar proof in the present case. One of them is that in the case of Rieffel's deformation the group R2d carrying a dual cocycle was abelian, so the deformation still carried an action of the same group. This is no longer the case for the groups considered by Bieliavsky and Gayral, where we can only hope that Aθ carries an action of GΩθ . References [1] S. Baaj and G. Skandalis, Unitaires multiplicatifs et dualit´e pour les produits crois´es de C∗-alg`ebres, Ann. Sci. ´Ecole Norm. Sup. (4) 26 (1993), no. 4, 425 -- 488. [2] S. Baaj, G. Skandalis and S. Vaes, Non-semi-regular quantum groups coming from number theory, Comm. Math. Phys. 235 (2003), no. 1, 139 -- 167. [3] J. Bhowmick, S. Neshveyev and A. Sangha, Deformation of operator algebras by Borel cocycles, J. Funct. Anal. 265 (2013), no. 6, 983 -- 1001. [4] J. Bichon, A. De Rijdt and S. Vaes, Ergodic coactions with large multiplicity and monoidal equivalence of quantum groups, Comm. Math. Phys. 262 (2006), no. 3, 703 -- 728. 28 S. NESHVEYEV AND L. TUSET [5] P. Bieliavsky, Ph. Bonneau, F. D'Andrea, V. Gayral, Y. Maeda and Y. Voglaire, Multiplicative unitaries and locally compact quantum Kahlerian groups, in preparation; see also "Deformation of Kahlerian Lie groups", available at http://www.dma.unina.it/francesco.dandrea/Files/Caen2010.[slides].pdf. [6] P. Bieliavsky and V. Gayral, Deformation Quantization for Actions of Kahlerian Lie Groups, preprint arXiv:1109.3419v4 [math.OA]. [7] D. Bulacu, F. Panaite and F. Van Oystaeyen, Quasi-Hopf algebra actions and smash products, Comm. Algebra 28 (2000), no. 2, 631 -- 651. [8] A. Connes and G. Landi, Noncommutative manifolds, the instanton algebra and isospectral deformations, Comm. Math. Phys. 221 (2001), no. 1, 141 -- 159. [9] K. De Commer, Galois coactions for algebraic and locally compact quantum groups, PhD thesis, Katholieke Uni- versiteit Leuven, 2009. [10] K. De Commer, Galois objects and cocycle twisting for locally compact quantum groups, J. Operator Theory 66 (2011), no. 1, 59 -- 106. [11] K. De Commer, On a Morita equivalence between the duals of quantum SU (2) and quantum E(2), Adv. Math. 229 (2012), no. 2, 1047 -- 1079. [12] A. De Rijdt and N. Vander Vennet, Actions of monoidally equivalent compact quantum groups and applications to probabilistic boundaries, Ann. Inst. Fourier (Grenoble) 60 (2010), no. 1, 169 -- 216. [13] V.G. Drinfeld, Quantum groups, in: Proceedings of the International Congress of Mathematicians, Vol. 1, 2 (Berkeley, Calif., 1986), 798 -- 820, Amer. Math. Soc., Providence, RI, 1987. [14] M. Enock, Measured quantum groupoids in action, M´em. Soc. Math. Fr. (N.S.) No. 114 (2008), ii+150 pp. [15] P. Fima and L. Vainerman, Twisting and Rieffel's deformation of locally compact quantum groups: deformation of the Haar measure, Comm. Math. Phys. 286 (2009), no. 3, 1011 -- 1050. [16] K.C. Hannabuss and V. Mathai, Noncommutative principal torus bundles via parametrised strict deformation quantization, in: "Superstrings, geometry, topology, and C ∗-algebras", 133 -- 147, Proc. Sympos. Pure Math., 81, Amer. Math. Soc., Providence, RI, 2010. [17] P. Kasprzak, Rieffel deformation via crossed products, J. Funct. Anal. 257 (2009), no. 5, 1288 -- 1332. [18] P. Kasprzak, Rieffel deformation of group coactions, Comm. Math. Phys. 300 (2010), no. 3, 741 -- 763. [19] J. Kustermans and S. Vaes, Locally compact quantum groups, Ann. Sci. ´Ecole Norm. Sup. (4) 33 (2000), no. 6, 837 -- 934. [20] J. Kustermans and S. Vaes, Locally compact quantum groups in the von Neumann algebraic setting, Math. Scand. 92 (2003), no. 1, 68 -- 92. [21] M.B. Landstad, Ergodic actions of nonabelian compact groups, in: "Ideas and methods in mathematical analysis, stochastics, and applications" (Oslo, 1988), 365 -- 388, Cambridge Univ. Press, Cambridge, 1992. [22] M.B. Landstad and I. Raeburn, Twisted dual-group algebras: equivariant deformations of C0(G), J. Funct. Anal. 132 (1995), no. 1, 43 -- 85. [23] M.B. Landstad and I. Raeburn, Equivariant deformations of homogeneous spaces, J. Funct. Anal. 148 (1997), no. 2, 480 -- 507. [24] S. Majid, Quasi-∗ structure on q-Poincar´e algebras, J. Geom. Phys. 22 (1997), no. 1, 14 -- 58. [25] M.V. Movshev, Twisting in group algebras of finite groups, Funct. Anal. Appl. 27 (1993), 240 -- 244. [26] S. Neshveyev and L. Tuset, On second cohomology of duals of compact groups, Internat. J. Math. 22 (2011), no. 9, 1231 -- 1260. [27] S. Neshveyev, Smooth crossed products of Rieffel's deformations, preprint arXiv:1307.2016v1 [math.OA]. [28] W. Pusz and S.L. Woronowicz, A new quantum deformation of 'ax + b' group, Comm. Math. Phys. 259 (2005), no. 2, 325 -- 362. [29] M.A. Rieffel, Proper actions of groups on C∗-algebras, in: "Mappings of operator algebras" (Philadelphia, PA, 1988), 141 -- 182, Progr. Math., 84, Birkhauser Boston, Boston, MA, 1990. [30] M.A. Rieffel, Deformation quantization for actions of Rd, Mem. Amer. Math. Soc. 106 (1993), no. 506. [31] M.A. Rieffel, Integrable and proper actions on C∗-algebras, and square-integrable representations of groups, Expo. Math. 22 (2004), no. 1, 1 -- 53. [32] A. Sitarz, Twists and spectral triples for isospectral deformations, Lett. Math. Phys. 58 (2001), no. 1, 69 -- 79. [33] G. Skandalis, Duality for locally compact 'quantum groups' (joint work with Baaj), in: Mathematisches Forschungsinstitut Oberwolfach, Tagungsbericht 46/1991, "C∗-algebren", 20.10 -- 26.10.1991, p. 20. [34] P. Stachura, On the quantum 'ax+b' group, J. Geom. Phys. 73 (2013), 125 -- 149. [35] S. Vaes, A new approach to induction and imprimitivity results, J. Funct. Anal. 229 (2005), no. 2, 317 -- 374. [36] S. Vaes and L. Vainerman, Extensions of locally compact quantum groups and the bicrossed product construction, Adv. Math. 175 (2003), no. 1, 1 -- 101. [37] J.C. V´arilly, Quantum symmetry groups of noncommutative spheres, Comm. Math. Phys. 221 (2001), no. 3, 511 -- 523. [38] S. Wang, Deformations of compact quantum groups via Rieffel's quantization, Comm. Math. Phys. 178 (1996), no. 3, 747 -- 764. DEFORMATION OF C∗-ALGEBRAS 29 [39] S. Wang, Rieffel type discrete deformation of finite quantum groups, Comm. Math. Phys. 202 (1999), no. 2, 291 -- 307. [40] A. Wassermann, Automorphic actions of compact groups on operator algebras, Ph.D. Dissertation, University of Pennsylvania (1981). [41] A. Wassermann, Ergodic actions of compact groups on operator algebras. I. General theory, Ann. of Math. 130 (1989), 273 -- 319. [42] S.L. Woronowicz, From multiplicative unitaries to quantum groups, Internat. J. Math. 7 (1996), no. 1, 127 -- 149. [43] S.L. Woronowicz and S. Zakrzewski, Quantum 'ax + b' group, Rev. Math. Phys. 14 (2002), no. 7 -- 8, 797 -- 828. [44] M. Yamashita, Deformation of algebras associated to group cocycles, preprint arXiv:1107.2512v1 [math.OA]. [45] G. Zeller-Meier, Produits crois´es d'une C∗-alg`ebre par un groupe d'automorphismes, J. Math. Pures Appl. (9) 47 (1968), 101 -- 239. E-mail address: [email protected] Department of Mathematics, University of Oslo, P.O. Box 1053 Blindern, NO-0316 Oslo, Norway E-mail address: [email protected] Department of Computer Science, Oslo and Akershus University College of Applied Sciences, P.O. Box 4 St. Olavs plass, NO-0130 Oslo, Norway
1301.3640	1	1301	2013-01-16T10:07:01	Operator synthesis and tensor products	[ "math.OA" ]	We show that Kraus' property $S_{\sigma}$ is preserved under taking weak* closed sums with masa-bimodules of finite width, and establish an intersection formula for weak* closed spans of tensor products, one of whose terms is a masa-bimodule of finite width. We initiate the study of the question of when operator synthesis is preserved under the formation of products and prove that the union of finitely many sets of the form $\kappa \times \lambda$, where $\kappa$ is a set of finite width, while $\lambda$ is operator synthetic, is, under a necessary restriction on the sets $\lambda$, again operator synthetic. We show that property $S_{\sigma}$ is preserved under spatial Morita subordinance. En route, we prove that non-atomic ternary masa-bimodules possess property $S_{\sigma}$ hereditarily.	math.OA	math	OPERATOR SYNTHESIS AND TENSOR PRODUCTS G.K. ELEFTHERAKIS AND I.G. TODOROV Abstract. We show that Kraus' property Sσ is preserved under taking weak* closed sums with masa-bimodules of finite width, and establish an intersection formula for weak* closed spans of tensor products, one of whose terms is a masa- bimodule of finite width. We initiate the study of the question of when operator synthesis is preserved under the formation of products and prove that the union of finitely many sets of the form κ × λ, where κ is a set of finite width, while λ is operator synthetic, is, under a necessary restriction on the sets λ, again operator synthetic. We show that property Sσ is preserved under spatial Morita subordinance. En route, we prove that non-atomic ternary masa-bimodules possess property Sσ hereditarily. 1. Introduction Operator synthesis was introduced by W.B. Arveson in his seminal paper [1] as an operator theoretic version of the notion of spectral synthesis in Harmonic Analysis, and was subsequently developed by J. Froelich, A. Katavolos, J. Ludwig, V.S. Shulman, N. Spronk, L. Turowska and the authors [8], [12], [16], [25], [26], [27], [28], [29], among others. It was shown in [12], [27] and [20] that, for a large class of locally compact groups G, given a closed subset E of G, there is a canonical way to produce a subset E∗ of the direct product G × G, so that the set E satisfies spectral synthesis if and only if the set E∗ satisfies operator synthesis. Thus, the well-known, and still open, problem of whether the union of two sets of spectral synthesis satisfies spectral synthesis can be viewed as a special case of the problem asking whether the union of two operator synthetic sets is operator synthetic. Another problem in Harmonic Analysis asks when the product of two sets of spectral synthesis is again synthetic. The analogous question in the operator theory setting is closely related to property Sσ, introduced by J. Kraus in [17]. It is widely recognised that functional analytic tensor products display a larger degree of subtlety than the algebraic ones, the reason for this being the fact that they are defined as the completion of the algebraic tensor product of two objects (say, operator algebras, or operator spaces) with respect to an appropriate topology. Therefore, it is usually not an easy task to determine the intersection of two spaces, both given as completed tensor products. Such issues give rise to a number of important concepts in Operator Algebra Theory, e.g. exactness [23]. Property Sσ is instrumental in describing such intersections, and is closely related to a number of important approximation properties. In particular, it was shown in [18] to be equivalent to the σ-weak approximation property, while in [13], an equivalence of Date: 12 January 2013. 1 2 G.K. ELEFTHERAKIS AND I.G. TODOROV when a group von Neumann algebra VN(G) possesses property Sσ was formulated in terms of an approximation property of the underlying group G. In this paper, we initiate the study of the question of when the direct product of two operator synthetic sets is operator synthetic. Furthermore, we combine the two stands of investigation highlighted in the previous two paragraphs by studying the question of when the union of direct products of operator synthetic sets is operator synthetic. The setting of operator synthesis is provided by the theory of masa-bimodules (see [1], [9], [25] and [26]). A prominent role in our study is played by the masa-bimodules of finite width. This class is a natural extension of the class of CSL algebras of finite width, which was introduced in [1] as a far reaching, yet tractable, generalisation of nest algebras [4]. It was shown in [14] that CSL algebras of finite width possess property Sσ. However, this class has for long remained the main example of operator spaces known to have this property. We note that the question of whether every weak* closed masa-bimodule possesses property Sσ is still open (see [18]). It should be noted that masa-bimodules of finite width have been studied in a number of other contexts. They include as a subclass the masa-bimodules which are ternary rings of operators [29], a class of operator spaces that has been studied extensively for the purposes of Operator Space Theory [3]. The supports of masa- bimodules of finite width (called henceforth sets of finite width) are precisely the sets of solutions of systems of inequalities, and were shown in [25] and [28] to be operator synthetic, providing in this way the largest single class of sets that are known to satisfy operator synthesis. It was shown in [8] that the union of an operator synthetic set and a set of finite width is operator synthetic. In [24], this line of investigation was continued by showing that masa-bimodules of finite width satisfy a rank one approximation property, and a large class of examples of sets of operator multiplicity was exhibited within this class. They were the motivating example for the introduction and study of I-decomposable masa-bimodules in [8]. The weak* closed masa-bimodules are precisely the weak* closed invariant sub- spaces of Schur multipliers or, equivalently, of weak* continuous (completely) bounded masa-bimodule maps. The projections in the algebra of all Schur multipli- ers, called henceforth Schur idempotents, were at the core of the methods developed in [8] in order to address the union problem, as well as the closely related problem of the reflexivity of weak* closed spans. Here we significantly extend the techniques whose development was initiated in [8] by establishing an intersection formula involving tensor products and applying it to the study of the product and union problems described above. Simultaneously, we initiate the study of the question of whether property Sσ is preserved under taking weak* closed spans. The paper is organised as follows. After gathering some preliminary notions and results in Section 2, we address in Section 3 the preservation problem for property Sσ outlined in the previous paragraph, showing that the class of spaces possessing Sσ is closed under taking weak* closed sums with masa-bimodules of finite width (Theorem 3.10). As a consequence, the weak* closed span of any finite number of OPERATOR SYNTHESIS AND TENSOR PRODUCTS 3 masa-bimodules of finite width possesses property Sσ. En route, we give a sufficient condition for a ternary masa-bimodule to possess Sσ hereditarily (Theorem 3.7). In Section 4, we establish the intersection formula B1 j1 ⊗ U1 + · · · + Br jr ⊗ Ur = (∩j1B1 j1) ⊗ U1 + · · · + (∩jrB1 jr ) ⊗ Ur, (1) \j1,...,jr valid for all masa-bimodules Bp jp of finite width and all weak* closed spaces of operators Up, p = 1, . . . , r, jp = 1, . . . , mp (Corollary 4.21). In Section 5, we formalise the relation between property Sσ and the problem for the synthesis of products (see Corollary 5.4). As part of Proposition 5.3, we establish a subspace version of the relation between Fubini products and the algebra tensor product formula discussed in [17]. These results, along with the formula (1), are used to show that the union of finitely many products κi × λi, where the sets κi are of finite width, while λi are operator synthetic sets satisfying certain necessary restrictions, is operator synthetic (Theorem 5.9). Finally, in Section 6, we show that property Sσ is preserved under spatial Morita subordinance. As a corollary, we obtain that if L1 and L2 are isomorphic CSL's then the CSL algebra AlgL1 possesses property Sσ if and only if AlgL2 does so. It is natural to wonder whether our results are valid for the more general class consisting of intersections of I-decomposable spaces introduced in [8]. We note that this class contains properly the class of masa-bimodules of finite width. Progress in this direction would rely on the answer of the question of whether the approximately I-injective masa-bimodules (that is, the intersections of descending sequences of ranges of uniformly bounded Schur idempotents) satisfy property Sσ; this question, however, is still open. 2. Preliminaries In this section, we collect some preliminary notions and results that will be needed in the sequel. If H and K are Hilbert spaces, we denote by B(H, K) the space of all bounded linear operators from H into K, and write B(H) = B(H, H). The space B(H, K) is the dual of the ideal of all trace class operators from K into H, and can hence be endowed with a weak* topology; we note that this is the weakest topology on B(H, K) with respect to which the functionals ω of the form ω(T ) = (T ξk, ηk), T ∈ B(H, K), ∞Xk=1 where (ξk)k∈N ⊆ H and (ηk)k∈N ⊆ K are square summable sequences of vectors, are continuous. In the sequel, we denote by U the weak* closure of a set U ⊆ B(H, K). Throughout the paper, H, K, H1, K1, H2 and K2 will denote Hilbert spaces. Let V ⊆ B(H1, H2) and U ⊆ B(K1, K2) be weak* closed subspaces. We denote by V ¯⊗U the weak* closed subspace of B(H1 ⊗ K1, H2 ⊗ K2) generated by the operators of the form S ⊗ T , where S ∈ V and T ∈ U . Here, H ⊗ K is the Hilbertian tensor product of H and K, and we use the natural identification B(H1 ⊗ K1, H2 ⊗ K2) ≡ B(H1, H2) ¯⊗B(K1, K2). 4 G.K. ELEFTHERAKIS AND I.G. TODOROV We will use some basic notions from Operator Space Theory; we refer the reader to the monographs [3], [5], [22] and [23] for the relevant definitions. If X is a linear space, we denote by id the identity map on X . The range of a linear map φ on X is denoted by Ran φ. As customary, the map φ is called idempotent if φ◦φ = φ; we let φ⊥ = id −φ. If X1 and X2 are subspaces of X , we set X1+X2 = {x1+x2 : xi ∈ Xi, i = 1, 2}. If Vi ⊆ B(H1, H2) and Ui ⊆ B(K1, K2) are weak* closed subspaces, i = 1, 2, and φ : V1 → V2 and ψ : U1 → U2 are completely bounded weak* continuous maps, then there exists a (unique) completely bounded weak* continuous map φ ⊗ ψ : V1 ¯⊗U1 → V2 ¯⊗U2 such that φ ⊗ ψ(A ⊗ B) = φ(A) ⊗ ψ(B), A ∈ V1, B ∈ U1 [3]. In the case U1 = U2 = B(K1, K2), we write throughout the paper φ = φ⊗id. We denote by V∗ the space of all weak* continuous functionals on V. If V ⊆ B(H1, H2) is a weak* closed subspace of operators and ω ∈ V∗ then we set Rω = ω; thus, Rω : V ¯⊗B(K1, K2) → B(K1, K2) is the Tomiyama's right slice map corresponding to ω (here we have use the natural identification C ¯⊗B(K1, K2) ≡ B(K1, K2)). We note that Rω(A ⊗ B) = ω(A)B, A ∈ V, B ∈ B(K1, K2). If, further, U ⊆ B(K1, K2) is a weak* closed subspace, the Fubini product F(V, U ) of V and U is the subspace of V ¯⊗B(K1, K2) given by F(V, U ) = {T ∈ V ¯⊗B(K1, K2) : Rω(T ) ∈ U , for all ω ∈ V∗}. If ξ ∈ H1 and η ∈ H2, we let ωξ,η be the vector functional on B(H1, H2) given by ωξ,η(A) = (Aξ, η), A ∈ B(H1, H2); we use the same symbol to denote the restriction of ωξ,η to the subspace U ⊆ B(H1, H2). It is easy to notice that V ¯⊗U ⊆ F(V, U ). The subspace V is said to possess property Sσ if F(V, U ) = V ¯⊗U for all weak* subspaces U ⊆ B(K1, K2) and all Hilbert spaces K1, K2. This notion was introduced by Kraus in [17], where he showed that B(K1, K2) possesses property Sσ. (We note that Kraus considered the case K1 = K2; however, it is easy to see that one can state both the definition and the result in terms of two Hilbert spaces.) From this fact, one can easily derive the formula F(V, U ) = (V ¯⊗B(K1, K2)) ∩ (B(H1, H2) ¯⊗U ). Now suppose that H1 and H2 are separable Hilbert spaces and D1 ⊆ B(H1) and D2 ⊆ B(H2) are maximal abelian selfadjoint algebras (for brevity, masas). A linear map φ on B(H1, H2) is called D2, D1-modular, or a masa-bimodule map when D1 and D2 are clear from the context, provided φ(BXA) = Bφ(X)A, for all X ∈ B(H1, H2), A ∈ D1 and B ∈ D2. We call the completely bounded weak* continuous D2, D1-modular maps on B(H1, H2) Schur maps (relative to the pair (D1, D2)); a Schur map that is also an idempotent is called a Schur idempotent. This terminology is natural in view of the fact that Schur maps correspond precisely to Schur multipliers, provided a particular coordinate representation of D1 and D2 is chosen. We refer the reader to [8] for details; we will return to this perspective in Section 5. A D2, D1-bimodule, or simply a masa-bimodule when D1 and D2 are understood from the context, is a subspace V ⊆ B(H1, H2) such that BXA ∈ V whenever X ∈ V, A ∈ D1 and B ∈ D2. Masa-bimodules will be assumed to be weak* closed throughout the paper; they are precisely the weak* closed subspaces invariant OPERATOR SYNTHESIS AND TENSOR PRODUCTS 5 under all Schur maps (see [8, Proposition 3.2]). A weak* closed masa-bimodule M is called ternary [16], [29] if M is a ternary ring of operators, that is, if T S∗R ∈ M whenever T, S, R ∈ M (see also [3]). It is not difficult to see that every ternary masa-bimodule is the intersection of a descending sequence of ranges of contractive Schur idempotents; this fact will be used extensively hereafter. It is easy to notice that the ternary masa-bimodules acting on a single Hilbert space which are unital algebras are precisely the von Neumann algebras with abelian commutant. A nest on a Hilbert space H is a totally ordered family of closed subspaces of H that contains the intersection and the closed linear span (denoted ∨) of any if its subsets. A nest algebra is the subalgebra of B(H) of all operators leaving invariant each subspace of a given nest. A nest algebra bimodule is a subspace V ⊆ B(H1, H2) for which there exist nest algebras A ⊆ B(H1) and B ⊆ B(H2) such that BVA ⊆ V. All nest algebra bimodules will be assumed to be weak* closed. A D2, D1-bimodule V is said to have finite width if it is of the form V = V1 ∩ · · · ∩ Vk, where each Vj is a D2, D1-bimodule that is also a nest algebra bimodule. The smallest k with this property is called the width of V. We note that every ternary masa-bimodule has width at most two [16]. If each Vj is a nest algebra then V is called a CSL algebra of finite width [1]. It was shown in [17] that von Neumann algebras with abelian commutant possess property Sσ and in [14] that every CSL algebra of finite width possesses property Sσ. Suppose that V ⊆ B(H1, H2) is a nest algebra bimodule. It was shown in [10] that there exist nests N1 ⊆ B(H1) and N2 ⊆ B(H2) and an increasing ∨-preserving map ϕ : N1 → N2 such that V = {X ∈ B(H1, H2) : XN = ϕ(N )XN, N ∈ N1}. Let {Pi}i∈N ⊆ N1 be a (countable) subset dense in N1 in the strong operator topology such that the set {ϕ(Pi)}i∈N is dense in N2, and Fn = {0, P1, P2, . . . , Pn, I} = {0 < N1 < N2 < · · · < Nn < I}. Set N0 = 0 and Nn+1 = I and let φn, ψn : B(H1, H2) → B(H1, H2) be the Schur idempotents (relative to any pair (D1, D2) with N1 ⊆ D1 qnd N2 ⊆ D2) given by φn(X) = nXi=0 ψn(X) = X0≤i<j≤n (ϕ(Ni+1) − ϕ(Ni))X(Ni+1 − Ni), X ∈ B(H1, H2), (ϕ(Ni+1) − ϕ(Ni))X(Nj+1 − Nj), X ∈ B(H1, H2), and Mn and Wn be the ranges of φn and ψn, respectively. We have that φnψn = 0, Wn ⊆ V ⊆ Wn + Mn, Wn ⊆ Wn+1, Mn+1 ⊆ Mn, n ∈ N, and ∩∞ n=1Mn ⊆ V. We will call the family (φn, ψn, Mn, Wn)n∈N a decomposition scheme for V. Decomposition schemes were first explicitly used (although not referred to as such) in [8] for the study of reflexivity and synthesis problems. We note that, if ψn,p, p = 1, . . . , n, is given by (ϕ(Ni+1) − ϕ(Ni))X(Nj+1 − Nj), X ∈ B(H1, H2), ψn,p(X) = Xj−i=p 6 G.K. ELEFTHERAKIS AND I.G. TODOROV then ψn =Pn p=1 ψn,p, ψn,pψn,q = 0 if p 6= q and kψn,pk = 1, p = 1, . . . , n. 3. Stability under summation with modules of finite width The main result in this section is Theorem 3.10 and the associated Corollary 3.11, which show that tensor product formulas are preserved under taking weak* closed sums with masa-bimodules of finite width. En route, we establish a sufficient condition for a ternary masa-bimodule to possess Sσ hereditarily (Theorem 3.7). We begin with some lemmas. Lemma 3.1. Let φ be a weak* continuous completely bounded linear map on B(H1, H2) and V, Vi ⊆ B(H1, H2), U , Ui ⊆ B(K1, K2) be weak* closed subspaces, i = 1, . . . , n. Suppose that V is invariant under φ. i=1 φ(Vi) ¯⊗Ui. In particular, φ leaves V ¯⊗U (i) We have φ(Pn i=1 Vi ¯⊗Ui) ⊆ Pn invariant. (ii) If φ is an idempotent then (Ran φ ¯⊗U ) ∩ (V ¯⊗U ) = φ(V) ¯⊗U = φ(V ¯⊗U ). In particular, ranges of Schur idempotents possess property Sσ. Proof. (i) Fix i ∈ {1, . . . , n} and suppose that T ∈ Vi ¯⊗Ui. Then T can be approxi- j=1 Aj ⊗Bj, where Aj ∈ Vi, Bj ∈ Ui, j = 1, . . . , k; therefore, φ(T ) can be approximated in the weak* topology j=1 φ(Aj) ⊗ Bj, where Aj ∈ Vi, Bj ∈ Ui, j = 1, . . . , k. Hence, φ(T ) ∈ φ(Vi) ¯⊗Ui. The conclusion now follows from the linearity and the weak* continuity of φ. mated in the weak* topology by operators of the formPk by operators of the formPk (ii) Since φ is an idempotent, φ(V) is weak* closed. By (i), φ(V ¯⊗U ) ⊆ φ(V) ¯⊗U while, since φ(V) ⊆ V, we have φ(V) ¯⊗U ⊆ (Ran φ ¯⊗U ) ∩ (V ¯⊗U ). Suppose that T ∈ (Ran φ ¯⊗U ) ∩ (V ¯⊗U ). Then, by (i), φ⊥(T ) = 0 and hence T = φ(T ) ∈ φ(V ¯⊗U ). Finally, if T ∈ (Ran φ ¯⊗B(K1, K2)) ∩ (B(H1, H2) ¯⊗U ) then, by the previous paragraph, T = φ(T ) ∈ φ(B(H1, H2)) ¯⊗U ; thus, Ran φ pos- sesses Sσ. (cid:3) Lemma 3.2. Let φ be a weak* continuous completely bounded idempotent acting on B(H1, H2), V ⊆ B(H1, H2) be a weak* closed subspace invariant under φ, U ⊆ B(K1, K2) be a weak* closed subspace and W ⊆ B(H1 ⊗ K1, H2 ⊗ K2) be a weak* closed subspace invariant under φ. Then V ¯⊗U + W ∩ Ran φ ¯⊗U + W = (Ran φ ∩ V) ¯⊗U + W. Proof. Suppose By Lemma 3.1 and the invariance of W under φ, we have that φ⊥(T ) ∈ W; similarly, T ∈ V ¯⊗U + W ∩ Ran φ ¯⊗U + W. φ(T ) ∈ φ(V) ¯⊗U + W. OPERATOR SYNTHESIS AND TENSOR PRODUCTS It follows that T = φ⊥(T ) + φ(T ) ∈ (Ran φ ∩ V) ¯⊗U + W. The converse inclusion is trivial. 7 (cid:3) Lemma 3.3. Let V ⊆ B(H1, H2) (resp. U ⊆ B(K1, K2)) be a weak* closed subspace and φ (resp. ψ) be a weak* continuous completely bounded map on B(H1, H2) (resp. B(K1, K2)). Then (φ ⊗ ψ)(F(V, U )) ⊆ F(φ(V), ψ(U )). Moreover, if φ and ψ are idempotents that leave V and U , respectively, invariant, then (φ ⊗ ψ)(F(V, U )) = F(φ(V), ψ(U )). Proof. By Lemma 3.1, (φ ⊗ ψ)(V ¯⊗B(K1, K2)) = φ ◦ (id ⊗ψ)(V ¯⊗B(K1, K2)) ⊆ φ(V ¯⊗B(K1, K2)) ⊆ φ(V) ¯⊗B(K1, K2); similarly, Hence, (φ ⊗ ψ)(B(H1, H2) ¯⊗U ) ⊆ B(H1, H2) ¯⊗ψ(U ). (φ ⊗ ψ)(F(V, U )) = (φ ⊗ ψ)((V ¯⊗B(K1, K2)) ∩ (B(H1, H2) ¯⊗U )) ⊆ (φ(V) ¯⊗B(K1, K2)) ∩ (B(H1, H2) ¯⊗ψ(U )) = F(φ(V), ψ(U )). Now suppose that φ and ψ are idempotents that leave V and U , respectively, invariant. Then, clearly, F(φ(V), ψ(U )) ⊆ F(V, U ). On the other hand, if T ∈ F(φ(V), ψ(U )) = (φ(U ) ¯⊗B(K1, K2)) ∩ (B(H1, H2) ¯⊗ψ(V)) then φ ⊗ ψ(T ) = (id ⊗ψ)(φ ⊗ id)(T ) = T, and hence T = φ ⊗ ψ(T ) ∈ (φ ⊗ ψ)(F(V, U )). Thus, F(φ(V), ψ(U )) ⊆ φ ⊗ ψ)(F(V, U )); the converse inclusion follows from the previous paragraph. (cid:3) Suppose that Hk, k ∈ N, are Hilbert spaces and let H = ⊕k∈NHk. Then every operator T ∈ B(H) has an operator matrix representation T = (Ti,j), where Ti,j ∈ B(Hj, Hi). Given a family X = (Xi,j)i,j∈N, where Xi,j ⊆ B(Hj, Hi) is a weak* closed subspace, we let eX = {T = (Ti,j) ∈ B(H) : Ti,j ∈ Xi,j, i, j ∈ N}. It can be readily verified that eX is weak* closed. Moreover, it follows directly from its definition that if Y = (Yi,j)i,j∈N is another such family, Zi,j = Xi,j ∩ Yi,j and Z = (Zi,j)i,j∈N, then (2) eZ = eX ∩ eY. 8 G.K. ELEFTHERAKIS AND I.G. TODOROV Lemma 3.4. Let H k and U ⊆ B(K1, K2) be a weak* closed subspace. 2 , k ∈ N, be Hilbert spaces, H1 = ⊕k∈NH k 1 , H k 1 , H2 = ⊕k∈NH k 2 , (i) If Vi,j ⊆ B(H j 1, H i 2) is a weak* closed subspace, i, j ∈ N, V = (Vi,j)i,j∈N and P = (F(Vi,j, U ))i,j∈N, then F(eV, U ) = eP. (ii) If Vk is a weak* closed subspace of B(H k ⊕k∈NF(Vk, U ). 1 , H k 2 ), k ∈ N, then F(⊕k∈NVk, U ) = Proof. (i) Let ǫi,j be the evaluation at the (i, j)-entry of a matrix (Xl,m)l,m∈N. We identify Hi ⊗ Ki with ⊕∞ i ⊗ Ki), i = 1, 2. We claim that, for every τ ∈ B(K1, K2)∗, we have k=1(H k (3) Lτ (ǫi,j(T )) = ǫi,j(Lτ (T )), T ∈ B(H1 ⊗ K1, H2 ⊗ K2). Indeed, if T = A ⊗ B, where A = (Ai,j)i,j∈N ∈ B(H1, H2) and B ∈ B(K1, K2), then Lτ (ǫi,j(T )) = τ (B)Ai,j = ǫi,j(τ (B)A) = ǫi,j(Lτ (T )), and the general case follows by linearity and weak* continuity. A similar argument shows that, for every ω ∈ B(H j 2)∗, we have 1, H i (4) Rω(ǫi,j(T )) = Rω◦ǫi,j (T ), T ∈ B(H1 ⊗ K1, H2 ⊗ K2). every τ ∈ B(K1, K2)∗, while (4) shows that Rω(Ti,j) ∈ U for every ω ∈ B(H j Now suppose that T = (Ti,j) ∈ F(eV, U ). Then (3) shows that Lτ (Ti,j) ∈ Vi,j for i, j ∈ N. It follows that Ti,j ∈ F(Vi,j, U ), i, j ∈ N; thus, F(eV, U ) ⊆ eP. Conversely, suppose that T ∈ eP. Then Ti,j ∈ F(Vi,j, U ) for every i and j. Given τ ∈ B(K1, K2)∗, (3) implies that Lτ (T ) ∈ eV. On the other hand, letting (Ei,j)i,j be the standard matrix unit system, (4) shows that, if ω ∈ B(H1, H2)∗ and ωi,j ∈ B(H j 2)∗ is given by ωi,j(A) = ω(A ⊗ Ei,j), then 1, H i 1 , H i 2)∗, NXi,j=1 Rωi,j (ǫi,j(T )) ∈ U ; NXi,j=1 Rω(T ) = lim N→∞ Rω(ǫi,j(T ) ⊗ Ei,j) = lim N→∞ thus, T ∈ F(eV, U ). every i, j and k, then eV and ⊕k∈NVk do so as well. follows from [17, Proposition 1.11]. (ii) is a special case of (i) obtained by letting Vi,j = {0} if i 6= j. (cid:3) Corollary 3.5. In the notation of Lemma 3.4, if Vi,j and Vk have property Sσ for We note that the conclusion regarding the direct sum in the last corollary also It was shown in [17] that every von Neumann algebra with abelian commutant possesses property Sσ. We will shortly show that the same holds for ternary masa- bimodules. We first need a lemma. Lemma 3.6. Assume that H1 = ℓ2, (ei)i∈N is its standard orthonormal basis, and H2 is a Hilbert space. Let Pi be the rank one projection whose range is spanned by ei, and Qi be a projection on H2, i ∈ N. The space V = {T ∈ B(H1, H2) : T Pi = QiT Pi, i ∈ N} OPERATOR SYNTHESIS AND TENSOR PRODUCTS 9 possesses property Sσ. P∞ Proof. Let U ⊆ B(K1, K2) be a weak* closed subspace and T ∈ F(V, U ). Then T = i=1 T (Pi ⊗ I) (where the series converges in the weak* topology); it hence suffices to show that T (Pi⊗I) ∈ V ¯⊗U for each i. However, since T ∈ V ¯⊗B(K1, K2), we have that T (Pi ⊗ I) = (Qi ⊗ I)T (Pi ⊗ I). But (Qi ⊗ I)T (Pi ⊗ I) ∈ (QiB(H1, H2)Pi) ¯⊗U , and the latter space is contained in V ¯⊗U . It follows that T ∈ V ¯⊗U . (cid:3) Theorem 3.7. Let D1 ⊆ B(H1) and D2 ⊆ B(H2) be masas and M ⊆ B(H1, H2) be a ternary D2, D1-bimodule. Then M possesses property Sσ. If, moreover, M does not contain subspaces of the form B(EH1, F K1), where E ∈ D1 and F ∈ D2 are non-zero non-atomic projections, then every masa-bimodule V with V ⊆ M possesses property Sσ. Proof. We have (see, e.g., [29]) that, up to unitary equivalence, M =(cid:0)⊕m j=1Mlj ,kj (D)(cid:1) ⊕(cid:16)⊕l j=1B(EkH1, FkH2)(cid:17) , where m, l ∈ N ∪ {∞}, D is the multiplication masa of L∞(0, 1) acting on L2(0, 1) and Ek (resp. Fk) is a projection in D1 (resp. D2). Since D possesses Sσ [17], Corollary 3.5 implies that M does so as well. 2 , where M0 ⊆ ⊕m Suppose that M does not contain subspaces of the form B(EH1, F K1), where E ∈ D1 and F ∈ D2 are non-zero non-atomic projections. Then, for each k, either Ek is totally atomic or Fk is such. Let V ⊆ M be a masa-bimodule. Then V = M0 ⊕ W1 ⊕ W ∗ j=1Mlj ,kj (D), while W1 and W2 have the form described in Lemma 3.6. We have that M0 = ⊕m p,q ∈ Dj p,q is itself a (continuous) masa and hence possesses property Sσ [17]. It follows from Corollary 3.5 that M0 has Sσ. On the other hand, every masa-bimodule contained in W1 or W ∗ 2 has the form described in Lemma 3.6. By Corollary 3.5 and Lemma 3.6, V has property Sσ. (cid:3) p,q) : Aj p,q ⊆ D, for all p = 1, . . . , lj and q = 1, . . . , kj. We have that Dj j=1Uj , where Uj = {(Aj p,q}, and Dj Lemma 3.8. Let M, V ⊆ B(H1, H2) be masa-bimodules with M ternary, and let U ⊆ B(K1, K2) be a weak* closed subspace. Then (M ¯⊗U ) ∩ (V ¯⊗U ) = (M ∩ V) ¯⊗U . Proof. Up to unitary equivalence, M = (⊕m j=1Mlj ,kj (D)) ⊕(cid:16)⊕k i=1B(EiH1, FiH2)(cid:17) , i=1 and (Fi)k where k, m ∈ N ∪ {∞}, (Ei)k i=1 are families of mutually orthogonal projections belonging to the corresponding masas, while D is a continuous masa. Let Ma = ⊕k j=1Mlj ,kj (Dj) be its continuous part, both naturally identified with subspaces of M. We have a natural identification i=1B(EiH1, FiH2) be the atomic part of M and Mc = ⊕m M ¯⊗U = (Ma ¯⊗U ) + (Mc ¯⊗U ) and (M ¯⊗U ) ∩ (V ¯⊗U ) = ((Ma ¯⊗U ) ∩ (V ¯⊗U )) + ((Mc ¯⊗U ) ∩ (V ¯⊗U )) 10 G.K. ELEFTHERAKIS AND I.G. TODOROV (for the second identity, we use the fact that, if θ is the Schur idempotent given by θ(X) = QXP , where P = ∨iEi and Q = ∨iFi, then θ leaves V ¯⊗U invariant and maps M ¯⊗U onto Ma ¯⊗U ). Let φ be the map on B(H1, H2) given by φ(X) = P∞ contractive Schur idempotent and so, by Lemma 3.1, we have i=1 FiXEi. Then φ is a (Ma ¯⊗U ) ∩ (V ¯⊗U ) = (Ran φ ¯⊗U ) ∩ (V ¯⊗U ) = φ(V) ¯⊗U = (Ma ∩ V) ¯⊗U ⊆ (M ∩ V) ¯⊗U . Suppose that T ∈ (Mc ¯⊗U ) ∩ (V ¯⊗U ). Then Lτ (T ) ∈ Mc ∩ V and Rω(T ) ∈ U for all τ ∈ B(K1, K2)∗ and all ω ∈ B(H1, H2)∗. Thus, T ∈ F(Mc ∩ V, U ) = (Mc ∩ V) ¯⊗U ⊆ (M ∩ V) ¯⊗U , where the equality follows from Theorem 3.7, applied to the masa-bimdule Mc. We hence showed that (M ¯⊗U ) ∩ (V ¯⊗U ) ⊆ (M ∩ V) ¯⊗U ; the converse inclusion is trivial. (cid:3) Lemma 3.9. Let (Mn)n∈N be a descending sequence of ternary masa-bimodules in B(H1, H2) and M = ∩n∈NMn. If U ⊆ B(K1, K2) is a weak* closed subspace then ∩n∈N(Mn ¯⊗U ) = M ¯⊗U . Proof. The inclusion M ¯⊗U ⊆ ∩n∈N(Mn ¯⊗U ) is trivial. Suppose that T ∈ Mn ¯⊗U for each n. Then Lτ (T ) ∈ Mn for all n, and so Lτ (T ) ∈ M, for all τ ∈ B(K1, K2)∗. On the other hand, Rω(T ) ∈ U for all ω ∈ B(H1, H2)∗. It follows that T ∈ F(M, U ) and since M possesses property Sσ (Theorem 3.7), we conclude that T ∈ M ¯⊗U . (cid:3) We are now ready to prove the main result of this section. Theorem 3.10. Let V ⊆ B(H1, H2) be a masa-bimodule, B ⊆ B(H1, H2) be a masa-bimodule of finite width and U ⊆ B(K1, K2) be a weak* closed subspace. If F(V, U ) = V ¯⊗U then F(V + B, U ) = V + B ¯⊗U . In particular, if V possesses property Sσ then V + B does so as well. Proof. We use induction on the length k of B. If k = 0, that is, B = {0}, the statement is trivial. Suppose that it holds for masa-bimodules of length at most k, let B be a masa-bimodule of length k, and let A be a nest algebra bimodule. Let (φn, ψn, Mn, Wn)n∈N be a decomposition scheme for A, and set θn = id −(φn +ψn). Suppose T ∈ F(V + (B ∩ A), U ) and write T = φn(T ) + ψn(T ) + θn(T ). Since A ⊆ Ran(φn + ψn), we have that θn(A) = {0} and hence, by Lemma 3.3, θn(T ) ∈ F(V, U ) = V ¯⊗U . By Lemma 3.3 and the fact that ψn(B) = Ran ψn ∩ B, we have ψn(T ) ∈ F(ψn(V + (B ∩ A)), U ) = F(ψn(V) + ψn(B ∩ A), U ) = F(ψn(V) + ψn(B), U ) = ψn(F(V + B, U )) = ψn(V + B ¯⊗U ) = ψn(V + B) ¯⊗U ⊆ V + ψn(B) ¯⊗U ) ⊆ V + (B ∩ A) ¯⊗U . OPERATOR SYNTHESIS AND TENSOR PRODUCTS 11 On the other hand, by Lemmas 3.3 and 3.8 and the facts that φn(V) = V ∩ Mn and φn(B) = B ∩ Mn, we have φn(T ) ∈ F(φn(V) + (φn(B) ∩ Mn), U ) = (V ∩ Mn + B ∩ Mn ¯⊗B(K1, K2)) ∩ (B(H1, H2) ¯⊗U ) = (V + B ∩ Mn ¯⊗B(K1, K2)) ∩ (B(H1, H2) ¯⊗U ) = (V + B ¯⊗B(K1, K2)) ∩ (Mn ¯⊗B(K1, K2)) ∩ (B(H1, H2) ¯⊗U ). Let S be a weak* cluster point of ( φn(T ))n∈N and set M = ∩n∈NMn. Using Lemmas 3.8, 3.9 and the inductive assumption, we have S ∈ (V + B ¯⊗B(K1, K2)) ∩ (M ¯⊗B(H1, H2)) ∩ (B(H1, H2) ¯⊗U ) = F(V + B, U ) ∩ F(M, U ) = (M ¯⊗U ) ∩ (V + B ¯⊗U ) = (M ∩ V + B) ¯⊗U . Every ternary masa-bimodule has finite width (in fact, it is the intersection of two nest algebra bimodules [16]) and hence, by [8, Theorem 2.10], we have that M ∩ V + B ⊆ V + M ∩ V + B = V + (B ∩ M) ⊆ V + (B ∩ A). It follows that S ∈ V + (B ∩ A) ¯⊗U . On the other hand, by the second paragraph of the proof, T − S = lim n→∞ ( ψn(T ) + θn(T )) ∈ V + (B ∩ A) ¯⊗U . Hence, T = S + (T − S) ∈ V + (B ∩ A) ¯⊗U , and the proof is complete. (cid:3) The next corollary is an immediate consequence of Theorem 3.10. It extends the fact, established in [14], that CSL algebras of finite width possess property Sσ. Corollary 3.11. If Bi, i = 1, . . . , n, are masa-bimodules of finite width, then B1 + · · · + Bn has property Sσ. 4. Intersections and spans In this section, we establish an intersection formula for weak* closures of spans of subspaces of the form B ¯⊗U , where B is a masa-bimodule of finite width (see Theorem 4.4 and Corollary 4.21). This result will be used in Section 5 to study questions about operator synthesis. We fix masas D1 ⊆ B(H1) and D2 ⊆ B(H2). All Schur idempotents we consider are relative to the pair (D1, D2) and act on B(H1, H2). We will say that a sequence (ψn)n∈N of Schur idempotents is nested if Ran ψn+1 ⊆ Ran ψn for all n ∈ N. Before formulating the main result of this section, Theorem 4.4, we state three propositions which will be needed in its proof. We first recall that the ranges of contractive Schur idempotents on B(H1, H2) are ternary masa-bimodules of the form ⊕kB(EkH1, FkH2), where (Ek)k ⊆ D1 and (Fk)k ⊆ D2 are families of mutually orthogonal projections (see, e.g., [15]). 12 G.K. ELEFTHERAKIS AND I.G. TODOROV Notation. For the rest of this section, we let Bi ⊆ B(H1, H2) be a masa-bimodule of finite width, U , V, Ui be weak* closed subspaces of B(K1, K2), i = 1, . . . , r, and i=1 Bi ¯⊗Ui. W =Pr Proposition 4.1. Let (ψi)i∈N be a nested sequence of contractive Schur idempo- tents. (i) Let (φi)i∈N be a nested sequence of contractive Schur idempotents. Then the subspaces ∩k,i(Ran ψi ∩ Ran φk) ¯⊗U + Ran φk ¯⊗V + W and ((∩i Ran ψi) ∩ (∩k Ran φk)) ¯⊗U + (∩k Ran φk) ¯⊗V + W coincide. (ii) Let B be a nest algebra bimodule. Then ∩iRan ψi ¯⊗U + B ¯⊗V + W = (∩i Ran ψi) ¯⊗U + B ¯⊗V + W. (iii) Let B be a nest algebra bimodule. Then the subspaces ∩iRan ψi ¯⊗U + (Ran ψi ∩ B) ¯⊗V + W and (∩i Ran ψi) ¯⊗U + ((∩i Ran ψi) ∩ B) ¯⊗V + W coincide. (iv) Let B be a masa-bimodule of finite width. Then the subspaces ∩iRan ψi ¯⊗U + (Ran ψi ∩ B) ¯⊗V + W and (∩i Ran ψi) ¯⊗U + ((∩i Ran ψi) ∩ B) ¯⊗V + W coincide. (v) ∩iRan ψi ¯⊗U + W = (∩i Ran ψi) ¯⊗U + W. We note that part (ii) of the previous proposition is more general than (v); how- ever, for the purpose of its proof it will be convenient to formulate these statements separately. Proposition 4.2. Let M be a ternary masa-bimodule and C be a nest algebra bimodule. Then M ¯⊗U + W ∩ C ¯⊗U + W = (M ∩ C) ¯⊗U + W. Proposition 4.3. Let M be a ternary masa-bimodule and C be a masa-bimodule of finite width. Then M ¯⊗U + W ∩ C ¯⊗U + W = (M ∩ C) ¯⊗U + W. Theorem 4.4. Let Cj ⊆ B(H1, H2) be a masa-bimodule of finite width, j = 1, . . . , m. Then ∩m j=1Cj ¯⊗U + W = (∩m j=1Cj) ¯⊗U + W. OPERATOR SYNTHESIS AND TENSOR PRODUCTS 13 on the number r of terms in the sum W = Pr The proof of the above results will be given simultaneously, using induction i=1 Bi ¯⊗Ui and will be split into a number of lemmas. The first series of steps, namely Lemmas 4.5 -- 4.12, provide the base of the induction. We will refer to the statements in Proposition 4.1 by their corresponding numbers (i) -- (v). It will be convenient to assume that W = {0} when r = 0. Given the notation in Proposition 4.1, throughout the proofs, we will set for brevity N = ∩i Ran ψi and R = ∩k Ran φk. The proofs of the lemmas in this section will all use the following idea: Let Ω be a weak* closed subspace of operators and (ρn)n∈N be a nested sequence of contractive idempotents with ∩∞ n=1 Ran ρn ⊆ Ω. In order to prove that a certain operator T belongs to Ω, it suffices to show that ρ⊥ n (T ) ∈ Ω for each n ∈ N. Indeed, letting S be a weak* cluster point of the sequence (ρn(T ))n∈N, we have that S ∈ ∩∞ n (T ), n ∈ N, show that T − S is a weak* cluster point of the sequence (ρ⊥ n (T ))n∈N, and hence it belongs to Ω; therefore, T = S + (T − S) ∈ Ω. n=1 Ran ρn ⊆ Ω. On the other hand, the identities T = ρn(T ) + ρ⊥ Lemma 4.5. Proposition 4.1 (i) holds if r = 0. Proof. Let and fix Ω = (N ∩ R) ¯⊗U + R ¯⊗V X ∈ ∩k,i(Ran ψi ∩ Ran φk) ¯⊗U + Ran φk ¯⊗V. By Lemma 3.1, ψ⊥ i (X) ∈ Ran φk ¯⊗V for all k, i. By Lemma 3.9, ψ⊥ i ∈ N. On the other hand, for all i ∈ N, we have, by Lemma 3.1, i (X) ∈ R ¯⊗V ⊆ Ω, ψi(X) ∈ (Ran ψi ∩ Ran φi) ¯⊗U + (Ran ψi ∩ Ran φi) ¯⊗V = (Ran ψi ∩ Ran φi) ¯⊗(U + V). By Lemma 3.9, any weak* cluster point S of the sequence ( ψi(X))i∈N belongs to (N ∩ R) ¯⊗(U + V), a subset of Ω. Thus, X = (X − S) + S ∈ Ω. (cid:3) Lemma 4.6. Proposition 4.1 (ii) holds if r = 0. Proof. Let Ω = N ¯⊗U + B ¯⊗V and fix X ∈ ∩iRan ψi ¯⊗U + B ¯⊗V. Let (φk, θk, Mk, Zk)k∈N be a decomposition scheme for B. By Lemma 3.1, φk(X) ∈ ∩i(Ran ψi ∩ Mk) ¯⊗U + Mk ¯⊗V. If S is a weak* cluster point of the sequence ( φk(X))k∈N, then S ∈ ∩k,i(Ran ψi ∩ Mk) ¯⊗U + Mk ¯⊗V. By Lemma 4.5, S ∈ Ω. It hence suffices to prove that φ⊥ Observe that φ⊥ k (X) ∈ ∩iRan ψi ¯⊗U + Zk ¯⊗V. k (X) ∈ Ω for all k ∈ N. 14 G.K. ELEFTHERAKIS AND I.G. TODOROV Using Lemmas 3.1 and 3.9, we see that k ( φ⊥ θ⊥ k (X)) ∈ ∩i(Ran ψi ¯⊗U ) = N ¯⊗U ⊆ Ω. Write θk =Pk By Lemma 4.5, p=1 θk,p, where θk,p is a contractive Schur idempotent whose range is contained in B (see the last paragraph of Section 2). Then θk,p( φ⊥ k (X)) ∈ ∩i(Ran ψi ∩ Ran θk,p) ¯⊗U + Ran θk,p ¯⊗V. θk,p( φ⊥ k (X)) ∈ N ¯⊗U + Ran θk,p ¯⊗V ⊆ Ω. Hence, φ⊥ is complete. k (X) = θ⊥ k ( φ⊥ k (X)) +Pk p=1 θk,p( φ⊥ k (X)) ∈ Ω for all k ∈ N and the proof (cid:3) Lemma 4.7. Proposition 4.1 (iii) holds if r = 0. Proof. Let Ω = N ¯⊗U + (N ∩ B) ¯⊗V and fix X ∈ ∩iRan ψi ¯⊗U + (Ran ψi ∩ B) ¯⊗V. Let (φk, θk, Mk, Zk)k∈N be a decomposition scheme for B. By Lemma 3.1, φ⊥ k (X) ∈ ∩iRan ψi ¯⊗U + (Ran ψi ∩ Zk) ¯⊗V. By Lemmas 3.1 and 3.9, k ( φ⊥ θ⊥ k (X)) ∈ ∩i(Ran ψi ¯⊗U ) = N ¯⊗U ⊆ Ω, and by Lemmas 3.1, 3.2 and 3.9, θk( φ⊥ k (X)) ∈ ∩i(cid:0)(Ran ψi ∩ Zk) ¯⊗U + V(cid:1) ⊆ (cid:0)∩i(cid:0)Ran ψi ¯⊗U + V(cid:1)(cid:1) ∩(cid:0)Zk ¯⊗U + V(cid:1) = (N ¯⊗U + V) ∩ (Zk ¯⊗U + V) = (N ∩ Zk) ¯⊗U + V ⊆ Ω; thus, φ⊥ k (X) ∈ Ω. On the other hand, by Lemma 3.1, φk(X) ∈ (Ran ψk ∩ Ran φk) ¯⊗U + V, k ∈ N. Therefore, if S is a weak* cluster point of the sequence ( φk(X))k∈N, then, by Lemma 3.9, S ∈ (N ∩ R) ¯⊗U + V ⊆ Ω. The proof is complete. (cid:3) Lemma 4.8. Proposition 4.1 (iv) holds if r = 0. Proof. Let Ω = N ¯⊗U + (N ∩ B) ¯⊗V. We use induction on the width n of B. If n = 1, the conclusion follows from Lemma 4.7. Suppose that the statement holds for masa-bimodules of width at most n − 1 and let B = ∩n l=1Al, where Al is a nest algebra bimodule, l = 1, . . . , n. Fix X ∈ ∩iRan ψi ¯⊗U + (Ran ψi ∩ B) ¯⊗V. OPERATOR SYNTHESIS AND TENSOR PRODUCTS 15 By the inductive assumption, X belongs to both and N ¯⊗U + (N ∩ (∩n−1 l=1 Al)) ¯⊗V N ¯⊗U + (N ∩ An) ¯⊗V. Let (φk, θk, Mk, Zk)k∈N be a decomposition scheme for An. For a fixed k, we have that φ⊥ k (X) belongs to the intersection of the spaces N ¯⊗U + (N ∩ (∩n−1 l=1 Al)) ¯⊗V and N ¯⊗U + (N ∩ Zk) ¯⊗V. By Lemma 3.1, k ( φ⊥ θ⊥ k (X)) ∈ N ¯⊗U ⊆ Ω and θk( φ⊥ since Zk ⊆ An. Thus, k (X)) ∈ N ¯⊗U + (N ∩ (∩n−1 i=1 Al) ∩ Zk) ¯⊗U ⊆ Ω, φ⊥ k (X) = θk( φ⊥ k (X)) + θ⊥ k ( φ⊥ k (X)) ∈ Ω. Let S be a weak* cluster point of ( φk(X))k∈N. Then S ∈ ∩k((Ran ψk ∩ Ran φk) ¯⊗U + (Ran ψk ∩ Ran φk ∩ (∩n−1 l=1 Al)) ¯⊗V). By the inductive assumption, the latter space coincides with l=1 Al)) ¯⊗V), (N ∩ R) ¯⊗U + (N ∩ R) ∩ (∩n−1 which is a subset of Ω since R ⊆ An. Lemma 4.9. Proposition 4.1 (v) holds if r = 1. (cid:3) Proof. Let Ω = N ¯⊗U + W. We set B = B1 and V = U1, and use induction on the width n of B. For n = 1 the conclusion follows from Lemma 4.6. Suppose that the statement holds if the length of B does not exceed n−1 and assume that B = ∩n l=1Al where Al is a nest algebra bimodule, l = 1, . . . , n. Let (φk, θk, Mk, Zk)k∈N be a decomposition scheme for An. Fix By the inductive assumption, X belongs to the intersection of the spaces X ∈ ∩iRan ψi ¯⊗U + B ¯⊗V. N ¯⊗U + (∩n−1 By Lemma 3.1, for a fixed k, φ⊥ N ¯⊗U , and hence θ⊥ k ( φ⊥ θk( φ⊥ k (X)) ∈ Ω. On the other hand, k (X)) ∈ N ¯⊗U + ((∩n−1 l=1 Al) ∩ Zk) ¯⊗V ⊆ Ω l=1 Al) ¯⊗V and N ¯⊗U + An ¯⊗V. k (X) ∈ N ¯⊗U + Zk ¯⊗V. Therefore, θ⊥ k ( φ⊥ k (X)) ∈ since Zk ⊆ An. Thus, φ⊥ k (X) ∈ Ω, for all k ∈ N. It hence suffices to prove that, if S is the limit of a subsequence ( φkl(X))l∈N, then S ∈ Ω. Let S′ be a weak* cluster point of the sequence ( ψkl( φkl(X)))l∈N; then S′′ = S − S′ is a weak* cluster point of ( ψ⊥ kl ( φkl(X)))l∈N. By Lemma 3.1, ψk( φk(X)) ∈ (Mk ∩ Ran ψk) ¯⊗U + ((∩n−1 l=1 Al) ∩ (Mk ∩ Ran ψk)) ¯⊗V, 16 while G.K. ELEFTHERAKIS AND I.G. TODOROV k ( φk(X)) = φk( ψ⊥ ψ⊥ k (X)) ∈ B ¯⊗V ⊆ Ω, k ∈ N. On the other hand, Lemma 4.8 implies that S′ ∈ ((∩kMk) ∩ N ) ¯⊗U + ((∩n−1 l=1 Al) ∩ (∩kMk) ∩ N ) ¯⊗V ⊆ Ω. Thus, S = S′ + S′′ ∈ Ω. Lemma 4.10. Proposition 4.2 holds if r = 1. (cid:3) Proof. Set B = B1 and V = U1, let Ω = (M ∩ C) ¯⊗U + B ¯⊗V and fix X ∈ M ¯⊗U + B ¯⊗V ∩ C ¯⊗U + B ¯⊗V. Let (ψk, θk, Ran ψk, Zk)k∈N be a decomposition scheme for C. We have that ψ⊥ k (X) ∈ M ¯⊗U + B ¯⊗V ∩ Zk ¯⊗U + B ¯⊗V. By Lemma 3.2, ψ⊥ k (X) ∈ (M ∩ Zk) ¯⊗U + B ¯⊗V ⊆ (M ∩ C) ¯⊗U + B ¯⊗V. On the other hand, by Lemmas 3.1 and 3.2, ψk(X) ∈ (Mk ∩ Ran ψk) ¯⊗U + B ¯⊗V for all k ∈ N. Lemma 4.9 shows that, if S is a weak* cluster point of the sequence ( ψk(X))k∈N, then S ∈ (∩k(Mk ∩ Ran ψk)) ¯⊗U + B ¯⊗V. Since ∩k(Mk ∩ Ran ψk) ⊆ M ∩ C, we conclude that S ∈ Ω. The proof is complete. (cid:3) Lemma 4.11. Proposition 4.3 holds if r = 1. Proof. Write B = B1 and V = U1. We use induction on the width n of C. If n = 1, the statement reduces to Lemma 4.10. Suppose that the statement holds for all masa-bimodules C of width at most n − 1 and let C = ∩n l=1Cl, where Cl is a nest algebra bimodule, l = 1, . . . , n. Fix X ∈ M ¯⊗U + B ¯⊗V ∩ C ¯⊗U + B ¯⊗V. Let (ψk, θk, Ran ψk, Zk)k∈N be a decomposition scheme for Cn and recall that N = ∩k Ran ψk. There exists a descending sequence (Mk)k∈N of ranges of contractive Schur idempotents such that M = ∩k∈NMk. Observe that, by the inductive as- sumption, ψ⊥ k (X) ∈ M ¯⊗U + B ¯⊗V ∩ (∩n−1 l=1 Cl) ¯⊗U + B ¯⊗V ∩ Zk ¯⊗U + B ¯⊗V = (M ∩ (∩n−1 l=1 Cl) ¯⊗U + B ¯⊗V ∩ (Zk ¯⊗U + B ¯⊗V). By Lemma 3.2, Observe that ψ⊥ k (X) ∈ (M ∩ C) ¯⊗U + B ¯⊗V. ψk(X) ∈ Mk ¯⊗U + B ¯⊗V ∩ (∩n−1 l=1 Cl) ¯⊗U + B ¯⊗V ∩ Ran ψk ¯⊗U + B ¯⊗V OPERATOR SYNTHESIS AND TENSOR PRODUCTS 17 and so, by Lemma 3.2, ψk(X) ∈ (Mk ∩ Ran ψk) ¯⊗U + B ¯⊗V ∩ (∩n−1 l Cl) ¯⊗U + B ¯⊗V, k ∈ N. By Lemma 4.9, if S is a weak* cluster point of ( ψk(X))k∈N, then S ∈ (M ∩ N ) ¯⊗U + B ¯⊗V ∩ (∩n−1 l=1 Cl) ¯⊗U + B ¯⊗V. By the inductive assumption, S ∈ (M ∩ N ∩ (∩n−1 l=1 Cl)) ¯⊗U + B ¯⊗V ⊆ (M ∩ C) ¯⊗U + B ¯⊗V, since N ⊆ Cn. The proof is complete. Lemma 4.12. Theorem 4.4 holds if r = 1. (cid:3) Proof. Since each masa-bimodule of finite width is the finite intersection of nest algebra masa-bimodules, we may assume, without loss of generality, that Cj is a nest algebra bimodule, j = 1, . . . , m. Set B = B1, V = U1 and C = ∩m j=1Cj. We use induction on m. For m = 1, the statement is trivial; suppose it holds if the j=1Cj ¯⊗U + B ¯⊗V. Let number of given bimodules is at most m − 1 and fix X ∈ ∩m (φk, θk, Mk, Zk)k∈N be a decomposition scheme for Cm and set M = ∩k∈NMk. Since X ∈ Cm ¯⊗U + W, Lemma 3.1 implies that By the inductive assumption and the invariance of ∩m−1 j=1 Cj under φk, we have that φ⊥ k (X) ∈ Zk ¯⊗U + W. Hence, by Lemma 3.2, φ⊥ k (X) ∈ (∩m−1 j=1 Cj) ¯⊗U + W. φ⊥ k (X) ∈ ((∩m−1 i=1 Ci) ∩ Zk) ¯⊗U + B ¯⊗V ⊆ C ¯⊗U + B ¯⊗V, k ∈ N. On the other hand, φk(X) ∈ Mk ¯⊗U + B ¯⊗V ∩ (∩m−1 j=1 Cj) ¯⊗U + B ¯⊗V, k ∈ N. If S is a weak* cluster point of the sequence ( φk(X))k∈N, by Lemma 4.9 we have S ∈ M ¯⊗U + B ¯⊗V ∩ (∩m−1 j=1 Cj) ¯⊗U + B ¯⊗V. By Lemma 4.11, S ∈ (M ∩ (∩m−1 i=1 Ci)) ¯⊗U + B ¯⊗V ⊆ C ¯⊗U + B ¯⊗V. (cid:3) We next establish the induction step for the proofs of Propositions 4.1 -- 4.3 and Theorem 4.4; this is done in Lemmas 4.13 -- 4.20 below. To this end, we assume that the statements in Proposition 4.1 (i) -- (iv) hold if the space W has r − 1 summands, while Proposition 4.1 (v), Proposition 4.2, Proposition 4.3 and Theorem 4.4 hold if W has r summands. Lemma 4.13. Proposition 4.1 (i) holds if the space W has r terms. 18 G.K. ELEFTHERAKIS AND I.G. TODOROV Proof. Let and fix Ω = (N ∩ R) ¯⊗U + R ¯⊗V + W X ∈ ∩k,i(Ran ψi ∩ Ran φk) ¯⊗U + Ran φk ¯⊗V + W. i (X) ∈ Ran φk ¯⊗V + W for all k, i ∈ N. By the inductive assump- By Lemma 3.1, ψ⊥ tion concerning Proposition 4.1 (v), On the other hand, for all k ∈ N we have ψ⊥ i (X) ∈ R ¯⊗V + W ⊆ Ω, i ∈ N. ψk(X) ∈ (Ran ψk ∩ Ran φk) ¯⊗U + (Ran ψk ∩ Ran φk) ¯⊗V + W = (Ran ψk ∩ Ran φk) ¯⊗(U + V) + W. Let S be a weak* cluster point of ( ψk(X))k∈N. Once again by the inductive as- sumption concerning Proposition 4.1 (v), S ∈ (N ∩ R) ¯⊗(U + V) + W. Thus, S ∈ Ω and the proof is complete. (cid:3) Lemma 4.14. Proposition 4.1 (ii) holds if the space W has r terms. Proof. Let and fix Ω = N ¯⊗U + B ¯⊗V + W X ∈ ∩iRan ψi ¯⊗U + B ¯⊗V + W. Let (φk, θk, Ran φk, Zk)k∈N be a decomposition scheme for B and observe that, by Lemma 3.1, we have φk(X) ∈ ∩i(Ran ψi ∩ Ran φk) ¯⊗U + Ran φk ¯⊗V + W. Letting S be a weak* cluster point of the sequence ( φk(X))k∈N, we have that S ∈ ∩k,i(Ran ψi ∩ Ran φk) ¯⊗U + Ran φk ¯⊗V + W. By Lemma 4.13 and the fact that R = ∩k Ran φk ⊆ B, we have that S ∈ Ω. So it suffices to prove that φ⊥ k (X) ∈ Ω for all k ∈ N. Note that, by Lemma 3.1, φ⊥ k (X) ∈ ∩iRan ψi ¯⊗U + Zk ¯⊗V + W. By the inductive assumption concerning Proposition 4.1 (v) and Lemma 3.1 again, k ( φ⊥ θ⊥ k (X)) ∈ ∩iRan ψi ¯⊗U + W = N ¯⊗U + W ⊆ Ω. p=1 θk,p, where each θk,p is a contractive Schur idempotent whose Write θk = Pk range is contained in B. We have that θk,p( φ⊥ k (X)) ∈ ∩i(Ran ψi ∩ Ran θk,p)) ¯⊗U + (Ran θk,p) ¯⊗V + W. By Lemma 4.13 (applied in the case of a constant sequence of maps with term θk,p), we have θk,p( φ⊥ k (X)) ∈ N ¯⊗U + (Ran θk,p) ¯⊗V + W ⊆ Ω, p = 1, . . . , k. OPERATOR SYNTHESIS AND TENSOR PRODUCTS 19 It follows that θk( φ⊥ Ω. k (X)) ∈ Ω and hence φ⊥ k (X) = θk( φ⊥ k (X)) + θ⊥ k ( φ⊥ k (X)) ∈ (cid:3) Lemma 4.15. Proposition 4.1 (iii) holds if the space W has r terms. Proof. We let and fix Ω = N ¯⊗U + (N ∩ B) ¯⊗V + W X ∈ ∩iRan ψi ¯⊗U + (Ran ψi ∩ B) ¯⊗V + W. Let (φk, θk, Mk, Zk)k∈N be a decomposition scheme for B and observe that φ⊥ k (X) ∈ ∩iRan ψi ¯⊗U + (Ran ψi ∩ Zk) ¯⊗V + W. Using the inductive assumption concerning Proposition 4.1 (v), we have θ⊥ k ( φ⊥ k (X)) ∈ ∩iRan ψi ¯⊗U + W = N ¯⊗U + W ⊆ Ω. p=1 θk,p as in the proof of Lemma 4.14; then Write θk =Pk θk,p( φ⊥ k (X)) ∈ ∩i(Ran ψi ∩ Ran θk,p) ¯⊗U + (Ran ψi ∩ Ran θk,p) ¯⊗V + W = ∩i(Ran ψi ∩ Ran θk,p) ¯⊗(U + V) + W. By the inductive assumption concerning Proposition 4.1 (v), we have that θk,p ( φ⊥ k (X) ∈ Ω. Let S be a weak* cluster point of the sequence ( φk(X))k∈N. Since k (X)) ∈ Ω for each p = 1, . . . , k. It follows that φ⊥ φk(X) ∈ ∩k(Ran ψk ∩ Mk) ¯⊗U + V + W, by the same inductive assumption once again, S ∈ (N ∩ (∩kMk)) ¯⊗U + V + W ⊆ Ω. (cid:3) Lemma 4.16. Proposition 4.1 (iv) holds if the space W has r terms. Proof. Let Ω = N ¯⊗U + (N ∩ B) ¯⊗V + W. We use induction on the width n of B. The case n = 1 reduces to Lemma 4.15. Suppose that the statement holds for masa-bimodules of width at most n − 1 and let B = ∩n l=1Cl where every Cl is nest algebra bimodule, l = 1, . . . , n. Fix X ∈ ∩iRan ψi ¯⊗U + (Ran ψi ∩ B) ¯⊗V + W. By the inductive assumption, X belongs to the intersection of N ¯⊗U + (N ∩ (∩n−1 l=1 Cl)) ¯⊗V + W and N ¯⊗U + (N ∩ Cn) ¯⊗V + W. Let (φk, θk, Mk, Zk)k∈N be a decomposition scheme for Cn. For a fixed k ∈ N, we have that φ⊥ k (X) belongs to the intersection of N ¯⊗U + (N ∩ (∩n−1 l=1 Cl)) ¯⊗V + W and N ¯⊗U + (N ∩ Zk) ¯⊗V + W. 20 G.K. ELEFTHERAKIS AND I.G. TODOROV By Lemma 3.2, i=1 Ci) ∩ Zk) ¯⊗U + W ⊆ Ω. Let S be a weak* cluster point of the sequence ( φk(X))k∈N; we have φ⊥ k (X) ∈ N ¯⊗U + (N ∩ (∩n−1 S ∈ ∩k(Ran ψk ∩ Mk) ¯⊗U + (Ran ψk ∩ Mk ∩ (∩n−1 l=1 Cl)) ¯⊗V + W. By the inductive assumption, the latter space is equal to (N ∩ (∩kMk)) ¯⊗U + (N ∩ (∩kMk) ∩ (∩n−1 l=1 Cl)) ¯⊗V + W which is contained in Ω since ∩kMk ⊆ Cn. (cid:3) Lemma 4.17. Proposition 4.1 (v) holds if the space W has r + 1 terms. Proof. Let Ω = N ¯⊗U + W. Set B = Br+1, V = Ur+1 and W0 =Pr i=1 Bi ¯⊗Ui. We use induction on the width n of B. For n = 1 the conclusion follows from the inductive assumption concerning Proposition 4.1 (ii). Assume that it holds when the length of B does not exceed n − 1, and suppose that B = ∩n l=1Cl where Cl is a nest algebra bimodule, l = 1, . . . , n. Let (φk, θk, Mk, Zk)k∈N be a decomposition scheme for Cn. Fix X ∈ ∩iRan ψi ¯⊗U + B ¯⊗V + W0. By assumption, X ∈ N ¯⊗U + (∩n−1 l=1 Cl) ¯⊗V + W0 ∩ N ¯⊗U + Cn ¯⊗V + W0. For a fixed k ∈ N, we have φ⊥ k (X) ∈ N ¯⊗U + (∩n−1 l=1 Cl) ¯⊗V + W0 ∩ N ¯⊗U + Zk ¯⊗V + W0. By Lemma 3.2, φ⊥ ( φk(X))k∈N, and S′ and S′′ be weak* cluster points of ( ψ⊥ (X)))k∈N, respectively, such that S = S′ + S′′. k (X) ∈ Ω. Let S be a weak* cluster point of the sequence k ( φk(X)))k∈N and ( ψk( φk We have It follows that φk(X) ∈ (Mk ∩ N ) ¯⊗U + ((∩n−1 l=1 Cl) ∩ Mk) ¯⊗V + W0. k ( φk(X)) ∈ ((∩n−1 ψ⊥ l=1 Cl) ∩ Mk) ¯⊗V + W0, k ∈ N, and hence, by the inductive assumption concerning Proposition 4.1 (v), S′ ∈ (∩n−1 l=1 Cl) ¯⊗V + W0 ∩ (∩kMk) ¯⊗V + W0. It follows from the inductive assumption concerning Proposition 4.3 that S′ ∈ Ω. On the other hand, S′′ ∈ ∩k(Mk ∩ Ran ψk) ¯⊗U + ((∩n−1 l=1 Cl) ∩ Mk ∩ Ran ψk) ¯⊗V + W0. Since kφkψkk ≤ 1 for each k ∈ N, Lemma 4.16 implies that S′′ ∈ Ω. It now follows that S ∈ Ω. (cid:3) Lemma 4.18. Proposition 4.2 holds if the space W has r + 1 terms. OPERATOR SYNTHESIS AND TENSOR PRODUCTS 21 Proof. Let Ω = (M ∩ C) ¯⊗U + W and fix X ∈ M ¯⊗U + W ∩ C ¯⊗U + W. Let (ψk, θk, Ran ψk, Zk)k∈N be a decomposition scheme for C. Let (φk)k∈N be a nested sequence of contractive Schur idempotents such that M = ∩∞ k=1Mk, where Mk = Ran φk, k ∈ N. We first observe that, by Lemma 3.2, θk(X) ∈ (Zk ∩ M) ¯⊗U + W ⊆ Ω, k ∈ N. On the otehr hand, by Lemma 3.1, ( ψk + θk)⊥(X) ∈ W ⊆ Ω and φ⊥ k (X) ∈ W ⊆ Ω, k ∈ N. It follows that φ⊥ point of ( φk( ψk(X)))k∈N. Since k ( ψk(X)) = ψk( φ⊥ k (X)) ∈ Ω for each k. Let S be a weak* cluster φk( ψk(X)) ∈ (Mk ∩ Ran ψk) ¯⊗U + W for all k ∈ N, we have, by Lemma 4.17, that S belongs to ∩k(Mk ∩ Ran ψk) ¯⊗U + W, which is a subset of Ω. Since X = φk( ψk(X)) + φ⊥ k ( ψk(X)) + θk(X) + ( ψk + θk)⊥(X), k ∈ N, it now follows that X ∈ Ω. (cid:3) Lemma 4.19. Proposition 4.3 holds if the space W has r + 1 terms. Proof. We use induction on the width n of C. If n = 1 the statement reduces to Lemma 4.18. Suppose that the statement holds for all masa-bimodules C of width not exceeding n − 1 and let C = ∩n l=1Cl, where Cl is a nest algebra bimodule, l = 1, . . . , n. Fix X ∈ M ¯⊗U + W ∩ C ¯⊗U + W and let (ψk, θk, Ran ψk, Zk)k∈N be a decomposition scheme for Cn. We also assume that M = ∩kMk where every (Mk)k∈N is descending sequence of ranges of con- tractive Schur idempotents. Using Lemma 3.2 and the inductive assumption, we obtain ψ⊥ k (X) ∈ M ¯⊗U + W ∩ (∩n−1 l=1 Cl) ¯⊗U + W ∩ Zk ¯⊗U + B ¯⊗V = M ∩ (∩n−1 = (M ∩ (∩n−1 l=1 Cl) ¯⊗U + W ∩ Zk ¯⊗U + W l=1 Cl) ∩ Zk) ¯⊗U + W ⊆ (M ∩ C) ¯⊗U + W. Using Lemma 3.2 again, we have ψk(X) ∈ Mk ¯⊗U + W ∩ (∩n−1 l=1 Cl) ¯⊗U + W ∩ Ran ψk ¯⊗U + W = (Mk ∩ Ran ψk) ¯⊗U + W ∩ (∩n−1 l=1 Cl) ¯⊗U + W, for every k ∈ N. By Lemma 4.17, if S is a weak* cluster point of ( ψk(X))k∈N, then S ∈ (M ∩ N ) ¯⊗U + W ∩ (∩n−1 l=1 Cl) ¯⊗U + W. By the inductive assumption, S ∈ (M ∩ N ∩ (∩n−1 l=1 Cl)) ¯⊗U + W ⊆ (M ∩ C) ¯⊗U + W. 22 G.K. ELEFTHERAKIS AND I.G. TODOROV The proof is complete. (cid:3) Lemma 4.20. Theorem 4.4 holds if the space W has r + 1 terms. Proof. It suffices to prove that if C1, . . . , Cm are weak* closed nest algebra bimodules and C = ∩m i=1Ci, then where C = ∩m i=1Ci ¯⊗U + W = C ¯⊗U + W, i=1Ci. We use induction on m. Suppose that ∩m ∩m−1 i=1 Ci ¯⊗U + W = (∩m−1 i=1 Ci) ¯⊗U + W and fix X ∈ ∩m for Cm. Using the inductive assumption and Lemmas 3.1 and 3.2, we have i=1Ci ¯⊗U + W. Let (ψk, θk, Mk, Zk)k∈N be a decomposition scheme ψ⊥ k (X) ∈ ((∩m−1 i=1 Ci) ∩ Zk) ¯⊗U + W ⊆ C ¯⊗U + W. On the other hand, ψk(X) ∈ Mk ¯⊗U + W ∩ (∩m−1 i=1 Ci) ¯⊗U + W. Thus, if S is a weak* cluster point of the sequence (ψk(X))k∈N then, by Lemma 4.17, we have that S ∈ M ¯⊗U + W ∩ (∩m−1 i=1 Ci) ¯⊗U + W. By Lemma 4.19, S ∈ (M ∩ (∩m−1 i=1 Ci)) ¯⊗U + W ⊆ C ¯⊗U + W. The proof is complete. (cid:3) Lemmas 4.5 -- 4.20 conclude the proof of Propositions 4.1 -- 4.3 and Theorem 4.4. The following statement, which is an equivalent formulation of Theorem 4.4, follows from that theorem by a straightforward induction on r. Corollary 4.21. Let r, l1, . . . , lr ∈ N, {Bi j=1 be a family of masa bimodules of finite width and Ui be a weak* closed subspace of B(K1, K2), i = 1, . . . , r. Set Bi = ∩li j, i = 1, . . . , r. Then j=1Bi j}li \j1,...,jr B1 j1 ¯⊗U1 + · · · + Br jr ¯⊗Ur = B1 ¯⊗U1 + · · · + Br ¯⊗Ur. 5. Operator synthesis of unions of products In this section we apply the results from Sections 3 and 4 to study questions about operator synthesis. We start by recalling the main definitions regarding the notion of operator synthesis. Let (X1, µ1) and (X2, µ2) be standard measure spaces, that is, the measures µ1 and µ2 are regular Borel measures with respect to some Borel structures on X1 and X2 arising from complete metrizable topologies. Let H1 = L2(X1, µ1) and H2 = L2(X2, µ2). For a function ϕ ∈ L∞(X1, µ1), let Mϕ be the (bounded) OPERATOR SYNTHESIS AND TENSOR PRODUCTS 23 operator on H1 given by Mϕf = ϕf , f ∈ L2(X1, µ1); similarly define Mψ for ψ ∈ L∞(X2, µ2). Let D1 = {Mϕ : ϕ ∈ L∞(X1, µ1)}. We have that D1 is a masa; we define D2 ⊆ B(H2) similarly. We need several facts and notions from the theory of masa-bimodules [1], [9], [25]. A subset E ⊆ X1 × X2 is called marginally null if E ⊆ (M1×X2)∪(X1×M2), where µ1(M1) = µ2(M2) = 0. We call two subsets E, F ⊆ X1 × X2 marginally equivalent (and write E ∼= F ) if the symmetric difference of E and F is marginally null. A set κ ⊆ X1 × X2 is called ω-open if it is marginally equivalent to a (countable) union of the form ∪∞ i=1αi × βi, where αi ⊆ X1 and βi ⊆ X2 are measurable, i ∈ N. The complements of ω-open sets are called ω-closed. An operator T ∈ B(H1, H2) is said to be supported on κ if Mχβ T Mχα = 0 whenever (α × β) ∩ κ ∼= ∅. (Here χγ stands for the characteristic function of a measurable subset γ.) Given an ω-closed set κ ⊆ X1 × X2, let Mmax(κ) = {T ∈ B(H1, H2) : T is supported on κ}. The space Mmax(κ) is a reflexive masa-bimodule in the sense that Ref(Mmax(κ)) = Mmax(κ) where, for a subspace U ⊆ B(H1, H2), we let its reflexive hull [19] be the subspace Ref(U ) = {T ∈ B(H1, H2) : T x ∈ U x, for all x ∈ H1}. We note two straightforward properties of the reflexive hull that will be used in the sequel: it is monotone (U1 ⊆ U2 implies Ref(U1) ⊆ Ref(U2)) and idempotent (Ref(Ref(U )) = Ref(U )). It was shown in [9] that every reflexive masa-bimodule is of the form Mmax(κ) for some, unique up to marginal equivalence, ω-closed set κ ⊆ X × Y . If U is any masa-bimodule, then its support supp U is defined to be the ω-closed set κ ⊆ X × Y such that Ref(U ) = Mmax(κ). The masa-bimodule Mmax(κ) is the largest, with respect to inclusion, (weak* closed) masa-bimodule with support κ (see [9]). As an extension of Arveson's work on commutative subspace lattices [1], it was shown in [25] that if κ is an ω-closed set, then there exists a smallest, with respect to inclusion, (weak* closed) masa-bimodule Mmin(κ) with support κ. The ω-closed subset κ ⊆ X × Y is called operator synthetic if Mmin(κ) = Mmax(κ). The roots of the notion of operator synthesis lie in Harmonic Analysis -- it is an operator theoretic version of the well-known concept of spectral synthesis. We refer the reader to [1] for a relevant discussion, and to [25] for the formal relation between the two concepts, which will be briefly summarised at the end of the section. The supports of masa-bimodules of finite width will be called sets of finite width. A set κ ⊆ X1 × X2 is of finite width precisely when it is the set of solutions of a system of (finitely many) measurable function inequalities, that is, precisely when it has the form κ = {(x, y) ∈ X1 × X2 : fk(x) ≤ gk(y), k = 1, . . . , n}, where fk : X1 → R and gk : X2 → R are measurable functions, k = 1, . . . , n (see, e.g., [28]). It was shown in [25] and [28] that sets of finite width are operator synthetic. 24 G.K. ELEFTHERAKIS AND I.G. TODOROV In this section, we will be concerned with the question of when operator synthesis is preserved under unions of products. Suppose that (Y1, ν1), (Y2, ν2) is another pair of standard measure spaces, Ki = L2(Yi, νi) and Ci is the multiplication masa of L∞(Yi, νi), i = 1, 2. Let U ⊆ B(H1, H2) be a D2, D1-module and V ⊆ B(K1, K2) be a C2, C1-module. Then the subspace U ¯⊗V is a D2 ¯⊗C2, D1 ¯⊗C1-module, and hence its support is a subset of (X1 × Y1) × (X2 × Y2). The "flip" ρ : (X1 × X2) × (Y1 × Y2) → (X1 × Y1) × (X2 × Y2), given by ρ(x1, x2, y1, y1) = (x1, y1, x2, y2), xi ∈ Xi, yi ∈ Yi, i = 1, 2, is thus needed in order to relate supp(U ¯⊗V) to (supp U ) × (supp V). Indeed, it was shown in [21] that (5) supp(U ¯⊗V) = ρ(supp U × supp V). It was observed in [24, Lemma 4.19] that (6) Mmin(ρ(κ × λ)) = Mmin(κ) ¯⊗Mmin(λ), whenever κ ⊆ X1 × X2 and λ ⊆ Y1 × Y2 are ω-closed sets. Remark 5.1. If κ ⊆ X1 × X2 and λ ⊆ Y1 × Y2 are non-marginally null ω-closed sets such that ρ(κ × λ) is operator synthetic, then both κ and λ are operator synthetic. Indeed, suppose that T ∈ Mmax(κ), and let 0 6= S ∈ Mmin(λ). Then T ⊗ S ∈ Mmax(ρ(κ × λ)) and, by assumption and identity (6), T ⊗ S ∈ Mmin(κ) ¯⊗Mmin(λ). It now easily follows that T ∈ Mmin(κ). Thus, κ is operator synthetic and by symmetry λ is so as well. Remark 5.2. Let G and H be locally compact groups. A problem in Harmonic Analysis asks when, given closed sets E ⊆ G and F ⊆ H satisfying spectral syn- thesis, the set E × F satisfies spectral synthesis as a subset of the direct product G × H. We refer the reader to [11] for the definition of the notion of spectral synthesis and other basic concepts and results from non-commutative Harmonic Analysis. Analogues of identities (5) and (6) in the setting of Harmonic Analysis can be formulated as follows. Let VN(G) ⊆ B(L2(G)) (resp. VN(H) ⊆ B(L2(H))) be the von Neumann algebra of G (resp. H), and note that VN(G) ¯⊗ VN(H) can be naturally identified with VN(G × H). The Harmonic Analysis analogue of masa-bimodules are invariant spaces; these are subspaces X ⊆ VN(G) that are annihilators of ideals of the Fourier algebra A(G) of G. Given an invariant space X ⊆ VN(G), one may define its support supp X as the null set of its preannihilator in A(G). It is not difficult to see that if X ⊆ VN(G) and Y ⊆ VN(H) are invariant spaces, then supp(X ¯⊗Y) = (supp X ) × (supp Y) and that, given any closed sub- set E ⊆ G, there exists a largest (resp. smallest) invariant space Xmax(E) (resp. Xmin(E)) with support E, and Xmin(E) ¯⊗Xmin(F ) = Xmin(E × F ). The next proposition describes the connection between operator synthesis and tensor product formulas. OPERATOR SYNTHESIS AND TENSOR PRODUCTS 25 Proposition 5.3. Let U ⊆ B(H1, H2) and V ⊆ B(K1, K2) be masa-bimodules with supports κ ⊆ X1 × X2 and λ ⊆ Y1 × Y2, respectively. Then (7) and (8) supp F(U , V) = ρ(κ × λ) F(Mmax(κ), Mmax(λ)) = Mmax(ρ(κ × λ)). Moreover, if κ and λ are operator synthetic, then the following statements are equivalent: ρ(κ × λ) is operator synthetic; (i) (ii) F(Mmax(κ), Mmax(λ)) = Mmax(κ) ¯⊗Mmax(λ); (iii) F(Mmin(κ), Mmin(λ)) = Mmin(κ) ¯⊗Mmin(λ). Proof. We have that F(U , V) = (U ¯⊗B(K1, K2)) ∩ (B(H1, H2) ¯⊗V), and hence supp F(U , V) = supp(U ¯⊗B(K1, K2)) ∩ supp(B(H1, H2) ¯⊗V). By (5), the support of U ¯⊗B(K1, K2) (resp. B(H1, H2) ¯⊗V) is ρ(κ × (Y1 × Y2)) (resp. ρ((X1 × X2) × λ)). Identity (7) now readily follows. To establish (8) note that F(Mmax(κ), Mmax(λ)) and Mmax(ρ(κ × λ)) are both reflexive and, by (7), have equal supports. Suppose that κ and λ are operator synthetic. (ii)⇔(i) Using [24, Lemma 4.19] for the first equality below and identity (8) for the last one, we have Mmin(ρ(κ × λ)) = Mmin(κ) ¯⊗Mmin(λ) = Mmax(κ) ¯⊗Mmax(λ) ⊆ F(Mmax(κ), Mmax(λ)) = Mmax(ρ(κ × λ)). If the inclusion in the above chain is equality then we have that Mmin(ρ(κ×λ)) = Mmax(ρ(κ × λ)), in other words, that ρ(κ × λ) is operator synthetic. Conversely, if ρ(κ × λ) is operator synthetic then we must have equalities throughout. (iii)⇔(i) follows similarly from the chain Mmin(ρ(κ × λ)) = Mmin(κ) ¯⊗Mmin(λ) ⊆ F(Mmin(κ), Mmin(λ)) = (Mmin(κ) ¯⊗B(K1, K2)) ∩ (B(H1, H2) ¯⊗Mmin(λ)) = (Mmax(κ) ¯⊗B(K1, K2)) ∩ (B(H1, H2) ¯⊗Mmax(λ)) = F(Mmax(κ), Mmax(λ)) = Mmax(ρ(κ × λ)). (cid:3) Corollary 5.4. Let κ ⊆ X1 ×X2 be an operator synthetic ω-closed set. If Mmax(κ) has property Sσ then ρ(κ × λ) is operator synthetic for every operator synthetic ω- closed set λ ⊆ Y1 × Y2. Proof. Immediate from Proposition 5.3 (ii)⇔(i). (cid:3) It follows from Corollary 5.4 that if κ is a set of finite width then ρ(κ × λ) is operator synthetic whenever λ is so. In fact, we have the following stronger result. 26 G.K. ELEFTHERAKIS AND I.G. TODOROV Corollary 5.5. Let κ ⊆ X1 × X2 and λ ⊆ Y1 × Y2 be operator synthetic sets and κ′ ⊆ X1 × X2 be an ω-closed set of finite width. If ρ(κ × λ) is operator synthetic then so is ρ((κ ∪ κ′) × λ). Proof. Let V = Mmax(κ), B = Mmax(κ′) and U = Mmax(λ). It is straightforward to check that the support of V + B is κ ∪ κ′. By [8, Corollary 4.2], κ ∪ κ′ is operator synthetic, and hence Mmax(κ ∪ κ′) = V + B. By Proposition 5.3, F(Mmax(κ), Mmax(λ)) = Mmax(κ) ¯⊗Mmax(λ). By Theorem 3.10, F(Mmax(κ ∪ κ′), Mmax(λ)) = Mmax(κ ∪ κ′) ¯⊗Mmax(λ). By Proposition 5.3, ρ((κ ∪ κ′) × λ) is operator synthetic. (cid:3) Our next aim is Theorem 5.9, for whose proof we will need some auxiliary lem- mas. Lemma 5.6. Let U ⊆ B(H1, H2) be a masa-bimodule and φ be a Schur idempotent acting on B(H1, H2). Then φ(Ref(U )) = Ref(φ(U )) = Ran φ ∩ Ref(U ). Proof. By [8, Proposition 3.3], Ref(U ) coincides with the space of all operators X ∈ B(H1, H2) such that ψ(X) = 0 whenever ψ is a Schur idempotent annihilating U . Fix T ∈ Ref(U ) and let θ be a Schur idempotent on B(H1, H2) such that θ(φ(U )) = {0}. Then θ ◦ φ(T ) = 0 and hence φ(T ) ∈ Ref(φ(U )); we thus showed that φ(Ref(U )) ⊆ Ref(φ(U )). Now suppose that T ∈ φ(Ref(U )); then clearly T ∈ Ran φ and, by the previous paragraph, T ∈ Ref(φ(U )) ⊆ Ref(U ). Thus, φ(Ref(U )) ⊆ Ran φ ∩ Ref(U ). On the other hand, if T ∈ Ran φ∩Ref(U ) then T = φ(T ) ∈ φ(Ref(U )); hence, φ(Ref(U )) = Ran φ ∩ Ref(U ). By [8, Proposition 3.3], Ran φ is reflexive and since reflexivity is preserved by intersections, the previous paragraph implies that φ(Ref(U )) is reflexive. Since φ(U ) ⊆ φ(Ref(U )), we have Ref(φ(U )) ⊆ Ref(φ(Ref(U ))) = φ(Ref(U )), and the proof is complete. (cid:3) Lemma 5.7. Let φi be a Schur idempotent, κi ⊆ X1 × X2 be the support of Ran φi, and λi ⊆ Y1 × Y2 be an ω-closed set, i = 1, . . . , r. Suppose that ∪p k=1λmk is operator synthetic, whenever 1 ≤ m1 < m2 < · · · < mp ≤ r. Then the set ρ(∪r i=1κi × λi) is operator synthetic. Proof. Set κ = ρ(∪r i=1κi × λi), Ui = Mmin(λi) and W = Mmin(κ). By (5), the i=1κi × λi); by the minimality property of W and the fact that the sets κi and λi are operator synthetic, we have that support of the masa-bimodulePr Mmin(κi) ¯⊗Mmin(λi) is ρ(∪r i=1 W = Mmin(κi) ¯⊗Mmin(λi) = Mmax(κi) ¯⊗Mmax(λi). rXi=1 rXi=1 For each i = 1, . . . , r, let φ1 2 · · · φǫr {1, . . . , r}, let φM = φǫ1 1 φǫ2 i = φi and φ−1 r , where ǫi = 1 if i ∈ M and ǫi = −1 if i 6∈ M . i , and for each subset M of i = φ⊥ OPERATOR SYNTHESIS AND TENSOR PRODUCTS 27 Fix T ∈ Ref(W); we will show that T ∈ W. This will then imply that W = Ref(W), and hence that ρ(∪r i=1κi × λi) is operator synthetic. We have T =P φM (T ), where the sum is taken over all subsets M of {1, . . . , r}. By Lemma 5.6, φM (T ) ∈ Ref( φM (W)) and hence By Lemma 3.1, φM (Ran φi ¯⊗Ui) = Ran φM ¯⊗Ui if i ∈ M , and φM (Ran φi ¯⊗Ui) = {0} otherwise. Thus, φM (Ran φi ¯⊗Ui)! . rXi=1 T ∈ Ref XM T ∈ RefXM Ran φM ¯⊗Xi∈M Ui! . We have that φM φN = 0 if M 6= N . The assumption concerning the synthesis of the finite unions of the sets λj implies that Pi∈M Ui = Mmax(∪i∈M λi); we may thus assume that the maps φi, i = 1, . . . , r have the property that φiφj = 0 if i 6= j. We now proceed by induction on r. If r = 1, the statement follows from Lemma 3.1 and Corollary 5.4. Assume that the statement holds if the number of the given terms is at most r − 1, and recall that T ∈ Ref(W). By Lemma 5.6 and the inductive assumption, φ⊥ r (T ) ∈ Ref( φ⊥ On the other hand, r (W)) ⊆ Ref r−1Xi=1 Mmax(κi) ¯⊗Mmax(λi)! ⊆ W. φr(T ) ∈ Ref(Mmax(κr) ¯⊗Mmax(λr)) = Mmax(κr) ¯⊗Mmax(λr) = Mmin(κr) ¯⊗Mmin(λr) ⊆ W (we have used Lemma 5.6 for the containment, and Corollary 5.4, Proposition 5.3 and the fact that Mmax(κr) has property Sσ for the first equality). Thus, T = φr(T ) + φ⊥ r (T ) ∈ W and the proof is complete. (cid:3) Lemma 5.8. Let κi ⊆ X1 × X2 be the support of a nest algebra bimodule, and let λi ⊆ Y1 × Y2 be an ω-closed set, i = 1, . . . , r. Suppose that ∪p k=1λmk is operator synthetic whenever 1 ≤ m1 < m2 < · · · < mp ≤ r. Then the set ρ(∪r i=1κi × λi) is operator synthetic. Proof. Set κ = ρ(∪r and W =Pr Ref(W) = Mmax(κ). i=1κi × λi). Let Bi = Mmax(κi), Ui = Mmax(λi), i = 1, . . . , r, i=1 Bi ¯⊗Ui. As in the proof of Lemma 5.7, W = Mmin(κ) and hence Let (φi,k, θi,k, Mi,k, Zi,k)k∈N be a decomposition scheme for Bi, i = 1, . . . , r. Set ψi,k = (φi,k + θi,k)⊥. For each subset M of {1, . . . , r}, a subset N of M , and indices = γ1 ◦ · · · ◦ γr, where γi = φi,ki if i ∈ N , k1, k2, . . . , kr ∈ N, we let γM,N k1,k2,...,kr 28 G.K. ELEFTHERAKIS AND I.G. TODOROV γi = θi,ki if i ∈ M \ N and γi = ψi,ki if i 6∈ M . Fix T ∈ Ref(W). Then, by Lemmas 3.1 and 5.6, where Yi,ki is equal to Mi,ki if i ∈ N , to Zi,ki if i ∈ M \ N and to {0} if i 6∈ M . Moreover, for all k1, k2, . . . , kr, we have that γM,N k1,k2,...,kr ¯⊗Ui! , Yi,ki (T ) ∈ Ref rXi=1 T = XM,N γM,N k1,k2,...,kr (T ), where the sum is taken over all subsets M and N of {1, . . . , n} with N ⊆ M . By Lemma 5.7, Since Zi,k ⊆ Bi for every k ∈ N, we have that γM,N k1,k2,...,kr (T ) ∈ γM,N k1,k2,...,kr (T ) ∈ Yi,ki ¯⊗Ui. Xi,ki ¯⊗Ui, rXi=1 rXi=1 where Xi,ki is equal to Mi,ki if i ∈ N , to Bi if i ∈ M \ N and to {0} if i 6∈ M . Let {(Mp, Np)}q k1,k2,...,kr−1,k′ r (T ))k′ p=1 be an enumeration of the pairs of sets (M, N ) with N ⊆ M ⊆ {1, . . . , r}. For every fixed r − 1-tuple (k1, . . . , kr−1) of indices, choose a weak* con- vergent subsequence (γM1,N1 (T ))kr∈N, and let γM1,N1 k1,k2,...,kr−1(T ) be its limit. Then choose a weak* convergent subse- quence (γM2,N2 r∈N, and let γM2,N2 k1,k2,...,kr−1(T ) be its limit. Continuing inductively, define, for each pair (M, N ), an operator γM,N k1,k2,...,kr−1(T ); by the choice of these operators, we have that r ∈N of the sequence (γM1,N1 r∈N of the sequence (γM1,N1 k1,k2,...,kr−1,k′′ r k1,k2,...,kr−1,k′ r (T ))k′′ k1,k2,...,kr (T ))k′ T = XM,N γM,N k1,k2,...,kr−1(T ). By Proposition 4.1 (v), γM,N k1,k2,...,kr−1(T ) ∈ r−1Xi=1 Xi,ki ¯⊗Ui + Br ¯⊗Ur. We now choose, as in the previous paragraph, for every r − 2-tuple (k1, . . . , kr−2) k1,k2,...,kr−1(T ))kr−1∈N such of indices, a weak* cluster point γM,N k1,k2,...,kr−2(T ) of (γM,N k1,k2,...,kr−2(T ), and use Proposition 4.1 (v) to conclude that that T =PM,N γM,N γM,N k1,k2,...,kr−2(T ) ∈ Xi,ki ¯⊗Ui + Br−1 ¯⊗Ur−1 + Br ¯⊗Ur. r−2Xi=1 OPERATOR SYNTHESIS AND TENSOR PRODUCTS 29 Continuing inductively, we conclude that T =PM,N γM,N (T ) ∈ W for all subsets N and M of {1, . . . , r} with N ⊆ M . Hence, T ∈ W and the proof is complete. (cid:3) (T ), where γM,N ∅ ∅ Theorem 5.9. Let κi ⊆ X1 × X2 be a set of finite width, and let λi ⊆ Y1 × Y2 be an ω-closed set, i = 1, . . . , r. Suppose that ∪p k=1λmk is operator synthetic whenever 1 ≤ m1 < m2 < · · · < mp ≤ r. Then the set ρ(∪r i=1κi × λi) is operator synthetic. Proof. Let κ = ρ(∪r j, where Bi j is a nest algebra bimodule, i = 1, . . . , r, j = 1, . . . , li. Let also Ui = Mmax(λi), i = 1, . . . , r. Fix i=1κi × λi), Bi = Mmax(κi) and write Bi = ∩li j=1Bi Bi ¯⊗Ui! . Bi ji ¯⊗Ui. T ∈ Mmax(κ) = Ref rXi=1 ¯⊗Ui! = T ∈ Ref rXi=1 rXi=1 rXi=1 T ∈ Bi ji Bi ¯⊗Ui = Mmin(κ). Lemma 5.8 implies that, for all j1, . . . , jr, we have By Corollary 4.21, (cid:3) Remark 5.10. In Theorem 5.9, the condition that ∪p k=1λmk be operator synthetic whenever 1 ≤ m1 < m2 < · · · < mp ≤ r cannot be omitted. Indeed, given such a choice of indices, fix a non-trivial subset of finite width κ, and let κmj = κ, j = 1, . . . , p, and κi = ∅ if i 6∈ {m1, . . . , mp}. If ρ(∪r j=1λmj )) is operator synthetic then, by Remark 5.1, ∪p i=1κi × λi) = ρ(κ × (∪p j=1λmj is operator synthetic. We conclude this section with an application of the previous results to spectral synthesis. Let G be a second countable locally compact group. By [20], a closed set E ⊆ G satisfies local spectral synthesis if and only if the set E∗ = {(s, t) ∈ G × G : st−1 ∈ E} is operator synthetic (here G is equipped with left Haar measure). We note that, in the case the Fourier algebra A(G) has an approximate identity, E is of local spectral synthesis if and only if it satisfies spectral synthesis (see [20]). Let R+ be the group of positive real numbers and ω : G → R+ be a continuous group homomorphism. For each t > 0, let Et ω = {x ∈ G : ω(x) ≤ t}; it is natural to call such a subset a level set. We have that (Et ω)∗ = {(x, y) ∈ G × G : ω(x) ≤ tω(y)} and hence the intersections of the form E = Et1 ω1 ∩ · · · ∩ Etk ωk 30 G.K. ELEFTHERAKIS AND I.G. TODOROV are a Harmonic Analysis version of sets of finite width: they have the property that the corresponding set E∗ is a set of finite width (see also [8]). Theorem 5.9 has the following immediate consequence. Corollary 5.11. Let G and H be second countable locally compact groups. Suppose that E1, . . . , Er are level sets in G and F1, . . . , Fr are closed subsets of H such that ∪p k=1Fmk is a set of local spectral synthesis whenever 1 ≤ m1 < m2 < · · · < mp ≤ r. Then the set ∪r i=1Ei × Fi is a set of local spectral synthesis of G × H. 6. Fubini products and Morita equivalence In this section, we show how tensor product formulas relate to the notion of spacial Morita equivalence introduced in [6]. For subspaces X and Y of operators, we let [X Y] =( kXi=1 XiYi : Xi ∈ X , Yi ∈ Y, i = 1, . . . , k, k ∈ N). We recall the following definition from [7]: Definition 6.1. Let A (resp. B ) be a weak* closed unital algebra acting on a Hilbert space H1 (resp. H2). We say that A is spatially embedded in B if there exist a B, A-bimodule X ⊆ B(H1, H2) and an A, B-bimodule Y ⊆ B(H2, H1) such that [X Y] ⊆ B and [YX ] = A. If, moreover, B = [X Y], we call A and B spatially Morita equivalent. We note that if two unital dual operator algebras A and B are weak* Morita equivalent in the sense of [2] then they have completely isometric representations α and β such that the algebras α(A) and β(B) are spatially Morita equivalent (this fact will not be used in the sequel). Theorem 6.1. Let H1, H2 and K be Hilbert spaces and A ⊆ B(H1) and B ⊆ B(H2) be weak* closed unital algebras. Suppose that A is spatially embedded in B and let U ⊆ B(K) be a weak* closed space such that F(B, U ) = B ¯⊗U . Then F(A, U ) = A ¯⊗U . Proof. Let X ⊆ B(H1, H2) and Y ⊆ B(H2, H1) be subspaces satisfying the con- ditions of Definition 6.1, and T ∈ A ¯⊗B(K) be such that Rφ(T ) ∈ U for all φ ∈ B(H1, H2)∗. Suppose that T = w∗- lim n An i ⊗ Sn i , mnXi=1 i )mn i=1 ⊆ A and (Sn where (An i=1 ⊆ B(K). Fix X1 ∈ X and Y1 ∈ Y and set S = (X1 ⊗ I)T (Y1 ⊗ I). For ψ ∈ B(H2)∗, let φ ∈ B(H1)∗ be given by φ(A) = ψ(X1AY1), A ∈ B(H1). Since i )mn Rφ(T ) = w∗- lim n mnXi=1 φ(An i )Sn i ∈ U , OPERATOR SYNTHESIS AND TENSOR PRODUCTS 31 we have that Rψ(S) ∈ U , for every ψ ∈ B(H2)∗. By our assumption, S ∈ B ¯⊗U . Therefore, (X1 ⊗ I)T (Y1 ⊗ I) ∈ B ¯⊗U for all X1 ∈ X and all Y1 ∈ Y. It follows that (Y2X1 ⊗ I)T (Y1X2 ⊗ I) ∈ A⊗U , for all X1, X2 ∈ X , Y1, Y2 ∈ Y. Since I ∈ A = [YX ], it follows that T ∈ A ¯⊗U . (cid:3) The following corollary is straightforward from Theorem 6.1. Corollary 6.2. Suppose that A and B are weak* closed unital operator algebras. (i) Suppose that A is spatially embedded in B. If B has property Sσ then so does A. (ii) Suppose that A and B are spatially Morita equivalent. Then A has property Sσ precisely when B does so. The last corollary, which is immediate from [7] and Corollary 6.2, concerns the inheritance of property Sσ in the class of CSL algebras; we refer the reader to [1] for the definition, relevant notation and theory of this class of algebras. Corollary 6.3. Let L1 and L2 be CSL's. (i) Suppose that φ : L1 → L2 is a strongly continuous surjective lattice homo- morphism. If Alg(L1) has property Sσ then so does Alg(L2). (ii) Suppose that φ : L1 → L2 is a strongly continuous lattice isomorphism. Then the algebra Alg(L1) has property Sσ if and only if Alg(L2) does so. References [1] W.B. Arveson, Operator algebras and invariant subspaces, Ann. Math. (2) 100 (1974), 433-532 [2] D. P. Blecher and U. Kashyap, Morita equivalence of dual operator algebras, J. Pure Appl. Algebra 212 (2008), 2401-2412 [3] D. P. Blecher and C. Le Merdy, Operator Algebras and Their Modules -- An Operator Space Approach, Oxford University Press, 2004 [4] K. R. Davidson, Nest Algebras, Longman Scientific & Technical, Harlow, 1988 [5] E. Effros and Z.-J. Ruan, Operator Spaces, Oxford University Press, 2000 [6] G. K. Eleftherakis, TRO equivalent algebras, Houston J. Math. 38 (2012), no. 1, 153-175 [7] G. K. Eleftherakis, Applications of operator space theory to nest algebra bimodules, Integral Eq. Operator Th. 72 (2012), no. 4, 577-595 [8] G. K. Eleftherakis and I. G. Todorov, Ranges of bimodule projections and reflexivity, J. Funct. Anal. 262 (2012), no. 11, 4891-4915 [9] J. A. Erdos, A. Katavolos, and V. S. Shulman, Rank one subspaces of bimodules over maximal abelian selfadjoint algebras, J. Funct. Anal. 157 (1998) no. 2, 554-587 [10] J.A. Erdos and S.C. Power, Weakly closed ideals of nest algebras, J. Operator Theory 7 (1982), no. 2, 219-235 [11] P. Eymard, L'alg`ebre de Fourier d'un groupe localment compact, Bull. Soc. Math. France 92 (1964), 181-236 [12] J. Froelich Compact operators, invariant subspaces and spectral synthesis, J. Funct. Anal. 81 (1988), 1-37 [13] U. Haagerup and J. Kraus, Approximation properties of group C-algebras and group von Neumann algebras, Trans. Amer. Math. Soc. 344 (1994), no. 2, 667-699 [14] A. Hopenwasser and J. Kraus, Tensor products of reflexive algebras II, J. London Math. Soc. (2) 28 (1983), 359-362 32 G.K. ELEFTHERAKIS AND I.G. TODOROV [15] A. Katavolos and V. Paulsen, On the ranges of bimodule projections, Canad. Math. Bull. 48 (2005) no. 1, 97-111 [16] A. Katavolos and I.G. Todorov, Normalizers of operator algebras and reflexivity, Proc. London Math. Soc. (3) 86 (2003), 463-484 [17] J. Kraus, The slice map problem for σ-weakly closed subspaces of von Neumann algebras, Trans. Amer. Math. Soc. 279 (1983), no. 1, 357-376 [18] J. Kraus, The slice map problem and approximation properties, J. Funct. Anal. 102 (1991), 116-155 [19] A.I. Loginov and V.S.Shulman, Hereditary and intermediate reflexivity of W-algebras, Izv. Akad. Nauk SSSR 39 (1975), 1260-1273; Math. USSR-Izv. 9 (1975), 1189-1201 [20] J. Ludwig and L. Turowska, On the connection between sets of operator synthesis and sets of spectral synthesis for locally compact groups, J. Funct. Anal. 233 (2006), 206-227 [21] M. McGarvey, L. Oliveira and I.G. Todorov, Normalisers of operator algebras and tensor product formulas, Rev. Math. Iber., to appear [22] V. I. Paulsen, Completely bounded maps and operator algebras, Cambridge University Press, 2002 [23] G. Pisier, Introduction to Operator Space Theory, Cambridge University Press, 2003 [24] V. S. Shulman, I.G. Todorov and L. Turowska, Sets of multiplicity and closable multi- pliers on group algebras, preprint [25] V. S. Shulman and L. Turowska, Operator synthesis I. Synthetic sets, bilattices and tensor algebras, J. Funct. Anal. 209 (2004), 293-331 [26] V.S. Shulman and L. Turowska, Operator synthesis II: Individual synthesis and linear operator equations J. Reine Angew. Math. 590 (2006), 143-187 [27] N. Spronk and L. Turowska, Spectral synthesis and operator synthesis for compact groups, J. London Math. Soc. (2) 66 (2002), 361-376 [28] I.G. Todorov, Spectral synthesis and masa-bimodules, J. London Math. Soc. (2) 65 (2002), 733-744 [29] I.G. Todorov, Synthetic properties of ternary masa-bimodules, Houston J. Math. 32 (2006), no. 2, 505-519 Department of Mathematics, University of Athens, Athens 157 84, Greece E-mail address: [email protected] Pure Mathematics Research Centre, Queen's University Belfast, Belfast BT7 1NN, United Kingdom E-mail address: [email protected]
1507.08270	4	1507	2016-02-15T15:30:38	Two-faced Families of Non-commutative Random Variables Having Bi-free Infinitely Divisible Distributions	[ "math.OA" ]	We study two-faced families of random variables having bi-free infinitely divisible distributions. We prove a limit theorem of the sums of bi-free two-faced pairs of random variables within a triangular array. Then, by using the full Fock space operator model, we show that a two-faced pair of random variables has a bi-free (additive) infinitely divisible distribution if and only if its distribution is the limit distribution in our limit theorem. Finally, we characterize the bi-free (additive) infinite divisibility of the distribution of a two-faced pair of random variables in terms of bi-free Levy processes.	math.OA	math	TWO-FACED FAMILIES OF NON-COMMUTATIVE RANDOM VARIABLES HAVING BI-FREE INFINITELY DIVISIBLE DISTRIBUTIONS MINGCHU GAO Abstract. We study two-faced families of non-commutative random variables having bi-free (additive) infinitely divisible distributions. We prove a limit theorem of the sums of bi-free two- faced families of random variables within a triangular array. As a corollary of our limit theorem, we get Voiculescu's bi-free central limit theorem. Using the full Fock space operator model, we show that a two-faced pair of random variables has a bi-free (additive) infinitely divisible distribution if and only if its distribution is the limit distribution in our limit theorem. Finally, we characterize the bi-free (additive) infinite divisibility of the distribution of a two-faced pair of random variables in terms of bi-free Levy processes. Key Words. Bi-free Probability, Limit Theorems, Infinitely Divisible Distributions, Bi-free Levy Processes. 2010 MSC 46L54 Introduction On a free product of Hilbert spaces with a specified unit vector there are two actions of the operators of the initial space, corresponding to a left and a right tensorial factorization respectively ([VDN]). In free probability, free random variables can be modeled as left or right actions on free products of spaces. When considering the algebra of left actions and the algebra of right actions on the free product space simultaneously, Voiculescu [Vo1] discovered a new phenomenon of freely independent family of two-faced families of random variables, which is called bi-free independence. This "two-faced" extension of free probability is called bi-free probability (or free probability for two-faced pairs) introduced in [Vo1] and [Vo2] recently. Bi-free independence is a more general notion than free independence. Classical independence and free independence can be treated as special cases of bi-freeness ([PS]). Voiculescu [Vo1] demonstrated that many results in free probability such as existence of free cumulants and the free central limit theorem, have direct analogues in the bi-free setting. Voiculescu [Vo1] showed the existence of bi-free cumulant polynomials, but did not give explicit formulas for the polynomials. Mastnak and Nica [MN] introduced combinatorial objects called bi-non-crossing partitions, associated a family of (l,r)-cumulants with the combinatorial objects, and conjectured that bi-freeness was equivalent to the vanishing of these mixed cumulants. This conjecture was later proved by Charlesworth, Nelson, and Skoufranis in [CNS1]. The same authors developed a theory of bi-freeness in an amalgamated setting in [CNS2]. In classical probability, infinitely divisible distributions appear as a broad generalization of the central limit theorem and the Poisson limit theorem: the limit distribution of the sums of i.i.d. random variables within a triangular array. Meanwhile, infinitely divisible distributions are closely related to Levy processes, a very important research area in probability theory ([PSa]). There has been a well-developed theory on infinitely divisible distributions in free probability: the limit distributions of free random variables within a triangular array, infinitely divisible distributions with respect to the additive free convolution, and free Levy processes have a perfect relation similar to those in classical probability ([NS]). In bi-free probability, Voiculescu [Vo1] defined centered bi-free Gaussian distributions and proved an algebraic bi-free central limit theorem, a bi-free version of semicircle distributions and the free central limit theorem, respectively (7.2 and 7.3 in [Vo1]). In 1 [Vo2], Voiculescu studied a special kind of two-faced families of random variables where algebraic relations between left and right random variables ensure that all moments can be computed from the "two-band" moments ϕ(LR), where L and R are monomials in left and, respectively, right random variables. A simplest example is bi-partite systems where left and right variables commute with each other. In [GHM], the authors studied infinitely divisible distributions in the bi-free probability setting for bi-partite systems. They derived a bi-free analogue of the Levy-Hincin formula for infinitely divisible distributions with respect to the additive bi-free convolution, constructed bi-free Levy processes corresponding to bi-free infinitely divisible distributions, and proved a bi-free limit theorem. In this paper, we study infinitely divisible distributions with respect to the additive bi-free convolution for general two-faced pairs of non-commutative random variables, i.e., we do not assume that the left and right random variables of the two-faced pair commute with each other. Besides this introduction, this paper contains 4 sections. In section 1 we review the basic knowl- edge on bi-free probability used in sequel. In Section 2, we prove limit theorems (Theorem 2.3 and Corollary 2.4) in bi-free probability, generalizing the limit theorem in free probability (Theorem 13.1 in [NS]) and the limit theorem in bi-free probability for bi-partite systems (Theorem 3.1 in [GHM]) to the general case of two-faced families of random variables. As a corollary of our limit theorems, we get Voiculescu's bi-free central limit theorem (Proposition 2.6). Section 3 is devoted to studying two-faced pairs of random variables having bi-free infinitely divisible distributions. We show that a two-faced pair (a, b) of self adjoint operators in a C∗-probability apace (A, ϕ) has a bi-free infinitely divisible distribution if and only if its distribution is the limit distribution of the sums of bi-free independent two-faced pairs of random variables within a triangular array (Theorem 3.7). Finally, in Section 4, we define bi-free Levy processes (Definition 4.1). Our definition is very similar to Definition 4.1 in [GHM]. But we do not assume that the system is bi-partite. We find that a bi-free infinitely divisible distribution is exactly the distribution of a1 = (al,1, ar,1) of a bi-free Levy process {at = (al,t, ar,t) : t ≥ 0} (Theorem 4.2), generalizing Theorem 4.2 in [GHM] to the general case of two-faced pairs of random variables. We refer the reader to [VDN] and [NS] for free probability, to [Vo1], [Vo2], [MN], [CNS1], and [CNS2] for bi-free probability, and to [KR] for operator algebras. Acknowledgement The author is grateful to the referee(s) for carefully reviewing the paper, providing many valuable suggestions, and finding out typos in the paper. The author would like to thank Dr. Paul Skoufranis of Texas A. and M. University and Prof. Guimei An of Nankai University, China, for pointing out mistakes and typos in the initial version of this paper. 1. Preliminaries In this section, we review some basic concepts and results in bi-free probability used in sequel. Free Products of vector spaces Let X be a vector space with a vector ξ, and a subspace X0 with co-dimension 1 such that X = Cξ ⊕ X0. Let L(X ) be the space of all linear operators on X . We use (X ,X0, ξ) to denote vector space X with the above decomposition property. Let (Xi,Xi,0, ξi), i ∈ I, be a family of vector spaces. The free product (X ,X0, ξ) = ∗i∈I (Xi,Xi,0, ξi) is defined as X = Cξ ⊕ X0,X0 = ⊕n≥1(⊕i16=i26=···6=inXi1,0 ⊗ Xi2,0 ⊗ ··· ⊗ Xin,0). We can define a linear functional φ : X → C by φ(ξ) = 1, ker(φ) = X0, then a linear functional ϕξ : L(X ) → C, ϕξ(T ) = φ(T ξ),∀T ∈ L(X ). The pair (L(X ), ϕξ) is a non-commutative probability space. A Tensor Product Factorization of Free Product Vector Spaces Let X = X0 ⊕ Cξ be the free product of a family {(Xi,X0,i, ξi) : i ∈ I} of vector spaces. For i ∈ I, define X0,i1 ⊗ X0,i2 ⊕ ··· ⊕ X0,in ), X (l, i) = Cξ ⊕Mn≥1 ( Mi16=i26=···6=in,i16=i 2 and X (r, i) = Cξ ⊕Mn≥1 ( Mi16=i26=···6=in,in6=i X0,i1 ⊗ X0,i2 ⊕ ··· ⊕ X0,in ). Define a unitary operator Vi : Xi ⊗ X (l, i) → X by Vi : ξi ⊗ ξ 7→ ξ,X0,i ⊗ ξ 7→ X0,i, ξi ⊗ (X0,i1 ⊗ X0,i2 ⊕ ··· ⊕ X0,in ) 7→ X0,i1 ⊗ X0,i2 ⊕ ··· ⊕ X0,in , X0,i ⊗ (X0,i1 ⊗ X0,i2 ⊕ ··· ⊕ X0,in ) 7→ X0,i ⊗ X0,i1 ⊗ X0,i2 ⊕ ··· ⊕ X0,in . We can define Wi : X (r, i) ⊗ Xi → X similarly. rightmost ρi(T ): An operator T ∈ L(Xi) can act on the free product space (X ,X0, ξ) from leftmost λi(T ), or from Bi-free independence. A two-faced family of random variables λi(T ) = Vi(T ⊗ I)V ∗i , ρi(T ) = Wi(I ⊗ T )W ∗i . (bb,bc) := ((bi)i∈I , (cj)j∈J ) is an ordered pair of two families of elements in a non-commutative probability space (A, ϕ). For a two-faced family (bb,bc) = ((bi)i∈I , (cj)j∈J ) of random variables in A, its distribution µbb,bc : ChXi, Yji ∈ I, j ∈ Ji → C is defined as µbb,bc(P ((Xi)i∈I , (Yj)j∈J )) = ϕ(P ((bi)i∈I , (cj)j∈J )),∀P ∈ ChXi, Yji ∈ I, j ∈ Ji. A pair of faces (or face-pair) in a non-commutative probability space (A, ϕ) is an ordered pair (B,C) of two unital subalgebras of A. Let π = ((Bk,Ck))k∈K be a family of pairs of faces in (A, ϕ). The joint distribution of π is the linear functional µπ : ∗k∈KBk ∗ Ck → C defined by µπ = ϕ ◦ α, where α : ∗k∈KBk ∗ Ck → A is the homomorphism such that αBk (x) = x,∀x ∈ Bk, αCk (x) = x,∀x ∈ Ck. Let z′ = ((b′i)i∈I , (c′j)j∈J ) and z′′ = ((b′′i )i∈I , (c′′j )j∈J ) be two two-faced families in a non- commutative probability space (A, ϕ). We say that z′ and z′′ are bi-free if there exist a free product (X , p, ξ) = (X ′, p′, ξ′) ∗ (X ′′, p′′, ξ′′) of vector spaces and homomorphisms j : j ∈ Ji → L(X ε), ε ∈ {′,′′ }, lε : Chbε such that T ε := (λε ◦ lε(bε same as that of z′ and z′′ in (A, ϕ). i : i ∈ Ii → L(X ε), rε : Chcε i )i∈I , ρε ◦ rε(cε j)j∈J ) with ε ∈ {′,′′ } have a joint distribution in (L(X ), ϕξ) An example of bi-free two-faced families of random variables. Let H = ⊕i∈IHi be the direct sum of complex Hilbert spaces, and F (H) = CΩ ⊕Ln≥1 H⊗n be the full Fock space. Let τH : B(F (H)) → C, τH(T ) = hT Ω, Ωi,∀T ∈ B(F (H)), be the vector state on B(F (H)) corresponding to the vacuum vector Ω ∈ F (H). Then (B(F (H)), τH) is a C∗-probability space. For f ∈ H, T ∈ B(H), ξ = ξ1 ⊗ ··· ⊗ ξn ∈ H ⊗ ··· ⊗ H , define } ntimes {z l(f )Ω = f, l(f )ξ = f ⊗ ξ, r(f )Ω = f, r(f )ξ = ξ ⊗ f, Λl(T )Ω = 0, Λl(T )ξ = (T ξ1) ⊗ ξ2 ⊗ ··· ⊗ ξn, Λr(T )Ω = 0, Λr(T )ξ = ξ1 ⊗ ··· ⊗ ξn−1 ⊗ (T ξn). Proposition 1.1 (Remark 3.5 in [GHM]). Let Bi and Ci be the C∗-algebras generated by {l(f ) : f ∈ Hi} ∪ {Λl(T ) : T ∈ B(H), THi ⊂ Hi, TH⊖Hi = 0} and {r(f ) : f ∈ Hi} ∪ {Λr(T ) : T ∈ B(H), THi ⊂ Hi, TH⊖Hi = 0}, respectively. Then {(Bi,Ci) : i ∈ I} is bi-free in (B(F (H)), τH). 3 Bi-free cumulants. Let χ = (h1, h2,··· , hn) ∈ {l, r}n. Let's record explicitly where are the occurrences of l and r in χ. {m : 1 ≤ m ≤ n, hm = l} = {ml(1) < ml(2) < ··· , < ml(u)}, {m : 1 ≤ m ≤ n, hm = r} = {mr(1) < mr(2) < ··· , < mr(n − u)}. Define a permutation sχ : {1, 2,··· , n} → {1, 2,··· , n}, sχ(i) = ml(i), if 1 ≤ i ≤ u; sχ(u + i) = mr(n − u + 1 − i), if 1 ≤ i ≤ n − u. For a subset V = {i1, i2,··· , ik} of the set [n] := {1, 2,··· , n}, a1 ··· , an ∈ A, define ϕV (a1,··· , an) = ϕ(ai1 ai2 ··· aik ). Let P(n) be the set of all partitions of [n]. For a partition π = {V1, V2,··· , Vd} ∈ P(n), we define ϕπ(a1,··· , an) := YV ∈π ϕV (a1,··· , an). Define P χ(n) = {sχ ◦ π : π ∈ N C(n)}, where N C(n) is the set of all non-crossing partitions of [n] (Lecture 9 in [NS]). Let (A, ϕ) be a non-commutative probability space. The bi-free cumulants (κχ : An → C)n≥1,χ∈{l,r}n of (A, ϕ) are defined by κχ(a1,··· , an) = Xπ∈P (χ)(n) ϕπ(a1,··· , an)µn(s−1 χ ◦ π, 1n) (1.1) for n ≥ 1, χ ∈ {l, r}n, a1,··· , an ∈ A, where µn is the Mobius function on N C(n) (Lecture 10 in [NS]). For a subset V = {i1, i2,··· , ik} ⊆ {1, 2,··· , n}, let χV be the restriction of χ on V . We define κχ,V (a1, a2,··· , an) = κχV (ai1 , ai2 ,··· , aik ). For a partition π = {V1, V2,··· , Vk} ∈ κχ,1n (a1,··· , an). The bi-free cumulants are determined by the equation P (χ), we define κχ,π =QV ∈π κχ,V (a1, a2,··· , an). Then the bi-free cumulant appeared in (1.1) is κχ,π(a1, a2,··· , an),∀a1,··· , an ∈ A, (1.2) ϕ(a1a2 ··· an) = Xπ∈P (χ)(n) for a χ : {1, 2,··· , n} → {l, r}n. Charlesworth, Nelson, and Skoufranis [CNS1] proved that two two-faced families z′ = ((z′i)i∈I , (z′j)j∈J ), z′′ = ((z′′i )i∈I , (z′′j )j∈J ) in a non-commutative probability space (A, ϕ) are bi-free if and only if κχ(zǫ1 α(1), zǫ2 α(2),··· , zǫn α(n)) = 0, (1.3) whenever α : {1, 2,··· , n} → IF J, χ : {1, 2,··· , n} → {l, r} such that α−1(I) = χ−1({l}), ǫ : [n] → {′,′′ }n is not constant, and n ≥ 2 (Theorem 4.3.1 in [CNS1]). 2. Bi-free limit theorems Our goal, in this section, is to prove a bi-free limit theorem, an analogue of Theorem 13.1 in [NS] in bi-free probability. Lemma 2.1. Let (A, ϕ) be a non-commutative probability space, and for each N ∈ N, {(al,N,1, ar,N,1), (al,N,2, ar,N,2),··· , (al,N,N , ar,N,N )} be a sequence of two-faced pairs of random variables in A. An index tuple {(i(1), i(2),··· , i(n)) : i(j) = 1, 2,··· , N, j = 1, 2,··· , n} corresponds to a partition π of {1, 2,··· , n} if The following statements are equivalent. 4 p ∼π q ⇔ i(p) = i(q), p, q = 1, 2,··· , n. (1) For all n ∈ N, all π ∈ P(n), and all χ : {1, 2,··· , n} → {l, r}, limits Nπκχ,σ(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)) lim N→∞ Xσ≤π,σ∈P (χ) exist, where {i(1), i(2),··· , i(n)} is an index tuple corresponding to partition σ. (2) The limits limN→∞Nπκχ,σ(aχ(1),N,i(1), aχ(1),N,i(2),··· , aχ(n),N,i(n)) exist, for all {i(1), i(2),··· , i(n)} correspond to partition σ, all χ : {1, 2,··· , n} → {l, r} all σ ∈ P (χ)(n), σ ≤ π, all π ∈ P(n), and all n ∈ N. (3) For each n ∈ N, χ : {1, 2,··· , n} → {l, r}, limit lim N→∞ N κχ,1n (aχ(1),N,i, aχ(2),N,i,··· , aχ(n),N,i) exists, where i = 1, 2,··· , N . Proof. (2) ⇒ (1) is obvious, since the sum in (1) is a sum of finite terms. (1) ⇒ (2). For every n ∈ N, π = 1n, π = 1, by (1.2), we have lim N→∞ Xσ≤π,σ∈P (χ) N Xσ∈P (χ) = lim N→∞ = lim N→∞ Nπκχ,σ(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)) κχ,σ(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)) N ϕ(aχ(1),N,iaχ(2),N,i ··· aχ(n),N,i) exists, i = 1, 2,··· , N . It implies that for a partition π = {V1, V2,··· , Vd}, N ϕVj (aχ(1),N,i, aχ(2),N,i,··· , aχ(n),N,i) lim N→∞ Nπϕπ(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)) dYj=1 Nπκχ,π(aχ(1),N,i(1), aχ(1),N,i(2),··· , aχ(n),N,i(n)) = lim N→∞ exists. From this, we have for π ∈ P (χ)(n), by (1.1), N κχ,V (aχ(1),N,i(1), aχ(1),N,i,··· , aχ(n),N,i) lim N→∞ = lim = lim N→∞YV ∈π N→∞YV ∈π Xσ∈P (χV )(V ) =YV ∈π Xσ∈P (χV )(V ) ( lim N→∞ Nσϕσ((aχ(1),N,i(1), aχ(1),N,i(2),··· , aχ(n),N,i(n))V )µV (s−1 χ (σ), 1V ) N Nσ N Nσ Nσϕσ((aχ(1),N,i(1), aχ(1),N,i(2),··· , aχ(n),N,i(n))V )µV (s−1 × lim N→∞ exists. Therefore, for σ ∈ P (χ), σ ≤ π, π ∈ P(n), χ : {1, 2,··· , n} → {l, r}n, the limit χ (σ), 1V )) lim N→∞ exists. Nπκχ,σ(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)) = lim N→∞ Nπ Nσ lim N→∞ Nσκχ,σ(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)) (2) ⇒ (3) is obvious. (3) ⇒ (2) is also obvious, by the proof of (1) ⇒ (2). 5 (cid:3) The following lemma is a bi-free probability version of Lemma 13.2 in [NS]. Using equations (1.1) and (1.2), we can get a proof as same as that of Lemma 13.2 in [NS]. Lemma 2.2. Let (A, ϕ) be a non-commutative probability space, and for each N ∈ N, {(al,N,1, ar,N,1), (al,N,2, ar,N,2),··· , (al,N,N , ar,N,N )} be a sequence of two-faced pairs of random variables in A.Then the following two statements are equivalent. (1) For each n ∈ N, χ : {1, 2,··· , n} → {l, r}, limit lim N→∞ N κχ,1n (aχ(1),N,i, aχ(2),N,i,··· , aχ(n),N,i) exists, where i = 1, 2,··· , N . (2) The limit lim N→∞ N ϕ(aχ(1),N,iaχ(1),N,i ··· aχ(n),N,i) exists, for every i = 1, 2,··· , N , χ : {1, 2,··· , n} → {l, r}, and n ∈ N. If the above limits exist, they are equal to one another, i. e., lim N→∞ N κχ,1n (aχ(1),N,i, aχ(2),N,i,··· , aχ(n),N,i) = lim N→∞ N ϕ(aχ(1),N,iaχ(1),N,i ··· aχ(n),N,i). For each N ∈ N, let {aN,1 = (al,N,1, ar,N,1), aN,2 = (al,N,2, ar,N,2),··· , aN,N = (al,N,N , ar,N,N )} be a bi-free family of N identically distributed two-faced fairs of random variables in a non- commutative probability (A, ϕ). It follows that, for n ∈ N, χ : {1, 2,··· , n} → {l, r}, the moments ϕ(aχ(1),N,i ··· aχ(n),N,i), 1 ≤ i ≤ N, are independent of i. Let Sh,N =PN i=1 ah,N,i, for h ∈ {l.r}, and SN = (Sl,N , Sr,N ) be the two-faced pair of random variables in (A, ϕ). Then we have the following limit theorem. Theorem 2.3. The sequence of two-faced pairs {SN : N ≥ 1} converges in distribution to a two- faced pair b = (bl, br) in a non-commutative probability space (B, φ), as N → ∞, i. e., SN distr −→ (bl, br), N → ∞, if and only if for each n ≥ 1, and χ : {1, 2,··· , n} → {l, r}, the limit N ϕ(aχ(1),N,1aχ(2),N,1 ··· aχ(n),N,1) lim N→∞ exists. (2.1) (2.2) Furthermore, if the limits exist, then the joint distribution of the limit pair b = (bl, br) is deter- mined in terms of bi-free cumulants by κχ(b) := κχ,1n (bχ(1), bχ(2),··· , bχ(n)) = lim N→∞ N ϕ(aχ(1),N,1aχ(2),N,1 ··· aχ(n),N,1), for χ : {1, 2,··· , n} → {l, r}. 6 Proof. We follow the idea in the proof of Theorem 13.1 in [NS]. For an n ≥ 1, χ : {1, 2,··· , n} → {l, r}, we have lim N→∞ ϕχ(SN ) := lim N→∞ ϕ(Sχ(1),N Sχ(2),N ··· Sχ(n),N ) NXi(1),i(2),··· ,i(n)=1 ϕ(aχ(1),N,i(1)aχ(2),N,i(2) ··· aχ(n),N,i(n)) N (N − 1)··· (N − π + 1)ϕ(aχ(1),N,i(1)aχ(2),N,i(2) ··· aχ(n),N,i(n)) = lim N→∞ = lim N→∞ Xπ∈P(n) = Xπ∈P(n) N→∞ Xπ∈P(n) × lim N→∞ = lim N (N − 1)··· (N − π + 1) ( lim N→∞ Nπϕ(aχ(1),N,i(1)aχ(2),N,i(2) ··· aχ(n),N,i(n))) Nπ Nπϕ(aχ(1),N,i(1)aχ(2),N,i(2) ··· aχ(n),N,i(n)), where {i(1), i(2),··· , i(n)} is an index tuple corresponding to partition π. By the discussion in the proof of Theorem 13.1 in [NS], {SN : N ≥ 1} converges in distribution if and only if lim N→∞ Nπϕ(aχ(1),N,i(1)aχ(2),N,i(2) ··· aχ(n),N,i(n)) exists for all π ∈ P(n), where {i(1), i(2),··· , i(n)} is an index tuple corresponding to partition π. For a π ∈ P(n), by (1.2) and (1.3), we have Nπϕ(aχ(1),N,i(1)aχ(2),N,i(2) ··· aχ(n),N,i(n)) Nπκχ,σ(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)) Nπκχ,σ(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)) lim N→∞ = lim N→∞ Xσ∈P (χ) N→∞ Xσ∈P (χ),σ≤π = lim By Lemmas 2.1 and 2.2, we get that (2.1) is equivalent to (2.2). If the existence of the limits is assumed, then ϕ(Sχ(1),N Sχ(2),N ··· Sχ(n),N ) Nπϕ(aχ(1),N,i(1)aχ(2),N,i(2) ··· aχ(n),N,i(n)) lim N→∞ = lim = lim N→∞ Xπ∈P(n) N→∞ Xπ∈P(n) Xσ∈P (χ),σ≤π N→∞ Xπ∈P χ(n) = Xπ∈P χ(n) lim N→∞ = lim Nπκχ,σ(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)) Nπκχ,π(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)) Nπκχ,π(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)) 7 On the other hand, by equation (2.1), for every n ∈ N, χ : {1, 2,··· , n} → {l, r} φ(bχ(1), bχ(2),··· , bχ(n)) = Xπ∈P (χ)(n) = Xπ∈P χ(n) = lim N→∞ κχ,π(bχ(1), bχ(2),··· , bχ(n)) ϕ(Sχ(1),N Sχ(2),N ··· Sχ(n),N ) lim N→∞ Nπκχ,π(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)). By the definition of bi-free cumulants, the cumulants are determined uniquely by (1.2) (Proposition 5.2 in [MN]). Therefore, κχ,π(bχ(1), bχ(2),··· , bχ(n)) = lim N→∞ Nπκχ,π(aχ(1),N,i(1), aχ(2),N,i(2),··· , aχ(n),N,i(n)). Especially, by Lemma 2.2, κχ(b) :=κχ,1n (bχ(1), bχ(2),··· , bχ(n)) = lim N→∞ = lim N→∞ N κχ,1n (aχ(1),N,1, aχ(2),N,1,··· , aχ(n),N,1) N ϕ(aχ(1),N,1aχ(2),N,1 ··· aχ(n),N,1). Without any essential difficulties, we can generalize the above limit theorem to the multidimen- (cid:3) sional case. r,N,m)j∈J ) be a two-faced family, and {aN,1, aN,2,··· , aN,N} be a bi-free sequence of identically distributed two-faced families of random variables in a non-commutative probability (A, ϕ), where I and J are disjoint index sets. It follows the moments For each N ∈ N and 1 ≤ m ≤ N , let aN,m = (a(i) that, for n ≥ 1, χ : {1, 2,··· , n} → {l, r}, and α : {1, 2,··· , n} → IF J, such that χ−1(l) = α−1(I), l,N,m)i∈I , (a(j) ϕ(a(α(1)) χ(1),N,ma(α(2)) χ(2),N,m ··· a(α(n)) m=1 a(i) χ(n),N,m), 1 ≤ m ≤ N, are independent of m. Let S(i) SN = ((S(i) the following limit theorem. l,N )i∈I , (S(j) r,N,m, for j ∈ J, and r,N )j∈J ) be the two-faced family of random variables in (A, ϕ). Then we have l,N,m, for i ∈ I, S(j) m=1 a(j) r,N = PN l,N = PN Corollary 2.4. The following two statements are equivalent. (1) There is a two-faced family b = ((bl,i)i∈I , (br,j)j∈J ) in a non-commutative probability space (B, φ) such that SN distr→ b, as N → ∞. χ−1(l) = α−1(I), the limit (2) For each n ≥ 1, χ : {1, 2,··· , n} → {l, r}, and α : {1, 2,··· , n} → IF J, such that exists. lim N→∞ N ϕ(a(α(1)) χ(1),N,1a(α(2)) χ(2),N,1 ··· a(α(n)) χ(n),N,1) If the existence of the limits is assumed, then we have κχ,1n (bχ(1),α(1), bχ(2),α(2),··· , bχ(n),α(n)) = lim N→∞ N ϕ(a(α(1)) χ(1),N,1a(α(2)) χ(2),N,1 ··· a(α(n)) χ(n),N,1). As an application of the above limit theorems, we can prove Voiculescu's bi-free central limit theorem (7.9 in [Vo1]) in a special (but most popular) case. First let's recall Voiculescu's centered bi-free Gaussian distributions. 8 Let's use our limit theorems to prove the bi-free central limit theorem. Definition 2.5 (7.3 in [Vo1]). Let I and J be two disjoint index sets. A two-faced family z = ((zi)i∈I , (zj)j∈J ) in a non-commutative probability space (A, ϕ) has a bi-free centered Gaussian distribution if its cumulants satisfy κα,1n(z) = 0, for all α : {1, 2,··· , n} → IF J, and n ∈ N, n 6= 2. k ) = 0,∀k ∈ IF J and n ∈ N. Let Proposition 2.6 (7.9 in [Vo1]). Let z(n) = ((z(n) identically distributed two-faced families in (A, ϕ) such that ϕ(z(n) SN,k = 1√NPN k , for k ∈ IF J, N ∈ N, and SN = ((SN,i)i∈I , (SN,j)j∈J ). Then )j∈J ), n ∈ N, be a bi-free sequence of )i∈I , (z(n) n=1 z(n) j i SN distr→ b, where b = ((bi)i∈I , (bj)j∈J ) is a two-faced family in a non-commutative probability space (B, φ) having a bi-free centered Gaussian distribution such that κ(bkbl) = ϕ(z(n) n ∈ N. Proof. By Corollary 2.4, we need to show that limN→∞ N ϕ( ),∀k, l ∈ IF J, and ) exists, for all k : k z(n) z(m) k(n)√N z(m) k(1)√N z(m) k(2)√N ··· l {1, 2,··· , n} → IF J, n ≥ 1, and m ∈ N. In fact, κk,1n (b) = lim N→∞ N ϕ( z(m) k(1)√N z(m) k(2)√N ··· z(m) k(n)√N ) = lim N→∞ N N n/2 ϕ(z(m) k(1)z(m) k(2) ··· z(m) k(n)) = δn,2ϕ(z(m) k(1) ··· z(m) k(n)). (cid:3) 3. Bi-free infinitely divisible distributions The goal of this section is to define and study bi-free infinitely divisible distributions in a more general setting than that in [GHM]: we do not require that the random variable in the left face commute with that in the right face of a two-faced pair of random variables. First let's give the definition. Definition 3.1. A two faced pair a = (al, ar) of self-adjoint operators in a C∗-probability space (A, ϕ) has a bi-free infinitely divisible distribution if for each N ∈ N, there is a bi-free sequence of N identically distributed two-faced pairs {(al,N,i, ar,N.i) : i = 1, 2,··· , N} of self-adjoint operators in a C∗-probability space (AN , ϕN ) such that SN := (Sl,N , Sr,N ) := ( al,N,i, ar,N,i) NXi=1 NXi=1 bf = in (AN , ϕN ) has a distribution same as that of a = (al, ar) in (A, ϕ). Lemma 3.2. For each N ∈ N, let HN = H ⊕ ··· ⊕ H N copies of H g ⊕ g ⊕ ··· ⊕ g √N f ⊕ f ⊕ ··· ⊕ f {z √N . For f, g ∈ H, T ∈ B(H), let } ∈ HN ,bT = T ⊕ T ⊕ ··· ⊕ T ∈ B(HN ). (B(F (HN )), τHN ) same as that of ∆ := {l(f ), r(f ), l(g)∗, r(g)∗, Λl(T ), Λr(T ) : f, g ∈ H, T ∈ B(H)} in (B(F (H), τH). Then b∆ := {l(bf ), l(bg)∗, r(bf ), r(bg)∗, Λl(bT ), Λr(bT ) : f, g ∈ H, T ∈ B(H)} has a distribution in Proof. For any b ∈ ∆, τH(b) = 0 = τHN (bb). For b0, b1,··· bn, bn+1 ∈ ∆, n ≥ 0, τH(b0 ··· bn+1) 6= 0 implies that b0 = χ(g)∗, bn+1 = χ(f ), f, g ∈ H, χ ∈ {l, r}, and b1 ··· bnf = f0 ∈ H. Then ∈ HN ,bg = τH(b0 ··· bn+1) = hb1 ··· bnf, gi. 9 Very similarly, τHN (bb0 ···bbn+1) 6= 0 implies that τHN (bb0 ···bbn+1) = hbb1 ···bbnbf ,bgi. Since b1 ··· bnf ∈ H, there are k operators {χ(i)(fi) : fi ∈ H, χ(i) ∈ {l, r}, i = 1, 2,··· , k} and k operators {χ(i)(fi)∗ : fi ∈ H, χ(i) ∈ {l, r}, i = 1, 2,··· , k} among {b1, b2,··· , bn}, 2k ≤ n. Let's prove (3.1) by induction in k. When k = 0, it is obvious that (3.1) holds true, because b1,··· , bn are Λχ(T ), T ∈ B(H), χ ∈ {l, r}. When k = 1, after performing actions of Λχ(T ), T ∈ B(H), χ ∈ {l, r}, we can assume that there is no Λχ(T ) in {b1,··· , bn}. Therefore, n = 2. Case I. b1 = l(f1)∗, b2 = l(f2). or b1 = r(f1)∗, b2 = r(f2). In this case, we have hb1 ··· bnf, gi = hbb1 ···bbnbf ,bgi. Case II. b1 = l(f1)∗, b2 = r(f2). hb1b2f, gi = hf2, f1ihf, gi = hbf2,bf1ihbf ,bgi = hbb1bb2bf ,bgi. hb1b2f, gi = hf, f1ihf2, gi = hbf ,bf1ihbf2,bgi = hbb1bb2bf ,bgi. Case III. b1 = r(f1)∗, b2 = l(f2). The discussion is same as that in case II. Suppose (3.1) holds true for k < K. Now we consider the case that k = K. After performing actions of Λχ(T ), T ∈ B(H), χ ∈ {l, r}, we can assume that b1,··· , bn are creation or annihilation operators. Choose the largest index i such that bi = l(fi)∗ or r(fi)∗, and bi+1 = l(fi+1) or r(fi+1). Then bj = χ(fj), j = i + 2,··· , n, χ ∈ {l, r}. Case I. bi = l(fi)∗, bi+1 = l(fi+1). Since l(ξ)r(ζ) = r(ζ)l(ξ), for all ξ, ζ ∈ H, we can write bi+1 ··· bn = l(fi+1)l(ξ1)··· l(ξp)r(ζq )··· r(ζ1), p + q = n− i− 1. Then, by the inductive hypothesis, we have hb1 ··· bnf, gi = hfi+1, fiihb1 ··· bi−1bi+2 ··· bnf, gi = hbfi+1,bfiihbb1 ···bbi−1bbi+2 ···bbnbf ,bgi = hbb1 ···bbnbf ,bgi. Case II. bi = r(fi)∗, bi+1 = l(fi+1), and q > 0. By inductive hypothesis, we have hb1 ··· bnf, gi =hb1 ··· bi−1r(fi)∗l(fi+1)l(ξ1)··· l(ξp)r(ζq)r(ζq−1)··· r(ζ1)f, gi =hb1 ··· bi−1l(fi+1)l(ξ1)··· l(ξp)r(fi)∗r(ζq)r(ζq−1)··· r(ζ1)f, gi =hζq, fiihb1 ··· bi−1l(fi+1)l(ξ1)··· l(ξp)r(ζq−1)··· r(ζ1)f, gi =hbζq,bfiihbb1 ···bbi−1l(bfi+1)l(bξ1)··· l(bξp)r(bζq−1)··· r(bζ1)bf ,bgi =hbb1 ···bbi−1l(bfi+1)l(bξ1)··· l(bξp)r(bfi)∗r(bζq)r(bζq−1)··· r(bζ1)bf ,bgi =hbb1 ···bbi−1r(bfi)∗l(bfi+1)l(bξ1)··· l(bξp)r(bζq)··· r(bζ1)bf ,bgi = hbb1 ···bbnbf ,bgi. hb1 ··· bnf, gi =hb1 ··· bi−1r(fi)∗l(fi+1)l(ξ1)··· l(ξp)f, gi Case III. bi = r(fi)∗, bi+1 = l(fi+1), and q = 0. =hb1 ··· bi−1l(fi+1)l(ξ1)··· l(ξp−1)[l(ξp)r(fi)∗f ], gi =hf, fiihb1 ··· bi−1l(fi+1)l(ξ1)··· l(ξp−1)ξp, gi =hbf ,bfiihbb1 ···bbi−1l(bfi+1)l(bξ1)··· l(bξp−1)bξp,bgi =hbb1 ···bbi−1l(bfi+1)l(bξ1)··· l(bξp−1)[l(bξp)r(bfi)∗bf ],bgi =hbb1 ···bbi−1r(bfi)∗l(bfi+1)l(bξ1)··· l(bξp)bf ,bgi = hbb1 ···bbnbf ,bgi. Case IV. bi = r(fi)∗, bi+1 = r(fi+1). The proof is same as that for Case I. Case V. bi = l(fi)∗, bi+1 = r(fi+1), and p = 0. The proof is same as that for Case III. Case VI. bi = l(fi)∗, bi+1 = r(fi+1), and p > 0. The proof is same as that for Case II. (cid:3) The following corollary provides a particular type of bi-free infinitely divisible distributions. 10 Corollary 3.3. For f, g ∈ H, λ1, λ2 ∈ C, and T1, T2 ∈ B(H), we have the following conclusions. (1) Let b = ((bl,1, bl,2, bl,3), (br,1, br,2, br,3)) = ((l(f ), l(g)∗, Λl(T1)), (r(f ), r(g)∗, Λr(T2))) and B the set of all operators in b. Then the bi-free cumulants of elements in B have the following form. Let χ be a map from {1, 2,··· , n} into {l, r}, for n ∈ N. Then κχ(bi) = 0,∀bi ∈ B. For n = 2, the cumulant κχ(b1, b2) 6= 0 only if b1 = l(g)∗, or r(g)∗, and b2 = l(f ), or r(f ). In this case, For n > 2, κχ(b1 ··· bn) 6= 0 only if b1 = l(g)∗, or r(g)∗, bn = l(f ), or r(f ), and b2, b3,··· , bn−1 are Λh(Ti), i = 1, 2, and h ∈ {l, r}. In this case, κχ(b1, b2) = hf, gi. κχ(b1, b2,··· , bn) = hb2 ··· bn−1f, gi. (2) Let T1, T2 ∈ B(H) be self adjoint, and λ1, λ2 ∈ R. Then the two-faced pair a = (al, ar) := (l(f ) + l(f )∗ + Λl(T1) + λ11, r(g) + r(g)∗ + Λr(T2) + λ21) has a bi-free infinitely divisible distribution. Proof. (1). For N ∈ N, by Lemma 3.2, b has a distribution same as bb := ((l(bf ), l(bg)∗, Λl(bT1)), (r(bf ), r(bg)∗, Λr(bT2))). Furthermore,bb is the bi-free sum ofbb1 + ···bbN , where the i-th component of bf . All summands have the joint distribution of bbi = ((l(bfi), l(bgi)∗, Λl((bT1)i), (r(bfi), r(bgi)∗, Λr((bT2)i)),bfi = , Λl(T1)), ( , Λr(T2))). (( , r(f ) √N , r(g)∗ √N 0 ⊕ ··· ⊕ 0 ⊕ f ⊕ 0 ⊕ ··· ⊕ 0 , √N l(f ) √N l(g)∗ √N For bi ∈ B, let bN,i be the corresponding element in BN := { l(f ) √N , l(g)∗ √N , Λl(T1), r(f ) √N , r(g)∗ √N , Λr(T2)}. For b1, b2,··· , bn ∈ B, we have limN→∞ N ϕN (bN,1 ··· bN,n) 6= 0 implies that bN,n = l(f )/√N or bN,n = r(f )/√N , and bN,1 = l(g)∗/√N or bN,1 = r(g)∗/√N , and others bN,2,··· bN,n−1 are Λh(Ti), i = 1, 2, h ∈ {l, r}. In this case, by Corollary 2.4, we have κχ(b1,··· , bn) = hb2b3 ··· bn−1f, gi. (2). Without loss of generality, we can assume λ1 = λ2 = 0. For any N ∈ N, by Lemma 3.2, a andba = (l(bf ) + l(bf )∗ + Λl(cT1), r(bg) + r(bg)∗ + Λr(cT2)) have the same distribution. Moreover, let where bfi,bgi,bTi are the i-th summands of the direct sum vectors bf ,bg, and operator bT , respectively. Then by Proposition 1.1,ba1,ba2,···baN are bi-free. It is obvious thatba =ba1 + ··· +baN . Hence, a bai = (l(bfi) + l(bfi)∗ + Λl(cT1i), r(bgi) + l(bgi)∗ + Λr(cT2i)), has a bi-free infinitely divisible distribution. (cid:3) The following result is a corollary of Proposition 6.4.1 in [CNS2]. Lemma 3.4. Let a1, a2,··· an be n random variables in a non-commutative probability space (A, ϕ), and n ≥ 2. If ai = 1, for some i, 1 ≤ i ≤ n, then for every χ : {1, 2,··· , n} → {l, r}, π ∈ P (χ)(n) and {1} is not a block of π, κχ,π(a1,··· , an) = 0. 11 Definition 3.5. Let (A, ϕ) be a non-commutative probability space, Υ := {κχ : χ : {1, 2,··· , n} → {l, r}, n ≥ 1} be the set of all bi-free cumulant polynomials of (A, ϕ). We say that the cumulant set of a two-faced pair a = (al, ar) is conditionally non-negative definite if for every sequence χi : {1, 2,··· , i} → {l, r}, i = 1, 2,··· , k, in Υ, and α1, α2,··· , αk ∈ C, kXn,m=1 αnαmκχn⊔χm (a) ≥ 0, where κχn⊔χm(a) = κχn⊔χm (aχn(1),··· , aχn(n), aχm(m),··· , aχm(1)), χn ⊔ χm : {1, 2,··· , n + m} → {l, r}, χn ⊔ χm(i) =(χn(i), χm(m + n − i + 1), if 1 ≤ i ≤ n; if n < i ≤ n + m. Definition 3.6. Let a = (al, ar) be a pair of two self-adjoint operators in a C∗-probability space (A, ϕ). We say Υ := {κχ(a) : χ : {1, 2,··· , n} → {l, r}, n ≥ 1} is conditionally bounded if for λ ∈ {l, r}, there exists a positive number L such that kXn,m=1 αnαmκ(χn∪λ)⊔(χm∪λ)(aχn(1),··· , aχn(n), aλ, aλ, aχm(m),··· , aχm(1)) ≤ L kXn,m=1 αnαmκχn⊔χm (a), ∀χn : {1, 2,··· , n} → {l, r}, αn ∈ C, n = 1, 2,··· , k, k ≥ 1. Theorem 3.7. Let a = (al, ar) be a two-faced pair of self-adjoint operators in a C∗-probability space (A, ϕ). The following statements are equivalent. (1) a has a bi-free infinitely divisible distribution. (2) the bi-free cumulant set Υ(a) := {κχ(a) : χ : {1, 2,··· , n} → {l, r}, n ≥ 1} is conditionally (3) a is the limit in distribution of a sequence of triangular arrays in the limit theorem (The- orem 2.3): For each N ∈ N, there is a bi-free family {(al,N,i,, ar,N,i) : i = 1, 2,··· , N} of identically distributed two-faced pairs of self-adjoint operators in a C∗-probability space (AN , ϕN ) such that κχ(a) = lim N→∞ N ϕN (aχ(1),N,iaχ(2),N,i ··· aχ(n),N,i),∀χ : {1, 2,··· , n} → {l, r}, k ≥ 1, non-negative definite and conditionally bounded. where 1 ≤ i ≤ N . Proof. (3) ⇒ (2). For k ∈ N, αi ∈ C, i = 1, 2,··· , k, and a sequence {κχi ∈ Υ(a) : i = 1, 2,··· , k}, by Theorem 2.3, we have αnαmκχn⊔χm(a) kXn,m=1 = lim N→∞ N kXm,n=1 kXn=1 = lim N→∞ N ϕN (( αnαmϕN (aχn(1),N,i ··· aχn(n),N,iaχm(m),N,i ··· aχm(1),N,i) αnaχn(1),N,i ··· aχn(n),N,i)( αmaχm(1),N,i ··· aχm(m),N,i)∗) ≥ 0. kXm=1 12 Now we show that Υ(a) is conditionally bounded. For αn ∈ C, aχn = aχn(1)aχn(2) ··· aχn(n), χn : {1, 2,··· , n} → {l, r}, n = 1, 2,··· , k, k ≥ 1, and λ ∈ {l, r}, by the previous argument, we have kXn,m=1 = lim N→∞ ≤ka2 kXn,m=1 =L N ϕN (( kXn=1 λk lim N→∞ N ϕN (( αnαmκ(χn∪λ)⊔(χm∪λ)(a) αnaχn(1),N,i ··· aχn(n),N,i)a2 λ( kXn=1 αnaχn(1),N,i ··· aχn(n),N,i)( kXm=1 αmaχm(1),N,i ··· aχm(m),N,i)∗) kXm=1 αmaχm(1),N,i ··· aχm(m),N,i)∗) αnαmκχn⊔χm(a). We get the desired result with L = ka2 λk. and Xr without constant terms. Let (2) ⇒ (1). Let ChXl, Xri be the set of all polynomials in two (non-commutative) variables Xl Xχ := Xχ(1)Xχ(2) ··· Xχ(n),∀χ : {1, 2,··· , n} → {l, r}, n ≥ 1. By (2), we can define an inner product on ChXl, Xri by a sesquilinear extension of hXχn , Xχmi = κχm⊔χn (a),∀χi : {1, 2,··· , i} → {l, r}, i = m, n, m, n ≥ 1. We, thus, get a Hilbert space H after dividing out the kernel and completion. After identifying ChXl, Xri with its image in H, we may treat elements of ChXl, Xri as vectors in H and operators on H. Consider the C∗-probability space B(F (H), τH) and the operators b = (bl, br) = (l(Xl) + l(Xl)∗ + Λl(Xl) + κ(al)1, r(Xr) + r(Xr)∗ + Λl(Xr) + κ(ar)1), where Xλ, λ ∈ {l, r}, is the right multiplication operator of Xλ on H, that is, Xλ(Xχ(1)Xχ(2) ··· Xχ(n)) = Xχ(1) ··· Xχ(n)Xλ, for Xχ(1) ··· Xχ(n) ∈ ChXl, Xri ⊂ H. By (2), Xλ is a bounded operator on H, therefore, Λ(Xλ) ∈ B(F (H)). Now we show that Λ(Xλ) is self adjoint on F (H). For χ : {1, 2,··· , n} → {l, r}, δ : {1, 2,··· , m} → {l, r}, we have hXχ(1) ··· Xχ(n), Λ(Xλ)∗Xδ(1) ··· Xδ(m)i =hΛ(Xλ)Xχ(1) ··· Xχ(n), Xδ(1) ··· Xδ(m)i =hXχ(1) ··· Xχ(n)Xλ, Xδ(1) ··· Xδ(m)i =κδ⊔(χ⊔λ)(aδ(1),··· , aδ(m), aλ, aχ(n),··· , aχ(1)) =hXχ(1) ··· Xχ(n), Xδ(1) ··· Xδ(m)Xλi =hXχ(1) ··· Xχ(n), Λ(Xλ)Xδ(1) ··· Xδ(m)i It implies that Λ(Xλ) = Λ(Xλ)∗. Thus, bl and br are self adjoint operators in B(F (H)). With the same method, we can prove that in decompositionba =ba1 +···+baN in Corollary 3.3 (2), eachbai is a pair of two self adjoint operators in a C∗-probability space for i = 1, 2,··· , N, N ∈ N. Hence, by Corollary 3.3, b = (bl, br) has a bi-free infinitely divisible distribution in sense of Definition 3.1. Finally, we show that a's distribution in (A, ϕ) is same as that of b in (B(F (H)), τH). It is obvious that κ(bχ) = κ(aχ), χ ∈ {l, r}. 13 For n ≥ 2, let eb = (bl, br) = ((bl,1 + bl,2 + bl,3), (br,1, br,2, br,3)) By Lemma 3.4, b and eb have the same bi-free cumulants κχ, for any map χ : {1, 2,··· , n} → {l, r}, n ≥ 2. For n = 2, by Corollary 3.3, we have := ((l(Xl) + l(Xl)∗ + Λl(Xl)), (r(Xr) + r(Xr)∗ + Λr(Xr))). κχ(bl, br) = κχ(bl,i, br,j) = hXr, Xli = κχ(al, ar) 3Xi,j=1 and, similarly, κ(br, bl) = κ(ar, al). For n > 2, χ : {1, 2,··· , n} → {l, r}, we have κχ(bχ(1), bχ(2),··· , bχ(n)) = κχ(bχ(1),α(1),··· , bχ(n),α(n)) 3Xα(1),α(2),··· ,α(n)=1 =hbχ(2),3 ··· bχ(n−1),3Xχ(n), Xχ(1)i =hΛχ(2)(Xχ(2))··· Λχ(n−1)(Xχ(n−1))Xχ(n), Xχ(1)i =hXχ(n)Xχ(n−1) ··· Xχ(2), Xχ(1)i =κχ(aχ(1), aχ(2),··· , aχ(n)). Hence, a has a bi-free infinitely divisible distribution. (1) ⇒ (3) is obvious. For each N ∈ N, choose a bi-free sequence {(al,N,1, ar,N,1), (al,N,2, ar,N,2)··· , (al,N,N , ar,N,N )} of identically distributed two-faced pairs of self-adjoint operators in a C∗-probability space (AN , ϕN ) i=1 ar,N,i) has a distribution same as that of a. It is trivial to see that SN converges in distribution to a, as N → ∞. (cid:3) such that SN = (PN i=1 al,N,i,PN In this section, we investigate the relation between bi-free Levy processes and bi-free infinitely 4. Bi-free Levy Processes divisible distributions. Let's give the definition of bi-free Levy processes first. Definition 4.1. A family {at = (al,t, ar,t) : t ≥ 0} of two-faced pairs of self-adjoint operators in a C∗-probability space (A, ϕ) is called a bi-free Levy process if it satisfies the following conditions. (1) a0 = (0, 0). (2) If 0 ≤ t1 < t2 < ··· < tn < ∞, then at2 − at1,··· , atn − atn−1 are bi-free, where at − as = (3) For 0 < s < t, the distribution of at − as depends only on t − s. (4) The distribution µt of at converges to 0, as t → 0+. (al,t − al,s, ar,t − ar,s). Theorem 4.2. Let a = (al, ar) and at = (al,t, ar,t) be two-faced pairs of self-adjoint operators in a C∗-probability space (A, ϕ). divisible distribution. (1) Let {at = (al,t, ar,t) : t ≥ 0} be a bi-free Levy process. Then a1 has a bi-free infinitely (2) If a = (al, ar) has a bi-free infinitely divisible distribution, then there is a bi-free Levy process {bt = (bl,t, br,t) : t ≥ 0} in a C∗-probability space (B, φ) such that a and b1 have the same distribution. Proof. For 0 < s, t, at+s = (at+s − as) + as. By the definition of bi-free Levy processes, (at+s − as) and as are bi-free, and at+s − as and at have the same distribution. Generally, for any n ∈ N, a1 = a1/n + (a1 − a1/n) = ··· = a1/n + 14 (a(i+1)/n − ai/n). n−1Xi=1 Hence, a1 has a bi-free infinitely divisible distribution. Let a = (al, ar) has a bi-free infinitely divisible distribution. By The proof of Theorem 3.7, we can choose a as a = (l(Xl) + l(Xl)∗ + Λl(Xl) + κ(al)1, r(Xr) + r(Xr)∗ + Λl(Xr) + κ(ar)1) on the C∗-probability space (B(F (H)), τH), where H is the Hilbert space obtained from the polyno- mial set ChXl, Xri by a special sesquilinear form defined in the proof of Theorem 3.7. The following construction is adapted from the proof of Theorem 4.2 in [GHM] (originally, from [GSS]). Define a new Hilbert space K = L2(R+, dx) ⊗ H, where R+ = [0,∞). For a Borel set I ⊂ R+, let χI be the characteristic function of I in L2(R+, dx). MI be the multiplication operator of χI on L2(R+, dx). Define Xl,t = χ[0,t) ⊗ Xl, Xr,t = χ[0,t) ⊗ Xr, Al,t = M[0,t) ⊗ M (Xl), Ar,t = M[0,t) ⊗ M (Xr), where M (Xχ) : H → H is the right multiplication operator of Xχ on H, χ ∈ {l, r}. Then consider the operators bt = (l(Xl,t) + l(Xl,t)∗ + Λl(Al,t) + tκ(al), r(Xr,t) + r(Xr,t)∗ + Λr(Ar,t) + tκ(ar)), t > 0, b0 = (0, 0) on the C∗-probability space B(F (K), τK). For 0 < s < t, define Xα,s,t = χ[s,t) ⊗ Xα, Aα,s,t = M[s,t) ⊗ M (Xα), α ∈ {l, r}. Then bt − bs := (cl, cr) := (cl,1 + cl,2 + cl,3 + (t − s)κ(al), cr,1 + cr,2 + cr,3 + (t − s)κ(ar)) = (l(Xl,s,t) + l(Xl,s,t)∗ + Λl(Al,s,t) + (t− s)κ(al), r(Xr,s,t) + r(Xr,s,t)∗ + Λr(Ar,s,t) + (t− s)κ(ar)). Let 0 < t0 < t1 < ··· < tn. Define K0 = L2([0, t0), dx) ⊗ H,Ki = L2([ti−1, ti), dx) ⊗ H, i = 1, 2,··· , n. Then K =Ln i=0 Ki. Let Bi and Ci be the unital C∗-algebras generated by {l(f ) : f ∈ Ki}∪{Λl(T ) : T ∈ B(K), TKi ⊂ Ki, TK⊖Ki = 0} and {r(f ) : f ∈ Ki} ∪ {Λr(T ) : T ∈ B(K), TKi ⊂ Ki, TK⊖Ki = 0}, respectively. By Proposition 1.1, {(Bi,Ci) : 0 ≤ i ≤ n} is bi-free in (B(F (K)), τK). Since Xα,t0 = χ[0,t0) ⊗ Xα ∈ K0, and Aα,t0 = M[0,t0) ⊗ M (Xα) : K0 → K0, and Aα,t0K⊖K0 = 0, for α ∈ {l, r}, we have bt0 ∈ (B0,C0). Very similarly, bt1 − bt0 ∈ (B1,C1),··· , btn − btn−1 ∈ (Bn,Cn). It implies that bt0, bt1 − bt0,··· , bn − bn−1 are bi-free. It is obvious that κ(cχ) = (t−s)κ(aχ), χ ∈ {l, r}. For n ≥ 2, by Corollary 3.3, κχ(cχ(1),··· , cχ(n)) 6= 0 if and only if κχ(aχ(1),··· , aχ(n)) 6= 0. In this case, we have κχ(cχ(1),··· , cχ(n)) =hcχ(2),3 ··· cχ(n−1),3Xχ(n),s,t, Xχ(1),s,ti =κχ(aχ(1), aχ(2),··· , aχ(n))hχ[s,t), χ[s,t)iL2(R+,dx) =κχ(aχ(1), aχ(2),··· , aχ(n))(t − s). The above discussion also shows that κχ(bt) = tκχ(a),∀χ : {1, 2,··· , n} → {l, r}, n ≥ 1. Thus, µt → 0, as t → 0+. It follows that {bt : t ≥ 0} is a bi-free Levy process, and κχ(b1) = κχ(a), for all χ : {1, 2,··· , n} → {l, r}, n ≥ 1. (cid:3) References [CNS1] I. Charlesworth, B. Nelson, and P. Skoufranis. On Two-faced families of non-commutative random variables, to appear in Canadian J. of Math., 2015, 26 pages. [CNS2] I. Charlesworth, B. Nelson, and P. Skoufranis. Combinatorics for bi-freeness with amalgamation. Commun. Math. Phys. 338, 801-847 (2015). [GSS] P. Glockner, M. Schurmann, and R. Speicher. Realization of free white noise. Arch. Math., Vol. 58(1992), 407-416. [GHM] Y. Gu, H. Huang, and J. Mingo. An analogue of the Levy-Hinchin formula for bi-free infinitely divisible distributions. arXiv:1501.05369v2 [math.OA], 9 July 2015. [KR] R. Kadison and J. Ringrose. Fundamentals of the theory of operator algebras. Graduate Studies in MAth. Vol. 16, AMS, 1997. 15 [MN] M. Mastnak and A. Nica. Double-ended queues and joint moments of left-right canonical operators on full Fock spaces. Internat. J. Math. 26 (2015), no. 2, 1550016, 34 pp. [NS] A. Nica and R. Speicher. Lectures on Combinatorics for Free Probability, LMS Lecture Notes 335, Cambridge University Press, 2006. [PSa] K. Sato. Levy Processes and infinitely divisible distributions. Cambrisge University Press, 1999. [PS] P. Skoufranis. Independences and partial R-transforms in bi-free probability, to appear in Annales de l'Institut Henri Poincare (H) Probabilites et Statistiques (2015), 31 pages. [Vo1] D. Voiculescu. Free Probability form Pairs of faces I. Comm. Math. Phys. 332 (2014), no. 3, 955-980. [Vo2] D. Voiculescu. Free Probability form Pairs of faces II. arXiv:1308.2035v1, [Math.OA], August, 2013. [VDN] D. Voiculescu, K. Dykema, and A. Nica. Free Random variables. CRM Monograph Series, Vol. 1, AMS, 1992. DEPARTMENT OF MATHEMATICS, LOUISIANA COLLEGE, PINEVILLE, LA 71359, USA, EMAIL: [email protected] 16
1109.4473	1	1109	2011-09-21T04:27:27	$K$-theory of Furstenberg transformation group $C^*$-algebras	[ "math.OA" ]	The paper studies the $K$-theoretic invariants of the crossed product $C^{*}$-algebras associated with an important family of homeomorphisms of the tori $\Bbb{T}^{n}$ called {\em Furstenberg transformations}. Using the Pimsner-Voiculescu theorem, we prove that given $n$, the $K$-groups of those crossed products, whose corresponding $n\times n$ integer matrices are unipotent of maximal degree, always have the same rank $a_{n}$. We show using the theory developed here, together with two computing programs - included in an appendix - that a claim made in the literature about the torsion subgroups of these $K$-groups is false. Using the representation theory of the simple Lie algebra $\frak{sl}(2,\Bbb{C})$, we show that, remarkably, $a_{n}$ has a combinatorial significance. For example, every $a_{2n+1}$ is just the number of ways that 0 can be represented as a sum of integers between $-n$ and $n$ (with no repetitions). By adapting an argument of van Lint (in which he answered a question of Erd\"os), a simple, explicit formula for the asymptotic behavior of the sequence $\{a_{n}\}$ is given. Finally, we describe the order structure of the K_{0}-groups of an important class of Furstenberg crossed products, obtaining their complete Elliott invariant using classification results of H. Lin and N. C. Phillips.	math.OA	math	K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS KAMRAN REIHANI Abstract. The paper studies the K-theoretic invariants of the crossed product C ∗-algebras associated with an important family of homeomorphisms of the tori Tn called Furstenberg transformations. Using the Pimsner-Voiculescu theorem, we prove that given n, the K-groups of those crossed products, whose corresponding n × n integer matrices are unipotent of maximal degree, always have the same rank an. We show using the theory developed here, together with two computing programs - included in an appendix - that a claim made in the literature about the torsion subgroups of these K-groups is false. Using the representation theory of the simple Lie algebra sl(2, C), we show that, remarkably, an has a combinatorial significance. For example, every a2n+1 is just the number of ways that 0 can be represented as a sum of integers between −n and n (with no repetitions). By adapting an argument of van Lint (in which he answered a question of Erdos), a simple, explicit formula for the asymptotic behavior of the sequence {an} is given. Finally, we describe the order structure of the K0-groups of an important class of Furstenberg crossed products, obtaining their complete Elliott invariant using classification results of H. Lin and N. C. Phillips. Contents Introduction 1. 2. K-groups of C(Tn) ⋊α Z 3. Anzai transformation group C ∗-algebras An,θ 4. A Poincar´e type of duality 5. The rank an of the K-groups of An,θ 6. Combinatorial properties of the sequence {an} 6.1. Connections with representation theory of sl(2, C) 6.2. Generating functions for the sequence {an} 7. The positive cone of K0(Fθ,f ) 8. Concluding remarks Appendices Appendix A. The Smith normal form Appendix B. Nilpotent and unipotent linear mappings Appendix C. Endomorphisms and derivations of exterior algebras Appendix D. Representation theory of sl(2, C) Appendix E. Some computer codes and a counterexample Appendix F. Applications to C ∗(Dn) References 2 5 7 10 11 12 12 14 18 21 22 22 23 24 25 26 27 29 2000 Mathematics Subject Classification. Primary: 19K14, 19K99, 46L35, 46L80, Secondary: 05A15, 05A16, 05A17, 15A36, 17B10, 17B20, 37B05, 54H20. 1 2 KAMRAN REIHANI 1. Introduction Furstenberg transformations were introduced in [9] as the first examples of homeomorphisms of the tori, which under some necessary and sufficient conditions are minimal and uniquely ergodic. In some sense, they generalize the irrational rotations on the circle. They also appear in certain applications of ergodic theory to number theory (e.g. in Diophantine approximation [8]), and sometimes are called skew product transformations or compound skew translations of the tori. The terminology "Furstenberg transformation group C ∗-algebra" is what we would like to use in this paper to call the crossed products associated with Furstenberg transformations, and we will denote them by Fθ,f . There have been several contributions to the computations of K-theoretic invariants for some examples of these C ∗-algebras in the literature (see [16, 18, 25, 30, 39] to name a few). However, a more general study of such invariants for these C ∗-algebras has not been available to the best of our knowledge. Remark 1.1. In independent (unpublished) work [16], R. Ji studied the K-groups of the C ∗-algebras Fθ,f (denoted by AFf,θ in there) associated with the descending affine Furstenberg transformations (denoted by Ff,θ in there) on the tori. He comments that "explicitly computing the K-groups of C(Tn) ⋊K Z [AFf,θ for θ = 0] is still not an easy matter". Moreover, he gives no information about the ranks of the K-groups or order structure of K0 in general, which are studied in the present paper. As we shall see in Remark 1.7 below, the claim that he makes about the form of the torsion subgroup of K∗(Fθ,f ) is unfortunately not correct. From the C ∗-algebraic point view, when a Furstenberg transformation is minimal and uniquely ergodic, the associated transformation group C ∗-algebra is simple and has a unique tracial state with a dense tra- cial range of the K0-group in the real line. Because of this, these C ∗-algebras fit well into the classification program of G. Elliott by finding their K-theoretic invariants. In fact, in the class of transformation group C ∗-algebras of uniquely ergodic minimal homeomorphisms on infinite compact metric spaces, K-theory is a complete invariant. More precisely, suppose that X is an infinite compact metric space with finite cov- ering dimension and h : X → X is a uniquely ergodic minimal homeomorphism, and put A := C(X) ⋊h Z. Let τ be the trace induced by the unique invariant probability measure. Then τ is the unique tracial state on A. Let τ∗ : K0(A) → R be the induced homomorphism on K0(A) and assume that τ∗K0(A) is dense in R. Then the 4-tuple (K0(A), K0(A)+, [1A], K1(A)) is a complete algebraic invariant (called the Elliott invariant of A) [21, Corollary 4.8]. In this case, A has stable rank one, real rank zero and tracial topological rank zero in the sense of H. Lin [19]. The order on K0(A) is also determined by the unique trace τ , in the sense that an element x ∈ K0(A) is positive if and only if either x = 0 or τ∗(x) > 0 [22, 29]. This implies, in particular, that the torsion subgroup of K0(A) contributes nothing interesting to the order information. In other words, the order on K0(A) is determined by the order on the free part. We will study the order structure of K0(Fθ,f ) in Section 7. In order to compute the K-groups of a crossed product of the form C(Tn) ⋊α Z in general, we make use of the algebraic properties of K∗(C(Tn)) in Section 2. More precisely, K∗(C(Tn)) is an exterior algebra over Zn with a certain natural basis, and the induced automorphism α∗ on K∗(C(Tn)) is in fact a ring automorphism, which makes computations much easier. In fact, it is shown in Theorem 2.1 that the problem of finding the K-groups of the transformation group C ∗-algebra of a homeomorphism of the n-torus is completely computable in the sense that one only needs to calculate the kernels and cokernels of a finite number of integer matrices. These K-groups are finitely generated with the same rank (see Corollary 2.2). In the special case of Anzai transformation group C ∗-algebras An,θ associated with Anzai transformations on the n-torus, we denote this common rank by an, which we will study in detail in this paper. It is proved in Theorem 5.1 that an is the common rank of the K-groups of a larger class of trans- formation group C ∗-algebras, including the C ∗-algebras associated with Furstenberg transformations on Tn. We describe an as the constant term in a certain Laurent polynomial (Theorem 6.7). Then we study K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 3 the combinatorial properties of the sequence {an}, which leads to a simple asymptotic formula. To present the results and proofs of this paper we need some definitions about transformations on the tori and the corresponding C ∗-crossed products. Throughout this paper, Tn denotes the n-dimensional torus with coordinates (ζ1, ζ2, . . . , ζn). Definition 1.2. An affine transformation on Tn is given by β(ζ1, ζ2, . . . , ζn) = (e2πit1 ζ b11 1 . . . ζ b1n n , e2πit2ζ b21 1 . . . ζ b2n n , . . . , e2πitn ζ bn1 1 . . . ζ bnn n ), where t := (t1, t2, . . . , tn) ∈ (R/Z)n and B := [bij]n×n ∈ GL(n, Z). We identify the pair (t, B) with β. Note that any automorphism of Tn followed by a rotation can be expressed in such a fashion. The set of affine transformations on Tn form a group Aff(Tn), which can be identified with the semidirect product (R/Z)n ⋊ GL(n, Z). More precisely, for two affine transformations β = (t, B) and β′ = (t′, B′) on Tn, we have β ◦ β′ = (t + Bt′, BB′) and β−1 = (−B−1t, B−1). (In the expression Bt, t is a column vector, but for convenience we write it as a row vector.) We remind the reader of an important fact before giving the next definition. Consider homotopy classes of continuous functions from Tn to T. It is well known that in each class there is a unique "linear" function (ζ1, . . . , ζn) 7→ ζ b1 n for some b1, . . . , bn ∈ Z. More precisely, every continuous function f : Tn → T can be written as 1 . . . ζ bn f (ζ1, . . . , ζn) = e2πig(ζ1 ,...,ζn)ζ b1 1 . . . ζ bn n , for some continuous function g : Tn → R and unique integer exponents b1, . . . , bn. In particular, the cohomotopy group π1(Tn) is isomorphic to Zn. Following [8, p. 35], we denote the exponent bi, which is uniquely determined by the homotopy class of f , as bi = Ai[f ]. Definition 1.3. We define the following transformations in accordance with [16]. (a) A Furstenberg transformation ϕθ,f on Tn is given by ϕ−1 θ,f (ζ1, ζ2, . . . , ζn) = (cid:0)e2πiθζ1, f1(ζ1)ζ2, f2(ζ1, ζ2)ζ3, . . . , fn−1(ζ1, . . . , ζn−1)ζn(cid:1) , where θ is a real number, each fi : Ti → T is a continuous function with Ai[fi] 6= 0 for i = 1, . . . , n − 1, and f = (f1, . . . , fn−1). (b) An affine Furstenberg transformation α on Tn is given by 2 ζ3, . . . , ζ b1n α−1(ζ1, ζ2, . . . , ζn) = (e2πiθζ1, ζ b12 1 ζ2, ζ b13 1 ζ b23 1 ζ b2n 2 . . . ζ bn−1,n n−1 ζn), where θ is a real number and the exponents bij are integers and bi,i+1 6= 0 for i = 1, . . . , n − 1. (c) An ascending Furstenberg transformation α on Tn is given by 2 ζ3, . . . , ζ kn−1 α−1(ζ1, ζ2, . . . , ζn) = (e2πiθζ1, ζ k1 1 ζ2, ζ k2 n−1 ζn), where θ is a real number and the exponents ki are nonzero integers and ki ki+1 for i = 1, . . . , n−2. (d) In (c), if ki = 1 for i = 1, . . . , n − 1, the transformation is called an Anzai transformation σn,θ on Tn. Thus it is given by σ−1 n,θ(ζ1, ζ2, . . . , ζn) = (e2πiθζ1, ζ1ζ2, . . . , ζn−1ζn), where θ is a real number. We usually drop the indices n and θ and write only σ for more convenience. Note that one can easily verify that ϕθ,f is a homeomorphism. Also in the above definition, we have con- verted "descending", which is used in [16, Definition 2.16], to "ascending" since the order of coordinates there is opposite to ours. For certain Furstenberg transformations on Tn we have the following theorem. 4 KAMRAN REIHANI Theorem 1.4. ([9], 2.3) If θ is irrational, then ϕθ,f defines a minimal dynamical system on Tn. If in addition, each fi satisfies a uniform Lipschitz condition in ζi for i = 1, . . . , n − 1, then ϕθ,f is a uniquely ergodic transformation and the unique invariant measure is the normalized Lebesgue measure on Tn. In particular, every affine Furstenberg transformation defines a minimal and uniquely ergodic dynamical system if θ is irrational. As a conclusion, we have the following result for the Furstenberg transformation group C ∗-algebra Fθ,f := C(Tn) ⋊ϕθ,f Z as introduced in [16]. Corollary 1.5. Fθ,f = C(Tn) ⋊ϕθ,f Z is a simple C ∗-algebra for irrational θ. If in addition, each fi satisfies a uniform Lipschitz condition in ζi for i = 1, . . . , n − 1, then Fθ,f has a unique tracial state. Proof. For the first part, the minimality of the action as stated in the preceding theorem implies the simplicity of Fθ,f [4, 32]. For the second part, one can easily check that since θ is irrational, the action of Z on Tn generated by ϕθ,f is free. So there are no periodic points in Tn. This and the unique ergodicity of ϕθ,f yield the result [38, Corollary 3.3.10, p. 91]. (cid:3) Remark 1.6. Using the preceding corollary and much like the proof of Theorem 2.1 in [33], one can prove that for irrational θ, Fθ,f is in fact the unique C ∗-algebra generated by unitaries U, V1, . . . , Vn satisfying the commutator relations [U, V1] = e2πiθ, [U, V2] = f1(V1), . . . , [U, Vn] = fn−1(V1, . . . , Vn−1), (CR)f where [a, b] := aba−1b−1 and all other pairs of operators from U, V1, . . . , Vn commute. Remark 1.7. In [16, Proposition 2.17], R. Ji claims to have proved (∗) If ϕθ,f is an ascending Furstenberg transformation on Tn with the ascending sequence {k1, k2, . . . , kn−1}, then the torsion subgroup of K∗(Fθ,f ) is isomorphic to Zk1 ⊕ Z(m2) Z(mi) is the direct product of mi copies of the cyclic group Zki = Z/kiZ. , where the group ki ⊕ . . . ⊕ Z(mn−1) kn−1 k2 From this claim one would immediately deduce that the K-groups of the C ∗-algebra An,θ := C(Tn) ⋊σ Z generated by an Anzai transformation σ on Tn should be torsion-free. However, we will show that this is not true in general. This type of example first appears for n = 6, which seems already beyond hand calculation. (We admit that hand calculation would be the most convincing method to use; however, it is not practicable.) As the first counterexample we obtained by computer, we will see in Example E.1 that K1(A6,θ) ∼= Z13 ⊕ Z2 (also, see Example 3.3). In fact, the error in the proof of (∗) is in [16, p. 29, l.2]; there it is "clearly" assumed that using a matrix S in GL(2n, Z), one can delete all entries denoted by ⋆'s in K∗ − I, where K∗ is the 2n × 2n integer matrix corresponding to Fθ,f that acts on K∗(C(Tn)) = Λ∗Zn with respect to a certain ordered basis. This error arose originally from the general form of the matrix K∗ in [16, p. 27], which is not correct. R. Ji went on to use the torsion subgroup in (∗) as an invariant to classify the C ∗-algebras generated by ascending transformations and matrix algebras over them [16, Theorem 3.6]. We do not know whether those classifications holds. Question. Do there exist two different ascending Furstenberg transformations with the same parameter θ and with isomorphic transformation group C ∗-algebras? It is worth mentioning that explicit calculations of K-groups in terms of the parameters involved are possible in low dimensions (of the tori), and answer the question raised above negatively. However, such calculations in terms of the given integer parameters (exponents) of the ascending Furstenberg transfor- mation become quickly cumbersome and impossible in higher dimensions. We have used the computer codes in Appendix E for several numerical values of the parameters in higher dimensions, and we have not found any examples leading to the negative answer to this question yet. The torsion subgroups of K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 5 the K-groups are usually larger than what R. Ji claimed in [16, Proposition 2.17]; it seems likely that the torsion subgroups depend on the parameters involved in such a way that Ji's classification result is still true. We are currently investigating this problem. This paper is organized as follows. In Section 2, we review the general approach of exterior algebras for finding K-groups of transformation group C ∗-algebras of homeomorphisms of the tori. In Section 3, we apply this method to the important case of Anzai transformations and give the K-groups of their transformation group C ∗-algebras based on the tori of dimension up to 12 in Table 1 using the computer programs given in Appendix E. In Section 4, we establish a Poincar´e type of duality for the cokernels of integer matrices that leads to some interesting facts about the K-groups when the dimension of the underlying torus is odd. In Section 5, we focus on the rank an of the K-groups of Anzai transformations group C ∗-algebras based on the n-torus, and we show that an is, in fact, the rank of the K-groups of a large class of transformation group C ∗-algebras including those associated with Furstenberg transformations on the n-torus. In Section 6, we first uncover an interesting connection between studying an and the irreducible representations of the Lie algebra sl(2, C). This leads to a formula for an in terms of certain partitions of integers. Then we use this formula to show several interesting combinatorial properties of the sequence {an}. In Section 7, we study the order structure of the K0-group of a class of simple Furstenberg transformation group C ∗-algebras to make their Elliott invariants more accessible. The appendices at the end are provided mainly for self-containment of the paper, but we sometimes refer to them for the proof of some lemmas or propositions that are somewhat far from the main concepts and goals of this paper by nature. Appendix F contains some applications of the results of this paper to our earlier work [33]. 2. K-groups of C(Tn) ⋊α Z In this section, we describe a general method to compute the K-groups of C(Tn) ⋊α Z, where α is an arbitrary homeomorphism of Tn. (By abuse of notation, the automorphism α of C(Tn) is defined by α(f ) = f ◦ α−1 for f ∈ C(Tn).) To do this, we will pay special attention to the algebraic structure of K ∗(Tn) and how the induced automorphisms on it can be realized. Note that it is sufficient to consider the special case of "linear" homeomorphisms since, as stated before Definition 1.3, every continuous func- tion f : Tn → T is homotopic to a unique "linear" function (ζ1, . . . , ζn) 7→ ζ b1 n for some integer exponents b1, . . . , bn. Moreover, the K-groups of C(Tn) ⋊α Z depend (up to isomorphism) only on the homotopy class of α [3, Corollary 10.5.2]. 1 . . . ζ bn It is well known that K ∗(Tn) is a Z2-graded ring, and by the Kunneth formula (see [2, Corollary 2.7.15] or [36, Theorem 4.1]), it is an exterior algebra (over Z) on n generators, where the elements of even degree are in K 0(Tn) and those of odd degree are in K 1(Tn). The generators of this exterior algebra correspond to the generators of the dual group Zn of Tn [37, p. 185]. Indeed, in this case the Chern character is integral and gives the Chern isomorphisms ch : K ∗(Tn) −→ H ∗(Tn, Q) ch0 : K 0(Tn) −→ H even(Tn, Z), ch1 : K 1(Tn) −→ H odd(Tn, Z), where H ∗(Tn, Z) ∼= Λ∗ H k(Tn, Z) ∼= Λk the class in K1(C(Tn)) of the coordinate function zi : Tn → T given by Z(e1, . . . , en) is the ( Cech) cohomology ring of Tn under the cup product, and Z(e1, . . . , en). On the other hand, K ∗(Tn) ∼= K∗(C(Tn)). So by introducing ei := [zi]1, i.e. as a unitary element of of C(Tn) for i = 1, . . . , n, we have the isomorphisms K∗(C(Tn)) ∼= Λ∗ Z(e1, . . . , en) ∼= Λ∗Zn, which respect the canonical embedding of Zn. Moreover, these isomorphisms are unique since only zi(ζ1, . . . , ζn) = ζi 6 KAMRAN REIHANI the identity automorphism of the ring Λ∗Zn fixes each element of Zn. Now, we use the Pimsner-Voiculescu six term exact sequence [31] as the main tool for computing the K-groups of C(Tn) ⋊α Z. Let α∗(= K∗(α)) be the ring automorphism of K∗(C(Tn)) induced by α and let αi be the restriction of α∗ on Ki(C(Tn)) for i = 0, 1, and set A := C(Tn) ⋊α Z. Then we have the following exact sequence. K0(C(Tn)) α0−id−−−−→ K0(C(Tn)) 0−−−−→ K0(A) (1) expx K1(A) 1←−−−− K1(C(Tn)) ∂ y α1−id←−−−− K1(C(Tn)) Here,  : C(Tn) → A is the canonical embedding of C(Tn) in A, 0 := K0() and 1 := K1(). Also, from now on id denotes the identity function on each underlying set. As a result, we have the following short exact sequences (2) (3) 0 −→ coker(α0 − id) −→ K0(C(Tn) ⋊α Z) −→ ker(α1 − id) −→ 0, 0 −→ coker(α1 − id) −→ K1(C(Tn) ⋊α Z) −→ ker(α0 − id) −→ 0. Since all the groups involved are abelian and finitely generated, and ker(αi − id) is torsion-free (i = 0, 1), these short exact sequences split (since projective Z-modules are precisely free abelian groups), and we have (4) (5) K0(C(Tn) ⋊α Z) ∼= coker(α0 − id) ⊕ ker(α1 − id), K1(C(Tn) ⋊α Z) ∼= coker(α1 − id) ⊕ ker(α0 − id). So it suffices to determine the kernel and cokernel of (α0−id) and (α1−id) acting as endomorphisms on the finitely generated abelian groups Λeven , respectively. Note that from the isomorphisms (4) and (5), the K-groups of C(Tn) ⋊α Z are finitely generated abelian groups. Now, since α∗ is a ring homomorphism, it suffices to know the action of α∗ on e1, . . . , en. In fact, for a general basis element ei1 ∧ ei2 ∧ . . . ∧ eir of K∗(C(Tn)) ∼= Λ∗ Z (e1, . . . , en) ∼= Z2n−1 (e1, . . . , en) ∼= Z2n−1 Z(e1, . . . , en) we have and Λodd Z α∗(ei1 ∧ ei2 ∧ . . . ∧ eir ) = α∗(ei1 ) ∧ α∗(ei2 ) ∧ . . . ∧ α∗(eir ). Thus if we consider {e1, . . . , en} as the canonical basis of Zn and take α = α∗Zn, we have α∗ = ∧∗ α = r=1 ∧r α, α0 = ∧even α = ⊕r≥0 ∧2r α and α1 = ∧odd α = ⊕r≥0 ∧2r+1 α, where ∧i α is the i-th exterior ⊕n power of α, which acts on ΛiZn for i = 0, 1, . . . , n. Now, let α−1 = (f1, . . . , fn) and aji := Aj[fi], or in other words, assume that fi is homotopic to za1i for i = 1, . . . , n. So we can write : (ζ1, . . . , ζn) 7→ ζ a1i . . . ζ ani . . . zani n n 1 1 α∗(ei) = α∗[zi]1 = [α(zi)]1 = [zi ◦ α−1]1 = [fi]1 = [za1i 1 . . . zani n ]1 = n Xj=1 aji[zj]1 = n Xj=1 ajiej. Therefore α acts on Zn via the corresponding integer matrix A := [aij ]n×n ∈ GL(n, Z), α∗ acts on Λ∗Zn via ∧∗A, and we have the following isomorphisms K0(C(Tn) ⋊α Z) ∼= coker(α0 − id) ⊕ ker(α1 − id) = coker(⊕r≥0 ∧2r α − id) ⊕ ker(⊕r≥0 ∧2r+1 α − id), so we can write K0(C(Tn) ⋊α Z) ∼= Mr≥0 [coker(∧2r α − id) ⊕ ker(∧2r+1 α − id)], K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 7 and similarly K1(C(Tn) ⋊α Z) ∼= Mr≥0 [coker(∧2r+1 α − id) ⊕ ker(∧2r α − id)]. We summarize the arguments discussed above in the following theorem. Theorem 2.1. Let α be a homeomorphism of Tn and α ∈ Aut(Zn) be the restriction of α∗ to Zn (as above). Then α∗ = ∧∗ α = ⊕n r=1 ∧r α on K ∗(Tn) = Λ∗Zn and K0(C(Tn) ⋊α Z) ∼= Mr≥0 K1(C(Tn) ⋊α Z) ∼= Mr≥0 [coker(∧2r α − id) ⊕ ker(∧2r+1 α − id)], [coker(∧2r+1 α − id) ⊕ ker(∧2r α − id)]. Therefore in order to compute the K-groups of C(Tn) ⋊α Z, we must find the kernel and cokernel of ∧r α − id as an endomorphism of ΛrZn for r = 0, 1, . . . , n. Note that the matrix of ∧r α − id with respect to the canonical basis {ei1 ∧ . . . ∧ eir 1 ≤ i1 < . . . < ir ≤ n} with lexicographic order is An,r := ∧r A − I r), (n r(cid:1) (Ik is the identity matrix of order k - we often omit k whenever it is clear). So by computing the kernel and cokernel of An,r for r = 0, 1, . . . , n with appropriate tools (such as the Smith normal form), one can determine the K-groups of C(Tn) ⋊α Z. The author has written some Maple codes to handle such computations (see Appendix E) . which is an integer matrix of order (cid:0)n Corollary 2.2. The K-groups of C(Tn) ⋊α Z are finitely generated abelian groups with the same rank. Moreover, this common rank equals rank ker(∧∗ α − id) = n Xr=0 rank ker(∧r α − id). Proof. Use the previous proposition and note that for any ϕ ∈ End(Zn) one has rank ker ϕ = rank coker ϕ by the Smith normal form theorem (see Theorem A.2). (cid:3) Corollary 2.3. If α, β are homeomorphisms of Tn, whose corresponding integer matrices A, B ∈ GL(n, Z) are similar over Z, then Kj(C(Tn) ⋊α Z) ∼= Kj(C(Tn) ⋊β Z), (j = 1, 2). Proof. The assumption obviously implies that the automorphisms α and β are conjugate in Aut(Zn). This together with an easy application of the identity ∧r( φ ◦ ψ) = (∧r φ) ◦ (∧r ψ) (see Appendix C) imply that ∧r α and ∧r β (and therefore ∧r α − id and ∧r β − id) are conjugate in Aut(ΛrZn) for r = 0, 1, . . . , n. The result follows now from Theorem 2.1. (cid:3) 3. Anzai transformation group C ∗-algebras An,θ The simplest case of a Furstenberg transformation on an n-torus is an Anzai transformation σ, which was defined in part (d) of Definition 1.3. To study the K-groups of Anzai transformation group C ∗-algebras An,θ = C(Tn) ⋊σ Z using methods of the previous section, we will first need the "linearized" form of the corresponding affine homeomorphism σ−1, which is as follows So σ(ei) = ei−1 + ei for i = 1, . . . , n (e0 := 0). The matrix with respect to the canonical basis {e1, . . . , en} of Zn that corresponds to σ is the full Jordan block (ζ1, ζ2, . . . , ζn) 7→ (ζ1, ζ1ζ2, . . . , ζn−1ζn). 8 KAMRAN REIHANI Sn :=   1 0 1 . . . 0 ... 0 · · · 1 0 · · · 1 . . . 0 0 . . . 1 0 0 ... 0 1 1  n×n The following examples illustrates the methods described in the previous section for computing the K- groups of Anzai transformation group C ∗-algebras An,θ. Example 3.1. We compute the K-groups of A3,θ, which were computed in [39] by another method (the C ∗-algebra was denoted by A5,5 in there). In fact, the Chern character and noncommutative geometry were used in [39] to compute the kernel and cokernel of σi − id for i = 0, 1. However, we compute the kernel and cokernel of S3,r := ∧r S3 − I θ r) for r = 0, 1, 2, 3, where (3     S3 = 0 1 0 0 0 1 1 1 1 r = 0) S3,0 = ∧0S3 − I1 = [0]. So ker S3,0 = Z and coker S3,0 = Z/h0i ∼= Z. r = 1) So r = 2) So S3,1 = ∧1S3 − I3 =   1 0 0 1 0 1 1 0 1 −     1 0 0 1 0 0 0 0 1   =   0 0 0 1 0 0 1 0 0   ker S3,1 = {(x, y, z) ∈ Z3 y = z = 0} = (Z, 0, 0) ∼= Z, coker S3,1 = Z3/S3,1Z3 = Z3/he1, e2i ∼= Z. S3,2 = ∧2S3 − I3 =   1 0 0 1 1 1 1 0 1 −     1 0 0 1 0 0 0 0 1   =   0 0 0 1 1 0 1 0 0   ker S3,2 = {(x, y, z) ∈ Z3 y + z = z = 0} = (Z, 0, 0) ∼= Z, coker S3,2 = Z3/S3,2Z3 = Z3/he1, e2i ∼= Z. r = 3) S3,3 = ∧3S3 − I1 = [0]. So ker S3,3 = Z and coker S3,3 = Z/h0i ∼= Z. K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 9 Now, using Theorem 2.1 we have K0(A5,5 K1(A5,5 θ ) = K0(A3,θ) ∼= (coker S3,0 ⊕ coker S3,2) ⊕ (ker S3,1 ⊕ ker S3,3) ∼= Z ⊕ Z ⊕ Z ⊕ Z = Z4, θ ) = K1(A3,θ) ∼= (coker S3,1 ⊕ coker S3,3) ⊕ (ker S3,0 ⊕ ker S3,2) ∼= Z ⊕ Z ⊕ Z ⊕ Z = Z4. Notation 3.2. We let an := rank K0(An,θ) = rank K1(An,θ) and an,r := rank ker(∧r Sn − I) for r = 0, 1, . . . , n. From Corollary 2.2 we have an = rank ker(∧∗Sn − I) = n Xr=0 an,r. Example 3.3. Using the methods described in Section 2, we have obtained the K-groups of An,θ by computer for 1 ≤ n ≤ 12. The cases n = 1, 2, 3 have been calculated in the literature already: A1,θ = Aθ in [34]; A2,θ = A4 in [39]. However, there are no explicit computations for the higher dimensional cases starting with A4,θ = A6,10 as in [24] since hand calculations of kernels and cokernels of the maps become quickly impossible. Using the Maple codes given in Appendix E, we can find the kernel and cokernel of θ in [25]; and A3,θ = A5,5 θ θ for r = 0, 1, . . . , n and n = 1, . . . , 12 by means of the Smith normal form theorem (see Appendix A), and therefore we can compute the K-groups. The results are illustrated in Table 1, where Z(m) denotes the direct sum of m copies of the cyclic group Zk = Z/kZ. k Sn,r := ∧r Sn − I (n r) n 1 2 3 4 5 6 7 8 9 10 Z52 ⊕ Z(2) 3 ⊕ Z(2) 9 11 12 Z268 ⊕ Z(14) Z152 ⊕ Z(12) 13 ⊕ Z(4) Z90 ⊕ Z(4) 55 11 ⊕ Z(4) 26 ⊕ Z(4) 143 ⊕ Z(2) 1716 ⊕ Z(2) 286 Table 1. The K-groups of An,θ for 1 ≤ n ≤ 12 K0(An,θ) K1(An,θ) Z2 Z3 Z4 Z6 Z8 Z13 Z20 Z32 ⊕ Z(2) 8 Z2 Z3 Z4 Z6 Z8 Z13 ⊕ Z2 Z20 Z32 ⊕ Z(2) 18 Z52 ⊕ Z(2) 3 ⊕ Z(2) 9 Z90 ⊕ Z(2) 11 ⊕ Z99 ⊕ Z198 ⊕ Z2574 3432 ⊕ Z(2) 58344 Z268 ⊕ Z(4) 286 ⊕ Z(2) 4862 ⊕ Z(2) 68068 Z152 ⊕ Z(12) 13 ⊕ Z(4) 11 ⊕ Z(4) 26 ⊕ Z(6) 143 ⊕ Z(2) 286 an 2 3 4 6 8 13 20 32 52 90 152 268 Due to computational limitations, we do not have any results yet for n > 12, except for the sequence of ranks {an}, which we will study in detail in Sections 5 and 6. We will show the importance of 10 KAMRAN REIHANI this sequence in Section 5. Briefly, an is the common rank of the K-groups of a certain family of C ∗- algebras including Furstenberg transformation group C ∗-algebras Fθ,f based on Tn. Also, we will prove that {an} is a strictly increasing sequence (see Proposition 6.5). On the other hand, it seems that the K-groups of An,θ have torsion in general. The first example is K1(A6,θ); this is in fact because coker S6,3 = coker(∧3S6 − I20) ∼= Z3 ⊕ Z2 (see Example E.1 and Remark 1.7). Also, it is seen that the K0- and K1-groups are isomorphic for odd values of n in Table 1. In fact, this is true for more general cases (see Theorem 4.3). 4. A Poincar´e type of duality As stated in Theorem 2.1, the K-groups of a transformation group C ∗-algebras of the form C(Tn) ⋊α Z are completely determined by the corresponding homomorphism α ∈ Aut(Zn) and its exterior powers. From a computational point of view, we only need the cokernels of the maps involved, since we know that for any endomorphism on Zm, ker ∼= coker/tor(coker), where tor(G) denotes the torsion subgroup of the finitely generated abelian group G (see Appendix A). When det α = 1, we don't even need to compute all the cokernels. This is due to the following proposition, which establishes a Poincar´e type of duality between cokernels of certain integer matrices. We refer to Definition A.1 for the notion of equivalence of endomorphisms of Zn. Proposition 4.1 (Poincar´e duality). Let α ∈ SL(n, Z) (i.e. det α = 1). Then ∧r α − id and ∧n−r α − id are equivalent as endomorphisms of ΛrZn = Λn−rZn = Z(n r). Equivalently, coker(∧r α − id) is isomorphic to coker(∧n−r α − id) for r = 0, 1, . . . , n. Proof. We prove the equivalence of the endomorphisms for their corresponding integer matrices with respect to a certain basis. Let E = {e1, . . . , en} be a basis for Zn and set S = {1, 2, . . . , n}. For I = {i1, . . . , ir} ⊂ S with 1 ≤ i1 < . . . < ir ≤ n, put eI = ei1 ∧ . . . ∧ eir ∈ ΛrZn. Then Er := I = r} is a basis for ΛrZn. Let ω := e1 ∧ . . . ∧ en, which generates ΛnZn. We have {eI I ⊂ S , ∧0 α − id = 0, and ∧n α(ω) = α(e1) ∧ . . . ∧ α(en) = (det α)(e1 ∧ . . . ∧ en) = ω, so ∧n α − id = 0. Now, fix an r ∈ {1, . . . , n − 1}. For an arbitrary subset I ⊂ S with I = r, take J = E \ I = {j1, . . . , jn−r}, so J = n − r. Then eI ∧ eJ = (sgn µ) ω, in which µ ∈ Sn is the permutation that converts (1, 2, . . . , n) to (i1, . . . , ir, j1, . . . , jn−r). It is easily seen that µ = µ1 . . . µr, where µk is the permutation that takes ik from its position in (1, 2, . . . , n) to its new position in (i1, . . . , ir, j1, . . . , jn−r). One can see that µk is the combination of ik − (r − k + 1) number of transpositions (k = 1, . . . , r). Thus (−1)ik−(r−k+1) = (−1)ℓ(I)− r(r+1) 2 , r sgn µ = Yk=1 k=1 ik. Now, take m = (cid:0)n where ℓ(I) := Pr n−r(cid:1) and let Er = {eI1, . . . , eIm} be a basis for ΛrZn. Write En−r = {eJ1, . . . , eJm} as the basis for Λn−rZn such that Jk = E \ Ik for k = 1, . . . , m. From the above argument one can write r(cid:1) = (cid:0) n eIi ∧ eJj = (−1)ℓ(Ii)− r(r+1) 2 δij ω, since if i 6= j then Ii ∩ Jj 6= ∅ and eIi ∧ eJj = 0. Let A = [aij]m×m and B = [bij]m×m be the corresponding p=1 apieIp integer matrices of ∧r α and ∧n−r α with respect to Er and En−r, respectively. So ∧r α(eIi ) = Pm and ∧n−r α(eJj ) = Pm q=1 bqjeJq . What we want to show is that A − I is equivalent to B − I. We have m apibqj(−1)ℓ(Ip)− r(r+1) 2 δpqω. ∧n α(eIi ∧ eJj ) = (−1)ℓ(Ii)− r(r+1) 2 Therefore one obtains (6) δijω = ∧r α(eIi )^ ∧n−r α(eJj ) = Xp,q=1 m Xk=1 (−1)ℓ(Ik)−ℓ(Ii)akibkj = δij. K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 11 Now, if we set cij := (−1)ℓ(Ij)−ℓ(Ii)aji and C := [cij]m×m, then cij − δij = (−1)ℓ(Ij )−ℓ(Ii)(aji − δji). Therefore C − I is obtained from A − I by changing rows (and columns) and occasionally multiplying some rows (and columns) by −1. This means that C − I is equivalent to A − I. On the other hand, the equation (6) means that CB = I. So C − I = C(B − I)(−I) and B − I is also equivalent to C − I. Consequently, A − I is equivalent to B − I. (cid:3) Corollary 4.2. If det α = 1, then rank ker(∧r α − id) = rank ker(∧n−r α − id). Notation 3.2, we have an,r = an,n−r for r = 0, 1, . . . , n. In particular, using We are now ready to apply our Poincar´e duality to a the following K-theoretic result. Theorem 4.3. Let A := C(T2m−1) ⋊α Z be such that the corresponding homomorphism α satisfies det α = 1. Then K0(A) ∼= K1(A) as abelian groups, and the (common) rank of the K-groups of A is an even number. In particular, for every Furstenberg transformation group C ∗-algebra Fθ,f based on an odd-dimensional torus (e.g. A2m−1,θ), one has K0(Fθ,f ) ∼= K1(Fθ,f ). Proof. Combining Theorem 2.1 and Proposition 4.1, one obtains m−1 K0(A) ∼= K1(A) ∼= Mk=0 [coker(∧k α − id) ⊕ ker(∧k α − id)]. As a result, the rank of the K-groups of A is an even number since the ranks of the cokernel and kernel of an endomorphism coincide. Note that for Fθ,f the corresponding integer matrix of α is an upper triangular matrix with 1's on the diagonal. Thus det α = 1. (cid:3) 5. The rank an of the K-groups of An,θ In this section, we study some general properties of an, the (common) rank of the K-groups of Anzai transformation group C ∗-algebras based on Tn. We specify a family of C ∗-algebras, whose ranks of K-groups are given by the same sequence {an}. As an application, we characterize the rank of the K- groups of Furstenberg transformation group C ∗-algebras Fθ,f . In Appendix F, this study will have some applications to the classification of simple infinite dimensional quotients of the Heisenberg-type group C ∗-algebras C ∗(Dn), which were studied in an earlier work [33]. We remind the reader of some linear algebraic properties of nilpotent and unipotent matrices in Appendix B. We compare the ranks of the K-groups of a class of C ∗-algebras of the form C(Tn) ⋊α Z in the following theorem, which shows that the rank an of the K-groups of An,θ is somehow generic. Theorem 5.1. Let A = C(Tn) ⋊α Z, in which α is a homeomorphism of Tn, whose corresponding integer matrix A ∈ GL(n, Z) is unipotent of maximal degree (i.e. deg(A) = n). Then rank K0(A) = rank K1(A) = an = rank K0(An,θ) = rank K1(An,θ). In particular, the rank of the K-groups of any Furstenberg transformation group C ∗-algebra Fθ,f = C(Tn) ⋊ϕθ,f Z is equal to the rank of the K-groups of An,θ, namely, to an. Proof. Let α denote the restriction of α∗ to Zn and A be the corresponding matrix of α acting on Zn. Also, let Sn be the corresponding matrix for An,θ as denoted in Section 3. Since A is unipotent of maximal degree by assumption, and Sn is unipotent of maximal degree too, the matrices A and Sn are similar over C (see Corollary B.4). In fact, the Jordan normal form of A − I is precisely Sn − I. On the other hand, we know by Corollary 2.2 that the rank of the K-groups of A is equal to rank ker(∧∗A − I). Note that by the Smith normal form theorem (see Theorem A.2), rank ker(∧∗A − I) = dimC ker(∧∗A − I). The similarity of A and Sn implies the similarity of ∧∗A − I and ∧∗Sn − I as matrices acting on Λ∗Cn. So dimC ker(∧∗A − I) = dimC ker(∧∗Sn − I) = an, which yields the result. For the second part, note that the corresponding integer matrix of a Furstenberg transformation ϕθ,f on Tn is of the form 12 (♥)   KAMRAN REIHANI 1 0 1 . . . 0 ... 0 · · · b12 b13 · · · b23 . . . 0 0 . . . 1 0 b1n ... bn−2,n bn−1,n 1  n×n which is unipotent of maximal degree since bi,i+1 6= 0 for i = 1, . . . , n − 1 (see Definition 1.3 and Example B.5). Now, the proof of the first part yields the result. (cid:3) Remark 5.2. In the preceding theorem, the basis for Zn for the matrices involved is {e1, . . . , en}, where ei := [zi]1 as introduced at the beginning of Section 2. It is interesting to know that if α is an arbitrary unipotent automorphism of Zn, then there is a basis for Zn with respect to which the integer matrix A of α is of the form (♥) above (but not necessarily with bi,i+1 6= 0 for i = 1, . . . , n − 1, unless α is of maximal degree) [12, Theorems 16 and 18]. The unipotency of α has also important effects on the dynamics of the generated flow on Tn. For example, if α is an affine transformation on Tn and α is unipotent, then the dynamical system (Tn, α) has quasi-discrete spectrum [12, Theorem 19]. More generally, let α = (t, A) be an affine transformation on Tn and take Zp(A) = ker(Ap − id) ⊂ Zn for p ∈ N and consider the following conditions (1) Z1(A) = Zp(A), ∀p ∈ N, (2) t is rationally independent over Z1(A), i.e. j=1 tjkj is a rational number, then k = 0. Pn (3) Z1(A) 6= {0}, (4) A is unipotent. if k = (k1, . . . , kn) ∈ Z1(A) is such that ht, ki := Then (Tn, α) is ergodic with respect to Haar measure if and only if α satisfies the conditions (1) and (2) [12]. Moreover, if α satisfies the conditions (1) through (4), then the dynamical system (Tn, α) is minimal, uniquely ergodic with respect to Haar measure, and has quasi-discrete spectrum. Conversely, any minimal transformation on Tn with topologically quasi-discrete spectrum is conjugate to an affine transformation which must satisfy the conditions (1) through (4) [13]. The C ∗-algebras corresponding to such actions are therefore simple and have a unique tracial state. 6. Combinatorial properties of the sequence {an} Since an = Pn As mentioned before, one of our main goals is to describe an as the rank of the K-groups of An,θ. r=0 an,r, it makes sense to first study an,r. So we begin by finding some combinatorial properties of an,r, which is the rank of ker(∧r σ − id) for r = 0, 1, . . . , n, where σ is the automorphism of Zn corresponding to the Anzai transformation σ on Tn, and is represented by the integer matrix Sn as in the beginning of Section 3. In fact, we will show that an,r equals the number of partitions of [r(n + 1)/2] to r distinct positive integers not greater than n. To do this, we will use properties of the irreducible representations of the simple Lie algebra sl(2, C). 6.1. Connections with representation theory of sl(2, C). The automorphism σ is realized through its action on the basis {e1, . . . , en} of Zn, where ei := [zi] for i = 1, . . . , n as in Section 2, and we have σ(ei) = ei + ei−1 with e0 := 0. Therefore introducing a new endomorphism of Zn by ϕ := σ − id, we will get ϕ(ei) = ei−1. This is precisely a relation that may be recognized as part of the data of the canonical representation πn of the Lie algebra sl(2, C) on a complex vector space V with basis {e1, . . . , en}. More precisely, we have ϕ = πn(f ), K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 13 where the canonical representation πn and the (third) basis element f ∈ sl(2, C) are defined in Appendix D. The endomorphism ϕ induces a derivation on ΛrV , which is defined by D r ϕ(x1 ∧ . . . ∧ xr) = r Xi=1 x1 ∧ . . . ∧ ϕ(xi) ∧ . . . ∧ xr for r = 2, . . . , n and xi ∈ V , and by setting D 0 ϕ := 0, and D 1 ϕ := ϕ (see Appendix C). Then the following result states that an,r equals the nullity of the linear mapping D r ϕ. Proposition 6.1. Let σ be an Anzai transformation on Tn and σ∗ be the corresponding induced homo- morphism on K∗(C(Tn)) = Λ∗Zn. Let σ be the restriction of σ∗ to Zn and consider the linear mapping σ ⊗ 1 on V := Zn ⊗ C. Take ϕ = σ ⊗ 1 − id and D r ϕ as above. Then an,r = rank ker(∧r σ − id) = dim ker D r ϕ. Proof. Since ϕ is a nilpotent mapping, we can use Corollary C.4 to conclude that ∧r(σ ⊗ 1) − id ∼ D r ϕ. Therefore rank ker(∧r σ − id) = dim ker(∧r(σ ⊗ 1) − id) = dim ker D r ϕ. (cid:3) Notation 6.2. Let n, k, r be positive integers. Then P (n, r, k) denotes the number of partitions of k to r distinct positive integers not greater than n. In other words P (n, r, k) = card{(i1, . . . , ir) i1 + . . . + ir = k, 1 ≤ i1 < . . . < ir ≤ n}. By convention, we set P (n, 0, 0) = 1 and P (n, r, 0) = P (n, 0, k) = 0 for r, k ≥ 1. We are ready now to state the main result of this section. Theorem 6.3. With the above notation, an,r = P (n, r, [r(n+1)/2]), where [x] denotes the greatest integer not greater than x. In particular, an = n Xr=0 P (n, r, [ r(n + 1) 2 ]). Proof. Let πn : sl(2, C) → gl(V ) be the canonical representation of the Lie algebra sl(2, C) on the n- dimensional complex vector space V , and extend πn to πr n = πn. More precisely, for every X ∈ sl(2, C) define n : sl(2, C) → gl(ΛrV ) with π1 πr n(X)(v1 ∧ . . . ∧ vr) = (πn(X)v1) ∧ v2 ∧ . . . ∧ vr + . . . + v1 ∧ . . . ∧ vr−1 ∧ (πn(X)vr). This means that we have D r ϕ = πr n(f ) by the previous proposition. Following Weyl's theorem (see Theorem D.1), since the Lie algebra sl(2, C) is semisimple the representation πr n has to be completely reducible. This means we should have a decomposition ΛrV = ⊕N n-invariant irreducible subspaces of ΛrV . Moreover, the number N of such subspaces is equal to dim E0 + dim E1, where n(f ). In particular, an,r is the nullity of πr p=1Wp, where Wp's are some πr Ej = {v ∈ ΛrV πr n(h) v = j v}, (j = 0, 1) and h is the first basis element of sl(2, C) as in Appendix D (see Theorem D.3). On the other hand, the number N is equal to the nullity of πn r Wp is an irreducible representation of sl(2, C) on Wp, it is equivalent to the canonical representation of sl(2, C) on Wp by Theorem D.3. But the image of f in the canonical representation has a 1-dimensional kernel due to the part (c) of Proposition D.2. So the nullity of πn r (f ) counts the number of Wp's. Therefore r (f ). In fact, since πn an,r = dim ker πr n(f ) = dim E0 + dim E1. To compute the last two terms, note that using Proposition D.2 we have πn(h)ei = (2i − n − 1)ei, which leads to πr n(h)(ei1 ∧ . . . ∧ eir ) = (2(i1 + . . . + ir) − r(n + 1))ei1 ∧ . . . ∧ eir . 14 KAMRAN REIHANI So for even r(n + 1) we have E1 = {0} and dim E0 = P (n, r, r(n + 1)/2), and for odd r(n + 1) we have E0 = {0} and dim E1 = P (n, r, r(n+1)/2−1). To summarize, we have established the following equalities an,r = dim ker D r ϕ = dim ker πr n(f ) = N = dim E0 + dim E1 = P (n, r, [r(n + 1)/2]). The desired formula for an is immediate now by writing an = Pn r=0 an,r. (cid:3) Using the previous theorem, we can prove that {an} is a strictly increasing sequence. We need a lemma first. Lemma 6.4. P (n + 1, r, k + s) ≥ P (n, r, k) for s = 0, 1, . . . , r. Proof. For s = 0, the proof is clear. Now, let 1 ≤ s ≤ r and suppose that (j1, . . . , jr) is a partition of k such that 1 ≤ j1 < . . . < jr ≤ n. Define iq := jq for 1 ≤ q ≤ r−s and iq := jq +1 for r−s+1 ≤ q ≤ r. Then (i1, . . . , ir) is a partition of k + s and 1 ≤ i1 < . . . < ir ≤ n + 1. Thus P (n + 1, r, k + s) ≥ P (n, r, k). (cid:3) Proposition 6.5. {an} is a strictly increasing sequence. Proof. First, note that an,0 = an,n = P (n, 0, 0) = P (n, n, n(n + 1)/2) = 1, and from the previous theorem we have an = Pn a2m+1 = 1 + r=0 P (n, r, [r(n + 1)/2]). Fix m ∈ N, and get m Xr=0 P (2m + 1, 2r, 2rm + 2r) + m−1 Xr=0 P (2m + 1, 2r + 1, 2rm + 2r + m + 1), a2m = m Xr=0 P (2m, 2r, 2rm + r) + m−1 Xr=0 P (2m, 2r + 1, 2rm + m + r) = 1 + m−1 Xr=0 P (2m, 2r, 2rm + r) + m−1 Xr=0 P (2m, 2r + 1, 2rm + m + r), a2m−1 = m−1 Xr=0 P (2m − 1, 2r, 2rm) + m−1 Xr=0 P (2m − 1, 2r + 1, 2rm + m). Applying the previous lemma to the terms of the sums expressed above implies that a2m+1 > a2m > a2m−1. 6.2. Generating functions for the sequence {an}. In this part, we express the rank of the K-groups of An,θ as explicitly as possible. In fact, we present them as the constant terms in the polynomial expansions of certain functions. First of all, we need the following basic lemma. (cid:3) Lemma 6.6. Let P (n, r, k) denote the number of partitions of k to r distinct positive integers not greater i=1(1 + uti). than n. Then P (n, r, k) is the coefficient of urtk in the polynomial expansion of Fn(u, t) := Qn In other words, n P (n, r, k)urtk = Xr,k≥0 (1 + uti). Yi=1 Proof. n (1 + uti) = 1 + Yi=1 n Xr=1 X(i1,...,ir ) (uti1 ) . . . (utir ) = 1 + Xr=1Xk≥1 P (n, r, k)urtk = Xr,k≥0 P (n, r, k)urtk. n 1≤i1<...<ir ≤n (cid:3) Now, we have the following result for the rank an of the K-groups of An,θ. K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 15 Theorem 6.7. Let an = rank K0(An,θ) = rank K1(An,θ). Then for a nonnegative integer m we have (i) a2m+1 is the constant term in the Laurent polynomial expansion of m Yj=−m (1 + zj), (ii) a2m is the constant term in the Laurent polynomial expansion of m (1 + z) Yj=−m+1 (1 + z2j−1). Proof. We know that an = Pn P2m+1 r=0 P (2m + 1, r, r(m + 1)). Now, take y = utm+1 and use the preceding lemma to get r=0 an,r and an,r = P (n, r, [r(n + 1)/2]) by Theorem 6.3. We have a2m+1 = 2m+1 F2m+1(u, t) = F2m+1(yt−m−1, t) = Yi=1 (1 + yti−m−1) = Xr,k≥0 P (2m + 1, r, k)yrtk−r(m+1). In particular, we get the following identity for y = 1 2m+1 Yi=1 (1 + ti−m−1) = Xr,k≥0 P (2m + 1, r, k)tk−r(m+1), or equivalently, by setting z = t and j = i − m − 1 we have m Yj=−m (1 + zj) = Xr,k≥0 P (2m + 1, r, k)zk−r(m+1). In particular, the constant term in the Laurent polynomial expansion of Qm we take the sum of those terms for which k = r(m + 1) holds, namely j=−m(1 + zj) is obtained when 2m+1 Xr=0 P (2m + 1, r, r(m + 1)), which is precisely the expression for a2m+1. For part (ii), write a2m = 2m Xr=0 P (2m, r, [r(m + 1 2 )]) = m Xr=0 =: Am + Bm. P (2m, 2r, r(2m + 1)) + m−1 Xr=0 P (2m, 2r + 1, 2rm + m + r) Let us determine Am first. Note that using the preceding lemma we have 1 2 { 2m 2m (1 + uti) + Yi=1 Yi=1 (1 − uti)} = Xr,k≥0 P (2m, r, k){ 1 + (−1)r 2 }urtk = Xr,k≥0 P (2m, 2r, k)u2rtk. If we define y := u2t2m+1, we have the following identity 1 2 { 2m (1 + y Yi=1 which for y = 1 yields 1 2 ti−(m+ 1 2 )) + 2m Yi=1 (1 − y 1 2 ti−(m+ 1 2 ))} = Xr,k≥0 P (2m, 2r, k)yrtk−r(2m+1), 1 2 { 2m (1 + ti−(m+ 1 2 )) + Yi=1 2m Yi=1 (1 − ti−(m+ 1 2 ))} = Xr,k≥0 P (2m, 2r, k)tk−r(2m+1). 16 KAMRAN REIHANI Hence Am is the constant term in the polynomial expansion of 1 2 { Similarly, for Bm we have 2m 2m (1 + ti−(m+ 1 2 )) + Yi=1 (1 − ti−(m+ 1 2 ))}. Yi=1 1 2 { 2m 2m (1 + uti) − Yi=1 Yi=1 (1 − uti)} = Xr,k≥0 P (2m, r, k){ 1 − (−1)r 2 }urtk = Xr,k≥0 P (2m, 2r + 1, k)u2r+1tk. If we define y2 := u2t2m+1, we have the following identities 1 2 { 2m Yi=1 (1 + y 1 2 ti−(m+ 1 2 )) − 2m (1 − y Yi=1 = Xr,k≥0 1 2 ti−(m+ 1 2 ))} P (2m, 2r + 1, k)y2r+1tk−(2rm+r+m)− 1 2 = t− 1 2 Xr,k≥0 P (2m, 2r + 1, k)y2r+1tk−(2rm+r+m), which for y = 1 yields 1 2 t 2 { 2m (1 + ti−(m+ 1 2 )) − Yi=1 2m Yi=1 (1 − ti−(m+ 1 2 ))} = Xr,k≥0 P (2m, 2r + 1, k)tk−(2rm+r+m). Hence Bm is the constant term in the polynomial expansion of 1 2 t 2 { 2m (1 + ti−(m+ 1 2 )) − Yi=1 (1 − ti−(m+ 1 2 ))}. 2m Yi=1 Therefore a2m = Am + Bm is the constant term in the polynomial expansion of 1 2 { 2m 2m (1 + ti−(m+ 1 2 )) + Yi=1 (1 − ti−(m+ 1 2 )) + Yi=1 1 2 t 2 2m (1 + ti−(m+ 1 2 )) − Yi=1 1 2 t 2 2m (1 − ti−(m+ 1 2 ))}, Yi=1 or equivalently, the constant term in the polynomial expansion of 1 2 {(1 + z) 2m 2m (1 + z2i−(2m+1)) + (1 − z) Yi=1 (1 − z2i−(2m+1))}, Yi=1 which equals the constant term in the Laurent polynomial expansion of m (1 + z) Yj=−m+1 (1 + z2j−1). Thanks to this theorem, one can compute an for large values of n using a computer algebra program. Many more terms are also available online at OEIS (The Online Encyclopedia of Integer Sequences at www.oeis.org). Moreover, as the following corollaries suggest, such recognitions as constant terms of certain Laurent polynomials opens the door to finding even more interesting combinatorial properties of the sequence {an}, which have been of interest to Erdos, J. H. van Lint and R. C. Entringer to name a few (cf. [6, 23, 5]). Corollary 6.8. Let n be a nonnegative integer. (cid:3) K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 17 (i) The integer a2n+1 is the number of solutions of the equation k=n Xk=−n k ǫk = 0, where ǫk = 0 or 1 for −n ≤ k ≤ n. In other words, a2n+1 is the number of ways that a sum of integers between −n and n (with no repetitions) equals to 0. (ii) The integer a2n is the number of solutions of the equation k=n Xk=−n+1 (2k − 1) ǫk = 0 or 1, where ǫk = 0 or 1 for −n + 1 ≤ k ≤ n. In other words, a2n is the number of ways that a sum of half-integers between −n + 1/2 and n − 1/2 (with no repetitions) equals to 0 or 1/2. Proof. Using Theorem 6.7, the number a2n+1 is the constant term in the Laurent polynomial expansion of Qn k=−n(1 + zk), which is a finite sum of the form P A(n, m)zm. Obviously, the integer coefficient A(n, m) is the number of all possible combinations from the terms z−n, . . . , z0, . . . , zn, whose product makes a zm. In other words, by putting ǫk = 1 when zk contributes to such a product making a zm, and ǫk = 0 otherwise, we conclude that A(n, m) = #{(ǫ−n, . . . , ǫ0, . . . , ǫn) ∈ {0, 1}2n+1 : k=n Xk=−n k ǫk = m}. In particular, the constant term of the Laurent polynomial expansion is A(n, 0), and we have a2n+1 = A(n, 0). This proves part (i). Fort part (ii), we use the same idea for the Laurent polynomial expansion of m (1 + z) Yk=−m+1 (1 + z2k−1) = m Yk=−m+1 (1 + z2k−1) + z m Yk=−m+1 (1 + z2k−1) as suggested by part (ii) of Theorem 6.7. (cid:3) J. H. van Lint in [23] answered a question of Erdos by determining the asymptotic behavior of A(n, 0) = #{(ǫ−n, . . . , ǫ0, . . . , ǫn) ∈ {0, 1}2n+1 : k=n Xk=−n k ǫk = 0}. The idea in his proof is as follows. Since A(n, 0) is the constant term of the Laurent polynomial expansion k=−n(1 + zk), we can compute it as the Cauchy integral of Qn 1 2πi IC Qn z k=−n(1 + zk) dz, where C denotes the unit circle. By parameterizing C by z = e2ix for x ∈ [0, π], applying the elementary identity (1 + e2ikx)(1 + e−2ikx) = 4 cos2 kx, and a simple calculation we arrive at A(n, 0) = 22n+2 π π 2 Z 0 n Yk=1 cos2 kx dx. We can then proceed by estimating the integrand near and far from 0 using some elementary inequalities, 2 [23]. This will immediately give the which lead to the asymptotic formula A(n, 0) ∼ (3/π) asymptotic behavior of the sequence {a2n+1} by the previous corollary. One can adapt the arguments 2 22n+1n− 3 1 18 KAMRAN REIHANI used by J. H. van Lint to obtain a similar asymptotic behavior for the sequence {a2n} by estimating the corresponding integral 22n+2 π π 2 Z 0 cos2 x n cos2(2k − 1)x dx, Yk=1 2 22nn− 3 1 which leads to the asymptotic formula a2n ∼ (3/π) 2 . This gives rise to the following result. Corollary 6.9. an ∼ r 24 π 2nn− 3 2 when n → ∞. In particular, lim n→∞ an+1 an = 2. 7. The positive cone of K0(Fθ,f ) In this section, we generalize a result of Kodaka on the order structure of the group K0 of the crossed product by a Furstenberg transformation on the 2-torus [18, Theorem 5.2]. However, our approach is dif- ferent, and follows the general guidelines of [30, Lemma 3.1]. We remind the reader that for a C ∗-algebra A the positive cone of K0(A) is the set K0(A)+ = {[q] ∈ K0(A) : q ∈ P∞(A)}, where P∞(A) is the set of all projections in matrix algebras over A. Also, any positive trace τ on a C ∗-algebra A induces a group homomorphism τ∗ : K0(A) → R. As was indicated in the Introduction, when the Furstenberg transfor- mation ϕθ,f is minimal and uniquely ergodic, using the results of H. Lin and N. C. Phillips in [21] the transformation group C ∗-algebra Fθ,f is classifiable by its Elliott invariant, and the order of K0(Fθ,f ) is determined by the unique tracial state τ on Fθ,f [22, 29]. The fact that τ∗K0(Fθ,f ) = Z + Zθ was first proved in the unpublished thesis of R. Ji [16]. However, we will study the effect of the trace on the order structure of K0 using R. Exel's machinery of rotation numbers [7]. For a C ∗-algebra A, we denote by Up(A) the set of unitary elements of Mp(A). The following lemma is well known, but it is convenient to state and prove it for self-containment of the paper. Lemma 7.1. Let A and B be unital C ∗-algebras and let A ⊗ B denote their minimal tensor product. Suppose that u ∈ Up(A) and v ∈ Uq(B), and let φ : C(T2) :→ Mpq(A ⊗ B) be the unique homomorphism mapping the coordinate unitaries z1, z2 ∈ U (C(T2)) to the commuting unitaries u ⊗ 1q, 1p ⊗ v ∈ Upq(A ⊗ B), respectively. Let b(u, v) ∈ K0(A ⊗ B) denote the Bott element of u, v defined by K0(φ)(b), where β = [z1] ∧ [z2] is the Bott element in K0(C(T2)) so that K0(C(T2)) = Z[1] + Z β. Then τ∗(b(u, v)) = 0 for any tracial state τ on A ⊗ B. Proof. Since τ ◦ φ is a trace on T2, there exists a Borel probability measure µ on T2 such that (τ ◦ φ)(f ) = ZT2 f (x) dµ(x), f ∈ C(T2). Write β = [p] − [q], where p, q are appropriate projections in some matrix algebra over C(T2), so we have τ∗(b(u, v)) = τ∗(K0(φ)(β) = (τ ◦ φ)∗(β) = ZT2 Tr(p(x)) − Tr(q(x)) dµ(x). It is well known that for the Bott element b we have Tr(p(x)) − Tr(q(x)) = 0, namely, the projections p(x) and q(x) have the same rank for all x ∈ T2, and this common rank does not depend on x since T2 is connected (in fact, they are rank one projections). This can be proved either by a calculation of the traces of the projections p(x) and q(x) explicitly (cf. [1, p. 7]), or by using the naturality in the Kunneth formula for T2, which shows that the image under any point evaluation of β is zero. Briefly speaking, the map x 7→ Tr(p(x)) − Tr(q(x)) belongs to C(T2, Z), so it has to assume a constant integer, which we call dim β. In particular, dim β is invariant under the change of coordinate (ζ1, ζ2) 7→ (ζ1, ζ −1 2 ), whereas the naturality of the Kunneth homomorphism α1,1 : K1(C(T)) ⊗ K1(C(T)) → K0(C(T2)), which maps [z] ⊗ [z] to β, implies that the Bott element β will transform into −β under this change of coordinates since [z] ⊗ [z−1] = [z] ⊗ (−[z]) = −([z] ⊗ [z]). This means dim β = 0. (cid:3) K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 19 We denote by u the unitary in Fθ,f implementing the action generated by the transformation ϕθ,f on Tn with irrational parameter θ, and by z1 the unitary in C(Tn) defined by z1(ζ1, . . . , ζn) = ζ1 as in Section θ,f = e2πiθz1 so that C ∗(u, z1) ∼= Aθ, the irrational rotation algebra. 2. Then we have uz1u−1 = z1 ◦ ϕ−1 Let pθ ∈ C ∗(u, z1) be a Rieffel projection of trace θ as in [34]. It is obvious that τ∗([1]) = 1. On the other hand, since the restriction of τ on the C ∗-subalgebra Aθ ⊆ Fθ,f has to be the unique tracial state on Aθ, we have τ∗([pθ]) = θ. The main result of this section will show that all the essential information about the order structure of K0(Fθ,f ) is encoded in the embedding of Aθ in Fθ,f . Theorem 7.2. Let ϕθ,f be a minimal uniquely ergodic Furstenberg transformation on Tn with θ ∈ (0, 1) (e.g. when θ ∈ (0, 1) \ Q and each fi satisfies a uniform Lipschitz condition in ζi for i = 1, . . . , n − 1). f denote, respectively, the rank and the torsion subgroup of K0(Fθ,f ) so that K0(Fθ,f ) ∼= Let an and T 0 Zan ⊕ T 0 f . Then the isomorphism of K0(Fθ,f ) with this group can be chosen in such a way that (i) the unique tracial state τ on Fθ,f induces the map τ∗(a[1] + b[pθ], c, t) = a + b θ on K0(Fθ,f ) for all (a[1] + b[pθ], c, t) ∈ (Z[1] + Z[pθ]) ⊕ Zan−2 ⊕ T 0 f ∼= Zan ⊕ T 0 f , (ii) the positive cone K0(Fθ,f )+ can be identified with {(a[1] + b[pθ], c, t) ∈ (Z[1] + Z[pθ]) ⊕ Zan−2 ⊕ T 0 f : a + b θ > 0} ∪ {0}. Proof. The idea of the proof is to show that there exists a generating set for the finitely generated abelian group K0(Fθ,f ) including [1] and [pθ] such that the induced homomorphism τ∗ vanishes at all generators, except for [1] and [pθ] for which we have τ∗([1]) = 1 and τ∗([pθ]) = θ. Using Theorem 2.1 and setting α = ϕθ,f and αj = Kj(α) for j = 1, 2 we have K0(Fθ,f ) ∼= coker(α0 − id) ⊕ ker(α1 − id) = Mr≥0 [coker(∧2r α − id) ⊕ ker(∧2r+1 α − id)], where α is the restriction of α1 to the subgroup Z[z1] + . . . + Z[zn] of K1(C(Tn)) as in Section 2, and zj(ζ, . . . , ζn) = ζj of C(Tn) for j = 1, . . . , n. Note that by Definition 1.3, f = (f1, . . . , fn−1) consists of continuous functions fj−1 : Tj−1 → T for j = 2, . . . , n. First, we "linearize" each fj−1 by finding the unique "linear" function in the homotopy class of fj−1. This allows us to calculate α([zj]) by writing (ζ1 . . . , ζj−1) 7→ ζ b1j 1 . . . ζ bj−1,j j−1 , (bj−1,j 6= 0) α([zj]) = [zj ◦ ϕ−1 α([z1]) = [z1 ◦ ϕ−1 θ,f ] = [fj−1(z1, . . . , zj−1)zj] = [zb1j In other words, the integer matrix of α with respect to the basis {[z1] . . . , [zn]} of for j = 2, . . . , n. Zn is precisely in the form (♥) as in the proof of Theorem 5.1. Now, we can realize α0 = ∧even α and α1 = ∧odd α to calculate the K-groups of Fθ,f as in Section 2. j−1 zj] = b1j[z1] + . . . + bj−1,j[zj−1] + [zj], θ,f ] = [e2πiθz1] = [z1], . . . zbj−1,j 1 It is important to note that, by referring to the exact sequences (1), (2) and (3) in Section 1, the isomor- phic image of coker(α0 − id) in K0(Fθ,f ) is precisely the image im0 of K0(C(Tn)), and an isomorphic image of ker(α1 − id) in K0(Fθ,f ) is obtained by finding the image of a splitting (injective) homomor- phism s : ker(α1 − id) → K0(Fθ,f ) for the exact sequence (2) so that ∂ ◦ s = id on ker(α1 − id). Any such splitting homomorphism is obtained as follows: fix a basis {γ1, . . . , γq} for the free finitely gener- ated group ker(α1 − id) of rank q, and find elements ν(0) j = γj for j = 1, . . . , q. Then define s(Pj mjγj) = Pj mjν(0) for mj ∈ Z. Clearly, K0(Fθ,f ) = im 0 ⊕ im s. q ∈ K0(Fθ,f ) such that ∂ν(0) 1 , . . . , ν(0) j Now, since ∧0 α = id on Λ0Zn = Z, and ∧1 α = α on Λ1Zn = Zn, we can write the isomorphism K0(Fθ,f ) ∼= Z ⊕ ker(α − id) ⊕Mr≥1 [ker(∧2r+1 α − id) ⊕ coker(∧2r α − id)]. 20 KAMRAN REIHANI In fact, a single generator for the isomorphic image of Z in K0(Fθ,f ) is [1], and since bj−1,j 6= 0 for j = 2, . . . , n we have ker(α − id) = Ze1 = Z[z1]. It is easy to see that ∂([pθ]) = [z1] (see the proposition in the appendix of [31]). Therefore there exists a basis {γ1, . . . , γq} for ker(α1 − id) = ⊕r≥0 ker(∧2r+1 α − id) with γ1 = [z1] and a splitting homomorphism s : ker(α1 − id) → K0(Fθ,f ) with s([z1]) = [pθ]. Hence a single generator for the image of ker(α − id) in K0(Fθ,f ) is [pθ]. It remains to study the effect of τ∗ on the isomorphic image of ⊕r≥1coker(∧2r α − id), which contains the the torsion subgroup T 0 f , and the image of ⊕r≥1 ker(∧2r+1 α − id) in K0(Fθ,f ). For more convenience, set ej := [zj] for j = 1, . . . , n as in Section 2. First, we study the isomorphic image of ⊕r≥1coker(∧2r α − id). We show that τ∗ vanishes on this whole subgroup by showing, equivalently, that τ∗ vanishes on the image of the subgroup ⊕r≥1Λ2r Z (e1, . . . , en) ⊂ K0(C(Tn)) in K0(Fθ,f ) under the map 0. Let η = ei1 ∧. . .∧ei2r ∈ K0(C(Tn)) for some r ≥ 1 and 1 ≤ i1 < . . . < i2r ≤ n. We want to show that τ∗(0(η)) = 0, where 0 := K0() and  : C(Tn) → Fθ,f is the natural embedding in the structure of the crossed product Fθ,f = C(Tn) ⋊α Z. By Kunneth formula we have η = b(u, z), where u is a unitary in some matrix algebra over C(Tn−1) with [u] = ei1 ∧ . . . ∧ ei2r−1 ∈ K1(C(Tn−1)) and z is the canonical unitary in C(T) with [z] = ei2r ∈ K1(C(T)). By Lemma 7.1, we have τ∗(0(η)) = τ∗(K0()(η)) = (τ ◦ )∗(η) = (τ ◦ )∗(b(u, z)) = 0. j j Z 1 ) = θ, and τ∗(ν(0) j ) = γj for j = 1, . . . , q and ν(0) ) = kj for some kj ∈ Z for j = 2, . . . , q, where ν(0) Now, we study the isomorphic image of ⊕r≥1 ker(∧2r+1 α−id) in K0(Fθ,f ). We will show that τ∗ assumes only integer values on this whole subgroup. In other words, if {γ1, . . . , γq} is a basis for the ker(α0 − id) as above such that γ1 = [z1] is a basis for ker(α − id) and {γ2, . . . , γq} is a basis for ⊕r≥1 ker(∧2r+1 α − id), then τ∗(ν(0) 's are chosen in K0(Fθ,f ) so that ∂(ν(0) 1 = [pθ] as above. To demonstrate this, we prove that the determinant of any unitary representing an element in the subgroup ⊕r≥1Λ2r+1 (e1, . . . , en) is the constant function 1. Then the rotation number homomorphism ρµ α : ker(α1 − id) → T defined by R. Exel is the constant 1 on the subgroup ⊕r≥1 ker(∧2r+1 α − id) [7, Theorem VI.11], hence the trace will be integer-valued on this subgroup because exp(2πiτ∗(η)) = ρµ α ◦ ∂(η) for all η ∈ K0(Fθ,f ) [7, Theorem V.12]. To calculate the determinant on ⊕r≥1Λ2r+1 (e1, . . . , en), let γ = ei1 ∧ . . . ∧ ei2r+1 ∈ K1(C(Tn)), set η = ei1 ∧ . . . ∧ ei2r ∈ K0(C(Tn−1)), and write η = [p] − [q] as above. Then ei2r+1 = [z] for the canonical unitary z of C(T), and using the Kunneth formula we have γ = η ⊗ [z] = ([p] − [q]) ⊗ [z] = [p] ⊗ [z] + [q] ⊗ [z−1] = [((1 − p) ⊗ 1 + p ⊗ z)((1 − q) ⊗ 1 + q ⊗ z−1)] So, γ = [ω1ω2], where ω1 := (1 − p) ⊗ 1 + p ⊗ z and ω2 := (1 − q) ⊗ 1 + q ⊗ z−1 are unitaries in some matrix algebra of the same size over C(Tn). Since for all x ∈ Tn−1 the projections p(x) and q(x) have the same rank ρ as in the proof of the previous lemma, we have the following equivalence of projections in some matrix algebra Ml(C) Z p(x) ∼ 1ρ ⊕ 0l−ρ ∼ q(x), where 1m, 0m denote the identity and the zero matrix of order m, respectively, and ⊕ is the direct sum of matrices. This implies the following unitary equivalence of projections in M2l(C) p(x) ⊕ 0l ∼u 1ρ ⊕ 02l−ρ ∼u q(x) ⊕ 0l. In particular, we conclude the following unitary equivalence of unitary matrices for all (x, ζ) ∈ Tn−1 × T ω1(x, ζ) ⊕ 1l ∼u ζ1ρ ⊕ 12l−ρ, ω2(x, ζ) ⊕ 1l ∼u ζ −11ρ ⊕ 12l−ρ. Therefore Det ω1(x, ζ) = ζ ρ and Det ω2(x, ζ) = ζ −ρ, hence Det (ω1ω2)(x, ζ) = 1, for all (x, ζ) ∈ Tn−1 × T. This implies that Det∗(γ) = 1 ∈ [Tn, T], where [Tn, T] denotes the set of homotopy classes of continuous functions from Tn to T (see Definition VI.8 and Proposition VI.9 of [7]). Finally, by setting ν1 := ν(0) j −kj[1] for j = 2, . . . , q so that τ∗(ν1) = θ and τ∗(νj) = 0 for j = 2, . . . , q, we can form a generating set with the desired property for K0(Fθ,f ) by taking the union of {ν1, . . . νq} and a generating set including [1] for the isomorphic image of coker(α0 − id). This proves 1 = [pθ] and νk := ν(0) K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 21 part (i). For part (ii), we use part (i) together with the fact that the order on K0(Fθ,f ) is determined by the effect of the unique tracial state τ because Tn is a finite dimensional infinite compact metric space and ϕθ,f is a minimal homeomorphism of Tn (see Theorem 5.1(1) of [22] or Theorem 4.5(1) of [29]). (cid:3) Corollary 7.3. Let ϕθ,f be a minimal uniquely ergodic Furstenberg transformation on Tn as above. Then linearizing the functions fi : Ti → T in f = (f1, . . . , fi−1) does not change the isomorphism class of the transformation group C ∗-algebra Fθ,f . Proof. Since ϕθ,f is minimal, θ must be irrational. So the range of the unique tracial state (by unique ergodicity) on K0(Fθ,f ) is dense in R as it is Z+Zθ by the above argument. Benefiting from the results of [21], such C ∗-algebras are completely classifiable by their Elliott invariants, which remain unchanged (up to isomorphism) after the linearization process: linearizing does not change the isomorphism classes of the K-groups, and the previous theorem guarantees that the order structure of the group K0 is precisely the regular order inherited from R on Z + Zθ before and after linearization. (cid:3) 8. Concluding remarks I) The method used in Section 1 for computing K-groups of the transformation group C ∗-algebras of homeomorphisms of the tori may be extended to more general settings. Let G be a compact connected Lie group with torsion-free fundamental group π1(G). (It is well known that the fundamental group of such spaces are finitely generated and abelian, so being torsion-free means π1(G) ∼= Zl, for some l.) Some important examples are any finite Cartesian products of the groups S3, SO(2), Sp(n), U (n) and SU (n). Then K ∗(G) is torsion-free and can be given the structure of a Z2-graded Hopf algebra over the integers [14]. Moreover, regarded as a Hopf algebra, K ∗(G) is the exterior algebra on the module of the primitive elements, which are of degree 1. The module of the primitive elements of K ∗(G) may also be described as follows. Let U (n) denote the group of unitary matrices of order n and let U := ∪∞ n=1U (n) be the stable unitary group. Any unitary representation ρ : G → U (n), by composition with the in- clusion U (n) ⊂ U , defines a homotopy class β(ρ) in [G, U ] = K 1(G). The module of the primitive elements in K 1(G) is exactly the module generated by all classes β(ρ) of this type. If in addition, G is semisimple and simply connected of rank l, there are l basic irreducible representations ρ1, . . . , ρl, whose maximum weights λ1, . . . , λl form a basis for the character group T of the maximal torus T of G and the classes β(ρ1), . . . , β(ρl) form a basis for the module of the primitive elements in K 1(G) and K ∗(G) = Λ∗(β(ρ1), . . . , β(ρl)). In any case, to compute K∗(C(G) ⋊α Z) it is sufficient to determine the homotopy classes of α ◦ ρ for the irreducible representations ρ of G in terms of β(ρ)'s. II) There is a relation between the K-theory of transformation group C ∗-algebras of the homeomorphisms of the tori and the topological K-theory of compact nilmanifolds. In fact, let α = (t, A) be an affine transformation on Tn satisfying the conditions (1) through (4) in Remark 5.2. Then it has been shown in [12] that α is conjugate (in the group of affine transformations on Tn) to the transformation α′ = (t′, A′), where A′ has an upper triangular matrix, whose bottom right k × k corner is the identity matrix Ik and t′ = (0, . . . , 0, t′ k). The transformation α′ is called a standard form for α [28]. Assume that α is given in standard form. Then J. Packer associates an induced flow (R, N/Γ) to the flow (Z, Tn) generated by α, where N is a simply connected nilpotent Lie group of dimension n + 1, the discrete group Γ is a cocompact subgroup of N , and the action of R is given by translation on the left by exp sX for s ∈ R and some X ∈ n, the Lie algebra of N . One of the most important facts is that the C ∗-algebra C(N/Γ) ⋊β R corresponding to the induced flow is strongly Morita equivalent to C(Tn) ⋊α Z [28, Proposition 3.1]. Consequently, one has 1, . . . , t′ (7) Ki(C(Tn) ⋊α Z) ∼= Ki(C(N/Γ) ⋊β R) ∼= K 1−i(N/Γ); i = 0, 1. 22 KAMRAN REIHANI The second isomorphism here, is the Connes' Thom isomorphism. So the K-theory of C(Tn) ⋊α Z is converted to the topological K-theory of the compact nilmanifold N/Γ. Following the proof of Proposition 3.1 in [28], one can conclude that for the special case of Anzai transformations we can take N = Fn−1 (the generic filiform Lie group of dimension n + 1) and Γ = Dn−1, which were defined in [33]. On the other hand, following [35, Theorem 3.6], one has the isomorphism (8) Ki(C ∗(Γ)) ∼= K i+n+1(N/Γ); i = 0, 1. Combining (7) and (8) one gets (9) Ki(C(Tn) ⋊α Z) ∼= K i+1(N/Γ) ∼= Ki+n(C ∗(Γ)); i = 0, 1. Using the above isomorphisms, one can relate the algebraic invariants of the involved C ∗-algebras and topological information of the corresponding nilmanifold. For example, since N/Γ is a classifying space for Γ, one has the following isomorphisms (10) H ∗ dR(N/Γ) ∼= H ∗(N/Γ, R) ∼= H ∗(Γ, R) ∼= H ∗(N, R) ∼= H ∗(n, R), dR(N/Γ) denotes the de Rham cohomology of the manifold N/Γ, H ∗(N/Γ, R) denotes the Cech where H ∗ cohomology of N/Γ with coefficients in R, H ∗(Γ, R) denotes the group cohomology of Γ with coeffi- cients in the trivial Γ-module R, H ∗(N, R) denotes the Moore cohomology group of N (as a locally compact group) with coefficients in the trivial Polish N -module R, and H ∗(n, R) denotes the cohomology of the Lie algebra n with coefficients in the trivial n-module R. Now, using the Chern isomorphisms ch0 : K 0(N/Γ) ⊗ Q → H even(N/Γ, Q) and ch1 : K 1(N/Γ) ⊗ Q → H odd(N/Γ, Q), one concludes that the even and odd cohomology groups stated in (10) are all isomorphic to Rk, where k is the (common) rank of the K-groups of C(Tn) ⋊α Z as in Corollary 2.2. As an example, if N = Fn−1, Γ = Dn−1, and n = fn−1, then the even and odd cohomology groups stated in (10) are all isomorphic to Ran , where an is the rank of the K-groups of An,θ that was studied in detail in Sections 5 and 6. Conversely, one may use the topological tools for N/Γ to get some information about C(Tn) ⋊α Z and C ∗(Γ). For example, we know that N/Γ as a compact nilmanifold can be constructed as a principal T-bundle over a lower dimensional compact nilmanifold [10]. Then we can compute the topological K-groups of N/Γ using the six term Gysin exact sequence [17, IV.1.13, p. 187]. As an example, one can see that Fn−1/Dn−1 is a principal T-bundle over Fn−2/Dn−2, and the corresponding Gysin exact sequence is in fact the topological version of the Pimsner-Voiculescu exact sequence for the crossed product An,θ ∼= An−1,θ ⋊α Z as in Theorem 2.1(d) in [33]. Acknowledgment. I would like to thank Graham Denham, George Elliott, Herve Oyono-Oyono, and Tim Steger for very helpful discussions. I would like to thank N. C. Phillips for his thoughtful suggestions and for bringing the possibility of generalizing Lemma 3.1 of [30] to my attention. I am grateful to Alan T. Paterson for reading the paper and suggesting several helpful comments. I would also like to thank Paul Milnes and the Department of Mathematics of the University of Western Ontario, where some parts of this work were done. Appendix A. The Smith normal form The Smith normal form is a very important tool for studying integer matrices. We refer to [27] for this interesting topic and its applications. Definition A.1. Let α, β ∈ End(Zm). We say that α is equivalent to β over Z (and write α equiv β) if there exist u, v ∈ Aut(Zm) such that u ◦ α ◦ v = β. Similarly, if A and B are integer m × m matrices, A is equivalent to B if there exist U, V ∈ GL(m, Z) such that UAV = B. Recall that α equiv β, if and only if cokerα ∼= coker β, if and only if α and β have the same Smith normal form. Also, A equiv B if and only if B is obtainable from A by a finite number of elementary operations. An elementary operation on an integer matrix is one of the following types: interchanging two rows (or two columns), adding an integer multiple of one row (or column) to another, and multiplying a row (or K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 23 column) by −1. Now, we recall a fundamental theorem for integer matrices (cf. [27, p. 26]). Theorem A.2 (Smith Normal Form). Let A be an m × m integer matrix. Then A is equivalent to a diagonal matrix diag(d1, . . . , dr, 0, . . . , 0), where r is the rank of A, and the integers d1, . . . , dr satisfy didi+1 for i = 1, . . . , r − 1. Appendix B. Nilpotent and unipotent linear mappings Definition B.1. Let V be a (complex) vector space. A mapping ǫ ∈ EndCV is called nilpotent (respec- tively, unipotent) if ǫk = 0 (respectively, (ǫ − id)k = 0) for some positive integer k. The minimum value of k with this property is called the degree of ǫ, denoted deg(ǫ). As an example, every upper (respectively, lower) triangular matrix with zeros on the diagonal is nilpotent. Also, the matrix Sn defined in Section 2 is a unipotent matrix of degree n. Note that all eigenvalues of a nilpotent (respectively, unipotent) matrix are zeros (respectively, ones). In particular, every unipotent matrix is invertible, and every unipotent endomorphism is an automorphism. Corollary B.2. Let V be a finite dimensional complex vector space and ǫ be a nilpotent (respectively, unipotent) endomorphism of V . Then deg(ǫ) is equal to the maximum order of its Jordan blocks. Proof. It suffices to prove the statement for the nilpotent case. Since all the eigenvalues of ǫ are zero, each Jordan block is a zero matrix of order one or is of the form   0 0 0 ... 0 1 0 . . . · · · 0 · · · 1 . . . 0 0 . . . 0 0 0 ... 0 1 0   which is a nilpotent matrix and its degree is the same as its order, which is greater than 1. The rest of proof is clear. (cid:3) Definition B.3. Let V be a finite dimensional complex vector space and ǫ ∈ EndCV be nilpotent (respectively, unipotent). We say that ǫ is of maximal degree if deg(ǫ) = dim V . Corollary B.4. Let V be an n-dimensional complex vector space and ǫ ∈ EndCV be nilpotent (respec- tively, unipotent). Then deg(ǫ) ≤ n. If deg(ǫ) = n, then the Jordan normal form of ǫ is the full Jordan block of order n with 0's on the diagonal. In particular, all nilpotent (respectively, unipotent) matrices of maximal degree acting on V are similar. Proof. Use the preceding corollary and the Jordan normal form theorem. (cid:3) Example B.5. Let B = [bij]n×n be any upper triangular matrix, whose diagonal elements are all zeros (respectively, ones), and whose entries bi,i+1 for i = 1, . . . , n − 1 are all nonzero. Then B is nilpotent i=1 bi,i+1, which is a nonzero number. Then one can easily see that Bn = 0 (respectively, (B − I)n = 0) and Bn−1 (respectively, (B − I)n−1) is a matrix with b appearing on the upper-right corner and zeros elsewhere. So deg(B) = n, i.e. B is of maximal degree. (respectively, unipotent) of maximal degree. In fact, let n be the order of B and let b := Qn−1 Lemma B.6. Let V be a complex vector space and ǫ ∈ EndCV be nilpotent of degree k. Then exp(ǫ) is unipotent of degree k. Moreover, exp(ǫ) − id is similar to ǫ. 24 KAMRAN REIHANI Proof. For the first part, we know that exp(ǫ) − id = ǫ + ǫ2/2! + . . . + ǫk−1/(k − 1)! = ǫω, where ω := id+ǫ/2!+. . .+ǫk−2/(k−1)! commutes with ǫ and is invertible since it is unipotent. So, (exp(ǫ)−id)r = (ǫω)r = ǫrωr for all positive integers r. Thus exp(ǫ) − id is unipotent with the same degree of ǫ. For the second part, using the Jordan normal form of ǫ, it is sufficient to prove the statement for the special case when ǫ is a Jordan block with zeros on the diagonal. Since in this case ǫ is of maximal degree, by the first part, exp(ǫ) − id is also of maximal degree. Therefore they are similar by Corollary B.4. (cid:3) Appendix C. Endomorphisms and derivations of exterior algebras We refer to the Chapter 5 of [11] for general properties of exterior algebras and mappings between them. Let V be a (complex) vector space and φ : V → V be a linear mapping. Then φ can be extended in a unique way to a homomorphism ∧∗ φ : Λ∗V → Λ∗V such that ∧∗ φ(1) = 1, yielding ∧∗ φ(x1 ∧ . . . ∧ xp) = φ(x1) ∧ . . . ∧ φ(xp), (xi ∈ V ). Also, φ can be extended in a unique way to a derivation D ∗ φ : Λ∗V → Λ∗V , yielding D ∗ φ(x1 ∧ . . . ∧ xp) = p Xi=1 x1 ∧ . . . ∧ φ(xi) ∧ . . . ∧ xp (p ≥ 2, xi ∈ V ). We define ∧r φ := ∧∗ φΛr V and D r φ := D ∗ φΛr V as induced linear mappings on the r-th exterior power of V for r ≥ 0. Then we have ∧∗ φ = Mr≥0 ∧r φ , D ∗ φ = Mr≥0 D r φ . One can easily show that ∧∗( φ ◦ ψ) = (∧∗ φ) ◦ (∧∗ ψ) and D ∗([ φ, ψ]) = [D ∗ φ, D ∗ φ] (cf. equations (5.20) and (5.25) in [11]). Lemma C.1. With the above notation, if φ : V → V is nilpotent, then ∧r φ and D r φ are also nilpotent for r ≥ 1. If V is finite dimensional, then D ∗ φ is nilpotent. Proof. Assume that φt = 0 for some t ∈ N. We have (∧r φ)t(x1 ∧ . . . ∧ xr) = φt(x1) ∧ . . . ∧ φt(xr) = 0. So, (∧r φ)t = 0, which means that ∧r φ is nilpotent. For D r φ, we know that D r φ(x1 ∧ . . . ∧ xr) = i=1 x1 ∧ . . . ∧ φ(xi) ∧ . . . ∧ xr and one can easily deduce that Pr D r φp(x1 ∧ . . . ∧ xr) = Xi1+...+ir =p p! (i1)! . . . (ir)! φi1 x1 ∧ . . . ∧ φir xr . Now since i1 + . . . + ir = p, there exists an ij with ij ≥ p/r. So, if p ≥ rt then φij = 0 and D r φp = 0. Thus D r φ is nilpotent. (ij ≥0) For the next part, let m := dim V . Since D ∗ φ = Lm (D ∗ φ)mt = 0, hence D ∗ φ is nilpotent, too. r≥0 D r φ and φ0 = 0, from the first part we have (cid:3) Lemma C.2. Let φ : V → V be a nilpotent linear mapping. Then exp(D ∗ φ) = ∧∗ exp( φ) on Λ∗V . Proof. We have exp(D ∗ φ)(x1 ∧ . . . ∧ xr) = Xp≥0 = Xp≥0 1 p! 1 p! D r φp(x1 ∧ . . . ∧ xr) p! (i1)! . . . (ir)! ( Xi1+...+ir =p (ij ≥0) φi1 x1 ∧ . . . ∧ φir xr) K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 25 φi1 x1 ∧ . . . ∧ φir xr 1 (i1)! . . . (ir)! = Xij ≥0 φi1 x1) ∧ . . . ∧ (Xir ≥0 = (Xi1≥0 = (exp( φ)x1) ∧ . . . ∧ (exp( φ)xr) = ∧∗ exp( φ)(x1 ∧ . . . ∧ xr), 1 (i1)! 1 (ir)! φir xr) which yields the result. Note that all sums in the above equalities are finite according to the previ- ous lemma. (cid:3) Remark C.3. The nilpotency of φ is not necessary in the preceding lemma. In fact, one may use the definition of exp : gl(Λ∗V ) → GL(Λ∗V ). More precisely, define s : R → GL(Λ∗V ) by s(t) = ∧∗ exp(t φ). s(0) = D ∗ φ), and we Then one may check that s is the 1-parameter subgroup generated by D ∗ φ (i.e. have s(1) = ∧∗ exp( φ). Corollary C.4. Let φ : V → V be a nilpotent linear mapping, and set ǫ := φ + id. Then ∧rǫ − id is similar to D r φ for r ≥ 0. Proof. We know from Lemma B.6 that exp( φ) − id is similar to φ, hence exp( φ) is similar to φ + id = ǫ. So ∧rǫ − id ∼ ∧r exp( φ) − id = exp(D r φ) − id ∼ D r φ. (cid:3) We refer to Section II.7 of [15] for studying the irreducible representations of the Lie algebra sl(2, C). Appendix D. Representation theory of sl(2, C) Let sl(2, C) denote the special linear Lie algebra over C2 defined by sl(2, C) := {a ∈ M2(C) Tr(a) = 0}. It is well known that sl(2, C) is a 3-dimensional simple complex Lie algebra. One can check that B := {h :=   is a basis for this Lie algebra. 1 0 0 −1   , e :=   0 1 0 0   , f :=   0 1 0 0 }   The following theorem is the foundation of representation theory of semisimple Lie algebras including sl(2, C). It is stated and proved in the Subsection II.6.3 of [15]. Theorem D.1 (Weyl). Every finite dimensional representation of a semisimple Lie algebra is completely reducible, namely, it can be decomposed into a direct sum of irreducible representations. Proposition D.2. Let V be an n-dimensional complex vector space with a basis {e1, . . . , en}. Then the following equalities (for i = 1, . . . , n) (a) πn(h)ei = (2i − n − 1)ei; (b) πn(e)ei = i(n − i)ei+1; (en+1 := 0) (c) πn(f )ei = ei−1; (e0 := 0). define a representation πn : sl(2, C) → gl(V ). We call πn defined in the previous proposition the canonical representation of sl(2, C) on V associated with the basis {e1, . . . , en}. We recall the following theorem from Section II.7 of [15]. Theorem D.3. Let πn be the representation described above. Then 26 KAMRAN REIHANI (i) πn is an irreducible representation of sl(2, C). (ii) Any n-dimensional irreducible representation of sl(2, C) is equivalent to πn. (iii) Let V be a finite dimensional sl(2, C)-module and define Vα = {v ∈ V h.v = α v} for α ∈ C. Then V decomposes into a direct sum of irreducible submodules (Weyl), and in any such decomposition, the number of summands is precisely dim V0 + dim V1. Appendix E. Some computer codes and a counterexample To compute the K-groups of C(Tn) ⋊α Z, one should first compute the kernels and cokernels of the following integer matrices (11) Ar := ∧r A − I (n r) for r = 0, 1, . . . , n, where A is the integer matrix corresponding to α acting on Zn. We have written two Maple codes to obtain this goal. The first one is an auxiliary procedure called exterior(r,A), which computes the r-th exterior power of a given n × n integer matrix A for r = 1, . . . , n as follows (note that ∧0A := I1) > exterior:=proc(r,A) > local n,N,Q,E,i,j; > n:=linalg[rowdim](A); > N:=binomial(n,r); > Q:=combinat[choose](n,r); > E:=array(1..N,1..N); > for i from 1 to N do > for j from 1 to N do > E[i,j]:=linalg[det](linalg[submatrix](A,Q[i],Q[j])); > od; > od; > RETURN(evalm(E)); > end; The second code, which calls the first one, is called po(A). It lists the factorized characteristic polynomials of the Smith normal forms of Ar for r = 0, 1, . . . , n given a matrix A > po:=proc(A) > local r,n,x,p; > n:=linalg[rowdim](A); > print(x); > for r from 1 to n do > p:=factor(linalg[charpoly](linalg[ismith](exterior(r,A)-1),x)); > print(p); > od; > end; These factorized polynomials encode the diagonal entries of the Smith normal forms of the matrices Ar for r = 0, 1, . . . , n. In particular, one can easily find the kernels and cokernels of these matrices. To see this, observe that the kernel of an m × m integer matrix is isomorphic to a torsion-free finitely gener- ated abelian group, whose rank is the number of zeros on the diagonal of its Smith normal form, and the cokernel is isomorphic to ⊕m j=1Zkj , where k1, . . . , km are the diagonal entries of the Smith normal form. The following example computes the K-groups of A6,θ, which was mentioned in Remark 1.7 as a coun- terexample for [16, Proposition 2.17]. K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 27 Example E.1. K0(A6,θ) ∼= Z13 and K1(A6,θ) ∼= Z13 ⊕ Z2. Proof. Using the procedure po(A) for the matrix A := S6 as defined at the beginning of Section 3 we get po(A); x x(x − 1)5 x3(x − 1)12 x3(x − 2)(x − 1)16 x3(x − 1)12 x(x − 1)5 x Consequently, we have the following table representing the kernels and cokernels of S6,r := ∧r S6 − I(6 r) for r = 0, 1, . . . , 6. r ker S6,r coker S6,r 0 1 2 3 4 5 6 Z Z Z3 Z3 Z3 Z Z Z Z Z3 Z3 ⊕ Z2 Z3 Z Z Now, using Theorem 2.1 we have K0(A6,θ) ∼=(coker S6,0 ⊕ coker S6,2 ⊕ coker S6,4 ⊕ coker S6,6) ⊕ (ker S6,1 ⊕ ker S6,3 ⊕ ker S6,5) ∼=(Z ⊕ Z3 ⊕ Z3 ⊕ Z) ⊕ (Z ⊕ Z3 ⊕ Z) = Z13, K1(A6,θ) ∼=(coker S6,1 ⊕ coker S6,3 ⊕ coker S6,4) ⊕ (ker S6,0 ⊕ ker S6,2 ⊕ ker S6,4 ⊕ ker S6,6) ∼=(Z ⊕ Z3 ⊕ Z2 ⊕ Z) ⊕ (Z ⊕ Z3 ⊕ Z3 ⊕ Z) = Z13 ⊕ Z2. (cid:3) Appendix F. Applications to C ∗(Dn) This appendix is devoted to some applications of the results of this paper to the C ∗-algebras studied in [33], in which the authors promised to completely classify the simple infinite dimensional quotients of the group C ∗-algebra of their interest C ∗(Dn) by their K-theoretic invariants in a later work. In this context, the discrete group Dn is a higher dimensional analogue of the discrete Heisenberg group H3, and is defined by Dn = h x, y0, y1, . . . , yn xy0 = y0x, yiyj = yjyi for 0 ≤ i, j ≤ n, [x, yj] = yj−1 for 1 ≤ j ≤ n i, where [x, y] := xyx−1y−1. The group Dn can be represented as a semidirect product Zn+1 ⋊η Z, where the group homomorphism η : Z → GL(n + 1, Z) is such that η(k) is the matrix, whose (i, j)-entry is given by (cid:0) k j−i(cid:1) as defined in [33]. This realization allows us to study K-theory of C ∗(Dn). 28 KAMRAN REIHANI Proposition F.1. Ki(C ∗(Dn)) ∼= Ki(An+1,θ) for i = 0, 1. In particular, rank K0(C ∗(Dn)) = rank K1(C ∗(Dn)) = an+1. Proof. Since Dn ∼= Zn+1 ⋊η Z, so C ∗(Dn) ∼= C ∗(Zn+1) ⋊η Z ∼= C(Tn+1) ⋊η Z and the integer matrix corresponding to η is the (n + 1) × (n + 1) matrix Mn introduced in [33], which is precisely the matrix Sn+1 defined in Section 5 to describe the linear structure of Anzai transformations on Tn+1. The rest of the proof follows from the Theorem 2.1. (cid:3) As some higher-dimensional analogues of the irrational rotation algebras Aθ, all simple infinite-dimensional quotients of C ∗(Dn) have been classified in Theorem 3.2 of [33]. These consist of the C ∗-algebras An,θ for some irrational parameter θ, and a few more classes of C ∗ algebras denoted by A(n) , which are of the form C(Yi ×Tn−i)⋊φi Z for some suitable finite sets Yi and minimal homeomorphisms φi for i = 1, 2, . . . , n−1. Then it is proved in Theorem 4.8 of [33] that A(n) is a certain transformation group C ∗-algebra of some affine Furstenberg transformation on Tn−i. We conclude the following results. ), where B(n) ∼= MCi(B(n) i i i i Corollary F.2. Let A be a simple infinite dimensional quotient of C ∗(Dn). Then rank K0(A) = i ∈ {0, 1, . . . , n − 1} that is uniquely determined by the isomorphism rank K1(A) = an−i for some A ∼= C(Yi × Tn−i) ⋊φi Z as in Theorem 3.2 of [33]. Proof. It is proved that A is isomorphic to a matrix algebra over a Furstenberg transformation group C ∗-algebra B(n) ) for j = 0, 1. The rest of the proof is clear from the preceding theorem. (cid:3) on Tn−i for some suitable i ∈ {0, 1 . . . , n} [33, Theorem 4.8]. So Kj(A) ∼= Kj(B(n) i i We saw in Proposition 6.5 that {an} is a strictly increasing sequence. Therefore the preceding corollary is a first step towards the classification of the simple infinite dimensional quotients of C ∗(Dn) by means of K-theory. But as is seen, the rank of the K-groups alone can not distinguish the algebras at the same "level" (i.e. those algebras that are included in the same class, but with different values of the parameters). The other powerful K-theoretic invariant that helps us do this is the trace invariant, i.e. the range of the unique tracial state acting on the K0-group. Proposition F.3. Suppose A ∼= C(Yi × Tn−i) ⋊φi Z is a simple infinite dimensional quotient of C ∗(Dn) as in Theorem 3.2 in [33]. Then A has a unique tracial state τ and τ∗K0(A) = 1 (Z + Zϑi), where Ci Ci = Yi and e2πiϑi = ζi = (−1)Ci+1ηCi i as in Lemma 4.6 and Theorem 4.8 in [33]. Proof. Following Theorem 4.8 in [33], A is isomorphic to MCi(B(n) is a simple Furstenberg transformation group C ∗-algebra with the irrational parameter ζi = (−1)Ci+1ηCi . i By Corollary 1.5, B(n) ) = Z + Zϑi, where e2πiϑi = ζi (see [16, Theorem 2.23] or Theorem 7.2). Thus A has the unique tracial state τ = ( 1 Tr) ⊗ τ , in which Ci Tr is the usual trace on MCi(C), and so τ∗K0(A) = 1 (cid:3) Ci has a unique tracial state τ . Moreover, τ∗K0(B(n) (Z + Zϑi) [16, Lemma 3.5]. ) = MCi(C) ⊗ B(n) , where B(n) i i i i i Finally, we can characterize all simple infinite-dimensional quotients of C ∗(Dn). Proposition F.4. An,θ ∼= An′,θ ′ if and only if n = n′ and there exists an integer k such that θ = k ± θ′. More generally, let A(n) ∼= C(Yi × Tn−i) ⋊φi Z be a simple infinite dimensional quotient of C ∗(Dn) with the structure constants λ, µ1, . . . , µi as on p. 165-166 of [33], and let A(n′) i′ ∼= C(Y′ Z be a simple infinite dimensional quotient of C ∗(Dn′ ) with the structure constants λ′, µ′ i′ . Suppose that Ci = Yi and C ′ i′ × Tn′−i′ 1, . . . , µ′ i′ and i′ . Then A(n) ∼= A(n′) i′ = Y′ ) ⋊φ′ i i i i′ λ( Ci i+1)µ (Ci i ) 1 ( Ci i−1) 2 µ . . . µCi if and only if n − i = n′ − i′, Ci = C ′ i = λ′( C ′ −1) . . . µ′ (C ′ i′ )µ′ i′ +1)µ′ ( C ′ C ′ i′ i′ 1 2 i′ i′ i′ i′ or λ( Ci i+1)µ (Ci i ) 1 ( Ci i−1) 2 µ . . . µCi i = (λ′( C ′ i′ +1)µ′ i′ 1 i′ (C ′ i′ )µ′ 2 ( C ′ i′ i′ −1) . . . µ′ i′ C ′ i′ )−1. K-THEORY OF FURSTENBERG TRANSFORMATION GROUP C ∗-ALGEBRAS 29 Proof. Use the previous proposition and the fact that {an} is a strictly increasing sequence (see Propo- sition 6.5). Note that ζi = (−1)Ci+1ηCi i = λ( Ci i+1)µ (Ci i ) 1 ( Ci i−1) 2 µ . . . µCi i by the last equation on p. 171 of [33]. (cid:3) Remark F.5. Note that Ci = Yi is completely determined by the structural constants λ, µ1, . . . , µi−1 (which are roots of unity). More precisely, by calculations on p. 165-166 of [33], we have Ci = min{r ∈ N λr = λ(r 2)µr 1 = . . . = λ(r i)µ ( r i−1) 1 ( r i−2) 2 µ . . . µr i−1 = 1}. For an explicit example in lower dimensions, see [26, Lemma 5.4]. References [1] J. Anderson, W. Paschke, The rotation algebra, Houston J. Math. 15(1) (1989), 1 -- 26. [2] M. Atiyah, K-theory, W. A. Benjamin, New York, 1967. [3] B. Blackadar, K-theory for operator algebras, Second edition, Mathematical Sciences Research Institute Publications 5, Cambridge University Press, Cambridge, 1998. [4] E.G. Effros and F. Hahn, Locally compact transformation groups and C ∗-algebras, Mem. Amer. Math. Soc. No. 75, Providence, RI (1967). [5] R. C. Entringer, Representation of m as Pn [6] P. Erdos, Extremal problems in number theory, 1965 Proc. Sympos. Pure Math., Vol. VIII, pp. 181 -- 189, Amer. Math. k=−n ǫkk, Canad. Math. Bull. 11 (1968), 289 -- 293. Soc., Providence, R.I. [7] R. Exel, Rotation numbers for automorphisms of C ∗-algebras, Pacific J. Math. 127(7) (1981), 31 -- 89. [8] H. Furstenberg, Recurrence in ergodic theory and combinatorial number theory, Princeton University Press, Princeton, NJ, 1981. [9] H. Furstenberg, Strict ergodicity and transformation of the torus, Amer. J. Math. 83 (1961), 573 -- 601. [10] V. V. Gorbatsevich, A. L. Onishchik and E. B. Vinberg, Foundations of Lie theory and Lie transformation groups, Encyclopaedia of Mathematical Sciences 20, Springer-Verlag, Berlin Heidelberg, 1997. [11] W. Greub, Multilinear algebra, Second edition, Springer-Verlag, New York-Heidelberg, 1978. [12] F. J. Hahn, On affine transformations of compact abelian groups, Amer. J. Math. 85 (1963), 428 -- 446. [13] F. J. Hahn and W. Parry, Minimal systems with quasi-discrete spectrum, J. London, Math. Soc. 40 (1965), 300 -- 323. [14] L. Hodgkin, On the K-theory of Lie groups, Topology, 6 (1967), 1 -- 36. [15] J. E. Humphreys, Introduction to Lie algebras and representation theory, Second printing, revised, Graduate Texts in Mathematics 9, Springer-Verlag, New York-Berlin, 1978. [16] R. Ji, On the crossed product C ∗-algebras associated with Furstenberg transformations on tori, Ph.D. dissertation, State University of New York at Stony Brook, August 1986. [17] M. Karoubi, K-theory, Springer-Verlag, Berlin Heidelberg, 1978. [18] K. Kodaka, The positive cones of K0-groups of crossed products associated with Furstenberg transformations on the 2-torus, Proc. Edinburgh Math. Soc. (2) 43 (1), (2000), 167 -- 175. [19] H. Lin, classification of simple C ∗-algebras of tracial topological rank zero, Duke Math. J. 125(1) (2004), 91 -- 119. [20] Q. Lin, N. C. Phillips, Direct limit decomposition for C ∗-algebras of minimal diffeomorphisms, Operator algebras and applications, 107 -- 133, Adv. Stud. Pure Math., 38, Math. Soc. Japan, Tokyo, 2004. [21] H. Lin, N. C. Phillips, Crossed products by minimal homeomorphisms J. Reine Angew. Math. 641 (2010), 95 -- 122. [22] Q. Lin, N. C. Phillips, Ordered K-theory for C ∗-algebras of minimal homeomorphisms, Operator algebras and operator theory (Shanghai, 1997), 289 -- 314, Contemp. Math., 228, Amer. Math. Soc., Providence, RI, 1998. k=−N ǫkk, Proc. Amer. Math. Soc. 18 (1967), 182 -- 184. [23] J. H. van Lint, Representations of 0 as PN [24] P. Milnes, Groups and C ∗-algebras from a 4-dimensional Anzai flow, Math. Nachr. 235 (2002), 129 -- 141. [25] P. Milnes and S. Walters, Simple quotients of the group C ∗-algebra of a discrete 4-dimensional nilpotent group, Houston J. Math. 19 (1993), 615 -- 636. [26] P. Milnes and S. Walters, Simple infinite dimensional quotients of C ∗(G) for discrete 5-dimensional nilpotent groups G, Illinois J. Math. 41(2) (1997), 315 -- 340. [27] M. Newman, Integral matrices, Pure and Applied Mathematics 45, Academic Press, New York and London, 1972. [28] J. A. Packer, K-theoretic invariants for C ∗-algebras associated to transformations and induced flows, J. Funct. Anal. 67 (1986), 25 -- 59. [29] N. C. Phillips, Cancellation and stable rank for direct limits of recursive subhomogeneous algebras , Trans. Amer. Math. Soc. 359 (2007), 4625 -- 4652. [30] N. C. Phillips, Examples of different minimal diffeomorphisms giving the same C ∗-algebras, Israel J. Math. 160 (2007), 189-217. 30 KAMRAN REIHANI [31] M. Pimsner and D. Voiculescu, Exact sequences for K-groups and Ext-groups of certain crossed product C ∗-algebras, J. Operator Theory, 4 (1980), 93 -- 118. [32] S. C. Power, Simplicity of C ∗-algebras associated to minimal dynamical systems, J. London Math. Soc. 18 (1978), 534 -- 538. [33] K. Reihani and P. Milnes, Analysis on discrete cocompact subgroups of the generic filiform Lie groups, Acta. Math. Hung., 112 (1-2) (2006), 157 -- 179. [34] M. Rieffel, C ∗-algebras associated with irrational rotations, Pacific J. Math. 93(2) (1981), 415 -- 429. [35] J. Rosenberg, Group C ∗-algebras and topological invariants, in Operator algebras and group representations, vol. I, Pitman, London, 1984, pp. 95 -- 115. [36] C. Schochet, Topological methods for C ∗-algebras II: Geometric resolutions and the Kunneth formula, Pacific J. Math. 98 (2) (1982), 443-458. [37] J. L. Taylor, Banach algebras and topology, Algebras in analysis, (J. H. Williamson, ed.), Academic Press, London, 1975, pp. 118 -- 186. [38] J. Tomiyama, Invitation to C ∗-algebras and topological dynamics, Advanced Series in Dynamical Systems, vol. 3, World Scientific, Singapore, 1987. [39] S. Walters, K-groups and classification of simple quotients of group C ∗-algebras of certain discrete 5-dimensional nilpotent groups, Pacific J. Math. 202(2) (2002), 491 -- 509. Department of Mathematics, University of Kansas, Lawrence, KS 66045-7594 E-mail address: [email protected]
1710.08622	1	1710	2017-10-24T07:10:19	The Matricial Range of $E_{21}$	[ "math.OA", "math.FA" ]	The matricial range of the $2\times2$ matrix $E_{21}$ (i.e., the $2\times 2$ unilateral shift) is described very simply: it consists of all matrices with numerical radius at most $1/2$. The known proofs of this simple statement, however, are far from trivial and they depend on subtle results on dilations. We offer here a brief introduction to the matricial range and a recap of those two proofs, following independent work of Arveson and Ando in the early 1970s.	math.OA	math	THE MATRICIAL RANGE OF E21 MARTÍN ARGERAMI Abstract. The matricial range of the 2 × 2 matrix E21 (i.e., the 2 × 2 unilateral shift) is described very simply: it consists of all matrices with numerical radius at most 1/2. The known proofs of this simple state- ment, however, are far from trivial and they depend on subtle results on dilations. We offer here a brief introduction to the matricial range and a recap of those two proofs, following independent work of Arveson and Ando in the early 1970s. Contents 1. Introduction 2. Preliminaries 3. The Matricial Range 4. Unitary Dilations, Numerical Radius, and Matricial Range 5. Ando's Characterizations of the Numerical Radius 6. Toeplitz Matrices 7. Nilpotent Dilations and Matricial Range 8. Characterizations of the Numerical Radius Acknowledgements References 1 2 4 17 21 32 36 38 38 39 1. Introduction In 1969, W. Arveson published a striking paper in Acta Mathematica, called Subalgebras of C∗-algebras [2]. In this paper he developed a non- commutative analog of the Choquet theory for function spaces. His paper is a wonderful mixture of technical prowess and deep thinking about how to rightly generalize certain ideas about function spaces to the non-commutative setting. A key feature of his paper was the use of complete positivity as a non- commutative replacement for the role that positivity has in the commutative case. A few years later he published an equally remarkable paper [3]. Besides containing his essential Boundary Theorem, this paper defined the matricial range of an operator. As a consequence of an analysis of nilpotent dilations, he was able to explicitly characterize the matricial range of the 2× 2 matrix unit E21. This is one of the very few non-trivial (that is, non-normal) cases 1 2 MARTÍN ARGERAMI where the matricial range has been determined (the other significant one is the unilateral shift on an infinite-dimensional Hilbert space, which is more or less straightforward). Almost concurrently, Ando published his results characterizing the nu- merical range [1]. As a direct byproduct of his results one recovers the characterization of the matricial range of E21. The goal of this article is to describe Arveson and Ando's techniques, together with basic characterizations of the matricial range. The results we offer follow closely the originals, but several of the proofs are new. In the case of Ando, we have also strived to fill in the details from his very condensed arguments. 2. Preliminaries Throughout, H will be a Hilbert space, with inner product (cid:104)·,·(cid:105). We use B(H) to denote the (C∗, von Neumann) algebra of bounded operators on H; and K(H) for the compact operators. When dim H = n, we canonically identify H with Cn and B(H) with Mn(C), the set of n×n complex matrices. This is done by fixing an orthonormal basis {ξj} ⊂ H and considering the rank-one operators Ekjξ = (cid:104)ξ, ξj(cid:105) ξk, ξ ∈ H. These are called matrix units and they are characterized up to unitary equiv- alence (i.e., choice of the orthonormal basis) by the relations E∗ kj = Ejk, (2.1) EkjEh(cid:96) = δjh Ek(cid:96), z = 1}, and D = {z ∈ C : Ekk = I. We write T = {z ∈ C : z < 1} for the unit circle and unit disk respectively. When needed for clarity, we will write En kj kj ∈ Mn(C). Of particular importance will be, for each to emphasize that En n, the unilateral shift: (cid:21) k=1 S2 = E21 = , Sn = Ek+1,k = (cid:20)0 0 1 0 In the infinite-dimensional case, when {ξj}j∈N is an orthonormal basis of H, the associated shift operator S is the linear operator induced by S : ξj (cid:55)−→ ξj+1. An element x of a normed space is said to be contractive if (cid:107)x(cid:107) ≤ 1. A linear map φ : A → B between two C∗-algebras is positive if it maps positive elements to positive elements; it is completely positive if φ(n) is positive for all n ∈ N, with φ(n) the nth amplification φ(n) : Mn(B(H)) → Mn(B(K)), given by φ(n)(A) = [φ(Akj)]kj. We will mostly consider completely positive maps which are also unital; these are commonly named ucp (unital, completely n(cid:88) n−1(cid:88) k=1  0 1 0 0 0 1 0 ···  . 0 0 0 0 0 ··· 0 ··· 0 ··· ... 1 0 THE MATRICIAL RANGE OF E21 3 positive). The basics of completely positive and completely contractive maps are covered in many texts. We refer the reader to the following canonical three: [5, 15, 16]. We mention one explicit result that we will use: Proposition 2.1 (Choi). Let φ : Mn(C) → B(H) be a linear map. Then φ is completely positive if and only if  E11 ... En1 φ(n)   ≥ 0. ··· E1n ... ... ··· Enn The operator system generated by T ∈ B(H) is the space OS(T ) = span{I, T, T ∗}. More generally, an operator system is a unital selfadjoint subspace S of B(H). When one considers ucp maps as morphisms, an operator system S can be characterized by its sequence of positive cones Mn(S)+. Arveson's Extension Theorem [2, Theorem 1.2.3] guarantees that if S ⊂ B(H) is an operator system and φ : S → B(K) is completely posi- tive, there exists a completely positive extension φ : B(H) → B(K). For any fixed operator system S, the set of ucp maps S → B(K) is BW-compact, where the BW-topology is that given by pointwise weak-operator conver- gence. Given T ∈ B(H), its numerical range is the set W1(T ) = {f (T ) : f is a state} = {Tr(HT ) : H ≥ 0, Tr(H) = 1}. We note that the equality above is not entirely obvious, since the right-hand- side only accounts for the normal states. But since the normal states are the predual of B(H), any state is a weak∗ (that is, pointwise) limit of normal states; so, as W1(T ) is closed -- by an easy application of Banach -- Alaoglu -- , the set of all f (T ) where f runs over all the states, is the same as the set of all f (T ) where f runs over all the normal states. The numerical range is always compact and convex. The numerical radius of T is the number w(T ) = sup{λ : λ ∈ W1(T )}. Remark 2.2. The numerical range is classically defined as W (T ) = {(cid:104)T ξ, ξ(cid:105) : ξ ∈ H}. It turns out that W (T ) is always dense in W1(T ). It is also convex, as proven in the Toeplitz -- Hausdorff Theorem. The fact that W1(T ) is convex, on the other hand, follows from a straightforward computation. As it is common -- although not standard -- we will refer by "strong" con- vergence of a net, to convergence in the strong operator topology; and by "weak" convergence, to convergence in the weak operator topology. 4 MARTÍN ARGERAMI 3. The Matricial Range The matricial range of T ∈ B(H) is the sequence W(T ) = {Wn(T ) : n ∈ N}, where Wn(T ) = {ϕ(T ) : ϕ : OS(T ) → Mn(C) is ucp}. In light of Arveson's Extension Theorem, the matricial range of T does not change if we consider C∗(T ) or even B(H) as the domain of the ucp maps in the definition of Wn(T ). A classic survey on the topic is [12]. One is tempted to include the set W∞(T ) = {φ(T ) : φ : OS(T ) → B((cid:96)2(N)) is ucp } (or even higher-dimensional versions in the non-separable case) in the list {Wn(T ) : n ∈ N}. But we have the following: are equivalent: Proposition 3.1. Let S ∈ B(H), T ∈ B(K). The following statements (1) Wn(S) = Wn(T ) for all n ∈ N; (2) W∞(S) = W∞(T ). Proof. Assume first that Wn(S) = Wn(T ) for all n ∈ N. Let X ∈ W∞(S). So X = φ(S) ∈ B((cid:96)2(N)) for some ucp map φ. Let {Pj} be an increasing net of finite-dimensional projections with Pj → I strongly. Let k(j) be the rank of Pj. We can think of PjXPj ∈ Mk(j)(C). So PjXPj = Pjφ(S)Pj ∈ Wk(j)(S) = Wk(j)(T ). Then there exists a ucp map ψj : B(K) → Mk(j)(C) with ψj(T ) = PjXPj. Let ψ be a BW-cluster point of the net {ψj}. Then ψ(T ) = limj ψj(T ) = limj PjXPj = X, so X ∈ W∞(T ). We have proven that W∞(S) ⊂ W∞(T ), and exchanging roles we get the equality. Conversely, assume now that W∞(S) = W∞(T ). Fix n ∈ N. Let X ∈ Wn(S). By identifying Mn(C) with the "upper left corner" of B((cid:96)2(N)), we may assume X ∈ W∞(S) = W∞(T ). Then there exists a ucp map ψ : B(K) → B((cid:96)2(N)) with ψ(T ) = X. If P is the projection of rank n such that X = P XP , then P ψP can be seen as a ucp map B(K) → Mn(C). So X ∈ Wn(T ). We have proven that Wn(S) ⊂ Wn(T ), and now reversing roles (cid:3) we get Wn(S) = Wn(T ). We will also consider briefly the spatial matricial range Ws(T ) = {Ws n(T ) : n ∈ N}, where Ws n(T ) = {V ∗T V : V : Cn → H isometry}. The following result is due to Bunce -- Salinas [7, Theorem 2.5]. The form we use is taken from [4, pp. 335 -- 336]; the proof follows mostly [8, Lemma THE MATRICIAL RANGE OF E21 5 II.5.2], but we use a slightly sharper version of Glimm's Lemma than the one used by Davidson. Lemma 3.2 (Bunce -- Salinas). Let φ : B(H) → Mn(C) be ucp and such that φ(L) = 0 for every compact operator L, and let A ⊂ B(H) be a separable C∗-algebra. Then there exists a sequence of isometries Vk : Cn → H such that Vk → 0 weakly and (cid:107)φ(T ) − V ∗ k T Vk(cid:107) → 0, T ∈ A. (cid:2)ξ1 Proof. For a fixed orthonormal basis ξ1, . . . , ξn of Cn, consider the map Φ : Mn(A) → C given by (cid:3)(cid:124) ∈ (Cn)n. The fact that φ is ucp makes Φ a Φ(A) = (cid:104)φ(n)(A) ξ, ξ(cid:105), ξn where ξ = 1√ n state. By construction, Φ(A) = 0 if all entries of A are compact; and the compact operators of Mn(A) are precisely the matrices where all entries are compact. Thus Glimm's Lemma (see [8, Lemma II.5.1], but here we use the exact form of [6, Lemma 1.4.11]) applies to the C∗-algebra Mn(A) and the state Φ, and we get an orthonormal sequence vectors {ηk} ⊂ H n, where ··· Define linear maps Xk : Cn → H by Xkξj = and k big enough so that (cid:104)√ n ηk √ n ηk j (cid:105) − δh,j < ε, n ηk h, √ j . For ξ, η ∈ Cn, ε > 0, (cid:104)(X∗ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) k Xk − In)ξ, η(cid:105) = (cid:104)Xkξ, Xkη(cid:105) − (cid:104)ξ, η(cid:105) (cid:88) (cid:104)ξ, ξh(cid:105)(cid:104)η, ξj(cid:105) ((cid:104)√ (cid:88) = h,j (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n ηk j (cid:105) − δh,j) √ n ηk h, ≤ ε (cid:104)ξ, ξh(cid:105)(cid:104)η, ξj(cid:105) ≤ ε n2(cid:107)ξ(cid:107)(cid:107)η(cid:107). h,j It follows that (cid:107)X∗ for each k, let Xk = Vk(X∗ enough, X∗ k Xk − In(cid:107) < ε n2, so (cid:107)X∗ k Xk is invertible, so for such k, k Xk − In(cid:107) → 0 as k → ∞. Now, k Xk)1/2 be the polar decomposition. For k big V ∗ k Vk = (X∗ k Xk)−1/2X∗ k Xk(X∗ k Xk)−1/2 = In. 1 ηk n ··· (cid:3)(cid:124) ∈ H n and ηk =(cid:2)ηk (3.1) Asymptotically, {√ unital and so φ(n)(I ⊗ Ehj) = I ⊗ Ehj, Φ(A) = lim 1 , . . . , n ηk √ n ηk √ k→∞(cid:104)√ lim n ηk j , n ηk k→∞(cid:104)Aηk, ηk(cid:105). n} is orthonormal; indeed, using that φ is h(cid:105) = n lim k→∞(cid:104)(I ⊗ Ehj)ηk, ηk(cid:105) = nΦ(I ⊗ Ehj) = n(cid:104)φ(n)(I ⊗ Ehj) ξ, ξ(cid:105) = (cid:104)ξj, ξh(cid:105) = δj,h. 6 MARTÍN ARGERAMI k Xk)−1/2 − Hence Vk : Cn → H is an isometry. Also, (cid:107)Vk − Xk(cid:107) = (cid:107)Xk((X∗ In)(cid:107) → 0. And Vk → 0 weakly since the range of Vk is contained in the range of Xk, and {ηk} is orthonormal. We have, by (3.1), (cid:104)(φ(T ) − X∗ As the above convergence is in Mn(C), it also holds in norm. Thus k→∞ 0. k T Xk)ξh, ξj(cid:105) = n(cid:104)Φ(T ⊗ Ejh) ξ, ξ(cid:105) − n(cid:104)(T ⊗ Ejh)ηk, ηk(cid:105) −−−→ k→∞(cid:107)φ(T ) − X∗ k→∞(cid:107)φ(T ) − V ∗ k T Vk(cid:107) = lim k T Xk(cid:107) −−−→ k→∞ 0. lim (cid:3) Remark 3.3. Another set considered by Bunce -- Salinas [7] is the essential matricial range. This would be We n(T ) = {φ(T ) : φ : B(H) → Mn(C) ucp, φK(H) = 0}. By Lemma 3.2, We n(T ) ⊂ Ws n(T ). In analogy to the fact that the classical spatial numerical range is dense in the numerical range -- minus the fact that Ws n(T ) is often not convex -- we have the following result. The proof we provide does not follow the original argument. Proposition 3.4 (Bunce -- Salinas [7]). Let T ∈ B(H). Then (cid:88) Wn(T ) = {(cid:88) kXAk : X ∈ Ws A∗ n(T ), A1, A2, . . . ∈ Mn(C), kAk = In}. A∗ k k kA∗ k A∗ k = In. n(T ). In other words, the C∗-convex hull of the closure of the nth spatial matricial range of T equals the nth matricial range of T . When T is compact, the n(T ) is not needed. If C∗(T ) ∩ K(H) = {0}, then Wn(T ) = closure of Ws Ws n(T ) ⊂ Wn(T ). Now consider a ucp Proof. Fix n ∈ N. It is clear that Ws map φ : B(H) → Mn(C), and write φ = V ∗πV for a Stinespring dilation, where V : Cn → K is an isometry and π : B(H) → B(K) a representation. n(T ), We want to show that φ(T ) is of the form (cid:80) and(cid:80) kXAk with X ∈ Ws k A∗ Since π is bounded, its kernel is a closed two-sided ideal of B(H). Thus there are just two possibilities: either π is an isometry, or ker π = K(H). Assume first that π is an isometry. Then we can identify π(T ) with T ⊗ I: indeed, it is not hard to show that there exist a Hilbert space H0 and a unitary U : K → H ⊗ H0 such that π(T ) = U∗(T ⊗ I)U. Let {Fst} denote a set of matrix units for H0, corresponding to the canonical basis {fn} (note that H0 may be finite or infinite-dimensional). Then, with W : H⊗Cf1 → H the isometry W (ξ ⊗ λf1) = λξ, (cid:104)W ∗T W (ξ ⊗ λf1), η ⊗ µf1(cid:105) = λµ(cid:104)T ξ, η(cid:105) = (cid:104)(T ⊗ F11)(ξ ⊗ λf1), η ⊗ µf1(cid:105). For each s, let Hs ⊂ H be the subspace Hs = W (I ⊗ F1s)U V Cn. We obviously have dim Hs ≤ n. Let Rs : Cn → H be an isometry that contains (cid:32)(cid:88) s (cid:33) (cid:33) U V U V (cid:32)(cid:88) = V ∗U∗ (I ⊗ Fs1)(T ⊗ F11)(I ⊗ F1s) s V ∗U∗(I ⊗ F1s)∗W ∗T W (I ⊗ F1s)U V (cid:88) (cid:88) s = Hs in its range. So R∗ in its range. We have THE MATRICIAL RANGE OF E21 sRs = In, and RsR∗ s is a projection that contains Hs 7 φ(T ) = V ∗U∗(T ⊗ I)U V = V ∗U∗ T ⊗ Fss V ∗U∗(I ⊗ F1s)∗W ∗Rs (R∗ sT Rs) R∗ s = The convergence of the sum(cid:80) (cid:88) V ∗U∗(I ⊗ F1s)∗W ∗RsR∗ in Mn(C), so the convergence is in norm. Also, sW (I ⊗ F1j)U V = V ∗U∗ sW (I ⊗ F1s)U V. (cid:33) (cid:32)(cid:88) I ⊗ Fss U V s T ⊗ Fss is strong, but our last sum occurs s s As R∗ shown that φ(T ) ∈ {(cid:88) sT Rs ∈ Ws n(T ) for all s, and R∗ kXAk : X ∈ Ws A∗ sW (I ⊗ F1s)U V ∈ Mn(C), we have kAk = In} A∗ n(T ), A1, A2, . . . ∈ Mn(C), (cid:88) = V ∗U∗U V = In. k k n(T ). k T Vk → φ(T ). So φ(T ) ∈ Ws (note the lack of closure of the spatial matricial range). In the case where π = 0 on K(H), we have the same property for φ, and we may apply Lemma 3.2 to the separable C∗-algebra C∗(T ); that way, we obtain isometries Vk : Cn → H with V ∗ When T is compact, this last case does not apply, and so the closure of Ws n(T ) is not needed. When C∗(T ) ∩ K(H) = {0}, the quotient map ρ : B(H) → B(H)/K(H) is isometric on C∗(T ). Thus the map φ ◦ ρ−1 : ρ(C∗(T )) → Mn(C) is ucp and maps ρ(T ) to φ(T ). By Arveson's Extension Theorem, there exists φ : B(H)/K(H) → Mn(C), ucp, that extends φ ◦ ρ−1. Now φ ◦ ρ : B(H) → Mn(C) is a ucp map that annihilates K(H) and such that φ(ρ(T )) = φ(T ). By Lemma 3.2, φ(T ) ∈ Ws (cid:3) The matricial range was initially defined and studied by Arveson [3]. It is straightforward to check that each Wn(T ) is compact and C∗-convex (the latter in the sense of 2b in Theorem 3.5), and that Wm(X) ⊂ Wm(T ) for all X ∈ Wn(T ). He mentions, after the definition, that "it is not hard" to see that the aforementioned properties characterize the matricial range. The proof we know and write below (Theorem 3.5) is not "very hard", but it is not trivial either since it depends on Proposition 3.4, that itself depends n(T ). 8 MARTÍN ARGERAMI on Glimm's Lemma. Besides Arveson's characterization -- (2) below -- we in- clude a slightly more explicit characterization in terms of finite C∗-convex combinations. c > 0. Then the following statements are equivalent: Theorem 3.5. Let {Xn : n ∈ N} be a sequence of sets Xn ⊂ Mn(C), and (1) there exists a Hilbert space H and T ∈ B(H) such that (cid:107)T(cid:107) ≤ c and Xn = Wn(T ) for each n ∈ N; (2) the sequence {Xn} satisfies the following properties: (a) for each n, the set Xn is compact, and contained in the ball of (b) for each n, if X1, X2, . . . ⊂ Xn and A1, A2, . . . ∈ Mn(C) with radius c; (cid:80) k A∗ kAk = In, then(cid:80) k A∗ kXkAk ∈ Xn; (c) for each m, n ∈ N, Wm(Xn) ⊂ Xm. (3) the sequence {Xn} satisfies the following properties: (a) for each n, the set Xn is compact, and contained in the ball of (b) for each n, m, if X1, X2, . . . , Xr ⊂ Xn and A1, A2, . . . , Ar ∈ radius c; Mn×m(C) with(cid:80) kAk = Im, then(cid:80) k A∗ k A∗ kXkAk ∈ Xm. k A∗ kφn(·)Ak is ucp, and so(cid:80) Proof. (1) =⇒ (3) If Xn = Wn(T ), as (cid:107)φ(T )(cid:107) ≤ (cid:107)T(cid:107) ≤ c for any ucp map φ, it follows that Xn is contained in the ball of radius c. Pointwise-norm limits of ucp maps are ucp; so Xn is the image of the BW-compact set {φ : B(H) → Mn(C), ucp} under the (continuous) evaluation map φ (cid:55)−→ φ(T ), and thus compact. If we have a sequence X1, X2, . . . , Xr ⊂ Wn(T ), there exist ucp kAk = Im, the map φ :=(cid:80) (cid:80) maps φk with Xk = φk(T ). For a sequence A1, A2, . . . , Ar ⊂ Mm×n(C) with k A∗ kXkAk = φ(T ) ∈ Wm(T ) = Xm. (cid:80) (3) =⇒ (2) Fix n, X1, X2, . . . ⊂ Xn and matrices A1, A2, . . . ∈ Mn(C) with that (cid:80) k A∗ Choose (cid:96)0 such that (cid:107)I −(cid:80)(cid:96) kAk = In. If only finitely many Ak are nonzero, then we get directly kXkAk ∈ Xn. So assume that infinitely many Ak are nonzero. k A∗ kAk is invertible. By (3b),(cid:80)(cid:96) R(cid:96) =(cid:80)(cid:96) kAk(cid:107) < 1 for all (cid:96) > (cid:96)0. Then, for (cid:96) > (cid:96)0, −1/2 ∈ Xn k=1 A∗ is closed, (cid:80)∞ k=1(AkR (cid:96) −1/2 (cid:96) → In and so, since Xn for all (cid:96) > (cid:96)0. As R(cid:96) → In, we also have R kXkAk ∈ Xn. To check that Wm(Xn) ⊂ Xm, let Y ∈ Y = ψ(X). By the Kraus Decomposition, we may write ψ =(cid:80)r Wm(Xn); so there exist X ∈ Xn and ψ : Mn(C) → Mm(C), ucp, with with A1, . . . , Ar ∈ Mm×n(C). Then Y =(cid:80)r k · Ak, (2) =⇒ (1) For each n ∈ N, let {Qn,k}k∈N be a countable dense subset of k=1 A∗ kXAk ∈ Xm by (3b). Xn. Let H = (cid:96)2(N), where we index the canonical orthonormal basis as )∗XkAkR k=1 A∗ k=1 A∗ k=1 A∗ −1/2 (cid:96) k A∗ {ξn,k,j : n, k ∈ N, j = 1, . . . , n}. For each n, k ∈ N we denote the canonical basis in Cn by δ1, . . . , δn, and we define linear isometries Wn,k : Cn → H by THE MATRICIAL RANGE OF E21 9 Wn,kδj = ξn,k,j, j = 1, . . . , n. (cid:104)W ∗ n,kξm,(cid:96),h, δj(cid:105) = (cid:104)ξm,(cid:96),h, Wn,kδj(cid:105) = (cid:104)ξm,(cid:96),h, ξn,k,j(cid:105) = δm,n δ(cid:96),k δh,j = δm,n δ(cid:96),k (cid:104)δh, δj(cid:105), We have So Wn,kW ∗ It follows that Wn,kW ∗ vectors ξn,k,1, . . . , ξn,k,n. Also, n,kξm,(cid:96),h = δm,n δ(cid:96),k Wn,kδh = δm,n δ(cid:96),k ξn,k,h. n,k is the orthogonal projection onto the span of the so W ∗ n,kWn,k = In and W ∗ n,kWn(cid:48),k(cid:48)δj = W ∗ W ∗ (cid:88) (cid:88) n,kWn(cid:48),k(cid:48) = 0 if n (cid:54)= n(cid:48) or k (cid:54)= k(cid:48). Now define n,kξn(cid:48),k(cid:48),j = δn,n(cid:48)δk,k(cid:48)δj, Wn,kQn,kW ∗ n,k. T = n∈N k∈N n∈N (cid:88) (cid:88) (cid:32)(cid:88) k∈N The series is well-defined -- via strong convergence -- because the projections n,k} add to the identity of H. It is clear that (cid:107)T(cid:107) ≤ max{(cid:107)Qn,k(cid:107) : {Wn,kW ∗ n, k} < c. For any isometry V : Cm → H, we have n,k V ∗T V = V ∗Wn,kQn,kW ∗ n,kV. n∈N k∈N (cid:33) j A∗ (cid:88) Wn,kW ∗ n,kV = V ∗ j XjAj ∈ Xm. Claim: If Xj ∈ Xn(j), Aj ∈ Mn(j)×m(C), j ∈ N, then(cid:80) As(cid:88) (cid:88) V = V ∗IV = V ∗V = Im V ∗Wn,kW ∗ n∈N k∈N n,kV ∈ Mn×m(C) for each n, the Claim above implies that V ∗T V ∈ and W ∗ Xm. It follows that Ws n(T ) ⊂ Xn. Then the C∗-convexity (2b), compactness, and Proposition 3.4 give us Wn(T ) ⊂ Xn. From Qn,k = W ∗ n,kT Wn,k we obtain Qn,k ∈ Wn(T ) for all k ∈ N, since the map X (cid:55)−→ W ∗ n,kXWn,k is ucp B(H) → Mn(C); the density of {Qn,k} then shows that Xn ⊂ Wn(T ). Thus Wn(T ) = Xn for all n ∈ N. It remains to prove the Claim. We will prove the statement for a finite number of matrices, say 1 ≤ j ≤ r; if we prove that, then an argument with an R(cid:96) like in the proof of (3) =⇒ (2) shows the general result. So fix j ∈ N, with 1 ≤ j ≤ r. Consider the ucp map ψj : Mn(j)(C) → Mn(1)+···+n(r)(C) given by ψj1(X) ψj(X) =  ψj2(X) ... ψjr(X) 10 MARTÍN ARGERAMI where each ψjk : Mn(j)(C) → Mn(k)(C) is some ucp map, and ψjj is the identity, so ψjj(Xj) = Xj. By construction ψj is ucp, so (2c) implies that ψj(Xj) ∈ Xn(1)+···+n(r) for each j = 1, . . . , r. This implies, by (2b), that X1  = X2 ... Xr Now, by (2c), r(cid:88) A∗ j XjAj = j=1 since conjugation by(cid:2)A1  0  0 In(r) 0 0 0 0 ...  ψ1(X1) In(1)  ψr(Xr) 0 A1   ∈ Xm, In(r) ... ... ... ... + ··· In(1) 0  ∗X1 (cid:3)∗ is ucp. A2 ... Ar ··· Ar A1 ∈ Xn(1)+···+n(r). X2 + A2 ... Ar Xr (cid:3) In the conditions of Theorem 3.5, the radius of Wn(T ) is actually (cid:107)T(cid:107) for 2 (cid:107)T(cid:107) (as is all n ≥ 2. For n = 1, it is well-known that the radius could be 1 the case when T = E21). Concretely, define νn(T ) = sup{(cid:107)X(cid:107) : X ∈ Wn(T )}. Proposition 3.6 (Smith -- Ward [17]). Let T ∈ B(H). Then νn(T ) = (cid:107)T(cid:107) for all n ≥ 2. Proof. For any ucp map φ : B(H) → Mn(C), we have (cid:107)φ(T )(cid:107) ≤ (cid:107)T(cid:107), so νn(T ) ≤ (cid:107)T(cid:107). Conversely, fix ε > 0. Choose ξ ∈ H such that (cid:107)ξ(cid:107) = 1 and (cid:107)T ξ(cid:107) > (cid:107)T(cid:107) − ε. Let H0 be an n-dimensional subspace of H that contains ξ and T ξ. We define ψ : B(H) → Mn(C) by ψ(L) = PH0LPH0 (with the usual identification PH0B(H)PH0 (cid:39) Mn(C)). The map ψ is ucp, and νn(T ) ≥ (cid:107)ψ(T )(cid:107) = (cid:107)PH0T PH0(cid:107) ≥ (cid:107)T ξ(cid:107) > (cid:107)T(cid:107) − ε. As ε was arbitrary, we get νn(T ) = (cid:107)T(cid:107). (cid:3) The family of examples where W(T ) can be found explicitly is fairly small. The most notable example is W(E21), as will be established independently in Corollaries 5.6 and 7.2. A small generalization, to quadratic operators, THE MATRICIAL RANGE OF E21 11 . Hj = In (cid:88) j Wn(T ) = j=1 is considered in [19]. As could be expected, though, the case of normal operators is not hard. λjHj : k ∈ N, Hj ≥ 0, λj ∈ σ(T ), Proposition 3.7. Let T ∈ B(H) be normal, and n ∈ N. Then Proof. Since T is normal, we can identify C∗(T ) with C(σ(T )). Given any j=1 λjHj as above, we can find positive linear functionals fj j=1 fj(X)Hj is a unital positive linear map C∗(T ) → Mn(C). As the domain is abelian, ψ is completely positive [15, Theorems 3.9 and 3.11]. This shows the inclusion ⊃ above. Conversely, let φ : B(H) → Mn(C) be ucp. Fix ε > 0. By the j Pj =  k(cid:88) decomposition(cid:80)k (characters, actually) with fj(T ) = λj. The map ψ : X (cid:55)−→(cid:80)k Spectral Theorem, we can find projections P1, . . . , Pk ∈ B(H) with(cid:80) I and λ1 . . . , λk ∈ σ(T ) with (cid:107)T −(cid:80) (cid:107)φ(T ) −(cid:88) matrices of the form(cid:80) j λjHj as above. As we can do this for each ε > 0, we have shown that φ(T ) is a limit of (cid:3) We continue with another of Arveson's gems from [3]. Given T ∈ B(H), S ∈ B(K), we say that S is a compression of T if there exists a projection P ∈ B(H) such that S is unitarily equivalent to P TP H ∈ B(P H). This definition agrees with the usual use of the word "compression" or "corner", but it is important to emphasize the restriction aspect: for instance, in M2(C) the projection E11 is not a compression of I2 in the above sense; or, for another example, the matrix unit E11 ∈ M2(C) is a compression of E11 ∈ M3(C), but not viceversa. The important thing to note is that the unitary implementing the unitary equivalence maps K onto P H. j λjPj(cid:107) < ε. Then λjφ(Pj)(cid:107) = (cid:107)φ(T −(cid:88) λjPj)(cid:107) < ε. j j Theorem 3.8 (Arveson). Let T ∈ B(H), S ∈ B(K). The following state- ments are equivalent: (1) Wn(S) ⊂ Wn(T ) for all n ∈ N; (2) for each n ∈ N and A, B ∈ Mn(C), (3.2) (cid:107)A ⊗ I + B ⊗ S(cid:107) ≤ (cid:107)A ⊗ I + B ⊗ T(cid:107); (3) for each finite-dimensional projection P ∈ B(K), there exists a ∗- representation π : C∗(T ) → B(Hπ) such that P SP K ∈ B(P K) is unitarily equivalent to a compression of π(T ); (4) there exists a ∗-representation π : C∗(T ) → B(Hπ) such that S is a compression of π(T ). Moreover, 12 MARTÍN ARGERAMI (5) When S is normal, the above conditions are equivalent to σ(S) ⊂ W1(T ); (6) when T is compact and irreducible, the above conditions are equiva- lent to S being unitarily equivalent to a compression of T ⊗ I. Proof. (1) =⇒ (2) By repeating the argument in the first paragraph of the proof of Proposition 3.1, we can get a ucp map φ : OS(T ) → B(K) with φ(T ) = S. Then, for any A, B ∈ Mn(C), (cid:107)A ⊗ IK + B ⊗ S(cid:107) = (cid:107)A ⊗ φ(IH ) + B ⊗ φ(T )(cid:107) = (cid:107)φ(n)(A ⊗ IH + B ⊗ T )(cid:107) ≤ (cid:107)A ⊗ IH + B ⊗ T(cid:107). (2) =⇒ (3) Define a linear map φ : span{IH , T} → span{IK, S} by φ(IH ) = Ik, φ(T ) = S. The condition (3.2) now says that φ is completely contractive. By [15, Proposition 3.5], φ extends to a ucp map φ : span{IH , T, T ∗} → B(K), and by Arveson's Extension Theorem we may enlarge the domain of φ to be B(H). Consider a Stinespring Dilation φ = PKπK, where π : B(H) → B(Kπ) is a representation and K ⊂ Kπ. For any projection P ∈ B(K), since P K ⊂ K = PKKπ, P SP K = P PKSP K = P PKπ(T )P K = P π(T )P K. (3) =⇒ (4) Let {Pj} ⊂ B(K) be an increasing net of finite-rank projec- tions that converges strongly to I. By hypothesis, for each j there exist a unitary Vj : PjK → QjHj, a projection Qj ∈ B(Hj), and a representation πj : C∗(T ) → B(Hj) such that PjSP jK = V ∗ j Qjπj(T )Qj Hj Vj. Fix a state f of C∗(T ). Then the maps ψj : C∗(T ) → B(K) given by ψj(X) = V ∗ j Qjπj(X)Qj Hj Vj + f (X) (I − Pj) are ucp. Let ψ : C∗(T ) → B(K) be a BW-cluster point of the net {ψj}. It is clear that ψ(T ) = S. Now a Stinespring decomposition of ψ gives S as a compression of T . (4) =⇒ (1) By hypothesis, there is a ucp map ψ with S = ψ(T ). Given any X ∈ Wn(S), there exists a ucp map φ with φ(S) = X. Then X = φ(S) = φ ◦ ψ(T ) ∈ Wn(T ). (5) Assume S is normal. If σ(S) ⊂ W1(T ), then W1(S) ⊂ W1(T ) since W1(S) is the closed convex hull of σ(S). This allows us to show that the unital map ψ : OS(T ) → OS(S) with ψ(T ) = S, ψ(T ∗) = S∗ is positive: indeed, if αI + βT + γT ∗ ≥ 0 then W1(αI + βT + γT ∗) ⊂ [0,∞) and, as σ(αI + βS + γS∗) = {α + βλ + γ¯λ : λ ∈ σ(S)} ⊂ {α + βλ + γ¯λ : λ ∈ W1(T )} = {f (αI + βT + γT ∗) : f state} ⊂ [0,∞) THE MATRICIAL RANGE OF E21 13 we obtain that αI+βS+γS∗ ≥ 0 (here we use that S is normal to characterize positivity by its spectrum, and also for the first equality above). So ψ is a unital positive map; as C∗(S) is abelian, ψ is ucp. Now we can use this ψ to check that (2) or (4) hold. Conversely, if the equivalent conditions hold, we have W1(S) ⊂ W1(T ), and so σ(S) ⊂ W1(S) ⊂ W1(T ). (6)When T is compact and (4) holds, the representation π is nonzero on T , and so it has to be isometric on C∗(T ) -- as the only possible kernel is K(H). Because T is irreducible, K(H) ⊂ C∗(T ) (this is a more or less straightforward consequence of Kadison's Transitivity Theorem; see [8, The- It is well-known that in this orem I.10.4]). So π is isometric on B(H). situation π(T ) is unitarily equivalent to T ⊗ I. Conversely, if S is unitarily equivalent to a compression of T ⊗ I, we can recover (4). (cid:3) It was proven by Hamana [13] (see [9] for a bit of history and a proof within Arveson's framework) that every operator system admits a C∗-envelope. That is, given an operator system S, there exists a C∗-algebra A and a complete isometry j : S → A such that for any C∗-algebra B and any com- plete isometry ψ : S → B, there exists a C∗-epimorphism π : C∗(ψ(S)) → A with π ◦ ψ = j. That is, the following diagram commutes: (3.3) S ψ j C∗(ψ(S)) π epimorphism A It is straightforward to prove that for a given operator system S the C∗- algebra A above is determined up to isomorphism, and so one denotes it e (S) and names it the C∗-envelope of S. By taking the quotient by by C∗ the kernel of π (the Šilov ideal, in Arveson's terminology), we always have e (T ) is a quotient of C∗(T ). In the particular case where an opera- that C∗ tor system OS(T ) ⊂ B(H) has the property that π has trivial kernel (so, it is isometric), we say that T is first order (this was Arveson's original terminology; in later years he used the word reduced). Let us now draw some consequences from Arveson's result. Corollary 3.9. Let T ∈ B(H), S ∈ B(K). The following statements are equivalent: (1) W(S) = W(T ); (2) for all n ∈ N, for any A, B ∈ Mn(C), (cid:107)A ⊗ I + B ⊗ T(cid:107) = (cid:107)A ⊗ I + B ⊗ S(cid:107); (3) OS(S) (cid:39) OS(T ) via a complete isometry φ with φ(S) = T . If both S, T are irreducible, first order, and both C∗(S) and C∗(T ) contain a nonzero compact operator, then the above statements are also equivalent to (4) S and T are unitarily equivalent. / / ( ( MARTÍN ARGERAMI 14 Proof. The equivalences (1) ⇐⇒ (2) ⇐⇒ (3) follow directly from Theo- rem 3.8. The implication (4) =⇒ (3) is trivial. So assume that S, T are irreducible, first order, that their C∗-algebras contain nonzero compact op- erators, and that there exists a complete isometry φ : C∗(T ) → C∗(S) with φ(T ) = S. Considering the diagram (3.3) for S = OS(S) and ψ = φ−1, and for S = OS(T ) and ψ = φ respectively, one deduces that C∗ e (S) via an isomorphism π with π(T ) = S. As T is irreducible and C∗(T ) contains a compact operator, it follows that K(H) ⊂ C∗(T ) (as mentioned above, see [8, Corollary I.10.4]). Similarly, C∗(S) contains all compacts of B(K). Recall that we are assuming that both T and S are first order, so C∗(T ) = C∗ e (T ) and C∗(S) = C∗ Because π is an irreducible representation of C∗(T ) and J = K(H) ⊂ C∗(T ) with πJ (cid:54)= 0, we have that πJ is irreducible. Indeed, let ξ ∈ K, and consider the subspace π(J)ξ ⊂ K. Since J is an ideal, π(J)ξ is invariant for π(C∗(T )) = C∗(S); as C∗(S) is irreducible, it follows that π(J)ξ is K or 0. If it were 0, we would have ξ ∈ [π(J)K]⊥. But then π(J)K (cid:40) K and it is invariant for C∗(S) -- which is irreducible -- so π(J) = 0, a contradiction. Thus π(J)ξ = K for all ξ, and so π(J) has no reducing subspaces. e (T ) (cid:39) C∗ e (S). Because both the domain and the range of π contain their respective com- pact operators, π necessarily maps rank-one projections to rank-one pro- jections. We will show that this implies that π is implemented by unitary conjugation. Fix an orthonormal basis {ξj} of H. Choose a unit vector η1 ∈ π(ξ1 ⊗ ξ1)H, and define ηj = π(ξj ⊗ ξ1)η1. One then checks easily that {ηj} is orthonormal; and it has to be a basis, because a rank-one projection corresponding to a vector orthogonal to {ηj} would get brought back by π to a rank-one projection with range orthogonal to {ξj}, an impossibility. Now define a unitary U : H → K by U ξj = ηj. Then U∗π(ξj ⊗ ξk)U ξ(cid:96) = U∗π(ξj ⊗ ξk)η(cid:96) = U∗π(ξj ⊗ ξk)π(ξ(cid:96) ⊗ ξ1)η1 = δ(cid:96),k U∗π(ξj ⊗ ξ1)η1 = δ(cid:96),k U∗ηj = δ(cid:96),k ξj = (ξj ⊗ ξk)ξ(cid:96). Thus U∗πU is the identity on all rank-one operators; by linearity and con- tinuity, it is the identity on all of K(H). For an arbitrary X ∈ C∗(T ) and ξ ∈ H, let P be the rank-one projection with P ξ = ξ. Then U∗π(X)U ξ = U∗π(X)U P ξ = U∗π(X)U U∗π(P )U ξ = U∗π(XP )U ξ = XP ξ = Xξ. So U∗π(X)U = X, that is π(X) = U XU∗ for all X ∈ C∗(T ). In particular, (cid:3) S = π(T ) = U T U∗. THE MATRICIAL RANGE OF E21 15 Remark 3.10. The requirement in (3) above that φ(S) = T cannot be relaxed. For example, with (cid:20)1 0 (cid:21) 0 0 (cid:20)1 0 (cid:21) 0 2 , S = , T = we have OS(S) = OS(T ), but W1(S) = [0, 1] and W1(T ) = [1, 2]. Also, we refer to Corollary 3.15 for examples of operators with the same matricial range but very far from unitarily equivalent. Remark 3.11. The conditions in Corollary 3.9 do not imply the equality of the spatial matricial ranges Ws(S) and Ws(T ). The equality Ws(S) = Ws(T ) implies W(S) = W(T ) by Proposition 3.4, but the converse is not true. For instance consider H = K = (cid:96)2(N), take S to be the unilateral shift with respect to the canonical basis {ξk}, and let T be the unitary given by T ξk = γk ξk, where {γk} is a dense sequence in T. By Corollary 3.15 below, we have W(S) = W(T ). Any isometry V : C2 → H is given by V e1 = x, V e2 = y, where {x, y} ⊂ H is orthonormal; we will write Vx,y for such an isometry. It is not hard to check that, for any R ∈ B(H), (cid:20)(cid:104)x, Rx(cid:105) (cid:104)x, Ry(cid:105) (cid:21) . (cid:104)y, Rx(cid:105) (cid:104)y, Ry(cid:105) V ∗ x,yRVx,y = (cid:20)1 0 (cid:21) By choosing the sequence {γk} with γ1 = γ2 = 1 and taking x = ξ1, y = ξ2, we get I2 = 2(S), we have I2 (cid:54)∈ Ws x,yT Vx,y ∈ Ws 2(S). But, while I2 ∈ Ws x,ySVx,y for orthonormal x, y ∈ H, then Vx,yV ∗ V ∗ projection onto the span of {x, y}. So = V ∗ 0 1 2(T ). Indeed, if we had I2 = x,y = P , the orthogonal P = Vx,yV ∗ x,y = Vx,yI2V ∗ x,y = Vx,yV ∗ x,ySVx,yV ∗ x,y = P SP. This equality cannot hold, because we would have P S∗SP = P = P S∗P SP, which implies that P S∗(I − P )SP = 0, and so (I − P )SP = 0, from where SP = P SP = P . This would make x and y eigenvectors for S, a contradic- tion. It is easy to see, on the other hand, that I2 ∈ Ws 2(S), so it is not clear at first sight whether Ws n(T ) (cid:54)= Ws n(S) or not. Let us now specialize the above result to the case of matrices. We mention [11] for a different and detailed proof of the equivalence (3) ⇐⇒ (5) in Corollary 3.12 below. Corollary 3.12. Let n ∈ N and S, T ∈ Mn(C). The following statements are equivalent: (1) W(S) = W(T ); (2) Wn(S) = Wn(T ); 16 MARTÍN ARGERAMI (3) for all A, B ∈ Mn(C), (cid:107)A ⊗ I + B ⊗ T(cid:107) = (cid:107)A ⊗ I + B ⊗ S(cid:107); (4) OS(S) (cid:39) OS(T ) via a complete isometry φ with φ(T ) = S. If both S, T are irreducible, the above statements are also equivalent to (5) S and T are unitarily equivalent. Proof. The equivalences (1) ⇐⇒ (3) ⇐⇒ (4) ⇐⇒ (5) follow directly from Corollary 3.9 (note that Mn(C) is simple, so the irreducibility of S and T imply that they are first order). The implication (1) =⇒ (2) is trivial, so all that remains is to prove (2) =⇒ (1). Assume that Wn(S) = Wn(T ), and let X ∈ Wm(S). So there exists a ucp map ϕ : Mn(C) → Mm(C) with ϕ(S) = X. As S ∈ Wn(S) = Wn(T ), there exists a ucp map ψ : Mn(C) → Mn(C) with ψ(T ) = S. Then X = φ(S) = φ(ψ(T )) ∈ Wm(T ). It follows that Wm(S) ⊂ Wm(T ), and by reversing the roles of S and T we (cid:3) get equality. Remark 3.13. The reason one requires irreducibility for unitary equivalence in Corollary 3.12 is multiplicity: for an easy example, we can take S = E11, T = E11 + E22 in M3(C) and then W(S) = W(T ) but they are obviously not unitarily equivalent. We will later use some sophisticated ideas -- mainly by Arveson and by Ando both building on ideas related to dilations -- to calculate the matricial range of the 2 × 2 unilateral shift (Corollaries 5.6 and 7.2). The matricial range of the n × n unilateral shift is unknown for n ≥ 3, but the infinite- dimensional unilateral shift (and, a posteriori, proper isometries) can be tackled with a rather direct approach. We are grateful to D. Farenick for a simplification of our original argument. Proposition 3.14. Let T ∈ B(K) with (cid:107)T(cid:107) ≤ 1, and S ∈ B(H) the unilateral shift. Then there exists a ucp map ψ : B(H) → B(K) with ψ(S) = T . Proof. Since T is a contraction, we can construct a unitary U0 = T (I − T ∗T )1/2 (I − T T ∗)1/2 −T ∗ ∈ M2(B(K)). If U is a universal unitary (that is, σ(U ) = T), then C∗(U ) = C(T). As σ(U0) is a compact subset of T, there is a ∗-epimorphism (onto by Tietze's Extension Theorem) acting by restriction: π : C∗(U ) = C(T) → C(σ(U0)) = C∗(U0), with π(U ) = U0. Let ρ : B(H) → B(H)/K(H) be the quotient map. As S becomes a universal unitary in the Calkin algebra, C∗(ρ(S)) (cid:39) C(T) (cid:39) C∗(U ). Let γ : C∗(ρ(S)) → C∗(U ) be a ∗-isomorphism with γ(ρ(S)) = U. Then π ◦ γ : C∗(ρ(S)) → C∗(U0) is a ∗-epimorphism with π ◦ γ(ρ(S)) = U0; (cid:20) (cid:21) THE MATRICIAL RANGE OF E21 17 in particular, ucp. Let φ : M2(B(K)) → B(K) be the compression to the 1, 1 entry. Then ψ = φ ◦ π ◦ ρ : B(H) → B(K) is a ucp map with (cid:3) Corollary 3.15. Let S ∈ B(H) be a proper isometry, or a unitary with full spectrum T, and let T ∈ B(K). Then the following statements are equivalent: φ ◦ π ◦ ρ(S) = T. (1) there exists a ucp map φ : B(H) → B(K) such that φ(S) = T ; (2) (cid:107)T(cid:107) ≤ 1. In other words, W∞(S) = {T : (cid:107)T(cid:107) ≤ 1}, and so for all n ∈ N, Wn(S) = {A ∈ Mn(C) : (cid:107)A(cid:107) ≤ 1}. Proof. (1) =⇒ (2) Since (cid:107)S(cid:107) = 1 and φ is ucp, we get that (cid:107)T(cid:107) = (cid:107)φ(S)(cid:107) ≤ 1. (2) =⇒ (1) Let S0 be the unilateral shift. By Proposition 3.14 there exists a ucp map ψ with ψ(S0) = T . If S is a proper isometry, by the Wold Decomposition there exist unitaries U and W such that S = W ∗(U ⊕ (cid:76) j S0)W . Then S (cid:55)−→ W SW ∗ = U ⊕(cid:77) S0 (cid:55)−→ S0 ψ−→ T j is a ucp map. If S is a unitary with full spectrum, we can proceed as in the (cid:3) proof of Proposition 3.14 to get a ucp map φ with φ(S) = T . 4. Unitary Dilations, Numerical Radius, and Matricial Range Advances in operator theory -- concretely, about dilations -- gave between in late 1960s and the early 1970s several striking characterizations of the nu- merical radius and, as a byproduct, a characterization of the matricial range of the 2 × 2 unilateral shift. We will visit, in Sections 5 and 7 respectively, Ando's and Arveson's techniques built upon these theories. Given a group G, a function T : G → B(H) is said to be positive-definite for all functions ξ : G → H of finite support. It is not hard to see that (4.1) implies that T (s−1) = T (s)∗ for all s ∈ G. A particular case of a positive-definite function is given by a unitary rep- resentation. That is, a function U : G → B(H) such that U (e) = I, U (s) is a unitary for all s ∈ G, and U (st) = U (s)U (t) for all s, t ∈ G. It turns out that one can do a kind of GNS representation for a positive-definite function, and so all positive-definite functions arise from unitary representations. We will only need the particular case where G = Z. Theorem 4.1 (Sz.Nagy -- Foias,). Let H be a Hilbert space, and T : Z → B(H) with T (0) = I. The following statements are equivalent: if (4.1) (cid:88) (cid:88) s∈G t∈G (cid:104)T (t−1s)ξ(s), ξ(t)(cid:105) ≥ 0 18 MARTÍN ARGERAMI (1) There exists a Hilbert space K ⊃ H and a unitary U ∈ B(K) such that T (n) = PH U nH for all N ∈ Z. (2) T is positive-definite. (cid:3) Proof. This is [18, Theorem 7.1]. Remark 4.2. In the case where H = C2 and T = T (1) = E21, T (n) = 0 for all n ≥ 2, the unitary U can be obtained explicitly as the bilateral shift. Concretely, we take K = (cid:96)2(Z), and define U on the canonical basis by U ek = ek+1, extended by linearity and continuity (since U is isometric). If we identify H = C2 with span{e0, e1}, then PH U nH = En 12 for all n ∈ N (which simply means that PH UH = E12, PH U 2H = 0). Most considerations of the numerical radius will use the following elemen- tary characterization: are equivalent: Proposition 4.3. Let T ∈ B(H) with (cid:107)T(cid:107) ≤ 1. The following statements (1) w(T ) ≤ 1; (2) for all λ ∈ T, I + Re λT ≥ 0; (3) for all λ ∈ T, Re λT ≤ I; (4) for all z ∈ D, Re zT ≤ I Proof. (1) =⇒ (2) If w(T ) ≤ 1, then for any λ ∈ T and ξ ∈ H with (cid:107)ξ(cid:107) = 1, (cid:104)− Re λT ξ, ξ(cid:105) = Re(cid:104)(−λ)T ξ, ξ(cid:105) ≤ (cid:104)(−λ)T ξ, ξ(cid:105) = (cid:104)T ξ, ξ(cid:105) ≤ 1 = (cid:104)ξ, ξ(cid:105). Thus (cid:104)(I + Re λT )ξ, ξ(cid:105) ≥ 0. (2) =⇒ (3) This is a direct consequence of the fact that T = −T. (3) =⇒ (4) If z ∈ D, then z = rλ with 0 ≤ r < 1 and λ ∈ T. So Re zT = r Re λT ≤ rI ≤ I. (4) =⇒ (1) Given ξ ∈ H with (cid:107)ξ(cid:107) = 1, let λ ∈ T such that (cid:104)T ξ, ξ(cid:105) = λ(cid:104)T ξ, ξ(cid:105). Then (cid:104)T ξ, ξ(cid:105) = Re(cid:104)T ξ, ξ(cid:105) = Re λ(cid:104)T ξ, ξ(cid:105) Thus w(T ) ≤ 1. = (cid:104)Re λT ξ, ξ(cid:105) = lim r→1 (cid:104)Re rλT ξ, ξ(cid:105) ≤ (cid:104)ξ, ξ(cid:105) = 1. (cid:3) We state and prove a version of a characterization of unitary dilations, due to Sz.-Nagy and Foias, [18, Theorem 11.1]. In their terminology, we only consider 2-dilations. Theorem 4.4 (Sz.Nagy-Foias,). Let T ∈ B(H). The following statements are equivalent: 19 (1) there exists a Hilbert space K ⊃ H, and a unitary U ∈ B(K) such THE MATRICIAL RANGE OF E21 that (4.2) (2) w(T ) ≤ 1. T n = 2 PH U nH for all n ∈ N; Proof. (1) =⇒ (2) Fix z ∈ D. Since z < 1, the series below converges and we can manipulate it as follows: I + 2 zkU k = −I + 2 zkU k = −I + 2(I − zU )−1 = (I + zU )(I − zU )−1. ∞(cid:88) ∞(cid:88) k=1 Then (4.3) k=0 (IH − zT )−1 = IH + ∞(cid:88) k=1 zkT k = PH = PH (I + zU )(I − zU )−1 . ∞(cid:88) k=1 zkU k (cid:33)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)H (cid:32) I + 2 (cid:12)(cid:12)(cid:12)(cid:12)H For any ξ ∈ K, Re(cid:104)(I + zU )ξ, (I − zU )ξ(cid:105) = (1 − z2)(cid:107)ξ(cid:107)2 ≥ 0. In particular, with ξ = (I − zU )−1η, we obtain Re(cid:104)(I + zU )(I − zU )−1η, η(cid:105) ≥ 0, (4.4) When ξ ∈ H, we get from (4.3) and (4.4) that η ∈ K, z ∈ D. (cid:104)Re(IH − zT )−1ξ, ξ(cid:105) = Re(cid:104)(IH − zT )−1ξ, ξ(cid:105) = Re(cid:104)PH (I + zU )(I − zU )−1ξ, ξ(cid:105) = Re(cid:104)(I + zU )(I − zU )−1ξ, ξ(cid:105) ≥ 0. Replacing ξ with (IH − zT )ξ, the above becomes (cid:104)ξ, Re(IH − zT )ξ(cid:105) ≥ 0, z ( 1 z ∈ D, and so Re(IH − zT ) ≥ 0. Now Proposition 4.3 gives w(T ) ≤ 1. (2) =⇒ (1) Since w(T ) ≤ 1, we have σ(T ) ⊂ W1(T ) ⊂ D. Then I − zT = z I − T ) is invertible for all z ∈ D (when z = 0, I − zT = I). For any (cid:107)znT n(cid:107)1/n → spr (zT ) = z spr (T ) ≤ z < 1. for all n ≥ n0. Thus the series(cid:80)∞ In particular, for ε < 1−z, there exists n0 such that (cid:107)znT n(cid:107) ≤ (z+ε)n < 1 k=0 znT n is norm convergent and equal to (I − zT )−1. Let ξ ∈ H, and put η = (I − zT )−1ξ. By Proposition 4.3, we have Re(I − zT ) ≥ 0 for all z ∈ D; thus 0 ≤ Re(cid:104)η, (I − zT )η(cid:105) = Re(cid:104)(I − zT )−1ξ, (I − zT )(I − zT )−1ξ(cid:105) 20 MARTÍN ARGERAMI = Re(cid:104)(I − zT )−1ξ, ξ(cid:105). Define, for 0 ≤ r < 1 and 0 ≤ t ≤ 2π, Q(r, t) = I + 1 2 rk(eiktT k + e−iktT ∗k). ∞(cid:88) k=1 ∞(cid:88) k=0 (cid:88) n∈Z ∞(cid:88) This converges in norm by the argument with the spectral radius we just used above. Also, with z = r eit, (cid:104)Q(r, t)ξ, ξ(cid:105) = (cid:107)ξ(cid:107)2 + 1 2 Re(cid:104) zkT kξ, ξ(cid:105) = (cid:107)ξ(cid:107)2 + 1 2 Re(cid:104)(I − zT )−1ξ, ξ(cid:105) ≥ 0. Given a sequence {ξn}n∈Z with finite support, let us form ξ(t) = e−intξn. 0 (cid:90) 2π (cid:88) (cid:88) m,n∈Z 1 2π n∈Z = = (cid:90) 2π 0 (cid:88) Then (recall that the sum over n has finitely many nonzero terms, so the exchange with the integral is not an issue; for the sum over k, the convergence is uniform and the exchange is again possible) 0 ≤ 1 2π (cid:104)Q(r, t)ξ(t), ξ(t)(cid:105) dt e−i(n−m)t(cid:104)ξn + 1 2 rkeiktT kξn + rke−iktT ∗kξn, ξm(cid:105) dt (cid:104)ξn, ξn(cid:105) + 1 2 k=1 rn−m(cid:104)T n−mξn, ξm(cid:105) + 1 2 n>m n<m (cid:88) rm−n(cid:104)T ∗(m−n)ξn, ξm(cid:105) (for the last equality, note that the integrals will be nonzero only when −n + m + k = 0 and −n + m − k = 0; this fixes k, and we also have the restriction that k ≥ 1, which makes the case n = m vanish). Now define a function T : Z → B(H) by T (0) = I, T (n) = 1 2 T n, T (−n) = 1 2 T ∗n, n ≥ 1. We can rewrite the inequality above as (cid:104)T (n − n)ξn, ξn(cid:105) + (cid:88) rn−m(cid:104)T (n − m)ξn, ξm(cid:105) 0 ≤(cid:88) (cid:88) n∈Z + n<m n>m rm−n(cid:104)T (n − m)ξn, ξm(cid:105). 0 ≤ (cid:88) n,m∈Z (cid:104)T (n − m)ξn, ξm(cid:105), Noting that all sums are finite by hypothesis, we may take r (cid:37) 1, and then THE MATRICIAL RANGE OF E21 21 which shows that the function n (cid:55)−→ T (n) is positive-definite. By Theo- rem 4.1, there exists a Hilbert space K ⊃ H and a unitary U ∈ B(K) such that T (n) = PH U nH for all n ∈ Z. In particular, T n = 2T (n) = 2 PH U nH , n ∈ N. (cid:3) 5. Ando's Characterizations of the Numerical Radius The following surprising characterization is [1, Lemma 1]. We do not think that "lemma" is a fair word to describe it. The proof follows closely that of Ando, but we have tried to make it a bit clearer. Theorem 5.1. Let T ∈ B(H) with w(T ) ≤ 1. Then there exists X ∈ B(H)+, contractive (i.e., 0 ≤ X ≤ I) such that for all ξ ∈ H : η ∈ H (cid:104)Xξ, ξ(cid:105) = inf (5.1) , (cid:26)(cid:28)(cid:20) I . (cid:21)(cid:20)ξ (cid:21) η 1 2 T ∗ 1 2 T X (cid:21)(cid:29) (cid:20)ξ (cid:20)I − Y η (cid:27) (cid:21) (cid:27) Such X satisfies (cid:26) Y ∈ B(H) : 0 ≤ Y ≤ I, X = max (5.2) Proof. We assume that w(T ) ≤ 1. By Theorem 4.4 there exist a Hilbert space K ⊃ H and a unitary U in B(K) such that T k = 2PH U kH, k ∈ N. We will define a sequence {Xn} ⊂ B(H) in the following way: we start with X0 = I, and put Xn = PH (I − Qn)PH, where Qn is the projection onto k=1 U∗kH. Since by construction these subspaces are increasing on n, span(cid:83)n ≥ 0 1 2 T . 1 2 T ∗ Y we have I = X0 ≥ X1 ≥ X2 ≥ ··· ≥ 0. So the sequence converges strongly to a positive operator X ∈ B(H) -- note that we could have defined X directly, but it provides no obvious benefit to the proof. For ξ ∈ H, we have (cid:104)U kξ, ξ(cid:105) = (cid:104)U kξ, PH ξ(cid:105) = (cid:104)PH U kH ξ, ξ(cid:105) = 1 2 (cid:104)T kξ, ξ(cid:105), k ∈ N. Define A ∈ Mn+1(B(H)) by Akj = 2 T ∗(j−k), 2 T (k−j), We can write, for ξ ∈ H, and using ξ0 = ξ, (cid:104)Xnξ, ξ(cid:105) = (cid:104)(I − Qn)ξ, ξ(cid:105) = (cid:107)(I − Qn)ξ(cid:107)2 = (cid:107)ξ − Qnξ(cid:107)2 = dist (ξ, QnH)2 1 j = k k < j k > j 1 I, (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)ξ + n(cid:88) k=1 = inf U∗kξk : ξ1, . . . , ξn ∈ H  k=0 k,j=0  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) n(cid:88)  n(cid:88)  n(cid:88)  n(cid:88)  n(cid:88) k=0 k=0 k,j=0 < = inf ξj∈H = inf ξj∈H = inf ξj∈H 22 = inf = inf MARTÍN ARGERAMI U∗kξk : ξ1, . . . , ξn ∈ H (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2  (cid:104)U∗(k−j)ξk, ξj(cid:105) : ξ1, . . . , ξn ∈ H (cid:88) (cid:88) j<k j<k (cid:104)ξk, ξk(cid:105) + (cid:104)U∗(k−j)ξk, ξj(cid:105) + (cid:104)U (j−k)ξk, ξj(cid:105) (cid:104)ξk, ξk(cid:105) + (cid:104) 1 2 T ∗(k−j)ξk, ξj(cid:105) + (cid:104) 1 2 T (j−k)ξk, ξj(cid:105)  2 T ∗n ... ... 2 T ∗ I 1 = inf (cid:104)Ajkξk, ξj(cid:105) : ξ1, . . . , ξn ∈ H 1 2 T ∗ ··· 1 2 T ∗(n−1) 1 I ... I 1 2 T ... ... 2 T n 1 1 We can write the matrix A as  I T ∗ ··· ··· T ∗n ... 0 I ... ... ... I T ∗ 0 0 ··· 0 I ... I 1 2 T 2 T ∗ 0 ··· ...   2 T n−1··· I − 1 − 1 2 T I 0 ... 0 ... ··· ... 0 ... 0 I − 1 2 T ∗ 0 − 1 I 2 T A =  (cid:88) (cid:88) k<j k<j    ξ ξ1 ... ... ξn ξ ξ1 ... ... ξn  , 0 ··· ··· 0 0 ··· 0 I ... > . I T ... ... ... I 0 T n ··· ··· T I     > : ξ1, . . . , ξn ∈ H ξ ξ1 ... ... ξn , ξ1, . . . , ξn ∈ H},  .  As the invertible triangular block matrix preserves the first entry of the n- tuple vector, and as the infimum is taken over all n-tuples in H, we obtain (5.3) (cid:104)Xnξ, ξ(cid:105) = inf < ... I 1 1 2 T I 0 ... 0 ··· 0 1 2 T ∗ 0 ··· 0 ... 0 2 T ∗ I 1 2 T I ... ...   ξ ξ1 ... ... ξn  ,  (cid:3)(cid:124) ··· ξ1 ξn = inf{(cid:104)Rn ξ, ξ(cid:105) : ξ =(cid:2)ξ THE MATRICIAL RANGE OF E21 23 where we use Rn to denote the (n + 1) × (n + 1) tri-diagonal block-matrix from the previous line. We were able to remove the negative signs because the infimum is taken over all n-tuples ξ1, . . . , ξn ∈ H; in particular, we can use ξ1,−ξ2, ξ3,−ξ4, . . . , (−1)n+1ξn. Now we will use to our advantage the fact that each Rn contains Rn−1 in its lower right corner. We have (cid:104)Rn ξ, ξ(cid:105) = (cid:104)ξ, ξ(cid:105) + Re(cid:104)T ξ, ξ1(cid:105) + (cid:104)Rn−1 ξ, ξ(cid:105), ··· . Fix ε > 0; by (5.3), applied to n − 1 and ξ1, we ξn can choose ξ2, . . . , ξn ∈ H so that (cid:104)Rn−1 ξ, ξ(cid:105) ≤ (cid:104)Xn−1ξ1, ξ1(cid:105) + ε. Thus, (5.4) with ξ =(cid:2)ξ1 (cid:28)(cid:20) I 1 2 T ∗ 2 T Xn−1 1 (cid:3)(cid:124) (cid:21) (cid:21)(cid:20) ξ , ξ1 = (cid:104)ξ, ξ(cid:105) + Re(cid:104)T ξ, ξ1(cid:105) + (cid:104)Xn−1ξ1, ξ1(cid:105) ≥ (cid:104)ξ, ξ(cid:105) + Re(cid:104)T ξ, ξ1(cid:105) + (cid:104)Rn−1 ξ, ξ(cid:105) − ε = (cid:104)Rn ξ, ξ(cid:105) − ε ≥ (cid:104)Xnξ, ξ(cid:105) − ε. (cid:20) ξ (cid:21)(cid:29) ξ1 (cid:20) ξ ξ1 (cid:21)(cid:29) For an arbitrary ξ1, . . . , ξn, (cid:28)(cid:20) I 1 2 T ∗ 2 T Xn−1 1 (cid:21)(cid:20) ξ (cid:21) , ξ1 = (cid:104)ξ, ξ(cid:105) + Re(cid:104)T ξ, ξ1(cid:105) + (cid:104)Xn−1ξ1, ξ1(cid:105) ≤ (cid:104)ξ, ξ(cid:105) + Re(cid:104)T ξ, ξ1(cid:105) + (cid:104)Rn−1 ξ, ξ(cid:105) = (cid:104)Rn ξ, ξ(cid:105) It follows from (5.3) and the last two estimates (writing η instead of ξ1), that (5.5) (cid:104)Xnξ, ξ(cid:105) = inf Taking limit in (5.5), (cid:104)Xξ, ξ(cid:105) = inf (cid:26)(cid:28)(cid:20) I (cid:26)(cid:28)(cid:20) I 1 1 2 T ∗ 2 T Xn−1 1 2 T ∗ 1 2 T X (cid:21)(cid:20)ξ (cid:21) (cid:21) (cid:21)(cid:20)ξ η , (cid:21)(cid:29) (cid:20)ξ (cid:21)(cid:29) (cid:20)ξ η , η η (cid:27) (cid:27) , : η ∈ H . : η ∈ H as claimed in (5.1). We have, for all ξ, η ∈ H, (cid:28)(cid:20)X 0 (cid:21)(cid:20)ξ (cid:21) 0 0 η (5.6) Thus = (cid:104)Xξ, ξ(cid:105) ≤ (cid:20)ξ η , (cid:21)(cid:29) (cid:20)I − X 1 1 2 T (cid:21) 2 T ∗ X (cid:28)(cid:20) I 1 2 T ∗ 1 2 T X (cid:21)(cid:20)ξ (cid:21) η , (cid:21)(cid:29) (cid:20)ξ η . ≥ 0, (cid:20)I − Y 24 and X belongs to the set in (5.2). Now assume that 0 ≤ Y ≤ I and that . By as- ≥ 0; this we can write as MARTÍN ARGERAMI (cid:21) (cid:21) ≥ 1 sumption, X0 = I ≥ Y . Suppose that Xn−1 ≥ Y . Then, for each ξ ∈ H, 2 T ∗ Y 1 2 T (cid:20)Y 0 (cid:21) (cid:27) 0 0 (cid:20) I (cid:21) (cid:21)(cid:20)ξ 1 2 T , η (cid:20)ξ η 1 2 T ∗ (cid:21)(cid:29) Y (cid:26)(cid:28)(cid:20) I (cid:104)Xnξ, ξ(cid:105) = inf : η ∈ H 1 2 T ∗ 2 T Xn−1 1 = inf{(cid:104)ξ, ξ(cid:105) + Re(cid:104)T ξ, η(cid:105) + (cid:104)Xn−1η, η(cid:105) : η ∈ H} ≥ inf{(cid:104)ξ, ξ(cid:105) + Re(cid:104)T ξ, η(cid:105) + (cid:104)Y η, η(cid:105) : η ∈ H} (cid:21)(cid:29) (cid:26)(cid:28)(cid:20) I (cid:26)(cid:28)(cid:20)Y 0 1 2 T 1 (cid:20)ξ (cid:21) (cid:21)(cid:20)ξ 2 T ∗ (cid:21)(cid:29) (cid:20)ξ (cid:21) Y (cid:21)(cid:20)ξ η , η 0 0 η , η : η ∈ H : η ∈ H (cid:27) = inf ≥ inf (cid:27) = (cid:104)Y ξ, ξ(cid:105). It follows by induction that Xn ≥ Y for all n, and thus X ≥ Y . (cid:3) If the manipulations in the proof of Theorem 5.1 were not impressive enough, Ando keeps going at it, with the following striking characterization of the numerical radius: are equivalent: Theorem 5.2 (Ando [1]). Let T ∈ B(H). Then the following statements (1) w(T ) ≤ 1; (2) there exist Y, Z ∈ B(H) with Y selfadjoint, (cid:107)Y (cid:107) ≤ 1, and (cid:107)Z(cid:107) ≤ 1 such that T = (I + Y )1/2Z(I − Y )1/2. {Y = Y ∗ : ∃Z, (cid:107)Z(cid:107) ≤ 1 and T = (I + Y )1/2Z(I − Y )1/2} When the above conditions are satisfied, the set (5.7) admits a maximum Ymax and a minimum Ymin. The corresponding Zmax is isometric on the range of I − Ymax, and Zmin is isometric on the range of I + Ymin. Proof. (1) =⇒ (2). By Theorem 5.1 there exists a positive contraction X satisfying (5.1).We can write this as (cid:104)Xξ, ξ(cid:105) = inf{(cid:104)ξ, ξ(cid:105) + Re(cid:104)T ξ, η(cid:105) + (cid:104)Xη, η(cid:105) : η ∈ H}. Flipping a few terms around, we get (cid:104)(I − X)ξ, ξ(cid:105) = sup{− Re(cid:104)T ξ, η(cid:105) − (cid:107)X 1/2η(cid:107)2 : η ∈ H}. Since η moves over all of H, and since the second term is invariant if we replace η with λη for λ ∈ T, we get (5.8) (cid:107)(I − X)1/2ξ(cid:107)2 = sup{(cid:104)T ξ, η(cid:105) − (cid:107)X 1/2η(cid:107)2 : η ∈ H}. THE MATRICIAL RANGE OF E21 25 As we can write tη instead of η, we have shown that for a fixed η t ∈ R. (5.9) The discriminant inequality for this quadratic is (cid:107)(I − X)1/2ξ(cid:107)2 − t(cid:104)T ξ, η(cid:105) + t2(cid:107)X 1/2η(cid:107)2 ≥ 0, (cid:104)T ξ, η(cid:105)2 ≤ 4(cid:107)(I − X)1/2ξ(cid:107)2 (cid:107)X 1/2η(cid:107)2, 1 2 (cid:104)T ξ, η(cid:105) ≤ (cid:107)(I − X)1/2ξ(cid:107)(cid:107)X 1/2η(cid:107). which we write as (5.10) Because of the supremum in (5.8), for any given ξ there exists η such that the quadratic in (5.9) is arbitrarily close to zero. Thus the discriminant can (cid:40) 1 (cid:41) be made arbitrarily close to zero by such an η, and the inequality in (5.10) can be made arbitrarily close to an equality. So 2 (cid:104)T ξ, η(cid:105) (cid:107)X 1/2η(cid:107) : X 1/2η (cid:54)= 0 (cid:107)(I − X)1/2ξ(cid:107) = sup (5.11) . We now construct a densely-defined sesquilinear form in the following way. Let H0, H1 ⊂ H be the following dense (due to X being selfadjoint) linear manifolds: H0 = {ξ0 + (I − X)1/2ξ : ξ0 ∈ ker(I − X), ξ ∈ H}, H1 = {η0 + X 1/2η : η0 ∈ ker X, η ∈ H}. Then we define, on H0 × H1, a form (5.12) [ξ0 + (I − X)1/2ξ, η0 + X 1/2η] := (cid:104)T ξ, η(cid:105). 1 2 By (5.10) the above form is well-defined and, since the kernel and range of X are orthogonal to each other, [ξ0 + (I − X)1/2ξ, η0 + X 1/2η] = 1 2 (cid:104)T ξ, η(cid:105) ≤ (cid:107)(I − X)1/2ξ(cid:107)(cid:107)X 1/2η(cid:107) ≤ (cid:107)ξ0 + (I − X)1/2ξ(cid:107)(cid:107)η0 + X 1/2η(cid:107). The sesquilinear form is thus bounded with norm at most one: we can then extended it to all of H × H. By the Riesz Representation Theorem there exists a linear contraction Z ∈ B(H), with Zker(I−X) = 0 and Z∗ker X = 0, and such that (5.13) In particular, 1 Y = Y ∗ and (I + Y )1/2Z(I − Y )1/2 = (2X)1/2Z(2I − 2X)1/2 = 2X 1/2Z(I − X)1/2 = T. Using (5.11) and (5.13), for a fixed ξ ∈ H and ε > 0 there exists η ∈ H with 2 (cid:104)T ξ, η(cid:105) = (cid:104)Z(I − X)1/2ξ, X 1/2η(cid:105), 2 T = X 1/2Z(I − X)1/2. If we now let Y = 2X − I, then ξ, η ∈ H. 1 (cid:107)(I − X)1/2ξ(cid:107) ≤ 1 2 (cid:104)T ξ, η(cid:105) (cid:107)X 1/2η(cid:107) + ε = (cid:104)Z(I − X)1/2ξ, X 1/2η(cid:105) (cid:107)X 1/2η(cid:107) + ε 26 MARTÍN ARGERAMI ≤ (cid:107)Z(I − X)1/2ξ(cid:107) + ε. But Z is a contraction and we can do this for all ε > 0, so (cid:107)Z(I − X)1/2ξ(cid:107) = (cid:107)(I − X)1/2ξ(cid:107) for all ξ ∈ H; a fortiori, as we can replace ξ with (I − X)1/2ξ, and writing Zmax for the contraction Z we constructed, we get that (cid:107)Zmax(I − X)ξ(cid:107) = (cid:107)(I − X)ξ(cid:107), From I − X = 1 2 (I − Y ), we obtain that Zmax is isometric on the range of I − Y . This Y = 2X − I we constructed, that we will denote as Ymax, is the maximum of the set in (5.7). Indeed, if T = (I + Y0)1/2Z0(I − Y0)1/2 for selfadjoint Y0 and contractive Z0, then for all ξ ∈ H. (cid:20)I − 1 2 (I + Y0) 1 2 T 1 2 T ∗ 1 2 (I + Y0) (cid:21) = 1 2 = M (cid:21) R R∗ (cid:20)I − Y0 (cid:20) I Z0 (cid:20)(I − Y0)1/2 I + Y0 (cid:21) I 0 Z∗ M∗ ≥ 0, (cid:21) . (I + Y0)1/2 2 (I + Y0) ≤ X, which we can 0 0 2 where R = (I − Y0)1/2Z∗ 0 (I + Y0)1/2 and M = 1√ By the maximality of X in (5.2), we have 1 write as Y0 ≤ 2X − I = Ymax. All of the above can be done for T ∗, so there is a maximum, say Y∗, corre- sponding to T ∗. By taking the adjoint, we can rewrite any decomposition T = (I + Y )1/2Z(I − Y )1/2 as T ∗ = (I + (−Y ))1/2Z∗(I − (−Y ))1/2. It fol- lows that −Y ≤ Y∗ for all Y that give a decomposition of T . In other words, −Y∗ = Ymin. (2) =⇒ (1) If T = (I + Y )1/2Z(I − Y )1/2 for a contraction Z then, using the trivial number inequality ab ≤ 1 2 (a2 + b2), (cid:104)T ξ, ξ(cid:105) = (cid:104)Z(I − Y )1/2ξ, (I + Y )1/2ξ(cid:105) ≤ (cid:107)(I − Y )1/2ξ(cid:107)(cid:107)(I + Y )1/2ξ(cid:107) ((cid:107)(I − Y )1/2ξ(cid:107)2 + (cid:107)(I + Y )1/2ξ(cid:107)2) ((cid:104)(I − Y )ξ, ξ(cid:105) + (cid:104)(I + Y )ξ, ξ(cid:105)) = (cid:104)ξ, ξ(cid:105). (cid:3) ≤ 1 2 = 1 2 Remark 5.3. Let us find the above decomposition for the case T = 2E21. We have w(T ) = 1, so the above results apply. In light of Remark 4.2, we may take K = (cid:96)2(Z), U the bilateral shift, and H = span{e1, e2}. Then U∗H = span{e0, e1}, and U∗kH = span{e−n+1, e−n+2, . . . , e1}. span So, in Theorem 5.1, I−Qn =(cid:80)∞ k=n E−k,−k +(cid:80)∞ k=1 k=2 Ekk and PH = E11 +E22. We get that Xn = PH (I − Qn)PH = E22 for all n. So X = E22. Then n(cid:91) (cid:21) (cid:20)α β (cid:20)α β , 1 2 T ∗ 1 2 T X (cid:21)(cid:29) (cid:21)α β γ δ (cid:28) X <(cid:20) I THE MATRICIAL RANGE OF E21 27 = β2; and  , α β γ δ > =<1 0 0 1  0 1 0 0 0 0 0 0 1 0 0 1  , α β γ δ α β γ δ > = α+δ2+β2, with the infimum over γ, δ being β2 (achieved when δ = −α). If Y satisfies (cid:20)I − Y E12 (cid:21) E21 Y ≥ 0, we immediately get from diagonal entries that 0 ≤ Y ≤ I, and considering the 2 × 2 matrix formed by the corner entries, we have (cid:20)1 − Y11 1 (cid:21) ≥ 0. 1 Y22 This can only be satisfied if Y11 = 0, Y22 = 1. From Y11 = 0 and positivity, we obtain Y12 = Y21 = 0 (a zero in the main diagonal forces its column and row to be zero). Thus Y = E22 = X. We have shown that, in this example, the set in (5.2) consists of just X. Looking into Theorem 5.2, we have Y = 2X − I = E22 − E11. The proof in that theorem constructs Z via the bilinear form on H0 × H1. In this case, (I − X)1/2 = E11, X 1/2 = E22, so the form is (cid:20)(cid:20)α (cid:21) (cid:21)(cid:21) (cid:20)γ , δ β (cid:28) 2E21 (cid:20)γ δ (cid:21)(cid:29) (cid:20)γ = (cid:28) (cid:21)(cid:29) (cid:21) (cid:20)α (cid:20)α (cid:21) β , β E21 E11 , E22 . δ = 1 2 (cid:28) = (cid:21) (cid:20)α β (cid:21)(cid:29) (cid:20)γ , δ E21 Thus Z = E21. The maximal decomposition is then 2E21 = (I + Y )1/2Z(I − Y )1/2 = (2E22)1/2E21(2E11)1/2 and, as we said, Ymax = 2X − I = E22 − E11 = (cid:20)−1 0 (cid:21) 0 1 . The computation that showed that 2X − I is maximum in the proof of Theorem 5.2 implies that if X is unique, so is Y . Thus Ymin = Ymax in this case. Remark 5.4. If instead we consider T = E21, the situation is very different. It seems hard to follow the path all the way from Theorem 4.1 to Theorem 5.1 to find X explicitly. Still, from Theorem 5.1, we know that we need to find MARTÍN ARGERAMI 28 the maximum of those selfadjoint X such that 0 ≤ X ≤ I and 1/2 0 X11 X12 X21 X22 (cid:20)I − X 1 1 − X22 2 E12 X (5.14) 1 2 E21 0 0  ≥ 0. 4 E11 + E22 satisfies (5.14), and so X ≥ It is easy to check that X0 = 3 4 E11 + E22. We get immediately that X11 ≥ 3/4 and X22 = 1 (since I − X ≥ 0). Again from I − X ≥ 0, 3 = 0 0 (cid:21) 0 1/2 −X21 1 − X11 −X12 (cid:21) (cid:20)1 − X11 X12 (cid:21) (cid:20)1 − X11 1/2 ≥ 0, X12 0 ≥ 0. 1/2 1 and so X12 = 0 by the positivity. If we now look, in (5.14), at the 2 × 2 matrix formed by the corner entries, we have (cid:20)3/4 0 (cid:21) 0 1 . Then 1 − X11 ≥ 1/4, i.e., X11 ≤ 3/4. So X11 = 3/4, and X = Now Ymax = 2X − I = (cid:20)1/2 0 (cid:21) 0 1 . If we look again at the proof of Theorem 5.2, we have (I − X)1/2 = 1 X 1/2 = The maximal decomposition is then 2 E11, √ 2 E11 +E22. Writing (5.13) explicitly, we immediately get Z = E21. 3 (cid:20)(cid:112)3/2 0 (cid:21)(cid:20)0 0 (cid:21)(cid:20)1/ √ 1 0 0 √ 0 2 (cid:21) 2 0 0 . E21 = (I + Ymax)1/2Z(I − Ymax)1/2 = As opposed to the previous case, though, different decompositions are pos- sible. If we repeat the analysis above for T ∗ = E12, we find that X∗ is now 2 E22. Then, with E11 + 3 respect to our original T = E21, we have 4 E22. Its maximum Y∗ will be 2X∗ − I = E11 + 1 (cid:20)−1 (cid:21) Ymin = −Y∗ = 0 0 −1/2 . We can also get decompositions that do not come from Ymax nor Ymin. For a trivial one, take Y0 = 0, Z0 = E21. Then, of course, E21 = (I + 0)1/2E21(I − 0)1/2. Yet another fairly trivial decomposition can be found if Y = E22. 2 E22, and (I − Y )1/2 = E11. So we get the Then (I + Y )1/2 = E11 + decomposition √ (cid:20)1 0 (cid:21)(cid:20) 0 √ 0 2 √ 0 2 0 1/ (cid:21)(cid:20)1 0 (cid:21) . 0 0 E21 = The following result is a well-known matricial characterization of the nu- merical radius. It uses Theorem 5.2 in an essential way. alent: THE MATRICIAL RANGE OF E21 29 Corollary 5.5. Let T ∈ B(H). Then the following statements are equiv- (1) w(T ) ≤ 1/2; (2) there exists A ∈ B(H)+, with A ≤ I, such that (cid:20)A ≥ 0. (cid:21) T ∗ T I − A (cid:20)A (cid:28)(cid:20)A T ∗ T I − A T ∗ T I − A (cid:21) (cid:21)(cid:20)ξ (cid:21) (cid:21)(cid:29) (cid:20)ξ , η η Proof. If 0 ≤ ≥ 0, then for any ξ, η ∈ H = (cid:104)Aξ, ξ(cid:105) + (cid:104)(I − A)η, η(cid:105) + 2 Re(cid:104)T ξ, η(cid:105). Taking η = λξ for λ ∈ T such that (cid:104)T ξ, ξ(cid:105) = λ(cid:104)T ξ, ξ(cid:105), we get 0 ≤ (cid:107)ξ(cid:107)2 − 2(cid:104)T ξ, ξ(cid:105), implying w(T ) ≤ 1/2. Conversely, if w(T ) ≤ 1/2, by Theorem 5.2 there exist a selfadjoint con- traction Y and contraction Z with 2T = (I + Y )1/2Z(I − Y )1/2. Since Z is contractive, the matrix is positive; then (cid:21) (cid:21) (I + Y )1/2Z∗(I − Y )1/2 (cid:21)(cid:20) I Z∗ I − Y (cid:21)(cid:20)(I + Y )1/2 Z I 0 0 (I − Y )1/2 (cid:21) (cid:20)I + Y 2T 2T ∗ I − Y I + Y = (I − Y )1/2Z(I + Y )1/2 (cid:21) Z I (cid:20) I Z∗ (cid:20) (cid:20)(I + Y )1/2 (cid:20)A 0 = ≥ 0. 0 (I − Y )1/2 (cid:21) T ∗ T I − A Multiplying by 1/2 we get ≥ 0, where A = 1 2 (I + Y ). (cid:3) Our main use of this result is Corollary 5.6, characterizing the matricial range of E21. The use of Corollary 5.5 to characterize W(E21) is a very well-known result, though we are not aware of any published reference. alent: Corollary 5.6. For any T ∈ B(H), the following statements are equiv- (1) w(T ) ≤ 1/2; (2) there exists ϕ : M2(C) → B(H), ucp, with T = ϕ(E21). Proof. (1) =⇒ (2). By Corollary 5.5, there exists A ∈ B(H)+ with A ≤ I and ≥ 0. Now define a linear map ϕ : M2(C) → B(H) by (cid:20)A (cid:21) T ∗ T I − A ϕ(E11) = A, ϕ(E21) = T, ϕ(E12) = T ∗, ϕ(E22) = I − A. Then ϕ is unital, and by Choi's criterion (Proposition 2.1) it is completely positive, since (cid:18)(cid:20)E11 E12 (cid:21)(cid:19) E21 E22 ϕ(2) (cid:20)A T ∗ T I − A = ≥ 0. (cid:21) 30 MARTÍN ARGERAMI (2) =⇒ (1). Let A = ϕ(E11). As ϕ is ucp, Choi's criterion (Proposi- tion 2.1) implies that(cid:20)A (cid:21) T ∗ T I − A (cid:18)(cid:20)E11 E12 (cid:21)(cid:19) E21 E22 ≥ 0, = ϕ(2) and so by Corollary 5.5, w(T ) ≤ 1/2. (cid:3) We remark that the proof (2) =⇒ (1) in Corollary 5.6 can be achieved without appealing to Ando's nor Choi's results. Indeed, by using the Stine- spring dilation one can show that if φ is ucp, then W1(φ(T )) ⊂ W1(T ). We finish this section with another characterization due to Ando. The interesting information we find in its proof, is that the operator C below allows us to express a unitary 2-dilation of T explicitly. Theorem 5.7 (Ando [1]). Let T ∈ B(H). Then the following statements are equivalent: (1) w(T ) ≤ 1; (2) there exists C ∈ B(H), contractive, such that T = 2(I − C∗C)1/2C. Proof. (1) =⇒ (2) By Theorem 5.2, there exist Y, Z ∈ B(H), both con- tractive and Y selfadjoint, with Z isometric on the range of (I − Y )1/2, and satisfying T = (I + Y )1/2Z(I − Y )1/2. Let C = 1√ Z(I − Y )1/2. The fact that Z is isometric on the range of (I − Y )1/2 can be written as (I − Y )1/2Z∗Z(I − Y )1/2 = (I − Y )1/2(I − Y )1/2 = I − Y . Then I − C∗C = I − 1 Thus 2 (I − Y )1/2Z∗Z(I − Y )1/2 = I − 1 2 (I − Y ) = 1 2 (I + Y ). 2 2(I − C∗C)1/2C = (I + Y )1/2Z(I − Y )1/2 = T. (2) =⇒ (1) We will explicitly construct a unitary 2-dilation of T , and then the result will follow from Theorem 4.4. We first construct a unitary dilation k∈Z H as follows: I ⊗ Ek+1,k + C ⊗ E0,0 + (I − CC∗)1/2 ⊗ E0,−1 W of C on K =(cid:76) (cid:88) W = k≥1 k≤−2 + (I − C∗C)1/2 ⊗ E1,0 − C∗ ⊗ E1,−1. We encourage the reader to check that this is indeed a unitary; besides a decent amount of patience and care, the only non-trivial (but well-known) manipulation required is to note that C(I − C∗C)1/2 = (I − CC∗)1/2C. k∈Z I ⊗ Ek+1,k, we define U = S∗W 2. With S the bilateral shift S = (cid:80) This is again a unitary and it has the form U = I ⊗ Ek,k−1 + (I − C∗C)1/2 ⊗ E1,0 − C∗ ⊗ E1,−1 + C2 ⊗ E−1,0 (cid:88) k≥2 k≤−2 THE MATRICIAL RANGE OF E21 31 + C(I − CC∗)1/2 ⊗ E−1,−1 + (I − CC∗)1/2 ⊗ E−1,−2 + (I − C∗C)1/2C ⊗ E0,0 + (I − C∗C)1/2(I − CC∗)1/2 ⊗ E0,−1 − C∗ ⊗ E0,−2. 2 T n, (U n)0,−1 = T n−1(I−C∗C)1/2(I−CC∗)1/2, We claim that this U is a 2-dilation of T ; that is, that (U n)0,0 = 1 2 T n for all n ∈ N. We proceed by induction. Assume that, for a fixed n and for all (cid:96) ≥ 1, (U n)0,0 = 1 (U n)0,(cid:96) = 0. This clearly holds for n = 1, and so now we show that the above equalities for n imply the corresponding versions for n + 1. We will use repeatedly the equality C(I − C∗C)1/2 = (I − CC∗)1/2C. Then (recall that our hypothesis is that T = 2(I − C∗C)1/2C and that U(cid:96),0 (cid:54)= 0 and U(cid:96),−1 (cid:54)= 0 only when (cid:96) ∈ {−1, 0, 1}) (U n+1)0,−1 = (U n)0,−1U−1,−1 + (U n)0,0U0,−1 + (U n)0,1U1,−1 = T n−1(I − C∗C)1/2(I − CC∗)1/2C(I − CC∗)1/2 + 1 2 T n(I − C∗C)1/2(I − CC∗)1/2 = T n−1(I − C∗C)1/2C(I − C∗C)1/2(I − CC∗)1/2 = 1 2 T n(I − C∗C)1/2(I − CC∗)1/2 + 1 2 T n(I − C∗C)1/2(I − CC∗)1/2 + 1 2 T n(I − C∗C)1/2(I − CC∗)1/2 = T n(I − C∗C)1/2(I − CC∗)1/2. Also, (U n+1)0,0 = (U n)0,−1U−1,0 + (U n)0,0U0,0 + (U n)0,1U1,0 = T n−1(I − C∗C)1/2(I − CC∗)1/2C2 + 1 = T n−1(I − C∗C)1/2C(I − C∗C)1/2C + 1 2 T n 1 2 T 4 T n+1 = 1 4 T n+1 + 1 4 T n+1 = 1 2 T n+1. And, since for (cid:96) ≥ 1 the only k such that Uk,(cid:96) (cid:54)= 0 is k = (cid:96) + 1 (U(cid:96)+1,(cid:96) = I), (U n+1)0,(cid:96) = (U n)0,(cid:96)+1U(cid:96)+1,(cid:96) = (U n)0,(cid:96)+1 = 0. The induction is then complete: for all n ∈ N, we have (U n)0,0 = 1 2 T n. (cid:3) 32 MARTÍN ARGERAMI 6. Toeplitz Matrices The goal in this section is Theorem 6.5. As this is a matricial generaliza- tion of the classical Theorem 6.3, we present first the scalar version to fix ideas. 6.1. Scalar Matrices. The following is [15, Lemma 2.5]. Lemma 6.1 (Fejer-Riesz). Let τ be a trigonometric polynomial of the form τ (λ) =(cid:80)N−N anλn. If τ (λ) > 0 for all λ ∈ T, then there exists a polynomial p(z) =(cid:80)N n=0 pnzn such that τ (λ) = p(λ)2, λ ∈ T. Proof. Since τ (λ) > 0, we have N(cid:88) n=−N anλn = Re anλn (cid:33) (cid:32) N(cid:88) N(cid:88) n=−N N(cid:88) n=1 Re(anλn + a−nλ−n) Re(anλn) = Re(a0) + = n=−N = Re(a0) + (an + a−n)λn + (an + a−n)λ−n 2 N(cid:88) N(cid:88) n=1 = Re(a0) + an + a−n λn. n=−N 2 , and then an = a−n. Also, a0 ∈ R. It follows that, for all n, an = an+a−n If necessary, we may decrease N so that a−N (cid:54)= 0. Although we consider τ as a polynomial on T, its formula works of course for all z ∈ C. Define g(z) = zN τ (z). Note that g(0) = a−N (cid:54)= 0, so all roots of g are nonzero. Also, for λ ∈ T, g(λ) = λN τ (λ) (cid:54)= 0, so no zero of g is in T. We have 2 g(1/z) = 1 zN τ (1/z) = z−N anz−n = z−N a−nz−n N(cid:88) n=−N N(cid:88) n=−N N(cid:88) n=−N = z−N anzn = z−2N g(z). This implies that g(z) = 0 if and only if g(1/z) = 0. Since no zero is in T, we get that the zeroes of g are of the form z1, . . . , zN , 1/z1, . . . , 1/zN. Then g(z) = aN r(z)s(z), THE MATRICIAL RANGE OF E21 33 where r, s are the polynomials N(cid:89) r(z) = (z − zn), s(z) = n=1 n=1 N(cid:89) (z − 1/zn). These two polynomials are related by s(z) = (−1)N zN r(1/z) z1 ··· zN . Then, for λ ∈ T, τ (λ) = τ (λ) = λ−N g(λ) = g(λ) = aNr(λ)s(λ) (cid:12)(cid:12)(cid:12)(cid:12) (−1)N λ−N r(λ) (cid:12)(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)1/2 (cid:12)(cid:12)(cid:12) aN z1 ··· zN z1···zN (cid:12)(cid:12)(cid:12)(cid:12) r(z). (cid:12)(cid:12)(cid:12)(cid:12) r(λ)2. aN z1 ··· zN = aNr(λ) Thus τ (λ) = p(λ)2, where p(z) = (cid:3) Matricial versions of the Fejer-Riesz Lemma exist -- see for instance [10,14, 20] -- but we will not discuss them here. Recall from page 2 that we denote by Sn the n × n unilateral shift. Definition 6.2. An n × n Toeplitz matrix is a matrix T of the form n−1(cid:88) k=1 a−1 a−2 T = a0 In + akSk n + where ak ∈ C for all k. Graphically, this is ··· ... ... ... ... a1 T = a0 a1 a0 a1 ··· a−1 a−2 a−1 ... ... ... ... ··· a0 ... ... ... ... ··· a2 ... ... ... ... an−1  a2 ... ... ... ··· n−1(cid:88) a−kS∗ n k, k=1 ··· ··· ... ... ... ... ... ... ... a−2 a−1 a0 a2 ··· a1 a2 a0 a1  a−n+1 ... ... ... ... a−2 a−1 a0 If in particular T is Hermitian, i.e. T = T ∗, then n−1(cid:88) (6.1) T = a0 In + 2 Re akSk n, a0, a1, . . . , an−1 ∈ C. k=1 In the case of a Hermitian Toeplitz matrix we will write, when needed, a−k = ak. The following theorem is based on [15, Theorem 2.14]; we do not need to make use of this theorem, but we will use a matricial generalization, 34 MARTÍN ARGERAMI Theorem 6.5 and so the scalar proof might help some readers. Paulsen considers infinite sequences, which we don't need here. Theorem 6.3. Let T be a Hermitian Toeplitz matrix as in (6.1). Then the following statements are equivalent: (1) T is positive; (2) there exists a positive linear functional φ on C(T) such that ak = φ(zk), k = 0, 1, . . . , n − 1. τ (λ) = (cid:80)m Proof. (1) =⇒ (2) Consider the operator system S = span{zk : k ∈ Z} ⊂ C(T). Define a linear map φ : S → C by φ(zk) = ak, φ(z−k) = ak for k = 0, . . . , n − 1, and φ(zk) = 0 for k ≥ n. Let τ ∈ S be strictly positive, i.e. τ (λ) > 0 for all λ ∈ T. By Lemma 6.1, there exist p0, p1, . . . , pm such that k,j=0 pkpj λk−j. Assume, without loss of generality, that m ≥ n (we complete the list of pk with zeroes if it is not the case). Then, with the convention that a−k = ak, ak = 0 if k ≥ n, and with x = (p0, . . . , pm) (cid:124), m(cid:88) k,j=0 φ(τ ) = pkpj ak−j = (cid:104)(T ⊕ 0)x, x(cid:105) ≥ 0 by the positivity of T . For arbitrary positive τ, we have that for any ε > 0 the function τ(cid:48)(λ) = τ (λ) + ε is strictly positive, and so φ(τ ) + ε = φ(τ(cid:48)) ≥ 0 for all ε > 0, which implies that φ(τ ) ≥ 0. Thus φ is a positive linear functional on the operator system of the trigonometric polynomials; this implies that it is bounded and we can extend it by density to C(T). (2) =⇒ (1) Note that, since T is Hermitian, a−k = ak = φ(zk) = φ(z−k). Given x = (p0, . . . , pn−1) (cid:124) ∈ Cn, n−1(cid:88) n−1(cid:88) φ(zk−j)pkpj = φ (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n(cid:88) k=0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2 ≥ 0. zkpk k=0 j=0 (cid:3) Remark 6.4. The proof of (2) =⇒ (1) in Theorem 6.3 can also be achieved by using that φ is completely positive (due to the abelian domain) and then noting that n−1(cid:88) n−1(cid:88) k=0 j=0 (cid:104)T x, x(cid:105) = ak−jpkpj = T = In + 2 Re = φ(n) k=1 n−1(cid:88) 1 z 0 0 ... 0 0  φ(zk)Sk n z2 ··· ··· ···  ∗1 z 0 0 ... 0 0 zn−1 0 ... 0   ≥ 0. z2 ··· ··· ··· zn−1 0 ... 0 THE MATRICIAL RANGE OF E21 35 6.2. Block Toeplitz Matrices. An n× n block Toeplitz matrix is a matrix T of the form T = A0 ⊗ In + Ak ⊗ Sk n + A−k ⊗ S∗ n k, n−1(cid:88) k=1 n−1(cid:88) k=1 n−1(cid:88) where A−(n−1), . . . , A0, . . . , An−1 ∈ A for some C∗-algebra A. If in particular T is Hermitian, i.e. T = T ∗, then T = A0 ⊗ In + 2 Re (6.2) and we may write A−k = A∗ k. it in the proof of Theorem 7.1, with A = Mm(C). k=1 Ak ⊗ Sk n Theorem 6.5 is the block-matrix version of Theorem 6.3. We will later use Theorem 6.5. Let T be a Hermitian block Toeplitz matrix as in (6.2), with coefficients in the C∗-algebra A. Then the following statements are equivalent: (1) T is positive; (2) there exists a (completely) positive map φ : C(T) → A such that φ(zk) = Ak, k = 0, . . . , n − 1. (6.3) Proof. (1) =⇒ (2) Consider the operator system S = span{zk : k ∈ Z} ⊂ C(T). Define a linear map φ : S → A by φ(zk) = Ak, φ(z−k) = A∗ k for k = 0, . . . , n − 1, and φ(zk) = 0 for k ≥ n. Let τ ∈ S be strictly positive, i.e. τ (λ) > 0 for all λ ∈ T. By Lemma 6.1, there exist p0, p1, . . . , pm such that k,j=0 pkpj λk−j. Assume, without loss of generality, that m ≥ n (we complete the list of pk with zeroes if it is not the case). Then, with the convention that A−k = A∗ τ (λ) = (cid:80)m k, Ak = 0 if k ≥ n, m(cid:88) k,j=0 φ(τ ) = pkpj Ak−j = ∗ (cid:20)T  p0 I ... pm I (cid:21) p0 I ... pm I  ≥ 0 0m−n by the positivity of T . For arbitrary positive τ, we have that for any ε > 0 the function τ(cid:48)(λ) = τ (λ) + ε is strictly positive, and so φ(τ ) + ε = φ(τ(cid:48)) ≥ 0 for all ε > 0, which implies that φ(τ ) ≥ 0. Thus φ is a positive linear map in the operator system of the trigonometric polynomials; thus it is bounded, and it extends by density to a positive map on C(T) with range still contained in A. (2) =⇒ (1) Since C(T) is abelian, φ is completely positive. Consider a Stinespring dilation of φ, i.e. φ(f ) = V ∗π(f )V . Then, for each vector x = (ξ0, . . . , ξn−1) (cid:104)T x, x(cid:105) = (cid:104)V ∗π(zk−j)V ξk, ξj(cid:105) (cid:104)Ak−jξk, ξj(cid:105) = (cid:124) ∈ H n, n−1(cid:88) n−1(cid:88) n−1(cid:88) n−1(cid:88) k=0 j=0 k=0 j=0 MARTÍN ARGERAMI (cid:104)π(zj)∗π(zk)V ξk, V ξj(cid:105) 36 = = n−1(cid:88) n−1(cid:88) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)n−1(cid:88) k=0 j=0 k=0 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2 π(zk)V ξk ≥ 0. (cid:3) 7. Nilpotent Dilations and Matricial Range The following is a significant technical result by Arveson, characterizing those contractions that can be power-dilated to nilpotents. The proof does not follow the original; in particular, the argument that we offer for (4) =⇒ (1) is more algebraic and direct that Arveson's original. Theorem 7.1. [3, Theorem 1.3.1] Let T ∈ B(H) with (cid:107)T(cid:107) ≤ 1 and let n ∈ N with n ≥ 2. Then the following statements are equivalent: (1) There exists φ : Mn(C) → B(H), ucp, with φ(Sj n) = T j, j = 1, . . . , n − 1. unitarily equivalent to(cid:76) (4) I + 2 Re(cid:80)n−1 (2) There exists a Hilbert space K ⊇ H and N ∈ B(K) such that N is j Sn, and T j = PH N jH, j = 0, 1, . . . , n−1. (3) There exists a Hilbert space K ⊇ H and N ∈ B(K) such that (cid:107)N(cid:107) ≤ 1, N n = 0, and T j = PH N jH, j = 0, 1, . . . , n − 1. k=1 λkT k ≥ 0 for all λ ∈ T. Proof. (1) =⇒ (2) This is a straightforward consequence of Stinespring's Indeed, after writing φ(X) = PH π(X)H -- under the Dilation Theorem. usual identification of H with its range under V -- we can take N = π(Sn) (recall that any representation of Mn(C) is of the form X (cid:55)−→ X ⊗ I). (2) =⇒ (3) Trivial. (3) =⇒ (4) We have that T j = PH N jH for j = 0, 1, . . . , n − 1. As I + 2 Re(cid:80)n−1 compressions are (completely) positive, it is enough to prove the inequality k=1 λkN k ≥ 0 for all λ ∈ T. Now we can take advantage of the fact that N n = 0. Fix z ∈ C with z < 1. We have n−1(cid:88) zkN k = Re(cid:0)I + 2zN (I − zN )−1(cid:1) zkN k = I + 2 Re ∞(cid:88) I + 2 Re k=1 = Re(cid:0)[(I − zN ) + 2zN ](I − zN )−1(cid:1) k=1 = Re(I + zN )(I − zN )−1 (cid:0)(I + zN )(I − zN )−1 + (I − zN∗)−1(I + zN∗)(cid:1) = = 2(I − zN∗)−1(I − z2N∗N )(I − zN )−1 ≥ 0. 1 2 THE MATRICIAL RANGE OF E21 all r ∈ [0, 1). The positivity is preserved as r (cid:37) 1. Now given λ ∈ T, we have from above that I + 2 Re(cid:80)n−1 such that A0 ⊗ In + 2 Re(cid:80)n−1 k=1 rkλkT k ≥ 0 for (4) =⇒ (1) Define a linear map ψ : OS(I, Sn, . . . , Sn−1 ) → B(H) by n) = T j. We will show that ψ is ucp. If we take A0, . . . , An−1 ∈ Mm(C) ψ(Sj n ≥ 0, then by Theorem 6.5 there exists a completely positive map φ : C(T) → Mm(C) with φ(zk) = Ak for k = 0, . . . , n − 1. We have k=1 Ak ⊗ Sk 37 n ψ(m) A0 ⊗ In + 2 Re Ak ⊗ Sk n = A0 ⊗ ψ(In) + 2 Re Ak ⊗ ψ(Sk n) (cid:33) n−1(cid:88) k=1 (cid:32) n−1(cid:88) n−1(cid:88) k=1 k=1 n−1(cid:88) (cid:33) k=1 = φ(1) ⊗ I + 2 Re φ(zk) ⊗ T k (cid:32) (cid:33) = (φ ⊗ id) 1 ⊗ I + 2 Re zk ⊗ T k . The expression inside the brackets is a function C(T) → Mn(C). For each λ ∈ T it evaluates to 1 ⊗ I + 2 Re λk ⊗ T k = 1 ⊗ n−1(cid:88) n−1(cid:88) I + 2 Re ≥ 0. (cid:32) λkT k k=1 k=1 As φ is ucp, φ⊗ id is positive (this can be seen quickly by using Stinespring) and so the expression (φ ⊗ id) (·) above is positive; thus ψ(m) is positive for any m, i.e., completely positive. (cid:3) Now we can extend it via Arveson's Extension Theorem to Mn(C). In the case n = 2, Theorem 7.1 characterizes the matricial range of S2 = (cid:20)0 0 (cid:21) 1 0 E21 = . Indeed: Corollary 7.2. The extended matricial range of the 2× 2 unilateral shift consists of the set of all operators T ∈ B(H) (for any dimension of H) such that w(T ) ≤ 1/2. Proof. Note that w(T ) ≤ 1/2 implies that (cid:107)T(cid:107) ≤ 1, since (cid:107)T(cid:107) ≤ 2w(T ). (cid:3) Now combine the case n = 2 in Theorem 7.1 with Proposition 4.3. We remark that for n > 2, Theorem 7.1 does not characterize the matricial range of Sn, as one is considering the additional requirement that higher powers of Sn are mapped to the corresponding powers of T . Forty five years after the above results, the matricial range of Sn, n ≥ 3, has not been characterized. As far as we can tell, there is not even a conjecture of what it could be. Remark 7.3. Fix n ∈ N. The map γ : span{I, Sn, S2 } → C(T) given by γ(Sj n) = zj, where z is the identity function, is never completely contractive. Indeed, every power of the shift is mapped to a unitary. If γ n, . . . , Sn−1 n 38 were ucc, we would be able to extend it to a ucp map Mn(C) → X where X is the injective envelope of C(T). Then MARTÍN ARGERAMI γ(E11) = γ(S∗ n n−1Sn−1 n ) ≥ γ(Sn−1 )∗γ(Sn−1 n n ) = 1 = γ(I). Positivity then implies that γ(E22) = ··· = γ(Enn) = 0. So, for any k = 2, . . . , n, 0 ≤ γ(Ek1)γ(E1k) ≤ γ(Ek1E1k) = γ(Ekk) = 0, implying that γ(Ek1) = 0 for k ≥ 2. Then (cid:32)n−1(cid:88) (cid:33) γ(Sn) = γ Ek+1,k = 0, a contradiction. k=1 8. Characterizations of the Numerical Radius The following theorem requires no proof, as it just collects equivalences we proved in previous sections. It is a consequence of very subtle ideas. Theorem 8.1. Let T ∈ B(H). The following statements are equivalent: (1) w(T ) ≤ 1; (2) Re(λT ) ≤ I for all λ ∈ T; (3) there exists a Hilbert space K ⊃ H, and a unitary U ∈ B(K) such that T n = 2PH U nH for all n ∈ N; (4) there exists a Hilbert space K ⊃ H, and N ∈ B(K), unitarily equiv- alent to(cid:76) j E21, such that T = 2PH NH; (5) there exists a Hilbert space K ⊃ H, and N ∈ B(K) with (cid:107)N(cid:107) ≤ 1 and N 2 = 0, such that T = 2PH NH; (6) there exist contractions Y, Z ∈ B(H), with Y = Y ∗, such that T = (I − Y )1/2Z(I + Y )1/2; (cid:20) A (cid:21) (7) there exists A ∈ B(H), with 0 ≤ A ≤ I, such that ≥ 0; (8) there exists A ∈ B(H), with 0 ≤ A ≤ I, with T = 2(I − A∗A)1/2A; (9) there exists ϕ : M2(C) → B(H), ucp, with ϕ(E21) = 1 2T I − A 2T ∗ 2 T . Proof. Combine Proposition 4.3 with Theorem 4.4, Corollary 5.5, Theo- (cid:3) rem 5.7, and Theorem 7.1. Acknowledgements This work has been supported in part by the Discovery Grant program of the Natural Sciences and Engineering Research Council of Canada. I would also like to thank D. Farenick for his encouragement and for several useful suggestions. THE MATRICIAL RANGE OF E21 39 References ∗ 34:11 -- 15, 1973. [1] T. Ando. Structure of operators with numerical radius one. Acta Sci. Math. (Szeged), [2] W. Arveson. Subalgebras of C∗-algebras. Acta Math., 123:141 -- 224, 1969. [3] W. Arveson. Subalgebras of C∗-algebras. II. Acta Math., 128(3-4):271 -- 308, 1972. [4] W. Arveson. Notes on extensions of C -algebras. Duke Math. J., 44(2):329 -- 355, 1977. [5] D. P. Blecher and C. Le Merdy. Operator algebras and their modules -- an operator space approach, volume 30 of London Mathematical Society Monographs. New Se- ries. The Clarendon Press Oxford University Press, Oxford, 2004. Oxford Science Publications. [6] N. P. Brown and N. Ozawa. C∗-algebras and finite-dimensional approximations, vol- ume 88 of Graduate Studies in Mathematics. American Mathematical Society, Prov- idence, RI, 2008. [7] J. Bunce and N. Salinas. Completely positive maps on C∗-algebras and the left ma- [8] K. R. Davidson. C∗-algebras by example, volume 6 of Fields Institute Monographs. tricial spectra of an operator. Duke Math. J., 43(4):747 -- 774, 1976. American Mathematical Society, Providence, RI, 1996. [9] K. R. Davidson and M. Kennedy. The Choquet boundary of an operator system. Duke Math. J., 164(15):2989 -- 3004, 2015. [10] L. Ephremidze, G. Janashia, and E. Lagvilava. A simple proof of the matrix-valued Fejér-Riesz theorem. J. Fourier Anal. Appl., 15(1):124 -- 127, 2009. [11] D. Farenick. Arveson's criterion for unitary similarity. Linear Algebra Appl., [12] D. R. Farenick. Matricial extensions of the numerical range: a brief survey. Linear and Multilinear Algebra, 34(3-4):197 -- 211, 1993. [13] M. Hamana. Injective envelopes of operator systems. Publ. Res. Inst. Math. Sci., 435(4):769 -- 777, 2011. 15(3):773 -- 785, 1979. [14] H. Helson and D. Lowdenslager. Prediction theory and Fourier series in several vari- ables. Acta Math., 99:165 -- 202, 1958. [15] V. Paulsen. Completely bounded maps and operator algebras, volume 78 of Cambridge Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 2002. [16] G. Pisier. Introduction to operator space theory, volume 294 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 2003. [17] R. R. Smith and J. D. Ward. Matrix ranges for Hilbert space operators. Amer. J. Math., 102(6):1031 -- 1081, 1980. [18] B. Sz.-Nagy, C. Foias, H. Bercovici, and L. Kérchy. Harmonic analysis of operators on Hilbert space. Universitext. Springer, New York, second edition, 2010. [19] S.-H. Tso and P. Y. Wu. Matricial ranges of quadratic operators. Rocky Mountain J. [20] N. Wiener and E. J. Akutowicz. A factorization of positive Hermitian matrices. J. Math., 29(3):1139 -- 1152, 1999. Math. Mech., 8:111 -- 120, 1959. Department of Mathematics and Statistics, University of Regina, 3737 Wascana Pwy, Regina, SK S4S0A2, Canada E-mail address: [email protected]
1112.2769	1	1112	2011-12-13T01:28:34	Some inverse limits of Cuntz algebras	[ "math.OA", "math.KT" ]	We construct a nontrivial inverse system of Cuntz algebras $\{{\cal O}_{n}:2\leq n<\infty\}$, whose inverse limit is -isomorphic onto ${\cal O}_{\infty}$. By using this result, it is shown that the $K_{0}$-functor is discontinuous with respect to the inverse limit even if the limit is a C$^{}$-algebra.	math.OA	math	Some inverse limits of Cuntz algebras Katsunori Kawamura∗ College of Science and Engineering, Ritsumeikan University, 1-1-1 Noji Higashi, Kusatsu, Shiga 525-8577, Japan Abstract We construct a nontrivial inverse system of Cuntz algebras {On : 2 ≤ n < ∞}, whose inverse limit is ∗-isomorphic onto O∞. By using this result, it is shown that the K0-functor is discontinuous with respect to the inverse limit even if the limit is a C∗-algebra. Mathematics Subject Classifications (2010). 46M15; 46M40; 46L80 Key words. Cuntz algebra; pro-C∗-algebra; inverse limit; K-functor 1 Introduction In this paper, we construct a nontrivial inverse system of Cuntz algebras {On : 2 ≤ n < ∞}, whose inverse limit is O∞: On ∼= O∞. lim←− (1.1) In order to explain (1.1), we will recall inverse system of C∗-algebras and pro-C∗-algebra, and show the construction of the inverse system in this section. 1.1 Unital ∗-homomorphisms among Cuntz algebras Our study is motivated by well-known facts of unital ∗-homomorphisms among Cuntz algebras. Hence we start with their explanation in this sub- section. For unital C∗-algebras A and B, let Hom(A, B) denote the set of all unital ∗-homomorphisms from A to B and let K0(A) denote the K0-group of A [5]. ∗e-mail: [email protected]. 1 Fact 1.1 For 2 ≤ n ≤ ∞, let On denote the Cuntz algebra [11]. Then the following holds: (i) For 2 ≤ n ≤ ∞ and a unital C∗-algebra A, if f ∈ Hom(A, On), then the induced homomorphism f from K0(A) to K0(On) is surjective. (ii) For 2 ≤ m, n < ∞, Hom(Om, On) 6= ∅ if and only if there exists a positive integer k such that m = (n − 1)k + 1. (iii) For any 2 ≤ m < ∞, Hom(Om, O∞) = ∅. (iv) For any 2 ≤ n ≤ ∞, Hom(O∞, On) 6= ∅. (i) From Example 6.3.2 in [5], the class of the unit of On is the Proof. generator of K0(On). Since f maps the class of the unit of A to that of On, the statement holds. (ii) This holds from Lemma 2.1 in [28] (see also [14], p164,V.16.) In § 1.3, we will given concrete ∗-homomorphisms among Cuntz algebras. (iii) From [12], K0(On) ∼= Z/(n − 1)Z (2 ≤ n < ∞) and K0(O∞) ∼= Z. From these and (i), the statement holds. (iv) This will be shown by using concrete ∗-homomorphisms {fn,∞ : n ≥ 1} in (1.10). About homomorphisms among Cuntz algebras, more general results are known ([20], Lemma 7.1.) Since On is simple for each n, any unital ∗-homomorphism from On is injective, hence it is an embedding of On. For examples of Fact 1.1(ii), the following embeddings among O2, . . . , O8 (except endomorphisms) are illustrated as follows ([28], §2.1): Figure 1.2 ✶ O2 ✒ ❨ ✐ O6 O8 ✶ O3 ❨ O5 O7 O4 ✯ where an arrow "A → B" means a unital embedding of A into B. For 2 ≤ n < m ≤ 8, there is no unital ∗-homomorphism from Om to On if there is no oriented path from Om to On in Figure 1.2. 2 Remark 1.3 In p184 of [11], there is a statement about embeddings among Cuntz algebras as follows: O2 ⊃ O3 ⊃ O4 ⊃ . . . ⊃ O∞. (1.2) It is explained that (1.2) is given by using the induction for the construction of a certain unital embedding of O3 into O2. However, (1.2) never means unital embeddings because of Fact 1.1(ii) except "O2 ⊃ O3" and "⊃ O∞". On the other hand, there exist well-known two orders on the set N of all positive integers {1, 2, 3, . . .}. The first is the standard linear order ≤, that is, 1 ≤ 2 ≤ 3 ≤ · · · . The second is the order (cid:22) on N ([6], §1.11) defined as m (cid:22) n if m divides n. (1.3) This relation m (cid:22) n is usually written as mn in number theory [21]. Both (N, ≤) and (N, (cid:22)) are directed sets, but (N, (cid:22)) is not a totally ordered set, which is illustrated as the following directed graph (except relations n (cid:22) n): Figure 1.4 ✮ 4 ✮ ❥ ✙ 2 ✠ 3 6 1 ❥ q 7 5 By comparison with Figure 1.2, it is clear that Figure 1.4 is just the graph with inverse direction of Figure 1.2 by rewriting their suffix numbers. This idea is rigorously verified by using Fact 1.1(ii) and we can restate Fact 1.1(ii), (iii) and (iv) by using the order (cid:22) as follows. Corollary 1.5 Let N ≡ N ∪ {∞} and extend (cid:22) on N as ∞ (cid:22) ∞ and n (cid:22) ∞ for each n ∈ N. Then, for n, m ∈ N, Hom(Om+1, On+1) 6= ∅ if and only if n (cid:22) m (1.4) where n + ∞ means ∞ for convenience. Especially, (1.4) holds for each n, m ∈ N, and there exists no unital ∗-homomorphism from Om to On when n > m. From Corollary 1.5, if there exist the following unital inclusions On1+1 ⊃ On2+1 ⊃ On3+1 ⊃ · · · (1.5) for {ni ∈ N : i ≥ 1}, then we obtain order relations n1 (cid:22) n2 (cid:22) n3 (cid:22) · · · . It is well-known that the order (cid:22) is used in the theory of profinite groups [39]. A profinite group is defined as an inverse limit of finite groups. From this and Corollary 1.5, the following questions are inspired. 3 Question 1.6 (i) Does there exist an inverse system of Cuntz algebras {On+1 : 1 ≤ n < ∞} over the directed set (N, (cid:22)) with respect to embeddings in Figure 1.2? (ii) If such an inverse system in (i) is found, then what-like algebra is the inverse limit? (iii) If the answer to (ii) is given, then what is the meaning of this result? The purpose of this paper is to give answers to these questions. 1.2 Inverse limits of C∗-algebras In order to consider inverse limits of Cuntz algebras in Question 1.6, we recall previous works of inverse limits of C∗-algebras in this subsection. According to Phillips [37], Fragoulopoulou [18] and Joit¸a [25], inverse limits (or projective limits [13, 36]) of C∗-algebras were studied by differ- ent names in the literature as follows: b∗-algebras (Allan [1], Apostol [2]), LMC∗-algebras (Lassner [31], Schmudgen [43]), l.m.c.C∗-algebras (Mallios [34]), locally C∗-algebras (Inoue [23], Fragoulopoulou [17]), generalized op- erator algebras (Weidner [46, 47]), F∗-algebras (Brooks [10]), σ-C∗-algebras (Arveson [4]), or pro-C∗-algebras (Voiculescu [44]). Next, we recall definitions and basic facts, where notations are slightly changed from [37]. Let (D, ≤) be a directed set [30], that is, D is a non- empty set and ≤ is a binary relation on D which satisfies the conditions: For all a, b, c ∈ D, we have a ≤ a; if a ≤ b and b ≤ c, then a ≤ c; if a, b ∈ D, then there exists c ∈ D such that a ≤ c and b ≤ c. Such ≤ is called a preorder [6]. We call ≤ an order in this paper for simplicity of description. For every concrete directed set (D, ≤) in this paper, the order ≤ satisfies the antisymmetric law, that is, if a ≤ b and b ≤ a, then a = b. Definition 1.7 Let (D, ≤) be a directed set. (i) A data {(Ad, ϕd,e) : d, e ∈ D} is an inverse system (or projective system [44]) of C∗-algebras if Ad is a C∗-algebra for each d ∈ D and ϕd,e is a ∗-homomorphism from Ae to Ad when d ≤ e such that ϕd,e ◦ ϕe,f = ϕd,f when d ≤ e ≤ f , and ϕd,d = idAd. (ii) The inverse limit (A, {πd}d∈D) of an inverse system {(Ad, ϕd,e) : d, e ∈ D} of C∗-algebras is a topological ∗-algebra A and a ∗-homomorphism πd from A to Ad such that the following conditions hold: (a) ϕd,e ◦ πe = πd when d ≤ e, 4 (b) for any ∗-algebra B and ∗-homomorphisms {ηd}d∈D, ηd : B → Ad, which satisfy ϕd,e ◦ ηe = ηd when d ≤ e, there exists a unique ∗-homomorphism ψ from B to A such that πd ◦ ψ = ηd for each d ∈ D, (c) the topology of A is the weakest topology such that every πd is continuous. (Ad, ϕd,e) or lim←−(Ad, ϕd,e), and πd In this case, A is written as lim←−D is called the canonical homomorphism (or projection [39], canonical mapping [6].) (iii) A pro-C∗-algebra A is a complete Hausdorff topological ∗-algebra over C whose topology is determined by its continuous C∗-seminorms in the sense that a net {aλ} converges to 0 if and only if p(aλ) → 0 for every continuous C∗-seminorm p on A. (iv) An inverse limit of C∗-algebras over a countable directed set is called a σ-C∗-algebra. From "1.2. Proposition" in [37], a C∗-algebra A is a pro-C∗-algebra if and only if it is the inverse limit of an inverse system of C∗-algebras. Any C∗- algebra is a pro-C∗-algebra. For any inverse system {(Ad, ϕd,e) : d, e ∈ D}, the inverse limit lim←−(Ad, ϕd,e) is given as the following subset of the product set Qd∈D Ad: {(xd) ∈ Yd∈D Ad : ϕd,e(xe) = xd for all d, e ∈ D such that d ≤ e}. (1.6) In general, a pro-C∗-algebra is not a C∗-algebra (see "1.3. EXAMPLE" in [44]), but the inverse limit of a special inverse system of C∗-algebras is also a C∗-algebra as follows. Fact 1.8 Let {(Ad, ϕd,e) : d, e ∈ D} be an inverse system of C∗-algebras. We write lim←−(Ad, ϕd,e) as lim←− Ad for simplicity of description. (i) If ϕd,e is injective for each d, e, then lim←− Ad is a C∗-algebra. (ii) If {An : n ∈ N} is a sequence of inclusions of C∗-algebras such that An ⊃ An+1 for each n ∈ N, then the inclusion map ιn : An+1 ֒→ An induces an inverse system over (N, ≤) such that lim←− An is ∗-isomorphic onto Tn≥1 An as a C∗-algebra. lim←− Ad is a unital C∗-algebra. (iii) If Ad is unital and simple, and fd,e is unital for each d, e ∈ D, then 5 (iv) If there exists the maximal element ω of D, then lim←− Ad Proof. Since an injective ∗-homomorphism is an isometry, (i) holds from Definition 1.7(ii)(c). (ii) and (iii) hold from (i). (iv) holds from (1.6). ∼= Aω. When an inverse limit of C∗-algebras is a C∗-algebra, it is not interesting as a pro-C∗-algebra, but it does not mean the triviality of the inverse limit as a C∗-algebra. Next, we consider the inverse limit of Cuntz algebras in general setting. (About inductive (or direct) limits of Cuntz algebras, see [20, 33, 40].) Since any Cuntz algebra is unital and simple, the inverse limit of any inverse system of Cuntz algebras by unital ∗-homomorphisms is a unital C∗-algebra from Fact 1.8(iii). By Corollary 1.5 and Fact 1.8(iv), the following holds. Fact 1.9 Let ( N, (cid:22)) be as in Corollary 1.5, and let {(On(d)+1, ϕd,e) : d, e ∈ D} be an inverse system of Cuntz algebras over a directed set (D, ≤) such that {n(d) ∈ N : d ∈ D}. We assume that ϕd,e is unital for each d, e. (i) The map F : D → N; d 7→ F (d) ≡ n(d) (1.7) is an ordered set homomorphism from (D, ≤) to ( N, (cid:22)). (ii) If there exists the maximal element ω of D, then lim←− Remark that F in (1.7) is not injective in general. There are many endo- morphisms of Cuntz algebras [7, 8, 29]. On(d)+1 Let ( N, (cid:22)) be as in Corollary 1.5. Define the order (cid:22)c on the set {On+1 : n ∈ N} of all Cuntz algebras as "A (cid:22)c B if and only if Hom(A, B) 6= ∅." Then ({On+1 : n ∈ N}, (cid:22)c) is anti-isomorphic onto ( N, (cid:22)) as an ordered set with respect to the mapping n 7→ On+1. ∼= On(ω)+1. 1.3 An inverse system of Cuntz algebras In this subsection, we construct an example of inverse system of Cuntz algebras as an answer to Question 1.6(i). For the directed set (N, (cid:22)) in (1.3), define the inverse system {(Rn, fn,m) : n, m ∈ N} of C∗-algebras as follows: For 2 ≤ n < ∞, let s(n) n denote the Cuntz generators of On, that is, (s(n) n )∗ = I. For convenience, rewrite On+1 as j = δijI for i, j = 1, . . . , n and s(n) 1 )∗ + · · · + s(n) 1 , . . . , s(n) 1 (s(n) )∗s(n) i n (s(n) Rn ≡ On+1 (n ∈ N). (1.8) 6 This notation is reasonable with respect to the K-theory of Cuntz alge- bras and Corollary 1.5. (About such a notation, see also [3], which is not related to the inclusions in Figure 1.2.) Remark that Rn is generated by s(n+1) n+1 by definition. When n (cid:22) m and n 6= m, define the ∗- 1 homomorphism fn,m from Rm to Rn by , . . . , s(n+1) fn,m(s(m+1) nl+i ) ≡ (s(n+1) n+1 )ls(n+1) i fn,m(s(m+1) m+1 ) ≡ (s(n+1) n+1 ) m n (cid:18) l = 0, 1, . . . , m i = 1, . . . , n n − 1, (cid:19) , (1.9)   where (s(n+1) n+1 )0 means the unit of Rn for convenience, and define fn,n as the identity map idRn on Rn. A graphical explanation of (1.9) will be given in § 3.2. Then the following holds. Theorem 1.10 Let Rn and fn,m be as in (1.8) and (1.9), respectively. (i) The data {(Rn, fn,m) : n, m ∈ N} is an inverse system of C∗-algebras over the directed set (N, (cid:22)), that is, the relation fn,m ◦fm,l = fn,l holds when n (cid:22) m (cid:22) l. (ii) Let {s(∞) 1 , s(∞) 2 , . . .} denote the Cuntz generators of O∞. For n ∈ N, define the embedding fn,∞ of O∞ into Rn by fn,∞(s(∞) ln+i) ≡ (s(n+1) n+1 )ls(n+1) i (i = 1, . . . , n, l ≥ 0). (1.10) Then {fn,∞ : n ∈ N} satisfies fn,m ◦ fm,∞ = fn,∞ when n (cid:22) m. (1.11) (iii) Every fn,m in (1.9) and (1.10) is irreducible, where we state that f ∈ Hom(A, B) is irreducible if f (A) ′ ∩ B = CI. We illustrate relations of maps in Theorem 1.10 as the following com- mutative diagrams where we assume n (cid:22) m: 7 Figure 1.11 fm,∞ ✯ (cid:8) O∞ Rm = Om+1 fn,m (cid:9) fn,∞ ❄ ❥ Rn = On+1 f1,m f1,n ❥ ✯ R1 = O2 Remark 1.12 Every fn,m in (1.9) and (1.10) is unital and injective, but not surjective when n 6= m. Such ∗-homomorphisms were given in [26, 28], hence they are not new, but their relations of inverse system are new. The essential part of Theorem 1.10(i) is the construction of formulas in (1.9). The proof is given by simple algebraic calculation. We make a point that Corollary 1.5 itself does not show the existence of any inverse system of Cuntz algebras over the directed set (N, (cid:22)). Inversely, it is interesting question whether the existence of such an inverse system of Cuntz algebras with unital ∗-homomorphisms is shown only from Corollary 1.5 and general theory without use of concrete construction of ∗-homomorphisms. For an example of fn,m in (1.9), the ∗-homomorphism f1,2 : R2(= O3) → R1(= O2) is given as follows: f1,2(s1) = s1, f1,2(s2) = s2s1, f1,2(s3) = s2s2 (1.12) where s1, s2 and s1, s2, s3 denote Cuntz generators of O2 and O3, respec- tively. A similar ∗-homomorphism was given by Cuntz's original paper ([11], p183). For the first time, the author knew (1.12) from Akira Asada who con- structed f1,2, and (1.9) is a generalization of (1.12). 1.4 Inverse limits of Cuntz algebras In this subsection, we show the inverse limit of the inverse system {(Rn, fn,m) : n, m ∈ N} in Theorem 1.10(i), which is the answer to Question 1.6(ii). The problem is solved in slightly generalized setting. Theorem 1.13 Let {(Rn, fn,m) : n, m ∈ N} be as in Theorem 1.10. For a directed subset Λ of (N, (cid:22)), let O(Λ) denote the inverse limit lim←−Λ Rn of 8 the subsystem {(Rn, fn,m) : n, m ∈ Λ} of the inverse system {(Rn, fn,m) : n, m ∈ N}: O(Λ) ≡ lim←− Λ Rn. (1.13) Then the following holds: (i) Assume that Λ is an infinite totally ordered subset of (N, (cid:22)) such that Λ = {n1, n2, . . .} and n1 (cid:22) n2 (cid:22) · · · . For {fn,∞ : n ∈ N} in (1.10), define the ∗-homomorphism ψΛ from O∞ to O(Λ) by ψΛ(x) ≡ (fn1,∞(x), fn2,∞(x), . . .) (x ∈ O∞) (1.14) where O(Λ) is identified with the standard form (1.6). Then ψΛ is a ∗-isomorphism such that πn ◦ ψΛ = fn,∞ (n ∈ Λ) (1.15) where πn denotes the canonical homomorphism from O(Λ) to Rn. (ii) For a directed subset Λ of (N, (cid:22)), the following holds: O(Λ) ∼=   ON +1 (∃N = max Λ), O∞ (otherwise). Especially, when Λ = N, we obtain lim←− Rn ∼= O∞. (1.16) (1.17) The proof of Theorem 1.13 will be given in § 2. The crucial part of the proof is the surjectivity of ψΛ in (1.14). From Theorem 1.13(i) for Λ = N, the data (O∞, {fn,∞}n∈N) is the inverse limit of {(Rn, fn,m) : n, m ∈ N} in the sense of Definition 1.7(ii). As a counterview of Theorem 1.13(ii), we can say that O∞ can be written as an inverse limit of On's. From this and Definition 1.7(ii)(b), the following corollary immediately holds. Corollary 1.14 Let {fn,m : n, m ∈ N} and {fn,∞ : n ∈ N} be as in (1.9) and (1.10), respectively. If a ∗-algebra B and ∗-homomorphisms {ηn}n∈N, ηn : B → On+1, which satisfy fn,m ◦ ηm = ηn when n (cid:22) m, there exists a unique ∗-homomorphism ψ from B to O∞ such that fn,∞ ◦ ψ = ηn for each n ∈ N. 9 This is an essentially new universal property of O∞. About other universal properties of O∞, see Chapter 7 of [42]. Remark 1.15 (i) Here we discuss the choice of notations. Both nota- tions On and Rn are useful in different situations. The standard no- tations O2, O3, . . . are used in the construction of C∗-bialgebra [27]. With respect to the standard notations of Cuntz algebras, the order (cid:22) in (1.3) can be rewritten as follows: A new order ≪ on the set N≥2 ≡ {2, 3, 4, . . .} is defined as "n ≪ m if n−1 (cid:22) m−1," or " m−1 n−1 ∈ N." Let gn,m ≡ fn−1,m−1 for n, m ∈ N≥2. Then {(On, gn,m) : n, m ∈ N≥2} is an inverse system over the directed set (N≥2, ≪) which is isomorphic to the inverse system {(Rn, fn,m) : n, m ∈ N} over (N, (cid:22)). By using (N≥2, ≪), the formula (1.1) makes sense. (ii) In "otherwise" in (1.16), we assume that Λ is neither a cofinal nor a totally ordered subset of (N, (cid:22)). (iii) We consider inverse systems of C∗-subalgebras of Cuntz algebras in Theorem 1.10. Let γ(n+1) and η(n+1) denote the U (1)-gauge action and the standard torus (=Tn+1)-action on On+1, respectively. Let An and Cn denote the fixed-point subalgebras of On+1 with respect to γ(n+1) and η(n+1), respectively: An ≡ (On+1)U (1), Cn ≡ (On+1)Tn+1 . (1.18) Then we see that fn,m(Am) 6⊂ An even if n (cid:22) m, but fn,m(Cm) ⊂ Cn. In this way, {(Cn, fn,mCm) : n, m ∈ N} is an inverse system of unital abelian C∗-algebras over (N, (cid:22)). Since fn,mCm is also injective and unital, lim←−(Cn, fn,mCm) is also a unital abelian C∗-algebra. (iv) As analogy with previous works of inductive limits [32, 33, 40], the classification of inverse limits of Cuntz algebras is an interesting prob- lem. In the case of inductive limits, classification theorems are given by using K-theory. On the other hand, it seems that K-theory is no use for the case of inverse limits, which will be shown in § 1.5. In order to consider the problem, concrete examples are not sufficient yet. We will show other example of inverse system which limit is not a Cuntz algebra in § 3.1. (v) In the proof of "3.11 COROLLARY" in [12], O∞ is written as an inductive limit of C∗-subalgebras of On's as follows: For n ≥ 1, let C ∗(s(n+1) ) denote the C∗-subalgebra of On+1 generated by , . . . , s(n+1) 1 n 10 n 1 , . . . , s(n+1) n+1 . Remark C ∗(s(n+1) without s(n+1) , . . . , s(n+1) s(n+1) 1 because K0(C ∗(s(n+1) n a C ∗-subalgebra of C ∗(s(n+2) i i = 1, . . . , n. Then the inductive system {(C ∗(s(n+1) 1 n ≥ 1} over the directed set (N, ≤) is obtained: ) 6∼= On )) ∼= Z. Then it is identified with for n+1 ) by ιn(s(n+1) i , . . . , s(n+1) , . . . , s(n+2) ) ≡ s(n+2) , . . . , s(n+1) n 1 1 C ∗(s(2) 1 ) ⊂ C ∗(s(3) 1 , s(4) and its inductive limit is isomorphic onto O∞: 2 ) ⊂ C ∗(s(4) 1 , s(3) 2 , s(4) 3 ) ⊂ · · · , lim−→ C ∗(s(n+1) 1 , . . . , s(n+1) n ) ∼= O∞. This is quite a contrast to our result. n ), ιn) : (1.19) (1.20) 1.5 Discontinuity of K0 We discuss the (dis-) continuity of K0-functor of C∗-algebras as an answer to Question 1.6(iii). Related to the question of the continuity of K0-functor with respect to the inverse limit, we could not find similar results in the In § 3 of [38], the inverse limit for standard textbooks [5, 24, 41, 45]. 1- representable K-theory is considered for σ-C∗-algebras as the Milnor lim←− sequence. However, it is assumed that all maps in the inverse system are surjective ("3.2 THEOREM", [38]). Hence it is no use for our example. 1.5.1 Profinite groups In order to discuss K-groups of inverse limits of Cuntz algebras, we recall basic examples of profinite group (especially, they are procyclic groups ([39], § 2.7)) as follows. Let (N, (cid:22)) be as in (1.3). For n, m ∈ N, if n (cid:22) m, then the natural projection from Z/mZ onto Z/nZ induces an inverse system {Z/nZ : n ∈ N} of finite cyclic groups (especially, they are rings) over It is well-known that the inverse limit Z of {Z/nZ : n ∈ N} is (N, (cid:22)). called the Prufer ring [35]: Z ≡ lim←− Z/nZ. (1.21) Remark Z 6∼= Z. For a fixed prime number p, the subset {pn : n ≥ 1} of N is a directed subset of (N, (cid:22)) which is not cofinal. For the subsystem {Z/pnZ : n ≥ 1} of {Z/nZ : n ∈ N}, the pro-p group Zp is defined as follows [39]: (1.22) Remark that Zp is identified with a uncountable proper subgroup of Z ([39], Exercise 2.1.8). Zp ≡ lim←− Z/pnZ. 11 1.5.2 K0-functor is not continuous with respect to the inverse limit It is well-known that the K0-functor of C∗-algebras is continuous with re- spect to the inductive limit as the following sense [5]: K0(lim−→ An) ∼= lim−→ K0(An). (1.23) Since lim−→ An is always a C∗-algebra for any inductive system, (1.23) holds for any inductive system of C∗-algebra. On the other hand, the inverse limit of C∗-algebra is not always a C∗-algebra. Hence (standard) K-groups can not be defined on a pro-C∗-algebra in general. (K-groups are generalized for σ-C∗-algebras [38], see also [13, 46, 47].) From (1.1), the K0-group of lim←− On is well-defined. Then the following holds: K0(lim←− On) ∼= K0(O∞) ∼= Z 6∼= Z = lim←− Z/nZ ∼= lim←− K0(On). (1.24) This shows that K0-functor is discontinuous with respect to the inverse limit even if the limit is a C∗-algebra. Since Z is the profinite completion of Z [35], On) coincides with lim←− K0(On). However the profinite completion of K0(lim←− the profinite completion of K0(lim←− Opn+1) ∼= K0(O∞) does not coincide with lim←− K0(Opn+1) ∼= Zp from Theorem 1.13(ii) and (1.22). Hence the profinite completion does not always recover the continuity of the K0-functor with respect to the inverse limit. In § 2, we will prove Theorem 1.10 and Theorem 1.13. In § 3, we will show examples. 2 Proofs of main theorems In this section, we prove main theorems in § 1. 2.1 Proof of Theorem 1.10 In this subsection, we prove Theorem 1.10. By simple calculation, both (i) and (ii) are directly verified from (1.9) and (1.10). In order to prove (iii), we recall a lemma. Lemma 2.1 (i) Let A and B be unital C∗-algebras and let ρ be a unital ∗-homomorphism from A to B. If B is simple and there exists an irreducible representation π of B such that π ◦ ρ is also irreducible, then ρ is irreducible. 12 (ii) Let {s(n) i } denote Cuntz generators of On for 2 ≤ n ≤ ∞. Fix i ∈ {1, . . . , n}. Then there exists a unique state ω on On such that ω(s(n) ) = 1. Furthermore, ω is pure. i (i) This holds from Proposition 3.1 in [28]. Proof. (ii) The uniqueness holds by Cuntz relations. Let (Hω, πω, Ωω) denote the GNS triple by ω. By definition, we see πω(s(n) )Ωω = Ωω. From [9, 15, 16], such a cyclic representation πω exists and is irreducible. Hence ω is pure. i Let ωn denote the state on Rn = On+1 such that ωn(s(n+1) 1 ) = 1. (2.1) From Lemma 2.1(ii), the GNS representation πn by ωn is irreducible. When n (cid:22) m, we can verify ωn ◦ fn,m = ωm because fn,m(s(m+1) . Hence πn ◦fn,m is unitarily equivalent to πm, and πn ◦fn,m is also irreducible. From this and Lemma 2.1(i), Theorem 1.10(iii) is proved. ) = s(n+1) 1 1 2.2 Proof of Theorem 1.13 In this subsection, we prove Theorem 1.13 except a certain equality of C∗- subalgebras, which will be proved in § 2.3. In order to reduce the problem, we show a lemma. A subset E of a directed set (D, ≤) is cofinal if {e ∈ D : e ≥ d} ∩ E 6= ∅ for any d ∈ D. In Ad for the subsystem {(Ad, ϕd,e) : this case, it is known that lim←−E d, e ∈ E} of the inverse system {(Ad, ϕd,e) : d, e ∈ D} over (D, ≤). Then the following lemma holds. ∼= lim←−D Ad Lemma 2.2 For any countable directed set (D, ≤), there exists a totally ordered cofinal subset D0 of D. If the maximal element ω of D exists, then let D0 ≡ {ω}. If not, Proof. let D = {x1, x2, . . .}, where we do not assume x1 ≤ x2 ≤ · · · . Then we can inductively construct a subsequence y1, y2, . . . of x1, x2, . . . as follows: Let y1 = x1. For n ≥ 2, there always exists z ∈ D such that yn−1 ≤ z and xn ≤ z. Choose such an element z and define yn ≡ z. Then D0 ≡ {y1, y2, . . .} is a totally ordered cofinal subset of D. For example, {n! : n ∈ N} is a totally ordered cofinal subset of (N, (cid:22)). 13 Proof of Theorem 1.13. (i) Here we will prove the statement except a cer- tain equality. From Fact 1.8(iii), O(Λ) is a C∗-algebra. By the standard construction in (1.6), O(Λ) is given as follows: Rnk : fnk,nl(xnl) = xnk for each k ≤ l}. {(xn1 , xn2, . . .) ∈ Yk≥1 Define the set {Qn : 1 ≤ n ≤ ∞} of C∗-subalgebras of R1 by (2.2) Qn ≡ f1,n(Rn) (1 ≤ n < ∞), Q∞ ≡ f1,∞(O∞). (2.3) Then we see that Q1 = R1 = O2, Qn ∼= Rn = On+1 when 1 ≤ n < ∞, and Qm ⊂ Qn when n (cid:22) m from Theorem 1.10(i). On the other hand, from (1.11), Q∞ ⊂ Qn for each n ≥ 1. Since n1 (cid:22) n2 (cid:22) · · · , we obtain the following unital inclusions O2 = Q1 ⊃ Qn1 ⊃ Qn2 ⊃ Qn3 ⊃ · · · ⊃ Q∞. (2.4) Define π1 ≡ f1,n1 ◦ πn1. Then we see that π1( O(Λ)) = \n∈Λ Qn. (2.5) Since π1 is injective, O(Λ) and Tn∈Λ Qn are ∗-isomorphic, and the map f1,∞ : O∞ → Tn∈Λ Qn is well-defined. following diagram is commutative: In consequence, we see that the Figure 2.3 O∞ ψΛ ✲ O(Λ) (cid:8) f1,∞ ❘ π1 ❄ Qn \n∈Λ In order to prove the bijectivity of ψΛ, it is sufficient to show the bijectivity of f1,∞. Since f1,∞ is an injective ∗-homomorphism, it is sufficient to show that Q∞ = \n∈Λ Qn. 14 (2.6) We will prove (2.6) in § 2.3. (ii) When there exists the maximal element of Λ, the statement holds from Fact 1.9(ii). Assume that Λ has no maximal element. In this case, it is sufficient to assume the condition in (i) for Λ by Lemma 2.2. Hence the statement holds from (i). 2.3 Proof of (2.6) From the proof of Theorem 1.13(i), the problem is reduced to the relation (2.6) among C∗-subalgebras {Qn : 1 ≤ n ≤ ∞} of Q1(= R1 = O2) in (2.3). In this subsection, we prove (2.6). 2.3.1 Inclusions of free subsemigroups of Q1 In this subsubsection, we consider free subsemigroups of Q1 and their re- lations. We rewrite the Cuntz generators of Q1 as t1, t2 here. Let S(X) denote the subsemigroup of Q1 generated by a subset X of Q1. Define subsemigroups Kn, Ln of Q1 as Kn ≡ S({tn 2 }) (1 ≤ n < ∞), L1 ≡ S({t1, t2}), Ln ≡ S({t1, t2t1, . . . , tn−1 2 t1, tn 2 }) (2 ≤ n < ∞), (2.7) L∞ ≡ S({t1, tm 2 t1 : m ≥ 1}).   For each n ≥ 1, Kn is abelian, and Ln is a (non-unital) free semigroup of rank n + 1 [22]. Both Kn and Ln are subsemigroups of Qn. If m (cid:22) n, then Kn ⊂ Km and Ln ⊂ Lm. For each n ≥ 1, Ln ⊃ L∞. Furthermore, we see that Kn = {tn 2 , t2n 2 , t3n 2 , . . .}, L∞ = {xt1 : x ∈ L1} = L1t1. (2.8) Hence Kn ∩ L∞ = ∅. Since L1 = L1t1 ⊔ L1t2 with respect to the ending of each word, the decomposition Ln = (Ln ∩ L1t1) ⊔ (Ln ∩ L1t2) holds. Let Yn ≡ {u, xu : x ∈ L∞, u ∈ Kn}. Then Ln ∩ L1t2 = Yn and Ln ∩ L1t1 = L∞. Hence the following decomposition into disjoint subsets holds: Ln = L∞ ⊔ Yn (2 ≤ n < ∞). (2.9) 15 2.3.2 Decomposition of algebras into linear subspaces Let Kn, Ln be as in (2.7). From definitions of Qn and {fn,m, fn,∞ : n, m ∈ N}, we see that Qn = C ∗hLni (1 ≤ n ≤ ∞) (2.10) where C ∗hXi denote the C∗-subalgebra of Q1 generated by a subset X of Q1. As closed linear subspaces of Q1, Qn's can be written as follows: Qn = Linh{xy∗ : x, y ∈ Ln}i (1 ≤ n < ∞), Q∞ = Linh{I, xy∗ : x, y ∈ L∞}i. (2.11)   Remark that vectors in each generating set in (2.11) are not always linearly independent because of Cuntz relations. Lemma 2.4 (i) For n ≥ 1, (t2)n(t∗ 2)n = I −Pn−1 k=0(t2)kt1t∗ 1(t∗ 2)k. (ii) Define closed linear subspaces Vn and V ∗ n of Qn by Vn ≡ Linh{u, xu, xuy∗ : x, y ∈ L∞, u ∈ Kn}i, Then Vn ⊂ Vm and V ∗ V ∗ n ∩ Q∞ = {0}. n ⊂ V ∗ m when m (cid:22) n, Vn ∩ V ∗ V ∗ n ≡ {x∗ : x ∈ Vn}. (2.12) n = Vn ∩ Q∞ = (iii) For u, v ∈ Kn and x, y ∈ L∞, xuv∗y∗ ∈ Q∞ ⊕ Vn ⊕ V ∗ n . (iv) For each n ≥ 1, the following decomposition of Qn into closed linear subspaces holds: Qn = Q∞ ⊕ Vn ⊕ V ∗ n . (2.13) (v) For Λ in Theorem 1.13(i), Tn∈Λ Vn = Tn∈Λ V ∗ n = {0}. (i) By the Cuntz relations of Q1 = O2, the statement holds. Proof. (ii) By definition, the statement holds. (iii) From (ii), Q∞ ⊕Vn ⊕V ∗ xuv∗y∗ ∈ Q∞ from (i) and (2.11). Furthermore, from (i), tn(l+k) 2 − tnl tnl 1(t∗ holds for the case u 6= v. (iv) From (ii), Qn ⊃ Q∞ ⊕ Vn ⊕ V ∗ space spanned by the set n makes sense as a subspace of Qn. If u = v, then 2 )∗ = 2)j) ∈ Vn ⊕ Q∞ for l ≥ 1. Hence the statement also n . Since Qn is the closure of the linear j=0 tj 2 (Pnk−1 2t1t∗ (tnk 2 Mn ≡ {x, x∗, xy∗ : x, y ∈ Ln} (1 ≤ n < ∞), (2.14) 16 it is sufficient to show that Mn is a subset of Q∞ ⊕ Vn ⊕ V ∗ Mn is decomposed into the disjoint union as follows: n . From (2.9), Mn = {x, x∗, xy∗ : x, y ∈ L∞ ⊔ Yn} = Mn,1 ⊔ Mn,2, Mn,1 ≡ {x, x∗, xy∗ : x, y ∈ L∞}, (2.15) Mn,2 ≡ {u, u∗, uv∗, xu∗, ux∗ : x ∈ L∞, u, v ∈ Yn}. Since Q∞ = LinhMn,1 ∪ {I}i, it is sufficient to show that Mn,2 is a subset of Q∞ ⊕ Vn ⊕ V ∗ n . By the definition of Yn in (2.9), Mn,2 = Mn,2,1 ⊔ Mn,2,2, Mn,2,1 ≡ {u, u∗, xu, (xu)∗ : x ∈ L∞, u ∈ Kn}, (2.16) Mn,2,2 ≡ {xuy∗, xu∗y∗, xuv∗y∗ : x, y ∈ L∞, u, v ∈ Kn}. Then Mn,2,1 ⊂ Vn ⊕ V ∗ Hence Mn,2 ⊂ Q∞ ⊕ Vn ⊕ V ∗ n . (v) By definition, n by definition. From (iii), Mn,2,2 ⊂ Q∞ ⊕ Vn ⊕ V ∗ n . Vn = LinhKni ⊕ LinhL∞ · Kni ⊕ LinhL∞ · Kn · L∗ ∞i (2.17) where L∗ ∞ ≡ {x∗ : x ∈ L∞}. By assumption, Λ is an infinite subset of N. Vn = \n∈Λ Hence Tn∈Λ Kn = ∅. From this, (2.17) and (i), \n∈Λ LinhL∞ · Kni ⊕ \n∈Λ In a similar way, we obtain Tn∈Λ V ∗ LinhKni ⊕ \n∈Λ LinhL∞ · Kn · L∗ ∞i = {0}. (2.18) n = {0}. Hence the statement holds. From Lemma 2.4(iv) and (v), \n∈Λ Qn = Q∞ ⊕ (\n∈Λ Vn) ⊕ (\n∈Λ V ∗ n ) = Q∞. (2.19) Hence (2.6) is proved. 3 Examples In order to explain theorems in § 1 more, we show examples in this section. 17 3.1 Other inverse system In this subsection, we show other example of inverse system of Cuntz alge- 1 , . . . , s(n) bras. Let s(n) n denote the Cuntz generators of On. Fix an integer r ≥ 2. Let rn ≡ r2n−1 and rewrite Orn as Ar,n ≡ Orn (n ≥ 1). (3.1) Then r2 Ar,n by n = rn+1 for n ≥ 1. Define the ∗-homomorphism qn from Ar,n+1 to qn(s(rn+1) rn(i−1)+j ) ≡ s(rn) i s(rn) j (i, j = 1, . . . , rn). (3.2) Then {(Ar,n, qn) : n ≥ 1} is an inverse system of Cuntz algebras over the directed set (N, ≤). Remark Ar,n = O(rn−1)+1 for n ≥ 1. Here we verify Fact 1.9(i) for the sequence {rn − 1 : n ≥ 1}. Define the map F from (N, ≤) to (N, (cid:22)) by F (n) ≡ rn − 1 (n ∈ N). (3.3) Then F (n) = rn − 1 (cid:22) (r2n−1 − 1)(r2n−1 the map F satisfies the statement in Fact 1.9(i). + 1) = r2n − 1 = F (n + 1). Exactly, Ar,n of {(Ar,n, qn) : n ∈ N} is ∗- Proposition 3.1 The inverse limit lim←−n isomorphic onto the uniformly hyperfinite algebra U HFr of the Glimm type {rl : l ≥ 1} [19]: Ar,n ∼= U HFr. lim←− n (3.4) Proof. Let γ denote the U (1)-gauge action on Or = Or1 = Ar,1. For l ∈ Z, define A(l) r,1 ≡ {x ∈ Ar,1 : γz(x) = zlx for all z ∈ U (1)}. (3.5) Rewrite the Cuntz generators of Or = Or1 as t1, . . . , tr. By identifying Ar,n with the C∗-subalgebra (q1 ◦ q2 ◦ · · · ◦ qn−1)(Ar,n) of Ar,1, Ar,n's are rewritten as follows: Ar,n = C ∗h{tJ : J ∈ {1, . . . , r}2n−1 }i (n ≥ 1) (3.6) where tJ ≡ tj1 · · · tjm for J = (j1, . . . , jm). Then we obtain the following unital (rapidly decreasing) inclusions: Ar,1 ⊃ Ar,2 ⊃ Ar,3 ⊃ · · · . Define A(l) r,n = Ar,n ∩ A(l) r,1 (l ∈ Z, n ≥ 1). (3.7) (3.8) 18 r,n for each n, l, and A(l) r,n is the closure of Linh{tJ t∗ , J − K = l}i. When l 6= 0, Tn≥1 A(l) r,n and A(0) K : r,n = {0} r,n+1 = A(0) r,n r,n+1 ⊂ A(l) Then A(l) J, K ∈ Sa≥1{1, . . . , r}a×2n−1 because A(l) for each n, lim←−n r,n = {0} if 2n−1 ∤ l. Since Ar,n = Ll∈Z A(l) ∼= U HFr. Ar,n ∼= Tn≥1 Ar,n = A(0) r,1 Since K0(U HFr) is the group Z(r∞) ⊂ Q of all rational numbers whose denominators divide the generalized integer r∞ (§ 7.5, [5]), we see that K0(lim←− Ar,n) ∼= K0(U HFr) ∼= Z(r∞), lim←− K0(Ar,n) = lim←− K0(Orn) ∼= lim←− Z/(rn − 1)Z. (3.9) The former is countable, but the latter is not. This case also shows that the K0-functor is discontinuous with respect to the inverse limit. 3.2 Illustrations of embeddings In this subsection, we illustrate embeddings in (1.9) by using decomposi- tions of Hilbert spaces. When On acts on a Hilbert space, by identifying a generator si with the range of si, the following illustration is helpful in understanding s1, . . . , sn: Figure 3.2 H si ↓ s1H · · · siH · · · snH Recall that R1 = O2, R2 = O3, R4 = O5. Then f1,2, f1,4, f2,4 in (1.9) are given as follows: f1,2 : R2 → R1; f1,2(s(3) 1 ) = s(2) 1 , f1,4 : R4 → R1; f1,4(s(5) 1 ) = s(2) 1 , f1,2(s(3) 2 ) = s(2) 2 s(2) 1 , f1,2(s(3) 3 ) = (s(2) 2 )2. f1,4(s(5) 2 ) = s(2) 2 s(2) 1 , 19 f1,4(s(5) 3 ) = (s(2) 2 )2s(2) 1 , (3.10) (3.11) f1,4(s(5) 4 ) = (s(2) 2 )3s(2) 1 , f1,4(s(5) 4 ) = (s(2) 2 )4. f2,4 : R4 → R2; f2,4(s(5) 1 ) = s(3) 1 , f2,4(s(5) 2 ) = s(3) 2 , f2,4(s(5) 3 ) = s(3) 3 s(3) 1 , f2,4(s(5) 4 ) = s(3) 3 s(3) 2 , f2,4(s(5) 5 ) = (s(3) 3 )2. (3.12) (3.13) (3.14) From these, we can directly verify the identity f1,2 ◦ f2,4 = f1,4. From Figure 3.2, inclusions O2 ⊃ O3 ⊃ O5 by embeddings f1,2 and f2,4 are illustrated as follows: Figure 3.3 s(2) 1 s(2) 2 s(2) 2 s(2) 1 s(2) 2 s(2) 2 O2 O3 O5 s(3) 1 s(3) 2 s(3) 3 s(3) 1 s(3) 3 s(3) 3 s(3) 2 s(3) 3 s(3) 3 s(5) 1 s(5) 2 s(5) 3 s(5) 4 s(5) 5 By using a more rough analogy, Figure 3.3 shows that an embedding is represented as a refinement of a partition of a unit interval in the real line R. Then relations of inverse system in Theorem 1.10(i) mean that the set of such refinements is a directed set. Acknowledgment The author would like to Akira Asada for his idea of embeddings among Cuntz algebras. References [1] Allan, G. R.: On a class of locally convex algebras. Proc. London Math. Soc. 17 (1967), 91 -- 114. 20 [2] Apostol, C.: b∗-algebras and their representations. J. London Math. Soc. 3 (1971), 30 -- 38. [3] Araki, H., Carey, A. L., Evans, D. E.: On On+1. J. Operator Theory 12 (1984), 247 -- 264. [4] Arveson, W.: The harmonic analysis of automorphism groups. Proc. Sympos. Pure Math. 38(1) (1982), 199 -- 269. [5] Blackadar, B.: K-theory for operator algebras 2nd ed., Cambridge Uni- versity Press, 1998. [6] Bourbaki, N.: Theory of sets, Springer, 1970. [7] Bratteli, O., Jorgensen, P. E. T., Price, G. L.: Endomorphisms of B(H), in "Quantization of nonlinear partial differential equation" (W. Arveson et al., Eds.), Proc. Sympos. Pure Math. 59 (1996), 93 -- 138. [8] Bratteli, O., Jorgensen, P. E. T.: Endomorphisms of B(H) II. Finitely correlated states on On. J. Funct. Anal. 145 (1997), 323 -- 373. [9] Bratteli, O., Jorgensen, P. E. T.: Iterated function systems and permu- tation representations of the Cuntz algebra. Memoirs Am. Math. Soc. 139(663) (1999), 1 -- 89. [10] Brooks, R. M.: On representing F ∗-algebras. Pacific J. Math. 39 (1971), 51 -- 69. [11] Cuntz, J.: Simple C ∗-algebras generated by isometries. Commun. Math. Phys. 57 (1977), 173 -- 185. [12] Cuntz, J.: K-theory for certain C ∗-algebras. Ann. Math. 113(2) (1981), 181 -- 197. [13] Cuntz, J.: Bivariant K- and cyclic theories. E.M. Friedlander, D.R. Grayson ed., Handbook of K-theory vol.2, Springer, 2005, pp. 655 -- 702. [14] Davidson, K. R.: C∗-algebra by example, American Mathematical So- ciety, 1996. [15] Davidson, K. R., Pitts, D. R.: The algebraic structure of non- commutative analytic Toeplitz algebras. Math. Ann. 311 (1998), 275 -- 303. 21 [16] Davidson, K. R., Pitts, D. R.: Invariant subspaces and hyper-reflexivity for free semigroup algebras. Proc. London Math. Soc. (3) 78 (1999), 401 -- 430. [17] Fragoulopoulou, M.: Spaces of representations and enveloping L.M.C. ∗-algebras. Pacific J. Math. 95(1) (1981), 61 -- 73. [18] Fragoulopoulou, M.: Topological algebras with involution, North- Holland, 2005. [19] Glimm, J.: On a certain class of operator algebras. Trans. Amer. Math. Soc. 95 (1960), 318 -- 340. [20] Goodearl, K. R., Pardo, E., Wehrung, F.: Semilattices of groups and inductive limits of Cuntz algebras. J. reine angew. Math. 588 (2005), 1 -- 25. [21] Hardy, G. H., Wright, E. M.: An introduction to theory of numbers, 6th ed., Oxford University Press, 2008. [22] Howie, J. M.: Fundamentals of semigroup theory, Oxford Science Pub- lications, 1995. [23] Inoue, A.: Locally C∗-algebras. Mem. Fac. Sci. Kyushu Univ. 25 (1971), 197 -- 235. [24] Jensen, K. K., Thomsen, K.: Elements of KK-theory, Birkhauser Boston, 1991. [25] Joit¸a, M.: Frames of multipliers in tensor products of Hilbert modules over pro-C ∗-algebras. J. Math. Anal. Appl. 367 (2010), 522 -- 534. [26] Kawamura, K.: Polynomial embedding of Cuntz-Krieger algebra into Cuntz algebra. preprint RIMS-1391 (2003). [27] Kawamura, K.: C∗-bialgebra defined by the direct sum of Cuntz alge- bras. J. Algebra 319 (2008), 3935 -- 3959. [28] Kawamura, K.: Universal algebra of sectors. Int. J. Alg. Comput. 19(3) (2009), 347 -- 371. [29] Kawamura, K.: A generalization of de Bruijn graphs and classification of endomorphisms of Cuntz algebras by graph invariants. Semigroup Forum 81 (2010), 405 -- 423. 22 [30] Kelley, J. L.: General topology, Springer-Verlag New York Heidelberg Berlin, 1955. [31] Lassner, G.: Topological algebra of operators. Rep. Math. Phys. 3 (1972), 279 -- 293. [32] Li, L.: Classification of simple C ∗-algebras: Inductive limits of matrix algebras over trees, Memoirs Am. Math. Soc. 605 (1997), 1 -- 123. [33] Lin, H., Phillips, N. C.: Classification of direct limits of even Cuntz- circle algebras. Memoirs Am. Math. Soc. 565 (1995), 1 -- 116. [34] Mallios, A.: Hermitian K-theory over topological ∗-algebras. J. Math. Anal. Appl. 106 (1985), 454 -- 539. [35] Neukirch, J.: Algebraic number theory, Springer, 1999. [36] Palmer, T. W.: Banach algebras and the general theory of ∗-algebras vol.I, Cambridge University Press, 1994. [37] Phillips, N. C.: Inverse limits of C ∗-algebras. J. Operator Theory 19 (1988), 159 -- 195. [38] Phillips, N. C.: Representable K-theory for σ-C ∗-algebras. K-theory 3 (1989), 441 -- 478. [39] Ribes, L., Zalesskii, P.: Profinite groups, Springer, 2000. [40] Rørdam, M.: Classification of inductive limits of Cuntz algebras. J. reine angew. Math. 440 (1993), 175 -- 200. [41] Rørdam, M., Larson, F., Laustsen, N. J.: An introduction to K-theory for C ∗-algebras, Cambridge University Press, 2000. [42] Rørdam, M., Stormer, E.: Classification of nuclear C∗-algebras, En- tropy in operator algebras, Operator algebras and non-commutative geometry VII, Encyclopedia Math. Sci., 126, Springer, 2001. [43] Schmudgen, K.: Uber LMC∗-algebras. Math. Nachr. 68 (1975), 167 -- 182. [44] Voiculescu, D.: Dual algebraic structures on operator algebras related to free products. J. Operator Theory 17 (1987), 85 -- 98. [45] Wegge-Olsen, N. E.: K-theory and C∗-algebras, Oxford Science Publi- cations, 1993. 23 [46] Weidner, J.: KK-groups for generalized operator algebras. I. K-Theory 3 (1989), 57 -- 77. [47] Weidner, J.: KK-groups for generalized operator algebras. II. K- Theory 3 (1989), 79 -- 98. 24
1202.5977	2	1202	2012-03-09T16:29:30	Inverse semigroup C*-algebras associated with left cancellative semigroups	[ "math.OA" ]	To each discrete left cancellative semigroup $S$ one may associate a certain inverse semigroup $I_l(S)$, often called the left inverse hull of $S$. We show how the full and the reduced C-algebras of $I_l(S)$ are related to the full and reduced semigroup C-algebras for $S$ recently introduced by Xin Li, and give conditions ensuring that these algebras are isomorphic. Our picture provides an enhanced understanding of Li's algebras.	math.OA	math	INVERSE SEMIGROUP C ∗-ALGEBRAS ASSOCIATED WITH LEFT CANCELLATIVE SEMIGROUPS MAGNUS DAHLER NORLING Abstract. To each discrete left cancellative semigroup S one may as- sociate a certain inverse semigroup Il(S), often called the left inverse hull of S. We show how the full and the reduced C ∗-algebras of Il(S) are related to the full and reduced semigroup C ∗-algebras for S recently introduced by Xin Li, and give conditions ensuring that these algebras are isomorphic. Our picture provides an enhanced understanding of Li's algebras. Contents Introduction 1. Acknowledgements 2. Semigroups 2.1. Semigroups and algebraic orders 2.2. 2.3. The semilattice J(S), Clifford's condition and independence of Inverse semigroups constructible right ideals. 3. C ∗-theory 3.1. The C ∗-algebras of an inverse semigroup 3.2. The left regular representation and the left inverse hull of a left cancellative semigroup 3.3. Li's constructions of full C ∗-algebras for a left cancellative semigroup 3.4. Left reversible semigroups, left amenability and functoriality 3.5. Amenability and weak containment when S embeds into a group. References 1 3 3 3 5 6 10 10 14 20 23 27 30 1. Introduction In [17], Xin Li proposed a construction for the full C ∗-algebra C ∗(S) of a discrete left cancellative semigroup S. For a semigroup S that embeds into a group he also constructs a related C ∗-algebra called C ∗ s (S). The reason one considers left cancellative semigroups is that these are the semigroups that can be faithfully represented as semigroups of isometries on Hilbert 2010 Mathematics Subject Classification. Primary 46L05; Secondary 20M18, 20M99, 43A07, 06A12. Key words and phrases. C ∗-algebras, left cancellative semigroups, inverse semigroups, semilattices, amenability. 1 2 MAGNUS DAHLER NORLING spaces. For instance, one can represent S on ℓ2(S) by isometries in this case. This representation is called the left regular representation of S and generates what is called the Toeplitz algebra or reduced C ∗-algebra of S, denoted C ∗ r (S). (One could of course consider right cancellative semigroups instead). Murphy had previously constructed C ∗-algebras of left cancellative semi- groups, but these turned out to be very large. For instance, his C ∗-algebra associated to (Z+)2 is non-nuclear [21]. (See Li's article for more references). Li adds a few extra restrictions that make the algebras behave better. Espe- cially he shows that C ∗(S) generalizes two important types of C ∗-algebras: Nica's C ∗-algebras for quasilattice ordered groups from [22], and the Toeplitz algebras associated with the ring of integers in a number field [7]. s (S) and C ∗ Li also shows that a cancellative left reversible1 semigroup S is left amenable if and only if C ∗ r (S) are canonically isomorphic, but only given that the constructible right ideals of S satisfy a certain technical require- ment called independence. Note that Li's proof uses that left reversibility of a cancellative semigroup S implies that S embeds into a group and that there exists a character on C ∗ s (S). Let I(S) be the inverse semigroup of all partial bijections on S. For each s ∈ S, let λs : S → sS be given by λs(t) = st. Since S is left cancellative, each λs is a bijection. The set {λs}s∈S generates an inverse subsemigroup Il(S) ⊂ I(S) called the left inverse hull of S. We show that Il(S) is isomor- phic to an inverse semigroup V (S) of partial isometries generating C ∗ r (S). By considering the full and reduced C ∗-algebras of Il(S) as for instance defined in Paterson's book [27] we get surjective ∗-homomorphisms −→ C ∗ Λ0−→ C ∗ −→ C ∗ r,0(Il(S)) 0 (Il(S)) C ∗(S) r (S) h η The composition of these is the canonical ∗-homomorphism C ∗(S) → C ∗ r (S). The question of whether this is an isomorphism splits into three separate problems. When S embeds into a group, we get the decomposition C ∗(S) πs−→ C ∗ s (S) In particular, C ∗ s (S) and C ∗ ≃ 0 (Il(S)) −→ C ∗ 0 (Il(S)) are canonically isomorphic. r,0(Il(S)) Λ0−→ C ∗ h −→ C ∗ r (S) A semigroup S is said to satisfy Clifford's condition if for all s, t ∈ S, either sS ∩ tS = ∅ or sS ∩ tS = rS for some r ∈ S. Any semigroup that is the positive cone in one of Nica's quasilattice ordered groups satisfies Clifford's condition. The ax+b-semigroup over an integral domain R satisfies Clifford's condition if and only if every pair of elements in R has a least common multiple. If S satisfies Clifford's condition, η is an isomorphism and the constructible right ideals of S are independent. If S is cancellative and satisfies Clifford's condition, or if S embeds into a group and the constructible right ideals of S are independent, then h is an isomorphism. Using Milan's work [19] on weak containment for inverse semigroups we show that when S embeds into a group G, Λ0 is an isomorphism if and only if a certain Fell bundle over G associated to Il(S) is amenable. In the special 1A semigroup S is left reversible if for any s, t ∈ S, sS ∩ tS 6= ∅. This is also called the Ore condition. C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 3 case when S is left reversible, Λ0 is an isomorphism if and only if S is left amenable if and only if C ∗ 0 (Il(S)) is nuclear. In the first part of the article we recall the algebraic theory of semigroups and inverse semigroups, and also look at an algebraic partial order and see how it is related to Nica's quasilattice ordered groups. We show that many of the properties of the positive cone in these groups can be defined in a more general context and remark that the algebraic order is not essential for the theory to work. In the second part, we introduce the C ∗-algebras associated to S and Il(S), and prove the above stated results. In addition, we show that our construction generalizes a method used by Nica in [23] to construct the C ∗- algebra of a quasilattice ordered group from a certain inverse semigroup called a Toeplitz inverse semigroup. We also prove a functoriality result for the construction S 7→ G(S) when S is left reversible. Here G(S) is the maximal group homomorphic image of Il(S). We use this to show that the construction S 7→ C ∗(Il(S)) is functorial for homomorphisms into groups when S is left reversible. The construction S 7→ G(S) originates from Rees' proof of Ore's Theorem: that all cancellative left reversible semigroups are group embeddable. An account of this theorem can be found in vol I, p. 35 of [6] or in ch. 2.4 of [16]. Acknowledgements. We wish to thank Erik Bedos for fruitful discussions and for providing us with many helpful references. Thanks also to Mark Lawson for answering our questions about semigroups satisfying Clifford's condition. 2. Semigroups 2.1. Semigroups and algebraic orders. There are many sources on the algebraic theory of semigroups. See for instance [6] or [18] and the references therein. Definition 2.1.1 A semigroup is a set S together with a associative binary operation · : S × S → S written (s, t) 7→ st and an identity element2 1 ∈ S. That is, for all s, r, t ∈ S, s(rt) = (sr)t and 1s = s1 = s. Sometimes we write 1 = 1S. If S has an element z ∈ S such that zs = sz = z for all s ∈ S, we will write z = 0 = 0S. If S is a semigroup, define S0 = S if S already has a 0 element, and otherwise let S0 be the semigroup S ∪ {0} with extended multiplication rule s0 = 0s = 0 for all s ∈ S0. This choice of notation can be confusing for instance in the case of (Z+, +) where we have 1Z+ = 0, and where Z+ does not have an element 0Z+ in the sense of the above definition, but the notation is otherwise very convenient. (In our notation, Z+ denotes {0, 1, 2, . . .}, while N denotes {1, 2, . . .}). Definition 2.1.2 A homomorphism between semigroups S, S′ is a function f : S → S′ such that for all s, t ∈ S, f (st) = f (s)f (t) and f (1S) = 1S ′. The 2Usually, semigroups are not required to have identities, and semigroups with identities are called monoids. We will however only talk about monoids in this article, and we prefer to call them semigroups. 4 MAGNUS DAHLER NORLING homomorphism f is a 0-homomorphism if in addition f (0S) = 0S ′ (and this term is only defined for homomorphisms between semigroups with zeroes). Definition 2.1.3 A semigroup S is left cancellative if for every s, r, t ∈ S, sr = st implies r = t. Equivalently, for every s ∈ S, the map λs : S → sS given by λs(t) = st is bijective. In a left cancellative semigroup, if ss′ = 1, then ss′s = 1s = s1, so s′s = 1, that is every element with a left (or right) inverse is invertible. One can similarly define right cancellativity. S is cancellative if it is both left and right cancellative. Definition 2.1.4 A congruence on a semigroup S is an equivalence relation ∼ such that for all s, t, r,∈ S, s ∼ t implies sr ∼ tr and rs ∼ rt. One can show that S/ ∼ is again a semigroup and that there is a homomorphism S → S/ ∼ sending elements to equivalence classes. In fact, the homomorphism theorems for semigroups say that every surjective homomorphism can be constructed in this way. Definition 2.1.5 A subset X ⊂ S is a right ideal if for all t ∈ X and s ∈ S, ts ∈ X. For X ⊂ S and s ∈ S, define s−1(X) = {t : st ∈ X} and sX = {st : t ∈ X}. For simplicity, we will sometimes write s−1X for s−1(X). If X ⊂ S is a right ideal, then so are sX and s−1X. The right ideals on the form sS are called the principal right ideals of S. Let (cid:22) be the relation on S given by s (cid:22) t if there exists an r ∈ S such that s = tr. This relation is reflexive and transitive, so it gives a preorder on S. If it is antisymmetric, then it is a partial order called the algebraic order on S and we say that S is algebraically ordered. Note that (cid:22) is often written with the opposite symbol (cid:23) or ≥ (such as in Nica's work [22]), but this is just a matter of convenience. For instance we have 5 (cid:22) 4 in (Z+, +) with our notation. For s, t ∈ S, s (cid:22) t is easily seen to be equivalent to s ∈ tS and sS ⊂ tS. It is also equivalent to t−1({s}) 6= ∅, and if S ⊂ G where G is a group, it is equivalent to t−1s ∈ S. Note that 1S is a maximal element for (cid:22) and that if 0S exists it is a minimal element. Lemma 2.1.6 Let S be a left cancellative semigroup. Then S is alge- braically ordered if and only if 1 is the only invertible element in S. Proof. Suppose rS = tS for some r, t ∈ S. Then there are s, s′ ∈ S such that rs = t and ts′ = r. So ts′s = t. By left cancellation with t, this gives us s′s = 1. Then s has a left inverse, so it is invertible since S was left cancellative. If 1 is the only invertible element in S, s = 1 and this implies r = t. On the other hand, suppose there are s, s′ ∈ S with s′s = 1. Then s′sS ⊂ s′S ⊂ S = s′sS, so if S is algebraically ordered, s′ = 1. (cid:3) For instance when S is a subsemigroup of a group G, S is algebraically ordered if and only if S ∩ S−1 = {1}. C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 5 2.2. Inverse semigroups. Inverse semigroups are a large topic. See for instance [16] or [27] and references therein. In this section we will just give a short overview of the main concepts that we need. Definition 2.2.1 A semigroup P is an inverse semigroup if for every p ∈ P , there exists a unique element p∗ ∈ P such that pp∗p = p and p∗pp∗ = p∗. It follows from this uniqueness property of the ∗-operation that for any semigroup homomorphism f : P → Q between inverse semigroups, f (p∗) = f (p)∗ for any p ∈ P . Let L be the set of idempotents in the inverse semigroup P . Then L = {p∗p : p ∈ P} = {pp∗ : p ∈ P}. One can show that L is a commutative subsemigroup of P , so L is what is called a semilattice. Definition 2.2.2 A semilattice is a commutative semigroup where every element is idempotent. Lemma 2.2.3 Let L be a semilattice, and let a, b ∈ L. Then a (cid:22) b if and only if ba = a. Hence (cid:22) is a partial order on L. Proof. If a = ba, then a ∈ bL, so a (cid:22) b. Suppose a ∈ bL, so a = bc for some c ∈ L. Then a = aa = bca. So bca = bbca = ba, which implies a = ba. If a (cid:22) b and b (cid:22) a, then a = ba = ab = b. So (cid:22) is a partial order. (cid:3) It also follows that for a, b ∈ L, ab is the greatest lower bound of a and b. On the other hand, if L is a partially ordered set where any finite subset has a unique greatest lower bound and one defines ab to be the greatest lower bound of {a, b}, then L is a semilattice with the product (a, b) 7→ ab. We will later study partially ordered semigroups S, and for this it is useful to let s∧ t mean the greatest lower bound of s and t if it exists, while st means the already existing semigroup product of s and t. These two products only coincide if S is a semilattice. Remark 2.2.4 There exists a partial order defined on inverse semigroups called the natural partial order. It does not in general coincide with what we have called the algebraic order. We will not use the natural partial order explicitly in this paper. The perhaps most important example of an inverse semigroup is the semi- group I(X) of all partially defined bijective maps on some set X. By a partially defined bijective map on X, we mean a bijective function f : dom(f ) → ran(f ) where dom(f ) and ran(f ) are subsets of X. The prod- uct f g of f, g ∈ I(X) is defined such that dom(f g) = g−1(dom(f )) and f g(x) = f (g(x)) for all x ∈ dom(f g). Note that this product can re- sult in the empty function, which acts as a 0 for I(X). The ∗-operation is given by function inversion. For any f ∈ I(X), f ∗f = idom(f ) where idom(f ) : dom(f ) → dom(f ) is the identity map. The Wagner-Preston Theorem states that any inverse semigroup P can be faithfully represented as a subsemigroup of I(P ) as follows: Let τ : P → I(P ) be given such that for p ∈ P , dom(τ (p)) = {q ∈ P : p∗pq = q}, and define τ (p)(q) = pq for all q ∈ dom(τ (p)). Another important class of inverse semigroups are semigroups of partial isometries in a C ∗-algebra. Note that in general the product of two partial 6 MAGNUS DAHLER NORLING isometries does not have to be a partial isometry. Two partial isometries can be part of the same inverse semigroup if and only if their initial and final projections commute. The following concepts are very important in the theory of inverse semi- groups. Definition 2.2.5 An inverse semigroup P is E-unitary if for every p, q ∈ P , pq = q implies p ∈ L. It is E∗-unitary (also called 0-E-unitary) if for every p, q ∈ P , pq = q and q 6= 0 implies p ∈ L. Note that if P is an E-unitary inverse semigroup with 0, then it is a semilattice. Note also that we can assume without loss of generality that the q in either defintion is idempotent if we want to. Multiply the equation pq = q on the right with q∗. This gives us pqq∗ = qq∗ where qq∗ is idempotent. Recall that for any semigroup S, S0 = S if S already has a 0 element, and otherwise S0 is the semigroup S ∪ {0} with extended multiplication rule s0 = 0s = 0 for all s ∈ S0. Definition 2.2.6 A grading of the inverse semigroup P is a map ϕ : P 0 → G0, where G is a group, such that ϕ−1({0}) = {0} and for all p, q ∈ P , ϕ(pq) = ϕ(p)ϕ(q) as long as pq 6= 0. P is strongly E∗-unitary if it has a grading ϕ such that ϕ−1({1G}) = L \ {0}. Such a grading is sometimes said to be idempotent pure. Note that if ϕ : P 0 → G0 is a grading of P , L \ {0} ⊂ ϕ−1({1G}) always. Note also that if P is strongly E∗-unitary, then it is E∗-unitary. It turns out that if P does not have a 0, all these concepts are equivalent. Definition 2.2.7 Define a relation ∼ on P by p ∼ q if pr = qr for some r ∈ P (if and only if pr = qr for some r ∈ L). Then ∼ is a congruence, and P/ ∼ is a group denoted G(P ). Let αP : P → G(P ) be the quotient homomorphism. Then P is E-unitary if and only if α−1 P (1G(P )) = L. G(P ) is often called the maximal group homomorphic image of P . We will need the following lemma later: Lemma 2.2.8 Let f : P → Q be a surjective homomorphism between in- verse semigroups. Let L(P ) and L(Q) denote the respective semilattices of idempotents in P and Q. Suppose the restriction of f to L(P ) is an isomor- phism onto L(Q). Then f is an isomorphism if and only if f −1(L(Q)) = L(P ) Proof. The only if part is trivial. Suppose f −1(L(Q)) = L(P ). Let p, q ∈ P with f (p) = f (q). Then f (pq∗) = f (qq∗) ∈ L(Q), so by assumption pq∗ is idempotent. Since f is an isomorphism restricted to L(P ), pq∗ = qq∗, so q∗pq∗ = q∗. Similarly, f (q∗) = f (p∗), so q∗p = p∗p, and pq∗p = p. Thus p = q by the uniqueness property for these relations in an inverse semigroup. (cid:3) 2.3. The semilattice J(S), Clifford's condition and independence of constructible right ideals. We will be interested in the semilattice J(S) C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 7 of constructible right ideals in the left cancellative semigroup S given by We will actually see in Lemma 3.2.1 that t−1 j1 sj1 ··· t−1 jnj N \j=1 J(S) =  J(S) =(cid:8)t−1 sjnj S : N, nj ∈ N, sjk, tjk ∈ S  1 s1 ··· t−1 n snS : n ∈ N, si, ti ∈ S(cid:9) Here the semilattice product on J(S) is given by set intersection. To mo- tivate this study, we can reveal that J(S) is isomorphic to a semilattice of projections generating the diagonal subalgebra of the C ∗-algebra of the left regular representation of S (also called the Toeplitz algebra of S or C ∗ r (S)). It is also the semilattice of idempotents in the left inverse hull of S. We will establish these facts later. This semilattice plays an important part in Li's theory [17]. Li's J is the same as our J(S) ∪ {∅} ≃ J(S)0. Lemma 2.3.1 Let S be an algebraically ordered semigroup and let s, t ∈ S. If sS ∩ tS = rS for some r ∈ S, then s ∧ t exists and equals r. Conversely, if s ∧ t exists, then (s ∧ t)S = sS ∩ tS. Proof. First, suppose r′ (cid:22) s, t. Then r′S ⊂ sS ∩ tS = rS, so r′ (cid:22) r, and therefore r is the greatest lower bound of s and t, i.e. r = s ∧ t. Next, if s∧ t exists, then by definition (s∧ t)S ⊂ sS ∩ tS. Let r ∈ sS ∩ tS. Then r (cid:22) s, t, so r (cid:22) s ∧ t, and r ∈ (s ∧ t)S, so (s ∧ t)S = sS ∩ tS. Lemma 2.3.2 Let S be a semigroup, and let s1, . . . , sn ∈ S. If Sn rS for some r ∈ S, then rS = siS for at least one 1 ≤ i ≤ n. Proof. rS ⊂ Sn r ∈ sjS for some j. Then rS ⊂ sjS ⊂Sn Definition 2.3.3 We say that a semigroup S satisfies Clifford's condition 3 if for any s, t ∈ S, sS∩tS = ∅, or there exists an r ∈ S such that sS∩tS = rS. For instance, all free or free abelian semigroups satisfy Clifford's condition. i=1 snS is equivalent to r ∈ Sn i=1 snS which implies that (cid:3) i=1 siS = rS, so rS = sjS. i=1 siS = (cid:3) We will see more examples below. Definition 2.3.4 Following Li [17], we say that J(S) is independent or that the constructible right ideals of S are independent if for any X1, . . . , Xn, Y ∈ J(S), Sn i=1 Xn = Y implies that Xi = Y for at least one 1 ≤ i ≤ n. Proposition 2.3.5 Let S be a left cancellative semigroup. The following two conditions are equivalent. (i) S satisfies Clifford's condition (ii) For every s, t ∈ S with t−1(sS) nonempty, there is some r ∈ S such that t−1(sS) = rS. 3This is not the same concept as a Clifford semigroup. Clifford's condition is a term coined by Mark Lawson because it plays an important role in the construction of 0-bisimple inverse semigroups, and Clifford was the first to use this in [5]. See also [15] for more on this. 8 MAGNUS DAHLER NORLING These conditions impliy that J(S) ∪ {∅} = {sS : s ∈ S} ∪ {∅} and that J(S) is independent. If S is algebraically ordered, then (i) is equivalent to the following statement. (iii) Every pair of elements in S that have a common lower bound have a greatest lower bound. This implies that when S is an algebraically ordered semigroup satisfying Clifford's condition, (S0,∧) is a semilattice and is isomorphic as a semilattice to J(S) ∪ {∅} ≃ J(S)0. Proof. (i)⇒(ii): Since S is left cancellative, then for any X ⊂ S we have tt−1(X) = tS ∩ X and t−1(tX) = X. If t−1(sS) is nonempty, then so is tt−1(sS) = sS ∩ tS. Let q ∈ S be such that qS = sS ∩ tS. Since q ∈ tS, t−1({q}) is nonempty and contains a unique element r since S was left cancellative. Now we have rS = (t−1{q})S = {uv : u, v ∈ S, tu = q} = t−1{tuv : u, v ∈ S, tu = q} = t−1{qv : v ∈ S} = t−1(qS) = t−1(sS ∩ tS) = t−1(sS) gives that rS = siS for at least one i. (ii)⇒(i): If tS ∩ sS is nonempty, then so is tt−1(sS) and t−1(sS). By assumption, t−1(sS) = rS, so tS ∩ sS = (tr)S. That (i)+(ii) implies J(S)∪{∅} = {sS : s ∈ S}∪{∅} is a simple induction proof. That this again implies that J(S) is independent follows from Lemma 2.3.2: If Sn i=1 siS = Y ∈ J(S), Y = rS for some r ∈ S, and Lemma 2.3.2 (i)⇔(iii): Let s, t ∈ S. Then sS ∩ tS 6= ∅ if and only if there is some r ∈ S such that r (cid:22) s, t if and only if s and t have a common lower bound. By Lemma 2.3.1 s and t have a greatest lower bound s ∧ t if and only if sS ∩ tS = (s ∧ t)S. By going to S0, we have sS0∩ tS0 = {0} = 0S0 if and only if sS ∩ tS = ∅. Otherwise sS0 ∩ tS0 = rS0 for some r ∈ S. The isomorphism from (S0,∧) to J(S) ∪ {∅} is then constructed by sending s to sS for s ∈ S and 0 to ∅. This is injective since S was algebraically ordered. (cid:3) Definition 2.3.6 Let G be a group and S ⊂ G a subsemigroup. If S is algebraically ordered and generates G, it induces a partial order on all of G by g ≤ h iff g−1h ∈ S. Nica [22] calls (G, S) for quasilattice ordered if in addition any finite family of elements in G that have a common upper bound in S has a least common upper bound in S. S is called the positive cone in (G, S). Note that when restricted to S, ≤ is the same as our (cid:23). This shows that if S is a positive cone in a quasilattice orderd group, any pair in S that have a common lower bound in S with respect to (cid:22) have a greatest lower bound in S. So S satisfies Clifford's condition by Proposition 2.3.5 and it follows that J(S) is independent. Note that Li proves in [17] that the positive cones of the quasilattice ordered groups have independent constructible right ideals. C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 9 We can give a description of when the ax + b semigroup over an integral domain R satisfies Clifford's condition. The ax+b semigroup over R, denoted R⋊R× is defined to be the set R×R× with product (b, a)(d, c) = (b+ad, ac). Here R× = R \ {0}. The reason one considers integral domains is that the ax + b semigroups over these are left cancellative. Consider first the multiplicative semigroup (R×,·). This is a semigroup since R has no zero divisors. We see that for a, b ∈ R×, a (cid:23) b if and only if a divides b. Definition 2.3.7 A common multiple of a, b ∈ R× is an element c of R× that is divided by a and b. A least common multiple of a and b is a common multiple c such that if c′ is a common multiple of a and b, then c divides c′. It follows by a similar argument to that in Lemma 2.3.1 that aR×∩bR× = cR× if and only if c is a least common multiple of a and b. Note that since R is commutative, ab ∈ aR× ∩ bR× 6= ∅. So R× satisfies Clifford's condition if and only if every pair in R× has a least common multiple (see also Theorem 2.1 in [4]). Such an integral domain R is often called a GCD domain because one can show that every pair has a greatest common divisor if and only if every pair has a least common multiple. See [4] for a detailed discussion of GCD domains. They are also discussed in [2] where they are called pseudo- Bezout domains. The next lemma is stated without proof in Li's article, but we include it for completeness. Lemma 2.3.8 Let R be a ring. For any subrings I, J ⊂ R and b, d ∈ R, either (b+I)∩(d+J) = ∅ or there is some x ∈ R such that (b+I)∩(d+J) = x + I ∩ J. Proof. Suppose (b + I) ∩ (d + J) 6= ∅. Then there are y ∈ I and z ∈ J such that b+y = d+z. Write x = b+y = d+z. Then x+I = b+y +I = b+I and x+J = d+z+J = d+J. So (b+I)∩(d+J) = (x+I)∩(x+J) = x+I∩J. (cid:3) Proposition 2.3.9 Let R be an integral domain. Then R ⋊ R× satisfies Clifford's condition if and only if R is a GCD domain. Proof. We show that R ⋊ R× satisfies Clifford's condition if and only if R× satisfies Clifford's condition. Suppose R× satisfies Clifford's condition. Note that since R is a commutative ring, aR is an ideal of R for every a ∈ R. Let a, b, c, d ∈ R. Then (b, a)R ⋊ R× ∩ (d, c)R ⋊ R× = [(b + aR) × aR×] ∩ [(d + cR) × cR×] = [(b + aR) ∩ (d + cR)] × [aR× ∩ cR×] If this set is nonempty, (b + aR) ∩ (d + cR) is nonempty, so by the previous lemma there exists some x ∈ R satisfying (b + aR)∩ (d + cR) = x + aR∩ cR. Moreover, aR× ∩ cR× 6= ∅, so since R× satisfies Clifford's condition, there is some y ∈ R× satisfying aR× ∩ cR× = yR×. This also implies aR∩ cR = yR. So we get (b, a)R ⋊ R× ∩ (d, c)R ⋊ R× = (x + yR) × yR× = (x, y)R ⋊ R× 10 MAGNUS DAHLER NORLING Suppose R ⋊ R× satisfies Clifford's condition and let a, c ∈ R×. Since aR× ∩ cR× 6= ∅, (0, a)(R ⋊ R×) ∩ (0, c)(R ⋊ R×) = (aR ∩ cR) × (aR× ∩ cR×) 6= ∅ It follows that there exist y ∈ R× and x ∈ R (one may take x = 0) such that (0, a)(R ⋊ R×) ∩ (0, c)(R ⋊ R×) = (x, y)R ⋊ R× This implies that aR× ∩ cR× = yR×. (cid:3) Li shows that when R is a Dedekind domain, J(R ⋊ R×) is independent. Every Dedekind domain that is also a GCD domain is a principal ideal do- main. One way to see this is to use that every nontrivial ideal in a Dedekind domain R is on the form c−1(aR) for some c, a ∈ R. This is for instance proved in [17]. Note that Li denotes c−1(aR) as ((c−1a)· R)∩ R. This comes from viewing c−1a as an element of the field of fractions of R. Applying statement (ii) in Proposition 2.3.5 to the semigroup R× one can deduce that if R is also a GCD domain, any nontrivial ideal in R is on the form aR for some a ∈ R. This is the definition of a principal ideal domain. There exist Dedekind domains that are not principal ideal domains. An example of this is Z[√10] as seen on p. 407 in [13]. This shows that not every left cancellative semigroup with independent constructible right ideals satisfies Clifford's condition. On the other hand, every Dedekind domain is Noetherian (Theorem 6.10 in [13]), but not every GCD domain is Noetherian. So the integral domain R does not have to be a Dedekind domain for J(R ⋊ R×) to be independent. Examples of non-Noetherian GCD domains can be found in [4]. 3. C ∗-theory 3.1. The C ∗-algebras of an inverse semigroup. Let P be an inverse semigroup. We want to recall some common constructions for C ∗-algebras that are generated by representations of P by partial isometries. This is a short account of the theory. A more thorough account can for instance be found in [27] or [8]. One may construct such C ∗-algebras by associating them to certain groupoids, but we won't use this approach in the present paper. Let {δp}p∈P be the canonical basis of ℓ2(P ) satisfying hδp, δqi =(1 if p = q 0 otherwise Let CP be the vector space consisting of formal sums n for any n ∈ N, ai ∈ C and pi ∈ P . Define an involution on CP by aipi Xi=1 aipi!∗ n Xi=1 = n Xi=1 aip∗ i C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 11 and a product by n Xi=1 m aipi! Xj=1  bjqj  = n m Xi=1 Xj=1 aibjpiqj These operations make CP a ∗-algebra. The left regular representation of CP is defined to be the map Λ : CP → B(ℓ2(P )) given by Λ(p)δq =(δpq if p∗pq = q 0 otherwise Then Λ can be shown to be a faithful ∗-representation of CP . Define C ∗ to be the closure of the image of Λ with respect to the operator norm. r (P ) r (P ). C ∗(P ) is universal for representations of P by partial isometries. One way to construct the full C ∗-algebra of P is to show that CP is dense in the convolution algebra ℓ1(P ). One then lets C ∗(P ) be the univer- sal C ∗-enveloping algebra of the Banach ∗-algebra ℓ1(P ). The left regular representation Λ extends to a ∗-homomorphism Λ : C ∗(P ) → C ∗ If A is a C ∗-algebra, Piso(A) is the set of partial isometries in A and f : P → Piso(A) is a homomorphism onto a subsemigroup of Piso(A), then there is a ∗-homomorphism π : C ∗(P ) → A such that π(p) = f (p) for each p ∈ P . This implies that if P, Q are two inverse semigroups, then every homomorphism f : P → Q extends to a ∗-homomorphism πf : C ∗(P ) → C ∗(Q). Note that if P has a 0, then Λ(0)δ0 = δ0 and Λ(0)δp = 0 for p 6= 0. So Λ(0) 6= 0 is a one dimensional projection. This is undesireable in some of our later applications, but it is not too difficult go around the problem. If 0P exists, then C0P is an ideal in CP , so it is an ideal in C ∗(P ), and Λ(C0P ) is an ideal in C ∗ r (P )/Λ(C0P ). Since Λ sends a ∈ C ∗(P ) to Λ(C0P ) if and only if a ∈ C0P , Λ defines a ∗-homomorphism Λ0 : C ∗ r,0(P ). Moreover, if P and Q are inverse semigroups and f : P 0 → Q0 is a 0-homomorphism, then πf pushes down to πf,0 : C ∗ 0 (Q). It is important to note that if P is an inverse semigroup without 0, then C ∗ 0 (P ) = C ∗(P )/C0P and C ∗ 0 (P ) → C ∗ 0 (P ) → C ∗ r (P ). Let C ∗ 0 (P 0) ≃ C ∗(P ) and C ∗ r,0(P 0) ≃ C ∗ r (P ). r,0(P ) = C ∗ Definition 3.1.1 The inverse semigroup P is said to have weak containment if Λ : C ∗(P ) → C ∗ r (P ) is an isomorphism. Clearly Λ is an isomorphism if and only if Λ0 is an isomorphism. See [19] for a recent study of weak containment for inverse semigroups. Proposition 3.1.2 Let P be a commutative inverse semigroup. Then P has weak containment. Proof. This does for instance follow from Paterson's results in [25] since every commutative inverse semigroup P is a so-called Clifford semigroup and any subgroup of P has to be amenable. (cid:3) Corollary 3.1.3 Let P be an inverse semigroup, and let L be the subsemi- lattice of idempotents in P . Let D be the C ∗-subalgebra of C ∗ r (P ) generated r (L) and C ∗(L). A similar by Λ(L). Then D is canonically isomorphic to C ∗ result holds if we look at the subalgebra generated by Λ0(L) in C ∗ r,0(P ). 12 MAGNUS DAHLER NORLING Proof. Since L is commutative it has weak containment, so C ∗ r (L) ≃ C ∗(L) is universal for representations of L. Thus the norm CL gets from its rep- resentation on ℓ2(L) is greater than or equal to the one it gets from ℓ2(P ). But we also have that for a ∈ CL kΛ(a)kC ∗ r (P ) = sup{kΛ(a)δpk : p ∈ P} ≥ sup{kΛ(a)δpk : p ∈ L} = kΛ(a)kC ∗ r (L) (cid:3) Let L be a semilattice with 0 and let S be a set such that there is an injective 0-homomorphism f : L → 2S. Here 2S = {X : X ⊂ S} is given the structure of a semilattice by saying that the semigroup product is given by set intersection. This gives a representation µ : L → ℓ∞(S) by µ(a) = Ef (a) where EX is the characteristic function of X ⊂ S. Let C ∗(L; f ) be the C ∗-algebra generated by the image of µ. By the universality of C ∗ 0 (L) for 0- representations of L by commuting projections, there is a ∗-homomorphism 0 (L) → C ∗(L; f ) such that π(a) = µ(a) for each a ∈ L. We will say π : C ∗ that f is a maximal representation of L if this π is an isomorphism. The fact that π isn't always an isomorphism is related to the problem of finding the right way to map semilattices into Boolean algebras which is discussed by Exel in [11, 12]. Before we investigate this we want to discuss filters, which is a concept that is important in the representation theory of semilattices in the same way that characters are important in the representation theory of abelian C ∗-algebras. Definition 3.1.4 Let L be a semilattice with 0. A filter on L is a 0- homomorphism φ : L → {0, 1}. Here {0, 1} is given the structure of a semilattice with 1 · 0 = 0. An alternative view of filters on L is to define them to be subsets φ ⊂ L such that for all a, b ∈ L, (i) If a ∈ φ, then a (cid:22) b implies b ∈ φ. (ii) If a, b ∈ φ, then ab ∈ φ. (iii) 0 /∈ φ and 1 ∈ φ. Through the correspondence a ∈ φ ⇔ φ(a) = 1, one sees that these are equivalent definitions. The latter picture is the more traditional one. Now if ψ is a character on C ∗(L; f ) it defines a filter φ on L by saying that φ(a) = ψ(µ(a)) for each a ∈ L. Since µ(a) is always an idempotent, we have ψ(µ(a)) ∈ {0, 1}, so this φ well defined. Moreover, since µ(L) generates C ∗(L; f ), two characters on C ∗(L; f ) are equal if and only if their associated filters are equal. So the characters on C ∗(L; f ) are completely determined by their associated filters. In general, not every filter on L will extend to a character on C ∗(L; f ). Proposition 3.1.5 Let the setup be as above. The following conditions are equivalent (i) f : L → 2S is a maximal representation, i.e. π : C ∗ 0 (L) → C ∗(L; f ) is an isomorphism. C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 13 (ii) For every filter φ : L → {0, 1}, there is a character ψ on C ∗(L; f ) (iii) For all a1, . . . , an ∈ L, if there is some b ∈ L such that Sn i=1 f (ai) = such that ψ(µ(a)) = φ(a) for each a ∈ L. f (b), then ai = b for at least one 1 ≤ i ≤ n. Proof. (i)⇒(ii): Let φ : L → {0, 1} be a filter. If C ∗(L; f ) is isomorphic to 0 (L), then by the universal properties of this C ∗-algebra there is a nonzero C ∗ ∗-homomorphism C ∗(L; f ) → C ∗ 0 ({0, 1}) ≃ C extending φ. A nonzero ∗- homomorphism to C is exactly the definition of a character. (ii)⇒(iii): Let a1, . . . , an, b ∈ L be such thatSn i=1 f (ai) = f (b). Note that µ(b) ≤ Pn i=1 µ(ai). Let φ = {c ∈ L : b (cid:22) c}. This is a filter on L. Let ψ be the extending character. Then n 1 = ψ(µ(b)) ≤ ψ(µ(ai)) Xi=1 Then ψ(µ(ai)) = φ(ai) = 1 for at least one i . We have f (aib) = f (ai) ∩ f (b) = f (ai), so aib = ai, that is ai (cid:22) b. If φ(ai) = 1, then b (cid:22) ai by the definition of φ, so ai = b. (iii)⇒(i): This can be proved just like (i)⇒(ii) in Proposition 2.24 of [17], so we skip the proof. (cid:3) Corollary 3.1.6 Let S be a left cancellative semigroup. Then J(S) is independent (Definition 2.3.4) if and only if the inclusion ι : J(S)∪{∅} → 2S is a maximal representation. Proof. This follows from (iii) above and the definition of the independence of J(S). (cid:3) We will need the following proposition later. Proposition 3.1.7 Let P be an E∗-unitary inverse semigroup, and let L be its subsemilattice of idempotents. There exists a faithful conditional expectation Er,0 : C ∗ 0 (L) such that Er,0(Λ0(p)) = p if p ∈ L and Er,0(Λ0(p)) = 0 otherwise. Proof. Let Er : B(ℓ2(P )) → ℓ∞(P ) be the usual faithful conditional expec- tation given by r,0(P ) → C ∗ hEr(a)δq, δqi = haδq, δqi Here ℓ∞(P ) is viewed as a subalgebra of B(ℓ2(P )) represented by pointwise multiplication. First, if p ∈ L, then hΛ(p)δq, δri 6= 0 if and only if p∗pq = pq = q and pq = r which implies r = q. So Λ(p) ∈ ℓ∞(P ), and Er(Λ(p)) = Λ(p). In general, let p ∈ P and suppose hΛ(p)δq, δqi 6= 0 for some q ∈ P \ {0}. Due to Corollary 3.1.3, we now identify C ∗(L) with the closure of CL r (P ) → C ∗(L), and since Er(Λ(0)) = Λ(0), r,0(P ). We have that Er,0([a∗a]) = 0 if and only if Er(a∗a) = αΛ(0P ) for some α ∈ C if and only Then pq = q, so since P is E∗-unitary, p ∈ L. inside C ∗ Er,0 : C ∗ r (P ). We have Er : C ∗ r,0(P ) → C ∗ For any a ∈ C ∗ 0 (L) can be defined with the desired properties. r (P ), let [a] denote its image in C ∗ 14 MAGNUS DAHLER NORLING if ha∗aδ0P , δ0P i = kaδ0P k2 = α and kaδqk2 = 0 for all q ∈ P \ {0P}. This implies that a∗a = αΛ(0P ) and that [a∗a] = 0, so Er,0 is faithful. (cid:3) On the other hand we have the following lemma, which is also interesting. Lemma 3.1.8 Let A be a C ∗-algebra generated by an inverse semigroup P of partial isometries. Let L be the semilattice of idempotents in P , and let D be the subalgebra of A generated by L. If there exists a conditional expectation E : A → D such that E(p) = p if p ∈ L and E(p) = 0 otherwise, then P is E∗-unitary. Proof. Suppose p ∈ P and q ∈ L \ {0} satisfy pq = q. Then pq = q = E(q) = E(pq) = E(p)q since E is a conditional expectation. This implies that E(p) 6= 0, so p ∈ L. This shows that P is E∗-unitary. (cid:3) 3.2. The left regular representation and the left inverse hull of a left cancellative semigroup. From now on, S will always be a left cancellative semigroup unless something else is stated. Let {εs}s∈S be the orthogonal basis of ℓ2(S), where εs(t) = 1 if s = t, and 0 otherwise. The left regular representation of S is the semigroup homomorphism s 7→ Vs, where Vs : ℓ2(S) → ℓ2(S) is given by Vsεt = εst. Now, hV ∗ s εt, εri = 1 if t = sr and 0 otherwise, so εr V ∗ s εt = Xr∈t−1({s}) Since S is left cancellative, t−1({s}) is either a singleton or empty. It follows readily that Vs is an isometry for each s ∈ S. Let E : B(ℓ2(S)) → ℓ∞(S) be the conditional expectation given by hE(a)εs, εsi = haεs, εsi for each s ∈ S. Here we view ℓ∞(S) as a sub- algebra of B(ℓ2(S)) represented by pointwise multiplication. For a subset X ⊂ S, let EX ∈ ℓ∞(S) be the associated characteristic function. It is easy to check that for all s ∈ S and X ⊂ S (3.1) In particular, VsV ∗ s = EsS. r (S) be the C ∗-algebra generated by {Vs : s ∈ S}, and let Dr(S) be the commutative C ∗-algebra generated by {EX : X ∈ J(S)}. Note that C ∗ V ∗ s EXVs = Es−1(X) We will let C ∗ s = EsX, VsEXV ∗ r (S) is the closed linear span of the set V (S) = {V ∗ t1 Vs1 ··· V ∗ tnVsn : n ∈ N, s1, . . . , sn, t1, . . . , tn ∈ S} V (S) is itself a semigroup under composition of operators, and it is an inverse semigroup since it consists of partial isometries with commuting initial and final projections. To see this, note that if V = V ∗ tnVsn, then by repeatedly applying the relations in (3.1) we get that V V ∗ = EX with X = t−1 n snS. Similarly, V ∗V = EY with Y = s−1 1 s1 ··· t−1 1 t1S. The second assertion of the next lemma is also proved in [17]. n tn ··· s−1 t1Vs1 ··· V ∗ Lemma 3.2.1 The semilattice of idempotents in V (S) is isomorphic to the ∩-semilattice J := {t−1 1 s1 ··· t−1 n snS : n ∈ N, si, ti ∈ S} C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 15 Moreover J = J(S). Proof. We saw in the previous paragraph that the semilattice of idempotents in V (S) is {EX : X ∈ J}. For any X, Y ∈ J, EX EY = EX∩Y , so J has to be a ∩-semilattice and be isomorphic to {EX : X ∈ J}. Since J is therefore closed under ∩ it must be equal to J(S) as defined in subsection 2.3. (cid:3) The inverse semigroup V (S) will play an important role in the sequel. As we noted in the introduction it can be given a purely algebraic description. Let I(S) be the inverse semigroup of all bijective partially defined functions S → S. Define Il(S) to be the inverse subsemigroup of I(S) generated by the partial bijections {λs}s∈S where λs : S → sS is given by λs(t) = st. Lemma 3.2.2 There is a faithful representation ω : Il(S) → B(ℓ2(S)) such that ω(λs) = Vs for each s ∈ S. So ω is an isomorphism of Il(S) onto V (S). Proof. For any f ∈ Il(S), define ω(f ) : ℓ2(S) → ℓ2(S) by ω(f )εs =(εf (s) if s ∈ dom(f ) 0 otherwise where dom(f ) is the domain of f . Then ω(f ) ∈ B(ℓ2(S)) since f is injective, and for any s ∈ S, ω(λs) = Vs. For any f, g ∈ Il(S) we have that f = g if and only if dom(f ) = dom(g) and f (s) = g(s) for all s ∈ dom(f ). This happens if and only if ker ω(f ) = ker ω(g) and ω(f )εs = ω(g)εs for all s ∈ S. This is equivalent to ω(f ) = ω(g). So ω is an injective map. Now for any f, g ∈ Il(S) and s ∈ S, ω(f )ω(g)εs =(εf (g(s)) if s ∈ g−1(dom(f )) 0 otherwise so ω(f · g) = ω(f )ω(g). This shows that ω is a surjective homomorphism of Il(S) onto V (S), and thus it is an isomorphism. (cid:3) Il(S) is often called the left inverse hull of S and has previously been studied in several settings. Some recent information on it can be found in [18, 15, 14] and [16]. Using ω one can translate most statements about V (S) into statements about Il(S) and vice versa. Note that the semilattice of idempotents in Il(S) is {iX : X ∈ J(S)} ≃ J(S) where iX : X → X is the identity map on X. We will sometimes identify J(S) with {iX : X ∈ J(S)} from here on. Many of the ideas of this subsection are present in [17], but they are not expressed in terms of V (S) as an inverse semigroup. Using a simple induction argument (or deducing it from the proof of Lemma 3.2.2) we know that for any V ∈ V (S) and s ∈ S, V εs is either 0 or εt for some t ∈ S. Lemma 3.2.3 Let V ∈ V (S). Then E(V ) = V if and only if V is idempo- tent. Proof. By Lemma 3.2.1, V is idempotent if and only if V = EX for some X ∈ J(S). This implies E(V ) = V . Let V ∈ V (S), and suppose E(V ) = V . For every s ∈ S, hV εs, εsi is either 0 or 1, so V = E(V ) = EX where X = {s ∈ S : hV εs, εsi = 1}. Hence V = V 2. (cid:3) 16 MAGNUS DAHLER NORLING Corollary 3.2.4 Let a ∈ C ∗ r (S). Then E(a) = a if and only if a ∈ Dr(S). Proof. V (S) spans a dense subset of C ∗ r (S), and {EX : X ∈ J(S)} spans a dense subset of Dr(S). The result follows by the linearity and continuity of E. (cid:3) ρr(f (s)) = f (ρr(s)) Lemma 3.2.5 Let ρ be the right action of S on itself given by ρr(s) = sr for s, r ∈ S. Then for every f ∈ Il(S) and s ∈ dom(f ), (3.2) for all r ∈ S. Proof. Note that since dom(f ) = dom(f ∗f ) ∈ J(S) is a right ideal in S, s ∈ dom(f ) implies sr ∈ dom(f ) for all r ∈ S. Equation (3.2) clearly holds when f = λt for some t ∈ S. Suppose now that f = λ∗ t ) = tS. Then there is some q ∈ S such that tq = s and λ∗ t (s) = q. For any p ∈ S, λ∗ t (ρr(s)) = p if and only if sr = ρr(s) = tp. On the other hand, ρr(λ∗ t (s)) = qr = p if and only if tqr = tp. Since s = tq, this happens if and only if sr = tp. So for any p ∈ S, ρr(λ∗ t (ρr(s)) = p. This shows that ρr(λ∗ t (ρr(s)). t . Let s ∈ dom(λ∗ t (s)) = p if and only if λ∗ t (s)) = λ∗ Let f ∈ Il(S) be arbitrary. Then there are n ∈ N and s1, . . . , sn, t1, . . . , tn ∈ S such that f = λ∗ t1 λs1 ··· λ∗ tn λsn. Since s ∈ dom(f ), t1λs1 ··· λ∗ tnλsn(s) ∈ dom(λ∗ λ∗ tj λsj ··· λ∗ tj−1 λsj−1) for all 1 ≤ j ≤ n. So for all 1 ≤ j ≤ n since dom(λ∗ Starting with ρr(λ∗ tj λsj ··· λ∗ tn λsn(s)) ∈ dom(λ∗ t1 λs1 ··· λ∗ tj−1λsj−1) t1 λs1 ··· λ∗ tj−1 λsj−1) is a right ideal in S. we can then move ρr to the left one step at a time until we get λ∗ t1λs1 ··· λ∗ tnλsn(ρr(s)) ρr(λ∗ t1 λs1 ··· λ∗ tn λsn(s)) (cid:3) Corollary 3.2.6 Let f ∈ Il(S). For any s ∈ dom(f ), f λs = λf (s) Proof. Note first that dom f λs = dom λf (s) = S. By Lemma 3.2.5 we get that for any r ∈ S, f λs(r) = f (sr) = f (ρr(s)) = ρr(f (s)) = f (s)r = λf (s)(r) So f λs(r) = λf (s)(r) for all r ∈ S, hence f λs = λf (s). Lemma 3.2.7 V (S) is E∗-unitary if and only if for every V ∈ V (S) we have E(V ) = V or E(V ) = 0. Proof. Assume V (S) is E∗-unitary and let V ∈ V (S). We show that if V εr = εr for some r ∈ S, then E(V ) = V . Otherwise E(V ) is of course 0. Suppose V εr = εr. Using ω we can get from Corollary 3.2.6 that V Vr = Vr. Since V (S) is E∗-unitary this implies that V is idempotent, so E(V ) = V by Lemma 3.2.3. The converse statement follows from Lemma 3.1.8. (cid:3) (cid:3) C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 17 Note that this implies that if V (S) is E∗-unitary, then E(C ∗ Of course, V (S) is E∗-unitary if and only if Il(S) is E∗-unitary. r (S)) = Dr(S). Lemma 3.2.8 Let s, t ∈ S. The following conditions are equivalent. (i) s = t (ii) λsλ∗ (iii) λ∗ t λs = 1. t is idempotent. Proof. (i)⇒(ii) is trivial. multiplying with λt gives the desired equality. (ii)⇒(iii): We have λsλ∗ (iii)⇒(i): This implies that λsλ∗ t = λsλ∗ t λsλ∗ s we get that λ∗ t = λ∗ relations are unique for λ∗ λt(1) = t. t . Left multiplying with λ∗ s and right t λs = λs and λ∗ t λsλ∗ t = λ∗ t . Since these s. So λs = λt and s = λs(1) = (cid:3) Corollary 3.2.9 If Il(S) is E∗-unitary, S is cancellative. Proof. Let s, t, r, p ∈ S and assume that sr = tr = p. Since λsλr = λp, λ∗ sλp = λr. So λtλ∗ s is idempotent, so by the previous lemma s = t. Hence S is (right) cancellative. (cid:3) sλp = λtλr = λp. Since Il(S) is E∗-unitary λtλ∗ Lemma 3.2.10 Assume S is a subsemigroup of a group G. Let m, n ∈ N and si, ti, pj, qj ∈ S for 1 ≤ i ≤ n, 1 ≤ j ≤ m. Set f = λ∗ tnλsn and f ′ = λ∗ qmλpm, and assume f, f ′ 6= 0. If f = f ′ then the equality q1λp1 ··· λ∗ t1λs1 ··· λ∗ 1 s1 ··· t−1 t−1 1 p1 ··· q−1 (3.3) holds in G, where (·)−1 means taking inverses in G. Proof. Since f, f ′ f (r) = f ′(r) gives n sn = q−1 m pm 6= 0 we can pick some r ∈ dom(f ). Then the equality 1 s1 ··· t−1 t−1 n snr = q−1 1 p1 ··· q−1 m pmr where (·)−1 denotes the preimage by left multiplication in S. However this implies that the same relation holds in G, where (·)−1 now stands for the inverse operation in G. Cancelling with r we get (3.3). (cid:3) The proof of the next proposition uses techniques similar to those em- ployed by Jiang in [14]. Proposition 3.2.11 Let S be a left cancellative semigroup. Then S embeds into a group if and only if Il(S) is strongly E∗-unitary. Proof. Suppose first that S embeds into a group G. We omit writing the embedding homomorphism, and instead view S as a subsemigroup of G. Define a grading ϕ : Il(S)0 → G0 by ϕ(0) = 0 tnλsn) = t−1 ϕ(λ∗ t1 λs1 ··· λ∗ 1 s1 ··· t−1 n sn t1λs1 ··· λ∗ when λ∗ tnλsn 6= 0 This is well-defined because of Lemma 3.2.10. Suppose ϕ(f ) = 1 for some f ∈ Il(S). Then if f = λ∗ n sn = 1. So f (r) = 1 s1 ··· t−1 t1 λs1 ··· λ∗ tn λsn, t−1 18 MAGNUS DAHLER NORLING t−1 1 s1 ··· t−1 n snr = r for all r ∈ dom(f ). Hence f is idempotent, and ϕ is idempotent pure. This shows that Il(S) is strongly E∗-unitary. Now suppose Il(S) is strongly E∗-unitary by some idempotent pure grad- t )ϕ(λt), so t ) = 1, is idempotent, and by Lemma 3.2.8 s = t. This implies that the (cid:3) ing ϕ : Il(S)0 → G0. For any t ∈ S, 1 = ϕ(λ∗ ϕ(λt)−1 = ϕ(λ∗ so λsλ∗ t homomorphism S → G given by s 7→ ϕ(λs) is injective. r,0(Il(S)) and C ∗ t ). For any s, t ∈ S, if ϕ(λs) = ϕ(λt), then ϕ(λsλ∗ We want to find a relation between C ∗ t λt) = ϕ(λ∗ r (S). Lemma 3.2.12 Let T : ℓ2(S) → ℓ2(Il(S)) be the isometry defined by T εs = δλs s ∈ S Then T ∗Λ(f )T = ω(f ) for all f ∈ Il(S). Proof. Let f ∈ Il(S) and s ∈ dom(f ). Then s ∈ dom(f ∗f ), so f ∗f (s) = s. By Corollary 3.2.6, f ∗f λs = λs and f λs = λf (s). Now by the definition of Λ, T ∗Λ(f )T εs = T ∗Λ(f )δλs = T ∗δf λs = T ∗δλf (s) = εf (s) = ω(f )εs On the other hand, if s /∈ dom(f ) then s /∈ dom(f ∗f ). Thus f ∗f λs 6= λs and we get T ∗Λ(f )T εs = T ∗Λ(f )δλs = 0. So T ∗Λ(f )T εs = ω(f )εs for any s ∈ S. This shows that T ∗Λ(f )T = ω(f ) for any f ∈ Il(S). (cid:3) Corollary 3.2.13 There is a surjective ∗-homomorphism h : C ∗ C ∗ r (S) such that h(Λ0(f )) = ω(f ) for all f ∈ Il(S). Proof. Define h′ r (S) by h′(a) = T ∗aT . Then h′ is a ∗-homomorphism on the span of Λ(Il(S)). Since this span is dense in C ∗ r (Il(S)) and since h′ is continuous, it has to be a ∗-homomorphism on all of C ∗ r (Il(S)). Since h′ sends Λ(0) to 0 whenever 0 ∈ Il(S), it descends to a ∗-homomorphism h : C ∗ r (S) with the desired properties. (cid:3) Theorem 3.2.14 Suppose Il(S) is E∗-unitary. Then the map r (Il(S)) → C ∗ r,0(Il(S)) → C ∗ r,0(Il(S)) → : C ∗ h : C ∗ r,0(Il(S)) → C ∗ r (S) is an isomorphism if and only if J(S) is independent. The restriction of h to C ∗ Proof. Recall that Dr(S) is the diagonal subalgebra of C ∗ r (S) generated by J(S). Then Dr(S) is C ∗(J(S)0; ι) as described in Proposition 3.1.5 and the paragraphs before it. Here ι : J(S)0 → 2S is the inclusion map. 0 (J(S)) (which we can identify with the sub- algebra generated by the image of J(S) in C ∗ r,0(Il(S)) by Corollary 3.1.3) maps onto Dr(S), and this restriction must necessarily be equal to the map π as described in and before Proposition 3.1.5. According to this proposi- tion and Corollary 3.1.6, hC ∗ 0 (J(S)) is an isomorphism if and only if J(S) is independent. This means that if J(S) is not independent, h is not injective. Suppose J(S) is independent and Il(S) is E∗-unitary. By Proposition 3.1.7 there is a faithful conditional expectation Er,0 : C ∗ 0 (J(S)). As a consequence of Lemma 3.2.7, E(C ∗ r (S)) = Dr(S). Moreover for any V ∈ V (S), E(V ) = V if and only if V is idempotent, and E(V ) = 0 r,0(Il(S)) → C ∗ C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 19 Now assume that h(a) = 0 for some a ∈ C ∗ otherwise. Using the properties of Er,0 given in Proposition 3.1.7 it follows that E ◦ h = h ◦ Er,0 r,0(Il(S)). Then h(a∗a) = 0, so E(h(a∗a)) = 0 = h(Er,0(a∗a)). Since h is an isomorphism on the image of Er,0, Er,0(a∗a) = 0, so a∗a = a = 0 since Er,0 is faithful. This shows that h is an isomorphism. (cid:3) We can use the equality Dr(S) = C ∗(J(S)0; ι) to describe the characters on Dr(S). Proposition 3.1.5 and Corollary 3.1.6 imply that when J(S) is independent, the characters on C ∗(J(S)0; ι) are uniquely determined by the filters on J(S)0. When S is algebraically ordered and satisfies Clifford's condition, Proposition 2.3.5 tells us that (S0,∧) ≃ J(S)0. So in this case the characters on Dr(S) correspond to the filters on (S0,∧). It is not difficult to see that the filters on (S0,∧) are exactly what Nica calls non-void hereditary directed subsets of S in subsection 6.2 of [22]. This is sometimes called the Nica spectrum of S. In general, the set of characters on Dr(S) corresponds to some subset of the set of filters on J(S)0, but it is not always obvious what this subset is. Doing computations in Il(S) can be difficult, but if S satisfies Clifford's condition it becomes easier. Note that S satisfies Clifford's condition exactly when Il(S) is 0-bisimple. We will however not use this fact explicitly in this paper. See for instance [5, 15] or [14] for more information on this. Proposition 3.2.15 The following conditions are equivalent (i) S satisfies Clifford's condition. (ii) For all s, t ∈ S such that λ∗ (iii) Il(S) \ {0} = {λpλ∗ q : p, q ∈ S}. λ∗ t λs = λpλ∗ q. t λs 6= 0 there exist p, q ∈ S such that Proof. (i)⇒(ii): Let s, t ∈ S. If λ∗ t λs 6= 0, then sS ∩ tS 6= ∅, so sS ∩ tS = rS for some r ∈ S. Since r ∈ sS ∩ tS, p := t−1(r) and q := s−1(r) exist, and λtλp = λr, so we get that λ∗ sλr = λq. By the definition of r we have t λtλp = λp. Similarly λ∗ t λr = λ∗ λtλ∗ t λsλ∗ s = itSisS = irS = λrλ∗ r So multiplying from the left by λ∗ λ∗ t λs = (λ∗ t and from the right by λs we get t λr)(λ∗ rλs) = λpλ∗ q t λsλ∗ t λs = λpλ∗ (ii)⇒(i): Let s, t ∈ S such that sS ∩ tS 6= ∅. Then λ∗ t λsλ∗ t λs 6= 0, so there exists p, q ∈ S with λ∗ s = λtpλ∗ sq. This element is idempotent, so tp = sq by Lemma 3.2.8. Write r = tp = sq. This s = λrλ∗ gives λtλ∗ q. This means that λtλ∗ r, which is equivalent to sS ∩ tS = rS. (iii)⇒(ii) is trivial. It remains to prove (ii)⇒(iii): Let f ∈ Il(S) \ {0}. t1λs1 ··· λ∗ qn for some pn, qn ∈ S. Assume Then there exist n ∈ N and s1, . . . , sn, t1, . . . , tn ∈ S such that f = λ∗ By condition (ii) we have that λ∗ that for a given 1 ≤ j ≤ n there exist pj, qj ∈ S such that tnλsn = λpnλ∗ tnλsn. λ∗ tj λsj ··· λ∗ tnλsn = λpj λ∗ qj Then λ∗ tj−1 λsj−1λ∗ tj λsj ··· λ∗ tnλsn = λ∗ tj−1 λsj−1λpj λ∗ qj 20 MAGNUS DAHLER NORLING Now by condition (ii), there exist p, q ∈ S such that λ∗ tj−1λsj−1λpj = λ∗ Setting pj−1 = p and qj−1 = qqj, we get λ∗ tj−1λsj−1λpj λ∗ qj = λpj−1λ∗ qj−1 tj−1 λsj−1pj = λpλ∗ q By induction on j, λ∗ t1λs1 ··· λ∗ tnλsn = λp1λ∗ q1 This shows that any f ∈ Il(S)\{0} is on the form λpλ∗ q for some p, q ∈ S. (cid:3) Corollary 3.2.16 Suppose S satisfies Clifford's condition. Then Il(S) is E∗-unitary if and only if S is cancellative Proof. One implication was proved in Corollary 3.2.9. Suppose S is cancella- tive, and that f iX = iX for some nonzero f ∈ Il(S) and some nonempty X ∈ J(S). Then f λr = λr for any r ∈ X since X is a right ideal. Write f = λsλ∗ t λr and define q = λ∗ sλr(p). Then tq = sq = rp. By right cancellativity this gives t = s, so f is idempotent. (cid:3) t with s, t ∈ S. Then λ∗ sλr. Let p ∈ dom λ∗ t λr(p) = λ∗ t λr = λ∗ Corollary 3.2.17 If S is cancellative and satisfies Clifford's condition, h : r,0(Il(S)) → C ∗ C ∗ Proof. By the previous corollary, Il(S) is E∗-unitary. By Proposition 2.3.5 J(S) is independent, so Theorem 3.2.14 implies that h is an isomorphism. (cid:3) r (S) is an isomorphism. Any semigroup that is the positive cone in a quasilattice ordered group satisfies these conditions. Note however that a semigroup satisfying Clifford's condition is allowed to have nontrivial invertible elements. For instance (Z× Z+, +) satisfies Clifford's condition, but it is not algebraically ordered. 3.3. Li's constructions of full C ∗-algebras for a left cancellative semigroup. In [17], Li defines the full C ∗-algebra C ∗(S) of a left can- cellative semigroup S. The construction is as follows: C ∗(S) is the uni- versal C ∗-algebra generated by isometries {vs : s ∈ S} and projections {eX : X ∈ J(S)0} such that for all s, t ∈ S and X, Y ∈ J(S)0, eS = 1 vst = vsvt e∅ = 0 s (S) when S embeds into a group G: C ∗ vseX v∗ s = esX eX∩Y = eX eY Li also defines a C ∗-algebra C ∗ s (S) is the universal C ∗-algebra generated by isometries {vs : s ∈ S} and projec- tions {eX : X ∈ J(S)0} such that for all s, t ∈ S, vst = vsvt e∅ = 0 and whenever s1, . . . , sn, t1, . . . , tn ∈ S satisfy t−1 tnvsn = eX v∗ t1vs1 ··· v∗ 1 s1 ··· t−1 n sn = 1 in G, then 1 s1 ··· t−1 where X = t−1 and {eX : X ∈ J(S)0} ⊂ C ∗ there exists a surjective ∗-homomorphism πs : C ∗(S) → C ∗ s (S) s (S) satisfy the relations defining C ∗(S), so s (S) that sends n snS. Li then shows that {vs : s ∈ S} ⊂ C ∗ C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 21 s (S). The universal property of C ∗ vs ∈ C ∗(S) to vs ∈ C ∗ a canonical ∗-homomorphism C ∗ s ∈ S. Moreover, we have: Proposition 3.3.1 Suppose S embeds into a group G. There is a ∗- isomorphism κ : C ∗ s (S) also gives r (S) that sends vs to Vs for all 0 (Il(S)) such that κ(vs) = λs for each s ∈ S. s (S) → C ∗ s (S) → C ∗ Proof. The existence of a surjective ∗-homomorphism κ : C ∗ 0 (Il(S)) s (S). C ∗(Il(S)) is generated by {λs : follows from the universality of C ∗ s ∈ S} and {iX : X ∈ J(S)}, and these satisfy the given relations when projected down to C ∗ 0 (Il(S)). In particular, if s1, . . . , sn, t1, . . . , tn ∈ S satisfy t−1 1 s1 ··· t−1 n sn = 1 in G, then by the proof of Proposition 3.2.11, f := tnλsn is idempotent. So f = f f ∗ = iX with X = t−1 λ∗ t1λs1 ··· λ∗ n snS. Let V ′(S) be the subset of C ∗ 1 s1 ··· t−1 s (S) → C ∗ s (S) given by {v∗ t1 vs1 ··· v∗ tnvsn : n ∈ N, t1, . . . , tn, s1, . . . , sn ∈ S} s (S) and C ∗ Comparing the universal properties of C ∗ Li's relations guarantee that V ′(S) is actually an inverse semigroup. Using Lemma 2.8 of [17] (and mapping down to C ∗ s (S)), we get that for any v ∈ V ′(S), v∗v = eX and vv∗ = eY for some X, Y ∈ J(S), so v is a partial isometry. Moreover, any v, w ∈ V ′(S) have commuting initial and final projections. 0 (V ′(S)) gives that these two C ∗-algebras are canonically isomorphic. Hence κ is an isomor- phism if and only if its restriction to V ′(S) gives a semigroup isomorphism V ′(S) → Il(S) (note that the restriction of κ to V ′(S) is automatically a surjective semigroup homomorphism onto Il(S)). Let J ′(S) be the semilat- tice of idempotents in V ′(S). For any v ∈ J ′(S), v∗v = v, so v = eX for some X ∈ J(S), i.e. J ′(S) = {eX : X ∈ J(S)}. Hence κ restricts to an isomorphism J ′(S) → J(S) since it is injective on this set. t1vs1 ··· v∗ tnvsn ∈ V ′(S). Suppose κ(v) = λ∗ tn λsn is idempotent. Then t−1 1 s1 ··· t−1 n sn = 1 in G. This implies that v is idem- potent. Thus κ−1(J(S)) = J ′(S), so κV ′(S) is an isomorphism by Lemma 2.2.8. t1 λs1 ··· λ∗ Let v = v∗ (cid:3) Proposition 3.3.2 Let S be any left cancellative semigroup. There is a surjective ∗-homomorphism η : C ∗(S) → C ∗ 0 (Il(S)) such that η(vs) = λs for each s ∈ S. If S satisfies Clifford's condition, then η is an isomorphism. Proof. The existence of η follows as before from the universality of C ∗(S). Define V(S) ⊂ C ∗(S) to be t1 vs1 ··· v∗ tnvsn : n ∈ N, t1, . . . , tn, s1, . . . , sn ∈ S} {v∗ As in the previous proposition, V(S) is an inverse semigroup and it is suffi- cient to show that the restriction of η to V(S) is a semigroup isomorphism onto Il(S). Let J ′′(S) = {eX : X ∈ J(S)}. Then J ′′(S) is the semilattice of idempotents in V(S) and ηJ ′′(S) is an isomorphism onto J(S). q : p, q ∈ S}. The proof is almost identical to that in Proposition 3.2.15, and we will only show that for any s, t ∈ S with v∗ t vs 6= 0, First we show that V(S) \ {0} = {vpv∗ t vs 6= 0, there are p, q ∈ S with v∗ t vs = vpv∗ q . If v∗ 22 MAGNUS DAHLER NORLING vtv∗ t vsv∗ t vsv∗ s = vrv∗ then vtv∗ s 6= 0, so since ηJ ′′(S) is an isomorphism onto J(S), and since S satisfies Clifford's condition, there is some r ∈ S with sS ∩ tS = rS and (3.4) Since r ∈ sS∩tS, there are p, q ∈ S such that sq = r and tp = r. Then vsvq = vr, so vq = v∗ t vr. Now by multiplying equation (3.4) on the left with v∗ r vs = vpv∗ q . Consider v ∈ V(S) \ {0}, and suppose η(v) is idempotent. There are q, so p = q by Lemma 3.2.8. Thus p, which is idempotent. It follows from Lemma 2.2.8 that (cid:3) s vr. Similarly, vp = v∗ t and on the right with vs, we get that v∗ p, q ∈ S with v = vpv∗ v = vpv∗ ηV(S) is an isomorphism. q . Now η(v) = λpλ∗ q = vpv∗ t vs = v∗ t vrv∗ r It is now clear that the canonical map C ∗(S) → C ∗ C ∗(S) η −→ C ∗ 0 (Il(S)) Λ0−→ C ∗ r,0(Il(S)) r (S) factors as −→ C ∗ r (S) h When S embeds into a group we get the following factorization: C ∗(S) πs−→ C ∗ s (S) κ −→≃ C ∗ 0 (Il(S)) Λ0−→ C ∗ r,0(Il(S)) h −→ C ∗ r (S) Note that in this case η = κ ◦ πs, so πs is an isomorphism if and only if η is an isomorphism. Li asks when a semigroup homomorphism φ : S → R of left cancella- tive semigroups induces a ∗-homomorphism C ∗(S) → C ∗(R) by the formula vs 7→ vφ(s). We can give a partial answer. It induces a ∗-homomorphism C ∗ 0 (Il(S)) → C ∗ 0 (Il(R)) given by λs 7→ λφ(s) if and only if φ extends to a 0-homomorphism Il(S)0 → Il(R)0. Of course determining when this is the case may not be easy. See Corollary 3.4.11 for a result in this direction when S is left reversible and R is a group. C ∗(S) has the nice feature that it can be described as a crossed product by endomorphisms (see Lemma 2.14 of [17]). Li also shows that C ∗(S) general- izes Nica's C ∗-algebras for quasilattice ordered groups as well as the Toeplitz algebras associated with rings of integers [7]. Nica proved in [23] that his C ∗-algebra for the quasilattice ordered group (G, S) can be constructed as a C ∗-algebra of the Toeplitz inverse semigroup T (G, S). This can be explained by the next lemma as well as Proposition 3.3.2 above. For each g ∈ G, define βg : {s ∈ S : gs ∈ S} → {s ∈ S : g−1s ∈ S}, βg(s) = gs T (G, S) is then defined to be the inverse subsemigroup of I(S) generated by {βg}g∈G. Lemma 3.3.3 Let (G, S) be a quasilattice ordered group. Then Il(S)0 = T (G, S)0. Proof. Let g ∈ S. Then {s ∈ S : gs ∈ S} = S and {s ∈ S : g−1s ∈ S} = gS, so βg = λg. This shows that T (G, S) contains Il(S). g = {s ∈ S : g−1s ∈ S} = gS ∩ S is nonempty, so g−1s = t ∈ S for some s, t ∈ S. Then g ≤ s as defined in Definition 2.3.6. Moreover, 1−1s ∈ S, so 1 ≤ s. Thus s is a common upper bound for g and 1 in S. Then g and 1 have a least common upper bound r ∈ S since (G, S) is quasilattice ordered. Note that for any g ∈ G, βg−1 = β∗ g . If βg 6= 0, dom β∗ C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 23 Let p ∈ gS ∩ S. Using the same arguments as we did for s, we get g ≤ p and 1 ≤ p, so r ≤ p since r was a least common upper bound for g and 1. Then r (cid:23) p, so p ∈ rS. This shows that gS ∩ S ⊂ rS. However since g ≤ r, g−1r = u for some u ∈ S. Then r = gu, so rS = guS ∩ S ⊂ gS ∩ S. Now dom β∗ g = gS ∩ S = rS = dom λuλ∗ r Moreover for any v ∈ gS ∩ S, g (v) = βg−1(v) = g−1v = ur−1v = λuλ∗ β∗ r(v) So β∗ g = λuλ∗ r, and βg = λrλ∗ u. This shows that T (G, S)0 ⊂ Il(S)0. (cid:3) A more detailed discussion on the relationship between Il(S) and T (G, S) can be found in [14]. Nica's construction of a C ∗-algebra for T (G, S) uses a groupoid, but Milan explains in section 5 of [19] why this C ∗-algebra is isomorphic to C ∗ 0 (T (G, S)). 3.4. Left reversible semigroups, left amenability and functoriality. We still consider a left cancellative semigroup S unless something else is stated. Recall that S is left reversible if for any s, t ∈ S, sS ∩ tS 6= ∅. The next lemma is a slightly stronger version of Lemma 2.4.8 in [16]. Lemma 3.4.1 S is left reversible if and only if 0 /∈ Il(S) if and only if ∅ /∈ J(S). Proof. Clearly 0 /∈ Il(S) if and only if ∅ /∈ J(S). Moreover sS ∩ tS ∈ J(S) for all s, t ∈ S so ∅ /∈ J(S) implies that S is left reversible. Suppose S is left reversible, and let X, Y ⊂ S be nonempty right ideals. If s ∈ X and t ∈ Y , then sS ⊂ X and tS ⊂ Y , so sS ∩ tS ⊂ X ∩ Y , and X ∩ Y 6= ∅. Moreover, for any t ∈ S, tt−1X = X ∩ tS, so t−1X is nonempty. It follows by a simple induction argument that ∅ /∈ J(S). (cid:3) We include a short proof of the well known fact that left amenable semi- groups are left reversible. See also Proposition (1.23) in [26]. To be formal, a left invariant mean on S is a state µ on ℓ∞(S) such that for any s ∈ S and ξ ∈ ℓ∞(S), µ(ξ◦λs) = µ(ξ). S is left amenable if it has a left invariant mean. Right amenability is similarly defined. It is not difficult to show that a group or an inverse semigroup is left amenable if and only if it is right amenable, so left amenable groups and inverse semigroups are often just called amenable. Lemma 3.4.2 Let S be left amenable. Then for any left invariant mean µ on S, µ(EX) = 1 for all X ∈ J(S). This implies that S is left reversible. Proof. For convenience we set µ(X) = µ(EX ) for any X ⊂ S. Note that for any t ∈ S, EX ◦ λt = Et−1X. Thus µ(t−1X) = µ(X). Since t−1tX = X, we also have µ(tX) = µ(X). By Lemma 3.2.1, J(S) = {t−1 1 s1 ··· t−1 n snS : n ∈ N, si, ti ∈ S} So by a simple induction argument we get that µ(X) = µ(S) = 1 for all X ∈ J(S). As µ(∅) = 0, this shows that ∅ /∈ J(S). (cid:3) 24 MAGNUS DAHLER NORLING It is also well known that a cancellative left reversible semigroup embeds into a group G. This is one formulation of Ore's Theorem from [24]. The proof we present below in Theorem 3.4.3 is basically the same as Rees' proof that can be found in vol I, p. 35 of [6]. The reason we repeat it here is that it illustrates how Il(S) is related to G. See also ch. 2.4 of [16] where Lawson gives an account of this proof and shows that Il(S) is E-unitary when S is left reversible and cancellative. Suppose 0 /∈ Il(S). Then we can construct the maximal group homomor- phic image G(Il(S)) of Il(S) as described in Definition 2.2.7. For simplicity we write G(S) = G(Il(S)). Let αS : Il(S) → G(S) denote the quotient ho- momorphism. Let γS : S → G(S) be given by γS(s) = αS(λs). Then G(S) is generated by the cancellative semigroup γS(S). Theorem 3.4.3 Let S be left reversible. The following conditions are equivalent (i) S is cancellative. (ii) γS : S → G(S) is injective. (iii) S embeds into a group. (iv) Il(S) is E-unitary. Proof. (ii)⇒(iii) and (iii)⇒(i) are trivial. (iii)⇔(iv) follows from Proposition 3.2.11 (Il(S) is strongly E∗-unitary if and only if S is group embeddable) and the fact that an inverse semigroup without 0 is E-unitary if and only if it is strongly E∗-unitary. (i)⇒(ii): Since the map S ֒→ Il(S) is injective, we only have to prove that the homomorphism Il(S) → G(S) is injective on the set {λs : s ∈ S}. Let s, t ∈ S and suppose λs and λt map to the same element. Then by the definition of the congruence that was used to construct G(S) there is an X ⊂ J(S) such that λsiX = λtiX. Hence for any r ∈ X, sr = λs(r) = λt(r)=tr. By cancelling with r we get that s = t. (cid:3) Definition 3.4.4 Let P be semigroup. Recall that a subset X ⊂ P is said to be left thick if for any finite sequence s1, . . . , sn ∈ P , n The next proposition is related to Proposition (1.27) in [26]. siX 6= ∅ \i=1 Proposition 3.4.5 Let S be a subsemigroup of a group G such that S generates G. Then S is left reversible if and only if S is a left thick subset of G. Proof. For t ∈ S and X ⊂ S, t−1(X) = (t−1)X ∩ S where t−1(X) is the preimage inside S and (t−1)X is defined by multiplication inside G. So for any n ∈ N and si, ti ∈ S, n t−1 1 s1 ··· t−1 n snS = S ∩ (t−1 1 s1 ··· t−1 j sj)S \j=1 If S is left thick, this set is never empty, nor is any finite intersection of sets of this type, so ∅ /∈ J(S). On the other hand, for any g1 . . . gm ∈ G, write C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 25 gi = t−1 i,1 si,1 ··· t−1 i,ni m si,ni with si,j, ti,j ∈ S. Then \i=1 giS ⊃ S ∩ \j=1 \i=1 ni m (t−1 i,1 si,1 ··· t−1 i,j si,j)S If ∅ /∈ J(S), the right hand side is nonempty, and so is the left hand side, so S is a left thick subset of G. (cid:3) By a theorem of Mitchell [20], if S′ is a left thick subsemigroup of a of a semigroup S, then S′ is left amenable if and only if S is left amenable. Hence we get: Corollary 3.4.6 Let S be cancellative and left reversible. Then S is left amenable if and only if G(S) is amenable. We will now show that the assumption that S is right cancellative is redundant in the statement of Corollary 3.4.6. If P is any semigroup, let ≈ (or ≈P ) be the relation on P given by s ≈ t if there is some r ∈ P with sr = tr. From vol I, p.35 of [6] we have that if P is left reversible, ≈ is a congruence and P/ ≈ is a right cancellative semigroup. Proposition 1.25 of [26] states that P is left amenable if and only if P/ ≈ is left amenable. Note that ≈Il(S) is exactly the congruence on Il(S) one takes the quotient with to create G(S). Lemma 3.4.7 Let S be left reversible. Then γS(S) is isomorphic to S/ ≈S Proof. We show that γS(s) = γS(t) if and only if there is some r ∈ S such that sr = tr. First, if sr = tr, then λsλr = λtλr, so γS(s) = γS(t). On the other hand, if γS(s) = γS(t), there is some X ∈ J(S) such that λsiX = λtiX . Let r ∈ X. Then λsiX λr = λsλr = λtiXλr = λtλr, so by evaluating at 1 we get sr = tr. (cid:3) Corollary 3.4.8 Let S be a left cancellative left reversible semigroup. Then S is left amenable if and only if γS(S) is left amenable if and only if G(S) is amenable if and only if Il(S) is amenable. Proof. To show that the amenability of G(S) is equivalent to left amenability of γS(S), we need to show that γS(S) is left reversible. We have that for any s, t ∈ S, γS(s)γS (S) ∩ γS(t)γS(S) = γS(sS) ∩ γS(tS) ⊃ γS(sS ∩ tS) The right hand side is nonempty, so the left hand side must be nonempty as well. This proves that γS(S) is left reversible since any p ∈ γS(S) is on the form γS(s) for some s ∈ S. All the other equivalences are taken care of by the results we have developed so far: Since G(S) = Il(S)/ ≈Il(S), G(S) is amenable if and only if Il(S) is amenable. Since γS(S) ≃ S/ ≈S, γS(S) is left amenable if and only if S is left amenable. (cid:3) We conclude this subsection by showing that when S is left reversible the construction S 7→ G(S) is a generalization of the Grothendieck construction in that it is functorial. This is probably already known by specialists, but we give a proof here for completeness. Another way to prove it is to show that any homomorphism of S to a group can be extended to define a group 26 MAGNUS DAHLER NORLING homomorphic image of Il(S), and then use that G(S) is the maximal group homomorphic image of Il(S). Lemma 3.4.9 Let S be a subsemigroup of a group G such that S generates G, and let H be a group. If S is left thick in G, then every homomorphism φ : S → H uniquely extends to a homomorphism φ′ : G → H. Proof. Let t1, . . . , tn, s1, . . . , sn ∈ S. We want to define φ′(t−1 n sn) = φ(t1)−1φ(s1)··· φ(tn)−1φ(sn), so we need to show that this is a consistent definition. Let q1 . . . qm, p1 . . . pm ∈ S be such that 1 p1 ··· q−1 t−1 1 s1 ··· t−1 1 s1 ··· t−1 n sn = q−1 m pm Since S is left thick in G, n m S ∩ \i=1 (s−1 n tn ··· s−1 i (p−1 m qm ··· p−1 1 q1)S 6= ∅ So there exists an r ∈ S such that ui := t−1 vj := q−1 ti)S ∩ \j=1 i si ··· t−1 j pj ··· q−1 n snr ∈ S m smr ∈ S for all 1 ≤ i ≤ n and 1 ≤ j ≤ m. First, t−1 n snr = un, so snr = tnun, which implies that φ(sn)φ(r) = φ(tn)φ(un) and φ(tn)−1φ(sn)φ(r) = φ(un). Suppose now that for some 1 ≤ k ≤ n, Then since we get φ(tk)−1φ(sk)··· φ(tn)−1φ(sn)φ(r) = φ(uk) sk−1t−1 n snr = sk−1uk = tk−1uk−1 k sk ··· t−1 φ(tk−1)−1φ(sk−1)φ(uk) = φ(uk−1) Using induction, this implies that φ(t1)−1φ(s1)··· φ(tn)−1φ(sn)φ(r) = φ(u1). Similarly, φ(q1)−1φ(p1)··· φ(qm)−1φ(pm)φ(r) = φ(v1) = φ(u1), so by can- celling with φ(r) we see that φ′ is well defined. Uniqueness of φ′ is trivial since S generates G. (cid:3) Theorem 3.4.10 Let S be left reversible and let H be a group. Then every homomorphism φ : S → H gives rise to a unique homomorphism φ′ : G(S) → H such that φ′ ◦ γS = φ. Moreover for any left cancellative left reversible semigroup R and homo- morphism φ : S → R, there is a unique homomorphism φ′ : G(S) → G(R) such that φ′ ◦ γS = γR ◦ φ. Proof. First, we need to show that φ : S → H can be pushed down to a homomorphism γS(S) → H. If s, t, r ∈ S with sr = tr, then φ(s)φ(r) = φ(t)φ(r), so φ(s) = φ(t). This implies that φ is constant on the equivalence classes of ≈, hence there exists a homomorphism φ′′ : γS(S) → H such that φ′′γS = φ. By Lemma 3.4.9, φ′′ extends to a homomorphism φ′ : G(S) → H such that φ′ ◦ γS = φ. Uniqueness follows since the constructions φ 7→ φ′′ and φ′′ 7→ φ′ are unique, so if ψ : S → G(S) is another homomorphism with ψ ◦ γS = φ, then the restriction of ψ to γS(S) must be equal to φ′′. C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 27 If φ : S → R is a homomorphism, then we can apply the above construc- tion to γR◦φ : S → G(R) and thereby get the desired φ′ : G(S) → G(R). (cid:3) Corollary 3.4.11 Let S be left reversible and let G be a group. Then for every homomorphism φ : S → G there exists a ∗-homomorphism πφ : C ∗(Il(S)) → C ∗(G) such that πφ(λs) = λφ(s) for each s ∈ S. Proof. Consider a homomorphism φ : S → G. From Theorem 3.4.10 there exists a homomorphism φ′ : G(S) → G such that φ′ ◦ γS = φ. Then φ′ ◦ αS : Il(S) → G satisfies φ′(αIl(S)(λs)) = λφ(s) for each s ∈ S, so the existence of πφ follows from the universal property of C ∗(Il(S)). (cid:3) For exampe if S is left reversible we may consider the quotient homo- morphism αS : Il(S) → G(S) and obtain a surjective ∗-homomorphism πS : C ∗(Il(S)) → C ∗(G(S)). (Li also shows the existence of such a map s (S)). When S = Z+ this is the surjective part of the classical C ∗- from C ∗ extension 0 → K(ℓ2(Z+)) → C ∗ r (Z+) → C(T) → 0 where K(ℓ2(Z+)) are the compact operators on ℓ2(Z+), C ∗ C ∗(Il(Z+)) is the unique C ∗-algebra generated by a single isometry, and C(T) ≃ C ∗(Z) ≃ C ∗(G(Z+)). πS,r : C ∗ r (G(S)) as well. In general it would be interesting to have a description of the kernel of πS and πS,r. Nica [22] gives some neces- sary and sufficient conditions for C ∗ r (S) to contain the compacts when (G, S) is a quasilattice ordered group. By Proposition 1.4 in [8], there is actually always a canonical ∗-homomorphism r (Il(S)) → C ∗ r (Z+) ≃ C ∗ r (Il(Z+)) ≃ 3.5. Amenability and weak containment when S embeds into a group. In [17], Li shows that if S is left reversible, embeds into a group, and J(S) is independent, then S is left amenable if and only if the canonical map C ∗ s (S) → C ∗ r (S) is an isomorphism. Note that to recover this formulation of the result from from Li's statement, one has to use the fact that when S embeds into a group, S is left reversible if and only if there is a character on C ∗ s (S). This is proved in Li's Lemma 4.6. One also has to use that since S is left reversible, S is cancellative if and only if S embeds into a group. From Milan's article [19], we know that an E-unitary inverse semigroup P has weak containment if and only if G(P ) is amenable. Hence Theorem 3.4.3 and Corollary 3.4.6 give us: Theorem 3.5.1 A cancellative left reversible semigroup S is left amenable if and only if Il(S) has weak containment. This lets us recover Li's result. Corollary 3.5.2 Suppose S is left reversible, embeds into a group, and J(S) is independent. Then S is left amenable if and only if the canonical map C ∗ r (S) is an isomorphism. s (S) → C ∗ Proof. Theorem 3.4.3 and Theorem 3.2.14 imply together that h is an iso- morphism. Proposition 3.3.1 shows that κ is an isomorphism. Theorem 3.5.1 shows that when S is left reversible, Λ is an isomorphism if and only 28 MAGNUS DAHLER NORLING r (S). if S is left amenable. The composition of κ, Λ and h is the canonical map C ∗ s (S) → C ∗ (cid:3) Corollary 3.5.3 Suppose S is cancellative, left reversible, and satisfies Clif- ford's condition. Then the canonical map C ∗(S) → C ∗ r (S) is an isomorphism if and only if S is left amenable. Proof. By Corollary 3.4.2, h is an isomorphism, and by Proposition 3.3.2, η is an isomorphism. Since S is left reversible, Theorem 3.5.1 implies that S is left amenable if and only if Λ is an isomorphism. The composition of η, Λ and h is the canonical map C ∗(S) → C ∗ (cid:3) Remark 3.5.4 Corollary 3.4.8 does not imply that left amenability of S is equivalent to weak containment of Il(S) for any left reversible S. Without Il(S) being E-unitary, one also has to prove that the inverse semigroup H := α−1 S (1G(S)) has weak containment (see Theorem 2.4 and Corollary 2.5 in [19]). Milan's results do however give us that weak containment of Il(S) implies left amenability of G(S) and thus of S (for left reversible S). r (S). When S is not left reversible, S can't be left amenable, but Λ0 : C ∗ 0 (Il(S)) → C ∗ r,0(Il(S)) can still be an isomorphism. Nica shows in [22] that his version of n are canonically isomorphic. Here F+ the full and reduced C ∗-algebras for F+ n is the free semigroup on n generators. This implies that Λ0 is an isomorphism in this case. F+ n is easily seen to be not left reversible for n > 1. Milan [19] has developed a technique for determining weak containment of strongly E∗-unitary inverse semigroups P . Fixing an idempotent pure grading ϕ : P → G0, he defines Ag = span{p ∈ P : ϕ(p) = g} inside CP/C0P Bg = Ag inside C ∗ 0 (P ) Milan then shows that {Bg}g∈G is a Fell bundle over G and that P has weak containment if and only if this Fell bundle is amenable. Milan states this result for the universal grading of P , but the proof works for any idempotent pure grading. In our setting, Il(S) is strongly E∗-unitary if and only if S embeds into a group G. Recaling the idempotent pure grading ϕ : Il(S)0 → G0 constructed in Proposition 3.2.11 one sees that the associated Fell bundle {Bg}g∈G is given by (3.5) Bg = span{λ∗ Theorem 3.5.5 Suppose S embeds into a group G. Then Λ0 : C ∗ 0 (Il(S)) → C ∗ r,0(Il(S)) is an isomorphism if and only if the Fell bundle {Bg}g∈G defined in equation (3.5) is amenable. 1 s1 ··· t−1 t1λs1 ··· λ∗ n sn = g} tnλsn : t−1 in C ∗ 0 (Il(S)) Corollary 3.5.6 Suppose S embeds into a group and satisfies Clifford's condition. Then the canonical map C ∗(S) → C ∗ r (S) is an isomorphism if and only if the Fell bundle {Bg}g∈G given by t : st−1 = g} (3.6) Bg = span{λsλ∗ 0 (Il(S)) in C ∗ is amenable. C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 29 Proof. For semigroups satisfying Clifford's condition, one can use Proposi- tion 3.2.15 to deduce that the Fell bundles defined in equations (3.5) and (3.6) are the same. (cid:3) When (G, S) is a quasilattice ordered group, this expresses Nica's amenabil- ity of (G, S) in terms of the amenability of the Fell bundle defined in eq. (3.6), and is in view of Lemma 3.3.3 merely a restatement of Proposition 5.2 i [19]. In the case where S is a finitely generated free semigroup, amenability of the Fell bundle defined in eq. (3.6) may be deduced from [10]. However the proof one would thereby get from Corollary 3.5.6 that C ∗(S) → C ∗ r (S) is an isomorphism would not be simpler than Nica's original proof [22]. Li [17] shows that C ∗(S) and C ∗ r (S) are nuclear when S is countable, can- cellative and right amenable. The last two conditions imply that S embeds into an amenable group. We show that S does not have to be countable to prove that C ∗ s (S) is nuclear. Proposition 3.5.7 Suppose S embeds into an amenable group. Then Λ0 is an isomorphism, and C ∗ r (S) are nuclear. 0 (Il(S)) ≃ C ∗ s (S) and C ∗ Proof. From Theorem 4.7 in [9] we know that a Fell bundle over an amenable group satisfies the approximation property, and is thus amenable. Moreover, it was proved in [1] that a Fell bundle with nuclear unit fiber has nuclear cross-sectional C ∗-algebra if it also satisfies the approximation property. The unit fiber in {Bg}g∈G is the closure of the span of J(S) in C ∗ 0 (Il(S)), and is abelian. C ∗ 0 (Il(S)) (see Theorem 10.1.3 in [3]). (cid:3) r (S) is also nuclear since it is a quotient of C ∗ Corollary 3.5.8 Let S be cancellative and left reversible. Then S is left amenable if and only if C ∗ s (S) is nuclear. If in addition J(S) is independent, S is left amenable if and only if C ∗ 0 (Il(S)) ≃ C ∗ r (S) is nuclear. 0 (Il(S)) is nuclear implies S is left amenable. This was shown for C ∗ Proof. One implication follows from Proposition 3.5.7 since a cancellative left amenable S embeds into the amenable group G(S). It remains to see that C ∗ s (S) in Proposition 4.17 in [17] with an argument analogous to the following: C ∗(G(S)) is a quotient of C ∗ 0 (Il(S)) and is thus nuclear. This implies that G(S) is amenable and that S is left amenable. (See Theorem 10.1.3 and 2.6.8 in [3]). Assume that J(S) is independent and C ∗ r (S) is nuclear. By Theorem 3.2.14 C ∗ 0,r(Il(S)) and is thus nuclear. This implies that G(S) is amenable (Theorem 2.6.8 in [3]). (cid:3) r (G(S)) is a quotient of C ∗ 0,r(Il(S)) ≃ C ∗ r (S). C ∗ Remark 3.5.9 Note that when S is left reversible, but not right cancella- tive, nuclearity of C ∗ 0 (Il(S)) still implies that S is left amenable. Corollary 3.5.10 Let R be a GCD domain. Then the canonical ∗-homomorphism C ∗(R ⋊ R×) → C ∗ Proof. As Li remarks [17], R ⋊ R× embeds into an amenable group, so Λ0 0 (Il(R ⋊ R×)) is nuclear by Proposition 3.5.7. By is an isomorphism and C ∗ r (R ⋊ R×) is an isomorphism and C ∗(R ⋊ R×) is nuclear. 30 MAGNUS DAHLER NORLING Corollary 3.2.17, h is an isomorphism, and by Proposition 3.3.2 and 2.3.9, η is an isomorphism. (cid:3) Note that if one also assumes that R is a Dedekind domain, the previous corollary is weaker than the results given in [7, 17] since not all rings of integers or Dedekind domains are GCD domains. References [1] Fernando Abadie-Vicens. Tensor products of Fell bundles over discrete groups. arXiv:funct-an/9712006v1, 1997. [2] Nicolas Bourbaki. Commutative Algebra. Hermann, Paris, France, 1972. [3] Nathaniel P. Brown and Narutaka Ozawa. C ∗-algebras and Finite Di- mensional Approximations, volume 88 of Graduate Studies in Mathe- matics. Amer. Math. Soc., Providence, RI, 2008. [4] Scott T. Chapman and Sarah Glaz. Non-Noetherian Commutative Ring Theory. Kluwer Academic Publishers, Dordrecht, The Netherlands, 2000. [5] A. H. Clifford. A class of d-simple semigroups. Amer. J. Math., 75:547 -- 556, 1953. [6] A. H. Clifford and G. B. Preston. The algebraic theory of semigroups. Number 7 in Mathematical Surveys. Amer. Math. Soc., Providence, RI, 1961. Two volumes. [7] Joachim Cuntz, Christopher Deninger, and Marcelo Laca. C ∗- algebras of Toeplitz type associated with algebraic number fields. arXiv:1105.5352v1, 2011. [8] J. Duncan and Alan L.T. Paterson. C ∗-algebras of inverse semigroups. Proc. Edinburgh Math. Soc., 28:41 -- 58, 1985. [9] Ruy Exel. Amenability for fell bundles. Journal fur die reine und ange- wandte Mathematik, 1997:41 -- 74, 1997. [10] Ruy Exel. Partial representations and amenable Fell bundles over free groups. Pacific J. Math, (192):39 -- 63, 2000. [11] Ruy Exel. Inverse semigroups and combinatorial C ∗-algebras. Bull. Braz. Math. Soc, 39:191 -- 313, 2008. [12] Ruy Exel. Tight representations of semilattices and inverse semigroups. Semigroup forum, 79:159 -- 182, 2009. [13] Thomas W. Hungerford. Algebra. Springer Science+Business Media, Inc., New York, NY, 1974. [14] Zhonghao Jiang. The structure of 0-bisimple strongly E∗-unitary inverse monoids. Semigroup Forum, 67:50 -- 62, 2003. [15] Mark V. Lawson. Constructing inverse semigroups from category ac- tions. Journal of pure and applied algebra, 137:57 -- 101, 1997. [16] Mark V. Lawson. Inverse semigroups: the theory of partial symmetries. World Scientific Publishing, River Edge, NJ, 1999. [17] Xin Li. Semigroup C ∗-algebras and amenability of semigroups. arXiv:1105.5539v2, 2011. [18] John Meakin. Groups and semigroups: connections and contrasts. http://www.math.unl.edu/ jmeakin2/groups and semigroups.pdf, 2011. Submitted for publication. C ∗-ALGEBRAS OF LEFT CANCELLATIVE SEMIGROUPS 31 [19] David Milan. C ∗-algebras of inverse semigroups: amenability and weak containment. J. Operator Theory, 63:317 -- 332, 2010. [20] Theodore Mitchell. Constant functions and left invariant means on semi- groups. Trans. Am. Math. Soc., 119:244 -- 261, 1965. [21] G. J. Murphy. C ∗-algebras generated by commuting isometries. Rocky Mountain Journal of Mathematics, 26:237 -- 267, 1996. [22] A. Nica. C ∗-algebras generated by isometries and Wiener-Hopf opera- tors. J. Operator Theory, 27:17 -- 52, 1992. [23] A. Nica. On a groupoid construction for actions of certain inverse semi- groups. Internat. J. Math., 5:349 -- 372, 1994. [24] Øystein Ore. Linear equations in non-commutative fields. Ann. of Math., 32:463 -- 477, 1931. [25] Alan L. T. Paterson. Weak containment and Clifford semigroups. Proc. Roy. Soc. Edinburgh, 81A:23 -- 30, 1978. [26] Alan L. T. Paterson. Amenability. Number 29 in Mathematical Surveys and Monographs. Amer. Math. Soc., Providence, RI, 1988. [27] Alan L. T. Paterson. Groupoids, Inverse Semigroups, and their Operator algebras, volume 170 of Progress in Mathematics. Birkhauser, Boston, MA, 1999. E-mail address: [email protected]
1309.2145	5	1309	2013-11-25T13:54:11	A nonseparable amenable operator algebra which is not isomorphic to a C*-algebra	[ "math.OA", "math.LO" ]	It has been a longstanding problem whether every amenable operator algebra is isomorphic to a (necessarily nuclear) C-algebra. In this note, we give a nonseparable counterexample. The existence of a separable counterexample remains an open problem. We also initiate a general study of unitarizability of representations of amenable groups in C-algebras and show that our method cannot produce a separable counterexample.	math.OA	math	Forum of Mathematics, Sigma (2018), vol. 1, e1, 11 pages doi:10.111 1 A NONSEPARABLE AMENABLE OPERATOR ALGEBRA WHICH IS NOT ISOMORPHIC TO A C∗-ALGEBRA YEMON CHOI, ILIJAS FARAH and NARUTAKA OZAWA † Department of Mathematics and Statistics University of Saskatchewan McLean Hall, 106 Wiggins Road Saskatoon, Saskatchewan, Canada S7N 5E6 Department of Mathematics and Statistics, York University, 4700 Keele Street, North York, Ontario, Canada, M3J 1P3, and Matematicki Institut, Kneza Mihaila 35, Belgrade, Serbia RIMS, Kyoto University, 606-8502 Japan Abstract It has been a longstanding problem whether every amenable operator algebra is isomorphic to a (necessarily nuclear) C∗-algebra. In this note, we give a nonseparable counterexample. The existence of a separable counterexample remains an open problem. We also initiate a general study of unitarizability of representations of amenable groups in C∗-algebras and show that our method cannot produce a separable counterexample. 2010 Mathematics Subject Classification: 47L30; 46L05 1. Introduction The notion of amenability for Banach algebras was introduced by B. E. Johnson ([Jo72]) in 1970s and has been studied intensively since then (see a more recent monograph [Ru02]). For several natural classes of Banach algebras, the amenability property is known to single out the "good" members of those classes. For example, B. E. Johnson's fundamental observation ([Jo72]) is that the Banach algebra L1(G) of a locally compact group G is amenable if and only † Corresponding author: Narutaka Ozawa, email: [email protected] c(cid:13) The Author(s) 2018. The online version of this article is published within an Open Access environment subject to the conditions of the Creative Commons Attribution licence <http://creativecommons.org/licenses/by/3.0/>. Y. Choi, I. Farah and N. Ozawa 2 if the group G is amenable. Another example is the celebrated result of Connes ([Co78]) and Haagerup ([Ha83]) which states that a C∗-algebra is amenable as a Banach algebra if and only if it is nuclear. In this paper, we are interested in the class of operator algebras. By an op- erator algebra, we mean a (not necessarily self-adjoint) norm-closed subalgebra of B(H), the C∗-algebra of the bounded linear operators on a Hilbert space H. It has been asked by several researchers whether every amenable operator alge- bra is isomorphic to a (necessarily nuclear) C∗-algebra. The problem has been solved affirmatively in several special cases: for subalgebras of commutative C∗- algebras ([Se77]), and subsequently for operator algebras generated by normal elements ([CL95]); for subalgebras of compact operators ([Wi95, Gi06]); for 1- amenable operator algebras (Theorem 7.4.18 in [BL04]); and for commutative subalgebras of finite von Neumann algebras ([Ch13]). Here we give the first counterexample to the above problem. In fact, our counterexample is a subalgebra of the homogeneous C∗-algebra (cid:96)∞(N, M2). Hence the result of [Se77] is actually quite sharp and the result of [Ch13] does not generalize to an arbitrary subalgebra of a finite von Neumann algebra. Theorem 1. There is a unital amenable operator algebra A which is not isomor- phic to a C∗-algebra. The algebra A is a subalgebra of (cid:96)∞(N, M2) with density character ℵ1, and is an inductive limit of unital separable subalgebras {Ai}i<ℵ1, each of which is conjugated to a C∗-subalgebra of (cid:96)∞(N, M2) by an invertible element vi ∈ (cid:96)∞(N, M2), such that supi (cid:107)vi(cid:107)(cid:107)v−1 i (cid:107) < ∞. Moreover, for any ε > 0, one can choose A to be (1 + ε)-amenable. Here, C-amenable means that the amenability constant is at most C (see Definition 2.3.15 in [Ru02]). One drawback of our counterexample is that it is inevitably nonseparable, as explained by Theorem 8 below, and the existence Neumann algebra (cid:81)∞ of a separable counterexample remains an open problem. We note that if such an example exists, then there is one among subalgebras of the finite von (cid:76)∞ subalgebra of(cid:81)∞ Indeed, by Voiculescu's theorem ([Vo91]), the cone C0((0, 1],A) of a separable operator algebra A can be realized as a closed 2.3.6 in [Ru02]), and its preimage A in(cid:81)∞ Mn. The cone of A is amenable (see Exercise the amenable algebra(cid:76)∞ Mn is an extension of the cone by Mn, hence A is amenable (see Theorem 2.3.10 in A is not isomorphic to a C∗-algebra, since it has A as a quotient and [Ru02]). every closed two-sided ideal in a C∗-algebra is automatically ∗-closed. n=1 Mn/ n=1 n=1 Mn. n=1 n=1 An amenable operator algebra 3 Acknowledgment. This joint work was initiated when the second and third au- thors participated in the workshop "C∗-Algebren" (ID:1335) held at the Mathe- matisches Forschungsinstitut Oberwolfach in August 2013. We are grateful to the organizers S. Echterhoff, M. Rørdam, S. Vaes, and D. Voiculescu, and the institute for giving the authors an opportunity of a joint work. We are also grate- ful to N. C. Phillips for useful conversations during the workshop and helpful remarks on the first version of this paper, and the third author would like to thank N. Monod and H. Matui for valuable conversations. Finally, we would like to thank S. A. White for his encouragement to include the last sentence of Theorem 1. The first author was supported by NSERC Discovery Grant 402153-2011. The second author was partially supported by NSERC, a Velux Visiting Pro- fessorship, and the Danish Council for Independent Research through Asger Tornquist's grant, no. 10-082689/FNU. The third author was partially supported by JSPS (23540233). 2. Proof of Theorem 1 Let C be a unital C∗-algebra, Γ be a group, and π: Γ → C be a representation, i.e., π(s) is invertible for every s ∈ Γ and π(st) = π(s)π(t) for every s, t ∈ Γ. The representation π is said to be uniformly bounded if (cid:107)π(cid:107) := sups (cid:107)π(s)(cid:107) < +∞. It is said to be unitarizable if there is an invertible element v in C such that Adv ◦π is a unitary representation. Here Adv(c) = vcv−1 for c ∈ C. The element v is called a similarity element. A well-known theorem of Sz.-Nagy, Day, Dixmier, and Nakamura -- Takeda states that every uniformly bounded representation of an amenable group Γ into a von Neumann algebra is unitarizable. In fact the latter property characterizes amenability by Pisier's theorem ([Pi07]). In particular, the operator algebra span π(Γ) generated by a uniformly bounded representation π of an amenable group Γ is an amenable operator algebra which is isomorphic to a nuclear C∗-algebra. See [Pi01] and [Ru02] for general information about uniformly bounded representations and amenable Banach algebras, respectively. Let us fix the notation. Let M2 be the 2-by-2 full matrix algebra, (cid:96)∞(N, M2) be the C∗-algebra of the bounded sequences in M2, and c0(N, M2) be the ideal of the sequences that converge to zero. We shall freely identify (cid:96)∞(N, M2) with (cid:96)∞(N) ⊗ M2, and (cid:96)∞(N, M2)/c0(N, M2) with C(N) ⊗ M2, where C(N) = (cid:96)∞(N)/c0(N). The quotient map from (cid:96)∞(N) (or (cid:96)∞(N) ⊗ M2) onto C(N) (or C(N) ⊗ M2) is denoted by Q. Lemma 2. Let Γ be an abelian group and π: Γ → C(N) ⊗ M2 be a uniformly bounded representation. Then the amenable operator algebra A := Q−1(span π(Γ)) ⊂ (cid:96)∞(N, M2) is isomorphic to a C∗-algebra if and only if π is unitarizable. Y. Choi, I. Farah and N. Ozawa 4 Proof. First of all, we observe that the operator algebra A is indeed amenable because it is an extension of an amenable Banach algebra span π(Γ) by the amenable Banach algebra c0(N, M2) (see Theorem 2.3.10 in [Ru02]). Suppose now that π is unitarizable and v ∈ C(N) ⊗ M2 has the property Adv ◦π is unitary. We may assume v is positive, by taking the positive component from its polar decomposition. Since v is invertible, we can choose a representing sequence vm, for m ∈ N of v such that each vm is positive and moreover 1/(cid:107)v−1(cid:107) ≤ vm ≤ (cid:107)v(cid:107) m (cid:107) ≤ (cid:107)v(cid:107)(cid:107)v−1(cid:107) for all m. for all m. In particular each vm is invertible and (cid:107)vm(cid:107)(cid:107)v−1 Now we have a representing sequence of an invertible lift v ∈ (cid:96)∞(N, M2) of v such that (cid:107)v(cid:107)(cid:107)v−1(cid:107) = (cid:107)v(cid:107)(cid:107)v−1(cid:107). Then vAv−1 = Q−1(span(Adv ◦π(Γ))) is a self- adjoint C∗-subalgebra of (cid:96)∞(N, M2). Conversely, suppose that A is isomorphic to a C∗-algebra, which is necessarily nuclear. Then thanks to the solution of Kadison's similarity problem for nuclear C∗-algebras (see Theorem 7.16 in [Pi01] or Theorem 1 in [Pi07]), there is v in the von Neumann algebra (cid:96)∞(N, M2) such that vAv−1 is a C∗-subalgebra. Let v = Q(v) ∈ C(N) ⊗ M2. Since Q(vAv−1) is a commutative C∗-subalgebra of C(N) ⊗ M2, for every s ∈ Γ, the element vπ(s)v−1 is normal with its spectrum in the unit circle, which implies that vπ(s)v−1 is unitary. The above proof uses the fact that every (not necessarily separable) amenable C-algebra is nuclear, as well as the solution to Kadison's similarity problem for nuclear C-algebras. The reader may appreciate a more elementary and self- contained proof. Assume θ is a bounded homomorphism of a unital C-algebra A into (cid:96)∞(N, M2). We need to prove that θ is similar to a -homomorphism. It suffices to show that every coordinate map is similar to a -homomorphism and that the similarities are implemented by a uniformly bounded sequence vn, for n ∈ N, of operators. Consider the restriction of θ to the unitary group G of A. At the n-th coordinate we have a bounded homomorphism from G to GL(2, C). Since a bounded subgroup of GL(2, C) is included in a compact subgroup, by a standard averaging argument we find vn such that Advn ◦θ is a unitary representation of G. The operators vn are easily seen to satisfy the required properties. (cid:104) 1 0 (cid:105) (cid:104) 1 0 2 , (cid:107)v(cid:107)(cid:107)v−1(cid:107) ≤ C} > 0 Proof of Theorem 1. We consider two 2-by-2 order 2 invertible matrices which are not simultaneously unitarizable. For instance, let s0 = . Then by compactness, one has ε(C) := inf{d(vs0v−1,U) + d(vs1v−1,U) : v ∈ M−1 (cid:105) 0 −1 and s1 = 1 −1 for every C > 0. Here U denotes the unitary group of M2. i i i ∩ El An amenable operator algebra : i ∈ ℵ1} and {E1 We shall need two families {E0 5 : i ∈ ℵ1} of subsets of N j is finite whenever (i, k) (cid:44) ( j, l) and (ii) these two families such that (i) Ek i \ F are not separated, in the sense that there is no F ⊆ N such that both E0 i ∩ F are finite for all i. The existence of such pair of families follows and E1 from [Lu47]. Luzin actually proved much more: he constructed a single family {Ei : i < ℵ1} of infinite subsets of N such that (i) Ei ∩ E j is finite whenever i (cid:44) j and (ii) whenever X ⊆ ℵ1 is such that both X and ℵ1 \ X are uncountable, then the families {Ei : i ∈ X} and {Ei : i ∈ ℵ1 \ X} cannot be separated (see Appendix B below for Luzin's proof). ) ∈ C(N) are mutually orthogonal. For each pair The projections pk i = Q(1Ek (i, k), we define sk i in C(N) ⊗ M2 by sk i = pk i∈ℵ1,k∈{0,1} Z/2Z and {ek Let Γ :=(cid:76) i ) ⊗ 1. i ⊗ sk + (1 − pk i} be its standard basis. Then the map ek (cid:55)→ sk extends to a uniformly bounded representation π: Γ → C(N) ⊗ M2 such that (cid:107)π(cid:107) = max{(cid:107)s0(cid:107),(cid:107)s1(cid:107)}. We claim that π is not unitarizable. Suppose for a contradiction that there is an invertible element v ∈ C(N) ⊗ M2 such that Adv ◦π is unitary. As in the proof of Lemma 2 we may assume v is positive and find a representing sequence vm, for m ∈ N, of an invertible lift of v such that (cid:107)vm(cid:107)(cid:107)v−1 m ,U) < ε/2}, and note that this set is disjoint i \ F0 is from F1 := {m : d(vms1v−1 infinite or such that E1 m ,U) < ε/2}. Therefore we have i such that E0 i \ F1 is infinite. If the former case applies, then m (cid:107) ≤ (cid:107)v(cid:107)(cid:107)v−1(cid:107) for all m. Let ε = ε((cid:107)v(cid:107)(cid:107)v−1(cid:107)). Now let F0 := {m : d(vms0v−1 i i i d(vns0v−1 n ,U) ≥ ε/2, lim sup n∈E0 i , n→∞ i \F1 is infinite contradicting the assumption that v unitarizes π. The case when E1 similarly leads to a contradiction. Thus, by Lemma 2, the preimage of span π(Γ) in (cid:96)∞(N, M2) is an amenable operator algebra which is not isomorphic to a C∗- algebra. Its density character is equal to ℵ1 = Γ. Let Γi be a countable subgroup of Γ and denote the separable algebra Q−1(span π(Γi)) by Ai. Theorem 8 below shows that Ai is similar inside (cid:96)∞(N, M2) to an amenable C∗-algebra, with a similarity element vi satisfying (cid:107)vi(cid:107)(cid:107)v−1 i (cid:107) ≤ (cid:107)π(cid:107)2. Furthermore, since every amenable C∗-algebra is 1-amenable by results of Haagerup ([Ha83]), Ai is (cid:107)π(cid:107)4-amenable. Now A is the inductive limit of the family (Ai) as Γi varies over all countable subgroups of Γ. Since each Ai is (cid:107)π(cid:107)4-amenable, a routine argument with approximate diagonals shows that (cid:105) A is also (cid:107)π(cid:107)4-amenable: for details see Proposition 2.3.17 in [Ru02]. Finally, we explain how our example can be modified to have arbitrarily but replace s1 small amenability constant. For 0 < t < 1, we keep s0 = (cid:104) 1 0 0 −1 (cid:104) 1 0 (cid:105) Y. Choi, I. Farah and N. Ozawa 6 t −1 with s1(t) = in our original construction. Denoting the resulting algebra by A(t), the previous arguments show that A(t) is (cid:107)s1(t)(cid:107)4-amenable, and (cid:107)s1(t)(cid:107) can be made arbitarily close to 1. We note that a set-theoretical study of the cohomological nature of gaps similar to Luzin's was initiated in [Ta95]. 3. Unitarizability of uniformly bounded representations In this section, we develop a general study of (non-)unitarizability. First, we shall deal with separable C∗-algebras. Let A be a unital C∗-algebra and θ be a ∗-automorphism on A. An element a ∈ A is called a cocycle if it satisfies (cid:107) n−1(cid:88) k=0 a := sup n≥1 θk(a)(cid:107) < +∞. It is inner (or a coboundary) if there is x ∈ A such that a = x − θ(x). We recall that the first bounded cohomology group (see [Mo01]) of the Z-module (A, θ) is defined as b(A, θ) = { cocycles }/{ inner cocycles }. H1 b is trivial (see Theorem 2.6 in [Or00]). We note that every cocycle is approximately inner. When A is abelian and θ corresponds to a minimal homeomorphism of its (cid:80)n−1 k=0 θk(a) satisfies an+1 = a + θ(an), the element xn := n−1(cid:80)n spectrum then H1 Indeed, since an := θ = Adu for a unitary element u ∈ A, and a ∈ A is a cocycle. Then, t =(cid:2) u au (cid:3) m=1 am satisfies (cid:107)xn(cid:107) ≤ a and (cid:107)a−(xn−θ(xn))(cid:107) ≤ 2n−1a. Suppose for a moment that θ is inner, is an invertible element in M2(A) such that tn = for n ≥ 1. Therefore supn∈Z (cid:107)tn(cid:107) ≤ 1 + a and t gives rise to a uniformly bounded representation πa of Z into A. Lemma 3. Let A, u, a, and πa as above. Then the uniformly bounded represen- tation πa is unitarizable if and only if a is inner. (cid:104) un anun (cid:105) 0 u 0 un See Lemma 4.5 in [Pi01] or [MO10] for the proof of this lemma. Proposition 4. Let A be a unital separable C∗-algebra and θ be a ∗-automorphism of A. Suppose that there are a (non-unital) θ-invariant C∗-subalgebra A0, a state φ on A0, and a sequence of natural numbers n(k) such that (φ ◦ θn(k))∞ converges to 0 pointwise on A0. Then, H1 b(A, θ) (cid:44) 0. k=1 An amenable operator algebra 7 Proof. By a standard Hahn -- Banach convexity argument, we construct an ap- n=0 of A0 such that 0 ≤ hn ≤ 1, hn+1hn = hn, and (cid:107)hn−θ(hn)(cid:107) < proximate unit (hn)∞ 2−n for all n. We note that φ(cid:48)(hn) → 1 for any state φ(cid:48) on A0. Taking a state ex- tension, we may assume that φ is defined on A. Since A is separable, passing to a subsequence, we may assume that φk := φ ◦ θn(k) converges pointwise to a state, say ψ, on A. (m( j))∞ every i ≤ j and φk( j+1)(hm( j)) < 2− j for every j. Let Set k(1) = 1. By induction, one can find strictly increasing sequences j=1 of natural numbers such that φk(i)(hm( j)) > 1 − 2− j for j=1 and (k( j))∞ ∞(cid:88) x = SOT- (hm(2 j) − hm(2 j−1)) ∈ A∗∗ . j=1 We extend θ and φ on A∗∗ by ultraweak continuity. One has a := x − θ(x) ∈ A, since it is a norm-convergent series in A0. By a telescoping argument, a is a cocycle. Suppose for the sake of obtaining a contradiction that a is inner and x−θ(x) = y − θ(y) for some y ∈ A. Then, y ∈ A and θ(x − y) = x − y. It follows that φk( j)(y) → ψ(y) and φk( j)(x − y) = φ(x − y). Hence the sequence (φk( j)(x))∞ converges. However, for j ≥ 1, j=1 φk(2 j)(x) ≥ φk(2 j)(hm(2 j) − hm(2 j−1)) ≥ 1 − 1 22 j − 1 22 j−1 and φk(2 j+1)(x) ≤ φk(2 j+1)( j(cid:88) i=1 ∞(cid:88) hm(2i)) + i= j+1 (1 − φk(2 j+1)(hm(2i−1))) ≤ 1 4 . Hence, the sequence (φk( j)(x))∞ j=1 does not converge, and we have a contradiction. An example of A0, φ and θ as in the statement of Proposition 4 are the ideal K of compact operators on B((cid:96)2(Z)), any one of its states, and the bilateral shift on (cid:96)2(Z). Lemma 5. For every unital separable C∗-algebra A which is not of type I, there is a unitary element u ∈ A such that H1 Proof. Let z be the bilateral shift on (cid:96)2(Z) and take a selfadjoint element √−1h). Let C ⊂ B((cid:96)2(Z)) be the unital C∗- h ∈ B((cid:96)2(Z)) such that z = exp( subalgebra generated by K and h, and let φ0 be the vector state at δ0. Since C is an extension of a commutative C∗-algebra by K, it is nuclear. By Kirchberg's b(A, Adu) (cid:44) 0. Y. Choi, I. Farah and N. Ozawa 8 theorem and Glimm's theorem in tandem (Corollary 1.4(vii) in [Ki95]), there are a unital C∗-subalgebra A1 of A and a surjective ∗-homomorphism π from √−1g) ∈ A1, A1 onto C. Let g ∈ A1 be a selfadjoint lift of h and let u := exp( which is a unitary lift of z. Then A0 = π−1(K) is an Adu-invariant subalgebra and the state φ = φ0 ◦ π satisfies φ ◦ (Adu)n → 0 pointwise on A0. Hence the result follows from Proposition 4. Combining Lemma 5 and Lemma 3, we arrive at the following theorem. Theorem 6. For every unital separable C∗-algebra A which is not of type I, there is a uniformly bounded representation of Z into M2(A) which is not unitarizable. Now, we shall deal with non-separable C∗-algebras. Our approach uses model theory of metric structures and the extension of Pedersen's techniques [Pe88] as presented in [FH13]. The following is Definition 1.1 from [FH13], with a misleading typo corrected. Definition 7. Given a C∗-algebra M, a degree-1 -polynomial with coefficients in M is a linear combination of terms of the form axb, ax∗b and a with a, b in M. A C∗-algebra M is said to be countably degree-1 saturated if for every countable family of degree-1 ∗-polynomials Pn( ¯x) with coefficients in M and variables xm, for m ∈ N, and every family of compact sets Kn ⊂ R, for n ∈ N, the following are equivalent (writing ¯b for (b1, b2, . . .) and M≤1 for the closed unit ball of M). 1. There are bm ∈ M≤1, for m ∈ N, such that (cid:107)Pn(¯b)(cid:107) ∈ Kn for all n. 2. For every N ∈ N there are bm ∈ M≤1, for m ∈ N, such that dist((cid:107)Pn(¯b)(cid:107), Kn) ≤ 1 N for all n ≤ N. A type {Pn( ¯x) ∈ Kn : n ∈ N} satisfying (1) is said to be realized in M and a type satisfying (2) is said to be consistent with (or approximately finitely realized in) M. Coronas of σ-unital C∗-algebras, in particular the Calkin algebra Q((cid:96)2) and C(N) ⊗ M2, as well as ultraproducts associated with nonprincipal ultrafilters on N, are countably degree-1 saturated ([FH13, Theorem 1.4]). In each of these cases, given a consistent type, a realization ¯b is assembled from the approximate realizations ¯bn, for n ∈ N and a carefully chosen, appropriately quasicentral n(en − en+1)1/2 ¯bn(en − en+1)1/2. See [FH13] for details and more examples of countably degree-1 saturated C∗-alge- bras. approximate unit en, for n ∈ N, as ¯b = (cid:80) An amenable operator algebra 9 Theorem 8. Let M be a unital countably degree-1 saturated C∗-algebra. Then, every uniformly bounded representation π: Γ → M of a countable amenable group Γ into M is unitarizable. Moreover a similarity element v can be chosen so that it satisfies (cid:107)v(cid:107)(cid:107)v−1(cid:107) ≤ (cid:107)π(cid:107)2. Proof. The proof is analogous to the standard one (see Theorem 0.6 in [Pi01]), modulo applying countable degree-1 saturation. Consider the type in variable x over M consisting of conditions (cid:107)x − x∗(cid:107) = 0, (cid:107)x(cid:107) ≤ (cid:107)π(cid:107)2, (cid:107)(cid:107)π(cid:107)2 − x(cid:107) ≤ (cid:107)π(cid:107)2 − (cid:107)π(cid:107)−2, and (cid:107)π(s)xπ(s)∗ − x(cid:107) = 0 for all s ∈ Γ. We now check that this type is consistent. Let (Fn)∞ n=1 be a Følner sequence of finite subsets of Γ. Then, (cid:88) t∈Fn hn = 1 Fn π(t)π(t)∗ , are positive elements in M such that (cid:107)π(cid:107)−2 ≤ hn ≤ (cid:107)π(cid:107)2 and (cid:107)π(cid:107)2 → 0 (cid:107)π(s)hnπ(s)∗ − hn(cid:107) ≤ Fn (cid:52) sFn Fn for every s ∈ Γ. Hence this type is consistent and by countable degree-1 saturation there is h ∈ M which realizes it. Therefore we have h = h∗, (cid:107)h(cid:107) ≤ (cid:107)π(cid:107)2, (cid:107)(cid:107)π(cid:107)2 − h(cid:107) ≤ (cid:107)π(cid:107)2 − (cid:107)π(cid:107)−2, and π(s)hπ(s)∗ = h for every s ∈ Γ. It follows that h is a positive element such that (cid:107)π(cid:107)−2 ≤ h ≤ (cid:107)π(cid:107)2 and the invertible elements h−1/2π(s)h1/2 satisfy (h−1/2π(s)h1/2)(h−1/2π(s)h1/2)∗ = h−1/2π(s)hπ(s)∗h−1/2 = 1, i.e., h−1/2π(s)h1/2 are unitary. Theorem 8 shows that the method used in the proof of Theorem 1 cannot be used to produce a separable counterexample. Appendix A. A correction for [Ch13] as a direct product(cid:81) We take the opportunity to fill a small gap in [Ch13]. The main result of that paper is only proved for commutative amenable subalgebras of σ-finite finite von Neumann algebras. It is then stated in [Ch13] that the general case follows from the σ-finite one because any finite von Neumann algebra M decomposes i Mi where each Mi is σ-finite. However, the example of the present paper shows that similarity to a C∗-algebra is not preserved by taking inductive limits, even with a uniform bound on the similarity elements, so more justification is needed. Instead, we may argue as follows. Let A be an amenable subalgebra of M and let Ai be its image under the projection M → Mi. Applying the main result of [Ch13] to each Ai, we obtain a uniformly bounded family vi ∈ Mi such that viAiv−1 is a commutative C∗-subalgebra of Mi. Take v to be the direct product of the vi. Then vAv−1 is an amenable subalgebra of the commutative C∗-algebra(cid:81) , and hence by [Se77] it is self-adjoint. i viAiv−1 i i Y. Choi, I. Farah and N. Ozawa 10 Appendix B. A construction of Luzin's gap For the reader's convenience we prove Luzin's theorem. Following von Neumann, we identify n ∈ N with the set {0, 1, . . . , n − 1}. We construct a family Ei, for i < ℵ1, of infinite subsets of N such that 1. Ei ∩ E j is finite whenever i (cid:44) j, and 2. for every i and every m ∈ N the set { j < i : E j ∩ Ei ⊆ m} is finite. Now let k(0) = 0 and k(n) = min Fn \ (k(n − 1) ∪(cid:83) The construction is by recursion. For a finite i let Ei = {2i(2k + 1) : k ∈ N}. Assume i < ℵ1 is infinite and the sets E j, for j < i were chosen to satisfy the requirements. Since i is countable, we can re-enumerate E j, for j < i as Fn, for n ∈ N. l<n Fl) for n ≥ 1. The sequence {k(n)} is strictly increasing and k(n) ∈ Fl implies n ≤ l. Therefore Ei = {k(n) : n ∈ N} is infinite and Ei ∩ Fn ⊆ {k(0), . . . , k(n)} is finite for all n. Finally, for any m ∈ N the set {n ∈ N : Fn ∩ Ei ⊆ m} ⊆ {n : k(n) < m} is finite. This describes the recursive construction of a family Ei, for i < ℵ1, satisfying (1) and (2). We claim that for any X ⊆ ℵ1 such that X and ℵ1 \ X are uncountable the families {Ei : i ∈ X} and {Ei : i ∈ ℵ1\X} cannot be separated. Assume otherwise, and fix F ⊆ N separating them. Since Ei \ F is finite for all i ∈ X, there is m ∈ N such that X(cid:48) = {i ∈ X : Ei \ F ⊆ m} is uncountable. By increasing m if necessary we can assure that Y(cid:48) = {i ∈ ℵ1 \ X : Ei ∩ F ⊆ m} is also uncountable. Pick i ∈ Y(cid:48) such that X(cid:48)(cid:48) = { j ∈ X(cid:48) : j < i} is infinite. Then for each j ∈ X(cid:48)(cid:48) we have E j ∩ Ei ⊆ (E j \ F) ∪ (Ei ∩ F) ⊆ m. But this contradicts (2). References [BL04] D. P. Blecher and C. Le Merdy; Operator algebras and their modules -- an operator space approach. London Mathematical Society Monographs. New Series, 30. Oxford Science Publications. The Clarendon Press, Oxford University Press, Oxford, 2004. [Ch13] Y. Choi; On commutative, operator amenable subalgebras of finite von Neumann 319. algebras. J. Reine Angew. Math. 678 (2013), 201 -- 222. [Co78] A. Connes; On the cohomology of operator algebras. J. Funct. Anal. 28 (1978), 248 -- [FH13] [CL95] 253. P. C. Curtis Jr., and R. J. Loy; A note on amenable algebras of operators. Bull. Austral. Math. Soc. 52 (1995), 327 -- 329. I. Farah and B. Hart; Countable saturation of corona algebras. C. R. Math. Rep. Acad. Sci. Canada 35 (2013), 35 -- 56. J. A. Gifford; Operator algebras with a reduction property. J. Aust. Math. Soc. 80 [Gi06] (2006), 297 -- 315. [Ha83] U. Haagerup; All nuclear C∗-algebras are amenable. Invent. Math. 74 (1983), 305 -- An amenable operator algebra 11 [Jo72] B. E. Johnson; Cohomology in Banach algebras. Memoirs of the American Mathemat- ical Society, No. 127. American Mathematical Society, Providence, R.I., 1972. [Ki95] E. Kirchberg; On subalgebras of the CAR-algebra. J. Funct. Anal. 129 (1995), 35 -- 63. [Lu47] N. Luzin; O qast(cid:31)h natural(cid:126)nogo r(cid:31)da, Izv. AN SSSR, seri(cid:31) mat. 11, (cid:125)5 (1947), 714 -- 722. Available at http://www.mathnet.ru/links/55625359125306fbfba1a9a6f07523a0/im3005.pdf [Mo01] N. Monod; Continuous bounded cohomology of locally compact groups, no. 1758, [MO10] N. Monod and N. Ozawa; The Dixmier problem, lamplighters and Burnside groups. J. [Or00] N. S. Ormes; Real coboundaries for minimal Cantor systems. Pacific J. Math. 195 Springer, 2001. Funct. Anal. 258 (2010), 255 -- 259. (2000), 453 -- 476. [Pe88] G. K. Pedersen; The corona construction, Operator Theory: Proceedings of the 1988 GPOTS-Wabash Conference (Indianapolis, IN, 1988), Pitman Res. Notes Math. Ser., vol. 225, Longman Sci. Tech., Harlow, 1990, pp. 49 -- 92. [Pi01] G. Pisier; Similarity problems and completely bounded maps. Second, expanded edi- tion. Includes the solution to "The Halmos problem". Lecture Notes in Mathematics, 1618. Springer-Verlag, Berlin, 2001. [Pi07] G. Pisier; Simultaneous similarity, bounded generation and amenability. Tohoku Math. J. (2) 59 (2007), 79 -- 99. Verlag, Berlin, 2002. [Ru02] V. Runde; Lectures on amenability. Lecture Notes in Mathematics, 1774. Springer- [Se77] M. V. Seınberg; A characterization of the algebra C(Ω) in terms of cohomology groups. Uspehi Mat. Nauk 32 (1977), 203 -- 204. [Ta95] D. E. Talayco; Applications of Cohomology to Set Theory I: Hausdorff gaps, Annals [Vo91] D. Voiculescu; A note on quasi-diagonal C∗-algebras and homotopy. Duke Math. J. 62 of Pure and Applied Logic 71 (1995), 69 -- 106. [Wi95] G. A. Willis; When the algebra generated by an operator is amenable. J. Operator (1991), 267 -- 271. Theory 34 (1995), 239 -- 249.
1904.03650	1	1904	2019-04-07T13:34:38	Geodesic neighborhoods in unitary orbits of self-adjoint operators of K+C	[ "math.OA" ]	We study the unitary orbit of a compact Hermitian diagonal operator with spectral multiplicity one under the action of the unitary group U_(K+C) of the unitization of the compact operators K(H)+C, or equivalently, the quotient U_(K+C)/ U_Diag(K+C). We relate this and the action of different unitary subgroups to describe metric geodesics (using a natural distance) which join end points. As a consequence we obtain a local Hopf-Rinow theorem. We also explore cases about the uniqueness of short curves and prove that there exist some of these that cannot be parameterized using minimal anti-Hermitian operators of K(H)+C.	math.OA	math	GEODESIC NEIGHBORHOODS IN UNITARY ORBITS OF SELF-ADJOINT OPERATORS OF K ` C TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA2 , 3 Abstract. In the present paper, we study the unitary orbit of a compact Hermitian diagonal operator with spectral multiplicity one under the action of the unitary group UK`C of the unitization of the compact operators KpHq ` C , or equivalently, the quotient UK`C{UDpK`Cq. We relate this and the action of different unitary subgroups to describe metric geodesics (using a natural distance) which join end points. As a consequence we obtain a local Hopf-Rinow theorem. We also explore cases about the uniqueness of short curves and prove that there exist some of these that cannot be parameterized using minimal anti-Hermitian operators of KpHq ` C. 1. Introduction Let A be a C-algebra, B a von Neumann subalgebra of A, and UA, UB its respective unitary groups. Theorem II of [7] or Theorem I-2 of [8] (and the Remark that follows it) imply that for every element ρ P UA{UB, and every tangent vector x P Tρ pUA{UBq, there exist a minimal lift Z P Aah of x (x " Zρ ´ ρZ and }Z} ď }Z ` D} for all D P B) such that the curves (1.1) γptq " etZρe´tZ are all the possible short curves (under a natural distance) starting at ρ with fixed initial velocity x. In this context, we will call this Z a minimal operator. Moreover in [8] local and global Hopf-Rinow theorems were proved in this context with additional hypothesis on the unitary groups involved. If the assumption of B being a von Neumann algebra is relaxed, Theorem I of [7] proves that every minimal lift Z still produces a short curve if B is only required to be a C-algebra. Nevertheless, in this case such a Z may not exist for every tangent vector x (see for example the discussion following Proposition 18 in [3] or the properties of Z2 defined in (4.2) in this paper). Therefore, if A and B are C -algebras the parameterization of minimal curves with arbitrary initial velocity is not known in general nor the existence of geodesic neighborhoods. The main objective of this work is the study of short curves of a particular example where the subalgebra B is not a von Neumann algebra. Denote with K`C the C-algebra obtained after the unitization of the compact operators in BpHq, that is tX P BpHq : X " K ` θI, K P K, θ P Cu. H will be a separable Hilbert space, and UK`C will denote the unitary operators of KpHq ` C. If we fix an orthonormal basis teiuiPN in H, we can consider matricial characterizations in BpHq an diagonal operators. Let b be a compact diagonal self-adjoint operator with spectral multiplicity one, and Ob the orbit (1.2) UK`C Ob " O b " tubu : u P UK`Cu This orbit has a structure of a smooth homogeneous space under the action of UK`C with the iden- tification Ob » UK`C{UDpK`Cq (see for example Lemma 1 of [4] and the discussion that follows it). As we comment in Remark 3.2 the homogeneous space Ob coincides with the orbit of b under the action of several other unitary subgroups. Moreover, a natural Finsler metric defined on the tangent Date: April 9, 2019. 2010 Mathematics Subject Classification. Primary: 22F30, 53C22. Secondary: 22E65, 47B15, 51F25, 58B20. Key words and phrases. unitary orbits, geodesic curves, minimality, Finsler metric. 1 2 TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA2,3 spaces (see 3.1) and a distance (see 3.6) in Ob is also preserved if we consider those different unitary subgroups. In this context, we analyze geodesic neighborhoods of Ob and cases of short curves satisfying initial conditions or connecting given endpoints that cannot obtained using minimal operators V P pKpHq ` Cqah. In Theorem 4.1 it is shown that short curves in Ob of the form (1.1) can be constructed using minimal operators Z P pKpHq ` DpBpHqqqah. These geodesics and a result from [7] allows us to prove in 5.4 a Hopf-Rinow local theorem for Ob. We also consider certain types of minimal operators with diagonal belonging to DpBpHqqzDpKpHq ` Cq and construct with them short curves γ in Ob (for the distance (3.6)) such that in a fixed neighborhood, there is not any curve δ of the form δptq " etV be´tV with V a minimal vector in pK ` Cqah that starts in b and ends in γptq (see 5.9). This means that there exist geodesics that cannot be obtained using minimal vectors V in pK ` Cqah. In the general context mentioned at the beginning of this section the previous results imply that if B is only required to be a C-algebra then, even when Hopf-Rinow type theorems can be obtained, there exist short curves in UA{UB that cannot be described using minimal elements of A. 2. Preliminaries Let BpHq be the algebra of bounded operators on a separable Hilbert space H, and KpHq and UpHq the compact and unitary operators respectively. If an orthonormal basis teiuiPN is fixed we can consider matricial representations of each A P BpHq, that is A " pAi,jqi,jPN " pxAei, ejyqi,jPN and diagonal operators which we denote with D pB pHqq. Any D P D pB pHqq fulfills xDei, ejy " 0 for every i ‰ j. With the preceding notation, we define columns (and similarly, rows) of any operator A P BpHq as cjpAq "ř8 i"1 xAei, ejy ej " pA1j, A2j, ...q Ă ℓ2, for each j P N. We call Uk the Fredholm subgroup of UpHq, defined as (2.1) Uk " tu P UpHq : u ´ I P KpHqu " tu P UpHq : D K P KpHqah, u " eKu, and the subgroup studied in [5]: (2.2) Uk,d " tu P UpHq : u ´ eD P KpHq for D P pD pB pHqqqahu " tu P UpHq : D K P KpHqah and pD pB pHqqqah , such that u " eK eDu, where I is the identity in BpHq and the superscript ah means anti-Hermitian as well as h means Hermitian. Consider the unitization of KpHq endowed with the norm K ` C " KpHq ` tλI : λ P Cu Ă BpHq, K ` λI}K`C " supt}KC ` λC} : C P KpHq, }C} " 1u, for any K ` λI P K ` C (here } } is the usual operator norm in BpHq). The } }K`C norm coincides with the operator norm in BpHq: K ` λIK`C " }K ` λI} (this follows after considering the multiplication pK ` λIqC for C " h b h P KpHq, with h P H and }h} " 1). The space pK ` C, K`Cq is a unital C -algebra with unit IK`C " 0 ` 1.I " I. Let D pK pHqq " D pB pHqq X KpHq and define the subspace of diagonal operators of K ` C, given by DpK ` Cq " D pK pHqq ` tλI : λ P Cu. GEODESIC NEIGHBORHOODS IN UNITARY ORBITS OF SELF-ADJOINT OPERATORS OF K ` C 3 It is apparent that DpK ` Cq is a unital C -subalgebra of K ` C and IK`C P DpK ` Cq. If u " K ` λI P pK ` Cq is a unitary operator, direct computations show that KK " K K and λ " 1. Therefore, there exists θ P R (λ " eiθ) such that u verifies that u ´ eiθI P KpHq. Then u P Uk,d (see (2.2)) and therefore there exists K0 P KpHqah such that u " eK0eiθI for the same θ (as seen in the proof of Proposition 3.3 in [5]). Moreover, it is apparent that if u " eK eiθI, with θ P R and K P KpHqah, then u " eiθI ``řně1 K n n! eiθI P UK`C, the unitary group of K ` C. Similar considerations can be made with the unitaries of DpK `Cq. If v P UDpK`Cq then v " d`eiθI with d P D pK pHqq and θ P R. This implies that dj,j ` eiθ " 1 for all j P N and therefore pd ` eiθIq " eiR with R a real diagonal matrix such that Rj,j Ñ θ when j Ñ 8. Conversely, if v " eiR with Rj,j P R and limjÑ8 Rj,j " θ, then v " eipR´θIqeiθI P UDpK`Cq because limjÑ8pRj,j ´ θq " 0 and therefore ipR ´ θIq P KpHqah. Then the unitary groups of K ` C and DpK ` Cq can be described as follows: (2.3) and UK`C " Uk.teiθI : θ P Ru " teK eiθI : K P KpHqah and θ P Ru " teK`iθI : K P KpHqah and θ P Ru UDpK`Cq " td ` eiθI : d P D pK pHqq , θ P R such that dj,j ` eiθI " 1u " ted0`iθI : d0 P D pK pHqqah and θ P Ru " teL0 : L0 P D pB pHqqah such that lim jÑ8 pL0qj,j " iθ with θ P Ru. 3. The homogeneous unitary orbit of a self-adjoint compact operator Given a subgroup U Ă UpHq we will denote with OU b the orbit of self-adjoint element b P KpHqh by a subgroup U Ă UpHq, that is OU b " tubu : u P Uu Let b " Diag ptbiuiPNq P DpKpHqhq denote the diagonal operator with the sequence tbiuiPN in its diagonal. We study the unitary orbit of b P KpHq Ă K ` C with bi ‰ bj for each i ‰ j under the action of UK`C: (3.1) UK`C Ob " O b " tubu : u P UK`Cu. Observe that it is apparent that the following inclusions hold for these subgroups of UpHq: Nevertheless the orbit of b under the three subgroups is the same set Ob because (3.2) eK be´K " eKeitI be´itI e´K " eK`itIbe´K´itI Uk ( UK`C ( Uk,d. for t P R and then OUk (see for example Remark 4.4 in [5]). Moreover, we will show further that the three of them share the same natural Finsler metric on the tangent spaces (as seen in Remark 4.5 in [5] for the cases of OUk Uk,d b " O b Uk,d b and O b ). Ob has a smooth structure as described in Lemma 1 in [4] and the comments that follow it. UK`C The isotropy at any c P Ob " O b is Ic " tu P UK`C : ucu " cu. In particular, Ib " tu P UK`C : ru, bs " 0u " UDpK`Cq. Remark 3.1. If c P Ob, the following identification can be made: TcOb -- pT UK`Cq1{pT Ibq1 " pK ` Cqah{pDpK ` Cqqah. 4 TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA2,3 Observe that pK ` Cqah{DpK ` Cqah " rXs : X " K0 ` iθ0I, K0 P KpHqah and θ0 P R( . where rXs is the class defined as Y P rXs iff Y " X ` d ` iθI for d P DpKpHqahq and θ P R. This quotient space is endowed with the usual quotient norm, that in this case for X " K0 ` iθ0I is }rXs} " }rK0 ` iθ0Is} " inf θPR; dPDpKpHqahq }K0 ` iθ0I ` d ` iθI}K`C " inf θPR; dPDpKpHqahq }K0 ` d ` iθI}K`C. In this context, a natural Finsler metric for any x P TbOb, x " Xb ´ bX, with X P pK ` Cqah is (3.3) }x}b " inft}Y } : Y P pK ` Cqah, Y b ´ bY " Xb ´ bXu " inf dPDpKpHqahq θPR }X ` d ` iθI}K`C " }X ` d ` iθI} inf dPDpKpHqahq θPR where X ` d ` iθI is any element of the class rXs " tY P pK ` Cqah : Y " X ` d ` iθI, for θ P R and d P DpKpHqahqu in pK ` Cqah{DpK ` Cqah. An element Y P BpHqah such that Y b ´ bY " x and }Y }K`C " }Y } " }rXs} " }x}b will be called a minimal lifting for x, and its diagonal will be a minimal diagonal approximant (or minimizing diagonal) for Y . This operator Y may not be compact nor unique (see [3]), and it will be called a minimal operator. Given any c " eK`itI be´K´itI " eK be´K P Ob (K P KpHqah, t P R) we can define the norm in its tangent space TcOb using that z " Zc ´ cZ P TcOb for Z " eKXe´K P KpHqah and Xb ´ bX P TbOb. Then the Finsler norm in TcOb is }z}c " }rZs} " inft}Y } : Y P pK ` Cqah, Y c ´ cY " Zc ´ cZu " }rXs}. Note that this norm is invariant under the action of UK`C. Remark 3.2. As it was mentioned before in (3.2), the following orbits are equal UK`C O b " OUk Uk,d b " O b for Uk ( UK`C ( Uk,d defined in (2.1), (2.3) and (2.2) respectively (see Proposition 4.1, 4.4 and Remark 4.7 in [5]). Let X P pK ` Cqah, X " K ` itI with K P KpHqah and t P R, then (3.4) inf DPDpBpHqahq }K ` D} " inf DPDpBpHqahq } Xhkkkikkkj K ` itI ` D} ď inf dPDpKpHqahq }X ` iθI ` d} " inf dPDpKpHqahq θPR }K ` iθI ` d} ď θPR inf dPDpKpHqahq }K ` d}. But since inf dPDpKpHqahq }K ` d} " inf DPDpBpHqahq all those infimums are equal. }K ` D}, [Prop. 5,[3]], the previous inequalities imply that This means that also the Finsler metric defined for TcOb in (3.3) coincides if we consider any of the quotients KpHqah{DpKpHqahq, pK ` Cqah{DpK ` Cqah or pKpHqah ` D pB pHqqahq{D pB pHqqah (see also Remark 4.5 in [5]). U pHq . This follows because if we suppose that for Remark 3.3. Observe that nevertheless Ob ( O b every X P BpHqah holds that eX be´X " eKbe´K for K P KpHqah, then e´K eX must be a diagonal unitary. Therefore, for all X P BpHqah we could write that eX " eKeD for D P D pB pHqqah which is known to be false (see for example Remark 3.7 in [5]). GEODESIC NEIGHBORHOODS IN UNITARY ORBITS OF SELF-ADJOINT OPERATORS OF K ` C 5 UK`C Consider piecewise smooth curves β : ra, bs Ñ O b . We define (3.5) (3.6) Lpβq "ż b a }β1ptq}βptq dt , and distpc1, c2q " inf tLpβq : β is smooth, βpaq " c1, βpbq " c2u UK`C as the rectifiable length of β and distance between two points c1, c2 P O b , respectively. 4. A short curve in Ob obtained acting with a minimal operator not belonging to pK ` Cqah{DpK ` Cqah Consider the following operator described as an infinite matrix: Zδ,γ " i ´δ 0 γ γ ´δ2 γ ´δ2 0 ´δ2 0 γ2 0 ´δ γ ´δ2 ´δ2 ´δ2 γ2 γ2 ´δ3 ´δ3 ´δ3 ´δ3 ´δ3 γ3 γ3 ... ... γ2 ´δ3 γ3 γ2 ´δ3 γ3 γ2 ´δ3 γ3 γ2 ´δ3 γ3 0 ´δ3 γ3 γ3 0 ... 0 γ3 ... γ3 ... γ3 ... γ2 γ3 ... . . . ‹‹‹‹‹‹‹‹‹‹‚ , with γ, δ P p0, 1q. Zδ,γ is a Hilbert-Schmidt operator which has been studied in [4] in its self-adjoint version. We recall here some of its properties. Let γ2 " δ and δ2 ă γ (for example γ " 1{2 and δ " 1{4), and denote with Z r1s δ,γ the operator defined by the matrix of Zδ,γ with zeros in the first column and row, with c1pZδ,γq the first column of Zδ,γ, and with D0 the (uniquely determined) diagonal matrix such that every pD0qi,i is chosen to satisfy cipZδ,γq K c1pZδ,γq for all i ‰ 1. Then (4.1) Zo " }Z r1s δ,γ ` D0} }c1pZδ,γq} pZδ,γ ´ Z r1s δ,γq ` Z r1s δ,γ is also a Hilbert-Schmidt operator with the property that D0 (constructed as mentioned before) is a minimal diagonal for Zo (see (5.1) in [4] for a detailed proof of these statements and the follow- ing comments). Moreover, it has been proved that D0 P DpBpHqahq is the unique bounded best approximant anti-Hermitian diagonal of Zo. D0 has the particular property that limkÑ8pD0q2k,2k ‰ limkÑ8pD0q2k`1,2k`1 and both limits are not null. Therefore D0 is not compact and we call it an oscillant diagonal. We will write with (4.2) Z2 " Zo ` D0 to denote the minimal operator constructed as above. Then (4.3) distpZo, DpKpHqahqq " }rZos}KpHqah{DpKpHqahq " }Z2} " }c1pZ2q} " }c1pZoq}. Theorem 4.1. Let b " Diag ptbiuiPNq P DpKpHqhq with bi ‰ bj for each i ‰ j. Consider the unitary UK`C defined in (3.1) and x " Zb ´ bZ P TbOb, for Z a minimal operator in KpHqah ` orbit O b D pB pHqqah. Then the uniparametric group curve γptq " etZbe´tZ has minimal length in the class of UK`C all curves in O b joining γp0q and γptq for each t P"´ π 2}Z}ı. 2}Z} , π 6 TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA2,3 Proof. This proof is a direct consequence of mentioned previous results, but we include here the citations and reasonings for the sake of clarity. UK`C for any t P R. Using Remark 3.2, we obtain that γptq P O b " By Theorem 4.2 in [5], γptq P OUk b Ob, @ t. Moreover, }x}b " }Zb ´ bZ}b " }rZs} " inf θPR; dPDpKpHqahq }Z ` d ` iθI} " }Z}, where the minimality of Z implies the last equality. Consider Pb " tubu : u P UpHqu, then by Theorem II in [7], since Z is minimal, the curve γ has 2}Z}. Since , that is minimal length over all the smooth curves in Pb that join γp0q " b and γptq, with t ď π UK`C b 2}Z}ı follows that γ is a short curve in O Ď Pb, then for each t0 P"´ π UK`C clearly O b 2}Z} , π where distpb, γpt0qq is the rectifiable distance between b and γpt0q defined in (3.6). (cid:3) L´γr0,t0s¯ " distpb, γpt0qq, Corollary 4.2. Let b " Diag ptbiuiPNq P DpKpHqhq with bi ‰ bj for each i ‰ j. Consider the unitary UK`C defined in (3.1), x " Zob ´ bZo P TbOb, for Zo defined in (4.1), with D0 its unique orbit O b minimizing diagonal, and Z2 " Zo ` D0 defined in (4.2). Then the uniparametric group curve γptq " etZ2 be´tZ2 has minimal length in the class of all curves UK`C in O b joining γp0q and γptq for each t P"´ π 2}Z2} , π 2}Z2}ı. Proof. If we consider Z " Z2 " Zo ` D0 in the statements of Theorem 4.1, then Z2 satisfies the conditions required and therefore the proof is apparent. (cid:3) The previous result will allow us to state that the converse of Theorem I in [7] does not necessary hold when the subalgebra considered (here DpK ` Cq) is not a von Neumann algebra. Let us describe the context of that article. Let A be a C-algebra and B a C-subalgebra, then a natural Finsler metric as the one in (3.3) is defined for the generalized flag P " UA{UB. If the element X P Aah is minimal for a tangent vector x P TppUA{UBq » T1pUAq{T1pUBq (that is: x " Xp ´ pX and }X} " inft}Y } : Y P Aah; Y p ´ pY " xu) then the curve γptq " LetX p has minimal length for t ď π 2}X} (where L is a left action on P) with the distance defined in (3.6). The following result shows that there might exist some minimal curves γ in these generalized flags P " UK`C{UDpK`Cq that are not of the form γptq " LetZ p for Z P Aah a minimal lifting of a tangent vector x P T1pUAq{T1pUBq. Remark 4.3. Let Z2 be the operator defined in (4.2). As it was mentioned in Corollary 4.2, the UK`C uniparametric curve γptq " etZ2be´tZ2 has minimal length in the class of all curves in Ob " O b joining γp0q and γptq for each t P"´ π 2}Z2} , π 2}Z2}ı. Therefore the curve γ is included in Ob with initial conditions γp0q " b, γ1p0q " x " Zob´bZo even for velocity vectors x P TbOb that do not have a minimal compact lifting K0 (recall that Zo ` D0 is not compact and D0 is its unique minimizing diagonal). Thus UK`C is an example of a group whose action on Ob has short curves that might not be described using minimal vectors Y P pK ` Cqah. This is not new (for instance, see Remark 4.7 in [5]), but in the present case K ` C and DpK ` Cq, whose anti-Hermitian elements are the Lie-algebras of UK`C and UDpK`Cq respectively, are unital C -algebras. We will develop some details of this situation in the next section. GEODESIC NEIGHBORHOODS IN UNITARY ORBITS OF SELF-ADJOINT OPERATORS OF K ` C 7 5. Neighborhoods of short curves defined by minimal vectors in UK`C{UDpK`Cq In this section we will consider the problem of the existence of a neighborhood around b P DiagpKpHqqh with bi,i ‰ bj,j for i ‰ j whose elements can be joined with b with a short curve of the form γptq " etZ be´tZ for some minimal anti-Hermitian element Z P pK ` Cqah and t in some interval. Recall here Z2 " Zo ` D0 defined in (4.2) where Zo P KpHqah is the Hilbert-Schmidt operator defined in (4.1) and D0 is its unique minimizing diagonal with the property that D0 has subsequences that converge to two different (not null) limits as described in the previous section. Moreover, Z2 satisfies, (1) }Z2} " }c1pZ2q}, (2) c1pZ2q1 " pZ2q1,1 " 0 and (3) c1pZ2qj " pZ2qj,1 ‰ 0 for all j ‰ 1. Lemma 5.1. Let γptq " etZbe´tZ Ă Ob with Z P BpHqah a minimal operator with unique minimizing diagonal, and consider a curve δptq " etV be´tV , with V P BpHqah another minimal operator such that δ1p0q " V b ´ bV " γ1p0q " Zb ´ bZ. Then it must be Z " V . Proof. Since δ1p0q " γ1p0q, then δ1p0q " V b ´ bV " γ1p0q " Zb ´ bZ and therefore V ´ Z commutes with b. Then V ´ Z P DpBpHqahq which implies that Z and V must be equal outside their diagonals. Then, since DiagpZq is the only minimizing diagonal for Z and V is also a minimal operator then DiagpV q " DiagpZq which implies that V " Z. (cid:3) Remark 5.2. Observe that if we apply the previous lemma to the case where Z " Z2 defined in (4.2) and δ and V satisfy the assumptions of the lemma, then in particular V R pK ` Cqah. This is a direct consequence of the fact that Z2 has two different non zero diagonal limits, something that V P pK ` Cq cannot satisfy. Lemma 5.3. Let Z " KZ ` DiagpZq with KZ P KpHqah, DiagpZq P D pB pHqqah be a minimal operator and γ : r0, π 2}Z}s Ñ Ob the short curve defined as γptq " etZ be´tZ (see Theorem 4.1) and let minimal operator with KV P KpHqah and DiagpV q P D pB pHqqah. Moreover, suppose that there exists 2}V }ı Ñ Ob be another short curve defined by δpsq " esV be´sV for V " KV ` DiagpV q another }Z} ı and s1 P´0, π 2}V }ı such that γpt1q " δps1q. δ :"0, π t1 P´0, logp2q{8 Then et1Z " es1V e´ Diagps1V q`Diagpt1Zq and }s1V } " }t1Z}. Proof. Note that t1 satisfies }t1Z} ă plog 2q{8, and then t1Z is sufficiently close to zero in the sense of Definition 2.1 of [5]. Now consider the equality γpt1q " δps1q Then es1V be´s1V " et1Zbe´t1Z. be´s1V et1Z " e´s1V et1Zb which implies that b commutes with e´s1V et1Z. Therefore e´s1V et1Z is diagonal and unitary. Then there exists D P D pB pHqqah such that e´s1V et1Z " eD. Observe that since Z is a minimal operator the length of γ restricted to r0, t1s is }t1Z} and if δ is a short curve then the length of δ must coincide with }t1Z} (see Theorem 4.1 in [7]). Since the 8 TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA2,3 length of δ restricted to r0, s1s equals }s1V } because V is a minimal operator, then }t1Z} " }s1V }. Also t1Z is sufficiently close to zero (thus s1V ), so we can apply Proposition 3.11 and Corollary 3.12 of [5] to obtain (5.1) eD " e´s1V et1Z " e´s1V `Diagps1V q´Diagps1V qet1Z´Diagpt1Zq`Diagpt1Zq " eK´Diagps1V q`Diagpt1Zq for K P KpHqah with DiagpKq " 0. Then, since }t1Z} " }s1V } ă plog 2q{8, then }D} " } logpe´s1V et1Zq} ď ´1{2 log`2 ´ e2}t1Z}`2}s1V } ă π (see some of the Baker-Campbell-Hausdorff series bounds in [2] or [10]). This implies that D " K ´ Diagps1V q ` Diagpt1Zq because eD " eK´Diagps1V q`Diagpt1Zq and both anti-Hermitian exponents have norm less than π (see for example Corollary 4.2 iii) of [6]). But, since DiagpKq " 0 and D P D pB pHqqah, then (5.2) and K " 0. D " ´ Diagps1V q ` Diagpt1Zq. (cid:3) UK`C Theorem 5.4. (Local Hopf-Rinow theorem) There exists Wb Ă Ob " O a neighborhood (with b the distance defined in (3.6)) of b P DpKpHqhq with bi,i ‰ bj,j for i ‰ j, such that for every ρ P Wb there exists a short curve γ in Ob that joins b with ρ, and γ : r0, 1s Ñ Wb Ă Ob, γptq " etpKρ`Dρqbe´tpKρ`Dρq, with Kρ P KpHqah, Dρ P D pB pHqqah, }Kρ}, }Dρ}, }Kρ ` Dρ} ă logp2q operator in KpHqah ` D pB pHqqah. 4 , and pKρ ` Dρq a minimal Proof. If we consider the isotropy compact generalized flag manifold P " UpHq{DpUpHqq then for ρ0 P P there exists a neighborhood Vρ0 " tLuρ0 : u " eX, for X P BpHqah, }X} ă π{2u where a local Hopf-Rinow theorem holds (see Theorem II-1 and Example 1 of [8]). That is, for every ρ P Vρ0 there exist a minimal operator X P BpHqah with }X}ăπ{2 and a minimal uniparametric group curve γ : r0, 1s Ñ P, γptq " LetX ρ0 joining γp0q " ρ0 and γp1q " ρ. The generalized flag manifold P " UpHq{DpUpHqq can be identified with the unitary orbit of b P DpKpHqhq with bi,i ‰ bj,j as well as its tangent spaces as we have done with Ob in Remark 3.1: U pHq Tc O b » T1 UpHq{T1 D pUpHqq " BpHqah{D pB pHqqah Since we are using the adjoint action L, then Lub " ubu in this context. And if we consider ρ0 " b we can conclude that for any ρ " eK be´K P pOb X Vbq with K P KpHqah (see (3.2)), there exists a minimal operator Z P BpHqah with }Z} ă π{2, such that U pHq γ : r0, 1s Ñ O b , γptq " etZ be´tZ, with γp0q " ρ0 " b and γp1q " ρ " eKbe´K " eZbe´Z. (5.3) Note that in this case eZ cannot be any element of UpHq because eK be´K " eZbe´Z implies that eZ " eKeD for D P D pB pHqqah, and therefore eZ P Uk,d. Moreover, we will show that choosing a smaller neighborhood Z can be written as Z " K 1 ` D1, with K 1 P KpHqah, D1 P D pB pHqqah. In order to obtain this last assertion recall from Lemma 3.14 of [5] that there exists ε0 ą 0 such that if u P Uk,d satisfies }u ´ 1} ă ε0, then there exist K 1 P KpHqah and D1 P pD pB pHqqqah with u " eK 1`D1 for }K 1}, }D1}, }K 1 ` D1} ă log 2 4 Then define the neighborhood of b in Ob: (see Definition 2.1 and the proof of Lemma 3.14 of [5]). Wb " tubu : u " eZ P Uk,d, Z P BpHqah, }Z} ă logp1 ` ε0qu GEODESIC NEIGHBORHOODS IN UNITARY ORBITS OF SELF-ADJOINT OPERATORS OF K ` C 9 with ε0 from Lemma 3.14 of [5]. Note that Wb ( pVb XObq (´Vb X O U pHq b ¯. It is apparent that if u " eZ P Uk,d, with Z P BpHqah, }Z} ă logp1 ` ε0q, then }eZ ´ 1} ď e}Z} ´ 1 ă ε0. Applying the mentioned lemma, this implies that in this case there exist K 1 P KpHqah and D1 P D pB pHqqah with u " eK 1`D1 for }K 1}, }D1}, }K 1 ` D1} ă log 2 is the short curve γptq " etZbe´tZ Ă Vb for Z P BpHqah a minimal operator such that γp1q " eK be´K P Ob (for K P KpHqah), then it must be eZ " eK eD P Uk,d. Moreover, since }eZ ´ 1} ă ε0, there exist K 1 P KpHqah and D1 P D pB pHqqah satisfying U pHq 4 . Following the discussion after (5.3), if γ : r0, 1s Ñ O b eZ " eK 1`D1 , for }K 1 ` D1} ă log 2 4 ùñ Z " K 1 ` D1 U pHq because Z and K 1 ` D1 have norm smaller than π. Then the entire curve γ : r0, 1s Ñ O is b UK`C . In this case, being Z " K 1 ` D1, the distance from ρ0 " b to ρ " eZbe´Z included in Ob " O b U pHq (see (3.4)). Then γptq " is the same either if we consider the Finsler metrics in Ob or in O b etZbe´tZ " etpK 1`D1qbe´tpK 1`D1q, γ : r0, 1s Ñ Ob defines a short curve between b and ρ " eK be´K " epK 1`D1qbe´pK 1`D1q, with K 1 ` D1 " Z a minimal operator of KpHqah ` D pB pHqqah. The statement of the theorem follows after substituting K 1 " Kρ and D1 " Dρ. The element ρ " eK be´K P Wb was chosen arbitrarily, so we have proved that Wb is a geodesic (cid:3) neighborhood of b in Ob. Remark 5.5. Observe that the unitary eKρ`Dρ P Uk,d mentioned in the previous theorem might not belong to K ` C, but γptq " etpKρ`Dρqbe´tpKρ`Dρq P Ob for every t (see Remark 3.2). Remark 5.6. Let c " eK0be´K0 P Ob, with K0 P KpHqah. The action Lupcq " ucu, for u P UK`C is invariant for the distance defined in Ob and therefore the previous result also holds in this case. That is, there exists a geodesic neighborhood Wc of c such that every ρ P Wc is joined with c by short curves included in Ob of the form LeK0 γ (for γ the curve described in Theorem 5.4). In the work mentioned, the minimality of a curve γ : r0, π Here we recall some results and the notation used in [7] and state its translation to the particular 2}Z} s Ñ Ob, example we are studying. γptq " etZbe´tZ , with Z be a minimal lifting of x P TbOb, was proved using a unitary reflection r0 in a Hilbert space with certain properties (see Definition 2.4 in [7]), a representation of A in BpHq with particular properties (in our case the identity representation verifies them) and a map F : Ob Ñ GrpHq (GrpHq is the Grassmann manifold of H) defined by F pubuq " ur0u. The unitary reflection r0 used there is (5.4) r0pxq "" x, if x P Sb ´x, if x P SK b , where Sb is the closure of Ω " tx P H : x " U ξ, for U a diagonal in UK`Cu and ξ P H certain vector satisfying Definition 4.1 of [7]. Now, if we consider the particular case in which Z is a minimal operator such that }Z} " }cj0pZq}, cj0pZqj0 " Zj0,j0 " 0 and cj0pZqj " Zj,j0 ‰ 0 for all j ‰ j0 (see the example of (4.2) and Lemma 6.1) we can be much more specific about r0 and ξ. After the corresponding translation to this case ξ " i ej0, Sb " gentξu and r0pxq "" x , ´x , (5.5) if x P gentξu " gentej0u if x P gentξuK " gentej0uK . Moreover, ξ fulfills Definition 4.1 in [7], since Z 2ξ " ´}Z}2ξ, r0pξq " ξ and r0pZξq " ´Zξ. Therefore, γptq " etZbe´tZ minimizes length between the points γp0q " b and γptq if 0 ď t ď π 2}Z} . 10 TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA2,3 In this context, the map Fξ : Ob Ñ S Ă H, Fξpubuq " ur0upξq, where S is the unit sphere of H, reduces length. That is, if δ : r0, t0s Ñ Ob and v : r0, t0s Ñ S , vptq " Fξpδptqq then (see Corollary 3.4 in [7]) ℓpvq ď Lpδq with L defined in (3.5) and ℓ the length in S . The following result is an application of Theorem 4.1 and Lemma 4.2 in [7] to our context. Proposition 5.7. Let Z and V be minimal operators of pKpHq ` DpHqqah and consider the following curves in Ob (defined in (3.1)) γptq " etZ be´tZ and δptq " etV be´tV , t ą 0. Suppose additionally that there exists 0 ď t0 ď min! π 2}Z2} ; π 2}V }) such that γpt0q " δpt0q. Then, following the previous notation, (1) wptq " Fξpγptqq and vptq " Fξpδptqq both are geodesics in the sphere S Ă H and ℓpwq " Lpγq " Lpδq " ℓpvq. (2) γ and δ minimize length between the points b and γptq and δptq respectively, if 0 ď t ď t0. (3) vptq " etV r0e´tV pξq " etZ r0e´tZpξq " wptq, for 0 ď t ď t0 and r0 the unitary reflection defined in (5.4). Proof. As it was mentioned before, items (1) and (2) are a direct consequence of Theorem 4.1 and Lemma 4.2 in [7]. If γpt0q " δpt0q, wpt0q and vpt0q match. Since geodesics for fixed ending points in the unit sphere S are unique (maximum circles), then wptq " vptq, for all 0 ď t ď t0. (cid:3) Observe that the assumption of existence of such t0 is possible even if γpt1q " δpt2q for t1 ‰ t2, since δptq can be re-scaled defining δptq " et t2 t1 V be´t t2 t1 V , for t P„0, }V }and then δpt1q " γpt1q. π 2 t2 t1 Lemma 5.8. Let b P DpKqh with bi,i ‰ bj,j for i ‰ j, Z, V P pKpHq ` DpBpHqqqah be minimal operators such that for some j0 P N }Z} " }cj0pZq}, cj0pZqj0 " Zj0,j0 " 0 and cj0pZqj " Zj,j0 ‰ 0 for all j ‰ j0. Moreover, if there exists t0 with 0 ă t0 ď π 2}Z} that satisfies et0Zbe´t0Z " es0V be´s0V , for s0 P"0, π 2}V }ı, then s0cj0pV q " t0cj0pZq, that is, the jth 0 column of Z is a multiple of the jth 0 column of V . Proof. We can suppose that j0 " 1. First, consider W " s0 t0 V , γptq " etZ be´tZ and δ0ptq " etW be´tW for t P"0, π 2}Z}ı. Then γpt0q " es0V be´s0V " et0s0{t0V be´t0s0{t0V " δ0pt0q. Moreover, since also W is a minimal operator, then γ and δ0 are short curves in the interval r0, t0s and then the length of γr0,t0s equals that of δ0r0,t0s . That is, t0}Z} " t0}W } (see for example Theorem 4.1 of [7]) and therefore in particular (5.6) }Z} " }W }. Using the preceding notations of Proposition 5.7, the assumptions made here imply that vptq " Fξpγptqq " Fξpδ0ptqq " wptq, for t P r0, t0s. Then their derivatives coincide for every t P r0, t0s (5.7) where ξ " ie1, η " c1pZq w1ptq "`W etW r0e´tW ´ etW r0e´tW W pξq "`ZetZr0e´tZ ´ etZr0e´tZZ pξq " v1ptq. }c1pZq} and the reflection r0 are defined in (5.5) and fulfill r0pξq " ξ and r0pηq " ´η. GEODESIC NEIGHBORHOODS IN UNITARY ORBITS OF SELF-ADJOINT OPERATORS OF K ` C 11 Then, if we evaluate (5.7) in t " 0 w1p0q " pW r0 ´ r0W q pξq " pZr0 ´ r0Zq pξq " v1p0q. Hence, since r0pξq " ξ and r0pc1pZqq " r0p}c1pZq}ηq " }c1pZq} r0pηq " ´}c1pZq}η " ´c1pZq, (5.8) W r0pξq ´ r0W pξq " Zr0pξq ´ r0Zpξq W pie1q ´ r0W pie1q " Zpie1q ´ r0Zpie1q ic1pW q ´ ir0pc1pW qq " ic1pZq ` r0 pic1pZqq ipI ´ r0q pc1pW qq " i2c1pZq On the other hand, if we consider the decomposition H " gentξu ' pgentξuqK then the identity operator I and r0 can be matricially described as I "1 0 0 1 and r0 "1 0 ´1 , 0 respectively. Then, (5.8) implies that pI ´ r0qpc1pW qq " 2pc1pW q ´ W1,1e1q " 2c1pZq and then (5.9) c1pW q ´ W1,1e1 " c1pZq and }c1pZq} " }Z} " }W } (see (5.6)). This implies that }c1pW q ´ W1,1e1} " }W }, and therefore W1,1 " 0, since otherwise }c1pW q} ą }W }, which is a contradiction. Therefore, returning to (5.9) we obtain that c1pW q " c1pZq that implies that c1 s0 t0 V " c1pZq which ends the proof. Next, we obtain the second main result of this section. (cid:3) Theorem 5.9. Let b P DpKqh with bi,i ‰ bj,j for i ‰ j, Z P pKpHq ` DpBpHqqqah be a minimal operator such that for some j0 P N, holds that }Z} " }cj0pZq}, cj0pZqj0 " Zj0,j0 " 0, cj0pZqj " Zj,j0 ‰ 0 for all j ‰ j0, the sequence tDiagpZqj,jujPN has more that one not null accumulation points, and γptq " etZ be´tZ, for t P´0, log 2 8}Z}¯. 2}V }ı Then there is not any minimal operator V P pK ` Cqah such that δptq " etV be´tV , t P "0, π operator V P pK ` Cqah such that δps0q " es0V be´s0V " γpt0q, for s0 P ´0, π 2}V }ı and t0 P ´0, log 2 8}Z}¯. Proof. As done before, we are going to prove only the case j0 " 1. Suppose that there exists a minimal satisfies γpt0q " δps0q for t0, s0 in the respective domains. Note that in particular t0 ă log 2 8}Z} ă π 2}Z} which implies that γ is a short curve in all its domain. Applying Lemma 5.3, ´ Diagps0V q ` Diagpt0Zq P D pB pHqqah is such that (5.10) e´s0V et0Z " e´ Diagps0V q`Diagpt0Zq and }t0Z} " }s0V }. Then, the fact that Z satisfies }Z} " }c1pZq} (see (4.3)) implies that V must fulfill that (5.11) }V } " t0 s0 }c1pZq} Using Lemma 5.8 follows that t0c1pZq " s0c1pV q, and therefore (5.11) implies that }V } " t0 s0 t0 s0››› s0 t0 c1pV q››› " }c1pV q}. }c1pZq} " 12 TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA2,3 Then Lemma 6.1 implies that c1pV q is orthogonal to every other column of V . This property also c1pV q} " c1pV q }c1pV q} , holds for c1pZq and the columns of Z. Recall the notation ξ " ie1 and η " c1pZq and consider }c1pZq} " s0 t0 } s0 t0 c1pV q ξ ` η " ie1 ` " ie1 ` c1pZq }c1pZq} c1pV q }c1pV q} P H. A direct computation shows that ξ ` η is an eigenvector of Z and V of the eigenvalue i}Z} " i}c1pZq} " i s0 }V } (see for example the proof of Theorem 2 in [4] for the self-adjoint t0 case). The previous comments imply that }c1pV q} " i s0 t0 eDpξ ` ηq " e´s0V et0Zpξ ` ηq " e´i 2 s 0 t0 }V }eit0}Z}pξ ` ηq " eipt0´s0q}Z}pξ ` ηq where in the last equality we used }V } " t0 s0 using (5.10) we write }Z} and the series expansion of the exponentials. Then, (5.12) eDpξ ` ηq " e´ Diagps0V q`Diagpt0Zqpξ ` ηq " eipt0´s0q}Z}pξ ` ηq. Therefore, considering the equality in each entry of (5.12) we obtain ep´s0V `t0Zqj,j pξ ` ηqj " eipt0´s0q}Z}pξ ` ηqj for all j P N. The fact that pξ ` ηqj ‰ 0 for all j P N implies that ep´s0V `t0Zqj,j " eipt0´s0q}Z} for every j. Since we are supposing that t0 P´0, log 2 8}Z}ı, then the exponent p´s0V ` t0Zqj,j is small enough and we can conclude that p´s0V ` t0Zqj,j " ´s0Vj,j ` t0pZqj,j " ipt0 ´ s0q}Z} for all j P N. But this is a contradiction since we are supposing DiagpV q " d ` iθI, with d P KpHqah, θ P R, and we know that DiagpZq has more than one (not null) limit. Therefore, a minimal operator V P pK ` Cqah cannot form a curve δptq " etV be´tV that crosses γ for t ą 0 in a certain small enough neighborhood of b. (cid:3) Corollary 5.10. If we consider Z2 as defined in (4.2) for every neighborhood Xb of b in Ob there exist elements pet0Z2be´t0Z2q P Xb such that there is not any short curve of the form δptq " etV be´tV with V P pK ` Cqah that joins b with et0Z2be´t0Z2. In fact, this is true for etZ2be´tZ2, for every t in certain interval. Proof. Observe that the operator Z2 satisfies every assumption needed by the operator Z in Theorem 5.9.Then the proof is a direct application of the previous theorem. (cid:3) Remark 5.11. Note that the situation mentioned in the previous corollary applies to the geodesic neighborhood Wb obtained in Theorem 5.4 even when in that case every element of Wb is reached by a short curve. 6. Appendix In this section we include various results concerning minimal anti-Hermitian operators in BpHqah. Lemma 6.1. For a given fixed orthonormal basis of H, let V P BpHqah be such that there exists j0 P N that satisfies }V } " }cj0pV q} (where cjpV q is the j th column of the corresponding matrix representation of V in the fixed basis). Then (6.1) cj0pV q K cjpV q, @j ‰ j0. If pcj0pV qqj0 " Vj0,j0 " 0 then V is a minimal operator. GEODESIC NEIGHBORHOODS IN UNITARY ORBITS OF SELF-ADJOINT OPERATORS OF K ` C 13 Moreover, if cj0pV qj " Vj,j0 ‰ 0 for all j ‰ j0, then V has a unique minimizing diagonal defined by (6.2) Vj,j " ´AcjpV qqj, cj0pV qqjE Vj,j0 , for j ‰ j0 where ckpXqql P H a gentelu is the element obtained after taking off the lth entry of ckpXq P H. Proof. Note that we can suppose that j0 " 1 to simplify the notation. Similar considerations could be done for the jth 0 column. In the matrix representation corresponding to the fixed orthonormal basis tejujPN, we can consider Observe that }x} " 1, and then it must hold }V x} ď }V }. Let us consider f : R Ñ R such that x " cosptqe1 ` sinptqej, for j ‰ 1. f ptq " }V pcosptqe1 ` sinptqejq}2 " } cosptqV pe1q ` sinptqV pejq}2 " } cosptqc1pV q ` sinptqcjpV q}2 " xcosptqc1pV q ` sinptqcjpV q, cosptqc1pV q ` sinptqcjpV qy " cos2ptq}c1pV q}2 ` sin2ptq}cjpV q}2 ` 2 cosptq sinptq Rexc1pV q, cjpV qy. (6.3) Then f 1ptq " ´ 2 sinptq cosptq}c1pV q}2 ` 2 sinptq cosptq}cjpV q}2` ` 2`cosptq2 ´ sinptq2 Rexc1pV q, cjpV qy " sinp2tq`}cj}2 ´ }c1}2 ` 2 cosp2tq Rexc1pV q, cjpV qy. Then, if Rexc1pV q, cjpV qy ą 0 f 1p0q " 2Rexc1pV q, cjpV qy ą 0 and f p0q " }c1pV q}2 and then f 1pt1q ą 0 for some t1 ą 0, which implies that f pt1q ą }c1pV q}2, a contradiction. On the other hand, if Rexc1pV q, cjpV qy ă 0 f 1p0q " 2Rexc1pV q, cjpV qy ă 0 and f p0q " }c1pV q}2 and then f 1pt2q ă 0 for some t2 ă 0, which implies that f pt2q ą }c1pV q}2, a contradiction. Therefore it must be Rexc1pV q, cjpV qy " 0. Now consider z " cosptqe1 ` i sinptqej P H, that also satisfies }z} " 1. Then following the it can be proved that 0 " steps we used in the case of x " cosptqe1 ` sinptqej but using z, Rep´iqxc1pV q, cjpV qy " Imxc1pV q, cjpV qy. In order to prove the last part of the lemma observe that (1) as proved in the first part of this lemma, c1pV q K cjpV q for all j ‰ 1, (2) and the assumptions (a) c1pV q1 " V1,1 " 0, (b) c1pV qj " Vj,1 ‰ 0 for j ‰ 1 (c) and the equality }V } " c1pV q Then the proof of the minimality of V follows applying Theorem 2.2 from [1] substituting A with BpHq, B with DpBpHqq, ρ with the identity, ξ with i e1 and Z with V . Note that we only need assumptions (1), (2)(a) and (2)(c) to prove that V 2pi e1q " ´}V }2i e1 and that xV pi e1q, Dpi e1qy " xi c1pV q, i D1,1y " 0 in order to fulfill the assumptions of Theorem 2.2 from [1]. ››››pV ` Dq c1pV q }c1pV q}›››› " (6.4) 1 }V c1pV q ` Dc1pV q} " }c1pV q} "››››´}c1pV q}e1 ` 1 }c1pV q} Dc1pV q›››› ą }c1pV q} " }V }. 1 }c1pV q}››´}c1pV q}2e1 ` Dc1pV q›› 14 TAMARA BOTTAZZI 1 AND ALEJANDRO VARELA2,3 The equality (6.2) follows after the condition c1pV q K cjpV q for j ‰ 1 and the fact that c1pV qj " Vj,1 ‰ 0 for those j. Moreover, if we consider V ` D, for D ‰ 0, and D1,1 ‰ 0, follows that }c1pV ` Dq} " }c1pV q ` D1,1e1} ą }c1pV q} " }V } and therefore V ` D cannot be minimal. Now suppose D1,1 " 0. Direct computations show that In the previous strict inequality we have used (2)(a), (2)(b), D ‰ 0 and D1,1 " 0. Then }V ` D} ą }V } for D ‰ 0, which implies that the diagonal defined in (6.2) is the only (cid:3) possible minimizing diagonal of V . Another way to prove equation (6.1) of the first part of the previous Lemma 6.1 is using Corollary 6.3 of the following theorem. Theorem 6.2 (Sain, [9]). Let H1, H2 be Hilbert spaces and T P BpH1, H2q. Given any x P H1, }T x} " }T } if and only if the following two conditions are satisfied: i) xx, yy " 0 implies that xT x, T yy " 0, ii) supt}T y} : }y} " 1, xx, yy " 0u ď }T x}. Corollary 6.3. Consider H " H1 " H2 and V P BpHqah. Then there exists j0 P N such that }V } " }V pej0q} " }cj0pV q}, if and only if i) xej0, ejy " 0 implies that xV ej0, V ejy " xcj0pV q, cjpV qy " 0 for each j P N, j ‰ j0, ii) supt}cjpV q} : j P Nu ď }cj0pV q}. Proof. The proof is a direct consequence of Theorem 6.2 after observing that in item i) of the corollary is equivalent to say that xej0, yy " 0 implies that xV ej0, V yy " 0 for any y P H. (cid:3) References [1] E. Andruchow, L. E. Mata-Lorenzo, A. Mendoza, L. Recht, and A. Varela. Minimal matrices and the corresponding minimal curves on flag manifolds in low dimension. Linear Algebra Appl., 430(8-9):1906 -- 1928, 2009. [2] D. Beltit¸a. Smooth homogeneous structures in operator theory, volume 137 of Chapman & Hall/CRC Monographs and Surveys in Pure and Applied Mathematics. Chapman & Hall/CRC, Boca Raton, FL, 2006. [3] T. Bottazzi and A. Varela. Best approximation by diagonal compact operators. Linear Algebra Appl., 439(10):3044 -- 3056, 2013. [4] T. Bottazzi and A. Varela. Minimal length curves in unitary orbits of a Hermitian compact operator. Differential Geom. Appl., 45:1 -- 22, 2016. [5] T. Bottazzi and A. Varela. Unitary subgroups and orbits of compact self-adjoint operators. Studia Math., 238(2):155 -- 176, 2017. [6] E. Chiumiento. On normal operator logarithms. Linear Algebra Appl., 439(2):455 -- 462, 2013. [7] C. E. Dur´an, L. E. Mata-Lorenzo, and L. Recht. Metric geometry in homogeneous spaces of the unitary group of a C-algebra. I. Minimal curves. Adv. Math., 184(2):342 -- 366, 2004. [8] C. E. Dur´an, L. E. Mata-Lorenzo, and L. Recht. Metric geometry in homogeneous spaces of the unitary group of a C-algebra. II. Geodesics joining fixed endpoints. Integral Equations Operator Theory, 53(1):33 -- 50, 2005. [9] D. Sain. On extreme contractions and the norm attainment set of a bounded linear operator. ArXiv, 2017. [10] V. S. Varadarajan. Lie groups, Lie algebras, and their representations, volume 102. Springer Science & Business Media, 2013. 1 Sede Andina, Universidad Nacional de R´ıo Negro, (8400) S.C. de Bariloche, Argentina 2Instituto Argentino de Matem´atica "Alberto P. Calder´on", Saavedra 15 3er. piso, (C1083ACA) Buenos Aires, Argentina GEODESIC NEIGHBORHOODS IN UNITARY ORBITS OF SELF-ADJOINT OPERATORS OF K ` C 15 3Instituto de Ciencias, Universidad Nacional de Gral. Sarmiento, J. M. Gutierrez 1150, (B1613GSX) Los Polvorines, Argentina E-mail address: [email protected], [email protected]
1608.02086	3	1608	2016-11-01T11:34:15	$C^*$-algebras generated by the paths semigroup	[ "math.OA", "math-ph", "math.FA", "math-ph" ]	In this paper we study the structure of the $C^$-algebra, generated by the representation of the paths semigroup on a partially ordered set (poset) and get the net of isomorphic $C^$-algebras over this poset. We construct the extensions of this algebra, such that the algebra is an ideal in that extensions and quotient algebras are isomorphic to the Cuntz algebra.	math.OA	math	S. Grigoryan, T. Grigoryan, E. Lipacheva, A. Sitdikov C ∗-ALGEBRA GENERATED BY THE PATH SEMIGROUP Kazan State Power Engineering University, Krasnosel'skaya Str., 51, Kazan, Tatarstan, 420066, Russia E-mail address: [email protected], [email protected], [email protected], airat−[email protected] Received July 4, 2016 Abstract. In this paper we study the structure of the C ∗-algebra, generated by the representation of the path semigroup on a partially ordered set (poset) and get a net of isomorphic C ∗-algebras over this poset. We construct the extensions of this algebra, such that the algebra is an ideal in that extensions and quotient algebras are isomorphic to the Cuntz algebra. 1. Introduction In the algebraic approach to the quantum field theory [1] (the algebraic quantum field theory) the physical content of the theory is encoding by a collection of C ∗-algebras of observables A = {Ao}o∈K indexed by elements of a partially ordered set K (poset) [2]. The poset K is a non- empty set with a binary relation ≤ which is reflexive, antisymmetric and transitive. A net of C ∗-algebras over the poset K is the pair (A, γ)K, where γ = {γo′o : Ao → Ao′}o≤o′ are -morphisms fulfilling the net relations γo′′o = γo′′o′ ◦ γo′o for all o ≤ o′ ≤ o′′ ∈ K. If we consider the poset K as a category in which objects are elements of K and morphisms are arrows (o, o′) for all o ≤ o′ ∈ K, then the net of C ∗-algebras represents a covariant functor from a poset K to category of unital C ∗-algebras with -morphisms (see for example [3, 4]). More precisely we have a net of C ∗-algebras for an 2010 Mathematical Subject Classification. 46L05. Key words and phrases. C ∗-algebra, partially ordered set, partial isometry operator, inverse semigroup, left regular representation, Cuntz algebra. 3 4 GRIGORYAN, GRIGORYAN, LIPACHEVA, SITDIKOV upward directed poset and in the event of non-upward directed we obtain a precosheaf of C ∗-algebras [5 -- 7]. In this paper we give an algebraic notion of a path on a poset K which turns out to be relevant to the point of view on a path as a sequence of 1- simplices. We introduce the path semigroup S on the given poset K and construct a new C ∗-algebra C ∗ red(S) generated by the representation of S. We consider both an upward directed set K and non-upward directed. The present paper is addressed to detailed study of the path semigroup S and the C ∗-algebra C ∗ red(S). We construct the net of isomorphic C ∗- algebras {Aa, γba, a ≤ b}a,b∈K over the poset K, where Aa are restrictions of the algebra C ∗ red(S) on Hilbert subspaces and γba : Aa → Ab are - isomorphisms, such that γcb ◦ γba = γca for all a ≤ b ≤ c ∈ K. In the last section we consider extensions C ∗ red,∞(S) of the algebra C ∗ red,n(S) and also in C ∗ red(S) and C ∗ red(S). We prove that C ∗ red,∞(S). We show that quotient algebras C ∗ red(S) is an ideal in C ∗ red(S) are isomorphic to the Cuntz algebra. red,∞(S)/C ∗ Several works in recent years have addressed the C ∗-algebras generated by the left regular representations of semigroups with reduction [8] and by the representations of an inverse semigroup [9 -- 11]. In the paper [12] have shown that the Cuntz algebra can be represented as a C ∗-crossed product by endomorphisms of the CAR algebra. red,n(S) and C ∗ red,n(S)/C ∗ 2. Path semigroup In this section we define the path semigroup S on a partially ordered set K. The semigroup S is an inverse semigroup and has subgroups Ga corresponding to loops which start and end at the same point a ∈ K. Let K be a partially ordered set with binary relation ≤ satisfying reflexivity, antisymmetry and transitivity conditions. We call the set K a poset. Elements a and b are called comparable on K if a ≤ b or b ≤ a. We say that the poset K is upward directed if for every pair a, b ∈ K there exists c ∈ K, such that a ≤ c and b ≤ c. We call an ordered pair of comparable elements a and b on K an elementary path. We denote it by (b, a) if b ≤ a and by (b, a) if b ≥ a and say that a is a starting point of p, and b is an ending point. We use the notation ∂1p = a to denote the starting point of p and ∂0p = b to denote the ending point. For an elementary path p = (b, a) we define the inverse path p−1 = (a, b). For p = (b, a) the inverse path is p−1 = (a, b). Obviously, (p−1)−1 = p. Finally we call the pair (a, a) = (a, a) = ia a trivial path. C ∗-ALGEBRA GENERATED BY THE PATH SEMIGROUP 5 Let p1, . . . , pn be elementary paths, such that ∂0pi−1 = ∂1pi for i = 2, . . . , n. We define a path p as the sequence p = pn ∗ pn−1 ∗ . . . ∗ p1. The starting point of p is ∂1p = ∂1p1 and the ending point is ∂0p = ∂0pn. For every path p = pn ∗ pn−1 ∗ . . . ∗ p1 the inverse path is p−1 = p−1 1 ∗ p−1 2 ∗ . . . ∗ p−1 n with ∂1p−1 = ∂0p and ∂0p−1 = ∂1p. Let us consider the set of all paths on K. We denote an empty path 0 as a formal symbol. The empty path 0 has neither the starting point nor the ending point. We define a semigroup structure on the set of all paths with the empty path by extending the operation "∗" to multiplication as p ∗ q =(cid:26) p ∗ q 0 for all paths p and q. if p 6= 0, q 6= 0 and ∂1p = ∂0q, otherwise The poset K is called connected if for all a, b ∈ K there exists a path p, such that ∂0p = a, ∂1p = b. Throughout the rest of this article we assume K be a connected set. We call the set of all paths on K with the empty path a path semigroup and denote it by S if for all a, b, c ∈ K, such that a ≤ b ≤ c, the following axioms hold: 1. (a, b) ∗ (b, c) = (a, c); 2. (c, b) ∗ (b, a) = (c, a); 3. (b, a) ∗ (a, b) = ib, 4. (a, b) ∗ ib = (a, b), 5. (b, a) ∗ ia = (b, a), 6. ia ∗ ia = ia. It is easy to see that path semigroup S has the following useful prop- (a, b) ∗ (b, a) = ia; ia ∗ (a, b) = (a, b); ib ∗ (b, a) = (b, a); erties: 1) for every p ∈ S, such that ∂0p = a, ∂1p = b, p−1 ∗ p = ib, p ∗ p−1 = ia; 2) for every p ∈ S, such that ∂0p = a, ∂1p = b, ia ∗ p = p ∗ ib = p; 3) for all p, q ∈ S, such that ∂0q = ∂1p, (p ∗ q)−1 = q−1 ∗ p−1; 4) for all p, q, s ∈ S if p ∗ q = p ∗ s 6= 0 or q ∗ p = s ∗ p 6= 0 then q = s; so the path semigroup S is a semigroup with a reduction. 6 GRIGORYAN, GRIGORYAN, LIPACHEVA, SITDIKOV Thus, we can write elements of S as follows: (1) p = (a2n, a2n−1) ∗ . . . ∗ (a3, a2) ∗ (a2, a1) ∗ (a1, a0). Here elementary paths of type (a, b) and (a, b) alternate with each other. Note that there exists a variety of representations of type (1) for a path p. Our definition of the path turns out to be in full accordance with the definition given in [4]. The multiplication (ai+1, ai) ∗ (ai, ai−1) is 1- simplex with support ai where elements ai−1, ai, ai+1 are 0-simplices (see definitions of 0-simplex and 1-simplex in [2 -- 4]). Three elements a, c, x ∈ K, such that a, c ≤ x, form a 1-simplex de- noted by with support x. An inverse 1-simplex is [axc] = (a, x) ∗ (x, c) [cxa] = (c, x) ∗ (x, a) with the same support. In general a 1-simplex depends on the support. But for example if x, y ∈ K are comparable elements then (2) [axc] = [ayc]. Indeed for x ≤ y we observe [ayc] = (a, y)∗(y, c) = (a, x)∗(x, y)∗(y, x)∗(x, c) = (a, x)∗ix∗(x, c) = [axc]. In Lemma 4 we show that a 1-simplex does not depend from the support if the poset is upward directed. Therefore, one can rewrite the path (1) as a sequence of 1-simplices: p = [a2n a2n−1 a2n−2] ∗ . . . ∗ [a2 a1 a0]. Let us recall the definition of an inverse semigroup (for details see [13 -- 15]). Let S be a semigroup. Elements a, b ∈ S are called mutual inverses if a = aba, b = bab. The semigroup S is called an inverse semigroup if for every a ∈ S there exists a unique inverse element b ∈ S. We use the following theorem in the proof of Lemma 1. Theorem 1 ( [15] ). For a semigroup S in which every element has an inverse, uniqueness of inverses is equivalent to the requirement that all idempotents in S commute. Lemma 1. The path semigroup S is an inverse semigroup. C ∗-ALGEBRA GENERATED BY THE PATH SEMIGROUP 7 Proof. Let p ∈ S be a path with a starting point ∂1p = a and an ending point ∂0p = b. For every p there is an inverse path p−1, such that p ∗ p−1 ∗ p = ib ∗ p = p, p−1 ∗ p ∗ p−1 = ia ∗ p−1 = p−1. Hence, p and p−1 are mutual inverses elements. For every a ∈ K we have ia ∗ ia = ia and ia ∗ ib = 0 for all a 6= b. Therefore the set {ia}a∈K forms a commutative subsemigroup of idempotents in the path semigroup S. Hence, by Theorem 1 the path semigroup S is an inverse semigroup. (cid:3) Lemma 2. If for some 1-simplices [axb] and [byc] there exists z ∈ K, such that x, y ≤ z, then [axb] ∗ [byc] = [azc]. Proof. We have [axb] ∗ [byc] = (a, x) ∗ (x, b) ∗ (b, y) ∗ (y, c) = (a, x) ∗ (x, z) ∗ (z, x) ∗ (x, b) ∗ (b, y) ∗ (y, z) ∗ (z, y) ∗ (y, c) = (a, z) ∗ (z, b) ∗ (b, z) ∗ (z, c) = (a, z) ∗ (z, c) = [azc]. (cid:3) Corollary 1. If for some 1-simplices [axb], [byc] and [azc] there exists w ∈ K, such that x, y, z ≤ w, then [axb] ∗ [byc] = [azc]. Proof. Using the Lemma 2 and the equality (2) we have [axb] ∗ [byc] = [awc] = [azc]. (cid:3) In the works [3,4] there exists the notion of an elementary deformation of a path. They say that a path admits an elementary deformation if one can replace some section [axb] ∗ [byc] of the path with [azc] and vice versa. It is possible in the conditions of the Corollary 1. If we can obtain a path q ∈ S from some path p ∈ S by a finite number of elementary deformations then according to the Lemma 2 and the Corollary 1 we have the equality q = p. We say that p ∈ S is a loop if ∂0p = ∂1p. Let us denote by Ga the set of all loops that start and end in the point a. Lemma 3. The following statements hold: 1) the set Ga is a subgroup in S with a unit ia; 2) each path p generates isomorphism between groups Ga and Gb if ∂0p = a, ∂1p = b; 3) if p, q ∈ S and ∂0p = ∂0q = a, ∂1p = ∂1q = b, then there exist g1 ∈ Ga and g2 ∈ Gb, such that p = g1 ∗ q = q ∗ g2. 8 GRIGORYAN, GRIGORYAN, LIPACHEVA, SITDIKOV Proof. 1) The first statement is obvious. 2) Define a map γp : Ga → Gb in the following way: γp(g) = p−1gp, where g ∈ Ga. One can check that γp is an isomorphism. 3) It is easy to see that the statement holds for g1 = p ∗ q−1 ∈ Ga and (cid:3) g2 = q−1 ∗ p ∈ Gb. Lemma 4. If the poset K is an upward directed set then the following statements hold: 1) for all a, b, x, y ∈ K if a, b ≤ x and a, b ≤ y then [axb] = (a, x) ∗ (x, b) = (a, y) ∗ (y, b) = [ayb]; for simplicity let us omit supports and denote a 1-simplex by [a, b]; 2) [a, b] ∗ [b, c] = [a, c] for all a, b, c ∈ K; 3) for every p ∈ S if ∂0p = a and ∂1p = b then p = [a, b]; 4) if g ∈ Ga then g = ia and the group Ga is a trivial group. Proof. 1) As the poset K is upward directed set then there exists z ∈ K, such that x, y ≤ z. Hence, we have [axb] = (a, x) ∗ (x, b) = (a, x) ∗ (x, z) ∗ (z, x) ∗ (x, b) = (a, z) ∗ (z, b) = = (a, y) ∗ (y, z) ∗ (z, y) ∗ (y, b) = (a, y) ∗ (y, b) = [ayb]. 2) It follows from Lemma 2. 3) It follows from 2). 4) For every g ∈ Ga we have g = [a, an] ∗ . . . ∗ [a2, a1] ∗ [a1, a]. Using 2) several times, one gets g = [a, a1]∗[a1, a] = [a, a] = (a, a)∗(a, a) = ia. (cid:3) 3. C ∗-algebra C ∗ red(S) In this section we define the C ∗-algebra C ∗ red(S) generated by the rep- resentation of the path semigroup S and obtain the net of isomorphic C ∗-algebras (Aa, γba, a ≤ b)a,b∈K over the poset K, where γba : Aa → Ab are -isomorphisms satisfying the identity γcb ◦ γba = γca for a ≤ b ≤ c. Let us consider a Hilbert space l2(S) =(f : S → C Xp∈S f (p)2 < ∞) with an inner product hf, gi = Pp∈S f (p)g(p). A family of functions {ep}p∈S is an ortonormal basis of l2(S) where ep(p′) = δp,p′ is a Kronecker symbol. Let B(l2(S)) be the algebra of all linear bounded operators acting on l2(S). C ∗-ALGEBRA GENERATED BY THE PATH SEMIGROUP 9 Define a representation π : S → B(l2(S)) by π(p) = Tp where Tpeq =(cid:26) ep∗q 0 if ∂1p = ∂0q, otherwise. Note that π is the left regular representation and coincides with the Vagner representation of an inverse semigroup (see the definition of the Vagner representation in [14]). We have hTpeq, eri 6= 0 if and only if p ∗ q = r or q = p−1 ∗ r. Hence, hTpeq, eri = heq, Tp−1eri . Define the adjoint operator T ∗ p = Tp−1. In Lemma 5 we show that operators Tp and T ∗ p are partial isometric operators. Given a ∈ K we define Sa = {p ∈ S ∂0p = a}. Thus l2(S) can be written as l2(S) = ⊕ a∈K l2(Sa). Lemma 5. The following statements hold: 1) for every p ∈ S, such that ∂0p = a, ∂1p = b, the operator Tp is a p is an inverse mapping mapping from l2(Sb) to l2(Sa) and the operator T ∗ from l2(Sa) to l2(Sb); 2) for every p ∈ S, such that ∂0p = a, ∂1p = b, operators Ia = TpT ∗ p and Ib = T ∗ p Tp are projectors on l2(Sa) and l2(Sb) respectively; 3) for every g ∈ Ga the operator Tg is a unitary operator on l2(Sa); 4) for all p, q ∈ S, such that ∂0p = ∂0q = a, ∂1p = ∂1q = b, there exist g1 ∈ Ga and g2 ∈ Gb, such that Tp = Tg1Tq = TqTg2. Proof. 1) We observe that Tpeq = ep∗q if ∂0q = b and Tpeq = 0 otherwise. Since ∂0(p ∗ q) = a then Tp : l2(Sb) → l2(Sa). Similarly, T ∗ p : l2(Sa) → l2(Sb). 2) It is easy to see that Iaeq = TpT ∗ p eq = ep∗p−1∗q = eq if ∂0q = a and Iaeq = 0 otherwise. Therefore, Ia is a projector on l2(Sa). Similarly, one can prove that Ib is a projector on l2(Sb). 3) We have Tg : l2(Sa) → l2(Sa) and TgT ∗ g ep = eg∗g−1∗p = ep, T ∗ g Tgep = ep for every p ∈ Sa. Hence, Tg is a unitary operator. 4) This statement follows from the Lemma 3 (item 3). (cid:3) Let us denote by C ∗ red(S) a uniformly closed subalgebra of B(l2(S)) generated by operators Tp for every p ∈ S. Obviously the set of finite linear combinations of operators Tp, p ∈ S, is dense in C ∗ red(S). Given a ∈ K we denote Sa = {p ∈ S ∂1p = a}. Thus we have again l2(S) = ⊕ a∈K l2(Sa). 10 GRIGORYAN, GRIGORYAN, LIPACHEVA, SITDIKOV Theorem 2. The following statements hold: 1) the algebra C ∗ 2) C ∗ red(S) = ⊕ a∈K red(S) is irreducible on l2(Sa) for every a ∈ K; C ∗ red(S)l2(Sa) and every operator A ∈ C ∗ red(S) can be represented as A = ⊕ a∈K Aa where Aa = Al2(Sa); 3) if the group Ga is non-trivial then C ∗ red(S)l2(Sa) doesn't contain com- pact operators. Proof. 1) The set {ep, ∂1p = a}p∈S is a basis of l2(Sa). For all p1, p2 ∈ Sa and p = p2 ∗ p−1 red(S) is irreducible on l2(Sa). 1 we have Tpep1 = ep2. It means that the algebra C ∗ 2) This statement follows from the fact that for every p ∈ S operator Tp maps the space l2(Sa) onto itself for every a ∈ K. 3) Let p ∈ Sa, g ∈ Ga and g 6= ia. Consider the sequence xn = ep∗gn . Since g ∗ g 6= g elements of the sequence where gn = g ∗ g ∗ . . . ∗ g n {z } {xn} are pairwise orthogonal. If A ∈ C ∗ red(S)l2(Sa) is a compact operator αiepi where pi ∈ Sa and αi are complex coefficients. Referral to the fact that A is approximated by finite linear combinations of operators Tq, q ∈ S, and αiepi∗g. Similarly then kAxnk → 0 as n → ∞. On the other hand Aep = Pi to the equality Tqep∗g = eq∗p∗g we obtain Aep∗g = Pi Aep∗gn = Pi (cid:18)Pi αi2(cid:19)1/2 αiepi∗gn for all n. Therefore, for every n we have kAxnk = > 0. Hence, A is not a compact operator. Theorem 3. Let K be an upward directed set. Then the following state- ments hold: (cid:3) 1) for every p ∈ S, such that ∂0p = a, ∂1p = b, we have Tp = T[a,b]; 2) for every a ∈ K the algebra C ∗ red(S)l2(Sa) coincides with the algebra of all compact operators on B(l2(Sa)); red(S) is non-unital. 3) the algebra C ∗ Proof. 1) This statement follows from the Lemma 4. 2) The set {e[c,a]}c∈K is a basis of l2(Sa). For every operator Tp we 6= 0 if and only if ∂1p = c. Hence, Tp = T[b,c] for some b have Tpe[c,a] and T[b,c]e[c,a] = e[b,a]. Therefore, Tpl2(Sa) is a one dimensional operator. So C ∗-algebra C ∗ red(S)l2(Sa) coincides with the algebra of all compact operators on B(l2(Sa)). 3) By the Theorem 2 for every element A ∈ C ∗ If the algebra C ∗ Aa where Aa ∈ C ∗ red(S)l2(Sa). red(S) we have A = red(S) has the unit ⊕ a∈K C ∗-ALGEBRA GENERATED BY THE PATH SEMIGROUP 11 I then Ia = Il2(Sa) is a compact operator in the infinite dimensional Hilbert space. This is a contradiction. (cid:3) Given a ∈ K we denote Aa = C ∗ red(S)l2(Sa). Theorem 4. There exists the set of -isomorphisms {γba, a ≤ b}a,b∈K: γba : Aa → Ab, such that γcb ◦ γba = γca for all a, b, c ∈ K and a ≤ b ≤ c. And we obtain a net of isomorphic C ∗-algebras {Aa, γba, a ≤ b}a,b∈K over the poset K. Proof. Define a unitary operator Uab : l2(Sa) → l2(Sb) for all a, b ∈ K, a ≤ b, by for every q ∈ Sa. Then U ∗ operator. Obviously, U ∗ define a mapping γba : Aa → Ab by Uabeq = eq∗(a,b) ab = Uba : l2(Sb) → l2(Sa) is the adjoint ab = idl2(Sb). Let us abUab = idl2(Sa) and UabU ∗ γba(A) = UabAU ∗ ab for every A ∈ Aa. One can check that γba is the -isomorphism. It remains to check the equality γcb ◦ γba = γca for a ≤ b ≤ c. We observe that (γcb ◦ γba)(A) = γcb(γba(A)) = UbcUabAU ∗ abU ∗ bc for every A ∈ Aa. Otherwise UbcUabeq = Ubceq∗(a,b) = eq∗(a,b)∗(b,c) = eq∗(a,c) = Uaceq for every q ∈ Sa and similarly U ∗ (γcb ◦ γba)(A) = γca(A) for every A ∈ Aa. abU ∗ bcep = U ∗ acep for every p ∈ Sc. So (cid:3) Remark 1. The set of isomorphisms {γba, a ≤ b}a,b∈K can be extended from elementary paths to 1-simplices {γ[bxa], a, b ≤ x}a,b,x∈K by γ[bxa] = γ−1 xb ◦ γxa, so that they satisfy 1-cocycle identity [4]: γ[cyb] ◦ γ[bxa] = γ[cza] for [cyb] ∗ [bxa] = [cza]. Extending the set {γ[bxa], a, b ≤ x}a,b,x∈K to paths we get the set of isomorphisms {γp}p∈S satisfying the equality γp2 ◦ γp1 = γp2∗p1 for all p1, p2 ∈ S and ∂0p1 = ∂1p2. 4. Extensions of the C ∗-algebra C ∗ red(S) In this section we consider the extensions of the algebra C ∗ red(S), such that this algebra is an ideal in that extensions and quotient algebras are isomorphic to the Cuntz algebra. 12 GRIGORYAN, GRIGORYAN, LIPACHEVA, SITDIKOV Let K be an upward directed countable set. By the lemma 4 for every path p ∈ S, such that ∂0p = a, ∂1p = b, we have p = [a, b]. Let us represent the set K as a finite union of countable disjoint sets K = Ei, n [i=1 We define one-to-one mappings φi : Ei → K, i = 1, . . . , n, and opera- T[x,φi(x)], i = 1, . . . , n. tors Tφi : l2(S) → l2(S) in the following way: where EiT Ej = ∅ for i 6= j. Tφi = Mx∈Ei [x,φi(x)] = Mx∈Ei φi = Mx∈Ei T ∗ T ∗ . An adjoint operator of the operator Tφi is The following equalities hold: T ∗ φiTφi = id; T ∗ φiTφj = 0, i 6= j; T[φi(x),x] =Mx∈K T[x,φ−1 i (x)] TφiT ∗ φi = id. n Xi=1 Indeed every basis element has a form e[a,b]. Therefore, φiTφie[a,b] = T ∗ T ∗ φiT[φ−1 i (a),a]e[a,b] = T ∗ φie[φ−1 i (a),b] = = T[a,φ−1 i (a)]e[φ−1 i (a),b] = e[a,b]. Analogously, since EiT Ej = ∅ we have T ∗ then φiTφj e[a,b] = 0. Finally if a ∈ Ek n Xi=1 TφiT ∗ φi! e[a,b] = Tφk T[φk(a),a]e[a,b] = Tφke[φk(a),b] = = T[a,φk(a)]e[φk(a),b] = e[a,b]. Let us consider a uniformly closed subalgebra of B(l2(S)) generated by red,n(S). The red(S). It is operators Tp, p ∈ S, and Tφi, i = 1, . . . , n. Denote it by C ∗ algebra C ∗ an extension of algebra C ∗ red,n(S) is unital. Hence, it doesn't coincide with C ∗ red(S). Moreover the following lemma holds. Lemma 6. The algebra C ∗ red(S) is an ideal in C ∗ red,n(S). Proof. We have TφiT[a,b] = T[x,b] for some x ∈ K and T[a,b]Tφi = T[a,y] for some y ∈ K. Since every element A ∈ C ∗ red(S) can be approximated by finite linear combinations of operators T[a,b] then TφiA and ATφi ∈ C ∗ (cid:3) red(S). C ∗-ALGEBRA GENERATED BY THE PATH SEMIGROUP 13 Let us recall the definition of the Cuntz algebra. The finite Cuntz algebra On is a C ∗-algebra generated by isometries s1, . . . , sn satisfying to the following conditions: s∗ i sj = δijid, sis∗ i = id. n Xi=1 The infinite Cuntz algebra O∞ is a C ∗-algebra generated by s1, s2, . . . and relations for every n ∈ N. s∗ i sj = δijid, sis∗ i ≤ id n Xi=1 Theorem 5. There exist an isomorphism C ∗ a short exact sequence red,n(S)/C ∗ red(S) ∼= On and 0 → C ∗ red(S) id→ C ∗ red,n(S) π→ On → 0, where id is an embedding map and π is a quotient map. Proof. Equivalence classes [Tφi] = Tφi + C ∗ tors of the quotient algebra C ∗ operators satisfying the following identity: red,n(S)/C ∗ red(S), i = 1, . . . , n, are genera- red(S). These classes are isometric [Tφi][T ∗ φi] = id. n Xi=1 Due to the universality of the Cuntz algebra we observe that C ∗ red,n(S)/C ∗ red(S) ∼= On. Now let us represent the set K as a countable union of disjoint count- (cid:3) able sets: K = Ei ∞ [i=1 and define operators Tφi : l2(S) → l2(S) in the following way: Tφi = Mx∈Ei T[x,φi(x)], i = 1, 2, . . . . By applying the reasoning used above one can prove the following equal- ities: T ∗ φiTφi = id; T ∗ φiTφj = 0, i 6= j; for every n ∈ N. TφiT ∗ φi ≤ id n Xi=1 14 GRIGORYAN, GRIGORYAN, LIPACHEVA, SITDIKOV Let us denote by C ∗ red,∞(S) the uniformly closed subalgebra of B(l2(S)) generated by operators Tp, p ∈ S, and Tφi, i = 1, 2, . . .. Similarly to the Lemma 6 the algebra C ∗ red(S) is an ideal in C ∗ red,∞(S) and for the infinite Cuntz algebra the following theorem holds. Theorem 6. There exist an isomorphism C ∗ a short exact sequence red,∞(S)/C ∗ red(S) ∼= O∞ and 0 → C ∗ red(S) id→ C ∗ red,∞(S) π→ O∞ → 0, where id is an embedding map and π is a quotient map. Acknowledgements. We thank E. Vasselli for helpful comments which have led to significant improvements. References [1] R. Haag, Local quantum Physics: Fields, particles, algebras (Springer-Verlag, Berlin, 1992). [2] G. Ruzzi, Homotopy of posets, net cohomology and superselection sectors in globally hyperbolic space-times, Rev. Math. Phys., 17, 1021-1070 (2005). [3] J.E. Roberts, More lectures in algebraic quantum field theory, in: S. Doplicher, R. Longo (Eds.), Noncommutative Geometry. C.I.M.E. Lectures, Martina Franca, Italy, 2000, Springer-Verlag (2003). [4] G. Ruzzi, E. Vasselli, A new light on nets of C-algebras and their representa- tions, Comm. Math. Phys., 312, 655-694 (2012). [5] E. Vasselli, Presheaves of symmetric tensor categories and nets of C-algebras, Journal of Noncommutative Geometry, 9, 121-159 (2015). [6] E. Vasselli, Presheaves of superselection structures in curved spacetimes, Comm. Math. Phys., 335, 895-933 (2015). [7] R. Brunetti, G. Ruzzi, Quantum charges and space-time topology: The emer- gence of new superselection sectors, Comm. Math. Phys., 287, 523-563 (2009). [8] M.A. Aukhadiev, S.A. Grigoryan, E.V. Lipacheva, Infinite-dimensional com- pact quantum semigroup, Lobachevskii Journal of Mathematics, 32 (4), 304-316 (2011). [9] M.A. Aukhadiev, V.H. Tepoyan, Isometric representations of totally ordered semigroups, Lobachevskii Journal of Mathematics, 33 (3), 239-243 (2012). [10] S.A. Grigoryan, V.H. Tepoyan, On isometric representations of the perforated semigroup, Lobachevskii Journal of Mathematics, 34 (1), 85-88 (2013). [11] T.A. Grigoryan, E.V. Lipacheva, V.H. Tepoyan, On the extension of the Toeplitz algebra, Lobachevskii Journal of Mathematics, 34 (4), 377-383 (2013). [12] M.A. Aukhadiev, A.S. Nikitin, A.S. Sitdikov, Crossed product of the canonical anticommutative relations algebra in the Cuntz algebra, Russian Mathematics, 58 (8), 71-73 (2014). [13] A.H. Clifford, G.B. Preston, The algebraic theory of semigroups V.1. (AMS, 1964). C ∗-ALGEBRA GENERATED BY THE PATH SEMIGROUP 15 [14] Alan L.T. Paterson, Groupoids, Inverse Semigroups, and their Operator Algebras (Birkhauser, 1998). [15] V.V. Vagner, Generalized groups (Russian), Doklady Akad. Nauk SSSR 84, 1119- 1122 (1952).
1102.0304	1	1102	2011-02-01T21:40:27	An Operator Space duality theorem for the Fourier-Stieltjes algebra of a locally compact groupoid	[ "math.OA" ]	It is a well-known result of Eymard that the Fourier-Stieltjes algebra of a locally compact group $G$ can be identified with the dual of the group $\cs$ $C^{}(G)$. A corresponding result for a locally compact groupoid $G$ has been investigated by Renault, Ramsay and Walter. We show that the Fourier-Stieltjes algebra $B_{\mu}(G)$ of $G$ (with respect to a quasi-invariant measure $\mu$ on the unit space $X$ of $G$) can be characterized in operator space terms as the dual of the Haagerup tensor product $\ov{L^{2}(X,\mu)}^{r}\otimes_{hA}C^{}(G,\mu)\otimes_{h A}L^2(X,\mu)^c$ and as the space of completely bounded bimodule maps $CB_{A}(C^{}(G,\mu),B(L^2(X,\mu)))$, where $A=C_{0}(X)$ and $C^{}(G,\mu)$ is the groupoid $\cs$ obtained from those $G$-representations associated with $\mu$. A similar but different result has been given by Renault, but our proof is along different lines, and full details are given. Examples illustrating the result are discussed.	math.OA	math	AN OPERATOR SPACE DUALITY THEOREM FOR THE FOURIER-STIELTJES ALGEBRA OF A LOCALLY COMPACT GROUPOID ALAN L. T. PATERSON 1. Introduction This paper studies a version of the Fourier-Stieltjes algebra for groupoids. Before considering this in more detail, and for motivation, let us first con- sider the (much studied) Fourier-Stieltjes algebra B(G) for a locally compact group. Specializing even further, let us start with a locally compact abelian group G with dual group G. The Fourier algebra, A(G), and the Fourier- Stieltjes algebra B(G) can then be defined in commutative harmonic analysis as L1( G) and M ( G) respectively. These, of course, are commutative convo- lution Banach algebras, and using the inverse Fourier transform, one usually regards A(G), B(G) as commutative Banach algebras of continuous bounded functions on the group G itself under pointwise multiplication and a certain norm (not the supremum norm). Under this identification, positive mea- sures on M ( G) correspond to positive definite functions on G, and these span B(G). Further, point masses on G correspond to the characters on G. The positive definite functions on G in turn correspond, by integration against L1(G)-functions, to the states of C ∗(G). B(G) as the span of the positive definite functions on G then translates over to B(G) as the span of the states on C ∗(G), and in fact identifies as a Banach space with the dual C ∗(G)∗. In terms of the original definition, i.e. taking B(G) = M ( G) and using C0( G) = C ∗(G), this just amounts to saying that C0( G)∗ = M ( G). Turning to the non-abelian locally compact group case, the set G of classes of irreducible representations no longer forms a group, the irreducible repre- sentations being usually no longer one-dimensional. So it is not clear how to define B(G) in terms of some "M ( G)". However, Pierre Eymard ([9]) was able to show that much of the theory of A(G) and B(G) goes through in suitable form provided we stay with functions on G and omit any mention of "L1( G)" and "M ( G)". So B(G) is taken to be the span of the positive definite functions on the group G. (Note that all such functions are contin- uous and bounded on G - as we will see later, a technical difficulty arises Date: December, 2010. 1991 Mathematics Subject Classification. 43A32. Key words and phrases. Fourier-Stieltjes algebra, locally compact groupoids, positive definite functions, operator spaces, completely bounded maps, Haagerup tensor product. 1 2 ALAN L. T. PATERSON at this point in the locally compact groupoid case since measurable positive definite functions need not be continuous ([24]).) To define the norm for B(G) one regards, as above, positive definite func- tions on the group as states on C ∗(G) and so can simply identify their span B(G) with the dual space C ∗(G)∗. However, it is not immediately clear from this definition of B(G) what its algebra product is, or how its norm can be defined more concretely (rather than just as a Banach space dual norm). Another disadvantage of staying with the equality "B(G) = C ∗(G)∗ is that an early result of Walter ([30]) showed that B(G) is associated with completely bounded maps on C ∗(G), which suggests that there is an op- erator space side to B(G) whose ramifications are not brought out if we regard it only as a dual Banach space. Fortunately there are three other ways that one can regard the equality, which clarify these issues and which are pertinent for the groupoid case. The first way is (at first sight) pointless. One writes B(G) = (C ⊗ C ∗(G) ⊗ C)∗. However, it turns out that this tensor product formulation, in an operator space context, is fundamental for the study of the groupoid case. (This was first observed by Jean Renault in [28].) Indeed, even in the group case, the equality B(G) = (C ⊗ C ∗(G) ⊗ C)∗ becomes more informative when it is replaced by the corresponding equality of operator spaces, precisely, are respectively the Hilbert B(G) = (C space C with its column operator space structure and the conjugate C of C with its row operator space structure, and the Haagerup tensor product is used. (A brief introduction to the operator space theory used in the paper is given in the second section.) In the groupoid case, C above for the group case gets replaced by an L2-space on the unit space of the groupoid. ⊗h C ∗(G) ⊗h Cc)∗ where Cc, C r r Reverting back to our discussion of the equality B(G) = C ∗(G)∗ for the group case (in terms of "normal" Functional analysis rather than operator space theory), the second way gives a formula for calculating the norm of B(G). (We know what B(G) is as an algebra - the span of the positive definite functions on the group - so the problem is to give a concrete way of calculating the C ∗(G)∗-norm of the elements of B(G).) To do this, it is helpful to realize the elements of B(G) as "coefficients" of (unitary) rep- resentations of G as follows. By the Gelfand-Naimark-Segal construction, every positive definite function φ on G gives rise to a representation L of G on some Hilbert space H, and there is a vector ξ ∈ H such φ(g) = hLgξ, ξi for all g ∈ G. More generally, using the polarization identity and direct sums of representations, for any φ ∈ B(G), there exists a representation L of G on some Hilbert space H and elements ξ, η ∈ H such that for all g ∈ G, φ(g) = hLgξ, ηi = (ξ, η)(g). Considering all possible ways of representing φ as some (ξ, η) gives a way to define the norm on B(G). Indeed, using the Jordan decomposition for A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 3 continuous linear functionals on a C ∗-algebra, Eymard showed that (1.1) kφk = inf{kξkkηk : φ = (ξ, η)} the inf being taken over all ways of writing φ = (ξ, η) over all possible representations of G. This way of looking at B(G) also gives the product of two elements of B(G) by just tensoring representations. There is yet a third way to regard B(G) and this is as an algebra of completely bounded operators on C ∗(G). This was proved by Walter in [30]. Indeed, if φ ∈ B(G), then pointwise multiplication by φ, f → f φ, takes L1(G) to itself. In fact, this map is continuous under the C ∗(G)-norm restricted to L1(G), and since L1(G) is dense in C ∗(G), extends to a bounded linear map Tφ on C ∗(G). Walter showed that Tφ in fact is a completely bounded map from C ∗(G) into C ∗(G) and the completely bounded norm kTφkcb of Tφ equals kφk. We might now ask if the theory of B(G) can be extended to the locally compact groupoid case, and if there are correspondingly three ways of re- garding B(G). Justification for the study of B(G) for a locally compact groupoid G lies in the fact that many C ∗-algebras arise naturally as gener- ated by groupoids, so that for such C ∗-algebras, groupoids play the role that the group does for group C ∗-algebras. Examples of this include transforma- tion group C ∗-algebras, the operator algebras associated with an equivalence relation, the C ∗-algebra of the holonomy groupoid of a foliation, and graph C ∗-algebras. Given the importance of B(G) in the study of harmonic analy- sis on groups, in particular, for duality, it is reasonable to expect an equally fundamental role for B(G) in the groupoid case. Despite it being "early days", there is a small but growing literature on B(G) (and the related space A(G)) for locally compact groupoids. The first study in this direc- tion seems to have been made by Walter ([30]) who examined the case where G = {1, 2, . . . , n}×{1, 2, . . . , n}. Other papers include [15, 16, 18, 19, 24, 28]. In the case of a group, there is, of course, only one unit - the identity element e - but in the case of a groupoid there are usually many units. In the case of a second countable, locally compact groupoid G with unit space X and left Haar system {λx}, one obtains the C ∗-algebras generated by the convolution algebra Cc(G) by combining two ingredients, first a special kind of probability measure µ on X - called a quasi-invariant measure - with a measurable Hilbert bundle on which the groupoid acts measurably. In the group case, there is only one quasi-invariant measure, that of the point mass at e, and we don't need to mention it explicitly in group representation theory. In the groupoid case, one integrates the G-action against a kind of "product measure" λµ = RX λx dµ(x) (including a "modular function" contribution) to obtain a ∗-representation π of Cc(G) on the Hilbert space of L2-cross-sections of the Hilbert bundle. With µ fixed, the largest C ∗-algebra that can be obtained in this way will be called C ∗(G, µ). This C ∗-algebra has to be distinguished from the C ∗-algebra C ∗ µ(G) of [28] (defined in the third section of the present paper) which is usually much larger than C ∗(G, µ). In 4 ALAN L. T. PATERSON this paper, we will be concerned only with C ∗(G, µ) (though see later in this Introduction). The "largest" C ∗-algebra obtained in this way, as µ ranges over the set of quasi-invariant measures on X, is the full C ∗-algebra of the groupoid, C ∗(G), the canonical C ∗-algebra associated with G. The two papers that are fundamental for the present work are those of Renault ([28]) and of Ramsay and Walter ([24]), and both define their B(G) in a way similar to that of the group case, i.e. as the span of a sutiable class of positive definite functions on G but with certain identifications de- termined by the quasi-invariance. (The group notion of positive definiteness itself extends simply to the groupoid case.) Renault in [28] fixes a quasi- invariant measure µ and uses essentially bounded λµ-measurable positive definite functions, whereas Ramsay and Walter, in [24] uses bounded Borel positive definite functions, any two of which are identified if there is a Borel set N off which the two functions are equal and is such that λµ(N ) = 0 for all quasi-invariant measures µ on X. Continuity for positive definite functions is not assumed, as one might have expected, basically because in general there are not enough of them to allow the representation theory of G to be used effectively. In this paper we work, as does [28], in the context of a second countable locally compact groupoid G with left Haar system {λx} and with a given quasi-invariant measure µ on the unit space X of G, and write Bµ(G) in place of B(G) to emphasize the dependence of that space on µ. The Borel approach of [24] is closely related to this but differs in having to consider all quasi-invariant measures at once rather than fixing on one throughout. The relation between to the two theories will not be considered here. We saw earlier that in the group case, B(G)-functions can be represented in the form (ξ, η) for ξ, η in a representation Hilbert space of the group. A similar result holds in the groupoid case. Indeed ([24, §3], [28, 1.1]) the Hilbert space is replaced by a G-Hilbert bundle H = {Hx}x∈X and L is a Borel unitary action of G on H. (In particular, each Lg is unitary from Hs(g) onto Hr(g).) One defines a Bµ(G)-function (α, β) : G → C for two essentially bounded sections α, β of H by taking: (α, β)(g) = hLgαs(g), βr(g)i. As in the group case, every Bµ(G)-function can be realized as some (α, β) for some G-Hilbert bundle H. Furthermore, Renault shows that, also as in the group case, we obtain an involutive Banach algebra structure on Bµ(G) by using pointwise products and by defining for φ ∈ Bµ(G), kφk = inf{kαkkβk : φ = (α, β)}. This raises the natural question: is this a dual norm? That is, is there some natural Banach space Z such that Bµ(G) (with its norm) is the Banach space dual of Z. In the groupoid case, Renault solves this, using operator space theory, with a beautiful idea. In the group case (so µ = δe) and with φ = (ξ, η) (ξ, η ∈ H In the group case above, Z would be just C ∗(G). A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 5 as above), we identified φ as a linear functional θφ on L1(G) by setting θφ(F ) = hπ(F )ξ, ηi where π is the integrated form of the representation L of G on H. Then θφ extends to a continuous linear functional on C ∗(G) and the identification of Bµ(G) with C ∗(G)∗ follows. When we try to do this in the groupoid case, H becomes a G-Hilbert bundle H (fibered over the unit space X) and π becomes a unitary G-representation L on H. One integrates up the groupoid representation L to get a representation π of the Banach ∗-algebra LI(G) (a groupoid analogue of L1(G) for the group case) on the Hilbert space L2(X, H, µ) of square integrable sections of H, and replaces ξ, η of the group case by essentially bounded sections α, β of the bundle H. So the natural way to define a linear functional θφ on LI (G) is presumably to take "θφ(F ) = hπ(F )α, βi". But this leads to a difficulty since the α, β in the formula for θφ are essentially bounded elements of the Hilbert space L2(X, H, µ), and these are special elements of this Hilbert space. We want to involve the whole of this Hilbert space of L2-sections, not just the L∞- ones. Renault solves this difficulty by multiplying both α, β by functions a, b in L2(X, µ) and this gives two general L2-sections of H to which we can apply π(F ). One obtains a linear functional, not on C ∗ µ(G), but, taking the a, b into account, on L2(X, µ) ⊗ C ∗ µ(G) ⊗ L2(X, µ) by defining θφ(b ⊗ F ⊗ a) = hπ(F )(aα), bβi µ(G) ⊗ L2(X, µ), or rather a with φ = (α, β). To interpret L2(X, µ) ⊗ C ∗ module version of it, operator space ideas - in particular, the Haagerup (In the group case, L2(X, µ) = C which tensor product - are required. brings us back to our earlier discussion.) Also, as commented earlier, we will be using the smaller C ∗-algebra C ∗(G, µ) rather than the algebra C ∗ µ(G) used by Renault. The C ∗-algebra C ∗(G, µ) relates to the fundamental C ∗-algebras associated with a locally compact groupoid, in particular, to the full C ∗-algebra, C ∗(G), of the groupoid. We note that in the C ∗(G, µ)-context, C0(X) acts as the diagonal algebra, µ(G), the diagonal algebra is L∞(X, µ). The author whereas in the case of C ∗ has been unable to relate in a satisfactory way [28, Theorem 2.1, Proposi- tion 2.2] to the main theorem of this paper, Theorem 3. (In this regard, see the second Note of the third section and Example 3 of the last section of the paper.) Theorem 3 generalizes the three ways of looking at B(G) for the group case as discussed earlier in this Introduction. The proof itself is inspired by the proof of Renault's theorem for C ∗ µ(G), but the essential ideas are simplified by using module versions of a theorem by Effros and Ruan and the generalized Stinespring theorem. An effort has been made to give complete details of the proof. In particular, certain technical prob- lems involved in making the proof of Theorem 3 work are resolved in the preliminary propositions, Proposition 4, Proposition 5 and Proposition 6. The paper closes by discussing four examples illustrating the theorem. In 6 ALAN L. T. PATERSON particular, the fourth example shows how the complete boundedness theory of Schur products ([20]) fits into the groupoid framework of the theorem. The importance of operator space theory in noncommutative harmonic analysis on a locally compact group G was highlighted by the remarkable result of Ruan, described in [7, 16.2], that while the Fourier algebra A(G) need not be amenable as a Banach algebra when G is amenable, neverthe- less A(G) is operator amenable if and only if G is amenable. The work of this paper, complementing the work of Renault in [28], emphasizes the even more fundamental, pervasive, role that operator space theory plays in noncommutative harmonic analysis on locally compact groupoids. For this reason, a brief survey of the operator space theory that we will need for proving Theorem 3 is covered in the second section. The author is grateful to Zhong-Jin Ruan and Roger Smith for gener- ous help with the subject of operator space theory, and especially to Arlan Ramsay for many valuable conversations. 2. Preliminaries on operator space theory This section discusses the basic ideas and results from operator space theory that we will require later in the paper for application within the groupoid context. For ease of reading, an attempt has been made to make the discussion as self-contained as possible. The main references for this section are the books: [7] of Effros and Ruan, [20] of Paulsen, [21] of Pisier, and [2] of Blecher and Le Merdy. A (concrete) operator space can conveniently be defined as a subspace X of some B(H, K) (or equivalently, of some B(H)) where H, K are Hilbert spaces. Each of the spaces Mn(X) of n × n matrices with entries in X then has its natural norm when identified with a subspace of B(Hn, Kn): for T = [Ti,j] ∈ Mn(X), and an n-column vector [hj] in Hn we just use j=1 Tk,jhj to get a column vector in Kn. Of course, the spaces of rectangular matrices Mm,n(X) also have operator space norms k.km,n, and these can be reduced to the square case by adding or dropping rows or columns of zeros. We write k.kn = k.kn,n. matrix multiplication: (T h)k = Pn A remarkable abstract characterization of operator spaces - in which we are given a Banach space X and norms k.kn assigned to the spaces Mn(X) and you can tell, purely from two axioms for the norms k.kn, when X is a con- crete operator space - was given by Ruan. This is the representation theorem for operator spaces (e.g. [7, Theorem 2.3.5]). The two axioms are, first, that for x ∈ Mm(X), y ∈ Mn(X), we have kx ⊕ ykm+n = max{kxkm, kykn}, and second, that for α, β ∈ Mn and x ∈ Mn(X), we have kαxβkn ≤ kαkkxknkβk. One easily checks that for a concrete operator space, both axioms are sat- isfied. While the operator space structure on a Banach space X is given by a sequence of norms {k.kn} satisfying Ruan's axioms, it is often helpful to specify this structure by a concrete, Hilbert space realization. A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 7 If A is a C ∗-algebra then it becomes an operator space by representing it faithfully on a Hilbert space. One gets the same norm on Mn(A) whatever choice of representation is made, so this gives a canonical operator space structure on A. Every Banach space E is an operator space in at least one way: identify E canonically with a subspace of the commutative C ∗-algebra C(E∗ 1 is the unit ball of E∗ with the relative weak∗-topology. 1), where E∗ The class of operator spaces (X, {k.kn}) is a category in a natural way. To define the morphisms, let X, Y be operator spaces and Φ : X → Y be a linear map. Then in the obvious way, Φ gives rise to a linear map Φn : Mn(X) → Mn(Y ) by applying Φ to matrix entries. The map Φ is called completely bounded if kΦkcb = supn kΦnk < ∞. The set of completely bounded maps from X to Y is a Banach space (CB(X, Y ), k.kcb), and this is defined to be M or(X, Y ) in the category of operator spaces. A completely bounded map Φ : X → Y is called completely contractive if kΦkcb ≤ 1 and is called a complete isometry if each Φn : Mn(X) → Mn(Y ) is isometric. The key to the power of operator space theory lies in the fact that its morphisms, the completely bounded maps, are far more special than just Banach space morphisms (bounded linear maps) - in fact, in certain contexts, they are close to algebraic "homomorphisms". The remarkable theorem below, making sense of the last assertion, is due to Wittstock, Haagerup and Paulsen, after earlier work by Stinespring and Arveson. (There is an important analogue of Theorem 1 for bilinear completely bounded maps due to Christensen, Sinclair, Paulsen and Smith (e.g. [7, Corollary 9.4.5], [2, Theorem 1.5.7]) but we omit this since it will not be needed in the paper.) Theorem 1. Let B be a C ∗-algebra, H be a Hilbert space and Φ : B → B(H) be a completely bounded map. Then there exists a Hilbert space L, a ∗- representation π of B on L and bounded linear operators S, T : H → L such that for all b ∈ B, (2.1) Φ(b) = S∗π(b)T. Further, kΦkcb ≤ kSkkT k, and L, π, S and T can be taken so that kΦkcb = kSkkT k. Conversely, any linear map Φ : B → B(H) which has the form of the right-hand side of (2.1) is a completely bounded map. Operator spaces behave well under natural constructions. Obviously, ev- ery linear subspace V of an operator space X inherits from X an operator space structure. If V is also closed in X, then for each n, there is the nat- ural quotient norm k.kn on Mn(X/V ) = Mn(X)/Mn(V ), and under these norms, X/V is an operator space. The Banach space completion of an op- erator space is an operator space in the natural way (since the closure of a subspace of some B(H, K) is also such a subspace). If X, Y are operator spaces then CB(X, Y ) itself is an operator space, k.kn in this case being the norm obtained by identifying Mn(CB(X, Y )) with CB(X, Mn(Y )). Futher, the dual space X ∗ can be identified with the operator space CB(X, C), and 8 ALAN L. T. PATERSON so is a natural operator space. (Note ([7, Corollary 2.2.3]) that for f ∈ X ∗, kf kcb = kf k.) A Hilbert space H itself is, of course, not immediately given as an operator space, i.e. as a subspace of some B(H′, K′), but can be made into one in a number of different ways. Two particular operator space structures for H will concern us: these are the column Hilbert space Hc and the row Hilbert space Hr. In each case, we identify the operator space concretely in the form B(K, K′) for appropriate Hilbert spaces K, K′. First, we identify H with B(C, H), where ξ ∈ H is identified with the linear map Tξ that sends the scalar λ to λξ. To explain the "column" terminology let us suppose that H is finite-dimensional. Let {e1, . . . , en} be an orthonormal basis for H, n (i.e. Cn with the standard inner and identify, through this basis, H as ℓ2 product). Then Tξ is the n × 1 column matrix whose ith entry is hξ, eii. Further ([2, p.46]) if H, K are Hilbert spaces, then CB(Hc, Kc) = B(H, K) canonically. Second, since Hilbert spaces are reflexive, we can identify H with the dual of H∗ to obtain the row Hilbert space Hr. In more detail, identify H∗ with the conjugate Hilbert space H and let η → η be the canonical conjugation from H to H. So η : H → C is given by: η(ξ) = hξ, ηi. Note that scalar multiplication in H is given by: for λ ∈ C, λ.η = λη. We obtain the row Hilbert operator space Hr by associating any ξ ∈ H with the map Sξ ∈ B(H, C) that takes η ∈ H to hξ, ηi. Taking {e1, . . . , en} as the (orthonormal) basis for H, each Sξ is realized as the (1 × n)-row matrix with entries hξ, eii. This is why Hr r is called the row Hilbert space. If, instead of Hr, we look at H , then we obtain a nice duality relation between that operator space and Hc. In fact, since H = B(H, C) = B(H, C), we obtain r (2.2) (Hc)∗ = H r as operator spaces. While the above discussion of the column and row Hilbert operator spaces was within the finite-dimensional context, it can be carried through for general Hilbert spaces ([2, 1.2.23]). We now discuss operator modules for operator algebras. The discussion is based on the detailed account given in the book by Blecher and LeMerdy ([2, Chapter 3] ). An earlier account of some of this is also given in the mem- oir of Blecher, Muhly and Paulsen ([3, Chapter 2]). These accounts cover the case of general operator algebras, but we will only need to consider the C ∗-algebra case. In addition to the operator module concept discussed below, there are two other notions of "operator module" that are useful, as- sociated respectively with the Haagerup and the projective tensor products, but we will not have occasion to use them. So let A be a C ∗-algebra with, of course, its canonical operator space structure. Let X be a left Banach A-module that is an operator space. Then X is called a concrete left operator A-module if X is given as a closed linear subspace of some B(K, H) with the operator space structure inherited A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 9 from B(K, H), and is such that the module multiplication - a bilinear map - is given by a ∗-representation θ of A on the Hilbert space H with θ(A)X ⊂ X. (So for T ∈ B(K, H), a.T = θ(a)◦T .) An (abstract) left operator A-module is an X that is completely isometrically A-isomorphic to a concrete one. Right operator A-modules and operator A-bimodules are defined in the obvious ways. It is easy to check that if θ and π are ∗-representations of A on the Hilbert spaces H, K respectively, then B(K, H) itself is a concrete operator A-bimodule, the module actions being the natural ones in which, for a, b ∈ A and T ∈ B(K, H), aT b = θ(a) ◦ T ◦ π(b). With θ as above, the Hilbert space H is a left A-module with module action given by: a.ξ = θ(a)ξ. The column Hilbert operator space Hc can be realized as a space of rank 1 operators on H ([2, p.11]), and it easily follows that Hc is a left operator A-module ([2, Proposition 3.1.7]). Further, there is a dual right module action of A on H∗ = H given by: η.a = π(a∗)η, and r H is a right operator A-module. Given operator A-bimodules W, Z, let CBA,A(Z, W ), which we shall usu- ally abbreviate to CBA(Z, W ), be the operator space (under the structure that it inherits as a subspace of the operator space CB(Z, W )) of completely bounded A-bimodule maps Φ : Z → W (so that a completely bounded map Φ : Z → W belongs to CBA(Z, W ) if and only if aΦ(z)a′ = Φ(aza′) for all z ∈ Z and all a, a′ ∈ A). The next result is a module version of Theorem 1. Proposition 1. Let A be a commutative C ∗-subalgebra of the multiplier algebra M (B) of a C ∗-algebra B (so that B is an operator A-bimodule). Let ρ : A → B(H) be a representation of A on a Hilbert space H (so that B(H) is an operator A-bimodule). Let Φ ∈ CBA(B, B(H)) be completely bounded. Then there exists a representation π of B on a Hilbert space L (determining canonically a representation π of M (B) and hence of A on L) and bounded left A-module maps S : H → L, T : H → L, such that for all b ∈ B, Φ(b) = S∗π(b)T. Further, S and T can be taken to satisfy kΦkcb = kSkkT k. Proof. By Theorem 1, there exists a representation π of B on a Hilbert space L and bounded, linear maps S′, T ′ : H → L such that for all b ∈ B, Φ(b) = (S′)∗π(b)T ′. By extending ρ to the multiplier algebra of A, the Hilbert space H is a left M (A)-module. A general result (unpublished) of Roger Smith for any C ∗-subalgebra A of M (B) gives that S′, T ′ can be taken to be A-module maps S, T . In our special situation a simple averaging procedure gives the result as follows. Let A1 ⊂ M (B) be A if 1 ∈ A and A + C1 otherwise. Let U be the unitary group of A1. Since Φ is a left A-module map, for every u ∈ U , b ∈ B, Φ(ub) = uΦ(b). (In the following, we write u in place of ρ(u), π(u) for ease of notation.) So u(S′)∗u−1π(b)T ′ = (S′)∗π(b)T ′, and 10 ALAN L. T. PATERSON with m an invariant mean on U , Φ(b) = (Z u(S′)∗u−1 dm(u))π(b)T ′ = S∗π(b)T ′ where u1Su−1 1 = S for all u1 ∈ U . Since U spans A1, S is an A-module map. Similarly, using the fact that Φ is a right A-module map, T = If we choose S′, T ′ so that kΦkcb = kS′kkT ′k, then the same holds with S, T in place of S′, T ′ since kSk ≤ kS′k, kT k ≤ kT ′k. R u(T ′)u−1 dm(u) is also an A-module map and Φ(b) = S∗π(b)T . (cid:3) As in Banach space theory, tensor products play an important role in operator space theory. For our purposes, we will only require the Haagerup tensor product. There are other important and interrelated operator space tensor products - see, for example, [7, Part 2], [2, 1.5]. (In particular, in the context of Theorem 2 below, the Haagerup tensor product is closely related to the projective tensor product.) So let X, Y be operator spaces. For v ∈ Mn,r(X), w ∈ Mr,n(Y ), define v ⊙ w ∈ Mn(X ⊗ Y ) in terms of "matrix multiplication": (v ⊙ w)i,j = rXk=1 vi,k ⊗ wk,j. The norm for u ∈ Mn(X ⊗ Y ) is defined: kukn h = inf{kvkkwk : u = v ⊙ w, v ∈ Mn,r(X), w ∈ Mr,n(Y ), r ≥ 1}. (It is not hard to show that for every u above, there always is a v and w such that u = v ⊙ w.) One can show that X ⊗ Y under the norms k.kn h satisfies Ruan's axioms, and so is an operator space, and so therefore is its completion X ⊗h Y . This completion is called the Haagerup tensor product of X and Y . The theory of Haagerup tensor products is described in detail in [7, Chapter 9]. Some of the remarkable properties of this tensor product are ([7, 9.2]) that it is associative, that tensor products of complete contractions remain complete contractions, and that it is both projective and injective in the appropriate senses. Tensor products are a means for linearizing bilinear maps; the bilinear maps that are linearized by the Haagerup tensor product are those that are completely bounded. To define this notion, let X, Y, Z be operator spaces, and let φ : X × Y → Z be a bilinear map. The norm kφk of φ is defined: kφk = sup{kφ(x, y)k : x ∈ X, y ∈ Y, kxk ≤ 1, kyk ≤ 1}. For each n, define a bilinear map φn : Mn(X) × Mn(Y ) → Mn(Z), also in terms of "matrix multiplication", by setting (2.3) φn(u, v)i,j = nXk=1 φ(ui,k, vk,j). Let kφkcb = supn≥1 kφnk. (The norms on Mn(X), Mn(Y ) are given, of course, by the operator space structures on X and Y .) Then φ is called A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 11 completely bounded if kφkcb < ∞. These are the bilinear maps that deter- mine in the natural way the elements of CB(X ⊗h Y, Z), and vice versa, corresponding norms being the same. The bilinear map φ is said to be completely contractive if kφkcb ≤ 1. The next fundamental result, due to Effros and Ruan ([7, Proposition 9.3.3]), is proved using the equality of the Haagerup and projective tensor products on certain tensor products one of whose factors is a column/row Hilbert space (e.g. [8], [4], [1], [7, Proposition 9.3.1], [2, Proposition 1.5.14]). Theorem 2. Let H, K be Hilbert spaces and Z be an operator space. Then r (H ⊗h Z ⊗h Kc)∗ ∼= CB(Z, B(K, H)) completely isometrically as operator spaces. This identification associates r ⊗h Z ⊗h Kc with a mapping Tθ ∈ a continuous linear functional θ on H CB(Z, B(K, H)) given by: (2.4) hTθ(z)(ξ), ηi = θ(η ⊗ z ⊗ ξ) for η ∈ H r , z ∈ Z, ξ ∈ Kc. We now describe the theory of module Haagerup products developed in [2, 3.4] and [3, Chapter 2]. Let A be an operator algebra, X a right operator A-module and Y a left operator A-module. As noted earlier, the Haagerup tensor product X ⊗h Y linearizes completely bounded bilinear maps on X × Y . The module Haagerup tensor product X ⊗hA Y , then, should linearize those completely bounded bilinear maps φ that respect the module actions of A, i.e. such that φ(xa, y) = φ(x, ay) for all x ∈ X, y ∈ Y and a ∈ A. It can be defined in a natural, universal way. Concretely, X ⊗hA Y is realized as the quotient (X ⊗h Y )/N where N is the closure of the subspace of X ⊗h Y spanned by tensors of the form xa ⊗ y − x ⊗ ay. The module Haagerup tensor product is itself an operator space since quotients of operator spaces are operator spaces. If now Y is an operator A-bimodule, X a right operator A-module and Z a left operator A-module,then X ⊗hA Y is a right operator A-module under the natural action induced by: (x ⊗ y)a = x ⊗ (ya), and similarly, Y ⊗hA Z is a left operator A-module. Next the module Haagerup tensor product is - see [2, Theorem 3.4.10]) for a more general result - associative in the following sense: for such X, Y, Z, canonically (X ⊗hA Y ) ⊗hA Z ∼= X ⊗hA (Y ⊗hA Z) completely isometrically. We denote this common value by X ⊗hA Y ⊗hA Z. The following is a special case of [2, Lemma 3.4.6]. Proposition 2. Let A be a C ∗-algebra and X be a left operator A-module that is essential (in the sense that AX = X). Then A ⊗hA X is completely isometrically isomorphic to X as left operator A-modules under the natural map a ⊗ x → ax. 12 ALAN L. T. PATERSON The following module version of Theorem 2 is a special case of [2, Corol- lary 3.5.11]. Proposition 3. Let A be a C ∗-algebra with ∗-representations on the Hilbert spaces H, K so that, in particular, Kc is a left operator A-module and (H)r is a right operator A-module, and B(K, H) is an operator A-bimodule. Let Z be an operator A-bimodule. Then (2.5) r (H ⊗hA Z ⊗hA Kc)∗ ∼= CBA(Z, B(K, H)) completely isometrically, with the same identification θ → Tθ as in Theo- rem 2. 3. The groupoid Fourier algebras Bµ(G) We first survey very briefly the basic theory of locally compact (Hausdorff) groupoids. (For more information, see, for example, [26, 17, 13].) This can be spelled out as follows. Let G be a locally compact groupoid with left Haar system {λx}x∈X where X = G0 = {g ∈ G : g2 = g} is the unit space of G. So algebraically, G is a small category with inverses. The range and source maps from G to X are denoted by r, s: r(g) = gg−1, s(g) = g−1g. Note that the product g1g2 of elements g1, g2 of G is defined if and only range and source match up, i.e. r(g2) = s(g1). Let Gx = {g ∈ G : r(g) = x}. Multiplication and inversion in G are continuous, and the range and source maps are both continuous and open. Turning to the left Haar system {λx}, each λx is a positive, regular Borel measure whose support is Gx, {λx} is invariant in the sense that for g ∈ G, gλs(g) = λr(g). Further, we require that for every F ∈ Cc(G) - the space of continuous, complex-valued functions on G with compact support - the function x → RGx F (g) dλx(g) is continuous. One writes λx = (λx)−1 defined on Gx = {g ∈ G : s(g) = x}. The space Cc(G) is a convolutive ∗-algebra with operations given by: F ∗ F ′(g) =ZGx x∈XZGx kF kI = max{sup A norm k.kI is defined on Cc(G) by: F (h)F ′(h−1g) dλx(h), F (g) dλx(g), sup x∈XZGx F ∗(g) = F (g−1). F (g) dλx(g)} < ∞. Then Cc(G) is a normed ∗-algebra under the I-norm. We next survey some of the basic theory of representations of G on mea- surable Hilbert bundles. For a probability measure µ on the unit space X of G, we integrate up to get a regular Borel measure ν = λµ: ν =ZX λx dµ(x) interpreted in the natural way. Let ν−1 be the regular Borel measure on G given by: ν−1(A) = ν(A−1). Then µ is called quasi-invariant if ν is equivalent to ν−1. The Radon-Nikodym derivative dν/dν−1 on G is called A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 13 the modular function for µ, and is denoted by ∆, and the symmetrized version ν0 of ν is given by: dν0 = ∆−1/2dν. (In the locally compact group case, we don't notice the quasi-invariant measure µ involved in the group's representation theory since there is only one unit in that case - the identity element e - and µ has to be the point mass at e and can be left implicit.) For the representation theory of the groupoid G, we consider G-Hilbert bundles. A G-Hilbert bundle is a pair (L, H) where H is a Borel Hilbert bundle and for each g ∈ G, we are given a unitary Lg : Hs(g) → Hr(g) such that the map g → Lg is a groupoid homomorphism which is Borel measurable in the sense that for any pair v, w of Borel sections of H, the map g → hLgvs(g), wr(g)i is Borel measurable on G. Note that no quasi- invariant measure is involved in the notion of a G-Hilbert bundle. The trivial G-Hilbert bundle (cf. [17, p.93]) is X × C where each Lg : Cs(g) → Cr(g) is the identity map on C. Let (L, H) be a G-Hilbert bundle and µ a quasi-invariant measure on G. We refer to the triple (L, H, µ) as a representation triple. Let L2(H, µ) be the Hilbert space of square integrable sections of H. The groupoid represen- tation L integrates up to give a ∗-representation, denoted by πL,µ, or simply by πL or even π, of the convolution ∗-algebra Cc(G) on L2(H, µ) where, for ξ, η ∈ L2(H, µ), (3.1) (3.2) This can be conveniently contracted to: hπ(F )ξ, ηi =Z F (g)hLgξs(g), ηr(g)i dν0(g). π(F )ξ(x) =ZGx F (g)Lgξs(g)∆−1/2(g) dλx(g). A deep theorem of Renault ([27, 13]) - see [14, Appendix B] for a complete proof of the theorem (covering even the locally Hausdorff case) - gives that every I-norm continuous representation of Cc(G) on a Hilbert space is the integrated form of some representation triple (L, H, µ). The measure µ is determined as follows. The representation π extends in a natural way to a representation of A = C0(X) on L2(H, µ), and one takes µ to be a probability measure on the spectrum of π(A) ⊂ X that is basic ([5, Part 1, Chapter 7]). (By basic, one means that a subset W of X is µ-null if and only if it is null for every spectral measure νξ,η for π(A) (ξ, η ∈ L2(H, µ)).) Trivially, basic measures are determined uniquely up to equivalence. Also, such a measure determines up to isomorphism the Hilbert bundle H ([5, Part II, Chapter 6, Theorem 2]). We obtain a C ∗-seminorm k.kµ on Cc(G) by defining (3.3) kF kµ = sup kπL(F )k the sup being taken over all representation triples (L, H, µ) (µ fixed). Then C ∗(G, µ) is just the enveloping C ∗-algebra associated with (Cc(G), k.kµ), i.e. the completion of Cc(G)/ ker k.kµ or equivalently, the C ∗-algebra generated 14 ALAN L. T. PATERSON by the image of Cc(G) under the direct sum of all such πL's. Using the sep- arability of Cc(G) under the I-norm, we can realize C ∗(G, µ) as the closure of the image π(Cc(G)) for a representation π coming from some specific rep- resentation triple (L, H, µ). When we take the direct sum of representations of Cc(G) with the µ's allowed to vary as well, the C ∗-algebra that we obtain is the full C ∗-algebra, C ∗(G), and Cc(G) is faithfully embedded in it. This gives the largest C ∗-algebra norm on Cc(G). A similar argument to that above shows that there is a special µ for which this full C ∗-algebra norm is the same as k.kµ. So C ∗(G) is one of these C ∗(G, µ)'s. (There is also a reduced C ∗-algebra C ∗ red(G) but this is not usually of the form C ∗(G, µ) (since it is associated with particular groupoid representations) and we will not have occasion to use it in this paper.) In [28], Renault used a C ∗-algebra C ∗ µ(G) bigger than C ∗(G, µ) but defined similarly, which we can formulate as follows. Let µ be a quasi-invariant measure on X and LI (G) be the space of Borel measurable functions F : F (g) dλx(g) are µ-essentially bounded. Two functions in LI (G) are identified if for each x, they agree λx and λx almost everywhere. The norm k.kI is defined on LI(G) by: G → C such that the maps x → RGxF (g) dλx(g), x → RGx kF kI = max{ess supx∈XZGx F (g) dλx(g), ess supx∈XZGx F (g) dλx(g)} < ∞. Then as for Cc(G), LI(G) is a normed ∗-algebra under convolution, and contains Cc(G). Again with the quasi-invariant measure µ fixed, each rep- resentation triple (L, H, µ) integrates (as above) over LI(G) to give an LI(G)-continuous ∗-representation. Define the C ∗-algebra seminorm k.kµ on LI(G) as in (3.3) and form the enveloping C ∗-algebra of (LI (G), k.kµ). This C ∗-algebra is C ∗ µ(G), and it contains C ∗(G, µ) as a C ∗-subalgebra. We now turn to the Fourier-Stieltjes algebras for G. This subject is investigated in the papers [24, 28]. With the group case in mind, the natural way to define a Fourier-Stieltjes algebra B(G) for a groupoid is first to define the set P (G) of positive definite functions on G, and then define B(G) to be the span of P (G). One would like to be able to require the functions in P (G) to be continuous as they are in the group case, but as shown in [24, p.364], it can happen that there are continuous functions in B(G) that are not in the span of the continuous P (G) functions, and there are also completeness problems with B(G) if we restrict to continuous functions only. Instead, we follow Ramsay and Walter ([24, Definition 3.1]) and require the functions φ in P (G) to be Borel measurable and bounded, with positive definiteness taking the form that for all x ∈ X and f ∈ Cc(G), (3.4) f (g1)f (g2)φ(g−1 2 g1) dλx(g1) dλx(g2) ≥ 0. ZGxZGx Again following [24], we define B(G) to be the span of P (G) (in the vector It follows from [24, Lemma 3.2] space of scalar-valued functions on G). A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 15 (and is easy to check directly) that if (L, H) is a G-Hilbert bundle and α, β are bounded Borel sections of H, then the function φ on G, defined by: φ(g) = hLgαs(g), βr(g)i, belongs to B(G). We write (α, β) for this function φ. The other direction is more subtle. It follows from [24, Theorem 3.5] that if φ ∈ B(G), then there exists a G-Hilbert bundle (L, H) and bounded Borel sections α, β of H such that φ = (α, β) (3.5) λµ-a.e. for every quasi-invariant measure µ on X. (This result that B(G)- functions can be taken to be for the form (α, β) is the natural groupoid version of the well-known group result of Eymard that B(G)-functions are of the form g → hLgα, βi for some unitary representation L of G on a Hilbert space of which α, β are elements, but in the groupoid case, it only applies up to quasi-invariance.) When φ ∈ P (G), we can, also as in the group case, take α = β. From the preceding, it is clear that if we want to regard the B(G) func- tions as of the form (α, β) then we need to identify two such functions that coincide λµ-a.e. for one or more quasi-invariant measures µ. There are two natural possibilities in the present situation. For the first, we fix a quasi- invariant measure µ and we identify two functions in B(G) if they agree λµ- almost everywhere on G. We will call B(G) with this identification Bµ(G) to emphasize its dependence on µ. This is B(G) as it is studied by Renault in [28], and will be further studied in the present paper. On the other hand, following Ramsay and Walter ([24]) we can identify two functions in B(G) if they agree λµ-almost everywhere for every quasi-invariant measure µ on G. This B(G) is intrinsic to the groupoid with no preference given to a specific quasi-invariant measure. As one might expect, there is a close relationship between B(G) and the Bµ(G)'s but we will not explore this connection in the present paper. B(G) (and Bµ(G)) are algebras under pointwise opera- tions on G. Indeed, by definition, B(G) is a vector space, and the product of two functions is given by taking tensor products of G-representations: if φ = (α, β), φ′ = (α′, β′) in B(G), then φφ′ = (α ⊗ α′, β ⊗ β′). It is shown in [24, 28] that B(G) and Bµ(G) are Banach algebras under natural norms. We will consider only Bµ(G) in this paper. To define the norm k.kµ on Bµ(G) used by Renault ([28, Proposition 1.4]), we first have to define the norm used for bounded Borel sections α of a Hilbert bundle H over X. We take kαk to be ess supx∈Xkα(x)k. (Note that x → kα(x)k is a Borel function, and that ess sup is taken with respect to the measure µ.) The norm kφkµ of φ ∈ Bµ(G) is the natural analogue of the Eymard norm on B(G) in the group case, and is defined by: kφkµ = inf kαkkβk, the inf being taken over all possible ways of representing φ = (α, β) over all possible representation triples (L, H, µ). A result of Renault ([28, Propo- sition 1.4]) - see also [19, Proposition 5] - gives that Bµ(G) is a Banach 16 ALAN L. T. PATERSON ∗-algebra under pointwise operations, its involution φ → φ∗ being given by: (We note that (α, β)∗ = (β, α).) The non-trivial part φ∗(g) = φ(g−1). of Renault's argument for this result lies in showing that Bµ(G) is a Ba- nach space. His method uses a variant of Paulsen's "off-diagonalization" technique in the "positive definite" setting for groupoids. (That Bµ(G) is a Banach space will also be an immediate consequence of Theorem 3 be- low.) The next result relates Hilbert bundles and the different Bµ(G)'s with respect to absolute continuity. Proposition 4. Let K be a Hilbert space and π′ : C ∗(G) → B(K) be a representation. Let π′ be the integrated form of the triple (L, H, µ′). Let µ be a quasi-invariant measure on X with µ′ ≪ µ, and π be the integrated form of the triple (L, H, µ). Then the map R, where Rξ′ = ξ′(dµ′/dµ)1/2, is a C0(X)-module isometric map from L2(X, H, µ′) into L2(X, H, µ). Further, for F ∈ C ∗(G), (3.6) R∗π(F )R = π′(F ). Proof. It is elementary that R is a C0(X)-module isometric map from L2(X, H, µ′) into L2(X, H, µ). To prove (3.6), it is sufficient to prove that for ξ′, η′ ∈ L2(X, H, µ′), (3.7) hπ(F )ξ, ηi = hπ′(F )ξ′, η′i where ξ = Rξ′, η = Rη′. To this end, let ν, ν−1, ∆, ν0 be as earlier for the representation (L, H, µ), and ν′, ν′−1, ∆′, ν′ 0 be the corresponding measures and function for the representation (L, H, µ′). For the sake of brevity, let p = dµ′/dµ. Then ν′ = R λx dµ′(x) = R λxp(x) dµ(x) so that dν′(g) = p(r(g))dν(g). Similarly, d((ν′)−1)(g) = p(s(g))d(ν−1)(g), and so ∆′(g) = dν′/d((ν′)−1)(g) = dν′/d(ν)(g) × dν/d(ν−1)(g) × (d((ν′)−1)/d(ν−1)(g))−1 = p(r(g)p(s(g))−1∆(g). So dν′ 0/dν0(g) = ∆′(g)−1/2∆(g)1/2(dν′/dν)(g) = [p(r(g))p(s(g)−1∆(g)]−1/2∆(g)1/2p(r(g)) = p(r(g))1/2p(s(g))1/2). A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 17 So for F ∈ Cc(G), s(g), η′ hπ′(F )ξ′, η′i =Z F (g)hLgξ′ Z F (g)hLgξs(g), ηr(g)ip(s(g))−1/2p(r(g))−1/2p(r(g))1/2p(s(g))1/2 dν0(g) =Z F (g)hLgξs(g), ηr(g)i dν0(g) = hπ(F )ξ, ηi. r(g)i dν′ 0(g) = (cid:3) In the main theorem of the paper below we will be faced with the follow- ing situation. We have two representations π, π′ of C ∗(G) with π′ factor- ing through π. Let µ, µ′ be the quasi-invariant measures on X associated respectively with π, π′. It is tempting to conjecture that µ′ is absolutely continuous with respect to µ. If that were the case, we could then apply the preceding proposition to replace µ′ by µ and so simplify the proof of the theorem. However, the conjecture is false, even in very simple cases. For example, suppose that G is the unit space groupoid [0, 1], π the (faithful) multiplication representation of C ∗(G) = C([0, 1]) on L2([0, 1], µ), where µ is Lebesgue measure restricted to [0, 1], and π′ the point evaluation at 0. Then µ′ is the point mass at 0, and this is mutually singular with respect to µ. The following two propositions will enable us to get round this difficulty, and in the proof of the theorem, be able to replace µ′ by a measure that is absolutely continuous with respect to µ. 0, µ′ 0 ⊥ ν and that ν′ 1 ≪ ν so that ν′ 0 + ν′ i =R λx dµ′ i(x) for i = 0, 1. Then ν′ = ν′ Proof. Let ν′ =R λx dµ′(x), ν′ 1 be the Lebesgue decomposition of µ′ with respect to µ (so that µ′ 1 ≪ µ). Then µ′ Proposition 5. Let µ, µ′ be quasi-invariant measures on X and let µ′ = µ′ 0 + µ′ 0 ⊥ µ and µ′ 1 are multiples of quasi-invariant measures on X. 0 + ν′ 1. It is obvious that ν′ 1 is the Lebesgue decomposition of ν′ with respect to ν. We have to show that ν′ i)−1 for i i = 0, 1. To this end, since µ, µ′ are quasi-invariant, we can write (ν′)−1 = f ′ν′, ν′ = (f ′)−1(ν′)−1 and ν−1 = f ν, ν = f −1ν−1 for appropriate Borel func- tions f ′, f , where for all x, 0 < f ′(x), f (x) < ∞. Then ν′ = (f ′)−1(ν′)−1 = 1)−1 ≪ (f ′)−1(ν′ 0)−1 ⊥ ν and (f ′)−1ν−1. Since (f ′)−1ν−1 ∼ (f ′)−1f ν, we obtain (f ′)−1(ν′ 1)−1 is the Lebesgue de- (f ′)−1(ν′ composition of ν′ with respect to ν. Since ν′ = ν′ 1 is also the Lebesgue decomposition of ν′ with respect to ν, and the decomposition is unique, we get (f ′)−1(ν′ 1 are multiples of quasi-invariant measures. (cid:3) 1)−1 ≪ ν. So ν′ = (f ′)−1(ν′ 0)−1 ⊥ (f ′)−1ν−1, (f ′)−1(ν′ 1)−1 = ν′ 1. Hence µ′ 0)−1 = ν′ 0 and (f ′)−1(ν′ 0)−1 + (f ′)−1(ν′ 1)−1, and (f ′)−1(ν′ ∼= (ν′ 0)−1 + (f ′)−1(ν′ 0 + ν′ 0, µ′ Proposition 6. Let µ be a quasi-invariant measure on X and (L, H, µ) be a representation triple for G with integrated form π : C ∗(G) → B(L2(H, µ)). Let µ = c0µ0 + c1µ1 where µ0, µ1 are mutually singular, quasi-invariant measures on X, and c0, c1 are non-negative real numbers. For i = 0, 1, let 18 ALAN L. T. PATERSON Hi be the Hilbert bundle H with inner product h, ii,x on each Hx the given scalar product scaled by ci. (0Hx is taken to be the zero Hilbert space.) For i = 0, 1 let πi be the integrated form of the triple (L, H, µi). Then L2(H, µ) = L2(H0, µ0) ⊕ L2(H1, µ1), (3.8) the L2(H, µi)'s are invariant subspaces of π(C ∗(G)), for each i, πi is the subrepresentation of π given by restriction to L2(H, µi) and π = π0 ⊕ π1. Proof. Since µ0 ⊥ µ1, there exist disjoint Borel subsets A, B of X such that X = A ∪ B and µ0(B) = 0 = µ1(A). It is easy to check (3.8). Explicitly, for ξ ∈ L2(H, µ) let ξA, ξB be respectively the restrictions of ξ to A, B. The Hilbert space direct sum decomposition of (3.8) is given by the map ξ → ξA + ξB. The decomposition π = π0 ⊕ π1 is trivial if either c0 = 0 or c1 = 0, so that we can suppose that c0 > 0, c1 > 0. Let ν =R λx dµ(x) and for i = 0, 1, νi =R λx dµi(x). (So ν0 in this proof does not have its usual meaning.) Then ν = c0ν0 + c1ν1. Further, G is the disjoint union r−1A ∪ r−1B, and ν0(r−1B) = 0 = ν1(r−1A). Further, (ν0)−1(s−1B) = ν0(r−1B) = 0, and since ν0 ∼ ν−1 0 , we also have ν0(s−1B) = 0. It follows that both of the sets s−1A∩ r−1B, r−1A∩ s−1B are ν-null. Also ν0 "lives on" Ar,s = r−1A ∩ s−1A, and similarly, ν1 on Br,s = r−1B ∩ s−1B. Note that (Ar,s)−1 = Ar,s, (Br,s)−1 = Br,s. Let ∆0, ∆1 be the modular functions for µ0, µ1. Then for F ∈ Cc(G), c0Z F dν0 =ZAr,s F dν =ZAr,s F ∆ dν−1 = c0ZAr,s F ∆ d(ν0)−1 so that ∆0 = χAr,s∆, and similarly, ∆1 = χBr,s∆. Then using the above, hπ(F )ξ, ηi =Z F (g)hLgξs(g), ηr(g)i∆−1/2(g) dν(g) =Z F (g)[hLg(ξA)s(g), (ηA)r(g)i + hLg(ξA)s(g), (ηB)r(g)i +hLg(ξB)s(g), (ηA)r(g)i+hLg(ξB)s(g), (ηB)r(g)i][(∆0)−1/2(g)+(∆1)−1/2(g)] dν(g) =Z F (g)hLg(ξA)s(g), (ηA)r(g)i0∆−1/2 0 (g) dν0(g)+ Z F (g)hLg(ξB)s(g), (ηB)r(g)i1∆−1/2 1 The proposition now follows. (g) dν1(g). (cid:3) Let A = C0(X). Let Aµ be the image of A as multiplication operators in B(L2(X, µ)). The operator norm of f ∈ Aµ is the L∞-norm: kf k = kf k∞,µ. Now let (L, H, µ) be a representation triple of G and πL the integrated form of L on L2(H, µ). So πL is a homomorphism from C ∗(G) into B(L2(H, µ)). We will also write πL : A → B(L2(H, µ)) for the diagonal representation of A on L2(X, µ): πL(f )ξ(x) = f (x)ξ(x) for x ∈ X and ξ ∈ L2(H, µ). We note A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 19 that this π can be regarded as a homomorphism with domain Aµ rather than A - it will not usually be an isomorphism on Aµ since some of the Hilbert spaces Hx can be 0. However, in the case of the trivial G-Hilbert bundle (L, X × C), the representation πL associated with the representation triple (L, X × C, µ) is an isomorphism on Aµ. It follows that the image of A in C ∗(G, µ) is also isomorphic to Aµ. From (3.2), we see (cf. [26, p.59]) that for f, f ′ ∈ A and F ∈ Cc(G), π((f ◦ r)F (f ′ ◦ s)) = π(f )π(F )π(f ′). We can regard (an image) of Aµ as contained in the multiplier algebra M (B) of B = C ∗(G, µ), and it follows, almost by definition (§2), that C ∗(G, µ) is an operator A-bimodule. Also H = L2(X, H, µ) is a Hilbert A-module - and so, in particular, a left A-module - so that from the discussion preceding Proposition 1, Hc is a left operator A-module and H a right operator A- module. So the module Haagerup tensor product r Xµ = L2(X, µ) r ⊗hA C ∗(G, µ) ⊗hA L2(X, µ)c in Theorem 3 below makes sense. The following theorem is inspired by analogous results in [28]. Theorem 3. Let µ be a quasi-invariant measure on X and let B = C ∗(G, µ). Then Bµ(G) is a Banach ∗-algebra, the following three Banach spaces are canonically isometrically isomorphic, and the operator spaces in (1) and (2) are completely isometrically isomorphic. (1) X∗ µ; (2) the space CBA(B, B(L2(X, µ))) of completely bounded, A-bimodule maps from B into the operator algebra of bounded linear maps on L2(X, µ); (3) Bµ(G). Proof. For the first claim, we know that Bµ(G) is a commutative normed ∗- algebra. That Bµ(G) is a Banach space (which is originally due to Renault) follows from the rest of the theorem, since X ∗ µ (or CBA(B, B(L2(X, µ))) are Banach spaces. We now prove the rest of the theorem. By Proposition 3 (with H = K = L2(X, µ)), the operator space X∗ µ is com- pletely isometrically isomorphic to the operator space CBA(B, B(L2(X, µ))) under the map θ → Tθ given in (2.4). So (1) is completely isometrically iso- morphic to (2). We now show that the Banach spaces of (2) and (3) are isometrically isomorphic. Let Φ ∈ CBA(B, B(L2(X, µ))). By Proposition 1 (with H = L2(X, µ)) there exists a Hilbert space L with a representation π′ of B on L - and also of A ⊂ M (B) - and A-module maps S, T : L2(X, µ) → L such that for all F ∈ B, (3.9) Φ(b) = S∗π′(F )T. Furthermore, we can assume that (3.10) kSkkT k = kΦk. Since π′ is non-degenerate, B is separable and the ranges of S, T are separable (since L2(X, µ) is). So we can take L to be separable. Then π′ is 20 ALAN L. T. PATERSON 1, µ′ 1)). the integrated form of some representation triple (L′, H′, µ′) (and so we can identify L with L2(H′, µ′)). We now want to replace µ′ by a quasi-invariant measure which is absolutely continuous with respect to µ. To this end, let µ′ = µ0 + µ1 be the Lebesgue decomposition of µ′ with respect to µ. So µ0 ⊥ µ and µ1 ≪ µ. By Proposition 5, µ0 = c0µ′ 1 where the ci's are non-negative real numbers and the µ′ i's are quasi-invariant measures. We will suppose that both ci's are positive, the easier cases where one or other is 0 being left to the reader. Then µ′ 1 ≪ µ. By Proposition 6, 0) ⊕ L2(H′ we have the direct sum L2(H′, µ′) = L2(H′ 1). We now 1) (and also, of course, T (L2(X, µ)) ⊂ claim that S(L2(X, µ)) ⊂ L2(H′ L2(H′ 0 ⊥ µ and µ′ 0, µ′ 0, µ1 = c1µ′ 1, µ′ 1, µ′ X = A ∪ B and µ′ 1 ≪ µ, we also have µ′ To prove this, let A, B be disjoint measurable subsets of X such that 1(A) = 0. 0(B) = 0 = µ(A). Since µ′ F ∈ L2(X, µ). This in turn is equivalent to showing that for any compact V be an open subset containing C and let f ∈ Cc(X) with 0 ≤ f ≤ 1, f = 0 outside V and f = 1 on C. Then kχCSF k ≤ kf SF k = kS(f F )k ≤ It is sufficient then to show that (cid:13)(cid:13)(SF )A)(cid:13)(cid:13)2 = 0 (in L2(H′, µ′)) for all subset C of A, kχCSF k2 = RC kSF (x)k2 dµ′(x) = 0. To prove that, let kSk(R f (x)2 kF (x)k2 dµ(x))1/2 ≤ kSk(RV kF (x)k2 dµ(x))1/2. Since µ(C) = RVn 0, there exists a decreasing sequence of open sets Vn in X containing C such that µ(Vn) < 1/n. The dominated convergence theorem then gives that kF (x)k2 dµ(x) → 0, and it follows that kχCSF k2 = 0 as required. By Proposition 6, we can take µ′ to be µ′ 1. For the purpose of applying 1 and L = L′. Using Proposition 6, (3.6) and Proposition 4, take H = H′ (3.9), we can replace (L′, H′, µ′, π′) by the quadruple (L, H, µ, π) and S, T by R ◦ S, R ◦ T . In particular, we now have L = L2(H, µ). Since T commutes with the C0(X) actions, it is decomposable and so is given by a measurable family {Tx : x ∈ X} of essentially bounded linear op- erators with Tx : C → Lx. (Here, of course, L2(X, µ) trivially disintegrates as R Cx dµ(x), and Lx = Hx.) Similarly, S is also given by a measurable family {Sx : x ∈ X} where each Sx : C → Lx. Since µ is a finite measure, 1 ∈ L∞(X, µ) ⊂ L2(X, µ). Then β = S(1) and α = T (1) belong to L2(H, µ). We now claim that α, β ∈ L∞(H, µ) and that for a, b ∈ L2(X, µ), we have S(b) = bβ and T (a) = aα. (Strictly, we should write π(b)β, π(a)α in place of bβ, aα.) To prove this, since S is a C0(X)-module map, we have S(b) = S(b.1) = bS(1) = bβ for b ∈ Cc(X). Suppose that β does not belong to L∞(H, µ). Then for each positive integer n, An = {x ∈ X : kβ(x)k > n} has positive µ-measure. By regularity, there exists a compact subset Cn of An with µ(Cn) > 0 and an open subset Vn of X containing Cn such that µ(Vn) < (3/2)µ(Cn). Let an ∈ Cc(X) be such that 0 ≤ an ≤ 1, an = 1 on Cn and an A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 21 vanishes outside Vn. Then for all n, kSk((3/2)µ(Cn))1/2 ≥ kSk(Z χVn(x) dµ(x))1/2 ≥ kSkkankL2(X,µ) ≥ kS(an)k ≥ (ZCn an(x)2kβ(x)k2 dµ(x))1/2 ≥ nµ(Cn)1/2. This is impossible. So β, α ∈ L∞(H, µ). Since S is continuous on, and Cc(X) is dense in, L2(X, µ), we obtain that for all a, b ∈ L2(X, µ), T a = aα and Sb = bβ (pointwise multiplication). Note that kaαk2 ≤ kak2kαk, kbβk2 ≤ kbk2kβk, and that (3.11) kT k = kαk, kSk = kβk. Let φ ∈ Bµ(G) be given by: φ = (α, β). Then for F ∈ Cc(G) and a, b ∈ L2(X, µ) and noting that aα, bβ ∈ L2(H, µ), we have, using (3.9) and (3.1), hΦ(F )a, bi = hπ(F )aα, bβi =Z F (g)hLgαs(g), βr(g)ia(s(g))b(r(g)) dν0(g). So (3.12) hΦ(F )a, bi =Z F (g)φ(g)a(s(g))b(r(g)) dν0(g). The function φ is determined λµ a.e. (equivalently ν0-a.e.) since the set of functions on G of the form F (a ◦ s)(b ◦ r) (a, b ∈ A) equals Cc(G). Since the elements of Bµ(G) are identified λµ-a.e., Φ determines the element φ ∈ Bµ(G). The map Φ → φ is obviously linear, and is one-to-one into Bµ(G) since, from (3.12), φ determines Φ. To show that the map Φ is also onto Bµ(G), let φ1 = (α1, β1) ∈ Bµ(G) where α1, β1 are L∞-sections for some representation triple (L1, H1, µ). Let π1 be the integrated form of L1. Then, from (3.12) and the argument leading up to it, we can define the Φ1 : B → B(L2(X, µ)), corresponding to φ1, and independently of the choice of α1, β1, by: To see that Φ1(F ) ∈ B(L2(X, µ)), hΦ1(F )a, bi = hπ1(F )aα1, bβ1i. hΦ(F )a, bi ≤ kF kC ∗(G,µ)kaα1k2kbβ1k2 ≤ [kF kC ∗(G,µ)kα1kkβ1k]kak2kbk2, and so (3.13) kΦk ≤ kα1kkβ1k. We claim that Φ1 ∈ CBA(B, B(L2(X, µ)). To see this, by Theorem 1, Φ1 ∈ CB(B, B(L2(X, µ))) since it is of the form: F → S∗ 1π1(F )T1 where S1(b) = bβ1, T1(a) = aα1. (We note that T1, S1 are bounded, and com- in the S1-case, for f ∈ A, mute with the A-action since, for example, 22 ALAN L. T. PATERSON f S1(b) = f bβ1 = S1(f b).) Then Φ1 ∈ CBA(B, B(L2(X, µ)) since, for ex- 1π1(f )π1(F )π1(f ′)T1 = ample, for f, f ′ ∈ A, F ∈ Cc(G), Φ1(f F f ′) = S∗ f S∗ 1π1(F )T1f ′ = f Φ1(F )f ′. Last, we have to show that kΦk = kφkµ. To prove this, by (3.10) and the definition of kφkµ, kΦk = kSkkT k = kαkkβk ≥ kφkµ. On the other hand, taking the infimum over all pairs (α1, β1) for which φ = (α1, β1) in (3.13) gives kΦk ≤ kφkµ, so that kΦk = kφkµ as asserted. (cid:3) Notes (1) It follows (as in [28]) from the preceding theorem that Bµ(G) can be made into an operator space by transferring to it the operator space structure of X ∗ µ or CBA(B, B(L2(X, µ))). (We get the same operator space structure on Bµ(G) whichever of these two operator spaces we choose.) Nevertheless, an unsatisfactory feature of the preceding theorem is that while two of the µ, CBA(B, B(L2(X, µ)))) are operator spaces a pri- three spaces involved (X ∗ ori, this is not the case for the third space Bµ(G). In [28, 2.3], Renault sketches an Mn-vector-valued theory of Bµ(G) to obtain a Banach algebra Bµ(G, Mn) which can be used to give an operator space structure for Bµ(G). (2) Earlier in this section, we defined the C ∗-algebra C ∗ µ(G) used by Renault in [28]. Renault gives there a "C ∗ µ(G)" version of the preceding theorem, with the C ∗-algebra A = C0(X) replaced by A′ = L∞(X). The result in [28] corresponding to Theorem 3 asserts: (3.14) (L2(X, µ) r ⊗hA′ C ∗ µ(G) ⊗hA′ L2(X, µ)c)∗ = CBA′(C ∗ µ(G), B(L2(X, µ))) = Bµ(G). From their definitions, C ∗(G, µ) is a C ∗-subalgebra of C ∗ A = C0(X) ֒→ A′ = L∞(X, µ), the restriction map µ(G), and since R : CBA′(C ∗ µ(G), B(L2(X, µ))) → CBA(C ∗(G, µ), B(L2(X, µ))) is a norm-decreasing A-bimodule map. So if these two spaces of completely bounded module maps are "the same", R has to be a C0(X)-module isomor- phism. I do not know if this is true or not. One can show, as follows, that R is an onto map. To see this, factorize Φ ∈ CBA(C ∗(G, µ), B(L2(X, µ))) as in Proposition 1, so that for some representation π of C ∗(G, µ) on a Hilbert space L and A-module maps S, T : L2(X, µ) → L, we have, for all F ∈ Cc(G), Φ(F ) = S∗π(F )T . Arguing as in the proof of Theorem 3, we can take π to be the integrated form of a representation triple (L, H, µ). Integrating up (L, H, µ) over LI (G) yields a representation π′ of C ∗ µ(G) on L and also of L∞(X, µ) as multipliers (extending the representation of C0(X) A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 23 on L). Note that S, T are L∞(X, µ)-module maps since Aµ is strong op- erator dense in L∞(X, µ). So Φ′, given by Φ′(F ) = S∗π′(F )T , belongs to CBA′(C ∗ µ(G), B(L2(X, µ))) and R(Φ′) = Φ. The C ∗-algebra C ∗ µ(G) is, in general, more difficult to determine than C ∗(G, µ) (cf. Example 3 of the next section). A difficulty with C ∗ µ(G) is that the image of Cc(G) is not usually dense in it, and indeed it is not usually separable. (In particular, disintegration theory does not apply to its general representations.) The next section examines some examples of Theorem 3, and relates them to other operator space theoretic results. 4. Examples (1) Let G be a locally compact group. Since the identity element e of G is the only unit in G, µ = δe is the only quasi-invariant measure for G. Then C ∗(G, µ) = C ∗(G), and Theorem 3 reduces to: r (C ⊗h C ∗(G) ⊗h Cc)∗ = CB(C ∗(G), C) = Bµ(G). Examining these three spaces in turn, C in the first of these is treated as a Hilbert space in the obvious way. As an operator space, Cc is realized as B(C, C), so that its operator space structure is the same as the canonical C ∗-algebra operator space structure on C. On the other hand, as in §2, C as an operator space is identified with B(C, C) so that again this is the same as C with its canonical C ∗-algebra operator structure. By Proposition 2, ⊗h C ∗(G) ⊗h Cc = C ∗(G) so that its dual is C ∗(G)∗. For the second of C these, CB(C ∗(G), C) = B(C ∗(G), C) = C ∗(G)∗ again. For the third, Bµ(G) is the Banach algebra B(G) with Eymard's norm, and it is a theorem of Eymard that B(G) = C ∗(G)∗. r r (2) Let G = X. So G is a unit groupoid, and every probability measure µ on X is quasi-invariant. For convenience we take µ to have support X. Each λx is the point mass at x, and since the support of µ is X, the I-norm on Cc(X) is just the sup-norm. It follows that the representations of G are just the continuous representations of C0(X) so that C ∗(G, µ) = C0(X) = A. Theorem 3 then becomes: (L2(X, µ) r ⊗hA A ⊗hA L2(X, µ)c)∗ = CBA(A, L2(X, µ)) = Bµ(G). We show directly that each of these spaces is L∞(X, µ). r For the first of these, by Proposition 2 and the associativity of the module ⊗hA L2(X, µ). Haagerup tensor product, it is just the dual of L2(X, µ) ⊗h L2(X, µ)c)/N where N is the Now L2(X, µ) closure of the subspace spanned by elements of the form ηf ⊗ ξ − η ⊗ f ξ , ξ ∈ L2(X, µ)c. Note that ηf = f η. By [7, Proposition where η ∈ L2(X, µ) 9.3.4], ⊗hA L2(X, µ)c = (L2(X, µ) r r r L2(X, µ) ⊗h L2(X, µ)c = T, r 24 ALAN L. T. PATERSON the operator space of trace class operators on L2(X, µ). (One gives ([7, Theorem 3.2.3]) T the operator space structure that it inherits as the dual of the C ∗-algebra of compact operators on L2(X, µ).) This identification sends simple tensors to finite rank operators: η ⊗ξ → Tη,ξ, where Tη,ξ(ζ) = hζ, ηiξ. ⊗hA L2(X, µ)c)∗ = N ⊥ in T∗ = B(L2(X, µ)). We now identify (L2(X, µ) The duality between T and B(L2(X, µ)) is given by: hS, Tη,ξi = T r(STη,ξ) = T r(Tη,Sξ) = hSξ, ηi. For S to vanish on N , we require that for all ξ, η ∈ L2(X, µ), f ∈ C0(X), r 0 = T r(S(T ηf ,ξ − Tη,f ξ)) = hηf , Sξi − hη, Sf ξi = hη, (f S − Sf )ξi, i.e. that S should commute with C0(X) regarded as a C ∗-algebra of multipli- cation operators on L2(X, µ).. Taking the weak operator closure of C0(X) in B(L2(X, µ)) gives that S commutes with L∞(X, µ), and since L∞(X, µ) is a ⊗hAL2(X, µ)c)∗ = L∞(X, µ) masa in B(L2(X, µ)), we obtain that (L2(X, µ) as an operator space. r For the second of these, by a module version of the Arveson-Wittstock Hahn-Banach theorem ([2, Corollary 3.6.3], any Φ ∈ CBA(A, B(L2(X, µ))) extends to a completely bounded A-bimodule map Φ from the unitization A+ of A into B(L2(X, µ)). Let TΦ = Φ(1). Then Φ(a) = aTΦ = TΦa in B(L2(X, µ). Then Tφ is uniquely determined by Φ, and as in the previous paragraph, TΦ ∈ L∞(X, µ). Define Γ : CBA(A, L2(X, µ)) → L∞(X, µ) by: Γ(Φ) = TΦ. Then Γ is a linear isometry. Next, for any T ∈ L∞(X, µ), the map a → aT belongs to CBA(A, B(L2(X, µ))) using elementary algebra and [7, Proposition 2.2.6], and so Γ is a complete isometry. For the third of these, noting that λµ = µ, the elements of P (G) are the classes of Borel measurable, bounded, positive functions on X, and so the span Bµ(G) of the image of P (G) in L∞(X, µ) is just L∞(X, µ). It is easy to check that the Bµ(G)-norm is the same as the L∞-norm. Then X = P and A = c0. Take µ = P∞ (3) Take the locally compact groupoid G to be the discrete groupoid P × Z, a bundle of groups each isomorphic to Z, over the set P of positive integers. n=1 2−nδn. Then L2(X, µ) ∼= ℓ2. Also, every measure on X is absolutely continuous with respect to µ, and so all of the representation triples for G can be taken to be of the It follows that C ∗(G, µ) = C ∗(G). Next, Cc(G) is just form (L, H, µ). the algebra of complex-valued functions F on G with finite support and is the commutative Banach ∗-algebra c0(ℓ1(Z)) with pointwise convolution multiplication. The characters of this Banach algebra are of the form F → with I-norm given by: kF k = maxn∈PPm∈Z F (n, m), and so its closure χ(F (n)) where χ ∈ bZ = T, the circle group, and we obtain that C ∗(G, µ) = Turning to Bµ(G), it is the set of bounded functions φ on G such that φn = φGn ∈ B(Gn). Then for each n, there exists a Hilbert space Hn and αn, βn ∈ Hn such that φn = (αn, βn) with kφnk close to kαnkkβnk. We can c0(C(T)). A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 25 take α, β as bounded sections of the Hilbert bundle {Hn} over P, and it is simple to check that kφk = supn≥1 kφnk and that Bµ(G) = ℓ∞(B(Z)) = ℓ∞(M (T)). Theorem 3 then states that (4.1) r (ℓ2 ⊗hc0 c0(C(T)) ⊗hc0 (ℓ2)c)∗ = CBc0(c0(C(T)), B(ℓ2)) = Bµ(G). Here, c0(C(T)) = c0 ⊗ C(T) is a c0-module by multiplying by c0 ⊗ 1. We now look more closely at the terms in (4.1). Starting with the middle one, T : c0(C(T)) → B(ℓ2) belongs to CBc0(c0(C(T)), B(ℓ2)) if and only if it is completely bounded and commutes with the multiplication representation of c0 on ℓ2. Then the range of T commutes also with the masa ℓ∞ ⊂ B(ℓ2), and so can be identified with a subspace of ℓ∞. For h ∈ C(T)) and i ≥ 1 let hi ∈ c0(C(T)) be given by: hi(n) = h if n = i and is 0 otherwise. Since T is a c0-module map, T hi ∈ ℓ∞ has all its components 0 except possibly for its ith entry, αi(h). Then αi is a continuous linear functional on C(T) and so is determined by a measure µ(i) ∈ M (T). The map T → {µ(n)} is a linear isometry onto Bµ(G) = ℓ∞(M (T)). This gives the identity of the middle and final terms in (4.1). We now sketch how one identifies Bµ(G) with X∗ µ. We start with φ ∈ ℓ∞(M (T)) = Bµ(G). Then the continuous linear functional Sφ on ⊗hc0 c0(C(T)) ⊗hc0 (ℓ2)c associated with φ in (4.1) is given by: ℓ2 r Sφ(η ⊗ F ⊗ ξ) =Xi≥1 ηnφ(n)(F (n))ξn for ξ, η ∈ ℓ2 and F ∈ c0(C(T)). The map θ → Tθ of (2.4) in the present case is as follows. With respect to the standard basis on ℓ2, Tθ is just the diagonal operator whose nth diagonal entry is θ(en ⊗ F ⊗ en). We now make some comments about C ∗ µ(G). The Banach ∗-algebra LI (G) is easily seen to be ℓ∞(ℓ1(Z)) with pointwise convolution. We write its el- ements as sequences {gn} of ℓ1(Z)-elements such that the map n → gn is bounded. Taking the sup norm over the characters of c0(C(T)) (canonically extended to characters of ℓ∞(C(T))) identifies C ∗ µ(G) with the closure in ℓ∞(C(T)) of its subalgebra Q = {{bgn} : gn ∈ ℓ1(Z), sup kgnk1 < ∞}. Of µ(G) does contain C ∗(G, µ) = c0(C(T)), and also all maps from course, C ∗ P to C(T) that have finite range. However, C ∗ µ(G) is a proper subalgebra of ℓ∞(C(T)). I am grateful to Colin Graham for suggesting the following argument. We use a well-known Banach space result: if X, Y are Banach spaces and T : X → Y is a bounded linear map that is not onto Y , then for all n, there exists fn in the unit ball U of Y with the property that whenever g ∈ X is such that kT (g) − fnk < 1/2, then kgk > n. (It can be proved by supposing the contrary and then showing that for every f ∈ U , there exists g ∈ X, the sum of a geometric series in X, such that T (g) = f , contradicting the fact that T is not onto.) We apply this result in the case where X = ℓ1(Z), Y = C(T), and T = F, the Fourier transform g → g. It is well-known (see, for example, [25, Chapter 5, 4.5], [12, (37.19)]) that F 26 ALAN L. T. PATERSON has dense range but is not onto. Then for each n ≥ 1, there exists fn in the unit ball U of C(T) with the property that whenever g ∈ ℓ1(Z) is such that kg − fnk < 1/2, then kgk1 > n. Let F ∈ ℓ∞(C(T)) be given by: F (n) = fn. Then by construction kF − {bgn}k ≥ 1/2 in ℓ∞(C(T)) for all {bgn} ∈ Q so that F /∈ C ∗ µ(G). (4) This example gives a groupoid interpretation of a remarkable result, described below, concerning Schur products. The result is due to Haagerup ([10]). (See also [30] for an early groupoid approach to Schur products.) The theory of Schur products and complete boundedness is treated in detail in the book of Paulsen ([20, Chapter 8]). For A ∈ Mn, define a linear map SA : Mn → Mn by: SA(B)i,j = Ai,jBi,j. So SA(B) = A ∗ B, the Schur product of the matrices A, B. By finite dimensionality, SA is a bounded linear operator on Mn. The following definitive theorem is [20, Theorem 8.7]. Theorem 4. The following three statements are equivalent: (1) kSAk ≤ 1 (2) kSAkcb ≤ 1 (3) there exist 2n vectors ξj, ηi (1 ≤ j, i ≤ n) in the unit ball of some Hilbert space such that Ai,j = hξj, ηii for all i, j). We need to reformulate this for A's with kSAk arbitrary. A straight- forward scaling argument, left to the reader, gives the following desired reformulation. Theorem 5. The following three numbers are equal: (1) kSAk (2) kSAkcb (3) infξ,η,H maxi,j kξjkkηik where, for some Hilbert space H, ξ = {ξj}1≤j≤n, η = {ηi}1≤i≤n belong to Hn and are such that Ai,j = hξj, ηii. We now interpret the equivalence of (2) and (3) of this theorem in terms of Theorem 3 above for the groupoid G = Gn = {1, 2, . . . , n} × {1, 2, . . . , n} (an equivalence relation groupoid). The product and inverse for Gn are respectively given by: (i, j)(j, k) = (i, k) and (i, j)−1 = (j, i). The unit space X is the diagonal {(i, i) : 1 ≤ i ≤ n}, which we usually identify with {1, 2, . . . , n}. Then r(i, j) = i, s(i, j) = j, Gi = {(i, j) : 1 ≤ j ≤ n}. We take λi to be counting measure on Gi. The example Gn is discussed in [19, Example, §5] in the context of the continuous Fourier-Stieltjes algebra of a groupoid. Since Gn is discrete, Gn-Hilbert bundles in the Borel sense of this paper are the same as continuous Gn-Hilbert bundles in the sense of [19]. In the latter context, a quasi-invariant measure is not directly involved, but i=1 δi in the present context, we need such a measure µ. We take µ = Pn n Pi,j δ(i,j) sum of the point masses over i divided by n. Then λµ = , which is equivalent to counting measure on Gn. Identifying Cc(G) with Mn - in , the n A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 27 which f ∈ Cc(G) is identified with the matrix which has f (i, j) at the (i, j)th place - we see that C ∗(G, µ) = Mn under the usual matrix multiplication. Theorem 3 then assumes the form: ⊗hDn Mn ⊗hDn (ℓ2 n)c)∗ ∼= CBDn(Mn, Mn) ∼= Bµ(Gn) (4.2) r (ℓ2 n n is just Cn with the usual inner product divided by n. Here, A = where ℓ2 C0(X) is identified with Dn, the C∗-subalgebra of diagonal matrices in Mn, the multiplier action of A on Mn being given by matrix multiplication on the right and left. We start by looking at Bµ(G). Let H 6= 0 be a Hilbert space. Then {1, 2, . . . , n} × H is a G-Hilbert bundle with action given by: (i, j)(j, ζ) = (i, ζ). Fix a unit vector ζ ∈ H and also fix i, j. Define sections α, β of the bundle by setting αi = ζ = βj, all other values of α, β being defined to be 0. Then with φ = (α, β), we have φ(l, m) = h(l, m)(m, αm), (l, βl)i = hαm, βli, which equals 1 if (l, m) = (j, i) and is 0 otherwise. So the matrix φ = ej,i. It follows that as a vector space, Bµ(G) = Mn, the space of all functions F : G → C. However, multiplication in Bµ(G) is pointwise, so that Bµ(G) acts on itself (as Mn) by Schur multiplication. To determine the norm on Bµ(Gn), we have to consider an arbitrary G-Hilbert bundle H over X. This is just a collection of Hilbert spaces {Hi}1≤i≤n with unitaries Li,j : Hj → Hi satisfying Li,j ◦ Lj,k = Li,k and L−1 i,j = L∗ i,j = Lj,i. We have to consider sections α, β of this bundle and determine, for a given φ ∈ Bµ(G), what is inf kαkkβk over all possible ways of representing φ = (α, β). Given such a bundle and such sections α, β, let H = H1 and let ξj = L1,jαj, ηi = L1,iβi. Then hξj, ηii = hL1,jαj, L1,iβii = hLi,jαj, βii = φ(i, j), and kαk = maxj kξjk, kβk = maxi kηik. So kφk is exactly the same as the norm of the matrix A = φ given in (3) of Theorem 5. Conversely, given φ = A ∈ Mn and vectors ξj, ηi as in (3) of Theorem 5, we can construct a corresponding G-representation and sections α, β such that φ = (α, β). So (3) defines the norm on Bµ(G). Turning next to the middle expression of (4.2), it is easy to prove ([20, Exercise 4.4]) that CBDn(Mn, Mn) = {SA : A ∈ Mn}, so that the identity of CBDn(Mn, Mn) with Bµ(G) is equivalent to the identity of (2) and (3) of Theorem 5. I do not know of a charaterization of the norm of Xµ, the module Haagerup tensor product whose dual is the first space in (4.2). The equivalence of (1) and (2) in Theorem 5 follows in complete groupoid gener- ality from [28, Proposition 2.3] - the proof for the special case of G = Gn in Theorem 5 appears in the first two paragraphs of [20, Theorem 8.7]. (This is also proved in [30, Proposition 8], where estimates for kSAk are obtained.) Paulsen ([20, Corollary 8.8]) obtains an ℓ2-version of the above Theo- rem 4 (so that B(ℓ2) replaces Mn = B(ℓ2 n) and A is an infinite matrix). In this case, the groupoid G would be P × P where P is the set of positive integers. In a more general context, where G = X × X, Renault ([28, Ex- ample 1.3]) points out that in this case, Bµ(G) is the space of functions φ on G for which there exists a Hilbert space H and α, β ∈ L∞(X, H, µ) such 28 ALAN L. T. PATERSON that φ(x, y) = hα(x), β(y)i (x, y ∈ X), and that these functions φ are the Hilbertian functions as defined by Grothendieck. References [1] D. P. Blecher, Tensor products of operator spaces II., Canad. J. Math. 44(1992), 75-90. [2] D. P. Blecher and C. Le Merdy, Operator algebras and their modules an operator space approach, London Mathematical Society Monographs. New Series, 30. Oxford Science Publications. The Clarendon Press, Oxford University Press, Oxford, 2004. [3] D. P. Blecher, P. S. Muhly and V. I. Paulsen, Categories of Operator Modules (Morita Equivalence and Projective Modules), Memoirs of the American Mathematical Society, Number 681, Providence, R. I., 2000. [4] D. P. Blecher and V. I. Paulsen, Tensor products of operator spaces, J. Funct. Anal., 99(1991), 262-292. [5] J. Dixmier, Von Neumann Algebras, North-Holland Publishing Company, Amster- dam, 1981. [6] J. Dixmier, C ∗-algebras, North-Holland Publishing Company, Amsterdam, 1977. [7] E. G. Efros and Z-J. Ruan, Operator Spaces, London Mathematical Society Mono- graphs. New Series, 23. Oxford Science Publications. The Clarendon Press, Oxford University Press, Oxford, 2000. [8] E. G. Efros and Z-J. Ruan, Self-duality for the Haagerup tensor product and Hilbert space factorizations., J. Funct. Anal. 100(1991), 257-284. [9] P. Eymard, L'alg`ebre de Fourier d'un groupe localement compact, Bull. Soc. Math. France 92(1964), 181-236. [10] U. Haagerup, Decomposition of completely bounded maps on operator algebras, Un- published notes, (1980). [11] P. Hahn, Haar measure for measure groupoids, Trans. Amer. Math. Soc. 242(1978), 1-33. [12] E. Hewitt and K. A. Ross, Abstract Harmonic Analysis, Volume II, Springer-Verlag, Berlin-Heidelberg-New York, 1970. [13] P. S. Muhly, Coordinates in operator algebra, CBMS Conference Lecture Notes (Texas Christian University 1990), 1999. In continuous preparation. [14] P. S. Muhly and D. P. Williams, Renault's Equivalence Theorem for groupoid crossed products, NYJM Monographs, 3. State University of New York, University at Albany, Albany, NY, 2008. 87 pp. [15] K. J. Oty, Fourier-Stieltjes algebras of r-discrete groupoids, J. Operator Theory 41(1999), 175-197. [16] K. J. Oty and A. Ramsay, Actions and coactions of measured groupoids on W ∗- algebras, J. Operator Theory 56(2006), 199-217. [17] A. L. T. Paterson, Groupoids, inverse semigroups and their operator algebras, Progress in Mathematics, Vol. 170, Birkhauser, Boston, 1999. [18] A. L. T. Paterson, The Fourier-Stieltjes and Fourier algebras for locally compact groupoids, Contemporary Mathematics 321(2003), 223-237. [19] A. L. T. Paterson, The Fourier algebra for locally compact groupoids, Canadian J. Math. 56(2004), 1259-1289. [20] V. I. Paulsen, Completely Bounded Maps and Operator Algebras, Cambridge Studies in Advanced Mathematics 78. Cambridge University Press, Cambridge, 2002. [21] G. Pisier, Introduction to Operator Space Theory, London Mathematical Society Lec- ture Note Series. 294. Cambridge University Press, Cambridge, 2003. [22] A. Ramsay, Nontransitive quasi-orbits in Mackey's analysis of group extensions, Acta Math. 137(1976), 17-48. [23] A. Ramsay, Topologies for measured groupoids, J. Funct. Anal., 47(1982), 314-343. A DUALITY THEOREM FOR FOURIER-STIELTJES ALGEBRAS 29 [24] A. Ramsay and M. E. Walter, Fourier-Stieltjes algebras of locally compact groupoids, J. Funct. Anal., 148(1997), 314-367. [25] H. J. Reiter, Classical Harmonic Analysis and Locally Compact Groups, Oxford Math- ematical Monographs. Oxford: Oxford University Press, 1968. [26] J. N. Renault, A groupoid approach to C ∗-algebras, Lecture Notes in Mathematics, Vol. 793, Springer-Verlag, New York, 1980. [27] J. N. Renault, Repr´eesentations des produits crois´es d'alg`ebres de groupoıdes, J. Op- erator Theory, 18(1987), 67-97. [28] J. N. Renault, The Fourier algebra of a measured groupoid and its multipliers, J. Funct. Anal., 145(1997), 455-490. [29] M. Walter, W ∗-algebras and nonabelian harmonic analysis, J. Funct. Anal. 11(1972), 17-38. [30] M. Walter, Dual algebras, Math. Scand. 58(1986), 77-104. Department of Mathematics, Campus Box 395, University of Colorado,, Boulder, Colorado 80309-0395, USA E-mail address: [email protected]
1706.06951	1	1706	2017-06-21T15:16:11	C*-Algebra Distance Filters	[ "math.OA" ]	We use non-symmetric distances to give a self-contained account of C*-algebra filters and their corresponding compact projections, simultaneously simplifying and extending their general theory.	math.OA	math	C∗-ALGEBRA DISTANCE FILTERS TRISTAN BICE AND ALESSANDRO VIGNATI Abstract. We use non-symmetric distances to give a self-contained account of C-algebra filters and their corresponding compact projections, simultaneously simplifying and extending their general theory. Introduction Quantum filters were introduced by Farah and Weaver to analyze pure states on C∗-algebras and various conjectures concerning them, like Anderson's conjecture and the Kadison-Singer conjecture (which has since become the Marcus-Speilman- Srivastava theorem -- see [MSS15]). They were also considered more recently in [BW16] in relation to quantum analogs of certain large cardinals, and they even make an appearance much earlier in [AP92] as faces of the positive unit ball. While their basic theory was fleshed out in [Bic13a] (as 'norm filters') and [FW13] (and in a forthcoming book by Farah), there remained some fundamental questions which we aim to address in this paper. The first such question is why they should be considered as filters at all. Filters in the classical sense are defined from a transitive relation (as the downwards directed upwards closed subsets), but in general there is no such relation defining quantum Indeed, it can even happen that every maximal quantum filter in a C∗- filters. algebra fails to be a filter in the traditional order theoretic sense -- see [Bic13a, Corollary 6.6]. While it might be intuitively clear that quantum filters are the 'right' quantum analog, and their utility in analyzing states justifies their study, regardless of whether they are considered as filters or not, a more precise connection to order theory would of course be desirable. The key here is to replace classical transitive relations with 'continuous' ones. These are the non-symmetric distances, binary functions D to [0,∞] satisfying the continuous version of transitivity, namely the triangle inequality D(x, y) ≤ D(x, z) + D(z, y). The first order sentences defining classical filters also have continuous versions, with the quantifiers ∀ and ∃ replaced by suprema and infima respectively. Then quantum filters are indeed the continuous filters with respect to the appropriate distance d on the positive unit ball A1 +, namely This simple observation allows for a markedly different approach to the theory. d(a, b) = ka − abk . 2010 Mathematics Subject Classification. 06A75, 46L05, 46L85, 54E99. Key words and phrases. filter, C-algebra, compact projection, non-symmetric distance. The first author has been supported by a WCMCS grant at IMPAN (Poland). The second author is partially supported by a York University Susan Mann Scholarship. 1 2 TRISTAN BICE AND ALESSANDRO VIGNATI In §1 we start off by examining the relationship between various distances and distance-like functions. As with metrics, uniform equivalence plays a fundamental role. We move on to d-filters in §2, using these relationships to provide charac- terizations using the distance, order, multiplicative and convex structure of A1 +. In §3 we then show how d-filters in A represent compact projections in A∗∗ (just as hereditary C∗-subalgebras in A represent open projections in A∗∗). We finish by examining interior containment of compact projections and its relation to the reverse Hausdorff distance on d-filters. 1. Distances We will deal with a number of binary functions D from some set X to [0,∞]. We view these as 'generalized' or 'continuous' relations on X. More precisely, any D : X × X → [0,∞] defines a classical relation D ⊆ X × X by xDy ⇔ D(x, y) = 0.1 We say that the function D quantifies the relation D. Conversely, every relation R ⊆ X × X has a trivial quantification given by its characteristic function, which we also denote by R, specifically R(x, y) =(0 if xRy ∞ otherwise. We define the composition D ◦ E of any D, E : X × X → [0,∞], by (D ◦ E)(x, y) = inf z∈X (D(x, z) + E(z, y)). Note when R and S are relations (identified with their characteristic functions), x(R ◦ S)y ⇔ ∃z ∈ X (xRzSy), so ◦ extends the usual composition of classical relations.2 Moreover, we always have D ◦ E ⊆ D ◦ E. We say that D is E-invariant when D = D ◦ E = E ◦ D. Definition 1.1. D is a distance if (△) D ≤ D ◦ D. On a C∗-algebra A, the only distance usually considered is the metric given by e(x, y) = kx − yk . Indeed, metrics are precisely the symmetric distances quantifying the equality re- lation. However, our thesis is that one should also consider various non-symmetric distances on C∗-algebras which quantify other important order relations like a ≪ b a ≤ b ⇔ ⇔ a = ab. b − a ∈ A+. 1Note we are using the standard infix notation xRy to mean (x, y) ∈ R. 2In other words, the category Rel of classical relations forms a wide subcategory of GRel, the category of generalized relations -- see [Bic17, §1] for more details. C∗-ALGEBRA DISTANCE FILTERS 3 Here A+ denotes the positive elements in A, while Asa, Ar and A=r will denote the self-adjoints, r-ball and r-sphere respectively. We also consider A embedded canonically in its enveloping von Neumann algebra A∗∗ and set eA = A + C1. So if A is unital then eA = A, otherwise eA is the unitization of A (see [Bla17, II.1.2]). Proposition 1.2. (1.1) (1.2) d(a, b) = ka − abk h(a, b) = k(a − b)+k is an e-invariant distance on A1 + quantifying ≪ . is an e-invariant distance on Asa quantifying ≤ . Proof. (1.1) For a, b, c ∈ A1 +, we have kak ,(cid:13)(cid:13)b⊥(cid:13)(cid:13) ≤ 1, where b⊥ = 1 − b(∈ eA), so d(a, b) =(cid:13)(cid:13)ab⊥(cid:13)(cid:13) =(cid:13)(cid:13)a(c⊥ + c)b⊥(cid:13)(cid:13) ≤(cid:13)(cid:13)ac⊥(cid:13)(cid:13)(cid:13)(cid:13)b⊥(cid:13)(cid:13) + kak(cid:13)(cid:13)cb⊥(cid:13)(cid:13) ≤ d(a, c) + d(c, b). As e quantifies equality, we immediately have d ◦ e, e ◦ d ≤ d. Conversely, d(a, b) = ka − abk ≤ ka − ck + kc − cbk = e(a, c) + d(c, b). d(a, b) = ka − abk ≤ ka − ack + kac − abk ≤ d(a, c) + e(c, b). (1.2) Denote the space of quasistates on A by Q = (A∗1 + = positive linear func- (1.3) tionals in the dual unit ball) and recall that, for a ∈ Asa, ka+k = sup φ∈Q φ(a). Thus for all a, b, c ∈ Asa, we have h(a, b) ≤ sup φ(c − b) = h(a, c) + h(c, b).3 Now as ka+k ≤ kak, we have h ≤ e so h ≤ h ◦ h ≤ h◦ e, e◦ h. But the reverse inequalities are again immediate, as e quantifies equality. φ(a − c) + sup φ∈Q φ∈Q (cid:3) Basic relationships between C∗-algebra distances reveal aspects of C∗-algebraic structure. Here are some required for our investigation of C∗-algebra filters. Proposition 1.3. On A1 +, (1.4) (1.5) h ≤ 2d. d2 ≤ d ◦ h. d2 ≤ h ◦ d. (1.6) In fact, in (1.5) and (1.6), we can even take ◦ in Asa. Proof. (1.4) For a, b ∈ A1 +, bab ≤ b2 ≤ b so h(a, b) ≤ h(a, bab) + h(bab, b) = ka − babk ≤ ka − abk + kab − babk ≤ ka − abk + ka − bakkbk ≤ 2d(a, b). 3One might naively use (a + b)+ ≤ a+ + b+ instead, but this only holds for commutative A. d(a, b)2 =(cid:13)(cid:13)ab⊥2a(cid:13)(cid:13) ≤(cid:13)(cid:13)ab⊥a(cid:13)(cid:13) =(cid:13)(cid:13)(ab⊥a)+(cid:13)(cid:13) ≤(cid:13)(cid:13)(ac⊥a)+(cid:13)(cid:13) +(cid:13)(cid:13)(a(b⊥ − c⊥)a)+(cid:13)(cid:13) ≤(cid:13)(cid:13)ac⊥(cid:13)(cid:13)kak + k(c − b)+k ≤ d(a, c) + h(c, b). + and c ∈ Asa, ≤(cid:13)(cid:13)b⊥ab⊥(cid:13)(cid:13) =(cid:13)(cid:13)(b⊥ab⊥)+(cid:13)(cid:13) ≤(cid:13)(cid:13)(b⊥(a − c)b⊥)+(cid:13)(cid:13) +(cid:13)(cid:13)(b⊥cb⊥)+(cid:13)(cid:13) ≤ k(a − c)+k +(cid:13)(cid:13)b⊥(cid:13)(cid:13)(cid:13)(cid:13)cb⊥(cid:13)(cid:13) ≤ h(a, c) + d(c, b). (1.6) Likewise, for a, b ∈ A1 d(a, b)2 =(cid:13)(cid:13)b⊥a2b⊥(cid:13)(cid:13) 4 TRISTAN BICE AND ALESSANDRO VIGNATI (1.5) First note that, for any a ∈ A1 and b ∈ Asa, (1.7) k(aba)+k = inf aba≤ckck ≤ inf b≤c kacak ≤ kab+ak ≤ kb+k . As k(a + b)+k ≤ ka+k + kb+k (see (1.3)), for all a, b ∈ A1 + and c ∈ Asa, (cid:3) Another important quantification of ≤ on A+ is given by p(a, b) = inf 0≤c≤bka − ck . Often p(a, b) = h(a, b), e.g. if ab = ba, a ≤ b, b ≤ a or if a and b are projections. However, p and h do not coincide in general. Proposition 1.4. On M2+, p 6= h. In fact, p is not even a distance. 1 1(cid:21) and b =(cid:20)4 Proof. Let a =(cid:20)1 1 So (a − tb) has eigenvalues 1 − 2t ± √4t2 + 1) and hence 0(cid:21) so a − tb =(cid:20)1 − 4t 1 det(a − tb − λ) = (1 − 4t − λ)(1 − λ) − 1 = λ2 + (4t − 2)λ − 4t. 1(cid:21) and 0 0 1 ka − tbk =(1 − 2t + √4t2 + 1 1 − 2t − √4t2 + 1 2 b(cid:13)(cid:13) = √2. On the other hand, h(a, b) is Thus p(a, b) = inf t∈[0,1] ka − tbk =(cid:13)(cid:13)a − 1 the positive eigenvalue for t = 1, i.e., √5 − 1 so Let c = a + (b − a)+ so p(a, c) = 0 and p(c, b) = kc − bk = k(a − b)+k = √5 − 1, as a, b ≤ c. Thus for t ≤ 1 for t ≥ 1 h(a, b) < p(a, b). 2 2 . p(a, b) > p(a, c) + p(c, b). (cid:3) Even when two metrics differ, it often suffices to show they are uniformly equiv- alent. If F and G are functions F, G : X → [0,∞] we define F(x). F w G sup ⇔ 0 = lim r→0 G(x)≤r C∗-ALGEBRA DISTANCE FILTERS 5 Equivalently, F w G if and only if, for all Y ⊆ X, G(y) = 0 ⇒ inf inf y∈Y y∈Y F(y) = 0. We call F and G uniformly equivalent, written F ≈ G, when F w G w F. Theorem 1.5. On A1 Proof. First we show that h ≤ p on A1 +, h ≤ p w h, so h is uniformly equivalent to p. +. For any c ≤ b, (1.3) yields + × A1 h(a, b) = sup φ∈Q φ(a − b) ≤ sup φ∈Q φ(a − c) ≤ ka − ck . Also a − b ≤ (a − b)+ so a − (a − b)+ ≤ b and h(a, b) = ka − (a − (a − b)+)k so h(a, b) = inf c≤bka − ck ≤ inf 0≤c≤bka − ck = p(a, b). Next we need to show that Take a, b ∈ A1 ∀ǫ > 0 ∃δ > 0 ∀a, b ∈ A1 + and let z = b + (a − b)+ so a, b ≤ z ∈ A+ and + h(a, b) < δ ⇒ p(a, b) < ǫ. 2√b → u, for some u ∈ A, by [Ped79, Lemma 1.4.4]. As √a( 1 n + z)− 1 √a( 1 2√z → √a, we have √a( 1 n + z)− 1 n + z)− 1 2 (√b − √z) → u − √a. 2(cid:13)(cid:13)(cid:13) ≤ 1 so 2(cid:13)(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13)√a( 1 n + a)− 1 n + z)− 1 By the continuity of the continuous functional calculus, we have some function O on +. [0, 2] with limt→0 O(t) = 0 such that(cid:13)(cid:13)√x − √y(cid:13)(cid:13) ≤ O(kx − yk), for all x, y ∈ A1 For all n, we have(cid:13)(cid:13)(cid:13)√a( 1 √b − √z(cid:13)(cid:13)(cid:13) ≤ O(kb − zk) = O(k(a − b)+k) = O(h(a, b)). (cid:13)(cid:13)u − √a(cid:13)(cid:13) ≤(cid:13)(cid:13)(cid:13) ka − u∗uk ≤(cid:13)(cid:13)a − u∗√a + u∗√a − u∗u(cid:13)(cid:13) ≤(cid:13)(cid:13)√a − u∗(cid:13)(cid:13) +(cid:13)(cid:13)√a − u(cid:13)(cid:13) ≤ 2O(h(a, b)). As in the proof of [Ped79] Proposition 1.4.10, we have u∗u ≤ b, and Thus p w h. (cid:3) Another quantification of ≤ on A1 + is given by n(a, b) = inf a≤c≤1kc − bk . For unital A, n(a, b) = p(b⊥, a⊥) and h(a, b) = h(b⊥, a⊥) so it immediately follows that h and n are also uniformly equivalent. Actually, this extends to non-unital A. Corollary 1.6. On A1 + × A1 +, h ≤ n w h, so h is uniformly equivalent to n. Proof. We can essentially apply the proof of Theorem 1.5 with a and b replaced by b⊥ and a⊥ respectively. First, for any c ≥ a, (1.3) yields h(a, b) = sup φ∈Q φ(a − b) ≤ sup φ∈Q φ(c − b) ≤ kc − bk . Also a − b ≤ (a − b)+ so a ≤ b + (a − b)+ and h(a, b) = kb + (a − b)+ − bk so h(a, b) = inf a≤ckc − bk ≤ inf 0≤c≤bkc − bk = n(a, b). Now let z = b⊥ + (a − b)+ and u = lim√a⊥( 1 a ≤ (u∗u)⊥ and(cid:13)(cid:13)(u∗u)⊥ − b(cid:13)(cid:13) ≤ 2O(h(a, b)). Note that (u∗u)⊥ ∈ A even when A 2√b⊥. As before, we have n + z)− 1 6 TRISTAN BICE AND ALESSANDRO VIGNATI is not unital, as then π(z) = 1 = π(u) and hence π((u∗u)⊥) = 0, where π is the (cid:3) character on the unitization eA with kernel A. Restricting to projections P = {p ∈ A : p ≪ p∗} allows for a stronger result. First we need the following lemma, adapted from [10113], which strengthens the standard result that close projections are unitarily equivalent (see [Bla17] II.3.3.5). Lemma 1.7. If p, q ∈ P and kp − qk < 1 then p and q can be exchanged by a symmetry(=self-adjoint unitary), i.e. we have u ∈ eAsa with u2 = 1 and up = qu. Proof. Let a = p + q − 1 ∈ eAsa so ap = qp = qa and aq = pq = pa. Thus a2p = aqp = pqp = pqa = pa2. Also, a2 = pq + qp − p − q + 1 = 1 − (p − q)2 so (cid:13)(cid:13)1 − a2(cid:13)(cid:13) = kp − qk2 < 1 and hence a2 is invertible. Thus we may set u = aa−1 so u2 = 1, as a ∈ Asa. Also, as p commutes with a2 and hence with a−1, (cid:3) up = aa−1p = apa−1 = qaa−1 = qu. Corollary 1.8. If p, q ∈ P and kp − qk < 1 then h(p, q) = kp − qk. Proof. if e(p, q) < 1 then Lemma 1.7 yields an automorphism a 7→ uau of A ex- changing p and q so k(p − q)+k = k(q − p)+k and hence kp − qk = k(p − q)+k ∨ k(q − p)+k = k(p − q)+k = h(p, q). (cid:3) Corollary 1.9. On P, d = h = p = n. Proof. If d(p, q) < 1 then qp is well-supported so we have a range projection [qp] = f (qpq) ∈ A (for continuous f on [0, 1] that is 1 on σ(qpq)). By [Bic13b] §2.3, d(p, q) = p − [qp] ≥ p(p, q) ≥ h(p, q) and so 0 = p(q − [qp]) = (p − [qp])(q − [qp]) (p − q)+ = (p − [qp])+ + ([qp] − q)+ = (p − [qp])+ and hence h(p, q) = h(p, [qp]) = p − [qp], by Corollary 1.8. While if d(p, q) = 1 then 1 = pq⊥p = (pq⊥p)+ so, by (1.7), 1 = (pq⊥p)+ = (p(p − q)p)+ ≤ (p − q)+ = h(p, q) ≤ p(p, q) ≤ p − q ≤ 1. Thus d = h = p on P. A similar argument with [p⊥q⊥]⊥ ∈ A applies to n. (cid:3) Alternatively we could have noted that, by reverting to a C-subalgebra if nec- essary, it suffices to prove Corollary 1.9 when A is generated by p and q. As every irreducible representation of a C-algebra generated by a pair of projections is on a Hilbert space of dimension at most 2, for d = h it suffices to consider A = C or M2, which can be done with some elementary calculations. For our characterizations of d-filters, we will also need to consider a number of In general, for D : X × X → [0,∞], we define unary functions defined from d. xD, Dy : X → [0,∞] by fixing the left/right coordinate, i.e. xD(y) = D(x, y) = Dy(x). Proposition 1.10. For any a, b ∈ A1 + and ǫ ∈ (0, 1), d(aba) ≈ da + db ≈ d(ǫa + (1 − ǫ)b). C∗-ALGEBRA DISTANCE FILTERS 7 Proof. First note daba ≤ 2da + db and hence d(aba) w da + db, as d(c, aba) = kc − cabak ≤ kc − cak + kca − cbak + kcba − cabak ≤ 2d(c, a) + d(c, b). Next, as aba ≤ a2 ≤ a, d(c, a)2 ≤ d(c, aba) + h(aba, a) = d(c, aba) and d(c, b)2 =(cid:13)(cid:13)cb⊥2c(cid:13)(cid:13) ≤(cid:13)(cid:13)cb⊥c(cid:13)(cid:13) =(cid:13)(cid:13)c2 − cbc(cid:13)(cid:13) ≤ kc − cabakkck + kca − ck kback + kcbkkac − ck ≤ d(c, aba) + 2d(c, a) ≤(cid:13)(cid:13)c2 − cabac(cid:13)(cid:13) + kcabac − cback + kcbac − cbck ≤ d(c, aba) + 2pd(c, aba). Thus da + db w d(aba) and hence da + db ≈ d(aba). In particular, for any n ∈ N, setting a = b above and using (1.5) yields (1.8) Also supx∈[0,1](ǫx + 1 − ǫ)n(1 − x) ≤ 1 nǫ so da w dan w da3n w da. da − 1 nǫ ≤ da − d((ǫa + 1 − ǫ)n, a) by (△) by (1.8) ≤ d(ǫa + 1 − ǫ)n w d(ǫa + 1 − ǫ) w d(ǫa + (1 − ǫ)b) by (1.5). As n was arbitrary, da w d(ǫa + (1 − ǫ)b) and, by symmetry, db w d(ǫa + (1 − ǫ)b). On the other hand, d(ǫa + (1 − ǫ)b) w da + db follows from d(c, ǫa + (1 − ǫ)b) = kc − c(ǫa + (1 − ǫ)b)k = kǫc − ǫca + (1 − ǫ)c − (1 − ǫ)cb)k ≤ ǫ kc − cak + (1 − ǫ)kc − cbk = ǫd(c, a) + (1 − ǫ)d(c, b). (cid:3) A slightly better substitute for multiplication on A1 + than aba might be a ⊙ b = √ab√a. For if ab = ba then a ⊙ b = ab and (a ⊙ b) ⊙ c = a ⊙ (b ⊙ c). In particular, ⊙ is left alternative, i.e., (a ⊙ a) ⊙ b = a ⊙ (a ⊙ b) and also right distributive, i.e., a ⊙ (b + c) = a ⊙ b + a ⊙ c. We can also quantify ≪ using ⊙ by f (a, b) = ka − a ⊙ bk . The advantage of f over d is that it determines h in a natural way. Proposition 1.11. We have d ≈ f = f ◦ h and (1.9) h(a, b) = sup c∈A1 + (f (c, b) − f (c, a)). Proof. First note that, for a, b ∈ A1 + (cid:13)(cid:13)ab⊥(cid:13)(cid:13)2 =(cid:13)(cid:13)b⊥a2b⊥(cid:13)(cid:13) ≤(cid:13)(cid:13)a2b(cid:13)(cid:13) . 8 TRISTAN BICE AND ALESSANDRO VIGNATI Thus, as binary functions on A1 +, d(a, b) =(cid:13)(cid:13)ab⊥(cid:13)(cid:13) ≈(cid:13)(cid:13)√ab⊥(cid:13)(cid:13) ≈(cid:13)(cid:13)(cid:13)√a√b⊥(cid:13)(cid:13)(cid:13) ≈(cid:13)(cid:13)(cid:13)√a√b⊥(cid:13)(cid:13)(cid:13) As in the proof of (1.5), we have 2 = c(a, b). f (a, b) =(cid:13)(cid:13)a ⊙ b⊥(cid:13)(cid:13) =(cid:13)(cid:13)(a ⊙ b⊥)+(cid:13)(cid:13) ≤(cid:13)(cid:13)(a ⊙ c⊥)+(cid:13)(cid:13) +(cid:13)(cid:13)(a ⊙ (b⊥ − c⊥))+(cid:13)(cid:13) ≤(cid:13)(cid:13)(a ⊙ c⊥)(cid:13)(cid:13) + k(c − b)+k ≤ f (a, c) + h(c, b). (f (d, a) − f (d, b)). Conversely, take a, b ∈ A1 Thus f = f ◦ h and h(a, b) ≥ supd∈A1 +. If h(a, b) = 0 then the reverse inequality is immediate from h(a, b) = f (0, b)−f (0, c). Otherwise, for any ǫ > 0, we can take a pure state φ on A with h(a, b) < φ(a−b)+ǫ. By [AAP86, Proposition 2.2], we have c ∈ A1 + with + φ(c) = 1 and kc ⊙ a − cφ(a)k < ǫ. Thus(cid:13)(cid:13)c ⊙ a⊥ − cφ(a⊥)(cid:13)(cid:13) < ǫ so(cid:13)(cid:13)c ⊙ a⊥(cid:13)(cid:13) <(cid:13)(cid:13)cφ(a⊥)(cid:13)(cid:13) + ǫ = φ(a⊥) + ǫ and h(a, b) < φ(a − b) + ǫ = φ(b⊥ − a⊥) + ǫ = φ(c ⊙ b⊥) − φ(a⊥) + ǫ <(cid:13)(cid:13)c ⊙ b⊥(cid:13)(cid:13) −(cid:13)(cid:13)c ⊙ a⊥(cid:13)(cid:13) + 2ǫ. = f (c, b) − f (c, a) + 2ǫ. As ǫ > 0 was arbitrary, we are done. (cid:3) The drawback of f is that it may not be a distance. Indeed, by (1.9), f is a distance if and only if h ≤ f . But by Corollary 1.9, for projections p, q ∈ A, h(p, q) = d(p, q) =qkpq⊥pk =pf (p, q). So h (cid:2) f whenver we have projections p, q ∈ A with 0 < f (p, q) < 1. A more involved argument could even show that h (cid:2) f whenever A is non-commutative. 2. Filters Definition 2.1. For D : X×X → [0,∞], define the following conditions on Y ⊆ X. (D-filter) (D(c, b) + D(c, a)) = 0. (D-directed) (D-closed) (D-initial) (D-coinitial) (D-cofinal) c∈Y c∈Y c∈Y a, b ∈ Y ⇔ inf a, b ∈ Y ⇒ inf b ∈ Y ⇐ inf b ∈ Y ⇒ inf b ∈ X ⇒ inf c ∈ X ⇒ inf b∈Y c∈Y c∈Y (D(c, b) + D(c, a)) = 0. D(c, b) = 0. D(c, b) = 0. D(c, b) = 0. D(c, b) = 0. For any operation • : X n → X, we also call Y ⊆ X •-closed if •[Y n] ⊆ Y . C∗-ALGEBRA DISTANCE FILTERS 9 This terminology covers a number of familiar concepts from metric, order and + are increasing C∗-algebra theory. For example, d-cofinal ≥-directed subsets of A1 approximate units in the usual sense, when considered as self-indexed nets. If D is a metric, D-coinitial/cofinal means dense while D-closed means closed, with respect to the usual ball topology defined by D. The other terms become trivial -- specifically, arbitrary subsets are D-initial, while the empty and one-point subsets are the only D-directed subsets. In particular, for C∗-algebras, e-closed/coinitial means norm closed/dense in the usual sense. On the other hand, for any order relation ≤ (again identified with its charac- teristic function), ≤-closed means upwards closed, ≤-directed means downwards directed and ≤-cofinal means cofinal in the usual sense. In particular, ≤-filters are the usual order-theoretic filters and, more generally, D-filter ⇔ D-directed and D-closed. We now characterize the C∗-algebra filters considered in [Bic13a] in various ways. + is a norm filter, according to [Bic13a] Definition 3.1, if First we recall that F ⊆ A1 inf k∈N a1,...,ak∈F(cid:13)(cid:13)a1a2 . . . akb⊥(cid:13)(cid:13) = 0 ⇒ b ∈ F. Also recall that a subset C of a vector space X is convex if ǫa + (1 − ǫ)b ∈ C whenever a, b ∈ C and ǫ ∈ (0, 1) and F ⊆ C is a face of C if converse also holds, i.e. if, for all a, b ∈ C and ǫ ∈ (0, 1), a, b ∈ F ⇔ ǫa + (1 − ǫ)b ∈ F 2 or any other fixed element of (0, 1)). (for faces it actually suffices to take ǫ = 1 Theorem 2.2. For F ⊆ A1 +, the following are equivalent. (1) F is a d-filter. (2) F is a d-initial h-filter. (3) F is the norm closure of a d-initial ≤-filter. (4) F is norm closed, ≤-closed and ⊙-closed.4 (5) F is norm closed, ≤-closed, 2-closed and convex. (6) F is a norm closed d-cofinal face. (7) F is a norm filter. If A is separable or commutative, they are also equivalent to the following. (8) F is the norm closure of a ≪-filter. Proof. (1)⇔(2) As h w d, d-directed ⇒ h-directed. h-closed ⇒ d-closed. If F is d-initial then inf b∈F h(b, a) = inf b∈F inf c∈F (d(c, b) + h(b, a)) = inf c∈F (d ◦ h)(c, a). So 0 = inf b∈F h(b, a) implies 0 = inf b∈F d(b, a), as d w d ◦ h, which yields the converse implications. 4Note (4)⇒(7) eliminates the real rank zero hypothesis from [Bic13a] Proposition 3.5. 10 TRISTAN BICE AND ALESSANDRO VIGNATI (1)⇒(5) As da2 w da, d(ǫa + (1 − ǫ)b) w da + db, and h ≤ e quantifies ≤, d-filter ⇒ 2-closed and convex. h-closed ⇒ norm closed and ≤-closed. (5)⇒(1) For any a ∈ A1 (2.1) Thus (1)⇒(5) now follows from (1)⇒(2). +, d(a2n , a) → 0 as d(an, a) =(cid:13)(cid:13)ana⊥(cid:13)(cid:13) ≤ sup 2 (b + c), db + dc w da, d(a2n x∈[0,1] Taking a = 1 xn(1 − x) ≤ 1/n. , a) → 0 and h w n yields 2-closed and convex ⇒ d-directed. norm closed and ≤-closed ⇒ h-closed. As h w d, h-closed implies d-closed. (1)⇒(4) As d(a ⊙ b) w d√a + db w da + db, d-filter ⇒ ⊙-closed. By (1)⇒(2), any d-filter is h-closed and hence norm closed and ≤-closed. (4)⇒(1) Taking c = a ⊙ b, da + db w d√a + db w dc and d(c2n , c) → 0 yields ⊙-closed ⇒ d-directed. As d w h w n, norm closed and ≤-closed implies d-closed. (7)⇒(1) We immediately see that norm filter ⇒ d-closed. For any a, b ∈ F , (cid:13)(cid:13)(aba)n(aba)⊥(cid:13)(cid:13) → 0, by (2.1), so aba ∈ F . Thus ∈ F too so, as da + db w d(aba), (aba)3n (1)⇒(7) Assume F ⊆ A1 As F is d-directed, for any a1, . . . , ak ∈ F and ǫ > 0, we can find a ∈ F with d(a, aj) ≤ ǫ, for all j ≤ k, and hence norm filter ⇒ d-directed. + is a d-filter and take b ∈ A1 inf k∈N a1,...,ak∈F(cid:13)(cid:13)a1a2 . . . akb⊥(cid:13)(cid:13) = 0. + with k + ak)b⊥(cid:13)(cid:13) d(a, b) =(cid:13)(cid:13)a(a⊥ ≤(cid:13)(cid:13)aa⊥ k(cid:13)(cid:13) +(cid:13)(cid:13)aakb⊥(cid:13)(cid:13) k−1 + ak−1)akb⊥(cid:13)(cid:13) ≤ ··· ≤ ǫ +(cid:13)(cid:13)a(a⊥ ≤ kǫ +(cid:13)(cid:13)a1a2 . . . akb⊥(cid:13)(cid:13) . Thus inf a∈F d(a, b) = 0. As F is d-closed, b ∈ F so F is a norm filter. (6)⇒(4) Assume F is a norm closed face of A1 +. We first claim where [a, b] = {c ∈ A1 and 1 2 (c + d) = 1 a, b ∈ F ⇒ [a, b] ⊆ F, + : a ≤ c ≤ b}. For if c ∈ [a, b] then d = a + b− c ∈ A1 + 2 (a + b) ∈ F and hence c ∈ F . Next we claim that a ∈ F ⇒ f (a) ⊆ F, C∗-ALGEBRA DISTANCE FILTERS 11 for any continuous f on [0, 1] taking 0 to 0 and 1 to 1, as in [AP92, Lemma 2.1]. For let g(a) = (2a− 1)+ and h(a) = f (a⊥)⊥. Then 1 2 (g(a) + h(a)) = a so g(a), h(a) ∈ F . Thus gn(a), hn(a) ∈ F , for all n, and hence Now we claim F is ⊙-closed, as in the proof of [AP92, Theorem 2.9]. For given a, b ∈ F and ǫ ∈ (0, 1), let cǫ = (ǫ√a + (1 − ǫ)b) ∈ F so f (a) ∈[[gn(a), hn(a)] ⊆ F. ǫcǫ√acǫ + (1 − ǫ)cǫbcǫ = c3 ǫ ∈ F To see that F is ≤-closed when F is also d-cofinal, take a ∈ A1 and hence cǫbcǫ ∈ F . As ǫ → 1, cǫbcǫ → a ⊙ b ∈ F , as required. + with a ≥ b ∈ F and (cn) ⊆ F with d(a, cn) → 0. Then cn ⊙ b ∈ F and cn ⊙ b ≤ cn ⊙ a ≤ cn so cn ⊙ a → a ∈ F , by our first claim above. (1)⇒(6) Take a d-filter F . By (1)⇒(5), F is norm closed and convex. As F is d- initial, d-closed and da + db w d(ǫa + (1− ǫ)b), F is also a face of A1 +. As A + and ǫ > 0 we have c ∈ A1 has an approximate unit, for any a, b ∈ A1 + with d(a, c), d(b, c) < ǫ. In particular this applies to b ∈ F and, as n w h w d, we may take c ≥ b. As F is ≤-closed, c ∈ F which shows that F is d-cofinal. (3)⇒(2) As d and h are e-invariant and n w h, for any F ⊆ A1 +, F is d-initial ⇒ F is d-initial. F is ≤-directed ⇒ F is h-directed. F is ≤-closed ⇒ F is h-closed. (1)⇒(3) Take a d-filter F and assume first that A is unital. Consider the invertible elements G of F . For every a ∈ F and ǫ > 0, (1 − ǫ)a + ǫ ∈ G so G = F . As F is ≤-closed and d-initial, so is G. It only remains to show that G is ≤-directed. So take a, b ∈ G. For some ǫ > 0 and a′, b′ ∈ A1 +, we have a = ǫ + (1 − ǫ)a′ and b = ǫ + (1 − ǫ)b′. As F is a face of A1 + containing 1, a′, b′ ∈ F . As F is h-directed, we have c′ ∈ F with h(c′, a′), h(c′, b′) < ǫ/2. Letting c = ǫ + (1 − ǫ)c′ ∈ F , we thus have h(c, a), h(c, b) < ǫ/2. Thus d = c − (c − a)+ − (c − b)+ is an invertible element of A1 + (as c ≥ ǫ and kc − dk < ǫ) and we further claim that d ∈ F and hence d ∈ G. Indeed, for any δ > 0, we have e ∈ F with d(e, a), d(e, b), d(e, c) < δ. Thus ke(c − a)+k ≤ ke(c − a)k < 2δ and ke(c − b)+k < 2δ so d(e, d) < 5δ. As δ > 0 was arbitrary and F is d-closed, we have d ∈ F , as required. + by taking the (upwards) ≤-closure. By what we just proved, the invertible elements G′ of F ′ are a ≤-filter with G′ = F ′. In particular, G′ is d-coinitial in F ′. Thus, for any a ∈ F , aG′a is d-coinitial in F . Hence the ≤-closure G of aG′a in A1 If A is not unital then first extend F to a d-filter F ′ in eA1 + with G = F , by Corollary 1.6. + is a d-initial ≤-filter in A1 (1)⇒(8) If A is separable then we can take dense (an) ⊆ F and let a =P 2−nan ∈ F . As noted above, f (a) ∈ F , for any continuous f on [0, 1] taking 0 to 0 and 1 to 1. Thus by choosing such (fn) converging pointwise to 0 everywhere except at 1 and satisfying f1 ≫ f2 ≫ . . ., the (upwards) ≪-closure G of (fn(a)) is a ≪-filter with F = G. Now take G = {a ∈ F : a ≫ b ∈ F}. 12 TRISTAN BICE AND ALESSANDRO VIGNATI Again, if a ∈ F then f (a) ∈ F , for any continuous f on [0, 1] taking 0 to 0 and 1 to 1. In particular, for any ǫ > 0 we can take f (x) = (1 + ǫ)x∧ 1 and g(x) = (ǫ−1(x−1)+1)+ so f (a) ≫ g(a) ∈ F and hence f (a) ∈ G. As ǫ → 0, f (a) → a so F = G. Likewise, if a ≫ b ∈ F then a ≫ f (b) ≫ g(b) ∈ F so G = {a ∈ F : a ≫ b ≫ c ∈ F}. If A is commutative, a ≫ b and a′ ≫ b′ implies aa′ ≫ bb′. For a, a′ ∈ G, we have b, b′, c, c′ ∈ F with a ≫ b ≫ c and a′ ≫ b′ ≫ c′ so bb′ ≫ cc′ ∈ F and hence a, a′ ≫ bb′ ∈ G, i.e. G is ≪-directed and hence a ≪-filter. (8)⇒(3) As d quantifies ≪ and ≪ ◦ ≤ ⊆ ≪ ⊆ ≤, ≪-initial ⇒ d-initial. ≪-closed and ≪-initial ⇒ ≤-closed. ≪-directed ⇒ ≤-directed. (cid:3) In (4) and (5), we could not replace '≤-closed' with '≪-closed'. For example, the norm closure C of the convex combinations of the functions xn in C([0, 1]), for n ∈ N, satisfies these conditions -- as every f ∈ C is positive on (0, 1], C is vacuously ≪-closed -- however C is not ≤-closed, being bounded above by x. Although we could replace 'norm closed and ≤-closed' with 'h-closed' or 'd-closed'. Furthermore, not every d-filter is the norm closure of a ≪-filter. Indeed, if this were the case then, for any non-unital A, the d-filter {1 − a : a ∈ A1 + would be the norm closure of a ≪-filter F . Then 1 − F would be an 'almost idempotent' approximate unit of A in the sense of [Bla17, II.4.1.1]. However, a C∗-algebra was recently constructed in [BK17] that does not possess such an approximate unit. All ω1-unital C∗-algebras have them though, by another result from [BK17], so (8) could be extended to any A with a dense subset of size ≤ ω1. +} in eA1 We can at least say a bit more in the commutative case. Proposition 2.3. If A is commutative and F ⊆ A1 G = {a ∈ F : a ≫ b ∈ F} + is a d-filter then is the unique ≪-filter with F = G. Proof. The only thing left to show is uniqueness. By the Gelfand represtentation, we may assume that A = C0(X) for some locally compact Hausdorff X, so Take a d-filter F ⊆ A1 f ≪ g + and let ⇔ X \ f −1{0} ⊆ g−1{1}. C = \f ∈F f −1{1}. For any ≪-filter G with G = F , we must also have C =Tg∈G g−1{1}. Otherwise, we could pick some x ∈ Tg∈G g−1{1} \ C and f ∈ F with f (x) 6= 1 and then kf − gk ≥ g(x) − f (x) = 1 − f (x) > 0, for all g ∈ G, contradicting G = F . + with C ⊆ f −1{1}◦. For every x ∈ X \ f −1{1}◦, we have gx ∈ G Take f ∈ A1 with gx(x) 6= 1. Thus we can pick arbitrary g ∈ G and cover the compact set g−1[ 1 x1 {1}, . . . , X \ g−1 2 , 1]\ f −1{1}◦ with finitely many open sets X \ g−1 xk {1}. As G is ≪-directed, we have some h ∈ G with h ≪ g, gx1, . . . , gxk and hence X \ h−1{0} ⊆ g−1{1} ∩ g−1 x1 {1} ∩ . . . ∩ g−1 xk {1} ⊆ f −1{1}◦ ⊆ f −1{1}, C∗-ALGEBRA DISTANCE FILTERS i.e., h ≪ f . Thus f ∈ G, as G is ≪-closed, so {f ∈ F : f ≫ g ∈ F} ⊆ {f ∈ A1 + : C ⊆ f −1{1}◦} ⊆ G. Conversely, G ⊆ {f ∈ F : f ≫ g ∈ F}, as G is a ≪-filter contained in F . 13 (cid:3) This does not extend to non-commutative A, i.e. G = {a ∈ F : a ≫ b ∈ F} may fail to be a ≪-filter and F may contain various dense ≪-filters. For example, consider A = C([0, 1], M2) and take everywhere rank 1 projections p, q ∈ A with p(0) = P = q(0) but p(x) 6= q(x), for all x > 0. Also take continuous fn on [0, 1] with f1 ≫ f2 ≫ . . . andTn f −1{1} = {0}. Then the ≪-closures F, G ⊆ A1 + of (fnp) and (fnq) respectively are distinct ≪-filters with F = G = {a ∈ A1 + : a(0) ≥ P}. Definition 2.4. We say Y generates a D-filter F ⊆ X if F is the smallest D-filter containing Y . We call X a D-semilattice if every Y ⊆ X generates a D-filter. For posets, ≤-semilattices are precisely the meet semilattices in the usual sense. Proposition 2.5. If ≤ is a partial order on X then X is a ≤-semilattice ⇔ every x, y ∈ X has an infimum x ∧ y ∈ X. Proof. If every x, y ∈ X has an infimum x ∧ y ∈ X then the ≤-closure of the ∧- closure of any Y ⊆ X is the ≤-filter generated by Y , so X is a ≤-semilattice. If some x, y ∈ X have no infimum then \z≤x,y {w ∈ X : z ≤ w} is an intersection of ≤-filters containing x and y but no lower bound of x and y. Thus {x, y} does not generate a ≤-filter and hence X is not a ≤-semilattice. Proposition 2.6. A1 + is a d-semilattice. Proof. For any B ⊆ A1 +, let D be the ⊙-closure of B, so D is d-directed and every d-filter containing B must contain D. Let F by the d-closure of D, so F is a d- filter and every d-filter containing B and hence D contains F , i.e., F is the d-filter generated by B. (cid:3) Definition 2.7. We call Y ⊆ X D-centred if, for all y1, . . . , yk ⊆ Y , (cid:3) inf x∈X D(x, y1) + . . . + D(x, yk) = 0. If ≤ is a partial order on X then Y ⊆ X is ≤-centred if and only if every finite subset of Y has a lower bound in X, i.e., if and only if Y is centred in the usual order theoretic sense. + rather than the positive unit ball A1 As with filters, we see that the centred subsets of C∗-algebras considered in [Bic13a] are precisely the d-centred subsets, this time in the positive unit sphere A=1 + is norm centred, according to [Bic13a, Definition 2.1], if the multiplicative closure of C is contained in the unit sphere. Proposition 2.8. For C ⊆ A=1 (1) C is d-centred in A=1 + . (2) C is h-centred in A=1 + . +. Specifically, recall that C ⊆ A=1 + , the following are equivalent. 14 TRISTAN BICE AND ALESSANDRO VIGNATI (3) C is norm centred. (4) C is generates a proper d-filter in A1 +. Proof. (1)⇒(2) Immediate from h ≤ 2d. (2)⇒(1) If kak = 1 then kank = 1 and, for any b ∈ A1 +, d(an, b)2 ≤ d(an, a) + h(a, b) → h(a, b). (1)⇒(3) If C is d-centred then, for any a1, . . . , ak ∈ C and ǫ > 0, we have b ∈ A=1 + with d(b, a1) + . . . + d(b, ak) < ǫ so 1 =(cid:13)(cid:13)b(a1 + a⊥ 1 )(cid:13)(cid:13) 2 )a1(cid:13)(cid:13) + d(b, a1) ≤(cid:13)(cid:13)b(a2 + a⊥ ≤(cid:13)(cid:13)b(a3 + a⊥ 3 )a2a1(cid:13)(cid:13) + d(b, a2) + d(b, a1) ≤ . . . ≤ kbak . . . a1k + ǫ ≤ kak . . . a1k + ǫ. As ǫ > 0 was arbitrary, C is norm centred. (3)⇒(4) If the multiplicative closure of C is contained in the unit sphere then the same goes for the closure D of C under the operation (a, b) 7→ aba. The same then applies to the d-closure F of D so, in particular, F is proper. As in Proposition 2.6, F is the d-filter generated by C. (4)⇒(1) We show that a d-filter F ⊆ A1 + is proper if and only if it is contained in the positive unit sphere. For if a ∈ F and kak < 1 then an ∈ F so, for any b ∈ A1 +, d(an, b) ≤ kank = kakn → 0, and hence b ∈ F , as F is d-closed. If C is contained in such a d-filter F then, for any c1, . . . , cn ∈ C, inf d(a, c1) + . . . + d(a, ck) = 0, d(a, c1) + . . . + d(a, ck) ≤ inf a∈F a∈A=1 + as F is d-directed, i.e., C is d-centred in A=1 + . (cid:3) In particular, the maximal d-centred subsets are precisely the maximal proper d-filters. These were the original quantum filters defined by Farah and Weaver to study pure states. Pure states correspond to minimal projections in A∗∗ and, more generally, d-filters correspond to the compact projections A∗∗ introduced by Akemann (which was touched on briefly in [Bic13a, Corollary 3.4]). This is the connection we explore next. 3. Compact Projections + defined by any projection p ∈ A∗∗, that is, Let ↑ p denote the upper set in A1 Definition 3.1. A projection p ∈ A∗∗ is compact if p =^ ↑ p. ↑ p = {a ∈ A1 + : p ≤ a}. Note that for p to be compact it is implicit that ↑ p is non-empty. C∗-ALGEBRA DISTANCE FILTERS 15 Theorem 3.2. We have mutually inverse bijections F 7→^ F between compact projections p ∈ A∗∗ and d-filters F ⊆ A1 projections p, q ∈ A∗∗ and corresponding d-filters F, G ⊆ A1 +, (3.1) p 7→ ↑ p d(a, b). and d(p, q) = sup b∈G inf a∈F +. Moreover, for compact + implies p ≪ b, as d2 ≤ d ◦ h on A1 Proof. Take a projection p ∈ A∗∗ and consider ↑ p. If p = pa then pa2 = pa = p, i.e., p ≪ a implies p ≪ a2 so ↑ p is 2-closed. Likewise, if p = pa, p = pb and ǫ ∈ (0, 1) then p(ǫa + (1 − ǫ)b) = ǫp + (1 − ǫ)p = p, i.e., p ≪ (ǫa + (1 − ǫ)b) so ↑ p is convex. Also p ≪ a ≤ b ∈ A1 +, so ↑ p is ≤-closed. Finally, if an → a and p ≪ an, for all n, then d(p, a) = lim d(p, an) = 0, i.e., p ≪ a so ↑ p is norm closed and thus a d-filter, by Theorem 2.2. Conversely, take a d-filter F ⊆ A1 + which, by (2.2), contains a dense ≤-filter F ′. The pointwise infimum of F ′ on Q (recall that Q is the space of quasistates on A) is an affine function and thus defines an element p ∈ A∗∗. As ≤ on A∗∗ sa is determined by Q, p =V F ′ =V F =V ↑ p. As p takes Q to [0, 1], p is positive and has norm at most 1, i.e., p ∈ A∗∗1 sup a∈F i.e., for all a ∈ F , p ≪ a so √p ≪ a. Thus √p ≤V F = p so p is a projection. projection q = V G which is a pointwise infimum of G on Q. Then pG′p is also ≤-directed so pqp =V pG′p =V pGp is also a pointwise infimum on Q and hence Now take another d-filter G containing a dense ≤-filter G′ and defining a compact + . As F is d-initial, d(p, a)2 ≤ sup (h(p, b) + d(b, a)) = 0, φ(pq⊥p) inf b∈F a∈F d(p, q)2 =(cid:13)(cid:13)pq⊥p(cid:13)(cid:13) = sup φ∈Q φ(pqp) φ∈Q = φ(q) − inf = φ(q) − inf = sup φ∈Q,b∈G φ(pb⊥p) φ(pbp) (3.2) φ∈Q,b∈G = sup φ(pb⊥2p) φ∈Q,b∈G For (3.2), note that(cid:13)(cid:13)ab⊥(cid:13)(cid:13)2 as G is d-initial and d-closed, b ∈ G if and only if b⊥2⊥ ∈ G, i.e., {b⊥ : b ∈ G} = {b⊥2 : b ∈ G}. Fix b ∈ G and define weak* continuous fa : Q → [0, 1] by b∈G = sup d(p, b)2. b∈G(cid:13)(cid:13)pb⊥2p(cid:13)(cid:13) = sup =(cid:13)(cid:13)ab⊥2a(cid:13)(cid:13) ≤(cid:13)(cid:13)ab⊥2(cid:13)(cid:13) ≤(cid:13)(cid:13)ab⊥(cid:13)(cid:13) so db ≈ db⊥2⊥. Thus, fa(φ) = (φ(b⊥ab⊥) −(cid:13)(cid:13)b⊥pb⊥(cid:13)(cid:13))+. a∈F(cid:13)(cid:13)b⊥ab⊥(cid:13)(cid:13) . a∈F ′(cid:13)(cid:13)b⊥ab⊥(cid:13)(cid:13) ≤(cid:13)(cid:13)b⊥pb⊥(cid:13)(cid:13) ≤ inf inf Then (fa)a∈F ′ is downwards directed in the product ordering on [0, 1]Q and con- verges to 0 pointwise. As Q is weak* compact, Dini's theorem says (fa)a∈F ′ must actually converge uniformly to 0 on Q and hence 16 TRISTAN BICE AND ALESSANDRO VIGNATI As F is 2-closed and √ -closed (as F is ≤-closed), a∈F(cid:13)(cid:13)b⊥ab⊥(cid:13)(cid:13) = inf a∈F(cid:13)(cid:13)b⊥a2b⊥(cid:13)(cid:13) = inf Thus, together with the above we have d(a, b)2 = inf inf a∈F a∈F ′(cid:13)(cid:13)b⊥ab⊥(cid:13)(cid:13) =(cid:13)(cid:13)b⊥pb⊥(cid:13)(cid:13) = d(p, b)2. (3.3) d(p, q) = sup b∈G d(p, b) = sup b∈G inf a∈F d(a, b). Now note that (3.4) sup b∈G inf a∈F d(a, b) = 0 ⇔ G ⊆ F. Indeed, the d-initiality of F yields ⇐, while the fact F is d-closed yields ⇒. Com- injective on d-filters. Thus the given maps are bijections, as required. bined with (3.3), this shows that p = q implies F = G, i.e., the map F 7→V F is (cid:3) In the above proof, we used dense ≤-filter subsets of d-filters in a couple of places, but this was not absolutely necessary. Indeed, one could verify directly that pointwise infimums on Q of h-directed subsets are affine and hence define elements of A∗∗. Likewise, Dini's theorem can be generalized to h-directed subsets and even h-Cauchy nets -- see [Bic16, Theorem 1]. The gist of Theorem 3.2 is that compact projections in A∗∗ can be more con- cretely represented by d-filters in A, and this extends to various relations or func- tions one might consider. For example, from (3.1) and (3.4) we immediately see that, for compact projections p, q ∈ A∗∗ and corresponding d-filters F, G ⊆ A1 +, p ≤ q ⇔ F ⊇ G. Likewise, as kp − qk = max{(cid:13)(cid:13)pq⊥(cid:13)(cid:13) ,(cid:13)(cid:13)qp⊥(cid:13)(cid:13)}, (3.1) yields kp − qk = max(sup d(a, b), sup a∈F inf b∈G inf a∈F b∈G d(b, a)), i.e., the metric on compact projections corresponds to the Hausdorff metric on d-filters. We can also show that the natural quantification of orthogonality on compact projections is determined by the corresponding d-filters. Theorem 3.3. For compact p, q ∈ A∗∗ and d-filters F = ↑ p and G = ↑ q, kpqk = inf a∈F,b∈Gkabk . Proof. Let r = inf a∈F,b∈G kabk. As p ≪ F and q ≪ G, we immediately have kpqk2 = kqpqk ≤ inf a∈F(cid:13)(cid:13)qa2q(cid:13)(cid:13) = inf a∈F kaqak ≤ inf a∈F,b∈G(cid:13)(cid:13)ab2a(cid:13)(cid:13) = r2. Conversely, take a dense ≤-filter F ′ ⊆ F and, for any a ∈ F ′, consider Qa = {φ ∈ Q : φ[G] = {1} and φ(a) ≥ r2}. By [Bic13a, Theorem 2.2], each Qa is non-empty. SoTa∈F ′ Qa is a directed inter- section of non-empty weak* compact subsets and we thus have some φ ∈Ta∈F As φ[G] = {1}, φ(q) = 1 and hence kpqk2 = kqpqk ≥ φ(qpq) = φ(p) ≥ r2. Qa. (cid:3) A natural question to ask is if the infimum above is actually a minimum. Question 1 ([Bic15]). Do we always have a ∈ F and b ∈ G with kpqk = kabk? C∗-ALGEBRA DISTANCE FILTERS 17 When pq = 0 the answer is yes, by Akemann's non-commutative Urysohn lemma -- see [Ake71, Lemma III.1]. However, we feel that a truly non-commutative Urysohn lemma should apply to compact projections that do not commute. Dual to compact projections, we have open projections. Specifically, let ↓ p denote the lower set in A1 + defined by any projection p ∈ A∗∗, i.e. ↓ p = {a ∈ A1 + : a ≤ p}. Definition 3.4. A projection p ∈ A∗∗ is open if p =W ↓ p and closed if p⊥ is open. For D, E : X × X → [0,∞], define r∈R ((rD) ◦ E), eD ◦ E = sup D ◦ E ≤ eD ◦ E ≤ D ◦ E. C w D ⇒ eC ◦ E ≤ eD ◦ E. So (eD ◦ E)(x, y) = inf{r : ∀ǫ > 0 ∃z ∈ X (xDz < ǫ and zEy < r)}. In particular, The following proof was inspired by interpolation arguments introduced by Brown -- see [Bro88] -- and adapted by Akemann and Pedersen -- see [AP92] (although the distance-like functions they used were never formalized as such). Theorem 3.5. Assume p ∈ A∗∗ is a closed projection. On A1 (3.5) +, we have p(≪ ◦ e) w pd Proof. inf a∈A1 + d(p, a) > 0 then (3.5) holds vacuously on A1 If inf a∈A1 d(p, a) = 0. We first claim a weakened form of (3.5) on A1 + +. So assume +, namely In more classical terms, we are claiming that (3.6) p(ed ◦ e) w pd. ∀ǫ > 0 ∃δ > 0 ∀a ∈ A1 + d(p, a) < δ ⇒ ∀γ > 0 ∃b ∈ A1 + (d(p, b) < γ and ka − bk < ǫ). To see this, take ǫ > 0. By Theorem 1.5, we can take δ > 0 such that h(a, b) < 4δ ⇒ p(a, b) < 1 2 ǫ, (3.7) for all a, b ∈ A1 d(p, a) = 0 and + with d(a, u), d(p, u) < γ2. By A1 + is d-directed, for any γ > 0, we have u ∈ A1 Corollary 1.6, n w h ≤ 2d so we may also assume a ≤ u (alternatively, use [Bro15, Proposition 1]) and hence u − a ∈ A1 +. We may further assume γ2 < δ so +. Now take a ∈ A1 + with d(p, a) < δ. As inf a∈A1 + h(u − a, p⊥) ≤ 2d(u − a, p⊥) = 2 k(u − a)pk ≤ 2(k(u − ua)pk + k(ua − a)pk) ≤ 2(kp − apk + kua − ak) = 2(d(p, a) + d(a, u)) < 4δ. 18 TRISTAN BICE AND ALESSANDRO VIGNATI As u − a− p⊥ is the pointwise infimum on Q of (u − a− c)p⊥≥c∈A1 again yields c ∈ A1 i.e., we have d ∈ A1 , Dini's theorem + with c ≤ p⊥ and h(u − a, c) < 4δ. By (3.7), p(u − a, c) < 1 2 ǫ, + with ku − a − dk < 1 2 ǫ and d ≤ c. Setting b = (u − d)+, + ku − d − bk = k(d − u)+k ≤ k(d + a − u)+k ≤ kd + a − uk < 1 2 ǫ so ka − bk ≤ ka + d − uk + ku − d − bk < ǫ. As pd = 0 and u − d ≤ b, d(p, b)2 ≤ d(p, u − d) + h(u − d, b) = d(p, u) < γ2, thus proving (3.6). Now (3.5) is saying the same thing as (3.6), just with d(p, b) < γ strengthened to p ≪ b. To prove this, we iterate (3.6). First take δn > 0 satisfying (3.6) with ǫ replaced by ǫ/2n, for any fixed ǫ > 0. So for any a1 ∈ A1 + with d(p, a1) < δ1, we + with d(p, an+1) < δn+1 and kan − an+1k < ǫ/2n. can recursively take an+1 ∈ A1 Thus (an) has a limit b ∈ A1 + with d(p, b) ≤ d(p, an) + e(an, b) → 0, i.e., p ≪ b. Also ka1 − bk <P ǫ/2n = ǫ, thus proving (3.5). result is available. Specifically, if a ∈ A1 d(p, a) = r < 1 then p ≪ f (a) ∈ A1 0 to 0 and [1 − r, 1] to 1. Thus inf p≪b∈A1 As one might expect, in the commutative case an easier proof of a stronger + and p is a projection with ap = pa and +, for any continuous function f on [0, 1] taking + ka − bk = r, i.e. (cid:3) (≪ ◦ e)(p, a) = d(p, a). + In particular, for any projection p, inf a∈A1 +,ap=pa d(p, a) must be 0 or 1. However, d(p, a). For example, let A be the C-subalgebra of this is not true for inf a∈A1 C([0, 1], M2) with f (0) ∈ CQ for some fixed rank one Q ∈ M2. Take any other rank one P ∈ M2 and define p on [0, 1] by p(0) = Q and p(x) = P otherwise. This represents a closed projection in A∗∗ (as the atomic representation is faithful on closed projections -- see [Ped79, Theorem 4.3.15]) with inf a∈A1 d(p, a) = kP − Qk, which can be anywhere between 0 and 1. Now we can show that 'compact' is the same as 'closed and bounded'. Indeed, this is usually taken as the definition, i.e., compact projections are usually defined as closed projections satisfying some notion of boundedness, like p ≤ a ∈ A+ (see [Ake71, Definition II.1]) or p ≪ a ∈ A1 + (see [OrT11, §3.5]). We also mention some other boundedness notions below in Proposition 3.8. However, we feel this obscures the duality between compact and open projections and it is more natural to define them independently via ↑ p and ↓ p respectively, as done here. Corollary 3.6. A projection p ∈ A∗∗ is compact if and only if p is closed and + d(p, a) = 0. inf a∈A1 + Proof. If p is compact then ↑ p is non-empty so certainly inf a∈A1 see that p is closed, consider + d(p, a) = 0. To B = {a ∈ A : ap = 0 = pa}, which is immediately seen to be a (hereditary) C-(sub)algebra. So B1 + is d-directed + in A∗∗ which is a projection and also a pointwise supremum on Q. For any φ ∈ Q with φ(p) = 0, we have an ∈ ↑ p with φ(an) → 0. We also have bn ∈ A with φ(bn) → kφk so a⊥ n ) → kφk (by and hence has a supremumW B1 the GNS construction). Thus p⊥ =W B1 + is open so p is closed. n ∈ B1 + and φ(a⊥ n ba⊥ n ba⊥ C∗-ALGEBRA DISTANCE FILTERS 19 If p is closed then, by (3.5), (3.8) d(p, a) = 0 inf a∈A1 + ⇒ ∃a ∈ A1 + (p ≪ a). + Alternatively, note inf a∈A1 d(p, a) = 0 implies the facial support {φ ∈ Q : φ(p) = 1} is weak* closed in Q so [AAP89, Lemma 2.4] yields (3.8). In any case, we can take a ∈ ↑ p. For all b ∈ ↓ p⊥, we then have ab⊥a ∈ ↑ p. Also, as p is closed, i.e., p⊥ is open, we have p = p⊥⊥ = (W ↓ p⊥)⊥ =V(↓ p⊥)⊥ so p = apa =^ a(↓ p⊥)⊥a =^{ab⊥a : b ∈ ↓ p⊥} ≥^ ↑ p ≥ p. (for the second equality note, as φ(a · a) ∈ Q whenever a ∈ A1 + and φ ∈ Q, inf c∈C φ(c) = φ(d), for all φ ∈ Q, implies inf c∈C φ(aca) = φ(ada), for all φ ∈ Q). Thus p is compact. (cid:3) For (3.8), it is crucial for p to be closed. Theorem 3.7. It is possible to have open p ∈ A∗∗ with ∄a ∈ A1 d(p, a) = 0 inf a∈A1 + but + (p ≪ a). Proof. We consider a variant of the non-regular open dense projection considered in [AB15, Example 4], where A is a C-subalgebra of C([0, 1],B(H)), i.e., the continuous functions from [0, 1] to B(H) for a separable infinite dimensional Hilbert space H. First let (xn) be a countable dense subset of (0, 1) (actually, it suffices to have inf n xn = 0), let (en) be an orthonormal basis for H and let (Pn) be the rank 1 projections onto (Cen). Define pn : [0, 1] → B(H) by if x > rn if x ≤ rn. Let Q be the projection onto Cv, for v =P 2−nen, and define q : [0, 1] → B(H) by pn(x) =(Pn q(x) =(Q if x > 0 if x = 0. 0 0 Let B and C be the hereditary C-subalgebras of K = C([0, 1],K(H)) defined by p =W pn and p′ = p ∨ q, i.e. B = pKp ∩ K and C = p′Kp′ ∩ K. Let A be the C-subalgebra of C([0, 1],B(H)) generated by C and the constant projection Q⊥. Let an = Q⊥ + fnq ∈ A, for some continuous function fn on [0, 1] with fn(0) = 0 and fn(x) = 1, for all x ∈ [ 1 d(pB, a) = 0 for pB =W B1 d(b, an) = d(p, an) = kp(1 − fn)Qk → 0, + ∈ A∗∗. However, for each x ∈ (0, 1), n , 1]. Then i.e., inf a∈A1 sup b∈B {a(x) : a ∈ A} = C1 + K(q′Hq′). Thus if p ≤ a ∈ A1 + then, for all x ∈ (0, 1), we must have a(x) = 1 − f (x)q′(x), where q′(x) = (p ∨ q − p)(x) and f is some continuous function on [0, 1]. But q′ is discontinuous at each rn, so the only way a could be continuous is if f (rn) = 0 so a(rn) = 1 and hence qa(rn) = Q, for all n. But then continuity yields qa(0) = Q, + 20 TRISTAN BICE AND ALESSANDRO VIGNATI contradicting the fact qa(0) = 0, by the definition of A. Thus there is no a ∈ A1 with pB ≪ a. + (cid:3) There are several other boundedness conditions on p that one might consider. However, they are all equivalent, even in a more general context. Proposition 3.8. For any a ∈ A1 +, r > 1 and C-subalgebra B ⊆ A, TFAE. (1) ∃b ∈ Br + (a ≤ b). (2) inf b∈Bsa h(a, b) = 0. (3) inf b∈B1 d(a, b) = 0. (4) inf b∈B d(a, b) = 0. + Proof. We immediately have (1)⇒(2) and (3)⇒(4). (2)⇒(3) By (1.6) (and the existence of an approximate unit for B in B1 +), h(a, c). (h(a, c) + d(c, b)) = inf inf b∈B1 + d(a, b)2 ≤ inf c∈Bsa inf b∈B1 + c∈Bsa (4)⇒(2) If d(a, bn) → 0, i.e., abn → a and hence b∗ h(a, b∗ nbn) ≤ h(a, b∗ nabn) ≤ e(a, b∗ nabn → a, then nabn) → 0. (3)⇒(1) See [Ake70, Theorem 1.2]. (cid:3) Another relation on compact relations one might like to quantify is 'interior containment'. Specifically, define the interior p◦ of any projection p ∈ A∗∗ to be the largest open projection below p, i.e. p◦ =_ ↓ p. c(p, q) = kp − pq◦k . We quantify the interior containment relation p ≤ q◦ by the distance By Akemann's non-commutative Urysohn lemma -- see [Ake71, Lemma III.1] -- p ≤ q◦ is equivalent to ∃a ∈ A1 + (p ≤ a ≤ q). For commutative A, this means c(p, q) = inf a∈F d(a, b), sup b∈G where F = ↑ p and G = ↑ q. However, this does not extend non-commutative A and in general c can behave quite badly with respect to the metric e. Theorem 3.9. It is possible to have compact p, q ∈ A∗∗ with p (cid:2) q◦, (3.9) d(a, b) = 0 but inf a∈F sup b∈G where F = ↑ p and G = ↑ q. It is also possible that c 6w e◦ c on compact projections. Proof. Take open p ∈ A∗∗ as in Theorem 3.7. Consider p =W ↓ p in eA∗∗ and let q =W A1 + ∈ eA so 0 = inf a∈↓q d(p, a) = inf a∈↓q supb∈↓p d(p, a) but ∄a ∈ ↓ q (p ≪ a). In unital C-algebras, compact ⇔ closed and d(a, b) = d(b⊥, a⊥) so 0 = inf a∈↑q⊥ sup b∈↓p⊥ d(a, p⊥), even though there no a ∈ ↑ q⊥ with a ≪ p⊥, i.e., q⊥ (cid:2) p⊥◦. For c 6w c ◦ e, let A = ([0, 1], M2). Let Pθ be the projection onto C(sin θ, cos θ), Pθ =(cid:20) sin2 θ sin θ cos θ sin θ cos θ cos2 θ (cid:21) . C∗-ALGEBRA DISTANCE FILTERS 21 For ǫ ≥ 0, consider the compact projections pǫ represented by pǫ(x) =(Pǫ sin(1/x) 1 if x > 0 if x = 0 (this is a projection in the atomic representation of A rather than the universal representation A∗∗ but again this does not matter as the atomic representation is faithful on open and closed projections, by [Ped79, Theorem 4.3.15]). So p(x) is a rank 1 projection which 'wiggles' with amplitude ǫ and increasing frequency as x → 0. This means that, for ǫ > 0, any a ∈ A with a ≤ pǫ must satisfy a(0) = 0 so ǫ (x) =(Pr sin(1/θ) 0 p◦ if x > 0 if x = 0. Also let p be the compact projection defined by p(x) = P0, for all x ∈ [0, 1], so 0 = p. For all ǫ > 0, c(p, pǫ) ≥ kp(0) − p(0)p◦ p◦ ǫ (0)k = kP0k = 1 even though as ǫ → 0. (cid:3) (c ◦ e)(p, pǫ) ≤ c(p, p0) + e(p0, pǫ) = e(p0, pǫ) = kP0 − Pǫk → 0, In other words, c fails to be e-invariant in a strong way. This suggests that the 'reverse Hausdorff distance' inf a∈F supb∈G d(a, b) may actually be the more natural extension of interior containment to non-commutative A. This is especially so if one wants to consider d-filters in a domain theoretic way -- see [Bic17]. We finish by showing that this distance can also be calculated from h. Theorem 3.10. For any d-filters F, G ⊆ A1 +, d(a, b) = inf a∈F inf a∈F sup b∈G sup b∈G h(a, b). Proof. As in the proof of Theorem 3.2, for q =V G ∈ A∗∗ we have Let aS ∈ A∗∗ denote the spectral projection of a ∈ A1 and consider d(a, b) = inf a∈F h(a, b) = inf a∈F d(a, q) sup b∈G sup b∈G inf a∈F inf a∈F and h(a, q). + corresponding to S ⊆ [0, 1] Note P and F are coinitial in each other, with respect to both d and h, i.e. P = {a[1−ǫ,1] : a ∈ F and ǫ > 0}. 0 = sup a∈F inf p∈P d(p, a) = sup p∈P inf a∈F d(p, a) = sup a∈F inf p∈P h(p, a) = sup p∈P inf a∈F h(a, p). Thus inf a∈F Likewise inf a∈F h(a, q) = inf p∈P a∈F,p∈P (d(a, p) + d(p, q)) = inf p∈P d(a, q) ≤ inf ≤ inf h(p, q). Now simply note that, by Corollary 1.9, d(p, a) + d(a, q) = inf a∈F d(a, q). d(p, q) p∈P,a∈F inf p∈P d(p, q) = inf p∈P h(p, q). (cid:3) 22 TRISTAN BICE AND ALESSANDRO VIGNATI References [10113] User 1015. unitarily If two equivalent. projections are close, then are http://mathoverflow.net/questions/151439/if-two-projections-are-close-then-they-are-unitarily-equivalent. MathOverflow, 2013. they URL: [AAP86] Charles A. Akemann, Joel Anderson, and Gert K. Pedersen. Excising states of C ∗- algebras. Canad. J. Math., 38(5):1239 -- 1260, 1986. doi:10.4153/CJM-1986-063-7. [AAP89] Charles A. Akemann, Joel Anderson, and Gert K. Pedersen. Approaching infinity in C ∗-algebras. J. Operator Theory, 21(2):255 -- 271, 1989. [AB15] Charles A. Akemann and Tristan Bice. Hereditary C∗-subalgebra lattices. Adv. Math., 285:101 -- 137, 2015. doi:10.1016/j.aim.2015.07.027. [Ake70] Charles A. Akemann. Approximate units and maximal abelian C ∗-subalgebras. Pacific J. Math., 33:543 -- 550, 1970. [Ake71] Charles A. Akemann. A Gelfand representation theory for C ∗-algebras. Pacific J. Math., 39:1 -- 11, 1971. URL: http://projecteuclid.org/euclid.pjm/1102969765. [AP92] Charles A. Akemann and Gert K. Pedersen. Facial structure in operator algebra theory. Proc. London Math. Soc. (3), 64(2):418 -- 448, 1992. doi:10.1112/plms/s3-64.2.418 . [Bic13a] Tristan Bice. Filters in C-algebras. Canad. J. Math., (3):485 -- 509, 2013. doi:10.4153/CJM-2011-095-4. [Bic13b] Tristan Bice. The projection calculus. Munster J. Math., 6:557 -- 581, 2013. URL: http://wwwmath.uni-muenster.de/mjm/vol_6/mjm_vol_6_17.pdf . [Bic15] Tristan Bice. Extending akemann's non-commutative urysohn lemma. MathOverflow, 2015. URL: http://mathoverflow.net/questions/222116/extending-akemanns-non-commutative-urysohn-lemma. [Bic16] Tristan Bice. Semicontinuity in ordered Banach spaces, 2016. arXiv:1604.03154. [Bic17] Tristan Bice. Distance domains, 2017. arXiv:1704.01024. [BK17] Tristan Bice and Piotr Koszmider. ≪-Unital C-algebras. 2017. (in preparation). [Bla17] Bruce Blackadar. Operator algebras: Theory of C-algebras and von Neumann algebras. 2017. URL: http://wolfweb.unr.edu/homepage/bruceb/Cycr.pdf. [Bro88] Lawrence G. Brown. Semicontinuity and multipliers of C*-algebras. Canad. J. Math., 40(4):865 -- 988, 1988. doi:10.4153/CJM-1988-038-5. [Bro15] Lawrence G. Brown. Some directed subsets of C ∗-algebras and semicontinuity theory. Proc. Amer. Math. Soc., 143(9):3895 -- 3899, 2015. doi:10.1090/proc12744. [FW13] and Nik Weaver. Quantum measurable [BW16] David P. Blecher arXiv:1607.08505. Ilijas In Appalachian ciety https://www.math.cmu.edu/~ eschimme/Appalachian/FarahWofseyNotes.pdf. Series. Cambridge University Lecture Note Eric Wofsey. Set Theory: 2006-2012, operator Set theory and Farah and Press, London Mathematical cardinals, 2016. algebras. So- 2013. URL: [MSS15] Adam W. Marcus, Daniel A. Spielman, and Nikhil Srivastava. Interlacing families II: Mixed characteristic polynomials and the Kadison-Singer problem. Ann. of Math. (2), 182(1):327 -- 350, 2015. doi:10.4007/annals.2015.182.1.8. [OrT11] Eduard Ortega, Mikael Rørdam, semigroup and comparison of open projections. J.Funct. Anal., 260(12):3474 -- 3493, 2011. doi:10.1016/j.jfa.2011.02.017 . and Hannes Thiel. The Cuntz [Ped79] Gert K. Pedersen. C ∗-algebras and their automorphism groups, volume 14 of London Mathematical Society Monographs. Academic Press Inc. [Harcourt Brace Jovanovich Publishers], London, 1979. Institute of Mathematics of the Polish Academy of Sciences, Warsaw, Poland E-mail address: [email protected] URL: http://www.tristanbice.com Department of Mathematics and Statistics, York University, Toronto, Ontario, Canada E-mail address: [email protected] URL: http://www.automorph.net/avignati
1108.0049	1	1108	2011-07-30T07:36:40	Free Products of Generalized RFD C*-algebras	[ "math.OA" ]	If $k$ is an infinite cardinal, we say a C-algebra $\mathcal{A}$ is residually less than $k$ dimensional, $R_{<k}D,$ if the family of representations of $\mathcal{A}$ on Hilbert spaces of dimension less than $k$ separates the points of $\mathcal{A}.$ We give characterizations of this property, and we show that if $\{\mathcal{A}_{i}:i\in I\} $ is a family of $R_{<k}D$ algebras, then the free product $\underset{i\in I}{\ast}\mathcal{A}_{i}$ is $R_{<k}D$. If each $\mathcal{A}_{i}$ is unital, we give sufficient conditions, depending on the cardinal $k$, for the free product $\underset{i\in I}{\ast_{\mathbb{C}}}\mathcal{A}_{i}$ in the category of unital C-algebras to be $R_{<k}D$. We also give a new characterization of RFD, in terms of a lifting property, for separable C*-algebras.	math.OA	math	Free Products of Generalized RFD C-algebras Don Hadwin June 21, 2018 Abstract If k is an infinite cardinal, we say a C-algebra A is residually less than k dimensional, R<kD, if the family of representations of A on Hilbert spaces of dimension less than k separates the points of A. We give char- acterizations of this property, and we show that if {Ai : i ∈ I} is a family of R<kD algebras, then the free product ∗ Ai is R<kD. If each Ai is i∈I unital, we give sufficient conditions, depending on the cardinal k, for the Ai in the category of unital C-algebras to be R<kD. We free product ∗C i∈I also give a new characterization of RFD, in terms of a lifting property, for separable C-algebras. 1 Introduction A C-algebra A is residually finite dimensional ( RF D ) if the collection of all finite-dimensional representations of A separate the points of A; equivalently, if there is a direct sum of finite-dimensional representations of A with zero kernel. It is clear that every commutative C-algebra is RFD. Man-Duen Choi [4] showed that free group C-algebras are RFD. Ruy Exel and Terry Loring [6] proved that the free product of two RFD algebras is RFD. The class of RFD C-algebras plays an important role in the theory of C-algebras, e.g., [1], [2], [3], [4], [5], [6], [7], [11], [10]. In this paper we introduce a related notion. Suppose k is an infinite cardinal. We say that a C-algebra A is residually less than k-dimensional, conveniently denoted by R<kD, if the class of representations of A on Hilbert spaces of dimension less than k separates the points of A; equivalently, if there is a direct sum of such representations that has zero kernel. Note that when k = ℵ0, we have R<kD is the same as RF D. We give characterizations of R<kD algebras that show that the free product of an arbitrary collection of R<kD C-algebras is R<kD. We also give conditions that ensure that the free product (amalgamated over C) of unital C-algebras in the category of unital C-algebras is R<kD; this always happens when each of the algebras has a one-dimensional unital representation. 1 The proofs rely on a simple result (Lemma 1) and results of the author [8], [9] on approximate unitary equivalence and approximate summands of nonsep- arable representations of nonseparable C-algebras. Suppose k and m are infinite cardinals. We say that a C-algebra A is m- generated if is generated by a set with cardinality at most m. For each cardinal s, we let Hs be a Hilbert space whose dimension is s. If π : A → B (H) is a ∗-homomorphism, we say that the dimension of π is dim π = dim H. We define Repk (A) to be the set of all representations π : A → B (Hs) for some s < k. If A is a C-algebra, then A+ denotes the C-algebra obtained by adding a unit to A (that is different from the unit in A if A is unital). We end this section with our key lemma. Suppose H is a Hilbert space and P is a projection in B (H). We define MP = P B (H) P . Then MP is a unital C- algebra, but the unit is P , not 1. However, MP is a C-subalgebra of B (H). A unitary element of MP is an operator U ∈ B (H) such that U U ∗ = U ∗U = P , and is the direct sum of a unitary operator on P (H) with 0 on P (H)⊥. If P 6= 1, a unitary operator in MP is never unitary in B (H). We use the symbol ∗-SOT to denote the ∗-strong operator topology. Lemma 1 Suppose {Pα} is a net of projections in B (H) such that Pα → 1 ( ∗-SOT ) and let B =({Tα} ∈Yα MPα : ∃T ∈ B (H) , Tα → T ( ∗ -SOT )) , and J = {{Tα} ∈ B : Tα → 0 ( ∗ -SOT )} , and define π : B → B (H) by π ({Tα}) = ( ∗ -SOT)- lim α Tα. Then 1. B is a unital C-algebra, 2. J is a closed two-sided ideal in B, 3. If T ∈ B, then π ({PαT Pα}) = T, 4. π is a unital surjective ∗-homomorphism 5. If U ∈ B (H) is unitary, then there is a unitary {Uα} ∈ B such that π ({Uα}) = U. 2 Proof. Statements (1)-(4) are easily proved. To prove (5), note that if U ∈ B (H) is unitary, then there is an A = A∗ ∈ B (H) such that U = eiA. We can easily choose Aα = A∗ α for each α so that π ({Aα}) = A. Thus, if Uα = eiAα (in MPα), then {Uα} is unitary in B and π ({Uα}) = U . Here is a simple application that gives the flavor of our results. Corollary 2 Every free group is RFD. Proof. Suppose F is a free group and A = C∗ (F) = C ∗ ({Ug : g ∈ F}) . Choose a Hilbert space H and a faithful representation ρ : A → B (H). Choose a net {Pα} of finite-rank projections such that Pα → 1 ( ∗-SOT ). Applying Lemma 1 we have, for each g ∈ F, we can find a unitary element {Ug,α} in B so that π ({Ug,α}) = Ug. For each α, we have a unitary group representation σα : F → MPα defined by σα (g) = Ug,α. By the definition of C ∗ (F), there is a ∗-homomorphism τα : A → Mα such that τα (Ug) = Ug,α. It follows that τ : A → B define by τ (Ug) = {Ug,a} is a ∗-homomorphism such that π ◦ τ = ρ. Hence the direct sum of the τα's is faithful, which shows that A is RF D. The following corollary is from [3, Exercise 7.1.4]. Corollary 3 Every C-algebra is a ∗-homomorphic image of an RFD C-algebra. Proof. Suppose A is a C-algebra. We can assume that A ⊆ B (H) for some Hilbert space H. Choose a net {Pα} of finite-rank projections converging ∗- strongly to 1, and let B,J and π be as in Lemma 1. Then B, and thus π−1 (A), is RF D and π(cid:0)π−1 (A)(cid:1) = A. 2 R<kD Algebras We now prove our main results on R<kD C-algebras. The following two lemmas contain the key tools. Lemma 4 Suppose ℵ0 ≤ k ≤ m, and A is R<kD and m-generated. Then ⊕ 1. We can write Hm = Xλ with CardΛ = m, and such that, for every λ ∈ Λ, dim Xλ < k and there is a unital representation πλ : A+ → B (Xλ) Xπλ is such that the representation π : A+ → B (Hm) defined by π = faithful. Moreover, this can be done so that, for each λ0 ∈ Λ, we have Card ({λ ∈ Λ : πλ ≈ πλ0}) = m. ⊕ Xλ∈Λ 3 2. It is possible to choose the decomposition in (1) so that, for each cardinal s < k, there is a λ ∈ Λ such that dim Xλ = s. Proof. Since A is R<kD, there is a direct sum of representations in Repk (A) whose direct sum is faithful. Suppose D is a generating set for A and Card (D) ≤ m. We can replace D by the ∗-algebra over Q + iQ generated by D without making the cardinality exceed m. For each a ∈ D we can choose a direct sum of countably many summands from our faithful direct sum that preserves the norm of a. Hence, by choosing ℵ0Car (D) summands, we get a direct sum that is isometric on D and thus isometric on A. Since ℵ0Car (D) ≤ m. we can replace this last direct sum with a direct sum of m copies of itself and get a direct sum on a Hilbert space with dimension m. We can replace this Hilbert space with Hm and get a decomposition as in (1). to get (2) note that, since A+ has a unital one-dimensional representation, we know that, for every cardinal s < k. there is a representation of A+ of dimension s. If we take one such representation for each s < k and take a direct sum of m copies of all of them, we get a representation that has has dimension at most m, so we add this as a summand to the representation we constructed satisfying (1) . Lemma 5 Suppose A is a C-algebra and k ≤ m are infinite cardinals and D Xπλ as is a generating set for A. Suppose we can write Hm = in part (1) of Lemma 4. If ρ : A+ → B (Hm) is a unital representation, then, for every ε > 0, every finite subset W ⊆ D and every finite subset E ⊆ Hm, there is a finite subset F ⊆ Λ, such that, for every finite set G with F ⊆ G ⊆ Λ, Xλ, then there is a unitary U ∈ if QG is the orthogonal projection onto Xλ and π = ⊕ ⊕ Xλ∈Λ ⊕ QGB (Hm) QG such that, for every a ∈ W and e ∈ E, we have Xλ∈G G Xλ∈G "ρ (a) − U ∗ k[ρ (a) − U ∗ Gπ (a) UG] ek =(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) < ε. πλ! (a) UG# e(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Proof. It follows that if a ∈ A and a 6= 0, then rankπ (a) = m = rank (π ⊕ ρ) (a). Hence, by [8], π is approximately unitarily equivalent to π ⊕ ρ. However, by [9], ρ is a point-∗-SOT limit of representations unitarily to ρ. Hence there is a net {Uα} of unitary operators in B (Hm) such that, for every a ∈ A, ( ∗ -SOT) lim α U ∗ απ (a) Uα = ρ (a) . However, the net {QF : F ⊆ Λ, F is finite} is a net of projections converging ∗-strongly to 1. Hence, by Lemma 1, each Uα is a ∗-SOT limit of unitaries in the union of QF B (Hm) QF (F ⊆ Λ, F is finite). The result now easily follows. 4 Theorem 6 Suppose ℵ0 ≤ k ≤ m, and A is m-generated with a generating set G with CardG ≤ m. The following are equivalent. 1. A is R<kD. 2. There is a faithful unital ∗-homomorphism ρ : A+ → B (Hm) such that, for every ε > 0, every finite subset E ⊆ Hm and every finite subset W ⊆ G, there is a projection P ∈ B (Hm) and a unital ∗-homomorphism τ : A → MP = P B (Hm) P such that, for every e ∈ E and every a ∈ W we have k[τ (a) − ρ (a)] ek < ε. 3. There is a faithful unital representation ρ : A+ → B (Hm) and a net {Pα} of projections in B (Hm), each with rank less than k, such that Pα → 1 (∗- SOT) and such that, for each α, there is a representation πα : A → MPα such that, for every a ∈ A, we have πα (a) → ρ (a) ( ∗ -SOT). 4. For every unital representation ρ : A+ → B (Hm) there is a net {Pα} of projections in B (Hm), each with rank less than k, such that Pα → 1 (∗- SOT) and such that, for each α, there is a representation πα : A → MPα such that, for every a ∈ A, we have πα (a) → ρ (a) ( ∗ -SOT). Proof. (2) =⇒ (1) Let A be the set of triples (ε, E, W ) ordered by (≥,⊆,⊆). If α = (ε, E, W ) let τα : A → PαB (Hm) Pα guaranteed by (2). Since G = G∗ we have ( ∗ -SOT) lim α τα (a) = ρ (a) for every a ∈ G. Since ρ and each τα is a ∗-homomorphism, the set of a ∈ A for which (∗-SOT)limα τα (a) = ρ (a) is a unital C-algebra and is thus A+. Hence, for every a ∈ A+, we have kak = kρ (a)k ≤ sup{kτα (a)k : α ∈ A} . Therefore the direct sum of the τα's is faithful and (1) is proved. (3) =⇒ (2). This is obvious. (4) =⇒ (3). It is clear that we need only show that there is a faithful unital representation ρ : A+ → B (Hm). Suppose τ : A+ → B (M ) is an irreducible representation, and suppose D is a generating set with Card (D) ≤ m. Let A0 be the unital ∗-subalgebra of A+ over the field Q + iQ of complex rational numbers. Then A0 is norm dense in A and CardA0 = CardD ≤ m. Suppose f ∈ M is a unit vector. Since τ is irreducible, τ (A0) f must be dense in M . Suppose B is an orthonormal basis for M , and, for each e ∈ B let Ue be the 5 open ball centered at e with radius √2/2. Each Ue must intersect the dense set τ (A0) f , and since the collection {Ue : e ∈ B} is disjoint, we conclude that dim M = CardB ≤ Cardτ (A0) f ≤ Card (A0) ≤ m. ⊕ ⊕ We know that for every x ∈ A0 there is an irreducible representation τx : A+ → Mx ≤ m · m = m, there is a B (Mx) such that kτx (x)k = kxk. Since dim representation ρ : A+ → B (Hm) that is unitarily to a direct sum of m copies of Xx∈A0 τx. Hence ρ is isometric on the dense subset A0, which implies ρ is faithful. Xλ∈Λ (1) =⇒ (3). Since A is R<kD, we can choose a decomposition Hm = Xx∈A0 Xλ ⊕ ⊕ and representation π = Lemma 5. Xλ∈Λ πλ as in part (1) of Lemma 4. Now (3) follows from We see that the class of R<kD algebras is closed under arbitrary free prod- ucts in the nonunital category of C-algebras. Theorem 7 Suppose k is an infinite cardinal and {Aι : i ∈ I} is a family of R<kD C-algebras. Then the free product ∗i∈IAi is R<kD. Proof. Choose an infinite cardinal m ≥ k + Xi∈I generated by G = (cid:20)Si∈IAi(cid:21)\ {0} ⊆ ∗i∈IAi, clearly ∗i∈IAi is m-generated. Choose a set Λ with Card (Λ) = m and let S be the set of cardinals less than k. Write Card (Ai). Since ∗i∈IAi is ⊕ Hm =X s∈SXλ∈Λ Xs,λ where dim Xs,λ = s for every s ∈ S and λ ∈ Λ. It follows that, for each i ∈ I, we can find a representation πi : Ai → B (Hm) such that ⊕ πi =X s∈SXλ∈Λ πi s,λ satisfying (1) and (2) of Lemma 4. Suppose ε > 0, E ⊆ Hm is finite and W ⊆ G is finite. We can write W as a disjoint union of Wi1 , . . . , Win with Wi = W ∩ Ai. Let ρi be the restriction of ρ to Ai. Applying Lemma 5 to Aij and ρij and πij for 1 ≤ j ≤ n, we can find one finite subset G ⊆ S × Λ so that if P is the projection on X Xs,λ, then there are unitary operators (s,λ)∈G ⊕ 6 Ui1, . . . , Uin ∈ MP = P B (Hm) P so that, for 1 ≤ j ≤ n, a ∈ Wj, e ∈ E, we have Define τij : A+ ij → MP by (cid:13)(cid:13)(cid:2)ρij (a) − U ∗ ij πij (a) Uij(cid:3) e(cid:13)(cid:13) < ε. τij (a) = U ∗ ijπij (a) Uij, and for i ∈ I\ {i1, . . . , in} define τi : Ai → MP by τi (a) = P πi (a) P . Then, by the definition of free product, there is a representation τ : ∗i∈IA+ i → MP such that τAi = τi for every i ∈ I. It follows that, for every e ∈ E and every a ∈ W , k[ρ (a) − τ (a)] ek < ε. It follows from part (2) of Lemma 6 that ∗i∈IAi is R<kD. Corollary 8 Suppose k is an infinite cardinal and {Aι : i ∈ I} is a family of R<kD C-algebras such that each Ai has a one-dimensional unital representa- tion. Then the unital free product ∗C i∈IAi is R<kD. Proof. This follows from the fact that if τi : Ai → C is a unital ∗-homomorphism for each i ∈ I, then ∗C i∈IAi is ∗-isomorphic to (cid:18) ∗i∈I ker τi(cid:19)+ . Without the condition on unital one-dimensional representations, the pre- n∈NMn (C) is not RF D ( = R<ℵ0 D ), ceding corollary is false. For example, ∗C even though each Mn (C) is RF D. The reason is that each unital represen- tation of the free product must be injective on each Mn (C) and must have infinite-dimensional range. call an infinite cardinal k a limit cardinal, if k is the supremum of all the cardinals less than k. However, there is something we can say about the general situation. If k is a limit cardinal, the cofinality of k is the smallest cardinal s for which there is a set E of cardinals less than k whose supremum is k. Clearly, the cofinality of k is at most k. If k is not a limit cardinal, then there is a cardinal s such that k is the smallest cardinal larger than s, and if E is a set of cardinals less than k, then sup (E) ≤ s < k. Theorem 9 Suppose k is an infinite cardinal and {Aι : i ∈ I} is a family of unital R<kD C-algebras. Then 1. If k is a limit cardinal and Card (I) is less than the cofinality of k, then the free product ∗C i∈IAi is R<kD. 7 i∈IAi is R<kD. 2. If k is not a limit cardinal, then the free product ∗C Proof. (1) . Choose m ≥ k +Xi∈I Card (Ai), and choose a set Λ with Card (Λ) = m. Using Lemma ?? we can, for each i ∈ I, find a faithful representation πi = Xλ∈Λ πλ,i so that dim πi = m and, for every i ∈ I and λ ∈ Λ, we have dim πλ,i < k. Since Card (I) is less than the cofinality of k, we have, for each λ ∈ Λ, a cardinal sλ < k such that supi∈I dim πλ,i ≤ sλ. If we replace each πλ,i with a direct sum of sλ copies of itself, we get a new decomposition which we will denote by the same names such that, for each i and each λ we have dim πλ,i = sλ. Hence we may write direct sum decompositions of the πi's with Xλ where dim Xλ = sλ for every respect to a common decomposition Hm = Xλ∈Λ λ ∈ Λ. The rest now follows as in the proof of Theorem 7. proof of part (1) with sλ = s for every λ ∈ Λ. (2) If k is not a limit cardinal, there is a largest cardinal s < k. Repeat the Remark 10 We cannot remove the condition on Card (I) in part (1) of The- orem ??. Suppose k is a limit cardinal and I is a set of cardinals less than k whose cardinality equals the cofinality of k and such that sup (I) = k. For each infinite cardinal m, choose a set Λm with cardinality m, and let Sm denote the universal unital C-algebra generated by {vλ : λ ∈ Λm} with the conditions 1. v∗ λvλ = 1 for every λ ∈ Λm, 2. vλ1 v∗ λ1 vλ2 v∗ λ2 = 0 for λ1 6= λ2 in Λm. Since Sm is m-generated, it follows that every irreducible representation of Sm is at most m-dimensional (see the proof of (4) =⇒ (3) in Theorem 6). Hence Sm is separated by m-dimensional representations. On the other hand, if π is a unital representation of Sm, then {π (vλ;v∗ λ) : λ ∈ Λm} is an orthogonal family of nonzero projections, which implies that the dimension of π is at least m.It follows that each Ss is R<kD for s ∈ I. However, any unital representation π of the free product ∗C s∈ISs must induce a unital representation of each Ss, so its dimension is at least sup I = k. Hence ∗C s∈ISs is not R<kD. 3 Separable RFD Algebras In this section we show that for a separable C-algebra being RFD is equivalent to a lifting property. 8 Suppose {e1, e2, . . .} is an orthonormal basis for a Hilbert space ℓ2, and, for each integer n ≥ 1, let Pn be the projection onto sp ({e1, . . . , en}). Let Mn = PnB(cid:0)ℓ2(cid:1) Pn for n ≥ 1, and, following Lemma 1, let ∞ B =({Tn} ∈ Mn : ∃T ∈ B(cid:0)ℓ2(cid:1) with Tn → T ( ∗ -SOT )) , Yn=1 J = {{Tn} ∈ B : Tn → 0 ( ∗ -SOT )} . and let Then, by Lemma 1, we have that B is a unital C-algebra, J is a closed ideal in B and π ({Tn}) = ( ∗ -SOT)- lim n→∞ Tn defines a unital surjective ∗-homomorphism from B to B (H) whose kernel is J . We can now give our characterization of RFD for separable C-algebras. Theorem 11 Suppose A is a separable C-algebra. The following are equiva- lent. 1. A is RFD 2. For every unital ∗-homomorphism ρ : A+ → B(cid:0)ℓ2(cid:1) there is a unital ∗- homomorphism τ : A+ → B such that π ◦ τ = ρ. Proof. The implication (2) =⇒ (1) is clear. (1) =⇒ (2). Suppose A = C∗ ({a1, a2, . . .}) is RF D and ρ : A+ → B(cid:0)ℓ2(cid:1) is a unital ∗-homomorphism. It follows from Theorem 6 that there is an increasing sequence {nk} of positive integers and unital ∗-homomorphisms τk : A → Mnk such that k[τk (aj) − ρ (aj)] eik < 1/k for 1 ≤ i, j ≤ k. It follows that τnk (a) → ρ (a) ( ∗-SOT ) for every a ∈ A+. If nk < n < nk+1 we define τn : A+→ Mn by τn (a) =  τnk (a) β (a) . . . β (a) ,   where β : A+ → C is the unique ∗-homomorphism with ker β = A, relative to the decomposition Pn(cid:0)ℓ2(cid:1) = Pnk(cid:0)ℓ2(cid:1) ⊕ Ce1+nk ⊕ ··· ⊕ Ce−1+nk+1. It is easily seen that τn (a) → ρ (a) ( ∗-SOT ) for every a ∈ A+. If we define τ : A → B by we see that π ◦ τ = ρ. τ (a) = {τn (a)} , 9 Acknowledgement 12 The author wishes to thank Tatiana Shulman and Terry Loring for bringing the question answered by Theorem 11 to his attention. Thinking about this question led to the discovery of all the results in this pa- per. References [1] R. J. Archbold, On residually finite-dimensional C-algebras, Proc. Amer. Math. Soc. 123 (1995) 2935-2937. [2] N. P. Brown, On quasidiagonal C-algebras, arXiv: math/0008181 (2000) [3] N. P. Brown and N. Ozawa, C-algebras and finite-dimensional approxi- mations,Graduate Studies in Math. 88, Amer. Math. Soc. (2008). [4] Man-Duen Choi, The full C-algebra of the free group on two generators, Pacific J. Math 87 (1980) 41-48. [5] M. Dardalat, On the approximation of quasidiagonal C-algebras, J. Funct. Anal. 167 (1999) 69-78. [6] R. Exel and T. Loring, Finite-dimensional representations of free product C-algebras, Internat. J. Math. 3 (1992) 46-476. [7] K. R. Goodearl and P. Menal, Free and residually finite-dimensional C- algebras, J. Funct. Anal. 90 (1990) 391-410. [8] D. Hadwin, Nonseparable approximate equivalence,Trans. Amer. Math. Soc. 266 (1981), no. 1, 203 -- 231. [9] D. Hadwin, An operator-valued spectrum, Indiana Univ. Math. J. [] [10] D. Hadwin, Qihui Li, Weihua Li, and Junhao Shen, MF-traces and topo- logical free entropy dimension, preprint. [11] Huaxin Lin, Residually finite-dimensional and AF-embeddable C*- algebras, Proc. Amer. Math. Soc.129 (2000) 1689-1696. 10
1706.02349	1	1706	2017-06-07T19:18:53	Connes' Embedding Problem and Winning Strategies for Quantum XOR Games	[ "math.OA", "quant-ph" ]	We consider quantum XOR games, defined in [11], from the perspective of unitary correlations defined in [7]. We show that Connes' embedding problem has a positive answer if and only if every quantum XOR game has entanglement bias equal to the commuting bias. In particular, the embedding problem is equivalent to determining whether every quantum XOR game $G$ with a winning strategy in the commuting model also has a winning strategy in the approximate finite-dimensional model.	math.OA	math	CONNES' EMBEDDING PROBLEM AND WINNING STRATEGIES FOR QUANTUM XOR GAMES SAMUEL J. HARRIS Abstract. We consider quantum XOR games, defined in [11], from the perspective of uni- tary correlations defined in [7]. We show that Connes' embedding problem has a positive answer if and only if every quantum XOR game has entanglement bias equal to the com- muting bias. In particular, the embedding problem is equivalent to determining whether every quantum XOR game G with a winning strategy in the commuting model also has a winning strategy in the approximate finite-dimensional model. 1. Introduction A long-standing open problem in the theory of von Neumann algebras, known as Connes' embedding problem [4], asks whether every weakly separable II1 factor can be ap- proximately embedded into the hyperfinite II1 factor in a way that approximately preserves the trace. Recently, several connections between the embedding problem and questions in quantum information theory have been exhibited. The most notable connection is the equiv- alence of Connes' embedding problem to the weak Tsirelson problem regarding probabilistic correlations in a separated system [5, 8, 9]. This problem asks whether every probabilistic quantum bipartite correlation in finite inputs and finite outputs in the commuting model can be approximated by correlations in the finite-dimensional model [13]. These quantum bipartite correlations naturally correspond to strategies that can be used in two-player, finite input-output non-local games. Such games have been instrumental in exhibiting separations between the probabilistic correlation sets in the various models. For example, the well-known CHSH game [2] is a famous example of a game for which the maximum winning probability of the game for Alice and Bob is higher if they have access to entanglement than if they play the game using classical methods. Recent work by W. Slofstra [12] shows that there is a non-local game with a winning strategy in the approximate finite- dimensional model, but no winning strategy in the finite-dimensional model. Therefore, there are input and output sets for which the set of correlations arising from the finite- dimensional model is not closed. Considering these advances, a natural question is whether the weak Tsirelson problem can be described in terms of a certain class of (possibly extended) non-local games and winning strategies in the commuting model and the approximate finite- dimensional model. In other words, is Connes' embedding problem equivalent to the assertion that a certain class of extended non-local games with winning commuting strategies must also have winning approximate finite-dimensional strategies? In this paper, we show that the answer to this question is affirmative, and that the class of quantum XOR games (defined in [11]) is rich enough to detect the solution to the embedding problem. In particular, determining whether every quantum XOR game with winning commuting strategy has a winning approximate finite-dimensional strategy is equivalent to Connes' embedding problem. If one considers the analogous weak Tsirelson 1 2 SAMUEL J. HARRIS problem related to unitary correlation sets [7], the coherent embezzlement games from [11] can be used to show that the set of unitary correlations in the finite-dimensional model are not closed [3, 7], as long as the unitaries have size at least 2. In light of these facts, it is plausible that studying quantum XOR games may be a reasonable plan of attack for solving the embedding problem. The paper is organized as follows. In Section 2, we give a brief overview of quantum XOR games from [11] and the notion of bias for these games. We also show the correspon- dence between bias and linear functionals on Mn ⊗ Mn that are contractive with respect to the unitary correlation norms from [7]. In Section 3, we use Lemma 5 to reduce the Tsirelson problem for unitary correlations to self-adjoint unitary correlations. This allows us to prove Corollary 7, which states the equivalence of the embedding problem to the problem of winning strategies for quantum XOR games in the commuting and approximate finite-dimensional models. 2. Preliminaries and Bob) satisfyingPs,t Rs,t = 1. The referee gives question s to Alice and question t to We briefly recall the definition of a quantum XOR game with two parties, Alice and Bob. More information can be found in [11]. Loosely speaking, a quantum XOR game is a generalization of a classical XOR game. In the classical case, the referee has a list of n possible questions {1, ..., n}, and the set of possible answers for Alice and Bob is {0, 1}. For each pair s, t ∈ {1, ..., n}, there is some associated number Rs,t ∈ [−1, 1] (known to Alice Bob with probability Rs,t. If Rs,t ≥ 0, then Alice and Bob must respond with the same bit; if Rs,t < 0, then they must respond with different bits. In a quantum XOR game, the questions are now given as states (i.e., unit vectors) on a certain Hilbert space. In particular, the referee sends some state on Cn ⊗ Cn to Alice and Bob, where Alice has access to the left copy of Cn and Bob has access to the right copy of Cn. Every quantum XOR game of size n is associated with a self-adjoint matrix M ∈ Mn ⊗ Mn with kMk1 ≤ 1, where k · k1 denotes the trace norm. Conversely, every self-adjoint matrix M ∈ Mn ⊗ Mn with kMk1 ≤ 1 is associated to a quantum XOR game G of size n [11]. For the sake of simplicity, we will always consider the case where kMk1 = 1. For our purposes, a quantum XOR game G of size n can be described as follows (see i=1 ⊆ Cn ⊗ Cn be an orthonormal basis. Let [11] for a more general definition): i=1 pi = 1, and let ci ∈ {0, 1} for each 1 ≤ i ≤ n2. With probability pi, the referee prepares the state ϕi ∈ Cn ⊗ Cn. Alice and Bob may use their own Hilbert space H and observables A ∈ B(Cn ⊗H) and B ∈ B(H⊗ Cn) (i.e., self-adjoint unitaries) such that A⊗ In and In ⊗ B commute in B(Cn ⊗H⊗ Cn). They may also prepare their space in the state ψ ∈ H. Based on the application of A and B to the state ψ, Alice and Bob return outcomes a ∈ {0, 1} and b ∈ {0, 1} respectively. If ci = 0, then Alice and Bob's output bits must be equal; if ci = 1, their output bits must be distinct. (If Alice and Bob are working in the tensor product model, then there must be a decomposition H = HA ⊗HB where A ∈ B(Cn ⊗HA) and B ∈ B(HB ⊗ Cn), and where ψ ∈ HA ⊗HB is a state. Moreover, the operator (A⊗ In)(In⊗ B) is replaced with A⊗ B.) The matrix M ∈ Mn ⊗ Mn associated i=1(−1)cipiϕiϕ∗i , where ϕiϕ∗i denotes the rank one orthogonal projection of Cn ⊗ Cn onto span {ϕi} [11]. Before we further consider possible strategies for Alice and Bob for quantum XOR games, we recall the definitions of some of the unitary correlation sets given in [7]. The set with this quantum XOR game is M =Pn2 p1, ..., pn2 ∈ [0, 1] be such that Pn2 let {ϕi}n2 CONNES' EMBEDDING PROBLEM AND QUANTUM XOR GAMES 3 of quantum correlations Bq(n, n) ⊆ Mn ⊗ Mn is given by the set of all X = (X(i,j),(k,ℓ)) ∈ Mn ⊗ Mn of the form X(i,j),(k,ℓ) = h(Uij ⊗ Vkℓ)ψ, ψi, where U = (Uij) ∈ Mn(B(HA)) and V = (Vkℓ) ∈ Mn(B(HB)) are unitary, HA and HB are finite-dimensional Hilbert spaces, and ψ ∈ HA ⊗ HB is a unit vector. The set of quantum spatial correlations Bqs(n, n) ⊆ Mn ⊗ Mn is defined similarly, only that we no longer assume that HA and HB are finite-dimensional. The set of quantum approximate cor- relations is given by Bqa(n, n) = Bq(n, n) = Bqs(n, n). The set of quantum commuting correlations is given by the set of all X ∈ Mn ⊗ Mn of the form X = (hUijVkℓψ, ψi)(i,j),(k,ℓ), where U = (Uij) ∈ Mn(B(H)) and V = (Vkℓ) ∈ Mn(B(H)) are unitary, ψ ∈ H is a unit vector, and UijVkℓ = VkℓUij for all i, j, k, ℓ. Since U and V are unitary, it also follows that the set {Uij, U∗ij}n i,j=1 commutes with the set {Vkℓ, V ∗kℓ}n k,ℓ=1. We have the containments Bq(n, n) ⊆ Bqs(n, n) ⊆ Bqa(n, n) ⊆ Bqc(n, n), ∀n ≥ 2, and the latter two sets are compact in Mn ⊗ Mn [7]. In fact, for t ∈ {qa, qc}, the set Bt(n, n) is the unit ball of a reasonable cross-norm on Mn ⊗ Mn [7, Theorem 4.8]; we will denote this norm by k· kt, and we will denote the Banach space (Mn ⊗ Mn,k· kt) by Mn ⊗t Mn. Finally, we will let k · k∗t denote the dual norm of k · kt on (Mn ⊗ Mn)∗. For a quantum XOR game G and t ∈ {q, qa, qc}, we define a t-strategy for Alice and Bob to be a correlation X ∈ Bt(n, n). Instead of working with maximum success probabilities in different models, it is con- venient to work with a related quantity, known as the bias. To ease notation, whenever H and K are Hilbert spaces and H is finite-dimensional, we will denote by TrH the operator Tr ⊗ idK acting on B(H ⊗ K), where Tr denotes the unnormalized trace on B(H). With this in hand, the entanglement bias (or the quantum bias) of a quantum XOR game G with associated matrix M is given by ω∗q (G) = sup{hTrCn⊗Cn[(A ⊗ B)(M ⊗ IHA⊗HB )]ψ, ψi}, where the supremum is taken over all finite-dimensional Hilbert spaces HA and HB, unit vectors ψ ∈ HA ⊗ HB, and observables A ∈ B(Cn ⊗ HA) and B ∈ B(HB ⊗ Cn) (that is, self-adjoint unitaries) [11]. In the supremum above, we are identifying M ⊗ IHA⊗HB with the matrix M′ ∈ Mn2(B(HA ⊗ HB)) given by M′ = (M(i,j),(k,ℓ)IHA⊗HB )(i,j),(k,ℓ). Similarly, we define the commuting bias of a quantum XOR game G with associated matrix M to be ω∗qc(G) = sup{hTrCn⊗Cn[((A ⊗ ICn)(ICn ⊗ B)(M ⊗ IH)]ψ, ψi, where the supremum is taken over all Hilbert spaces H, unit vectors ψ ∈ H, and self-adjoint unitaries A = (Aij) ∈ B(Cn ⊗ H) and B = (Bkℓ) ∈ B(H ⊗ Cn) such that (A ⊗ In)(In ⊗ B) = (In ⊗ B)(A ⊗ In) as operators on Cn ⊗ H ⊗ Cn. Here, we are identifying M ⊗ IH with the operator M′ ∈ Mn2(B(H)) given by M′ = (M(i,j),(k,ℓ)IH). Adapting the proof of [3, Proposition 3.1], since the matrix (A ⊗ In)(In ⊗ B) is given by (AijBkℓ)(i,j),(k,ℓ), it follows that A⊗In commutes with In⊗B if and only if AijBkℓ = BkℓAij for all i, j, k, ℓ. In particular, since A = A∗ and B = B∗, we have Aij = A∗ji and Bkℓ = B∗ℓk. Thus, the self-adjoint unitaries 4 SAMUEL J. HARRIS i,j=1 k,ℓ=1. A and B satisfy (A ⊗ In)(In ⊗ B) = (In ⊗ B)(A ⊗ In) if and only if the set {I} ∪ {Aij}n ∗-commutes with the set {I} ∪ {Bkℓ}n One may view both of these notions of bias as twice the difference between the max- imum success probability in the corresponding model and the success probability from the random strategy (i.e. Alice and Bob respond randomly, regardless of the input). (The argu- ment from [11] for entanglement bias can be adapted in the obvious way for the commuting bias.) Thus, if p is the maximum success probability of winning a quantum XOR game G using t-strategies (for t ∈ {q, qc}), then ω∗t (G) = 2p − 1. We will also consider the above definitions of bias that arise from omitting the assump- tion that A and B are self-adjoint, while keeping the other assumptions intact. In particular, we may consider the notions of bias given with respect to the unitary correlation sets defined above. In this context, we also consider the bias of a particular strategy, which has the same definition but is denoted by ω∗t (G, U, V, ψ) for a specific strategy (U, V, ψ) whose correlation matrix X is in Bt(n, n). If X is the correlation associated with (U, V, ψ), then we also let ω∗t (G, X) = ω∗t (G, U, V, ψ). We note that ω∗t (G, U, V, ψ) is C-valued. A perfect t-strategy is a t-strategy X for which ω∗t (G, X) = 1. It will follow from Theorem 4 that there is a perfect t-strategy for the quantum XOR game G if and only if there is a t-strategy arising from observables for which the probability that Alice and Bob win is 1. We remark that there is a natural correspondence between bias for quantum XOR games and self-adjoint linear functionals on Mn ⊗ Mn. The easiest way to see this is using unitary correlation sets. Indeed, suppose that X = (hUijVkℓψ, ψi)(i,j),(k,ℓ) ∈ Bqc(n, n). Then since M is self-adjoint, ω∗qc(G, X) = hTrCn⊗Cn[(UijVkℓ)(i,j),(k,ℓ)(M(i,j),(k,ℓ)IH)(i,j),(k,ℓ)]ψ, ψi UipVkqM(p,j),(q,ℓ)! ψ, ψ+ =*TrCn⊗Cn nXp,q=1 = Xi,j,p,q = Xi,j,p,q = Xi,j,k,ℓ = Tr(M X). M(j,i),(ℓ,k)X(i,j),(k,ℓ), hUipVjqM(p,i),(q,j)ψ, ψi M(p,i),(q,j)hUipVjqψ, ψi Thus, the quantum XOR game G defines the linear functional Tr(M·) : Mn ⊗ Mn → C which gives ω∗qc(G, X) for each X ∈ Bqc(n, n). An analogous argument holds for ω∗q (G, X) whenever the corresponding correlation matrix X lies in Bq(n, n). A helpful fact is that for t ∈ {q, qs, qa, qc}, the self-adjoint t-correlations arise from t-strategies involving observables. Proposition 1. Let t ∈ {q, qs, qa, qc} and X = X∗ ∈ Bt(n, n). Then there is a Hilbert space H, self-adjoint unitaries U ∈ B(Cn ⊗ H) and V ∈ B(H ⊗ Cn) and a unit vector ψ ∈ H such that X(i,j),(k,ℓ) = hUijVkℓψ, ψi for all i, j, k, ℓ. If t = qs, then we may take H = HA ⊗HB and U ∈ B(Cn⊗HA) and V ∈ B(HB ⊗ Cn) such that X(i,j),(k,ℓ) = h(Uij ⊗ Vkℓ)ψ, ψi for all i, j, k, ℓ. Moreover, if t = q, then we may take HA and HB to be finite-dimensional. Finally, if t = qa, then X = limm→∞ Y (m), where Y (m) ∈ Bq(n, n) is a q-strategy involving observables. CONNES' EMBEDDING PROBLEM AND QUANTUM XOR GAMES 5 Proof. We first let t = q, so that X = (h(Rij ⊗ Skℓ)ψ, ψi)(i,j),(k,ℓ) for unitaries R = (Rij) ∈ B(Cn ⊗ HA), S = (Skℓ) ∈ B(HB ⊗ Cn), finite-dimensional Hilbert spaces HA and HB, and a unit vector ψ ∈ HA ⊗ HB. Let Uij = (cid:20) 0 Rij 0 (cid:21). Per- forming a canonical shuffle [10, p. 97] on the unitaries (cid:20) 0 R 0(cid:21) ∈ M2(Mn(B(HA))) and (cid:20) 0 S S∗ 0(cid:21) ∈ M2(Mn(B(HB))), we see that U = (Uij) and V = (Vkℓ) are self-adjoint unitaries (U, V,eψ), where eψ = 1√2(cid:2)ψ 0 0 ψ(cid:3)t in Mn(B(HA ⊕ HA)) and Mn(B(HB ⊕ HB)), respectively. Hence, we obtain a q-strategy 0 (cid:21) and Vkℓ = (cid:20) 0 . Using the fact that U∗ij = Uji and V ∗kℓ = Vℓk, R∗ji S∗ℓk Skℓ R∗ 1 2h(R∗ji ⊗ S∗ℓk)ψ, ψi = X(i,j),(k,ℓ). 1 2h(Rij ⊗ Skℓ)ψ, ψi + The proof for t = qs is similar. For t = qc, we assume that R ∈ B(Cn⊗H) and S ∈ B(H⊗Cn) are unitaries and ψ ∈ H is a unit vector such that (R ⊗ In)(In ⊗ S) = (In ⊗ S)(R ⊗ In) and X(i,j),(k,ℓ) = hRijSkℓψ, ψi for all i, j, k, ℓ. We let h(Uij ⊗ Vkℓ)eψ,eψi = Uij = UijVkℓ = 0 Rij R∗ji 0 0 0 0 R∗ji 0 0 0 0 0 0 Rij 0  and Vkℓ = 0 0 S∗ℓk 0 0 0 0 S∗ℓk Skℓ 0 0 0 0 Skℓ 0 0  . A calculation shows that 0 0 0 0 0 0 RijSkℓ  = VkℓUij, (U ⊗ In)(In ⊗ V ) = (In ⊗ V )(U ⊗ In). Letting eψ = 1√2(cid:2)ψ 0 0 ψ(cid:3)t so that U = (Uij) ∈ B(Cn ⊗ H) and V = (Vkℓ) ∈ B(H ⊗ Cn) are self-adjoint unitaries with , it readily follows that R∗jiS∗ℓk R∗jiSkℓ RijS∗ℓk 0 0 0 0 0 1 2hRijSkℓψ, ψi + 1 2hR∗jiS∗ℓkψ, ψi = X(i,j),(k,ℓ). hUijVkℓψ, ψi = 0 Thus, the proposition holds for t = qc. The last statement about qa correlations immediately follows from the t = q case. (cid:3) The next proposition, combined with convexity of each Bt(n, n) [7], guarantees that Re(X) ∈ Bt(n, n) whenever X ∈ Bt(n, n). Proposition 2. Let n ≥ 2 and X = (X(i,j),(k,ℓ))(i,j),(k,ℓ) ∈ Bt(n, n) for t ∈ {q, qs, qa, qc}. Then X∗ ∈ Bt(n, n). Proof. Suppose that U = (Uij) ∈ Mn(B(H)) and V = (Vkℓ) ∈ Mn(B(H)) are unitary and ψ ∈ H is a unit vector such that UijVkℓ = VkℓUij for all i, j, k, ℓ and Then hUijVkℓψ, ψi = X(i,j),(k,ℓ). X∗ = (X (j,i),(ℓ,k)) = (hUjiVℓkψ, ψi) = (hψ, UjiVℓkψi) = (hU∗jiV ∗ℓkψ, ψi), 6 SAMUEL J. HARRIS using the fact that U∗ijV ∗kℓ = V ∗kℓU∗ij for all i, j, k, ℓ. It follows that X∗ = (h(U∗)ij(V ∗)kℓψ, ψi)(i,j),(k,ℓ) ∈ Bqc(n, n). A similar argument gives the desired result when t ∈ {q, qs}. The same result follows for t = qa by using the density of Bq(n, n) in Bqa(n, n). Corollary 3. For n ≥ 2 and t ∈ {loc, q, qs, qa, qc}, let k · k∗t denote the dual norm on Mn ⊗ Mn with respect to the normed space Mn ⊗t Mn. Then k · kt is a ∗-norm; i.e., if kMk∗t ≤ 1, then kM∗k∗t ≤ 1. Proof. Let M ∈ Mn ⊗ Mn be such that kMkt ≤ 1. The linear functional on Mn ⊗ Mn with respect to M is given by (cid:3) If g is the linear functional on Mn ⊗ Mn with respect to M∗, then g((X(i,j),(k,ℓ))) = Tr(XM∗) = Tr(M X∗) = Tr(X∗M) = f (X∗). f ((X(i,j),(k,ℓ))) = Tr(XM). Since k · kt is a ∗-norm, it follows that kM∗kt ≤ 1, as required. Using Propositions 1 and 2, we obtain an equivalent description of bias, which allows us to use the theory of unitary correlations. Theorem 4. Let G be a quantum XOR game of size n with associated matrix M ∈ Mn⊗Mn. Then (cid:3) Proof. Since every observable is a self-adjoint unitary, it is clear that ω∗qc(G) is at least the quantity given in the theorem statement; thus, we need only establish the reverse inequality. Using the fact that Bqc(n, n) is compact [7], we may choose X ∈ Bqc(n, n) such that sup{Tr(M Z) : Z ∈ Bqc(n, n)} = Tr(M X). Suppose that X = hUijVkℓψ, ψi where U = (Uij) ∈ B(Cn ⊗ H) and V = (Vkℓ) ∈ B(H ⊗ Cn) are unitaries, ψ ∈ H is a unit vector and (U ⊗In)(In⊗V ) = (In⊗V )(U ⊗In). By multiplying the unitary U by some λ ∈ T if necessary, we may assume that Tr(M X) = ω∗qc(G, X) = sup{Tr(M Z) : Z ∈ Bqc(n, n)}. Since Bqc(n, n) is convex [7], it follows that Y := 1 Y is represented by self-adjoint observables. Finally, we see that 2 (X + X∗) ∈ Bqc(n, n). By Proposition 1, Similarly, we have ω∗qc(G) = sup{Tr(M X) : X ∈ Bqc(n, n)}. ω∗q (G) = sup{Tr(M X) : X ∈ Bq(n, n)}. ω∗qc(G, Y ) = Tr(cid:18)1 2 (X + X∗)M(cid:19) (Tr(XM) + Tr(X∗M)) (Tr(XM) + Tr(XM)) = = 1 2 1 2 = ω∗qc(G), using the fact that M is self-adjoint. This establishes the reverse inequality for the com- muting case. For the entanglement bias, the same argument shows that if X ∈ Bq(n, n) with ωq(G, X) = α ∈ [0, 1], then Y := 1 2 (X + X∗) ∈ Bq(n, n) satisfies ωq(G, Y ) = α. Using CONNES' EMBEDDING PROBLEM AND QUANTUM XOR GAMES 7 Proposition 1 and taking the supremum over all such strategies, the result holds for the entanglement bias. (cid:3) 3. Main Results In this section, we connect the embedding problem with commuting and entanglement bias for quantum XOR games. The first step is showing that, when considering Connes' em- bedding problem, it is enough to consider self-adjoint elements of Bqc(m, m) and Bqa(m, m) for all m ≥ 2. Lemma 5 allows for this reduction. Lemma 5. Let X ∈ Mn ⊗ Mn and t ∈ {qa, qc}. Then X ∈ Bt(n, n) if and only if W := 0 0 X 0 0 0 0 0 0 0 0 0 X∗ 0 0 0  ∈ Bt(2n, 2n). and R(a,b),(c,d) − W(a,b),(c,d) < ε for all 1 ≤ a, b, c, d ≤ 2n. Since Bqa(2n, 2n) = Bq(2n, 2n), we see that W ∈ Bqa(2n, 2n). Proof. Suppose that t = qc and X ∈ Bqc(n, n). Then there are unitaries U = (Uij), V = (Vkℓ) ∈ Mn(B(H)) and a vector ψ ∈ H of norm 1 such that, for all i, j, k, ℓ, we have UijVkℓ = VkℓUij and X(i,j),(k,ℓ) = hUijVkℓψ, ψi. (Uij) X 0 0 0 With Z = (UijV ∗ℓkψ, ψ) ∈ Bqc(n, n), we see by Proposition 2 that hU∗jiVkℓψ, ψi hUijV ∗ℓkψ, ψi 0 0 0 X∗ 0 0 0 0 0 0 (Vkℓ) (U∗ji) (V ∗ℓk) W ′ ∈ Bqc(2n, 2n), where Let eU = (cid:18) 0 0 (cid:19) ∈ M2n(B(H)), which are unitary. The entries of eU commute with the entries of eV , so with eU , eV and ψ, we obtain  .  ∈ Bqc(2n, 2n). 0 (cid:19) ∈ M2n(B(H)) and eV = (cid:18) 0 W ′ = WZ := W ′ = A similar argument using the unitaries i(Uij) and −i(Vkℓ) shows that W−Z ∈ Bqc(2n, 2n). By convexity, we obtain W = 1 2(WZ + W−Z) ∈ Bqc(2n, 2n). If t = qa and ε > 0, then there is Y ∈ Bq(n, n) such that X(i,j),(k,ℓ) − Y(i,j),(k,ℓ) < ε for all i, j, k, ℓ. The above argument shows that 0 0 X 0 Z 0 0 0 0 0 0 Z∗ X∗ 0 0 0 R = 0 0 Y 0 0 0 0 0 0 0 0 0 Y ∗ 0 0 0  ∈ Bq(2n, 2n), 8 SAMUEL J. HARRIS Conversely, suppose that W = 0 0 X 0 0 0 0 0 0 0 0 0 X∗ 0 0 0  is in Bq(2n, 2n). Let HA and HB be finite-dimensional Hilbert spaces, U ∈ M2n(B(HA)) and V ∈ M2n(B(HB)) be unitaries, and ψ ∈ HA ⊗ HB be a unit vector such that for all 1 ≤ a, b, c, d ≤ 2n, we have W(a,b),(c,d) = hUab ⊗ Vcdψ, ψi. Let i,j=1 ∈ Mn(B(HA)) and T = (Vk,n+ℓ)n k,ℓ=1 ∈ Mn(B(HB)). S = (Ui,(j+n))n Then S and T are contractions. Applying the Halmos dilation and performing a canonical shuffle, we obtain unitaries eU ∈ Mn(B(H(2) Sij B )), where (√I − SS∗)ij A )) and eV ∈ Mn(B(H(2) (√I − S∗S)ij (cid:21) (√I − T ∗T )kℓ (cid:21) . −S∗ji −T ∗ℓk Tkℓ (√I − T T ∗)kℓ A ⊗ H(2) B gives eUij =(cid:20) eVkℓ =(cid:20) ∈ H(2) and similarly Taking eψ =(cid:2)ψ 0 0 0(cid:3)t X = (h(Sij ⊗ Tkℓ)ψ, ψi)(i,j),(k,ℓ) = (h(eUij ⊗eVkℓ)eψ,eψi)(i,j),(k,ℓ) ∈ Bq(n, n). Since Bq(m, m) is dense in Bqa(m, m) for all m ≥ 2, the converse follows for t = qa. Finally, assume that W ∈ Bqc(2n, 2n), and let U, V ∈ M2n(B(H)) be unitaries and ψ ∈ H be a unit vector such that W(a,b),(c,d) = hUabVcdψ, ψi for all 1 ≤ a, b, c, d ≤ 2n. As before, let S = (Ui,(j+n))n k,ℓ=1. We use an argument similar to the proof of [6, Proposition 4.6]. First, we let i,j=1 and T = (Vk,n+ℓ)n (√I − T ∗T )kℓ −T ∗ℓk (cid:21) ∈ B(H(2)). Cij =(cid:20)Sij 0 0 Sij(cid:21) ∈ B(H(2)) and Dkℓ =(cid:20) (√I − T T ∗)kℓ i,j=1 commutes with the set {Tij, T ∗ij}n Tkℓ Since the set {Sij, S∗ij}n i,j=1, it follows that, by examining polynomials in T and T ∗, the set {Sij, S∗ij}n i,j=1 commutes with each entry of Dkℓ and D∗kℓ. Therefore, the set {Cij, C∗ij}n k,ℓ=1, while C = (Cij) is a contraction and D = (Dkℓ) is a unitary. Performing a similar dilation on C and replacing Dkℓ i,j=1 commutes with the set {Dkℓ, D∗kℓ}n 0 Dkℓ(cid:21), we obtain unitaries A = (Aij) and B = (Bkℓ) in Mn(B(H(4))) such that the with(cid:20)Dkℓ (1, 1)-block of Aij is Sij and the (1, 1)-block of Bkℓ is Tkℓ. Letting eψ =(cid:2)ψ 0 0 0(cid:3)t ∈ H(4), we see that 0 which completes the proof. (cid:3) The following theorem shows that it is enough to consider self-adjoint elements of X = (hAijBkℓeψ,eψi)(i,j),(k,ℓ) ∈ Bqc(n, n), Mn ⊗ Mn for the embedding problem. Theorem 6. The following are equivalent. (1) Connes' Embedding Problem has a positive answer. (2) Bqa(n, n) = Bqc(n, n) for all n ≥ 2. CONNES' EMBEDDING PROBLEM AND QUANTUM XOR GAMES 9 (3) Mn ⊗qa Mn = Mn ⊗qc Mn isometrically for all n ≥ 2. (4) For every n ≥ 2 and X = X∗ ∈ Mn ⊗ Mn with kXkqc = kXkMn⊗minMn = 1, we have (5) For every n ≥ 2 and M = M∗ ∈ Mn ⊗ Mn such that kMk1 = kMk∗qc = 1, we have kXkqa = 1. kMk∗qa = 1. Proof. The equivalence of (1), (2) and (3) is proven in [7]. Clearly 3 implies 4 and 5. We will show that (5) implies (2); the proof that (4) implies (2) is similar. Let A ∈ Mn⊗ Mn. By the proof that (2) implies (1), it suffices to know that the following holds for all n ≥ 2: whenever Y ∈ Bqc(n, n) is diagonal with diagonal entries τ (uiu∗j ) for some unitaries u1, ..., un in a unital C∗-algebra A and a tracial state τ on A, we have that Y ∈ Bqa(n, n). In particular, there are entries in such Y equal to 1. Therefore, if k · k denotes the operator norm on Mn ⊗ Mn, then 1 ≤ kY k ≤ kY kqc = 1. Now, assume that (2) fails. Then there is Y ∈ Bqc(n, n) with operator norm 1 such that Y 6∈ Bqa(n, n). By Lemma 5, there is n ≥ 2 and some X = X∗ ∈ M2n ⊗ M2n such that kXkqc = kXk = 1 but kXkqa > 1. By the Hahn-Banach Theorem, we have 1 = sup{g(X) : g ∈ (M2n ⊗ M2n)∗, kgk∗Mn⊗minMn = 1} < sup{f (X) : f ∈ (M2n ⊗ M2n)∗, kfk∗qa = 1}. Therefore, there is g ∈ (M2n⊗M2n)∗ such that kgk∗qa = 1 but g(X) > 1. A scaling argument shows that we may choose ε > 0 and g ∈ (M2n ⊗ M2n)∗ such that kgk∗Mn⊗minMn = 1 = g(X) and kgk∗qa = 1− ε. By multiplying g by some z ∈ T if necessary, we may assume that g(X) = 1. Let M ∈ M2n ⊗ M2n be the matrix such that g(Y ) = Tr(Y M) for all Y ∈ M2n ⊗ M2n. Since X = X∗, we see that g∗(X) = Tr(XM∗) = Tr(M X) = Tr(XM) = g(X) = 1. Since k · k∗qa is a ∗-norm, we obtain kg∗k∗qa ≤ 1 − ε and kg∗k∗Mn⊗minMn = 1 = g∗(X). Thus, f = Re(g) := g+g∗ is a functional with kfk∗qa ≤ 1 − ε and f (X) = 1 = kfk∗Mn⊗minMn. The associated matrix to f is M +M ∗ , which is self-adjoint. By the contrapositive, (5) implies (2). (cid:3) 2 2 As a corollary, we can describe the embedding problem in terms of perfect strategies for quantum XOR games. Corollary 7. The following are equivalent. (1) Connes' embedding problem has a positive answer. (2) For every n ≥ 2 and for every quantum XOR game G of size n with associated matrix (3) Whenever n ≥ 2 and G is a quantum XOR game of size n with a perfect qc-strategy, M ∈ Mn ⊗ Mn, we have ω∗qa(G) = ω∗qc(G). there is also a perfect qa-strategy for G. Proof. For t ∈ {qa, qc}, the quantity ω∗t (G) is the norm of a self-adjoint linear functional on Mn ⊗t Mn. In particular, if Connes' embedding problem holds, then by Theorem 6, Bqa(n, n) = Bqc(n, n) for all n, so that ω∗qa(G) = ω∗qc(G) for all quantum XOR games G. Hence, (1) implies (2). Clearly (2) implies (3), so it remains to show that (3) implies (1). If 10 SAMUEL J. HARRIS (1) fails, then by Theorem 6, there is some n ≥ 2, M = M∗ ∈ Mn ⊗ Mn, and 0 < ε < 1 such that 1 − ε = kMk∗qa < kMk∗qc = kMk∗Mn⊗minMn = 1. This implies that kMk1 = 1. Then there is a quantum XOR game G with associated matrix M [11]. By the choice of M, for every correlation X = (X(i,j),(k,ℓ))(i,j),(k,ℓ) ∈ Bqa(n, n), we have Hence, there is no perfect qa-strategy for G. Meanwhile, there is Y ∈ Bqc(n, n) with ω∗qc(G; Y ) = 1, so that Y is a perfect qc-strategy for G. This completes the proof. (cid:3) ω∗qa(G; X) ≤ 1 − ε. References [1] L. Brown, Ext of certain free product C ∗-algebras, J. Operator Theory 6 (1981), 135 -- 141. [2] J.F. Clauser, M.A. Horne, A. Shimony, and R.A. Holt, Proposed experiment to test local hidden-variable theories, Phys. Rev. Lett. 23 (1969), no. 15, 880 -- 884. [3] R. Cleve, L. Liu, and V.I. Paulsen, Perfect embezzlement of entanglement, J. Math. Phys. 58 (2017), no. 1, 012204, 18pp. [4] A. Connes, Classification of injective factors. Cases I I1, I I∞, I I Iλ, λ 6= 1, Ann. Math. (2) 104 (1976), [5] T. Fritz, Tsirelson's problem and Kirchberg's conjecture, Rev. Math. Phys. 24 (2012), no. 5, 1250012, no. 1, 73 -- 115. 67 pp. [6] S.J. Harris, A non-commutative unitary analogue of Kirchberg's conjecture, arXiv:1608.03229 [math.OA] (2016). [7] S.J. Harris and V.I. Paulsen, Unitary correlation sets, Integral Equations and Operator Theory (to appear). [8] M. Junge, M. Navascues, C. Palazuelos, D. Perez-Garcia, V.B. Scholz, and R.F. Werner, Connes em- bedding problem and Tsirelson's problem, J. Math. Phys. 52 (2011), no. 1, 012102, 12 pp. [9] N. Ozawa, About the Connes embedding conjecture: algebraic approaches, Jpn. J. Math. 8 (2013), no. 1, 147 -- 183. [10] V.I. Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in Advanced Mathe- matics, vol. 78, Cambridge University Press, Cambridge, 2002. [11] O. Regev and T. Vidick, Quantum XOR games, ACM Trans. Comput. Theory 7 (2015), no. 4, Art. 15, 43 pp. [12] W. Slofstra, The set of quantum correlations is not closed, arXiv:1703.08618 (2017). [13] B.S. Tsirelson, Some results and problems on quantum Bell-type inequalities, Hadronic J. Suppl. 8 (1993), no. 4, 329 -- 345. University of Waterloo Department of Pure Mathematics Waterloo, Ontario Canada N2L 3G1 E-mail address: [email protected]
1912.10331	1	1912	2019-12-21T20:27:50	Gelfand transforms and boundary representations of complete Nevanlinna--Pick quotients	[ "math.OA", "math.CV", "math.FA" ]	The main objects under study are quotients of multiplier algebras of certain complete Nevanlinna--Pick spaces, examples of which include the Drury--Arveson space on the ball and the Dirichlet space on the disc. We are particularly interested in the non-commutative Choquet boundaries for these quotients. Arveson's notion of hyperrigidity is shown to be detectable through the essential normality of some natural multiplication operators, thus extending previously known results on the Arveson--Douglas conjecture. We also highlight how the non-commutative Choquet boundaries of these quotients are intertwined with their Gelfand transforms being completely isometric. Finally, we isolate analytic and topological conditions on the so-called supports of the underlying ideals that clarify the nature of the non-commutative Choquet boundaries.	math.OA	math	GELFAND TRANSFORMS AND BOUNDARY REPRESENTATIONS OF COMPLETE NEVANLINNA -- PICK QUOTIENTS RAPHAEL CLOU ATRE AND EDWARD J. TIMKO Abstract. The main objects under study are quotients of multiplier algebras of certain complete Nevanlinna -- Pick spaces, examples of which include the Drury -- Arveson space on the ball and the Dirichlet space on the disc. We are particularly interested in the non-commutative Choquet boundaries for these quotients. Arveson's notion of hyperrigidity is shown to be detectable through the essential normality of some natural multiplication operators, thus extend- ing previously known results on the Arveson -- Douglas conjecture. We also highlight how the non-commutative Choquet boundaries of these quotients are intertwined with their Gelfand transforms being completely isometric. Fi- nally, we isolate analytic and topological conditions on the so-called supports of the underlying ideals that clarify the nature of the non-commutative Choquet boundaries. 1. Introduction The symbiotic relationship between analytic function theory on the open unit disc D ⊂ C and operator theory has long been fruitfully exploited. The standard setting is that of the Hardy space H 2 1 consisting of those analytic functions on D with square summable Taylor coefficients at the origin, and the main actor is the operator Mz : H 2 1 of multiplication by the variable z. Given an inner function θ, we can consider the so-called "model operator", obtained as the following compression 1 → H 2 Zθ = P(θH2 1 )⊥ Mz(θH2 1 )⊥. This operator and the various operator algebras related to it contain a striking amount of information about the function θ. One key to unraveling this information is an additional piece of structure that the Hardy space enjoys: the Szego kernel k1(z, w) = (1 − zw)−1, z, w ∈ D is its reproducing kernel. The potential lying within these concepts and their interactions was uncovered early on by Sarason in his seminal paper [58], wherein it was shown how a complete description of the operators commuting with Zθ is equivalent to several classical interpolation results, once the function θ is chosen appropriately. In this instance, Date: December 24, 2019. 2010 Mathematics Subject Classification. 47L55, 46E22, 47A13. RC was partially supported by an NSERC Discovery Grant. EJT was partially supported by a PIMS postdoctoral fellowship. 1 2 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO operator theory clarifies a function theory problem. In the reverse direction, opera- tor theoretic information can be extracted from the fine function theoretic proper- ties of the symbol θ. Indeed, the model operators Zθ, and operator-valued versions thereof, are the building blocks for a very robust classification theory of Hilbert space contractions (see [60],[14]). In the aforementioned developments a unifying central figure is what is now called the Commutant Lifting Theorem [59], which identifies the commutant of Zθ as the weak-∗ closed unital operator algebra that it generates. This operator algebra can also be regarded as the quotient M1/θM1, where M1 denotes the multiplier algebra of H 2 1 . We then see that the quotient algebra M1/θM1 carries information that is relevant both from the function theoretic and the operator theoretic points of view. Operator algebraic properties of these quotients were further leveraged in [19] to sharpen some classification results. More precisely, a key point there was whether or not the identity representation lies in the non-commutative Choquet boundary. An earlier foray into such questions can be found in Arveson's work in [6] and [7], which undertook a careful study of the norm-closed subalgebra of M1/θM1 generated by Zθ. Although the foregoing discussion is set in the one-variable world of the unit disc, there is an analogous multivariate version of the story that takes places on the open unit ball Bd ⊂ Cd, where d is some positive integer. This higher-dimensional setting has been a fertile ground for many years for research in both operator theory and function theory, and it continues to be to this day. Here, the classical Hardy space H 2 1 is replaced with an appropriate generalization, called the Drury -- Arveson space and denoted by H 2 d . The reproducing kernel of H 2 kd(z, w) = (1 − hz, wiCd )−1, d is z, w ∈ Bd and the multiplier algebra is denoted by Md. The intense research activity sur- rounding this space stems partly from the fact that it is universal (in a precise technical sense) for both the purposes of function theory [2] and of operator theory [9]. A commutant lifting theorem is known in this context as well [13]. Furthermore, mirroring the univariate developments, there is a robust yet still developing corre- sponding classification theory for commuting row contractions using weak-∗ closed ideals a ⊂ Md along with the associated d-tuples Za = (P[aH2 d ]⊥Mz1[aH2 d ]⊥, . . . , P[aH2 d ]⊥Mzd[aH2 d ]⊥) as building blocks (see [48],[9],[52],[25],[26] and the references therein). For these reasons, it is desirable to understand the structure of the quotients Md/a, as well as that of the norm-closed subalgebras generated therein by Za. Once again, the precise question we will explore is whether their identity representations belong to their respective non-commutative Choquet boundaries. We note that there are some important qualitative differences between the uni- variate and the multivariate worlds. For instance, it is readily seen that when d = 1, the model operator Zθ is always essentially unitary, in the sense that I − Z ∗ θ Zθ and I − ZθZ ∗ θ are compact [6, Theorem 3.4.2]. The corresponding statement when d > 1 is not obvious; whether or not it even holds is the content of the celebrated Arveson -- Douglas essential normality conjecture (see [36] and [33] for some recent progress GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 3 on this problem). Interestingly, our methodology here is in some ways connected to this subtle question, as was first realized in [43]. Throughout the paper, we work in a more general context, one which attempts to distill the salient distinguishing feature of the Drury -- Arveson space. Accordingly, we will be considering Hilbert spaces of analytic functions on the ball Bd with a complete Nevanlinna -- Pick reproducing kernel that is unitarily invariant, along with their multiplier algebras and their quotients. As we explain below, there are many important examples of such spaces, such as the Drury -- Arveson space and the Dirichlet space on the disc. The impetus to treat these two spaces (and many more) simultaneously is provided by recent advances in the connection between their function theory and the associated operator theory [15],[24],[27] building on ideas from [1],[5],[40]. With that being said, we emphasize that the setting of the Drury -- Arveson space and quotients of its multiplier algebra were the original motivation for the work undertaken in this paper, and many of our results are already new in this concrete setting. We now describe the organization of the paper and state our main results. In Section 2, we gather some background material on operator algebras, boundary rep- resentations and their C∗-envelopes. The main results of this section are Theorem 2.4 and Corollary 2.5, which highlight a connection between the identity represen- tation enjoying a certain unique extension property and the Gelfand transform of a commutative unital operator algebra being completely isometric. In Section 3, we define what we call maximal regular unitarily invariant repro- ducing kernel Hilbert spaces with the complete Nevanlinna -- Pick property, along with some natural algebras of multipliers associated with them. We also prove some preliminary facts, many of which are undoubtedly known to experts but do not seem to appear explicitly in the literature. Let H be such a space on the unit ball Bd and let M(H) denote its multiplier algebra. Fix some weak-∗ closed ideal a ⊂ M(H) and let Ha = H ⊖ [aH]. We define a weak-∗ closed, commutative unital operator algebra on Ha as Ma = {PHaMψHa : ψ ∈ M(H)}. We also let Aa ⊂ Ma denote the norm-closed subalgebra generated by the com- pressed coordinate multipliers z1, . . . , zd, where zj = PHaMzjHa, 1 ≤ j ≤ d. Let ∆(Aa) denote the space of characters of Aa and let g : Aa → C(∆(Aa)) be the Gelfand transform, so that (g(b))(χ) = χ(b), b ∈ Aa, χ ∈ ∆(Aa). Our investigation throughout is predicated on determining whether the identity representation is a boundary representation for the algebras Aa and Ma. Our analysis bifurcates into two strands, one for Aa and another for Ma. The need for a separate analysis of each case is not entirely unexpected, as even in the most classical situation where H is the Hardy space on the disc, the phenomena highlighted in [6],[7] and [19] are vastly different, and the methods of proof are quite distinct. The knowledge that the identity representation is a boundary representation is expected have some applications in operator theory. Indeed, it could lead to sharper versions of the results of [25] and [26], in the spirit of what was done in [19] 4 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO in the single-variable context. A more thorough examination of this prospect will be undertaken in a future paper. In Section 4, we analyze the identity representation of Ta = C∗(Aa) and wish to determine when it is a boundary representation for Aa. We start by identifying the character space of both Aa and Ta in Theorem 4.1 and use it to calculate the essential norm of elements in Aa in Theorem 4.2. Leveraging these pieces of information, we obtain the following characterization (Theorem 4.6). Theorem 1.1. Let H be a maximal regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal. If Ha has dimension greater than one, then the following statements are equivalent. (i) The Gelfand transform of Aa is completely isometric. (ii) The C∗-envelope of Aa is commutative. (iii) The identity representation of Ta is not a boundary representation for Aa. We note here that the case where H is the Hardy space on the unit disc was investigated by Arveson, who completely characterized the ideals a for which Aa has a completely isometric Gelfand transform. The details appear in Example 4 below. We show in Example 5 that our general situation is more complicated. Arveson's notion of hyperrigidity for operator algebras is also relevant for our purposes in Section 4. Determining whether an operator algebra is hyperrigid is typically quite difficult, as it requires checking that every representation of the generated C∗-algebra has a certain unique extension property. Nevertheless, hyper- rigidity of operator algebras has generated significant research activity recently (see for instance [43],[31],[24],[21],[20],[45],[57] and the references therein) prompted by a compelling yet still unresolved conjecture of Arveson [11]. It is therefore note- worthy that our approach also allows us to characterize when the algebra Aa is hyperrigid (see Corollary 4.5). Theorem 1.2. Let H be a maximal regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal. The following statements are equivalent. (i) The algebra Aa is hyperrigid in Ta. (ii) The d-tuple (z1, . . . , zd) is essentially normal and the Gelfand transform of Aa is not completely isometric. This theorem and its proof are inspired by [43, Theorem 4.12] which focuses on the Drury -- Arveson space. It should be noted that the results of [43, 44] hold for more general spaces than the Drury -- Arveson space, see Section 5 therein. However, as explained on page 1731 of [24], the overlap with our present setting consists only of the Drury -- Arveson space. In the second part of Section 4, we turn our attention to the whole algebra Ma, and analyze when the identity representation of C∗(Ma) is a boundary representa- tion for it. We obtain an analogue of Theorem 1.1 in this context (Corollary 4.7). Examples 6, 7, 8 and 9 put the various conditions therein in perspective. In Section 5, we find conditions under which the identity representation of Ta necessarily is a boundary representation for Aa, or equivalently, under which the Gelfand transform of Aa is not completely isometric. For instance, we show that this is almost always the case when a is a homogeneous ideal (Theorem 5.1), unless Aa can be identified with the classical disc algebra. Another instance where this GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 5 can be verified is when the reproducing kernel of the space H fails to be maximal (Theorem 5.2). For the rest of the section, we focus on the maximal situation and seek to exploit some analytic boundary assumptions on the so-called support of the ideal a, which was introduced in [27]. Given a compact subset K ⊂ Cd, we say that it is an approximation set if the polynomials are uniformly dense in the continuous functions on K. The following is Corollary 5.4. Theorem 1.3. Let H be a maximal regular unitarily invariant complete Nevanli- nna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal such that Ha has dimension greater than one and supp a ∩ Sd is an approximation set. Then, the Gelfand transform of Aa is not completely isometric. We recall that in one variable, a classical theorem of Lavrentiev states that if K ⊂ C is a compact subset, then the polynomials are dense in C(K) if and only if K has no interior and connected complement. A similar characterization of this density property in several variables appears to be a (difficult) open problem in approximation theory [46]. Some sufficient conditions are known however, and can be used to describe concrete situations where the previous theorem applies (see Corollary 5.5). In Corollary 5.9, we adapt the arguments found in [44] to sharpen Theorem 1.2 in the presence of some geometric restrictions on the zero set of the ideal a, and bring its statement in line with that found in [44]. Finally, in Section 6, we seek to identify sufficient conditions on the ideal a in order for the Gelfand transform of Ma not to be completely isometric. We show that the presence of some isolated points in the character space of Ma accomplishes this (Theorem 6.6). Theorem 1.4. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal whose support has an isolated point λ ∈ Bd. Then, the following statements hold. (i) The algebra Ma contains a non-zero idempotent finite-rank operator and the identity representation of C∗(Ma) is a boundary representation for Ma. More- over, the Gelfand transform of Ma is not completely isometric if Ha has di- mension greater than one. (ii) If supp a consists of more than one point, then the Gelfand transform of Ma is not isometric. As an application, we obtain the following (Corollary 6.7). Theorem 1.5. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal with the property that its zero set has an isolated point. Then, the following statements are equivalent. (i) The ideal a is the vanishing ideal of a single point. (ii) The space Ha is one-dimensional. (iii) The Gelfand transform of Ma is completely isometric. (iv) The Gelfand transform of Ma is isometric. This result is consistent with the findings of [18, Theorem 4.2] for the classical Hardy space on the unit disc, in which case the statement of the previous theorem is valid for any weak-∗ closed ideal. Whether or not this stronger statement holds in the generality discussed here calls for further investigation. 6 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO 2. Operator algebras, boundary representations and C∗-envelopes Let B(H) denote the C∗-algebra of bounded linear operators on some Hilbert space H. By an operator algebra we simply mean a subalgebra A ⊂ B(H). We denote by A′ the commutant of A, so that A′ = {X ∈ B(H) : Xa = aX, a ∈ A}. For each n ≥ 1, we let Mn(A) denote the algebra of n × n matrices with entries in A. When A = C, we write Mn instead of Mn(C). Generally, we view Mn(A) as a subalgebra of B(H(n)) in the obvious way (where H(n) is the n-fold direct sum of H), and thus obtain a norm on Mn(A). If E is a Hilbert space and τ : A → B(E) is a linear map, then for every n ≥ 1 we have an induced linear map τ (n) : Mn(A) → B(E (n)) defined as τ (n)([aij ]) = [τ (aij )], [aij] ∈ Mn(A). The map τ is said to be completely contractive or completely isometric whenever τ (n) is contractive or isometric, respectively, for each positive integer n. The reader may consult [50] for more details on the topic. Starting with a unital operator algebra A ⊂ B(H), we can consider the C∗- algebra C∗(A) that it generates inside of B(H). However, this C∗-algebra is typi- cally not invariant under completely isometric isomorphisms of A. In an attempt to construct the smallest possible C∗-algebra generated by a completely isometric copy of A, Arveson introduced in [6] non-commutative versions of the Choquet and Shilov boundaries for uniform algebras. We briefly recall some aspects of this construction, and the reader may consult [50, Chapter 15],[17, Chapter 4], or [10] for details. Let π : C∗(A) → B(Hπ) be a unital ∗-homomorphism. The restriction πA is a unital completely contractive map on A, and by virtue of Arveson's extension theorem it thus admits an extension to a unital completely contractive map on C∗(A). Of course, π itself is such an extension, but there may well be others in general. Accordingly, we say that π has the unique extension property with respect to A if π is the unique completely contractive linear extension of πA to C∗(A). The following result is found in [24, Theorem 3.4] and is based on a result from [54]; it gives a simple concrete criterion that detects the unique extension property. In the statement, any infinite sum is to be interpreted as being convergent in the strong operator topology. of commuting operators with the property that PT ∈F T T ∗ ≤ I. Let A ⊂ B(H) Theorem 2.1. Let H be a Hilbert space and let F ⊂ B(H) be a countable collection denote the unital norm-closed algebra generated by F . Let ρ : A → B(Hρ) be a unital completely contractive homomorphism with the property that ρ(T ) is normal for every T ∈ F and ρ(T )ρ(T )∗ = I. XT ∈F Then, there is a unital ∗-homomorphism π : C∗(A) → B(Hρ) with the unique extension property with respect to A such that πA = ρ. If π is an irreducible ∗-representation of C∗(A), then it is said to be a boundary representation for A whenever it has the unique extension property with respect GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 7 to A. The collection of boundary representation is usually thought of as the non- commutative Choquet boundary of A. In identifying boundary representations, the following general principle is often useful. It is entirely standard, but we record it here for ease of reference throughout the paper. Lemma 2.2. Let A ⊂ B(H) be a unital operator algebra such that C∗(A) contains the ideal K(H) of compact operators. Let q : C∗(A) → C∗(A)/K(H) be the quotient map. Let π : C∗(A) → B(Hπ) be a unital ∗-homomorphism. Then, there exists a closed subspace H0 ⊂ Hπ that is reducing for π(C∗(A)) along with a cardinal number κ, a unitary operator U : H0 → H(κ) and a unital ∗-homomorphism σ : C∗(A)/K(H) → B(H⊥ 0 ) such that π(X) = U ∗X (κ)U ⊕ (σ ◦ q)(X), X ∈ C∗(A). Proof. See the discussion at the bottom of page 15 along with Corollary 1 on page 20 in [8]. (cid:3) When every unital ∗-representation of C∗(A) has the unique extension property with respect to A, then we say that A is hyperrigid in C∗(A). This property was introduced in [11], where it was conjectured that it holds whenever every irreducible ∗-representation of C∗(A) is a boundary representation for A. This conjecture is still open at the time of this writing. Regardless of the status of the hyperrigidity conjecture, boundary representa- tions are known to always be plentiful: there always exists a set F of boundary metric on A, as shown in [34],[10],[30]. More generally, let F be a set of unital ∗-representations of C∗(A) that all have the unique extension property with respect to A, and such that the map non-commutative version of the Shilov boundary for A. We define the C∗-envelope of A to be representations for A with the property that the map Lπ∈F π is completely iso- ΠF =Lπ∈F π is completely isometric on A. The set F can be used to construct a Note that ker ΠF is a closed two-sided ideal of C∗(A) for which e(A) = ΠF (C∗(A)). C∗ e(A) ∼= C∗(A)/ ker ΠF C∗ It can be verified that C∗ and that the associated quotient map C∗(A) → C∗(A)/ ker ΠF is completely iso- (Here and throughout, the symbol ∼= denotes the relation of ∗- metric on A. isomorphism.) We often identify C∗ e(A) is invariant under unital completely isometric iso- morphisms of A [6, Theorem 2.1.2]. Furthermore, the C∗-envelope can be inter- preted as the essentially unique smallest C∗-algebra generated by a copy of A, in the sense that it is a quotient of every other such C∗-algebra [34, Theorem 4.1]. This object was first proven to exist by Hamana [41] using different techniques. e(A) with C∗(A)/ ker ΠF . When the identity representation of C∗(A) has the unique extension property with respect to A, then the set F above can be chosen to simply consist of this e(A) ∼= C∗(A). We thus see that it single representation. So in this case we have C∗ is desirable to detect when the identity representation enjoys the unique extension property. For this purpose, the best tool is the following result, known as Arveson's boundary theorem [7, Theorem 2.1.1]. 8 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO quotient map. Then, the following statements are equivalent. Theorem 2.3. Let A ⊂ B(H) be a unital operator algebra such that C∗(A) contains the ideal K(H) of compact operators. Let q : C∗(A) → C∗(A)/K(H) denote the (i) The identity representation of C∗(A) is a boundary representation for A. (ii) The map q is not completely isometric on A. We now recall some terminology. Let B be a unital commutative Banach algebra. We denote the space of its characters by ∆(B). The Gelfand transform of B is the unital contractive homomorphism defined as g : B → C(∆(B)) (g(b))(χ) = χ(b) for every b ∈ B, χ ∈ ∆(B). When we wish to emphasize the dependence on the algebra, we sometimes write gB for the Gelfand transform. The main result of this section shows that whether or not the Gelfand transform of a commutative unital operator algebra is completely isometric can be detected via the C∗-envelope. Theorem 2.4. Let A ⊂ B(H) be a norm-closed unital commutative operator alge- bra. The following statements are equivalent. (i) The Gelfand transform of A is completely isometric. (ii) The C∗-envelope of A is commutative. (iii) There is a closed two-sided ideal J ⊂ C∗(A) such that C∗(A)/J is commutative and the quotient map qJ : C∗(A) → C∗(A)/J is completely isometric on A. Proof. (i) ⇒ (ii): Let g : A → C(∆(A)) denote the Gelfand transform of A, which we assume is completely isometric. Because the C∗-envelope is invariant under unital completely isometric isomorphisms of A, we see that C∗ e(g(A)). e(g(A)) is a quotient of C∗(g(A)) ⊂ C(∆(A)), so it is But we also know that C∗ commutative. e(A) be a unital, completely isometric homomorphism, e(A) which exists by construction of the C∗-envelope. By assumption we have that C∗ is commutative, so that for n ≥ 1 and A ∈ Mn(A) we find (ii) ⇒ (i): Let Π : A → C∗ e(A) ∼= C∗ We conclude that the Gelfand transform g is completely isometric. described above. (ii) ⇒ (iii): This follows immediately from the construction of the C∗-envelope (iii) ⇒ (ii): Assume that there is a closed two-sided ideal J ⊂ C∗(A) such that C∗(A)/J is commutative and the quotient map qJ : C∗(A) → C∗(A)/J is completely isometric on A. It then follows that C∗ e(qJ(A)) is a quotient of C∗(A)/J and hence it is commutative. e(A) ∼= C∗ (cid:3) We extract a consequence that will be useful for our purposes. Corollary 2.5. Let H be a Hilbert space with dimension greater than one. Let A ⊂ B(H) be a norm-closed unital commutative operator algebra. Assume that C∗(A) contains the ideal K(H) of compact operators. Consider the following statements. kAk = kΠ(n)(A)k = max{kχ(n)(Π(n)(A))k : χ ∈ ∆(C∗ e(A))} = max{k(χ ◦ Π)(n)(A)k : χ ∈ ∆(C∗ ≤ max{kχ(n)(A)k : χ ∈ ∆(A)} = kg(n)(A)k. e(A))} GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 9 (i) The Gelfand transform of A is completely isometric. (ii) The identity representation of C∗(A) is not a boundary representation for A. Then, we have that (i) ⇒ (ii). If we assume in addition that C∗(A)/K(H) is com- mutative, then we have (i) ⇔ (ii). Proof. (i) ⇒ (ii): We note that C∗(A) is irreducible and not commutative since it contains the compact operators and H has dimension greater than one. If the e(A) ∼= identity representation of C∗(A) is a boundary representation for A, then C∗ C∗(A), which is not commutative. Therefore, the Gelfand transform of A is not completely isometric by Theorem 2.4. Assume now that C∗(A)/K(H) is commutative and that the identity represen- tation of C∗(A) is not a boundary representation for A. By Theorem 2.3, we see that the quotient map q : C∗(A) → C∗(A)/K(H) is completely isometric on A. Since C∗(A)/K is assumed to be commutative, it follows from Theorem 2.4 that the Gelfand transform of A is completely isometric. Thus (i) ⇔ (ii) in this case. (cid:3) 3. Unitarily invariant kernels and their multiplier algebras 3.1. Unitarily invariant kernels with the complete Nevanlinna -- Pick prop- erty. Although the theory of reproducing kernels makes sense in great generality, our focus in this paper will be fairly concrete, so we choose to only recall the rele- vant definitions in the more specialized setting that we will work in. More details on the general theory can be found in [3] or [51]. Throughout, the underlying domain will be the open unit ball Bd ⊂ Cd, for some integer d ≥ 1. A kernel is a function k : Bd × Bd → C such that the complex matrix [k(λi, λj )]i,j is non-negative definite for any finite subset {λ1, . . . , λn} ∈ Bd. We will always assume that k is a unitarily invariant kernel, in the sense that it satisfies the following additional conditions. (a) The kernel is normalized at 0, in the sense that k(0, w) = 1, w ∈ Bd. (b) The function k is analytic in the first variable and co-analytic in the second. (c) We have that k(U z, U w) = k(z, w), for every unitary operator U : Cd → Cd. z, w ∈ Bd As shown in [42, Lemma 2.2], these properties force the kernel to be of the form k(z, w) = anhz, win Cd , z, w ∈ Bd ∞Xn=0 for some sequence (an) of non-negative real numbers such that a0 = 1. Associated to the kernel k is a Hilbert space Hk of analytic functions on Bd which is spanned by the vectors {kw : w ∈ Bd}, where Furthermore, these distinguished vectors have the usual reproducing property kw(z) = k(z, w), z, w ∈ Bd. for every f ∈ Hk and every w ∈ Bd. hf, kwiHk = f (w) 10 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO We say that k is a regular unitarily invariant kernel if in addition we have that an > 0 for every n ≥ 0 and lim n→∞ an an+1 = 1. As explained in [24, Section 2], this last condition insures in particular that the natural domain of definition of the functions in Hk is Bd, and not some larger set. 0} form an orthogonal basis for H, where It follows from [40, Proposition 4.1] that the monomials {zα : α ∈ Nd Here, we use the standard multi-index notation, so that if α = (α1, . . . , αd) ∈ Nd then 0 N0 = {0, 1, 2, . . .}. α! = α1!α2!··· αd! and α = α1 + α2 + ··· + αd. Moreover, given a d-tuple r = (r1, . . . , rd) of elements in some unital commutative ring we set rα = rα1 1 rα2 2 ··· rαd d . Among the regular unitarily invariant kernels, we will focus on those that have the complete Nevanlinna -- Pick property, in the sense that there is a sequence (bn) of non-negative numbers with the property that 1 k(z, w) 1 − = ∞Xn=1 bnhz, win, z, w ∈ Bd. Classically, the complete Nevanlinna -- Pick property is defined in terms of the solv- ability of some interpolation problem, but this is equivalent to the previous def- inition in our context [42, Lemma 2.3]. A calculation reveals that our standing assumption that a0 = 1 implies a1 = b1 > 0. Moreover, it is readily seen that n=1 bn = 1. This is seen n=1 an diverges. The vast majority of n=1 bn ≤ 1. We will say that the kernel is maximal ifP∞ P∞ to be equivalent to the fact that the seriesP∞ our results will require this maximality condition. Since the kernel k and the Hilbert space Hk uniquely determine one another, we will often speak of maximal, regular, unitarily invariant complete Nevanlinna -- Pick spaces, rather than kernels. The following example shows that there are many meaningful instances of such spaces. Example 1. A standard family of examples of maximal, regular, unitarily invariant kernels on Bd with the complete Nevanlinna -- Pick property is given by {ks : −1 ≤ s}, where ks(z, w) = (n + 1)shz, win, z, w ∈ Bd. ∞Xn=0 See [24, Section 2.4] for details. We note that choosing s = 0 yields the Drury- Arveson kernel mentioned in the introduction. Moreover, when d = 1 and s = −1, we obtain the famous Dirichlet kernel [35]. (cid:3) 3.2. Multiplier algebras and representations. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space. Recall that a function ψ : Bd → C is a multiplier of H if ψf ∈ H for every f ∈ H. In this case, the associated multiplication operator Mψ : H → H is bounded. Throughout, we identify the GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 11 function ψ with Mψ whenever convenient. In particular, we may view the algebra M(H) of all multipliers of H as an operator algebra on H, and we define kψkM(H) = kMψkB(H), ψ ∈ M(H). It is a standard fact that M(H)′ = M(H) in our context, so in particular M(H) is weak-∗ closed. The vectors {kw : w ∈ Bd} span a dense subset of H and they satisfy M ∗ ψkw = ψ(w)kw for every w ∈ Bd and every ψ ∈ M(H). One consequence of this is that (1) k[ψij]kMn(M(H)) ≥ sup z∈Bd k[ψij (z)]kMn for every n ≥ 1 and every [ψij] ∈ Mn(M(H)). Another useful consequence is that a bounded net (ψi)i of multipliers converges to ψ in the weak-∗ topology of M(H) if and only if it converges to ψ pointwise on Bd. Combining [42, Lemma 2.3 and Proposition 6.4], we see that the algebra M(H) contains the polynomials. Denoting the standard coordinate functions by z1, . . . , zd, we thus have Mz1, . . . , Mzd ∈ B(H) and we put Mz = (Mz1, . . . , Mzd ). The following lemma gathers some additional facts that we require throughout. Lemma 3.1. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space with kernel k. Let (bn) be the sequence of non-negative real numbers such that The following statements hold. (i) The row operator is contractive. In fact, the operator 1 − k(z, w) bnhz, win, z, w ∈ Bd. 1 = ∞Xn=1 b1/2 α (cid:18)α! α! (cid:19)1/2 ∞Xn=1 bn Xα=n I − M α z!α≥1 n! α! M α z M α∗ z is the orthogonal projection onto C1. M α (ii) The function 1 − 1/kλ is a contractive multiplier of H for every λ ∈ Bd. . Then I− RR∗ is positive and bounded by [24, Lemma 5.2]. Thus, the row operator R is indeed contractive. Next, let λ, µ ∈ Bd. We compute Proof. (i) Consider the row operator R =(cid:18)b1/2 α (cid:16) α! α!(cid:17)1/2 z  kλ, kµ+ =1 − *I − = 1 − µαhkλ, kµi bn Xα=n bnhµ, λin! k(µ, λ) = 1. z(cid:19)α≥1 bn Xα=n ∞Xn=1 M α z M α∗ n! α! α λ n! α! ∞Xn=1 ∞Xn=1 12 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO On the other hand, if we denote by E ∈ B(H) the orthogonal projection onto C1, then we find Since the vectors {kλ : λ ∈ Bd} span a dense subset of H, we conclude that hEkλ, kµi = hkλ, 1ih1, kµi = 1. (ii) Consider the row operator We find M α z M α∗ z . n! α! E = I − bn Xα=n α! (cid:19)1/2 ∞Xn=1 Λ = b1/2 λαI!α≥1 α (cid:18)α! ΛΛ∗ = ∞Xn=1 bnkλk2n! I ≤ ∞Xn=1 ∞Xn=1 M α z n! α! . bn Xα=n ΛR∗kµ =(cid:18)1 − 1 k(λ, µ)(cid:19) kµ bn! I ≤ I. Recall now that we proved above that RR∗ ≤ I, so λ RΛ∗ = α is a contractive multiplier. A straightforward verification shows that for every µ ∈ Bd, which implies that M1−1/kλ = RΛ∗ is a contractive multiplier. (cid:3) We let A(H) denote the norm-closed subalgebra of M(H) generated by the polynomials. Further, we denote by T(H) the C∗-subalgebra of B(H) generated by A(H). We now identify the character spaces of these algebras. Because of (1), it is readily seen that the functions in A(H) extend to be con- tinuous on Bd. In particular, for each λ ∈ Bd there is a character ελ : A(H) → C defined as ελ(ϕ) = ϕ(λ), ϕ ∈ A(H). Conversely, if χ : A(H) → C is a character and if we put λ = (χ(Mz1), . . . , χ(Mzd )), ∆(A(H)) = {ελ : λ ∈ Bd}. then λ ∈ Bd; this can be seen as in the proof of [24, Lemma 5.3] using part (i) of Lemma 3.1. Since the polynomials are dense in A(H), we conclude that χ = ελ. Thus, if we let ∆(A(H)) denote the character space of A(H), then we find (2) We now turn to the character space of T(H). We denote by Sd the unit sphere in Cd, i.e. the topological boundary of Bd. Theorem 3.2. Let H be a regular unitarily invariant complete Nevanlinna -- Pick is compact and the C∗-algebra T(H) contains the ideal K(H) of compact operators on H. Furthermore, there is a ∗- isomorphism Φ : T(H)/K(H) → C(Sd) such that space. Then, the operator I −Pd j=1 Mzj M ∗ zj Φ(Mϕ + K(H)) = ϕSd , ϕ ∈ A(H). GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS Proof. This follows from [40, Proposition 4.3 and Theorem 4.6]. 13 (cid:3) 3.3. Ideals and quotients. In this subsection we introduce the main operator algebras that we study in the paper, and prove some relevant preliminary facts about them. We first point out a very useful correspondence between weak-∗ closed ideals and invariant subspaces of M(H), which can be found in [27, Theorem 2.2]. Theorem 3.3. Let H be a unitarily invariant complete Nevanlinna -- Pick space. Given a weak-∗ closed ideal a ⊂ M(H), the set R(a) = span{Mϕf : ϕ ∈ a, f ∈ H} is a closed invariant subspace for M(H). Given a closed subspace N ⊂ H that is invariant for M(H), the set ι(N ) = {ψ ∈ M(H) : ran Mψ ⊂ N} is a weak-∗ closed ideal of M(H). Moreover, we have that ι(R(a)) = a, R(ι(N )) = N for every weak-∗ closed ideal a ⊂ M(H) and every closed subspace N ⊂ H that is invariant for M(H). Let H be a unitarily invariant complete Nevanlinna -- Pick space and let a ⊂ M(H) be a weak-∗ closed ideal. Let [aH] = span{Mϕf : ϕ ∈ a, f ∈ H} and Ha = H⊖ [aH]. If a is a proper ideal of M(H), then Ha is necessarily non-zero, as we show next. Lemma 3.4. Let H be a unitarily invariant complete Nevanlinna -- Pick space and let a ⊂ M(H) be a weak-∗ closed ideal. The following statements hold. (i) The ideal a is proper if and only if PHa 1 6= 0. (ii) The ideal a is maximal if and only if Ha is one-dimensional. Proof. (i) If a = M(H), then clearly Ha = {0} and PHa1 = 0. Conversely, if a is proper then 1 /∈ a, whence 1 /∈ [aH] by Theorem 3.3. This is equivalent to PHa1 6= 0. (ii) If a is proper but not maximal, then M(H)/a has dimension at least two. Hence, there is ϕ ∈ a such that the cosets 1 + a and ϕ + a are linearly independent in M(H)/a. Thus, if c, d ∈ C are not both zero, then c1 + dϕ /∈ a, which in turn implies by Theorem 3.3 that c1 + dϕ /∈ [aH]. We conclude that PHa1 and PHaϕ are linearly independent in Ha, so that Ha has dimension greater than one. Conversely, if a is maximal then we have M(H) = C1 + a. Using that M(H) is dense in H, we infer that C1 + a is dense in H as well. This implies that CPHa 1 = PHa(C1 + a) is dense in Ha, so that Ha is one-dimensional. (cid:3) Observe that Ha is a co-invariant subspace for M(H). Let Za = (z1, . . . , zd) denote the d-tuple of commuting operators on Ha where 1 ≤ j ≤ d. zj = PHaMzjHa, Consider now the map Γ : M(H) → B(Ha) defined as Γ(ψ) = PHa MψHa, ψ ∈ M(H). 14 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO It is readily verified that Γ is unital completely contractive homomorphism that is also weak-∗ continuous. It follows from Theorem 3.3 that ker Γ = a. The map Γ should be thought of as the "M(H)-functional calculus" in the sense of [15], and thus we use the notation Throughout the paper, we will use the notation Ma = Γ(M(H)). Thus, ψ(Za) = Γ(ψ), ψ ∈ M(H). Ma = {ψ(Za) : ψ ∈ M(H)}. This algebra will be one of our main objects of study. It follows from the Com- mutant Lifting Theorem [13, Theorem 5.1] that M′ a = Ma. In particular, Ma is closed in the weak-∗ topology of B(Ha). Invoking [27, Theorem 2.3], we see that Γ induces a unital, completely isometric and weak-∗ homeomorphic isomorphism Γa : M(H)/a → Ma such that Γa(Mψ + a) = ψ(Za), ψ ∈ M(H). We denote the character space of Ma by ∆(Ma). Following [27, Theorem 3.8], we define the support of a to be the set supp a = {(χ(z1), . . . , χ(zd)) : χ ∈ ∆(Ma)}. If M(H) is assumed to satisfy the Corona property, then a more concrete function theoretic description of the support can be given, see [27, Corollary 3.12]. In general, it can be shown that supp a ∩ Bd coincides with the zero set {λ ∈ Bd : ψ(λ) = 0 for every ψ ∈ a}, see [27, Theorem 3.4] for details. Next, we define the second operator algebra on which we focus throughout. We put so that Aa = Γ(A(H)) norm Aa = {ϕ(Za) : ϕ ∈ A(H)} norm . We also set Ta = C∗(Aa). We show below that Ta contains the ideal K(Ha) of compact operators, and we denote the corresponding quotient by Oa = Ta/K(Ha). The following generalizes [24, Lemma 7.1]. Lemma 3.5. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space with kernel k. Let (bn) be the sequence of non-negative real numbers such that 1 k(z, w) 1 − = bnhz, win, z, w ∈ Bd. ∞Xn=1 Let a ⊂ M(H) be a proper weak-∗ closed ideal. Then, the following statements hold. (i) The operator I −Pd (ii) We have that j=1 zjz∗ j is compact. PHaPC1Ha = I − ∞Xn=1 bn Xα=n n! α! Z α a Z α∗ a and this operator is compact and non-zero. GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 15 (iii) The C∗-algebra Ta is irreducible and contains the ideal of compact operators on Ha. Proof. (i) Observe that a = PHa I − I − ZaZ ∗ which is compact by Theorem 3.2. (ii) Put Q = I − MzkM ∗ zk!Ha dXk=1 ∞Xn=1 bn Xα=n n! α! Z α a Z α∗ a . Then Q = PHaPC1Ha by Lemma 3.1(i), so that Q has rank at most one. Moreover, if Q = 0 then PHa PC1 = 0, and so PHa 1 = 0, which is impossible by Lemma 3.4 since a is a proper ideal. (iii) Let P ∈ B(Ha) be a self-adjoint projection with P 6= I that commutes with Aa, and hence with Ma. By the Commutant Lifting Theorem [13, Theorem 5.1], there is ψ ∈ M(H) with ψ 6= 1 and kψk ≤ 1 such that P = ψ(Za). By the maximum modulus principle, we see that the sequence (ψn)n converges to 0 pointwise on Bd, and thus also in the weak-∗ topology of M(H). In particular, we infer that (Γ(ψn))n converges to 0 in the weak-∗ topology of B(Ha). Since P = P n = Γ(ψn), n ≥ 1 we conclude that P = 0. This shows that Aa is irreducible, and hence so is Ta. Therefore, Ta is an irreducible C∗-algebra containing a non-zero compact operator (by (ii)), and hence it must contain all compact operators on Ha [8, Corollary 2 page 18]. (cid:3) In light of the identication of M(H)/a with Ma, it is natural to wonder if Aa can be identified in an analogous fashion with some quotient of A(H). This is possible in some special situations. For instance, [24, Corollary 8.3] can be used to show that if a is a so-called homogeneous ideal, then Aa can be identified completely isometrically with A(H)/(a ∩ A(H)). But this identification does not hold for general ideals. Indeed, even in the classical case of the Hardy space on the unit disc there are non-trivial weak-∗ closed ideals a such that a ∩ A(H) is zero, yet the natural map from A(H) onto Aa is not isometric [7, Corollary 1 page 291]. We end this section with an important fact. Recall that a commuting d-tuple U = (U1, . . . , Ud) of operators is said to be a spherical unitary if U1, . . . , Ud are all normal andPd j=1 UjU ∗ j = I. Lemma 3.6. Let H be a maximal regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal. Let ρ : Aa → B(E) be a unital completely contractive representation with the property that (ρ(z1), . . . , ρ(zd)) is a spherical unitary. Then, there is a unital ∗-homomorphism π : Ta → B(E) with the unique extension property with respect to Aa such that πAa = ρ. Proof. The proof follows that of [24, Proposition 6.1]. For each multi-index α with α ≥ 1, we let Tα = b1/2 α (cid:16) α! α!(cid:17)1/2 Z α a . Then, we find Xα≥1 TαT ∗ α ≤ I 16 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO by Lemma 3.5(ii). On other hand, we see that ρ(Tα)ρ(Tα)∗ = Xα≥1 = ∞Xn=1 ∞Xn=1 bn Xα=n bn nXj=1 n! α! ρ(Za)αρ(Za)∗α ρ(zj )ρ(zj )∗ n = ∞Xn=1 bnI = I. Note also that each ρ(Tα) is normal by virtue of the Fuglede -- Putnam theorem. Because Aa is the norm-closed unital algebra generated by {Tα : α ≥ 1}, we may apply Theorem 2.1 to finish the proof. (cid:3) We note here that the maximality condition in the previous theorem cannot be removed; see [24, Theorem 6.2(b)]. 4. The connection between Gelfand transforms and boundary representations The main goal of this section is to establish a link between the operator algebras Aa and Ma having a completely isometric Gelfand transform, and the identity representation being a boundary representation for them. As announced in the introduction, we will deal with the two algebras separately. 4.1. The norm-closed algebra Aa. We first focus on Aa. For this purpose, we heavily draw from [43] and adapt some of the results therein. Before proceeding, we introduce some notation. Given a compact subset K ⊂ Cd, we denote its polynomially convex hull by bK, so that λ ∈ bK if and only if for every polynomial p. We say that K is polynomially convex when bK = K. It is well-known that if K ⊂ Bd, then bK ∩ Sd = K ∩ Sd; this can readily be verified upon considering the polynomial p(λ) ≤ max w∈K p(w) pζ = (1 + hz, ζi) 1 2 which peaks at ζ ∈ Sd. We now identify the character spaces of Aa and Ta, thereby obtaining versions [43, Proposition 2.1] and [24, Theorem 7.3] applicable to our context. Given a subset S ⊂ A(H), we write ZBd (S) = {λ ∈ Bd : ϕ(λ) = 0 for every ϕ ∈ S}. Moreover, if S ⊂ M(H) we write ZBd (S) = {λ ∈ Bd : ψ(λ) = 0 for every ψ ∈ S}. Theorem 4.1. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal. Let ∆(Aa) and ∆(Ta) denote the character spaces of Aa and Ta respectively. Then, the following statements hold. (i) For each λ ∈ \supp a, there is a unique character τλ ∈ ∆(Aa) such that τλ(ϕ(Za)) = ϕ(λ), ϕ ∈ A(H). GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 17 Moreover, we have that (ii) We have that ∆(Aa) = {τλ : λ ∈ \supp a}. {(χ(z1), . . . , χ(zd)) : χ ∈ ∆(Aa)} ⊂ ZBd (a ∩ A(H)). (iii) Assume that H is maximal. Let χ ∈ ∆(Aa) and set ζ = (χ(z1), . . . , χ(zd)). If ζ ∈ Sd, then there exists a unique πζ ∈ ∆(Ta) such that πζ(ϕ(Za)) = χ(ϕ(Za)) = ϕ(ζ), ϕ ∈ A(H). (iv) Assume that H is maximal. If Ha is one-dimensional, then ∆(Aa) = ∆(Ta). If Ha has dimension greater than one, then ∆(Ta) = {πζ : ζ ∈ supp a ∩ Sd}. Proof. (i) By [49, Theorem I.3.10 (vii)] and by our definition of the support, we see that {(χ(z1), . . . , χ(zd)) : χ ∈ ∆(Aa)} = \supp a. Furthermore, by density of the polynomials in A(H), a character χ ∈ ∆(Aa) is uniquely determined by its values on z1, . . . , zd. Both statements follow from this. (ii) We first make a preliminary observation. Let w ∈ supp a. By (i), we have that τw ∈ ∆(Aa). Hence ϕ(w) = τw(ϕ(Za)) = 0, ϕ ∈ a ∩ A(H). This shows that supp a ⊂ ZBd λ ∈ \supp a such that χ = τλ. It follows that (a ∩ A(H)). Next, let χ ∈ ∆(Aa). By (i), there is ϕ(λ) ≤ max w∈supp aϕ(w), ϕ ∈ A(H) whence λ ∈ ZBd (a ∩ A(H)) by the previous observation. We conclude that {(χ(z1), . . . , χ(zd)) : χ ∈ ∆(Aa)} ⊂ ZBd (a ∩ A(H)). (iii) Note that dXj=1 χ(zj )2 = kζk2 Cd = 1, so we may invoke Lemma 3.6 to see that there is a unique πζ ∈ ∆(Ta) such that πζ(ϕ(Za)) = χ(ϕ(Za)) = ϕ(ζ), ϕ ∈ A(H). (iv) If Ha is one-dimensional, then clearly Aa = Ta. Suppose therefore that Ha has dimension greater than one. We wish to show that ∆(Ta) = {πζ : ζ ∈ supp a ∩ Sd}. One inclusion follows from (iii). To see the reverse inclusion, let χ ∈ ∆(Ta) and put ζ = (χ(z1), . . . , χ(zd)). Then, χAa ∈ ∆(Aa) so that ζ ∈ \supp a and χAa = τζ by (i). By Lemma 3.5, we j=1 zj z∗ j see that Ta contains the ideal K(Ha) of compact operators and that I −Pd and ζ ∈ Sd. Thus kζk2 Cd = χ(zj)2 = 1 dXj=1 ζ ∈ \supp a ∩ Sd = supp a ∩ Sd 18 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO is compact. Since Ha has dimension greater than one, it follows from Lemma 2.2 that K(Ha) ⊂ ker χ, whence in light of the remark preceding the theorem. Finally, the equality χAa = τζ implies χ = πζ since Ta = C∗(Aa). (cid:3) We now aim to calculate the essential norm of an element of Aa, using the same argument as in [43, Theorem 3.3]. To this end, recall that Ta always con- tains the ideal K(Ha) of compact operators by Lemma 3.5, and that we write Oa = Ta/K(Ha). We use the symbols gB to denote the Gelfand transform of a commutative unital Banach algebra B. Theorem 4.2. Let H be a maximal regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal and let q : Ta → Oa denote the quotient map. Let n ≥ 1 and let A = [ϕij (Za)]ij ∈ Mn(Aa). Then, we have that (3) If we assume in addition that Ha has dimension greater than one, then we have kq(n)(A)k ≤ kg(n) (A)k ≤ kg(n) (A)k. Aa Ta kq(n)(A)k = kg(n) Ta (A)k = kg(n) Oa (q(n)(A))k = kg(n) q(Aa)(q(n)(A))k. In particular, the Gelfand transform of q(Aa) is completely isometric in this case. Proof. In what follows, when χ is a character of an algebra containing z1, . . . , zd, we put χ(Za) = (χ(z1), . . . , χ(zd)), with similar natural abbreviations being used where appropriate. We may assume that Oa ⊂ B(E) for some Hilbert space E. Lemma 3.5(i) that It follows from I = q(zj )q(zj )∗. dXj=1 By [12, Proposition 2], we see that there is a Hilbert space E ′ containing E along with a spherical unitary U = (U1, . . . , Ud) acting on E ′ and leaving E co-invariant, with the property that U ∗ j E = q(zj )∗ for every 1 ≤ j ≤ d. Requiring that U ∗ be minimal among the normal dilations of q(Za)∗, we can arrange that by virtue of [53]. In turn, as explained in [54, Section 3] this implies σTa(U ∗) ⊂ σTa(q(Za)∗) On the other hand, we always have that σTa(U ) ⊂ σTa(q(Za)). σTa(q(Za)) ⊂ {χ(q(Za)) : χ ∈ ∆(q(Aa))} by [49, Proposition IV.25.3]. Now, given χ ∈ ∆(q(Aa)), we see that χ ◦ q ∈ ∆(Aa) and so σTa(q(Za)) ⊂ {χ(Za) : χ ∈ ∆(Aa)}. GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 19 In particular, we obtain σTa(q(Za)) ∩ Sd ⊂ {χ(Za) : χ ∈ ∆(Aa)} ∩ Sd ⊂ {χ(Za) : χ ∈ ∆(Ta)} where the last equality follows Theorem 4.1(iii). Since by [29, Proposition 7.2], we find σTa(U ) = {χ(U ) : χ ∈ ∆(C∗(U ))} ⊂ Sd (4) σTa(U ) ⊂ σTa(q(Za)) ∩ Sd ⊂ {χ(Za) : χ ∈ ∆(Ta)} ⊂ {χ(Za) : χ ∈ ∆(Aa)}. Observe now that q(n)(A) = [q(ϕij (Za))]ij = [(ϕij (q(Za))]ij = PE (n) [ϕij (U )]ijE (n) so we find kq(n)(A)k ≤ k[ϕij(U )]ijk. Next, we know that Mn(C∗(U )) is ∗-isomorphic to C(∆(C∗(U )), Mn) and k[ϕij(U )]ijkMn(C∗(U)) = max{k[ϕij(χ(U ))]ijkMn : χ ∈ ∆(C∗(U ))} = max{k[ϕij(ζ)]ijkMn : ζ ∈ σTa(U )}. Combining these facts with (4), we have kq(n)(A)k ≤ max{k[ϕij(χ(Za))]ijk : χ ∈ ∆(Ta)} ≤ max{k[ϕij(χ(Za))]ijk : χ ∈ ∆(Aa)}, which is (3). If Ha has dimension greater than one, then every character of Ta must annihilate K(Ha) by Lemma 2.2. Thus, given χ ∈ ∆(Ta), we can find a character bχ ∈ ∆(Oa) such that χ =bχ ◦ q. This implies that (A)k = k(gOa ◦ q)(n)(A)k ≤ k(gq(Aa) ◦ q)(n)(A)k ≤ kq(n)(A)k. kg(n) Ta Applying (3), we conclude that kq(n)(A)k = kg(n) Ta in this case. (A)k = k(gOa ◦ q)(n)(A)k = k(gq(Aa) ◦ q)(n)(A)k (cid:3) We note that the second (stronger) assertion of the previous theorem fails when Ha is one-dimensional. Indeed, in that case we have q(I) = 0 while the identity map is a character of Ta. Furthermore, we mention that the method of proof above does not typically adapt to apply to Ma, in view of the following example. Example 2. Let d ≥ 2 and let H be the Drury -- Arveson space on Bd. Let a = {0} so that H = Ha and Ma = M(H). Let q : C∗(M(H)) → C∗(M(H))/K(H) be the quotient map. By [37, Theorem 1.1], there is ψ ∈ M(H) such that kq(Mψ)k > sup z∈Bd ψ(z). Next, let χ : q(M(H)) → C be a character. Then, χ ◦ qM(H) is a character of M(H), so by the Corona theorem [28], we see that χ(q(Mψ)) ≤ sup z∈Bd ψ(z). We conclude that kq(Mψ)k > max{χ(q(Mψ)) : χ ∈ ∆(q(M(H))}. 20 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO Before we give some consequences of Theorem 4.2, we require a maximum mod- ulus principle for Ma that is of independent interest. Theorem 4.3. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal and let λ ∈ ZBd(a). Let M = [ψij (Za)]ij ∈ Mn(Ma) be such that kMk = k[ψij(λ)]ijkMn . Then, there are unit vectors v = (v1, . . . , vn) and w = (w1, . . . , wn) in Cn with the property that (cid:3) nXj=1 vj ψij(Za) = kMkwiIHa, 1 ≤ i ≤ n. Proof. Put M (λ) = [ψij(λ)]ij . By assumption, we may choose unit vectors v, w ∈ Cn with the property that hM (λ)v, wiCn = kMk. Because λ ∈ ZBd (a), we have kλ ∈ Ha. Define unit vectors ξ, η ∈ H(n) kkλk(cid:19) . η =(cid:18)w1 ξ =(cid:18)v1 kλ kkλk kλ kkλk , . . . , wn , . . . , vn a by kλ We compute kλ kkλk(cid:19) , nXi,j=1 nXi,j=1 1 hM ξ, ηiH(n) a = = kkλk2 vjwihψij (Za)kλ, kλiHa vjwiψij (λ) = hM (λ)v, wiCn = kMk. By the Cauchy-Schwarz inequality, we infer that M ξ = kMkη. In other words, for each 1 ≤ i ≤ n we find which is equivalent to  ψij(Za)vj kλ = kMkwikλ nXj=1 vjψij − kMkwi kλ ∈ [aH]. nXj=1 nXj=1 nXj=1 nXj=1 ψij (Za)vj = kMkwiIHa vj ψij − kMkwi ∈ [aH] vjψij − kMkwi ∈ a From Lemma 3.1(ii), we know that 1/kλ is a multiplier so in fact for every 1 ≤ i ≤ n. In turn, by Theorem 3.3, this implies that whence GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS for every 1 ≤ i ≤ n. 21 (cid:3) We now illustrate the importance in the previous theorem that λ be an interior zero of the ideal; the conclusion fails if λ is simply an interior point of the ball. Example 3. Let H be the Hardy space on the unit disc. Let θ ∈ M(H) be an inner function such that if a = θM(H) then there is ψ ∈ M(H) with ψ /∈ a + C1. Choose distinct points w1, w2 ∈ D such that ψ(w1), ψ(w2), θ(w1), θ(w2) are all non-zero. Moreover, choose c1, c2 ∈ C such that ψ(w1) + c1θ(w1) > sup z∈D ψ(z) = kMψk and ψ(w2) + c2θ(w2) = 0. Next, let ϕ ∈ A(H) be such that ϕ(w1) = c1, ϕ(w2) = c2. We find (ψ + θϕ)(w1) > kMψk ≥ kψ(Za)k and (ψ + θϕ)(w2) = 0. Let r = max{w1,w2}, so that 0 < r < 1. Since the function ψ + θϕ is continuous on the closed disc of radius r centred at the origin, we conclude that there is w ∈ D such that (ψ + θϕ)(w) = kψ(Za)k. Nevertheless, we see that ψ(Za) is not a constant multiple of the identity, since ψ /∈ a + C1. (cid:3) As a first application of Theorem 4.2, we compare the norm of the Gelfand transform of an element of Aa to its essential norm in some extremal cases. Corollary 4.4. Let H be a maximal regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal with the property that Ha is infinite-dimensional. Let q : Ta → Oa be the quotient map. Let A = [ϕij(Za)]ij ∈ Mn(A(H)). Then, we have that kAk = max{kχ(n)(A)kMn : χ ∈ ∆(Aa)} if and only if kAk = kq(n)(A)k. Proof. If kAk = kq(n)(A)k, then Theorem 4.2 implies that kAk = max{kχ(n)(A)kMn : χ ∈ ∆(Aa)}. Conversely, assume that the previous equality holds. By Theorem 4.1(i), we see that kAk = max{k[ϕij(λ)]kMn : λ ∈ \supp a}. Let λ ∈ \supp a and choose unit vectors v, w ∈ Cn such that k[ϕij(λ)]kMn = h[ϕij (λ)]v, wi. Since the polynomials are dense in A(H), by definition of the polynomial convex hull we conclude that k[ϕij(λ)]kMn ≤ max µ∈supp ah[ϕij (µ)]v, wi ≤ max µ∈supp ak[ϕij(µ)]kMn . This shows that kAk = max{k[ϕij(λ)]kMn : λ ∈ supp a}. 22 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO Choose λ ∈ supp a such that kAk = k[ϕij(λ)]ijkMn. Assume first that λ ∈ supp a ∩ Sd. By Theorem 4.1(iii), there exists π ∈ ∆(Ta) such that π(Za) = λ and kAk = k[ϕij(λ)]ijkMn = kπ(n)(A)kMn ≤ kg(n) Ta (A)k ≤ kAk. Ta (A)k. That is, kAk = kg(n) Invoking Theorem 4.2, we conclude that kAk = kq(n)(A)k as desired. Thus, it remains only to deal with the case where λ ∈ supp a ∩ Bd = ZBd (a). In this case, it follows from Theorem 4.3 that there are unit vectors v, w ∈ Cn with the property that nXj=1 nXj=1 vjϕij (Za) = kAkwiIHa , 1 ≤ i ≤ n. Therefore, for each χ ∈ ∆(q(Aa)) we find vj(χ ◦ q)(ϕij (Za)) = kAkwi, 1 ≤ i ≤ n or h(χ ◦ q)(n)(A)v, wi = kAk. Since v and w are unit vectors, this forces k(χ ◦ q)(n)(A)kMn = kAk whenever χ ∈ ∆(q(Aa)). Finally, because Ha is infinite-dimensional, the algebra q(Aa) is non-zero and it has at least one character, so that we necessarily have kAk = kq(n)(A)k. (cid:3) It seems relevant given the previous result to mention that we generally do not have that kq(b)k = max{χ(b) : χ ∈ ∆(Aa)} for every b ∈ Aa. Indeed, counter-examples can easily be constructed even on the classical Hardy space on the disc using ideals with support equal to {0, 1}. Our second application of Theorem 4.2 is one of our main results. It is a gen- eralization of part of [43, Proposition 4.10 and Theorem 4.12] and uses essentially the same scheme of proof. Corollary 4.5. Let H be a maximal regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal. The following statements are equivalent. (i) The algebra Aa is hyperrigid in Ta. (ii) The d-tuple Za is essentially normal and the identity representation of Ta is a boundary representation for Aa. Proof. The statement is trivial when Ha is one-dimensional, so we assume without loss of generality that Ha has dimension greater than one. Let q : Ta → Oa be the quotient map. (i) ⇒ (ii): Assume that Aa is hyperrigid in Ta. In particular, the identity repre- sentation of Ta is a boundary representation for Aa; we recall that it is irreducible by Lemma 3.5(iii). Moreover, q(Aa) is hyperrigid in Oa by [11, Corollary 2.2]. This means that the identity representation of Oa has the unique extension property e(q(Aa)) ∼= Oa as explained in Section 2. On the with respect to q(Aa), whence C∗ other hand, it follows from Theorem 4.2 that the Gelfand transform of q(Aa) is GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 23 completely isometric. We conclude that C∗ e(q(Aa)) is commutative by Theorem 2.4, and thus so is Oa. This says precisely that Za is essentially normal. (ii) ⇒ (i): Assume that the d-tuple Za is essentially normal and the identity representation of Ta is a boundary representation for Aa. Now, the unique extension property is preserved upon taking direct sums [11, Proposition 4.4]. Thus, by virtue of Lemma 2.2, to establish the hyperrigidity of Aa it suffices to check that if σ is a unital ∗-representation of Oa, then σ ◦ q has the unique extension property with j=1 q(zj)q(zj )∗ by Lemma 3.5(i). Therefore, we must also have that respect to Aa. To see this, we notice that I =Pd I = dXj=1 (σ ◦ q)(zj )(σ ◦ q)(zj )∗. Moreover, because Za is essentially normal by assumption, (σ ◦ q)(zj ) must be normal for each 1 ≤ j ≤ d. The fact that σ ◦ q has the unique extension property with respect to Aa now follows immediately from Lemma 3.6. (cid:3) In light of the previous theorem, we see that it is of great interest to be able to determine when the identity representation of Ta is a boundary representation for Aa. The following is one of our main results, and it addresses this question. Theorem 4.6. Let H be a maximal regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal. If Ha has dimension greater than one, then the following statements are equivalent. (i) The Gelfand transform of Aa is completely isometric. (ii) The C∗-envelope of Aa is commutative. (iii) The identity representation of Ta is not a boundary representation for Aa. If Ha is one-dimensional, then Aa = Ta ∼= C and every irreducible ∗-representation of Ta is a boundary representation for Aa. Proof. When Ha is one-dimensional, the statement is trivial. We suppose hence- forth that Ha has dimension greater than one. operators by Lemma 3.5(iii), this follows immediately from Corollary 2.5. (i) ⇔ (ii): This is immediate from Theorem 2.4. (i) ⇒ (iii): Since Ha has dimension greater than one and Ta contains the compact (iii) ⇒ (i): It follows from Theorem 2.3 that the quotient map q : Ta → Oa is completely isometric on Aa. Theorem 4.2 then implies that the Gelfand transform of Aa is completely isometric. (cid:3) We now look at some concrete examples. First, we address the classical situation of the Hardy space on the unit disc, which was completely elucidated by Arveson in [6, 7]. Example 4. Let H = H 2 1 be the Hardy space on the unit disc. In this special case, the algebras A(H) and M(H) can be identified completely isometrically with classical algebras of functions. We briefly hint at the well-known details. Recall that H ∞(D) is the algebra of all bounded holomorphic functions on D. It is a Banach algebra when equipped with the norm kψk = sup λ∈Dψ(λ), ψ ∈ H ∞(D). 24 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO The disc algebra A(D) is the closed subalgebra of H ∞(D) consisting of those con- tinuous functions on D that are holomorphic on D. By Inequality (1), we see that there is a unital, injective, contractive homomorphism Θ : M(H) → H ∞(D) such that In fact, it is known that this embedding is isometric and surjective. As we show next, more is true. Θ(Mψ) = ψ, ψ ∈ M(H). Let T ⊂ C denote the unit circle, let m denote Lebesgue measure on T and let Identifying a function in H ∞(D) with its [m]-almost everywhere E = L2(T, m). defined boundary values on T, we see that A(D) ⊂ H ∞(D) ⊂ L∞(T, m), so that in particular A(D) and H ∞(D) are operator algebras. Up to unitary equivalence we have H ⊂ E and Mψ = ψ(U )H, ψ ∈ M(H) where U ∈ B(E) denotes the unitary operator of multiplication by the coordinate ζ. In particular, given [ψij ]ij ∈ Mn(M(H)), we see that k[Mψij ]ijkMn(M(H)) ≤ k[ψij]ijkMn(L∞(T,m)) λ∈Dk[ψij(λ)]ijkMn. = sup The converse inequality always holds, by Inequality (1). Thus, the map Θ is a unital completely isometric isomorphism and Θ(A(H)) = A(D). Now, A(D) ⊂ C(D) and H ∞(D) ⊂ L∞(T, m), and C(D) and L∞(T, m) are unital commutative C∗-algebras. It is then well known that the Gelfand transforms of A(H) and M(H) must be isometric, and hence completely isometric by [50, Theorem 3.9]. This corresponds to the situation where a = {0}. We now consider non-trivial quotients. It follows from Beurling's theorem and Theorem 3.3 that every non-zero weak-∗ closed ideal of M(H) is a principal ideal generated by an inner function. Let θ ∈ M(H) be such an inner function and let a = θM(H). Assume that Ha has dimension greater than one. Let K ⊂ D be the closure of the zero set of θ, along with the points on the unit circle T contained in an arc across which θ cannot be continued holomorphically. Then, K = supp a in this case [14, Theorem 2.4.11]. It follows from [7, Corollary 1 page 291] that the identity representation of Ta is a boundary representation for Aa if and only if K ∩ T is a proper subset of the circle. By Theorem 4.6, we conclude that the Gelfand transform of Aa is completely isometric if and only if K ∩ T = T. (cid:3) For the Drury -- Arveson space in several variables, the situation is more compli- cated. Example 5. Let H = H 2 2 be the Drury -- Arveson space on B2. Let a = {0}. Then, Ha = H and Aa = A(H). It is known (see [9, Theorem 3.3]) that the quotient map q : Ta → Oa is not completely isometric on Aa in this case, whence the identity representation of Ta is a boundary representation for Aa by Theorem 2.3. By Theorem 4.6, we conclude that the Gelfand transform of Aa is not completely isometric. Next, let b ⊂ M(H) be the weak-∗ closed ideal generated by z2. In this case, 1 ), which has a completely isometric Gelfand (cid:3) Ab is unitarily equivalent to A(H 2 transform as seen in Example 4. GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 25 In Section 5, we will explore further concrete conditions on a that guarantee that the Gelfand transform of Aa is not completely isometric. In light of Corollary 4.5 and Theorem 4.6, this will complement [43, Theorem 4.12] (see [44] for updated details). 4.2. The full algebra Ma. Up to now, we devoted our attention to the norm- closed unital algebra Aa generated by z1, . . . , zd inside of Ma. We now turn our focus to the "full" algebra Ma, and examine when its Gelfand transform is com- pletely isometric. The first step we take in our analysis of Ma is the following, which provides an analogous tool to Theorem 4.6. Note that C∗(Ma) trivially con- tains Ta = C∗(Aa), and in particular it contains the ideal of compact operators on Ha by Lemma 3.5(iii). Corollary 4.7. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal such that Ha has dimension greater than one. Consider the following statements. (i) The Gelfand transform of Ma is completely isometric. (ii) The C∗-envelope of Ma is commutative. (iii) There is a closed two-sided ideal J ⊂ C∗(Ma) such that C∗(Ma)/J is commu- tative and the quotient map qJ : C∗(Ma) → C∗(Ma)/J is completely isometric on Ma. (iv) The identity representation of C∗(Ma) is not a boundary representation for Ma. Then we have If we assume in addition that C∗(Ma)/K(Ha) is commutative, then we have (i) ⇔ (ii) ⇔ (iii) ⇒ (iv). (i) ⇔ (ii) ⇔ (iii) ⇔ (iv). Proof. This is a special case of Theorem 2.4 and Corollary 2.5. (cid:3) We remark that one reason that the previous statement is not as complete as that of Theorem 4.6 is that the analogue of Theorem 4.2 fails for Ma (see Example 2). We also note that while it may be tempting, in light of Theorem 3.2, to guess that C∗(Ma)/K(Ha) is always commutative, this is unfortunately not the case (see [37]). Corollary 4.7 thus raises interesting questions regarding quotients of C∗(Ma), which we explore in the following examples. Example 6. Let d ≥ 2 and let H = H 2 d be the Drury-Arveson space on Bd. Let a = {0} so that H = Ha and Ma = M(H). Then, C∗(M(H))/K(H) is not commutative by [37, Theorem 1.2]. Let q : C∗(M(H)) → C∗(M(H))/K(H) be the quotient map. We saw in Example 2 that the Gelfand transforms of q(M(H)) and M(H) are not isometric. Moreover, q is not completely isometric on M(H) as explained in Example 5. Hence the identity representation of C∗(M(H)) is a boundary representation for M(H) by virtue of Theorem 2.3. (cid:3) Next, we consider non-trivial ideals on the Hardy space. Example 7. Let H = H 2 1 be the Hardy space on the unit disc. As explained in Example 4, every weak-∗ closed ideal of M(H) is a principal ideal generated by some inner function. Let θ ∈ M(H) be such an inner function and let a = θM(H). It was shown in [47, Theorem 1] that Ma contains a non-zero compact operator, 26 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO so that the identity representation of C∗(Ma) is a boundary representation for Ma by Theorem 2.3. Moreover, the Gelfand transform of Ma is never isometric when Ha has dimension greater than one, by [18, Theorem 4.2]. (cid:3) We make an elementary remark. Let n ≥ 1 and let A = [ϕij (Za)]ij ∈ Mn(Aa). We see that max{kχ(n)(A)kMn : χ ∈ ∆(Aa)} = max{k[ϕij(λ)]ijkMn : λ ∈ \supp a} ≥ max{k[ϕij(λ)]ijkMn : λ ∈ supp a} = max{kχ(n)(A)kMn : χ ∈ ∆(Ma)} by definition of supp a and by Theorem 4.1(i). Therefore, if the Gelfand transform of Ma is completely isometric, then so is that of Aa. The converse does not generally hold; see Examples 4 and 7. The phenomena described in the previous example are not representative of the general multivariate situation, as we illustrate next. Example 8. Let H = H 2 2 be the Drury -- Arveson space on B2. Let a = {0}. We saw in Example 5 that the Gelfand transform of Aa is not completely isometric, and so neither is that of Ma by the remark above. Next, let b ⊂ M(H) be the weak-∗ closed ideal generated by z2. In this case, Mb is unitarily equivalent to M(H 2 1 ), which has a completely isometric Gelfand transform as seen in Example 4. (cid:3) We close this section by giving an example where an ideal J fitting into the framework of Corollary 4.7 can be identified. Example 9. Let H = H 2 1 be the Hardy space on the unit disc. Let a = {0} so that H = Ha and Ma = M(H). In that case, there is a unital completely isometric isomorphism Θ : M(H) → H ∞(D) such that Θ(Mψ) = ψ, ψ ∈ M(H) (see Example 4 for details). Let W ⊂ B(H) be the C∗-algebra generated by the Toeplitz operators with symbols in L∞(T, m), so that M(H) ⊂ W (see [32, Chapter 7] for details). Let C ⊂ W denote the commutator ideal of W. It follows from [32, Theorem 7.11] that there is a ∗-isomorphism ρ : W/C → L∞(T, m) such that ρ(Mψ + C) = ψ, ψ ∈ M(H). Next, let J = C ∩ C∗(M(H)). A standard theorem for C∗-algebras [16, Corollary II.5.1.3] implies that there is a ∗-isomorphism Ξ : C∗(M(H))/J → (C∗(M(H)) + C)/C such that Let qJ : C∗(M(H)) → C∗(M(H))/J be the quotient map. We then note that Ξ(Mψ + J) = Mψ + C, ψ ∈ M(H). qJM(H) = Ξ−1 ◦ ρ−1 ◦ Θ so that qJ is completely isometric on M(H). Moreover, W/C is commutative by choice of C, and thus so is C∗(M(H))/J = Ξ−1((C∗(M(H)) + C)/C). GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 27 Applying Corollary 4.7, we can now recover the classical fact that the Gelfand transform of M(H) is completely isometric and that the identity representation of C∗(M(H)) is not a boundary representation for M(H). (cid:3) It is known that the ideal J above contains the compact operators [32, Propo- sition 7.12], but the containment is proper. Indeed, it is readily seen that Mθ is not essentially normal whenever θ is an inner function but not a finite Blaschke product, so in particular C∗(M(H))/K(H) is not commutative. 5. The Gelfand transform of Aa often fails to be completely isometric In the previous section, we saw that the determining whether the identity repre- sentation of Ta is a boundary representation for Aa is equivalent to deciding whether the Gelfand transform of Aa fails to be completely isometric. In this section, we seek concrete conditions that address the latter property. First, we improve on part of [24, Theorem 8.8], removing the essential normality condition that was required therein. Recall that a weak-∗ closed ideal a ⊂ M(H) is said to be homogeneous if it is the weak-∗ closure of a polynomial ideal generated by some collection of homogeneous polynomials. Theorem 5.1. Let H be a maximal regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed homogeneous ideal such that Ha has dimension greater than one. If Aa is not completely isometrically isomor- phic to the disc algebra A(D), then the Gelfand transform of Aa is not completely isometric. Proof. By Theorems 2.3 and 4.6, we must show that the quotient map q : Ta → Oa is not completely isometric on Aa. For A = [ϕij (Za)]ij ∈ Mn(Aa), it follows from Theorems 4.1 and 4.2 that kq(n)(A)k = max{kχ(n)(A)kMn : χ ∈ ∆(Ta)} ≤ max{k[ϕij(ζ)]ijkMn : ζ ∈ ZBd ≤ kAk. (a ∩ A(H)) ∩ Sd} Since we assume that Aa is not completely isometric to the disc algebra, we may now argue as in the proof of [24, Theorem 8.8] to see that q is not completely isometric on Aa. (cid:3) Up to now, our attention has been mostly devoted to maximal spaces. The next result completely settles the non-maximal case. Theorem 5.2. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space which is not maximal. Let a ⊂ M(H) be a proper weak-∗ closed ideal such that Ha has dimension greater than one. Then, the Gelfand transform of Aa is not completely isometric. Proof. By assumption, there is a sequence (bn) of non-negative numbers such that n=1 bn < 1 and with the property that P∞ 1 k(z, w) 1 − = ∞Xn=1 bnhz, win, z, w ∈ Bd. 28 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO n=1 bn. For every λ ∈ Bd, we see that For convenience, put r =P∞ ∞Xn=1 Consider the row operator R =(cid:18)b1/2 3.5(ii) that bnkλk2n ≤ r < 1. α (cid:16) α! α!(cid:17)1/2 Z α a(cid:19)α≥1 . It follows from Lemma RR∗ = I − PHa PC1Ha. Since Ha has dimension at least two, we can find two orthogonal unit vectors f, g ∈ Ha. If f (0) 6= 0, then we may consider the non-zero vector h = g − g(0) f (0) f ∈ Ha which satisfies h(0) = 0. Thus, we conclude that Ha always contains a non-zero vector h such that PC1h = 0. We find so that kRk = 1. Hence, there is N ≥ 1 such that the truncated row RR∗h = (I − PHaPC1Ha )h = h RN = b1/2 α (cid:18)α! a!1≤α≤N α! (cid:19)1/2 Z α satisfies kRNk > √r. On the other hand, given χ ∈ ∆(Aa), by Theorem 4.1(ii) there is λ ∈ ZBd (a ∩ A(H)) such that α (cid:18)α! α! (cid:19)1/2 a!!1≤α≤N Z α (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) χ b1/2 2 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) = ≤ NXn=1 ∞Xn=1 n! α!λα2 bn Xα=n bnkλk2n ≤ r. We conclude that Gelfand transform of Aa is not completely isometric. (cid:3) Next, we explore "size" restrictions on the boundary portion of the support of a. Our approach hinges on the following fact, which generalizes a result on page 444 of [55]. To simplify the statement, we introduce some terminology. Given a compact subset K ⊂ Bd, we say that it is an approximation set if the polynomials are uniformly dense in the continuous functions on K. Moreover, given a unital contractive representation ρ of A(H), we say that K is a spectral set for ρ(A(H)) if kρ(ϕ)k ≤ max w∈K ϕ(w), ϕ ∈ A(H). Lemma 5.3. Let H be a maximal regular unitarily invariant complete Nevanlinna -- Pick space with kernel k. Let (bn) be the sequence of non-negative real numbers such that 1 k(z, w) 1 − = ∞Xn=1 bnhz, win, z, w ∈ Bd. GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 29 Let ρ : A(H) → B(E) be a unital contractive homomorphism such that ρ(A(H)) admits an approximation set K ⊂ Sd as a spectral set. Then, (ρ(z1), . . . , ρ(zd)) is a spherical unitary and ∞Xn=1 bn Xα=n n! α! ρ(zα)ρ(zα)∗ = I. Proof. Let P(K) ⊂ C(K) denote the space of polynomials. Because K is a spectral set for ρ(A(H)), we find kρ(p)k ≤ max w∈K p(w), p ∈ P(K). On the other hand, since K is an approximation set, by density we obtain a unital contractive homomorphism π : C(K) → B(E) such that p ∈ P(K). π(p) = ρ(p), Thus, π must be a ∗-homomorphism [50, Proposition 2.11] and because K ⊂ Sd we see that so that (ρ(z1), . . . , ρ(zd)) is a spherical unitary. Hence where the last equality follows by maximality of H. We can now apply the previous theorem to obtain a sufficient condition for the Gelfand transform of Aa not to be completely isometric. Corollary 5.4. Let H be a maximal regular unitarily invariant complete Nevanli- nna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal such that Ha has dimension greater than one and supp a ∩ Sd is an approximation set. Then, the Gelfand transform of Aa is not completely isometric. Proof. Assume that the quotient map q : Ta → Oa is completely isometric on Aa. By Theorems 4.1(iv) and 4.2, we see that see that supp a ∩ Sd is a spectral set for Applying Lemma 5.3 we find {ϕ(Za) : ϕ ∈ A(H)}. ∞Xn=1 bn Xα=n n! α! Z α a Z α∗ a = I. But this contradicts Lemma 3.5(ii). Thus, q is not completely isometric on Aa, and the conclusion follows from Theorems 2.3 and 4.6. (cid:3) dXj=1 ρ(zj)ρ(zj)∗ = π zj2 = I dXj=1 ρ(zj)ρ(zj )∗ bn ∞Xn=1 dXj=1 ∞Xn=1 ρ(zα)ρ(zα)∗ = bnI = I = ∞Xn=1 bn Xα=n n! α! n (cid:3) 30 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO 3 and a = hz1i. We note that the Gelfand transform of Aa can certainly fail to be completely isometric without supp a∩ Sd being an approximation set: this happens for instance when H = H 2 There are known concrete sufficient conditions that guarantee that a subset is an approximation set. We refer the reader to the papers quoted in the proof for the precise definitions of the properties appearing in the statement below. Corollary 5.5. Let H be a maximal regular unitarily invariant complete Nevanli- nna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal such that Ha has dimension greater than one. Let K = supp a ∩ Sd. Then, the Gelfand transform of Aa is not completely isometric whenever one of the following conditions holds. (i) K ⊂ Rd. (ii) K is polynomially convex and has zero 2-Hausdorff measure. (iii) K is a polynomially convex real analytic subvariety of Cd. (iv) K is totally null for the ball algebra. Proof. In view of Corollary 5.4, it suffices to show that K is an approximation set whenever one of these conditions hold. See [46, pages 95, 97 and 116] for (i),(ii) and (iii) and see [56, Theorem 10.1.2] for (iv). (cid:3) We close this section with a generalization of results from [44] and [43]. Therein, geometric conditions on ZBd (a) are given that ensure that the identity represen- tation of Ta is a boundary representation for Aa. Our generalization will lead to a refinement of Corollary 4.5, thus bringing the resulting statement closer to that found in [44]. Though the argument we present here is closely modeled on that of [44], we sketch the details for the benefit of the reader so that the required modifications of the argument are evident. For 0 < ν ≤ 1, we define a kernel k(ν) on Bd by the formula k(ν)(z, w) = 1 (1 − hz, wi)ν = 1 + νhz, wi + ∞Xn=2 ν(ν + 1)··· (ν + n − 1) n! hz, win. For the remainder of the section, we will focus entirely on the kernels k(ν) and on the associated reproducing kernel Hilbert spaces of holomorphic functions Hν. We make some preliminary observations. It follows from [42, Lemma 2.3 and Section 4] that among maximal regular unitarily invariant completely Nevanlinna -- Pick spaces on Bd, the scale {Hν : 0 < ν ≤ 1} consists precisely of those that are well behaved with respect to composition with automorphisms of the ball. More precisely, if Θ : Bd → Bd is a conformal automorphism, then there is a unitary operator VΘ : Hν → Hν such that VΘMψ = Mψ◦ΘVΘ, ψ ∈ M(Hν). For each 1 ≤ j ≤ d, we let θj : Bd → C be corresponding component of Θ, so that Θ = (θ1, . . . , θd). It follows from [56, Theorem 2.2.5] and [27, Theorem 2.10] that θj ∈ A(Hν ). We also find that VΘMxj V ∗ Θ, then we see that b is another proper weak-∗ closed ideal. Moreover, it is readily verified Next, let a ⊂ M(Hν ) be a proper weak-∗ closed ideal. If we let b = VΘaV ∗ Θ = Mθj . GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 31 that VΘHa = Hb. Consider the unitary operator W = VΘHa : Ha → Hb. Writing Za = (z1, . . . , zd), we compute that W zjW ∗ = θj(Zb) for every 1 ≤ j ≤ d. In turn, this implies that WAaW ∗ = Ab. We record a useful technical fact. Lemma 5.6. Let 0 < ν ≤ 1 and let a ⊂ M(Hν ) be a proper weak-∗ closed ideal. Let Θ : Bd → Bd be a conformal automorphism. Put b = VΘaV ∗ Θ and λ = Θ(0). Then, if and only if (C(z1 − λ1) + . . . + C(zd − λd)) ∩ a = {0} (Cz1 + . . . + Czd) ∩ b = {0}. C such that r = Pd θ1, . . . , θd ∈ A(Hν ). By definition of b, we see thatPd Proof. Let r ∈ (C(z1 − λ1) + . . . + C(zd − λd)) ∩ a. Thus, there are c1, . . . , cd ∈ j=1 cj(zj − λj). As before, we write Θ = (θ1, . . . , θd) with j=1 cj(θj − λj ) = r ◦ Θ ∈ b. Now, by [56, Theorem 2.2.5], there is a µ ∈ Bd such that the function (1 − hz, µi)θj is a degree-one polynomial for every 1 ≤ j ≤ d. We infer that (1 − hz, µi)(r ◦ Θ) = cj((1 − hz, µi)θj − λj(1 − hz, µi)) dXj=1 is a degree-one polynomial in b. On the other hand, because r vanishes at λ, this degree-one polynomial must vanish at 0, and thus must lie in (Cz1 + . . . + Czd)∩ b. The converse can be proved similarly. (cid:3) an ideal such that 0 ∈ ZBd (J ) and Tϕ∈J ker ϕ′(0) = {0}. Then, 0 is an isolated Next, we need a standard generalization of the fact that a function whose deriv- ative is injective at a point must itself be injective near that point. In what follows, given a holomorphic function ϕ : Bd → C, we view its derivative ϕ′(0) as a linear map from Cd to C. Lemma 5.7. Let A be an algebra of holomorphic functions on Bd. Let J ⊂ A be point of ZBd (J ). Proof. Let G denote the ring of germs of analytic functions at 0. Given a function g analytic on a neighborhood of 0, we let [g] denote the corresponding germ. Because G is Noetherian, the ideal h[ϕ] : ϕ ∈ J i is generated by [ϕ1], . . . , [ϕm] for some ϕ1, . . . , ϕm ∈ J [39, Theorem II.E.3]. j=1 ker ϕ′ j (0). Let ϕ ∈ J . Correspondingly, choose j=1[gj][ϕj ]. As 0 ∈ Suppose v ∈ Cd lies in Tm g1, . . . , gd analytic on a neighborhood of 0 such that [ϕ] = Pm ZBd (J ), we find ϕ1(0) = . . . = ϕm(0) = 0 and thus ϕ′(0)v = gj(0)ϕ′ j(0)v = 0. mXj=1 By assumption, we infer that v = 0. If we let D : Cd → Cm be the matrix defined , then we see that D is injective. Hence, there is a d× m complex as D = ϕ′ ϕ′ 1(0) 2(0) ... m(0) ϕ′  32 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO matrix C such that CD = Id. Let Ψ = C ϕ1 ϕ2 ... ϕm  . If we write Ψ = ψ1 ψ2 ... ψd  then we have that ψ1, . . . , ψd ∈ J and Ψ′(0) = CD = Id. Thus, Ψ is injective near 0, which implies that 0 is an isolated point of ZBd (ψ1, . . . , ψd), and hence of ZBd (J ). (cid:3) The crucial step of the argument can now be taken, following [44]. Theorem 5.8. Let 0 < ν ≤ 1 and let a ⊂ M(Hν) be a proper weak-∗ closed ideal. Assume that there is a point λ ∈ ZBd(a) that is not an isolated point of ZBd (a). Let q : Ta → Oa denote the quotient map. Then, q is not completely isometric on Aa under either of the following sets of conditions: (i) 0 < ν < 1, or (ii) ν = 1 and (C(z1 − λ1) + . . . + C(zd − λd)) ∩ a = {0}. In particular, the Gelfand transform of Aa is not completely isometric under these conditions. Proof. Because λ is not an isolated point of ZBd(a), there are infinitely many points in ZBd (a). In particular, the infinite set {kµ : µ ∈ ZBd (a)} is linearly independent in Ha, so that Ha is infinite-dimensional. Therefore, we focus on showing that q is not completely isometric on Aa, since the statement about the Gelfand transform follows immediately from this, by virtue of Theorems 2.3 and 4.6. We start with a reduction. Let Θ : Bd → Bd be a conformal automorphism such that Θ(0) = λ. Let b = VΘaV ∗ Θ. We observed in the discussion preceding Lemma 5.6 that Aa and Ab are unitarily equivalent, so it is equivalent to show that the quotient map Tb → Ob is not completely isometric on Ab. It follows from Lemma 5.6 that we may as well assume that λ = 0. We make one more reduction before starting the proof. Invoke Lemma 5.7 to operator U : Cd → Cd such that U e1 = v, where e1 = (1, 0, . . . , 0) ∈ Cd. Let Ω : Bd → Bd be the conformal automorphism defined as Ω(z) = U z. Let c = VΩaV ∗ Ω . can thus assume that this last statement is true of the ideal a itself. infer the existence of a unit vector v ∈Tϕ∈a ker ϕ′(0). Choose a constant unitary Then, we see that e1 ∈Tϕ∈c ker ϕ′(0). Arguing as in the previous paragraph, we 0 ∈ ZBd (a) and e1 ∈Tϕ∈a ker ϕ′(0), we have z1 ∈ [aHν]⊥ = Ha. Also, it follows from [42, Remark 7.1(iv)] that kz1k2 = 1/ν. Next, because 0 ∈ ZBd (a) we see that 1 ∈ Ha so that Having made these reductions, we start the argument. We note first that because kz1k2 ≥ kz11k2 = kz1k2 = 1/ν where we recall that we use the notation Za = (z1, . . . , zd). In case (i), we see that kz1k > 1. On the other hand, it readily follows from Theorems 4.1 and 4.2 that kq(z1)k ≤ 1, so that q is not even isometric on Aa in this case. For the rest of the proof, we focus on case (ii), and so we assume that ν = 1 and that (Cz1 + . . . + Czd) ∩ a = {0}. The argument here is identical to that of [44]. The fact that z2, . . . , zd /∈ a implies that z21, . . . , zd1 are all non-zero vectors in Ha GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 33 by Theorem 3.3. Let C be the column operator (z1, . . . , zd). Then kCk2 ≥ dXj=1 kzj1k2 = 1/ν + dXj=2 kzj1k2 > 1/ν ≥ 1. Once again, Theorems 4.1 and 4.2 show that the essential norm of C is at most 1, so indeed qAa is not completely isometric in this case as well. (cid:3) We can now refine Corollary 4.5 in some special cases. Corollary 5.9. Let 0 < ν ≤ 1 and let a ⊂ M(Hν) be a proper weak-∗ closed ideal. Assume that there is a point λ ∈ ZBd (a) that is not an isolated point of ZBd (a). Assume also that either (i) 0 < ν < 1, or (ii) ν = 1 and (C(z1 − λ1) + . . . + C(zd − λd)) ∩ a = {0}. Then, Aa is hyperrigid in Ta if and only if Za is essentially normal. Proof. Combine Theorem 5.8 with Theorem 2.3 and Corollary 4.5. (cid:3) 6. Isolated points in the character space In this final section, much in the spirit of Section 5, we seek concrete conditions that detect whether or not the Gelfand transform is completely isometric. More precisely, we aim to leverage the presence of isolated points in the character space. We start with the norm-closed algebra Aa, where the situation is more transparent. Theorem 6.1. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal. Assume that \supp a consists of more than one point and has an isolated point λ ∈ Bd. Then, the Gelfand transform of Aa is not isometric. Proof. By Theorem 4.1(i), there is a unique character τλ ∈ ∆(Aa) such that τλ(Za) = λ, and this character is easily seen to be an isolated point of ∆(Aa). By the Shilov idempotent theorem [38, Corollary III.6.5], there is an idempotent element b ∈ Aa such that τλ(b) = 1 and γ(b) = 0 for every γ ∈ ∆(Aa)\{τλ}. An ap- proximation argument then yields a multiplier θ ∈ A(H) such that θ(λ) > 1/2 and γ(θ(Za)) < 1/2 for every γ ∈ ∆(Aa) \ {τλ}. Since ∆(Aa) is not a mere singleton, we infer that θ(Za) is not a constant multiple of the identity, so kθ(Za)k > θ(λ) by Theorem 4.3. Hence, kθ(Za)k > max{χ(θ(Za)) : χ ∈ ∆(Aa)} and the Gelfand transform of Aa is not isometric. (cid:3) Before giving a sufficient condition in order for the previous result to be appli- cable, we need a preliminary observation. Recall that a sequence Λ ⊂ Bd is said to be interpolating for M(H) if the restriction map ρ : M(H) → ℓ∞(Λ) defined as ρ(ψ) = ψΛ, ψ ∈ M(H) is surjective. These sequences have recently been characterized in [4]. We will say that a subset Ω ⊂ Bd is a zero set for A(H) if there is a closed ideal J ⊂ A(H) such that Ω = ZBd (J ). 34 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO Lemma 6.2. Let H be the Drury -- Arveson space on Bd and let Λ ⊂ Bd be an interpolating sequence for M(H). Assume that Λ is a zero set for A(H). Let a = {ψ ∈ M(H) : ψΛ = 0}. Then, supp a = Λ and thus supp a is polynomially convex. Proof. We know that supp a = {χ(z1, . . . , zd) : χ ∈ ∆(Ma)} so that supp a consists of those points λ = (λ1, . . . , λd) ∈ Cd such that (z1 − λ1I)Ma + . . . + (zd − λdI)Ma 6= Ma. It then readily follows from [26, Corollary 4.7] that supp a = Λ. Thus, supp a is a zero set for A(H), and hence is automatically polynomially convex. (cid:3) We can now give an application of Theorem 6.1. Recall that a closed subset K ⊂ Sd is said to be Ad-totally null if µ(K) = 0 for every Ad-Henkin measure µ on Sd (see [22] for details). Corollary 6.3. Let H be the Drury -- Arveson space on Bd and let Λ ⊂ Bd be an interpolating sequence for M(H) consisting of more than one point. Assume that Λ ∩ Sd is Ad-totally null. Let a = {ψ ∈ M(H) : ψΛ = 0}. Then, the Gelfand transform of Aa is not isometric. Proof. It follows from [23, Corollary 5.13] that Λ is a zero set for A(H). Apply Lemma 6.2 to see that \supp a = supp a = Λ. The proof is complete upon applying Theorem 6.1, as an interpolating sequence cannot have an accumulation in Bd. (cid:3) For the remainder of the section we will turn out attention to Ma. To begin we collect some relevant information about the character space of Ma, which is typically more complicated than that of Aa. Lemma 6.4. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal and let λ ∈ ZBd(a). The following statements hold. (i) There is a unique character τλ ∈ ∆(Ma) such that λ = (τλ(z1), . . . , τλ(zd)). Furthermore, this character is weak-∗ continuous and satisfies τλ(ψ(Za)) = ψ(λ), ψ ∈ M(H). (ii) The point λ is an isolated point of ZBd (a) if and only if τλ is an isolated point of ∆(Ma). Proof. (i) Let ελ be the weak-∗ continuous evaluation character on M(H) corre- sponding to λ. Recall that the map Γ : M(H) → Ma defined as Γ(ψ) = ψ(Za), ψ ∈ M(H) is a weak-∗ continuous unital surjective homomorphism. Moreover, we have that ker Γ = a. Because λ ∈ ZBd (a), we see that a ⊂ ker ελ. Thus, there is a weak-∗ continuous character τλ ∈ ∆(Ma) such that τλ ◦ Γ = ελ. Then, we have λ = (τλ(z1), . . . , τλ(zd)). GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 35 To show that τλ is unique, let χ ∈ ∆(Ma) such that λ = (χ(z1), . . . , χ(zd)). Then, χ ◦ Γ ∈ ∆(M(H)) and ((χ ◦ Γ)(Mz1), . . . , (χ ◦ Γ)(Mzd)) = λ. It follows from [42, Proposition 8.5] that χ ◦ Γ = ελ, whence χ = τλ since Γ is surjective. (ii) Assume that λ ∈ ZBd (a) is an isolated point. Choose r > 0 small enough so that B(λ, r) ⊂ Bd and B(λ, r) ∩ ZBd (a) = {λ}. To see that τλ is an isolated point of ∆(Ma), consider U = {χ ∈ ∆(Ma) : χ(Za) ∈ B(λ, r)} which is a weak-∗ open neighbourhood of τλ in ∆(Ma). We claim that U = {τλ}. Indeed, let χ ∈ U and put µ = χ(Za). Then, we see that µ ∈ Bd by choice of r. But we also have that µ ∈ supp a by definition of the support, so that µ ∈ supp a ∩ Bd = ZBd (a). Hence µ ∈ ZBd (a) ∩ B(λ, r) = {λ}. Invoking (i) we infer that τλ = χ. ψ1, . . . , ψn ∈ M(H) and ε > 0 such that Conversely, assume that τλ is an isolated point of ∆(Ma). Thus, there are {χ ∈ ∆(Ma) : χ(ψj (Za)) − τλ(ψj(Za)) < ε, 1 ≤ j ≤ d} = {τλ}. Choose δ > 0 small enough so that if µ ∈ Bd satisfies kµ − λk < δ, then ψj(µ) − ψj(λ) < ε, 1 ≤ j ≤ d. By (i), we conclude that if µ ∈ ZBd (a) ∩ B(λ, δ) then τµ(ψj(Za)) − τλ(ψj(Za)) < ε, 1 ≤ j ≤ d whence τµ = τλ and µ = λ. Therefore λ is an isolated point of ZBd (a). (cid:3) We mention in passing one basic consequence of the previous lemma. Roughly speaking, it says that if the Gelfand transform of Ma is isometric, then the support of a cannot be contained in the open ball. This will be useful below. Lemma 6.5. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal such that Ha has dimension greater than one and supp a ⊂ Bd. Then, the Gelfand transform of Ma is not isometric. Proof. Note that if a + C1 = M(H), then a is maximal and thus Ha is one- dimensional by Lemma 3.4. Hence, we may assume that there is ψ ∈ M(H) that does not lie in a + C1. It thus follows from Theorem 4.3 that kψ(Za)k > ψ(λ) for every λ ∈ ZBd (a). Next, let χ ∈ ∆(Ma) and let µ = χ(Za). By assumption, we see that µ ∈ supp a ⊂ Bd, so that µ ∈ ZBd (a). But using Lemma 6.4(i) we find χ(ψ(Za)) = ψ(µ). By the first paragraph, we thus conclude that χ(ψ(Za)) < kψ(Za)k. Since χ ∈ ∆(Ma) was arbitrary, we conclude that the Gelfand transform of Ma is not isometric. (cid:3) 36 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO The main thrust for what is to come is given by the following, which is the analogue of Theorem 6.1 in the context of Ma. Theorem 6.6. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal. Assume that supp a has an isolated point λ ∈ Bd. Then, the following statements hold. (i) The algebra Ma contains a non-zero idempotent finite-rank operator and the identity representation of C∗(Ma) is a boundary representation for Ma. More- over, the Gelfand transform of Ma is not completely isometric if Ha has di- mension greater than one. (ii) If supp a consists of more than one point, then the Gelfand transform of Ma is not isometric. Proof. (i) By [27, Theorem 3.8], we know that the Taylor spectrum of Za coincides with {χ(Za) : χ ∈ ∆(Ma)} = supp a. By assumption, λ is an isolated point of supp a, and hence of the Taylor spectrum of Za. Using [27, Theorem 2.1(iii)], we obtain a non-zero idempotent E ∈ {Za}′′ such that the Taylor spectrum of Zaran E consists only of {λ}. On the other hand, we may invoke [27, Theorem 3.1 and Lemma 4.14] to see that there are commuting nilpotent operators N1, . . . , Nd on ran E such that zjran E = λjIran E + Nj, 1 ≤ j ≤ d. Now, since PHa 1 is cyclic for Za, the vector EPHa1 is cyclic for Zaran E, and it is readily seen that this forces ran E to be finite-dimensional. Finally, we may apply the Commutant Lifting Theorem [13, Theorem 5.1] to obtain ω ∈ M(H) such that E = ω(Za). Then, ω(Za) has finite-rank. By virtue of Theorem 2.3, we conclude that the identity representation of C∗(Ma) is a boundary representation for Ma. Applying Corollary 4.7, we also see that the Gelfand transform of Ma is not completely isometric whenever Ha has dimension greater than one. (ii) To see that the Gelfand transform of Ma is not isometric when supp a consists of more than one point, we argue as in the proof of Theorem 6.1. Let τλ ∈ ∆(Ma) be as in Lemma 6.4(i). By Lemma 6.4(ii), we see that τλ is an isolated point of ∆(Ma). By the Shilov idempotent theorem, there is a multiplier θ ∈ M(H) such that θ(λ) = 1 and γ(θ(Za)) = 0 for every γ ∈ ∆(Ma) \ {τλ}. Since ∆(Ma) is not a singleton, we infer that θ(Za) is not a constant multiple of the identity, so kθ(Za)k > θ(λ) by Theorem 4.3. Hence, kθ(Za)k > max{χ(θ(Za)) : χ ∈ ∆(Ma)} and the Gelfand transform of Ma is not isometric. (cid:3) We note that the topological restrictions on the support of the ideal in Theorems 6.1 and 6.6 cannot simply be removed entirely as Examples 5 and 8 illustrate. Furthermore, the condition that supp a has an isolated point that lies in the interior of the unit ball is important for the proof Theorem 6.6(i) to work, as we show next. Example 10. Let H = H 2 1 be the Hardy space on the unit disc. Let θ ∈ M(H) be the singular inner function z − 1(cid:19) , θ(z) = exp(cid:18) z + 1 z ∈ D. GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 37 Let a = θM(H). It is well known that the spectrum of Za is supp a = {1} in this case, so that the multiplier ω from the proof of Theorem 6.6(i) can be taken to be the constant function 1. Thus, ω(Za) = I here. But this operator is not compact since Ha is easily seen to be infinite-dimensional. (cid:3) As mentioned in Example 7, in the classical case where H is the Hardy space on the disc, the Gelfand transform Ma is never isometric, unless Ha is one-dimensional [18, Theorem 4.2]. We suspect that this a reflection of a more general phenomenon that we will explore further in upcoming work. We substantiate our suspicion in the following result. Corollary 6.7. Let H be a regular unitarily invariant complete Nevanlinna -- Pick space. Let a ⊂ M(H) be a proper weak-∗ closed ideal with the property that ZBd (a) has an isolated point. Then, the following statements are equivalent. (i) The ideal a is the vanishing ideal of a single point. (ii) The space Ha is one-dimensional. (iii) The Gelfand transform of Ma is completely isometric. (iv) The Gelfand transform of Ma is isometric. Proof. (i) ⇒ (ii): We assume that there is λ ∈ Bd with the property that a = {ψ ∈ M(H) : ψ(λ) = 0}. Then, kλ ∈ Ha. Next, we note that Ckλ is co-invariant for M(H), so by Theorem 3.3 there is a weak-∗ closed ideal b ⊂ M(H) such that Ckλ = Hb and a ⊂ b. On the other hand, the equality [bH] = (Ckλ)⊥ implies that every multiplier in b vanishes at λ, whence b ⊂ a. We conclude that Ha = Hb = Ckλ. (ii) ⇒ (i): It follows from Lemma 3.4 that a is a maximal ideal, so there is a character χ ∈ ∆(Ma) such that a = ker χ. By assumption, there is λ ∈ ZBd (a). Let τλ ∈ ∆(Ma) be the corresponding character of evaluation at λ (see Lemma 6.4(i)). Then, we have which forces χ = τλ and a = ker τλ as desired. ker χ = a ⊂ ker τλ (ii) ⇒ (iii),(iii) ⇒ (iv) : This is trivial. (iv) ⇒ (ii): By Theorem 6.6(ii), we see that supp a consists of only one point lying in Bd. But then Lemma 6.5 forces Ha to be one-dimensional. (cid:3) The previous corollary can be seen as a partial multivariate analogue of [18, Theorem 4.2]. We close the paper with a result that illustrates that the topological condition of having an isolated zero can be removed entirely in some cases, leveraging our generalization of [44]. Corollary 6.8. Let 0 < ν ≤ 1 and let Hν be the space corresponding to the kernel k(ν)(z, w) = z, w ∈ Bd. 1 (1 − hz, wi)ν , Let a be a proper weak-∗ closed ideal of M(Hν) and let λ ∈ ZBd (a). Assume that supp a consists of more than one point and that either (i) 0 < ν < 1, or (ii) ν = 1 and (C(z1 − λ1) + . . . + C(zd − λd)) ∩ a = {0}. 38 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO Then, the Gelfand transform of Ma is not completely isometric. If, in addition, supp a is polynomially convex, then the Gelfand transform of Aa is not completely isometric. Proof. First suppose that λ is not an isolated point of ZBd (a). Then, it follows from Theorem 5.8 that the Gelfand transform of Aa is not completely isometric. As explained in the remark following Example 7, the Gelfand transform of Ma is not completely isometric either. Next, suppose that λ is an isolated point of ZBd (a). Since supp a consists of more than one point, it follows from Theorem 6.6(ii) that the Gelfand transform of Ma is not isometric. If supp a is polynomially convex, then we see that λ is also an isolated point of \supp a, and it now follows from Theorem 6.1 that the Gelfand transformation of Aa is not completely isometric. (cid:3) References [1] Jim Agler, The Arveson extension theorem and coanalytic models, Integral Equations Oper- ator Theory 5 (1982), no. 5, 608 -- 631. MR697007 (84g:47011) [2] Jim Agler and John E. McCarthy, Complete Nevanlinna-Pick kernels, J. Funct. Anal. 175 (2000), no. 1, 111 -- 124. MR1774853 [3] , Pick interpolation and Hilbert function spaces, Graduate Studies in Mathematics, vol. 44, American Mathematical Society, Providence, RI, 2002. MR1882259 (2003b:47001) [4] Alexandru Aleman, Michael Hartz, John E. McCarthy, and Stefan Richter, Interpolating sequences in spaces with the complete pick property, Int. Math. Res. Not. IMRN 12 (2019), 3832 -- 3854. MR3973111 [5] C.-G. Ambrozie, M. Englis, and V. Muller, Operator tuples and analytic models over general domains in Cn, J. Operator Theory 47 (2002), no. 2, 287 -- 302. MR1911848 (2004c:47013) [6] William Arveson, Subalgebras of C ∗-algebras, Acta Math. 123 (1969), 141 -- 224. MR0253059 (40 #6274) [7] [8] [9] [10] [11] , Subalgebras of C ∗-algebras. II, Acta Math. 128 (1972), no. 3-4, 271 -- 308. MR0394232 (52 #15035) , An invitation to C ∗-algebras, Springer-Verlag, New York-Heidelberg, 1976. Graduate Texts in Mathematics, No. 39. MR0512360 , Subalgebras of C ∗-algebras. III. Multivariable operator theory, Acta Math. 181 (1998), no. 2, 159 -- 228. MR1668582 (2000e:47013) , The noncommutative Choquet boundary, J. Amer. Math. Soc. 21 (2008), no. 4, 1065 -- 1084. MR2425180 (2009g:46108) , The noncommutative Choquet boundary II: hyperrigidity, Israel J. Math. 184 (2011), 349 -- 385. MR2823981 [12] Ameer Athavale, On the intertwining of joint isometries, J. Operator Theory 23 (1990), no. 2, 339 -- 350. MR1066811 [13] Joseph A. Ball, Tavan T. Trent, and Victor Vinnikov, Interpolation and commutant lifting for multipliers on reproducing kernel Hilbert spaces, Operator theory and analysis (Amsterdam, 1997), 2001, pp. 89 -- 138. MR1846055 [14] Hari Bercovici, Operator theory and arithmetic in H∞, Mathematical Surveys and Mono- graphs, vol. 26, American Mathematical Society, Providence, RI, 1988. MR954383 (90e:47001) [15] Kelly Bickel, Michael Hartz, and John E. McCarthy, A multiplier algebra functional calculus, Trans. Amer. Math. Soc. 370 (2018), no. 12, 8467 -- 8482. MR3864384 [16] B. Blackadar, Operator algebras, Encyclopaedia of Mathematical Sciences, vol. 122, Springer- Verlag, Berlin, 2006. Theory of C ∗-algebras and von Neumann algebras, Operator Algebras and Non-commutative Geometry, III. MR2188261 [17] David P. Blecher and Christian Le Merdy, Operator algebras and their modules -- an operator space approach, London Mathematical Society Monographs. New Series, vol. 30, The Claren- don Press, Oxford University Press, Oxford, 2004. Oxford Science Publications. MR2111973 [18] Raphael Clouatre, Completely bounded isomorphisms of operator algebras and similarity to complete isometries, Indiana Univ. Math. J. 64 (2015), no. 3, 825 -- 846. MR3361288 GELFAND TRANSFORMS OF NEVANLINNA -- PICK QUOTIENTS 39 [19] , Unitary equivalence and similarity to Jordan models for weak contractions of class C0, Canad. J. Math. 67 (2015), no. 1, 132 -- 151. MR3292697 [20] Raphael Clouatre, Non-commutative peaking phenomena and a local version of the hyper- rigidity conjecture, Proc. Lond. Math. Soc. (3) 117 (2018), no. 2, 221 -- 245. MR3851322 [21] , Unperforated pairs of operator spaces and hyperrigidity of operator systems, Canad. J. Math. 70 (2018), no. 6, 1236 -- 1260. MR3850542 [22] Raphael Clouatre and Kenneth R. Davidson, Duality, convexity and peak interpolation in the Drury-Arveson space, Adv. Math. 295 (2016), 90 -- 149. MR3488033 [23] Raphael Clouatre and Kenneth R. Davidson, Ideals in a multiplier algebra on the ball, Trans. Amer. Math. Soc. 370 (2018), no. 3, 1509 -- 1527. MR3739183 [24] Raphael Clouatre and Michael Hartz, Multiplier algebras of complete Nevanlinna-Pick spaces: dilations, boundary representations and hyperrigidity, J. Funct. Anal. 274 (2018), no. 6, 1690 -- 1738. MR3758546 [25] Raphael Clouatre and Edward J. Timko, Cyclic row contractions and rigidity of invariant subspaces, J. Math. Anal. Appl. 479 (2019), no. 2, 1906 -- 1938. MR3987939 [26] [27] , Row contractions annihilated by interpolating vanishing ideals, to appear in Inter- national Mathematics Research Notices, arXiv:1902.06826 (2019). , The spectra of constrained commuting row contractions, preprint arXiv:1911.03525 (2019). [28] Serban Costea, Eric T. Sawyer, and Brett D. Wick, The Corona theorem for the Drury- Arveson Hardy space and other holomorphic Besov-Sobolov spaces on the unit ball in Cn, Anal. PDE 4 (2011), no. 4, 499 -- 550. [29] Ra´ul E. Curto, Applications of several complex variables to multiparameter spectral theory, Surveys of some recent results in operator theory, Vol. II, 1988, pp. 25 -- 90. MR976843 [30] Kenneth R. Davidson and Matthew Kennedy, The Choquet boundary of an operator system, Duke Math. J. 164 (2015), no. 15, 2989 -- 3004. MR3430455 [31] , Choquet order and hyperrigidity for function systems, preprint arXiv:1608.02334 (2016). [32] Ronald G. Douglas, Banach algebra techniques in operator theory, Second, Graduate Texts in Mathematics, vol. 179, Springer-Verlag, New York, 1998. MR1634900 [33] Ronald G. Douglas, Xiang Tang, and Guoliang Yu, An analytic Grothendieck Riemann Roch theorem, Adv. Math. 294 (2016), 307 -- 331. MR3479565 [34] Michael A. Dritschel and Scott A. McCullough, Boundary representations for families of representations of operator algebras and spaces, J. Operator Theory 53 (2005), no. 1, 159 -- 167. MR2132691 (2006a:47095) [35] Omar El-Fallah, Karim Kellay, Javad Mashreghi, and Thomas Ransford, A primer on the Dirichlet space, Cambridge Tracts in Mathematics, vol. 203, Cambridge University Press, Cambridge, 2014. MR3185375 [36] Miroslav Englis and Jorg Eschmeier, Geometric Arveson-Douglas conjecture, Adv. Math. 274 (2015), 606 -- 630. MR3318162 [37] Quanlei Fang and Jingbo Xia, Multipliers and essential norm on the Drury-Arveson space, Proc. Amer. Math. Soc. 139 (2011), no. 7, 2497 -- 2504. MR2784815 [38] Theodore W. Gamelin, Uniform algebras, Prentice-Hall, Inc., Englewood Cliffs, N. J., 1969. MR0410387 [39] Robert C. Gunning and Hugo Rossi, Analytic functions of several complex variables, Prentice- Hall, Inc., Englewood Cliffs, N.J., 1965. MR0180696 [40] Kunyu Guo, Junyun Hu, and Xianmin Xu, Toeplitz algebras, subnormal tuples and rigidity on reproducing C[z1, . . . , zd]-modules, J. Funct. Anal. 210 (2004), no. 1, 214 -- 247. MR2052120 (2005a:47007) [41] Masamichi Hamana, Injective envelopes of operator systems, Publ. Res. Inst. Math. Sci. 15 (1979), no. 3, 773 -- 785. MR566081 (81h:46071) [42] Michael Hartz, On the isomorphism problem for multiplier algebras of Nevanlinna-Pick spaces, Canad. J. Math. 69 (2017), no. 1, 54 -- 106. MR3589854 [43] Matthew Kennedy and Orr Moshe Shalit, Essential normality, essential norms and hyper- rigidity, J. Funct. Anal. 268 (2015), no. 10, 2990 -- 3016. MR3331791 [44] , Corrigendum to "Essential normality, essential norms and hyperrigidity" [J. Funct. Anal. 268 (2015) 2990 -- 3016] [ MR3331791], J. Funct. Anal. 270 (2016), no. 7, 2812 -- 2815. MR3464058 40 RAPHA EL CLOU ATRE AND EDWARD J. TIMKO [45] Se-Jin Kim, Hyperrigidity of C ∗-correspondences, preprint arXiv:1905.10473 (2019). [46] Norm Levenberg, Approximation in CN , Surv. Approx. Theory 2 (2006), 92 -- 140. MR2276419 [47] Berrien Moore III and Eric Nordgren, On transitive algebras containing C0 operators, Indiana Univ. Math. J. 24 (1974/75), 777 -- 784. MR0361846 (50 #14291) [48] V. Muller and F.-H. Vasilescu, Standard models for some commuting multioperators, Proc. Amer. Math. Soc. 117 (1993), no. 4, 979 -- 989. MR1112498 (93e:47016) [49] Vladimir Muller, Spectral theory of linear operators and spectral systems in Banach algebras, Second, Operator Theory: Advances and Applications, vol. 139, Birkhauser Verlag, Basel, 2007. MR2355630 (2008g:47013) [50] Vern Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in Ad- vanced Mathematics, vol. 78, Cambridge University Press, Cambridge, 2002. MR1976867 (2004c:46118) [51] Vern I. Paulsen and Mrinal Raghupathi, An introduction to the theory of reproducing kernel Hilbert spaces, Cambridge Studies in Advanced Mathematics, vol. 152, Cambridge University Press, Cambridge, 2016. MR3526117 [52] Gelu Popescu, Operator theory on noncommutative varieties, Indiana Univ. Math. J. 55 (2006), no. 2, 389 -- 442. MR2225440 (2007m:47008) [53] Mihai Putinar, Spectral inclusion for subnormal n-tuples, Proc. Amer. Math. Soc. 90 (1984), no. 3, 405 -- 406. MR728357 [54] Stefan Richter and Carl Sundberg, Joint extensions in families of contractive commuting operator tuples, J. Funct. Anal. 258 (2010), no. 10, 3319 -- 3346. MR2601618 (2011a:47019) [55] Frigyes Riesz and B´ela Sz.-Nagy, Functional analysis, Dover Books on Advanced Mathemat- ics, Dover Publications, Inc., New York, 1990. Translated from the second French edition by Leo F. Boron, Reprint of the 1955 original. MR1068530 [56] Walter Rudin, Function theory in the unit ball of Cn, Classics in Mathematics, Springer- Verlag, Berlin, 2008. Reprint of the 1980 edition. MR2446682 (2009g:32001) [57] Guy Salomon, Hyperrigid subsets of Cuntz-Krieger algebras and the property of rigidity at zero, J. Operator Theory 81 (2019), no. 1, 61 -- 79. MR3920688 [58] Donald Sarason, Generalized interpolation in H∞, Trans. Amer. Math. Soc. 127 (1967), 179 -- 203. MR0208383 (34 #8193) [59] B´ela Sz.-Nagy and Ciprian Foia¸s, Dilatation des commutants d'op´erateurs, C. R. Acad. Sci. Paris S´er. A-B 266 (1968), A493 -- A495. MR0236755 (38 #5049) [60] B´ela Sz.-Nagy, Ciprian Foias, Hari Bercovici, and L´aszl´o K´erchy, Harmonic analysis of op- erators on Hilbert space, enlarged, Universitext, Springer, New York, 2010. MR2760647 (2012b:47001) Department of Mathematics, University of Manitoba, Winnipeg, Manitoba, Canada R3T 2N2 E-mail address: [email protected] E-mail address: [email protected]
1612.03549	2	1612	2017-05-15T07:46:13	A relative tensor product of subfactors over a modular tensor category	[ "math.OA", "math-ph", "math-ph" ]	We define and study a certain relative tensor product of subfactors over a modular tensor category. This gives a relative tensor product of two completely rational heterotic full local conformal nets with trivial superselection structures over a common chiral representation category. In particular, we have a new realization of fusion rules of modular invariants. This also gives a mathematical definition of a composition of two gapped domain walls between topological phases.	math.OA	math	A relative tensor product of subfactors over a modular tensor category∗ Yasuyuki Kawahigashi† Graduate School of Mathematical Sciences The University of Tokyo, Komaba, Tokyo, 153-8914, Japan and Kavli IPMU (WPI), the University of Tokyo 5-1-5 Kashiwanoha, Kashiwa, 277-8583, Japan e-mail: [email protected] September 17, 2018 Abstract We define and study a certain relative tensor product of subfactors over a modular tensor category. This gives a relative tensor product of two completely rational het- erotic full local conformal nets with trivial superselection structures over a common chiral representation category. In particular, we have a new realization of fusion rules of modular invariants. This also gives a mathematical definition of a composition of two gapped domain walls between topological phases. 1 Introduction The theory of subfactors due to Jones [21] has been a very powerful tool in conformal field theory. We study some aspects of full conformal field theory from a viewpoint of subfactors and modular tensor categories. (We consider only unitary modular tensor categories in this paper.) We are interested in a subfactor N ⊂ M with finite Jones index [M : N ]. In conformal field theory, it is often useful to formulate a subfactor N ⊂ M in terms of a Q-system Θ = (θ, w, x) where θ is an endomorphism of a type III factor N with separable predual and w ∈ Hom(id, θ), x ∈ Hom(θ, θ2) as in [31]. When θ is an object of an abstract modular tensor category C, we say Θ is a Q-system on C. (Note that any modular tensor category is realized as a subcategory of End(N ) for a type III factor N .) It is also often called a C ∗-Frobenius algebra on C. When we have x = ε(θ, θ)x, where ε denotes the braiding, we say that the Q-system Θ is local. It is also often said that it is commutative. We say Θ is Lagrangian if we have (dim θ)2 = dim C. (See [11, page 153] for the origin of this ∗Keywords: conformal field theory, modular tensor category, modular invariant, subfactor; MSC: 81T40, 46L37, 18D10 †Supported in part by Research Grants and the Grants-in-Aid for Scientific Research, JSPS. 1 terminology.) See [22] and references therein for more on subfactors and tensor categories. Our basic reference on modular categories is [2]. See [14] for basics of subfactor theory. Let {A(I)} be a completely rational local conformal net in the sense of [26], [24], and let C be the Doplicher-Haag-Roberts representation category of {A(I)}. (It is a modular tensor category by [26].) A maximal full conformal field theory in the sense of [25] is given by a local Lagrangian Q-system on C ⊠ Copp as in [25], where "opp" means the opposite modular tensor category for which the braiding is reversed. (Also see [4, Proposition 6.7].) Let θ = Lλ∈Irr(C),µ∈Irr(Copp) Zλµλ ⊠ ¯µ be the object of such a Q-system on C ⊠ Copp, where "Irr" means the set of equivalence classes of simple objects in the modular tensor category. The matrix Z = (Zλµ) is then a modular invariant in the sense that it commutes with the S- and T -matrices arising from C as in [4, Proposition 6.6]. Suppose we have two µν ). Then the matrix product Z 1Z 2 clearly satisfies such modular invariants (Z 1 the properties of the modular invariant except for the normalization condition Z00 = 1 where 0 denotes the identity object of the modular tensor category C. It is sometimes possible to have a decomposition Z 1Z 2 = Pi Z 3,i into modular invariants Z 3,i. Such decomposition rules of matrix products have been studied under the name of fusion rules of modular invariants in [13], [15, Section 3.1], [16, Remark 5.4 (iii)]. We have a machinery of α-induction for subfactors as in [32], [6], [7], [8],[9], and it produces a modular invariant It gives a Q-system on C ⊠ Copp as in [35], and this is a general form of a as in [7]. maximal full conformal field theory on C ⊠ Copp as in [4, Proposition 6.7]. The results in [15, Section 3.1], [16, Remark 5.4 (iii)] say that a braided product of Q-systems on C gives a fusion rule of the corresponding Q-systems on C ⊠ Copp. In this way, we indirectly have an irreducible decomposition of a certain relative tensor product of two local irreducible Lagrangian Q-systems on C ⊠ Copp. λµ) and (Z 2 One typical example of such fusion rules is given as follows. Let C be the modular tensor category corresponding to the WZW-model SU (2)17. Then by [34, Page 202] (and also by [27, Theorem 2.1] and [4, Proposition 6.7]), we have exactly three irreducible local Lagrangian Q-systems on C ⊠Copp and they are labeled with A17, D10, E7 as in [10]. (These labels are for the modular invariant matrices. The label A17 corresponds to the identity matrix.) Their nontrivial fusion rules are as follows by [13, Section 5.1], [16, Remark 5.4 (iii)]. D10 ⊗ D10 = 2D10, D10 ⊗ E7 = E7 ⊗ D10 = 2E7, E7 ⊗ E7 = D10 ⊕ E7. We would like to extend this relative product to the irreducible local Lagrangian Q- in this paper where C1, C2, C3 can be different. This systems on C1 ⊠ Copp setting corresponds to a heterotic full conformal field theory. 2 and C2 ⊠ Copp 3 The author thanks M. Bischoff, L. Kong, R. Longo, K.-H. Rehren and Z. Wang for useful discussions and comments. Parts of this work were done at Instituto Superior T´ecnico, Universidade de Lisboa and Microsoft Research Station Q at Santa Barbara. The author thanks the both institutions for their hospitality. 2 2 A relative tensor product of Q-systems We consider a Q-system Θ = (θ, w, x) where θ is an endomorphism of a type III factor N with separable predual and w ∈ Hom(id, θ), x ∈ Hom(θ, θ2). We adapt [5, Definition 3.8], which means that such a Q-system corresponds to an inclusion N ⊂ M where M may not be a factor. We have N ′ ∩ M = C if and only if the Q-system Θ is irreducible. We recall the following proposition in [33]. (Also see [12, Proposition 3.7, Corollary 3.8].) Proposition 2.1 Let Θ = (θ, w, x) be an irreducible local Q-system where θ is of the form Lλ∈Irr(C1),µ∈Irr(Copp ) Zλµλ ⊠ ¯µ for some modular tensor categories C1, C2. Then it is Lagrangian if and only if we have the modular invariance property SC1 Z = ZSC2 and TC1 Z = ZTC2 for the matrix Z = (Zλµ), where SC1, SC2 , TC1 , TC2 are the S-matrix for C1, S-matrix for C2, T -matrix for C1 and T -matrix for C2, respectively. 2 This was first raised as a problem in [36, Section 3] in the context of full conformal field theory, and proved by Muger [33] and an unpublished manuscript of Longo and the author. This is valid in a general context of a modular tensor category. (Also see [4, Proposition 5.2].) Let (θ, w, x) be a Q-system where θ is of the form Lλ∈Irr(C1),µ∈Irr(C1),ν∈Irr(C2) Zλµν λ⊠µ⊠ ν for some modular tensor categories C1, C2. By applying the functor T to the C1 component as in [28, Section 4.1], [4, Section 4.2], we obtain a new Q-system T (Θ) = (T (θ), wT , xT ) where T (θ) = Lλ∈Irr(C1),µ∈Irr(C1),ν∈Irr(C2) Zλµν λµ ⊠ ν. (We need the braiding structure of C1 in order to define xT .) Note that even if Θ is irreducible, T (Θ) is not irreducible in general. Let (θ, w, x) be another Q-system where θ is of the form Lλ∈Irr(C1),µ∈Irr(C2) Zλµλ ⊠ µ for some modular tensor categories C1, C2. By applying [20, Corollary 3.10], we have a new Q-system (θ1, w1, s1) with θ = Lλ∈Irr(C1) Zλ0λ where 0 denotes the identity object of C2. We call it the restriction of Θ to C1. Now let Θ1 = (θ1, w1, x1) and Θ2 = (θ2, w2, x2) be Q-systems where on C1 ⊠ Copp 2 and θ1 = M Z 1 λµλ ⊠ ¯µ λ∈Irr(C1),µ∈Irr(Copp 2 ) θ2 = M Z 2 µν µ ⊠ ¯ν µ∈Irr(C2),ν∈Irr(Copp 3 ) on C2 ⊠ Copp 3 for some modular tensor categories C1, C2, C3. Let Θ1 ⊠ Θ2 be the tensor product of the two Q-systems for which the object is given by λ∈Irr(C1),µ∈Irr(Copp 2 M ),µ′∈Irr(C2),ν∈Irr(Copp Z 1 λµZ 2 µ′ν λ ⊠ ¯µ ⊠ µ′ ⊠ ¯ν. ) 3 By applying the T functor to the C2 components, we obtain a new Q-system whose object is λ∈Irr(C1),µ∈Irr(Copp 2 M ),µ′∈Irr(C2),ν∈Irr(Copp λµZ 2 Z 1 µ′νλ ⊠ ¯µµ′ ⊠ ¯ν. ) 3 3 By restricting this Q-system to C1 ⊠ Copp 3 , we obtain a new Q-system whose object is M Z 1 λµZ 2 µν λ ⊠ ¯ν. λ∈Irr(C1),µ∈Irr(C2),ν∈Irr(Copp 3 ) Definition 2.2 We call the above Q-system the relative tensor product of Θ1 and Θ2 over C2 and write Θ1 ⊗C2 Θ2. From the definition, it is easy to see the following. Proposition 2.3 The relative tensor product operation is associative. To apply this notion to a full conformal field theory, we need the following. Proposition 2.4 If two Q-systems are both local, then the relative tensor product Θ1 ⊗C2 Θ2 is also local. Proof. For notational simplicity, we may treat C1 ⊠ Copp category, so we simply write C1 for C1 ⊠ Copp category Vec of finite dimensional Hilbert spaces. 3 as a single modular tensor as if C3 were the trivial modular tensor 3 Locality of the tensor product Q-system Θ1 ⊠ Θ2 is represented as in Fig. 1. (We follow the graphical convention of [7, Section 3], but compose morphisms from the bottom to the top, which is a converse direction to the one in [7, Section 3].) In this picture, the triple points on the left hand side denote x1, x2, x2, respectively. The second braiding on the right hand side is reversed because we have Copp for this component. 2 λ ❖ λ′ ✗ ¯µ ❖ ¯µ′ ✗ µ1 ❖ µ′ 1 ✗ λ ■ λ′ ✒ ¯µ ■ ¯µ′ ✒ µ1 ■ µ′ 1 ✒ L ⊠ ⊠ = L ⊠ ⊠ ✻ λ′′ ✻ ¯µ′′ ✻ µ′′ 1 ✻ λ′′ ✻ ¯µ′′ ✻ µ′′ 1 Figure 1: Locality (1) From Fig. 1, we connect the wires ¯µ′′ and µ′′ 1, the wires ¯µ and µ1, and the wires ¯µ′ 1 on the both hand sides so that the wires connecting ¯µ and µ1 go over the ones 1. Then we obtain Fig. 2. Then the Reidemeister move II on the most and µ′ connecting ¯µ′ and µ′ right picture of Fig. 2 produces Fig. 3. Fig. 3 represents the locality of Θ1 ⊗C2 Θ2. (cid:3) 4 µ µ′ ■ ■ ❘ µ′′ µ µ′ ■ ■ ❘ µ′′ λ ❖ λ′ ✗ L ⊠ ✻ λ′′ µ µ′ ■ ■ ❘ µ′′ λ′ λ ■ ✒ = L ⊠ ✻ λ′′ Figure 2: Locality (2) λ ❖ λ′ ✗ L ⊠ ✻ λ′′ µ µ′ ■ ■ ❘ µ′′ λ′ λ ■ ✒ = L ⊠ ✻ λ′′ Figure 3: Locality (3) We consider the irreducible decomposition Θ1 ⊗C2 Θ2 = Li Θi 3, which is a finite sum. By [6, Corollary 3.6], this coincides with the corresponding factorial decomposition M = Li Mi where the Q-system Θ1 ⊗C2 Θ2 corresponds to an inclusion N ⊂ M and the one Θi 3 corresponds to N ⊂ Mi. We first list the following lemma. See [11, Definition 5.1] for the definition of Witt equivalence. Lemma 2.5 Let C1, C2 be Witt equivalent modular tensor categories and Θ = (θ, w, x) be an irreducible local Q-system where θ is of the form Lλ∈Irr(C1),µ∈Irr(Copp ) Zλµλ ⊠ ¯µ. Then there exists an irreducible local Lagrangian Q-system Θ = (θ, w, x) where θ is of the form Lλ∈Irr(C1),µ∈Irr(Copp ) and Z00 = Z00 = 1 where 0 denotes the identity objects of C1 and C2. Zλµλ ⊠ ¯µ with Zλµ ≥ Zλµ for all λ ∈ Irr(C1), µ ∈ Irr(Copp 2 ) 2 2 Proof. Let C be the modular tensor category arising as the ambichiral category from the Q-system Θ as in [8, Theorem 4.2]. (Note that the ambichiral objects correspond to 5 dyslectic/local modules in the terminology of [11], [12].) By [9, Corollary 4.8], C1 ⊠ Copp is Witt equivalent to C, which means that C is Witt equivalent to the trivial modular tensor category Vec. By [23, Theorem 2.4], we have an irreducible local Lagrangian Q-system on C. Composing this with the original Q-system Θ, we have an irreducible local Lagrangian Q-system Θ on C1 ⊠ Copp . It has the modular invariance property by Proposition 2.1, and Zλµ ≥ Zλµ and Z00 = Z00 = 1 are clear. (cid:3) 2 2 Theorem 2.6 If the Q-systems Θ1 and Θ2 are both Lagrangian, so is each Θi 3. λµZ 2 λν = Pµ Z 1 Proof. Set Z 3 λν λ ⊠ ¯ν be the object for Θi 3. By Proposition 2.1, being Lagrangian for Θi 3 is equivalent to modular invariance property SC1 Z 3,i = Z 3,iSC3 and TC1 Z 3,i = Z 3,iTC3 for Z 3,i, where SC1 , SC3 , TC1 , TC3 are the S-matrix for C1, S-matrix for C3, T -matrix for C1 and T -matrix for C3, respectively. µν and let Lλ∈Irr(C1),ν∈Irr(Copp 3 ) Z 3,i 3 ) Z 3,i Note that C1 and C2 are Witt equivalent, and so are C2 and C3. Hence C1 and C3 3 whose object is Z 3,i = λν , where each are also Witt equivalent and each Θi Lλ∈Irr(C1),ν∈Irr(Copp Z 3,iTC3 by Proposition 2.1. By Lemma 2.5, we may write Z 3,i Z 3,i λν is a non-negative integer. Z 3,i = Z 3,iSC3 and TC1 λν = Z 3,i 3 has a Lagrangian extension Θi λν λ ⊠ ¯ν by Lemma 2.5 and we have SC1 λν + Z 3,i Since the matrix Pi Z 3,i also has the modular invariance property, the matrix Z 3 = λν SC3,ν0 = Z 3 Z 3,i also has the modular invariance property. This implies Pλν SC1,0λ Z 3 00, Pi but Z 3 λν = 0 for all λ ∈ Irr(C1) and ν ∈ Irr(C3). This proves the modular invariance property SC1Z 3,i = Z 3,iSC3 and TC1 Z 3,i = Z 3,iTC3 for Z 3,i, as desired. (cid:3) 00 = 0 and SC1,0λ > 0, SC3,ν0 > 0. We thus have Z 3 00 = Pi Z 3,i Note that the use of modular invariance in the last paragraph of the above proof is the same as in [17, p. 726 (5.2)]. This relative tensor product of Q-systems looks similar to that of bimodules, but the example of the A17-D10-E7 modular invariants mentioned in the Introduction shows that their fusion rules do not give a fusion category since the rigidity axiom is not satisfied. We have interpreted an irreducible local Lagrangian Q-system on C1 ⊠ C2 as a gapped domain wall between topological phases represented with C1 and C2 in [23, Definition 3.1]. (See [18], [19], [30] for physical treatments of gapped domain walls.) From this viewpoint, the above relative tensor product gives a mathematical definition of the composition of gapped domain walls mentioned in [30, Fig. 1 (d)]. (Note that irreducibility of a Q-system is called stability of a gapped domain wall in [30].) A mathematical definition of such a composition has been studied in [29], [1]. It would be interesting to compare the above definition with theirs. Another construction of fusion product with some formal similarity has been defined in [3]. It would be interesting to find direct relations to their construction. 6 References [1] Y. Ai, L. Kong and H. Zheng, Topological orders and factorization homology, arXiv:1607.08422. [2] B. Bakalov and A. Kirillov, Jr., "Lectures on tensor categories and modular func- tors", American Mathematical Society, Providence (2001). [3] A. Bartels, C. L. Douglas and A. Henriques, Conformal nets III: Fusion of defects, to appear in Mem. Amer. Math. Soc. [4] M. Bischoff, Y. Kawahigashi and R. Longo, Characterization of 2D rational local con- formal nets and its boundary conditions: the maximal case, Doc. Math. 20 (2015), 1137 -- 1184. [5] M. Bischoff, Y. Kawahigashi, R. Longo and K.-H. Rehren, "Tensor categories and endomorphisms of von Neumann algebras -- with applications to quantum field theory", Springer Briefs in Mathematical Physics, 3, Springer, (2015). [6] J. Bockenhauer and D. E. Evans, Modular invariants, graphs and α-induction for nets of subfactors I. Comm. Math. Phys. 197 (1998), 361 -- 386. [7] J. Bockenhauer, D. E. Evans and Y. Kawahigashi, On α-induction, chiral projectors and modular invariants for subfactors, Comm. Math. Phys. 208 (1999), 429 -- 487. [8] J. Bockenhauer, D. E. Evans and Y. Kawahigashi, Chiral structure of modular in- variants for subfactors, Comm. Math. Phys. 210 (2000), 733 -- 784. [9] J. Bockenhauer, D. E. Evans and Y. Kawahigashi, Longo-Rehren subfactors arising from α-induction, Publ. Res. Inst. Math. Sci. 37 (2001), 1 -- 35. [10] A. Cappelli, C. Itzykson and J.-B. Zuber, The A-D-E classification of minimal and A(1) 1 conformal invariant theories, Comm. Math. Phys. 113 (1987), 1 -- 26. [11] A. Davydov, M. Muger, D. Nikshych and V. Ostrik, The Witt group of non- degenerate braided fusion categories, J. Reine Angew. Math. 677 (2013), 135 -- 177. [12] A. Davydov, D. Nikshych and V. Ostrik, On the structure of the Witt group of braided fusion categories, Selecta Math. 19 (2013), 237 -- 269. [13] E. D. Evans, Fusion rules of modular invariants, Rev. Math. Phys. 14 (2002), 709 -- 731. [14] D. E. Evans and Y. Kawahigashi, "Quantum Symmetries on Operator Algebras", Oxford University Press, Oxford (1998). [15] D. E. Evans and P. Pinto, Subfactor realisation of modular invariants, Comm. Math. Phys. 237 (2003), 309 -- 363. [16] J. Fuchs, I. Runkel and C. Schweigert, TFT construction of RCFT correlators. I. Partition functions, Nuclear Phys. B 646 (2002), 353 -- 497. 7 [17] T. Gannon, WZW commutants, lattices and level-one partition functions, Nuclear Phys. B 396 (1993), 708 -- 736. [18] L.-Y. Hung and Y. Wan, Ground state degeneracy of topological phases on open surfaces, Phys. Rev. Lett. 114 (2015), 076401. [19] L.-Y. Hung and Y. Wan, Generalized ADE classification of gapped domain walls, J. High Energy Phys. 2015 (2015), 120. [20] M. Izumi, R. Longo and S. Popa, A Galois correspondence for compact groups of automorphisms of von Neumann algebras with a generalization to Kac algebras, J. Funct. Anal. 155 (1998), 25 -- 63. [21] V. F. R. Jones, Index for subfactors, Invent. Math. 72 (1983), 1 -- 25. [22] Y. Kawahigashi, Conformal field theory, tensor categories and operator algebras, J. Phys. A 48 (2015), 303001, 57 pp. [23] Y. Kawahigashi, A remark on gapped domain walls between topological phases, Lett. Math. Phys. 105 (2015), 893 -- 899. [24] Y. Kawahigashi and R. Longo, Classification of local conformal nets. Case c < 1, Ann. of Math. 160 (2004), 493 -- 522. [25] Y. Kawahigashi and R. Longo, Classification of two-dimensional local conformal nets with c < 1 and 2-cohomology vanishing for tensor categories, Comm. Math. Phys. 244 (2004), 63 -- 97. [26] Y. Kawahigashi, R. Longo and M. Muger, Multi-interval subfactors and modularity of representations in conformal field theory, Comm. Math. Phys. 219 (2001), 631 -- 669. [27] Y. Kawahigashi, R. Longo, U. Pennig and K.-H. Rehren, Classification of non-local chiral CFT with c < 1, Comm. Math. Phys. 271 (2007), 375 -- 385 [28] L. Kong and I. Runkel, Morita classes of algebras in modular tensor categories, Adv. Math. 219 (2008), 1548 -- 1576. [29] L. Kong and H. Zheng, The center functor is fully faithful, arXiv: 1507.00503. [30] T. Lan, J. Wang and X.-G. Wen, Gapped domain walls, gapped boundaries and topological degeneracy, Phys. Rev. Lett. 114 (2015), 076402. [31] R. Longo, A duality for Hopf algebras and for subfactors, Comm. Math. Phys. 159 (1994), 133 -- 150. [32] R. Longo and K.-H. Rehren, Nets of Subfactors, Rev. Math. Phys. 7 (1995), 567 -- 597. [33] M. Muger, On superselection theory of quantum fields in low dimensions, in XVIth International Congress on Mathematical Physics, (2010), 496 -- 503, World Sci. Publ. 8 [34] V. Ostrik, Module categories, weak Hopf algebras and modular invariants, Trans- form. Groups8 (2003), 177 -- 206. [35] K.-H. Rehren, Canonical tensor product subfactors, Commun. Math. Phys. 211 (2000) 395 -- 406. [36] K.-H. Rehren, Locality and modular invariance in 2D conformal QFT, in Mathe- matical physics in mathematics and physics (Siena, 2000), (2001), 341 -- 354. 9
1210.7581	3	1210	2013-11-11T05:46:14	Continuous minimax theorems	[ "math.OA", "math.PR" ]	In classical matrix theory, there exist useful extremal characterizations of eigenvalues and their sums for Hermitian matrices (due to Ky Fan, Courant-Fischer-Weyl and Wielandt) and some consequences such as the majorization assertion in Lidskii's theorem. In this paper, we extend these results to the context of self adjoint elements of finite von Neumann algebras, and their distribution and quantile functions. This work was motivated by a lemma in a paper by Voiculescu and Bercovici, that described such an extremal characterization of the distribution of a self-adjoint operator affiliated to a finite von Neumann algebra - suggesting a possible analogue of the classical Courant-Fischer-Weyl minmax theorem, for a self adjoint operator in a finite von Neumann algebra. It is to be noted that the only von Neumann algebras considered here have separable pre-duals.	math.OA	math	Continuous minmax theorems Madhushree Basu∗ V.S. Sunder† Abstract In classical matrix theory, there exist useful extremal characterizations of eigenval- ues and their sums for Hermitian matrices (due to Ky Fan, Courant-Fischer-Weyl and Wielandt) and some consequences such as the majorization assertion in Lidskii’s the- orem. In this paper, we extend these results to the context of self adjoint elements of finite von Neumann algebras, and their distribution and quantile functions. This work was motivated by a lemma in [BV93] that described such an extremal characterization of the distribution of a self-adjoint operator affiliated to a finite von Neumann algebra - suggesting a possible analogue of the classical Courant-Fischer-Weyl minmax theorem, for a self adjoint operator in a finite von Neumann algebra. It is to be noted that the only von Neumann algebras considered here have separable pre-duals. 1 Introduction This paper is arranged as follows: in Section 2, we prove an extension of the classical minmax theorem of Ky Fan’s ( [Fan49]) in a von-Neumann algebraic setting for self adjoint operators having no atoms in their distributions, and then, as an application of the above, in Section 3, we give a few applications of the previous section. First we state and prove an exact analogue of the Courant-Fischer-Weyl minmax theorem ( [CH89]) for operators inside II1 factors hav- ing no eigenvalues. It is interesting to note that classically Courant-Fischer-Weyl minmax theorem came before Ky Fan’s theorem for Hermitian matrices whereas the order of events is reversed in our proofs. Then, as an application of our version of Courant-Fischer-Weyl min- max theorem, we prove that that if a self adjoint operator with no eigenvalues, is dominated by another such operator (both being inside a II1 factor), then their respective quantile functions are dominated one by the other in the same order. Finally we discuss a continuous analogue of Lidskii’s theorem - a majorization-type inequality between eigenvalues of sum of Hermitian matrices and the sum of eigenvalues of the summand matrices, discussions and proofs of the finite dimensional version of which can be found in [Lid50], [Lid82], [Wie55]. In Section 4, we state and prove, for operators with no eigenvalues in II1 factors, a continuous ∗Indian Statistical Institute, Bangalore; email: [email protected] †The Institute of Mathematical Sciences, Chennai; e-mail: [email protected] 1 analogue of Wielandt’s minmax theorem ( [Wie55]), the classical version of which gives an extremal characterization of arbitrary sums of eigenvalues of Hermitian matrices. Similar continuous analogues of minmax-type results have been worked out earlier, for example in [FK86] and [Hia87], in the general context of defining trace-measurablity of operators affiliated to von Neumann algebras and for generalizing the concept of majorization in von Neumann algebras. However in these papers, the emphasis have been on positive operators and the rigorous proofs for the von Neumann algebraic adaptation of minmax- type results, corresponded to singular values of Hermitian matrices. On the other hand, our proofs are simple, independent of the approach of these papers, deal explicitly with self adjoint (as against positive) operators in certain von Neumann algebras and correspond to eigenvalues (as against singular values) of Hermitian matrices in the finite dimensional case. Moreover as far as we know, unlike former works on this topic, our formulations, for the particular case of finite dimensional matrix algebras, give the exact statements of the classical Ky Fan’s, Courant-Fischer-Weyl’s and Wielandt’s theorems. However in the continuous case, our results are restricted by non-atomicity property of the distribution of the self adjoint operator under consideration as well as by certain properties of the underlying von Neumann algebra. In order to describe our results, which are continuous analogues of certain inequalities that appear as part of the set of inequalities mentioned in Horn’s conjecture ( [Hor62]), it will be convenient to re-prove the well-known fact that any monotonic function with appropriate one-sided continuity is the distribution function of a random variable X - which can in fact be assumed to be defined on the familiar Lebesgue space [0, 1) equipped with the Borel σ-algebra and Lebesgue measure. (We adopt the convention of [BV93] that the distribution function Fµ of a compactly supported probability measure1 µ defined on the σ-algebra BR of Borel sets in R, is left-continuous; thus Fµ(x) = µ((−∞, x)). Proposition 1.1. If F : R → [0, 1] is monotonically non-decreasing and left continuous and if there exists α, β ∈ R with α < β such that F (t) = 0, for t ≤ α and F (t) = 1 for t ≥ β, (1.1) then there exists a monotonically non-decreasing right-continuous function X : [0, 1) → R such that F is the distribution function of X, i.e., F (t) = m({s : X(s) < t}), where m denotes the Lebesgue measure on [0, 1). Moreover range(X) ⊂ [α, β]. Proof. Define X : [0, 1) → R by X(s) = inf{t : F (t) > s} = inf{t : t ∈ Es} , (1.2) 1Actually Bercovici and Voiculescu considered possibly unbounded self-adjoint operators affiliated to M , so as to also be able to handle probability measures which are not necessarily compactly supported, but we shall be content with the case of bounded a ∈ M , having a compactly supported probability measure as its distribution. 2 where Es = {t ∈ R : F (t) > s} ∀s ∈ [0, 1). (The hypothesis 1.1 is needed to ensure that Es is a non-empty bounded set for every s ∈ [0, 1) so that, indeed X(s) ∈ R.) First deduce from the monotonicity of F that s1 ≤ s2 ⇒ Es2 ⊂ Es1 ⇒ X(s1) ≤ X(s2) and hence X is indeed monotonically non-decreasing. The definition of X and the fact that F is monotonically non-increasing and left contin- uous are easily seen to imply that Es = (X(s), ∞), and hence, it is seen that X(s) < t ⇔ ∃t0 < t such that F (t0) > s ⇔ F (t) > s (since F is left-continuous) (1.3) Hence, if t ∈ R m({s ∈ [0, 1) : X(s) < t}) = m([0, F (t)) = F (t), proving the required statement. (1.4) Moreover, if for any s ∈ [0, 1), X(s) < α, then by definition of X, ∃ t′ < α such that F (t′) > s ≥ 0, a contradiction to the first hypothesis in 1.1. On the other hand, if for any s ∈ [0, 1), X(s) > β, then by 1.3, s ≥ F (β) = 1 (by the second hypothesis in 1.1), a contradiction. Hence indeed range(X) ⊂ [α, β]. ✷ This function X is known as quantile function 2 of the distribution F . If F = Fµ for a probability measure µ on R, then X is denoted as Xµ. X can also be thought of as an element of L∞(R, µ), where µ is a compactly supported probability measure on R such that µ = m ◦ X −1 and supp µ ⊂ [α, β]. We will elaborate on this later in Proposition 2.1. Given a self-adjoint element a in a von Neumann algebra M and a (usually faithful normal) tracial state τ on M, define µa(E) := τ (1E(a)) (1.5) (for the associated scalar spectral measure) to be the distribution of a. Since τ is positivity preserving, µa indeed turns out to be a probability measure on R. For simplicity we write Fa, Xa instead of Fµa, Xµa (to be pedantic, one should also indicate the dependence on (M, τ ), but the trace τ and the M containing a will usually be clear.) Note that only the abelian von Neumann subalgebra A generated by a and τ A are relevant for the definition of Fa and Xa. For M, a, τ as above, it was shown in [BV93] that 1 − Fµa(t) = max{τ (p) : p ∈ P(M), pap ≥ ta}. (1.6) 2This function acts as the inverse of the distribution function at every point that is not an atom of the probability measure µ. 3 Example 1.2. Let M = Mn(C) with τ as the tracial state on this M. If a = a∗ ∈ M has j distinct eigenvalues λ1 < λ2 < · · · < λn, then Fa(t) = 1 n 1(λj ,λj+1]. We see that the distinct numbers less than 1 in the range of Fa are attained at the n distinct eigenvalues of a, and further that equation 1.6 for t = λj says that n − j + 1 is the largest possible dimension of a subspace W of Cn such that haξ, ξi ≥ λj for every unit vector ξ ∈ W . In other words equation 1.6 suggests a possible extension of the classical Courant- Fischer minmax theorem for a self adjoint operator in a von Neumann algebra, involving its distribution. n {j : λj < t} = Pn j=1 It is also true and not hard to see that the right side of equation 1.6 is indeed a maximum (and not just a supremum), and is in fact attained at a spectral projection of a; i.e., the two sides of equation 1.6 are also equal to max{τ (p) : p ∈ P(A), pap ≥ ta}, where A is the abelian von Neumann subalgebra generated by a. 2 Our version of Ky Fan’s theorem In this section we wish to proceed towards obtaining non-commutative counterparts of the classical Ky Fan’s minmax theorem formulated for appropriate self-adjoint elements of ap- propriate finite von Neumann algebras. Proposition 2.1. Let (Ω, B, P ) be a probability measure, and suppose Y : Ω → R is an essentially bounded random variable. Let σ(Y ) = {Y −1(E) : E ∈ BR} and let µ = P ◦ Y −1 be the distribution of Y . Then, for any s0 ∈ Fµ(R), we have Y dP : Ω0 ∈ σ(Y ), P (Ω0) ≥ s0} f0dµ : E ∈ BR, µ(E) ≥ s0} Xµdm : G ∈ σ(Xµ), m(G) ≥ s0} inf{ZΩ0 = inf{ZE = inf{ZG = Z s0 0 Xµdm, (2.1) where f0 = idR and m denotes Lebesgue measure on [0,1). Proof. The version of the change of variable theorem we need says that if (Ωi, Bi, Pi), i = 1, 2 are probability spaces and T : Ω1 → Ω2 is a measurable function such that P2 = P1 ◦ T −1, then ZΩ2 gdP2 = ZΩ1 g ◦ T dP1 , (2.2) for every bounded measurable function g : Ω2 → R. 4 For every Ω0 ∈ σ(Y ), which is of the form Y −1(E) for some E ∈ BR, set G = X −1 µ (E). Notice, from equations 1.3 and 1.4 that m ◦ X −1 µ (−∞, t) = µ({s ∈ [0, 1) : Xµ(s) < t}) = µ({s ∈ [0, 1) : s < Fµ(t)}) = Fµ(t) = µ(−∞, t) ; i.e. m◦X −1 µ = µ = P ◦Y −1. Now, set g = 1E ·f0. Since g ◦Y = 1E ◦Y ·Y = 1Y −1(E)Y = 1Ω0Y , and (similarly) g ◦ Xµ = 1GXµ, we see that the first two equalities in 2.1 are immediate consequences of two applications of the version stated in equation 2.2 above, of the ‘change of variable’ theorem. As for the last, if G ∈ B[0,1) with m(G) ≥ s0, then write I = G∩[0, s0), J = [0, s0)\I, K = G \ I and note that G = I ` K, [0, s0) = I ` J (where ` denotes disjoint union, and K = G \ [0, 1) ⊂ [s0, 1). So we may deduce that ZG Xµdm − Z s0 0 Xµdm = ZK Xµdm − ZJ Xµdm ≥ Xµ(s0)m(K) − Xµ(s0)m(J) ≥ 0 , since s1 ∈ J, s2 ∈ K ⇒ s1 ≤ s0 ≤ s2 ⇒ Xµ(s1) ≤ Xµ(s0) ≤ Xµ(s2) (by the monotonicity of Xµ), and m(K) ≥ m(J). Thus, we see that inf{ZG Xµdm : G ∈ σ(Xµ), m(G) ≥ s0} ≥ Z s0 0 Xµdm , while conversely, inf{ZG Xµdm : G ∈ BR, m(G) ≥ s0} ≤ Z[0,s0) Xµdm = Z s0 0 Xµdm , thereby establishing the last equality in 2.1. ✷ Theorem 2.2. Let a be a self-adjoint element of a von Neumann algebra M equipped with a faithful normal tracial state τ . Let A be the von Neumann subalgebra generated by a in M and P(M) be the set of projections in M. Then, for all s ∈ Fa(R), inf{τ (ap) : p ∈ P(M), τ (p) ≥ s} = inf{τ (ap) : p ∈ P(A), τ (p) ≥ s} = Z s 0 Xadm (2.3) (hence the infima are attained and are actually minima), if either: 5 1. (‘continuous case’) µa has no atoms, or 2. (‘finite case’) M = Mn(C) for some n ∈ N and a has spectrum {λ1 < λ2 < · · · < λn}. Proof. We begin by noting that in both the cases, the last equality in 2.3 is an immediate consequence of Proposition 2.1. Moreover the set {τ (ap) : p ∈ P(A), τ (p) ≥ s} being contained in {τ (ap) : p ∈ P(M), τ (p) ≥ s}, it is clear that inf{τ (ap) : p ∈ P(A), τ (p) ≥ s} ≥ inf{τ (ap) : p ∈ P(M), τ (p) ≥ s}. So we just need to prove that inf{τ (ap) : p ∈ P(A), τ (p) ≥ s} ≤ inf{τ (ap) : p ∈ P(M), τ (p) ≥ s}. (2.4) 1. (the continuous case) Due to the assumption of µa being compactly supported and having no atoms, it is clear that Fa is continuous and that Fa(R) = [0, 1]. Under the standing assumption of separability of pre-duals of our von Neumann alge- bras, the hypothesis of this case implies the existence of a probability space (Ω, B, P ) and a map π : A → L∞(Ω, B, P ) such that R π(x)dP = τ (x) ∀x ∈ A, Y := π(a) is a random variable and π is an isomorphism onto L∞(Ω, σ(Y ), P ). We shall establish the first equality of 2.3 by showing that if p′ ∈ P(M) and τ (p′) = s, then τ (ap′) ≥ min{τ (ap) : p ∈ P(A), τ (p) ≥ s}. For this, first note that since τ is a faithful normal tracial state on M, there exists a τ -preserving conditional expectation E : M → A. Then τ (ap′) = τ (aE(p′)) = Z Y ZdP, where Z = π(E(p′)). Since E is linear and positive, it is clear that 0 ≤ Z ≤ 1 P − a.e. So it is enough to prove that inf{ZΩ = inf{ZE Y ZdP : 0 ≤ Z ≤ 1,Z ZdP ≥ s} Y dP : E ∈ B, P (E) ≥ s}. For this, it is enough, thanks to the Krein-Milman theorem (see, e.g. [KM40]), to note that K = {Z ∈ L∞(Ω, B, P ) : 0 ≤ Z ≤ 1,R ZdP ≥ s} is a convex set which is compact in the weak* topology inherited from L1(Ω, B, P ), and prove that the set ∂e(K) of its extreme points is {1E : P (E) ≥ s}. For this, suppose Z ∈ K is not a projection, Clearly then P ({Z ∈ (0, 1)}) > 0, so there exists ǫ > 0 such that P ({ǫ < Z < 1−ǫ}) > 0. Since µa, and hence P has no atoms, we may find disjoint Borel subsets E1, E2 ⊂ {Z ∈ (ǫ, 1−ǫ)} such that P (E1) = P (E2) > 0. If we now set Z1 = Z + ǫ(1E1 −1E2) and Z2 = Z + ǫ(1E2 −1E1), it is not hard to see that Z1, Z2 ∈ K, Z1 6= Z2 and Z = 1 2(Z1 + Z2) showing that Z /∈ ∂e(K) , thereby proving 2.4. 6 abelian self-adjoint subalgebra of Mn(C). Recall that in this case, Fa(t) = 1 2. (the finite case) Since a has distinct eigenvalues λ1 < λ2 < · · · < λn, A is a maximal n {j : n1(λj ,λj+1]. It then follows that Fa(R) = { j n : 0 ≤ j ≤ n} and that n ) and 2.3 is then (after multiplying by n) precisely the statement n , j j λj < t} = Pn Xa = Pn j=1 j=1 λj1[ j−1 of Ky Fan’s theorem (in the case of self-adjoint matrices with distinct eigenvalues): For 1 ≤ j ≤ n, inf{τ (ap) : p ∈ P(Mn(C)), rank(p) ≥ j} = inf{τ (ap) : p ∈ P(A), rank(p) ≥ j} = 1 n j Xi=1 j n λi = Z 0 Xa(s)ds. It suffices to prove the following: inf{τ (ap) : p ∈ P(A), rank(p) ≥ j} ≤ inf{τ (ap) : p ∈ P(Mn(C)), rank(p) ≥ j}. For this, begin by deducing from the compactness of P(Mn(C)) that there exists a p0 ∈ P(Mn(C)) with rank(p0) ≥ j such that τ (ap0) ≤ τ (ap) ∀p ∈ P(Mn(C)) with rank(p) ≥ j. We assert that any such minimizing p0 must belong to A. The assumption that A is a masa means we only need to prove that p0a = ap0. For this pick any self-adjoint x ∈ Mn(C), and consider the function f : R → R defined by f (t) = τ (eitxp0e−itxa). Since clearly eitxp0e−itx ∈ P(M) and rank(eitxp0e−itx) = rank(p0) ≥ j, for all t ∈ R, we find that f (t) ≥ f (0) ∀t. As f is clearly differentiable, we may conclude that f ′(0) = 0. Hence, 0 = τ (ixp0a − ip0xa) = i(τ (xp0a) − τ (p0xa)) = i(τ (xp0a) − τ (xap0)), so that τ (x(p0a − ap0)) = 0 for all x = x∗ ∈ M, and indeed ap0 = p0a as desired. ✷ Case 1 of Theorem 2.2 is our continuous formulation of Ky Fan’s result while Case 2 only captures the classical Ky Fan’s theorem for the case of distinct eigenvalues. However the general case of non-distinct eigenvalues can also be deduced from our proof, as we show in the following corollary: Corollary 2.3. Let a be a Hermitian matrix in Mn(C) with spectrum {λ1 ≤ · · · ≤ λn}, where not all λjs are necessarily distinct. Then for all j ∈ {1, · · · , n}, min{τ (ap) : p ∈ P(Mn(C)), rank(p) ≥ j} = 1 n j Xi=1 λi. 7 Proof. We may assume that a is diagonal. Let A1 be the set of all diagonal matrices, so that A ( A1. Pick a(m) = diag(λ(m) s are all distinct and limm→∞ λ(m) j = λj ∀1 ≤ j ≤ n. Then the already established case of Theorem 2.2 in the case of distinct eigenvalues shows that for all p ∈ P(Mn(C)) with rank(p) ≥ j, we have n ) ∈ A1 such that λ(m) , · · · , λ(m) , λ(m) j 1 2 τ (ap) = lim m→∞ τ (a(m)p) ≥ lim m→∞ 1 n j Xi=1 λ(m) i Xi=1 The above, along with the fact that τ (apj) = 1 = λi j 1 n i=1 λi, where pj is the obvious diagonal projection, completes our proof of Ky Fan’s theorem for Hermitian matrices in full generality. ✷ n Pj Remark 2.4. It is not difficult to see that equation 2.3 holds even if we replace the inequality τ (p) ≥ s with equality. Remark 2.5. Notice that the hypothesis and hence the conclusion, of the ‘continuous case’ of Theorem 2.2 are satisfied by any self-adjoint generator of a masa in a II1 factor. 3 Applications of our version of Ky Fan’s theorem In this section we discuss three applications of our version of Ky Fan’s theorem, but we first identify a necessary definition for many of our ‘continuous cases’ , as well as a lemma which we will need to use at a later stage. Definition 3.1. Given a self-adjoint element a in a finite von Neumann algebra M, we say that we are in the continuous case if for B ∈ {M, A} (with A the von Neumann subalgebra generated by a in M) and p ∈ P(B), we have {τ (r) : r ∈ P(B), r ≤ p} = [0, τ (p]. (This assumption for B = A amounts to requiring that µa has no atoms) Lemma 3.2. With M, a, A in the continuous case as above, suppose t0 < t1 ∈ R, F (t1) − F (t0) = δ > 0 and let r0 = 1[t0,∞)(a) and q0 = 1[t0,t1)(a). Then r0, q0 ∈ P(A), τ (r0) = 1 − F (t0), τ (q0) = δ and q0 ≤ r0 and τ (aq0) = min q∈P(A) τ (aq) = min q∈P(M ) τ (aq) q≤r0 τ (q)=δ q≤r0 τ (q)=δ Proof. If we consider any other q ∈ A, with q ≤ r0 and τ (q) = δ, then q is of the form 1E(a), where E ⊂ [t0, ∞) with µa(E) = δ. Arguing as in the proof of Proposition 2.1, 8 t0 t dµ(t) Z t1 ⇒Z F (t1) t dµ(t) ≤ ZE X(s) ds ≤ ZF (E) F (t0) X(s) ds ⇒τ (aq0) ≤ τ (aq). (cid:16)r0Mr0, τ (·) To prove the same for any q ≤ r0, first we note that since r0 ∈ W ∗({a}), (M0, τ0) := τ (r0)(cid:17) is also a von Neumann algebra with equipped with a faithful normal tracial state and a0 := r0ar0 is a self adjoint element with a continuous distribution µ0 (with respect to τ0) in it. Let the von Neumann subalgebra generated by a0 in M0 be A0. Then M0 and A0 satisfy the same ‘continuity hypotheses as M and A. Any q ≤ r0 with τ (q) = δ can be thought of as q ∈ P(M0) with τ0(q) = δ τ (r0), and conversely. Now as in the proof of the continuous case of Theorem 2.2 we can assume that there exists a non-atomic probability space (Ω0, B0, P0) and a map π0 : A0 → L∞(Ω0, B0, P0) such that R π0(x)dP0 = τ0(x) ∀x ∈ A0, Y0 := π0(a0) and π0 is an isomorphism onto L∞(Ω0, σ(Y0), P0). We proceed exactly as we did in the proof of the ‘continuous case’ of Theorem 2.2 to show that min{R Y0Z0 dP0 : Z0 ∈ L∞(Ω0, B0, P0), 0 ≤ Z0 ≤ 1,R Z0dP0 = δ τ (r0)} is indeed attained and the minimizing contractions can only be of the form 1E ∈ L∞(Ω0, B0, P0) for E ∈ σ(Y0), P0(E) = δ τ (r0). Thus we have τ0(a0q0) = min q∈P(M0) τ0(q)=δ0 τ0(a0q) τ (a0q) τ (r0) = min q∈P(M ) q≤r0 τ (r0) = δ τ (q) τ (r0) τ (aq), since r0 commutes with a and any q ≤ r0. ⇒ τ (a0q0) τ (r0) ⇒τ (aq0) = min q∈P(M ) q≤r0 τ (q)=δ Remark 3.3. By considering −a in place of a, for instance, we clearly have the following dual to Lemma 3.2: With M, a, A, t0, t1 as in Lemma 3.2, let p := 1(−∞,t1)(a) and q := 1[t0,t1)(a) ≤ p. Then, p, q ∈ P(A), τ (p) = F (t1), τ (q) = δ, and ✷ τ (aq) = max q∈P(A) τ (aq) = max q∈P(M ) τ (aq) q≤p τ (q)=δ q≤p τ (q)=δ 9 We now proceed to our generalization of the classical Courant Fischer-Weyl minmax theorem: Theorem 3.4. Let a be a self adjoint element of a von Neumann algebra M equipped with a faithful normal tracial state τ . Let t0 and t1 ∈ R such that t0 < t1 and Fa(t1) − Fa(t0) =: δ > 0. Then Z Fa(t1) Fa(t0) Xa(s) ds = sup r∈P(M ) τ (r)≥1−Fa(t0) inf q∈P(M ) q≤r τ (q)=δ τ (aq), (3.1) if either 1. we are in the ‘continuous case’; or 2. (‘finite case’) M is a type In factor for some n ∈ N and a has spectrum {λ1 < λ2 < · · · < λn}. Moreover there exists r0 ∈ P(A) ≤ P(M) with τ (r0) ≥ 1 − F (t0) such that Z Fa(t1) Fa(t0) Xa(s) ds = min q∈P(M ) τ (aq), q≤r0 τ (q)=δ so that the supremum is actually maximum. Proof. For simplicity we write F and X for Fa and Xa respectively. 1. (the continuous case) For proving “≤”, deduce, from Lemma 3.2 that Z F (t1) F (t0) X(s) ds ≤ sup r∈P(M ) τ (r)≥1−F (t0) inf q∈P(M ) q≤r τ (q)=δ τ (aq). (3.2) For “≥”, let us choose any projection r with τ (r) ≥ 1 − F (t0). Let r1 = 1(−∞,t1)(a). Then τ (r1) = F (t1) ⇒ τ (r1 ∧ r) ≥ F (t1) − F (t0) = δ. Hence, by the hypothesis in this continuous case, ∃ q1 ≤ r ∧ r1 with τ (q1) = δ. Now consider the II1 factor (M1, τ1) := (cid:16)r1Mr1, τ (·) tracial state on M1. Then q1 can be thought of as a projection in P(M1) with τ1(q1) = τ (r1). Note that q0 = 1[t0,t1)(a) ≤ r1. τ (r1)(cid:17), where τ1 is a faithful normal δ 10 As above a1 := r1ar1 is a self adjoint element with continuous distribution in M1. So we can consider our version of Ky Fan’s theorem in M1 (Theorem 2.2) (also see Remark 2.4): X(s)ds R F (t0) 0 τ (r1) = τ1(a(r1 − q0)) = min q∈P(M1) τ1(q)= F (t0) τ (r1) τ1(aq). (using the fact that a, q0 and q ∈ P(M1) commute with r1.) Subtracting both sides from τ1(a1) and writing q′ for r1 − q in the index, we can rewrite it as: R F (t1) F (t0) X(s) ds τ (r1) = or equivalently, max q′∈P(M1) τ1(q′)= F (t1)−F (t0) τ (r1) τ1(aq′), = δ τ (r1) Z F (t1) F (t0) X(s) ds = max q′∈P(M ) τ (aq′). q′≤r1 τ (q′)=δ Now using the fact that q1 ≤ r ∧ r1, we have: Z F (t1) F (t0) X(s) ds = max q′∈P(M ) τ (aq) ≥ τ (aq1) ≥ inf q∈P(M ) τ (aq), q′≤r1 τ (q′)=δ q≤r τ (q)=δ thus, and using the fact that our choice of r was arbitrary with τ (r) ≥ 1 − F (t0), we have: Z F (t1) F (t0) X(s) ds ≥ sup r∈P(M ) τ (r)≥1−F (t0) inf q∈P(M ) q≤r τ (q)=δ τ (aq). (3.3) Equations 3.2 and 3.3 together give us the required equality. 11 2. (the finite case) Notice that if we set t0 = λi, t1 = λi+j, δ = j n , where i, j ∈ {1, · · · , n} such that i + j − 1 ≤ n, equation 3.2 translates to: λi + λi+1 + · · · + λi+j−1 = sup r∈P(Mn(C)) T r(r)≥n−i+1 T r(aq), inf q∈P(Mn(C)) q≤r T r(q)=j where T r is the sum of the diagonal entries of matrices. For the inequality “≤” we prove, λi + λi+1 + · · · + λi+j−1 = T r(aq0) = min q∈P(Mn(C)) T r(aq), q≤r0 T r(q)=j where r0 = 1{λi,λi+1,··· ,λn}(a) and q0 = 1{λi,λi+1,··· ,λi+j−1}(a), by first showing that any minimizing projection below r0 has to commute with r0ar0, and then using the fact that with distinct eigenvalues r0ar0 generates a masa in r0Mn(C)r0, concluding that minimizing projections have to be spectral projections (see the exactly similar proof of the finite case of Theorem 2.2). For proving “≥”, we start with an arbitrary projection r with T r(r) ≥ n − i + 1 and note that if we define r1 := 1{λ1,··· ,λi+j−1}(a), then ∃ q1 ≤ r ∧ r1 such that T r(q1) = j. Now we proceed using Ky Fan’s theorem for finite dimensional Hermitian matrix r1ar1 in r1Mn(C)r1, exactly as in the above proof of the continuous case of this theorem. ✷ Remark 3.5. Theorem 3.4 can equivalently be stated as: Z F (t1) F (t0) X(s) ds = inf p∈P(M ) τ (p)≥F (t1) sup q≤p τ (q)=δ τ (aq), Moreover we can get the classical Courant-Fischer-Weyl minmax theorem for Hermitian involving non-distinct eigenvalues as well) from the above matrices in full generality (i.e. theorem in exactly similar manner as in Corollary 2.3. The classical Courant-Fischer-Weyl minmax theorem has a natural corollary that says if a, b are Hermitian matrices in Mn(C) such that a ≤ b (i.e. b−a is positive semi-definite), and if {α1 ≤ · · · ≤ αn} and {β1 ≤ · · · ≤ βn} are their spectra respectively, then αj ≤ βj for all j ∈ {1, · · · , n}. As expected, Theorem 3.4 leads us to the same corollary for the ‘continuous case’: 12 Corollary 3.6. Let M be a II1 factor equipped with faithful normal tracial state τ . a, b ∈ M such that a = a∗, b = b∗ and µa, µb have no atoms. Then If a ≤ b ⇒ Xa ≤ Xb. Proof. Notice that since a ≤ b and τ is positivity preserving, we have τ (xax∗) ≤ τ (xbx∗). (3.4) (3.5) for all x ∈ M. Fix 0 ≤ s0 < s1 < 1. By our assumptions on a and b, µa, µb are compactly supported probability measures with no atoms. Hence Fa and Fb are continuous functions with range(Fa) = range(Fb) = [0, 1]. Thus ∃ ta 1 ∈ R such that s0 = Fa(ta 0) = Fb(tb 0) and s1 = Fa(ta 1) = Fb(tb 0, ta 1, tb 0, tb 1). Now using Theorem 3.4 Z s1 s0 Xa dm = sup r∈P(M ) τ (r)≥1−Fa(ta 0 ) = sup r∈P(M ) τ (r)≥1−Fa(ta 0 ) ≤ = sup r∈P(M ) τ (r)≥1−Fb(tb 0) sup r∈P(M ) τ (r)≥1−Fb(tb 0) inf q∈P(M ) q≤r τ (r)=s1−s0 inf q∈P(M ) q≤r τ (r)=s1−s0 inf q∈P(M ) q≤r τ (r)=s1−s0 inf q∈P(M ) q≤r τ (r)=s1−s0 = Z s1 s0 Xb dm. τ (aq) τ (qaq) τ (qbq), by the inequality 3.5 τ (bq) This proves that ZI Xa dm ≤ ZI Xb dm (3.6) for any interval I = [s0, s1) ⊂ [0, 1), and in fact for any I ∈ A := {⊔k sj 1 < 1, k ∈ N}. j=1[sj 0, sj 1) : 0 ≤ sj 0 < But A is an algebra of sets which generates the σ-algebra B[0,1). Thus for any Borel E ⊂ [0, 1), there exists a sequence {In : n ∈ N} ⊂ A such that µ(In∆E) → 0. Recall from Proposition 1.1 that our quantile functions of self adjoint elements of von Neumann algebras are elements of L∞([0, 1), B[0,1), m). We may hence deduce from the 13 sentence following equation. (3.6)that if E, In are the previous paragraph, we have: ZE Xadm = lim Xadm Xadm ≤ lim n→∞ZIn n→∞ZIn = ZE Xbdm. As E ∈ B[0,1) was arbitrary, this shows that, Xa ≤ XBm − a.e.; as Xa, Xb are continuous by our hypotheses, this shows that indeed Xa ≤ Xb. ✷ Finally, we discuss a continuous analogue of Lidskii’s majorization result. By Theorem 2.2, we have the following lemma: Lemma 3.7. If M is a von Neumann algebra with a faithful normal tracial state τ on it, then for a = a∗, b = b∗ ∈ M with µa, µb non-atomic and for all s ∈ [0, 1), Moreover, Z s 0 Xa+b dm ≥ Z s 0 (Xa + Xb) dm. Z 1 0 Xa+b dm = Z 1 0 (Xa + Xb) dm. Proof. Recall from our proof of Theorem 2.2 that there exists a projection p ∈ P(M) (in fact in the von Neumann algebra generated by a + b) such that τ (p) ≥ s and Z s 0 Xa+b dm = τ ((a + b)p) = τ (ap) + τ (bp) ≥ inf{τ (ap′) : p′ ∈ P(M), τ (p′) ≥ s} + inf{τ (bp′) : p′ ∈ P(M), τ (p′) ≥ s} = Z s = Z s 0 0 Xa dm +Z s 0 Xb dm (Xa + Xb) dm. Finally, it is clear (from our change-of-variable argument in Proposition 2.1 for instance) that for any c = c∗ ∈ M, we have R 1 Z 1 0 Xa+b dm = τ (a + b) = τ (a) + τ (b) = Z 1 0 Xcdm = τ (c) and hence Xa dm +Z 1 0 Xb dm = Z 1 0 0 (Xa + Xb) dm. ✷ 14 The above is an analogue of the fact that for n × n Hermitian matrices a, b, with their eigenvalues λ1(a) ≤ · · · ≤ λn(a) and λ1(b) ≤ · · · ≤ λn(b), for all k ∈ {1, · · · , n − 1}, and k Xj=1 n Xj=1 λj(a + b) ≥ λj(a + b) = k Xj=1 n Xj=1 λj(a) + λj(a) + k Xj=1 n Xj=1 λj(b), λj(b), i.e. λ(a) + λ(b) is majorized by λ(a + b) in the sense of [HLP29]. We consider the definition of majorization in the continuous context (see for example, [Sak85]) as follows: Definition 3.8. For a = a∗, b = b∗ in a von Neumann algebra M with a faithful normal 0 Xb dm for all s ∈ [0, 1) tracial state τ on it, a is said to be majorized by b if R s and R 1 0 Xa dm = R 1 0 Xb dm. When this happens, we simply write Xa ≺ Xb. 0 Xa dm ≤ R s Then, Lemma 3.7 can be written as: which gives a version of the continuous analogue of Lidskii’s theorem. Xa+b ≺ Xa + Xb, The study of majorization and its von Neumann algebraic analogue is vast (see for exam- ple, [Kam83], [Hia87]) and closely related to the minmax-type results but we will not discuss it further in this paper. 4 Continuous version of Wielandt’s minmax principle In this section we state and prove a continuous analogue of Wielandt’s minmax theorem. The classical matrix formulation of Wielandt’s theorem is obtained by taking δ = 1 n when M = Mn(C), but we shall not repeat the kind of reasoning given in the case of the Courant- Fischer-Weyl theorem in the finite-dimensional case where our assumptions of our ‘continuous case’ are not valid. We shall be content with formulating and proving the continuous case. We make the standing ‘continuity assumption’ of Definition 3.1 throughout this section. Thus our results are valid for any von Neumann algebra that admits a faithful normal tracial state and has the above-mentioned property. Our version of Wielandt’s theorem is as follows: Theorem 4.1. Let F, X be the distribution and quantile function of a. Let δj ∈ R+ and tj 0, tj 0 < tk−1 1 ≤ tk 1, j = 1, · · · , k, be points in the spectrum of a such that t1 0) = δj, for all j. Then 1 ≤ · · · ≤ tk−1 0 and F (tj 0 < t1 1 ≤ t2 0 < t2 1) − F (tj 15 k Xj=1 Z[F (tj 0),F (tj 1)) X(s) ds = inf pj∈P(M ) p1≤···≤pk τ (pj)≥F (tj 1) sup qj∈P(M ) qj≤pj τ (qj)=δj qj⊥qi for j6=i k Xj=1 τ (aqj). Moreover, ∃p1 ≤ · · · ≤ pk with pj ∈ P(A) ⊂ P(M), for which there exist mutually orthogonal projections qj ≤ pj, τ (qj) = δj , ∀j such that k Xj=1 Z[F (tj 0),F (tj 1)) X(s) ds = max qj≤pj τ (qj )=δj qj⊥qi k Xj=1 τ (aqj); The following lemmas lead to the proof of the theorem above: Lemma 4.2. Let (M, τ ) be as above. Consider, for any k ≥ 2, k−1} ⊂ P(M), 1, · · · , q′ {r1, r2, · · · rk; q′ r1 ≥ · · · ≥ rk, τ (rj) ≥ δk + · · · + δj ∀ 1 ≤ j ≤ k, q′ j ≤ rj ∀1 ≤ j ≤ k − 1, q′ sq′ τ (q′ t = 0 ∀1 ≤ s < t ≤ k − 1, j) = δj ∀1 ≤ j < k − 1. Then there exist mutually orthogonal projections qj ≤ rj ∀1 ≤ j ≤ k in M, such that j=1 qj ≥ Pk−1 Pk j=1 q′ j, and τ (qj) = δj ∀1 ≤ j ≤ k. Proof. The proof follows by induction. For k = 2, choose q2 ≤ r2 such that τ (q2) = δ2. Let e = q2 ∨ q′ 1. Then τ (e) ≤ τ (q2) + τ (q′ 1) = δ2 + δ1 and e ≤ r1. But by the hypothesis for k = 2, τ (r1) ≥ δ2 + δ1. Hence by the ‘standing continuity assumption’, there exists f ∈ P(M) such that e ≤ f ≤ r1 and τ (f ) = δ2 + δ1. In particular q2 ≤ e ≤ f ; thus f − q2 ∈ P(M) with trace δ1. Choose q1 = f − q2. Then qj ≤ rj with trace δj for j = 1, 2 and q1 + q2 = f ≥ e ≥ q′ 1, as required. Suppose now, for the inductive step, that this result holds with k replaced by k − 1, and that r1, · · · , rk, q1, · · · , qk−1 are as in the statement of the Lemma. By induction hypothesis - applied to {r2, · · · , rk; q′ k−1} ⊂ P(M) - there exist mutually orthogonal projections q2, · · · , qk in M such that qj ≤ rj and τ (qj) = δj, ∀2 ≤ j ≤ k and 2, · · · , q′ k Xj=2 qj ≥ k−1 Xj=2 q′ j 16 (4.1) . Let e2 = q2 + · · · + qk and e = e2 ∨ q′ 1. Then τ (e) ≤ τ (e2) + τ (q′ But τ (r1) ≥ δk + · · · + δ1; thus (by the ‘standing continuity assumption’) there exists f ∈ P(M) such that e ≤ f ≤ r1 and τ (f ) = δk + · · · + δ1. In particular e2 ≤ e ≤ f ; thus f − e2 ∈ P(M) with trace δ1. 1) = (δk + · · · + δ2) + δ1 and e ≤ r1. Choose q1 = f − e2. Then q1 ≤ r1 and q1 ⊥ qj for 2 ≤ j ≤ k. Moreover, q1 + q2 + · · · + qk = f ≥ e = e2 ∨ q′ 1 = ≥ = k Xj=2 k−1 X2 X1 k−1 qj ∨ q′ 1 q′ j ∨ q′ 1 by equation 4.1 q′ j, thus completing the proof of the inductive step. ✷ Lemma 4.2 can be rewritten as: Lemma 4.3. Let (M, τ ) be as above. Suppose δj ∈ R+, and {r1 ≥ · · · ≥ rk} ⊂ P(M) such that τ (rj) ≥ δk + · · · δj, ∀j = 1, · · · , k and suppose we are given (k − 1) mutually orthogonal projections q′ j) = δj ∀ j = 1, · · · , k − 1. Let j such that q′ j ≤ rj and τ (q′ e′ = q′ 1 + · · · + q′ k−1 ≤ r1. Then there exist projections q ≤ r1 − e′, qj ≤ rj ∀ j = 1, · · · , k, such that τ (q) = δk and τ (qj) = δj ∀ j, {qj : 1 ≤ j ≤ k} pairwise mutually orthogonal and which is also a projection below r1. q + e′ = q1 + · · · + qk, Proof. Use Lemma 4.2 and choose q = (q1 + · · · + qk) − e′. ✷ Before proceeding further, we state a short but useful result: Lemma 4.4. For (M, τ ) as above and r, e ∈ P(M), τ (r ∧ e⊥) ≥ τ (r) − τ (e) where, of course, e⊥ = 1 − e. 17 Proof. as required. 1 + τ (r ∧ e⊥) ≥ τ (r ∨ e⊥) + τ (r ∧ e⊥) = τ (r) + 1 − τ (e) ✷ The above results lead to the following lemma: Lemma 4.5. Let (M, τ ), tj 1, δj be as in Wielandt’s theorem. Let {r1 ≥ · · · ≥ rk} and {p1 ≤ · · · ≤ pk} be sets of projections in M such that τ (pj) ≥ F (tj 0) for all 1 ≤ j ≤ k. Then there exist mutually orthogonal projections qj ≤ rj and mutually orthogonal projections qj ≤ pj such that τ (qj) = τ (qj) = δj ∀ j and q1 + · · · + qk = q1 + · · · + qk. 1), τ (rj) ≥ 1 − F (tj 0, tj Proof. The proof is by induction. For k = 1, deduce from Lemma 4.4 that τ (p1 ∧ r1) ≥ τ (p1) − τ (r⊥ 1 ) ≥ τ (p1) − 1 + τ (r1) ≥ F (t1 = F (t1 = δ1, 1) − 1 + 1 − F (t1 0) 1) − F (t1 0) and thus (by our standing ‘continuity assumption) there exists a projection q1 = q1 ≤ p1 ∧ r1 of trace δ1. For the inductive step, assume p1 ≤ · · · ≤ pk, r1 ≥ · · · rk are as in the lemma and that the lemma is valid with k replaced by k − 1. By the induction hypothesis applied to p1 ≤ · · · ≤ pk−1, r1 ≥ · · · ≥ rk−1, there are mutually orthogonal projections q′ j ≤ rj and mutually orthogonal projections qj ≤ pj such that τ (q′ j) = τ (qj) = δj for all j = 1, · · · , k − 1 and Pk−1 j=1 q′ j = Pk−1 j=1 qj =: e′, say. Then e′ ≤ pk−1 ≤ pk. Let ℓj = rj ∧ pk, ∀ j = 1, · · · , k. Then ℓk ≤ · · · ≤ ℓ1. An application of Lemma 4.4, as seen above in the k = 1 case, gives: τ (ℓj) ≥ F (tk ≥ F (tk = δk + · · · + δj ∀ j = 1, · · · , k. 1) − F (tj 0) 0) + F (tk−1 1) − F (tk 1 ) − F (tk−1 0 ) + · · · + F (tj 1) − F (tj 0) Now by Lemma 4.3 - applied with ℓj in place of rj - we may conclude that ∃ q ≤ ℓ1−e′, qj ≤ ℓj (≤ rj) with τ (q) = δk, τ (qj) = δj ∀ j and qj⊥qi ∀ j 6= i, such that q + e′ = q1 + · · · + qk. But q+e′ = q+q1+· · ·+qk−1, where qj ≤ pj ∀ j = 1, · · · , k−1 and q ≤ ℓ1−e′ ≤ ℓ1 = r1∧pk. Choosing qk = q, the proof of the inductive step is complete. ✷ 18 Now we are ready to prove Theorem 4.1. Proof. For “≥” : we take pj := 1(−∞,tj 3.3 that for each 1 ≤ j ≤ k, we do have 1)(a) and qj := 1[tj 1)(a) ≤ pj, and deduce from Remark 0,tj 0) Z F (tj F (tj 0) Xadm = τ (aqj) = max q∈P(M ) τ (aq) . q≤pj τ (q)=δ As the qj.1 ≤ j ≤ k are mutually perpendicular, we see that k Xj=1 0) Z F (tj F (tj 0) Xadm = k Xj=1 max qj∈P(M ) qj≤pj τ (qj)=δ τ (aqj) k Xj= τ (aqj) ≥ sup qj∈P(M ) qj≤pj τ (qj )=δ qi⊥qj for i6=j and in particular, k Xj=1 Z[F (tj 0),F (tj 1)) X(s) ds ≥ inf pj∈P(M ) p1≤···≤pk τ (pj)≥F (tj 1) sup qj∈P(M ) qj≤pj τ (qj )=δj qj⊥qi for j6=i k Xj=1 τ (aqj). For proving “≤” here, let us choose any p1 ≤ · · · ≤ pk such that pj ∈ P(M) and τ (pj) ≥ F (tj 1). Let rj = 1[tj Now by Lemma 4.5, there exist mutually orthogonal projections qj ≤ rj and mutually orthogonal projections qj ≤ pj with τ (qj) = τ (qj) = δj such that q1 + · · · + qk = q1 + · · · + qk. 0,∞)(a) ∀ j = 1, · · · , k. Then r1 ≥ · · · ≥ rk with τ (rj) = 1 − F (tj 0). Notice that by our version of Ky Fan’s theorem, τ (aqj) ≥ inf q∈P(M ) q≤rj τ (q)=δj τ (aq) = Z F (tj 1) F (tj 0) X(s) ds. Hence, k Xj=1 1) Z F (tj F (tj 0) τ (aqj) = k Xj=1 τ (aqj) X(s) ds ≤ k Xj=1 19 (since q1 + · · · + qk = q1 + · · · + qk), where qj ∈ P(M), qj ≤ pj with τ (qj) = δj and qj⊥qi. Hence, k Xj=1 Z[F (tj 0),F (tj 1)) X(s) ds ≤ sup qj∈P(M ) qj≤pj τ (qj )=δj qj⊥qi for j6=i k Xj=1 τ (aqj). As the p1 ≤ · · · ≤ pk were chosen arbitrarily, the proof of the theorem is complete. ✷ Acknowledgment It is a pleasure to record our appreciation of the very readable book [Bha97] by Rajendra Bhatia, whose proof of the matrix case of Wielandt’s theorem we could assimilate and adapt to the continuous case. We also wish to thank Manjunath Krishnapur and Vijay Kodiyalam for helpful discussions. 5 References [Bha97] Rajendra Bhatia, Matrix analysis, Graduate Texts in Mathematics, vol. 169, Springer-Verlag, New York, 1997. MR 1477662 (98i:15003) [BV93] Hari Bercovici and Dan Voiculescu, Free convolution of measures with unbounded support, Indiana Univ. Math. J. 42 (1993), no. 3, 733–773. MR 1254116 (95c:46109) [CH89] R. Courant and D. Hilbert, Methods of mathematical physics. Vol. II, Wiley Clas- sics Library, John Wiley & Sons Inc., New York, 1989, Partial differential equa- tions, Reprint of the 1962 original, A Wiley-Interscience Publication. MR 1013360 (90k:35001) [Fan49] Ky Fan, On a theorem of Weyl concerning eigenvalues of linear transformations. I, Proc. Nat. Acad. Sci. U. S. A. 35 (1949), 652–655. MR 0034519 (11,600e) [FK86] Thierry Fack and Hideki Kosaki, Generalized s- numbers of τ -measurable operators, Pacific J. Math. 123 (1986), no. 2, 269–300. [Hia87] Fumio Hiai, Majorization and stochastic maps in von Neumann algebras, J. Math. Anal. Appl. 127 (1987), no. 1, 18–48. MR 904208 (88k:46076) [HLP29] G. H. Hardy, J. E. Littlewood, and G. Polya, Some simple inequalities satisfied by convex functions, Messenger Math. 58 (1929), 145–152. [Hor62] Alfred. Horn, Eigenvalues of sums of Hermitian matrices, Pacific J. Math 12 (1962), 225–241. MR 0140521 (25 #3941) 20 [Kam83] Eizaburo Kamei, Majorization in finite factors, Math. Japon. 28 (1983), no. 4, 495–499. MR 717521 (84j:46086) [KM40] M. Krein and D. Milman, On extreme points of regular convex sets, Studia Math. 9 (1940), 133–138. MR 0004990 (3,90a) [Lid50] V. B. Lidskii, On the characterization numbers of the sum and product of symmet- ric matrices, Doklady Akad. Nauk SSSR (N.S.) 75 (1950), 769–772. MR 0039686 (12,581h) [Lid82] B. V. Lidskii, A polyhedron of the spectrum of the sum of two Hermitian matrices, Funktsional. Anal. i Prolozhen.. 16 (1982), no. 2, 76–77. MR 659172 (83k:15009) [Sak85] Y¯uji Sakai, Weak spectral order of Hardy, Littlewood and P´olya, J. Math. Anal. Appl. 108 (1985), no. 1, 31–46. MR 791129 (87k:26026) [Wie55] Helmut Wielandt, An extremum property of sums of eigenvalues, Proc. Amer. Math. Soc. 6 (1955), 106–110. MR 0067842 (16,785a) 21
1207.6923	2	1207	2013-03-15T14:47:09	The K-theory of some reduced inverse semigroup C*-algebras	[ "math.OA" ]	We use a recent result by Cuntz, Echterhoff and Li about the K-theory of certain reduced C-crossed products to describe the K-theory of C_r(S) when S is an inverse semigroup satisfying certain requirements. A result of Milan and Steinberg allows us to show that C*_r(S) is Morita equivalent to a crossed product of the type handled by Cuntz, Echterhoff and Li. We apply our result to graph inverse semigroups and the inverse semigroups of one-dimensional tilings.	math.OA	math	THE K-THEORY OF SOME REDUCED INVERSE SEMIGROUP C ∗-ALGEBRAS. MAGNUS DAHLER NORLING Abstract. We use a recent result by Cuntz, Echterhoff and Li about the K-theory of certain reduced C ∗-crossed products to describe the K-theory of C ∗ r (S) when S is an inverse semigroup satisfying certain requirements. A result of Milan and Steinberg allows us to show that C ∗ r (S) is Morita equivalent to a crossed product of the type handled by Cuntz, Echterhoff and Li. We apply our result to graph inverse semigroups and the inverse semigroups of one-dimensional tilings. 1. Introduction A semigroup is a set P with an associative binary operation. It is a monoid if it has an identity element. An inverse semigroup is a semigroup S where for every s ∈ S there is a unique element s∗ ∈ S satisfying ss∗s = s s∗ss∗ = s∗. We will assume that every inverse semigroup we are working with has a zero element 0 = 0S and that S is countable. Let E = E(S) = {ss∗ : s ∈ S} be the set of idempotents in S. Then E is a commutative idempotent semigroup, i.e. a semilattice. Let S× = S\{0}, and if S does not have an identity element let S1 = S ∪ {1} with the obvious operations. If S has an identity element, let S1 = S. See [14, 21] for general references on inverse semigroups. Every inverse semigroup comes equipped with a natural partial order given by s ≤ t if there is some e ∈ E such that s = et, equivalently, there is some f ∈ E such that s = tf . This is also equivalent to s = ss∗t and to s = ts∗s.The inverse semigroup S is said to be 0-F -inverse if every element in S× is beneath a unique maximal element with respect to the natural partial order. In this case, let M (S) denote the set of maximal elements in S. Let G be a group. A morphism [12] from S× to G (sometimes called grading of S) is a map σ : S× → G satisfying σ(st) = σ(s)σ(t) whenever st 6= 0 Note that σ(e) = 1 for every e ∈ E. The morphism σ is said to be idempotent pure if σ(s) = 1 implies s ∈ E for any s ∈ S. It is easy to see that if S is 0-F -inverse and σ : S× → G is a morphism that is injective on M (S), then σ is idempotent pure. 2010 Mathematics Subject Classification. Primary 46L05. Key words and phrases. C ∗-algebras, inverse semigroups, K-theory. 1 2 MAGNUS DAHLER NORLING Associated to every inverse semigroup S there is a C ∗-algebra C ∗(S) that is universal for all ∗-representations of S. Similarly there is a reduced C ∗- algebra C ∗ r (S) which is the image of the left regular representation Λ : C ∗(S) → B(ℓ2(S)). We use here the convention that all ∗-representations of S are to send the 0-element of S to 0. This convention coincides with the one used in [19]. See otherwise [21] for an introduction to C ∗-algebras associated with inverse semigroups. The main purpose of the paper is to provide a formula describing the K-theory of C ∗ r (S). The key ingredient will be a recent result from [4] concerning the K-theory of some crossed products C0(Ω) ⋊r G where G is a group and Ω is a locally compact totally disconnected (i.e. zero-dimensional) Hausdorff space. We will prove the following main theorem by showing that C ∗ r (S) is strongly Morita equivalent to such a crossed product under suitable conditions. Theorem 1.1 Let S be a 0-F -inverse semigroup, and suppose there exists a group G and a morphism σ : (S1)× → G that is injective on the set of maximal elements in S1. Let ≈ be the equivalence relation on E given by e ≈ f if there is some s ∈ S such that s∗s = e and ss∗ = f . Let eE be the set E×/ ≈. For each e ∈ E, let Ge = {σ(s) : ss∗ = s∗s = e}. If G is a-T-menable [3], then K∗(C ∗ r (S)) ≃ M[e]∈ eE K∗(C ∗ r (Ge)). Many important inverse semigroups are 0-F inverse. One example is the left inverse hull associated to a submonoid og a group satisfying the Toeplitz condition of [17]. Using this we will show that our main theorem generalizes [4, Corollary 4.9] which describes the K-theory of the reduced C ∗-algebra of such monoids. Other examples are graph inverse semigroups [22] and some tiling and point set inverse semigroups [12]. The reduced C ∗-algebra of the graph inverse semigroup is the Toeplitz C ∗-algebra of the graph as defined in [23]. We calculate the K-theory for some of these examples. In general, the reduced C ∗-algebra of S is not the only one of interest, but also its tight C ∗-algebra as defined in [5]. For instance, the Cuntz-Krieger graph C ∗-algebras and some of the tiling C ∗-algebras [8] are the tight C ∗-algebras of their respective inverse semigroups. The tight C ∗-algebra is a quotient of the reduced C ∗-algebra in the stated examples. We hope that the results of this paper may be useful in the future when trying to describe the K-theory of these tight C ∗-algebras. 2. Preliminaries on the construction of the Morita enveloping action In this and the following section we will fix a 0-F -inverse semigroup S, and we will suppose that there exists a group G and a morphism σ : (S1)× → G that is injective on M (S1). Note that if S is any inverse semigroup and σ : S× → G is a morphism such that σ−1(g) has a unique maximal element for each g ∈ G, then S is 0-F -inverse. THE K -THEORY OF SOME REDUCED INVERSE SEMIGROUP C ∗-ALGEBRAS. 3 Our exposition in the first part of this section follows [21, 5]. Let S be an inverse semigroup and let E be its semilattice of idempotents. A character on E is a homomorphism φ : E → {0, 1} that satisfies φ(0) = 0 and φ(e) = 1 for at least one e ∈ E. Here {0, 1} is considered a semilattice in the obvious way. inherited from the product topology on {0, 1}E . The locally compact space Let bE ⊂ {0, 1}E be the set of all characters on E with the relative topology bE is called the spectrum of E. For each e ∈ E, let De = {φ ∈ bE : φ(e) = 1}. Then De is a compact open subset of bE. For each e ∈ E, let φe be the φe(f ) =(1 if e ≤ f character given by 0 otherwise. (2.1) For any e, f ∈ E× it is easy to show that φe = φf if and only if e = f . Recall that there is an action θ of S by partial homeomorphisms on bE given by θs : Ds∗s → Dss∗, θs(φ)(e) = φ(s∗es), φ ∈ Ds∗s, e ∈ E. One can show that θs is a homeomorphism for each s ∈ S, and θ−1 s = θs∗. In [19] the partial action β of G on bE is defined as follows θs. βg = [σ(s)=g Since we are working with a 0-F -inverse semigroup and since σ is injective on M (S1), this expression can be simplified. Let sg be the uniqe element in M (S1) satisfying σ(sg) = g. Then βg = θsg . We now recall from [1] the construction of the Morita enveloping action (Ω, G, τ ) of a partial action (X, G, β). Define an equivalence relation ∼ on G × X by (g, x) ∼ (h, y) if x ∈ dom(βh−1g) and βh−1g(x) = y. Define Ω = ΩX = (G × X)/ ∼ with the quotient topology, and let [g, x] denote the equivalence class of (g, x) ∈ G × X. Let G act on Ω by τg([h, x]) = [gh, x]. By [1, Theorem 1.1], the map ι : x 7→ [1, x] defines a homeomorphism X → ι(X), with ι(X) open in Ω. Moreover, for all g ∈ G and x ∈ d(θg), (2.2) ι(θg(x)) = τg(ι(x)) It also follows that the orbit of ι(X) in Ω is all of Ω. We will from here on omit writing ι, and view X as a subset of Ω. Note in that [19] Milan and Steinberg assume that G is the universal group of S, but we want a more general result and do not want to demand that G is the universal group. Because of this we have to redo their proof that C ∗ r (S) is strongly Morita equivalent to C0(Ω) ⋊r G using a result from [13]. The universal groupoid Gu of S is the groupoid of germs for θ [5], and can be defined as follows. For s, t ∈ S× and φ ∈ Ds∗s, ψ ∈ Dt∗t, let (s, φ) ∼ (t, ψ) 4 MAGNUS DAHLER NORLING if and only if φ = ψ and there exists u ≤ s, t such that φ ∈ Du∗u. Let [s, φ] denote the equivalence class of (s, φ). Then define Gu = {[s, φ] : s ∈ S×, φ ∈ Ds∗s}, G(2) u = {([s, φ], [t, ψ]) ∈ Gu × Gu : φ = θt(ψ)} For ([s, φ], [t, ψ]) ∈ G(2) u , let [s, φ][t, ψ] = [st, ψ], [s, φ]−1 = [s∗, θs(φ)] By [5, Proposition 4.11], the unit space G(0) u sending [e, φ] to φ, where e ∈ E and φ ∈ De. The d and r maps from Gu to G(0) can be identified with bE by u = bE are given by d([s, φ]) = φ, r([s, φ]) = θs(φ). For each s ∈ S and open U ⊂ Ds∗s, let Θ(s, U ) = {[s, φ] : φ ∈ U } The collection {Θ(s, U ) : s ∈ S, U ⊂ Ds∗s is open} is a basis for the topology on Gu. Then for each s ∈ S, the restrictions d : Θ(s, Ds∗s) → Ds∗s, r : Θ(s, Ds∗s) → Dss∗ are homeomorphisms. The reduced C ∗-algebra of S is isomorphic to the reduced C ∗-algebra of Gu [21]. In [13], a continuous cocycle on a groupoid G is defined to be a continuous groupoid homomorphism ρ : G → H, where H is a group. The cocycle ρ is said to be (i) faithful if the map G 7→ G(0)×H×G(0) given by γ 7→ (r(γ), ρ(γ), d(γ)) is injective. (ii) closed if the map γ 7→ (r(γ), ρ(γ), d(γ)) is closed. (iii) transverse if the map H ×G → H ×G given by (γ, g) 7→ (gρ(γ), d(γ)) is open. If ρ is faithful, closed and transverse, one defines the enveloping action of ρ as follows. Let ∼ be the equivalence relation on H × G(0) given by (g, x) ∼ (h, y) if there exists a γ ∈ G with d(γ) = y, r(γ) = x and ρ(γ) = h−1g. Let Ω be the space (H × G(0))/ ∼ with the quotient topology, and let τ be the action of H on Ω given by τg([h, x]) = [gh, x]. By [13, Lemma 1.7 and Theorem 1.8], G is Morita equivalent to the transformation groupoid H ⋉ Ω, and so C ∗ r (G) is strongly Morita equivalent to C0(Ω) ⋊τ,r H. Let Gu be the universal groupoid of S and let σ : (S1)× → G be the morphism specified earlier. Define a map ρ : Gu → G by ρ([s, φ]) = σ(s). To see that this is well defined, note that if 0 6= u ≤ s, t, then σ(u) = σ(s) = σ(t). It is straightforward to check that ρ is a continuous cocycle. Inspection reveals that the enveloping action of ρ is the Morita enveloping action of β. THE K -THEORY OF SOME REDUCED INVERSE SEMIGROUP C ∗-ALGEBRAS. 5 To see this, note that the equivalence relation one obtains on G × bE is the same in either case. It remains to prove the following proposition. The proof closely follows the proof of [13, Propositions 3.6 and 3.9]. Proposition 2.1 The cocycle ρ : Gu → G is faithful, closed and transverse. Proof. We prove faithfulness first. Let γ = [s, φ] ∈ Gu. Then (r(γ)ρ(γ), d(γ)) = (θs(φ), σ(s), φ). Suppose (θs(φ), σ(s), φ) = (θt(ψ), σ(t), ψ). Then φ = ψ and σ(s) = σ(t). Since φ ∈ Dt∗t ∩ Ds∗s = Dt∗ts∗s, t∗ts∗s 6= 0. So ts∗ 6= 0, and σ(ts∗) = σ(t)σ(s)−1 = σ(t)σ(t)−1 = 1. Thus ts∗ is idempotent. Let u = ts∗s. Then 0 6= u ≤ s, t and u∗u = s∗st∗t, so φ ∈ Du∗u. It follows that [s, φ] = [t, ψ]. Next we prove closedness. Let X ⊂ Gu be closed. Then {g} × {r(γ), d(γ)) : γ ∈ X, ρ(γ) = g}. (ρ, r, d)(X) = [g∈σ(S×) Since G is discrete, we only have to show that {(d(γ), r(γ)) : γ ∈ X, ρ(γ) = g} is closed for each g ∈ σ(S×). If g = 1, then this set is just (X ∩bE)2. Since bE is a closed subset of Gu, this is closed. If g 6= 1, then σ−1(g) has a maximal element sg, and [s, φ] = [sg, φ] for each [s, φ] ∈ Gu with σ(s) = g, so {(d(γ), r(γ)) : γ ∈ X, ρ(γ) = g} = {(d(γ), r(γ)) : γ ∈ X ∩ Θ(s, Ds∗ gsg )} Since d and r are homeomorphisms of Θ(sg, Ds∗ the assertion follows. gsg ) onto closed subsets of bE, Finally, we prove transversity. Since G is discrete it is sufficient to show that the set {d(γ) : γ ∈ Gu, ρ(γ) = g} = [s∈σ−1(g) Ds∗s is open for each g ∈ σ(S×), but this is obviously true. (cid:3) Remark 2.2 It is shown in [19] that Morita equivalence with the enveloping action holds for a slightly larger class of inverse semigroups than the one we have considered here, but this restriction is needed in some of the proofs of the next section. 3. Proof of the main theorem Recall that a locally compact Hausdorff space is totally disconnected if and only if it has a basis of compact open subsets. For a totally disconnected space X, let Uc(X) be the family (actually a Boolean algebra) of compact open subsets of X. In [4], a subfamily V ⊂ Uc(X) is said to be a generating family for Uc(X) if Uc(X) is the smallest family of compact open sets that contains V and is closed under finite intersections, finite unions and taking difference sets (i.e. V generates Uc(X) as a Boolean algebra). The family V i=1 Un implies that U = Uj for some 1 ≤ j ≤ n. If V is an independent generating family is said to be independent if for any U, U1, . . . , Un ∈ V, U =Sn 6 MAGNUS DAHLER NORLING for Uc(X) and V ∪ {∅} is closed under finite intersections, then V is said to be a regular basis for X. For notational convenience we will require that ∅ /∈ V. The following theorem is a part of the statement of [4, Corollary 3.14]. Theorem 3.1 ([4]) Let X be a separable totally disconnected locally com- pact Hausdorff space with an action α of a group G and suppose that X has a G-invariant regular basis V. Suppose also that G satisfies the Baum-Connes conjecture with coefficients in C0(X) and C0(V). Then K∗(C0(X) ⋊α,r G) ≃ M[v]∈V\G K∗(C ∗ r (Gv)) where V \ G is the set of orbit classes in V and Gv is the stabilizer group of v ∈ V. We will use Theorem 3.1 to prove Theorem 1.1. The way we will do it is to apply Theorem 3.1 to the enveloping action (Ω, G, τ ) of (bE, G, β). We will show that (3.1) V = {τg(De) : e ∈ E×, g ∈ G} is a G-invariant regular basis for Ω. It is shown in [7] that if G is a-T-menable (i.e. it satisfies the Haagerup property), then it satisfies the strong Baum- Connes conjecture with arbitrary separable coefficients. The statement of Theorem 1.1 could have been made more general by not requiring that G is a-T-menable, but only that it satisfies the Baum-Connes conjecture with coefficients in C0(Ω) and C0(V). However, that would have made the state- ment more complicated by making explicit reference to the enveloping action instead of just utilizing it in the proof. The next proposition will allow us to translate statements about the partial action into statements about the enveloping action. Proposition 3.2 Let X ⊂ Y be locally compact Hausdorff spaces such that X is totally disconnected and open in Y . Let α be an action of a discrete group G on Y such thatSg∈G αg(X) = Y . Let J be a collection of compact open subsets of X. Suppose αg(p)∩ X ∈ J ∪ {∅} for every p ∈ J and g ∈ G. Let W = {αg(p) : g ∈ G, p ∈ J }. Then (i) If J ∪ {∅} is closed under finite intersections, so is W ∪ {∅}. (ii) If J is independent, so is W. (iii) If J is a generating family for Uc(X), then W is a generating family for Uc(Y ) and Y is totally disconnected. So if J is a regular basis for X, then W is a G-invariant regular basis for Y . Proof. (i) It is sufficient to show that the intersection of two elements in W is in W ∪{∅}. Let p, q ∈ J ∪{∅} and g, h ∈ G. Let r = αg−1h(q)∩X ∈ J ∪{∅}. Then αg(p) ∩ αh(q) = αg(p ∩ αg−1h(q)) = αg(p ∩ (αg−1h(q) ∩ X)) = αg(p ∩ r). Since J ∪ {∅} is closed under finite intersections, p ∩ r ∈ J ∪ {∅}, so αg(p ∩ r) ∈ W ∪ {∅}. THE K -THEORY OF SOME REDUCED INVERSE SEMIGROUP C ∗-ALGEBRAS. 7 (ii) Let p, p1, . . . , pn ∈ J and let g, g1, . . . , gn ∈ G. Suppose Then αgi(pi) = αg(p). αg−1gi(pi) = p. n[i=1 n[i=1 For each i, αg−1gi(pi) ⊂ p ⊂ X, so αg−1gi(pi)∩X = αg−1gi(pi) ∈ J . It follows by the independence of J that p = αg−1gj (pj) for at least one 1 ≤ j ≤ n, so αgj (pj) = αg(p). (iii) First, any element in W is compact open in Y since X is open in Y . Let U ⊂ Y be compact open. We have U = U ∩ [g∈G αg(X) = [g∈G αg(αg−1(U ) ∩ X). and αg−1(U ) ∩ X is open in Y for each g ∈ G. By the compactness of U , there is some finite subset F ⊂ G such that U = [g∈F αg(αg−1(U ) ∩ X). For each g ∈ G, αg−1(U )∩X is compact open in X. So since J is a generating family for Uc(X), it follows that W is a generating family for Uc(Y ). As X is totally disconnected, any open set in X is a union of elements in Uc(X). It follows by a similar argument that any open set in Y is a union of elements in Uc(Y ), so Y is totally disconnected. (cid:3) g . gsg. Lemma 3.3 For any e ∈ E and g ∈ σ(S×) we have τg(De) ∩ bE = Dsges∗ Moreover τg(De) ⊂ bE if and only if e ≤ s∗ Proof. For any g ∈ σ(S×) and φ ∈ bE, we have that [g, φ] ∈ bE if and only if [g, φ] = [1, ψ] for some ψ ∈ bE if and only if φ ∈ Ds∗ Then τg(De) ∩ bE = θsg (De ∩ Ds∗ gsg , so τg(De) = θsg(De) ⊂ bE. Suppose τg(De) ⊂ bE. Then φ ∈ Ds∗ gsg and θsg (φ) = ψ. gsg ). By [5, Equation (10.3.1)] this set is gsg for each φ ∈ De, so De ⊂ Ds∗ gsg, then De ⊂ Ds∗ Given any e ∈ E× and finite subset Z ⊂ eE = {f ∈ E : f ≤ e}, let equal to Dsges∗ g. If e ≤ s∗ gsg and (cid:3) e ≤ s∗ gsg. D(e,Z) = De \ [z∈Z Dz! . Note that since Dx is compact open for each x ∈ E, so is D(e,Z). Recall from [21, Section 4.3] that the family T = {D(e,Z) : e ∈ E×, Z ⊂ eE is a finite subset} is a basis for the topology on bE. We now see that any compact open subset of bE can be written as a finite union of elements in T : Indeed, since T is a basis for the topology on bE, any open set U in in bE can be written as a union 8 MAGNUS DAHLER NORLING of elements in T . These elements form a cover for U , so if U is compact, one can rewrite U as a finite union of such elements. Proposition 3.4 The family V defined in equation (3.1) is a G-invariant regular basis for Ω. Proof. If we can show that the family J = {De : e ∈ E×} is a regular basis We know that J ∪ {∅} is closed under finite intersections, since for any e, f ∈ E, De ∩ Df = Def (and D0 = ∅). To show that J is independent, i=1 Dei = De. For each i, Dei ⊂ De, so ei ≤ e. Moreover, φe ∈ Dej for at least one 1 ≤ j ≤ n. Then φe(ej) = 1, so e ≤ ej by the definition of φe. Thus e = ej and De = Dej . Finally, we for bE, the result will follow from Proposition 3.2 and Lemma 3.3. suppose that e, e1, . . . , en ∈ E× satisfySn need that J is a generating set for Uc(bE). This follows from the fact that any compact open subset of bE is a finite union of elements in T . Recall from Theorem 1.1 that we defined a relation ≈ on E× by e ≈ f if there is some s ∈ S× such that s∗s = e and ss∗ = f . (cid:3) Lemma 3.5 Let e, f ∈ E×. Then De and Df are in the same G-orbit in V if and only if e ≈ f . Proof. Suppose g ∈ G satisfies τg(De) = Df . Then by Lemma 3.3 e ≤ s∗ and sges∗ ss∗ = sgees∗ g = f . Let s = sge. Then s∗s = es∗ gsge = e since e ≤ s∗ gsg gsg, and g = sges∗ Suppose to the contrary that there is some s ∈ S such that s∗s = e and ss∗ = f . Let g = σ(s). Then s ≤ sg. It follows that sge = sgs∗s = s, so e ≤ s∗ (cid:3) gsg, and f = ss∗ = sges∗ g. So τg(De) = Df . g = f . Given e ∈ E×, let Ge = {g ∈ σ(S×) : e ≤ s∗ gsg, sges∗ g = e} = {σ(s) : s ∈ M (S1), e ≤ s∗s, ses∗ = e} = {σ(s) : s ∈ S×, e ≤ s∗s, ses∗ = e} = {σ(s) : s ∈ S×, ss∗ = s∗s = e}. We leave to the reader to prove that these definitions are equivalent. Lemma 3.6 Let e ∈ E×. Then Ge is the stabilizer group of De ∈ V. Proof. Assume τg(De) = De. Then τg(De) ⊂ bE, so by Lemma 3.3 e ≤ s∗ gsg, g. The converse implication follows from the same (cid:3) De = Dsges∗ lemma. g and e = sges∗ We can now complete the proof of the main theorem. Since G is a-T- menable, Theorem 3.1 together with Proposition 3.4 gives us that K∗(C ∗ r (S)) ≃ K∗(C0(Ω) ⋊r G) ≃ M[v]∈V\G K∗(C ∗ r (Gv)) For any v ∈ V we have that v = τg(De) for some e ∈ E×, so [v] = [De]. Thus it is sufficient to take a direct sum over {[De] : e ∈ E×}, which can be identified with eE by Lemma 3.5. The rest follows from Lemma 3.6. THE K -THEORY OF SOME REDUCED INVERSE SEMIGROUP C ∗-ALGEBRAS. 9 4. Connections to submonoids of groups Given a set X, let I(X) be the inverse monoid of all partial bijections on X. Here a partial bijection is a bijection f : d(f ) → r(f ) with d(f ), r(f ) ⊂ X, f ∗ = f −1 and the product is given by composition wherever it makes sense. This product may result in the empty function, which is the 0-element of I(X). We will now explain how the "0-F -inverse" condition for inverse semi- groups is related to the Toeplitz condition for subsemigroups of groups as defined in [17, 4]. Let P be a submonoid of the (discrete) group G. For each p ∈ P , let λp : P → pP be the map given by λp(q) = pq. Then λp ∈ IP since P is left cancellative. Let I λ P be the inverse subsemigroup of IP generated by {λp : p ∈ P } ∪ {0}. Since P is a monoid, I λ P is called the left inverse hull of P [18]. Each idempotent e ∈ E(I λ P ) is the identity function on some subset X ⊂ P . Let J = {d(e) : e ∈ E(I λ P )}. Then it is easy to show that E(I λ P ) and (J , ∩) are isomorphic as semilattices. P is also a monoid. I λ Let {εp : p ∈ P } be the canonical basis for ℓ2(P ). For each p ∈ P , let r (P ) to be the C ∗-algebra It was shown in [20, Theorem 3.2.14] that r (P ) given by Vp ∈ B(ℓ2(P )) be given by Vpεq = εpq. Define C ∗ generated by {Vp : p ∈ P }. if J is independent, there is an isomorphism C ∗ Λ(λp) 7→ Vp. P ) → C ∗ r (I λ As shown in [20, Proposition 3.2.11], one can define an idempotent pure P )× → G by σ(λp) = p. Moreover, any group H and P → H gives rise to an embedding P → H morphism σ : (I λ idempotent pure morphism σ : I λ given by p 7→ σ(λp). Note that for any f ∈ I λ P and p ∈ d(f ), (4.1) f (p) = σ(f )p Let {εg : g ∈ G} be the canonical basis for ℓ2(G). Let Ug ∈ B(ℓ2(G)) be given by Ugεh = εgh (i.e. g 7→ Ug is the left regular representation of G). Let EP be the orthogonal projection of ℓ2(G) onto the subspace ℓ2(P ). By [17, Definition 4.1], P ⊂ G satisfies the Toeplitz condition if for any g ∈ G with EP UgEP 6= 0 there exist p1, . . . , pn, q1, . . . , qn ∈ P such that EP UgEP = V ∗ q1Vp1 · · · V ∗ qnVpn. Proposition 4.1 Let P be a submonoid of the group G. Then P ⊂ G satisfies the Toeplitz condition if and only if I λ P is 0-F -inverse and σ is injective on the set of maximal elements in I λ P . Proof. For g ∈ G, let αg : (g−1)P ∩ P → gP ∩ P be given by αg(p) = gp. Then αg ∈ IP . We claim that P ⊂ G satisfies the Toeplitz condition if and P for each g ∈ G. To verify this, let ω : IP → B(ℓ2(P )) be only if αg ∈ I λ given by ω(f )εp =(εf (p) if p ∈ d(f ) 0 otherwise. Note that ω is injective and that ω(λp) = Vp for each p ∈ P . Moreover, for any g ∈ G, it is easy to verify that ω(αg) = EP UgEP . Fix g ∈ G and assume without loss of generality that αg 6= 0. Suppose P satisfies σ(f ) = g. By P . Clearly σ(αg) = g. Suppose f ∈ I λ αg ∈ I λ 10 MAGNUS DAHLER NORLING equation (4.1) it follows that f ≤ αg. This implies that αg is the unique maximal element of σ−1(g). Conversely, suppose I λ P is 0-F -inverse and that σ is injective on M (I λ P ). Let f ∈ σ−1(g) be the maximal element. If we can show that d(f ) = d(αg) = (g−1)P ∩ P , then equation (4.1) gives us that f = αg. We already noted that equation (4.1) implies that d(f ) ⊂ d(αg), so let p ∈ (g−1)P ∩ P . Then there is some q ∈ P such that p = g−1q, so g = qp−1. Then σ(λqλ∗ p) = g and p ∈ pP = d(λqλ∗ p) ⊂ d(f ), so we are done. (cid:3) p). Since f is maximal in σ−1(g), p ∈ d(λqλ∗ The computation of the K-theory for C ∗ r (P ) in the case when J is in- dependent and P ⊂ G satisfies the Toeplitz condition is done in details in [4]. The same computation could now be obtained using Theorem 1.1 and Proposition 4.1. 5. Graph inverse semigroups We will use [23] as our main source on graphs and graph C ∗-algebras. A (discrete) directed graph E = (E 0, E 1, s, r) consists of countable sets E 0, E 1 and functions s, r : E 1 → E 0. The elements of E 0 are called the vertices of E, while the elements of E 1 are called the edges of E. A finite path in E is either a single vertex or a finite string µn . . . µ1 of edges such that r(µk) = s(µk+1) for any 0 ≤ k ≤ n − 1. Let E ∗ be the set of all finite paths in E. For a path µ = µn · · · µ1 ∈ E ∗, define s(µ) = s(µ1) and r(µ) = r(µn). If µ is a single vertex, then define s(µ) = r(µ) = µ. One can define a partial product on E ∗ as follows. If µ and ν are strings of edges, then their product µν is given by concatenation if the resulting string is a path, otherwise the product is undefined. If µ is a string of edges and v is a vertex, then µv = µ if s(µ) = v and vµ = µ if r(µ) = v. If v, w are vertices, then vw = v if v = w. In [22] (see also [15]) the graph inverse semigroup SE is defined to be the set SE = {(µ, ν) ∈ E ∗ × E ∗ : s(µ) = s(ν)} ∪ {0} with all products not involving 0 given by (µ, ν)(α, β) = (µ, βν′) if ν = αν′ (µα′, β) if α = να′ 0 otherwise. The ∗-operation is given by (µ, ν)∗ = (ν, µ). For (µ, ν) ∈ SE , we have (µ, ν)∗(µ, ν) = (ν, µ)(µ, ν) = (ν, ν) This shows that E(SE )× = {(ν, ν) : ν ∈ E ∗}. We have (µ, ν) ≤ (α, β) if and only if there is some ρ ∈ E ∗ such that µ = αρ and ν = βρ. In particular we get that for v ∈ E 0 we have (µ, µ) ≤ (v, v) if and only if µ = vµ if and only if r(µ) = v. The maximal elements of S1 E are M (S1 E ) = {1} ∪ {(µ, ν) : µ and ν have no common initial segment.} THE K -THEORY OF SOME REDUCED INVERSE SEMIGROUP C ∗-ALGEBRAS. 11 Every element is beneath such a maximal element, so S1 Let F be the free group on the alphabet E 1. Define h : E ∗ → F by E is 0-F -inverse [16]. h(µ) =(1 if µ is a vertex, µ otherwise. Define σ : (S1 is a morphism and that it is injective on the maximal elements of S1 next lemma lets us identify ^E(SE ) = E(SE )/ ≈ with E 0. E )× → F by σ((µ, ν)) = h(µ)h(µ)−1. It is easy to check that σ E . The Lemma 5.1 Let µ, ν ∈ E ∗. Then (µ, µ) ≈ (ν, ν) if and only if s(µ) = s(ν). Proof. Fix µ, ν ∈ E ∗. Suppose s(µ) = s(ν). Then (ν, ν) = (µ, ν)∗(µ, ν) and (µ, µ) = (µ, ν)(µ, ν)∗. Conversely, suppose (ν, ν) ≈ (µ, µ). Then there exists (α, β) ∈ SE such that (ν, ν) = (α, β)∗(α, β) = (β, β). and (µ, µ) = (α, β)(α, β)∗ = (α, α). It follows that ν = β and µ = α, so s(µ) = s(α) = s(β) = s(ν). (cid:3) Lemma 5.2 For each µ ∈ E ∗, G(µ,µ) is the trivial group. Proof. Suppose (α, β) ∈ SE satisfies (µ, µ) = (β, α)(α, β) = (α, β)(β, α). Then clearly α = β = µ, so σ((α, β)) = h(µ)h(µ)−1 = 1. (cid:3) It is possible to show that C ∗ r (SE ) is canonically isomorphic to the Toeplitz C ∗-algebra of E as defined in [23]. Thus the next proposition also follows from [2, Theorem 1.1]. Proposition 5.3 We have K0(C ∗ Proof. By Lemma 5.2, Ge = {1} for each e ∈ E(SE ). Then K0(C ∗ K0(C) = Z and K1(C ∗ [3]. The result now follows from Lemma 5.1 and Theorem 1.1. r (Ge)) = r ({1})) = K1(C) = 0. Since F is free, it is a-T-menable (cid:3) r (SE )) =LE 0 Z and K1(C ∗ r (SE )) = 0. 6. Tiling inverse semigroups The tiling inverse semigroups were first introduced in [10, 11]. A tile in Rn is a subset of Rn homeomorphic to the closed unit ball. A partial tiling is a collection of tiles that have pairwise disjoint interiors. The support of a partial tiling is the union of its elements. A tiling is a partial tiling with support equal to Rn. A patch is a finite partial tiling. For any tile t and x ∈ Rn, let t + x be the translation of t by x. For any partial tiling P , let P + x = {t + x : t ∈ P }. Let T be a tiling. Let M be the set of subpatches P of T such that P has connected support. For P, Q ∈ M and t1, t2 ∈ P , r1, r2 ∈ Q, we say that (t1, P, t2) ∼ (r1, Q, r2) if there is some x ∈ Rn such that t1 + x = r1, t2 + x = r2 and P + x = Q. The equivalence class of (t1, P, t2) is denoted [t1, P, t2], and is called a doubly pointed pattern class. Let ST = {[t1, P, t2] : P ∈ M, t1, t2 ∈ P } ∪ {0} We define an inverse semigroup structure on ST . Let [t1, P, t2], [r1, Q, r2] ∈ ST . Whenever there are x, y ∈ Rn such that P + x and Q + y are patches in T and t2 + x = r1 + y, define [t1, P, t2][r1, Q, r2] = [t1 + x, (P + x) ∪ (Q + y), r2 + y]. 12 MAGNUS DAHLER NORLING All other products are defined to be 0. The ∗-operation is given by [t1, P, t2]∗ = [t2, P, t1] ST is called the connected tiling semigroup of T . If one drops the requirement that the elements of M have connected support, one gets the tiling semigroup ΓT . It was shown in [6] that when T is so-called strongly aperiodic, repetitive and has finite local complexity (see the original paper for definitions), the tiling C ∗-algebra AT from [9, 8] is the tight C ∗-algebra of ST . We do not know if the reduced C ∗-algebra of ST or ΓT has been previously studied. : P ∈ M, t ∈ P }. Clearly, E(ST ) = {[t, P, t] It is easy to see that [t, P, t] ≤ [r, Q, r] if and only if there is an x ∈ Rn such that r = t + x and Q ⊂ P + x. Say that two partial tilings are congruent if one is a translation of the other. Let L be the set of congruence classes of elements in M. The next lemma lets us identify ^E(ST ) (respectively ^E(ΓT )) with L. Lemma 6.1 Let P, Q ∈ M and t ∈ P , r ∈ Q. Then [t, P, t] ≈ [r, Q, r] if and only if P and Q are congruent. Proof. Suppose P and Q are congruent. Then there is some x ∈ Rn such that Q = P + x. So t + x ∈ Q. Moreover, it is easy to check that [r, Q, r] = [t + x, Q, r]∗[t + x, Q, r] and [t + x, Q, r][t + x, Q, r]∗ = [t + x, Q, t + x] = [t, Q − x, t] = [t, P, t]. So [t, P, t] ≈ [r, Q, r]. Conversely, suppose [t, P, t] ≈ [r, Q, r]. Then there is some R ∈ M and a, b ∈ R such that [t, P, t] = [a, R, b]∗[a, R, b] = [b, R, b], [r, Q, r] = [a, R, b][a, R, b]∗ = [a, R, a] It follows that P and Q are both congruent to R and thus to each other. (cid:3) ST is not 0-F -inverse in general. It is 0-F -inverse when n = 1 [12, Propo- sition 4.2.1]. In this case, the universal grading of ST is well understood. Lemma 6.2 For each e ∈ E(ST )×, Ge is the trivial group. Proof. Let P ∈ M and t ∈ P . Suppose that R ∈ M and a, b ∈ R satisfy [t, P, t] = [a, R, b]∗[a, R, b] = [b, R, b], [t, P, t] = [a, R, b][a, R, b]∗ = [a, R, a] This shows that a = b, so [a, R, b] is idempotent, and σ([a, R, b]) = 1 regard- less of what σ is. (cid:3) Proposition 6.3 Let T be a one-dimensional tiling. Then K0(C ∗ r (ST )) = Z and K1(C ∗ r (ST )) = 0. LL Proof. By [12, Corollary 4.2.4], the universal group of ST is free, so it is a-T- menable.To show that ST is 0-F -inverse it remains to check that the universal grading σ is injective on the set of nonidempotent maximal elements in ST . One has then to go through the actual construction of σ in [12]. This requires THE K -THEORY OF SOME REDUCED INVERSE SEMIGROUP C ∗-ALGEBRAS. 13 some work, but is not hard. We leave this verification to the reader. The rest now follows from Lemmas 6.1, 6.2 and Theorem 1.1. (cid:3) Note also that for any n-dimensional tiling T , the tiling semigroup ΓT has the property that any element is beneath a maximal element [12]. If one can find a good group morphism for ΓT , one will get similar results for the K-theory of C ∗ In [12] Kellendonk and Lawson also showcase other 0-F -inverse semigroups, such as inverse semigroups of point sets in Rn and point set semigroups of model sets, where such morphisms exist. r (ΓT ). References [1] Fernando Abadie. Enveloping actions and Takai duality for partial ac- tions. J. of Funct. Anal., 197:14 -- 67, 2003. [2] Bernhard Burgstaller. The K-theory of certain C ∗-algebras endowed with gauge actions. J. Funct. Anal., 256:1693 -- 1707, 2009. [3] P.-A. Cherix, M. Cowling, P. Jolissaint, P. Julg, and A. Valette. Groups with the haagerup property. In Gromov's a-T-menability, volume 197 of Progress in mathematics, Basel, 2001. Birkhauser. [4] Joachim Cuntz, Siegfried Echterhoff, and Xin Li. On the K-theory of crossed products by automorphic semigroup actions. Preprint (2012) arXiv:1205.5412v1. [5] R. Exel. Inverse semigroups and combinatorial C ∗-algebras. Bull. Braz. Math. Soc., 39:191 -- 313, 2008. [6] R. Exel, D. Gonçalves, and C. Starling. The tiling C ∗-algebra viewed as a tight inverse semigroup algebra. Semigroup forum, 84:229 -- 240, 2012. [7] N. Higson and G. Kasparov. E-theory and KK-theory for groups which act properly and isometrically on Hilbert space. Invent. Math., 144:23 -- 74, 2001. [8] J. Kellendonk and I. F. Putnam. Tilings, C ∗-algebras, and K-theory. CRM Monograph Series, 13:177 -- 206, 2000. [9] Johannes Kellendonk. Non commutative geometry of tilings and gap labelling. Rev. Math. Phys, 7:1133 -- 1180, 1995. [10] Johannes Kellendonk. The local structure of tilings and their integer group of covariants. Commun. Math. Phys., 187:115 -- 157, 1997. [11] Johannes Kellendonk. Topological equivalence of tilings. J. Math. Phys., 38:1823 -- 1842, 1997. [12] Johannes Kellendonk and Mark V. Lawson. Universal groups for point- sets and tilings. J. of Alg., 276:462 -- 492, 2004. [13] M. Khoshkam and G. Skandalis. Regular representations of groupoid C ∗-algebras and applications to inverse semigroups. J. Reine Angew. Math., 546:47 -- 72, 2002. [14] Mark V. Lawson. Inverse semigroups, the theory of partial symmetries. World Scientific Publishing, Singapore, 1998. [15] Mark V. Lawson. Constructing inverse semigroups from category ac- tions. J. Pure and Appl. Alg., 137:57 -- 101, 1999. [16] Mark V. Lawson. E∗-unitary inverse semigroups. In Semigroups, algo- rithms, automata and languages, pages 195 -- 214, River Edge, NJ, 2001. World Sci. Publishing. 14 MAGNUS DAHLER NORLING [17] Xin Li. Nuclearity of semigroup C ∗-algebras and the connection to amenability. Preprint (2012) arXiv:1203.0021v2. [18] John Meakin. Groups and semigroups: connections and contrasts. In Groups St Andrews, number 340 in London Math. Soc. Lecture Note Series, pages 357 -- 400. Cambridge University Press, 2005. [19] David Milan and Benjamin Steinberg. On inverse semigroup C ∗-algebras and crossed products. Preprint (2011) arXiv:1104.2304v1. [20] Magnus Dahler Norling. Inverse semigroup C ∗-algebras associated with left cancellative semigroups. Preprint (2012) arXiv:1202.5977v2. To appear in Proc. Edin. Math. Soc. [21] Alan L. T. Paterson. Groupoids, Inverse Semigroups, and their Operator algebras, volume 170 of Progress in Mathematics. Birkhauser, Boston, MA, 1999. [22] Alan L. T. Paterson. Graph inverse semigroups, groupoids and their C ∗-algebras. J. Oper. Th., 48:645 -- 662, 2002. [23] Iain Raeburn. Graph Algebras. Number 103 in Regional Conference Series in Mathematics. Amer. Math. Soc., Providence, RI, 2005. Magnus Dahler Norling, Institute of Mathematics, University of Oslo, P.b. 1053 Blindern, 0316 Oslo, Norway. E-mail address: [email protected]
1005.2849	1	1005	2010-05-17T09:26:31	Relatively independent joinings and subsystems of W*-dynamical systems	[ "math.OA" ]	Relatively independent joinings of W-dynamical systems are constructed. This is intimately related to subsystems of W-dynamical systems, and therefore we also study general properties of subsystems, in particular fixed point subsystems and compact subsystems. This allows us to obtain characterizations of weak mixing and relative ergodicity, as well as of certain compact subsystems, in terms of joinings.	math.OA	math	RELATIVELY INDEPENDENT JOININGS AND SUBSYSTEMS OF W-DYNAMICAL SYSTEMS ROCCO DUVENHAGE Abstract. Relatively independent joinings of W-dynamical sys- tems are constructed. This is intimately related to subsystems of W-dynamical systems, and therefore we also study general prop- erties of subsystems, in particular fixed point subsystems and com- pact subsystems. This allows us to obtain characterizations of weak mixing and relative ergodicity, as well as of certain compact sub- systems, in terms of joinings. 1. Introduction In the study of the general mathematical structure of quantum dy- namical systems and quantum statistical mechanics, the operator alge- braic approach has proven very valuable. In particular the framework of von Neumann algebras, in which case the dynamical system is called a W-dynamical system (see Section 2 for a precise definition), provides a natural arena for ergodic theory which extends the classical measure theoretic framework. Refer to [6, 7] for an account of many aspects of these topics. The first aim of this paper is to study the extension of the concept of a relatively independent joining of two dynamical systems in classi- cal ergodic theory to the noncommutative framework of W-dynamical systems. It is essentially a generalized way of forming the tensor prod- uct of two systems which takes into account a common subsystem of the systems. A clear exposition of the classical case can be found in [15, Chapter 6]. Since the early works [14, 30], joinings have been a useful tool in classical ergodic theory (again see [15]). This paper, building on [9, 10], is part of a programme to systematically develop the theory of joinings in the noncommutative case. The goal is to eventually have a similarly useful tool for noncommutative dynamical systems. Joinings have in fact already gradually found some use in noncommu- tative dynamical systems, in particular related to dynamical entropy [31] where a special case of joinings appears (also see [25, Section 5.1]). Date: 2010-5-15. 2000 Mathematics Subject Classification. Primary 46L55. Key words and phrases. W-dynamical systems; relatively independent joinings; subsystems. 1 2 ROCCO DUVENHAGE Early work related to joinings in a noncommutative setting goes back to at least [4], where disjointness was studied in the context of a non- commutative extension of topological dynamics (see in particular [4, Definition 2.9 and Section 5]). Since subsystems form an integral part of relatively independent join- ings, we also study the properties of particular subsystems, namely the fixed point subsystem and the compact subsystem generated by the eigenoperators of the dynamics. These subsystems are related to the properties of W-dynamical systems, namely to ergodicity and weak mixing respectively. For subsystems more generally we will see in Sec- tion 2 and onward that the modular group of the state of the W- dynamical system in question plays an essential role in the definition and application of subsystems relevant to relatively independent join- ings. This fits in very naturally with the physically relevant case in which the dynamics itself is given by the modular group of the state, but we will see that the framework is much more widely applicable. Note that subsystems have already proven useful in the study of non- commutative dynamical systems; see for example [3, Sections 3 and 4] and [28, Definition 2.9 and Lemma 2.10] for recent work. After summarizing the basic framework in Section 2, we construct relatively independent joinings of two W-dynamical systems in Section 3 and derive a few useful facts regarding them. In Sections 4 and 5 we then study relative ergodicity and compact subsystems respectively, to illustrate how relatively independent joinings fit into the theory of W- dynamical systems. In the latter two sections the fixed point subsystem and compact subsystem generated by the eigenoperators respectively play a central role. In Section 6 we conclude the paper by studying further properties of this compact subsystem when the W-dynamical system is ergodic. 2. Basic definitions, notations and background We use the same basic definitions as in [9, 10]. For convenience we summarize them here, along with some additional definitions. Simul- taneously this fixes notations that will be used throughout the rest of the paper. Some related background material, in particular regard- ing Tomita-Takesaki theory (or modular theory) is also discussed. A general notation that we use often is B(X) to denote the space of all bounded linear operators X → X on a normed space X. The identity element of a group will be indicated by 1. In the remainder of this paper W-dynamical systems are referred to simply as "systems" and they are defined as follows: Definition 2.1. A system A = (A, µ, α) consists of a faithful normal state µ on a (necessarily σ-finite) von Neumann algebra A, and a rep- resentation α : G → Aut(A) : g 7→ αg of an arbitrary group G as JOININGS AND SUBSYSTEMS 3 ∗-automorphisms of A, such that µ ◦ αg = µ for all g. We call the system A an identity system if αg = idA for all g where idA : A → A is the identity mapping, while we call it trivial if A = C1A where 1A (often denoted simply as 1) is the unit of A. In the rest of the paper the symbols A, B, F and R will denote sys- tems (A, µ, α), (B, ν, β), (F, λ, ϕ) and (R, ψ, ρ) respectively, all making use of actions of the same group G. Definition 2.2. A joining of A and B is a state ω (i.e. a positive linear functional with ω(1) = 1) on the algebraic tensor product A ⊙ B such that ω (a ⊗ 1B) = µ(a), ω (1A ⊗ b) = ν(b) and ω ◦ (αg ⊙ βg) = ω for all a ∈ A, b ∈ B and g ∈ G. The set of all joinings of A and B is denoted by J (A, B). We call A disjoint from B when J (A, B) = {µ ⊙ ν}. We will also have occasion to use a more general concept, namely if A and B are von Neumann algebras with faithful normal states µ and ν respectively, then a coupling of (A, µ) and (B, ν) is a state ω on A ⊙ B such that ω (a ⊗ 1B) = µ(a) and ω (1A ⊗ b) = ν(b). The modular group of a faithful normal state µ on a von Neumann algebra A will be denoted by σµ, and its elements by σµ for every t t ∈ R. The cyclic representation of A obtained from µ by the GNS construction will be denoted by (Hµ, πµ, Ωµ); this notation will in fact also be used in the case of an arbitrary state on a unital ∗-algebra. The associated modular conjugation will be denoted by Jµ, and we let jµ : B(Hµ) → B(Hµ) : a 7→ Jµa∗Jµ for which we note that j −1 µ = jµ. We will also use the notation γµ : A → Hµ : a 7→ πµ(a)Ωµ even in the case of unital ∗-algebras. The dynamics α of a system A can be represented by a unitary group U on Hµ defined by extending It satisfies Ugγµ(a) := γµ(αg(a)). Ugπµ(a)U ∗ g = πµ(αg(a)) for all g ∈ G; also see [6, Corollary 2.3.17]. The unitary representation of β will be denoted by V . Definition 2.3. We call F a subsystem of A if there exists an injective unital ∗-homomorphism ζ of F onto a von Neumann subalgebra of A such that µ ◦ ζ = λ and αg ◦ ζ = ζ ◦ ϕg for all g ∈ G. If ζ(F ) is invariant under σµ, i.e. σµ t (ζ(F )) = ζ(F ) for all t ∈ R, then F is called a modular subsystem of A. If furthermore ζ : F → A is surjective, then we say that ζ is an isomorphism of dynamical systems, and the systems A and F are isomorphic. 4 ROCCO DUVENHAGE The symbol ζ will be used uniformly for this purpose. We call it the imbedding of F into A. Without loss one could assume ζ to be an inclusion, so F ⊂ A, and we will often do so. When the system F is a subsystem of B, the notation η : F → B instead of ζ will be used. If F is a (modular) subsystem of both A and B, then we call it a common (modular) subsystem of A and B. Important nontrivial examples of modular subsystems will be studied in Sections 4 and 5. In the meantime we note that in classical ergodic theory all subsystems (also known as "factors") are of course modular. More generally, if the state µ of our system A happens to be a trace (i.e. µ(ab) = µ(ba) for all a, b ∈ A), then again all subsystems are modular. Also, in the physically relevant situation where the dynamics α of A is the modular group of µ, then any subsystem of A is automatically modular. A standard fact from Tomita-Takesaki theory related to modular subsystems is the following (a proof of which is contained in [35, Section 10.2] for example): Lemma 2.4. Let µ be a faithful normal state on a von Neumann alge- bra A. Let F be a von Neumann subalgebra of A, invariant under σµ. Setting HF := γµ(F ), we have JµHF = HF . When we construct relatively independent joinings of A and B over a common modular subsystem F in the next section, one of the key tricks (also used in [12]) is to work with the "commutant" B of B. Given a system B, this means the following: We set B := πν(B)′ and then carry the state and dynamics of B over to B in a natural way using jν by defining a state ν and ∗-automorphism βg on B by ν(b) := ν ◦ π−1 ν ◦ jν(b) for all g ∈ G. From Tomita-Takesaki theory one has that VgJν = JνVg (see [9, Construction 3.4]). It follows that ν ◦ jν(b) and βg(b) := jν ◦ πν ◦ βg ◦ π−1 and ν(b) = hΩν, bΩνi βg(b) = VgbV ∗ g for all b ∈ B and g ∈ G. unitary representation of β is the same as that of β, namely V . In particular the latter tells us that the 3. Relatively independent joinings Throughout this section we consider two systems A and B which have a common modular subsystem F. We are going to construct the relatively independent joining of A and B over F. More precisely it will be a joining of A and B. Without loss we can assume ζ to be an inclusion, so F ⊂ A. JOININGS AND SUBSYSTEMS 5 Since F is a modular subsystem of B, we obtain a modular subsystem F =(cid:16) F , λ, ϕ(cid:17) of B in the following way (as is easily checked): Set F := jν ◦ πν ◦ η(F ) ⊂ B and let λ := ν F and ϕg := βg F for all g ∈ G. Since F and F are invariant under the modular groups of µ and ν respectively we know by Tomita-Takesaki theory (see for example [36, Theorem IX.4.2]) that we have unique conditional expectations and D : A → F D : B → F such that λ◦D = µ and λ◦ D = ν. (This corresponds to disintegrations in the classical case; see [15, Section 6.1].) Setting Hη := γν(η(F )), we obtain a well-defined unital ∗-homomorphism πη : F → B(Hη) : a 7→ πν(η(a))Hη and one can check that (Hη, πη, Ων) is a cyclic representation of (F, λ) and hence unitarily equivalent [6, Theorem 2.3.16] to (Hλ, πλ, Ωλ). In other words there is a unique unitary operator uη : Hλ → Hη such that uηΩλ = Ων and u∗ ηπη(a)uη = πλ(a) for all a ∈ F . From Lemma 2.4, for any a ∈ F we have aHη ⊂ Hη, and therefore we obtain a well-defined injective ∗-homomorphism κη : F → B(Hλ) : a 7→ u∗ η(aHη)uη. It is then easily shown from the definition of F that κη( F ) ⊂ πλ(F )′, and hence δ : πλ(F ) ⊙ κη( F ) → B(Hλ) defined as the linear extension of πλ(F ) × κη( F ) → B(Hλ) : (a, b) 7→ ab is a well-defined unital ∗-homomorphism. We now introduce the diagonal state of λ as the state ∆λ : F ⊙ F → C defined by ∆λ(c) := hΩλ, δ ◦ (πλ ⊙ κη) (c)Ωλi for all c ∈ F ⊙ F . This enables us to define a linear functional µ ⊙λ ν on A ⊙ B by µ ⊙λ ν := ∆λ ◦ (D ⊙ D). (If we did not assume ζ to be an inclusion mapping, one would simply replace D by ζ −1 ◦ D.) This completes the basic construction. The next step is to show that µ ⊙λ ν is a joining of A and B. In fact we will show a bit more in the next proposition. We first define a special class of joinings of A and B: 6 ROCCO DUVENHAGE Definition 3.1. Any ω ∈ J(cid:16)A, B(cid:17) with ωF ⊙ F = ∆λ is called a joining of A and B over F. The set of all such ω is denoted Jλ(cid:16)A, B(cid:17). then µ ⊙λ ν ∈ Jλ(cid:16)A, B(cid:17). Proposition 3.2. If F is a common modular subsystem of A and B, Proof. From the uniqueness of D it follows that ϕ−1 g ◦ D ◦ αg = D and therefore D ◦ αg = ϕg ◦ D. Similarly D ◦ βg = ϕg ◦ D. Denot- ing the unitary representation of ϕ on Hλ by W , note that for a ∈ F we have u∗ ηπν(βg(η(a)))uη = πλ(ϕg(a)) = Wgπλ(a)W ∗ Wg. Therefore κη ( ϕg(b)) = Wgκη(b)W ∗ ηVgHη uηπλ(a)u∗ g . Letting this act on the cyclic vector Ωλ, we obtain u∗ ηVgHη uη = u∗ g for b ∈ F . For a ∈ A and b ∈ B it follows that ηVgHη uη = µ ⊙λ ν(cid:16)αg ⊙ βg(a ⊗ b)(cid:17) = DΩλ, πλ (ϕg(D(a))) κη(cid:16) ϕg( D(b))(cid:17) ΩλE = DΩλ, Wgπλ (D(a)) W ∗ g Wgκη(cid:16) D(b)(cid:17) W ∗ g ΩλE = µ ⊙λ ν(a ⊗ b) are conditional expectations, it follows that D ⊙ D is positive (see [35, and therefore µ ⊙λ ν ◦(cid:16)αg ⊙ βg(cid:17) = µ ⊙λ ν as required. Since D and D p. 119] for example). It is now easily seen that µ ⊙λ ν ∈ J(cid:16)A, B(cid:17). Jλ(cid:16)A, B(cid:17). independent joining of A and B over F. If Jλ(cid:16)A, B(cid:17) = {µ ⊙λ ν}, From the definition of µ ⊙λ ν it then immediately follows that µ ⊙λ ν ∈ (cid:3) Definition 3.3. The joining µ ⊙λ ν of A and B is called the relatively then A and B are called disjoint over F. In the remainder of this section we study simple but useful properties of this relatively independent joining. These properties will be used in subsequent sections of the paper. Proposition 3.4. If F = C, it follows that µ ⊙λ ν = µ ⊙ ν and Jλ(cid:16)A, B(cid:17) = J(cid:16)A, B(cid:17). Proof. This follows directly from the definitions above. (cid:3) Proposition 3.5. Let F and R both be common modular subsystems of A and B, and F a subsystem of R, with inclusions F ⊂ R ⊂ A giving the corresponding imbeddings of the subsystems into A, while θ : R → B and η = θF give the imbeddings into B. Then Jψ(cid:16)A, B(cid:17) ⊂ Jλ(cid:16)A, B(cid:17). Furthermore, if µ ⊙ψ νR⊙ R = µ ⊙λ νR⊙ R, then R = F . JOININGS AND SUBSYSTEMS 7 Proof. For the first part we only need to show that ∆ψF ⊙ F = ∆λ. Given (Hψ, πψ, Ωψ), we can assume that (Hλ, πλ, Ωλ) is given by Hλ = γψ(F ), πλ(a) = πψ(a)Hλ for a ∈ F , and Ωλ = Ωψ, without changing ∆λ, because of unitary equivalence between different cyclic representa- tions of the same state. We can also define κθ : R → B(Hψ) in the same way as κη above, with the corresponding unitary operator denoted by uθ instead of uη. Note that uθγλ(a) = πθ(a)Ων = πη(a)Ων = uηγλ(a) for a ∈ F , from which follows that κθ(b)Hλ = κη(b) for b ∈ F , hence ∆ψF ⊙ F = ∆λ. Assuming µ ⊙ψ νR⊙ R = µ ⊙λ νR⊙ R, we have (in Hψ) hπψ(a)Ωψ, κθ(b)Ωψi = µ ⊙ψ ν(a∗ ⊗ b) = µ ⊙λ ν(a∗ ⊗ b) = µ ⊙λ ν(a∗ ⊗ D(b)) = µ ⊙ψ ν(a∗ ⊗ D(b)) =Dπψ(a)Ωψ, κθ( D(b))ΩψE for any a ∈ R and b ∈ R (with R of course defined in an analogous way to F ). Since Ωψ is cyclic for πψ, and therefore separating for κθ( R) ⊂ πψ(R)′, it follows that κθ(b) = κθ( D(b)). Therefore bHθ = D(b)Hθ where Hθ := γν(θ(R)), and hence bΩν = D(b)Ων. Since Ων is separating for B we conclude that b = D(b) ∈ F and therefore R = F . It follows that R = F as required. (cid:3) Finally we consider a Hilbert space characterization of relatively in- dependent joinings. We will work in the following setting: Let ω be a coupling of (A, µ) and (cid:16) B, ν(cid:17) as defined in Section 2. Then we can "imbed" (Hµ, πµ, Ωµ) and (Hν, πν, Ων) into (Hω, πω, Ωω) in a nat- ural way (see [9, Construction 2.3] for explicit details; also note that (Hν, πν, Ων) is unitarily equivalent to (Hν, ι B, Ων) but not necessarily equal). In particular Hµ and Hν are then subspaces of Hω and the cor- responding cyclic vectors are equal to Ωω. We set H ′ λ = γµ(F ) ⊂ Hµ and Hλ = γν( F ) ⊂ Hν, λ = Hλ as follows: H ′ this condition is satisfied one has H ′ We are particularly interested in the case where ωF ⊙ F = ∆λ. are both contained H⊙ := γω(cid:16)F ⊙ F(cid:17). Consider x ∈ H⊙ ⊖ H ′ If λ and Hλ λ and a sequence bn ∈ F ⊙ F such that γω(bn) → x, then for every a ∈ F it follows from ωF ⊙ F = ∆λ that 0 = hγµ(a), xi = hγλ(a), x′i where the limit x′ := limn→∞ δ ◦ (πλ ⊙ κη) (bn)Ωλ exists in Hλ by unitary equivalence between the cyclic representations (Hλ, δ ◦ (πλ ⊙ κη) , Ωλ) and(cid:0)H⊙, πω(·)H⊙, Ωω(cid:1) of(cid:16)F ⊙ F , ∆λ(cid:17). Therefore x′ = 0, and so x = 0 (Hλ, πλ, Ωλ) to be (cid:0)H ′ , Ωµ(cid:1), with a corresponding change in by the same unitary equivalence. Hence H ′ λ = H⊙. In a similar fashion Hλ = H⊙ proving the claim. By unitary equivalence one can choose λ, πµ(·)H ′ λ 8 ROCCO DUVENHAGE κη, without changing the state ∆λ . In conclusion we therefore have Hλ = Hλ when ωF ⊙ F = ∆λ. In this setting we now have the following result: Proposition 3.6. Suppose that ω is a coupling of (A, µ) and (cid:16) B, ν(cid:17) such that ωF ⊙ F = ∆λ. Then ω = µ⊙λ ν if and only if any of the follow- ing three equivalent conditions are satisfied: (Hµ ⊖ Hλ) ⊥ (Hν ⊖ Hλ), (Hµ ⊖ Hλ) ⊥ Hν, or Hµ ⊥ (Hν ⊖ Hλ). Proof. The equivalence of the three conditions is easily verified. So assume ω = µ ⊙λ ν and consider any x ∈ Hµ ⊖ Hλ and y ∈ Hν, as well as sequences an ∈ A and bn ∈ B such that γµ (an) → x and γν (bn) → y. Then hx, γν(bn)i = lim m→∞ ω (a∗ m ⊗ bn) = lim m→∞ µ ⊙λ ν (a∗ m ⊗ bn) m))(cid:16)κη ◦ D(bn)(cid:17) ΩλE m))(cid:16)κη ◦ D(cid:16) D(bn)(cid:17)(cid:17) ΩλE =DΩλ, (πλ ◦ D(a∗ =DΩλ, (πλ ◦ D(a∗ =Dx, γν(cid:16) D(bn)(cid:17)E = 0 and therefore hx, yi = 0. It follows that (Hµ ⊖ Hλ) ⊥ Hν. Conversely, assume that (Hµ ⊖ Hλ) ⊥ (Hν ⊖ Hλ) and consider any a ∈ A and b ∈ B. Let P and P respectively be the projections of Hµ and Hν on Hλ. Then for any a ∈ A and b ∈ B ω (a ⊗ b) = hγµ(a∗), γν(b)i =DP γµ(a∗), P γν(b)E =Dγµ (D(a∗)) , γν(cid:16) D(b)(cid:17)E = ω(cid:16)D(a) ⊗ D(b)(cid:17) = ∆λ(cid:16)D(a) ⊗ D(b)(cid:17) = µ ⊙λ ν (a ⊗ b) which is sufficient. (cid:3) We will also unitarily represent the dynamics of B on Hν, and we denote this representation by V . This representation is used in Sections 4 and 5 instead of V on Hν to fit into the setting of Proposition 3.6. 4. Relative ergodicity In this section and the next we study how the relatively independent joinings constructed in the previous section relate to properties of sys- tems. In particular in this section we consider relative ergodicity, which is a simple generalization of ergodicity. See for example [15, Section 6.6] for a discussion of the classical case. JOININGS AND SUBSYSTEMS 9 The relevant terminology and definitions are as follows: The fixed point algebra of a system A (or, put differently, of α) is defined as Aα := {a ∈ A : αg(a) = a for all g ∈ G} and this gives an identity subsystem Aα of A by simply restricting the state of A to Aα . We call Aα the fixed point subsystem of A. Similarly, we call H U µ := {x ∈ Hµ : Ugx = x for all g ∈ G} the fixed point space of U. Definition 4.1. Let F be a subsystem of A. We call A ergodic relative to F if Aα ⊂ ζ(F ). If F in Definition 4.1 is the trivial system, then A is simply called ergodic. We note the following important facts: Proposition 4.2. The fixed point subsystem Aα is a modular subsys- tem of A. Furthermore, H U µ = γµ(Aα). Proof. By [36, Corollary VIII.1.4] αg ◦ σµ which σµ t (Aα) = Aα follows. t = αg ◦ σµ◦αg t = σµ t ◦ αg from µ = γµ(Aα). We now show that H U It is clear that γµ(Aα) ⊂ H U µ . The converse follows from the Kov´acs-Szucs mean ergodic theorem [6, Proposition 4.3.8] (also see [21]) using a argument similar to that of [6, Theorem 4.3.20] (also see [18]): Let P be the projection of Hµ onto H U µ is cyclic for πµ(A)′, there exists a unique normal (πµ ◦ αg ◦ π−1 µ )g∈G invariant projection E : πµ(A) → πµ(Aα). Furthermore, E has the property that Ea is the unique element of πµ(A) such that (Ea) P = P aP , for every a ∈ πµ(A). Hence (Ea)Ωµ = P aΩµ from which we obtain γµ(Aα) ⊃ H U µ . (cid:3) µ , then, since Ωµ ∈ H U Now for the main result of this section: Theorem 4.3. Let the identity system F be a modular subsystem of A. If A and B are disjoint over F for B = Aα, then A is ergodic relative to F. On the other hand, if A is ergodic relative to F, then A and B are disjoint over F for every identity system B which has F as a modular subsystem. Proof. Without loss we assume that ζ is an inclusion, so F ⊂ A. Now, suppose A is not ergodic relative to F, in other words F is strictly contained in Aα. Note that F is a modular subsystem of Aα, since F and Aα are modular subsystems of A. Then apply Proposition 3.5 with B = R = Aα to obtain µ ⊙ψ ν 6= µ ⊙λ ν. But both these joinings are contained in Jλ(cid:16)A, B(cid:17) by Propositions 3.2 and 3.5, so A and B are not disjoint over F. 10 ROCCO DUVENHAGE Conversely, assume that A is ergodic relative to F, and let B be any identity system which has F as a modular subsystem. We are going to apply Proposition 3.6. So consider any ω ∈ Jλ(A, B). From this joining we obtain a conditional expectation operator Pω : Hν → Hµ (i.e. hx, Pωyi = hx, yi for all x ∈ Hµ and y ∈ Hν) such that UgPω = Pω Vg (see [9, Proposition 2.4]). Then for any b ∈ B we have UgPωγν(b) = Pωγν(b) , since B is an identity system. Since F = Aα, it follows from Proposition 4.2 that Pωγν(b) ∈ γµ(Aα) = Hλ. Noting that Pω is the projection of Hω onto Hµ restricted to Hν, we conclude that Hν ⊥ (Hµ ⊖ Hλ), and hence ω = µ ⊙λ ν by Proposition 3.6. (cid:3) This result contains [10, Theorem 2.1] as a special case by Proposi- tion 3.4, namely A is ergodic if and only if it is disjoint (over the trivial system) from all identity systems. Also note that in Theorem 4.3, A being ergodic relative to F means exactly that Aα = ζ(F ), i.e. F is isomorphic to Aα, since F is assumed to be an identity system. 5. Compact subsystems In this section we study an analogue of the previous section in terms of compact systems instead of identity systems. We therefore proceed by first building up some background regarding compact subsystems, which includes extending certain results from [26, Section 4] to more general group actions in a way (see Lemma 5.3 and Theorem 5.4 below) that is relevant for our intended applications. Our proofs of these extensions follow the basic plans of those in [26], but many of the details differ. (To avoid any confusion, we note that our terminology regarding topologies on B(H) differs slightly from that of [26] and furthermore we do not always use the same topologies that [26] used in their results; we use the terminology of [6, Section 2.4.1], in particular the weak topology on B(H) is generated by hx, (·)yi for x, y ∈ H.) The additional definitions that we use in this section are the following: Definition 5.1. A system A is compact if the orbit UGx is totally bounded (i.e. UGx is compact) in Hµ for every x ∈ Hµ. An eigenvector of U is an x ∈ Hµ\{0} such that there is a function, called its eigen- value, χ : G → C such that Ugx = χ(g)x for all g ∈ G. The set of all eigenvalues is denoted by σA. Let H0 denote the Hilbert subspace of Hµ spanned by the eigenvectors of U. We will call u ∈ A\{0} an eigenoperator of α if there is a function χ : G → C, its eigenvalue, such that αg(u) = χ(g)u for all g ∈ G. As already mentioned, our immediate goal is to study compact sub- systems of a given system. We start with an analogue of the first part of Proposition 4.2: Proposition 5.2. For a system A, denote by AK the von Neumann subalgebra of A generated by the eigenoperators of α. Then αg(AK) = JOININGS AND SUBSYSTEMS 11 AK, which allows us to define a subsystem AK = (cid:0)AK, µK, αK(cid:1) of A (in terms of an inclusion as the imbedding into A) by setting µK := µAK and αK g (a) := αg(a) for all a ∈ AK. The system AK is a compact modular subsystem of A. Proof. Let S be the ∗-algebra generated by the eigenoperators of α. Then clearly αg(S) = S ⊂ AK. But AK = S ′′ , so S is σ-weakly dense in AK, while αg is σ-weakly continuous for each g. It follows that αg(AK) = AK. Next we prove that the subsystem AK so obtained, is compact. Note that for any eigenoperator u of α, γµ(u) is an eigenvector of U. Fur- thermore, it is easily seen that if u and v are eigenoperators of α, then u∗ is also eigenoperator of α, while uv is either zero or an eigenop- erator of α. It follows that γµ(S) ⊂ H0, hence γµ(AK) ⊂ H0. It is easy to show from the definition of H0 that each of its elements has a totally bounded (i.e. relatively compact) orbit under U. Also note µ := γµ(AK), which is contained in H0, with the unitary representation of αK given by the restriction of Ug to H K that (cid:0)AK, µ(cid:1) can be cyclically represented on H K µ . Therefore AK is indeed compact. Lastly note that as in Proposition 4.2, for any eigenoperator u of α t (θg(u)) = χ(g)σµ t (u) so t (AK) = AK, i.e. AK is (cid:3) with eigenvalue χ, we have that θg(σµ σµ t (u) ∈ S. It follows that σµ a modular subsystem of A. t (u)) = σµ t (S) = S, hence σµ In order to go further, we first prove a technical result which is a version of [26, Lemma 4.1] appropriate for our needs: Lemma 5.3. Let H be a Hilbert space and T a weakly closed vector subspace of B(H). Assume that there is a unit vector Ω ∈ H such that T ′Ω = H. Let G be an arbitrary group and Wg : H → H a linear isometry for every g ∈ G such that WgWh = Wgh for all g, h ∈ G, and with Ka := {WgaΩ : g ∈ G} compact in H for every a ∈ T . Endow B(T ) with the topology of pointwise strong convergence generated by the seminorms Θ 7→ kΘ(a)xk where a ∈ T and x ∈ H. Let Γ be the set of all linear contractions Θ ∈ B(T ) such that Θ(a)Ω ∈ Ka for all a ∈ T , and view it as a topological subspace of B(T ). Then Γ is a compact group with composition as its operation and the identity map on T as its identity. Proof. Let us first consider the topology on Γ. For a ∈ T, a′ ∈ T ′, x ∈ H and Θ1, Θ2 ∈ B(T ) we have kΘ1(a)x − Θ2(a)xk ≤ 2 kak kx − a′Ωk + ka′k kΘ1(a)Ω − Θ2(a)Ωk from which (together with T ′Ω = H) it follows that the topology on B(T ) is in fact generated by the seminorms Θ 7→ kΘ(a)Ωk where a ∈ T , 12 ROCCO DUVENHAGE hence we have a simplified description of the topology on Γ. Now consider the compact space K := Ya∈T Ka and the function f : Γ → K : Θ 7→ (Θ(a)Ω)a∈T which is easily shown to be injective because of T ′Ω = H. From the simplified description of the topology on Γ described above and the definition of the product topology on K it is clear that f : Γ → f (Γ) is a homeomorphism. We now show that f (Γ) is closed: Let x = (x(a))a∈T be in the closure of f (Γ). Let (Θι)ι be a net in Γ such that (f (Θι))ι converges to x, then from the inequality above (Θι(a)y)ι is a Cauchy net in H for every y ∈ H and a ∈ T . This allows us to define Θ(a)y := limι Θι(a)y and since each Θι is a contraction, Θ(a) ∈ B(H) and kΘ(a)k ≤ kak . By this definition (Θι(a))ι converges strongly and therefore also weakly to Θ(a), but T is weakly closed therefore Θ(a) ∈ T . Since Θι(a)Ω ∈ Ka and Ka is closed, it again follows from the definition of Θ(a) that Θ(a)Ω ∈ Ka for all a ∈ T . This shows that Θ ∈ Γ. But by our choice of (Θι)ι we have x(a) = limι Θι(a)Ω = Θ(a)Ω so x = f (Θ) ∈ f (Γ) and therefore f (Γ) is indeed closed. It follows that f (Γ) is compact, since K is, and therefore Γ is com- pact. Also note that Γ is Hausdorff. Now we start to prove the group structure on Γ. Note that W1 = 1, since WgW1 = Wg while Wg is injective. It follows that idT ∈ Γ. For any Θ1, Θ2 ∈ Γ we have Θ1 ◦ Θ2(a)Ω ∈ KΘ2(a) ⊂ {WgWhaΩ : g, h ∈ G} = Ka and therefore Θ1 ◦ Θ2 ∈ Γ. Next we show the continuity of this product on Γ. For Θ ∈ Γ we have Θ(a)Ω ∈ Ka ⊂ {x ∈ H : kxk = kaΩk}, since each Wg is an isometry. Therefore we have a well-defined linear isometry WΘ : T Ω → T Ω such that WΘ(aΩ) = Θ(a)Ω, so in effect WΘ represents Θ on T Θ. Using WΘ one can show that k(Θ1 ◦ Θ′ ≤ kΘ′ 1)(a)Ω − (Θ2 ◦ Θ′ 2)(a)Ωk 2(a)Ωk + kΘ1 (Θ′ 1(a)Ω − Θ′ 2(a)) Ω − Θ2 (Θ′ 2(a)) Ωk for all Θ1, Θ2, Θ′ product on Γ is indeed continuous. 1, Θ′ 2 ∈ Γ and a ∈ T . From this it follows that the We now complete the proof that Γ is a group by considering in- verses. We have already shown that Γ is a semigroup, so for Θ ∈ Γ JOININGS AND SUBSYSTEMS 13 it follows that Θn ∈ Γ for every n ∈ N = {1, 2, 3, ...}. Further- more W n Θ(aΩ) = Θn(a)Ω ∈ Ka, so since Ka is compact, the orbit (W n Θ(aΩ))n∈N is relatively compact (and therefore totally bounded) for every a ∈ T . Since WΘ is an isometry, it follows from [26, Corollary 9.10] (and remarks made just after it) that for every ε > 0 and every finite set E ⊂ T there exists an n(E, ε) ∈ N such that (cid:13)(cid:13)Θn(E,ε)(a)Ω − aΩ(cid:13)(cid:13) < ε for all a ∈ E. Defining (E, ε) ≤ (E ′, ε′) to mean E ⊂ E ′ and ε ≥ ε′, by using the description of the topology given at the beginning of this we obtain a net (cid:0)Θn(E,ε)(cid:1)(E,ε) in Γ which is seen to converge to idT proof. Since Γ is compact, the net (cid:0)Θn(E,ε)−1(cid:1)(E,ε) has a limit point in Γ, say Θ′, and it is then not too difficult to show that Θ ◦ Θ′ = idT = Θ′ ◦ Θ. In other words every element of Γ has an inverse in Γ, and therefore Γ is a group. By the single theorem appearing in [11], it follows that Γ is a topo- (cid:3) logical group, which completes the proof. Now we can state and prove the basic result regarding compact sub- systems, which is the version of [26, Theorem 4.2] that we will use. In this result and the rest of the section our group G is assumed to be abelian, although this assumption is unnecessary in the first part of the proof of the next theorem, as will be indicated. Theorem 5.4. Consider a system A with G assumed to be abelian. Set T := {a ∈ πµ(A) : aΩµ ∈ H0} ξ(a) := hΩµ, aΩµi τg(a) := UgaU ∗ g for all a ∈ T and g ∈ G. Then T := (T, ξ, τ ) is a subsystem of A which is isomorphic to AK with the isomorphism given by π−1 µ T : T → AK. It follows that AK is the largest compact subsystem of A in the sense that AK contains the image of the algebra of any compact subsystem F of A (under the imbedding of F in A). Furthermore, T Ωµ = H0, so T can be cyclically represented on H0 and σT = σA. Proof. The notation M = πµ(A), θg : M → M : a 7→ UgaU ∗ g , H = Hµ and Ω = Ωµ will be used in what follows. We divide the proof into a number of parts. (i) First we show (see also [26, Proposition 3.2]) that for any χ ∈ σA we have where MχΩ = Hχ Mχ := {a ∈ M : θg(a) = χ(g)a for all g ∈ G} 14 and ROCCO DUVENHAGE Hχ := {x ∈ H : Ugx = χ(g)x for all g ∈ G} . To do this, let Pχ be the projection of H on Hχ. Note that χ(g) = 1 and χ(g)χ(h) = χ(gh). It follows that g 7→ χ(g)Ug is also a unitary group on H, and therefore by the Alaoglu-Birkhoff mean ergodic the- orem [6, Proposition 4.3.4] (also see [1]) we know that Pχ is in the strong closure of the convex hull of nχ(g)Ug : g ∈ Go. Thus there is a net (mι)ι given by mι = nιXj=1 wιjχ(gιj)Ugιj with wι1 + ... + wιnι = 1 and wιj ≥ 0, which converges strongly to Pχ. Now for any a ∈ M the net (cι)ι given by cι(a) = nιXj=1 wιjχ(gιj)θgιj (a) of B(H) is weakly compact [6, Proposition 2.4.2]. Since M is a von Neumann algebra, it is weakly closed, so the limit of this subnet, say has a weakly convergent subnet, say (cid:0)cι(κ)(a)(cid:1)κ, since the unit ball a0, is in M. Since (cid:0)mι(κ)(cid:1)κ converges strongly to Pχ and UgΩ = Ω, we have that for any x ∈ H, so PχaΩ = a0Ω. It follows that θg (a0) Ω = χ(g)a0Ω, but Ω is separating for M, hence a0 ∈ Mχ. We remark that one can in fact go further: Taking b ∈ M ′ we have cι(κ)(a)bΩ = bmι(κ)aΩ → bPχaΩ = a0bΩ since cι(κ)(a) ∈ M. Since M ′Ω is dense in H, it is now straightforward to show that bΩ can be replaced by any x ∈ H, in other words(cid:0)cι(κ)(a)(cid:1) actually converges strongly to a0. We will only need the convergence cι(κ)(a)Ω → a0Ω, though. Now consider any x ∈ Hχ and ε > 0, and let a ∈ M be such that kaΩ − xk < ε. Then (cid:13)(cid:13)(cid:13)χ(g)θg(a)Ω − x(cid:13)(cid:13)(cid:13) = kUgaΩ − χ(g)xk < ε so (cid:13)(cid:13)cι(κ)(a)Ω − x(cid:13)(cid:13) < ε. But by the convergence above, there is a κ such that (cid:13)(cid:13)a0Ω − cι(κ)(a)Ω(cid:13)(cid:13) < ε, so ka0Ω − xk < 2ε. This proves that MχΩ is dense in Hχ as required, since clearly MχΩ ⊂ Hχ. (ii) Now we prove a number of properties of T . First, since UgH0 ⊂ g Ω = Ω, we clearly have τg(T ) = H0 from the definition of H0 while U ∗ θg(T ) = T . It follows that τg ∈ B(T ). hx, PχaΩi = lim κ (cid:10)x, mι(κ)aΩ(cid:11) = lim κ (cid:10)x, cι(κ)(a)Ω(cid:11) = hx, a0Ωi JOININGS AND SUBSYSTEMS 15 Next, set S := span [χ∈σA Mχ where "span" means finite linear combinations. From (i) it follows that SΩ = H0. It is also readily verified that S is a ∗ -algebra contained in T ; in fact, S is the ∗-algebra generated by the eigenoperators of θ. In particular T Ω = H0. Lastly we show that T is weakly closed, so consider any a in the weak closure of T , and a net (aι) in T converging weakly to a. Since M is weakly closed, a ∈ M. Furthermore, from the definition of T , for every x ∈ H ⊥ 0 we have 0 = hx, aιΩi → hx, aΩi, so aΩ ∈ H0. Therefore, again from the definition of T , we have a ∈ T as required. (iii) Everything so far in the proof holds for non-abelian G as well, but in the rest of the proof we do make use of the fact that G is abelian, since we are going to work with its Bohr compactification. In particular we now show that T is a von Neumann algebra using the group Γ given by Lemma 5.3 applied to T with W = U. Note that all the requirements in Lemma 5.3 are satisfied: M ′ ⊂ T ′ so T ′Ω is dense in H, and for every a ∈ T the closure Ka of the corresponding orbit in H is indeed compact, since each element of H0 has a totally bounded (i.e. relatively compact) orbit under U (in fact H0 contains all such elements [5, Lemma 6.6], but this fact will only be used in (iv) below). Assign the discrete topology to G. Since G is abelian, we can imbed it into its Bohr compactification ¯G. Let us denote this canonical imbedding by ι : G → ¯G. It is an injective group homomorphism with ι(G) dense in ¯G. The dual map of this is a group isomorphism ι : b¯G → bG. For χ ∈ bG, set h·, χi := ι−1(χ), then it follows that hι(g), χi = χ(g). Note that since τg ∈ B(T ) is an isometry, it follows that τ : G → Γ : g 7→ τg is a well-defined group homomorphism (and it is continuous, since G is discrete) which by the universal property of the Bohr compactification [13, Proposition (4.78)] can be extended to a continuous group homomorphism ¯τ : ¯G → Γ : g 7→ ¯τg which means that ¯τι(g) = τg for all g ∈ G. Since Γ → H : Θ 7→ Θ(a)y is continuous, so is ¯G → H : g 7→ ¯τg(a)y for all a ∈ T and y ∈ H. Denoting the normalized Haar measure on the compact group ¯G by m, we can for every a ∈ T and χ ∈ bG define a unique aχ ∈ B(H) by requiring hx, aχyi =Z ¯G hg, χi hx, ¯τg(a)yi dm(g) for all x, y ∈ H. If f is a weakly continuous linear functional on B(H) it follows (see for example [8, Theorem V.3.9]) that it is a finite linear 16 ROCCO DUVENHAGE combination of such hx, (·)yi forms, hence (1) f (aχ) =Z ¯G hg, χif (¯τg(a)) dm(g) so if f is zero on T we have f (aχ) = 0 and therefore by [8, Corollary V.3.12] aχ ∈ T since T is weakly closed. For g ∈ G it follows from the definition of aχ that g x, ¯τh(a)U ∗ g y(cid:11) dm(h) g y(cid:11) =Z ¯G hh, χi(cid:10)U ∗ hh, χi(cid:10)x, ¯τι(g)h(a)y(cid:11) dm(h) hι(g)−1h, χi hx, ¯τh(a)yi dm(h) g x, aχU ∗ hx, τg(aχ)yi = (cid:10)U ∗ = Z ¯G = Z ¯G = hι(g)−1, χiZ ¯G hh, χi hx, ¯τh(a)yi dm(h) = χ(g) hx, aχyi and therefore τg(aχ) = χ(g)aχ which means that aχ ∈ S, since aχ ∈ T ⊂ M. We now use this result to show that S is weakly dense in T , so take any a ∈ T . Let f be any weakly continuous linear functional on B(H) which is zero on the weak closure a bG of the span ofnaχ : χ ∈ bGo. From (1) and the Plancherel theorem it follows that g 7→ f (¯τg(a)) is zero in L2( ¯G). But this function is continuous by arguments as above, so f (¯τg(a)) = 0 for all g ∈ ¯G, for if not, then f would be non-zero on an open neighbourhood of some g ∈ ¯G, contradicting the fact that the Haar measure of a non-empty open set is non-zero. In particular for g = 1 we find f (a) = 0. This means that a ∈ a bG and therefore S is weakly dense in T . Since S is a ∗-algebra containing the identity operator, we conclude that T is a von Neumann algebra by the von Neumann density theorem [6, Corollary 2.4.15]. (iv) By (ii) and (iii) we know that T is indeed a subsystem of A. It has σT = σA, since it can be cyclically represented on H0 (because of T Ωµ = H0) with the dynamics given by UgH0. Since S above is the ∗-algebra generated by the eigenoperators of θ, we have S ⊂ πµ(AK) ⊂ T . Hence T = πµ(AK), since πµ(AK) is a von Neumann algebra and therefore weakly closed. From this the isomorphism between T and AK follows. If F is any compact subsystem of A, and a ∈ F , then by definition of compactness of a system the orbit of γµ(ζ(a)) under U is totally bounded (keep in mind that (F, λ) can be cyclically represented on γµ(ζ(F )), with the corresponding unitary representation of ϕ given by the restriction of Ug to this space), thus πµ(ζ(a))Ωµ ∈ H0 according to [5, Lemma 6.6]. In other words πµ(ζ(a)) ∈ T , and therefore ζ(a) ∈ AK which means that AK is the largest compact subsystem of A. (cid:3) JOININGS AND SUBSYSTEMS 17 Note that the sole reason we needed to assume that G is abelian in this theorem, is that we use its Bohr compactification in the proof. Whereas Aα is by definition the largest identity subsystem of A, we have seen above that AK is the largest compact subsystem of A, at least when G is abelian. Clearly Aα ⊂ AK. We mention that a very simple version of this in the context of noncommutative topological dynamics was also discussed in [23, Section 2] and [4, Definition 1.2]. In analogy to Proposition 4.2, Theorem 5.4 says that H0 = γµ(AK). Before we return to relatively independent joinings we note the fol- lowing generalization of [5, Theorem 6.8] (also see [26, Proposition 5.4]). Note that in terms of Definition 5.1's notation, we call A weakly mixing if dim H0 = 1. Corollary 5.5. A system A with G assumed abelian, is weakly mixing if and only if AK is the trivial system. Proof. If AK is trivial, then so is T in Theorem 5.4, so H0 = T Ωµ = CΩµ. Conversely, if A is weakly mixing, then T Ωµ = CΩµ, but Ωµ is separating for πµ(A) and therefore for T , so T = C, hence AK is trivial. (cid:3) In addition to the assumptions in this corollary, in [5, Theorem 6.8] it was also assumed that A is ergodic. Now we finally return to relatively independent joinings, namely an analogue of Theorem 4.3. The proof is very similar to that of Theorem 4.3. Theorem 5.6. Assume that G is abelian, and let F be a compact mod- ular subsystem of A. If A and B are disjoint over F for B = AK, then F is isomorphic to AK. On the other hand, if F is isomorphic to AK, then A and B are disjoint over F for every compact system B which has F as a modular subsystem. Proof. Assume without loss that F is imbedded in A by inclusion. Suppose F is not isomorphic to AK, then F is strictly contained in AK by Theorem 5.4. Note that F is a modular subsystem of AK, since F and AK are modular subsystems of A. Now apply Proposition 3.5 with B = R = AK to obtain µ ⊙ψ ν 6= µ ⊙λ ν. But both these joinings are not disjoint over F. are contained in Jλ(cid:16)A, B(cid:17) by Propositions 3.2 and 3.5, so A and B AK. Consider any ω ∈ Jλ(cid:16)A, B(cid:17). For any eigenvector y ∈ Hν of V Conversely, assume that F is isomorphic (and therefore equal) to with eigenvalue χ ∈ σ B we have UgPωy = Pω Vgy = χ(g)Pωy with Pω as in the proof of Theorem 4.3. So Pωy = 0 or χ ∈ σA. Therefore Pωy ∈ H0 = Hλ with H0 as in Definition 5.1 and Hλ as given 18 ROCCO DUVENHAGE in the setting described before Proposition 3.6. Since B has discrete spectrum (i.e. Hν is spanned by the eigenvectors of V ) as remarked in [10, Proposition 2.6] (also see [22, Section 2.4] for the general theory), it follows that Pωy ∈ Hλ for all y ∈ Hν. Since Pω is the projection of Hω onto Hµ restricted to Hν, it follows that Hν ⊥ (Hµ ⊖ Hλ), and hence ω = µ ⊙λ ν by Proposition 3.6. (cid:3) In the case of trivial F this theorem gives a variation on [10, Theo- rems 2.7 and 2.8] in that the latter assumed ergodicity of the systems involved, but did not require G to be abelian, namely: Corollary 5.7. Assuming G is abelian, A is weakly mixing if and only if it is disjoint from all compact systems. Proof. Use trivial F in Theorem 5.6, together with Proposition 3.4 and Corollary 5.5. (cid:3) As in Section 4, these results illustrate how relatively independent joinings relate to the structure of W-dynamical systems, in particular to certain types of subsystems. 6. Ergodicity and compactness In this concluding section we turn away from joinings and focus on subsystems. The goal is to study AK, which was defined in Proposition 5.2, when A is ergodic, i.e. when Aα is trivial. This is related to, and largely motivated by [19, Theorem 1.3] (see Corollary 6.2 below). Remember that a state µ on a von Neumann algebra A is called tracial if µ(ab) = µ(ba) for all a, b ∈ A. The following result strength- ens the fact that AK is a modular subsystem of A when the latter is ergodic, and note that in this result we need not assume G is abelian: Theorem 6.1. Let A be ergodic. Then AK is contained in the fixed point algebra of the modular group σµ of µ. Furthermore, µK is then tracial, and AK therefore a finite von Neumann algebra. Proof. Let u be any eigenoperator of α, with eigenvalue χ. Then as in the proof of Proposition 4.2 αg(σµ t (αg(u)) = χ(g)σµ t (u)) = σµ Since A is ergodic and u and σµ t (u) are eigenoperators with the same norm and the same eigenvalue, it follows from [34, Lemma 2.1(3)] that there is a number, say Λ(t) ∈ C, with Λ(t) = 1, such that t (u). σµ t (u) = Λ(t)u for all t ∈ R. Now we show that σµ t (u) = u (also see the proof of [10, Lemma 5.2]). It follows from the group property of σµ that Λ(s + t) = Λ(s)Λ(t) for all s, t ∈ R, and since t 7→ hx, πµ(σµ t (u))yi is continuous for all x, y ∈ Hµ, it follows that t 7→ Λ(t) is continuous. Therefore Λ(t) = eiθt JOININGS AND SUBSYSTEMS 19 for all t ∈ R for some θ ∈ R; see for example [29, p. 12]. Denoting the modular operator associated with (πµ(A), Ωµ) by ∆, it follows that ∆itπµ(u)Ωµ = πµ(σµ t (u))Ωµ = eiθtπµ(u)Ωµ, hence by the definition of Jµ∆1/2 (see for example [6, Section 2.5.2]) Jµπµ(u)∗Ωµ = Jµ(cid:0)Jµ∆1/2(cid:1) πµ(u)Ωµ = ∆1/2πµ(u)Ωµ = eθ/2πµ(u)Ωµ and by taking the norm both sides we conclude that eθ/2 = 1, since αg(uu∗) = χ(g)2 uu∗ = uu∗ and αg(u∗u) = u∗u, implying uu∗ = u∗u ∈ C1\{0} by ergodicity and the fact that uu∗ and u∗u are both positive. Therefore θ = 0. This proves that σµ t (u) = u for all t ∈ R. Then by the definition of AK in Proposition 5.2, it is indeed contained in the fixed point algebra of the modular group σµ, since the latter algebra is a von Neumann algebra containing all eigenoperators of α, as just shown. Since AK is a modular subsystem of A, the modular group σµK of µK is simply given by the restriction of σµ t = idAK for all t by what we have shown above. It follows that the modular operator t to AK. Hence σµK operator. Therefore by the basic definitions of Tomita-Takesaki theory, associated with the cyclic representation of (cid:0)AK, µK(cid:1) is the identity and writing the cyclic representation of (cid:0)AK, µK(cid:1) as (H, π, Ω) and the corresponding modular conjugation as J, we have JaΩ = a∗Ω for all a ∈ π(AK). From this it follows that µK is tracial, namely for any a, b ∈ AK we have µ(ab) = hΩ, π(ab)Ωi = hπ(a)∗Ω, π(b)Ωi = hJπ(a)Ω, π(b)Ωi = hJπ(b)Ω, π(a)Ωi = hΩ, π(ba)Ωi = µ(ba) and the existence of a faithful (numerical) trace on AK implies that it is a finite von Neumann algebra (see for example [20, p. 505]). (cid:3) To clarify the connection with [19, Theorem 1.3], we deduce it as a corollary from this theorem (but we only state it for an inverse tem- perature of −1): Corollary 6.2. Suppose that the dynamics α of A is given by the modular group σµ, i.e. G = R and αt = σµ for all t ∈ R. If A is t ergodic, it then follows that it is weakly mixing. Proof. By Theorem 6.1 AK ⊂ Aα, but Aα = C, since A is ergodic, therefore AK = C. Since R is an abelian group, it follows from Corol- lary 5.5 that A is weakly mixing. (cid:3) One may wonder whether more generally, when Aα is not necessarily trivial, one has AK = Aα if α is the modular group of µ. However on 20 ROCCO DUVENHAGE A = B(H) with H a finite dimensional Hilbert space, one has the sim- ple counter example given by the Gibbs state (at inverse temperature −1) µ(a) = Tr(eha)/Tr(eh) where h ∈ A is hermitian, in which case αt(a) = eihtae−iht is the modular group of µ; the system A so obtained is compact but not in general an identity system, so AK = A 6= Aα. Another corollary of Theorem 6.1 is the following (also see [27, Proposition 7.2]): Corollary 6.3. If a system A is ergodic and compact, and G is abelian, then µ is necessarily tracial and A a finite von Neumann algebra. Proof. By Theorem 5.4 AK = A, and the result then follows from Theorem 6.1. (cid:3) Results of the nature of Corollary 6.3 have of course been well stud- ied. However, our assumption that the system is compact makes Corol- lary 6.3 a relatively easy result. In the literature much more difficult results have been obtained; see [17, 32, 33, 34, 24, 2, 27, 16] for the development. The results of this section illustrate how subsystems can be applied to derive interesting properties of systems, and show how aspects of the work above fit into previous literature on noncommutative dynamical systems. Acknowledgments I thank Anton Stroh for useful discussions, and Richard de Beer for pointing the paper [19] out to me long ago. This work was supported by the National Research Foundation of South Africa. References [1] L. Alaoglu and G. Birkhoff, General ergodic theorems, Ann. of Math. (2) 41, (1940). 293 -- 309. MR0002026 (1,339a) [2] S. Albeverio and R. Høegh-Krohn, Ergodic actions by compact groups on C- algebras, Math. Z. 174 (1980), 1 -- 17. MR0591609 (82h:46087) [3] T. Austin, T. Eisner and T. Tao, Nonconventional ergodic averages and multiple recurrence for von Neumann dynamical systems, arXiv:0912.5093v4 [math.OA]. [4] D. Avitzour, Noncommutative topological dynamics. II, Trans. Amer. Math. Soc. 282 (1984), 121 -- 135. MR 0728705 (85i:46089) [5] C. Beyers, R. Duvenhage and A. Stroh, The Szemer´edi property in ergodic W-dynamical systems, J. Operator Theory, to appear, arXiv:0709.1557v1 [math.OA]. [6] O. Bratteli and D. W. Robinson, Operator algebras and quantum statistical mechanics 1, second edition, Springer-Verlag, New York, 1987. MR 0887100 (88d:46105) [7] O. Bratteli and D. W. Robinson, Operator algebras and quantum statisti- cal mechanics 2, second edition, Springer-Verlag, Berlin, 1997. MR1441540 (98e:82004) JOININGS AND SUBSYSTEMS 21 [8] N. Dunford and J. T. Schwartz, Linear operators. Part I. General theory, With the assistance of William G. Bade and Robert G. Bartle. Reprint of the 1958 original. Wiley Classics Library. A Wiley-Interscience Publication. John Wiley & Sons, Inc., New York, 1988. MR1009162 (90g:47001a) [9] R. Duvenhage, Joinings of W-dynamical systems, J. Math. Anal. Appl. 343 (2008), 175 -- 181. arXiv:0802.0827v1 [math.OA]. MR 2409467 (2009b:46135) [10] R. Duvenhage, Ergodicity and mixing of W-dynamical systems in terms of joinings, arXiv:0803.2147 [math.OA]. [11] R. Ellis, A note on the continuity of the inverse, Proc. Amer. Math. Soc. 8 (1957), 372 -- 373. MR0083681 (18,745d) [12] F. Fidaleo, An ergodic theorem for quantum diagonal measures, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 12 (2009), 307 -- 320. MR 2541399 [13] G. B. Folland, A course in abstract harmonic analysis, Studies in Advanced Mathematics. CRC Press, Boca Raton, FL, 1995. [14] H. Furstenberg, Disjointness in ergodic theory, minimal sets, and a problem in Diophantine approximation, Math. Systems Theory 1 (1967), 1 -- 49. MR 0213508 (35 #4369) [15] E. Glasner, Ergodic theory via joinings, Mathematical Surveys and Mono- graphs 101, American Mathematical Society, Providence, RI, 2003. MR 1958753 (2004c:37011) [16] R. Høegh-Krohn, M. B. Landstad and E. Størmer, Compact ergodic groups of automorphisms, Ann. of Math. (2) 114 (1981), 75 -- 86. MR0625345 (82i:46097) [17] N. M. Hugenholtz, On the factor type of equilibrium states in quantum sta- tistical mechanics, Comm. Math. Phys. 6 (1967), 189 -- 193. MR0225177 (37 #772) [18] A. Z. Jadczyk, On some groups of automorphisms of von Neumann algebras with cyclic and separating vector, Comm. Math. Phys. 13 (1969) 142 -- 153. MR0264936 (41 #9525) [19] V. Jaksi´c and C.-A. Pillet, A note on eigenvalues of Liouvilleans, J. Statist. Phys. 105 (2001), 937 -- 941. MR1869571 (2002j:46085) [20] R. V. Kadison and J. R. Ringrose, John R. Fundamentals of the theory of operator algebras. Vol. II. Advanced theory, Corrected reprint of the 1986 orig- inal. Graduate Studies in Mathematics, 16. American Mathematical Society, Providence, RI, 1997. [21] I. Kov´acs and J. Szucs, Ergodic type theorems in von Neumann algebras, Acta Sci. Math. (Szeged) 27 (1966) 233 -- 246. MR0209857 (35 #753) [22] U. Krengel, Ergodic theorems, Walter de Gruyter & Co., Berlin, 1985. MR 0797411 (87i:28001) [23] D. Laison and G. Laison, Topological dynamics on C-algebras, Trans. Amer. Math. Soc. 204 (1975), 197 -- 205. MR0358365 (50 #10831) [24] R. Longo, Notes on algebraic invariants for noncommutative dynamical sys- tems, Comm. Math. Phys. 69 (1979), 195 -- 207. MR0550019 (80j:46108) [25] S. Neshveyev and E. Størmer, Dynamical entropy in operator algebras, Ergeb- nisse der Mathematik und ihrer Grenzgebiete, 3. Folge, A Series of Modern Surveys in Mathematics, Vol. 50, Springer-Verlag, Berlin, 2006. MR2251116 (2007m:46108) [26] C. P. Niculescu, A. Stroh and L. Zsid´o, Noncommutative extensions of classical and multiple recurrence theorems, J. Operator Theory 50 (2003), 3 -- 52. MR 2015017 (2004k:46123) [27] D. Olesen, G. K. Pedersen and M. Takesaki, Ergodic actions of compact abelian groups, J. Operator Theory 3 (1980), 237 -- 269. 22 ROCCO DUVENHAGE [28] S. Popa, Cocycle and orbit equivalence superrigidity for malleable actions of w-rigid groups, Invent. Math. 170 (2007), no. 2, 243 -- 295. [29] W. Rudin, Fourier analysis on groups, Interscience Tracts in Pure and Applied Mathematics, No. 12 Interscience Publishers (a division of John Wiley and Sons), New York-London, 1962. MR0152834 (27 #2808) [30] D. J. Rudolph, An example of a measure preserving map with minimal self- joinings, and applications, J. Analyse Math. 35 (1979), 97 -- 122. MR 0555301 (81e:28011) [31] J.-L. Sauvageot and J.-P. Thouvenot, Une nouvelle d´e finition de l'entropie dynamique des syst`emes non commutatifs, Comm. Math. Phys. 145 (1992), 411 -- 423. MR 1162806 (93d:46113) [32] E. Størmer, Types of von Neumann algebras associated with extremal invariant states, Comm. Math. Phys. 6 (1967), 192 -- 204. MR0225178 (37 #773) [33] E. Størmer, Automorphisms and invariant states of operator algebras, Acta Math. 127 (1971), 1 -- 9. MR0383097 (52 #3978) [34] E. Størmer, Spectra of ergodic transformations, J. Funct. Anal. 15 (1974), 202 -- 215. MR 0377544 (51 #13715) [35] S¸. Stratila, Modular theory in operator algebras, Editura Academiei Repub- licii Socialiste Romania, Bucharest, Abacus Press, Tunbridge Wells, 1981. MR 0696172 (85g:46072) [36] M. Takesaki, Theory of operator algebras. II, Encyclopaedia of Mathematical Sciences, 125. Operator Algebras and Non-commutative Geometry, 6. Springer- Verlag, Berlin, 2003. MR1943006 (2004g:46079) Department of Physics, University of Pretoria, Pretoria 0002, South Africa E-mail address: [email protected]
1312.0269	4	1312	2016-01-15T22:28:48	Double-ended queues and joint moments of left-right canonical operators on full Fock space	[ "math.OA" ]	We follow the guiding line offered by canonical operators on the full Fock space, in order to identify what kind of cumulant functionals should be considered for the concept of bi-free independence introduced in the recent work of Voiculescu. By following this guiding line we arrive to consider, for a general noncommutative probability space (A, phi), a family of "(l,r)-cumulant functionals" which enlarges the family of free cumulant functionals of the space. In the motivating case of canonical operators on the full Fock space we find a simple formula for a relevant family of (l,r)-cumulants of a (2d)-tuple (A_1, ..., A_d, B_1, ..., B_d), with A_1, ... , A_d canonical operators on the left and B_1, ... , B_d canonical operators on the right. This extends a known one-sided formula for free cumulants of A_1, ..., A_d, which establishes a basic operator model for the R-transform of free probability.	math.OA	math	Double-ended queues and joint moments of left-right canonical operators on full Fock space Mitja Mastnak 1 Alexandru Nica 1 Abstract We follow the guiding line offered by canonical operators on the full Fock space, in order to identify what kind of cumulant functionals should be considered for the con- cept of bi-free independence introduced in the recent work of Voiculescu. By following this guiding line we arrive to consider, for a general noncommutative probability space (A, ϕ), a family of "(ℓ, r)-cumulant functionals" which enlarges the family of free cu- mulant functionals of the space. In the motivating case of canonical operators on the full Fock space we find a simple formula for a relevant family of (ℓ, r)-cumulants of a (2d)-tuple (A1, . . . , Ad, B1, . . . , Bd), with A1, . . . , Ad canonical operators on the left and B1, . . . , Bd canonical operators on the right. This extends a known one-sided for- mula for free cumulants of A1, . . . , Ad, which establishes a basic operator model for the R-transform of free probability. Keywords: bi-free probability, canonical operators, double-ended queues, bi-non-crossing partitions, bi-free cumulants. Mathematics Subject Classification 2000: Primary 46L54; Secondary 68R05. 1. Introduction In this paper 2 we follow the guiding line offered by a special type of canonical operators on the full Fock space, in order to identify what kind of cumulant functionals should be considered for the concept of bi-free independence introduced by D. Voiculescu [8], [9]. In Section 1.1 we give a general (rather informal) description of what is the above mentioned "guiding line"; the main point of the description is that we are upgrading from (I) to (II) in the diagrams displayed on the next page. 1.1 From (bi-)free independence to partitions, via canonical operators. The concept of free independence for noncommutative random variables was pinned down by D. Voiculescu in the 1980's ([6], [7]), with inspiration from natural families of generators for algebras associated to free groups. An important further idea brought by R. Speicher [5] was that calculations with freely independent random variables can be efficiently done by using the lattices N C(n) of non-crossing partitions, and free cumulant functionals based on these lattices. The main point of the cumulant approach is that free independence is equivalent to a condition (many times easier to verify than the original definition) of "vanishing of mixed free cumulants" for the random variables in question. What is the guiding line which allows one to start from the original definition of free independence, based on free groups, and "find" the N C(n)'s? Here we pursue (with due 1Research supported by a Discovery Grant from NSERC, Canada. 2This is the electronic version of an article published in International Journal of Mathematics 26 (2015), Issue 02, paper 1550016. DOI: 10.1142/S0129167X15500160, c(cid:13)World Scientific Publishing Company. 1 acknowledgement that there is more than one answer to the question just asked) the line provided by a special type of "canonical" operators on the full Fock space Td over Cd. These are certain power series in creation/annihilation operators on Td, which were invented in [7] for d = 1, then extended in [3] to general d. The fact that they play a role in free probability is not so surprising, since Td itself is in a certain sense an incarnation of the free monoid on d generators. But now, the way how products of canonical operators act on the vacuum- vector of Td is encoded by the action of inserting and removing objects in a well-known gadget from theoretical computer science, called lifo-stack (lifo = abbreviation for "last- in-first-out"). Lifo-stacks are closely connected (through the intermediate of some lattice paths called Lukasiewicz paths) to non-crossing partitions, so overall we get a diagram like this: (I) (cid:18) free independence for d random variables (cid:19) −→ (cid:18) −→ (cid:18) lifo- operators on full Fock space (cid:19) d-tuple of canonical stacks (cid:19) −→ (cid:18) non-crossing partitions (cid:19) . In 2013, Voiculescu [8], [9] started to study the concept of bi-free independence for d pairs of random variables in a noncommutative probability space. This too is a concept inspired from looking at algebras associated to free groups, but its definition is in some sense indirect -- it ultimately boils down to a special way of representing the 2d random variables in question on a free-product space. In order to advance the study of bi-free independence, it is of obvious importance (even more so than it was the case for plain free independence) to be able to re-phrase it as a vanishing condition on mixed cumulants, for a suitable construction of cumulant functionals. The candidate for how to construct such "bi-free cumulant functionals" will come out of an upgrade of diagram (I), which tells us what lattices of partitions to use in the stead of N C(n)'s. The upgrade is as follows: (II)(cid:18) bi-free independence for d pairs of random variables (cid:19) (1) (2) −→ (cid:18) double-ended queues operators on full Fock space, d on left and d on right (2d)-tuple of canonical −→   −→ (cid:18) bi-non-crossing (cid:19) (3) partitions (cid:19) .   (We have numbered the three arrows which appear in diagram (II), in order to discuss them separately in the next subsection.) The lattices of partitions obtained in (II) are indexed not only by a positive integer n, but also by a tuple χ = (h1, . . . , hn) where every hi is either the letter ℓ (for "left") or the letter r (for "right"). Throughout the paper, the notation used for such a lattice of partitions will be P (χ)(n). For every n ∈ N and χ ∈ {ℓ, r}n we have that P (χ)(n) is a collection of partitions of {1, . . . , n}. If χ = (ℓ, ℓ, . . . , ℓ) or if χ = (r, r, . . . , r), then P (χ)(n) = N C(n), but for arbitrary χ ∈ {ℓ, r}n we generally have P (χ)(n) 6= N C(n). 1.2 Discussion of the three arrows in diagram (II). Discussion of the connection " (1) −→". Let d be a fixed positive integer, and let us also fix an orthonormal basis e1, . . . , ed for Cd. Recall that the full Fock space over Cd is Td = C ⊕ Cd ⊕ (Cd ⊗ Cd) ⊕ · · · ⊕ (Cd)⊗n ⊕ · · · (1.1) 2 In Td we have a preferred orthonormal basis, namely {ξvac} ∪ {ei1 ⊗ · · · ⊗ ein n ≥ 1, 1 ≤ i1, . . . , in ≤ d}, (1.2) where ξvac is a fixed unit vector in the first summand C on the right-hand side of (1.1). For every 1 ≤ i ≤ d we denote by Li, Ri ∈ B(Td) the left-creation and respectively the right-creation operator on Td defined by the vector ei. These are isometries which act on the preferred basis by Li(ξvac) = Ri(ξvac) = ei and by Li(ei1 ⊗ · · · ⊗ ein) = ei ⊗ ei1 ⊗ · · · ⊗ ein, Ri(ei1 ⊗ · · · ⊗ ein) = ei1 ⊗ · · · ⊗ ein ⊗ ei, for n ≥ 1 and 1 ≤ i1, . . . , in ≤ d. Suppose we are given a polynomial f without constant term in non-commuting indeter- minates z1, . . . , zd. Write it explicitly as f (z1, . . . , zd) = ∞Xn=1 dXi1,...,in=1 α(i1,...,in)zi1 · · · zin, (1.3) where the α's are in C (and ∃ no ∈ N such that α(i1,...,in) = 0 for n > no). For 1 ≤ i ≤ d, let Ai be the operator in B(Td) defined as follows: Ai := L∗ i(cid:16)I + ∞Xn=1 dXi1,...,in=1 α(i1,...,in)Lin · · · Li1(cid:17). (1.4) We will say that (A1, . . . , Ad) is the d-tuple of left canonical operators with symbol f . The reason to care about this d-tuple is that it provides one of the possible approaches (historically the first) to the R-transform, a fundamental tool used in free probability. More precisely: if we endow B(Td) with the vacuum-state ϕvac (i.e. with the vector-state defined by the vector ξvac from (1.2)), then the R-transform of (A1, . . . , Ad) with respect to ϕvac is 3 the given f . For a detailed presentation of how this goes, see Lecture 21 of [4]. Let's also note that, as explained in that lecture of [4] (Theorem 21.4 there), the derivation of the R-transform of (A1, . . . , Ad) relies solely on the fact that L1, . . . , Ld used in Equation (1.4) form a free family of Cuntz isometries. The latter fact means, by definition, that the Li's are isometries with L∗ i Lj = 0 for i 6= j, and that one has ϕvac(Li1 · · · LimL∗ j1 · · · L∗ jn) = 0 for all non-negative integers m, n with m + n ≥ 1 and all i1, . . . , im, j1, . . . , jn ∈ {1, . . . , d}. Since the creation operators R1, . . . , Rd on the right also form a free family of Cuntz isometries, it follows that canonical d-tuples of operators could be equivalently constructed by using creation and annihilation operators on the right side (instead of the left side favoured in Equation (1.4)). For the present paper it is important to also write explicitly this second set of formulas. So let g be a polynomial without constant term in the same z1, . . . , zd, and let us write it explicitly as g(z1, . . . , zd) = ∞Xn=1 dXi1,...,in=1 β(i1,...,in)zi1 · · · zin (1.5) 3 So what we have here is a canonical construction which produces a d-tuple of operators with prescribed R-transform. In this discussion f could be a formal power series in z1, . . . , zd, in which case A1, . . . , Ad would live in a suitable algebra of formal operators on Td, as described for instance on pp. 344-346 of [4]. For the sake of not complicating the notations, we will stick to f being a polynomial. 3 (with β's in C, and where ∃ no such that β(i1,...,in) = 0 for n > no). For 1 ≤ i ≤ d, we put Bi := R∗ i(cid:16)I + ∞Xn=1 dXi1,...,in=1 β(i1,...,in)Rin · · · Ri1(cid:17) ∈ B(Td). (1.6) Then (B1, . . . , Bd) is called the d-tuple of right canonical operators with symbol g, and has the property that its R-transform with respect to ϕvac is the polynomial g we started with. If taken in isolation, the Bi's from Equation (1.6) would merely duplicate what the Ai's from (1.4) are already doing. What is interesting is to consider the combined (2d)-tuple of Ai's and Bi's -- this provides a significant example of d pairs of left/right variables, as one wants to study in bi-free probability. Discussion of the connection " (2) −→". We now examine the values of ϕvac on monomials made with the canonical operators A1, . . . , Ad, B1, . . . , Bd that were considered above. Let us first recap the one-sided case, where we look at an expectation ϕvac(Aj1 · · · Ajn) = hAj1 · · · Ajnξvac , ξvaci, for some n ≥ 1 and 1 ≤ j1, . . . , jn ≤ d. If we replace each of Aj1, . . . , Ajn by using (1.4), and if we expand the ensuing product of sums, then we arrive to act on ξvac with products of the form j1(Li1,1 · · · Li1,m(1)) · · · L∗ L∗ jn(Lin,1 · · · Lin,m(n) ), (1.7) where m(1), . . . , m(n) ≥ 0 are such that m(1)+· · ·+m(n) = n. The action of such a product on ξvac can be followed intuitively by thinking of how a collection of n balls 1(cid:13), . . . , n(cid:13) moves through a lifo-stack -- the balls go into the stack in 4 groups (creation) and are taken out of the stack one by one (annihilation). j 's become Ri's and R∗ What happens when we upgrade to a monomial which contains both some Ai's and some Bi's? We can still proceed in the same way as above, only that in (1.7) some of the Li's and L∗ j 's. The intuition of n balls moving through a device like a stack continues to work, but sometimes (when we have "R" instead of "L") the balls must go in or out "by the other end of the stack". This is precisely the device which in theoretical computer science goes under the name of double-ended queue, or deque for short -- see e.g. Section 2.2 of [2]. Some pictures showing how we think about deques and how we use them in the present paper appear on pages 11-12 in Section 3. Discussion of the connection " (3) −→". The set of partitions P (χ)(n). Once again, let us first recap the one-sided case, where we look at n balls moving through a lifo-stack. Every possible scenario of how the balls move through the stack has associated to it a certain partition of {1, . . . , n}, which we call "output-time partition" (see Definition 3.3, Example 3.4.2 below). When we pursue the discussion started in (1.7), the concrete formula obtained for ϕvac(Aj1 · · · Ajn) comes out as a sum indexed by all possible output-time partitions. But the last-in-first-out rule of the stack forbids output-time partitions from having crossings! What has come out is precisely a summation formula over N C(n) (as we knew it should). What happens when we upgrade to the case of combined Ai's and Bi's? We now have n balls moving through a deque. We still have the concept of output-time partition associated to a scenario for how the balls move through the deque, but this partition may now have 4 The sizes of the groups of balls going in the stack are m(n), . . . , m(2), m(1). Here we ignore for the moment some relations that must also be imposed in between the indices i1,1, . . . , in,m(n) and j1, . . . , jn. 4 crossings, due to the interference between left and right moves (e.g. it is possible that some of the balls 1(cid:13), . . . , n(cid:13) enter the deque by one side and exit by the other). If we fix a tuple χ ∈ {ℓ, r}n, then the set P (χ)(n) of bi-non-crossing partitions cor- responding to χ will be defined in Section 3 of the paper as the set of all partitions of {1, . . . , n} which can arise as output-time partition for a deque-scenario compatible with χ. (See details in Definition 3.5.) In the case when χ happens to be either (ℓ, . . . , ℓ) or (r, . . . , r), then the deque is reduced to a lifo-stack, and P (χ)(n) is equal to N C(n). For general χ ∈ {ℓ, r}n, it follows immediately from the definition that P (χ)(n) has the same cardinality as N C(n), but P (χ)(n) is generally different from N C(n) itself. In Section 4 of the paper we will find (Theorem 4.10) an alternative description for P (χ)(n). It says that P (χ)(n) = {σχ · π π ∈ N C(n)}, where σχ is a specific (con- cretely described) permutation of {1, . . . , n} associated to χ, and where the action of a permutation σ on a partition π is defined in the natural way (if π = {V1, . . . , Vk}, then σ · π = {σ(V1), . . . , σ(Vk)}). This alternative description of P (χ)(n) is very useful for con- crete calculations. It also gives immediately the fact that, with respect to reverse refinement order, P (χ)(n) is a lattice isomorphic to N C(n). 1.3 From partitions to cumulant functionals. In both classical and free probability theory, the standard method to introduce cumu- lant functionals goes by writing a so-called "moment-cumulant" formula, where moments are expressed in terms of cumulants via summations over a suitable family of lattices of partitions. In particular, free cumulant functionals (as introduced by Speicher [5]) have a moment-cumulant formula based on non-crossing partitions. To be specific, let a non- commutative probability space (A, ϕ) be given. The free cumulant functionals of (A, ϕ) are defined as the family of multilinear functionals ( κn : An → C )∞ n=1 which is uniquely determined by the requirement that for every n ≥ 1 and a1, . . . , an ∈ A we have: (M-FC) ϕ(a1 · · · an) = Xπ∈N C(n)(cid:16) YV ∈π κV ( (a1, . . . , an) V ) (cid:17). By the same token, the families of partitions P (χ)(n) discussed at the end of section 1.2 can be used to define a concept of cumulant functionals, as follows. Let a noncommutative probability space (A, ϕ) be given. There exists a family of multilinear functionals (cid:16) κχ : An → C(cid:17)n≥1, χ∈{ℓ,r}n which is uniquely determined by the requirement that for every n ≥ 1, χ ∈ {ℓ, r}n and a1, . . . , an ∈ A we have: (M-F2C) ϕ(a1 · · · an) = Xπ∈P (χ)(n)(cid:16) YV ∈π κχV ( (a1, . . . , an) V ) (cid:17). The explanation of various notational details, the easy proof of existence/uniqueness, and a bit of further discussion around the functionals κχ are given in Section 5 of the paper. We will refer to 5 these functionals as the (ℓ, r)-cumulant functionals associated to the 5 In view of the results obtained in [1], after the present paper was first circulated, it is justified to also refer to the κχ's as "bi-free cumulant functionals" associated to (A, ϕ). 5 noncommutative probability space (A, ϕ). It is immediate that they provide an enlargement of the family of free cumulants (κn)∞ n=1 associated to the same space, in the respect that κn = κ( ℓ, . . . , ℓ {z }n ), n ≥ 1. ) = κ( r, . . . , r {z } n 1.4 Back to canonical operators: main result of the paper. The concept of (ℓ, r)-cumulants was introduced in section 1.3 for a general noncommu- tative probability space (A, ϕ). Here we go back to the special space (B(Td), ϕvac) from section 1.2. Let (A1, . . . Ad) and (B1, . . . Bd) be d-tuples of left (respectively right) canonical operators with symbols f and g, as in Equations (1.4) and (1.6). We are interested in the (ℓ, r)-cumulants of the combined (2d)-tuple (A1, . . . , Ad, B1, . . . , Bd). The point we want to establish is that a relevant family of such (ℓ, r)-cumulants simply consists of coefficients of either the polynomial f (an α(i1,...,in) from Equation (1.3)) or the polynomial g (a β(i1,...,in) from Equation (1.5)). For the statement of our theorem it is convenient to use a unified notation for Ai's and Bi's: Ai =: Ci;ℓ and Bi =: Ci;r, for 1 ≤ i ≤ d. Theorem. Consider all the notations pertaining to A1, . . . , Ad, B1, . . . , Bd that were introduced above. Let n be a positive integer and let χ = (h1, . . . , hn) be in {ℓ, r}n. Let us record explicitly where are the occurrences of ℓ and of r in χ:   {m 1 ≤ m ≤ n, hm = ℓ} =: {mℓ(1), . . . , mℓ(u)} with mℓ(1) < · · · < mℓ(u), {m 1 ≤ m ≤ n, hm = r} =: {mr(1), . . . , mr(v)} with mr(1) < · · · < mr(v). Then for every i1, . . . , in ∈ {1, . . . , d} we have κχ(Ci1;h1, . . . , Cin;hn) =  α(imr (v),...,imr (1),imℓ(1),...,imℓ(u)), if hn = ℓ, β(imℓ (u),...,imℓ(1),imr (1),...,imr (v)), if hn = r. (1.8) Remark. The above theorem generalizes the result from [3] which says that the canonical left d-tuple (A1, . . . , Ad) has R-transform R(A1,...,Ad) = f . Indeed, if the tu- ple χ from the theorem is set to be (ℓ, . . . , ℓ) ∈ {ℓ, r}n, then Equation (1.8) says that κn(Ai1 , . . . , Ain ) = α(i1,...,in), ∀ 1 ≤ i1, . . . , in ≤ d. So then R(A1,...,Ad)(z1, . . . , zd) := = ∞Xn=1 ∞Xn=1 dXi1,...,in=1 dXi1,...,in=1 κn(Ai1, . . . , Ain)zi1 · · · zin α(i1,...,in)zi1 · · · zin = f (z1, . . . , zd), as claimed. (By setting χ = (r, . . . , r) we could, of course, also infer from the above theorem that R(B1,...,Bd) = g.) 6 Example. The upshot of the theorem is that every mixed moment of the Ai's and Bi's is written as a straight sum (no signs or coefficients!) indexed by P (χ)(n), where every term of the sum is a product of α's and β's. For a concrete example, say that we are interested in the mixed moment ϕvac(Ai1Bi2Ai3Bi4), for some given i1, i2, i3, i4 ∈ {1, . . . , d}. In the above theorem we take n = 4 and χ = (ℓ, r, ℓ, r), and get: ϕvac(Ai1Bi2Ai3Bi4) = κ(ℓ,r,ℓ,r)(Ai1, Bi2, Ai3 , Bi4) + κ(ℓ)(Ai1) κ(r,ℓ,r)(Bi2, Ai3, Bi4) + · · · + κℓ(Ai1) κr(Bi2) κℓ(Ai3 ) κr(Bi4) = β(i3,i1,i2,i4) + α(i1) β(i3,i2,i4) + · · · + α(i1) β(i2) α(i3) β(i4), a sum of 14 terms corresponding to the 14 partitions in P (ℓ,r,ℓ,r)(4). (As is easily checked, P (ℓ,r,ℓ,r)(4) is different from N C(4) -- it contains the crossing partition { {1, 3}, {2, 4} }, and misses the non-crossing partition { {1, 4}, {2, 3} }.) Note that the nice feature of getting a summation formula without signs and coefficients is due precisely to the fact that we are using the tuple χ = (ℓ, r, ℓ, r) coming from the mixed moment that we want to calculate. If we tried for instance to evaluate the same mixed moment via a sum indexed by P (r,r,r,r)(4) = N C(4) (which would just be the usual free cumulant expansion), then we would run from the very beginning into the term κ(r,r,r,r)(Ai1, Bi2, Ai3, Bi4 ) = κ4(Ai1, Bi2, Ai3 , Bi4) = β(i3,i1,i2,i4) + α(i1,i3)β(i2,i4) − α(i2,i3)β(i1,i4), and the nice structure in the formula for ϕvac(Ai1Bi2Ai3Bi4) (a straight sum of products) would only emerge after going through some more complicated expressions, and doing can- cellations between terms. 1.5 Vanishing mixed (ℓ, r)-cumulants, and a question. Suppose the polynomials f and g considered in section 1.2 are of the form f (z1, . . . , zd) = f1(z1) + · · · + fd(zd) and g(z1, . . . , zd) = g1(z1) + · · · + gd(zd), (1.9)   where f1, . . . , fd, g1, . . . , gd are polynomials of one variable. Then the formulas defining the canonical operators A1, . . . , Ad, B1, . . . , Bd simplify to 1 ≤ i ≤ d. (1.10) Ai = L∗ Bi = R∗   i (cid:0)I + fi(Li)(cid:1), i (cid:0)I + gi(Ri)(cid:1), The d pairs of operators (A1, B1), . . . , (Ad, Bd) appearing in Equations (1.10) give an example of bi-free family of pairs of elements of a noncommutative probability space, in the sense of Voiculescu [8]. On the other hand, in view of the formula for (ℓ, r)-cumulants provided by Equation (1.8), the special case of f, g considered in (1.9) can be equivalently described via the requirement that κχ(cid:0)Ci1;h1, . . . , Cin;hn(cid:1) = 0 whenever ∃ 1 ≤ p < q ≤ n such that ip 6= iq.   7 This coincidence is in line with the fact that various forms of independence for noncom- mutative random variables which are considered in the literature have a combinatorial incarnation expressed in terms of the vanishing of some mixed cumulants. It is in fact tempting to make a definition and ask a question, as follows. Definition. Let (A, ϕ) be a noncommutative probability space, and let (cid:0) κχ : An → C(cid:1)n≥1,χ∈{ℓ,r}n be the family of (ℓ, r)-cumulant functionals of (A, ϕ). Let a1, b1, . . . , ad, bd be in A. We say that the pairs (a1, b1), . . . , (ad, bd) are combinatorially-bi- free to mean that the following condition is fulfilled: denoting ci;ℓ := ai and ci;r := bi for 1 ≤ i ≤ d, one has κχ(cid:0)ci1;h1, . . . , cin;hn(cid:1) = 0 whenever n ≥ 2, i1, . . . , in ∈ {1, . . . , d}, χ = (h1, . . . , hn) ∈ {ℓ, r}n and ∃ 1 ≤ p < q ≤ n such that ip 6= iq. (1.11)   Question. Is it true that combinatorial-bi-freeness is equivalent to the (representation theoretic) concept of bi-freeness introduced by Voiculescu in [8]? After the first version of the present paper was circulated, the above question was found to have an affirmative answer, in the paper [1] by I. Charlesworth, B. Nelson and P. Skoufranis. 1.6 Organization of the paper. Besides the present introduction, the paper has five other sections. After some review of background in Section 2, we introduce the sets of partitions P (χ)(n) in Section 3, via the idea of examining double-ended queues. The alternative description of P (χ)(n) via direct bijection with N C(n) is presented in Section 4. In Section 5 we introduce the family of (ℓ, r)-cumulant functionals κχ that are associated to a noncommutative probability space. Finally, in Section 6 we prove the main result of the paper, giving the formula for (ℓ, r)- cumulants of canonical operators that was announced in Equation (1.8) above. 2. Background on partitions and on Lukasiewicz paths Definition 2.1. [Partitions of {1, . . . , n}.] Let n be a positive integer. 1o We will let P(n) denote the set of all partitions of {1, . . . , n}. A partition π ∈ P(n) is thus a set π = {V1, . . . , Vk} where V1, . . . , Vk (called the blocks of π) are non-empty sets with Vi ∩ Vj = ∅ for i 6= j and with ∪k i=1Vi = {1, . . . , n}. 2o On P(n) we consider the partial order given by reverse refinement; that is, for π, ρ ∈ P(n) we will write "π ≤ ρ" to mean that for every block V ∈ π there exists a block W ∈ ρ such that V ⊆ W . The minimal and maximal partition with respect to this partial order will be denoted as 0n and 1n, respectively: 0n :=n {1}, . . . , {n}o, 1n :=n {1, . . . , n}o. (2.1) 8 3o Let τ be a permutation of {1, . . . , n} and let π = {V1, . . . , Vk} be in P(n). We will use the notation "τ · π" for the new partition τ · π := { τ (V1), . . . , τ (Vk) } ∈ P(n). 4o Let π be a partition in P(n). By the opposite of π we will mean the partition πopp := τo · π ∈ P(n), where τo is the order-reversing permutation of {1, . . . , n} (with τo(m) = n + 1 − m for every 1 ≤ m ≤ n). 5o A partition π ∈ P(n) is said to be non-crossing when it is not possible to find two distinct blocks V, W ∈ π and numbers a < b < c < d in {1, . . . , n} such that a, c ∈ V and b, d ∈ W . The set N C(n) of all non-crossing partitions in P(n) will play a significant role in this paper; for a review of basic some facts about it we refer to Lectures 9 and 10 of [4]. Let us record here that N C(n) is one of the many combinatorial structures counted by Catalan numbers, one has N C(n) = Cn := (2n)!/n!(n + 1)! (the n-th Catalan number). Definition 2.2. [Lukasiewicz paths.] 1o We will consider paths in Z2 which start at (0, 0) and make steps of the form (1, i) with i ∈ N ∪ {−1, 0}. Such a path, λ =(cid:0) (0, 0), (1, j1), (2, j2), . . . , (n, jn)(cid:1), (2.2) is said to be a Lukasiewicz path when it satisfies the conditions that jm ≥ 0 for every 1 ≤ m ≤ n and that jn = 0. 2o For every n ≥ 1, we will use the notation Luk(n) for the set of all Lukasiewicz paths with n steps. For a path λ ∈ Luk(n) written as in Equation (2.2), we will refer to the vector ~λ = (j1 − 0, j2 − j1, . . . , jn − jn−1) ∈ (N ∪ {−1, 0})n (2.3) as to the rise-vector of λ. It is immediate how λ can be retrieved from its rise-vector; moreover, it is immediate that a vector (q1, . . . , qn) ∈ (N ∪ {−1, 0})n appears as rise-vector ~λ for some λ ∈ Luk(n) if and only if its satisfies the conditions that q1 + · · · + qm ≥ 0 for every 1 ≤ m ≤ n, and q1 + · · · + qn = 0. (2.4) Remark and Notation 2.3. [The surjection Ψ and the bijection Φ.] Let n be a positive integer. 1o Let π = {V1, . . . , Vk} be a partition in P(n). Consider the vector (q1, . . . , qn) ∈ ( N ∪ {−1, 0} )n where for 1 ≤ m ≤ n we put qm :=(cid:26) Vi − 1, −1, if m = min(Vi) for an i ∈ {1, . . . , k} , otherwise. (2.5) It is immediately seen that (q1, . . . , qn) satisfies the conditions listed in (2.4), hence it is the rise-vector of a uniquely determined path λ ∈ Luk(n). 9 2o We will denote by Ψ : P(n) → Luk(n) (also denoted as Ψn, if needed to clarify what is n) the map which acts by Ψ(π) := λ, π ∈ P(n), with λ obtained out of π via the rise-vector described in (2.5). 3o The map Ψ : P(n) → Luk(n) introduced above has the remarkable property that its restriction to N C(n) gives a bijection from N C(n) onto Luk(n); for the verification of this [4], Proposition 9.8. We will denote by Φ : Luk(n) → N C(n) the inverse of fact, see e.g. this bijection. That is: for every λ ∈ Luk(n), we define Φ(λ) to be the unique partition in N C(n) which has the property that Ψ( Φ(λ) ) = λ. The bijection Φ confirms the well-known fact that the set of paths Luk(n) has the same cardinality Cn (Catalan number) as N C(n). 3. Double-ended queues and the sets of partitions P (χ)(n) Definition and Remark 3.1. [Description of a deque device.] We will work with a device called double-ended queue, or deque for short, which is used in the study of information structures in theoretical computer science (see e.g. [2], Section 2.2). We will think about this device in the way depicted in Figures 1, 2 below, and described as follows. Let n be a fixed positive integer, and suppose we have n labelled balls 1(cid:13) , . . . , n(cid:13) which have to move from an input pipe into an output pipe (both depicted vertically in the figures), by going through a deque pipe (depicted horizontally). The deque device operates in discrete time: it goes through a sequence of states, recorded at times t = 0, t = 1, . . . , t = n, where at time t = 0 all the n balls are in the input pipe and at t = n they are all in the output pipe. Compared to the discussion in Knuth's treatise [2], we will limit the kinds of moves 6 that a deque can do, and we will require that: For every 0 ≤ i ≤ n − 1, the deque device moves from its state at time t = i to its state at time t = i + 1 by performing either a "left-p move" or a "right-p move", where p ∈ N ∪ {0}.     The description of a left-p move is like this: "Take the top p balls that are in the input pipe, and insert them into the deque pipe, from the left. Then take the leftmost ball in the deque pipe and insert it at the bottom of the output pipe." The description of a right-p move is analogous to the one of a left-p move, only that the words "left" and "leftmost" get to be replaced by "right" and respectively "rightmost". 6 In [2] the deque may perform any sequence of "insertions and deletions at either end of the queue", where the term "insertion" designates the operation of moving some balls from the input pipe into the deque pipe, and "deletion" refers to moving some balls from the deque pipe into the output pipe. In the present paper we will only allow the special moves described in Definition 3.1, which match, in some sense, the creation and annihilation performed by canonical operators on a full Fock space. 10 As is clear from the description of a left-p (or right-p) move, the number p ∈ N ∪ {0} used in the move (t = i) (t = i + 1) is subjected to the restriction that there exist p balls (or more) in the input pipe at time t = i. Note also the additional restriction that p = 0 can be used in the move (t = i) (t = i + 1) only if the device has at least 1 ball in the deque pipe at time t = i. output → pipe 2(cid:13) 5(cid:13) ր → deque pipe input pipe 1(cid:13) 3(cid:13) 4(cid:13) 6(cid:13) 7(cid:13) 8(cid:13) 9(cid:13) Figure 1: Say that n = 9. Here is a possible state of the deque device at time t = 2. 11 2(cid:13) 5(cid:13) 8(cid:13) 7(cid:13) 6(cid:13) 1(cid:13) 3(cid:13) 4(cid:13) 9(cid:13) Figure 2: The deque device from Figure 1, at time t = 3, after performing a left-3 move. ****************************************************************** 2(cid:13) 5(cid:13) 4(cid:13) 1(cid:13) 3(cid:13) 6(cid:13) 7(cid:13) 8(cid:13) 9(cid:13) Figure 3: The deque device from Figure 1, at time t = 3, after performing a right-0 move. 12 Definition and Remark 3.2. [Deque-scenarios.] Let n be the same fixed positive integer as in Definition 3.1, and consider the deque device described there. We assume that at time t = 0 the n balls are sitting in the input pipe in the order 1(cid:13) , . . . , n(cid:13) , counting top-down. From the description of the moves of the device, it is clear that for every 0 ≤ i ≤ n there are exactly i balls in the output pipe at time t = i. In particular, all n balls find themselves in the output pipe at time t = n (even though they may not be sitting in the same order as at time t = 0). We will use the name deque-scenario to refer to a possible way of moving the n balls through the deque device, according to the rules described above. Every deque-scenario is thus determined by an array of the form (cid:18) p1 h1 · · · · · · hn (cid:19) , pn (3.1) with p1, . . . , pn ∈ N ∪ {0} and h1, . . . , hn ∈ {ℓ, r}; this array simply records the fact that in order to go from its state at time t = i − 1 to its state at time t = i, the device has executed a left-pi move, if hi = ℓ a right-pi move, if hi = r , 1 ≤ i ≤ n. Let us observe that the top line of the array in (3.1) must satisfy the inequalities     p1 + · · · + pi ≥ i, ∀ 1 ≤ i ≤ n, where for i = n we must have p1 + · · · + pn = n. (3.2) This is easily seen by counting that at time t = i there are i balls in the output pipe and n − (p1 + · · · + pi) balls in the input pipe, which leaves a difference of n −(cid:16)i + n − (p1 + · · · + pi)(cid:17) = (p1 + · · · + pi) − i balls that must be in deque pipe. (And of course, the number of balls found in the deque pipe at t = i must be ≥ 0, with equality for t = n.) It is easy to see that conversely, every array as in (3.1) with p1, . . . , pn satisfying (3.2) will define a working deque-scenario -- the inequalities p1 + · · · + pi ≥ i ensure that we never run into the situation of having to "move a ball out of the empty deque pipe". Thus, as a mathematical object, the set of deque-scenarios can be simply introduced as the set of arrays of the kind shown in (3.1), and where (3.2) is satisfied. Moreover, let us observe that condition (3.2) can be read as saying that the n-tuple (p1 − 1, . . . , pn − 1) ∈(cid:0)N ∪ {−1, 0}(cid:1)n (3.3) is the rise-vector of a uniquely determined Lukasiewicz path λ, as reviewed in Section 2. We will then refer to the deque-scenario described by the array (3.1) as the deque-scenario determined by (λ, χ), where λ ∈ Luk(n) has rise-vector given by (3.3) and χ is the n-tuple (h1, . . . , hn) ∈ {ℓ, r}n from the second line of (3.1). 13 Definition and Remark 3.3. [Output-time partition associated to a deque-scenario.] We consider the same notations as above and we look at the deque-scenario determined by (λ, χ), where λ ∈ Luk(n) has rise-vector ~λ = (p1−1, . . . , pn−1) and where χ = (h1, . . . , hn) ∈ {ℓ, r}n. Let i ∈ {1, . . . , n} be such that pi > 0, and consider the i-th move of the deque device (the move that takes the device from its state at t = i − 1 to its state at t = i). In that move there is a group of pi balls (namely those with labels from p1 + . . . + pi−1 + 1 to p1 + . . . + pi) which leave together the input pipe. These balls arrive in the output pipe one by one, at various later times, which we record as t(i) 1 < t(i) 2 < · · · < t(i) pi . (3.4) Observe that in particular we have t(i) 1 = i; indeed, it is also part of the i-th move of the device that the ball with label p1 + · · · + pi goes from the deque pipe into the output pipe. Let us make the notation Ti := {t(i) 1 , . . . , t(i) pi }, where t(i) 1 , . . . , t(i) pi are from (3.4). In the preceding paragraph we have thus constructed a set Ti ⊆ {1, . . . , n} for every 1 ≤ i ≤ n such that pi > 0. It is clear from the construction that for every such i we have It is also clear that the sets Ti = pi and min(Ti) = i. {Ti 1 ≤ i ≤ n such that pi > 0} (3.5) (3.6) form together a partition of π ∈ P(n). We will refer to this π as the output-time partition associated to the pair (λ, χ). Example 3.4. 1o A concrete example: say that n = 5, that λ ∈ Luk(5) has rise-vector ~λ = (2, −1, 1, −1, −1), and that χ = (r, ℓ, ℓ, r, ℓ). In the deque-scenario associated to this pair (λ, χ), there are two groups of balls that are moved from the input pipe into the deque pipe: first group consists of 1(cid:13) , 2(cid:13) , 3(cid:13) , which arrive in the deque pipe at time t = 1; the second group consists of 4(cid:13) , 5(cid:13) , which arrive in the deque pipe at time t = 3. The final order of the balls in the output pipe (counting downwards) is 3(cid:13) , 1(cid:13) , 5(cid:13) , 2(cid:13) , 4(cid:13) , and the output-time partition associated to (λ, χ) is π = { {1, 2, 4}, {3, 5} } ∈ P(5). 2o Let n be a positive integer, and consider the n-tuple χℓ := (ℓ, . . . , ℓ) ∈ {ℓ, r}n. For any λ ∈ Luk(n), the deque-scenario determined by λ and χℓ is what one might call a "lifo-stack process" (where lifo is a commonly used abbreviation for last-in-first-out). It is easy to see that the output-time partition associated to the pair (λ, χℓ) is the non-crossing partition Φ(λ), where Φ : Luk(n) → P(n) is as reviewed in Remark 2.3.3. A similar statement holds if instead of χℓ we use the n-tuple χr := (r, . . . , r); that is, the output-time partition associated to (λ, χr ) is the same Φ(λ) ∈ N C(n) as above. 14 Definition 3.5. 1o Consider a pair (λ, χ) where λ ∈ Luk(n) and χ ∈ {ℓ, r}n, and let π ∈ P(n) be the output-time partition associated to (λ, χ) in Definition 3.3. We will denote this partition π as "Φχ(λ)". 2o Let χ be an n-tuple in {ℓ, r}n. The notation introduced in 1o above defines a function Φχ : Luk(n) → P(n). We define P (χ)(n) := {Φχ(λ) λ ∈ Luk(n)} ⊆ P(n). (3.7) Proposition 3.6. Let n be a positive integer, let χ be an n-tuple in {ℓ, r}n, and consider the function Φχ : Luk(n) → P(n) introduced in Definition 3.5. 1o Φχ is injective, hence it gives a bijection between Luk(n) and P (χ)(n). 2o Let Ψχ : P (χ)(n) → Luk(n) be the function inverse to Φχ. Then Ψχ is the restriction to P (χ)(n) of the canonical surjection Ψ : P(n) → Luk(n) that was reviewed in Remark 2.3.2. Proof. Both parts of the proposition will follow if we can prove that Ψ ◦ Φχ is the identity map on Luk(n). Thus given a path λ ∈ Luk(n) and denoting Φχ(λ) =: π, we have to show that Ψ(π) = λ. But the latter fact is clear from the observation made in (3.5) of Remark 3.3. (cid:4) Remark 3.7. Let n be a positive integer. 1o From Proposition 3.6 and the fact that Luk(n) = Cn (n-th Catalan number), it follows that P (χ)(n) = Cn for every χ ∈ {ℓ, r}n. 2o Suppose that χ = (ℓ, . . . , ℓ). The discussion from Example 3.4.2 shows that in this case we have P (χ)(n) = N C(n). Similarly, we also have P (χ)(n) = N C(n) in the case when χ = (r, . . . , r). 3o If n ≤ 3, then it is clear from cardinality considerations that P (χ)(n) = P(n) = N C(n), no matter what χ ∈ {ℓ, r}n we consider. For n ≥ 4, cardinality considerations now show that P (χ)(n) is a proper subset of P(n). It is usually different from N C(n). (For instance the output-time partition from Example 3.4.1 is not in N C(5), showing that χ = (r, ℓ, ℓ, r, ℓ) ∈ {ℓ, r}5 has P (χ)(5) 6= N C(5).) Some general properties of the sets of partitions P (χ)(n) will follow from their alternative description provided in the next section. 4. An alternative description for P (χ)(n) In this section we put into evidence a bijection between P (χ)(n) and N C(n) which is implemented by the action of a special permutation σχ of {1, . . . , n}. The main result of the section is Theorem 4.10. We will arrive to it by observing a certain construction of partition in N C(n) -- the "combined-standings partition" associated to a pair (λ, χ) ∈ Luk(n) × {ℓ, r}n, which is introduced in Definition 4.3. 15 Definition 4.1. Consider a pair (λ, χ) where λ ∈ Luk(n) and χ = (h1, . . . , hn) ∈ {ℓ, r}n, and let π ∈ P(n) be the output-time partition associated to (λ, χ) in Definition 3.3. Let us record explicitly where are the occurrences of ℓ and of r in χ: {m 1 ≤ m ≤ n, hm = ℓ} =: {mℓ(1), . . . , mℓ(u)} with mℓ(1) < · · · < mℓ(u), (4.1) {m 1 ≤ m ≤ n, hm = r} =: {mr(1), . . . , mr(v)} with mr(1) < · · · < mr(v).   1o Suppose that in (4.1) we have u 6= 0. We define a partition ρλ,χ;ℓ ∈ P(u) by the following prescription: two numbers q, q′ ∈ {1, . . . , u} are in the same block of ρλ,χ;ℓ if and only if the numbers mℓ(q), mℓ(q′) ∈ {1, . . . , n} belong to the same block of π. The partition ρλ,χ;ℓ will be called the left-standings partition associated to (λ, χ). 2o Likewise, if in (4.1) we have v 6= 0, then we define a partition ρλ,χ;r ∈ P(v) via the prescription that q, q′ ∈ {1, . . . , v} belong to the same block of ρλ,χ;r if and only if mr(q), mr(q′) are in the same block of π. The partition ρλ,χ;r will be called the right- standings partition associated to (λ, χ). Remark and Notation 4.2. Consider the framework of Definition 4.1. It will help the subsequent discussion if at this point we introduce some more terminology, which will also clarify the names chosen above for the partitions ρλ,χ;ℓ and ρλ,χ;r . 1o Same as in Section 3, we will think of the numbers in {1, . . . , n} as of moments in time. We will say that t ∈ {1, . . . , n} is a left-time (respectively a right-time) for χ to mean that ht = ℓ (respectively that ht = r). If t is a left-time for χ, then the unique q ∈ {1, . . . , u} such that t = mℓ(q) will be called the left-standing of t in χ. Likewise, if t is a right-time for χ, then the unique q ∈ {1, . . . , v} such that t = mr(q) will be called the right-standing of t in χ. 2o Let the rise-vector of λ be ~λ = (p1 − 1, . . . , pn − 1), with p1, . . . , pn ∈ N ∪ {0}. The numbers in the set I := {1 ≤ i ≤ n pi > 0} (4.2) will be called insertion times for (λ, χ). Recall that the output-time partition π associated to (λ, χ) has its blocks indexed by I; indeed, Equation (3.6) in Definition 3.3 introduces this partition as π = {Ti i ∈ I}. With a slight abuse of notation, ρλ,χ;ℓ and ρλ,χ;r from Definition 4.1 can be written as ρλ,χ;ℓ = {Vi i ∈ I} and ρλ,χ;r = {Wi i ∈ I} (4.3) (4.4) where for every i ∈ I we put Vi := {1 ≤ q ≤ u mℓ(q) ∈ Ti} and Wi := {1 ≤ q ≤ v mr(q) ∈ Ti}. (4.5) (Every Vi is a block of ρλ,χ;ℓ unless Vi = ∅, and every Wi is a block of ρλ,χ;r unless Wi = ∅. Note that Vi and Wi cannot be empty at the same time, since Vi + Wi = Ti = pi > 0.) 16 Definition 4.3. We continue to consider the framework of Definition 4.1 and of Notation 4.2. For every i ∈ I let us denote (n + 1) − Wi := {n + 1 − q q ∈ Wi} ⊆ {u + 1, . . . , n}. The partition will be called the combined-standings partition associated to (λ, χ). ρλ,χ := {Vi ∪ ( (n + 1) − Wi ) i ∈ I} (4.6) (4.7) Remark 4.4. The blocks of the partition ρλ,χ are indexed by the same set I of insertion times that was used to index the blocks of the output-times partition π = {Ti i ∈ I} in Notation 4.2. Moreover, we have Vi ∪ ( (n + 1) − Wi ) = Vi + Wi = Ti, ∀ i ∈ I; this shows that it must be possible to go between π and ρλ,χ via the action of some suitably chosen permutation of {1, . . . , n}. We next make the easy yet significant observation that the permutation in question can be picked so that it only depends on χ (even though each of π and ρλ,χ depends not only on χ, but also on λ). Definition 4.5. Let χ be a tuple in {ℓ, r}n. We associate to χ a permutation σχ of {1, . . . , n} defined (in two-line notation for permutations) as σχ :=(cid:18) 1 mℓ(1) u · · · u + 1 · · · mℓ(u) mr(v) · · · · · · mr(1) (cid:19) , n (4.8) where mℓ(1) < · · · < mℓ(u) and mr(1) < · · · < mr(v) are as in Definition 4.1 (the lists of occurrences of "ℓ" and "r" in χ). In (4.8) we include the possibility that v = 0 (when u = n and σχ is the identity permutation), or that u = 0 (when v = n and σχ(m) = n + 1 − m for every 1 ≤ m ≤ n). Lemma 4.6. Consider a pair (λ, χ) ∈ Luk(n)×{ℓ, r}n, and let π ∈ P(n) be the output-time partition associated to (λ, χ) in Definition 3.3. We have where ρλ,χ and σχ are as in Definitions 4.3 and 4.5, respectively, and where the action of a permutation on a partition is as reviewed in Definition 2.1.3. σχ · ρλ,χ = π, (4.9) Proof. We use the notations established earlier in this section. Clearly, (4.9) will follow if we prove that σχ(cid:0) Vi ∪ ( (n + 1) − Wi )(cid:1) = Ti, ∀ i ∈ I. (4.10) 17 Let us fix an i ∈ I for which we verify that (4.10) holds. Since the sets Vi∪( (n+1)−Wi ) ) and Ti have the same cardinality, it suffices to verify the inclusion "⊆" of the equality. And indeed, referring to how the permutation σχ is defined in Equation (4.8), we have: q ∈ Vi ⇒ σχ(q) = mℓ(q) ∈ Ti, and q ∈ (n + 1) − Wi ⇒ σχ(q) = mr(n + 1 − q) ∈ Ti (where the fact that mr(n + 1 − q) ∈ Ti comes from Equation (4.5), used for the element n + 1 − q ∈ Wi). Thus both σχ(Vi) and σχ( (n + 1) − Wi ) are subsets of Ti, and (4.10) follows. (cid:4) Example 4.7. Consider (same as in Example 3.4.1) the concrete case when n = 5, χ = (r, ℓ, ℓ, r, ℓ), and λ ∈ Luk(5) has rise-vector ~λ = (2, −1, 1, −1, −1). As found in Example 3.4.1, the output-time partition associated to this (λ, χ) is π = { {1, 2, 4}, {3, 5} }. The set of insertion times for (λ, χ) of this example is I = {1, 3}; in order to illustrate the system of notation from Equation (4.3), we then write π as π = {T1, T3}, with T1 = {1, 2, 4} and T3 = {3, 5}. The left-times for χ are mℓ(1) = 2, mℓ(2) = 3, mℓ(3) = 5, and the right-times are mr(1) = 1, mr(2) = 4. Since mℓ(1) ∈ T1 and mℓ(2), mℓ(3) ∈ T3, we get (in reference to the notations from Equations (4.4) and (4.5)) that V1 = {1}, V3 = {2, 3}, hence ρλ,χ;ℓ = { {1}, {2, 3} } ∈ P(3). For the right-times we have mr(1), mr(2) ∈ T1, giving us that W1 = {1, 2}, W3 = ∅, hence ρλ,χ;r = { {1, 2} } ∈ P(2). The combined-standings partition ρλ,χ associated to (λ, χ) has blocks V1 ∪ (6 − W1) = {1} ∪ {4, 5} and V3 ∪ (6 − W3) = {2, 3} ∪ ∅, hence ρλ,χ = { {1, 4, 5}, {2, 3} }. Finally, the permutation associated to χ is σχ =(cid:18) 1 2 3 4 5 2 3 5 4 1 (cid:19) . As explained in the proof of Lemma 4.6, we have that σχ( {1, 4, 5} ) = {1, 2, 4} = T1 and σχ( {2, 3} ) = {3, 5} = T3, leading to the equality σχ · ρλ,χ = π. Our next goal is to prove that the combined-standings partition ρλ,χ always is a non- crossing partition. In order to obtain this, we first prove a lemma. 18 Lemma 4.8. Consider the framework and notations of Definition 4.3. Let us denote the maximal element of I by j, and let us consider the block S = Vj ∪(cid:0) (n + 1) − Wj(cid:1) of the partition ρλ,χ. Then S is an interval-block (i.e. S = [t′, t′′] ∩ Z for some t′ ≤ t′′ in {1, . . . , n}). Proof. The conclusion of the lemma is clear if S = 1, so we will assume that S ≥ 2, i.e. that pj ≥ 2. The maximal insertion time j considered in the lemma is either a left-time or a right- time for χ. We will write the proof by assuming that j is a left-time (the case of a right-time is analogous). We denote the left-standing of the time j as q; recall from Notation 4.2 that this amounts to j = mℓ(q). In view of the above assumptions, the deque-scenario associated to (λ, χ) has the follow- ing feature: in the j-th move of the deque device, the last pj balls of the input pipe (with i=1 pi and n) are inserted into deque pipe from the left, and during the same move, the ball n(cid:13) goes into the output pipe. Thus the configuration of balls residing in the deque-pipe at time j is labels between 1 +Pj−1 n'(cid:13) n"(cid:13) · · · s(cid:13) x(cid:13) · · · y(cid:13) , (4.11) where n′ = n − 1, n′′ = n − 2, . . . , s = 1 +Pj−1 i=1 pi, and where " x(cid:13), . . . , y(cid:13)" is the (possibly empty) configuration of balls that were in the deque pipe at time j −1. Let us also note here that each of the remaining moves of the deque device ((j + 1)-th move up to n-th move) is either a left-0 move or a right-0 move, since the input pipe was emptied at the j-th move. Due to our assumption that j is a left-time, it is certain that Vj 6= ∅ (we have in any case that Vj ∋ q). But Wj may be empty, and we will discuss separately two cases. Case 1. Wj = ∅. In this case all the balls n'(cid:13), . . . , s(cid:13) exit the deque-pipe by its left side. Some of the balls x(cid:13), . . . , y(cid:13) may also exit the deque-pipe by its left side, but they can only do so after all of n'(cid:13), . . . , s(cid:13) are out of the way. This immediately implies that the times when n'(cid:13), . . . , s(cid:13) exit the deque-pipe must have consecutive 7 left-standings. It follows that in this case we have S = Vj = {q, q + 1, . . . , q + pj − 1}, and hence S is an interval-block of ρλ,χ. Case 2. Wj 6= ∅. In this case some of the balls n'(cid:13), . . . , s(cid:13) (at least one and at most pj − 1 of them) exit the deque-pipe by its right side. We observe it is not possible to find s ≤ a < b ≤ n − 1 such that the ball a(cid:13) exits the deque-pipe by its left side while b(cid:13) exits by the right-side. (Indeed, assume by contradiction that this would be the case. In the picture n'(cid:13) · · · b(cid:13) · · · a(cid:13) · · · s(cid:13) x(cid:13) · · · y(cid:13) , one of the two balls a(cid:13), b(cid:13) must be the first to exit the deque-pipe -- but that's not possible, since the other ball will block it.) As a consequence, there must exist a label c ∈ {s, . . . , n−1} such that the balls s(cid:13), . . . , c(cid:13) (i.e. the balls with labels in [s, c] ∩ Z) exit the deque-pipe by the right side, while the balls with labels in (c, n − 1] ∩ Z (if any) exit by the left side. We next observe that all the balls x(cid:13), . . . , y(cid:13) from the picture in (4.11) must exit the deque-pipe by its right side. This follows via the same kind of "blocking" argument as in 7 Note that the times themselves when n'(cid:13), . . . , s(cid:13) exit the deque-pipe don't have to be consecutive, because they may be interspersed with some right-times used by balls from x(cid:13), . . . , y(cid:13). The "consecutive" claim is only in reference to left-standings. 19 the preceding paragraph. (Say e.g. that x(cid:13) wants to exit by the left -- then out of the two balls s(cid:13) and x(cid:13), none can be the first to exit the deque-pipe, because it would be blocked by the other.) Based on the above tallying of how the balls from the picture in (4.11) exit the deque- pipe, a moment's thought shows that the set Wj ⊆ {1, . . . , v} must consist of the c − s + 1 largest numbers in {1, . . . , v} and that, likewise, the set Vj must be the sub-interval {q, . . . , u} of {1, . . . , u}. Then (n + 1) − Wj comes to {u + 1, . . . , u + (c − s + 1)}, and the union S = Vj ∪ ( (n + 1) − Wj ) is an interval-block of ρλ,χ, as required. (cid:4) Proposition 4.9. Let n be a positive integer, and let (λ, χ) be a pair in Luk(n) × {ℓ, r}n. The combined-standings partition ρλ,χ introduced in Definition 4.3 is in N C(n). Proof. We proceed by induction on n. The base case n = 1 is clear, so we focus on the induction step: we fix an integer n ≥ 2, we assume the statement of the proposition holds for pairs in Luk(m) × {ℓ, r}m whenever 1 ≤ m ≤ n − 1, and we prove that it also holds for pairs in Luk(n) × {ℓ, r}n. Let us then fix a pair (λ, χ) in Luk(n) × {ℓ, r}n, for which we will prove that ρλ,χ is in N C(n). We denote χ = (h1, . . . , hn), and we denote the rise-vector of λ as ~λ = (p1 − 1, . . . , pn − 1). Besides ρλ,χ, we will also work with the output-time partition π ∈ P(n) associated to (λ, χ), and we will use the same notations as earlier in the section: ρλ,χ = {Vi ∪ (n + 1) − Wi i ∈ I} and π = {Ti i ∈ I}, where I = {1 ≤ i ≤ n pi > 0}, the set of insertion times for (λ, χ). We will assume that I ≥ 2 (if I = 1 then clearly ρλ,χ = 1n ∈ N C(n)). Same as in Lemma 4.8, we put j := max(I); we thus have pj ≥ 1 and pj+1 = · · · = pn = 0. Let us put m := n − pj =Pj−1 there exists i < j with pi > 0), and also m < n (since pj > 0). We consider the m-tuple i=1 pi. Then m > 0 (because the assumption I ≥ 2 means χo := χ ( {1, . . . , n} \ Tj ) ∈ {ℓ, r}m (that is, χo = (ht1 , . . . , htm ), where one writes {1, . . . , n} \ Tj = {t1, . . . , tm} with t1 < · · · < tm). On the other hand, let us consider the Lukasiewicz path λo ∈ Luk(m) determined by the requirement that ~λo = (p1 − 1, . . . , pn − 1) ( {1, . . . , n} \ Tj ). It is easily seen that the combined-standings partition ρλo,χo ∈ P(m) associated to (λo, χo) is obtained from ρλ,χ ∈ P(n) by removing the block Vj ∪ ( (n + 1) − Wj ) of ρλ,χ, and then by re-naming the elements of the remaining blocks of ρλ,χ in increasing order. (Indeed, for this verification all one needs to do is ignore the last group of pj balls which moves through the pipes of the deque device, in the deque-scenario determined by (λ, χ).) Now, the block Vj ∪( (n+1)−Wj ) removed out of ρλ,χ is an interval-block, by Lemma 4.8. On the other hand, the partition ρλo,χo is in N C(m), due to our induction hypothesis. Thus the partition ρλ,χ ∈ P(n) is obtained via the insertion of an interval-block with pj(= n − m) elements into a partition from N C(m). This way of looking at ρλ,χ readily implies that ρλ,χ ∈ N C(n), and concludes the proof. (cid:4) 20 It is now easy to prove the main result of this section, which is stated as follows. Theorem 4.10. Let χ be a tuple in {ℓ, r}n, and let the set of partitions P (χ)(n) ⊆ P(n) be as in Definition 3.5. Then P (χ)(n) can also be obtained as P (χ)(n) =(cid:8)σχ · π π ∈ N C(n)(cid:9) ⊆ P(n), with σχ as in Definition 4.5. (4.12) Proof. We will show, equivalently, that (cid:8)σ−1 · π π ∈ P (χ)(n)} = N C(n). Since on both sides of the latter equality we have sets of the same cardinality Cn, it suffices to verify the inclusion "⊆". But "⊆" is clear from Lemma 4.6 and Proposition 4.9, since for π = Φχ(λ) with λ ∈ Luk(n) we get σ−1 (cid:4) · π = ρλ,χ ∈ N C(n). χ χ Corollary 4.11. Let n be a positive integer and let χ be a tuple in {ℓ, r}n. 1o P (χ)(n) contains the partitions 0n and 1n (from Notation 2.1.2), and also contains all the partitions π ∈ P(n) which have n − 1 blocks. 2o The bijection N C(n) ∋ π 7→ σχ · π ∈ P (χ)(n) from Theorem 4.10 is a poset iso- morphism, where on both N C(n) and P (χ)(n) we consider the partial order "≤" defined by reverse refinement. 3o (P (χ)(n), ≤) is a lattice. The meet operation "∧" of P (χ)(n) is described via block- intersections -- the blocks of π1 ∧ π2 are non-empty intersections V1 ∩ V2, with V1 ∈ π1 and V2 ∈ π2. Proof. 1o This follows from the fact that the set {0n, 1n} ∪ {π ∈ P(n) π has n − 1 blocks} is contained in N C(n) and is sent into itself by the action of σχ (no matter what the permutation σχ is). 2o This is an immediate consequence of the observation that the partial order by reverse refinement is preserved by the action of either σχ or σ−1 χ . 3o The fact that (P (χ)(n), ≤) is a lattice follows from 2o, since (N C(n), ≤) is a lattice. The description of the meet operation of P (χ)(n) holds because the meet operation of N C(n) is given by block-intersections, and because the action of σχ on partitions respects block-intersections. (cid:4) Remark 4.12. 1o For every positive integer n, the permutations associated to the (ℓ, r)- words (ℓ, . . . , ℓ) and (r, . . . , r) are σ(ℓ,...,ℓ) :=(cid:18) 1 2 · · · n 1 2 · · · n (cid:19) , σ(r,...,r) :=(cid:18) 1 1 (cid:19) . 2 · · · n n n − 1 · · · When plugged into Theorem 4.10, this gives P (ℓ,...,ℓ)(n) = P (r,...,r)(n) = N C(n), a fact that had already been noticed in Remark 3.7.2. 21 2o Say that n = 4 and that χ = (ℓ, r, ℓ, r), with associated partition σχ =(cid:18) 1 2 3 4 1 3 4 2 (cid:19) . Theorem 4.10 gives, via an easy calculation, that P (χ)(4) contains all the partitions of {1, 2, 3, 4} with the exception of { {1, 4} , {2, 3} } (in agreement with the description of this particular P (χ)(n) that was mentioned in section 1.2 of the introduction). In the sequel there will be instances when we will need to "read in reverse" a tuple from {ℓ, r}n. We conclude the section with an observation about that. Definition 4.13. For every n ≥ 1 and χ = (h1, . . . , hn) ∈ {ℓ, r}n, the tuple will be called the opposite of χ. χopp := (hn, . . . , h1) Proposition 4.14. Let n be a positive integer, let χ be a tuple in {ℓ, r}n, and consider the opposite tuple χopp. Then the sets of partitions P (χ)(n) and P (χopp )(n) are related by the formula P (χopp )(n) = {πopp π ∈ P (χ)(n)}, (4.13) where the opposite πopp of a partition π ∈ P(n) is as considered in Definition 2.1.4. Proof. Let τo be the order-reversing permutation of {1, . . . , n} that was considered in Defi- nition 2.1.4 (τo(m) = n + 1 − m for 1 ≤ m ≤ n). On the other hand let u ∈ {0, 1, . . . , n} be the number of occurrences of the letter ℓ in the word χ, and let us consider the permutation τu :=(cid:18) 1 u u − 1 · · · 1 n 2 · · · u u + 1 · · · n − 1 · · · u + 2 u + 1 (cid:19) . n (4.14) (Note that if u happens to be 0, then the permutation τ0 defined in (4.14) coincides, fortu- nately, with the permutation τo that had been considered above.) Let π be a partition in P(n) \ N C(n). Let V, W be two distinct blocks of π which cross, and let a < b < c < d be numbers such that a, c ∈ V and b, d ∈ W . We leave it as an exercise to the reader to check via a case-by-case discussion that the numbers τu(a), τu(b), τu(c), τu(d) ∈ {1, . . . , n} (despite not being necessarily in increasing order) ensure the existence of a crossing between the blocks τu(V ) and τu(W ) of the partition τu · π ∈ P(n). The argument in the preceding paragraph shows that {τu · π π ∈ P(n) \ N C(n)} ⊆ P(n) \ N C(n). A cardinality argument forces the latter inclusion to be an equality, and then from the fact that τu sends P(n) bijectively onto itself it also follows that we have {τu · π π ∈ N C(n)} = N C(n). (4.15) 22 Now let us consider the positions of the letters ℓ and r in the words χ and χopp. By tallying these positions and plugging them into the formulas for the permutations σχ and σχopp (as in Definition 4.5), one immediately finds that σχopp = τo σχ τu. (4.16) So then we can write: P (χopp )(n) = {σχopp · π π ∈ N C(n)} (by Theorem 4.10) = {τoσχτu · π π ∈ N C(n)} (by Eqn.(4.16) = {τoσχ · π′ π′ ∈ N C(n)} (by Eqn.(4.15) = {τo · π′′ π′′ ∈ P (χ)(n))} (by Theorem 4.10), and this establishes the required formula (4.13). (cid:4) 5. (ℓ, r)-cumulant functionals In this section we introduce the family of (ℓ, r)-cumulant functionals associated to a non- commutative probability space. In order to write in a more compressed way the summation formula defining these functionals, we first introduce a notation. Notation 5.1. [Restrictions of n-tuples.] Let X be a non-empty set, let n be a positive integer, and let (x1, . . . , xn) be an n-tuple in X n. For a subset V = {i1, . . . , im} ⊆ {1, . . . , n}, with 1 ≤ m ≤ n and 1 ≤ i1 < · · · < im ≤ n, we will denote (x1, . . . , xn) V := (xi1, . . . , xim) ∈ X m. The next definition uses this notation in two ways: • for X = A (algebra of noncommutative random variables); • for X = {ℓ, r}, when we talk about the restriction χ V of a tuple χ ∈ {ℓ, r}n. Proposition and Definition 5.2. [(ℓ, r)-cumulants.] Let (A, ϕ) be a nocommutative probability space. There exists a family of multilinear func- tionals (cid:16) κχ : An → C(cid:17)n≥1, χ∈{ℓ,r}n which is uniquely determined by the requirement that ϕ(a1 · · · an) =Pπ∈P (χ)(n)(cid:16) QV ∈π κχV ( (a1, . . . , an) V ) (cid:17), for every n ≥ 1, χ ∈ {ℓ, r}n and a1, . . . , an ∈ A. (5.1)   These κχ's will be called the (ℓ, r)-cumulant functionals of (A, ϕ). 23 Proof. For n = 1 we define κ(ℓ) = κ(r) = ϕ. We then proceed recursively, where for every n ≥ 2, every χ ∈ {ℓ, r}n and every a1, . . . , an ∈ A we put κχ(a1, . . . , an) = ϕ(a1 · · · an) − Xπ∈P (χ)(n) (cid:16) YV ∈π κχV ( (a1, . . . , an) V ) (cid:17). (5.2) π6=1n It is immediate that (5.2) defines indeed a family of multilinear functionals which fulfil (5.1). The uniqueness part of the proposition is also immediate, by following the (obligatory) recursion (5.2). (cid:4) Remark 5.3. Let (A, ϕ) be a noncommutative probability space, let (κn)∞ of free cumulant functionals of (A, ϕ), and let (cid:0) κχ : An → C(cid:1)n≥1, χ∈{ℓ,r}n be the family of (ℓ, r)-cumulant functionals introduced in Definition 5.2. n=1 be the family 1o As noticed in Remark 3.7.2, one has P (ℓ,...,ℓ)(n) = P (r,...,r)(n) = N C(n). By plugging this fact into the recursion (5.2) which characterizes the functionals κχ, one immediately obtains the fact (already advertised in the introduction) that κ( ℓ, . . . , ℓ ) = κ( r, . . . , r ) = κn, ∀ n ≥ 1. {z }n n {z } 2o If n ≤ 3, then we actually have κχ = κn for every χ ∈ {ℓ, r}n. This comes from the fact, observed in Remark 3.7.3, that P (χ)(n) = N C(n) when n ≤ 3, no matter what χ ∈ {ℓ, r}n we consider. 3o For n ≥ 4, the functionals κχ with χ ∈ {ℓ, r}n are generally different from κn. Say for instance that χ = (ℓ, r, ℓ, r) ∈ {ℓ, r}4, then the difference between the lattices N C(4) and P (χ)(4) leads to the fact that for a1, . . . , a4 ∈ A we have κ(ℓ,r,ℓ,r)(a1, . . . , a4) = κ4(a1, . . . , a4) +κ2(a1, a4)κ2(a2, a3) − κ2(a1, a3)κ2(a2, a4). Remark 5.4. Let us also record here a formula, concerning (ℓ, r)-cumulants, which is related to the reading of (ℓ, r)-words in reverse (i.e. to looking at χ versus χopp, as in Definition 4.13 and Proposition 4.14). Suppose that (A, ϕ) is a ∗-probability space. Then, with(cid:0) κχ : An → C(cid:1)n≥1, χ∈{ℓ,r}n denoting the family of (ℓ, r)-cumulant functionals of (A, ϕ), one has κχ(a∗ 1, . . . , a∗ n) = κχopp (an, . . . , a1), for every n ≥ 1, χ ∈ {ℓ, r}n and a1, . . . , an ∈ A. (5.3)   The verification of (5.3) is easily done by induction on n, where one relies on the bijections P (χ)(n) ∋ π 7→ πopp ∈ P (χopp )(n), for n ≥ 1 and χ ∈ {ℓ, r}n, that were observed in Proposition 4.14. (The proof of the induction step starts, of course, by writing that ϕ(a∗ n) and ϕ(an · · · a1) is then expanded into (ℓ, r)-cumulants, in the way described in Definition 5.2.) n) = ϕ(an · · · a1); each of the moments ϕ(a∗ 1 · · · a∗ 1 · · · a∗ 24 6. (ℓ, r)-cumulants of canonical operators In this section we prove the theorem announced in section 1.4 of the Introduction. We will adopt the framework and notations of the theorem -- so we are dealing with the d-tuples (A1, . . . , Ad) and (B1, . . . , Bd) of left and respectively right canonical operators on Td, which were defined in Equations (1.3) -- (1.6) of section 1.2 by starting from two non-commutative polynomials f (z1, . . . , zd) and g(z1, . . . , zd). Recall that the coefficients of zi1 · · · zin in the polynomials f and g are denoted as α(i1,...,in) and as β(i1,...,in), respectively. In the formula claimed by the theorem we used the unified notation Ai =: Ci;ℓ and Bi =: Ci;r, for 1 ≤ i ≤ d. (6.1) In order to give a concise re-statement of that formula, let us also introduce a unified notation for the relevant coefficients α and β, as follows. Definition 6.1. [Bi-words and bi-mixtures of coefficients.] Let n be a positive integer. 1o The elements of the set {1, . . . , d}n × {ℓ, r}n will be called bi-words of length n. 2o Let (ω; χ) be a bi-word of length n, where ω = (i1, . . . , in) ∈ {1, . . . , d}n and χ = (h1, . . . , hn) ∈ {ℓ, r}n. We denote γ(ω; χ) :=  α(imr (v),...,imr (1),imℓ(1),...,imℓ(u)), if hn = ℓ, β(imℓ (u),...,imℓ(1),imr (1),...,imr (v)), if hn = r, (6.2) where mℓ(1) < · · · < mℓ(u) and mr(1) < · · · < mr(v) record the lists of occurrences of ℓ and of r in χ (same convention of notation as in Definition 4.1). We will refer to γ(ω; χ) as the bi-mixture of α's and β's corresponding to the bi-word (ω; χ). The result we want to prove can then be stated as follows. Theorem 6.2. For every n ≥ 1 and every χ = (h1, . . . , hn) ∈ {ℓ, r}n, ω = (i1, . . . , in) ∈ {1, . . . , d}n, one has κχ(Ci1;h1, . . . , Cin;hn) = γ(ω; χ). (6.3) The remaining part of the section is devoted to the proof of Theorem 6.2. The proof will go by formalizing, in Lemma 6.6 below, the intuitive idea that the action of A1, . . . , Ad, B1, . . . , Bd on the vacuum vector ξvac ∈ Td is closely related to the deque-scenarios from Section 3 of the paper. In order to state Lemma 6.6, we need the concept (related to the one from Definition 6.1.2) of what is a "reverse-bi-mixture" of coefficients α and β. 25 Definition 6.3. Let n be a positive integer and let (ω; χ) be a bi-word of length n, where ω = (i1, . . . , in) and χ = (h1, . . . , hn). We will denote eγ(ω; χ) :=  α(imr (1),...,imr (v),imℓ(u),...,imℓ(1)), if h1 = ℓ β(imℓ (1),...,imℓ(u),imr (v),...,imr (1)), if h1 = r, (6.4) where mℓ(1) < · · · < mℓ(u) and mr(1) < · · · < mr(v) record the lists of occurrences of ℓ and of r in χ (same convention of notation as in Definition 4.1 and in Definition 6.1). We will refer to eγ(ω; χ) as the reverse-bi-mixture of α's and β's corresponding to the bi-word (ω; χ). Remark 6.4. It is obvious that the reverse-bi-mixtures which were just introduced are related to the bi-mixtures from Definition 6.1 by the formula γ(ω; χ) =eγ(ωopp; χopp), where for χ = (h1, . . . , hn) and ω = (i1, . . . , in) we put χopp := (hn, . . . , h1) (same as in Definition 4.13) and ωopp := (in, . . . , i1). Let us also record an immediate extension of Equation (6.5), namely that for every non-empty set T ⊆ {1, . . . , n} we have γ((ω; χ) T ) =eγ( (ωopp ; χopp) (n + 1) − T ), with (n + 1) − T := {n + 1 − t t ∈ T }. [The restrictions of bi-words that have appeared in Equation (6.6) are defined by the same convention as used in Notation 5.1 -- e.g. we have (6.5) (6.6) (ω; χ) T := (ω T ; χ T ) ∈ {1, . . . , d}m × {ℓ, r}m, where m is the number of elements of T .] In the statement of Lemma 6.6 we will also use the following notation. Notation 6.5. We denote X0;ℓ = X0;r = I (identity operator);   Xp;ℓ =Pd Xp;r =Pd i1,...,ip=1 α(i1,...,ip)Lip · · · Li1, i1,...,ip=1 β(i1,...,ip)Rip · · · Ri1, for p ≥ 1; for p ≥ 1. (6.7) The canonical operators Ai, Bi that we are dealing with can then be written as Ai = L∗ i ∞Xp=0 Xp;ℓ, Bi = R∗ i ∞Xp=0 Xp;r, for 1 ≤ i ≤ d. (6.8) (The sums in (6.8) are actually finite, since Xp;ℓ = Xp;r = 0 for p large enough.) 26 It will be convenient to use a "unified left-right notation" of the Equations (6.8), as follows. We already have a unified notation for Ai and Bi (the Ci;h from Equation (6.1)), and let us also denote Li =: Si,ℓ, Ri =: Si,r, for 1 ≤ i ≤ d. Then (6.8) can be put in the form Ci;h = S∗ i;h ∞Xp=0 Xp;h, for 1 ≤ i ≤ d and h ∈ {ℓ, r}. (6.9) Lemma 6.6. Let n be a positive integer, and consider the following items: • an n-tuple ω = (i1, . . . , in) ∈ {1, . . . , d}n; • an n-tuple χ = (h1, . . . , hn) ∈ {ℓ, r}n; • a Lukasiewicz path λ ∈ Luk(n) with rise-vector denoted as ~λ = (p1 − 1, . . . , pn − 1), where p1, . . . , pn ∈ N ∪ {0}. Let π ∈ P (χ)(n) be the output-time partition associated to (λ, χ) in Definition 3.3. Then we have X ∗ p1;h1Si1;h1 · · · X ∗ pn;hnSin;hnξvac = c ξvac, where c = YT ∈πeγ( (ω; χ) T ). (6.10) In Equation (6.10), the operators Xp;h and Si;h are as in Notation 6.5, and the coefficients eγ are reverse-bi-mixtures, as in Definition 6.3. Proof. Let {j1 < j2 < . . . < jt} = {ipi > 0} let π = {Tj1, . . . , Tjt}, where for r = 1, . . . , t, we have that Tjr denotes the block of the output-time partition corresponding to time jr, i.e., the block whose minimal element is jr. We abbreviate k = jt. We proceed by induction on t. We first deal with the base case. If t = 1 then we must have k = 1, p1 = pk = n, p2 = . . . = pn = 0, and π = {T1} = {{1, . . . , n}}. If we denote {mℓ(1) < . . . mℓ(u)} = {ihi = ℓ} and {mr(1) < . . . < mr(v)} = {ihi = r} as in Definition 6.3, then we have X ∗ p1;h1Si1;h1 · · · X ∗ = X ∗ = X ∗ pn;hnSin;hnξvac n;h1 (Si1;h1 · · · Sin;hnξvac) n;h1emℓ(1) ⊗ . . . ⊗ emℓ(u) ⊗ emr(v) ⊗ . . . ⊗ emr (1) = eγ ((ω, χ)T1)ξvac = cξvac. Now assume that t > 1 and that the conclusion of the lemma holds for all smaller values of t. Let denote the unique increasing bijection. We abbreviate f : {1, 2, . . . , n − pk} → {1, . . . , n} \ Tk bω =(cid:0)if (1), . . . , if (n−pk)(cid:1) , bχ =(cid:0)hf (1), . . . , hf (n−pk)(cid:1) , 27 and we also denote ~v :=(cid:0)pf (1)−1, . . . , pf (n−pk)−1(cid:1). Letbλ be the Lukasiewicz path associated to ~v, and let bπ ∈ P ( bχ)(n − pk) be the output-time partition associated to (bλ,bχ). We now note, as is implicit in the discussions in Sections 3 and 4, that Details of this observation are left to the reader. Observe now that by the induction hy- pothesis we have f (bπ) = {T1, . . . , Tjt−1}. X ∗ pf (1),hf (1) Sif (1);hf (1) · · · X ∗ pf (n−pk ),hf (n−pk ) Sif (n−pk );hf (n−pk) =bcξvac, where bc = YT ∈bπeγ( (bω;bχ) T ) = YT ∈π,T 6=Tkeγ( (ω; χ) T ). Let us list elements of the set {k, . . . , n} as dk, . . . , dn by listing left elements first in the increasing order followed by the right elements in the decreasing order, i.e., the order in the list dk, . . . , dn respects the order from the list mℓ(1), . . . , mℓ(u), mr(v), . . . , mr(1). Now note that we have pn;hnSin;hnξvac X ∗ pk;hk Sik;hk · · · X ∗ = X ∗ = X ∗ (Sik;hk · · · Sin;hnξvac) = pk;hk pk;hk edk ⊗ . . . ⊗ edn αdn,...,dk ξvac βdk ,...,dnξvac αdk+pk −1,...,dk edk+pk βdn+1−pk   = eγ ((ω, χ)Tk)Sif (k);df (k) · · · Sif (n−pk);df (n−pk )ξvac. ⊗ . . . ⊗ edn ,...,dnedk ⊗ . . . ⊗ edn−pk , if pk = n − k + 1 and hk = ℓ , if pk = n − k + 1 and hk = r , if pk < n − k + 1 and hk = ℓ , if pk < n − k + 1 and hk = r Hence we have X ∗ p1;h1Si1;h1 · · · X ∗ = X ∗ p1;h1Si1;h1 · · · X ∗ pk−1;hk−1Sik−1;hk−1X ∗ pk;hkSik;hk · · · X ∗ pn;hnSin;hnξvac pk−1;hk−1(cid:16)eγ ((ω, χ)Tk) Sif (k);df (k) · · · Sif (n−pk );df (n−pk )ξvac(cid:17) pk−1;hk−1Sif (k);df (k) · · · Sif (n−pk );df (n−pk )ξvac pf (n−pk );hf (n−pk ) Sif (n−pk );hf (n−pk) pf (1);hf (1) Sif (1);hf (1) · · · X ∗ p1;h1Si1;h1 · · · X ∗ = eγ ((ω, χ)Tk)X ∗ = eγ ((ω, χ)Tk)X ∗ = eγ ((ω, χ)Tk) · YT ∈π,T 6=Tkeγ( (ω; χ) T )ξvac = YT ∈πeγ( (ω; χ) T )ξvac = cξvac. This concludes the induction step. (cid:4) Example 6.7. For clarity, let us follow the preceding lemma in the concrete case (also discussed earlier, in Examples 3.4.1 and 4.7) where n = 5, χ = (r, ℓ, ℓ, r, ℓ), and λ ∈ Luk(5) has rise-vector ~λ = (2, −1, 1, −1, −1). As found in Example 3.4.1, the output-time partition 28 associated to this (λ, χ) is π = { {1, 2, 4}, {3, 5} }. Let us also fix a tuple ω = (i1, . . . , i5) ∈ {1, . . . , d}5. We have eγ(cid:0) (ω; χ) {1, 2, 4}(cid:1) =eγ(cid:0) (i1, i2, i4); (r, ℓ, r)(cid:1) = β(i2,i4,i1) and eγ(cid:0) (ω; χ) {3, 5}(cid:1) = eγ(cid:0) (i3, i5); (ℓ, ℓ)(cid:1) = α(i5,i3). The constant c from Equation (6.10) is thus c = β(i2,i4,i1) α(i5,i3), and the formula claimed by the lemma should come to X ∗ 3;rRi1X ∗ 0;ℓLi2X ∗ 2;ℓLi3X ∗ 0;rRi4X ∗ 0;ℓLi5ξvac = c ξvac for this particular value of c. And indeed, let us record how ξvac travels when we apply to it the operators listed on the left-hand side of the above equation: we get ξvac 7→ Li5ξvac = ei5 7→ Ri4ei5 = ei5 ⊗ ei4 7→ X ∗ 7→ Li2(αi5,i3 ei4) = αi5,i3 ei2 ⊗ ei4 7→ X ∗ 2;ℓLi3(ei5 ⊗ ei4) = X ∗ 3;rRi1(αi5,i3 ei2 ⊗ ei4) = αi5,i3 X ∗ 2;ℓ ei3 ⊗ ei5 ⊗ ei4 = αi5,i3 ei4 3;r (ei2 ⊗ ei4 ⊗ ei1) = αi5,i3 · βi2,i4,i1 ξvac, as claimed. Proposition 6.8. Let n be a positive integer and let (ω; χ) be a bi-word of length n, where ω = (i1, . . . , in) and χ = (h1, . . . , hn). We have ϕvac(Ci1;h1 · · · Cin;hn) = Xπ∈P (χ)(n)(cid:16)YT ∈π γ( (ω; χ) T )(cid:17). (6.11) where the bi-mixtures "γ" on the right-hand side of the equation are as introduced in Defi- nition 6.1. Proof. Write each of Ci1;h1, . . . , Cin;hn as a sum in the way indicated in Equation (6.9) of Notation 6.5, then expand the ensuing product of sums; we get ϕvac(Ci1;h1 · · · Cin;hn) = ∞Xp1,...,pn=0 term(p1,...,pn), (6.12) where for every p1, . . . , pn ∈ N ∪ {0} we put term(p1,...,pn) := ϕvac(S∗ i1;h1 Xp1;h1 · · · S∗ in;hnXpn;hn) (6.13) = hS∗ i1;h1 Xp1;h1 · · · S∗ in;hn Xpn;hn ξvac , ξvaci. We will proceed by examining what n-tuples (p1, . . . , pn) ∈ (N ∪ {0})n may contribute a non-zero term in the sum from (6.12). So let p1, . . . , pn be in N ∪ {0}. We make the following observations. • If there exists m ∈ {1, . . . , n} with pm + · · · + pn < (n + 1) − m, then term(p1,...,pn) = 0. Indeed, if such an m exists then it is immediately seen that im;hmXpm;hm · · · S∗ S∗ in;hnXpn;hn ξvac = 0, which makes the inner product from (6.13) vanish. 29 • If p1 + · · · + pn > n, then term(p1,...,pn) = 0. Indeed, in this case the vector S∗ Xpn;hn ξvac is seen to belong to the subspace Xp1;h1 · · · S∗ in;hn i1;h1 span{ej1 ⊗ · · · ⊗ ejq 1 ≤ j1, . . . , jq ≤ d} ⊆ Td, where q = (p1 + · · · + pn) − n > 0. The latter subspace is orthogonal to ξvac, and this again makes the inner product from (6.13) vanish. The observations made in the preceding paragraph show that a necessary condition for term(p1,...,pn) 6= 0 is that pm + · · · + pn ≥ (n + 1) − m, ∀ 1 ≤ m ≤ n, where for m = 1 we must have p1 + · · · + pn = n.   This says precisely that the tuple (pn − 1, . . . , p1 − 1) is the rise-vector of a uniquely de- termined path λ ∈ Luk(n). Hence the sum on the right-hand side of (6.12) is in fact, in a natural way, indexed by Luk(n). Now let us fix a path λ ∈ Luk(n), where (consistent to the above) we denote the rise- vector of λ as ~λ := (pn − 1, . . . , p1 − 1). If we put then Equation (6.13) can be re-written in the form epm := pn+1−m, ehm := hn+1−m, eim := in+1−m, 1 ≤ m ≤ n, term(p1,...,pn) = hξvac , X ∗ ep1;eh1 Sei1;eh1 · · · X ∗ epn;ehn Sein;ehn ξvac i, where on the right-hand side we are in the position to invoke Lemma 6.6. The lemma must be used in connection to the path λ and the tuples χopp = (eh1, . . . ,ehn), ωopp = (ei1, . . . ,ein). If we also denote the application of Lemma 6.6 takes us to: eπ := Φχopp (λ) (output-time partition associated to λ and χopp), term(p1,...,pn) = YeT ∈eπ eγ( (ωopp ; χopp) eT ). Finally, we note that when eT runs among the blocks of eπ, the set (n + 1) − eT runs among the blocks of the opposite partition eπopp. Thus, in view of the relation between γ's and eγ's observed in Remark 6.4, we arrive to the formula term(p1,...,pn) = γ( (ω; χ) T ). YT ∈( Φχopp (λ) )opp The overall conclusion of the above discussion is that we have ϕvac(Ci1;h1 · · · Cin;hn) = Xλ∈Luk(n) YT ∈( Φχopp (λ) )opp γ( (ω; χ) T ). The only thing left to verify is, then, that the set of partitions (cid:8)(cid:0) Φχopp (λ)(cid:1)opp λ ∈ Luk(n)(cid:9) coincides with P (χ)(n). But this is indeed true, since {Φχopp (λ) λ ∈ Luk(n)} = P (χopp )(n) (by the definition of P (χopp )(n)), and in view of Proposition 4.14. (cid:4) 30 6.9. Proof of Theorem 6.2. We verify the required formula (6.3) by induction on n. For n = 1 we only have to observe that κ(ℓ)(Ai) = γ( (i); (ℓ) ), ∀ 1 ≤ i ≤ d (both the above quantities are equal to α(i)), and that κ(r)(Bi) = γ( (i); (r) ), ∀ 1 ≤ i ≤ d (both quantities equal to β(i)). Induction step: consider an n ≥ 2, suppose the equality in (6.3) has already been verified for all bi-words of length ≤ n − 1, and let us fix a bi-word (ω; χ) of length n, for which we want to verify it as well. Write explicitly ω = (i1, . . . , in) and χ = (h1, . . . , hn), with 1 ≤ i1, . . . , in ≤ d and h1, . . . , hn ∈ {ℓ, r}. The joint moment ϕvac(Ci1;h1 · · · Cin;hn) can be expressed as a sum over P (χ)(n) in two ways: on the one hand we have it written as in Equation (6.11) of Proposition 6.8, and on the other hand we can write it by using the moment↔cumulant formula (5.1) which was used to introduce the (ℓ, r)-cumulants in Definition 5.2: ϕvac(Ci1;h1 · · · Cin;hn) = Xπ∈P (χ)(n)(cid:16) YV ∈π κχV ( (Ci1;h1, . . . , Cin;hn) V ) (cid:17). (6.14) The induction hypothesis immediately gives us that, for every π 6= 1n in P (χ)(n), the term indexed by π in the two summations that were just mentioned (right-hand side of (6.11) and right-hand side of (6.14)) are equal to each other. When we equate these two summations and cancel all the terms indexed by π 6= 1n in P (χ)(n), we are left precisely with κχ(Ci1;h1, . . . , Cin;hn) = γ(ω; χ), as required. (cid:4) Acknowledgements This research work was started while the authors were participating in the focus program on free probability at the Fields Institute in Toronto, in July 2013. The uplifting atmosphere and the support of the Fields focus program are gratefully acknowledged. We also express our thanks to the anonymous referee who pointed to us the importance of re-writing the introduction in a way which better shows the motivation of the paper. References [1] I. Charlesworth, B. Nelson, P. Skoufranis. On two-faced families of non-commutative random variables. Preprint, March 2014, available at arxiv.org/abs/1403.4907. [2] D.E. Knuth. The Art of Computer Programming, Volume 1: Fundamental Algorithms, 2nd edition, Addison-Wesley, 1973. [3] A. Nica. R-transforms of free joint distributions and non-crossing partitions, Journal of Functional Analysis 135 (1996), 271 -- 296. 31 [4] A. Nica, R. Speicher. Lectures on the combinatorics of free probability, London Math- ematical Society Lecture Note Series 335, Cambridge University Press, 2006. [5] R. Speicher. Multiplicative functions on the lattice of noncrossing partitions and free convolution, Mathematische Annalen 298 (1994), 611 -- 628. [6] D. Voiculescu. Symmetries of some reduced free product C ∗-algebras, in Operator Alge- bras and Their Connections with Topology and Ergodic Theory (H. Araki, C.C. Moore, S. Stratila and D. Voiculescu, editors), Springer Lecture Notes in Mathematics Volume 1132, Springer Verlag, 1985, pp. 556-588. [7] D. Voiculescu. Addition of certain noncommuting random variables, Journal of Func- tional Analysis 66 (1986), 323 -- 346. [8] D. Voiculescu. Free probability for pairs of faces I, Communications in Mathematical Physics 332 (2014), 955-980. [9] D. Voiculescu. Free probability for pairs of faces II: 2-variables bi-free partial R- transform and systems with rank ≤ 1 commutation. Preprint, August 2013, available at arxiv.org/abs/1308.2035. Mitja Mastnak Department of Mathematics and Computing Science, Saint Mary's University, Halifax, Nova Scotia B3H 3C3, Canada. Email: [email protected] Alexandru Nica Department of Pure Mathematics, University of Waterloo, Waterloo, Ontario N2L 3G1, Canada. Email: [email protected] 32
1305.2843	2	1305	2013-11-07T17:14:33	Irreducible Induced Representations of Fell Bundle C*-Algebras	[ "math.OA" ]	We give precise conditions under which irreducible representations associated to stability groups induce to irreducible representations for Fell bundle C-algebras. This result generalizes an earlier result of Echterhoff and the second author. Because the Fell bundle construction subsumes most other examples of C-algebras constructed from dynamical systems, our result percolates down to many different constructions including the many flavors of groupoid crossed products.	math.OA	math	IRREDUCIBLE INDUCED REPRESENTATIONS OF FELL BUNDLE C ∗-ALGEBRAS MARIUS IONESCU AND DANA P. WILLIAMS Abstract. We give precise conditions under which irreducible representations associated to stability groups induce to irreducible representations for Fell bundle C ∗-algebras. This result generalizes an earlier result of Echterhoff and the second author. Because the Fell bundle construction subsumes most other examples of C ∗-algebras constructed from dynamical systems, our result percolates down to many different constructions including the many flavors of groupoid crossed products. . A O h t a m [ 2 v 3 4 8 2 . 5 0 3 1 : v i X r a 1. Introduction One of the fundamental tasks in any study of the ideal structure of C∗-algebras associated to dynamical systems is to construct a suitably large class of irreducible representations. In the type I case, it makes sense to try to find representatives of all equivalence classes of irreducible representations (for example, see [20, Theo- rem 8.16]). In general, in the presence of suitable amenability, it is only reasonable to try to construct enough irreducible representations to account for all primitive ideals. The quintessential example is the Gootman-Rosenberg-Sauvagoet Theorem (see [20, §8.3] for a precise statement and further references): the GRS-Theorem says that for a separable C∗-dynamical system α : G → Aut A with G amenable, every primitive ideal in A ⋊α G is induced in an appropriate sense from a stability group GP = { s ∈ G : s · P = P } with respect to the induced action of G on Prim A for some P ∈ Prim A. Motivated in part by the GRS-Theorem and by results in the case where the ac- tion of G on Prim A is smooth, Echterhoff and the second author have conjectured that every separable C∗-dynamical system (A, G, α) satisfies the Effros-Hahn In- duction Property (EHI) which asserts that if P ∈ Prim A and J ∈ Prim(A ⋊α GP ) with Res J = P , then IndG (See [2, §2] for precise definitions and additional details.) Although the validity of the conjecture is open in general -- even when G is amenable and the GRS Theorem holds -- it was shown in [2] to hold in a wide variety of cases including all separable sys- tems with A of type I. Moreover, in all cases in which the conjecture is known to hold, a stronger property holds, called strong-EHI, which asserts that if ρ ⋊ π is an irreducible representation of A ⋊α GP with ker ρ = P , then IndG GP (ρ ⋊ π) is an GP J is a primitive ideal in A ⋊α G. Date: 3 May 2013. 2000 Mathematics Subject Classification. 46L05; 46L55. Key words and phrases. Fell bundle, irreducible representation, ideal structure, Fell bundle C ∗-algebra. Both Authors were supported by individual grants from the Simons Foundation. Dana would also like to thank Marius and his colleagues at Colgate for a very pleasant and productive visit. 1 2 MARIUS IONESCU AND DANA P. WILLIAMS irreducible representation of A ⋊α G. The key observation in [2] concerning group C∗-dynamical systems is that strong-EHI always holds if, in addition, ρ is assumed to be a homogeneous representation (as defined in, for example, [20, Definition G.1]) -- this is the content of [2, Theorem 1.7]. Our goal here is to extend the results on inducing irreducible representations in [2], and [2, Theorem 1.7] in particular, to other sorts of dynamical system con- structions built not only on groups but on groupoids. At first glance, there are a horrifying number of potential targets for such an analysis. For example, there are groupoid C∗-algebras with or without a cocycle, and more generally, one could con- sider the C∗-algebras associated to twists over groupoids (also called T-groupoids). There are also Green twisted dynamical systems, groupoid dynamical systems and even twisted versions of groupoid dynamical systems to name a few of the most important. Fortunately, as described in detail in [9, §3] or [10, §2], all these vari- ants are subsumed using the C∗-algebra of a separable Fell bundle p : B → G over a locally compact groupoid G with a Haar system. In this event, the sec- tions A = Γ0(G(0); B) form a C∗-algebra and the groupoid G acts continuously on Prim A. Any representation L of C∗(GP , B) is associated to a representation π of the C∗-algebra A. Our main theorem (Theorem 4.1) says that if L is irreducible, ker π = P and π is homogeneous, then IndG GP L is irreducible. This result extends [2] and we will illustrate how it "trickles down" to other dynamical systems settings in Section 6. Our proof requires an intermediate result which is of considerable interest on its own. Namely if p : B → G is a separable Fell bundle over a locally compact groupoid G with Haar system, then we show that if u ∈ G(0), if G(u) = { x ∈ G : r(x) = u = s(x) } is the stability group of u in G and if L is an irreducible representation of C∗(G(u), B), then IndG G(u) L is an irreducible representation of C∗(G, B) (Theorem 2.1). This result is a direct generalization of [6, Theorem 5] where the result is proved for groupoid C∗-algebras (so that B is the trivial bundle B = G × C). In fact the proof is disarmingly similar to that in [6], but extra care must be taken to account for the rather significant difference between scalar-valued sections of a trivial bundle and Banach space-valued sections of potentially highly nontrivial Fell bundles. Combining Theorem 2.1 with the usual induction in stages allows us to reduce the proof of our main theorem to the more comforting setting of a Fell bundle over a group (rather than a groupoid). Our paper is organized as follows. We start in Section 2 with a very brief review of induced representations of Fell bundle C∗-algebras and prove our generalization, Theorem 2.1, of [6, Theorem 5]. In Section 3 we give the precise definition of the strong Effros-Hahn Induction property in the Fell bundle setting. In Section 4, we give our proof of the Main Theorem taking advantage of induction in stages and Theorem 2.1 to reduce to the case that G is a group. In Section 5 we see that the additional hypothesis of homogeneity is automatically satisfied in the case that points are locally closed in the C∗-algebra A = Γ0(G(0); B) associated to the Fell bundle p : B → G. While this includes many interesting classes of algebras, it in particular applies any time A is of type I. In Section 6 we examine how the Fell bundle result applies to the examples of groupoid dynamical systems and their twisted counterparts. It is worth noting that special cases of the latter include Green twisted systems in the case where G is a group, and the C∗-algebras of twists or T-goupoids when B is a trivial line bundle. IRREDUCIBLE REPRESENTATIONS 3 Assumptions. Throughout, p : B → G will be a saturated, separable Fell bundle over a locally compact groupoid G as defined in [10]. Thus p : B → G is an upper semicontinuous Banach bundle over a second countable locally compact groupoid G such that its continuous sections, vanishing at infinity on G, Γ0(G; B), form a separable Banach space with respect to the supremum norm. Furthermore, all our groupoids are assumed to be second countable, locally compact and Hausdorff. When G is a groupoid, it will be assumed to have a Haar system {λu}u∈G(0). We will write A = Γ0(G(0); B) for the C∗-algebra of B over G(0). We then follow [10, §1] to make the compactly supported continuous sections Γc(G; B) into a ∗- algebra with C∗-completion C∗(G, B). When dealing with any sort of Banach bundle p : B → X, we will use a roman font, B(x), to indicated the fibre over x ∈ X together with its Banach space structure. When we have the need to work with a Fell bundle over a group -- in particular, when we restrict a Fell bundle over a groupoid to a subgroup -- we will, for the sake of consistency, treat the underlying group as a groupoid as regards our conventions with modular functions.1 (See [8, §1.5] for an elaboration on this.) 2. Inducing from G(u) In this section we prove the following generalization of [6, Theorem 5] to Fell bundle C∗-algebras. Theorem 2.1. Suppose that p : B → G is a separable Fell bundle over a locally compact groupoid with a Haar system. Let u ∈ G(0) and let G(u) := { x ∈ G : r(x) = u = s(x) } be the stability group at u. Suppose that L is an irreducible representation of C∗(G(u), B). Then IndG G(u) L is an irreducible representation of C∗(G, B). Our proof of Theorem 2.1 follows that of [6, Theorem 5] very closely. Because the notation for the convolution of sections, inner products and actions is virtually identical to the scalar-valued case, there are parts where the proof can be used mutatis mutandis from [6]. While this is one of the benefits of the Fell bundle formalism, Theorem 2.1 is a highly nontrivial generalization of the scalar version, and we have tried to be careful below to point out the places where we have had to adjust from working with scalar-valued functions to sections of a nontrivial Banach bundle. At the same time, it seemed prudent to retain enough of the original argument from the scalar case that the exposition remains readable. 2.1. Induced Representations of Fell Bundle C ∗-Algebras. We begin by recalling the construction of induced representations for Fell bundles over groupoids from [19, §4.1]. Let q : B → G be a separable Fell bundle and assume that H is a closed subgroupoid of G. Let qH : BH → H be the Fell bundle obtained by restriction to H. Then GH(0) = s−1(H (0)) is an (H G, H)-equivalence, where H G is the imprimitivity groupoid (GH(0) ∗s GH(0) )/H. If σ : H G → G is the continuous map given by σ([x, y]) = xy−1, the pull-back Fell bundle σ∗q : σ∗B → H G is the Fell bundle σ∗B = { ([x, y], b) : [x, y] ∈ H G, b ∈ B, σ([x, y]) = q(b) } with bundle map σ∗q([x, y], b) = [x, y]. Then E = q−1(GH(0) ) is a σ∗B − BH -equivalence with 1The issue is that in the convolution algebra, the involution for groupoids has no modular function (since groupoids don't have modular functions until a quasi-invariant measure is picked). The modular function then reappears in the integrated forms of representations. 4 MARIUS IONESCU AND DANA P. WILLIAMS the left action of σ∗B given by ([x, y], b) · e = be if q(e) = yh, the right action of BH given by e · b = eb, and the left and right inner products on E ∗s E given by he , f i = ([q(e), q(f )], ef ∗) and he , f i = e∗f. σ∗B BH Therefore Γc(GH(0) ; E ) is a pre-imprimitivity bimodule with actions and inner prod- ucts determined by F · ϕ(z) = ZG ϕ · g(z) = ZH (h) = ZG hϕ , ψi([x, y]) = ZH hϕ , ψi ⋆ ⋆ F ([z, y])ϕ(y) dλs(z)(y), ϕ(zh)g(h−1) dαs(z)(h), ϕ(y)∗ψ(yh) dλr(h)(y), ϕ(xh)ψ(yh)∗ dαs(x)(h). The completion X = X G H is a C∗(H G, σ∗B) − C∗(H, BH )-imprimitivity bimodule. If L is a representation of C∗(H, BH ), then we write X -- Ind L for the represen- tation of C∗(H G, σ∗B) induced via X. Recall (see, for example, [13, Proposition 2.66]) that X -- Ind L acts on the completion HInd L of X ⊙ HL with respect to via (cid:0)ϕ ⊗ h ψ ⊗ k(cid:1) = (cid:0)L(hψ , ϕi ⋆ )h k(cid:1)HL (X -- Ind L)(F )(ϕ ⊗ h) = F · ϕ ⊗ h. The induced representation of C∗(G; B) acts on HInd L by H L)(f )(ϕ ⊗ h) = f ∗ ϕ ⊗ h, (IndG where f ∗ ϕ(z) = RG f (y)ϕ(y−1z) dλr(z)(y) for f ∈ Γc(G, B) and ϕ ∈ Γc(GH(0) , E ). 2.2. The Proof of Theorem 2.1. For the proof, we consider the case H = G(u) for some u ∈ G(0). Then H (0) = {u} and GH(0) = Gu. Let L be an irreducible representation of C∗(G(u), B). Since X is a C∗(G(u)G, σ∗B) − C∗(G(u), BG(u))-imprimitivity bimodule, [13, Corollary 3.32] implies that X -- Ind L is an irreducible representation of C∗(G(u)G, σ∗B). To prove the theorem, we just need to see that any T in the commutant of IndG G(u) L is a scalar multiple of the identity. It will suffice to see that any such T commutes with (X -- Ind L)(F ) for all F ∈ Γc(G(u)G, σ∗B). Hence, given such an F , we need to produce a net {fi} in Γc(G, B) such that (IndG G(u) L)(fi) → (X -- Ind L)(F ) in the weak operator topology. We will arrange that this net is uniformly bounded in the k · kI -norm on Γc(G, B) -- so that the net {(IndG G(u) L)(fi)} is uniformly bounded in B(HInd L). Then we just have to arrange that (cid:0)(IndG G(u) L)(fi)(ϕ ⊗ h) ψ ⊗ k(cid:1) → (cid:0)(X -- Ind L)(F )(ϕ ⊗ h) ψ ⊗ k(cid:1) for all ϕ, ψ ∈ Γc(Gu, E ) and h, k ∈ HInd L. As in the proof of [6, Theorem 5], the following lemma is the essential ingredient in our proof. Lemma 2.2. Suppose that F ∈ Γc(G(u)G, σ∗B). Then there is a compact set CF in G such that for each compact set K ⊂ Gu there is an fK ∈ Γc(G, B) such that IRREDUCIBLE REPRESENTATIONS 5 (a) fK(zy−1) = F ([z, y]) for all (z, y) ∈ K × K, (b) supp fK ⊂ CF and (c) kfKkI ≤ kF kI + 1. In order to prove lemma 2.2, we need a version of [6, Lemma 7] for semicontinuous functions. Lemma 2.3. Suppose that f is a non-negative upper semicontinuous function on G with compact support and that K ⊂ G is a compact set such that ZK f (x) dλu(x) ≤ M for all u ∈ G(0). Then there is a neighborhood V of K such that ZV f (x) dλu(x) ≤ M + 1 for all u ∈ G(0). Proof. Let K1 a compact neighborhood of K and {Vn} a countable fundamental system of neighborhoods of K in K1 such that Vn+1 ⊂ Vn. Assuming to the contrary that no V as prescribed in the lemma exists, then we can find a sequence {un} ⊂ G(0) such that ZVn f (x) dλun (x) > M + 1. As in [6], we can assume that un → u0. The dominated convergence theorem implies that ZVn f (x) dλu0 (x) → ZK f (x) dλu0 (x). In particular, there is an n1 such that ZVn1 f (x) dλu0 (x) ≤ M + 1 2 . Let W1 be an open set such that K ⊂ W1 ⊂ W 1 ⊂ Vn1 . Then 1W 1 semicontinuous and f is upper ZW 1 f (x) dλu0 ≤ M + 1 2 . Let 0 < ε < 1 for all x ∈ W 1 and 2 . Using [10, Lemma 3.4] we can find g ∈ C+ c (G) such that f (x) ≤ g(x) ZW 1 f (x) dλu0 (x) ≤ ZG g(x) dλu0 (x) < M + 1 2 + ε. Let W be open such that K ⊂ W ⊂ W ⊂ W1 and let f0 ∈ C+ f0W = g, f0 ≤ g, and supp f0 ⊂ W1. Then c (G) be such that ZG f0(x) dλu0 (x) < M + 1 2 + ε. However, since {λu} is a Haar system, ZG f0(x) dλum (x) → ZG f0(x) dλu0 (x). 6 MARIUS IONESCU AND DANA P. WILLIAMS Therefore, for large n, we have that RG f0(x) dλum (x) < M + 1. However, for large n, we have Vn ⊂ W and therefore ZG f0(x) dλun (x) ≥ ZVn ≥ ZVn f0(x) dλun (x) = ZVn g(x) dλun (x) f (x) dλun (x) > M + 1. This leads to a contradiction and the proof is complete. (cid:3) Proof of Lemma 2.2. The map (z, y) 7→ zy−1 is continuous on Gu × Gu and factors through the orbit map π : Gu × Gu → G(u)G. Moreover, the continuous map σ : G(u)G → G defined via σ([z, y]) = zy−1 is injective. We let CF be a compact neighborhood of σ(supp F ). Fix a compact set K ⊂ Gu. Then the restriction of σ to the compact set π(K × K) is a homeomorphism. Using the vector-valued Tietze Extension Theorem ([10, Proposition A.5]) we can find a section fK ∈ Γc(G, B) such that supp fK ⊂ CF and such that fK(zy−1) = F ([z, y]) for all (z, y) ∈ K × K. Let KG := σ(cid:0)π(K × K)(cid:1) ⊂ G. If ZKG k fK(y)k dλw(y) 6= 0, then KGT Gw 6= ∅. Therefore there is z ∈ K such that r(z) = w. Then by left invariance ZKG k fK(y)k dλw(y) = ZG = ZG = ZG 1KG(zy)k fK(zy)k dλw(y) 1KG(zy−1)k fK(zy−1)k dλu(y) 1KG(zy−1)kF ([z, y])k dλu(y) ≤ kF kI. Similarly, if ZKG k fK(y−1)k dλw(y) 6= 0, then as before there is a z ∈ K such that r(z) = w and ZKG k fK(y−1)k dλw(y) = ZG = ZG 1KG(zy)k fK(y−1z−1)k dλw(y) 1KG(zy−1)k fK(yz−1)k dλu(y) which, since K −1 G = KG, is = ZG ≤ ZG 1KG(yz−1)k fK(yz−1)k dλu(y) kF ([y, z])k dλu(y) ≤ kF kI. IRREDUCIBLE REPRESENTATIONS 7 Using Lemma 2.3, we can find a neighborhood V of KG contained in CF such that both ZV k fK(y)k dλw(y) and ZV k fK(y−1)k dλw(y) are bounded by kF kI + 1 for all w ∈ G(0). Since KG is symmetric we can assume that V = V −1 as well. Using the vector-valued Tietze extension theorem, we let fK be any element of Γc(G, B) such that fK = fK on KG, supp fK ⊂ V and kfK(x)k ≤ k fK(x)k everywhere. Then fK satisfies the conclusion of the Lemma. (cid:3) Proof of Theorem 2.1. For each K ⊂ Gu, let fK be as in Lemma 2.3. Then { fK } and { (IndG G(u) L)(fK) } are nets indexed by increasing K. Notice that (2.1) (cid:0)(IndG G(u) L)(fK)(ϕ ⊗G(u) h) ψ ⊗G(u) k(cid:1) − (cid:0)(X -- Ind L)(F )(ϕ ⊗G(u) h) ψ ⊗G(u) k(cid:1) = (cid:0)L(cid:0)hψ , fK ∗ ϕ − F · ϕi ⋆(cid:1)h k(cid:1) Furthermore, using the invariance of the Haar system on G, we can compute as follows: (2.2) hψ , fK ∗ ϕi ⋆ (s) = ZG ψ(x)∗fK ∗ ϕ(xs) dλu(x) = ZGZG ψ(x)∗fK(xz−1)ϕ(zs) dλu(z) dλu(x), while on the other hand, hψ , F · ϕi ⋆ (s) = ZG ψ(x)∗F · ϕ(xs) dλu(x) (2.3) = ZGZG = ZGZG = ZGZG ψ(x)∗F(cid:0)[xs, z](cid:1)ϕ(z) dλu(z) dλu(x) ψ(x)∗F(cid:0)[x, zs−1](cid:1)ϕ(z) dλu(z) dλu(x) ψ(x)∗F(cid:0)[x, z](cid:1)ϕ(zs) dλu(z) dλu(x). Notice that supphψ , ϕi ⋆ have ⊂ (supp ψ)(supp ϕ). Since supp fK ⊂ CF for all K, we supp fK ∗ ϕ ⊂ (supp fK)(supp ϕ) ⊂ CF (supp ϕ). Therefore if (2.2) does not vanish, then we must have s ∈ (supp ψ)CF (supp ϕ). Therefore there is a compact set K0 -- which does not depend on K -- such that both (2.2) and (2.3) vanish if s /∈ K0. Thus if s ∈ K0 and if K ⊃ (supp ψ) ∪ (supp ϕ)K −1 0 , then the integrand in (2.2) and (2.3) are both zero or we must have (x, z) ∈ K × K. Therefore we can replace fK(xz−1) by F(cid:0)[x, z](cid:1), and then fK ∗ ϕ − F · ϕ is the zero section whenever K contains (supp ψ) ∪ (supp ϕ)K −1 0 . Therefore the left-hand side of (2.1) is eventually zero, and the theorem follows. (cid:3) The proof of the following theorem is similar to that of [6, Theorem 4] and we will omit it. The result will be useful in Section 4. 8 MARIUS IONESCU AND DANA P. WILLIAMS Theorem 2.4 (Induction in stages). Suppose that q : B → G is a separable Fell bundle over a second countable locally compact Hausdorff groupoid G and that H and K are closed subgroupoids of G with H ⊂ K. Assume that H, K, and G have Haar systems. If L is a representation of C∗(H, B), then are equivalent representations of C∗(G, B). K(cid:0)IndK H L(cid:1) IndG H L and IndG 3. The Strong Effros-Hahn Induction Property for Fell Bundles Recall that, for a Fell bundle q : B → G over a groupoid G, A = Γ0(G(0); B) is a C∗-algebra that we call the C∗-algebra of B over G(0). Then A is a C0(G(0))- algebra and we let σA : Prim A → G(0) be the associated structure map (see, for example, [20, §C.1]). If u ∈ G(0), let pu : A → A(u) be the quotient map with kernel I(u). Note that if P ∈ Prim A, then u = σA(P ) is the unique u ∈ G(0) such that P ⊃ I(u). Moreover, Prim A is naturally identified with the disjoint union of the Prim A(u) [20, Proposition C.5]. Recall that G acts on Prim A via the Rieffel correspondence hx : Prim A(s(x)) → Prim A(r(x)) (see [7, §2]). Thus, if P ∈ Prim A and if x ∈ G is such that σA(P ) = s(x), then x · P = hx(P ). Since the G-action hx also maps I(s(x)) to I(r(x)) (see [13, Proposition 3.24]), the stability group GP of P is a subgroup of the stability group G(u), where u = σA(P ). Suppose now that q : B → G is a separable Fell bundle over a group G.2 If L : C∗(G, B) → B(W) is a representation, then there is a strictly continuous, nondegenerate ∗-homomorphism π : B → B(W) such that (3.1) L(f ) = ZG π(cid:0)f (s)(cid:1)∆G(s)− 1 2 dµG(s). This is a consequence of [8, Lemma 1.3]. (The appearance of the modular function in (3.1) is a consequence of our convention of treating G as a groupoid: see [8, Remark 1.5].) Notice that πA is a representation of A on W. Let I be an ideal in A = B(e). Then we say that L or π has kernel I if ker(πA) = I. In analogy with [2, Definition 1.1], we make the following definition. Definition 3.1. We say that a Fell bundle q : B → G over a groupoid G with C∗-algebra A over G(0) satisfies the strong Effros-Hahn induction property (strong- EHI) if given P ∈ Prim A and an irreducible representation L of C∗(GP , B) with kernel P , then IndG GP L is irreducible. 4. The strong-EFI for homogeneous representations In this section we prove our main result, which asserts that the strong-EHI prop- erty holds under the additional assumption that the restriction of the representation to A is homogeneous. Theorem 4.1. Let q : B → G be a saturated, separable Fell bundle over a lo- cally compact groupoid G and suppose that P ∈ Prim A where A = Γ0(G(0); B) is the associated C∗-algebra over G(0). Suppose that L is an irreducible represen- tation of C∗(GP , BGP ) which is the integrated form of π : BGP → B(W) with 2Notice that in this case the underlying Banach bundle is continuous [1, Lemma 3.30]. IRREDUCIBLE REPRESENTATIONS 9 πA homogeneous with kernel P . Then IndG C∗(G, B). GP L is an irreducible representation of Remark 4.2. Let L be a representation of C∗(GP , B) with kernel P . Then Theo- rem 2.4 implies that IndG L(cid:1) are equivalent representations, where u = σA(P ). Together with Theorem 2.1, this shows that the irreducibility of IndG(u) GP L. Hence, in order to show that a Fell bundle satisfies strong-EHI, it will suffice to consider the case where G is a group. G(u)(cid:0)IndG(u) GP L and IndG L implies that of IndG GP GP In view of Remark 4.2, we will assume for the remainder of this section that G is a group. Hence, we fix a Fell bundle q : B → G over a group G and let A = B(e) be the corresponding C∗-algebra. To start, let H be any closed subgroup of G. Fix Haar measures µ and ν on G and H, respectively. Let ρ : G → (0, ∞) be a continuous function satisfying (4.1) ρ(xh) = ∆H (h) ∆G(h) ρ(x) for all x ∈ G and h ∈ H. For convenience later, we normalize ρ so that ρ(e) = 1. Then there is a quasi- invariant measure ¯µ on G/H (see, for example [13, Lemma C.2]) such that (4.2) ZG f (x)ρ(x) dµ(x) = ZG/H ZH f (xh) dν(h) d¯µ( x) for all f ∈ Cc(G). In fact, ¯µ is quasi-invariant when viewed as a measure on the unit space of the transformation groupoid G × G/H (which we identify with G/H). Recall that two elements (x, yH) and (z, wH) in G × G/H are composable provided that wH = x−1yH and (x, yH)(z, x−1yH) = (xz, yH). The inverse of (x, yH) is (x−1, x−1yH). It follows that s(x, yH) = x−1yH and r(x, yH) = yH. The Haar system on the transformation groupoid is given by λ = {µ × ǫyH}yH∈G/H . Lemma 4.3. The modular function on the transformation groupoid G × G/H with respect to the quasi-invariant measure ¯µ defined in (4.2) is (4.3) δ(x, yH) = ∆G(x) ρ(y) ρ(x−1y) . Proof. Let F ∈ Cc(G × G/H). Then λ(F ) = ZG/H ZG = ZG/H ZG F (x, yH) dµ(x) d¯µ( y) F (x−1, yH)∆G(x−1) dµ(x) d¯µ( y) which, using Fubini and [13, Lemma C.2], is = ZG/H ZG/H ZH F ((xh)−1, (xh)−1y)∆G((xh)−1) ρ((xh)−1y) 1 = ZG/H ZG F (x−1, x−1yH)∆G(x−1) ρ(y) ρ(x−1y) ρ(y) dν(h)d¯µ( y)d¯µ( x) ρ(xh) dµ(x) d¯µ( y). 10 MARIUS IONESCU AND DANA P. WILLIAMS It follows that δ((x, yH)−1) = δ(x−1, x−1yH) = ∆G(x−1) ρ(x−1y) ρ(y) . The conclusion follows. (cid:3) We now fix a representation L of C∗(H, B) on W, and assume that it is the integrated form of π : BH → B(W) as in (3.1). Note that the map [x, y] 7→ (xy−1, xH) allows us to identify the imprimitivity groupoid H G = (G × G)/H with the transformation groupoid G × G/H. Then we can complete Γc(G; B) to B) and C∗(H, B), where an imprimitivity bimodule X between C∗(G × G/H, pr∗ 1 pr1 : G × G/H is the projection onto the G-factor. Hence pr∗ B = { (y, xH, b) : 1 b ∈ B(y) }. Then the induced representations X -- Ind L of C∗(G × G/H, pr∗ B) and 1 IndG H L of C∗(G, B) both act on the completion of with respect to the pre-inner product Γc(G; B) ⊙ W (cid:0)f ⊗ξ g ⊗ η(cid:1) = (cid:0)L(hg , f i ⋆ )ξ η(cid:1) ⋆ 2 dν(h) (h)(cid:1)ξ η(cid:1)∆H (h)− 1 = ZH(cid:0)π(cid:0)hg , f i = ZH ZG(cid:0)π(cid:0)g(x−1)∗f (x−1h)(cid:1)ξ η(cid:1)∆H (h)− 1 = ZH ZG(cid:0)π(cid:0)g(x)∗f (xh)(cid:1)ξ η(cid:1)∆G(x)−1∆H (h)− 1 = ZH ZG/H ZH(cid:0)π(cid:0)g(xr)∗f (xrh)(cid:1)ξ η(cid:1)∆G(xr)−1∆H (h)− 1 2 dµ(x) dν(h) 2 dµ(x) dν(h) 2 ρ(xr)−1 dν(r) d¯µ( x) dν(h) (4.4) = ZG/H(cid:0)f ⊗ ξ g ⊗ η(cid:1)xH d¯µ( x), where we define (4.5) (cid:0)f ⊗ ξ g ⊗ η(cid:1)xH := ZH ZH(cid:0)π(cid:0)g(xr)∗f (xh)(cid:1)ξ η(cid:1)∆G(xr)−1∆H (r−1h)− 1 2 ρ(xr)−1 dν(h) dν(r). As we described in Section 2.1, the induced representations are given by (X -- Ind L)(F )(g ⊗ η) = F · g ⊗ η and (IndG H L)(f )(g ⊗ η) = f ∗ g ⊗ η, where F · g(x) = ZG F (y, xH)g(y−1x) dµ(y) and f ∗ g(x) = ZG f (y)g(y−1x) dµ(y). We will need the following technical lemma. Lemma 4.4. For each x ∈ G, the sesquilinear form (· ·)xH is positive semi- definite on Γc(G; B) ⊙ W and therefore a pre-inner product. IRREDUCIBLE REPRESENTATIONS 11 Proof. It is not hard to check that f 7→ (f ⊗ ξ f ⊗ ξ) is continuous in the inductive limit topology. It follows from [11, Lemma 6.1] that if {aκ} is an approximate unit for A and f ∈ Γc(G; B), then aκ · f → f in the inductive limit topology. Hence n (cid:16) Xi=1 n n n (4.6) Xj=1 fi ⊗ ξi (cid:12)(cid:12) fj ⊗ ξj(cid:17) = lim κ (cid:16) aκ · fj ⊗ ξj(cid:17). Since B(x) is an A -- A-imprimitivity bimodule, sums of the form Pr crc∗ r, with each cr ∈ B(x), are dense in A+ (see [11, Lemma 6.3]). Since each aκ ≥ 0 in A, this implies that we can replace aκ by a sum Pr cκ r ∈ B(x) and still aκ · fi ⊗ ξi (cid:12)(cid:12) have (4.6) hold. Xi=1 r (cκ r )∗ with cκ Xj=1 But then since ρ(e) = 1 we have ∆G(xr)−1ρ(xr)−1∆H (r−1h)− 1 2 = ∆G(x)−1∆H (hr)− 1 2 ρ(x)−1, and where (cid:0)Xi aκ · fi ⊗ ξi Xj ZH ∆G(x)− 1 r = Xi ηκ aκ · fj ⊗ ξj(cid:1)xH = Xr (ηκ r ηκ r ) 2 ∆H (h)− 1 2 π((cκ r )∗fi(xh))ξi dν(h). Since Pk(ηk ηk) is nonnegative, it follows from (4.6) that our form is a pre-inner product as claimed. (cid:3) In view of Lemma 4.4, the (Hausdorff) completion of Γc(G; B) ⊙ W with respect to (· ·)xH is a Hilbert space V(xH). We denote the image of f ⊗ ξ in V(xH) by f ⊗xH ξ. Note that the class of f ⊗xH ξ depends only on f xH. Furthermore, in view of the Tietze Extension Theorem for upper semicontinuous Banach bundles [10, Proposition A.5], every f0 ∈ Γc(xH; B) is the restriction of some f ∈ Γc(G; B). Using [20, Proposition F.8], we can form a Borel Hilbert bundle G/H ∗ V , and identify the space of both induced representations with the direct integral L2(G/H ∗ V , ¯µ). We will write f ⊗ ξ for the element in L2(G/H ∗ V , ¯µ) given by f ⊗ ξ(xH) = f ⊗xH ξ. If b ∈ B(y) ⊂ B, then we define an operator ιB(b) on the vector space Γc(G; B) by (4.7) ιB(b)f (x) = ∆G(y) 1 2 bf (y−1x). By [8, Lemma 1.2 and Remark 1.5], ιB extends to a homomorphism, also denoted ιB, of B into M (C∗(G, B)). H L is the integrated form of Π : B → B(L2(G/H ∗ V , ¯µ)) given by It follows that IndG (4.8) Π(b)(f ⊗ ξ) = ιB(b)f ⊗ ξ. Note that Π is not decomposible; that is, it does not operate fibrewise on L2(G/H ∗ V , ¯µ). Nevertheless, it does play nice with the fibres and is related to a Borel ∗-functor for pr1 Lemma 4.5. If b ∈ B(y) ⊂ B, B(V(y−1xH), V(xH)) of norm at most kbk given by then we get an operator π(y, xH, b) ∈ B (see [10, Definition 4.5]). π(y, xH, b)(f ⊗y−1xH ξ) = π0(y, xH, b)(f ) ⊗xH ξ, where π0(y, xH, b)(f )(xh) = δ(y, xH) 1 2 bf (y−1xh). 12 MARIUS IONESCU AND DANA P. WILLIAMS (Recall that δ is the modular function on G × G/H for the quasi-invariant measure ¯µ -- see Lemma 4.3.) Proof. Using δ(y, xH) = ∆G(y) ρ(x) ρ(y−1xh) for any h ∈ H, we first check that ρ(xh) ρ(y−1x) as well as the observation that ρ(x) ρ(y−1x) = (cid:0)π(y, xH, b)(f ⊗ ξ) g ⊗ η(cid:1)xH = (cid:0)f ⊗ ξ π(y−1, y−1xH, b∗(cid:1)(g ⊗ η))y−1xH . However, in A, we have kbk21 A − b∗b ≥ 0. Hence there is a c ∈ A such that kbk21 A − b∗b = c∗c. Thus, extending π(e, xH, a) to all a ∈ A in the usual way, we have kbk2kf ⊗ ξk2 y−1xH − kπ(y, xH, b)(f ⊗ ξ)k2 xH = kbk2(cid:0)f ⊗ ξ f ⊗ ξ(cid:1)y−1xH −(cid:0)π(y, xH, b)(f ⊗ ξ) π(y, xH, b)(f ⊗ ξ)(cid:1)xH = kbk2(cid:0)f ⊗ ξ f ⊗ ξ(cid:1)y−1xH −(cid:0)f ⊗ ξ π(e, y−1xH, b∗b)(f ⊗ ξ(cid:1)y−1xH = (cid:0)π(e, y−1xH, c)(f ⊗ ξ) π(e, y−1xH, c)(f ⊗ ξ)(cid:1)y−1xH ≥ 0. Thus π(y, xH, b) is a well-defined operator from V(y−1xH) to V(xH) for any x, and has norm at most kbk. (cid:3) Now it is not hard to see that π is a Borel ∗-functor on pr1 B, and more signifi- cantly, X -- Ind L is the integrated form of π -- see [10, Proposition 4.10]. At this point, we want to fix a primitive ideal P ∈ Prim A and replace H by GP , where GP is the stability group for the G action on Prim A induced using the Rieffel correspondence: see [7, §2] and Section 3 for details. In particular, if x ∈ G, then (4.9) B(x) · P = (x · P ) · B(x). We let ρxGP be the representation of A in V(xGP ) given by a 7→ π(e, xGP , a). While a priori the ρxGP do not seem related to πA, they in fact are intimately so. Lemma 4.6. The representation ρeGP defined above is equivalent to πA. Proof. We claim that the map U : V(GP ) → W defined via U (f ⊗eGP ξ) = ZeGP ∆H (h)− 1 2 π(f (h))ξ dν(h), is a unitary that intertwines ρeGP with πA. Let f ⊗ ξ and g ⊗ η be two elements in Γc(G, B) ⊙ W. Then ∆H (h)− 1 (cid:0)U (f ⊗eGP ξ) (cid:12)(cid:12) U (g ⊗eGP η)(cid:1) = (cid:16)ZGP = ZGP ZGP ZGP(cid:0)π(f (h))ξ (cid:12)(cid:12) π(g(r))η(cid:1)∆H (h)− 1 2 π(f (h))ξ dν(h) (cid:12)(cid:12)(cid:12) which, using the fact that πB is a ∗-homomorphism and Equation (4), equals = ZGP ZGP(cid:0)π(g(r)∗f (h))ξ (cid:12)(cid:12) η(cid:1)∆G(r)−1ρ(r)−1∆H (r−1h) dν(h) dν(r) ∆H (r)− 1 2 π(g(r))η dν(r)(cid:17) 2 ∆H (r)− 1 2 dν(h) dν(r) IRREDUCIBLE REPRESENTATIONS 13 = (cid:0)f ⊗eGP ξ (cid:12)(cid:12) g ⊗eGP η(cid:1)eGP . The surjectivity of U follows from the fact that there are enough sections and the fact that π is nondegenerate (see [18, Section 3]). Thus U is indeed a unitary that clearly intertwines ρeGP with πA. (cid:3) Since each B(x) is an A -- A-imprimitivity bimodule, we can form the induced representation B(x) -- Ind(ρeGP ) of A on B(x) ⊗A V(eGP ). Lemma 4.7. For each x ∈ G, B(x) -- Ind(ρeGP ) is equivalent to ρxGP . Proof. Define U : B(x) ⊙ V(eGP ) → V(xGP ) by U (c ⊗ f ⊗eGP ξ) = ∆G(x) 1 2 ρ(x) 1 2 (fc ⊗xGP ξ), where fc is any section in Γc(G; B) satisfying fc(xh) = cf (h). To check that U is isometric one can proceed as follows: (cid:0)U (c ⊗ f ⊗eGP ξ) U (b ⊗ g ⊗eGP η)(cid:1)xGP = ∆G(x)ρ(x)(cid:0)fc ⊗xGP ξ gb ⊗xGP η(cid:1)xGP = ZGP ZGP(cid:0)π(gb(xr)∗fc(xh))ξ η(cid:1)∆G(x)∆G(xr)−1∆H (r−1h)− 1 2 ρ(x)ρ(xr)−1dν(h)dν(r) = ZGP = ZGP ZGP(cid:0)π(g(r)∗b∗cf (h))ξ η(cid:1)∆G(r)−1∆H (r−1h)− 1 ZGP(cid:0)π(g(r)∗ π0(e, eGP , b∗c)(f )(h))ξ η(cid:1)∆G(r)−1∆H (r−1h)− 1 2 2 ρ(r)−1dν(h)dν(r) ρ(r)−1dν(h)dν(r) = (cid:0)ρeGP (b∗c)(f ⊗eGP ξ) g ⊗eGP η(cid:1)eGP = (cid:0)c ⊗ f ⊗eGP ξ b ⊗ g ⊗eGP η(cid:1). Clearly U intertwines B(x) -- Ind(ρeGP )) with ρxGP . Notice that { c ⊗ f (h) : f ∈ Γc(G, B), c ∈ B(x) } is dense in B(x) ⊗A B(h) ≃ B(xh). Therefore the set of fc(xh) where fc(xh) = cf (h), f ∈ Γc(G, B), and c ∈ B(x) is dense in B(xh). [10, Lemma A.4] implies that U has dense range. (cid:3) Of course, just as with group dynamics (see [2, §2.3]), the homogeneity of πA will play a crucial role. As in [2, Remark 1.5], we will use [16, Lemme 1.5] or [3, Theorem 1.4] to characterize homogeneous representations ρ of A as those that have the property that for any ideal I in A such that I 6⊂ ker ρ, ρ(I)Hρ = Hρ. Alternatively, note that ρ is homogeneous if given any approximate identity { ui } for an ideal I 6⊂ ker ρ, we have ρ(ui) → IHρ in the strong operator topology. We need the following general Morita Equivalence result which we have not seen elsewhere.3 Proposition 4.8. Let A and B be C∗-algebras and let X be an A -- B-imprimitivity bimodule. Then π is a homogeneous representation of B if and only if X -- Ind π is a homogeneous representation of A. 3For other such results, see [5]. 14 MARIUS IONESCU AND DANA P. WILLIAMS Proof. By symmetry, it suffices to show that π homogeneous implies that X -- Ind π is homogeneous. So, we suppose that π is homogeneous and that K is an ideal in A with K 6⊂ ker(X -- Ind π). Let { ui } be an approximate unit for K. As remarked above, it will suffice to see that (X -- Ind π)(ui) → I in the strong operator topology. By the Rieffel correspondence ([13, §3.3]), we can assume that K = X -- Ind J for an ideal J in B with J 6⊂ ker π. Since π is homogeneous, π(J)Hπ = Hπ. Hence it will suffice to show that ui · x ⊗ π(b)h → x ⊗ π(b)h in the Hilbert space X ⊗B Hπ for all x ∈ X, b ∈ J and h ∈ Hπ. However, X · J is a K -- J-imprimitivity bimodule (see [13, Proposition 3.25]). Hence for any y ∈ X · J, we have ui · y → y in X (for example, see equation (2.5) in the proof of [13, Corollary 2.7]). So in X ⊗B Hπ, we have (cid:0)ui · x ⊗ π(b)h ui · x ⊗ π(b)h(cid:1) = (cid:0)π(cid:0)hui · x , ui · xi = (cid:0)π(cid:0)hui · x · b , ui · x · bi ⋆(cid:1)π(b)h π(b)h(cid:1) ⋆(cid:1)h h(cid:1), which converges to (cid:0)π(cid:0)hx , xi ⋆(cid:1)π(b)h π(b)h(cid:1) = (cid:0)x ⊗ π(b)h x ⊗ π(b)h(cid:1) since x · b ∈ X · J. Then, using similar calculations, we can see that kui · x ⊗ π(b)h − x ⊗ π(b)hk2 converges to zero as required. = (cid:0)ui · x ⊗ π(b)h − x ⊗ π(b)h ui · x ⊗ π(b)h − x ⊗ π(b)h(cid:1) (cid:3) Proposition 4.9. Suppose that L has kernel P and that πA is homogeneous. Then, for all x ∈ G, ρxGP is a homogeneous representation of A with kernel x · P . Proof. By Lemma 4.6 and our assumptions on L and π, we have that ρeGP is homogeneous with kernel P . The rest follows from Lemma 4.7 and Proposition 4.8. (cid:3) We have set things up so that the representation of A given by ΠA (where Π is defined in (4.8)) is the direct integral (4.10) Z ⊕ G/GP ρxGP d¯µ on L2(G/GP ∗ V , ¯µ). Recall that the diagonal operators, ∆(G/GP ∗ V , ¯µ), on L2(G/GP ∗ V , ¯µ) are, by definition, the multiplication operators determined by bounded Borel functions on G/GP (see [20, Definition F.13]). In this case, ∆(G/GP ∗ V , ¯µ) are exactly the operators of the form M (ϕ) with ϕ ∈ L∞(G/GP ) and M (ϕ)(f ⊗ ξ) = ϕ · f ⊗ π, where ϕ · f (x) = ϕ(xGP )f (x). We have worked fairly hard to see that (4.10) is an ideal center decomposition (see [20, Definition G.18 and Theorem G.20]). As in the proof of [2, Theorem 1.7], the essential feature we require from this observation is that M (L∞(G/GP )) lies in the center of the commutant of Π(A) so that (4.11) Π(A)′ ⊂ M (L∞(G/GP ))′ ⊂ M (C0(G/GP ))′. IRREDUCIBLE REPRESENTATIONS 15 Lemma 4.10. Let f ∈ Γc(G; B) and ϕ ∈ Cc(G/GP ). Then we get a section f ⊗ ϕ in Γc(G × G/GP ; pr∗ 1 B) via f ⊗ ϕ(y, xH) = ϕ(xH)f (y). Such sections span a dense subspace of Γc(G × G/GP ; pr1 topology. B) in the inductive limit Proof. Let G be the span of sections of the form f ⊗ ϕ. Since g · (f ⊗ ϕ) = (g · f ) ⊗ ϕ and ψ · (f ⊗ ϕ) = f ⊗ (ψ · ϕ) for g ∈ Cc(G), f ∈ ΓC (G; B), ψ, ϕ ∈ Cc(G/GP ), it follows that G is closed under multiplication by functions in Cc(G × G/GP ). Notice that, for (y, xH) ∈ G × G/GP , we have that pr∗ 1 Therefore, the set { f ⊗ ϕ(y, xH) f ⊗ ϕ ∈ G } is dense in pr∗ 1 variation of [10, Lemma A.4] implies the result. B(y, xH) = B(y). B(y, xH). A small (cid:3) Proof of Theorem 4.1 in the case G is a group. Suppose that T ∈ B(L2(G/GP ∗ V , ¯µ) is in the commutant of IndG GP L. It suffices to see that T is a scalar operator. But T must also commute with Π(b) = (IndG GP L)(cid:0)ιB(b)(cid:1) for all b ∈ B. In partic- ular, T belongs to Π(A)′. Hence T commutes with M (ϕ) for all ϕ ∈ Cc(G/GP ) by (4.11). But (X -- Ind L)(f ⊗ ϕ) = M (ϕ)(IndG GP L)(f ). Thus T commutes with (X -- Ind L)(f ⊗ ϕ). By Lemma 4.10, this means T commutes with the irreducible representation X -- Ind L. Thus T is a scalar. (cid:3) 5. The Type I Case In this section, we prove the appropriate analogue of [2, Lemma 3.2]. Recall that if p : B → G is a Fell bundle and if P is a primitive ideal of A = Γ0(G(0); B), then P ⊃ I(u) for u = σA(P ). In particular, we can view P as a primitive ideal of the quotient A(u) = B(u). Of course, P is locally closed in Prim A if and only if it is locally closed when viewed as a primitive ideal of A(u). Proposition 5.1. Let p : B → G be a separable, saturated Fell bundle over a locally compact groupoid G and let P be a primitive ideal in Prim A where A = Γ0(G(0); B) is the associated C∗-algebra. Suppose that P is locally closed in Prim A and that L is an irreducible representation of C∗(GP , B) which is the integrated form of the strongly continuous map π : BGP → B(V) such that ker πA(u) = P where u = σA(P ). Then πA(u) is a homogeneous representation of A(u). We'll need the following standard observation from Morita theory. Lemma 5.2. Suppose that X is an A -- B-imprimitivity bimodule. Then the Rieffel correspondence h : I (B) → I (A) preserves arbitrary intersections. Proof. Notice that Hence the result follows once we establish that h(I) = { a ∈ J : a · X ⊂ X · I }. \ X · IJ = X ·\ IJ . But this can be proved exactly as in the proof of [20, Lemma 5.19]. (cid:3) 16 MARIUS IONESCU AND DANA P. WILLIAMS Proof of Proposition 5.1. We may as well replace B by BGP and A by A(u). Let ρ = πA. As in the proof of [2, Lemma 3.2], it will suffice to see that if I is any ideal in A such that I 6⊂ P , then ρ(I)V = V. Since ρ(I) = ρ(I + P ), we can assume I properly contains P . If x ∈ GP , let x · I := hx(I). Then x · I properly contains P . Since any ideal in A is the intersection of those primitive ideals which contain it, K := \x∈GP x · I ⊃ \P ′∈Prim A P ′⊃P P ′6=P P ′ := J. Since P is locally closed in Prim A, [2, Lemma 3.1] implies that J properly contains P . Hence so does K. Moreover K must be GP -invariant as defined in [7, §3.1]: if P ′ is a primitive ideal containing K, then using Lemma 5.2, we have y · P ′ = hy(P ′) ⊃ \x∈GP hy(x · I) = \y∈GP hxy(I) = K. Thus we can identify C∗(GP , BK ) with an ideal of C∗(GP , B) as in [7, Lemma 3.5]. We claim (5.1) To see this, let w ∈ (π(K)V)⊥. Then for any h ∈ V, L(cid:0)C∗(GP , BK)(cid:1)V ⊂ π(K)V. (5.2) (cid:0)L(f )h w(cid:1) = ZGP ∆Gp (s)− 1 2(cid:0)π(f (s))h w(cid:1) ds. But f (s) ∈ B(s) · K = (s · K) · B(s) = K · B(s). Hence π(f (s))h ∈ π(K)V for each h. Thus the integrand in (5.2) is zero and (cid:0)L(f )h w(cid:1) = 0 for all w ∈ (π(K)V)⊥. Thus (5.1) follows. Next we claim that (5.3) L(cid:0)C∗(GP , BK)(cid:1)V 6= {0}. To see this, note that as K properly contains P , there is an a ∈ K and h ∈ V such that π(a)h 6= 0. But as Fell bundles always have sufficiently many sections, there is an f ∈ Γc(GP ; BK) such that f (e) = a. Since π is strongly continuous, there is a neighborhood V of e in GP such that for all s ∈ V we have and such that kπ(f (s))h − π(a)hk < 1 4 kπ(a)hk, ∆GP (s)− 1 2 ≤ 2. Let ϕ ∈ C+ c (GP ) be such that supp ϕ ⊂ V and ZGP ∆GP (s)− 1 2 ϕ(s) ds = 1. Then ϕ · f ∈ Γc(GP ; BK ) and ZGP kL(ϕ · f )h − π(a)f k = (cid:13)(cid:13)(cid:13) ∆GP (s)− 1 < 2 kπ(a)hk = 1 4 2 ϕ(s)(cid:0)π(f (s)h − π(a)h(cid:1) ds(cid:13)(cid:13)(cid:13) kπ(a)hk. 1 2 Thus L(ϕ · f )h 6= 0. This establishes (5.3). IRREDUCIBLE REPRESENTATIONS However, as L is assumed irreducible, (5.3) and (5.1) imply that Thus ρ(I)V = V and ρ is homogeneous as claimed. L(cid:0)C∗(GP , BK)(cid:1)V = V ⊂ π(K)V. 17 (cid:3) Theorems 5.1 and 4.1 imply immediately the following generalization of [2, The- orem 1.8]. Proposition 5.3. Suppose that p : B → G is a separable, saturated Fell bundle over a locally compact groupoid G such that all points are locally closed in Prim A. Then p : B → G satisfies strong-EHI. In particular, if A is of type I, then p : B → G satisfies the strong-EHI. Remark 5.4. Even if A is not of type I, there are a number of general situations where Prim A necessarily has locally closed points. For example, as we observe in [2, Proposition 1.8], this is case in any subquotient of the group C∗-algebra of an almost connected Lie group. Nevertheless, there are separable C∗-algebras for which this property fails (see [2, Example 3.3]). 6. Examples 6.1. Groupoid dynamical systems. Let G be a locally compact Hausdorff groupoid with Haar system { λu }u∈G(0). Let π : A → G(0) be an upper semicontinuous C∗-bundle over G(0) and let A = Γ0(G(0), A ). Assume that (A , G, α) is a groupoid dynamical system (see, [11, Definition 4.1]). Recall from [10, §2] that one can define a Fell bundle p : B → G, where B := r∗A = { (a, x) : π(a) = r(x) } is the pull-back of A via the range map r. The multiplication on B is defined via for example, if (x, y) ∈ G(2), and the involution is given by (a, x)(b, y) := (aαx(b), xy), (a, x)∗ := (α−1 x (a∗), x−1). Then Γc(G, r∗A ) is a ∗-algebra with respect to the operations f ∗ g(x) = ZG f (y)αy(cid:0)g(y−1x)(cid:1) dλr(x)(y) and f ∗(x) = αx(cid:0)f (x−1)∗(cid:1). The crossed product A ⋊α G is the enveloping C∗-algebra of Γc(G; r∗A ). Recall (see, for example, [11, Definition 7.5]) that a unitary representation of a groupoid G with Haar system { λu }u∈G(0) is a triple (µ, G(0) ∗ H, L) consisting of a quasi-invariant measure µ on G(0), a Borel Hilbert bundle G(0) ∗ H over G(0) and a Borel homomorphism U : G → Iso(G(0) ∗ H) such that U (x) = (r(x), Lx, s(x)), where Iso(G(0) ∗ H) is the isomorphism groupoid. A covariant representation (M, µ, G(0) ∗ H, U ) of (A , G, α) ([11, Definition 7.9]) consists of a unitary representation (µ, G(0) ∗ H, U ) of G and a C0(G(0))-linear representation M : A → B(L2(G(0) ∗ H, µ)) such that there are representations Mu : A → B(H(u)) so that M (a)h(u) = Mu(a)(h(u)) for µ-almost all u, 18 MARIUS IONESCU AND DANA P. WILLIAMS and such that there is a ν := λ ◦ µ-null set N such that for all x /∈ N , UxMs(x)(b) = Mr(x)(αx(b))Ux for all b ∈ A(s(x)). Recall from [11, Proposition 7.11] that if (M, µ, G(0) ∗ H, U ) is a covariant repre- sentation of (A , G, α), then there is a k · kI-norm decreasing ∗-representation L of Γc(G; r∗A ) called the integrated form of the covariant representation given by L(f )h(u) = ZG Mu(f (x))Uxh(s(x))∆(x)− 1 2 dλu(x). Conversely, given any representation L of A ⋊αG, there is a covariant representation (M, µ, G(0) ∗ H, U ) such that L is equivalent to the corresponding integrated form ([11, Theorem 7.12]). Notice that if G is a group and (ρ, U ) is a covariant representation of (A, G, α), then, since we are treating G as a groupoid, the corresponding integrated form is given by the formula ρ ⋊ U (f ) = ZG ρ(f (s))Us∆(s)− 1 2 dµ(s) (compare this formula against, for example, [20, Equation (2.19)]). If H is a closed subgroupoid of G and L is a representation of A ⋊aH H, then the construction of induced representations from [17, §4.1] (see also Section 2 above) gives us that the induced representation IndG H L acts on the completion of X ⊙ HL by (IndG H L)(f )(ϕ ⊗ h) = f ∗ ϕ ⊗ h, where f ∗ ϕ(z) = RG f (y)αy(cid:0)ϕ(y−1z)(cid:1) dλr(z)(y). Our Theorem 2.1 seems to be new for this set-up. Theorem 6.1. Assume that (A , G, α) is a groupoid dynamical system. Let u ∈ G(0) and suppose that L is an irreducible representation of A ⋊αG(u) G(u). Then IndG G(u) L is an irreducible representation of A ⋊α G. Recall from [7, Remark 3] that the G action on Prim A is the same as the usual one: x · P = αx(P ). Our main theorem (Theorem 4.1) becomes: Theorem 6.2. Let (A , G, α) be a groupoid dynamical system. Let P ∈ Prim A and let (ρ, U ) be a covariant representation of (A, GP , αGP ). Assume that ker ρ = P and that ρ ⋊ U is irreducible. If either A is type I or if ρ is homogeneous, then IndG GP (ρ ⋊ U ) is irreducible. In particular we recover the main result of [2]. 6.2. Green-Renault's twisted groupoid dynamical systems. Suppose that G(0) → S i−→ Σ j −→ G → G(0) is a groupoid extension of locally compact groupoids over G(0) where S is a group bundle of abelian groups admitting a Haar system. We view S as a closed sub- groupoid of Σ. We assume that we have a groupoid dynamical system (A , Σ, α), so that π : A → G(0) = Σ(0) is an upper semicontinuous C∗-bundle. We also need an element χ ∈ Qs∈S M (A(r(s))) such that (s, a) 7→ χ(s)a IRREDUCIBLE REPRESENTATIONS 19 is continuous from S ∗ A to A , and such that αs(a) = χ(s)aχ(s)∗ for all (s, a) ∈ S ∗ A , and χ(σsσ−1) = ασ(χ(s)) for (σ, s) ∈ Σ(2). Following [14] and [15], we call (G, Σ, A ) a twisted groupoid dynamical system. As in [10, Example 2.5], we define an S-action on r∗A = {(a, σ) : π(a) = r(σ)} by (a, σ) · s := (aχ(s)∗, sσ). The associated Fell bundle is then B := r∗A /S with the map p : B → G, p([a, σ]) = j(σ). The operations in B are defined via (see [10, Example 2.5] for details): if (j(σ), j(τ )) ∈ G(2) and [a, σ][b, τ ] := [aασ(b), στ ] [a, σ]∗ = [α−1 σ (a∗), σ−1]. To define a section of B, we need a continuous function f : Σ → A such that f (σ) ∈ A(r(σ)) and f (sσ) = f (σ)χ(s)∗ for all s ∈ S and σ ∈ Σ such that (s, σ) ∈ Σ(2). The corresponding section is given by f (j(σ)) = [f (σ), σ]. Replacing f with f , the ∗-operations on Σ are given by and f ∗(σ) = ασ(f (σ−1)∗). f ∗ g(σ) = ZG f (τ )ατ(cid:0)g(τ −1σ)(cid:1) dλr(j(σ))(τ ) As described in [10, Example 2.10], the completion is Renault's C∗(G, Σ, A , λ) from [14] and [15]. A covariant representation (M, µ, G(0) ∗ H, U ) of (G, Σ, A ) ([14, Definition 3.4]) consists of a unitary representation (µ, G(0) ∗ H, U ) of Σ and a C0(G(0))-linear representation of M : A → B(L2(G(0) ∗ H), µ) so that there are representations Mu : A → B(H(u)) such that M (a)h(u) = Mu(a)(h(u)) for µ-almost all u, and such that there is a µ-conull set V such that UxMs(x)(a) = Mr(x)(αx(a))Ux for all x ∈ ΣV and a ∈ A(s(x)), and Us = Mr(s)(χ(s)) for all s ∈ SV . If (M, µ, G(0)∗H, U ) is a covariant representation of (G, Σ, A ), then [14, Proposition 3.5] (see also [10, Proposition 4.10]) implies that there is a k · kI -norm decreasing ∗-representation L of Γc(G, Σ, A ), called the integrated form of the covariant rep- resentation, given by L(f )h(u) = ZG Mu(f (x))Uxh(s(x))∆(j(x))−1/2dλu(j(x)). Conversely, [14, Theorem 4.1] and [10, Theorem 4.13] imply that every representa- tion of C∗(G, Σ, A ) is equivalent to the integrated form of a covariant representa- tion. Notice that if Σ is a group and S is an abelian subgroup of Σ then G = Σ/S. We recover the twisted dynamical systems of Green [4] and ¯Da. ng Ngo. c [12]. If (ρ, U ) is a covariant representation of (A, G, α) that preserves χ, i.e. ρ(χ(s)) = U (s) for all 20 MARIUS IONESCU AND DANA P. WILLIAMS s ∈ S, then, since we are treating groups as groupoids, the integrated form ρ ⋊χ U of (ρ, U ) is ρ ⋊χ U (f ) = ZG ρ(f (s))U (s)∆( s)− 1 2 dµ( s). If H is a closed subgroupoid of G and L is a representation of C∗(H, Σ, A ), then the induced representation IndG H acts on the completion of X ⊙ HL (see Section 2 for details) via IndG H (L)(f )(ϕ ⊗ h) = f ∗ ϕ ⊗ h, where f ∗ ϕ(z) = RG f (y)αy(g(y−1z)) dλr(j(z))(y). As an immediate consequence of Theorem 2.1 we obtain the following: Theorem 6.3. Assume that (G, Σ, A ) is a twisted groupoid dynamical system. Let u ∈ G(0) and suppose that L is an irreducible representation of C∗(G(u), Σ(u), A ). Then IndG G(u) L is an irreducible representation of C∗(G, Σ, A ). Using [7, Lemma 2.1] one obtains that the G action on Prim A is the same as the usual one. This fact, together with 4.1, implies the following: Theorem 6.4. Let (G, Σ, A ) be a twisted groupoid dynamical system. Let P ∈ Prim A and let (ρ, U ) be a covariant representation of (GP , ΣP , A ) that preserves χ and is such that ker ρ = P and ρ ⋊χ U is irreducible. If either A is of type I or if ρ is homogeneous, then IndG GP ρ ⋊χ U is irreducible. References [1] Alcides Buss, Ralf Meyer, and Chenchang Zhu, A higher category approach to twisted actions on C ∗-algebras, Proc. Edinb. Math. Soc. (2) 56 (2013), 387 -- 426. [2] Siegfried Echterhoff and Dana P. Williams, Inducing primitive ideals, Trans. Amer. Math. Soc 360 (2008), 6113 -- 6129. [3] Edward G. Effros, A decomposition theory for representations of C ∗-algebras, Trans. Amer. Math. Soc. 107 (1963), 83 -- 106. [4] Philip Green, The local structure of twisted covariance algebras, Acta Math. 140 (1978), 191 -- 250. [5] Astrid an Huef, Iain Raeburn, and Dana P. Williams, Properties preserved under Morita equivalence of C ∗-algebras, Proc. Amer. Math. Soc. 135 (2007), 1495 -- 1503. [6] Marius Ionescu and Dana P. Williams, Irreducible representations of groupoid C ∗-algebras, Proc. Amer. Math. Soc. 137 (2009), 1323 -- 1332. [7] , Remarks on the ideal structure of Fell bundle C ∗-algebras, Houston J. Math. 38 (2012), 1241 -- 1260. [8] S. Kaliszewski, Paul S. Muhly, John Quigg, and Dana P. Williams, Coactions and Fell bun- dles, New York J. Math. 16 (2010), 315 -- 359. [9] Paul S. Muhly, Bundles over groupoids, Groupoids in analysis, geometry, and physics (Boul- der, CO, 1999), Contemp. Math., vol. 282, Amer. Math. Soc., Providence, RI, 2001, pp. 67 -- 82. [10] Paul S. Muhly and Dana P. Williams, Equivalence and disintegration theorems for Fell bun- dles and their C ∗-algebras, Dissertationes Math. (Rozprawy Mat.) 456 (2008), 1 -- 57. [11] , Renault's equivalence theorem for groupoid crossed products, NYJM Monographs, vol. 3, State University of New York University at Albany, Albany, NY, 2008. Available at http://nyjm.albany.edu:8000/m/2008/3.htm. [12] Nghiem ¯Da. ng Ngo. c, Produits crois´es restreints et extensions of groupes, 1977. Notes, Paris. [13] Iain Raeburn and Dana P. Williams, Morita equivalence and continuous-trace C ∗-algebras, Mathematical Surveys and Monographs, vol. 60, American Mathematical Society, Providence, RI, 1998. [14] Jean Renault, Repr´esentation des produits crois´es d'alg`ebres de groupoıdes, J. Operator The- ory 18 (1987), 67 -- 97. IRREDUCIBLE REPRESENTATIONS 21 [15] , The ideal structure of groupoid crossed product C ∗-algebras, J. Operator Theory 25 (1991), 3 -- 36. [16] Jean-Luc Sauvageot, Id´eaux primitifs induits dans les produits crois´es, J. Funct. Anal. 32 (1979), 381 -- 392. [17] Aidan Sims and Dana P. Williams, Renault's equivalence theorem for reduced groupoid C ∗- algebras, J. Operator Theory 68 (2012), 223 -- 239. [18] [19] , Amenability for Fell bundles over groupoids, Illinios J. Math. (2013), in press. (arXiv:math.OA.1201.0792). , An equivalence theorem for reduced Fell bundle C ∗-algebras, New York J. Math. 19 (2013), 159 -- 178. [20] Dana P. Williams, Crossed products of C ∗-algebras, Mathematical Surveys and Monographs, vol. 134, American Mathematical Society, Providence, RI, 2007. Department of Mathematics, Colgate University, Hamilton, NY 13346 E-mail address: [email protected] Department of Mathematics, Dartmouth College, Hanover, NH 03755-3551 E-mail address: [email protected]
1204.1841	1	1204	2012-04-09T09:49:50	Strong skew commutativity preserving maps on von Neumann algebras	[ "math.OA", "math.RA" ]	Let ${\mathcal M}$ be a von Neumann algebra without central summands of type $I_1$. Assume that $\Phi:{\mathcal M}\rightarrow {\mathcal M}$ is a surjective map. It is shown that $\Phi$ is strong skew commutativity preserving (that is, satisfies $\Phi(A)\Phi(B)-\Phi(B)\Phi(A)^=AB-BA^$ for all $A,B\in{\mathcal M}$) if and only if there exists some self-adjoint element $Z$ in the center of ${\mathcal M}$ with $Z^2=I$ such that $\Phi(A)=ZA$ for all $A\in{\mathcal M}$. The strong skew commutativity preserving maps on prime involution rings and prime involution algebras are also characterized.	math.OA	math	STRONG SKEW COMMUTATIVITY PRESERVING MAPS ON VON NEUMANN ALGEBRAS XIAOFEI QI AND JINCHUAN HOU Abstract. Let M be a von Neumann algebra without central summands of type I1. Assume that Φ : M → M is a surjective map. It is shown that Φ is strong skew commutativity preserving (that is, satisfies Φ(A)Φ(B) − Φ(B)Φ(A)∗ = AB − BA∗ for all A, B ∈ M) if and only if there exists some self-adjoint element Z in the center of M with Z 2 = I such that Φ(A) = ZA for all A ∈ M. The strong skew commutativity preserving maps on prime involution rings and prime involution algebras are also characterized. 2 1 0 2 r p A 9 ] . A O h t a m [ 1 v 1 4 8 1 . 4 0 2 1 : v i X r a 1. Introduction Let R be a -ring. For A, B ∈ R, denote by [A, B]∗ = AB − BA∗ the skew Lie product. This product AB − BA∗ is found playing an important role in some research topics. Let B be a fixed element in R. The additive map Φ : R → R defined by Φ(A) = AB − BA∗ for all A ∈ R is a Jordan -derivation, that is, it satisfies Φ(A2) = Φ(A)A∗ + AΦ(A). The notion of Jordan -derivations arose naturally in Semrl's work [10, 11] investigating the problem of representing quadratic functionals with sesquilinear functionals. Motivated by the theory of rings (and algebras) equipped with a Lie product [T, S] = T S − ST or a Jordan product T ◦ S = T S + ST , Moln´ar in [9] studied the skew Lie product and gave a characterization of ideals in B(H) in terms of the skew Lie product. It is shown in [9] that, if N ⊆ B(H) is an ideal, then N = span{AB − BA∗ : A ∈ N , B ∈ B(H)} = span{AB − BA∗ : A ∈ B(H), B ∈ N }. In particular, every operator in B(H) is a finite sum of AB − BA∗ type operators. Later, Bresar and Fonsner [2] generalized the above results in [9] to rings with involution in different ways. Recall that a map Φ : R → R is skew commutativity preserving if, for any A, B ∈ R, [A, B]∗ = 0 implies [Φ(A), Φ(B)]∗ = 0. The problem of characterizing linear (or additive) bijective maps preserving skew commutativity had been studied intensively on various algebras (see [4, 5] and the references therein). More specially, we say that a map Φ : R → R is strong skew commutativity preserving (briefly, SSCP) if [Φ(A), Φ(B)]∗ = [A, B]∗ for all A, B ∈ R. SSCP maps are also called strong skew Lie product preserving maps in [6]. We prefer to the 2010 Mathematical Subject Classification. 47B49, 46L10. Key words and phrases. Von Neumann algebras, prime rings, general preserving maps, skew Lie products. This work is partially supported by the National Natural Science Foundation of China (11171249, 11101250). 1 2 XIAOFEI QI AND JINCHUAN HOU terminology SSCP because many authors call the maps satisfying [Φ(A), Φ(B)] = [A, B] strong commutativity preserving maps. It is obvious that a strong skew commutativity preserving map must be skew commutativity preserving, but the inverse is not true generally. In [6], Cui and Park proved that, if R is a factor von Neumann algebra, then every SSCP surjective map Φ on R has the form Φ(A) = Ψ(A) + h(A)I for every A ∈ R, where Ψ : R → R is a linear bijective map satisfying Ψ(A)Ψ(B) − Ψ(B)Ψ(A)∗ = AB − BA∗ for A, B ∈ R and h is a real functional on R with h(0) = 0; particularly, if R is a type I factor, then Φ(A) = cA + h(A)I for every A ∈ R, where c ∈ {−1, 1}. In the present paper, we show further that, in the above result, h = 0 and Ψ(A) = A for all A or Ψ(A) = −A for all A. In fact, the purpose of the present paper is to give a characterization of the SSCP maps on prime -rings or on general von Neumann algebras without central summands of type I1. And the improvement of the above result in [6] is an immediate consequence of our results. Before embarking on the main results, we need some notations. Let A be a -ring. Denote by Z(A) the center of A. Observe that Z ∈ Z(A) if and only if Z ∗ ∈ Z(A). ZS(A) = {A ∈ Z(A) : A = A∗}. An element A ∈ A is symmetric (respectively, skew symmetric) if A∗ = A (respectively, if A∗ = −A). If A is a unital -algebra with unit I over a field F, let S (resp. K) be the set of its symmetric (resp. skew symmetric) elements. Then every element A ∈ A can be written as A = S + K, where S ∈ S and K ∈ K. Moreover, this decomposition is unique. We say that the involution ∗ is of the first kind if FI ⊆ S; equivalently, ∗ is an F-linear map. Otherwise, ∗ is said to be of the second kind (see [3]). Recall that a ring A is prime if, for any A, B ∈ A, AAB = 0 implies A = 0 or B = 0. A von Neumann algebra M is a subalgebra of some B(H), the algebra of all bounded linear operators acting on a complex Hilbert space H, which satisfies the double commutant property: M′′ = M, where M′ = {T : T ∈ B(H) and T A = AT ∀A ∈ M} and M′′ = {M′}′. For the theory of von Neumann algebras, ref. [7]. The rest part of this paper is organized as follows. In Section 2, we discuss the question in a pure algebraic setting. Let A be a unital prime -ring. Assume that Φ : A → A is a general strong skew commutativity preserving surjective map. We show that, if A contains a nontrivial symmetric idempotent element, then Φ has the form Φ(A) = A+f (A) for all A ∈ A, or Φ(A) = −A+f (A) for all A ∈ A, where f : A → ZS(A) is an arbitrary map (Theorem 2.1). Particularly, if A is a -algebra and if the involution ∗ is of the second kind, then Φ : A → A is strong skew commutativity preserving if and only if Φ(A) = A for all A ∈ A, or Φ(A) = −A for all A ∈ A (Theorem 2.10). As an application, we obtain a characterization of strong skew STRONG SKEW COMMUTATIVITY PRESERVING MAPS 3 commutativity preserving general surjective maps on factor von Neumann algebras (Theorem 2.11) which improves the main results in [6]. Section 3 is devoted to characterizing the strong skew commutativity maps on general von Neumann algebras. We prove that, if A is a von Neumann algebra without central summands of type I1, then a map Φ : A → A is strong skew commutativity preserving if and only if Φ(A) = ZA for all A ∈ A, where Z ∈ ZS(A) with Z 2 = I (Theorem 3.1). It is clear that Theorem 2.11 above is also an immediate consequence of this result. 2. SSCP maps on prime rings with involution In this section, we discuss the question of characterizing the strong skew commutativity preserving maps on prime rings with involution ∗. The following is our main result in this section. Theorem 2.1. Let A be a unital prime -ring with the unit I. Assume that A contains a nontrivial symmetric idempotent P and Φ : A → A is a surjective map. If Φ is strong skew commutativity preserving, that is, Φ(A)Φ(B) − Φ(B)Φ(A)∗ = AB − BA∗ for all A, B ∈ A, then there exists a map f : A → ZS(A), the set of central symmetric elements, such that Φ(A) = A + f (A) for all A ∈ A, or Φ(A) = −A + f (A) for all A ∈ A. To prove Theorem 2.1, we need a lemma. Let A be a prime ring. Denote by Q = Qml(A) the maximal left ring of quotients of A. Note that the center C of Q is a field which is called the extended centroid of A (see [1, 3] for details). Moreover, Z(A) ⊆ C. The following result is well-known. Lemma 2.2. ([3, Theorem A.7]) Let A be a prime ring, and let Ai, Bi, Cj, Dj ∈ Qml(A) be such that n X i=1 AiXBi = m X j=1 CjXDj for all X ∈ A. If A1, . . . , An are linearly independent over C, then each Bi is a C-linear combination of D1, . . . , Dm. Similarly, if B1, . . . , Bn are linearly independent over C, then each Ai is a C- linear combination of C1, . . . , Cm. In particular, for A, B ∈ Qml(A), if AXB = BXA for all X ∈ A, then A and B are C-linear dependent. We will prove Theorem 2.1 by a series of lemmas. Assume in the sequel that Φ : A → A is a SSCP surjective map. Lemma 2.3. Φ(ZS(A)) = ZS(A). Proof. Take any Z ∈ ZS(A). Then for any T ∈ A, we have [Φ(Z), Φ(T )]∗ = [Z, T ]∗ = 0. So Φ(Z)Φ(T ) = Φ(T )Φ(Z)∗. By the surjectivity of Φ, we get Φ(Z)X = XΦ(Z)∗ for all X ∈ A. (2.1) 4 XIAOFEI QI AND JINCHUAN HOU Take X = I in Eq.(2.1), we get Φ(Z) = Φ(Z)∗. This and Eq.(2.1) imply that Φ(Z) ∈ ZS(A). On the other hand, if there exists some A ∈ A such that Φ(A) = Z = Z ∗ ∈ ZS(A), then we get [A, S]∗ = [Φ(A), Φ(S)]∗ = 0 for all S ∈ A. Hence AS = SA∗ holds for all S, which means that A = A∗ ∈ ZS(A), that is, ZS(A) ⊆ Φ(ZS(A)), completing the proof. (cid:3) Lemma 2.4. For any T, S ∈ A, there exists an element ZT,S ∈ ZS(A) depending on T, S such that Φ(T + S) = Φ(T ) + Φ(S) + ZT,S. Proof. For any T, S, R ∈ A, we have [Φ(T + S) − Φ(T ) − Φ(S), Φ(R)]∗ = [Φ(T + S), Φ(R)]∗ − [Φ(T ), Φ(R)]∗ − [Φ(S), Φ(R)]∗ = [T + S, R]∗ − [T, R]∗ − [S, R]∗ = 0. By the surjectivity of Φ and the above equation, one sees that ZT,S = Φ(T +S)−Φ(T )−Φ(S) ∈ ZS(A) holds for all T, S ∈ A. (cid:3) In the following, we will use the technique of Peirce decomposition. By the assumption, we can take a symmetric nontrivial idempotent element P in A. Set A11 = P AP , A12 = P A(I − P ), A21 = (I − P )AP and A22 = (I − P )A(I − P ). Then A = A11 + A12 + A21 + A22. It is clear that A∗ ij = Aji, i, j = 1, 2. For an element Sij ∈ A, we always mean that Sij ∈ Aij. Lemma 2.5. Φ(P )∗ = Φ(P ) and Φ(I − P )∗ = Φ(I − P ). Moreover, there exist elements α, β, µ ∈ C with α 6= 0 such that Φ(P ) = αP + µI and Φ(I − P ) = α(I − P ) + βI. Proof. For any A ∈ A, it is easy to check that [P, [P, [P, A]∗]∗]∗ = [P, A]∗. So we have [P, [P, [Φ(P ), Φ(A)]∗]∗]∗ = [Φ(P ), Φ(A)]∗. It follows from the surjectivity of Φ that [P, [P, [Φ(P ), A]∗]∗]∗ = [Φ(P ), A]∗ for all A ∈ A. Write Φ(P ) = S11 + S12 + S21 + S22. Then the above equation becomes P AS ∗ 11 + P AS ∗ 12 − P AS ∗ 21 − P AS ∗ 22 −S11AP − S12AP + S21AP + S22AP (2.2) = S21A + S22A − AS ∗ 21 − AS ∗ 22 for all A ∈ A. Taking A = A12 in Eq.(2.2), we get A12S ∗ 12 = S21A12 = 0, that is, P A(I − P )S ∗ 12 = S21P A(I − P ) = 0 for all A ∈ A. It follows from the primeness of A that S ∗ 12 = S21 = 0, and so S12 = S21 = 0. Now let A = A11 in Eq.(2.2), and we get A11S ∗ 11 = S11A11. This implies that S11 = S ∗ 11 by taking A11 = P . So P AS11 = S11AP holds for all A ∈ A. It follows from Lemma 2.2 that S11 = λP for some λ ∈ C. STRONG SKEW COMMUTATIVITY PRESERVING MAPS 5 Similarly, taking A = A22 in Eq.(2.2), and one can obtain S22 = S ∗ 22 and S22 = µ(I − P ) for some µ ∈ C. Hence Φ(P ) = S11 + S22 = S ∗ 11 + S ∗ 22 = Φ(P )∗ and Φ(P ) = S11 + S22 = λP + µ(I − P ) = αP + µI, where α = λ − µ ∈ C. It is obvious that µI ∈ C. Note that Φ(I) ∈ ZS(A), Φ(I) − Φ(P ) − Φ(I − P ) ∈ ZS(A) and ZS(A) ⊆ C. So Φ(I − P )∗ = Φ(I − P ) and there exists an element β ∈ C such that Φ(I − P ) = α(I − P ) + βI. Finally, we still need to prove that α 6= 0. On the contrary, if α = 0, then Φ(P ) = µI ∈ C. Since Φ(P ) ∈ A, it follows that Φ(P ) ∈ ZS(A). By Lemma 2.2, we get P ∈ ZS(A), which is impossible as A is prime. The proof is finished. (cid:3) Note that C is a field as A is prime ([3, Theorem A.6]). So α ∈ C is invertible. In the following, let λ = α−1 ∈ C. Also note that the unit 1 of C is the same to the unit I of A Lemma 2.6. For any Aij ∈ Aij, we have Φ(Aij) = λAij, 1 ≤ i 6= j ≤ 2. Proof. Take any A12 ∈ A12 and let Φ(A12) = S11 +S12+S21+S22. Since [Φ(P ), Φ(A12)]∗ = [P, A12]∗ = A12, by Lemma 2.5, we get αS12 − αS21 = A12, which implies that S21 = 0 and S12 = α−1A12 = λA12. For any B ∈ A, by the surjectivity of Φ, there exists an element T = T11+T12+T21+T22 ∈ A such that Φ(T ) = B. Since [B, Φ(A12)]∗ = [Φ(T ), Φ(A12)]∗ = [T, A12]∗, we have BS11 + B(λA12) + BS22 − S11B ∗ − λA12B ∗ − S22B ∗ = T11A12 + T21A12 − A12T ∗ 12 − A12T ∗ 22. (2.3) Multiplying by P from the right in Eq.(2.3), one gets BS11 − S11B ∗P − λA12B ∗P − S22B ∗P = −A12T ∗ 12. Replacing B by (I − P )BP in the above equation, we have (I − P )BP S11 = −A12T ∗ 12, which implies that (I − P )BP S11 = 0 for all B ∈ A. It follows from the primeness of A that S11 = 0. Similarly, replacing B by P B(I − P ) and multiplying by I − P from the left in Eq.(2.3), one can show that S22 = 0. Hence we obtain Φ(A12) = S12 = λA12. The proof of Φ(A21) = λA21 is similar, and we omit it here. (cid:3) Lemma 2.7. For any Aii ∈ Aii, we have Φ(Aii) = αAii, i = 1, 2. Proof. Still, we only need to prove that the lemma is true for A11. Take any A11 ∈ A11 and let Φ(A11) = S11 + S12 + S21 + S22. Since [Φ(P ), Φ(A11)]∗ = [P, A11]∗ = 0, by Lemma 2.5, we have αS11 + αS12 − S11(αP ) − S21(αP ) = 0. 6 XIAOFEI QI AND JINCHUAN HOU This implies αS12 = 0, an so S12 = 0. On the other hand, since [Φ(I − P ), Φ(A11)]∗ = [I − P, A11]∗ = 0, by Lemma 2.5 again, one gets αS21 = 0, which implies that S21 = 0. For any B12 ∈ A12, by Lemma 2.6, we have [λB12, Φ(A11)]∗ = [Φ(B12), Φ(A11)]∗ = [B12, A11]∗ = 0, that is, λB12S22 − S11(λB12)∗ − S22(λB12)∗ = 0. Note that λB12 ∈ A12. It follows that λB12S22 = 0, which implies B12S22 = 0. That is, P B(I − P )S22 = 0 for all B ∈ A. As A is prime, we get S22 = 0. For any B21 ∈ A21, by Lemma 2.6, we have [λB21, Φ(A11)]∗ = [Φ(B21), Φ(A11)]∗ = [B21, A11]∗, that is, λB21S11 − S11(λB21)∗ = B21A11 − A11B ∗ 21. This forces B21(λS11 − A11) = (λS11 − A11)B ∗ 21 = 0. So we get (I − P )BP (λS11 − A11) = 0 for each B ∈ A. It follows from the primeness of A that λS11 = A11. Hence Φ(A11) = S11 = λ−1A11 = αA11, completing the proof. Lemma 2.8. α = λ, and hence α2 = 1 and α = ±1. (cid:3) Proof. For any A12 ∈ A12 and A21 ∈ A21, by the definition of Φ and Lemma 2.6, we have A12A21 = [A12, A21]∗ = [Φ(A12), Φ(A21)]∗ = [λA12, λA21]∗ = (λA12)(λA21) = λ2A12A21. It follows that (λ2 −1)A12A21 = 0. First fix A12. Then the equation becomes (λ2 −1)A12AP = 0 for all A ∈ A. Assume that λ2 − 1 6= 0. Since C is a field, we get that λ2 − 1 is invertible. So we have A12AP = 0 for all A ∈ A. Since A is prime, it follows that A12 = 0, that is, P A(I − P ) = 0 for all A ∈ A. This implies that either P = 0 or P = I, a contradiction. So λ2 = 1 and λ = α. Since C is a field, we see that α = 1 or −1, completing the proof. (cid:3) Lemma 2.9. There exists a map f : A → Z(A) such that Φ(A) = A + f (A) for all A ∈ A or Φ(A) = −A + f (A) for all A ∈ A. Proof. By Lemmas 2.6-2.8, we have proved that Φ(Aij) = Aij for all Aij ∈ Aij or Φ(Aij) = −Aij for all Aij ∈ Aij (i, j = 1, 2). Now, for any A = A11 + A12 + A21 + A22 ∈ A, by Lemma 2.4, there exists some element ZA ∈ ZS(A) such that Φ(A) − (Φ(A11) + Φ(A12) + Φ(A21) + Φ(A22)) = Φ(A11 + A12 + A21 + A22) − Φ(A11) − Φ(A12) − Φ(A21) − Φ(A22) = ZA. Define a map f : A → ZS(A) by f (A) = ZA for each A ∈ A. Then Φ(A) = A + f (A) for each A, or Φ(A) = −A + f (A) for each A. (cid:3) By applying Theorem 2.1, we give a characterization of SSCP maps on prime -algebras with the second kind involution. STRONG SKEW COMMUTATIVITY PRESERVING MAPS 7 Theorem 2.10. Let A be a unital prime -algebra over a field F with a nontrivial symmetric idempotent P . Assume that the involution ∗ is of the second kind and Φ : A → A is a surjective map. Then Φ is strong skew commutativity preserving (that is, Φ(A)Φ(B) − Φ(B)Φ(A)∗ = AB − BA∗ for all A, B ∈ A) if and only if Φ(A) = A for all A ∈ A or Φ(A) = −A for all A ∈ A. Proof. Obviously, the "if" part is true. For the "only if" part, by Theorem 2.1, there exists a map f : A → ZS(A) such that Φ(A) = A + f (A) for all A ∈ A, or Φ(A) = −A + f (A) for all A ∈ A. So, to complete the proof of the theorem, we only need to prove f ≡ 0. In fact, Φ is SSCP implies that f (B)(A − A∗) = 0 for all A, B ∈ A. (2.4) Since ∗ is of the second kind, there exists a nonzero ǫ ∈ F such that (ǫI)∗ = −ǫI. Thus, let A = ǫI in Eq.(2.4), we get 2ǫf (B) = 0, which implies f (B) = 0 for each B ∈ A as F is a field. The proof is finished. (cid:3) Recall that a von Neumann algebra M is called a factor if its center is trivial (i.e., Z(M) = CI). Note that factor von Neumann algebras are prime and ∗ is of the second kind. So, as an application of Theorem 2.10 to the factor von Neumann algebras case, we improve the main results of [6] immediately. Theorem 2.11. Let A be a factor von Neumann algebra. Assume that Φ : A → A is a surjective map. Then Φ is strong skew commutativity preserving if and only if Φ(A) = αA for all A ∈ A, where α ∈ {1, −1}. 3. SSCP maps on von Neumann algebras In this section, we will discuss the strong skew commutativity preserving maps on von Neumann algebras. The following is the main result of this section. Theorem 3.1. Let M be a von Neumann algebra without central summands of type I1. Assume that Φ : M → M is a surjective map. Then Φ is strong skew commutativity preserving if and only if there exists an element Z ∈ ZS(M) with Z 2 = I such that Φ(A) = ZA for all A ∈ A. Obviously, Theorem 2.11 is also an immediate consequence of the above result. We remark that the methods used in the proof of Theorem 2.1 are not valid here since the von Neumann algebras in Theorem 3.1 may not be prime. In order to overcome the difficulties caused by the absence of primeness, we need some deep results coming from the theory of von Neumann algebras. 8 XIAOFEI QI AND JINCHUAN HOU Let M be a von Neumann algebra and A ∈ M. Recall that the central carrier of A, denoted by A, is the intersection of all central projections P such that P A = A. If A is self-adjoint, then the core of A, denoted by A, is sup{S ∈ Z(M) : S = S ∗, S ≤ A}. Particularly, if A = P is a projection, it is clear that P is the largest central projection ≤ P . A projection P is said to be core-free if P = 0 [8]. It is easy to see that P = 0 if and only if I − P = I. The following lemmas are useful for our purpose, where Lemma 3.2 and Lemma 3.3 are proved in [8]. Lemma 3.2. ([8]) Let M be a von Neumann algebra without central summands of type I1. Then each nonzero central projection C ∈ M is the carrier of a core-free projection in M. Particularly, there exists a nonzero core-free projection P ∈ M with P = I. Lemma 3.3. ([8]) Let M be a von Neumann algebra. For projections P, Q ∈ M, if P = Q 6= 0 and P + Q = I, then T ∈ M commutes with P XQ and QXP for all X ∈ M implies that T ∈ Z(M). Lemma 3.4. Let M be a von Neumann algebra. Assume that P ∈ M is a projection satisfying P = 0 and P = I. Then, for any Z ∈ Z(M), ZP M(I − P ) = {0} implies Z = 0. Proof. Assume that M ⊆ B(H), where H is a Hilbert space, and assume that Z ∈ Z(M) with Z 6= 0 such that ZP M(I−P ) = {0}. Let Q be the projection onto the closure of the range of Z. It is clear that Q ∈ Z(M). So we may write M = QMQ⊕(I −Q)M(I −Q) = M1⊕M2. Thus, according to the space decomposition H = QH ⊕ (I − Q)H, we have A =   A1 0 0 A2   , Z =   Z1 0 0 0   , P =   P1 0 0 P2   , where A ∈ M is arbitrary, Ai, Pi ∈ Mi with Pi = P ∗ i = P 2 i (i ∈ {1, 2}) and Z1 ∈ Z(M1) is injective with dense range. It follows that ZP A(I − P ) =   Z1P1A1(I1 − P1) 0 0 0   . By the assumption, we get Z1P1A1(I1 − P1) = 0, which implies that P1A1(I1 − P1) = 0 for all A1 ∈ M1 as Z1 is injective. So P1A1 = P1A1P1 for each A1 ∈ M1. Thus, by the following claim, we have P1 ∈ Z(M1). Claim. Let N be a von Neumann algebra. Assume that P ∈ N is a projection satisfying P N (I − P ) = {0}. Then P ∈ Z(N ). In fact, for such P , write P1 = P and P2 = I − P . Denote Nij = PiN Pj, i, j ∈ {1, 2}. Then N = N11 +N12 +N21 +N22. For any A ∈ N , we have P A = P AP +P A(I −P ). It follows from the assumption P N (I − P ) = {0} that P A(I − P ) = 0 for any A ∈ N . Since A is arbitrary, STRONG SKEW COMMUTATIVITY PRESERVING MAPS 9 it is true that P A∗(I − P ) = 0 holds for any A ∈ N , which implies that (I − P )AP = 0 for any A ∈ N . Thus, for any A ∈ N , we must have A = P AP + (I − P )A(I − P ). Now it is clear that P A = AP for each A, that is, the claim is true. Let us go back to the proof of the lemma and let Q0 =   P1 0 0 0  . Obviously, Q0 is a central projection with Q0 ≤ P . Since P = 0, we have Q0 = 0, and so P1 = 0. This yields QP = 0, which implies I − Q ≥ P = I, a contradiction. Hence Z = 0 and the proof is completed. (cid:3) Now we are in a position to give our proof of Theorem 3.1. Proof of Theorem 3.1. Still, we only need to prove the "only if" part. By the same argument as that of Lemmas 2.3-2.4, one can obtain that Φ(ZS(M)) = ZS(M) (3.1) and Φ(T + S) = Φ(T ) + Φ(S) + ZT,S for all T, S ∈ M, (3.2) where ZT,S ∈ ZS(M) depending on T, S. By Lemma 3.2, there is a non-central core-free projection P with central carrier I. For such a P , by the definitions of core and central carrier, I − P is also core-free with I − P = I. Denote Mij = PiMPj (i, j ∈ {1, 2}), where P1 = P and P2 = I − P . Then M = M11 + M12 + M21 + M22. In all that follows, when writing Sij, it always indicates Sij ∈ Mij. We will complete the proof by several steps. Step 1. Φ(P ) = Φ(P )∗. For the identity operator I, by Eq.(3.1), there exists some Z ∈ ZS(M) such that Φ(Z) = I. Thus 0 = [P, Z]∗ = [Φ(P ), Φ(Z)]∗ = Φ(P ) − Φ(P )∗, which implies Φ(P ) = Φ(P )∗. Step 2. There exist elements Z1, Z2, Z3 ∈ ZS(M) with Z1 6= 0 such that Φ(P ) = Z1P + Z2 and Φ(I − P ) = Z1(I − P ) + Z3. Note that [P, [P, [P, A]∗]∗]∗ = [P, A]∗ holds for any A ∈ M. Thus for every A, we have [P, [P, [Φ(P ), Φ(A)]∗]∗]∗ = [Φ(P ), Φ(A)]∗. It follows from the surjectivity of Φ that [P, [P, [Φ(P ), A]∗]∗]∗ = [Φ(P ), A]∗ for all A ∈ M. (3.3) 10 XIAOFEI QI AND JINCHUAN HOU Write [Φ(P ), A]∗ = B. Then Eq.(3.3) becomes P B − 2P BP + BP = B. Multiplying by P and I − P from both sides in the above equation, respectively, one gets P BP = 0 and (I − P )B(I − P ) = 0. Therefore P (Φ(P )A − AΦ(P )∗)P = 0 and (I − P )(Φ(P )A − AΦ(P )∗)(I − P ) = 0 (3.4) hold for all A ∈ M. Write Φ(P ) = S11 + S12 + S21 + S22. Replacing A by P T (I − P ) for any T ∈ M in Eq.(3.4), we get P T (I − P )S ∗ 12 = 0 and S21P T (I − P ) = 0, which, together with Φ(P ) = Φ(P )∗, implies that P T (I − P )S21 = S21P T (I − P ) = 0 (3.5) holds for all T ∈ M. It is obvious that S21(I − P )T P = (I − P )T P S21 = 0. Hence, by Lemma 3.3, we see that S21 ∈ Z(M), which forces S21 = 0. Similarly, replacing A by (I − P )T P for any T ∈ M in Eq.(3.4) and applying Step 1, it is easily checked S12 = 0. Now taking A = P in Eq.(3.3), and by Step 1, one gets S11P T P = P T P S ∗ 11 = P T P S11, which means S11 ∈ P ZS(M). Symmetrically, by taking A = I − P in Eq.(3.3), we obtain S22 ∈ (I − P )ZS(M). Write S11 = Z11P and S22 = Z22(I − P ), where Z11, Z22 ∈ ZS(M). Then Φ(P ) = S11 + S22 = Z11P + Z22(I − P ) = (Z11 − Z22)P + Z22. Note that Φ(I) − Φ(I − P ) − Φ(P ) ∈ ZS(M) by Eqs.(3.1)-(3.2). So there exists some Z33 ∈ ZS(M) such that Φ(I − P ) = Z33 − Φ(P ) = Z33 − Z22 − (Z11 − Z22)P = (Z11 − Z22)(I − P ) + (Z33 − Z11). Let Z1 = Z11 − Z22, Z2 = Z22 and Z3 = Z33 − Z11. Thus Φ(P ) = Z1P + Z2 and Φ(I − P ) = Z1(I − P ) + Z3. Finally, we still need to prove that Z1 6= 0. On the contrary, if Z1 = 0, then Φ(P ) = Z2 ∈ Z(M). By Eq.(3.1), we get P ∈ ZS(M), which is impossible. Step 3. If Φ(T ) = P and Φ(S) = I − P , where T = T11 + T12 + T21 + T22 ∈ M and S = S11 + S12 + S21 + S22 ∈ M, then T12 = T21 = 0 and S12 = S21 = 0. STRONG SKEW COMMUTATIVITY PRESERVING MAPS 11 In fact, by the equation [P, T ]∗ = [Φ(P ), Φ(T )]∗ = [Φ(P ), P ]∗ and Step 2, one can get T12 = T21 = 0; by the equation [I − P, S]∗ = [Φ(I − P ), Φ(S)]∗ = [Φ(I − P ), I − P ]∗ and Step 2, one can get S12 = S21 = 0. Step 4. For any Aij ∈ Mij, we have Φ(Aij) ∈ Mij, 1 ≤ i 6= j ≤ 2. Moreover, Z1P Φ(A12)(I − P ) = A12 holds for all A12 ∈ M12 and Z1(I − P )Φ(A21)P = A21 holds for all A21 ∈ M21. We only check the assertion for A12, and the case of A21 is similarly dealt with. For any A12, write Φ(A12) = S11 + S12 + S21 + S22. By Step 3, there exists some T = T11 + T22 ∈ M such that Φ(T ) = P . Then [P, Φ(A12)]∗ = [T, A12]∗, that is, S12 − S21 = T11A12 − A12T ∗ 22. Multiplying by I − P and P from the left side and the right side respectively in the above equation, one gets S21 = (I − P )Φ(A12)P = 0. (3.6) On the other hand, since [Φ(A12), P ]∗ = [A12, T ]∗, by Eq.(3.6), we have S11 − S ∗ 11 = A12T22 − T22A∗ 12. Multiplying by P from both sides of the equation, it follows that S11 = S ∗ 11. (3.7) Similarly, by using the equation [Φ(A12), I − P ]∗ = [Φ(A12), Φ(S)]∗ = [A12, S]∗ and Step 3, it is easily checked that S22 = S ∗ 22. (3.8) Now, for any X ∈ M, by the surjectivity of Φ, there exists an element R = R11 + R12 + R21 + R22 ∈ M such that Φ(R) = X. Since [X, Φ(A12)]∗ = [Φ(R), Φ(A12)]∗ = [R, A12]∗, applying Eq.(3.6), we get XS11 + XS12 + XS22 − S11X ∗ − S12X ∗ − S22X ∗ = R11A12 + R21A12 − A12R∗ 12 − A12R∗ 22. (3.9) Replacing X by P B(I − P ) + (I − P )BP for any B ∈ M, and multiplying by I − P and P from the left and the right respectively in Eq.(3.9), one obtains (I − P )BP S11 = S22(I − P )B ∗P holds for all B ∈ M; (3.10) Replacing X by (I − P )BP for any B ∈ M, and multiplying by I − P and P from the left and the right respectively in Eq.(3.9), one gets (I − P )BP S11 = 0 holds for all B ∈ M. (3.11) 12 XIAOFEI QI AND JINCHUAN HOU Equivalently, (I − P )B ∗P S11 = 0 holds for all B ∈ M. (3.12) By Eqs.(3.7) and (3.12), we have S11P B(I − P ) = S ∗ 11P B(I − P ) = ((I − P )B ∗P S11)∗ = 0. (3.13) Moreover, it is obvious that P B(I − P )S11 = S11(I − P )BP = 0. This, together with Eqs.(3.11), (3.13) and Lemma 3.3, yields S11 ∈ ZS(M). Hence (3.14) (3.15) Combining Eq.(3.10) with Eq.(3.11), we get S11 = 0. S22(I − P )B ∗P = 0 for all B ∈ M. Equivalently, S22(I − P )BP = 0 for all B ∈ M. It follows from Eqs.(3.8) and (3.15) that P B(I − P )S22 = P B(I − P )S ∗ 22 = (S22(I − P )B ∗P )∗ = 0. Thus we have P B(I − P )S22 = S22P B(I − P ) = S22(I − P )BP = (I − P )BP (I − P )S22 = 0, which implies that S22 ∈ ZS(M) by Lemma 3.3, and so S22 = 0. (3.16) Combining Eqs.(3.6), (3.14) and (3.16), we see that Φ(A12) ∈ M12, as desired. Finally, since [Φ(P ), Φ(A12)]∗ = [P, A12]∗, by Step 2, it is easy to check that A12 = Z1S12 = Z1P Φ(A12)(I − P ). Step 5. For any Aii ∈ Mii, we have Φ(Aii) = Z1Aii, i = 1, 2. Still, we only prove that Φ(A11) = Z1A11 holds for all A11 ∈ A11. The case of i = 2 is checked similarly. Take any A11 ∈ M11 and any B ∈ M. Write Φ(A11) = S11 + S12 + S21 + S22. On the one hand, since 0 = [P B(I − P ), A11]∗ = [Φ(P B(I − P )), Φ(A11)]∗, by Step 4, we have P Φ(P B(I − P ))(I − P )S21 + P Φ(P B(I − P ))(I − P )S22 −S12(I − P )Φ(P B(I − P ))∗P − S22(I − P )Φ(P B(I − P ))∗P = 0, which implies that P Φ(P B(I − P ))(I − P )S21 = S12(I − P )Φ(P B(I − P ))∗P, STRONG SKEW COMMUTATIVITY PRESERVING MAPS 13 and P Φ(P B(I − P ))(I − P )S22 = 0 S22(I − P )Φ(P B(I − P ))∗P = 0. Multiplying by Z1 ∈ ZS(M) in the above three equations, and applying Step 4, one gets and P B(I − P )(I − P )S21 = S12(I − P )B ∗P, P B(I − P )S22 = 0 S22(I − P )B ∗P = 0 hold for all B ∈ M. Furthermore, Eq.(3.19) yields S22(I − P )BP = 0 for all B ∈ M. (3.17) (3.18) (3.19) (3.20) Note that (I − P )BP (I − P )S22 = 0 and S22P B(I − P ) = 0. These, together with Lemma 3.3, Eqs.(3.18) and (3.20), imply S22 ∈ Z(M). So we must have S22 = (I − P )Φ(A11)(I − P ) = 0. (3.21) On the other hand, by using the equation [A11, P B(I − P )]∗ = [Φ(A11), Φ(P B(I − P ))]∗, we get It follows that and A11P B(I − P ) = S11P Φ(P B(I − P ))(I − P ) +S21P Φ(P B(I − P ))(I − P ) −P Φ(P B(I − P ))(I − P )S ∗ 12. A11P B(I − P ) = S11P Φ(P B(I − P ))(I − P ), S21P Φ(P B(I − P ))(I − P ) = 0 Multiplying by Z1 in the above three equations leads to P Φ(P B(I − P ))(I − P )S ∗ 12 = 0. and (Z1A11 − S11)P B(I − P ) = 0, S21P B(I − P ) = 0 P B(I − P )S ∗ 12 = 0 for all B ∈ M. Eq.(3.24) implies that S12(I − P )B ∗P = 0, (3.22) (3.23) (3.24) (3.25) 14 and so XIAOFEI QI AND JINCHUAN HOU S12(I − P )BP = 0 for all B ∈ M. (3.26) Eq.(3.17) and Eq.(3.25) together yield P B(I − P )S21 = 0 for all B ∈ M. (3.27) It is obvious that (I − P )BP (I − P )S21 = S21(I − P )BP = 0. By Eqs.(3.23), (3.27) we get S21 ∈ Z(M) and hence S21 = (I − P )Φ(A11)P = 0. (3.28) Now consider the equation [(I − P )BP, A11]∗ = [Φ((I − P )BP ), Φ(A11)]∗. It follows from Step 4, Eqs.(3.21) and (3.28) that (I − P )BP A11 − A11P B ∗(I − P ) = (I − P )Φ((I − P )BP )P S11 +(I − P )Φ((I − P )BP )P S12 − S11P Φ(P B(I − P ))∗(I − P ). This implies that and (I − P )BP A11 = (I − P )Φ((I − P )BP )P S11 (I − P )Φ((I − P )BP )P S12 = 0. Multiplying by Z1 in the above two equations, we obtain and (I − P )BP (Z1A11 − S11) = 0 (I − P )BP S12 = 0. (3.29) (3.30) Since S12P B(I − P ) = P B(I − P )P S12 = 0, by Eqs.(3.26) and (3.30), we get S12 ∈ Z(M), and so S12 = P Φ(A11)(I − P ) = 0. (3.31) By Eqs.(3.22) and (3.29), also noting that (Z1A11 − S11)(I − P )BP = P B(I − P )(Z1A11 − S11) = 0, we see that Z1A11 − S11 ∈ Z(M), and hence Z1A11 = S11 = P Φ(A11)P. (3.32) Now it is clear by Eqs.(3.21), (3.28), (3.31) and (3.32) that Φ(A11) = Z1A11. Step 6. Z 2 1 = I. STRONG SKEW COMMUTATIVITY PRESERVING MAPS 15 For any A, B ∈ M, by the assumptions on Φ and Step 4, we have P A(I − P )BP = [P A(I − P ), (I − P )BP ]∗ = [Φ(P A(I − P )), Φ((I − P )BP )]∗ = P Φ(P A(I − P ))(I − P )Φ((I − P )BP )P. Multiplying by Z 2 1 in the above equation and applying Step 4 again, one gets Z 2 1 P A(I −P )(I − P )BP = P A(I − P )(I − P )BP . Let us fix A. Then the equation becomes (Z 2 1 − I)P A(I − P )(I − P )BP = 0 for all B ∈ M. (3.33) Similarly, by using the equation [(I − P )BP, P A(I − P )]∗ = [Φ((I − P )BP ), Φ(P A(I − P ))]∗, one obtains (I − P )BP (Z 2 1 − I)P A(I − P ) = (Z 2 1 − I)(I − P )BP P A(I − P ) = 0 for all B ∈ M. (3.34) By Lemma 3.3, Eqs.(3.33)-(3.34) imply that (Z 2 1 − I)P A(I − P ) ∈ Z(M). Hence (Z 2 1 − I)P A(I − P ) = 0 holds for all A ∈ M. Now by Lemma 3.4, one gets Z 2 1 = I. Step 7. Φ(A) = ZA for all A ∈ M, where Z ∈ ZS(M) with Z 2 = I. Therefore, Theorem 3.1 is true. By Steps 4-6, we have proved that Φ(Aij) = Z1Aij for all Aij ∈ Mij, where Z1 ∈ ZS(M) with Z 2 1 = I (i, j = 1, 2). Define Z = Z1. Now, by a similar argument to that in the proof of Lemma 2.9, one can show that there exists a map f : M → ZS(M) such that Φ(A) = ZA + f (A) for all A ∈ M. To complete the proof of the theorem, we have to prove f ≡ 0. Indeed, since [Φ(A), Φ(B)]∗ = [A, B]∗ for every A, B ∈ M, we have [ZA + f (A), ZB + f (B)]∗ = [A, B]∗. It follows that Zf (B)(A − A∗) = 0. As Z 2 = I, we get f (B)(A − A∗) = 0 for all A, B ∈ M. Take A = iI in the above equation, one gets 2if (B) = 0, and so f (B) = 0 for every B ∈ M, completing the proof of Theorem 3.1. (cid:3) References [1] K. I. Beidar, W. S. Martindale 3rd, A. V. Mikhalev, Rings with Generalized Identities, Marcel Dekker, Inc., New York-Basel-Hong Kong, 1996. [2] M. Bresar, M. Fsoner, On ring with involution equipped with some new product, Publ. Meth. Debrecen, 57 (2000), 121-134. [3] M. Bresar, M. A. Chebotar and W. S. Martindale III, Functional identities, Birkhauser Basel, 2007. [4] M. A. Chebotar, Y. Fong, P.-H. Lee, On maps preserving zeros of the polynomial xy − yx∗, Lin. Alg. Appl., 408 (2005), 230-243. 16 XIAOFEI QI AND JINCHUAN HOU [5] J. Cui, J. Hou, Linear maps preserving elements annihilated by a polymnomial XY − Y X †, Studia Math.174(2) (2006), 183-199. [6] J. Cui, C. Park, Maps preserving strong skew Lie product on factor von Neumann algebras, Acta Math. Sci., 32B(2) (2012), 531-538. [7] R. V. Kadison, J. R. Ringrose, Fundamentals of the Theory of Operator Algebras, Vol. I, Academic Press, New York, 1983, Vol. II, Academic Press, New York, 1986. [8] C. R. Miers, Lie isomorphisms of operator algebras, Pacific J. Math., 38 (1971), 717-735. [9] L. Moln´ar, A condition for a subspace of B(H) to be an ideal, Lin. Alg. Appl. 235 (1996), 229-234. [10] P. Semrl, Quadratic functionals and Jordan -derivations, Studia Math. 97 (1991), 157-165. [11] P. Semrl, On Jordan *-derivations and an application, Colloq. Math. 59 (1990) 241-251. (Xiaofei Qi) Department of Mathematics, Shanxi University, Taiyuan 030006, P. R. China. E-mail address: [email protected] (Jinchuan Hou) Department of Mathematics, Taiyuan University of Technology, Taiyuan 030024, P. R. of China E-mail address: [email protected]
1805.05637	1	1805	2018-05-15T08:28:51	The factor type of conservative KMS weights on graph C*-algebras	[ "math.OA" ]	We determine the factor generated by the GNS-representation defined by a KMS-weight for a generalized gauge action on a simple graph C*-algebra when the corresponding measure on the path space of the graph is conservative for the shift. This is an improvment of Theorem 3.2 in arXives 1412.6762.	math.OA	math	THE FACTOR TYPE OF CONSERVATIVE KMS WEIGHTS ON GRAPH C ∗-ALGEBRAS KLAUS THOMSEN 1. Introduction In certain quantum statistical models the observables are represented by the el- ements of a C ∗-algebra and the time-evolution by a one-parameter group of au- tomorphisms on the C ∗-algebra, [BR]. In such models the equilibrium states are characterized by the KMS-condition, [HWH], and much work has been devoted to the study of these KMS-states. As a result there are now large classes of C ∗-algebras and one-parameter groups for which the structure of the equilibrium states is com- pletely understood. An extremal KMS-state, and more generally also an extremal KMS-weight, gives rise in a canonical way to a representation of the C ∗-algebra which generates a von Neumann algebra factor. Some of the papers on KMS-states have included a determination of this factor for the extremal KMS-states. This is the case in [EFW], [BC], [O], [BF] [IKW], [N], [LN], [Y1], [LLNSW], [CPPR], [KW], [Th1], [Th2], [Th4],[Y2] and [I]. 1 The factor types give a natural way to distinguish between KMS states very similar to the type classification of non-singular transfor- mations in ergodic theory. The purpose with this paper is to determine the factor types for a class of KMS-weights and KMS-states that arise from generalized gauge actions on simple graph C ∗-algebras. In [Th3] the author studied such KMS-weights and among other things it was shown that they can be naturally divided into three classes depending on the properties of the measure they induce on the unit space of the groupoid underlying the C ∗-algebra. For the KMS-weights we consider here the measure is concentrated on the space of infinite paths in the graph and it is conser- vative with respect to the action of the shift. We call them therefore conservative KMS-weights. We determine their factor type when the graph C ∗-algebra is simple by calculating the Γ-invariant of Connes for the associated factors. As shown in [Th3] there is a bijective correspondence between the rays of KMS-weights and the KMS-states on a corner defined by a vertex in the graph, at least for the cases we consider in this paper, and we show therefore first that the Γ-invariant for the factor defined by an extremal KMS-weight is the same as the Γ-invariant for the factor obtained from the corresponding KMS-state on the corner. This is not surprising, but it is helpful because it shows that our calculation of the Γ-invariant covers two cases. Acknowledgement I am grateful to Johannes Christensen for discussions and help to eliminate mistakes. The work was supported by the DFF-Research Project 2 'Automorphisms and Invariants of Operator Algebras', no. 7014-00145B. Version: November 9, 2018. 1Despite some effort to make this a complete list, presumably it is not. 1 2 KLAUS THOMSEN 2. The setting 2.1. Factors from KMS-weights. Let A be a C ∗-algebra and A+ the convex cone of positive elements in A. A weight on A is a map ψ : A+ → [0,∞] with the properties that ψ(a + b) = ψ(a) + ψ(b) and ψ(λa) = λψ(a) for all a, b ∈ A+ and all λ ∈ R, λ > 0. By definition ψ is densely defined when {a ∈ A+ : ψ(a) < ∞} is dense in A+ and lower semi-continuous when {a ∈ A+ : ψ(a) ≤ α} is closed for all α ≥ 0. We will use [Ku], [KV1] and [KV2] as our source of information on weights, and as in [KV2] we say that a weight is proper when it is non-zero, densely defined and lower semi-continuous. Let ψ be a proper weight on A. Set Nψ = {a ∈ A : ψ(a∗a) < ∞} and note that N ∗ ψNψ = Span{a∗b : a, b ∈ Nψ} is a dense ∗-subalgebra of A, and that there is a unique well-defined linear map ψNψ → C which extends ψ : N ∗ N ∗ ψNψ ∩ A+ → [0,∞). We denote also this densely defined linear map by ψ. Let α : R → Aut A be a point-wise norm-continuous one-parameter group of automorphisms on A. Let β ∈ R. Following [C] we say that a proper weight ψ on A is a β-KMS weight for α when i) ψ ◦ αt = ψ for all t ∈ R, and ii) for every pair a, b ∈ Nψ ∩ N ∗ ψ there is a continuous and bounded function F defined on the closed strip Dβ in C consisting of the numbers z ∈ C whose imaginary part lies between 0 and β, and is holomorphic in the interior of the strip and satisfies that F (t) = ψ(aαt(b)), F (t + iβ) = ψ(αt(b)a) for all t ∈ R. 2 A β-KMS weight ψ with the property that will be called a β-KMS state. The following is Theorem 2.4 in [Th3]. sup{ψ(a) : 0 ≤ a ≤ 1} = 1 Theorem 2.1. Let α : R → Aut A be a point-wise norm-continuous one-parameter group of automorphisms on a C ∗-algebra A. Let p be a projection in the fixed point algebra of α such that p is full in A. For all β ∈ R the map ψ 7→ ψ(p)−1ψpAp is a bijection between the set of rays of β-KMS weights for α and the β-KMS states for the restriction of α to pAp. Given a proper weight ψ on a C ∗-algebra A there is a GNS-type construction consisting of a Hilbert space Hψ, a linear map Λψ : Nψ → Hψ with dense range and a non-degenerate representation πψ of A on Hψ such that • ψ(b∗a) = hΛψ(a), Λψ(b)i , a, b ∈ Nψ, and • πψ(a)Λψ(b) = Λψ(ab), a ∈ A, b ∈ Nψ, 2Note that we apply the definition from [C] for the action α−t in order to use the same sign convention as in [BR], for example. FACTOR TYPES 3 cf. [Ku], [KV1], [KV2]. A β-KMS weight ψ on A is extremal when the only β- KMS weights ϕ on A with the property that ϕ(a) ≤ ψ(a) for all a ∈ A+ are scalar multiples of ψ, viz. ϕ = sψ for some s > 0. The following is Lemma 4.9 in [Th1]. Lemma 2.2. Let A be a separable C ∗-algebra and α a pointwise norm-continuous one-parameter group of automorphisms on A. Let ψ be an extremal β-KMS weight for α. Then πψ(A)′′ is a factor. It is shown in Section 2.2 of [KV1] that a β-KMS weight ψ extends to a normal semi-finite faithful weight ψ on πψ (A)′′ such that ψ = ψ ◦ πψ, and that the modular group on πψ(A)′′ corresponding to ψ is the one-parameter group θ on πψ(A)′′ given by θt = α−βt , where α is the σ-weakly continuous extension of α defined such that αt◦πψ = πψ◦αt. By construction αt = Ad Ut, where Ut ∈ B(Hψ) is defined by UtΛψ(a) = Λψ(αt(a)) . In the setting of Theorem 2.1, let (πϕ, Hϕ, ξϕ) be the GNS-representation of the state ϕ on pAp defined such that ϕ(x) = ψ(p)−1ψ(x) . The modular automorphism group θ′ on πϕ(pAp)′′ corresponding to the vector state defined by ξϕ is given by θ′ t (πϕ(pap)) = πϕ (α−βt(pap)) = πϕ (pα−βt(a)p) . Lemma 2.3. In the setting of Theorem 2.1 there is an ∗-isomorphism πϕ (pAp)′′ ≃ πψ(p)πψ(A)′′πψ(p) (2.1) of von Neumann algebras which is equivariant with respect to θ and θ′. Proof. Let q ∈ B(Hψ) be the orthogonal projection on Λψ (pAp) and define a unitary W : Hϕ → qHψ such that W πϕ(x)ξϕ = ψ (p)− 1 2 Λψ(x) for x ∈ pAp. Conjugation by W gives an isomorphism πϕ (pAp)′′ ≃ qπψ(A)′′q. Since (cid:18)qHψ, πψ, Λψ(p)√ψ(p)(cid:19) is the GNS-triple of ϕ it follows from Corollary 5.3.9 in [BR] that Λψ(p) is separating for πψ(p)πψ(A)′′πψ(p). Since q commutes with πψ(p)πψ(A)′′πψ(p) and qHψ contains Λψ(p), the map m 7→ mq is an isomorphism πψ(p)πψ(A)′′πψ(p) → qπψ(A)′′q . We obtain then the isomorphism (2.1) as the composition of two isomorphisms, both of which are equivariant. (cid:3) When M is a σ-finite von Neumann algebra factor every normal faithful semi- finite weight on M comes together with a modular automorphism group θ = (θt)t∈R and the Connes-invariant Γ(M) is the intersection Γ(M) =\q Sp(qMq) , 4 KLAUS THOMSEN where we take the intersection over all projections q ∈ M fixed by θ, and Sp(qMq) de- notes the Arveson spectrum of the restriction of θ to qMq. In more detail, Sp(qMq) is defined as follows. For f ∈ L1(R) define a linear map θf : qMq → qMq such that θf (a) =ZR f (t)θt(a) dt . Then where Sp(qMq) =\(cid:8)Z(f ) : f ∈ L1(R), θf (qMq) = {0}(cid:9) , Z(f ) =(cid:26)r ∈ R : ZR eitrf (t) dt = 0(cid:27) . See [Co1]. In particular, when ψ is an extremal β-KMS weight on A we can calculate Γ (πψ(A)′′) by using the automorphism group θt = α−βt and we get the following immediate corollary to Lemma 2.3. Corollary 2.4. In the setting of Lemma 2.3, Γ(cid:0)πϕ (pAp)′′(cid:1) = Γ (πψ(A)′′) . 2.2. Generalized gauge actions on graph C ∗-algebras. Let G be a countable directed graph with vertex set GV and arrow set GAr. For an arrow a ∈ GAr we denote by s(a) ∈ GV its source and by r(a) ∈ GV its range. A vertex v which does not emit any arrow is a sink, while a vertex v which emits infinitely many arrows is called an infinite emitter. The set of sinks and infinite emitters in G is denoted by V∞. An infinite path in G is an element p ∈ (GAr)N such that r(pi) = s(pi+1) for all i. i=1 ∈ (GAr)n is defined similarly. The number of A finite path µ = a1a2 · · · an = (ai)n edges in µ is its length and we denote it by µ. A vertex v ∈ GV will be considered as a finite path of length 0. We let P (G) denote the (possibly empty) set of infinite paths in G and Pf (G) the set of finite paths in G. The set P (G) is a complete metric space when the metric is given by ∞ d(p, q) = 2−iδ(pi, qi) , Xi=1 where δ(a, a) = 0 and δ(a, b) = 1 when a 6= b. We extend the source map to P (G) such that s(p) = s(p1) when p = (pi)∞ i=1 ∈ P (G), and the range and source maps to Pf (G) such that s(µ) = s(a1) and r(µ) = r(aµ) when µ ≥ 1, and s(v) = r(v) = v when v ∈ GV . Associated to the finite path µ ∈ Pf (G) is the cylinder set (2.2) Z(µ) = {(pi)∞ i=1 ∈ P (G) : pj = aj, j = 1, 2,· · · ,µ} which is an open and closed set in P (G). In particular, when µ has length 0 and hence is just a vertex v, Z(v) = {p ∈ P (G) : s(p) = v} . The C ∗-algebra C ∗(G) of the graph G was introduced in this generality in [BHRS] as the universal C ∗-algebra generated by a collection Sa, a ∈ GAr, of partial isome- tries and a collection Pv, v ∈ GV , of mutually orthogonal projections subject to the conditions that 1) S∗ 2) SaS∗ aSa = Pr(a), ∀a ∈ GAr, a ≤ Ps(a), ∀a ∈ GAr, 3) Pa∈F SaS∗ a ≤ Pv for every finite subset F ⊆ s−1(v) and all v ∈ GV , and FACTOR TYPES 5 4) Pv =Pa∈s−1(v) SaS∗ For a finite path µ = (ai)µ a, ∀v ∈ GV \V∞. i=1 ∈ Pf (G) we set Sµ = Sa1Sa2Sa3 · · · Saµ . The elements SµS∗ projections ν , µ, ν ∈ Pf (G), span a dense ∗-subalgebra A in C ∗(G). The Pµ = SµS∗ µ will play an important role in the following. Lemma 2.5. C ∗(G) is a nuclear C ∗-algebra and π(C ∗(G))′′ is a hyperfinite von Neumann algebra for all non-degenerate representations π of C ∗(G). Proof. The nuclearity of C ∗(G) follows from [KP] when G is row-finite in the sense that #s−1(v) < ∞ for all v ∈ GV , and the general case follows then from [DT]. The second statement is a wellknown consequence of the first and goes back to [CE] and [Co2]. (cid:3) We describe next the necessary and sufficient conditions which G must satisfy for C ∗(G) to be simple. These conditions were identified by Szymanski in [Sz]. A loop in G is a finite path µ ∈ Pf (G) of positive length such that r(µ) = s(µ). We will say that a loop µ has an exit then #s−1(v) ≥ 2 for at least one vertex v in µ. A subset H ⊆ GV is hereditary when a ∈ GAr, s(a) ∈ H ⇒ r(a) ∈ H, and saturated when v ∈ GV \V∞, r(s−1(v)) ⊆ H ⇒ v ∈ H . In the following we say that G is cofinal when the only non-empty subset of GV which is both hereditary and saturated is GV itself. Theorem 2.6. (Theorem 12 in [Sz].) C ∗(G) is simple if and only if G is cofinal and every loop in G has an exit. A function F : GAr → R will be called a potential on G. Using it we can define a t (cid:1)t∈R on C ∗(G) such that pointwise norm-continuous one-parameter group αF =(cid:0)αF for all a ∈ GAr and for all v ∈ GV . An action of this sort is called a generalized gauge action; the gauge action itself being the one-parameter group corresponding to the constant function F = 1. To describe the KMS-weights for a generalized gauge action, extend F to a map F : Pf (G) → R such that F (v) = 0 when v ∈ GV , and t (Sa) = eiF (a)tSa αF t (Pv) = Pv αF F (µ) = F (ai) n Xi=1 A(β)v,w =Xa e−βF (a) when µ = (ai)n over GV such that i=1 ∈ (GAr)n. For β ∈ R, define the matrix A(β) = (A(β)v,w)v,w∈GV where we sum over arrows a ∈ GAr with s(a) = v and r(a) = w. As in [Th3] we say is almost A(β)-harmonic when that a non-zero non-negative vector ψ = (ψv)v∈GV • Pw∈GV A(β)v,wψw ≤ ψv, ∀v ∈ GV , and 6 KLAUS THOMSEN A(β)v,wψw = ψv, ∀v ∈ GV \V∞. The almost A(β)-harmonic vectors ψ for which A(β)v,wψw = ψv, ∀v ∈ GV , • Pw∈GV • Pw∈GV will be called A(β)-harmonic. In particular, when G is row-finite without sinks an almost A(β)-harmonic vector is automatically A(β)-harmonic. An almost A(β)- harmonic vector which is not A(β)-harmonic will be said to be a proper almost A(β)-harmonic vector. It was shown in [Th3] that an almost A(β)-harmonic vector ϕ gives rise to a β-KMS weight Wϕ characterized by the properties that S∗ µ ∈ NWϕ and Wϕ (SµS∗ ν ) =(0, e−βF (µ)ϕr(µ), µ 6= ν µ = ν when µ, ν ∈ Pf (G), and that all gauge invariant KMS-weights arise like this. Gen- erally there can be KMS-weights that are not gauge invariant, but as shown in Proposition 5.6 of [CT] all KMS-weights are gauge invariant when C ∗(G) is simple. Therefore, in the case that concerns us here, all KMS-weights arise from almost A(β)-harmonic vectors. Borrowing terminology from harmonic analysis we say that an almost A(β)-harmonic vector ϕ is minimal when it only dominates multiples of itself, i.e. when every almost A(β)-harmonic vector ϕ′ with the property that v ≤ ϕv for all v ∈ GV is of the form ϕ′ = λϕ for some λ > 0. Thus the minimal ϕ′ almost A(β)-harmonic vectors are those for which the corresponding β-KMS weight Wϕ is extremal. The minimal almost A(β)-harmonic vectors can be subdivided in various ways. Here we shall consider three fundamental classes. The first class consists of the minimal proper almost A(β)-harmonic vectors. The other two classes consist of the A(β)-harmonic vectors and are distinguished by the properties of the measures they define on P (G). To explain this observe that by Lemma 3.7 in [Th3] an A(β)- harmonic vector ϕ defines a Borel measure mϕ on P (G) such that mϕ(Z(µ)) = e−βF (µ)ϕr(µ) (2.3) for all µ ∈ Pf (G). The Borel measures m on P (G) that arise from A(β)-harmonic vectors in this way are characterized by the two properties that • m(Z(v)) < ∞ for all v ∈ GV , and • m (σ(B ∩ Z(a))) = eβF (a)m(B∩Z(a)) for every edge a ∈ GAr and every Borel subset B of P (G). Here σ denotes the shift on P (G), defined such that σ(p)i = pi+1. Non-zero Borel measures on P (G) with these two properties are called eβF -conformal, [Th5]. They are the measures that were called harmonic β-KMS measures in [Th3]. Let G be a cofinal graph. As in [Th1] and [Th3] we say that a vertex v ∈ GV is non-wandering when there is a finite path µ ∈ Pf (G) of positive length such that v = s(µ) = r(µ). When the set NWG of non-wandering vertexes is not empty it is a hereditary subset of GV , and together with the arrows NWAr = {a ∈ GAr : s(a) ∈ NWG} FACTOR TYPES 7 they constitute a strongly connected subgraph of G which we also denote by NWG, cf. Proposition 4.3 in [Th3]. Set and P (G)rec = \a∈N WAr P (G)wan = \v∈GV {p ∈ P (G) : pi = a for infinitely many i} {p ∈ P (G) : #{i ∈ N : s(pi) = v} < ∞ } . We say that an A(β)-harmonic vector ϕ is conservative when mϕ is concentrated on P (G)rec and that ϕ is dissipative when mϕ is concentrated on P (G)wan. When NWG is empty, P (G)rec = ∅ and P (G)wan = P (G), and hence all A(β)-harmonic vectors are dissipative. To see that we have introduced a genuine dichotomy we need the following. For strongly connected graphs it is Theorem 4.10 in [Th3]. Theorem 2.7. Let G be a cofinal digraph such that NWG 6= ∅. Every eβF -conformal measure m is concentrated either on P (G)rec or on P (G)wan, and • m is concentrated on P (G)rec if and only if v,v = ∞ A(β)n ∞ Xn=0 for one, and hence all v ∈ NWG, and • m is concentrated on P (G)wan if and only if v,v < ∞ Xn=0 for one, and hence all v ∈ NWG. A(β)n ∞ Proof. Consider an eβF -conformal measure m on P (G). For every µ ∈ Pf (G), set Z(µ)P (NWG) = (cid:8)(pi)∞ Note that since m is eβF -conformal, i=1 ∈ Z(µ) : (pi)∞ i=µ+1 ∈ P (NWG) (cid:9) . m (Z(µ)P (NWG)) = e−βF (µ)m (P (NWG) ∩ Z(r(µ))) . Combined with the observation that P (G) = [µ∈Pf (G) Z(µ)P (NWG) (2.4) In short, no eβF - by Proposition 4.3 in [Th3], it follows that m(P (NWG)) 6= 0. conformal measure annihilates P (NWG). Therefore all conclusions follow from (2.4) above and Theorem 4.10 in [Th3]. (cid:3) It follows that when G is cofinal every minimal A(β)-harmonic vector is either • a proper almost A(β)-harmonic vector, • a conservative A(β)-harmonic vector or • a dissipative A(β)-harmonic vector. In this paper we focus on the conservative minimal A(β)-harmonic vectors. We call the corresponding β-KMS weights conservative. The adjective 'minimal' is su- perfluous in connection with conservative β-KMS weights because of the following 8 KLAUS THOMSEN Theorem 2.8. Assume that C ∗(G) is simple. There is a conservative A(β)-harmonic vector if and only if • NWG 6= ∅, • P∞ n=0 A(β)n v,v = ∞, and • lim supn(cid:0)A(β)n v,v(cid:1) n = 1 1 for one and hence all v ∈ NWG. When it exists it is unique up to multiplication by scalars and it is minimal. Proof. As observed above there can not be any conservative eβF -conformal measure on P (G) unless NWG 6= ∅. Therefore all statements follow by combining Theorem 2.7 above with Proposition 4.9 and Theorem 4.14 in [Th3]. (cid:3) In particular, when C ∗(G) is simple the existence of a conservative KMS-weight for αF depends only on the restriction of F to the strongly connected subgraph NWG. For many generalized gauge actions there is at most one β-value for which the three conditions in Theorem 2.8 can be met, and we ask Question 2.9. Is there a strongly connected digraph with a potential function F such that there exists a conservative β-KMS weight for αF for more than one value of β? When G is strongly connected the gauge action on C ∗(G) has a conservative β- KMS weight if and only if β is the Gurevich entropy of G and G is recurrent in the sense of Ruette, [Ru]. 3. The factor type of a conservative β-KMS weight In the setting of Section 2.2, assume that NWG 6= ∅ and pick a vertex v in NWG. Then {βF (µ) − βF (µ′) : µ, µ′ ∈ Pf (G), r(µ) = r(µ′) = s(µ) = s(µ′) = v} is a subgroup of R which does not depend on the vertex v ∈ NWG. Let RG,F be the closure in R of this subgroup. Lemma 3.1. Assume that G is cofinal and that NWG 6= ∅. Let ψ be an extremal β-KMS weight for αF . Then πψ(C ∗(G))′′ is a hyperfinite factor and Γ (πψ(C ∗(G))′′) ⊆ RG,F . Proof. M = πψ(C ∗(G))′′ is hyperfinite by Lemma 2.5. In the following proof we suppress πψ in the notation and consider C ∗(G) as a subalgebra of M. We will show that R\RG,F ⊆ R\Γ(M). Let therefore r ∈ R\RG,F and choose a function f ∈ L1(R) whose Fourier transform f satisfies that f (t) = 0 for all t ∈ RG,F and f (r) 6= 0. Fix a vertex v ∈ NWG and let µ, ν ∈ Pf (G) be finite paths with s(µ) = s(ν) = v. We assume that r(µ) = r(ν) since SµS∗ ν is zero otherwise. Since v is wandering and NWG is strongly connected there is a finite path δ in G such that s(δ) = r(µ) = r(ν) and r(δ) = v. Then (3.1) It follows that θf (SµS∗ βF (µ) − βF (ν) = βF (µδ) − βF (νδ) ∈ RG,F . ν ) =ZR ν ) dt = f (β(F (µ) − F (ν)))SµS∗ f (t)θt(SµS∗ ν = 0 FACTOR TYPES 9 (cid:3) because β(F (µ) − F (ν)) ∈ RG,F . Since the elements of the form SµS∗ ν for some µ, ν ∈ Pf (G) with s(µ) = s(ν) = v span a strongly dense subspace of PvMPv we conclude that θf (PvMPv) = {0}. Since f (r) 6= 0 we conclude that r /∈ Sp(PvMPv), and hence r /∈ Γ(M). Lemma 3.2. Assume C ∗(G) is simple and that there is a β-KMS weight for αF . Let µ, ν, δ be finite paths in G such that δ > max{µ,ν} and F (µ) = F (ν). Then when µ 6= ν when µ = ν . ν Pδ =(0, Proof. If µ 6= ν and S∗ ν Sδ 6= 0, the relations defining C ∗(G) imply that a piece of µ or a piece of ν will be a loop κ of positive length in G such that F (κ) = 0. By Lemma 10.6 in [Th5] the existence of a β-KMS weight rules out the existence of such a loop. It follows that S∗ ν Sδ can only be non-zero when µ = ν. But S∗ δ Sµ 6= 0 implies that µ must be the initial piece of δ and similarly S∗ ν Sδ 6= 0 implies that ν must also be the initial piece of δ. Therefore S∗ ν Sδ 6= 0 implies that µ = ν, in which case PδSµS∗ δ = PδPµ. ν Pδ = SδS∗ PδSµS∗ δ SµS∗ δ SµS∗ δ SµS∗ δ SµS∗ µSδS∗ PδPµ, (cid:3) Theorem 3.3. Assume that C ∗(G) is simple and NWG not empty. Let ψ be a conservative β-KMS weight for the generalized gauge action αF . Then Γ(πψ(C ∗(G))′′) = RG,F . Proof. The proof is a further development of the proofs of Proposition 4.11 in [Th1] and Theorem 4.1 in [Th4]. As in the proof of Lemma 3.1 we suppress πψ in the notation and consider C ∗(G) as a subalgebra of M. The modular automorphism group θ on M defined by ψ is given by θt = αF −βt , where αF is the σ-weakly continuous extension of αF . Note that β 6= 0 by Propo- sition 2.3 in [Th5]. Let N ⊆ M be the fixed point algebra for θ and consider a vertex v ∈ NWG. By Lemma 3.1 it suffices to show that RG,F ⊆ Γ(M), and from Lemme 2.3.3 and Proposition 2.2.2 in [Co1] we see that it suffices for this to consider a non-zero central projection q ∈ PvNPv for some vertex v ∈ NWG, and show that βF (l) ⊆ Sp(qMq) when l is a loop in G such that s(l) = r(l) = v. Let ω be the state on PvMPv given by ω(a) = ψ(Pv)−1 ψ(a). Then ω is a faithful normal state which is a trace on PvNPv and we consider the corresponding 2-norm kakv =pω(a∗a) . By Kaplansky's density theorem there is an element f ∈ PvAPv such that 0 ≤ f ≤ 1 and kq − fkv is as small as we want. Then f is a linear combination of elements of the form SµS∗ ν where µ, ν ∈ Pf (G) and s(µ) = s(ν) = v. Note that lim R→∞ ν when F (µ) = F (ν) when F (µ) 6= F (ν) , θt(SµS∗ 0 0 with convergence in norm, and that 1 RZ R RZ R 1 0 ν ) dt =(SµS∗ =(cid:13)(cid:13)(cid:13)(cid:13) RZ R 1 0 q − (cid:13)(cid:13)(cid:13)(cid:13) θt(f ) dt(cid:13)(cid:13)(cid:13)(cid:13)v θt(q − f ) dt(cid:13)(cid:13)(cid:13)(cid:13)v ≤ kq − fkv 10 KLAUS THOMSEN by Kadisons Schwarz-inequality. By exchanging limR→∞ assume that N 0 θt(f ) dt for f we may 1 RR R f = λiSµiS∗ νi , Xi=1 where 0 ≤ λi ≤ 1 and F (µi) = F (νi) for all i. Let mψ be the eβF -conformal measure on P (G) defined by ψ and let m be the restriction of the measure mψ(Z(v))−1mψ to Z(v). Then m is a Borel probability measure on Z(v) ⊆ P (G) such that ω (Pµ) = m(Z(µ)) for all µ ∈ Pf (G) with s(µ) = v. For each k ∈ N we let Mk be the set of paths δ in G such that s(δ) = v and δ = k. Then Xδ∈Mk which implies that 1 = Pδ∈Mk Pδ, where the sum converges with respect to the 2-norm and hence also in the strong operator topology. It follows that we can define Qk : PvMPv → PvMPv such that ω(Pδ) = Xδ∈Mk m(Z(δ)) = 1 , Qk(m) = Xδ∈Mk PδmPδ . Then Qk is a positive linear map of norm one and Qk(q) = q. When k > max{µ,ν} it follows from Lemma 3.2 that Qk (SµS∗ ν ) =(Pµ, when µ = ν when µ 6= ν . 0, Thus, for some k large enough, we have that Qk(f ) is a linear combination N ′ Qk(f ) = λiPµi , Xi=1 where 0 ≤ λi ≤ 1 for all i. Using Kadisons Schwarz inequality again we find that kq − Qk(f )kv = kQk(q − f )kv ≤pω (Qk ((q − f )2)) = kq − fkv since ω ◦ Qk = ω for all k. By exchanging f with Qk(f ) for some k large enough we may assume that f = λiPµi . (3.2) N ′ Xi=1 Now observe that since ψ is conservative by assumption it follows that m is concen- trated on {p ∈ Z(v) : s(pi) = v for infinitely many i} . Fix one of the paths µi appearing in (3.2). Let Hj denote the set of finite paths δ of length j such that s(δ) = r(µi), r(δ) = v. Then, up to an m-null set, ∞ {Z(µiδ) : δ ∈ Hj} = Z(µi) . [j=1 We can therefore find a finite set Ki ⊆S∞ m (Z(µi)\ ⊔δ∈Ki Z(µiδ)) j=1 Hj such that FACTOR TYPES 11 is as small as we want. Note that (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) Pµi − Xδ∈Ki Pµiδ(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)v =pm (Z(µi)\ ⊔δ∈Ki Z(µiδ)) is then also small. Therefore, by exchanging Pδ∈Ki Pµiδ for Pµi we may assume that s(µi) = r(µi) = v for all v. Finally, since q is a projection, a standard argument, as in the proof of Lemma 12.2.3 in [KR], allows us to select a subset of the µi's and arrange, after a renumbering, that is a projection in PvAPv such that p = Pµi M Xi=1 kq − pkv ≤ ǫ , where ǫ > 0 can be chosen as small as we need. We choose ǫ > 0 so small that (3.3) (3.4) ω(q) − eF (l)βǫ − 3ǫ > 0 . For each µi from (3.3) we let wi ∈ PvAPv be the elements wi = SµiS∗ µil . Each wi is a partial isometry such that ∗ = Pµi , a) wiwi b) wi c) αF ∗wi = Pµil ≤ Pµi, and t (wi) = e−iF (l)twi for all t ∈ R. i=1 wi and note that w is a partial isometry. It follows from b) that wp = w and therefore from (3.4), c), b) and a) that Set w = PM ω(qwqw∗q) ≥ ω(wqw∗) − 2ǫ = eβF (l)ω(qw∗w) − 2ǫ ≥ eβF (l)ω(pw∗w) − eβF (l)ǫ − 2ǫ = eβF (l)ω(w∗w) − eβF (l)ǫ − 2ǫ = ω(ww∗) − eβF (l)ǫ − 2ǫ = ω(p) − eβF (l)ǫ − 2ǫ ≥ ω(q) − eβF (l)ǫ − 3ǫ . The choice of ǫ ensures that ω(qwqw∗q) > 0 and hence that qwq 6= 0. Since θt(qwq) = e−itβF (l)qwq for all t ∈ R, it follows from Lemme 2.3.6 in [Co1] that βF (l) ∈ Sp(qMq), as desired. (cid:3) In combination with Corollary 2.4 we get the following n=0 A(β)n Corollary 3.4. Assume that C ∗(G) is simple and that NWG 6= ∅. Let ϕ be a β-KMS state for the restriction of αF to PvC ∗(G)Pv for some v ∈ GV . When P∞ w,w = ∞ for some (and hence all) w ∈ NWG, the Connes invariant of πϕ(PvC ∗(G)Pv)′′ is Γ (πϕ(PvC ∗(G)Pv)′′) = RG,F . In an Appendix in [Th5] it is shown that in the setting of Theorem 3.3 the sub- group RG,F is never {0}. Therefore, thanks to Connes' classification in [Co2] and Haagerup's result in [Ha], we get the following Corollary 3.5. In the setting of Theorem 3.3 and Corollary 3.4 the factors πψ(C ∗(G))′′ and πϕ(PvC ∗(G)Pv)′′ are isomorphic; they are both isomorphic to the hyperfinite fac- tor of type IIIλ for some 0 < λ ≤ 1. 12 KLAUS THOMSEN Example 3.6. (Generalized gauge actions on O∞.) The Cuntz algebra O∞, [Cu], is the graph C ∗-algebra C ∗(G) when G is the countable digraph with one vertex and infinitely many arrows, an, n = 1, 2, 3,· · · . Since O∞ is unital all proper weights are bounded and can be normalized to states. Let {tn}∞ n=1 be a sequence of real numbers and define F : GAr → R such that F (an) = tn for all n. The gauge action, where tn = 1 for all n, was considered by Olesen and Pedersen who showed in [OP] that there are no KMS states for the gauge action. In general, it follows from [Th3] that a β-KMS state exists iff P∞ n=1 e−βtn ≤ 1 and that it is unique. There is a Xn=1 conservative β0-KMS state for αF iff ∞ e−β0tn = 1 . By Theorem 3.3 the Connes invariant of the corresponding factor is the closed sub- group of R generated by the numbers {β0ti}∞ i=1. It follows that the factor is always of type III and never of type III0 while all hyperfinite factors of type IIIλ for 0 < λ ≤ 1 can occur by varying the sequence {tn}. The KMS states for αF that are not conservative occur for values of β for whichP∞ n=1 e−βtn < 1 and they correspond to minimal proper almost A(β)-harmonic vectors, albeit of a somewhat degenerate kind since the vectors only have one coordinate. [BF] [BHRS] [BC] [BR] [CPPR] [C] [Co1] [Co2] [CE] [CT] [Cu] [DT] [EFW] [HWH] References S. D. Barreto and F. Fidaleo, On the structure of KMS states of disordered systems, Comm. Math. Phys. 250 (2004), 1-21. T. Bates, J. H. Hong, I. Raeburn, W. Szymanski, The ideal structure of the C ∗- algebras of infinite graphs, Illinois J. Math. 46 (2002), 1159 -- 1176. J.-B. Bost and A. Connes, Hecke algebras, type III factors and phase transitions with spontaneous symmetry breaking in number theory, Selecta Math. (N.S.) 1 (1995), 411-457. O. Bratteli and D.W. Robinson, Operator Algebras and Quantum Statistical Mechan- ics I + II, Texts and Monographs in Physics, Springer Verlag, New York, Heidelberg, Berlin, 1979 and 1981. A. Carey, J. Phillips, I. Putnam, A. Rennie, Families of type III KMS states on a class of C ∗-algebras containing On and QN , J. Funct. Anal. 260 (2011), 1637-1681. F. Combes, Poids associ´e `a une alg`ebre hilbertienne `a gauche, Compos. Math. 23 (1971), 49-77. A. Connes, Une classification des facteurs de type III, Ann. Sci. Ecole Norm. Sup. 6 (1973), 133-252. A. Connes, Classification of injective factors. Cases II1, II∞, IIIλ, λ 6= 1, Ann. of Math. (2) 104 (1976), 73-115. M. D. Choi and E. G. Effros, Separable nuclear C ∗-algebras and injectivity, Duke Math. J. 43 (1976), 309-322. J. Christensen and K. Thomsen, Diagonality of actions and KMS weights, J. Oper. Th. 76 (2016), 449-471. J. Cuntz, Simple C ∗-algebras generated by isometries, Comm. Math. Phys. 57 (1977), 173-185. D. Drinen and M. Tomforde, The C ∗-algebras of arbitrary graphs, Rocky Mountain J. Math. 35 (2005), 105-135. M. Enomoto, M. Fujii and Y. Watatani, KMS states for gauge action on OA, Math. Japon. 29 (1984), 607-619. R. Haag, M. Winnink and N.M. Hugenholtz, On the equilibrium states in quantum statistical mechanics, Comm. Math. Phys., 5 (1967), 215-236. FACTOR TYPES 13 [Ha] [I] [IKW] [KW] [KR] [KP] [Ku] [KV1] [KV2] [LN] [LLNSW] [N] [O] [OP] [Ru] [Sz] [Th1] [Th2] [Th3] [Th4] [Th5] [Y1] [Y2] U. Haagerup, Connes' bicentralizer problem and uniqueness of the injective factor of type III1, Acta Math. 158 (1987), 95 -- 148. M. Izumi, The flow of weights and the Cuntz-Pimsner algebras, Comm. Math. Phys., to appear. M. Izumi, T. Kajiwara and Y. Watatani, KMS states and branched points, Ergod. Th. Dyn. Syst. 27 (2007), 1887-1918. T. Kajiwara and Y. Watatani, KMS states on finite-graph C ∗-algebras, Kyushu J. Math. 67 (2013), 83-104. R.V. Kadison and J.R. Ringrose, Fundamentals of the Theory of Operator Algebras II, Academic Press, London 1986. A. Kumjian and D. Pask, C ∗-algebras of directed graphs and group actions, Ergod. Th. Dyn. Syst. 19 (1999), 1503-1519. J. Kustermans, KMS-weights on C ∗-algebras, arXiv: 9704008v1. J. Kustermans and S. Vaes, Weight theory for C ∗-algebraic quantum groups, arXiv:990163. J. Kustermans and S. Vaes, Locally compact quantum groups, Ann. Scient. ´Ec. Norm. Sup. 33 (2000), 837-934. M. Laca and S. Neshveyev, Type III1 equilibrium states of the Toeplitz algebra of the affine semigroup over the natural numbere, J. Func. Analysis 261 (2011), 169-187. M. Laca, N. Larsen, S. Neshveyev, A. Sims, S.B.G. Webster, Von Neumann algebras of strongly connected higher-rank graphs, Math. Ann. 363 (2015), 657-678. S. Neshveyev, von Neumann algebras arising from Bost-Connes type systems, Int. Math. Res. Not. IMRN 2011 (2011), 217-236. R. Okayasu, Type III factors arising from Cuntz-Krieger algebras, Proc. Amer. Math. Soc. 131 (2003), 2145-2153. D. Olesen and G.K. Pedersen, Some C ∗-dynamical systems with a single KMS-state, Math. Scand. 42 (1978), 111-118. S. Ruette, On the Vere-Jones classification and existence of maximal measure for countable topological Markov chains, Pac. J. Math. 209 (2003), 365-380. W. Szymanski, Simplicity of Cuntz-Krieger algebras of infinite matrices, Pac. J. Math. 122 (2001), 249-256. K. Thomsen, KMS weights on groupoid and graph C ∗-algebras, J. Func. Analysis 266 (2014), 2959 -- 2988. K. Thomsen, Exact circle maps and KMS states, Israel J. Math. 205 (2015), 397 -- 420. K. Thomsen, KMS weights on graph C ∗-algebras, Adv. Math. 309 (2017), 334-391. K. Thomsen, Phase transition in O2, Comm. Math. Phys. 349 (2017), 481-492. K. Thomsen, KMS weights, conformal measures and ends in digraphs, arXives 1612.04716 v2. D. Yang, Type III von Neumann algebras associated with 2-graphs, Bull. Lond. Math. Soc. 44 (2012), 675-686. D. Yang, Factoriality and type classification of k-graph von Neumann algebras, Proc. Edinb. Math. Soc. (2) 60 (2017), 499-518. E-mail address: [email protected] Department of Mathematics, Aarhus University, Ny Munkegade, 8000 Aarhus C, Denmark
1310.4639	1	1310	2013-10-17T09:50:08	Annihilators and Type Decomposition in C*-Algebras	[ "math.OA" ]	We initiate the study of annihilators in C-algebras, showing that they are, in many ways, the best C-algebra analogs of projections in von Neumann algebras. Using them, we obtain a type decomposition for arbitrary C*-algebras that is symmetric and completely consistent with the classical von Neumann algebra type decomposition. We also show that annihilators admit a very simple notion of equivalence that is again completely consistent with the notion of Murray-von Neumann equivalence in von Neumann algebras, sharing many of its general order theoretic properties.	math.OA	math	ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS TRISTAN BICE Abstract. We initiate the study of annihilators in C-algebras, showing that they are, in many ways, the best C-algebra analogs of projections in von Neumann algebras. Using them, we obtain a type decomposition for arbitrary C-algebras that is symmetric and completely consistent with the classical von Neumann algebra type decomposition. We also show that annihilators admit a very simple notion of equivalence that is again completely consistent with the notion of Murray-von Neumann equivalence in von Neumann algebras, sharing many of its general order theoretic properties. 1. Introduction 1.1. Motivation. It is no exaggeration to say that type decomposition and Murray- von Neumann equivalence of projections are absolutely fundamental to the theory of von Neumann algebras. These concepts have been extended to other operators in a couple of different ways (see below), and this has certainly led to some interesting theory. However, only very specific aspects of the von Neumann algebra theory have been generalized to C-algebras in this way. No doubt it was accepted that this is an unavoidable fact of life, that much of the von Neumann algebra theory simply can not be applied in any general way to the much larger class of C-algebras. In the present paper we show this to be false. By choosing generalizations appropri- ately, using annihilators rather than projections, a surprising amount of the basic von Neumann algebra theory does indeed extend fully to C-algebras. Let us first go back to the beginning and consider a von Neumann algebra A. Here, type decomposition is obtained by utilizing the projections P(A), and crucial to this is their order structure, specifically the fact P(A) is a complete orthomodular lattice. Also crucial is the fact that projections exist in abundance in an arbitrary von Neumann algebra. In C-algebras, on the other hand, there may be no non-zero projections whatsoever, even in the simple case. And even when they are plentiful, they may fail to form a lattice. Consequently, to prove results for C-algebras that generalize or are analogous to classical von Neumann algebra results, like those relating to type decomposition, involves finding an appropriate replacement for projections on which an appropriate analog of Murray-von Neumann equivalence can be defined. One example of this can be found in [CP79], where projections in a C-algebra A are replaced with positive elements and a, b ∈ A+ are said to be equivalent if 2010 Mathematics Subject Classification. Primary: 46L05. Key words and phrases. C-Algebras, Type Decomposition, Annihilators, Non-Commutative Topology, Ortholattices. This research has been supported by a CAPES (Brazil) postdoctoral fellowship through the program "Science without borders", PVE project 085/2012. 1 2 TRISTAN BICE there exists (xn) ⊆ A such that a =P xnx∗n and b =P x∗nxn, where the sums are norm convergent. The close relationship between traces and Murray-von Neumann equivalence classes in von Neumann algebras generalizes to positive operators with this notion of equivalence, as demonstrated in [CP79]. An analogous classification and even a decomposition of an arbitrary C-algebra into types I, II and III is also obtained in [CP79] Proposition 4.13. However, this decomposition is neither symmetric (the type III part can only be found in the quotient w.r.t. the type I part, not the other way around, and likewise the type II part is only obtained at the end as a quotient of quotients) nor consistent with the classical von Neumann algebra type classification (for example, B(l2) is a type I von Neumann algebra but not a type I C-algebra1 (see [Ped79] 6.1.2)). Furthermore, A+ will not be a lattice unless A is commutative and other natural order theoretic properties fail for A+ in general, for example the sum of two finite elements of A+ may be infinite (see [CP79] Corollary 7.10). Slight variants of the above can be obtained by changing the sums in the defini- tion of equivalence, e.g. by specifying that they must be finite, or allowing them to represent supremums that might not necessarily converge in norm. Another quite different example is given in [Cun77], where projections are replaced with arbitrary operators and a, b ∈ A are said to be equivalent if a / b and b / a, where a / b means a = cbd for some c, d ∈ A. Again, this leads to a natural notion of a finite and factorial (simple) C-algebra. The natural norm closed variant of this (i.e. cnbdn → a, for some (cn), (dn) ⊆ A) has also received considerable attention and has a strong relation to the dimension functions on A (see [BH82]). Although again the order structures so obtained are generally less tractable and there is a limit to how far the analogy to projections in von Neumann algebras can be pushed (there does not appear to even be a canonical decomposition into types in this case, for example). However, there is another quite different, but very natural, candidate to replace projections with in an arbitrary C-algebra A, one that seems to have been largely overlooked. Namely, we can use the left (or right) annihilator ideals, i.e. those of the form {a ∈ A : ∀b ∈ B(ab = 0)} for some B ⊆ A. Equivalently, we can use the hereditary C-subalgebras corresponding to these left ideals (see [Eff63] Theorem 2.4 or [Ped79] Theorem 1.5.2), which we refer to simply as annihilators. Indeed, the map p 7→ pAp is an order isomorphism from projections (with their canonical order p ≤ q ⇔ pq = p) to a subset of annihilators (ordered by inclusion). This map is even surjective whenever A is a von Neumann algebra (see Proposition 3.40 (1)) or, more generally, an AW-algebra (see [Ber72]), thus yielding a precise correspondence between projections and annihilators in this case.2 Unlike projections, though, the annihilators still exist in abundance in an arbi- trary C-algebra, thanks to the continuous functional calculus (see the discussion preceeding Lemma 3.10). And we can see from the outset that they have greater po- tential to fulfil the role of projections in a von Neumann algebra, as they also always 1Here, and here only, we are using the terminology of [CP79], where a C-algebra is said to be of type I if all its representations are type I. Other standard terms for such C-algebras are 'GCR' and 'postliminal'. Throughout the rest of this article we will use the term 'postliminal'. 2This correspondence is well known, as is the fundmental importance of studying projections in AW-algebras. Despite this, however, there does not seem to be any indication, either in [Ber72] or elsewhere in the operator algebra literature (except in [Arz13]), that it was ever thought the annihilators themselves might be of interest in more general contexts. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 3 form a complete lattice. Furthermore, there is a natural orthocomplementation on annihilators, something we do not have for arbitrary hereditary C-subalgebras (the collection of all hereditary C-subalgebras may not even be complemented). It turns out that this orthocomplementation is not always orthomodular (see Example 3.87), but we can prove a close approximation to orthomodularity (see §3.5) which allows much of the von Neumann algebra theory to be generalized fully to arbitrary C- algebras. In particular, we can obtain a type decomposition that is symmetric and completely consistent with the classical type decomposition of von Neumann al- gebras, together with a natural analog of Murray-von Neumann equivalence that is again completely consistent with the classical notion. The type decomposition itself can actually be obtained in a very general order theoretic context, and in the specific case of annihilators in a C-algebra, the definitions in (2.36), (2.37), (2.38), and (2.39) yield the following result. Theorem 1.1. For any C-algebra A we have orthogonal annihilator ideals AI, AII, AIII and AIV such that is an essential ideal in A. AI ⊕ AII ⊕ AIII ⊕ AIV When A is a von Neumann algebra, AI, AII and AIII are indeed the usual type I, II and III parts in the classical von Neumann algebra decomposition (see the comments after (2.39)). As every von Neumann algebra is an AW-algebra, i.e. every annihilator ideal is of the form pA for some central p ∈ P(A), a finite sum of annihilator ideals is again an annihilator ideal, coming from the sum of the corresponding projections. As the only essential annihilator ideal is the entire algebra itself, we have A = AI⊕AII⊕AIII. As P(A) is orthomodular, the extra type IV part is {0} here, although we do not know if the same is true for annihilators in C-algebras. Indeed, this paper puts us in much the same position as von Neumann himself was in at the early stages of his investigation into von Neumann algebras (see [vN30] and [MvN36]). Namely, we can decompose an arbitrary C-algebra into various types but do not know if all these potential types are actually realized by some C-algebra, most notably we do not know if there are any type IV C-algebras (just as von Neumann did not verify the existence of type III von Neumann algebras until later in [vN40]). It should now be clear that the annihilators in a C-algebra are of fundamental importance. The C-algebra theory required for the initial investigation presented here is not even particularly great, and this paper should be accessible to anyone familiar with the material in the relevant parts of the first few chapters of [Ped79]. Given this, it is very surprising that more papers analyzing the annihilator structure of C-algebras have not been written before and, in our opinion, such an analysis is well overdue. We know of only one such article, namely [Arz13], where a few ideas similar to those presented here are also discussed. However, there are simple counterexamples to [Arz13] Lemma 16.1 (see the discussion at the end of §3.2) which, unfortunately, is used repeatedly in [Arz13] and thus puts the results there into question. The key point is that care must be taken to distinguish arbitrary open projections from those that are also topologically regular, which amounts to distinguishing aribitrary hereditary C-subalgebras from annihilators. And there is no mention of an analog of Murray-von Neumann equivalence for annihilators in [Arz13], although analogs of Murray-von Neumann equivalence for arbitrary 4 TRISTAN BICE hereditary C-subalgebras have been considered before (see [OrT11] and [PZ00]). The difference between annihilators and arbitrary hereditary C-subalgebras may at first seem slight, but it turns out that it is the annihilators that are more amenable to attack by a fortuitous combination of non-commutative topology, order theory and algebra, as we proceed to demonstrate in this paper. It is really the order theory that is central here. Kaplansky initiated a program to isolate the algebraic structure of von Neumann algebras, and this is what al- lowed the von Neumann algebra theory to be generalized to AW-algebras (and, more generally, Baer -rings). All we are really doing is taking this a step further, isolating the order structure of projections in AW-algebras in such a way that the theory can be generalized to annihilators in C-algebras. Von Neumann himself isolated the order structure of projections in finite von Neumann algebras, those in which the projection lattice is modular, resulting in the elegant theory of continu- ous geometries (see [vN60]). Even when the projection lattice of a von Neumann algebra is not modular, it is still orthomodular, and this inspired the development of a large body of work on orthomodular lattices (see [Kal83]). Type decompositions have also been obtained for certain orthomodular lattices, namely the dimension lattices of [Loo55], and these have been successively generalized to various other contexts, like the espaliers in [GW05] and the effect algebras in [FP13]. However, these other contexts still assume something equivalent to orthomodularity in the ortholattice case, like unique orthogonal complements.3 We diverge from this previous work with the simple observation that separativ- ity, a significantly weaker assumption than orthomodularity, is sufficient for much of the development of the theory. And this is most fortunate, for we can only verify that the annihilators in an arbitrary C-algebra are separative (although in strong sense quite close to orthomodularity -- see Theorem 3.26 and Example 3.87). Also, we work with what we call type relations, which are again more general than the dimension relations in dimension lattices (e.g. they need not satisfy finite (orthog- onal) divisibility). Again, we see that these are sufficient for much of the theory to be developed which, yet again, is fortunate because we can only verify these weaker properties for what we believe to be the natural equivalence relation on annihilators generalizing Murray-von Neumann equivalence. We also make a number of other order theoretic observations and generalizations of our own that do not seem to have appeared elsewhere in the literature, even in the orthomodular case. Really, you could see this paper as bringing the theory of lattices and operator algebras back together after over half a century of divergence from von Neumann's seminal work in both fields, namely in continuous geometry and von Neumann algebras. 1.2. Outline. We start off in §2 by developing the theory of type decomposition in an abstract order theoretic context general enough to be applied later to anni- hilators in a C-algebra. We take [Kal83] as our primary reference, although we have to do things in greater generality as we are concerned with ortholattices that may not be orthomodular. In particular, we have to be careful to distinguish [p] 3This is no longer true for the pre-effect algebras in [CK12] and presumably type decomposition could also be done for the subclass of pre-effect algebras corresponding to separative complete ortholattices, and probably some more general subclass (centrality is discussed in [CK12], although they do not quite go as far as doing type decomposition). But we decided to stick to ortholattices rather than pre-effect algebras, as these are certainly sufficient for analyzing the annihliators in a C-algebra, and probably also more familiar to operator algebraists. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 5 from [p]p (see the discussion following Definition 2.5). As such, this section should be of independent interest to lattice theorists, although many of the new results are relatively straightforward generalizations of known results for orthomodular lattices. As mentioned above, the key observation here is that separativity, rather than orthomodularity, is sufficient to prove Theorem 2.22. In §3, we start by gathering together some relevant facts about projections and annihilators, and the relationship between them. In §3.1 we set the stage for our in- vestigation of the annihilators, defining them as the orthocompletion of a C-algebra A with respect to a certain preorthogonality relation. Next we introduce some stan- dard non-commutative topological terminology in §3.2 and show in Theorem 3.4 that the annihilators correspond precisely to the topologically regular open projec- tions. Then we make some important observations about spectral projections and projections in general in §§3.3 and 3.4. It is in §3.5 that we develop the theory needed to prove that the annihilators satisfy the all important property known as separativity. In fact, Theorem 3.26 says that the annihilators satisfy a strong form of separativity which is very close to orthomodularity. On the way, we also strengthen a result from [AE02], showing that non-regular open dense projections must, in fact, be as non-regular as possible (see Theorem 3.25 and the discussion at the start of §3.5). With separativity out of the way, we are free to apply the theory in §2 and investigate the interplay between the algebraic structure of a C-algebra A and the order structure of its annihilators [A]⊥. In §3.6, we see that the central annihilators are precisely the annihilator ideals and that the annihilators always have the relative centre property. Next, in §3.7, we define and investigate what we believe to be the natural analog of the fundamental notion of Murray-von Neumann equivalence. We then move on to discuss the abelian annihilators in §3.8, starting with Theorem 3.49 which says that A is commutative precisely when [A]⊥ is a Boolean algebra. We then extend the results of §3.8 to homogeneous annihilators in §3.9, and show that the annihilator notion of homogeneity is closely related to the more classical representation theoretic notion. In §3.10 we investigate C-algebras of continuous functions from topological spaces X to finite rank matrices Mn. More specifically we show how to represent hereditary C-subalgebras/open projections of such C-algebras by lower semicon- tinuous projection valued functions on X. The moral of the story here is that the theory of annihilators turns out to be the theory of these projection functions modulo nowhere dense subsets of X. Lastly, in §3.11, we give a number of exam- ples illustrating the subtle distinction between various order theoretic and algebraic notions. 1.3. Acknowledgements. The author would like to thank Charles Akemann and Dave Penneys for many helpful comments on earlier versions of this paper, as well as Vladimir Pestov, for giving the author the opportunity to pursue the research that lead to this paper. 2.1. Basic Definitions. 2. Order Theory Definition 2.1 (Relation Terminology). A relation R on a set S is • reflexive if sRs, for all s ∈ S. 6 TRISTAN BICE • transitive if sRt and tRu ⇒ sRu, for all s, t, u ∈ S. • symmetric if sRt ⇔ tRs, for all s, t ∈ S. • antisymmetric if sRt and tRs ⇒ s = t, for all s, t ∈ S. • annihilating if sRs ⇒ ∀t ∈ S(sRt), for all s ∈ S. • a preorder if R is reflexive and transitive. • a partial order if R is an antisymmetric preorder. • an equivalence relation if R is a symmetric preorder. • a preorthogonality relation if R is symmetric and annihilating. • an orthogonality relation if R is a preorthogonality relation and, for s, t ∈ S, ∀u ∈ S(s ⊥ u ⇔ t ⊥ u) ⇔ s = t. (2.1) complete if every S ⊆ P has a supremum and infimum, denoted by W S and V S Definition 2.2 (Partial Order Terminology). Let P be a partial order. We call P a lattice if every pair p, q ∈ P has a supremum (least upper bound) and infimum (greatest lower bound), denoted by p ∨ q and p ∧ q respectively. A lattice P is respectively. If P has a greatest element 1 and least element 0 then p and q are complementary if p ∨ q = 1 and p ∧ q = 0. If every element of P has a complement then P is complemented. For a preorder P, we write p < q to mean p ≤ q but q (cid:2) p, and we call S ⊆ P order-dense in P if ∀p ∈ P(p > 0 ⇒ ∃s ∈ S(0 < s ≤ p)), (2.2) and we call S join-dense in P if, for all p ∈ P (with p > 0), (2.3) p =_{q ∈ S : q ≤ p}. Definition 2.3 (Function Terminology). A function f on P is • involutive if f (f (p)) = p, for all p ∈ P. • a complementation if p and f (p) are complementary, for all p ∈ P. • order preserving if p ≤ q ⇒ f (p) ≤ f (q), for all p, q ∈ P. • antitone if p ≤ q ⇒ f (q) ≤ f (p), for all p, q ∈ P. • an orthocomplementation if f is an antitone involutive complementation. • an order isomorphism if f is 1-1, onto, and f and f−1 are order preserving. • an orthoisomorphism if, further, f (p⊥) = f (p)⊥, for all p ∈ P. A partial order P with a distinguished orthocomplementation is an orthoposet and, if P is also a lattice, an ortholattice. 2.2. Orthocompletions. We will be interested in a particular case of the following situation (see §3.1). We are given a relation ⊥ on a set S and, for T ⊆ S, define T ⊥ = {s ∈ S : ∀t ∈ T (t ⊥ s)} and [S]⊥ = {T ⊥ : T ⊆ S}. Also, for future reference, we make the following definition. Definition 2.4. We call T ⊆ S essential (w.r.t. ⊥) if T ⊥⊥ = S. For a collection of subsets T of S, we have (ST )⊥ = T{T ⊥ : T ∈ T }. In particular, this means infimums, w.r.t. the inclusion order, always exist in [S]⊥ are are simply given by intersections. Furthermore, S = ∅⊥ is the largest element of [S]⊥, while S⊥ = {s ∈ S : ∀t ∈ S(t ⊥ s)} is the smallest, and T 7→ T ⊥ is (1) antitone, (2) involutive on [S]⊥, if ⊥ is symmetric, and (3) an orthocomplementation on [S]⊥, if ⊥ is a preorthogonality relation. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 7 (1) is immediate, and for (2) note that symmetry implies T ⊆ T ⊥⊥, for all T ⊆ S, and thus, by (1), (T ⊥⊥)⊥ ⊆ T ⊥ ⊆ (T ⊥)⊥⊥, i.e. T ⊥ = T ⊥⊥⊥. Lastly, if ⊥ is also annihilating then T ∩T ⊥ = S⊥, for all T ∈ [S]⊥, and hence T ⊥∨T = T ⊥∨T ⊥⊥ = S so T and T ⊥ are complementary, i.e. [S]⊥ is a complete ortholattice. We call [S]⊥ the orthocompletion of S w.r.t. ⊥. In fact, we have really just proved a slightly more general version of [Mac64] Lemma 2.1, and what we have denoted by [S]⊥ is denoted in [Mac64] by L(S), where it is called the completion of S. As shown in [Mac64] Theorem 2.4, it really is the canonical completion by cuts when S itself is an orthoposet. We define the preorder ⊣ induced by ⊥ on S by s ⊣ t ⇔ {t}⊥ ⊆ {s}⊥. Note that ⊣ is a partial order if and only if ⊥ satisfies (2.1) and that (2.4) is an order preserving map from S (ordered by ⊣) to [S]⊥ (ordered by ⊆). If ⊥ is symmetric then ⊥ is involutive and hence we actually have s 7→ {s}⊥⊥ s ⊣ t ⇔ {s}⊥⊥ ⊆ {t}⊥⊥. If T ⊆ S is join-dense in S (see (2.3)) w.r.t. ⊣ then, by (a slight generalization of) [Mac64] Theorem 2.5, the map (2.5) U 7→ U ∩ T is an orthoisomorphism witnessing [S]⊥ ∼= [T ]⊥. Going in the other direction, given an orthoposet P, we can always define an orthogonality relation ⊥ by p ⊥ q ⇔ p ≤ q⊥. If P = [S]⊥, where ⊥ is a preorthogonality relation on S then, for T, U ∈ [S]⊥, the relation ⊥ on [S]⊥ defined in this way is related to the original relation ⊥ on S by Furthermore, for T ∈ [S]⊥, we can consider the restriction of ⊥ to T and then T ⊥ U ⇔ ∀t ∈ T∀u ∈ U (t ⊥ u). [T ]⊥ = [T ]T , according to Definition 2.5. 2.3. Relative Complements. Definition 2.5. Given an ortholattice P and p, q, r ∈ P we define r⊥p = r⊥ ∧ p and [q]p = {r⊥p : r ≤ p and r⊥p ≤ q}. In particular, [p]p = {r⊥p : r ≤ p}. We also drop the subscript when r = 1, i.e. [p] = [p]1 = {q : q ≤ p} and [q]p = [p]p ∩ [q]. It is important to note that [p] may not be an ortholattice (see the Hasse diagram below), in contrast to [p]p. Proposition 2.6. If P is an ortholattice and p ∈ P then [p]p is an ortholattice. If q ∈ [p]p then [q]q ⊆ [p]p while, for any q ∈ [p], we have (2.6) q⊥p⊥p =^{r ∈ [p]p : q ≤ r}. 8 TRISTAN BICE Proof. First note that for q, r ≤ p we have (q⊥ ∧ p) ∧ (r⊥ ∧ p) = (q ∨ r)⊥ ∧ p ∈ [p]p so infimums exist and agree with those in P. Next note that, when q ≤ p, we have q ≤ (p ∧ q⊥)⊥ ∧ p so q⊥ ≥ ((p ∧ q⊥)⊥ ∧ p)⊥ and p ∧ q⊥ ≥ ((p ∧ q⊥)⊥ ∧ p)⊥ ∧ p = ((p ∧ q⊥) ∨ p⊥) ∧ p ≥ p ∧ q⊥, so ⊥p is involutive and, therefore, actually characterizes the elements of [p]p, i.e. [p]p = {q = q⊥p⊥p : q ∈ P}. And if q ∈ [p]p and r ∈ [q]q then r⊥q ≤ p and hence r⊥q⊥p ∈ [p]p. Therefore r = r⊥q⊥q = r⊥q⊥p ∧ q ∈ [p]p, i.e. [q]q ⊆ [p]p. As ⊥ is order reversing, so is ⊥p so, in particular, supremums also exist. Also, if r ∈ [p]p and q ∈ [r] then r⊥p ≤ q⊥p and hence q⊥p⊥p ≤ r⊥p⊥p = r, thus verifying (2.6). Finally, for any q ∈ [p]p, we have q⊥p ∧ q ≤ q⊥ ∧ q ≤ 0 and hence also q⊥p ∨p q = 0⊥p = p. Thus q⊥p is a complement of q in [p]p, i.e. [p]p is an ortholattice with orthocomplement function ⊥p . (cid:3) On the other hand, [p] is always a sublattice of P, while [p]p may not be. Indeed, while infimums in [p]p agree with those in P, the same can not be said for supre- mums. For example, in the ortholattice represented by the following Hasse diagram (x ≤ y in such a diagram if and only if y appears above x and joined to it by lines), which appears as [Kal83] Figure 6.5, [p]p = {0, a⊥, c⊥, p} and hence a⊥ ∨p c⊥ = p, even though a⊥ ∨ c⊥ = b. Also, [p] = {0, a⊥, c⊥, p, b}, which does not possess any orthocomplement functions. Figure 1. 1 p b b⊥ p⊥ 0 a c⊥ c a⊥ 2.4. Order Types. Definition 2.7. A preorder P is separative if, for all p, q ∈ P, p (cid:2) q ⇒ ∃r ∈ P(0 < r ≤ p and r ∧ q = 0). We call an orthoposet P orthomodular if, for all p, q ∈ P, p ⊥ q ⇒ p ∨ q exists, and (2.7) A lattice P is modular if, for p, q, r ∈ P, q ≤ p ⇒ p = q ∨ (p ∧ q⊥) q ≤ p ⇒ p ∧ (q ∨ r) = q ∨ (p ∧ r). ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 9 A lattice P is distributive if, for all p, q, r ∈ P, and (2.8) p ∧ (q ∨ r) = (p ∧ q) ∨ (p ∧ r) A Boolean algebra is a distributive complemented lattice. p ∨ (q ∧ r) = (p ∨ q) ∧ (p ∨ r). Every element of a Boolean algebra in fact has a unique complement and the map taking each element to this unique complement is an orthocomplement function. In fact, an ortholattice is uniquely complemented if and only if it is a Boolean algebra, by [Kal83] §3 Proposition 7. For an ortholattice, we immediately have distributivity ⇒ modularity ⇒ orthomodularity ⇒ separativity. To see that the first two of these implications can not be reversed, it suffices to note that the subspaces of a Hilbert space H are modular (more generally, submod- ules of a module are modular, hence the name) but not distributive if dim(H) > 1, while the closed subspaces are orthomodular but not modular if dim(H) = ∞, by [Kal83] §5 Proposition 5. This last fact is actually key to showing that the projec- tions in an infinite AW-algebra are not modular (see [Kap55] Theorem on page 1). There are also finite ortholattices that illustrate these differences, for example the Chinese latern MO2 represented by [Kal83] Figure 2.1 11 is modular but not distributive, while the ortholattice in [Kal83] Figure 3.2 is orthomodular but not modular. For an example of an ortholattice that is separative but not orthomod- ular, consider the orthodouble of the 8 element Boolean algebra given in [Fla82] Figure 2b, as represented by the following Hasse diagram. a d b e c f 1 0 f⊥ c⊥ e⊥ b⊥ d⊥ a⊥ For an example of an ortholattice that is not even separative, just consider O6 in [Kal83] Figure 3.1. 2.5. Separativity. Separativity is fundamental to our later work because it is precisely what is required to turn order-density into join-density. Proposition 2.8. A preorder P is separative if and only if, for all S ⊆ P, S is join-dense ⇔ S is order-dense. Proof. Join-density certainly implies order-density, while if S is not join-dense then we can find p, q ∈ P with p (cid:2) q even though q ≥ r, for all r ∈ [p] ∩ S. If P is separative then we can find t ∈ P with 0 < t ≤ p and t ∧ q = 0, and hence there is no s ≤ t with 0 < s ∈ S, i.e. S is not order-dense. On the other hand, if P is not separative then we have p, q ∈ P with p (cid:2) q even though there is no r ≤ p with r > 0 and r ∧ q = 0. Now consider S = [q] ∪ {s ∈ P : s (cid:2) p}. If t (cid:2) p then t ∈ S, while if 0 < t ≤ p then t ∧ q 6= 0, i.e. there exists s > 0 with t ≥ s ∈ [q] ⊆ S. Thus S is order-dense, however, S ∩ [p] ⊆ [q] which, as p (cid:2) q, (cid:3) means that p 6=W S ∩ [p], so S is not join-dense. 10 TRISTAN BICE In a lattice P another equivalent of separativity is obtained if we replace p (cid:2) q in the definition of separativity with the apparently stronger condition q < p. For p (cid:2) q implies p ∧ q < p and hence we could find r ≤ p with 0 = (q ∧ p) ∧ r = q ∧ (p ∧ r) = q ∧ r. 2.6. Perspectivity. Definition 2.9. If P is an ortholattice, we say p, q ∈ P are (1) perspective if p and q have a common complement. (2) orthoperspective if p and q have a common orthogonal complement. (3) semiorthoperspective if p and q⊥ are complementary. These relations will be denoted by ∼p, ∼op and ∼sop respectively. The definition of perspectivity is perfectly valid in an arbitrary lattice with 1 and 0 and is fundamental to the theory of continuous geometries (see [vN60]). Note that p and q⊥ are complementary if and only if p⊥ and q are complementary, so semiorthoperspectivity is a symmetric relation. Semiorthoperspective p and q are sometimes said to be 'in position p′', while if p is also semiorthoperspective to q⊥ then they are 'in generic position' or 'in position p' (see [Ber72] §13 Definition 2 and Definition 3). Perspectivity is weaker than semiorthoperspectivity, however their transitive closures are the same. For if p and q have common complement r then p is semiorthoperspective to r⊥ which is in turn semiorthoperspective to q. Orthoperspectivity on the other hand is much stronger, and is often just equality (see Proposition 2.16). Note that p and q are orthoperspective if and only if (p∨q)⊥ is complementary to both p and q. Proposition 2.10. For a symmetric transitive relation ∼ on an ortholattice P, the following are equivalent. (1) ∼ is weaker than orthoperspectivity. (2) q ≤ p and q⊥ ∧ p = 0 ⇒ p ∼ q. (3) q ∼ q⊥p⊥p , for all p ∈ P and q ∈ [p]. Proof. (1)⇒(2) p ∨ q = p so p ∨ (p ∨ q)⊥ = 1 and q ∨ (p ∨ q)⊥ = (q⊥ ∧ p)⊥ = 1 so p and q (2)⇒(1) If p∨ (p∨ q)⊥ = 1 then p⊥∧ (p∨ q) = 0 so orthoperspectivity and (2) imply (2)⇒(3) q ≤ q⊥p⊥p and q⊥ ∧ q⊥p⊥p = q⊥p ∧ q⊥p⊥p = 0 so q ∼ q⊥p⊥p . (3)⇒(2) q⊥p⊥p = (q⊥ ∧ p)⊥p = 0⊥p = p so p ∼ q. are orthoperspective and hence p ∼ q. p ∼ p ∨ q ∼ q which, by transitivity, gives p ∼ q. (cid:3) (Transitivity only appears in (2)⇒(1) so orthoperspectivity satisfies (2) and (3).) Proposition 2.11. If P is an ortholattice, p, q ∈ P, [p]p = [p] and [q] = [q]q then Proof. Note that s = p′∧q′⊥ ≤ p and hence s ⊥ (q∧p⊥). But then s ⊥ (q∧p⊥)∨q′ = (q∧p⊥)∨(q∧p⊥)⊥q = q, as [q] = [q]q, and hence s ≤ p∧q⊥ which, as s ≤ (p∧q⊥)⊥, means s = 0. Now q′ ∧ p′⊥ = 0 follows by a symmetric argument. (cid:3) Corollary 2.12. If P is an ortholattice, p, q ∈ P, [p]p = [p] and [q] = [q]q then ∃r ∈ P(p ≤ r ∼sop q) ⇒ ∃s ∈ P(p ∼sop s ≤ q). p′ = (p ∧ q⊥)⊥p and q′ = (q ∧ p⊥)⊥q are semiorthoperspective. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 11 Proof. With p′ and q′ as above we have p′ ∼sop q′. But p ∧ q⊥ ≤ r ∧ q⊥ = 0 so p′ = 0⊥p = p. So, setting s = q′, we are done. (cid:3) 2.7. Finiteness. Definition 2.13. Given a symmetric relation ∼ on P, we call p ∈ P ∼-finite if p ∼ q ≤ p ⇒ p = q. If P is an ortholattice, we call p ∈ P ∼-orthofinite if p ∼ q ≤ p ⇒ p ∧ q⊥ = 0. We call ∼ itself (ortho)finite if all elements of P are ∼-(ortho)finite. Note that if ∼ is transitive and weaker than orthoperspectivity then, when p is not orthofinite, there exists q ≤ p with p ∼ q ∼ q⊥p⊥p < p, as q⊥p 6= 0, i.e. p is orthofinite ⇔ {p}∼ ∩ [p]p = {p}. Also, by definition, p is finite ⇔ {p}∼ ∩ [p] = {p}. Definition 2.14. We say ∼ is finitely additive if, whenever p ⊥ q, r ⊥ s, p ∼ r and q ∼ r, we have p ∨ q ∼ r ∨ s. Proposition 2.15. A finitely additive reflexive relation ∼ on an ortholattice P is orthofinite if and only if 1 is (ortho)finite. Proof. If ∼ is orthofinite then, in particular, 1 is (ortho)finite. If ∼ is not orthofi- nite, we have p ∼ q ≤ p with p ∧ q⊥ 6= 0. As ∼ is finitely additive and reflexive, we have 1 = p ∨ p⊥ ∼ q ∨ p⊥ even though (q ∨ p⊥)⊥ = p ∧ q⊥ 6= 0 so q ∨ p⊥ 6= 1. (cid:3) 2.8. Orthomodularity. Orthomodularity has a number of important equivalents (for more see [Kal83] §3 Theorem 2). Proposition 2.16. For an ortholattice P, the following are equivalent. (1) P is orthomodular. (2) Orthogonal complements are unique. (3) [p]p = [p], for all p ∈ P. (4) Orthoperspectivity is finite. (5) Orthoperspectivity is equality. (6) ∼-orthofiniteness and ∼-finiteness always coincide. (7) For all p, q, r ∈ P, q ≤ p and q ⊥ r ⇒ p ∧ (q ∨ r) = q ∨ (p ∧ r). p⊥ = q ∨ (q⊥ ∧ p⊥) = q, i.e. orthogonal complements are unique. p⊥ ∨ q ∨ (p⊥ ∨ q)⊥ = 1 so (2) r = p, showing that P is orthomodular. Proof. (1)⇒(2) If q ≤ p⊥ and p ∨ q = 1 then p⊥ ∧ q⊥ = 1⊥ = 0 so orthomodularity gives (2)⇒(1) Given q ≤ p let r = q ∨ (p ∧ q⊥) ≤ p. Then p⊥ ∨ r = p⊥ ∨ q ∨ (p ∧ q⊥) = (2)⇒(4) If q ≤ p = p⊥⊥ and 0 = q⊥ ∧ p = (q ∨ p⊥)⊥ then (2) gives q = p⊥⊥ = p. (4)⇒(2) If p ∈ P has an orthogonal complement q < p⊥ then p⊥ ∧ q⊥ = 0. (5)⇒(4) Equality is finite. (4)⇒(5) If p and q are orthoperspective then so are p and p ∨ q which, if orthoper- (4)⇒(6) q ≤ p and q⊥ ∧ p = 0 means p and q are orthoperspective which, from (4), (6)⇒(4) Orthoperspectivity is orthofinite. If it is not finite these notions differ. spectivity is finite, means p = p ∨ q. Likewise q = p ∨ q = p. gives p = q. Thus ∼-orthofiniteness implies ∼-finiteness. 12 TRISTAN BICE (5)⇒(3) q and q⊥p⊥p are orthoperspective so (5) gives q = q⊥p⊥p ∈ [p]p for q ∈ [p]. (3)⇒(4) If q < p but q⊥ ∧ p = 0 then q /∈ [p]p. (1)⇒(7) See [Kal83] §3 Theorem 5. (7)⇒(1) Immediate by setting r = q⊥. (cid:3) So if P is not orthomodular then we have [p] 6= [p]p, for some p ∈ P, and there is no reason to think that properties [p] inherits from P, like separativity, are necessarily inherited by [p]p. However, if we happen to know that [p]p is order-dense in [p] (which is true for the annihilators we will be interested in -- see the comment after Proposition 3.11), then [p]p will indeed be separative if P is. Proposition 2.17. If P is an orthomodular lattice and p, q ∈ P then p′ = (p ∧ q⊥)⊥p and q′ = (q ∧ p⊥)⊥q are maximal semiorthoperspective. Proof. Semiorthoperspectivity is immediate from Proposition 2.11 and Proposition 2.16. For maximality, note that if s > p′ then 0 6= s∧p′⊥ = s∧p∧q⊥, by orthomodularity. Thus s ∧ q⊥ 6= 0 so s could not be semiorthoperspective to anything in [q]. (cid:3) 2.9. Modularity. As shown in [Jac85] Theorem 8.4, P is modular if and only if, for p, q, r ∈ P, (2.9) p ∨ r = q ∨ r, p ∧ r = q ∧ r and p ≤ q ⇒ p = q. In particular, this means perspectivity is a finite relation on P. In fact, if P is an ortholattice then (2.10) modularity ⇔ perspectivity is finite. To see the converse, first note that if perspectivity is finite then so is orthoperspec- tivity and hence P is orthomodular, by Proposition 2.16. Now say p, q ∈ P satisfy the conditions on the left of (2.9) then set p′ = p ∧ (p ∧ r)⊥ and p′′ = p′ ∨ (p′ ∨ r)⊥ and likewise for q′ and q′′. Then we see that p′ ∧ r = 0 = q′ ∧ r and hence p′′ ∧ r = 0 = q′′ ∧ r, as well as p′′ ∨ r = 1 = q′′ ∨ r. So if perspectivity is finite then p′′ = q′′. As P is orthomodular, p′ ∨ r = p ∨ r = q ∨ r = q′ ∨ r, and so orthomodularity together with p′′ = q′′ implies p′ = q′ and hence orthomodularity together with p ∧ r = q ∧ r finally yields p = q. 2.10. The Centre. Definition 2.18. In an ortholattice P, we say s, t ∈ P commute if (2.8) holds whenever p, q, r ∈ {s, s⊥, t, t⊥}. We call p ∈ P central if it commutes with all q ∈ P. If P is complete and p ∈ P, we define the central cover c(p) of p by so cP = {p ∈ P : p is central}. We call p, q ∈ P very orthogonal if c(p) ⊥ c(q). The only non-trivial instances of (2.8), for p, q, r ∈ {s, s⊥, t, t⊥}, are of the form (2.11) (2.12) p ∧ q = p ∧ (p⊥ ∨ q) and p = (p ∧ q) ∨ (p ∧ q⊥), Given T ⊆ P we define c(p) =^{q ≥ p : q is central}. cT = {c(t) : t ∈ T} ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 13 for p ∈ {s, s⊥} and q ∈ {t, t⊥}, or p ∈ {t, t⊥} and q ∈ {s, s⊥}. Thus, the order theoretic definition of commutativity agrees with the algebraic definition for pro- jections in a C-algebra, by Proposition 3.13. In fact, for orthomodular lattices, each non-trivial instance of (2.8) is equivalent to every other and to 1 = (s ∧ t) ∨ (s⊥ ∧ t) ∨ (s ∧ t⊥) ∨ (s⊥ ∧ t⊥), by [Kal83] §3 Lemma 3 and Proposition 8. Even if P is not orthomodular, we still have the following important characterizations of centrality. Definition 2.19. Given partial orders P and Q we order P × Q coordinatewise, i.e. (p, q) ≤ (r, s) ⇔ p ≤ r and q ≤ s. If they are orthocomplemented, we make P × Q orthocomplemented by defining (p, q)⊥ = (p⊥, q⊥). Given an ortholattice P and p ∈ P, we say P is canonically isomorphic to [p] × [p⊥] to mean that [p] = [p]p, [p⊥] = [p⊥]p⊥ and the (order preserving) maps (q, r) 7→ q∨ r and q 7→ (q∧ p, q∧ p⊥), from [p] × [p⊥] to P and vice versa, are inverse to each other. Theorem 2.20. If P is an ortholattice then the following are equivalent for p ∈ P. (1) q = (q ∧ p) ∨ (q ∧ p⊥), for all q ∈ P. (2) p is central. (3) P is canonically isomorphic to [p] × [p⊥]. Proof. (1)⇒(3) See [Kal83] §3 Theorem 1 or [Mac64] Theorem 3.2. (3)⇒(2) Immediately verified by coordinatewise calculations. (2)⇒(1) Immediate from the definition of central. (cid:3) Note that (3) is important because it means we can now do calculations coordi- natewise in [p] × [p⊥] rather than P. For example, say we had p ∈ cP and q ∈ P and we want to show that (2.13) If equality fails here then we would have r ∈ cP with r ≥ p ∧ q with r (cid:3) p ∧ c(q). By replacing r with r∧ p∧ c(q) if necessary we may assume that r < p∧ c(q). Then r ∨ (p⊥ ∧ c(q)) ≥ (p ∧ q) ∨ (p⊥ ∧ q) = q even though c(p ∧ q) = p ∧ c(q). r ∨ (p⊥ ∧ c(q)) < (p ∧ c(q)) ∨ (p⊥ ∧ c(q)) = c(q), where we know the first inequality is strict because the inequality in the first coor- dinates in [p] × [p⊥] is strict. This contradicts the fact c(q) is the central cover of q and we are done. Still assuming p ∈ cP, we have q ∧ p = (q ∧ p⊥)⊥q ∈ [q]q. For any r ∈ [q]q, we have r ∧ (q ∧ p) = (r ∧ q) ∧ p = r ∧ p and r ∧ (q ∧ p)⊥q = r ∧ q ∧ (q ∧ p)⊥ = r ∧ (q ∧ p⊥) = r ∧ p⊥. {q ∧ p : p ∈ cP} ⊆ cq[q]q. So r = (r ∧ p) ∨ (r ∧ p⊥) = (r ∧ (q ∧ p)) ∨q (r ∧ (q ∧ p)⊥q ) and q ∧ p ∈ cq[q]q, i.e. (2.14) In general this inclusion can be strict (see Figure 1, where cb[b]b = [b] = {b, c⊥, a⊥, 0} even though cP = {1, 0}), although not for the annihilators in a C-algebra (see Corollary 3.33). This inclusion is also an equality for central elements, i.e. given p ∈ cP, we have [p] = [p]p and cP ∼= c[p] × c[p⊥] so (2.15) cp[p]p = c[p]. 14 TRISTAN BICE Another consequence of (3) is that, for any p ∈ cP, p⊥ is the unique complement of p so, in fact, cP is a Boolean algebra, by [Kal83] §3 Proposition 7. For orthomodular P, the converse also holds. Proposition 2.21. For orthomodular P, p ∈ cP if p⊥ is its only complement. Proof. If p is not central then q > (q ∧ p) ∨ (q ∧ p⊥), for some q ∈ P. Setting q′ = q∧ (q∧ p)⊥, we have p∧ q′ = 0 and, by orthomodularity, q′ > q∧ p⊥ ≥ q′ ∧ p⊥. This implies q′′ = q′ ∨ (p∨ q′)⊥ 6= p⊥ and, again by orthomodularity, p∧ q′′ = p∧ q′ = 0. Also p∨ q′′ = p∨ q′∨ (p∨ q′)⊥ = 1, so q′′ is a complement of p different from p⊥. (cid:3) For order theoretic type decompositions, what we actually need is an infinite version of (3), and the key extra condition required for this is separativity. Theorem 2.22. If P is a separative complete ortholattice, q ∈ P and (pα) ⊆ cP, (1) q ∧W pα =W q ∧ pα, and (2) q = (q ∧W pα) ∨ (q ∧ (W pα)⊥). Proof. Both statements follow from the fact This means that, by Corollary 2.23, given very orthogonal (pα) ⊆ P in a separative complete ortholattice P, we also have (2.18) [(pα)](pα) =Y[pα]pα . [_ pα]W pα ∼=Y[pα]pα . S = [^ p⊥α ] ∪[[pα] is join-dense in P. For this, it suffices to prove S is order-dense, by Proposition 2.8. Now note that, for any q ∈ P, q ∧ pα = 0 implies q ≤ p⊥α , as q = (q ∧ pα) ∨ (q ∧ p⊥α ). have s ∈ S with 0 < s ≤ q. (cid:3) So if this were true for all α, we would have q ≤V p⊥α . In any case, if q > 0, we So, by (2) above and Theorem 2.20 (1), if P is a separative complete ortholattice, (2.16) cP is a complete sublattice. [_ pα] ∼=Y[pα]. Another important consequence is the following. Corollary 2.23. For orthogonal (pα) ⊆ cP in a separative complete ortholattice P, (2.17) and f (q) =Y(pα ∧ q) Proof. Define f : [Wα pα] →Qα[pα] and g :Qα[pα] → [Wα pα] by g((qα)) =_ qα. pα ∧ (qα ∨ (Wβ6=α qβ)) ≤ pα ∧ (qα ∨ p⊥α ) = qα, so f ◦ g is the identity map. But Given a collection of ortholattices (Pα) and (pα), (qα) ∈Q Pα with (qα) ≤ (pα), Take (qα) ⊆ P with qα ≤ pα, for all α. Given that pα commutes with qα, we have g◦ f is also the identity map, by Theorem 2.22 (1) and thus g and f are (canonical) (cid:3) isomorphisms inverse to each other. α ∧ pα), i.e. we have (qα)⊥ ∧ (pα) = (q⊥α α ) ∧ (pα) = (q⊥α ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 15 ⇔ 2.11. Type Decomposition. Definition 2.24. Given a complete ortholattice P, we call T ⊆ P a type ideal if, whenever we have pairwise very orthogonal S ⊆ P, (2.19) If (2.19) holds for arbitrary S ⊆ T then T is a complete ideal. _ S ∈ T. S ⊆ T These type ideals correspond to the type-determining subsets defined in [FP10] §4 (6) in the context of effect algebras. Note that if T is a complete ideal then if T is a type ideal in a complete Boolean algebra P then T is actually a complete ideal. For then cP = P so if p ≤ q ∈ T then q = p∨ (p⊥∧ q) and hence p ∈ T . While T = [W T ], i.e. complete ideals are precisely those subsets of P of the form [p]. And given any (tα) ⊆ T we can define (very) orthogonal (sα) by sα = tα ∧ (Wβ<α tβ)⊥ and transfinite induction givesW tα =W sα ∈ T . The most common examples of type ideals come from the the type relations and type classes to be defined in the following sections. However, there is one important example we can give straight away, namely the equality type ideal P= of any ortholattice P defined by (2.20) This is indeed a type ideal, by (2.17) and (2.18), and P = P= if and only if P is orthomodular, by Proposition 2.16. P= = {p ∈ P : [p] = [p]p}. We call T and S complementary when (W T )∧ (W S) = 0 and (W T )∨ (W S) = 1. Proposition 2.25. Given a separative complete ortholattice P and a type ideal T ⊆ P, the following pairs are complementary complete ideals of cP. (2.21) (2.22) {p ∈ cP : c[p] ∩ T = {0}}. {p ∈ cP : [p] ∩ T = {0}}. T ∩ cP cT Proof. (2.21) As T is a type ideal, so is T ∩cP which, as cP is a complete Boolean algebra, means it is a complete ideal of cP. Letting q =W(T ∩ cP), we immediately have and and c[q⊥] ⊆ {p ∈ cP : c[p] ∩ T = {0}}, while if p ∈ cP satisfies c[p] ∩ T = {0} then, as p ∧ q ∈ T ∩ cP (because p ∧ q ≤ q and q ∈ T ∩ cP), we must have p ∧ q = 0 and hence p ∈ [q⊥], i.e. the inclusion above is an equality. (2.22) By (2.13) and (2.16), c(W pα) = W c(pα) whenever (pα) ⊆ P so cT is a type ideal and thus, as above, a complete ideal of cP. The fact that its complementary complete ideal is {p ∈ cP : [p] ∩ T = {0}} follows in a similar manner, with another application of (2.13). It immediately follows that any type ideal naturally gives rise to the following (cid:3) two type decompositions. Corollary 2.26. Given a type ideal T in a separative complete ortholattice P, there are unique pT , qT ∈ cP such that (2.23) and pT ∈ T qT ∈ cT and c[p⊥T ] ∩ T = {0}. [q⊥T ] ∩ T = {0}. (2.24) 16 TRISTAN BICE It follows immediately from Definition 2.24 that, if P is a complete ortholattice, κ is some (finite or infinite) cardinal and T ⊆ P is a type ideal, then so is (2.25) T κ = {_ S : S ⊆ T and T = κ}. If P is also separative then c(W S) =W c(S), for any S ⊆ P, which means cT = cT κ, as cT is a complete ideal by Proposition 2.25, i.e. qT = qT κ . It also means that pT κ ≤ qT , even though in this case we can have pT < pT κ. 2.12. Homogeneity. We now show how order-density yields homogeneous decom- positions. Definition 2.27. Assume P is a complete ortholattice. We say p ∈ P is T - S = κ, we say that p has order κ or that p is κ-T -homogeneous. We denote the κ-T -homogeneous elements by Tκ. We call p ∈ P T -subhomogeneous if there are we say that p is <κ-T -subhomogeneous. By κ-T -subhomogeneous we mean <κ+-T - subhomogeneous. We denote the T -subhomogeneous and <κ-T -subhomogeneous elements by Tsub and T<κ respectively homogeneous if there are orthogonal S ⊆ T ∩ [p] with p =W S and cS = {c(p)}. If very orthogonal T -homogeneous S with p =W S. If each s ∈ S has order < κ then If P is a separative complete ortholattice and T is a type ideal in P then so is Tsub, T<κ and Tκ, for each cardinal κ. Hence we get the type decompositions in Corollary 2.26, and the central <κ-T -subhomogeneous part is just the join of all the central λ-T -homogeneous parts, for λ < κ, i.e. 1 =_ tκ. pT<κ = _λ<κ pTλ and pTsub =_λ pTλ Also T ⊆ Tκ ⊆ T<κ+ ⊆ T κ (see (2.25)), so cT ⊆ cTκ ⊆ cT<κ+ ⊆ cT κ = cT , for all κ, and cTsub = cT , so pTsub ≤ qTsub = qT . We aim to show that pTsub = qT for suitable T . Also note that the order of a homogeneous element is not unique in general. For one thing, as long as 0 ∈ T , then 0 is κ-T -homogeneous, for all κ. However, this can also happen for non-zero elements, i.e. we can have Tλ ∩ Tκ 6= {0} for λ 6= κ. Theorem 2.28. If T is an order-dense type ideal in a separative complete ortho- lattice P then 1 is T -subhomogeneous. have non-zero t ∈ T ∩ [s] ⊆ Tα. This t would then be very orthogonal to tα so t ∨ tα ∈ Tα and c(t ∨ tα) > c(tα), contradicting the definition of tα. Thus Proof. By (2.22), we may recursively define tα ∈ Tα so that c(tα) =W cTα, where Tα is the type ideal given by Tα = T ∩[(Wβ<α tβ)⊥]. Let sα = c(tα)⊥∧Vβ<α c(tβ) ∈ cP, by (2.16). If s = sα ∧ (Wβ<α tβ)⊥ 6= 0 then, by the order density of T , we would sα = sα ∧Wβ<α tβ =Wβ<α(sα ∧ tβ). Also, for each β < α, Thus (W sα)⊥ =V c(tα) = 0 so the (sα) witness the subhomogeneity of 1. so each sα is homogeneous. As the (tα) are orthogonal, they must be eventually 0. (cid:3) Corollary 2.29. If T is an order-dense type ideal in a separative complete ortho- lattice P and Tλ ∩ Tκ ∩ cP = {0}, whenever λ 6= κ, then there are unique orthogonal (tκ) ⊆ cP with tκ ∈ Tκ, for all κ, and c(sα ∧ tβ) = sα ∧ c(tβ) = sα, ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 17 Proof. Existence follows immediately from Theorem 2.28 by joining all resulting ho- mogeneous elements of the same order. Uniqueness now follows from Corollary 2.26, as Tλ ∩ Tκ ∩ cP = {0}, for λ 6= κ, means that tκ = pTκ , for all κ. (cid:3) 2.13. Type Relations. Definition 2.30. A relation ∼ on a complete ortholattice P is a type relation if, whenever (qα), (rα) ⊆ P are such that (qα ∨ rα) are very orthogonal, (2.26) We call ∼ proper if p ∼ 0 ⇒ p = 0, for all p ∈ P. _ qα ∼_ rα. ∀α(qα ∼ rα) ⇔ For any complete ortholattice P, equality is a trivial example of a proper reflexive type relation, and it is of course the strongest reflexive relation on P. While (2.27) c(p) ≤ c(q) also defines a type relation, the weakest proper type relation. For if - is such a relation then p - q implies c(q)⊥ ∧ p - c(q)⊥ ∧ q = 0 so p ≤ c(q) and hence c(p) ≤ c(q). Thus c(p) = c(q) is the weakest symmetric proper type relation and this will coincide with equal- ity if and only if P = cP, i.e. if and only if P is a Boolean algebra, in which case they both represent the unique proper reflexive symmetric type relation on P. Slightly more interesting examples are given by perspectivity, orthoperspectivity and semiorthoperspectivity, which are type relations by Corollary 2.23. Any type relation ∼ naturally defines other type relations - and -rel by (2.28) (2.29) p - q ⇔ ∃r ∈ [q](p ∼ r), and p -rel q ⇔ ∃r ∈ [q]q(p ∼ r). In general -rel is stronger than -, and they coincide when P is orthomodular, by Proposition 2.16. Even in the non-orthomodular case, if ∼ is weaker than orthop- erspectivity (on P∼ at least), then p ∼ r ≤ q implies r ∼ r⊥q⊥q ≤ q and hence, if ∼ is also transitive, -rel and - again coincide. Proposition 2.31. For any reflexive type relation on a separative complete ortho- lattice P, the following subsets are type ideals. (1) The ∼-finite elements of P. (2) The ∼-orthofinite elements of P. Proof. (1) Say we have very orthogonal (pα) ⊆ P. IfW pα were not ∼-finite, we would haveW pα ∼W qα for some (qα) with qβ < pβ, for some β, by Corollary 2.23. for some qβ < pβ and thenW pα ∼W qα <W pα, where qα = pα for α 6= β, and henceW pα is not ∼-finite. As ∼ is a type relation, we would have pβ ∼ qβ, and hence pβ would not be ∼-finite. On the other hand, if pβ is not ∼-finite, for some β, then pβ ∼ qβ (2) Essentially the same proof as in (1). (cid:3) Theorem 2.32. If ∼ is a finite symmetric transitive type relation on a complete ortholattice P weaker than perspectivity then P is modular and ∼ coincides with ∼p. 18 TRISTAN BICE Proof. If P were not orthomodular, we would have [p] 6= [p]p, for some p ∈ P, and thus q < q⊥p⊥p , for any q ∈ [p] \ [p]p, even though q⊥p⊥p ∼ q, by Proposition 2.10, contradicting finiteness. Perspectivity must also be finite, being weaker than ∼, and this comibined with orthomodularity means P is modular, by (2.9). As P is complete, it is a continuous geometry, by [Kap55]. Following the proof of [Kap51] Theorem 6.6(c), we note that [vN60] Part III Theorem 2.7 means that perspectivity satisfies generalized comparison, i.e. for any q, r ∈ P, we have p ∈ cP (the centre we have defined is the same as the centre defined at the start of [vN60] Part III Chapter I, either by Theorem 2.20 (3) or by [vN60] Part I Theorem 5.4 and Proposition 2.21) with p ∧ q perspective to some s ≤ p ∧ r and p⊥ ∧ r perspective to some t ≤ p⊥ ∧ q. So if q ∼ r then s ∼ p ∧ q ∼ p ∧ r which, by finiteness, means s = p ∧ r and, likewise, t = p⊥ ∧ q. So p ∧ q and p ∧ r are perspective, as are p⊥ ∧ q and p⊥ ∧ r so, by Theorem 2.20 (3), q and r are also perspective. (cid:3) 2.14. Type Classes. Definition 2.33. Let C be the class of complete lattices with a relation ⊥. We call T ⊆ C a type class if it is closed under isomorphisms and, for (Pα) ⊆ T, (2.30) (Pα) ⊆ T ⇔ Y Pα ∈ T Proposition 2.34. If T is a type class and P is a separative complete ortholattice then the following subsets of P are type ideals. (2.31) PT = {p ∈ P : [p] ∈ T}. PTrel = {p ∈ P : [p]p ∈ T}. Proof. PT and PTrel are type ideals by (2.17) and (2.18) respectively. (cid:3) Any subclass of C consisting of all those complete lattices satisfying some uni- versally quantified sentence in the language {0,≤,⊥,∧,∨} will be a type class, like the following important examples -- (2.32) (2.33) (2.34) (2.35) D = {P ∈ C : P is a distributive}, M = {P ∈ C : P is a modular}, and O = {P ∈ C : P is a orthomodular}. (note that Proposition 2.16(7) characterizes orthomodularity just with an orthogo- nality relation ⊥, rather than an orthocomplementation ⊥, which we may not have for [p]). We also have the type class S = {P ∈ C : P is a separative}, although this does not lead to any interesting type decompositions in a separative complete ortholattice P because then P = PS and, even though we may have P 6= PSrel (see the comments after Proposition 2.16), we will still have cP ⊆ PSrel. Now, using the type ideals and type decompositions coming from several instances of (2.31) and (2.24) respectively, we can define (2.36) (2.37) (2.38) (2.39) pI = qPD, pII = q⊥PD ∧ qPM, pIII = q⊥PM ∧ qPO, and pIV = q⊥PO. If P is the projection lattice P(A) of a von Neumann algebra A then these projections do indeed correspond to those you get from the classical von Neumann ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 19 algebra type decomposition. Specifically, to see that pIA corresponds to the type I part, see the comments after Proposition 3.13, and to see that this even extends to annihilators in C-algebras, see Theorem 3.49. Also pIIA corresponds the classical type II part because a von Neumann algebra is finite if and only if its projection lattice is modular (see [Ber72] §34 Proposition 1 and [Kap55] Theorem on page 1). This fact can also be extended, at least partially, to annihilators in C-algebras, by Theorem 2.32 and Proposition 3.45. And P(A) is necessarily orthomodular, which means pIV = 0, so pIIIA also corresponds to the type III part in this case. We do not know if the same is true for the annihilators in a C-algebra, i.e. whether there exist any type IV C-algebras at all. The annihilators in such a C-algebra would be wildly different from any ortholattices seen before in operator algebras, as they would fail to be orthomodular in a very strong way. We can also use Corollary 2.26 and (2.25), to define pIn = pPDn, and pII1 = q⊥PD ∧ pPM. If P is again the projection lattice P(A) of a von Neumann algebra A then pInA and pII1A are again the type In and II1 part respectively in the classical von Neumann type decomposition. In this case, a supremum of finite projections is again finite, which means that pIn ≤ pPM and Pn M = PM. In an arbitrary separative complete ortholattice P we do not have any notion of Murray-von Neumann equivalence and so there is no reason to believe these facts remain true in general. For annihilators in C-algebras we will, however, define an analgous notion (see §3.7) which will enable us to prove the first of these facts (see Corollary 3.67(3)). We could have also used the relative type-ideals in (2.32), rather than those in (2.31), to define the above type classes, as given below. (2.40) (2.41) (2.42) (2.43) pIrel = qPDrel , pIIrel = q⊥PDrel ∧ qPMrel , pIIIrel = q⊥PMrel ∧ qPOrel , and pIVrel = q⊥POrel . If P is the projection lattice in a von Neumann algebra then, as this is orthomodular, we have [p] = [p]p (see Proposition 2.16 (3)) and this decomposition is exactly the same as the previous one. But for annihilators in C-algebras, it may well be different, although we do at least know that pIrel = pI in this case, by Theorem 3.49. Indeed, these relative versions using [p]p are really more natural for annihilators, as using [p] instead might result in the rather awkward situation that an annihilator B could be modular in the larger C-algebra A even though it is not modular in itself (i.e. [B] could be modular lattice even when [B]B is not), or vice versa. Although we could avoid this problem by using the even smaller type-ideal of P given by {p ∈ P : [p]p = [p] ∈ M} ⊆ PMrel ∩ PM. In fact, this is just the intersection of PM (or PMrel) with the equality type-ideal P= defined in (2.20). Taking intersections of the various resulting type-ideals we get from these type-classes would yield more type-ideals leading to even finer type decompositions. Although, whether all these potential types are actually realized by certain C-algebras, we do not know. 20 TRISTAN BICE 2.15. Boolean Elements. The following terminology comes from [Che91] Definition 2.35. An ortholattice P has the relative centre property if, for all p ∈ P, cp[p]p = {p ∧ q : q ∈ cP}. This is equivalent to saying c(q) ∧ p = cp(q) whenever q ∈ [p]p. Definition 2.36. A relation - on a complete ortholattice P satisfies generalized comparison if, for any q, r ∈ P, there exists p ∈ cP with p ∧ q - p ∧ r and p⊥ ∧ r - p⊥ ∧ q. Proposition 2.37. If ∼ is a symmetric type relation and -rel is defined as in (2.29) then -rel satisfies generalized comparison if and only if, for all q, r ∈ P, ∃u ∈ [q]q∃v ∈ [r]r(u ∼ v while q ∧ u⊥ is very orthogonal to r ∧ v⊥). (2.44) Proof. If -rel satisfies generalized comparison then, for any q, r ∈ P we have p ∈ cP, s ∈ [p ∧ r]r, t ∈ [p⊥ ∧ q]q such that p ∧ q ∼ s and p⊥ ∧ r ∼ t. If ∼ is a symmetric type relation then u = (p ∧ q) ∨q t = (p ∧ q) ∨ t ∼ s ∨ (p⊥ ∧ r) = s ∨r (p⊥ ∧ r) = v while q ∧ u⊥ ≤ p⊥ and r ∧ v⊥ ≤ p. Conversely, given such a u and v, setting p = c(r ∧ v⊥) gives p ∧ q ≤ u⊥q⊥q = u and hence and, likewise, p⊥ ∧ r = p⊥ ∧ v ∼ p⊥ ∧ u ≤ p⊥ ∧ q. p ∧ q = p ∧ u ∼ p ∧ v ≤ p ∧ r (cid:3) As noted after Definition 2.30, p ∼ q ⇔ c(p) = c(q) defines a (proper reflexive symmetric) type relation on any complete ortholattice P. To see that -rel then satisfies generalized comparison, simply take any q, r ∈ P, set u = q ∧ c(r) and v = r ∧ c(q) and note that c(u) = c(q) ∧ c(r) = c(v) while q ∧ u⊥ ≤ c(r)⊥ and r ∧ v⊥ ≤ c(q)⊥. The next result shows that, under suitable extra hypotheses, it is the only such relation on PDrel. In fact it shows that, if ∼ is a proper type relation on a complete Boolean algebra P then generalized comparison is equivalent to reflexivity, i.e. in this case generalized comparison just says ∼ is equality. Theorem 2.38. If P is a complete ortholattice with the relative centre property, ∼ is a proper symmetric type relation and -rel satisfies generalized comparison then (1) For p ∈ PDrel and q ∈ P, c(p) ≤ c(q) ⇔ p -rel q. (2) For p, q ∈ PDrel, c(p) = c(q) ⇔ p ∼ q. (3) When p, q ∈ PDrel and p ∼ q, ∼ is an orthoisomorphism on [p]p × [q]q. Proof. (1) The ⇐ part is immediate from the comments after (2.27). For the ⇒ part, note that, by generalized comparison, we have s ∈ [p]p and t ∈ [q]q with s ∼ t. As P has the relative centre property and every element of a Boolean algebra is central, we have c(p ∧ s⊥) ∧ p = cp(p ∧ s⊥) = p ∧ s⊥. As ∼ is a type relation, 0 = c(p ∧ s⊥) ∧ s ∼ c(p ∧ s⊥) ∧ t. But ∼ is also proper and hence c(p ∧ s⊥) ∧ t = 0 so c(p ∧ s⊥) ∧ q = 0 and c(p ∧ s⊥) ∧ c(q) = 0. As c(p) ≤ c(q), we must have p ∧ s⊥ = 0 and therefore p = s, i.e. p ∼ t ≤ q. (2) A symmetric argument yields q = t too, i.e. p ∼ q. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 21 (3) Say we had r, s ∈ [p]p and t ∈ [q]q with r ∼ t and s ∼ t. We would then have c(r) = c(t) = c(s) and hence r = cp(r) = c(r) ∧ p = c(s) ∧ p = cp(s) = s. Also, for any s ∈ [p]p, c(c(s) ∧ q) = c(s) ∧ c(q) = c(s) ∧ c(p) = c(s) and hence s ∼ c(s) ∧ q ∈ [q]q. Thus ∼ restricted to [p]p × [q]q is the function s 7→ c(s)∧q. It is clearly order preserving and a symmetric argument shows that the same is true of the inverse function t 7→ c(t)∧ p. Finally note that, as [p]p and [q]q are Boolean algebras, any order isomorphism is actually an orthoisomorphism. (cid:3) 3. Annihilators and Projections 3.1. Annihilators. Throughout, A denotes a C-algebra with positive elements A+ = {aa∗ : a ∈ A}, self-adjoint elements Asa = {a = a∗ : a ∈ A} (or Asa = {a − b : a, b ∈ A+}), unit ball A1 = {a : a ≤ 1} and projections P(A) = {a ∈ A : a = a∗ = a2}, where p ≤ q means p = pq, for p, q ∈ P(A). Consider the following relations on A. aLb ⇔ ab∗ = 0. a⊥b ⇔ aLb and aLb∗. a⊤b ⇔ a⊥b and a∗⊥b. Taking the orthocompletion of A w.r.t. any of these relations amounts to the same thing. To see why, first note that {a}⊥ = {a∗a}⊤ and {a}⊤ = {a, a∗}⊥ so [A]⊥ = [A]⊤ ⊆ {S ⊆ A : S = S∗} and, for any S ⊆ A with S = S∗(= {s∗ : s ∈ S}), we have S⊥ = S⊤. As ⊤ is a preorthogonality relation, it follows from §2.2 that [A]⊥ is a complete ortholattice. In fact, all we have used here is the fact that A is a -ring with proper involution. Also L is a preorthogonality relation on A and the map B 7→ B ∩ B∗ is an orthoisomorphism from [A]L to [A]⊥. Every element of [A]L is clearly a closed left ideal and so every element of [A]⊥ is a hereditary C-subalgebra of A (see [Ped79] Theorem 1.5.2). As C-algebras, rather than their left ideals, are our primary object of study, and the equivalence in §3.7 is slightly easier to define with ⊥ rather than ⊤, we shall focus on the relation ⊥. The elements of [A]⊥ will be called annihilators of A. Another thing to note is that the above relations all agree on Asa. The fact that {a}⊥ = {a∗a}⊥, shows that A+ is join-dense in A, w.r.t. to the preorder induced by ⊥, and thus [A]⊥ ∼= [A+]⊥. Indeed, it will often be convenient in proofs to use the elements of A+ rather than A. If A has real rank zero, then every hereditary C-subalgebra contains an approximate unit of projections, so every annihilator will be an annihilator of subset of projections, i.e. P(A) will be join-dense in A and We generalize this observation in Corollary 3.27. [A]⊥ ∼= [P(A)]⊥. 22 TRISTAN BICE One other thing worth pointing out is that we have something extra on [A]⊥ that we do not have for an arbitrary ortholattice. Specifically, we actually have a function ·· from [A]⊥ × [A]⊥ to [0, 1] which quantifies the degree of orthogonality of B, C ∈ [A]⊥ given by (3.1) BC = b∈B1 sup +,c∈C 1 + bc. The important properties of this function are that, for B, C, D ∈ [A]⊥, BC = CB, B 6= {0} ⇔ BB = 1, B ⊥ C ⇔ BC = 0, and C ⊆ D ⇒ BC ≤ BD. Indeed, these properties could be derived from the relevant properties of · · on + (and [A]⊥ ∼= [A1 +]⊥). If, + and the the fact that a ⊥ b ⇔ ab = 0, for a, b ∈ A1 A1 further, BC⊥ satisfies the triangle inequality, i.e. for all B, C, D ∈ [A]⊥, BD⊥ ≤ BC⊥ + CD⊥, then we naturally call · · an orthonorm. We do not know if (3.1) always yields an orthonorm on [A]⊥, although we show it does for certain C-algebras in Corollary 3.67(2). We can even use ·· to define the orthospectrum of B, C ∈ [A]⊥ by σ(BC) = {BD2 : D ∈ [C] and ((D ∨ B⊥) ∧ B) ⊥ ((D⊥B ∨ B⊥) ∧ B)}, except when B = C = A, in which case we define σ(AA) = {1} (see Theorem 3.19). 3.2. Non-Commutative Topology. From this point on, it will be convenient to assume that A is concretely represented faithfully and non-degenerately on some Hilbert space H. This is, of course, valid, thanks to the standard GNS construc- tion. One canonical choice would be the universal representation, which has the advantage that its projections distinguish all hereditary C-subalgebras of A. The same is also true for the atomic representation, by by [Ped79] Proposition 4.3.13 and Theorem 4.3.15. However, we are primarily concerned with annihilators, and the projections in any faithful representation of A still distinguish the annihilators (see Theorem 3.4). So, as long as we fix it throughout, any faithful non-degenerate representation will do. However, we will still use standard non-commutative topology terminology (see [Ake69], [Ake70] and [Ped79] 3.11.10) which, traditionally, has only be used with reference to projections in A∗∗. This restriction is certainly convenient for many of the proofs, and necessary for some of the results, but many of the results themselves are valid also for arbitrary representations, thanks to the following property of A∗∗. Theorem 3.1. Any representation π of A has a normal extension to A∗∗. Proof. See [Ped79] Theorem 3.7.7. (cid:3) Definition 3.2. We call p ∈ P(B(H)) open if p = sup(pα), for some increasing net (pα) ⊆ A1 +, and closed if p⊥ is open. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 23 Equivalently, a projection p is closed if and only if p = inf(pα) for some decreasing net (pα) ⊆ 1 − A1 +. And the supremum/infimum of an increasing/decreasing net in A coincides with the limit of that net in the strong (or weak) topology, so open and closed projections always lie in the double commutant A′′ of A. We will denote the sets of open and closed projections by P(A′′)◦ and P(A′′) respectively. Also note that the identity operator 1 is open, as we are dealing with a non-degenerate representation. + = (A + C1)1 + (or even eA1 +) = pB. So a projection is open precisely when it is the supremum of B1 If B is a C-subalgebra of A, then {b ∈ B+ : b < 1} is directed (see [Ped79] +) ∈ P(A′′)◦. Conversely, Theorem 1.4.2) and hence has a supremum pB = sup(B1 if p ∈ P(A′′)◦, then B = pAp ∩ A is a hereditary C-subalgebra of A with p = sup(B1 + for some hereditary C-subalgebra B of A.4 The open and closed projections can also be characterized as the limits of increasing and decreasing sequences respectively sa), a surprisingly non-trivial fact when A is not in eA1 unital (and hence A 6= eA). Another important fact is that the collection of open projections P(A′′)◦ is norm closed (see [Ped79] Proposition 3.11.9). Definition 3.3. We define the interior p◦ of p ∈ P(B(H)) by p◦ = sup(pAp∩ A)1 + and the closure by p = p⊥◦⊥. We call p topologically regular 5 if p = p◦ and denote the collection of all topologically regular open projections by P(A′′)◦. It follows that p◦ is the largest open projection satisfying p◦ ≤ p and p is the smallest closed projection satisfying p ≤ p. The existence of such projections also follows from the fact that the supremum of a collection of open projections is open and the infimum of a collection of closed projections is closed (see [Ake69] Proposition II.5 and combine it with Theorem 3.1 to obtain the result for arbitrary representations). Also note that the interior of any closed projection is in fact topologically regular. For if p = q◦, for some closed q ∈ P(B(H)), then p ≤ q so (3.2) p◦ ≤ q◦ = p ≤ p◦. One more important thing to note is the difference between complements of annihilators and their corresponding projections. Specifically, given B ∈ [A]⊥, we may have pB⊥ 6= p⊥B, and one could view the complications of extending projection results to annihilators as all stemming from this fact. Indeed, pB⊥ is open while p⊥B is closed, so they could not be equal unless they were clopen. This occurs when these projections, or their complements, lie in A itself, i.e. if B = pAp for some p ∈ A. In fact, if A′′ = A∗∗ (i.e. if we are dealing with the universal representation of A) and 1 ∈ A then the clopen projections are precisely those in A (see [Ake69] Proposition II.18). However, B⊥ = p⊥BAp⊥B ∩ A so, by definition, we do always have (3.3) pB⊥ = p⊥◦B . Theorem 3.4. There are order isomorphisms between [A]⊥ and P(A′′)◦ given by B 7→ pB and p 7→ pAp ∩ A 4Incidentally, the closed projections of A can similarly be characterized as the infimums of norm filters (see [Bic11] Corollary 3.4), although we will not have further occasion to refer to these. 5As far as we know, such projections have not been considered or named before. They are the analog of regular open subsets of a topological space, although we are averse to simply calling them regular, as this is already a standard term meaning something different (see [Ake69] Definition II.11 and the discussion at the start of §3.5). 24 TRISTAN BICE Proof. For B ∈ [A]⊥, pB = pB⊥⊥ = p⊥◦B⊥, by (3.3). But pB⊥ = sup(B⊥)1 + is open so p⊥B⊥ is closed and hence its interior, pB, is a topologically regular projection. Also, if pB = p = pC , for B, C ∈ [A]⊥, then B⊥ = p⊥Ap⊥ ∩ A = C⊥ and hence B = B⊥⊥ = C⊥⊥ = C, so the first map is injective. For topologically regular p, (p⊥◦Ap⊥◦ ∩ A)⊥ = p⊥◦⊥Ap⊥◦⊥ ∩ A = p⊥◦⊥◦Ap⊥◦⊥◦ ∩ A = p◦Ap◦ ∩ A = pAp ∩ A, so pAp ∩ A is indeed an annihilator. Also ppAp∩A = sup(pAp ∩ A)1 + = p, as p is open. This, combined with the injectivity of the first map, shows that these maps are indeed inverse to each other which, as they are immediately seen to be order (cid:3) preserving, means they are also order isomorphisms. When making order theoretic statements about projections we must now always take care to note whether they are with respect to P(A′′) or P(A′′)◦, and we shall adopt the convention that, by default, it is the order structure of P(A′′) being referred to unless otherwise specified, with subscripts for example. For while it follows from Theorem 3.4 that P(A′′)◦ is a complete lattice, it is not a sublattice of P(A′′). When p, q ∈ P(A′′)◦ satisfy pq = qp, we do have p ∧ q = p ∧P(A′′) ◦ q = pq, thanks to [Ake69] Theorem II.7, so infimums do at least agree in this case, but even commutativity does not imply that supremums agree. Moreover, as mentioned above, the orthocomplement functions in the two structures are different. However, there are relations between some order theoretic concepts in P(A′′) and the corresponding concepts in P(A′′)◦. For example, the concept of centrality coincides in the two structures (see Theorem 3.30 and the comments that follow) and the following result shows that commutativity is usually stronger in P(A′′) than in P(A′′)◦ (and it can be strictly stronger -- see the comments after Example 3.82). Proposition 3.5. If P(A′′)◦ is orthomodular then, for p, q ∈ P(A′′)◦, ∃r ∈ {p, p, p⊥, p⊥}∃s ∈ {q, q, q⊥, q⊥}(rs = sr) ⇒ p and q commute in P(A′′)◦. Proof. For any projections p and q in a C-algebra, pq = qp ⇔ pq⊥ = q⊥p, and any p and q in an orthomodular lattice commute if and only if p and q⊥ commute. Thus, without loss of generality, we may assume that pq = qp and hence pq = p∧P(A′′) ◦ q, by the comments above. As P(A′′)◦ is orthomodular, we have r ∈ P(A′′)◦ with pqr = 0, r ≤ p and p = pq ∨P(A′′) ◦ r = pq ∨ r◦. But then p⊥qr = qp⊥r = 0 and hence qr = 0. So r ≤ p∧P(A′′) ◦ q⊥P(A′′ ) ◦ and hence p = pq ∨P(A′′) ◦ r ≤ (p ∧P(A′′) ◦ q) ∨P(A′′) ◦ (p ∧P(A′′) ◦ q⊥P(A′′ ) ◦ ) ≤ p. As P(A′′)◦ is orthomodular, we are done, by the comments after Definition 2.18. (cid:3) In fact, the following result shows that orthomodularity itself is only an issue when q < p and pq 6= qp (which is indeed possible, by Example 3.81). Combining this argument, the proof of Proposition 3.5 and [Kal83] §3 Lemma 3, we get that, even without orthomodularity, if p, q ∈ P(A′′)◦ then ∀r, s ∈ {p, p⊥, q, q⊥}(rs = sr and rrs = rsr) ⇒ p and q commute in P(A′′)◦. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 25 Proposition 3.6. If p, q ∈ P(A′′)◦, q < p and pq = qp then p = q ∨P(A′′) Proof. Note q ∨ q⊥p ≥ q∨ q⊥p ≥ qp∨ q⊥p = p so p ≤ q ∨ q⊥p◦ = q∨P(A′′) ◦ q⊥p. ◦ q⊥p. (cid:3) We also note some elementary facts about ideals and their associated projections. Proposition 3.7. p ∈ P(A′′)◦ is central if and only if pAp ∩ A is an ideal. Proof. If pAp∩ A is an ideal in A then its weak closure is an ideal in A′′ containing p as its unit. Thus, for any a ∈ A we have ap = pap = pa so p ∈ A′. On the other hand, if p is central then, for any a ∈ A and b ∈ pAp ∩ A, we have ab = apbp = pabp ∈ pAp ∩ A and ba = pbpa = pbap ∈ pAp ∩ A. (cid:3) Proposition 3.8. If p ∈ P(A′′) is central then so are both p and p◦. Proof. If a ∈ A and ap = 0 then abp = apb = 0 = bap, for any b ∈ A, so {p}⊥ is an ideal. Thus p{p}⊥ = p⊥◦ is central, by Proposition 3.7. As p is central if and only if p⊥ is, we are done. (cid:3) In particular, if I is an ideal in A then pI ⊥ = p⊥I is central and hence I⊥ is also an ideal (although this can also be proved by elementary algebraic means). Of course, in the commutative case, all the concepts in this subsection correspond to their usual topological counterparts. Specifically, if A = C0(X) for some locally compact topological space then the atomic representation identifies every element of C0(X) with the multiplication operator on l2(X) it defines. Then A′′ is the set of all bounded multiplicaiton operators on l2(X), which can naturally be identified with l∞(X) in the same way. Under this identification, projections are just characteristic functions χS of subsets S of X, where χS is open, closed or topologically regular if and only if S is, in the topology of X (and note that the characteristic function of an open (closed) set is lower (upper) semicontinuous, a fact which will be generalized in §3.10). In particular, any open subset S of X that is open but not (topologically) regular corresponds to an open projection χS that is not topologically regular, which itself corresponds to a hereditary C-subalgebra that is not an annihilator, contradicting [Arz13] Lemma 16.1 (with e = f = χS). For example, we could have X = [−1, 1] and S = [−1, 0) ∪ (0, 1]. Incidentally, the atomic representation will usually not be the same as the uni- versal representation, even in the commutative case considered in the previous paragraph, illustrating that the universal representation may not always be the best to work with. Above, we could also consider the subrepresentation on l2(Y ), for some Y ⊆ X, which will still be faithful as long as Y is dense in X. This is sometimes nicer in some sense, for example if A = Cb(N) ∼= C(βN) where βN is the Stone- Cech compactification of N, we can consider the faithful subrepresentation on l2(N), which has the advantage that all projections in A′′ are clopen (as all sub- sets of N are). These facts provide some justification for our choice to work with arbitrary representations. 3.3. Spectral Projections. Some quite useful open and closed projections are the spectral projections of self-adjoint operators corresponding to open and closed 26 TRISTAN BICE intervals of R. First, we define continuous functions fr,s on R, for r < s, like so Also, for future reference, define fδ = fδ/2,δ, for δ > 0. Using the continuous functional calculus and the fact we can take infimums and supremums of monotone nets in A′′+, we further define, for any a ∈ Asa, fr,s(t) = 0 t − r s − r 1 for t ∈ (−∞, r) for t ∈ [r, s] for t ∈ (s,∞). fr,s(a), and a[s,∞) = inf a(s,∞) = sup r<s fs,r(a). r>s We can similarly define spectral projections aS for any open or closed (even Borel) subset S ⊆ R. As weak/strong limits of elements that commute with a, spectral projections also commute with a so a(s,∞)a = aa(s,∞) = a(s,∞)aa(s,∞) ∈ A′′sa, and likewise for a[s,∞). We also have the following important operator inequalities. aa(−∞,s] ≤ sa(−∞,s] and aa[s,∞) ≥ sa[s,∞). We also define [a] = (aa∗)(0,∞) = the projection onto R(a), the norm closure of the range of a. Note [a∗]⊥ = (a∗a)(−∞,0] is the projection onto N (a), the kernel of a. In fact, these inequalities (almost) uniquely define the spectral projections. Specif- ically, for any a ∈ B(H)+, a(−∞,s] and a(s,∞) are the unique complementary or- thogonal projections in B(H) such that hav, vi ≤ s, for all unit v ∈ R(a(−∞,s]), and hav, vi > s, for all unit v ∈ R(a(s,∞)). Using this fact we obtain the following result, which will be useful later on. Proposition 3.9. For any s, t with 0 ≤ s < t and a ∈ B(H), (aa∗)(s,t] = [a(a∗a)(s,t]]. Proof. First note that the map p 7→ [ap] preserves the orthogonality of spectral projections of a∗a. Specifically, if S and T are disjoint Borel subsets of R+ \ {0} then, as [a∗a(a∗a)S] ≤ (a∗a)S ⊥ (a∗a)T , we have [a(a∗a)S] ⊥ [a(a∗a)T ]. In particular, this holds for S = (0, s] and T = (s,∞] and the result will follow if we can show that (aa∗)(0,s] = [a(a∗a)(0,s]] and (aa∗)(s,∞] = [a(a∗a)(s,∞]]. Note that (aa∗){0} = [a]⊥ so, by the comments above, we just need to show haa∗av, avi ≤ shav, avi, for all v ∈ R((a∗a)(0,s]), and haa∗av, avi > shav, avi, for all non-zero v ∈ R((a∗a)(s,∞]). But we immediately have ha∗a(a∗a− s)v, vi ≤ 0, for all v ∈ R((a∗a)(0,s]), while also ha∗a(a∗a− s)v, vi > 0, for non-zero v ∈ R((a∗a)(s,∞]), (cid:3) so we are done. Proposition 3.12 below indicates why spectral projections are a convenient tool when dealing with annihilators. It also gives an idea of how plentiful they are. For example, assume A is infinite dimensional so we have a ∈ A+ with σ(a) infinite. Then define a sequence (fn) of continuous functions from σ(a) to [0, 1] with the sets (f−1 n (0, 1]) disjoint and non-empty. These give rise to infinitely many orthogonal ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 27 annihilators {fn(a)}⊥⊥ in A. then we get continuum many annihilators BN = {gN (a)}⊥⊥ that, while no longer orthogonal, are still far apart in the sense that pBN − pBM = 1, for all distinct M, N ⊆ N. In fact, if we let gN = Pn∈N 2−nfn, for N ⊆ N, First, though, we prove the following simple, but useful, algebraic lemma. Lemma 3.10. For any x ∈ A, a, b ∈ A+ and α, β, γ > 0, xaαbβaγ = 0 ⇔ xaαb = 0. Proof. Note that ycc∗ = 0 ⇔ yc = 0, for all y, c ∈ a, as ycc∗ = 0 gives 0 = ycc∗y∗ = (yc)(yc)∗0 and hence yc = 0. Applying this to y = xaαbβ and c = aγ/2 we see that xaαbβaγ = 0 implies xaαbβaγ/2 = 0. We may continue to halve the last exponent in this way until it gets below α, and then multiply on the right by a suitable exponent of a to make it actually equal α. Then applying the note again, this time with y = x and c = aαbβ/2, we get xaαbβ/2 = 0. By reducing the last exponent again, and increasing it at the end if necessary, we finally get xaαb = 0. (cid:3) The converse is similar. This lemma holds even if A is just an arbitrary -ring with proper involution, as long as you can also take positive square roots. Note that by iterating it we also get the same result for arbitrarily long sequences of (powers of) a's and b's. If x = 1 then we actually get the slightly better result aαbβaγ = 0 ⇔ ab = 0. However, for arbitrary x, we must keep the α exponent (e.g. if A = M2 ∼= B(C2), x is the projection onto C(2,−1), a =(cid:20)1 0 0 2(cid:21) and b is the projection onto C(1, 1) +, S ⊆ A+ and as = s, for all s ∈ S, then ab = b, for then xaαb = 0 if and only if α = 1). Proposition 3.11. If a ∈ A1 all b ∈ S⊥⊥. Proof. If we had 1 ∈ A then, as as = s ⇔ (1− a)s = 0, it would immediately follow that 1− a ∈ S⊥ and hence (1− a)b = 0, for all b ∈ S⊥⊥. Even if 1 /∈ A, we still have (1 − a)b(1 − a) ∈ S⊥, for any b ∈ A. Thus, if b ∈ S⊥⊥+ , then (1 − a)b(1 − a)b = 0 and hence b(1 − a) = 0 = (1 − a)b, by Lemma 3.10, i.e. ab = b. As any C-algebra (cid:3) is generated by its positive elements, this completes the proof. One immediate consequence of Proposition 3.11 is that, even if [B]B is strictly contained in [B], it is still order-dense in [B] (see (2.2)). In fact, for all b ∈ B+, we can find D ∈ [B]B contained bAb, the hereditary C-subalgebra generated by b. Just let D = {f2δ(b)}⊥B⊥B , for δ < b, and note that, as fδ(b)f2δ(b) = f2δ(b), we have d = fδ(b)d = fδ(b)dfδ(b) ∈ bAb, for all d ∈ D+. Proposition 3.12. If a ∈ A1 + and f (a) ∈ A, for some f : [0, 1] → [0, 1], then {f (a)}⊥⊥ ⊆ af −1(0,1]Aaf −1(0,1]. Proof. If 0 ∈ R\f−1(0, 1], take continuous g on [0, 1] with g−1{0} = R\f−1(0, 1], and hence g(0) = 0 so g(a) ∈ A. Also, g(a) ∈ {f (a)}⊥, so {f (a)}⊥⊥ ⊆ {g(a)}⊥ ⊆ g(a){0}Ag(a){0} = af −1(0,1]Aaf −1(0,1]. 28 TRISTAN BICE On the other hand, if 0 /∈ R\f−1{0} then take continuous g on [0, 1] with g−1{1} = R\f−1{0} and g(0) = 0, so g(a) ∈ A and g(a)b = b. By Proposition 3.11, g(a)c = c, for all c ∈ {b}⊥⊥, and hence {b}⊥⊥ ⊆ g(a){1}Ag(a){1} = af −1(0,1]Aaf −1(0,1]. (cid:3) 3.4. Projection Properties. We now point out some basic properties of projec- tions, useful in their own right, but also good to keep in mind as results that might admit generalization to the annihilators in some way. Indeed, §3.5 is devoted to proving some of these generalizations. The projections in any C-algebra are orthomodular (if A is not unital then q⊥ is not well-defined, but we can still interpret p∧ q⊥ as denoting the largest projection below q that annihilates p, when such a projection exists). In fact, we have the following order theoretic characterizations of commutatitivity, of which (3.4)⇒(3.6) implies orthomodularity (we should mention that the projections in a C-algebra do not always form a lattice, so (3.5) and (3.6) should be interpretted as saying the given supremums and infimums actually exist and satisfy the given identity). Proposition 3.13. For p, q ∈ P(A), the following are equivalent. (3.4) pq = qp. (3.5) (3.6) Proof. p ∧ q = p ∧ (p⊥ ∨ q). p = (p ∧ q) ∨ (p ∧ q⊥). (3.4)⇒(3.6) Note that p ∧ q = pq when (3.4) holds so p = pq + pq⊥ = (p ∧ q) ∨ (p ∧ q⊥). (3.6)⇒(3.4) As p ∧ q ≤ q, p ∧ q commutes with q. Likewise, p ∧ q⊥ commutes with q⊥ and hence q so, if (3.6) holds, p = (p ∧ q) + (p ∧ q⊥) commutes with q too. (3.4)⇒(3.5) By (3.4), p ∧ (p⊥ ∨ q) = p ∧ (p ∧ q⊥)⊥ = p(pq⊥)⊥ = p − pq⊥ = pq = p ∧ q. (3.5)⇒(3.6) As p ∧ q⊥ ≤ p, we have p ∧ (p⊥ ∨ q) = p ∧ (p ∧ q⊥)⊥ = p − p ∧ q⊥ which, if (cid:3) (3.5) holds, means p = p ∧ q + p ∧ q⊥ = (p ∧ q) ∨ (p ∧ q⊥). So, by Proposition 3.13, if A is a commutative C-algebra then every projection is central in P(A) (see Definition 2.18) and hence P(A) is a Boolean algebra (see the comments before Proposition 2.21). Conversely, if A is non-commutative C- algebra that is generated by its projections then pq 6= qp for some p, q ∈ P(A) and hence P(A) is not a Boolean algebra, again by Proposition 3.13. In particular, if A is a von Neumann (or AW-) algebra and p ∈ P(A) then [p](= [p]p, as P(A) is orthomodular) is a Boolean algebra precisely when pAp is commutative, i.e. precisely when p is an abelian projection. Thus if pI is obtained as in (2.36) (with P = P(A)) then pIA is indeed the type I part of A, in the classical von Neumann algebra type decomposition. Roughly speaking, the following result says that, for a projection p, being far from q⊥ (as in (1)) is equivalent to being close to a subprojection of q (as in (2) and (3)), where λ here quantifies this distance. Proposition 3.14. For p, q ∈ P(A) and λ ∈ [0, 1], the following are equivalent. (1) pq⊥2 ≤ λ. (2) pqp ≥ (1 − λ)p. (3) There exists r ∈ P(A) with r ≤ q and r − p2 ≤ λ. Proof. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 29 (1)⇔(2) Just note pq⊥2 ≤ λ ⇔ pq⊥p ≤ λp ⇔ (1 − λ)p ≤ pqp. (2)⇒(3) If λ = 1 let r = q. Otherwise inf(σ(qpq)\{0}) ≥ 1 − λ > 0 so r = [qp] ∈ A (2)⇐(3) Given such an r, simply note that pq⊥2 ≤ pr⊥2 ≤ p − r2 ≤ λ. and r − p2 ≤ λ. (cid:3) Similarly, the following says that a non-zero projection p can not be simultane- ously far from both another projection and its complement. In fact, (3.7) is still valid for p ∈ A+ with p = 1 (although q still has to be a projection). Proposition 3.15. For all p, q ∈ P(A) with p 6= 0, pq2 + pq⊥2 ≥ 1. (3.7) Moreover, pq2 + pq⊥2 = 1 if and only if σpAp(pqp) is a singleton. Proof. For (3.7), simply take v ∈ R(p)\{0} and note that v2 = qv2 + q⊥v2 ≤ (qp2 + q⊥p2)v2. If pq2 + pq⊥2 = 1 then, setting λ = 1 − pq⊥2, we have pq2 = λ and hence pqp ≤ λp ≤ pqp (using Proposition 3.14 (1)⇒(2) for the last inequality), i.e. pqp = λp and hence σpAp(pqp) = {λ}. Conversely, if σpAp(pqp) = {λ} for some λ ∈ [0, 1] then pqp = λp and pq⊥p = (1−λ)p so pq2+pq⊥2 = λ+(1−λ) = 1. (cid:3) Proposition 3.16. For p, q ∈ P(A), σ(pq⊥) \ {0, 1} = 1 − σ(pq) \ {0, 1} = σ(p⊥q) \ {0, 1}. Proof. From elementary spectral theory, we have σ(pq) = σ(ppq) = σ(pqp) and 1 − σ(pqp) \ {0, 1} = σ(p − pqp) \ {0, 1} = σ(pq⊥p) \ {0, 1}. (cid:3) Note that pq⊥ satisfies the triangle inequality, i.e. for p, q, r ∈ P(A), pr⊥ = p(q + q⊥)r⊥ ≤ pq⊥ + qr⊥. Also pq⊥ = 0 = qp⊥ ⇔ p = q so max(pq⊥,qp⊥) defines a metric on P(A). the next result shows that it in fact coincides with the usual metric on P(A), i.e. p − q. In particular, P(A) is complete in this metric. Proposition 3.17. For p, q ∈ P(A), p − q = max(pq⊥,p⊥q). In fact, p − q < 1 ⇒ pq⊥ = p − q = qp⊥. Proof. As p − q = pq⊥ + pq − pq − p⊥q = pq⊥ − p⊥q, and pq⊥(p⊥q)∗ = 0 = (pq⊥)∗p⊥q, we have p − q = max(pq⊥,p⊥q). So if pq⊥,p⊥q < 1 then, by Proposition 3.16, pq⊥2 = max(σ(pq⊥)) = max(σ(p⊥q)) = p⊥q2. (cid:3) The following result shows that the Sasaki projection (see [Kal83] §7 Lemma 13) defined by q takes any p to the range projection of qp. Proposition 3.18. For p, q ∈ P(B(H)), [qp] = (p ∨ q⊥) ∧ q 30 TRISTAN BICE Proof. As P(B(H)) is orthomodular, this is equivalent to p⊥ ∧ q = [qp]⊥ ∧ q. To see this note that, for any v ∈ R(p) and w ∈ R(q), (3.8) As w ∈ R(p⊥ ∧ q) if and only if the left expression is 0, for all v ∈ R(p), and w ∈ R([qp]⊥ ∧ q) if and only if the right expression is 0, for all v ∈ R(p), we are (cid:3) done. hv, wi = hv, qwi = hqv, wi. The previous results of this subsection, while rarely stated explicitly, are no doubt well known. The next result, however, might not be. It characterizes the spectrum of a product of projections purely in terms of the norm and the ortholattice structure of P(B(H)). Theorem 3.19. For p, q ∈ P(B(H)), (3.9) Proof. If θ ∈ σ(pq) \ {0} then, for any λ > θ, q(pqp)(0,λ]2 ∈ [θ, λ], while σ(pq) ∪ {0} = {qr2 : r ≤ p and ((r ∨ q⊥) ∧ q) ⊥ ((r⊥p ∨ q⊥) ∧ q)}. ((pqp)⊥p and ((pqp)(0,λ] ∨ q⊥) ∧ q = (qpq)(0,λ] (0,λ] ∨ q⊥) ∧ q = (pqp)(λ,1], by Proposition 3.9 (with a = qp) and Proposition 3.18 (noting that we have (pqp)⊥p (pqp)(λ,1] ∨ (p ∧ q⊥)). Conversely, say r ≤ p and (r ∨ q⊥) ∧ q = [qr] is orthogonal to (r⊥p ∨ q⊥) ∧ q = [qr⊥p ]. Thus [qr] ⊥ r⊥p (see (3.8)) so [pqr] = ([qr]⊥ ∧ p)⊥p ≤ r and this means (pqp)r = r(pqp)r = r(pqp). Thus r commutes with every spectral projection of pqp. Thus r ≤ (pqp)(0,qr2], otherwise s = r(pqp)(λ,1] 6= 0, for some λ > qr2, and q⊥s2 ≤ q⊥(pqp)(λ,1]2 ≤ 1 − λ so sqs ≥ λs, by Proposition 3.14, which gives qr2 ≥ qs2 = sqs ≥ λ > qr2, a contradiction. On the other hand, if θ < qr then r(pqp)(θ,1] 6= 0, otherwise we would have r ≤ (pqp)(0,θ] and hence qr2 ≤ q(pqp)(0,θ]2 ≤ θ < qr2, a contradiction. But then r(pqp)(θ,qr2] = r(pqp)(θ,1] 6= 0 so (θ,qr2] ∩ σ(pq) 6= ∅, for all θ < qr2, i.e. qr2 ∈ σ(pq). (cid:3) 3.5. Annihilator Separativity. The most fundamental difference between anni- hilators and projections is that, as shown in Example 3.87, the annihilators in an arbitrary C-algebra may not be orthomodular. Nonetheless, we can show that [A]⊥ is always separative, for arbitary C-algebra A, which will suffice to allow us to apply the theory in §2. In fact, we will prove a strong version of separativity that is quite close to orthomodularity. To see what this strong version is, note that with annihilators we can naturally quantify the degree of separativity. Specifically, for ǫ ∈ [0, 1], call [A]⊥ ǫ-separative if, for all B, C ∈ [A]⊥, (0,λ] = B $ C ⇒ ∃D ∈ [A]⊥({0} < D ⊆ C and BD ≤ ǫ). One immediately sees that if [A]⊥ is ǫ-separative, for any ǫ < 1, then it is separa- tive and, by Proposition 2.16, 0-separativity is equivalent to orthomodularity. So Theorem 3.26 really is as close as we can get to orthomodularity without actually verifying it. We can also use annihilators to separate arbitrary C-subalgebras B, C ⊆ A with B ⊆ C. Specifically, define the separation of B from C by D∈[C]\{{0}}BD. sepC (B) = inf ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 31 By the comments after Proposition 3.11, we could replace [C] with [C]C here with- out changing the value of sepC (B), so we might as well get rid of C, just assume we have a C-subalgebra B of A and write sep(B) for sepA(B). Let us also assume B = bAb, for b ∈ B1 +, and define xb = sup n≥0xb1/nx∗1/2 and γ(b) = inf x6=0xb/x, as in [AE02]. As xb1/nx∗1/2 = b1/(2n)x∗xb1/(2n)1/2 = √x∗xb1/n√x∗x1/2, we actually have γ(b) = inf a∈A1 + ab. And for any a ∈ A1 +, ab = sup n≥0ab1/n ≥ {f1−δ,1(a)}⊥⊥B − δ, and hence γ(b) ≥ sep(B). While for any C ∈ [A]⊥ and c ∈ C1 cb ≤ CB so γ(b) ≤ sep(B), i.e. +, we certainly have γ(b) = sep(B). It is immediate that sep(B) = 0 whenever B is not essential in A or, equivalently, when b is a zero-divisor. Whether it is still possible to have sep(B) < 1 when B is essential in A was an open problem (see [PZ00]), usually phrased as asking whether all open dense projections p (B is essential means pB is dense in that pB = 1) are regular 6 in the sense that ap = ap(= a when p is dense), for all a ∈ A. This was answered in the negative in [AE02] Proposition 3.4 where it was shown that γ(b) ≤ 4/5 for a particular non-zero divisor b. We improve on this in Theorem 3.25, showing that sep(B) is, in fact, always 1 or 0, i.e. open dense projections are always either regular or very non-regular. In order to prove these results, we first need the spectral projection inequalities contained in the lemmas below. Note that if c ∈ B(H)+ and p ∈ P(B(H)) then c ≤ p implies p⊥cp⊥ ≤ 0 and hence p⊥c = 0, i.e. c = pc. Lemma 3.20. For ǫ, λ > 0, there exists δ > 0 such that, whenever b, c ∈ B(H)1 +, c ≤ q ∈ P(B(H)) and bq2 ≤ λ + δ, we have (3.10) c[0,1−ǫ](cb2c)[λ−δ,1] ≤ ǫ. Proof. If ǫ ≥ 1 then (3.10) holds trivially, so assume ǫ < 1. For any v ∈ H, cv2 = cc[0,1−ǫ]v2 + cc(1−ǫ,1]v2 ≤ (1 − ǫ)2c[0,1−ǫ]v2 + c(1−ǫ,1]v2 ≤ v2 − ǫ(2 − ǫ)c[0,1−ǫ]v2 ≤ v2 − ǫc[0,1−ǫ]v2 (as ǫ ≤ 1). 6This concept was first introduced and investigated in [Tom60] Chapter 2 §2 under a somewhat different, but equivalent, definition. 32 TRISTAN BICE Now for v ∈ R((cb2c)[λ−δ,1]), (λ − δ)v2 ≤ hcb2cv, vi = bcv2 ≤ bq2cv2 ≤ (λ + δ)(v2 − ǫc[0,1−ǫ]v2), so (λ + δ)ǫc[0,1−ǫ]v2 ≤ 2δv2 and c[0,1−ǫ]v2 ≤ 2δv2/(λǫ), which immediately yields (3.10), for δ ≤ λǫ3/2. Corollary 3.21. For ǫ, λ > 0, there exists δ > 0 such that, whenever b, c ∈ B(H)1 +, c ≤ q ∈ P(B(H)) and bq2 ≤ λ + δ, we have (cid:3) (1 − c)(cb2c)[λ−δ,1] ≤ ǫ. Proof. Replacing ǫ with ǫ/√2 in Lemma 3.20, we obtain δ > 0 such that, for any v ∈ R((cb2c)[λ−δ,1]), (1 − c)v2 = (1 − c)c[0,1−ǫ/√2)v2 + (1 − c)c[1−ǫ/√2,1]v2 ≤ c[0,1−ǫ/√2)v2 + ǫ2c[1−ǫ/√2,1]v2/2 ≤ ǫ2v2/2 + ǫ2v2/2. (cid:3) The following result generalizes [Bic12] Lemma 5.3. Lemma 3.22. For ǫ, λ > 0, there exists δ > 0 such that, whenever b, c ∈ B(H)1 +, c ≤ q ∈ P(B(H)) and bq2 ≤ λ + δ, we have (3.11) b[0,√δ](cb2c)[λ−δ,1]2 ≤ 1 − λ + ǫ. Proof. Let δ > 0 be that obtained in Corollary 3.21 from replacing ǫ with ǫ/4. If necessary, replace δ with a smaller non-zero number so that we also have (1 − λ + δ + ǫ/2)/(1 − δ) ≤ 1 − λ + ǫ. Then, for all v ∈ R((cb2c)[λ−δ,1]), (λ − δ)v2 ≤ hcb2cv, vi = hb2cv, cvi ≤ hb2v, vi + ǫv2/2, by Corollary 3.21, = hb2b[0,√δ]v, vi + hb2b(√δ,1]v, vi + ǫv2/2 ≤ δhb[0,√δ]v, vi + hb(√δ,1]v, vi + ǫv2/2 = δb[0,√δ]v2 + (v2 − b[0,√δ]v2) + ǫv2/2, so (1 − δ)b[0,√δ]v2 ≤ (1 − λ + δ + ǫ/2)v2, and hence b[0,√δ]v2 ≤ (1 − λ + ǫ)v2. (cid:3) ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 33 With Lemma 3.22, we can already prove that [A]⊥ is separative (see the proof of Theorem 3.26, ignoring the last line). However, for ǫ-separativity, for arbitrary ǫ > 0, we need a couple more results. Lemma 3.23. For ǫ, λ > 0, there exists δ > 0 such that, whenever b, c ∈ B(H)1 +, p, q ∈ P(B(H)), b ≤ p, c ≤ q and pq2 ≤ λ + δ, we have (3.12) p(cb2c)[λ−δ,1]2 ≤ λ + ǫ. Proof. Let δ > 0 be that obtained in Corollary 3.21 with ǫ replaced with ǫ/4. If necessary, decrease δ so that δ ≤ ǫ/2. Then, for v ∈ R((cb2c)[λ−δ,1]), pv2 = hpv, pvi ≤ hpcv, pcvi + ǫv2/2 ≤ (pc2 + ǫ/2)v2 ≤ (λ + δ + ǫ/2)v2. ≤ (λ + ǫ)v2. (cid:3) Theorem 3.24. If we have ǫ > 0 and C-subalgebras B and C 6= {0} of A satisfying BC2 = λ < 1, then there exists D ∈ [A]⊥ with BD ≤ ǫ and CD2 ≥ 1 − λ − ǫ. Proof. Choose δ > 0 small enough that it satisfies Lemma 3.22 and Lemma 3.23 + and b ∈ B1 with ǫ replaced by some µ > 0, to be determined later. Take c ∈ C1 with bc2 > λ−δ/2. Let c′ = fλ−δ,λ−δ/2(cb2c) ∈ C and a = (1−fδ(b))c′2(1−fδ(b)), so + a = (1 − fδ(b))c′2 ≥ b{0}(cb2c)[λ−δ/2,1]2 ≥ 1 − [b](cb2c)[λ−δ,1]2, by (3.7) ≥ 1 − λ − µ, by (3.12). (3.13) In particular, a > 0 as long as µ < 1 − λ, and we may define a′ = a−1a. By (3.11), we have (3.14) b[0,√δ][c′]2 ≤ b[0,√δ](cb2c)[λ−δ,1]2 ≤ 1 − λ + µ and so, by Proposition 3.14, (3.15) [c′]b[δ,1][c′] ≥ [c′]b[√δ,1][c′] ≥ (λ − µ)[c′]. Now take b′ ∈ B1 (3.16) + with (1 − b′)fδ(b)) ≤ µ, and note that, as BC2 = λ, (3.17) c′b′c′ ≤ c′[c′][b′][c′]c′ ≤ λc′2. = b′ab′ = b′(1 − fδ(b))c′2(1 − fδ(b))b′ ≤ (b′ − b′fδ(b)b′)c′2 + 2µ, by (3.16), = c′(b′ − b′fδ(b)b′)2c′ + 2µ ≤ c′(b′ − b′fδ(b)b′)c′ + 2µ ≤ c′(λ − b′b[δ,1]b′)c′ + 2µ, by (3.17), ≤ c′(λ − [c′]b[δ,1][c′])c′ + 4µ, again by (3.16), ≤ c′(λ − λ + µ)c′ + 4µ, by (3.15) (and BC2 = λ) ≤ 5µ, and hence, b′a′2 ≤ b′√a′2 ≤ 5µ/(1 − λ − µ). 34 TRISTAN BICE Putting all this together, we have (1 − λ − µ)b′a′b′ ≤ ab′a′b′, by (3.13) Thus this inequality holds for all b′ in an approximate unit for B, and hence for all b′ ∈ B1 +, we have d − a′d ≤ 2µ and hence, for any b′ ∈ B1 +, +. If we let D = {f1−2µ,1(a′)}⊥⊥ ∈ [A]⊥ then, for any d ∈ D1 b′d2 ≤ b′a′d2 + 4µ ≤ b′a′2 + 4µ ≤ 5µ/(1 − λ − µ) + 4µ. Thus, so long as we choose µ > 0 at the start sufficiently small, this immediately gives BD ≤ ǫ. Also, [1 − fδ(b)]c′2 ≤ b[0,√δ][c′]2 ≤ 1 − λ + µ, by (3.14), so as long as we chose µ at least half as small as the δ obtained in Lemma 3.22 (from the given ǫ), we can also apply (3.11) with c and b replaced by 1 − fδ(b) and c′ respectively to get c′[0,√µ]a′[1−µ,1]2 ≤ c′[0,√2µ]((1 − fδ(b))c′2(1 − fδ(b)))1−λ−2µ2 ≤ λ + ǫ, and hence CD2 ≥ f√µ(c′)f1−2µ,1−µ(a′)2 ≥ c′[√µ,1]a′[1−µ,1]2 ≥ 1 − λ − ǫ. (cid:3) Note that Theorem 3.24 is a modulo-ǫ generalization of (3.7) to annihilators. Essentially just rephrasing the first part also immediately yields the following. Theorem 3.25. If B is a C-subalgebra of A with sep(B) < 1 then sep(B) = 0. With a little more work, Theorem 3.24 also gives us ǫ-separativity. Theorem 3.26. [A]⊥ is ǫ-separative, for all ǫ > 0. Proof. Take B, C ∈ [A]⊥ with B $ C, so we have c ∈ C1 + \ B. This means we have b ∈ B⊥+ with b = 1 and bc 6= 0, and hence b[c] 6= 0. Set q = [c], λ = bq2, take positive ǫ < λ and let δ > 0 be that obtained in Lemma 3.22. Note that we may now assume that bc2 > λ − δ by replacing c with fµ(c) for sufficiently small µ. Take s, s′ ∈ (λ − δ,bc2) with s < s′ and set D = {fs,s′(cb2c)}⊥⊥. Then we see that pD ≤ (cb2c)[s,1] ≤ (cb2c)[λ−δ,1] and pB ≤ b{0} ≤ b[0,√δ] so, by Lemma 3.22, BD2 ≤ b[0,√δ](cb2c)[λ−δ,1]2 ≤ 1 − λ + ǫ < 1. Now simply apply Theorem 3.24 to get another D ∈ [A]⊥ with BD ≤ ǫ. Corollary 3.27. If A has property (SP) then [A]⊥ ∼= [P(A)]⊥. (cid:3) ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 35 Proof. Property (SP) means that every hereditary C-subalgebra of A contains a non-zero projection (see [Bla94] Definition 6.1.1) which, by the comments af- ter Proposition 3.11, is equivalent to saying every annihilator contains a non-zero projection. This, in turn, is equivalent to saying P(A) is order-dense in A in the induced preorder. As [A]⊥ and hence A is separative, this is equivalent to saying P(A) is join-dense in A, by Proposition 2.8. Now [A]⊥ ∼= [P(A)]⊥, by (2.5). (cid:3) Note that [P(A)]⊥ is just the canonical completion by cuts of the orthomodular partial order P(A). However, this does not necessarily mean that [A]⊥ is orthomod- ular for A with property (SP), as orthomodularity is not necessarily preserved by completions. In fact, an example is even given in [Ada69] of a modular lattice such that its completion is not orthomodular, and Example 3.87, with X replaced by the Cantor space, even yields a real rank zero A such that [A]⊥ is not orthomodular. Replacing projections above with any essential ideal (for any C-algebra now), we get the same result (and more can be said in this case -- see Proposition 3.48). Proposition 3.28. If B contains an essential ideal in A then [A]⊥ ∼= [B]⊥. Proof. Take a ∈ A+. If B is an essential ideal, there exists b ∈ B+ with 0 6= ba ∈ B and hence {ba}⊥⊥ ⊆ {a}⊥⊥, i.e. ba is below a in the induced preorder. As a was arbitrary, B is order-dense in A in the induced preorder. The result now follows as (cid:3) in the proof of Corollary 3.27. So internally, essential ideals are the same as the C-algebra itself, as far as the annihilators are concerned. The following shows that the same is true externally. Proposition 3.29. If C is an essential ideal in a hereditary C-subalgebra B of A then B⊥ = C⊥. Proof. For a ∈ C⊥+ , b ∈ B+ and c ∈ C+ we have bc ∈ C and hence babc = 0. As c ∈ C+ was arbitrary, bab ∈ C⊥ ∩ B = {0} and hence ab = 0. As b ∈ B+ was arbitrary, a ∈ B⊥, i.e. C⊥ ⊆ B⊥, while the reverse inclusion is immediate. (cid:3) Note that if Proposition 3.29 were true for all hereditary C-subalgebras of a particular C-algebra A, not just ideals, then [A]⊥ would have to be orthomodular. 3.6. Annihilator Ideals. Now that we know the annihilators in an arbitrary C- algebra are separative, we immediately have the type decompositions given in §2.11. The only question that remains is whether these types can also be characterized algebraically in the same way as the projections appearing in the classical type decomposition of a von Neumann algebra. Our first task is to identify the central annihilators. Theorem 3.30. B ∈ [A]⊥ is central if and only if B is an ideal. Proof. We first show that, whenever B, D ∈ [A]⊥ and B is an ideal, (D ∩ B)⊥ ∩ (D ∩ B⊥)⊥ ⊆ D⊥. To see this, take any e /∈ D⊥, so there is d ∈ D such that ed 6= 0. But then there must also be b ∈ B or b ∈ B⊥ such that edb 6= 0 and hence edbd 6= 0. But then dbd ∈ D ∩ B and hence e /∈ (D ∩ B)⊥, or dbd ∈ D ∩ B⊥ and hence e /∈ (D ∩ B⊥)⊥. So e /∈ (D ∩ B)⊥ ∩ (D ∩ B⊥)⊥ and the inclusion is proved. The reverse inclusion is immediate and thus B ∈ c[A]⊥, by Theorem 2.20. 36 TRISTAN BICE + and b ∈ B1 Conversely, if B is not an ideal then we have a ∈ A1 + with ab /∈ B. As ba2b ∈ B we must have ab2a /∈ B (because B is hereditary so x ∈ B ⇔ x∗x ∈ B and xx∗ ∈ B -- see [Ped79] Theorem 1.5.2). If cc∗ ≤ λab2a for some λ ≥ 0 then c = abd, for some d ∈ B(H), by [Dou66] (alternatively, one could obtain a similar factorization in A with [Ped79] Proposition 1.4.5). Then bc = babd = 0 ⇔ c = abd = 0, by Lemma 3.10, and hence c /∈ B⊥ as long as c 6= 0. So, if we set C = {fδ(ab2a)}⊥⊥ for δ > 0 sufficiently small that fδ(ab2a) /∈ B, then C is not contained in B and cc∗ ≤ 2c2δ−1ab2a, for all c ∈ C, and hence C ∩ B⊥ = {0}. So we have (B ∧ C) ∨ (B⊥ ∧ C) = B ∩ C $ C and hence B is not central. (cid:3) In particular, B is central (in [A]⊥) if and only if pB is (in P(A′′)), by Proposition 3.7. This is perhaps a little surprising, considering that commutativity itself is not the same in [A]⊥ and P(A′′), as shown by the examples in §3.11. Another important thing to note is the difference between central covers in [A]⊥ and the central covers defined in [Ped79] 2.6.2. There, the central cover c(a) of an element a ∈ A′′sa is defined to be the smallest b ∈ (A′ ∩ A′′)sa with a ≤ b. If p ∈ P(A′′), this means that c(p) = cP(A′′)(p), i.e. central covers in this algebraic sense are the same for projections as central covers in the order theoretic sense with respect to P(A′′) (not P(A′′)◦). In general we have c(pB) ≤ pc(B), for B ∈ [A]⊥, but this inequality can be strict. We will primarily be concerned with central covers in [A]⊥ rather than A′′sa, although we would be remiss not to point out the following connections. Proposition 3.31. If B is a C-subalgebra of A then pc(B) = c(pB)◦. Proof. As c(B) is an ideal, pc(B) is a central projection and B ⊆ c(B) so pB ≤ pc(B) and hence c(pB) ≤ pc(B) which, in turn, gives c(pB)◦ ≤ p◦c(B) = pc(B). For the reverse inequality, first note that c(pB)◦ is central, by Proposition 3.8. So c(pB)◦Ac(pB)◦ ∩ A is an annihilator ideal, by Proposition 3.7, which certainly contains B and hence also c(B). Thus pc(B) ≤ c(pB)◦. (cid:3) Even if pc(B) 6= c(pB), the analogous notion of 'very orthogonal' is equivalent. Corollary 3.32. For B, C ∈ [A]⊥, the following are equivalent. (1) B and C are very orthogonal. (2) c(pB)c(pC ) = 0 (3) BAC = {0}. Proof. (1)⇒(2) As c(pB) ≤ pc(B), c(pB)c(pC ) ≤ pc(B)pc(C) ≤ c(B)c(C) = 0. (2)⇒(1) If c(pB)c(pC ) = 0 then c(pB) ≤ c(pC )⊥ so pB = p◦B ≤ c(pB)◦ ≤ c(pC )⊥◦. By Proposition 3.8, c(pC )⊥◦ is central. Also pC ≤ c(pC )◦, and this latter projection is central, again by Proposition 3.8. Thus we must in fact have c(pC )◦ = c(pC ), i.e. c(pC ) is open so c(pC )⊥ is closed and c(pC)⊥◦ is topologically regular, by (3.2). So we have pc(B) ≤ c(pC)⊥◦ and hence pc(B)c(pC) = 0. Now pc(B)pc(C) = 0 follows, by the same argument applied to pC, and thus c(B) ∩ c(C) = {0}. (2)⇒(3) For a ∈ A, b ∈ B and c ∈ C, bac = bc(pB)ac(pC )c = c(pB)c(pC )bac = 0. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 37 (3)⇒(2) If BAC = {0} then if I = ABA is the ideal generated by B we have IC = {0} and hence IC = {0}, so c(pB)pC ≤ pI pC = 0. But c(pB)pC = 0 means pC ≤ c(pB)⊥ and hence c(pC ) ≤ c(pB)⊥ so c(pB)c(pC ) = 0. (cid:3) The next result, applied to B ∈ [A]⊥, shows [A]⊥ has the relative centre property. Corollary 3.33. If B is a hereditary C-subalgebra of A then cB[B]⊥ = {B ∩ C : C ∈ c[A]⊥}. Proof. For one inclusion, take C ∈ c[A]⊥ and b ∈ (B+ ∩ C⊥)⊥B . Then, for any c ∈ C⊥+ , we have √bc ∈ B ∩ C⊥+ and hence bc = 0, i.e. b ∈ C⊥⊥ = C. As C is an ideal in A, B ∩ C is an ideal in B, and thus B ∩ C = (B ∩ C⊥)⊥B ∈ c[B]⊥. For the reverse inclusion, say we have C ∈ cB[B]⊥ which, by Corollary 3.32 (3), means that CBC⊥B = {0}. But, for any c ∈ C+, d ∈ C⊥B cad = √c(√ca√d)√d ∈ CBC⊥B = {0}. + and a ∈ A+, we have Thus C and C⊥B are very orthogonal (in [A]⊥) so c(C) ∩ B = C. We also have something of an algebraic substitute for Corollary 2.23. Proposition 3.34. Given hereditary C-subalgebras (Bα) with c(pBα )c(pBβ ) = 0, Proof. A C-subalgebra B is hereditary if and only if bab ∈ B for all a ∈ A and whenever α 6= β, the C-subalagebraL Bα they generate is also hereditary. b ∈ B (see [Bla06] II.5.3.9). But given a ∈ A andP bα ∈L Bα we have (X bα)a(X bα) = (X bαc(pBα))a(X c(pBα)bα) by centrality and orthogonality, = X bαc(pBα )ac(pBα)bα, = X bαabα ∈M Bα, by hereditarity of the (Bα). (cid:3) (cid:3) 3.7. Equivalence. Here we study the basic properties of the following relation, which we believe to the be the natural analog for annihilators in C-algebras of Murray-von Neumann equivalence. Definition 3.35. B, C ∈ [A]⊥ are equivalent, written B ∼ C, if {a}⊥⊥ = B and {a∗}⊥⊥ = C, for some a ∈ A. We write B - C if B ∼ D ⊆ C, for some D ∈ [A]⊥. For p, q ∈ P(A), let us write p ∼MvN q for the usual Murray-von Neumann if there is a (partial isometry) u ∈ A with u∗u = p and equivalence notion, i.e. uu∗ = q. Likewise, p -MvN q means p ∼MvN r ≤ q for some r ∈ P(A). Also, we shall say a A has polar decomposition if, for all a ∈ A, there is a partial isometry u ∈ A with a = ua, where a = √a∗a. Von Neumann algebras certainly have polar decomposition, but so too do AW-algebras, Rickart C-algebras and all C- algebra quotients of these (e.g. the Calkin algebra). The following result shows that, for any such C-algebra, the relation ∼ defined above really is a completely consistent extension of Murray-von Neumann equivalence. Proposition 3.36. If A has polar decomposition then, for p, q ∈ P(A), p ∼MvN q ⇔ pAp ∼ qAq p -MvN q ⇔ pAp - qAq. and 38 TRISTAN BICE If p ∼MvN q then the partial Proof. It suffices to prove the first equivalence. isometry witnessing this will also witness pAp ∼ qAq. Conversely, say we have a ∈ A with {a}⊥⊥ = pAp and {a∗}⊥⊥ = qAq and take a partial isometry u ∈ A with a = ua. Then we immediately see that the left annihilator of u is contained in the left annihilator of a, i.e. {u∗}⊥ ⊆ {a∗}⊥ and hence qAq = {a∗}⊥⊥ ⊆ {u∗}⊥⊥ = uAu∗, which gives uu∗ ≥ q. By replacing u with qu if necessary we may assume that uu∗ = q and hence {u∗}⊥⊥ = {a∗}⊥⊥. We also have a2 = a∗a = au∗ua and hence a(1 − u∗u)a = 0, which means (1 − u∗u)a = 0, i.e. a = u∗ua = u∗a = a∗u. This immediately gives {a}⊥⊥ ⊆ {u}⊥⊥, while if 0 = ab then 0 = √aa∗ab = a√a∗ab = aab = aa∗ub which, as {a∗}⊥ = {u∗}⊥, gives 0 = u∗ub, i.e. b ∈ {u}⊥. As b ∈ {a}⊥ was arbitrary, {a}⊥ ⊆ {u}⊥ and hence {u}⊥⊥ ⊆ {a}⊥⊥, i.e. u∗Au = {u}⊥⊥ = {a}⊥⊥ = pAp so u∗u = p too. (cid:3) Lemma 3.37. If a, b ∈ A and {b∗}⊥⊥ ⊆ {a}⊥⊥ then {ab}⊥⊥ = {b}⊥⊥. Proof. If c ∈ A+ and abc = 0 then abcb∗ = 0 so bcb∗ ∈ {a}⊥ ∩ {b∗}⊥⊥ ⊆ {a}⊥ ∩ {a}⊥⊥ = {0}. (cid:3) Thus {ab}⊥ ⊆ {b}⊥, while {b}⊥ ⊆ {ab}⊥ is immediate, so {ab}⊥ = {b}⊥. Corollary 3.38. ∼ and - are transitive relations. Proof. If a witnesses B - C and b witnesses C - D, then ab witnesses B - D, by Lemma 3.37. If - was actually ∼ here then one more application of Lemma 3.37 shows that ab witnesses B % D and hence, as - and % are witnessed by the same element ab of A, B ∼ D. (cid:3) Definition 3.39. A is anniseparable if B ∈ [A]⊥ ⇒ B = S⊥ for countable S ⊆ A. We call B ∈ [A]⊥ principal if there exists b ∈ B with {b}⊥⊥ = B. We say A is orthoseparable if every pairwise orthogonal subset of A+ is countable. We can see immediately that anniseparability ⇔ every annihilator is principal ⇔ ∼ is reflexive. + such that (sn)⊥⊥ = B, we can simply let b =P 2−nsn to get For, given (sn) ⊆ A1 {b}⊥⊥. It then immediately follows from Corollary 3.38 that ∼ is an equivalence relation and - is a preorder. Likewise, if we have orthogonal (Bn) ⊆ [A]⊥, orthogonal (Cn) ⊆ [A]⊥ and (an) ⊆ A witnesses their equivalence, i.e. {an}⊥⊥ = Bn and {a∗n}⊥⊥ = Cn, for all i.e. ∼ is countably additive. This yields the second implicaiton in n ∈ N, then a =P 2−nan witnesses the equivalence of B =W Bn and C =W Cn, separability ⇒ orthoseparability ⇒ additivity. Most C-algebras one encounters are anniseparable. Proposition 3.40. Each of the following conditions implies A is anniseparable. (1) A is an AW-algebra (e.g. a von Neumann algebra). (2) A is (norm) separable. (3) A is orthoseparable and [A]⊥ is orthomodular. Proof. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 39 (1) An AW-algebra is, by definition, a Baer -ring (see [Ber72] §4 Definition 1 and 2, and also Proposition 1 to see how this is equivalent to other definitions sometimes given for AW-algebras, like the one in [Ped79] 3.9.2), meaning that, for every S ⊆ A, R(S) = pA for some p ∈ P(A), and hence S⊥ = pAp = {p}⊥⊥. To see that von Neumann algebras are AW-algebras, simply note that, for any S ⊆ A, R(S) is a weakly closed right ideal and hence of the form pA, for some p ∈ P(A), by [Ped79] Proposition 2.5.4. (2) For any B ∈ [A]⊥, B = S⊥⊥, where S is any countable dense subset of B. (3) For any B ∈ [A]⊥, let S be a maximal pairwise orthogonal subset of B+. If we had S⊥⊥ < B then, by orthomodularity, we would have b ∈ B+ ∩ S⊥, contradicting the maximality of S. (cid:3) Proposition 3.41. If B, C ∈ [A]⊥ are principal and semiorthoperspective, B ∼ C. Proof. Take b, c ∈ A+ with {b}⊥⊥ = B and {c}⊥⊥ = C. Given d ∈ A+ with bcd = 0 we have cdc ∈ C ∩ B⊥ = {0} and hence cd = 0, i.e. {bc}⊥ = {c}⊥ and hence {bc}⊥⊥ = {c}⊥⊥ = C. A symmetric argument gives {cb}⊥⊥ = {b}⊥⊥ = B and hence bc witnesses B ∼ C. (cid:3) Corollary 3.42. If A is anniseparable, ∼ is weaker than perspectivity. Corollary 3.43. If A is anniseparable and B ∈ [A]⊥ then ∼B is ∼ on [B]B. Proof. If {b}⊥⊥,{b∗}⊥⊥ ∈ [B]B, for some b ∈ A (necessarily with b ∈ B), then {b}⊥B⊥B = {b}⊥⊥⊥B⊥B = {b}⊥⊥ and, likewise, {b∗}⊥B⊥B = {b∗}⊥⊥. Thus ∼ restricted to [B]B is stronger than ∼B. Conversely, if [A]⊥ is anniseparable then, as ∼ is weaker than perspectivity, which is itself weaker than orthoperspectivity, we have {b}⊥B⊥B ∼ {b}⊥⊥ ∼ {b∗}⊥⊥ ∼ {b∗}⊥B⊥B , for any b ∈ B. (cid:3) From Corollary 3.42, we see that the only thing stopping ∼ from being a dimen- sional equivalence relation (for anniseparable A) according to [Kal83] §11 Definition 1, is the potential lack of finite (orthogonal) divisibility. Although we can prove that ∼ satisfies a non-orthogonal version of divisibility. Proposition 3.44. For any a ∈ A, the map B 7→ (Ba)⊥⊥ is order and supremum preserving on [A]⊥. Also (Ba)⊥⊥ ∼ B, for all princpal B ∈ [{a∗}⊥⊥]. Proof. The given map is certainly order preserving, even for arbitrary subsets B of A. If we have B ⊆ [A]⊥ then _B∈B (Ba)⊥⊥ = (\B∈B (Ba)⊥)⊥ = (([B)a)⊥⊥ ⊆ ((_ B)a)⊥⊥. To see that this last inclusion can be reversed, take c ∈ ((SB)a)⊥+. This means aca∗ ∈ (SB)⊥ = (WB)⊥ and hence c ∈ ((WB)a)⊥, i.e. ((S B)a)⊥ ⊆ ((W B)a)⊥ and hence ((W B)a)⊥⊥ ⊆ ((SB)a)⊥⊥. Lastly, note that if B = {b}⊥⊥ ⊆ {a∗}⊥⊥, for some b ∈ B+, then {a∗b}⊥⊥ = B, by Lemma 3.37, and c ∈ {ba}⊥ ⇔ aca∗ ∈ {b}⊥ = B⊥ ⇔ c ∈ (Ba)⊥, i.e. {ba}⊥⊥ = (Ba)⊥⊥ and hence ba witnesses B ∼ (Ba)⊥⊥. (cid:3) 40 TRISTAN BICE 1 Note, however, that the map B 7→ (Ba)⊥⊥ may not preserve infimums, as the following example shows. Specifically, consider A = B(H), where H is a separable infinite dimensional Hilbert space with basis (en). Define a ∈ B(H)+ by aen = n2 en, so R(a) = H and hence {a}⊥⊥ = A, and let c be the projection onto Cv, where v = P 1 n en. Then b = c⊥ ⊥ c, and hence B = bAb = {b}⊥⊥ ⊥ {c}⊥⊥ = cAc = C even though we still have R(ab) = H and hence (Ba)⊥⊥ = {ba}⊥⊥ = {a}⊥⊥ = A, while (Ca)⊥⊥ 6= {0}. Proposition 3.45. If A is anniseparable and orthoseparable, ∼ is a type relation. Proof. Orthoseparability yields the ⇒ part of (2.26). For the ⇐ part, say we have B, C ∈ [A]⊥, B ∼ C and D ∈ c[A]⊥. By Proposition 3.44, we have E, F ∈ [C] with B ∩ D ∼ E, B⊥ ∩ D ∼ F and E ∨ F = C. As the closed ideal generated by any a ∈ A is the same as that generated by a∗, we have c(E) = c(B ∩ D) ⊆ B and c(F ) = c(B⊥ ∩ D) ⊆ B⊥. This means E = B ∩ C and F = B⊥ ∩ C. (cid:3) In particular, we get the type-decompositions in Corollary 2.26 coming from the type ideals of ∼-finite and ∼-orthofinite annihilators, by Proposition 2.31. Whether these agree with the type decompositions you get from the type ideals of modular or relatively modular annihilators (see Proposition 2.34 and (2.34)), we do not know. However, it does follow from Theorem 2.32 that if ∼ is finite on [A]⊥ then [A]⊥ is ∼⊥Fin is the type ideal of ∼-orthofinite annihilators in [A]⊥ then modular. So if [A]⊥ ∼⊥Fin ∩ [A]⊥Orel (see (2.35)) is the type ideal consisting of those B ∈ [A]⊥ for [A]⊥ which ∼ is finite on [B]B and [A]⊥ ∼⊥Fin ∩ [A]⊥Orel ⊆ [A]⊥Mrel. Proposition 3.44 also yields a Cantor-Schroeder-Bernstein theorem. Theorem 3.46. If A is anniseparable and B, C ∈ [A]⊥ then B - C - B ⇒ B ∼ C. Proof. Take b, c ∈ A with {b}⊥⊥ = B, {b∗}⊥⊥ ⊆ C, {c}⊥⊥ = C and {c∗}⊥⊥ ⊆ B. We will apply Tarski's fixed point theorem, as in [Ber72] §1 Theorem 1. Specifically, note that the map on [B] defined by is order-preserving so, as [B] is a complete lattice, it has a fixed point F . By Proposition 3.44, D 7→ ((Db∗)⊥C c∗)⊥B F ⊥B = ((F b∗)⊥C c∗)⊥B⊥B ∼ (F b∗)⊥C , and F ∼ (F b∗)⊥C⊥C , so B ∼ F ∨ F ⊥B ∼ (F b∗)⊥C⊥C ∨ (F b∗)⊥C ∼ C. (cid:3) The next result shows - satisfies generalized comparison (see Definition 2.36). Theorem 3.47. If A is anniseparable and orthoseparable then, for all B, C ∈ [A]⊥, there exists D ∈ c[A]⊥ with (3.18) and B ∩ D - C ∩ D C ∩ D⊥ - B ∩ D⊥. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 41 Proof. Let (an) be a maximal subset of A1 such that, for all distinct m, n ∈ N, and ama∗n = 0 = a∗man. a∗nan ∈ B, ana∗n ∈ C, Let a =Pn 2−nan, E = {a}⊥B⊥B ⊆ B and F = {a∗}⊥C⊥C ⊆ C. By Proposition 2.10 and Corollary 3.42, E ∼ {a}⊥⊥ ∼ {a∗}⊥⊥ ∼ F . By maximality, (B ∩ E⊥)A(C ∩ F ⊥) = {0} and (3.18) now follows from Proposition 2.37 and Corollary 3.32. (cid:3) Next we show that ∼ is the same in any sufficienlty large hereditary C-subalgebra B of A. By Proposition 3.28, it applies to any B containing an essential ideal of A. Proposition 3.48. If B is an anniseparable hereditary C-subalgebra of A and C 7→ B ∩ C is an injective map from [A]⊥ to [B]⊥ then, for D, E ∈ [A]⊥, D ∼ E ⇔ B ∩ D ∼B B ∩ E {b}⊥B⊥B = {b}⊥⊥ ∩ B. Proof. For any b ∈ B, B ∩ {b}⊥⊥ is the smallest element of {B ∩ C : C ∈ [A]⊥} containing b∗b. As C 7→ B ∩ C maps [A]⊥ to [B]⊥, [B]⊥ = {B ∩ C : C ∈ [A]⊥} so (3.19) So if {b}⊥B⊥B = B ∩ D and {b∗}⊥B⊥B = B ∩ E then {b}⊥⊥ = D and {b∗}⊥⊥ = E, by the injectivity of C 7→ B ∩ C, i.e. B ∩ D ∼B B ∩ E implies D ∼ E. On the other hand, as B is anniseparable, we have d ∈ B ∩ D+ such that B ∩ D = {d}⊥B⊥B = B ∩ {d}⊥⊥ and hence D = {d}⊥⊥, as C 7→ B ∩ C is injective. Likewise, we have e ∈ B ∩ E+ such that {e}⊥⊥ = E. If a ∈ A witnesses D ∼ E then so does dae, by Lemma 3.37. But B is a hereditary C-subalgebra so dae ∈ B and hence witnesses B ∩ D ∼B B ∩ E, again by (3.19). (cid:3) In particular, under the hypothesis of Proposition 3.48, A is also anniseparable. However, the converse fails in general. For example, if H is a non-separable Hilbert space then K(H) is not anniseparable, as K(H) is a non-principal annihilator in itself, even though K(H) is an essential ideal in the (necessarily anniseparable) von Neumann algebra B(H). In fact, if B is an essential ideal in anniseparable A then B is principal (in itself) ⇔ B is anniseparable. The ⇐ part is immediate, and for the ⇒ part simply note that if {b}⊥B⊥B = B, for some b ∈ B, then {b}⊥⊥ = A and, for any c ∈ A, we have bc ∈ B and {bc}⊥B⊥B = {c}⊥⊥ ∩ B, by (3.19). 3.8. Abelian Annihilators. Theorem 3.49. B ∈ [A]⊥ is commutative if and only if [B] is a Boolean algebra. Proof. First note that A is commutative if and only if every hereditary C-subalgebra of A is an ideal. For if A is commutative then so is A′′ and, in particular, P(A′′)◦, so all hereditary C-subalgebras of A are ideals. While if every hereditary C- subalgebra of A is an ideal then every open projection in A′′ is central. As P(A′′)◦ contains all spectral projections of elements of A+ corresponding to open subsets of R, it follows that span(P(A′′)◦) is dense in A and hence A is commutative. [B]B = cB[B]B is a Boolean algebra, so we only have to show that So if B is commutative then, by the previous paragraph and Theorem 3.30, (3.20) [B]B = [B]. 42 TRISTAN BICE + Given C ∈ [B], take b ∈ C⊥B⊥B . Then babc = bacb = 0, for all a ∈ C⊥+ and c ∈ C, so bab ∈ B ∩ C⊥ = C⊥B , but bab ∈ C⊥B⊥B too so ba = 0 and hence b ∈ C⊥⊥ = C, i.e. C⊥B⊥B = C and hence C ∈ [B]B. Conversely, if B is not commutative then, by the first paragraph and the other direction of Theorem 3.30, [B]B is not a Boolean algebra. It may be that [B] strictly contains [B]B, but infimums still agree in both structures, as do supremums when one of the annihilators is {0}, so the counterexample to distributivity in the proof (cid:3) of Theorem 3.30 still works in the possibly larger structure [B]. Definition 3.50. If B ∈ [A]⊥ is commutative it will be called abelian. If [A]⊥ con- tains no non-zero abelian annihilator ideals, A will be called properly non-abelian. If c(B) = A for some abelian B ∈ [A]⊥, we call A discrete. If A contains no non-zero abelian annihilators it will be called continuous. These definitions are consistent with classical von Neumann (or AW-)algebra It follows from Theorem 3.49 (and terminology (see [Ber72] §15 Definition 3). (3.20)) that the collection of abelian annihilators coincides with the type ideal [A]⊥D = [A]⊥Drel (see (2.31), (2.32) and (2.33) for this notation). It then follows from Theorem 3.47 Theorem 2.38 that when B and C are abelian annihilators in anniseparable or- thoseparable A, B ∼ C ⇔ c(B) = c(C). Also (2.23) gives us a decomposition of A into central abelian and properly non- abelian parts, while (2.24) gives us a decomposition into central discrete and countin- uous parts. We could actually obtain these decompositions in a more algebraic, rather than order theoretic way, using Proposition 3.34 and the following result. Theorem 3.51. If B is a commutative hereditary C-subalgebra then so is B⊥⊥. Proof. We first claim that every element of B commutes with every element of + and a ∈ B⊥⊥+ such that ab 6= ba. Then, for B⊥⊥. If not, we would have b ∈ B1 some ǫ > 0, we must have ab[ǫ,1] 6= b[ǫ,1]a and hence b[ǫ,1]a(1− b[ǫ,1]) = b[ǫ,1]ab[0,ǫ) 6= 6= 0 0. Thus, for some δ < ǫ sufficiently close to ǫ, we must have b[ǫ,1]ab[0,δ] and hence f (b)ag(b) 6= 0 where f = f(ǫ+δ)/2,ǫ and g = 1 − fδ,(ǫ+δ)/2. If we had g(b)af (b)2ag(b) ∈ B⊥ then, as a ∈ B⊥⊥, f (b)ag(b)a = 0 and hence f (b)ag(b) = 0, a contradiction. Thus f (b)ag(b)c 6= 0 for some c ∈ B. As B is hereditary and both f (b) and c are in B, this means that d = f (b)ag(b)c ∈ B and, likewise d∗ ∈ B. However, dd∗ ≤ λf (b)2 for some λ > 0 while d∗d ≤ λ′g(b)2 (note that b, c ∈ B so c commutes with b and hence with g(b) ∈ B + C1) for some λ′ > 0. As f (b)g(b) = 0, this means that d and d∗ do not commute, contradicting the fact B is commutative. Now the claim is proved, take any e, f ∈ B⊥⊥+ . Given any b ∈ B+, note that b(ef − f e) = b1/4eb1/2f b1/4 − b1/4f b1/2eb1/4 = 0, as b1/4eb1/4, b1/4f b1/4 ∈ B. Thus ef − f e ∈ B⊥ ∩ B⊥⊥ = {0} and hence, as e and f were arbitrary, B⊥⊥ is (cid:3) commutative. It now follows that the continuous C-algebras are, in fact, already well-known. Corollary 3.52. A is continuous if and only if it is antiliminal. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 43 Proof. A C-algebra is antiliminal if and only if it contains no commutative heredi- tary C-subalgebras (either by definition, as in [Ped79] 6.1.1, or by a theorem, as in [Li92] Propositions 13.3.11 and 13.3.12). So if A is antiliminal then it certainly will not contain any abelian annihilators, while conversely if A contains a commutative hereditary C-algebra B then B⊥⊥ is an abelian annihilator, by Theorem 3.51, and (cid:3) so A can not be continuous. And we now have our first simple application of decomposition. Corollary 3.53. If A is postliminal then it is discrete. Proof. Every C-subalgebra of a postliminal algebra is postliminal (see [Ped79] Proposition 6.2.9 or [Li92] Proposition 13.3.5) and, in particular, not antiliminal. Thus the continuous part of A must be {0} and A must be discrete. (cid:3) It should be mentioned that there is already a well-known postliminal-antiliminal decomposition theorem, which says that every C-algebra has a unique postliminal ideal I such that A/I is antiliminal (see [Ped79] Proposition 6.2.7 or [Li92] Propo- sition 13.3.6). The continous/antiliminal part ANI obtained from (2.24) (using the type ideal of abelian annihilators) is quite different, being a subalgebra rather than a quotient, and not just any subalgebra either, an annihilator ideal. Although it does naturally give rise to an antiliminal quotient. To see how, note that, as the relation ab = 0, for a, b ∈ A1 +, is liftable (see [Lor97] Proposition 10.1.10), whenever B ∈ c[A]⊥ and π : A → A/B⊥ is the canonical homomorphism, π ↾ B is injective and π(B) is an essential ideal in A/B⊥. Thus A/AI (where AI denotes the discrete part of A) contains π(ANI) as an essential antiliminal ideal. If A/AI contained a non-zero commutative hereditary C-subalgebra B then B ∩ π(ANI) would be a non-zero commutative hereditary C-subalgebra of π(ANI), a contradiction, so A/AI must also be antiliminal. As noted in §1.1, any infinite dimensional type I von Neumann factor will not be postliminal, so the converse of Corollary 3.53 fails in general. However, there is the possibility that they could be equivalent in the separable case. If this were true then it would show that the discrete-continuous annihilator decomposition really is stronger than the classical postliminal-antiliminal decomposition in the separable case, because you could obtain the latter as a corollary of the former by the comments in the previous paragraph. Here are some more consequences of Theorem 3.51 that will be important in the next subsection. Corollary 3.54. If B ∈ [A]⊥D, C ∈ [A]⊥ and B ∼ C then C ∈ [A]⊥D. Proof. Assume {a}⊥⊥ = B and {a∗}⊥⊥ = C. As B is commutative, we have ba∗ad = da∗ab and hence aba∗ada∗ = ada∗aba∗, for all b, d ∈ B. As B is hereditary, so is aBa∗ (because aBa∗AaBa∗ ⊆ aBa∗) and C = {a∗}⊥⊥ = {aa∗aa∗}⊥⊥ ⊆ aBa∗⊥⊥ ⊆ {a∗}⊥⊥ = C so, by Theorem 3.51, C is also commutative. (cid:3) Corollary 3.55. If A is discrete, [A]⊥D is order-dense in [A]⊥. Proof. Take B ∈ [A]⊥D with c(B) = A. For any other non zero C ∈ [A]⊥, we have c(C) ∧ c(B) = c(C) 6= 0 and hence bac 6= 0, for some b ∈ B a ∈ A and c ∈ C, by Corollary 3.32. So C′ ∼ B′ ⊆ B, where B′ = {cab}⊥⊥ and C′ = 44 TRISTAN BICE {bac}⊥⊥, which, as B and hence B′ is commutative, means C′ is also commutative, (cid:3) by Corollary 3.54. Before leaving the topic of abelian annihilators, let us point out that, while projection lattices are a complete isomorphism invariant for commutative AW- algebras (by Stone's representation theorem), the same is not true for annihilator lattices of C-algebras. Indeed, if X is any topological space with a countable basis of regular open subsets then [Cb(X)]⊥ ∼= [B∞]⊥ × {0, 1}n, where B∞ denotes the free Boolean algebra with infinitely many generators and n is the number of isolated points of X (see [Bir67] Ch IX §4 Theorem 3). So [C([0, 1])]⊥ is orthoisomorphic to [C([0, 1] × [0, 1])]⊥, for example, even though C([0, 1]) and C([0, 1] × [0, 1]) are not isomorphic C-algebras (as their spectrums [0, 1] and [0, 1]×[0, 1] are not homeomorphic, having dimension 1 and 2 respectively). 3.9. Homogeneous Annihilators. We have two competing notions of homogene- ity. One is classical, where A is said to be n-homogeneous (n-subhomogeneous), for some n ∈ N, if dim(Hπ) = n (dim(Hπ) ≤ n), for all π ∈ bA(= the set of all irre- ducible representations of A). The other is the natural one coming from the abelian annihilators, that is the notion of A being n-[A]⊥D-(sub)homogeneous, according to Definition 2.27. We show in this subsection that these concepts are closely related. Proposition 3.56. If A is n-[A]⊥-homogeneous then A is not (n−1)-subhomogeneous. Proof. Take orthogonal B1, . . . , Bn ∈ [A]⊥ with c(Bk) = A, for all k. This means that the ideal generated by each Bk is essential in A. So, taking any b1 ∈ B1 \ {0}, we can therefore find a1 ∈ A and b2 ∈ B2 with b1a1b2 6= 0. Continuing in this manner, we obtain a1, . . . , an−1 ∈ A and b1, . . . , bn with bk ∈ Bk, for all k, and c = b1a1b2a2 . . . ..an−1bn 6= 0. Let π be an irreducible representation of A on a Hilbert space H with π(c) 6= 0. Then π(bk) 6= 0, for all k, which, as the B1, . . . , Bn are orthogonal, means that dim(H) ≥ n. (cid:3) Definition 3.57. A map F on Asa, for each C-algebra A, is functorial if π ◦ F = F ◦ π whenever π : A → B is a (possibly non-unital) C-algebra homomorphism. Lemma 3.58. If we have rank one p1, . . . , pn ∈ P(Mn) with p1 ∨ . . . ∨ pn = 1 then there are functorial maps F = Fp1,...,pn and G = Gp1,...,pn with G(p1, . . . , pn) = p1 such that, whenever we have rank one q1, . . . , qn ∈ P(Mn) and G(q1, . . . , qn) 6= 0, q1 ∨ . . . ∨ qn = 1 = F (q1, . . . , qn). Proof. For projections p and q on a Hilbert space, σ(pq) = σ(p⊥q⊥) = σ(p⊥q⊥p⊥). So, if 0 < δ < 1 − sup(σ(pq) \ {1}), fδ(1 − p⊥q⊥p⊥) = 1 − (p⊥q⊥p⊥){1} = (p⊥ ∧ q⊥)⊥ = p ∨ q. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 45 So we get the functorial maps we want be letting Fp1 and Gp1 be the identity and recursively defining Fp1,...,pn(q1, . . . , qn) = fδ/2(q) Gp1,...,pn(q1, . . . , qn) = fδ(q)Gp1,...,pn−1(q1, . . . , qn−1), where and q = 1 − (1 − qn)(1 − Fp1,...,pn−1(q1, . . . , qn−1))(1 − qn) and δ = 1 − (p1 ∨ . . . ∨ pn−1)pn2. (cid:3) Theorem 3.59. If A = B1∨. . .∨Bn, for B1, . . . , Bn ∈ [A]⊥D, A is n-subhomogeneous. projection rkπ of rank at most one with π[Bk] = Crkπ. Let Proof. Assume that we have π ∈ bA with dim(Hπ) > n. For all k ≤ n, Bk is a commutative hereditary C-subalgebra of A and hence, for all π ∈ bA, there is a and pick π ∈ bA such that dim(Hπ) > n and m = rank(Wk≤m rkπ). We can then also pick b1, . . . , bm ∈Sk≤n Bk+ such that π(bk) is a (necessarily rank one) projection, for each k ≤ m, and rank(Wk≤m π(bk)) = m. Take any δ ∈ (0, 1) and set rkπ) : π ∈ bA and dim(Hπ) > n} m = max{rank(_k≤n bF = Fπ(b1),...,π(bm)(fδ/2(b1), . . . , fδ/2(bm)) bG = Gπ(b1),...,π(bm)(fδ/2(b1), . . . , fδ/2(bm)). and By irreducibility, we have a ∈ A+ with at least n + 1 points in σ(π(a)). Take continuous functions g1, . . . , gn+1 on R with disjoint supports that each have non- empty intersection with σ(π(a)). Now define c = (1 − bF )a0bGa1g1(a) . . . an+1gn+1(a)an+2fδ(b1) . . . an+m+1fδ(bm), where a0, . . . , an+m+1 ∈ A are chosen so that π(c) 6= 0 (which is possible by the irreducibility of π and the fact rank(π(bF )) < dim(Hπ)). all k ≤ n + 1, and hence dim(Hπ′ ) ≥ σ(π′(a)) > n. Likewise, for all k ≤ m, fδ(π′(bk)) 6= 0 and hence fδ/2(π′(bk)) is a rank one pro- Now say we have another π′ ∈ bA such that π′(c) 6= 0. Then gk(π′(a)) 6= 0, for jection. As we also have π′(bG) 6= 0, this means that π′(bF ) =Wk≤m fδ/2(π′(bk)) and rank(π′(bF )) = m which, by our choice of m, means π′(bF ) =Wk≤n rkπ′ . Thus 0 = π′(b(1 − bF )) = π′(bc), for all b ∈ Sk≤n Bk, and this also holds of course if lows that bc = 0, for all b ∈ Sk≤n Bk, i.e. cc∗ ∈ (B1 ∨ . . . ∨ Bn)⊥ = A⊥ = {0}, contradicting π(c) 6= 0. Theorem 3.60. A is n-[A]⊥D-homogeneous if and only if A contains an n-homogeneous essential ideal. π′(c) = 0. As π′ was arbitrary (and the atomic representation is faithful), it fol- (cid:3) Proof. Say A contains an n-homogeneous essential ideal B. Then B is liminal and hence discrete so B =W cB[B]⊥D. As [B]⊥D is order-dense in [B]⊥, by Corollary 3.55, we have B =W Cλ for (Cλ) ⊆ cB[B]⊥ with each Cλ being λ-[B]⊥D-homogeneous. If we had Cλ 6= {0}, for some λ 6= n, then we would have π ∈ bCλ with dim(Hπ) 6= n, by either Proposition 3.56 or Theorem 3.59. As Cλ is an ideal, this π extends to an 46 TRISTAN BICE irreducible representation of B, contradicting n-homogeneity. Thus B is n-[B]⊥D- homogeneous and hence A is n-[A]⊥D-homogeneous, by Proposition 3.28. Conversely, assume A = B1 ∨ . . . ∨ Bn, for orthogonal B1, . . . , Bn ∈ [A]⊥D with c(Bk) = A, for all k. This means that the closed ideal Ck generated by Bk is essential in A, for all k. As the intersection of a pair of essential ideals D and E is again essential (because then, for any a ∈ A \ {0}, we have e ∈ E with ae 6= 0, and then f ∈ F with aef 6= 0, where ef ∈ E ∩ F ), we see that C =T Ck is an essential ideal in A. If π ∈ bC then ker(π) ∩ Bk 6= 0, for each k, otherwise we would have C ⊆ ker(π) and π would be the zero representation. Thus dim(Hπ) ≥ n. But also C = B1 ∨C . . . ∨C Bk and hence dim(Hπ) ≤ n, by Theorem 3.59. (cid:3) Corollary 3.61. The following are equivalent. (1) A ∈ [A]⊥n D . (2) A ∈ [A]⊥D<n+1. (3) A is n-subhomogeneous. (4) There are orthogonal k-homogeneous ideals Bk with B1⊕ . . .⊕ Bn essential. Proof. (1)⇒(3) See Theorem 3.59. (3)⇒(2) If A is discrete but A /∈ [A]⊥D<n+1, then there exists non-zero B ∈ [A]⊥D(n+1). By Theorem 3.59, B has representations π with dim(Hπ) > n. By [Ped79] Proposition 4.1.8, the same is true of A and so A is not n-subhomogeneous. (2)⇒(1) Immediate. (2)⇒(4) Immediate from the 'only if' part of Theorem 3.60. (4)⇒(2) If (4) holds then, from the 'if' part of Theorem 3.60, we see that each B⊥⊥k is k-[B⊥⊥k ]⊥D-homogeneous and A = B⊥⊥1 ∨ . . . ∨ B⊥⊥n . (cid:3) We should mention that a more classical proof of (3)⇔(4) could be obtained by utilizing the Jacobson topology on qA(= {ker(π) : π ∈ bA} = the primitive ideal space -- see [Ped79] §4.1), specifically by using [Ped79] Theorem 4.4.6 and Proposition 4.4.10. Another thing qA can be used for is establishing a similar result about <ℵ0- subhomogeneity. Specifically, we call A <ℵ0-subhomogeneous if dim(Hπ) < ∞ for all π ∈ bA. This notion turns out to be different from <ℵ0-[A]⊥D-subhomogeneity (see Example 3.88), but we can at least show it is stronger. Proposition 3.62. If A is <ℵ0-subhomogeneous then A ∈ [A]⊥D<ℵ0 Proof. As A is < ℵ0-subhomogeneous, it is liminal and hence discrete. Thus if A /∈ [A]⊥D<ℵ0 we would have non-zero B ∈ [A]⊥Dℵ0 , meaning there are orthogonal (Bn) ⊆ [A]⊥D with c(Bn) = c(B), for all n. This means that the subsets of qB defined by On = {I ∈ qB : Bn \ I 6= {0}} are open dense in the Jacobson topology. Thus O =T On is non-empty, as qB is a Baire space (see [Ped79] Theorem 4.3.5), so B, and hence A, has an irreducible representation π which is not zero on all of Bn, for any n. Thus dim(Hπ) = ∞, contradicting the <ℵ0-subhomogeneity of A. (cid:3) Theorem 3.63. If A is <ℵ0-[A]⊥D-subhomogeneous then every essential hereditary C-subalgebra B of A contains an essential ideal. . ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 47 Proof. First assume that A is n-([A]⊥D-)subhomogeneous, for some n ∈ N. Let bAB = {π ∈ bA : dim(π[B]Hπ) < dim(Hπ)} and C = \π∈ bAB ker(π). This is immediately seen to be an ideal, and C ⊆ B because the atomic represen- tation is faithful on open projections, by [Ped79] Proposition 4.3.13 and Theorem 4.3.15. All we need to verify is that C⊥ = {0}. Assume to the contrary that we have non-zero c ∈ C⊥+ . Then c /∈ C so we have π ∈ bAB with π(c) 6= 0 and dim(π[B]Hπ) = m = max{dim(π′[B]Hπ′ ) : π′ ∈ bAB and π′(c) 6= 0}. As B is hereditary, so is π[B] and we can take b ∈ B+ with σ(π(b)\{0}) = m. Let g1, . . . , gm be continuous functions on R with disjoint supports contained in [δ,∞), for some δ > 0, that each have non-empty intersection with σ(π(b)). Now define a = (1 − fδ(b))a0g1(b)a1 . . . gm(b)amc ∈ C⊥, where a0, . . . , am ∈ A are chosen so that π(a) 6= 0. For the general case, take an increasing sequence of (annihilator) ideals (In) in Take any b′ ∈ B. As a ∈ C⊥ we have b′a ∈ C⊥ so, if b′a 6= 0, there exists all k ≤ m. This means σ(π′(b)\{0}) ⊆ [δ,∞) and π′(b′a) = π′(b′(1 − fδ(b))) = 0, by the definition of m, a contradiction. Thus b′a = 0 which, as b′ was arbitrary, means aa∗ ∈ B⊥ = {0}, contradicting π(a) 6= 0. π′ ∈ bAB with π′(b′a) 6= 0. But then π′(a) 6= 0 so π′(c) 6= 0 and gk(π′(b)) 6= 0, for A such thatS In is essential in A and In is n-subhomogeneous, for all n. As B is essential ideal Cn in In. But thenS Cn is an essential ideal inS In which, asS In is itself an essential ideal in A, meansS Cn is an essential ideal in A. essential in A, and In is an ideal in A, B ∩ In is essential in In (if a ∈ In+ then we have b ∈ B+ with ab 6= 0 6= abab and bab ∈ B ∩ In+), and hence contains an (cid:3) Corollary 3.64. If B ∈ [A]⊥ is <ℵ0-[B]⊥D-subhomogeneous then [B] = [B]B so [A]⊥D<ℵ0 ⊆ [A]⊥=. Proof. Take C ∈ [B]. As C is essential in C⊥B⊥B , we have an essential ideal D in C⊥B⊥B with D ⊆ C, by Theorem 3.63. Then C⊥B⊥B = C⊥B⊥B⊥⊥ = D⊥⊥ = C, (cid:3) by Proposition 3.29. Corollary 3.65. If B is a hereditary C-subalgebra of A and B ∈ [B]⊥Dn then B⊥⊥ ∈ [A]⊥Dn. , . . . , B⊥⊥n ∈ [A]⊥D, Proof. Say B1, . . . , Bn ∈ [B]⊥D witness B ∈ [B]⊥Dn. Then B⊥⊥1 ∈ [A]⊥Dn, by Corollary 3.33. We by Theorem 3.51, and C = B⊥⊥1 ∨ . . . ∨ B⊥⊥n do not immediately know that B ⊆ C, as supremums in B and A can differ, but C will contain the hereditary C-subalgebra generated by B1, . . . , Bn. As this hereditary C-subalgebra is essential in B, it contains an essential ideal D in B, by Theorem 3.63. By Proposition 3.29, we have B⊥⊥ = D⊥⊥ ⊆ C ⊆ B⊥⊥. (cid:3) Theorem 3.66. If B, C ∈ [A]⊥ are perspective and B ∈ [A]⊥Dn then C ∈ [A]⊥Dn. 48 TRISTAN BICE By Corollary 3.55, [A]⊥D is order-dense in [W c[A]⊥D] ⊇ [c(B)] = [c(C)]. By Theorem 3.51, Proof. It suffices to prove the result for semiorthoperspective B and C (see §2.6). [C]⊥D is order-dense in C so by Theorem 2.28, C is [C]⊥D-subhomogeneous. If C is not n-[C]⊥D-homogeneous then one of the non-zero homogeneous parts D has order < n or > n. In the first case, D = C ∩ c(D) and B ∩ c(D) are semiorthoperspec- In the second, we have E ∈ [D] that is (n + 1)-[C]⊥D-homogeneous. Then tive. [E]E = [E] and [B] = [B]B, by Corollary 3.64, and hence E is semiorthoperspec- tive to some F ∈ [B], by Corollary 2.12. By cutting down further by a central element, we may assume that F is m-homogeneous for some m ≤ n. So in either case, after some renaming, we have m-[A]⊥D-homogeneous B semiorthoperspective to n-[A]⊥D-homogeneous C with m < n. We show that this leads to a contradiction. Take orthogonal B1, . . . , Bm ∈ [A]⊥D with c(Bk) = c(B), for all k ≤ m. By Corollary 3.32, we have B′1 ⊆ B1 and B′2 ⊆ B2 such that B′1 ∼ B′2. We also have non-zero B′′2 ⊆ B′2 and B′′3 ⊆ B3 with B′′2 ∼ B′′3 . We then also have B′′1 ⊆ B′1 with B′′1 ∼ B′′2 , by Corollary 3.38. Continuing in this way, and again renaming, we get non-zero orthogonal B1 ∼ . . . ∼ Bm ∼ C1 ∼ . . . ∼ Cn in [A]⊥D such that B ∩ D = B1 ∨ . . .∨ Bm and C ∩ D = C1 ∨ . . .∨ Cn, where D = c(B1) = . . . = c(Cn). As D is central, replacing B and C with B ∩ D and C ∩ D we still see that B and D are semiorthoperspective. As they are also principal, we have B ∼ C, by Proposition 3.41. This means that C = D1∨. . .∨Dm, with Bk ∼ Dk, for all k ≤ m. By Corollary 3.54, each Dk will be abelian and hence C will be m-subhomogeneous, (cid:3) by Theorem 3.59, contradicting Proposition 3.56. Corollary 3.67. Assume A is <ℵ0-[A]⊥D-subhomogeneous. (1) Every open dense projection in A′′ is regular. (2) The function BC⊥, for B, C ∈ [A]⊥, satisfies the triangle inequality. (3) [A]⊥ is modular, and hence a continuous geometry. Proof. (1) Every dense p ∈ P◦(A′′) dominates some dense q ∈ P◦(A′′ ∩ A′), by Theorem 3.63. As central projections are necessarily regular (see [Eff63] §6), p must also be regular. b ∈ B1 (2) Take B, C, D ∈ [A]⊥ and q ∈ P◦(A′′ ∩ A′) with q ≤ pC + pC⊥. For every + and d ∈ D⊥1 + , we have bd = bdq = bqd ≤ b(pC + pC⊥)d ≤ bpC⊥ + pC d, (3) We prove that perspectivity is finite (see (2.10)). and hence BD⊥ ≤ BC⊥ + CD⊥. If not, we would have perspective B, C ∈ [A]⊥ with C < B. By cutting down by a central el- ement, we may assume that C is n-[A]⊥D-homogeneous, for some n ∈ N. By Corollary 3.64, C ∈ [B] = [B]B so C < B means we have non- zero b ∈ B+ ∩ C⊥. As c(B) = c(C), by perspectivity, the proof of Proposition 3.56 (starting with b instead of b1) shows that B is not n- subhomogeneous. However, B is n-[A]⊥D-homogeneous, by Theorem 3.66, and hence n-subhomogeneous, by Theorem 3.59, a contradiction. (cid:3) So, by Corollary 3.67(3) and [Kap55], the annihilators [A]⊥ in a < ℵ0-[A]⊥D- subhomogeneous C-algebra A form a continuous geometry. As far as we know, these kinds of continuous geometries have not been investigated before. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 49 3.10. Matrix Valued Functions. Definition 3.68. Assume X and Y are topological spaces. We denote the set of continuous functions from X to Y by C(X, Y ). If A is a C-algebra, we call f : X → A bounded if supx∈X f (x) < ∞. The C-algebras (with pointwise operations) of bounded and bounded continuous functions from X to A will be denoted by B(X, A) and Cb(X, A) respectively. We call f : X → P(A) lower semicontinuous at x ∈ X if ∀ǫ > 0∃ a neighbourhood Y of x s.t. ∀y ∈ Y (p(y)⊥p(x) < ǫ). We say that p is lower semicontinuous if it is lower semicontinuous at every x ∈ X and denote the set of all lower semicontinuous p by C◦(X,P(A)). We call p upper semicontinuous if p⊥ (i.e. the function defined by p⊥(x) = p(x)⊥, for all x ∈ X) is lower semicontinuous and denote the set of upper semicontinuous p by C(X,P(A)). By Proposition 3.17, p : X → P(A) will be continuous at x if and only if it is simultaneously upper and lower semicontinuous at x and hence C(X,P(A)) = C(X,P(A)) ∩ C◦(X,P(A)). Also note that if d : P(A) → R is continuous and monotone in the sense that p ≤ q ⇒ d(p) ≤ d(q) then, by Proposition 3.14 (1)⇒(3), d ◦ p will be lower semi- continuous in the usual sense (i.e. f−1(r,∞) is open, for all r ∈ R) if p is lower semicontinous according to Definition 3.68. In particular, d could be the restriction to P(A) of a dimension function on A. For f : X → A, [f ] denotes the map x 7→ [f (x)] = (f (x)f (x)∗)(0,∞) to P(A′′). Proposition 3.69. If f ∈ C(X, A+), x ∈ X and f (x) has finite spectrum then [f ] is lower semicontinuous at x. Proof. Say δ > 0, a, b ∈ A, a− b ≤ δ and λ ∈ σ(b). For unit v ∈ R(b{λ}) we have λa{0}v = a{0}(a − λ)v ≤ (a − λ)v ≤ (b − λ)v + δ = δ, and hence a{0}b{λ} ≤ δ/λ. As f (x)(0,∞) is a finite sum of spectral projections, this calculation shows that, for any ǫ > 0, by choosing a sufficientely small neigh- bourhood Y of x, we can ensure that f (y){0}f (x)(0,∞) ≤ ǫ, for all y ∈ Y . (cid:3) If f (x) has infinite spectrum then f can indeed fail to be lower semicontinuous at x. For example, let X = N∞ (the one point {∞} compactification of N) and A = K(H), where H is infinite dimensional separable with basis (en). Let pn denote the projection onto Cen, for each n ∈ N. Now consider f ∈ C(X, A+) defined by k=1 2−kp2k +P∞k=n+1 2−kp2k+1 and f (∞) = P∞k=1 2−kp2k. Note that k=1 p2k +P∞k=n+1 p2k+1 and [f ](∞) =P∞k=1 p2k so [f ](n)⊥[f ](∞) = f (n) = Pn [f ](n) =Pn 1, for all n ∈ N, and hence [f ] is not lower semicontinuous at ∞. Lemma 3.70. The map (p, q) 7→ [pq] is lower semicontinuous on {(p, q) ∈ P(A) × P(A) : inf(σ(pq) \ {0}) > 0}. Proof. If inf(σ(pq) \ {0}) > 0 then r = (qpq)(0,1] ∈ A, p⊥r < 1 and [pq] = [pr] ∈ A. By [Bic13] Lemma 2.3, for any ǫ > 0 we can choose δ > 0 so that p − p′,r − r′ ≤ δ implies [p′r′] − [pr] ≤ ǫ. But then p − p′,q − q′ ≤ δ implies r′ = (q′rq′)(0,1] ∈ A and r − r′ < δ (see [Bic13] (2.3)) and hence [p′q′]⊥[pq] ≤ [p′r′]⊥[pr] ≤ [p′r′] − [pr] ≤ ǫ. 50 TRISTAN BICE (cid:3) Lemma 3.71. The map (p, q) 7→ p ∨ q is lower semicontinuous on {(p, q) ∈ P(A) × P(A) : sup(σ(pq) \ {1}) < 1}. Proof. If sup(σ(pq) \ {1}) < 1 then r = (qpq)[0,1) ∈ A, pr < 1 and p ∨ q = p ∨ r ∈ A. By [Bic13] Lemma 2.4, for any ǫ > 0 we can choose δ > 0 so that p − p′,r − r′ ≤ δ implies p′ ∨ r′ − p ∨ r ≤ ǫ. But then p − p′,q − q′ ≤ δ implies r′ = (q′rq′)(0,1] ∈ A and r − r′ < δ (see [Bic13] (2.3)) and hence (p′ ∨ q′)⊥(p ∨ q) ≤ (p′ ∨ r′)⊥(p ∨ r) ≤ p′ ∨ r′ − p ∨ r ≤ ǫ. (cid:3) If p, q ∈ C◦(X,P(Mn)) then sup(σ(p(x)q(x)) \ {1}) < 1, for all x ∈ X, so it follows from Lemma 3.71 and Proposition 3.14 (1)⇒(3) that p∨ q ∈ C◦(X,P(Mn)) (where (p ∨ q)(x) = p(x) ∨ q(x), for all x ∈ X). Also, if (pn) ⊆ C◦(X,P(Mn)) is increasing (i.e. for any x ∈ X, (pn(x)) is (non-strictly) increasing in P(Mn)) then, for any x ∈ X, it follows thatWn pn(x) = pm(x), for some m ∈ N, which, as pm ≤ Wn pn and pm is lower semicontinuous, means thatWn pn is lower semicontinuous at x. As x ∈ X was arbitrary,W pn ∈ P(Mn). This applies equally well to transfinite sequences and thus, combining thes facts, we see that C◦(X,P(Mn)) is closed under taking arbitrary suprema, i.e. (3.21) P ⊆ C◦(X,P(Mn)) ⇒ _P ∈ C◦(X,P(Mn)). Given a Hilbert space H, we represent the C-algebra of bounded functions B(X,B(H)) from X to B(H) on Lx∈X H by defining f (vx)x∈X = (f (x)vx)x∈X, for any f ∈ B(X,B(H)) and (vx)x∈X ∈Lx∈X H. Thus, our assumption that A is represented concretely and faithfully on some Hilbert space H means we can identify Cb(X, A)′′ with a subset of B(X, A′′) and then use non-commutative topological concepts on B(X,P(B(H))) with reference to Cb(X, A). In fact, if X is a completely regular Hausdorff topological space then it is not difficult to see that Cb(X, A)′′ will be identified in this way with the entirety of B(X, A′′), i.e. Cb(X, A)′′ = B(X, A′′). Definition 3.72. For any function f between topological spaces we denote the set of points on which f is continuous by Cf . Proposition 3.73. If X is a completely regular topological space and p : X → P(A) then p◦(x) = p(x) = p(x), for all x ∈ C◦p . Proof. For any x ∈ C◦p we have a function g : X → [0, 1] with g(x) = 1 and g[X \ C◦p ] = {0}. Then gp (the map x 7→ g(x)p(x)) is continuous on X, gp ≤ p and g(x)p(x) = p(x). As p◦ = sup(pCb(X, A)p ∩ Cb(X, A))1 +, the first statement is proved. The second follows from this, Cp = Cp⊥ and p = p⊥◦⊥. (cid:3) Representing Mn canonically on Cn, we have M′′n = Mn, and thus we can identify Cb(X, Mn)′′ with a subset of B(X, Mn). The following result says that, under this identification, the open and closed projections in Cb(X, Mn)′′ are precisely the lower and upper semicontinous projection functions respectively. Proposition 3.74. If X is a normal regular topological space then C◦(X,P(Mn)) = P(Cb(X, Mn)′′)◦ C(X,P(Mn)) = P(Cb(X, Mn)′′) and ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 51 Proof. First note that [f ] ∈ C◦(X,P(Mn)) whenever f ∈ C(X, Mn), by Proposition 3.69. Thus, for any hereditary C-subalgebra B of Cb(X, Mn), we have sup B1 f ∈ B1 C◦(X,P(Mn)). that p(y)⊥p(x) < 1, for all y ∈ Y . For each m ∈ N, the set Ym = {y ∈ Y : dim(p(y)) ≤ dim(p(x)) + m} +} ∈ C◦(X,P(Mn)), by (3.21), which means that P◦(Cb(X, Mn)′′) ⊆ Given p ∈ C◦(X,P(A)) and x ∈ X, let Y be a closed neighbourhood of x such + =W{[f ] : is closed. The function defined by f0(y) = [p(y)p(x)] is continuous on Y0 and thus has a continuous extension g0 : X → M 1 n, by the Tietze extension theorem. The function defined by f1(y) = p(y)g0(y)∗g0(y)p(y) is continuous on Y1 and, con- tinuing to apply the Tietze extension theorem in this way we eventually obtain f = fn−dim(p(x)) ∈ C(Y, M2+) with f (y) ≤ p(y), for all y ∈ Y . As X is completely regular, we have a function h : X → [0, 1] with h(x) = 1 and h[X \ Y ] = {0}. The function fx defined by fx(y) = h(y)f (y), for all y ∈ Y , and fx(y) = 0, for y ∈ X\Y , is continuous on all of X. We thus have p = supx∈X fx ∈ P◦(Cb(X, Mn)′′) (cid:3) βX We suspect that normality here could not be replaced with complete regularity. To see this, all you would need is a continuous function f : Y → [0, 1], where Y is a closed subset of a completely regular space X such that Y is not locally compact and f has no extension to any Z ⊆ Y properly containing Y (βX here denotes the Stone- Cech compactification of X -- see [GJ60]). Identifying [0, 1] with a subspace of P(M2) and extending f to X by defining f (x) to be the identity of M2, for x /∈ Y , we see that f ∈ C◦(X,P(M2)). If f ∈ P◦(Cb(X, Mn)′′) then n+) with f = sup fn and hence f β = sup f β we would have a net (fn) ⊆ Cb(X, M 1 n (where f β n denotes the unique continuous extension of fn to βX) would be a lower \ Y . semicontinuous extension of f to βX. Then f would have to be 0 on Y \ Y is not closed and hence we would have But Y is not locally compact so Y f (y) = 0 for some y ∈ Y , a contradiction. Theorem 3.75. For every p ∈ C◦(X,P(Mn)), Cp is open dense. Proof. For finite rank projections q and r, dim(q) = dim(r) implies q − r = qr⊥ = rq⊥ and hence Cp = Cdim ◦p. As noted after Definition 3.68, dim◦p is lower semicontinuous and, as {0, . . . , n} is discrete, dim◦p will be constant, and hence continuous, on some open neighbourhood about any x ∈ Cdim ◦p. Now note that dim(p(x)) < n, for all x ∈ X \ Cp, by discontinuity. If X \ Cp contained an open set O, then we would have dim(p(x)) < n − 1, for all x ∈ O, again by discontinuity. But then dim(p(x)) < n − 2, for all x ∈ O, etc. and hence we eventually get dim(p(x)) = 0, for all x ∈ O, which shows that dim◦p is continuous (cid:3) on O, a contradiction, i.e. Cp is open dense. βX βX Definition 3.76. Given functions f, g on a topological space X, we define f =od g ⇔ {x ∈ X : f (x) = g(x)}◦ is dense, and f =d g ⇔ {x ∈ X : f (x) = g(x)} is dense. Note that =od is a transitive relation, while =d may not be. However, they coincide for the functions we are considering. Proposition 3.77. If X is a completely regular topological space, the following are equivalent, for p, q ∈ C◦(X,P(Mn)). 52 TRISTAN BICE (1) p = q. (2) Cp ∩ Cq ⊆ {x ∈ X : p(x) = q(x)}. (3) p =od q. (4) p =d q. If these conditions hold and p ≤ q then Cp ⊆ Cq, so Cp ⊆ {x ∈ X : p(x) = q(x)}. Proof. (1)⇒(2) Assume p(x) (cid:2) q(x) for some x ∈ Cp ∩ Cq. As X is completely regular, we have a continuous function f : X → [0, 1] such that f (x) = 1 and f [X \ Cq] = {0}. Letting g(y) = f (y)q(x)⊥, for all y ∈ X, we see that g is continuous and p(x)g(x) 6= 0, so g witnesses the fact q⊥◦ (cid:2) p⊥◦ and hence p = p⊥◦⊥ 6= q⊥◦⊥ = q. (2)⇒(3) Immediate from Theorem 3.75. (3)⇒(4) Immediate from the definitions. (4)⇒(1) If f ∈ p⊥C(X, Mn)p⊥ ∩ Cb(X, Mn)1 If p =d q then qf is 0 on a dense subset of X which, as qf is immediately seen to be lower semicontinuous, means qf is 0 on the entirety of X, i.e. f ∈ q⊥C(X, Mn)q⊥ ∩ Cb(X, Mn)1 +. Thus p⊥◦ = q⊥◦ and hence we have p = p⊥◦⊥ = q⊥◦⊥ = q. + then pf is 0 everywhere on X. For the last statement, assume that x ∈ Cp \ Cq, so there is an open neighbourhood Y of x such that dim(p(y)) = dim(p(x)), for all y ∈ Y . As x /∈ Cq, x must be a limit point of the open set Y ∩ {y ∈ X : dim(q(y)) > dim(q(x))} so, in particular, this set is not empty. As p(x) ≤ q(x), it is contained in {y ∈ X : p(y) 6= q(y)}, showing that this set has non-empty interior and p 6=d q. (cid:3) Say X is a completely regular topological space and we have a =od equivalence class E ⊆ C◦(X,P(Mn)). For any p ∈ E, we have p◦ ∈ E, by Proposition 3.73 and Theorem 3.75. And, for any other q ∈ E, we have p = q and hence p◦ = q◦, by Proposition 3.77, i.e. E contains precisely one topologically open projection. This will in fact be the maximum of E ∩P(Cb(X, Mn)′′)◦ and hence, if X is normal, the maximum of E, by Proposition 3.74. Proposition 3.78. If X is completely regular and p, q ∈ P(Cb(X, Mn)′′)◦, p and q commute in P(Cb(X, Mn)′′)◦ ⇔ pq =od qp. Proof. For any p, q ∈ P(Cb(X, Mn)′′)◦, we have p, q ∈ C◦(X,P(Mn)) by the first part of the proof of Proposition 3.74. We then have p∨q ∈ C◦(X,P(Mn)), by (3.21), and hence p ∨ q◦ =od p ∨ q, by Proposition 3.73 and Theorem 3.75. Likewise, we have p⊥◦ =od p and hence p ∧P(C b(X,Mn)′′) (3.22) ◦ q = (p⊥◦ ∨ q⊥◦)⊥◦ =od p ∧ q. Thus (p⊥◦ ∨ q⊥◦)⊥◦ ∨ (p⊥◦ ∨ q)⊥◦◦ =od (p ∧ q) ∨ (p ∧ q⊥). If p and q commute in P(Cb(X, Mn)′′)◦ then (3.22) becomes p =od (p∧ q)∨ (p∧ q⊥) which is equivalent to pq =od qp. If pq =od qp then p =od (p⊥◦ ∨ q⊥◦)⊥◦ ∨ (p⊥◦ ∨ q)⊥◦◦ hence p and q commute in P(Cb(X, Mn)′′)◦. which, as both sides here are topologically regular, means =od is actually = and (cid:3) As [A]⊥ and P(A′′)◦ are isomorphic (see Theorem 3.4), the relation ∼ we have on [A]⊥ can be seen as a relation on P(A′′)◦, which we will also denote by ∼, i.e. p ∼ q ⇔ ∃a ∈ A(p = [a]◦ and q = [a∗]◦). ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 53 Theorem 3.79. If X is a hereditarily paracompact Hausdorff topological space then, for p, q ∈ P(Cb(X, Mn)′′)◦, p ∼ q ⇔ dim◦p =od dim◦q Proof. If a ∈ A witnesses p ∼ q then (even if X is just completely regular), by Proposition 3.69, Proposition 3.73 and Theorem 3.75, dim◦p = dim◦[a]◦ =od dim◦[a] = dim◦[a∗] =od dim◦[a]◦ = dim ◦q. Yx. Specifically, let v(y) = p(x)p(y)pp(y)p(x)p(y)−1 On the other hand, if dim ◦p =od dim◦q then, for all x ∈ Cp ∩ Cq, dim(p(x)) = dim(q(x)). Let u be a partial isometry with u∗u = p(x) and uu∗ = q(x). Choose an open neighbourhood Yx of x with Yx ⊆ Cp∩Cq and p(x)−p(y),q(x)−q(y) < 1. This means p(x)p(y) has a polar decomposition for all y ∈ Yx and, moreover, the partial isometry appearing in this decomposition can be chosen continuously on , where −1 here denotes the quasi-inverse, so v(y) is a partial isometry with v(y)∗v(y) = p(y) and v(y)v(y)∗ = p(x), and v is continuous on Yx (alternatively, this can be derived from [Bla06] II.3.3.4). Likewise, define a continuous function w on Yx so that w(y)∗w(y) = q(x) and w(y)w(y)∗ = q(y), for all y ∈ Yx. Take a locally finite refinement (Zα) of resulting empty sets), we obtain a locally finite collection of disjoint open sets with (Yx)x∈Cp∩Cq . By replacing each Zα with Zα \Sβ<α Zβ (and then throwing out the S Zα = X. For each α pick x with Zα ⊆ Yx, choose a function gα : Zα → [0, 1] Defining f (z) = 0 for all z ∈ X \S Zα we see that the local finiteness of (Zα) with gα(x) = 1 and g[X \ Zα] = {0}, and set f (z) = g(z)w(y)uv(y), for all z ∈ Zα. implies f is continuous on all of X, i.e. f ∈ Cb(X, Mn). Furthermore, [f ] =od q and [f∗] =od p and so we must have [f ]◦ = q and [f∗]◦ = p. (cid:3) We now show that Theorem 3.79 can be used to prove a number of important facts about C-algebras of the form Cb(X, Mn). Corollary 3.80. If A is isomorphic to a C-algebra of the form Cb(X, Mn), for any hereditarily paracompact Hausdorff topological space X, then (1) A is anniseparable, (2) ∼ is finite on [A]⊥, (3) [A]⊥ is modular, and (4) ∼ coincides with perspectivity. Proof. (1) follows from Theorem 3.79 and the fact =od is reflexive. For (2), simply note that, for p, q ∈ P(Cb(X, Mn)′′)◦ with p ≤ q, dim◦p =od dim◦q implies p =od q and hence p = q, by Proposition 3.77. Now (3) and (4) follow from Theorem 2.32. (cid:3) Lastly, let us point out that any C-algebra of the form Cb(X, Mn) will actually be isomorphic to the C-algebra C(βX, Mn), where βX is the Stone- Cech com- pactification of X. So we could have restricted ourselves to compact X without restricting the class of C-algebras under consideration. However, as mentioned in §3.2, the Stone- Cech compactification can often be significantly harder to work with than X itself, which is why we chose not to do this (e.g. it is not clear to us that Theorem 3.79 would apply to βX if it applied to X, i.e. we do not know if the Stone- Cech compactification of a hereditarily paracompact space is again hereditarily paracompact). 54 TRISTAN BICE In fact, if X is compact Hausdorff then the representation of C(X, Mn) coming from the canonical representation of Mn on Cn is just the atomic representation of C(X, Mn). As every such space is normal and regular, Proposition 3.74 shows that lower semicontinous projection functions correspond precisely to the open pro- jections which, in turn, correspond precisely to the hereditary C-subalgebras of C(X, Mn), by [Ped79] Proposition 3.11.9, Proposition 4.3.13 and Theorem 4.3.15. Also, analogous theorems for C0(X, Mn), where X is locally compact, can also be proved in much the same way, or can be derived from the corresponding the- orems for C(X∞, Mn) (where X∞ is the one point compactification of X), using Proposition 3.28 and the fact that C0(X, Mn) is an essential ideal in C(X∞, Mn). Also, there is presumably some room for the results of this subsection to be generalized to non-trivial Hilbert and C-bundles (see [RW98]). Indeed, we could have derived Corollary 3.80(3) (which, combined with (4), also gives (2)) without reference to the topological space X, simply by using Corollary 3.67(3) and the fact Cb(X, Mn) ∼= C(βX, Mn) is n-homogeneous. At any rate, C-algebras of the form Cb(X, Mn) already allow us to create a number of instructive elementary examples, as we now show. 3.11. Examples. For use in the following examples, define Pθ ∈ P(M2) by Pθ =(cid:20)sin θ cos θ(cid:21)(cid:2)sin θ For p, q ∈ P(A′′)◦, we have cos θ(cid:3) =(cid:20) sin2 θ sin θ cos θ sin θ cos θ cos2 θ (cid:21) . p ≤ q ⇒ p ≤ q ⇔ p ≤ q ⇔ p ≤ q, and the first implication can not be reversed in general, even when A is commuta- tive. Slightly more worthy of note is the fact that p ≤ q does not even imply that p and q commute. Example 3.81 (p, q ∈ P(A′′)◦ with p < q but pq 6= qp). q(x) =(Pπ/4 Let A = C([0, 1], M2) and define for x ∈ [0, 1 2 ] for x ∈ ( 1 2 , 1] We immediately have p < q and for x ∈ [0, 1 2 ] for x ∈ ( 1 2 , 1] p(x) =(0 and P0 1 . p(x) =(0 2 ) = P0Pπ/4 6= Pπ/4P0 = qp( 1 2 ). P0 for x ∈ [0, 1 2 ) for x ∈ [ 1 2 , 1] , so pq( 1 Example 3.82 (p, q ∈ P(A′′)◦ with pq = qp, pq = qp and p q = q p but pq 6= qp). Let A = C([0, 1], M2) and define for x ∈ [0, 1 2 ) for x ∈ [ 1 2 , 1] p(x) =(1 P0 We immediately have pq( 1 and q(x) =(Pπ/4 1 for x ∈ [0, 1 2 ] for x ∈ ( 1 2 , 1] . 2 ) = P0Pπ/4 6= Pπ/4P0 = qp( 1 q(x) =(Pπ/4 for x ∈ [0, 1 2 ) for x ∈ [ 1 2 , 1] 1 , 2 ) and Let A = C([0, 1], M2) and define for x ∈ [0, 1 2 ) for x = 1 2 for x ∈ ( 1 2 , 1] P0 0 Pπ/4 p(x) = and q(x) =(P0 1 for x ∈ [0, 1 2 ] for x ∈ ( 1 2 , 1] . ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 55 so pq = qp. By concatenating the projection functions (and their orthocomplements) in Example 3.81 and Example 3.82 (with the functions taking the constant value 1 on open intervals in between) it is easy to see that one can obtain p, q ∈ P(C([0, 1], M2)′′)◦ having any combination of truth values for the statements (3.23) pq = qp, pq = qp, pq = qp and p q = q p, (although, if even one of the statements in (3.23) is true, then this is automatic from while still having pq =od qp, i.e. while still having p and q commute in P(C([0, 1], M2)′′)◦ Proposition 3.5 and the fact P(C([0, 1], M2)′′)◦ is orthomodular, by Corollary 3.80 (3)). While on the topic of commutativity, we mention the following natural question posed by Akemann -- if p < q are topologically regular open projections in A′′, for some separable C-algebra A, can we always find commuting a, b ∈ A that have range projections p and q respectively? The answer is no in general, even when p and q satisfy all the various combinations of commutativity in (3.23), as the following example shows. Example 3.83 (p, q ∈ P(A′′)◦, p < q but ab 6= ba when [a] = p and [b] = q). 2 )] 6= P0 and hence [b] 6= q. If we have a ∈ A with [a] = p then, for x 6= 1 We immediately see that p, q ∈ P(A′′)◦, p < q, and all the equations in (3.23) 2 , a(x) = λ(x)p(x), for hold. some non-zero λ(x) ∈ C. Thus if we have b ∈ A with ab = ba then, for x 6= 1 2 , b(x) ∈ Cp(x) + Cp(x)⊥. But (CP0 + CP ⊥0 ) ∩ (CPπ/4 + CP ⊥π/4) = C1 and hence b( 1 2 ) = λ1, for some λ ∈ C. In particular, [b( 1 A non-regular projection in a liminal C-algebra is given in [Ake70] Example I.2. The following example shows that non-regular projections even exist in homo- geneous C-algebras and, moreover, that they can still be topologically regular and even equivalent to a regular projection. In fact, the p and q given below are even Murray-von Neumann equivalent in A∗∗, and the partial isometry witnessing this will even yield an isomorphism between pAp ∩ A and qAq ∩ A, i.e p and q will even be Peligrad-Zsid´o equivalent (see [PZ00]). Example 3.84 (p, q ∈ P(A′′)◦ with dim◦p = dim◦q, p non-regular and q regular). Let A = C([0, 1], M2) and let p(x) = P(π/4) sin(1/x), for all x ∈ (0, 1], and set p(0) = 0. We immediately see that p ∈ P(A′′)◦ and p is identical to p except at 0, where we have p(0) = 1. If r(x) = Pπ/2, for all x ∈ [0, 1], then r ∈ P(A) and pr = r(0) = 1 > 1/√2 = pr, so p is not regular. Let q be a continuous function from (0, 1] to rank 1 projections in M2 such that {q(1/n) : n ∈ N \ {0}} is dense in the collection of rank 1 projections in M2. Extending q to a lower semicontinuous function on [0, 1] by defining q(0) = 0, we immediately see that q ∈ P(A′′)◦, while we also have a(0) ≤ supx∈(0,1] a(x)q(x) ≤ aq and hence aq = aq, for any a ∈ A, i.e. q is regular. 56 TRISTAN BICE On the other hand, it is easy to find examples of open projections that are regular but not topologically regular. Indeed, if A is commutative then every projection in A′′ is central and hence regular (as noted before [Eff63] Theorem 6.1) and so any non-regular open subset of the spectrum of A will represent such a projection. So, apart from the name, there does not appear to be any strong connection between regular and topologically regular projections. Example 3.85 (P(A′′)◦ is not (operator norm) closed). Let A = C([0, 1], M2). For each n ∈ N, define pn ∈ P(A′′)◦ by pn(x) = P(1/n) sin(1/x), for all x ∈ (0, 1], and set pn(0) = 0. Then pn converges to p∞, where p∞(x) = P0, for all x ∈ (0, 1], and again p∞(0) = 0. This p∞ is open (as it should be, because the set of open projections is always norm closed in A′′, by [Ped79] Proposition 3.11.9) but not topologically regular, as p∞(0) = P0 and p◦ ∞ = p∞. may not be complete even with respect to the orthometric. Note the above sequence also converges in the orthometric max(pq⊥◦,qp⊥◦) coming from the orthonorm (see Corollary 3.67(2)), but to a topologically regular open projection this time, namely p◦ ∞ = p∞. Thus the orthometric would appear to yield the more natural topology on P(A′′)◦, at least in this case. However, P(A′′)◦ Example 3.86 (P(A′′)◦ is not complete in the orthometric). Let A = C([0, 1], M2). Let g1(x) = sin(1/x), for all x ∈ (0, 1], and recursively define gn+1(x) = gn(2x), for x ∈ (0, 1 2 , 1]. For each n, let pn(x) = 0, when x = m2−n for some m ∈ N with m < 2−n, and k=1 4−kgk(x), for all other x ∈ [0, 1]. Now (pn) ⊆ P(A′′)◦ and (pn) is pn(x) = PPn Cauchy in the orthometric but, as dyadic rationals are dense in [0, 1], (pn) has no limit in P(A′′)◦. 2 ], and gn+1(x) = gn(2x − 1), for x ∈ ( 1 If we want to work with a complete metric space, we can of course just take the completion with respect to the orthometric (as long as · · is an orthonorm on [A]⊥). The orthonorm then extends continuously to this completion and we can then naturally extend the ordering too by defining B ≤ C ⇔ BC⊥ = 0. However, we do not know whether the order properties of the orthometric comple- tion are generally better or worse than those of [A]⊥ itself. The next example is important because it shows that all the work we did in §2, extending results about orthomodular lattices to separative ortholattices and carefully distinguishing [p] and [p]p, was in fact necessary for the theory to apply to arbitrary C-algebras. Example 3.87 ([A]⊥ is not orthomodular). Let A = C([0, 1],K(H)), where H is a separable infinite dimensional Hilbert space with basis (en), and let (sn) enumerate a dense subset of [0, 1]. Define Un : H → C2 by Unv = (hv, e2ni,hv, e2n+1i) and set pn(sn) = 0 and, for x 6= sn, pn(x) = U∗nP1/(x−sn)Un. Let p = P pn ∈ P(A)◦. Also let v = P(1/n)en ∈ H, let Q ∈ P(K(H)) be the projection onto Cv and define q ∈ P(A) by q(x) = Q, for all x ∈ [0, 1]. Note that p ∨ q > p. ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 57 We first claim that p ∨ q ∈ P(A)◦. To see this, note that pn(x) = pn(x), for x 6= sn, and pn(sn) = U∗nUn. Also p =P pn and p ∨ q = p ∨ q. Now take m ∈ N, let (xn) ⊆ [0, 1] be such that xn → sm and P1/(xn−sm) = P0, for all n. Then p ∨ q(xn) → p(sm)∨Q∨U∗mP0Um in the weak (and strong) operator topology, while if a ∈ A+ then a(xn) → a(sm) in norm. Thus if a ≤ p ∨ q then a(sm) ≤ p(sm) ∨ Q ∨ U∗mP0Um. But we could have also chosen (xn) such that P1/(xn−sm) = Pπ/2, for all n, and this would show that a(sm) ≤ p(sm) ∨ Q ∨ U∗mPπ/2Um. But (p(sm) ∨ Q ∨ U∗mP0Um) ∧ (p(sm) ∨ Q ∨ U∗mPπ/2Um) = p(sm) ∨ Q that is not <ℵ0-subhomogeneous). which, as m was arbitrary, shows that p ∨ q◦ = p ∨ q. We next claim that (p∨ q)∧ p⊥ = {0}, which will verify the non-orthomodularity of [A]⊥. If not, we would have non-zero r ∈ A+ with r ≤ p∨ q and r ⊥ p. For some n, we would then have r(sn) 6= 0 and hence r(sn) = λR, where R = p(sn)∨Q−p(sn) and λ 6= 0. But RU∗nP0Un 6= 0 and hence r(x)pn(x) 6= 0 for some x sufficiently close to sn. This contradicts the fact that rp = 0. Example 3.88 (A ∈ [A]⊥D<ℵ0 Let (en) be a basis for a separable Hilbert space H. For each n ∈ N, iden- tify Mn in the canonical way with PnB(H)Pn, where Pn is the projection onto span{e1, . . . , en}. Let A be the C-subalgebra of Qn Mn of sequences (an) con- verging to some a∞ ∈ B(H). Then each canonical copy of Mn in A will be an n-homogeneous annihilator ideal (the annihilator of the kernel of the nth coordinate representation in fact) and (Wn Mn)⊥ = {0} so A ∈ [A]⊥D<ℵ0 . But the representa- tion π that takes each (an) to its limit a∞ will map A onto K(H). Thus π is irre- ducible which, as H is infinite dimensional, means A is not <ℵ0-subhomogeneous. We have focused on examples of C-algebras of continuous functions, as these are the most tractable and already provide a good intuitive basis for working with annihilators. But there are undoubtedly many secrets to be gleaned from looking at the annihilator structure of all the various C-algebras under investigation in current research. We make the first tentative step in this direction with the following simple example. we have es,t = Pr∈{0,1}m esr,tr, where sr is the m + n length sequence obtained Proposition 3.89. The CAR algebra M2∞ is (purely) infinite. Proof. For each n ∈ N, let (es,t)s,t∈{0,1}n be the canonical matrix units of the canonical unital copy of M2n in M2∞. So, for each m, n ∈ N and s, t ∈ {0, 1}n, from simply concatenating s and r (and likewise for tr). We also have eq,res,t = 0 whenever r and s differ on their common domain. For each n ∈ N, let δn be the sequence of n 0s followed by a 1, and recursively define σn ∈ {0, 1}n+1 so that σj and σk differ on their common domain, for distinct j and k, and such that every finite sequence of 0s and 1s has some restriction or extension in (σn). Setting Thus, for any n ∈ N, we can find m ∈ N and (tr)r∈{0,1}n ⊆ {0, 1}m so that en =Pr∈{0,1}n ertr,rtr ≤ 22(m+n)a∗a and hence en ∈ {a}⊥⊥ with en = 1. Also, for any n ∈ N and b = Pr,s∈{0,1}n λr,ser,s ∈ M2n ⊆ M2∞ , we have enben = a =P∞n=1 2−neδn,σn , we have aa∗ = 2−2neδn,δn ∞Xn=1 while a∗a = 2−2neσn,σn . ∞Xn=1 58 TRISTAN BICE Pr,s∈{0,1}n λr,sertr,sts = b. As Sn M2n is dense in M2∞ this means that supe∈{a}⊥⊥,e=1 be = b for all b ∈ M2∞ so {a}⊥ = {a}⊥⊥⊥ = {0}, i.e. {a}⊥⊥ = M2∞. On the other hand, we immediately see that e1,1 ∈ {a∗}⊥ and so {a∗}⊥⊥ 6= M2∞ , so a witnesses the equivalence of M2∞ with a proper subannihila- tor. To see that any annihilator B is equivalent to a proper subannihilator just take any projection p ∈ B and note that it must be unitarily equivalent to some pro- jection in M2n, for some n ∈ N, which itself will have a subprojection unitarily equivalent to a matrix unit e ∈ M2n. But then {e}⊥⊥ is isomorphic to M2∞ and (cid:3) so the argument of the previous paragraph applies. This example might be a little surprising, given that the CAR algebra, and all other UHF algebras for that matter, are usually thought of as being very finite C- algebras, primarily due to the fact they possess a (unique bounded) faithful trace τ which, furthermore, always gives rise to the unique hyperfinite II1 factor as πτ [A]′′ (where πτ is the representation coming from the GNS construction applied with τ ). This might even lead some operator algebraists to dismiss the annihilators as clearly giving the 'wrong' notion of finiteness. Alternatively, you might blame the relation ∼ rather than the annihilators themselves, and if it turned out that the inclusion ordering on the annihilators in the CAR algebra is modular, despite being infinite w.r.t. ∼, then it might indeed be the better to focus on perspectivity rather than the ∼ relation, at least for some C-algebras. However, we would interpret Proposition 3.89 rather as merely showing that the close relationship between traces and projections in von Neumann algebras does not extend to annihilators in C- algebras. If the tracial structure is what you are interested in then you are better off looking at the positive elements of A under the equivalence notion given in [CP79], as mentioned in §1.1. In fact, traces give rise to dimension functions on A but these are quite different from the natural dimension functions you would consider on [A]⊥, even in the commutative case. To see this, note that dimension functions on C(X) correspond to measures µ on X, while annihilators correspond to regular open subsets O of X. Thus you would naturally think D(O) = µ(O) defines a dimension function on these regular open sets, but this is not the case. Consider, for example, X = [−1, 1] and µ = δ0, the point probability measure at 0. Then µ([−1, 0)) = 0 = µ((0, 1]) even though µ([−1, 0) ∨ (0, 1]) = µ([−1, 1]) = 1, so this function on regular open sets does not even satisfy the basic dimension function axiom D(N ∨ O) + D(N ∧ O) = D(N ) + D(O). (3.24) Even using the Lebesgue measure for µ instead would not help, as there are Cantor- like nowhere dense subsets of [−1, 1] with non-zero Lebesgue measure, and their complements can be expressed as the union of two disjoint regular open sets. Going up a dimension to [−1, 1] × [−1, 1], there even exists a pair of connected disjoint regular open sets whose complement is an Osgood curve (see [Osg03]), i.e. a Jordan curve of non-zero Lebesgue measure. Proposition 3.89 also illustrates the fact that taking direct limits, a common construction with C-algebras, is likely to produce a C-algebra that is infinite. This begs the question of whether there might exist other general constructions that can still produce C-algebras with the properties we want, but which remain finite. Indeed, do there exist any separable finite type II C-algebras at all? Surely ANNIHILATORS AND TYPE DECOMPOSITION IN C-ALGEBRAS 59 there must, and it is probably just a matter of examining enough examples through the lens of annihilators until one is found. Although there is the possibility that some new technique might be needed to create one or that somehow separability precludes their existence (for non-separable examples, we can of course just take any type II1 von Neumann algebra), which would be quite intriguing in either case. References [Ada69] D. H. Adams. The completion by cuts of an orthocomplemented modular lattice. Bull. Austral. Math. Soc., 1:279 -- 280, 1969. [AE02] Charles A. Akemann and Søren Eilers. Regularity of projections revisited. J. Operator Theory, 48(3, suppl.):515 -- 534, 2002. [Ake69] Charles A. Akemann. The general Stone-Weierstrass problem. J. Functional Analysis, 4:277 -- 294, 1969. [Ake70] Charles A. Akemann. Left ideal structure of C-algebras. J. Functional Analysis, 6:305 -- [Arz13] Farkhad Arzikulov. On 317, 1970. doi:10.1016/0022-1236(70)90063-7. classification the problem for C-algebras, 2013. [Ber72] arXiv:1002.4711v3. Sterling K. Berberian. Baer -rings. Springer-Verlag, New York, 1972. Die Grundlehren der mathematischen Wissenschaften, Band 195. [BH82] Bruce Blackadar and David Handelman. Dimension functions and traces on C ∗-algebras. J. Funct. Anal., 45(3):297 -- 340, 1982. doi:10.1016/0022-1236(82)90009-X. [Bic11] Tristan Bice. Filters in C-algebras. Canad. J. Math., 2011. doi:10.4153/CJM-2011-095-4. [Bic12] Tristan Bice. The order on projections in C ∗-algebras of real rank zero. Bull. Polish Acad. Sci. Math., 60(1):37 -- 58, 2012. doi:10.4064/ba60-1-4. [Bic13] Tristan Bice. The projection calculus. Munster J. Math., 2013. URL: http://wwwmath.uni-muenster.de/mjm/acc/Bice.pdf. [Bir67] Garrett Birkhoff. Lattice theory. Third edition. American Mathematical Society Collo- quium Publications, Vol. XXV. American Mathematical Society, Providence, R.I., 1967. [Bla94] Bruce Blackadar. Projections in C-algebras. In C-algebras: 1943 -- 1993 (San Anto- nio, TX, 1993), volume 167 of Contemp. Math., pages 130 -- 149. Amer. Math. Soc., Providence, RI, 1994. doi:10.1090/conm/167/1292013. [Bla06] B. Blackadar. Operator algebras, volume 122 of Encyclopaedia of Mathematical Sci- ences. Springer-Verlag, Berlin, 2006. Theory of C-algebras and von Neumann algebras, Operator Algebras and Non-commutative Geometry, III. [Che91] G. Chevalier. Around the relative center property in orthomodular lattices. Proc. Amer. [CK12] [CP79] Math. Soc., 112(4):935 -- 948, 1991. doi:10.2307/2048637. Ivan Chajda and Jan Kuhr. A generalization of effect algebras and ortholattices. Math. Slovaca, 62(6):1045 -- 1062, 2012. doi:10.2478/s12175-012-0063-4. Joachim Cuntz and Gert Pedersen. Equivalence and traces on C-algebras. J. Funct. Anal., 33(2):135 -- 164, 1979. doi:10.1016/0022-1236(79)90108-3. [Cun77] Joachim Cuntz. The structure of multiplication and addition in simple C-algebras. Math. Scand., 40:215 -- 233, 1977. URL: http://www.mscand.dk/article.php?id=2360. [Dou66] R. G. Douglas. On majorization, factorization, and range inclusion of operators on [Eff63] [Fla82] Hilbert space. Proc. Amer. Math. Soc., 17:413 -- 415, 1966. Edward G. Effros. Order ideals in a C-algebra and its dual. Duke Math. J., 30:391 -- 411, 1963. Jurgen Flachsmeyer. Note on orthocomplemented posets. II. In Proceedings of the 10th Winter School on Abstract Analysis (Srni, 1982), number Suppl. 2, pages 67 -- 74, 1982. URL: https://eudml.org/doc/221665. [FP10] David J. Foulis and Sylvia Pulmannov´a. Type-decomposition of an effect algebra. Found. Phys., 40(9-10):1543 -- 1565, 2010. doi:10.1007/s10701-009-9344-3. [FP13] David J. Foulis and Sylvia Pulmannov´a. Dimension theory for generalized effect algebras. Algebra Universalis, 69(4):357 -- 386, 2013. doi:10.1007/s00012-013-0237-0. 60 [GJ60] TRISTAN BICE Leonard Gillman and Meyer Jerison. Rings of continuous functions. The University Series in Higher Mathematics. D. Van Nostrand Co., Inc., Princeton, N.J.-Toronto- London-New York, 1960. [GW05] K. R. Goodearl and F. Wehrung. The complete dimension theory of partially ordered systems with equivalence and orthogonality. Mem. Amer. Math. Soc., 176(831):vii+117, 2005. doi:10.1090/memo/0831. [Jac85] Nathan Jacobson. Basic algebra. I. W. H. Freeman and Company, New York, second edition, 1985. [Kap51] [Kap55] [Kal83] Gudrun Kalmbach. Orthomodular lattices, volume 18 of London Mathematical Soci- ety Monographs. Academic Press Inc. [Harcourt Brace Jovanovich Publishers], London, 1983. Irving Kaplansky. Projections in Banach algebras. Ann. of Math. (2), 53:235 -- 249, 1951. Irving Kaplansky. Any orthocomplemented complete modular lattice is a continuous geometry. Ann. of Math. (2), 61:524 -- 541, 1955. B.R. Li. An Introduction to Operator Algebras. World Scientific Publishing Company Incorporated, 1992. [Li92] [Loo55] L. H. Loomis. The lattice theoretic background of the dimension theory of operator algebras. Mem. Amer. Math. Soc., 1955(18):36, 1955. [Lor97] Terry A. Loring. Lifting solutions to perturbing problems in C ∗-algebras, volume 8 of Fields Institute Monographs. American Mathematical Society, Providence, RI, 1997. [Mac64] M. Donald MacLaren. Atomic orthocomplemented lattices. Pacific J. Math., 14:597 -- 612, 1964. [MvN36] F. J. Murray and J. v. Neumann. On rings of operators. Ann. of Math. (2), 37(1):116 -- 229, 1936. doi:10.2307/1968693. [OrT11] Eduard Ortega, Mikael Rørdam, semigroup and comparison of open projections. J.Funct. Anal., 260(12):3474 -- 3493, 2011. doi:10.1016/j.jfa.2011.02.017 . and Hannes Thiel. The Cuntz [Osg03] William F. Osgood. A Jordan curve of positive area. Trans. Amer. Math. Soc., 4(1):107 -- 112, 1903. doi:10.2307/1986455. [Ped79] Gert K. Pedersen. C ∗-algebras and their automorphism groups, volume 14 of London Mathematical Society Monographs. Academic Press Inc. [Harcourt Brace Jovanovich Publishers], London, 1979. [PZ00] Costel Peligrad and L´aszl´o Zsid´o. Open projections of C-algebras: comparison and regularity. In Operator theoretical methods (Timi¸soara, 1998), pages 285 -- 300. Theta Found., Bucharest, 2000. Iain Raeburn and Dana P. Williams. Morita equivalence and continuous-trace C- algebras, volume 60 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 1998. [RW98] [Tom60] Minoru Tomita. Spectral theory of operator algebras. URL: [vN30] [vN40] [vN60] Math. J. Okayama Univ., 9:63 -- 98, I. http://www.math.okayama-u.ac.jp/mjou/mjou1-46/mjou_pdf/mjou_09/mjou_09_063.pdf. J. v. Neumann. Allgemeine Eigenwerttheorie Hermitescher Funktionaloperatoren. Math. Ann., 102(1):49 -- 131, 1930. doi:10.1007/BF01782338 . J. v. Neumann. On rings of operators. III. Ann. of Math. (2), 41:94 -- 161, 1940. URL: http://www.jstor.org/stable/1968823. John von Neumann. Continuous geometry. Foreword by Israel Halperin. Princeton Mathematical Series, No. 25. Princeton University Press, Princeton, N.J., 1960. 1959/1960. Federal University of Santa Catarina, Florianopolis, Brazil E-mail address: [email protected]
1807.03230	1	1807	2018-07-09T15:29:01	AF-embeddings of residually finite dimensional C*-algebras	[ "math.OA" ]	It is shown that a separable exact residually finite dimensional C-algebra with locally finitely generated (rational) even K-homology embeds in a uniformly hyperfinite C-algebra.	math.OA	math	AF-EMBEDDINGS OF RESIDUALLY FINITE DIMENSIONAL C-ALGEBRAS MARIUS DADARLAT Abstract. It is shown that a separable exact residually finite dimensional C-algebra with locally finitely generated (rational) K 0-homology embeds in a uniformly hyperfinite C ∗-algebra. 1. Introduction Kirchberg proved that any separable exact C-algebra embeds in the Cuntz algebra O2, see [6]. A related major open problem asks if any separable exact and quasidiagonal C-algebra embeds in an almost finite dimensional algebra (AF-algebra), see [1, Ch.8]. Most positive re- sults on AF-embeddability depend on the universal coefficient theorem in KK-theory (abbreviated UCT) [10], see for example [3], [8], [11]. A general result of Ozawa [7] shows that the cone over an exact separable C-algebra is AF-embeddable. Such cones are automatically quasidi- agonal by a theorem of Voiculescu [12]. While Ozawa's proof does not use the UCT explicitly, cones are contractible and in particular they do satisfy the UCT. Cones also play a key role in Rørdam's paper on purely infinite AH-algebras and AF-embeddings [9]. Indeed, Rørdam's C ∗-algebra A[0, 1], which he showed that contains the cone over O2 as a subalgebra, is itself an inductive limit of cones over matrix alge- bras and in particular it is KK-contractible. In a very recent paper [5], Gabe proves that a separable exact C ∗-algebra for which its primitive spectrum has no non-empty compact open subsets embeds in A[0, 1] and hence it is AF-embeddable. As far as I am aware, all previously known AF-embeddings results which do not assume the UCT factor through inductive limits of cones and in particular are not applicable to C ∗-algebras that contain nonzero projections. A C-algebra A is called residually finite-dimensional (abbreviated RFD) if the finite dimensional representations of A separate the points of A. We have shown in [3] that a separable, exact, RFD C-algebra M.D. was partially supported by NSF grant #DMS–1700086. 1 2 MARIUS DADARLAT which satisfies the UCT is AF-embeddable and in fact it even embeds in a UHF-algebra N∞ n=1 Mk(n)(C). In the present note we point out that the arguments of [3] can be adapted to obtain a UHF-embeddability re- sult for RFD C ∗-algebras which does not assume the UCT but requires (local) finite generation of the even K-homology, see Definition 2.1. Theorem 1.1. Let A be a separable exact residually finite dimensional C-algebra. If the rational K 0-homology of A is locally finitely gener- ated, then A embeds in a UHF-algebra. 2. Preliminaries Let A be a C ∗-algebra. A family D of C ∗-subalgebras of A is called exhaustive if for any finite subset F ⊂ A and any ε > 0 there exists D ∈ D such that F ⊂ε D, i.e. for each x ∈ F there is d ∈ D such that kx − dk < ε. Definition 2.1. We say that the K 0-homology of A is locally finitely generated if there is an exhaustive family D of C ∗-subalgebras of A such that for every D ∈ D the abelian group K 0(D) = KK(D, C) is finitely generated. If instead we require the weaker condition that each Q-vector space K 0(D) ⊗Z Q is finite dimensional, then we say that the rational K 0-homology of A is locally finitely generated. The case when the vector space K 0(A) ⊗Z Q is itself finite dimensional is an obvious first example. Let L(H) denote the linear operators acting on a separable Hilbert space H and let K(H) denote the compact operators. We make the identifications L(Ck) = K(Ck) ∼= Mk(C). The unitary group of Mk(C) is denoted by U(k). Let A be a C ∗-algebra, let F ⊂ A be a finite subset and let ε > 0. If ϕ : A → L(Hϕ) and ψ : A → L(Hψ) are two maps, we write ϕ ∼F ,ε ψ if there is a unitary v : Hϕ → Hψ such that kvϕ(a)v∗ − ψ(a)k < ε for all a ∈ F . If m is a positive integer and π is a representation, then mπ will denote the representation π ⊕ · · · ⊕ π (m-times). The infinite direct sum π ⊕ π ⊕ · · · is denoted by π∞. We need the following approximation result. Proposition 2.2 ([2, Prop.6.1]). Let A be a unital separable exact C ∗-algebra and let (χn)n≥1 be a sequence of unital representations of A that separates the elements of A and such that each representation in the sequence repeats itself infinitely many times. For any F ⊂ A AF-EMBEDDINGS OF RESIDUALLY FINITE DIMENSIONAL C-ALGEBRAS 3 a finite subset and any ε > 0 there is an integer r ≥ 1 such that if π = χ1 ⊕ χ2 ⊕ · · · ⊕ χr, then for any unital faithful representation σ : A → L(H) with σ(A) ∩ K(H) = {0}, one has σ ∼F ,ε π∞. Definition 2.3. Let A be a unital RFD C ∗-algebra. Let F ⊂ A be a finite subset and let ε > 0. A unital representation π : A → Mk(C) is called (F , ε)-admissible if there is a unital faithful representation σ : A → L(H) with σ(A) ∩ K(H) = {0} (H = Ck ⊕ Ck ⊕ · · · ) such that (1) kσ(a) − π∞(a)k < ε, for all a ∈ F . Remark 2.4. Note that if π is (F , ε)-admissible, then so is π ⊕ α for any unital finite dimensional representation α. Moreover kπ(a)k > If a unital C ∗-algebra A is separable exact and kak − ε for a ∈ F . RFD, then Proposition 2.2 guaranties the existence of (F , ε)-admissible representations for any finite set F ⊂ A and any ε > 0. The following proposition is crucial for our embedding result. It is based on a uniqueness theorem from [4]. Proposition 2.5. Let A be a unital separable exact RFD C ∗-algebra. Let F ⊂ A be a finite subset and let ε > 0. Then for any (F , ε)- admissible representation π : A → Mk(C) and any two unital represen- tations ϕ, ψ : A → Mr(C), such that [ϕ] = [ψ] ∈ K 0(A), there exist a positive integer M and a unitary u ∈ U(r + M k) such that (2) ku(ϕ(a) ⊕ M π(a))u∗ − ψ(a) ⊕ M π(a)k < 3ε, for all a ∈ F . Proof. Fix F , ε and π. Let σ be a unital faithful representation of A given by Definition 2.3. In particular, σ satisfies (1) and σ(A)∩K(H) = {0}. Since [ϕ] = [ψ] ∈ K 0(A), it follows that if we set Φ = ϕ ⊕ σ and Ψ = ψ ⊕ σ, then Φ(a) − Ψ(a) is a compact operator for all a and the class of the Kasparov triple (Φ, Ψ, 1) in KK(A, C) = K 0(A) vanishes since [Φ, Ψ, 1] = [ϕ ⊕ σ, ψ ⊕ σ, 1r ⊕ 1H] = [ϕ] − [ψ] = 0. Moreover, both Φ and Ψ are faithful representations whose images do not contain nonzero compact operators. This enables us to apply [4, Thm. 3.12] and obtain that Φ is asymptotically unitarily equivalent to Ψ via a continuous path of unitaries which are compact perturbations of the identity. In particular, there is a unitary v ∈ U(Cr ⊕ H) of the form v = 1 + x with x ∈ K(Cr ⊕ H) such that kv(ϕ(a) ⊕ σ(a))v∗ − ψ(a) ⊕ σ(a)k < ε, for all a ∈ F . Using (1) we obtain that (3) kv(ϕ(a) ⊕ π∞(a))v∗ − ψ(a) ⊕ π∞(a)k < 3ε, for all a ∈ F . 4 MARIUS DADARLAT Since π is a unital representation, it follows that the sequence of pro- jections pn = 1r ⊕ nπ(1) forms an approximate unit of K(Cr ⊕ H) and hence [pn, v] = [pn, x] → 0 as n → ∞. Since each pn commutes with both ϕ(a) ⊕ π∞(a) and ψ(a) ⊕ π∞(a), we obtain from (3) that k(pnvpn)(ϕ(a) ⊕ nπ(a))(pnvpn)∗ − ψ(a) ⊕ nπ(a)k < 3ε, for all a ∈ F and all sufficiently large n. Moreover one can perturb the almost unitary operator pnvpn to a unitary u satisfying (2) for a sufficiently large value of n denoted M. 3. Proof of Theorem 1.1 Without any loss of generality, we may assume that A is unital. We denote by Repfd(A) the set of unital finite dimensional representa- tions of A. Since A is separable and RFD, there is a sequence (χn)n≥1 in Repfd(A) which separates the points of A and such that each rep- resentation in the sequence repeats itself infinitely many times. Let n=1 be a dense sequence of elements of A and let εn = 2−n. Since (xn)∞ K 0(A) ⊗Z Q is locally finitely generated, for each n ≥ 1 there is a uni- tal C ∗-subalgebra An of A such that the Q-vector space K 0(An) ⊗Z Q is finite dimensional and Xn := {x1, ..., xn} ⊂εn An. Fix a finite set Fn = {an,1, ..., an,n} ⊂ An such that kxi − an,ik < εn for all 1 ≤ i ≤ n. Since K 0(An) ⊗Z Q is finite dimensional, its subspace Vn generated by all the classes {[χiAn] ⊗ 1 : i ≥ 1} must also be finite dimen- sional. Thus there is an integer rn such that Vn is generated by just {[χiAn] ⊗ 1 : 1 ≤ i ≤ rn}. Define πn ∈ Repfd(A) by πn = χ1 ⊕χ2 ⊕...⊕χrn . By Proposition 2.2, after increasing rn, if necessary, we can moreover arrange that πnAn is (Fn, εn)-admissible. With these choices, we are going to construct a sequence of unital representations γn : A → Mkn(C) such that for all n ≥ 1: (i) kn+1 = knmn for some positive integer mn, (ii) γn is unitarily equivalent to πn ⊕ αn for some αn ∈ Repfd(A), (iii) kγn+1(x) − mnγn(x)k < 5εn, for all x ∈ Xn. We will see that in fact each γn is unitarily equivalent to a representa- tion of the form q1χ1 ⊕ q2χ2 ⊕ ... ⊕ qrnχrn, for integers qi ≥ 0. Set γ1 = π1 ⊕ χ1. Suppose now that αi and γi were constructed for all i ≤ n such that the properties (i), (ii) and (iii) are satisfied. We construct αn+1 and γn+1 as follows. AF-EMBEDDINGS OF RESIDUALLY FINITE DIMENSIONAL C-ALGEBRAS 5 We need the following elementary observation. Suppose that G is an abelian group such that the vector space G ⊗Z Q is finite dimensional and it is spanned by g1 ⊗ 1, ..., gr ⊗ 1 with gi ∈ G. Then for any g ∈ G there are strictly positive integers p, m, q1, ..., qr, such that pg + q1g1 + · · · + qrgr = m(g1 + · · · + gr). By applying this observation to the abelian subgroup of K 0(An) generated by {[χiAn] : i ≥ 1} with gi = [χiAn], i = 1, ..., rn, one obtains strictly positive integers p, m, q1, ..., qrn, such that p[πn+1An] + rn X i=1 qi[χiAn] = m rn X i=1 [χiAn], in K 0(An). Set α′ n+1 = (p − 1)πn+1 ⊕ (cid:0) Lrn i=1 qiχi(cid:1) ⊕ mαn. Then [(πn+1 ⊕ α′ n+1)An] = rn X m[χiAn] + m[αnAn] = m[πnAn] ⊕ m[αnAn], i=1 and hence [(πn+1 ⊕ α′ n+1)An] = m[γnAn], using (ii). In particular, the representations πn+1 ⊕ α′ n+1 and mγn have the same dimension. Since πnAn is (Fn, εn)-admissible, so is γnAn as noted in Remark 2.4. By Proposition 2.5 applied to An, there is an integer M ≥ 1 such that (πn+1 ⊕ α′ n+1)An ⊕ M γnAn ∼Fn,3εn mγnAn ⊕ M γnAn. Set αn+1 = α′ n+1 ⊕ M γn, mn = m + M and γn+1 = πn+1 ⊕ αn+1. Then (4) γn+1An ∼Fn,3εn mnγnAn. Since kxi − an,ik < εn for all 1 ≤ i ≤ n, we deduce immediately from (4) that γn+1 ∼Xn,5εn mnγn. By conjugating γn+1 by a suitable unitary, we can arrange that kγn+1(x) − mnγn(x)k < 5εn, for all x ∈ Xn. Consider the UHF algebra B = lim −→ Mk(n)(C) and let ιn : Mk(n)(C) → B be the canonical inclusion. Having the sequence γn available, we construct a unital embedding γ : A → B by defining γ(x), x ∈ {x1, x2, . . . }, to be the limit of the Cauchy sequence (ιnγn(x))n≥1 and then extend to A by continuity. Note that γ is a ∗-homomorphism, since all γn are ∗-homomorphisms. Moreover, kγ(x)k = kxk for all x ∈ A since kγn(a)k > kak − εn for a ∈ Fn (by Remark 2.4) hence kγn(xi)k ≥ kxik − 3εn, as kan,i − xik < εn for 1 ≤ i ≤ n. Remark 3.1. It is clear from the proof that the conclusion of Theo- rem 1.1 holds under the weaker assumption that A admits a separating sequence of finite dimensional representations (χn)∞ n=1 such that for 6 MARIUS DADARLAT some exhaustive family D of C ∗-subalgebras of A, the vector subspace of K 0(D) ⊗Z Q spanned by {[χnD] ⊗ 1 : n ≥ 1} is finite dimensional for all D ∈ D. For example, this condition is satisfied if A is the suspen- sion of a separable exact RFD C ∗-algebra. Indeed, in that case one can choose a separating sequence (χn)∞ n=1 with the property that [χn] = 0 in K 0(A) for all n ≥ 1. Acknowledgements. The author would like to thank the referee for a close reading of the paper and for useful suggestions. References [1] N. P. Brown and N. Ozawa. C ∗-algebras and finite-dimensional approxima- tions, volume 88 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2008. [2] M. Dadarlat. Residually finite dimensional C ∗-algebras and subquotients of the CAR algebra. Math. Res. Lett., 8(4):545–555, 2001. [3] M. Dadarlat. On the topology of the Kasparov groups and its applications. J. Funct. Anal., 228(2):394–418, 2005. [4] M. Dadarlat and S. Eilers. Asymptotic unitary equivalence in KK-theory. K- Theory, 23(4):305–322, 2001. [5] J. Gabe. Traceless AF embeddings and unsuspended E-theory. arXiv:math/1804.08095 [math.OA], 2018. [6] E. Kirchberg. Exact C∗-algebras, tensor products, and the classification of purely infinite algebras. In Proceedings of the International Congress of Math- ematicians, Vol. 1, 2 (Zurich, 1994), pages 943–954, Basel, 1995. Birkhauser. [7] N. Ozawa. Homotopy invariance of AF-embeddability. Geom. Funct. Anal., 13(1):216–222, 2003. [8] N. Ozawa, M. Rørdam, and Y. Sato. Elementary amenable groups are quasidi- agonal. Geom. Funct. Anal., 25(1):307–316, 2015. [9] M. Rørdam. A purely infinite AH-algebra and an application to AF- embeddability. Israel J. Math., 141:61–82, 2004. [10] J. Rosenberg and C. Schochet. The Kunneth theorem and the universal coeffi- cient theorem for Kasparov's generalized K-functor. Duke Math. J., 55(2):431– 474, 1987. [11] A. Tikuisis, S. White, and W. Winter. Quasidiagonality of nuclear C ∗-algebras. Ann. of Math. (2), 185(1):229–284, 2017. [12] D. Voiculescu. A note on quasidiagonal C ∗-algebras and homotopy. Duke Math. J., 62(2):267–271, 1991. MD: Department of Mathematics, Purdue University, West Lafayette, IN 47907, USA E-mail address: [email protected] URL: http://www.math.purdue.edu/~mdd/
1506.03130	2	1506	2015-07-30T15:14:43	Poisson Processes in Free Probability	[ "math.OA" ]	We prove a multidimensional Poisson limit theorem in free probability, and define joint free Poisson distributions in a non-commutative probability space. We define (compound) free Poisson process explicitly, similar to the definitions of (compound) Poisson processes in classical probability. We proved that the sum of finitely many freely independent compound free Poisson processes is a compound free Poisson processes. We give a step by step procedure for constructing a (compound) free Poisson process. A Karhunen-Loeve expansion theorem for centered free Poisson processes is proved. We generalize free Poisson processes to a notion of free Poisson random measures (which is slightly different from the previously defined ones in free probability, but more like an analogue of classical Poisson random measures). Then we develop the integration theory of real-valued functions with respect to a free Poisson random measure, generalizing the classical integration theory to the free probability case. We find that the integral of a function (in certain spaces of functions) with respect to a free Poisson random measure has a compound free Poisson distribution. For centered free Poisson random measures, we can get a simpler and more beautiful integration theory.	math.OA	math	POISSON PROCESSES IN FREE PROBABILITY GUIMEI AN AND MINGCHU GAO Abstract. We prove a multidimensional Poisson limit theorem in free probability, and define joint free Poisson distributions in a non-commutative probability space. We define (compound) free Poisson process explicitly, similar to the definitions of (compound) Poisson processes in clas- sical probability. We proved that the sum of finitely many freely independent compound free Poisson processes is a compound free Poisson processes. We give a step by step procedure for con- structing a (compound) free Poisson process. A Karhunen-Loeve expansion theorem for centered free Poisson processes is proved. We generalize free Poisson processes to a notion of free Poisson random measures (which is slightly different from the previously defined ones in free probability, but more like an analogue of classical Poisson random measures). Then we develop the integration theory of real-valued functions with respect to a free Poisson random measure, generalizing the classical integration theory to the free probability case. We find that the integral of a function (in certain spaces of functions) with respect to a free Poisson random measure has a compound free Poisson distribution. For centered free Poisson random measures, we can get a simpler and more beautiful integration theory. Key Words. Free Probability, Free Poisson Processes, Integration with respect to free Poisson random measures. 2010 MSC 46L54 Introduction The theory of stochastic processes is a very important branch in classical probability with wide applications in engineering and finance ([DJ] and [TK]). In free probability theory, stochastic processes have been studied since 1990’s. The most popular and important stochastic process in classical probability is Brownian motion (the Wiener process). The counterpart of Brownian motion in free probability is the free Brownian motion. The free Brownian motion and stochastic analysis with respect to the free Brownian motion have been studied thoroughly ([PBi], [BS1], [BS2] etc.). Anshelevich [MA1] developed an integration theory of bi-processes with respect to (additive) non- commutative stochastic measures. Free infinite divisibility and free Levy processes and stochastic integration with respect to a free Levy process were studies in [BnT]. Certain stochastic differential equations driven by free Levy processes were studied in [MG1] and [MG2]. It is well known that Poisson distributions form a class of the most prominent processes in classical probability beyond normal distributions (Lecture 12 in [NS]), and free Poisson processes form a class of the most important processes with free increment in free probability after free Brownian motion ([MA]). But free Poisson distributions and processes have not been investigated thoroughly. In this paper, we study some interesting questions on free Poisson distributions and free Poisson processes. A free Poisson Limit Theorem. The counterpart of normal distributions in free probabil- ity is semicircle distributions. There is a semicircle limit theorem called free central limit theo- rem(Theorem 8.10 in [NS]). Very similarly, a free Poisson distribution can be realized as the limit in distribution of a sequence of simple distributions (Proposition 12.11, Definition 12.12 in [NS]). Nica and Speicher presented a multidimensional central limit theorem (Theorem 8.17 in [NS]). Roughly speaking, the theorem states that a joint semicircle distribution can be realized as the limit in distribution of a sequence In this paper, we proved a multidimensional free Poisson limit of families of random variables. 1 theorem (Theorem 2.4). Therefore, a joint free Poisson distribution can be defined as the limit in distribution of certain sequence of families of elements (Definition 2.6). Free Poisson processes. A construction of free Poisson process with all free cumulants equal to 1 was given in Section 4.2 in [MA], but no definition of free Poisson processes was give there. Anshelevich gave a description of free Poisson processes as “A process with stationary freely in- dependent increments such that the increments have free Poisson distributions is the free Poisson process” (4.2 in [MA2]). In this paper, we give a definition of free Poisson process (Definition 3.1), an analogue of a classical Poisson process. We provide a step-by-step procedure for constructing a free Poisson process (Theorem 3.2). Nica and Speicher gave the definition of compound free Poisson distributions in 12.16 of [NS]. We generalize free Poisson processes to the compound case (Definition 3.5), and give a similar procedure for constructing a compound free Poisson process (Theorem 3.6). In classical probability, the sum of two independent Poisson processes is a Poisson process (Section 2.3 in [RG]). We prove in this paper that the sum of finitely many freely independent compound free Poisson processes is a compound free Poisson process (Theorem 3.7), and conditions under which the sum of two freely independent free Poisson processes is a free Poisson process are given (Corollary 3.8). The Karhunen-Loeve expansion of a stochastic process is a significant result in classical stochastic processes ([DJ]). Roughly speaking, the expansion says that under certain conditions, a stochastic process can be represented as an infinite series of the products of random variables and deterministic functions Xt = Xiφi(t), 0 < t ≤ T, ∞Xi=1 where Xi, i = 1, 2, · · · , are uncorrelated random variables (E(XiXj) = δi,jλi), and {φi : i = 1, 2, · · · } is an orthonormal basis of L2([0, T ]), T > 0 (Theorem 5.3 or [AA]). In this paper, we present a Karhunen-Loeve expansion for a centered L2-continuous free Poisson process in a W ∗- probability space (A, ϕ) with precise formulas for φi(t) and λi (Theorem 4.5). Integration with respect to a free Poisson random measure. Stochastic integration with respect to a non-commutative stochastic measure was studied by several mathematicians. Anshelevich [MA1] defined a non-commutative stochastic measure as follows. Definition 0.1 (Definition 1 in [MA1]). A non-commutative stochastic measure is a map from the set of all finite half-open intervals I = [a, b) ⊂ [0, ∞) to the self-adjoint part of a W ∗-probability space (A, ϕ), I 7→ X(I), with three properties. (1) Additivity. I1 ∩ I2 = ∅, I1 ∪ I2 = J ⇒ X(I1) + X(I2) = X(J). (2) Stationary. The distribution of X(I) dependents only on I. (3) Free increments. If I1, I2, · · · , In are mutually disjoint intervals, then are freely independent. X(I1), X(I2), · · · , X(In) Then Anshelevich [MA1] defined the integral of a bi-process U in A ⊗ Aop with respect to a non-commutative stochastic measure ([MA1]). Glockner, Schurmann, and Speicher [GSS] gave a definition in a ∗-probability space similar to the above Definition 0.1, and named it a free white noise. Barndorff-Nielson and Thorjornsen [BnT] defined free Poisson random measures in a more general setting. Definition 0.2 (Definition 6.7 in [BnT]). Let (Θ, E, ν) be a measure space, and E0 = {E ∈ E : ν(E) < ∞}. A free Poisson random measure is a map M from E0 into the cone of all non-negative operators of a W ∗-probability space (A, ϕ) with the following properties. (1) ∀E ∈ E0, M (E) has a free Poisson distribution κn(M (E)) = ν(E), n = 1, 2, · · · . 2 (2) If E1, E2, · · · , En are mutually disjoint sets in E0, then M (E1), M (E2), · · · , M (En) are freely independent. (3) If E1, E2, · · · , En are mutually disjoint sets in E0, then M (∪n i=1Ei) =Pn i=1 M (Ei). The authors of [BnT] also gave an existence theorem for free Poisson random measures (Theorem 6.9 in [BnT]), and defined the integral of a L1(Θ, ν) function with respect to a free Poisson random measure (Definition 6.19 in [BnT]). A definition of free Poisson random measures, very similar to Definition 0.2 above, was given in [BP]. The authors of [BP] studied multiple integrals of a special kind of functions with respect to a free Poisson random measure, and proved a semicircle limit theorem for free Poisson multiple integrals (Theorem 4.1 in [BP]). In this paper, we define free Poisson random measures via a sightly different way from the others mentioned above in a W ∗-probability space. We do not require that operators XE, for E ⊂ R of finite measure, be non-negative, but self-adjoint only (Definition 5.1). Our definition of free Poisson random measures is more like an analogue to that in classical probability theory (Section 9.3 in [TK]). We define the integral X(f ) of a function f ∈ L1(R) ∩ L2(R) with respect to a free Poisson random measure (Theorem 5.4). We prove a limit and free stochastic integration exchange formula n→∞ZR lim fn(t)dXE(t) =ZR lim n→∞ fn(t)dXE(t) R(R) → L1(A, ϕ) is contractive (Theorem 5.7), where L1 If XE ≥ 0, for every E ⊂ R of finite measure, then the integration operator (Theorem 5.5). X : L1 R(R) is the space of all real-valued L1- functions on R. When we focus on L∞− = ∩n≥1Ln(R), we find that the integral X(f ) of f ∈ L∞− has a compound free Poisson distribution (Theorem 5.9). For a centered free Poisson random measure (Definition 6.1), the integration operator X is an isometry from L2(R) into L2(A, ϕ) (Lemma 6.3). The integration operator X with respect to a centered free Poisson random measure can be extended to a bounded operator from L1(R) into L1(A, ϕ) with norm less than or equal to 2 (Lemma 6.4). 1. Preliminaries In this section we recall some basic concepts and results in free probability used in sequel or mentioned previously. The reader is referred to [NS] and [VDN] for the basics on free probability, and to [KR] for operator algebras. Non-commutative Probability spaces. A non-commutative probability space is a pair (A, ϕ) consisting of a unital algebra A and a unital linear functional ϕ on A. When A is a ∗-unital algebra, ϕ should be positive, i. e. ϕ(a∗a) ≥ 0, ∀a ∈ A. A C∗-probability space (A, ϕ) consists of a unital C∗-algebra and a state ϕ on A. A W ∗-probability space (A, ϕ) consists of a finite von Neumann algebra A and a faithful normal tracial state ϕ on A. An element a ∈ A is called a (non- commutative) random variable. ϕ(an) is called the n-th moment of a, n = 1, 2, · · · . Let C[X] be the complex algebra of all polynomials of an indeterminate X. The linear function µa : C[X] → C, µa(P (X)) = ϕ(P (a)), ∀P ∈ C[X], is called the distribution (or law) of a. A sequence {an} of random variables an ∈ (An, ϕn) converges in distribution to a ∈ (A, ϕ) if lim n→∞ ϕn(am n ) = ϕ(am), ∀m ≥ 1. Joint Distributions. Let ChX1, X2, · · · , Xsi be the unital algebra freely generated by s non-commutative indeterminates X1, X2, · · · , Xs, and a1, a2, · · · , as ∈ A, where (A, ϕ) is a non- commutative probability space. The family {ϕ(ai1 ai2 · · · ain) : 1 ≤ i1 ≤ i2 ≤ · · · ≤ in ≤ s, n ≥ 1} is called the family of joint moments of a1, a2, · · · , as. The linear functional µ : ChX1, X2, · · · , Xsi → C defined by µ(P ) = ϕ(P (a1, a2, · · · , as)), ∀P ∈ ChX1, · · · , Xsi, 3 is called the joint distribution of a1, a2, · · · , as. Similar to the single variable case, we can define the limit in distribution of a sequence of families of random variables. Free independence. A family {Ai : i ∈ I} of unital subalgebras of a non-commutative probabil- ity space (A, ϕ) is freely independent (or free) if ϕ(a1a2 · · · an) = 0 whenever the following conditions are met: ai ∈ Al(i), ϕ(ai) = 0 for i = 1, 2, · · · , n, and l(i) 6= l(i + 1), for i = 1, 2, · · · , n − 1. A family {ai : i ∈ I} of elements is free if the unital subalgebras generated by ai’s are free. Non-crossing partitions. Given a natural number m ≥ 1, let [m] = {1, 2, · · · , m}. A partition π of [m] is a collection of non-empty disjoint subsets of [m] such that the union of all subsets in π is [m]. A partition π = {B1, B2, · · · , Br} of [m] is non-crossing if one cannot find two block Bi and Bj of π, and four numbers p1, p2 ∈ Bi, q1, q2 ∈ Bj such that p1 < q1 < p2 < q2. The collection N C(m), the number of non-crossing of all non-crossing partitions of [m] is denoted by N C(m). partitions of [m], is Cm = (2m)! m!(m+1)! , which is called the m-th Catalan number (Notation 2.9 in [NS]). The Mobius function. Let P be a finite partial ordered set (poset), and P (2) = {(π, σ) : π, σ ∈ P, π ≤ σ}. For two functions F, G : P (2) → C, we define the convolution F ∗ G by F ∗ G(π, σ) := Xρ∈P,π≤ρ≤σ F (π, ρ)G(ρ, σ). Let δ(π, σ) = 1, if π = σ; δ(π, σ) = 0, if π < σ. Then F ∗ δ(π, σ) = Xρ∈P,π≤ρ≤σ F (π, ρ)δ(ρ, σ) = F (π, σ), ∀F. It follows that δ is the unit of set of all functions on P (2) with respect to convolution ∗. The inverse function of the function ζ : P (2) → C, ζ(π, σ) = 1, ∀(π, σ) ∈ P (2), with respect to the convolution ∗ is called the Mobius function µP of P . Free Cumulants Let π, σ ∈ N C(n). We say π ≤ σ if each block (a subset of [n]) of π is completely contained in one of the blocks of σ. N C(n) is a poset by this partial order. The Mobius function of N C(n) is denoted by µn. The unital linear functional ϕ : A → C produces a sequence of multilinear functionals ϕn : An → C, ϕn(a1, a2, · · · , an) = ϕ(a1a2 · · · an), n = 1, 2, · · · . Let V = {i1, i2, · · · , is} ⊆ [n]. We define ϕV (a1, a2, · · · , an) = ϕ(ai1 ai2 · · · ais ). More generally, for i=1 ϕVi (a1, a2, · · · , an). a partition π = {V1, V2, · · · , Vr} ∈ N C(n), we define ϕπ(a1, a2, · · · , an) =Qr The n-th free cumulant of (A, ϕ) is the multilinear functional κn : An → C defined by κn(a1, a2, · · · , an) = Xπ∈N C(n) ϕπ(a1, a2, · · · , an)µn(π, 1n), where 1n = [n] is the single-block partition of [n]. Free cumulants κn : An → C and free independence have a very beautiful relation. Theorem 1.1 (Theorem 11.20 in [NS]). A family {ai : i ∈ I} of elements in (A, ϕ) is freely independent if and only if for all n ≥ 2 and all i(1), i(2), · · · , i(n) ∈ I, κn(ai(1)ai(2) · · · ai(n)) = 0 whenever there exist 1 ≤ l, k ≤ n with i(l) 6= i(k). Therefore, if a and b are freely independent, then κn(a + b) = κn(a + b, a + b, · · · , a + b) = kn(a) + kn(b). Semicircle elements. Let (A, ϕ) be a ∗-probability space. A self-adjoint element a ∈ A is a semicircle element (or has a semicircle distribution) if ϕ(an) = 2 πr2 Z r −r tnpr2 − t2dt, n = 1, 2, · · · , 4 where r is called the radius of a. When r = 2, ϕ(a2) = 1, we say a a standard semicircle element (or has a standard semicircle distribution). A semicircle element can be characterized by ϕ(a2k) = (r2/4)kCk, where Ck is the k-th Catalan number, and ϕ(a2k+1) = 0, k = 0, 1, 2, · · · , or by free cumulants κn(a) = δn,2 r2 4 ((11.13) in [NS]). 2. Multidimensional free Poisson distributions By the discussion in Page 203 and Exercise 12.22 of [NS], a classical Poisson distribution is the limit in distribution of a sequence of convolutions of Bernoulli distributions. In the point of view of random variables, we can restate it as follows. Let λ > 0, α ∈ R. For each N ∈ N, N > λ, let {bi,N : i = 1, 2, · · · , N } be a sequence of i.i.d. Bernoulli random variables such that Then the binomial random variable SN =PN i=1 bi,N has a binomial distribution P r(bi,N = 0) = 1 − λ N , P r(bi,N = α) = λ N . P r(SN = kα) = Ck N ( λ N )k(1 − λ N )N −k, k = 0, 1, 2, · · · , N , where Ck by elementary calculus, we can get N is the combination number (or the binomial coefficient). Let N → ∞, lim N→∞ P r(SN = kα) = λk k! e−λ = P r(P = kα), where P has a Poisson distribution P r(P = kα) = λk k! e−λ, k = 0, 1, 2, · · · . In non-commutative case, the free Poisson limit theorem (Proposition 12.11 in [NS]) says that a free Poisson distribution is the limit in distribution of a sequence of free convolutions of Bernoulli distributions. We want to restate it in the language of random variables. Let’s define Bernoulli random variables in a non-commutative probability space. Let (A, ϕ) be a non-commutative probability space. A Bernoulli random variable a ∈ A is a linear combination a = αp + β(1 − p), where α, β ∈ R, and p ∈ A is an idempotent (p2 = p) with 0 ≤ ϕ(p) ≤ 1. The classical interpretation of a Bernoulli random variable is that a is a random variable with two “values”: α and β, and P r(a = α) = ϕ(p), P r(a = β) = 1 − ϕ(p). In the free Poisson limit theorem, β = 0, ϕ(p) = λ N , N > λ. We can restate the free Poisson limit theorem as follows. Let λ > 0, α ∈ R. For N ∈ N, N > λ, let {αp1,N , αp2,N , · · · , αpN,N } be a free family of Bernoulli random variables such that ϕ(pi,N ) = λ i=1 αpi,N . Then N , i = 1, 2, · · · , N . Let SN =PN κm(SN ) = λαm, m = 1, 2, · · · . lim N→∞ Hence, we may restate the definition of free Poisson random variables as follows. Definition 2.1 (Proposition 12.11, Definition 12.12 [NS]). Let λ ≥ 0, α ∈ R, and (A, ϕ) a non- commutative probability space. A random variable a ∈ A has a free Poisson distribution if the free cumulants of a are κn(a) = λαn, ∀n ∈ N. In this section, we want to generalize the results on free Poisson distributions in Lecture 12 of [NS] to the multidimensional case. By the proof of Theorem 13.1 in [NS], we can modify the theorem slightly to be the following form. Proposition 2.2 (Theorem 13.1 and Lemma 13.2 in [NS]). Let {nk} be a sequence of natural numbers such that limk→∞ nk = ∞, and, for each natural number k, (Ak, ϕk) be a non-commutative probability space. Let I be an index set. Consider a triangular array of random variables, i. e., for each i ∈ I, 0 ≤ r ≤ nk, we have a random variable a(i) nk,r ∈ Ak. Assume that, for each k, the sets {a(i) (nk,nk)}i∈I are free and identically distributed. Then the following statements are equivalent. (nk,2)}i∈I , · · · , {a(i) (nk,1)}i∈I , {a(i) 5 (1) There is a family of random variables (bi)i∈I in some non-commutative probability space nk,nk )i∈I converges in distribution to (bi)i∈I , as nk,2 + · · · + a(i) nk,1 + a(i) (A, ϕ) such that (a(i) k → ∞. (2) For all n ≥ 1, and all i(1), i(2), · · · , i(n) ∈ I, the limits limk→∞ nkϕk(a(i(1)) nk,r · · · a(i(n)) nk,r ) exist, 1 ≤ r ≤ nk. (3) For all n ≥ 1, and all i(1), i(2), · · · , i(n) ∈ I, the limits limk→∞ nkκk n(a(i(1)) nk,r · · · a(i(n)) nk,r ) exist, 1 ≤ r ≤ nk, where κk n is the n-th free cumulant functional in Ak. Furthermore, if one of these conditions is satisfied, then the limits in (2) are equal to the corre- sponding limits in (3), and the joint distribution of the limit family (bi)i∈I is determined in terms of free cumulants by (n ≥ 1, i(1), i(2), · · · , i(n) ∈ I) κn(bi(1)bi(2) · · · bi(n)) = lim k→∞ nkϕk(a(i(1)) nk,r a(i(2)) nk,r · · · a(i(n)) nk,r ). We will use the following elementary result in sequel. Lemma 2.3. Let {ai,j : i, j = 1, 2, · · · } be a bi-index sequence of complex numbers. If sup{ai,j : i = 1, 2, · · · } = Mj < ∞, ∀j, then there exists a sequence (nk)k∈N of natural numbers such that limk→∞ nk = ∞, and limk→∞ an(k),j exists,∀j ∈ N. Proof. Since {ai,1 : i = 1, 2, · · · } is bounded, there is a sequence {i(k, 1) : k = 1, 2, · · · } of natural numbers such that limk→∞ ai(k,1),1 = a1, for some number a1. Consider the sequence {ai(k,1),2 : k = 1, 2, · · · }. Since the sequence is bounded, there is a subsequence {i(k, 2) : k = 1, 2, · · · } of {i(k, 1) : k = 1, 2, · · · } such that limk→∞ ai(k,2),2 = a2. But we also have limk→∞ ai(k,2),1 = a1. Continuing the process, we can obtain a bi-index sequence {i(k, l) : k, l = 1, 2, · · · } of natural numbers such that limk→∞ ai(k,l),j = aj, for j ≤ l. Let nk = i(k, k), for k = 1, 2, · · · . Then, for a j, and an ǫ > 0, there there exists a natural number K > j such that ai(k,j),j − aj < ǫ, ∀k > K. Note that {i(n, k); n = 1, 2, · · · } is a subsequence of {i(n, j) : n = 1, 2, · · · }. Thus, i(k, k) ≥ i(k, j), and ai(k,k) − aj < ǫ, ∀k > K. It means that limk→∞ ank,j = aj, ∀j. (cid:3) Theorem 2.4. Let {αi : i = 1, 2, · · · } be a sequence of real numbers, {λi ≥ 0}i∈N with λ = sup{λi : i ≥ 1} < ∞, and for each N ∈ N, N > λ, there be N freely independent and identically distributed sequences {p(i) 1,N }i∈N, {p(i) 2,N }i∈N, · · · , {p(i) N,N}i∈N j,N p(i(1)) of commutative projections on a C∗-probability space (AN , ϕN ), i. e., p(i(1)) j,N , ∀i(1), i(2) = 1, 2, · · · . Moreover, ϕN (p(i) N , i = 1, 2, · · · , r = 1, 2, · · · , N . Define a triangular family of sequences of random variables {a(i) j,N : i = 1, 2, · · · }, for j = 1, 2, · · · , N, N = 1, 2, · · · . Then there exists a family of random variables (bi)i∈N in a non-commutative probability space (A, ϕ) and a sequence {nk : k = 1, 2, · · · } of natural numbers such that limk→∞ nk = ∞ and (a(i) 1,nk nk,nk )i∈N converges to (bi)i∈N in distribution, as k → ∞. j,N = p(i(2)) j,N = αip(i) j,N p(i(2)) + · · · + a(i) r,N ) = λi + a(i) 2,nk Proof. For N, i(1), i(2), · · · , i(n), n ∈ N, let r,N ), 1 ≤ r ≤ N, and M (i(1), i(2), · · · , i(n)) = αi(1)αI(2) · · · αi(n) min{λi(j) : j = 1, 2, · · · , n}. Then f (N, i(1), i(2), · · · , i(n)) = N ϕN (a(i(1)) r,N · · · a(i(n)) r,N a(i(2)) f (N, i(1), i(2), · · · , i(n)) = N αi(1)αi(2) · · · αi(n)ϕN (p(i(1)) r,N p(i(2)) r,N · · · p(i(n)) r,N ) ≤N αi(1)αi(2) · · · αi(n) min{ϕN (p(i(j)) r,N ) : j = 1, 2, · · · n} =N αi(1)αi(2) · · · αi(n) min{ =M (i(1), i(2), · · · , i(n)), λi(j) N : j = 1, 2, · · · , n} 6 since ϕN : AN → C is positive, and p(i(1)) projections. r,N p(i(2)) r,N · · · p(i(n)) r,N ≤ min{p(i(j)) : j = 1, 2, · · · , n}, as Let Sm = {(i(1), i(2), · · · , i(n)) : i(1) + i(2) + · · · + i(n) = m, i(1), i(2), · · · , i(n) ∈ N}, for m ∈ N. Then for each m ∈ N, Sm is a finite set with Sm = km, and {Sm : m ∈ N} is a partition of the set {(i(1), i(2), · · · , i(n) : i(1), i(2), · · · , i(n), n ∈ N}. Define a bijective map γ : S1 → {1}, γ : Sm → {( m−1Xl=1 kl) + 1, ( m−1Xl=1 kl) + 2, · · · , mXl=1 kl}, m ≥ 2. For instance, γ((1, 1)) = 2, γ(2) = 3, γ(S2) = {2, 3}. It implies that γ({(i(1), i(2), · · · , i(n)) : i(1), i(2), · · · , i(n), n ∈ N}) =γ(S1) ∪ γ(S2) ∪ · · · ∪ γ(Sm) ∪ · · · ={j : j = 1, 2, 3, · · · }. Thus, {f (N, i(1), i(2), · · · , i(n)) : N, i(1), i(2), · · · , i(n), n ∈ N} = {f (N, γ−1(j)) : N, j ∈ N} is a bi-index sequence. By Lemma 2.3, there is a sequence (nk)k∈N of natural numbers such that limk→∞ nk = ∞, and f (nk, i(1), i(2), · · · , i(n)) converges as k → ∞, for every tuple (i(1), i(2), · · · , i(n)). By Theorem 2.2, there is a family of random variables (bi)i∈N in a non-commutative probability space (A, ϕ) such that ((a(i) (cid:3) 1,nk nk,nk )i∈N converges to (bi)i∈N in distribution. + · · · + a(i) + a(i) 2,nk Remark 2.5. (1) By Theorems 2.2 and 2.4, For each i ∈ N, κn(bi) = lim k→∞ (nkϕnk ((a(i) r,nk )n) = αn i λi, n = 1, 2, · · · . Hence, bi has a free Poisson distribution, for each i ∈ N. (2) If {p(i) r,N }i∈N is an orthogonal sequence of projections, for ∀N, r = 1, 2, · · · , N , then κn(bi(1)bi(2) · · · bi(n)) = lim k→∞ nkϕnk (αi(1) · · · αi(n)p(i(1)) r,nk · · · p(i(n)) r,nk ) = 0, whenever there are i(j) 6= i(l), 0 ≤ j, l ≤ n. This means that {bi : i ∈ N} is a free family of free Poisson random variables. A similar procedure of constructing a free family from an orthogonal one can be found in Example 12.19 in [NS]. We, therefore, define multidimensional free Poisson distributions as follows. Definition 2.6. A family of random variables (bi)i∈N in a non-commutative probability space (A, ϕ) has a joint free Poisson distribution if the family has a joint distribution same as the limit distri- bution in Theorem 2.4. 3. Free Poisson Processes An analogue of classical Poisson processes in free probability can defined as follows. Definition 3.1. For k ∈ N and α ∈ R. A family {Xt : t ≥ 0} of self-adjoint elements in a ∗-non- commutative probability space (A, ϕ) is a free Poisson process if it satisfies the following conditions. (1) X0 = 0. (2) For 0 ≤ t1 < t2 < · · · < tn < ∞, Xtn − Xtn−1, · · · , Xt2 − Xt1 form a freely independent family. (3) For 0 ≤ s < t, Xt − Xs has a free Poisson distribution with parameters λ = k(t − s) and α, that is, κn(Xt − Xs) = k(t − s)αn, n = 1, 2, · · · . (The most common case is that k = 1.) Construction We give a procedure for constructing a free Poisson process in a ∗-non-commutative probability space, for a real number α. 7 (1) For each natural number N , let t 7→ pt,N be a projection-values process [0, N ] → A, where (A, ϕ) be a W ∗-probability space. That is, for 0 ≤ t1 < t2 < · · · < tn ≤ N , {pt2,N − pt1,N , pt3,N − pt2,N , · · · , ptn,N − ptn−1,N } is an orthogonal family of projections, and ϕ(pt,N ) = t N ), for 0 ≤ s ≤ t ≤ N in the projection-valued process in Section 4.2 of [MA]. N , ∀0 ≤ t ≤ N . Actually, we can get such a process by letting pt − ps = p[ s N , t (2) Choose N free copies of the process in Step 1. That is, processes {pi t,N ) = t 1, 2, · · · , N , are N free families of random variables, and ϕ(pi t,N : 0 ≤ t ≤ N }, i = N i = 1, 2, · · · , N . t,N . Then the limit in distribution of at,N , as N → ∞, has free (3) Let at,N = αPN i=1 pi cumulants κn(at) = tαn, ∀n ≥ 1, by Proposition 12.11 in [NS]. (4) By Exercise 16.21 or Theorem 21.7 in [NS], there is a family {at : t ≥ 0} in A such that κn(at) = tαn, ∀n ≥ 1 (If necessary, we can expand A so that A contains all limit elements {at, t ≥ 0}). Now we show that {at : t ≥ 0} is a free Poisson process with parameter α. Theorem 3.2. The process {at : t ≥ 0} constructed via the above procedure is a free Poisson process in (A, ϕ). Proof. For 0 ≤ s < t, choose N > t, and consider at,N − as,N = αPN of Proposition 12.11 in [NS], the n-th free cumulant of at,N − as,N is (t − s)αn + O( 1 i=1(pi t,N − pi s,N ). By the proof N ). Thus, κn(at − as) = lim N→∞ ((t − s)αn + O( 1 N )) = (t − s)αn, ∀n ≥ 1. Moreover, by Remark 2.5, for 0 ≤ t1 < t2 < · · · < tn < ∞, atn − atn−1, · · · at2 − at1 form a freely independent family. (cid:3) Remark 3.3. (1) In Section 4.2 of [MA], Anshelevich constructed a free Poisson process as follows. For a projection-valued process I 7→ pI , from half-open intervals I ⊂ [0, 1] into projections in a W ∗-probability space(A, ϕ), and a standard semicircle element s ∈ A, which is free from {pI : I ⊂ [0, 1]}, I 7→ spI s is a free Poisson process with κn(spI s) = I. Let I = [0, t). Then κn(spI s) = t (t > 0). This is a special case of free Poisson processes with α = 1. (2) When combining the construction in Remark 1.9 in [NS1] and that in Section 4.2 in [MA], we can get a free Poisson process I 7→ spIs for a general semicircle element s ∈ A with radius r. The free Poisson process I 7→ spI s has n-th cumulant κn(spI s) = I r2n 4n . This is a free Poisson process with α = r2 4 > 0. (3) One can get a free Poisson process by our procedure for any α ∈ R. A generalized version of free Poisson distributions is the following compound free Poisson distri- butions. Definition 3.4 (Definition 12.16 in [NS]). Let ν be a compactly supported probability measure on R, and λ ≥ 0. A probability measure µ on R is called a compound free Poisson distribution with rate λ and jump distribution ν if the n-th free cumulant of µ is κn(µ) = λmn(ν), where mn(ν) is the n-th moment of measure ν, and n ≥ 1. Now we generalize the notion of free Poisson processes to a compound version. Definition 3.5. Let ν be a compactly supported probability measure on R and k ∈ N. A family {aI : I = [s, t) ⊂ [0, ∞)} of self-adjoint elements in a ∗-non-commutative probability space (A, ϕ) is a compound free Poisson process with respect to ν, if it satisfies the following conditions. (1) aI has a compound free Poisson distribution: κn(aI ) = kImn(ν), n ≥ 1. (2) If I1, I2, · · · , In are mutually disjoint half-open intervals, then aI1 , aI2, · · · , aIn form a freely independent family. 8 Construction For a compactly supported probability measure ν on R, the construction for a compound free Poisson process is very similar to that for a free Poisson process. (1) For each natural number N , choose a projection-valued process I 7→ pI,N , from half-open N , as we did intervals I ⊂ [0, N ] into a W ∗-probability space (A, ϕ) such that ϕ(pI,N ) = I in the previous construction for a free Poisson process. Let pt,N = p[0,t),N . (2) For a half-open interval I = [s, t), where 0 ≤ s < t, choose a natural number N > t, and a I,N ) = mn(ν), where AI = pI,N ApI,N , I,N ) = ϕ(pI,N ) , for x ∈ AI , and mn(ν) is the n-th moment of measure ν. Then ϕ(an self-adjoint element aI,N in (AI , ϕI ) such that ϕI (an ϕI (x) = ϕ(x) ϕ(pI,N )mn(ν) = t−s N mn(ν), n ≥ 1. (3) As the Step 2 in the previous construction, for each natural number N , choose N processes N mn(ν), n ≥ 1, i = {a(i) 1, 2, · · · , N . Also, the N processes are freely independent from each other. I,N : I = [s, t) ⊂ [0, N ]}, i = 1, 2, · · · , N , that is, ϕ((a(i) I,N )n) = I (4) Let bI,N = PN i=1 a(i) [NS], limN→∞ κn(bI,N ) = Imn(ν), n ≥ 1. I,N , I ⊂ [0, N ], and N = 1, 2, · · · . By the proof of Proposition 12.11 in (5) Let {bI : I = [s, t), 0 ≤ s < t < ∞} be the family of random variables with the distribution κn(bt) = (t − s)mn(ν), n ≥ 1. Theorem 3.6. {bI : I = [s, t), 0 ≤ s < t} constructed above is a compound free Poisson process. Proof. By the construction, κn(bI ) = Imn(ν), n ≥ 1, that is, bI has a compound free Poisson distribution. For a family {I1, I2, · · · , Ik} of mutually disjoint half-open intervals, choose a natural number Ij ,N ∈ AIj ,N = pIj ,N ApIj ,N , N such that Ij ⊂ [0, N ], j = 1, 2, · · · , k. For every 0 < r ≤ N , a(r) j = 1, 2, · · · , N . Thus, Qk j=1 a(r) Ij ,N = 0. By Theorem 13.1 in [NS], κl(bIi(1) , bIi(2) , · · · bIi(l) ) = lim N→∞ N ϕ( lYj=1 a(r) Ii(j),N ) = 0, if there are two disjoint intervals in {Ii(1), Ii(2), · · · , Ii(l)} ⊆ {I1, I2, · · · , Ik}, ∀1 < l ≤ k. It follows from Theorem 11.20 in [NS] that bI1 , bI2, · · · , bIk form a freely independent family. (cid:3) In classical probability, the sum of two independent Poisson processes is still a Poisson process (see Section 2.3 in [RG]). In free probability, we have a slightly different result. Theorem 3.7. Let {bI,i : I = [s, t) ⊂ [0, ∞)}, i = 1, 2, · · · , k, be a freely independent family of k compound free Poisson processes such that κn(bI,i) = Imn(νi), where νi is a compactly supported i=1 bI,i is a compound free Poisson probability measure on R, i = 1, 2, · · · , k. Then the sum bI =Pk process with distribution κn(bI ) = kImn(ν), n ≥ 1, where ν =Pk Proof. For a half-open interval I = [s, t), operators bI,1, bI,2, · · · , bI,k are freely independent. There- fore, νi k . i=1 κn(bI ) = kXi=1 κn(bI,i) = kXi=1 Imn(νi) = kI kXi=1 mn(νi) k = kImn(ν), n ≥ 1. For mutually disjoint intervals I1, I2, · · · , Il of [0, ∞) and n ≤ l, we have κn(bI1, bI2 , · · · , bIn ) = nX i1,i2,··· ,in=1;ip6=ip′ ,∀p6=p′ kXj1,j2,··· ,jn=1 κn(bIi1 ,j1 , bIi2 ,j2 , · · · , bIin ,jn ). It implies that bI1 , bI2, · · · , bIl are freely independent. (cid:3) 9 Corollary 3.8. Let {at,1 : t ≥ 0} and {at,2 : t ≥ 0} be two freely independent free Poisson processes with distributions κn(at,i) = tαn i , n ≥ 1, i = 1, 2. Then {at = at,1 + at,2 : t ≥ 0} is compound free Poisson process. Moreover, the sum process {at : t ≥ 0} of two non-zero free Poisson processes (that is, (α1)2 + (α2)2 6= 0) is a free Poisson process if and only if α1 = α2. Proof. The first conclusion that the sum of two freely independent free Poisson processes is a compound free Poisson process follows from the previous result, Theorem 3.7. More precisely, the sum process has a distribution κn(at) = 2tmn(ν), n ≥ 1, where ν = 1 2 (δα1 + δα2), and δα is the point mass distribution at number α (see Example 3.4.1 in [VDN]). Note that a compound free Poisson distribution is a free Poisson distribution if and only if the probability measure ν in the definition of compound free Poisson distributions is a point mass distribution δα. If α1 = α2, then δα1 + δα2 = 2δα, where α = α1 = α2. Therefore, κn(at,1 + at,2) = t(αn 1 + αn 2 ) = 2tαn, ∀t ≥ 0, n ≥ 1. It follows that the sum process is a free Poisson process. Suppose now that δα1 + δα2 = 2δα, for some α ∈ R, we will show that α1 = α2. Suppose α1 6= α2. If α1 > α2, then for t > 0 and all n ∈ N, we have κn(at,1 + at,2) = t(αn 1 + αn 2 ) = 2tαn ⇒ 1 + ( α2 α1 )n = 2( α α1 )n. Let n → ∞, we have the limit of the left side sequence is 1. On the other hand, if α > α1, the limit of the right side sequence is ±∞; if α < α1, the limit of the right side sequence is 0; if α = −α1, the limit of the right side sequence (−1)n does not exist. All these cases lead to a contradiction. It implies that α = α1. Thus, limn→∞ 2( α )n = 2. This contradicts to the fact that α1 the limit on the left is 1. Hence, the assumption α1 > α2 is wrong. Very similarly, we can prove that the assumptions α1 < α2 and α1 = −α2 lead contradictions. Therefore, if the sum is a free Poisson process, then α1 = α2. (cid:3) Remark 3.9. (1) A similar result about the sum of free Poisson distributions can be found in Exercise 12.25 in [NS]. (2) Combining the construction after the definition of free Poisson processes and the above corollary, we can construct a free Poisson process {at : t ≥ 0} such that κn(at) = ktαn, n ≥ 1, for k ∈ N and α ∈ R. 4. The Karhunen-Loeve expansion of a free Poisson Process 4.1. The Karhunen-Loeve expansion of a classical stochastic process. First let’s recall the Karhunen-Loeve Expansion of a classical stochastic process (Sections 2.3.6, 2.3.7 in [DJ] and [AA]). Definition 4.1 ([PB]). Let T ∈ R, T > 0. A function K(s, t) : [0, T ] × [0, T ] → R is called a kernel if (1) K is symmetric, that is, K(s, t) = K(t, s), ∀0 ≤ s, t ≤ T , and (2) K is non-negative definite, nXi,j=1 cicjK(ti, tj) ≥ 0, ∀t1, t2, · · · , tn ∈ [0, T ], c1, c2, · · · , cn ∈ R. nel. Then the integral operator TK : L2([0, T ]) → L2([0, T ]), TK(f )(x) = R T Theorem 4.2 (Mercer’s theorem, [PB]). Suppose K : [0, T ] × [0, T ] → R is a continuous ker- 0 K(x, y)f (y)dy, ∀f ∈ L2([0, T ]) is non-negative definite, and there is an orthonormal basis {φi : i = 1, 2, · · · } of L2([0, T ]) consisting of eigenfunctions of TK such that the corresponding eigenvalues {λi} are non-negative. 10 The eigenfunctions corresponding to non-zero eigenvalues are continuous on [0, T ] and the kernel K has the form K(s, t) = λiφi(s)φi(t), ∞Xi=1 where the convergence is absolute and uniform, that is, lim n→∞ sup K(s, t) − s,t∈[0,T ] nXi=1 λiφi(s)φi(t) = 0. Theorem 4.3 (Karhunen-Loeve, [AA]). Let Xt be a centered (i. e. E(Xt) = 0), mean square continuous (limt→t0 E((Xt − Xt0)2) = 0) stochastic process on a probability space (Ω, F , µ) with Xt ∈ L2(Ω, [0, T ]). Then there is an orthonormal basis {φi} of L2([0, T ]) such that for all t ∈ [0, T ], Xt = ∞Xi=1 Xiφi(t), where Xi = R T 1, 2, · · · . 0 Xtφi(t)dt, the convergence is in L2(Ω), and E(Xi) = 0, E(XiXj) = δi,jλi, i, j = 4.2. The non-commutative stochastic process case. Let {X(t) : t ≥ 0} be a free Poisson process in a W ∗-probability space (A, ϕ) with α ∈ R, k = 1 (See Definition 3.1). Suppose that the process is continuous with respect to k · k2 in A (which is called L2-continuous). For T > 0, let L2(0, T ; A) be the space of all L2-continuous non-commutative stochastic processes of self-adjoint operators in (A, ϕ). It is obvious that (s, t) 7→ ϕ(XsXt) is continuous on [0, T ] × [0, T ]. Thus, we 0 ϕ(X(t)Y (t))dt in L2(0, T ; A). We get a Hilbert space can define an inner product hX, Y i = R T L2(0, T ; A). Lemma 4.4. Let Xt be a free Poisson Process in a W ∗-probability space (A, ϕ) with distribution κn(Xt) = tαn, n = 1, 2, · · · , t ≥ 0, α ∈ R. Suppose Xt ∈ L2(0, T ; A). Then the second free cumulant k(s, t) := κ2(Xs, Xt) is a continuous kernel on [0, T ]. Proof. Suppose 0 < s ≤ t ≤ T . Then k(s, t) = κ2(Xs, Xt) = κ2(Xs, Xt − Xs + Xs) = κ2(Xs, Xs) = sα2 = k(t, s). Moreover, by Example 11.6 in [NS], k(s, t) = κ2(Xs, Xt) = ϕ(XsXt) − ϕ(Xs)ϕ(Xt) = ϕ((Xs − ϕ(Xs))(Xt − ϕ(Xt))) = ϕ(eXs eXt), where eXt = Xt − ϕ(Xt). nXi=1 It implies that for 0 < t1, t2, · · · , tn ≤ T and c1, c2, · · · , cn ∈ R, we have nXi,j=1 nXi,j=1 cicjk(ti, tj). 0 ≤ ϕ(( cicjϕ(eXti eXtj ) = nXi=1 ci eXti)( ci eXti)) = That is, k is non-negative definite. Note that the continuity of t 7→ Xt implies the continuity of t 7→ eXt. Thus, we can assume that Xt is centered, therefore, k(s, t) = ϕ(XsXt). It implies that k(s, t) : [0, T ] × [0, T ] → R is continuous. (cid:3) We are in the position to present the Karhunen-Loeve expansion of a free Poisson Process. Theorem 4.5. Let(Xt)t≥0 be an L2-continuous free Poisson process in a W ∗-probability space (A, ϕ) with distribution κn(Xt) = tαn, t ≥ 0, n = 1, 2, · · · , for some α ∈ R. Let eXt = Xt − tαI(t ≥ 0), λn = Then we have the Karhunen-Loeve expansion of eXt. 11 α2T 2 (n − 1/2)2π2 , φn(t) = ( 2 T )1/2 sin( (n − 1/2)πt T ), n ≥ 1. (1) ∞Xi=1 eXiφi(t), 0 < t ≤ T, 0 eXtφi(t)dt, the convergence is in L2(A, ϕ). eXt = where eXi =R T (2) ϕ(eXi eXj) = δi,jλi, ∀i, j. Proof. By Lemma 4.4 and Theorem 4.2, there is an orthonormal basis {φi(t) : i = 1, 2, · · · } of L2([0, T ]) and non-negative numbers λi, i = 1, 2, · · · . such that ∞Xi=1 λiφi(s)φi(t), ∀t, s ∈ [0, T ], where the convergence is absolute and uniform on [0, T ] × [0, T ]. Let’s prove the properties of Xi. For t ∈ [0, T ] and n ∈ N, we have ϕ(eXi) =R T 0 ϕ(eXs eXt) = ϕ(XsXt) − ϕ(Xs)ϕ(Xt) = k(s, t) = 0 ϕ(eXt)φi(t)dt = 0, and ϕ(eXi eXj) =Z T 0 Z T ϕ((eXt − ϕ(eXt eXs)φi(s)φj (t)dsdt =Z T 0 Z T nXi=1 ϕ(eXt eXi)φi(t) + (Z T nXi=1 nXi=1 nXi=1 eXiφi(t))2) = ϕ(eX 2 λiφi(t)2 + = k(t, t) − 2 = k(t, t) − 2 nXi=1 t ) − 2 0 nXi,j=1 k(s, t)φi(s)φj (t)dsdt = δi,jλi. ϕ(eXi eXj)φi(t)φj (t) ϕ(eXt eXs)φi(s)dsφi(t)) + 0 ϕ(eXi eXj)φi(t)φj (t) nXi,j=1 λiφi(t)2 → 0, as n → ∞, by Theorem 4.2. Moreover, the convergence is uniform on [0, T ]. By the proof of Lemma 5.4 and the example in Page 27 of [DJ], we can choose the eigenvalues and eigenfunctions as follows λn = α2T 2 (n − 1/2)2π2 , φn(t) = ( 2 T )1/2 sin( (n − 1/2)πt T ). (cid:3) 5. Integration with respect to free Poisson Random Measures A generalization of free Poisson processes is the following free Poisson random measures. Definition 5.1. Let α ∈ R. A free Poisson random measure is a map X from the set B0 of all Borel subsets with finite Lebesgue measure on the real line R into Asa, the space of all self-adjoint elements of a ∗ non-commutative probability space (A, ϕ), with the following properties. (1) XE has a free Poisson distribution with parameters α and E, for a set E ∈ B0, where E is the Lebesgue measure of E, that is, the n-th free cumulant κn(XE) = Eαn, n ≥ 1. (2) If E1, E2, · · · , Ek are mutually disjoint, then XE1, XE2, · · · , XEk are freely independent. (3) If E1, E2, · · · , Ek are mutually disjoint, then X∪n i=1 XEi. Remark 5.2. We restrict the map X to {[0, t) : 0 < t}, and let X0 = 0, then {0} ∪ {Xt = X[0,t) : t > 0} is a free Poisson process. Thus a free Poisson measure can be regarded as a generalization of a free Poisson process. i=1Ei =Pn For notational simplicity, we assume from now on that α = 1 in the definition of free Poisson measures. To study integration with respect to a free Poisson random measure, we need to deal with convergence problems. Thus, from now on, we assume that (A, ϕ) is a W ∗-probability space. 12 1 p , for A ∈ A. Let Lp(A, ϕ) be the completion of A with respect to the i=1 ciχEi be a simple function on R, where E1, E2, · · · , En are mutually disjoint Borel sets on R with finite measures, and c1, c2, · · · , cn are mutually distinct numbers. Define the integral of s with respect to XE as Define kAkp = (ϕ(Ap)) norm k · kp for p ≥ 1. Let s =Pk Then ϕ(X(s)) =Pn X(s) =ZR i=1 ciϕ(XEi ) =Pn nXi,j=1 ϕ(X ∗(s)X(s)) = cicjϕ(XEi XEj ) nXi=1 s(x)X(dx) := ciXEi. i=1 ciEi =RR s(x)dx, and cicjϕ(XEi XEj ) + ci2ϕ(X 2 Ei) =Xi6=j =Xi6=j =ZR nXi=1 nXi=1 cicjϕ(XEi )ϕ(XEj ) + ci2(κ2(XEi) + ϕ(XEi )2) s(x)dxZR s(x)dx +ZR s(x)2dx where supp(s) is the Lebesgue measure of the support set of function s. That is, ≤ksk2 1 + ksk2 2 ≤ (supp(s)2 + 1)ksk2 2, ϕ(X ∗(s)X(s)) ≤ ksk2 1 + ksk2 2 ≤ (supp(s)2 + 1)ksk2 2. (1) Hence, for a function f ∈ L2(R) with finite measure support Sf , we can define the integral of f with respect to a free Poisson random measure X as follows. Let {sn} be a sequence of simple functions such that ksn − f k2 → 0, as n → ∞, and the support of sn, for all n, are contained in Sf , the support of f . Then, we have kX(sn) − X(sm)k2 2 → 0, as both m and n approach ∞. It implies that we can define the integral as 2 ≤ (Sf 2 + 1)ksm − snk2 X(f ) :=ZR f (x)X(dx) = lim n→∞ X(sn) ∈ L2(A, ϕ), where the limit is taken in the norm k · k2 in A. It is obvious that the limit X(f ) is independent of the choice of sn. Moreover, for f ∈ L2(R) with Sf := supp(f ), Sf < ∞, let sn → f in k · k2, and supp(sn) ⊆ Ef . By Cauchy-Schwartz inequality, sn → f with respect to k · k1. It follows from (1) that ϕ(X(f )∗X(f )) = lim n→∞ ϕ(X(sn)∗X(sn)) ≤ lim n→∞ For a function f ∈ L2(R), we can define (ksnk2 1 + ksnk2 2) = kf k2 1 + kf k2 2. X(f, t) :=Z t −t f (x)X(dx), ∀t ≥ 0, X(f ) = lim t→∞ X(f, t), (2) (3) provided that the limit taken in the norm k · k2 in L2(A, ϕ) exists. By the above definition of integration, we have the following rudimental property. Proposition 5.3. Let f, g ∈ L2(R) such that X(f ) and X(g) exist. Then, for real numbers α, β, X(αf + βg) = αX(f ) + βX(g). Theorem 5.4. Let f ∈ L1(R) ∩ L2(R) be a real-valued function. Then X(f ) = limt→∞ X(f, t) exists. 13 Proof. For f ∈ L1(R) ∩ L2(R), let In = [−n, n], n = 1, 2, · · · . Then In ր R, and ZR f (x)dx = lim n→∞ZIn f (x)dx,ZR f (x)2dx = lim f (x)2dx. n→∞ZIn Thus, let fn(x) = f (x)χ[−n,n](x), n = 1, 2, · · · . Then kf − fnk1 + kf − fnk2 → 0, as n → ∞. By identity (2), we have kX(fn) − X(fm)k2 as n, m → ∞. Therefore, we have 2 = kX(fn − fm)k2 2 ≤ kfn − fmk2 1 + kfn − fmk2 2 → 0, X(f ) =ZR f (x)X(dx) = lim n→∞Z n −n f (x)X(dx) exists. (cid:3) Theorem 5.5. Let fn, n ≥ 1, f be real-valued functions in L1(R) ∩ L2(R) such that limn→∞(kfn − f k1 + kfn − f k2) = 0.Then limn→∞ kX(fn) − X(f )k2 = 0, that is, n→∞ZR lim fn(t)dXE(t) =ZR lim n→∞ fn(t)dXE(t). Proof. For every N ∈ N, we have (Z[−N,N ] lim n→∞ fn(t) − f (t) + fn(t) − f (t)2)dt ≤ lim n→∞ (ZR fn(t) − f (t) + fn(t) − f (t)2)dt = 0. For every ε > 0, there exists a number n0 ∈ N such that Let fn,N = χ[−N,N ](t)fn(t) and fN (t) = χ[−N,N ](t)f (t). By (2), we have kfn,N − fN k1 + kfn,N − fN k2 < ε, ∀n ≥ n0, N ∈ N. kX(fn,N ) − X(fN )kL2(A,ϕ) ≤ kfn,N − fN k1 + kfn,N − fN k2 < ε, ∀n ≥ n0, N ∈ N. Hence, for ε > 0, we have kX(fn) − X(f )kL2(A,ϕ) = lim N→∞ kX(fn,N ) − X(fN )kL2(A,ϕ) ≤ ε, ∀n ≥ n0. It follows that limn→∞ kX(fn) − X(f )kL2(A,ϕ) = 0. (cid:3) Remark 5.6. (1) Let a be a self-adjoint operator in a W ∗-probability space (A, ϕ) with a free Poisson distribution κn(a) = λ > 0, n = 1, 2, · · · . If 0 < λ < 1, by Remark 4.3 and Section 4.2 in [MA], we can choose a positive element sps having the same distribution, where s is a self-adjoint operator with the standard semicircle distribution, and p is a projection in A such that s and p are freely independent, and ϕ(p) = λ. Moreover, for a general λ > 0, by the construction in section 4.2 in [MA], we still can choose a positive operator a′ ∈ A such that a′ and a have the same distribution. (2) We know that if κm(a) > 0, ∀m ≥ 1, then ϕ(am) > 0, ∀m ≥ 1 ((11.6) and (11.8) in [NS]). But that having all positive moments doesn’t mean that a ∈ A is positive. For instance, let A = L∞([0, 1], dx), E1 = [0, 1/3], E2 = [2/3, 1], and f = 2χE1 − χE2 ∈ A is not non- negative. But R 1 0 f (x)ndx = 1/3(2n + (−1)n) > 0, ∀n = 1, 2, · · · . Theorem 5.7. If {XE : E ∈ B0} is a free Poisson random measure and XE ≥ 0, ∀E ∈ B0, then the integration with respect to the random measure is a contractive mapping from L1 R(R) into L1(A, ϕ), where L1 R(R) is the space of all real-valued L1-functions on R. i=1 ciχEi be a simple function with finite support ∪n i=1 Ei < ∞. i=1 En = Pn Proof. Let s = Pn By the definition, X(s) =Pn i=1 ciXEi. Thus, ϕ(X(s)) ≤ nXi=1 ciϕ(XEi ) = 14 nXi=1 ciEi = ksk1. (4) Let f ∈ L1 e. and in L1. Then, by (4), we define R(R), and choose a sequence {sn : n ≥ 1} of simple functions such that f = limn→∞ sn a. X(f ) =ZR f (x)X(dx) := lim n→∞ X(sn), where the limit is taken in L1(A, ϕ). Moreover, kX(f )k1 = ϕ(X(f )) = lim n→∞ ϕ(X(sn)) ≤ lim n→∞ ksnk1 = kf k1. (cid:3) When we consider real-valued functions in L1(R), the integration operator X : L1(R) → L1(A, ϕ) has the following monotonic property. Proposition 5.8. Suppose that XE ≥ 0, ∀E ∈ B0. If real-valued functions f ≤ g, and f, g ∈ L1(R), then X(f ) ≤ X(g) in L1(A, ϕ). To study the distribution of X(f ), we focus on real-valued functions in L∞− :=Tn≥1 Ln(R). T∞ n=1 Ln(A, ϕ), and X(f ) has a compound free Poisson distribution: Theorem 5.9. Let f ∈ L∞− be a real-valued function. If XE ≥ 0, ∀E ∈ B0, then X(f ) ∈ κm(X(f )) =ZR f (x)mdx, m = 1, 2, · · · . Proof. Suppose first that f ≥ 0. Then there exists a sequence {sn : n ≥ 1} of simple functions such that sn → f a. e., as n → ∞, and 0 ≤ sn(x) ≤ f (x) a. e.. By Lebesgue’s dominated convergence theorem, limn→∞ sn = f in Lm, for all m ≥ 1. It implies that sn(x)mdx =ZR On the other hand, for a simple function s =Pn n→∞ZR lim f (x)mdx, m = 1, 2, · · · . i=1 ciχEi , X(s) =Pn (5) i=1 ciXEi. It follows that κm(X(s)) = κm( ciXEi) = cm i κm(XEi) = sm(x)dx. (6) nXi=1 mXi=1 i Ei =ZR cm nXi=1 By (5) and (6), we have lim n→∞ κm(X(sn)) =ZR Moreover, for a real-valued simple function s =Pn f in Lm, for all m ≥ 1, we have RR sn1 − sn2m(x)dx → ∞, as n1, n2 → ∞, for all m ≥ 1. Let sn1 − sn2 =Pk i=1 ciχEi , s(x) =Pn f (x)mdx, m = 1, 2, · · · . i=1 ciχEi. Since limn→∞ sn = i=1 diχFi , we have (7) ZR sn1 − sn2m(x)dx = cimEi → 0, kXi=1 as n1, n2 → ∞, for all m ≥ 1. Therefore, by (6) we have κm(X(sn1 − sn2 )) = kXi=1 cimEi = ksn1 − sn2km m → 0, as, n1, n2 → ∞, for all m ≥ 1. By the moment-cumulant formals (11.7) and (11.8) in [NS], we have ϕ(X(sn1 ) − X(sn2 )m) → 0, 15 as n1, n2 → ∞, for all m ≥ 1. Note that 0 ≤ X(sn) ∈ A. Thus, there is a positive operator n=1 Ln(A, ϕ) such that ϕ(X(f )m) = limn→∞ ϕ(X(sn)m). It implies from (7) X(f ) ∈ L∞− := T∞ that κm(X(f )) = lim n→∞ κm(X(sn)) =ZR f (x)mdx, m = 1, 2, · · · . By Exercise 16.21 in [NS], there is a non-commutative probability space (D, ψ) and an element d ∈ D such that ψ(dm) =RR f (x)mdx, m = 1, 2, · · · . Then we have κm(X(f )) = ψ(dm), m = 1, 2, · · · , that is, X(f ) has a compound Poisson distribution with λ = 1 and measure ν, where mn(ν) = ψ(dn), n = 1, 2, · · · . For a general real-valued f ∈ L∞−, let f = f + − f −. Then f = f + + f −. We have X(f ) = n ) and n ) ⊂ supp(f −) are disjoint. X(f +) − X(f −) ∈ L∞−(A, ϕ). Let simple functions s+ X(s− It follows that n ) are freely independent, since supp(s+ n ) ⊂ supp(f +) and supp(s− n → f − a. e.. Then X(s+ n → f + and s− κm(X(f )) = κm(X(f +) − X(f −)) = lim n→∞ κm(X(s+ n ) − X(s− n )) κm(X(s+ n )) + (−1)m lim n→∞ κm(X(s− n )) f −(x)mdx =ZR f (x)mdx. It implies that X(f ) has a compound free Poisson distribution = lim n→∞ =ZR f +(x)mdx + (−1)mZR κm(X(f )) =ZR f (x)mdx, m = 1, 2, · · · . A random variable a in a non-commutative probability space (A, ϕ) is centered if ϕ(a) = 0. For 6. Centered Free Poisson Measures (cid:3) a random variable a ∈ A, ea := a − ϕ(a) is always centered. Definition 6.1. (1) If a random variable a ∈ (A, ϕ) has a free Poisson distribution κn(a) = and only if centered free Poisson random measure. λαn, n ≥ 1, then we say ea = a − ϕ(a) has a centered free Poisson distribution, that is, κn(ea) = λαn, n ≥ 2, and κ1(ea) = 0. (2) If XE, E ∈ B0, is a free Poisson random measure, we called eXE = XE − E, E ∈ B0, a Proposition 6.2. A random measure eXE, E ∈ B0, is a centered free Poisson random measure, if (1) eXE has a centered free Poisson distribution with parameter E, for a set E ∈ B0. (2) If E1, E2, · · · , Ek are mutually disjoint, then eXE1, eXE2, · · · , eXEk are freely independent. Proof. Suppose that eXE = XE − E. Then by Proposition 11.15 in [NS], for a1, a2, · · · , am ∈ A, m ≥ 2, we have κm(a1, a2, · · · , am) = 0, if there is at least one scalar element ai = αI, where α is a constant, and I is the unit in A. It follows that, for m ≥ 2, m ∈ N, is a free Poisson random measure. κm(eXE) = κm(XE − E, eXE, · · · , eXE) = κm(XE, eXE, · · · , eXE) = · · · = κm(XE) = E. It is obvious that {eXE1, · · · , eXEn} is a free family if E1, E2, · · · , En are mutually disjoint. Conversely, if eXE satisfies the two conditions, let XE = eXE + E. By the above discussion, XE Let s =Pn Lemma 6.3. keX(s)k2 i=1 ciχEi with Ei < ∞. Define eX(s) =RR s(x)eX(dx) :=Pn i=1 ci eXEi. 2 = ksk2 2. 16 (cid:3) Proof. keX(s)k2 cicjϕ(eXEi eXEj )) nXi,j=1 ci2(κ2(eXEi ) + ϕ(eXEi )2) 2 = ϕ(eX(s)∗ eX(s)) = ϕ( nXi=1 nXi=1 nXi=1 ci2ϕ(eX 2 ci2Ei = ksk2 2. Ei) = = = From the above lemma, we can extend the integration to L2(R). Let f be a real-valued functionin L2(R), and {sn : n ≥ 1} be a sequence of simple functions such that sn → f a. e. and in k · k2. (cid:3) Then define eX(f ) = limn→∞ eX(sn) ∈ L2(A, ϕ), where the limit is taken with respect to k · k2 of L2(A, ϕ). The operator eX : L2(R) → L2(A, ϕ) is isometric. Lemma 6.4. Let s = Pn keX(s)k1 ≤ 2ksk1. i=1 ciEi be a real-valued simple function, and XE ≥ 0, ∀E ∈ B0. Then Proof. keX(s)k1 = ϕ(eX(s)) ≤ nXi=1 ciϕ(XEi ) + nXi=1 ciEi = 2ksk1. (cid:3) Hence, we can extend the integration to L1(R). For a real-valued f ∈ L1(R), let sn be a real-valued simple functions such that sn → f a. e. and in L1(R). Then we define eX(f ) = limn→∞ eX(sn) ∈ L1(A, ϕ), where the limit is taken with respect to k · k1 of L1(A, ϕ). References [AA] A. Alexanderian. A brief note on the Karhunen-Loeve expansion. http://users.ices.utexas.edu/∼alen/articles/KL.pdf. [MA] M. Anshelevich. Free Stochastic measures via noncrossing partitions. Adv. Math. 155(2000), No.1, 154-179. [MA1] M. Anshelevich. Ito formula for free stochastic integrals. J. Funct. Anal., 188(2002), 292-315. [MA2] M. Abshelevich. Linearization coefficients for orthogonal polynomials using stochastic processes. Ann. of Probab., 33(2005), No.1, 114-136. [PB] P. Bartlett. Reproducing Kernel Hilbert spaces. http://www.cs.berkeley.edu/∼bartlett/courses/281b-sp08/7.pdf. [BnT] O. E. Barndorff-Nielson and S. Thorjornsen. Classical and free infinite divisibility and Levy processes. Lecture Note in Math. 1866 (2006), 33-160. Springer-Verlarg Berlin-Heidelberg. [PBi] P. Biane. Free Brownian motion, free stochastic calculus and random matrices. Free Probability Theory edited by D. Voiculescu, 1-19, Fields Institute Communications 12, AMS, 1997. [BS1] P. Biane and R. Speicher. Stochastic calculus with respect to free Brownian motion and analysis on Wigner spaces. Probab. Theory Related Fields 112(1998), no.3, 373-409. [BS2] P. Biane and R. Speicher.Diffusion, free entropy, and free Fisher information. Ann. Inst. H. Poincare Probab. Stat. 25 (2001), 581-606. [BP] S. Bourguin and G. Peccati. Semicircle limits on the free Poisson chaos: counterexample to a transfor principle. J. Funct. Anal., 267(2014), No. 4, 963-997. [RG] R.G. Gallager. Stachastic Processes: Theory for Applications. Cambridge University Press, 2013. [MG1] M. Gao. Free Markov processes and Stochastic differential equations in von Neumann algebras. Illinois J. Math. 52(2008), No. 1, 153-180. [MG2] M. Gao. Free Ornstein-Uhlenbeck processes. J. Math. Anal. Appl., 322(2006), 177-192. [GSS] P. Glockner, M. Schurmann and R. Speicher. Realization of free white noises. Arch. Math., Vol58(1992), 407-416. [DJ] D. Johnson. Statistical Signal Processing. http://elec531.blogs.rice.edu/files/2015/01/notes1.pdf. 17 [KR] R. Kadison and J. Ringrose. Fundamentals of the theory of operator algebras. Graduate Studies in MAth. Vol. 16, AMS, 1997. [TK] T. Kurts. Lecures on Stochastic Analysis. http://www.math.wisc.edu/∼kurtz/735/main735.pdf. [NS] A. Nica and R. Speicher. Lectures on Combinatorics for Free Probability, LMS Lecture Notes 335, Cambridge University Press, 2006. [NS1] A. Nica and R. Speicher. On the multiplication of free N -tuples of noncommutative random variables. Amer. J. of Math., 118(1996), 799-837. [VDN] D. Voiculescu, K. Dykema, and A. Nica. Free Random variables. CRM Monograph Series, Vol. 1, AMS, 1992. School of Mathematical Sciences and LPMC, Nankai University, Tianjin 300071, China E-mail address, Guimei An: [email protected] Department of Mathematics, Louisiana College, Pineville, LA 71359, USA E-mail address, Mingchu Gao: [email protected] 18
0810.0096	3	0810	2012-02-20T11:11:38	C*-Algebras over Topological Spaces: Filtrated K-Theory	[ "math.OA", "math.KT" ]	We define the filtrated K-theory of a C-algebra over a finite topological space X and explain how to construct a spectral sequence that computes the bivariant Kasparov theory over X in terms of filtrated K-theory. For finite spaces with totally ordered lattice of open subsets, this spectral sequence becomes an exact sequence as in the Universal Coefficient Theorem, with the same consequences for classification. We also exhibit an example where filtrated K-theory is not yet a complete invariant. We describe a space with four points and two C-algebras over this space in the bootstrap class that have isomorphic filtrated K-theory but are not KK(X)-equivalent. For this particular space, we enrich filtrated K-theory by another K-theory functor, so that there is again a Universal Coefficient Theorem. Thus the enriched filtrated K-theory is a complete invariant for purely infinite, stable C*-algebras with this particular spectrum and belonging to the appropriate bootstrap class.	math.OA	math	C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY RALF MEYER AND RYSZARD NEST Abstract. We define the filtrated K-theory of a C∗-algebra over a finite topo- logical space X and explain how to construct a spectral sequence that computes the bivariant Kasparov theory over X in terms of filtrated K-theory. For finite spaces with totally ordered lattice of open subsets, this spectral sequence becomes an exact sequence as in the Universal Coefficient Theorem, with the same consequences for classification. We also exhibit an example where filtrated K-theory is not yet a complete invariant. We describe two C∗-algebras over a space X with four points that have isomorphic filtrated K-theory without being KK(X)-equivalent. For this space X, we enrich filtrated K-theory by another K-theory functor to a com- plete invariant up to KK(X)-equivalence that satisfies a Universal Coefficient Theorem. 1. Introduction (1.1) 1.1. The UCT-problem. One of the main problems in the theory of C∗-algebras is the computation of the equivariant KK-theory of C∗-algebras endowed with some extra structure. Here we apply the general techniques developed in [6, 9] to the case of C∗-algebras with a non-trivial ideal lattice. The appropriate version of KK-theory is Kirchberg's generalisation of Kasparov theory to C∗-algebras over non-Hausdorff topological spaces (see [5]). Our goal is to compute it in terms of more manageable K-theoretic information, generalising the usual Universal Coefficient Theorem that computes Kasparov's original theory for C∗-algebras in the bootstrap class by an exact sequence Ext(cid:0)K∗+1(A), K∗(B)(cid:1) ֌ KK∗(A, B) ։ Hom(cid:0)K∗(A), K∗(B)(cid:1). The generalisation of the bootstrap class to the case of C∗-algebras with non- trivial ideal lattice was introduced and studied in [8]. Let us first recall some of the notation from [8]. Let X be a (usually non-Hausdorff) topological space. A C∗-algebra over X is a C∗-algebra A endowed with a continuous map Prim(A) → X. Let C∗alg(X) be the category of C∗-algebras over X; the morphisms in C∗alg(X) are given by X-equivariant (in obvious sense) ∗-homomorphisms. Taking Kirchberg's KK-groups as morphisms and the same objects, we get the category KK(X). It has a structure of a triangulated category (see [8]). For finite X, the bootstrap class B(X) is defined as the smallest subcategory of KK(X) that is closed under suspension, isomorphism, exact triangles, and direct sums and contains all objects with underlying C∗-algebra C. General methods from homological algebra suggest to study a homology the- ory H∗ for C∗-algebras over X, taking values in some Abelian category C. Under some mild assumptions, the machinery developed in [6, 9] yields an Adams type , ), with an E2-term expressed in terms spectral sequence which abuts to KK(X; of H∗. 2000 Mathematics Subject Classification. 19K35, 46L35, 46L80, 46M18, 46M20. The second author was supported by the German Research Foundation (Deutsche Forschungs- gemeinschaft (DFG)) through the Institutional Strategy of the University of Gottingen. 1 2 RALF MEYER AND RYSZARD NEST For classification purposes, we need, instead of a spectral sequence, a short exact sequence of the type (1.1): (1.2) ExtC(cid:0)H∗+1(A), H∗(B)(cid:1) ֌ KK∗(X; A, B) ։ HomC(cid:0)H∗(A), H∗(B)(cid:1), and a precise description of the range of H∗. In this case, given two C∗-algebras A and B over X that belong to the bootstrap class, an isomorphism of H∗(A) to H∗(B) lifts to a KK(X)-equivalence between A and B. The results of Eberhard Kirchberg then allow to lift this KK(X)-equivalence to a ∗-isomorphism A ∼= B, provided A and B are tight, purely infinite, stable, nuclear and separable; here tightness means that the maps Prim(A) → X and Prim(B) → X are homeomorphisms (see [5]). It is also shown in [8] that, in the case when X is finite, any object of the bootstrap class is KK(X)-equivalent to a tight, purely infinite, stable, nuclear, separable C∗-algebra over X. Hence the existence of an exact sequence of the form (1.2) for all objects of the bootstrap class leads to a complete classification of the tight, purely infinite, stable, nuclear, separable C∗-algebras over X in terms of their image under the functor H∗. 1.2. Main results. It is relatively easy to construct filtrations on KK which pro- duce spectral sequences which converge to KK-groups on the bootstrap category and whose E2-term involves only the K-theory of the quotients K∗(A/J) for the ideals J corresponding to minimal open subsets of X; an example is the filtration used in [8, Section 4.1]. However, this spectral sequence is not very useful for practical purposes, since it does not degenerate at the E2-level. The second dif- ferential involves, in particular, the K-theory of various subquotients I/J for the ideals I ⊂ J ⊂ A and the associated six-term exact sequences in K-theory (1.3) K0(cid:0)I(cid:1) K1(cid:0)J/I(cid:1) / K0(cid:0)J(cid:1) K1(cid:0)J(cid:1) / K0(cid:0)J/I(cid:1) K1(cid:0)I(cid:1). Also higher differentials do not vanish. To get a short exact sequence instead, we need to consider more sophisticated ho- mology theories. The homology theory analysed here is "filtrated K-theory," which is in some sense the second approximation to this spectral sequence. Roughly speak- ing, filtrated K-theory comprises the K-theory of various subquotients together with all canonical maps between these groups. We will make this definition precise later. The part of it which involves the exact sequences (1.3) appeared previously in the work of Gunnar Restorff [11] for Cuntz -- Krieger algebras and of Mikael Rørdam [13] and Alexander Bonkat [2] for extensions of C∗-algebras. The UCT theorem in the case when the ideal structure is given by I1 ⊳I2 ⊳A was obtained by Gunnar Restorff in his phd-thesis [12], where he introduced an invariant which is a particular case of filtrated K-theory. In this paper we prove the following Theorem 1.1. The filtrated K-theory satisfies the Universal Coefficient Theorem and is a complete invariant for C∗-algebras over those finite topological spaces with a totally ordered lattice of open subsets. Note that a C∗-algebra over a space of the type described in this result is essen- tially the same as a C∗-algebra A together with a finite increasing chain of ideals {0} = I0 ⊳ I1 ⊳ I2 ⊳ I3 ⊳ · · · ⊳ In−1 ⊳ In = A. We will also show that the spectral sequence associated to the filtrated K-theory does not collapse in general. Let (X, <) be the partially ordered set, where X = / / O O o o o o C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 3 {1, 2, 3, 4} with the partial order given by 1, 2, 3 < 4 and no further strict inequalities between 1, 2, 3. A C∗-algebra over this space is a C∗-algebra A together with an ideal I and a decomposition of A/I into a direct sum of three orthogonal ideals. Theorem 1.2. The filtrated K-theory over (X, <) does not satisfy the Universal Coefficient Theorem and is not a complete invariant. In fact, we give an explicit example of two C∗-algebras A and B over X in the bootstrap class that have isomorphic filtrated K-theory but are not KK(X)-equivalent. However, for the particular four-point space X, we still get a complete invari- ant and a Universal Coefficient Theorem as in (1.2), by adding another K-theory functor to filtrated K-theory. It is not clear how to construct such an enriched and still computable filtrated K-theory for general finite spaces. 1.3. The general machinery. Now we explain the general machinery behind our approach. Let us fix a finite topological space X. The first step is the correct definition of filtrated K-theory. The filtrated K-theory of a C∗-algebra A over X comprises the Z/2-graded Abelian groups K∗(cid:0)A(Y )(cid:1) for all locally closed subsets Y ⊆ X together with all natural transformations between these groups. The main issue here is to find all natural transformations. These natural transformations enter in the definition of the target category of the filtrated K-theory functor and thus influence the Hom and Ext terms that we expect in the Universal Coefficient Theorem. We can guess some of these natural transformations. If U is a relatively open subset of Y , then A(U ) is an ideal in A(Y ), with quotient A(Y )/A(U ) = A(Y \ U ). This C∗-algebra extension leads to a natural six-term exact sequence K0(cid:0)A(U )(cid:1) K1(cid:0)A(Y \ U )(cid:1) / K0(cid:0)A(Y )(cid:1) K1(cid:0)A(Y )(cid:1) / K0(cid:0)A(Y \ U )(cid:1) K1(cid:0)A(U )(cid:1). (1.4) These exact sequences provide three types of natural transformations associated to inclusions of open subsets, restriction to closed subset, and boundary maps. An obvious source for relations between these natural transformations are morph- isms of C∗-algebra extensions: since the six-term exact sequences in (1.4) are nat- ural, each natural morphism of extensions provides some commuting diagrams, which become relations between our generators. But do these obvious generators and relations already describe all natural trans- formations? This turns out to be the case for the spaces studied in this article -- both the positive and the negative examples. Although the authors know no counter- examples, we do not expect this to be so in general. The starting point for our study of filtrated K-theory is that the covariant func- for suitable C∗-algebras RY over X -- these are the representing objects. Our con- struction of RY yields commutative C∗-algebras, consisting of C0-functions on suit- able locally closed subspaces of the order complex of the partial order on X. The tors A 7→ K∗(cid:0)A(Y )(cid:1) are representable, that is, they are of the form KK∗(X; RY , A) Yoneda Lemma tells us that natural transformations from K∗(cid:0)A(Y )(cid:1) to K∗(cid:0)A(Z)(cid:1) correspond to KK∗(X; RZ , RY ) ∼= K∗(cid:0)RY (Z)(cid:1). These groups are easy enough to compute in the examples we consider, and turn out to be definable by the concrete generators and relations mentioned above. The natural transformations acting on filtrated K-theory form a Z/2-graded pre- additive category N T . A (countable) module over N T is, by definition, an additive / / O O o o o o 4 RALF MEYER AND RYSZARD NEST functor from N T to the category of (countable) Z/2-graded Abelian groups. By construction, the filtrated K-theory of any C∗-algebra over X is such a countable module. Let C be the category of countable N T -modules. This is an Abelian cat- egory, and filtrated K-theory is a stable homological functor FK from the Kasparov category KK(X) of C∗-algebras over X to C. It is easy to check that the functor FK : KK(X) → C is universal in the notation of [9]. General results on homological ideals in triangulated categories now pro- duce a cohomological spectral sequence that converges towards KK∗(X; A, B) if A belongs to the bootstrap class; its E2-term involves Extp The main issue is whether the Ext-groups Extp C(cid:0)FK(A), FK(B)(cid:1). C(cid:0)FK(A), FK(B)(cid:1) with p ≥ 2 vanish, so that our spectral sequence degenerates to an exact sequence of the desired form. This amounts to checking whether FK(A) has a projective resolution of length 1 in C. Already for the non-Hausdorff two-point space considered in [2, 13], the cat- egory C has infinite cohomological dimension, that is, there are objects that admit no projective resolution of finite length. But these objects do not belong to the range of the functor FK. If an N T -module A belongs to the range of FK, then there are exact sequences (1.5) · · · → A(U ) → A(Y ) → A(Y \ U ) → A(U ) → · · · for any Y ∈ LC(X), U ∈ LC(Y ) because of (1.4). But there are N T -modules without finite length projective resolutions. For totally ordered spaces, an object of C has a projective resolution of length 1 if and only if it has a projective resolution of finite length, if and only if the sequences (1.5) are exact, if and only if it is the filtrated K-theory of some separable C∗-algebra over X, which we can take in the bootstrap class. For the four-point counterexample considered in Section 5, we first find a torsion- free exact module that is not projective, and then use it to find an exact module without projective resolutions of length 1. Then we find two non-isomorphic objects of the bootstrap class with the same filtrated K-theory. The idea here is to consider a certain exact triangle ΣC → A → B → C, which splits on the level of filtrated K-theory, so that A ⊕ C and B have the same filtrated K-theory. But we can prove in our concrete example that A ⊕ C and B are not KK(X)-equivalent. A C∗-algebra over the four-point space X is a C∗-algebra A with a distinguished ideal I and a direct sum decomposition of A/I as a direct sum of three orthogonal ideals. Since both direct sums and extensions of C∗-algebras can be classified by filtrated K-theory, it is remarkable that the combination of both provides a counterexample. Incidentally, the space X op that corresponds to a C∗-algebra A with a distinguished ideal I and a direct sum decomposition of I as a direct sum of three orthogonal ideals also leads to a counterexample in a similar fashion. For the four-point space X above, there is essentially just one module that ought to be projective but is not. We can add another invariant to filtrated K-theory that corresponds to this offending module. Since this changes our whole category, it may lead to further offending modules, which would have to be added in a second step, and this could, in principle, go on forever. But in the concrete case at hand, we get projective resolutions of length 1 for all modules over the enriched filtrated K-theory. As a result, the enriched filtrated K-theory classifies objects of the bootstrap class over X up to KK(X)-equivalence, and it classifies purely infinite separable nuclear stable C∗-algebras with primitive ideal space X and simple subquotients in the bootstrap class. 1.4. Some basic notation. We shall use the following notation from [8]: ∈∈ we write x ∈∈ C for objects of a category C as opposed to morphisms; C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 5 X topological space, often assumed sober (see [14]); O(X) set of open subsets of X, partially ordered by ⊆; LC(X) set of locally closed subsets of X; LC(X)∗ set of connected, non-empty locally closed subsets of X; (cid:22) specialisation preorder on X, defined by x (cid:22) y ⇐⇒ {x} ⊆ {y} A C∗-algebra; Prim(A) primitive ideal space of A with hull -- kernel topology; I(A) set of closed ∗-ideals in A, partially ordered by ⊆; C∗alg(X) category of C∗-algebras over X with X-equivariant ∗-homomorphisms C∗sep(X) full subcategory of separable C∗-algebras over X; KK(X) Kasparov category of C∗-algebras over X: its objects are separable C∗-algebras over X, its set of morphisms from A to B is KK0(X; A, B); B(X) the bootstrap class in KK(X); Y extension functor C∗alg(Y ) → C∗alg(X) or KK(Y ) → KK(X) for a iX subset Y ⊆ X; ix abbreviation for iX X restriction functor C∗alg(X) → C∗alg(Y ) or KK(X) → KK(Y ) for a rY {x} for x ∈ X; locally closed subset Y ⊆ X; Σ suspension ΣA := C0(R, A). Roughly speaking, a space is sober if it can be recovered from the lattice O(X). It is explained in [8, §2.5] why we may restrict attention to such spaces. For finite spaces, sobriety is equivalent to the separation axiom T0, that is, two points are equal once they have the same closure. A C∗-algebra over X is pair (A, ψ) consisting of a C∗-algebra A and a continuous map ψ : Prim(A) → X. If X is sober, this is equivalent to a map ψ∗ : O(X) → I(A), U 7→ A(U ), that preserves finite infima and arbitrary suprema, that is, A(cid:18)\U∈F U(cid:19) = \U∈F A(U ), A(cid:18)[U∈S U(cid:19) = _U∈S A(U ) = XU∈S A(U ), where F ⊆ O(X) is finite and S ⊆ O(X) is arbitrary. In particular, this implies A(∅) = {0}, A(X) = A, and the monotonicity condition A(U ) ⊳ A(V ) for U ⊆ V . A ∗-homomorphism f : A → B between two C∗-algebras over X is X-equivariant if f(cid:0)A(U )(cid:1) ⊆ B(U ) for all U ∈ O(X). A subset Y ⊆ X is locally closed if and only if Y = U \ V for open subsets V, U ∈ O(X) with V ⊆ U . Then we define A(Y ) := A(U )/A(V ) for a C∗-algebra A over X; this does not depend on the choice of U and V by [8, Lemma 2.15]. If Y ⊆ X is locally closed and A is a C∗-algebra over Y , then we extend A to a Y A(Z) := A(Y ∩Z) for Z ∈ LC(X). Conversely, we can X B(Z) := B(Z) C∗-algebra iX restrict a C∗-algebra B over X to a C∗-algebra rY for all Z ∈ LC(Y ) ⊆ LC(X). X (B) over Y by rY Y A over X by iX The category KK(X) is triangulated, with exact triangles coming either from mapping cone triangles of X-equivariant ∗-homomorphisms or, equivalently, from semi-split C∗-algebra extensions over X (see [7, 8]). Here an extension is called semi-split if it splits by an X-equivariant completely positive contraction. The bootstrap class B(X) is the localising subcategory of KK(X) generated by the objects ixC for all x ∈ X. That is, it is the smallest class of objects containing these generators that is closed under suspensions, KK(X)-equivalence, semi-split extensions, and countable direct sums. 6 RALF MEYER AND RYSZARD NEST 2. Filtrated K-theory Let X be a finite topological space. We do not discuss filtrated K-theory for C∗-algebras over infinite spaces here. Definition 2.1. For a locally closed subset Y ⊆ X, we define a functor FKY : KK(X) → AbZ/2, FKY (A) := K∗(cid:0)A(Y )(cid:1). Here Ab denotes the category of Abelian groups and AbZ/2 denotes the category of Z/2-graded Abelian groups. For each Y ∈ LC(X), the functor FKY is stable and homological, that is, it intertwines the suspension on KK(X) with the translation functor on AbZ/2 (this functor shifts the grading), and if ΣC → A → B → C is an exact triangle in KK(X) -- this may, for instance, come from a semi-split extension A ֌ B ։ C -- then FKY (A) → FKY (B) → FKY (C) is an exact sequence in AbZ/2. The functors FKY together form the filtrated K-theory functor. But the latter also includes its target category, which we now define in a rather abstract way. Definition 2.2. For Y, Z ∈ LC(X), let N T ∗(Y, Z) be the Z/2-graded Abelian group of all natural transformations FKY ⇒ FKZ. The composition of natural transformations provides a product N T i(Y, Z) × N T j(W, Y ) → N T i+j(W, Z), f, g 7→ f ◦ g, which is associative and additive in each variable. We let N T be the Z/2-graded category whose object set is LC and whose morph- ism space Y → Z is N T ∗(Y, Z). The Abelian group structure on these morphism spaces turns this into a pre-additive category. Definition 2.3. A module over N T is a grading preserving, additive functor G : N T → AbZ/2. That is, it consists of a family of Z/2-graded Abelian groups GY = (GY,0, GY,1) for Y ∈ LC(X) and product maps N T i(Y, Z) × GY,j → GZ,i+j for all Y, Z ∈ LC(X), i, j ∈ Z/2; these product maps are associative, additive in each variable, and the identity transformations in N T (Y, Y ) act identically on GY for all Y ∈ LC(X). Let Mod(N T ) be the category of N T -modules. The morphisms in Mod(N T ) are the natural transformations of functors or, equivalently, families of grading preserving group homomorphisms GY → G′ Y that commute with the actions of N T . Let Mod(N T )c be the full subcategory of countable modules. By construction, the natural transformations FKY ⇒ FKZ in N T ∗(Y, Z) induce maps FKY (A) → FKZ (A) for all A ∈∈ KK(X). This turns(cid:0)FKY (A)(cid:1)Y ∈LC(X) into a module over N T . Furthermore, it is well-known that the K-theory of separable C∗-algebras such as A(Y ) for A ∈∈ KK(X) is countable. Definition 2.4. Filtrated K-theory is the functor FK = (FKY )Y ∈LC(X) : KK(X) → Mod(N T )c, A 7→(cid:16)K∗(cid:0)A(Y )(cid:1)(cid:17)Y ∈LC(X) . The target category Mod(N T )c is an important part of this definition because we will compute groups of morphisms and extensions in this category. Since A(∅) = {0} for all C∗-algebras over X, we have FK∅ = 0, so that ∅ is a zero object of N T . Therefore, G∅ vanishes for any N T -module. If Y ∈ LC(X) is not connected, that is, Y = Y1 ⊔ Y2 with two disjoint relat- ively open subsets Y1, Y2 ∈ O(Y ) ⊆ LC(X), then A(Y ) ∼= A(Y1) ⊕ A(Y2) for any C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 7 C∗-algebra A over X. Hence FKY (A) ∼= FKY1 (A) × FKY2(A). The natural trans- formations that implement this natural isomorphism correspond to a direct sum diagram Y ∼= Y1 ⊕ Y2 in N T . Therefore, any N T -module has GY ∼= GY1 ⊕ GY2 ; here we use the fact that a functor that is additive on morphisms is also additive on objects, even if the category in question is only pre-additive. Since X is finite, any locally closed subset is a disjoint union of its connected components. This corresponds to a direct sum decomposition Y ∼= Lj∈π0(Y ) Yj in N T . Therefore, we lose no information when we replace LC(X) by the subset LC(X)∗ of non-empty, connected, locally closed subsets. 2.1. The representability theorem. The representability theorem serves two purposes. We will first use it to describe the category N T . Later, we use it to construct geometric resolutions in KK(X). Theorem 2.5. Let X be a finite topological space. The covariant functors FKY for Y ∈ LC(X) are representable, that is, there are objects RY ∈∈ KK(X) and natural isomorphisms for all A ∈∈ KK(X), Y ∈ LC(X). KK∗(X; RY , A) ∼= FKY (A) = K∗(cid:0)A(Y )(cid:1) Before we prove this theorem in §2.2, we first describe the representing ob- jects RY explicitly, and we use this to describe the groups of natural transformations N T ∗(Y, Z) as K-theory groups of certain locally compact spaces. The construction of RY requires some preparation. We equip X with the spe- cialisation preorder (cid:22) as in [8, §2.7]; recall that x (cid:22) y if and only if {x} ⊆ {y}. Since the topological space X is finite, it carries the Alexandrov topology of the preorder (cid:22), that is, a subset Y ⊆ X is open if and only if x (cid:23) y ∈ Y implies x ∈ Y . Similarly, Y ⊆ X is closed if and only if x (cid:22) y ∈ Y implies x ∈ Y , and locally closed if and only if x (cid:22) y (cid:22) z and x, z ∈ Y implies y ∈ Y . Definition 2.6. Let (X, (cid:22)) be a partially ordered set. Its order complex is the geometric realisation of the simplicial set Ch(X) whose n-simplices are the chains x0 (cid:22) x1 (cid:22) · · · (cid:22) xn in X and whose face and degeneracy maps delete or double an entry of the chain. Equivalently, Ch(X) is the classifying space of the thin category that has object set X and a morphism x → y whenever x (cid:22) y. The order complex is the main ingredient in the construction of the representing objects RY for Y ∈ LC(X). The non-degenerate n-simplices in Ch(X) are the strict chains x0 ≺ · · · ≺ xn in X. We let SX be the set of all strict chains. For each I = (x0 ≺ · · · ≺ xn) ∈ SX , we let ∆I be a copy of ∆n; more formally, ∆I = {(t, I) t ∈ ∆n}. We also let ∆◦ I ⊆ ∆I be the corresponding open simplex ∆n \ ∂∆n. The space Ch(X) is obtained from the union`I∈SX ∆I by identifying ∆I with the corresponding face in ∆J whenever I, J ∈ SX satisfy I ⊆ J. Thus the underly- ing set of Ch(X) is a disjoint union For I ∈ SX , let min I and max I be the (unique) minimal and maximal elements in SX , respectively. We define two functions by mapping points in ∆◦ tions on Ch(X) because of (2.1). I to min I and max I, respectively. This well-defines func- m, M : Ch(X) → X (2.1) Ch(X) = aI∈SX ∆◦ I . 8 RALF MEYER AND RYSZARD NEST Lemma 2.7. If Y ⊆ X is closed, then m−1(Y ) is open and M −1(Y ) is closed in Ch(X). If Y ⊆ X is open, then m−1(Y ) is closed and M −1(Y ) is open. If Y ⊆ X is locally closed, then m−1(Y ) and M −1(Y ) are locally closed. Proof. First we show that M −1(Y ) is closed if Y is closed. If I ∈ SX satisfies max I ∈ Y , then max J ∈ Y for all J ⊆ I because max J (cid:22) max I ∈ Y . Hence ∆I ⊆ M −1(Y ) once M −1(Y ) ∩ ∆◦ I 6= ∅, so that M −1(Y ) ∩ ∆I is closed for all I ∈ SX ; this implies that M −1(Y ) is closed. A similar argument shows that m−1(Y ) is closed in Ch(X) if Y is open. Now the remaining assertions follow easily because the maps m−1 and M −1 commute with complements, unions, and intersections. (cid:3) More explicitly, if Y ⊆ X is open, then m−1(Y ) is the union of the simplices SX for all chains x0 ≺ x1 ≺ · · · ≺ xn with x0 ∈ Y and hence x0, . . . , xn ∈ Y . Thus Similarly, m−1(Y ) = Ch(Y ) if Y ⊆ X is open. M −1(Y ) = Ch(Y ) if Y ⊆ X is closed. Here we identify Ch(Y ) with a subcomplex of Ch(X) in the obvious way. Let X op be X with the topology for the reversed partial order ≻; that is, the open subsets of X op are the closed subsets of X, and vice versa. We may rephrase Lemma 2.7 as follows: Proposition 2.8. The map (m, M ) : Ch(X) → X op × X is continuous. Let be the C∗-algebra of continuous functions on Ch(X). Since R := C(cid:0)Ch(X)(cid:1) Prim R = Prim C(cid:0)Ch(X)(cid:1) ∼= Ch(X), S(Y, Z) := m−1(Y ) ∩ M −1(Z) ⊆ Ch(X); the map (m, M ) turns R into a C∗-algebra over X op × X. We abbreviate this is a locally closed subset of Ch(X) by Lemma 2.7 Definition 2.9. We let RY be the C∗-algebra over X with RY (Z) := R(Y op × Z) = C0(cid:0)S(Y, Z)(cid:1) for all Y, Z ∈ LC(X); here Y op denotes Y with the subspace topology from X op. Equivalently, we let RY be the restriction of R to Y op × X, viewed as a C∗-algebra over X via the coordinate projection Y op × X → X. We will prove the Theorem 2.5 for this choice of RY in §2.2. Taking this for granted, we use the concrete description of RY to compute the groups of natural transformations. By the Yoneda Lemma, natural transformations between the func- tors FKY come from morphisms between the representing objects. More precisely, (2.2) N T ∗(Y, Z) ∼= KK∗(X; RZ , RY ) ∼= FKZ (RY ) = K∗(cid:0)RY (Z)(cid:1) = K∗(cid:0)R(Y op × Z)(cid:1) = K∗(cid:0)m−1(Y ) ∩ M −1(Z)(cid:1) = K∗(cid:0)S(Y, Z)(cid:1). By the way, the universal property of Kasparov theory says that it makes no difference for the natural transformations FKY ⇒ FKZ whether we view these two functors as defined on C∗sep(X) or KK(X). But since RY only represents FKY on the level of KK(X), we get KK∗(X; RZ, RY ) and not the space of X-equivariant ∗-homomorphisms RZ → RY . C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 9 We describe S(Y, Z) more explicitly using the closure and boundary operations Z := {x ∈ X there is z ∈ Z with x (cid:22) z}, ∂Z := Z \ Z, Lemma 2.10. If Y, Z ∈ LC(X), then e∂Y := eY \ Y. eY := {x ∈ X there is y ∈ Y with x (cid:23) y}, Of course, Z is the closure of Z in X and eY is the closure of Y in X op. S(Y, Z) = Ch(eY ∩ Z)(cid:15)(cid:0)Ch(eY ∩ ∂Z) ∪ Ch(e∂Y ∩ Z)(cid:1). In particular, S(Y, Z) = Ch(Y ∩ Z) \ Ch(Y ∩ ∂Z) if Y is open, S(Y, Z) = Ch(eY ∩ Z) \ Ch(e∂Y ∩ Z) S(Y, Z) = Ch(Y ∩ Z) if Z is closed, if Y is open and Z is closed. Proof. Let x0 ≺ x1 ≺ · · · ≺ xn be a strict chain in X. The interior of the corres- ponding simplex belongs to S(Y, Z) if and only if x0 ∈ Y and xn ∈ Z. This implies xj ∈ eY and xj ∈ Z for all j, so that the simplex belongs to Ch(eY ∩Z). Furthermore, we neither have xj ∈ e∂Y ∩ Z for all j nor xj ∈ eY ∩ ∂Z for all j because x0 ∈ Y and xn ∈ Z. Thus the simplex belongs neither to Ch(eY ∩ ∂Z) nor to Ch(e∂Y ∩ Z). Conversely, if xj ∈ eY ∩ Z for all j and neither xj ∈ e∂Y ∩ Z for all j nor xj ∈ eY ∩ ∂Z for all j, then some xj must belong to Y ∩ Z and some xk must belong to eY ∩ Z. Since Y ∩ Z is closed in eY ∩ Z and eY ∩ Z is open in eY ∩ Z, this implies x0 ∈ Y and it is contained in Ch(eY ∩ Z)(cid:15)(cid:0)Ch(eY ∩ ∂Z) ∪ Ch(e∂Y ∩ Z)(cid:1). N T ∗(Y, Z) ∼= K∗(cid:0)S(Y, Z)(cid:1) ∼= K∗(cid:0)Ch(eY ∩ Z), Ch(eY ∩ ∂Z) ∪ Ch(e∂Y ∩ Z)(cid:1). xn ∈ Z. This shows that the interior of a simplex belongs to S(Y, Z) if and only if (cid:3) This is the K-theory of a finite CW-pair and hence is always finitely generated as an Abelian group. Lemma 2.10 and (2.2) yield If C is any finite simplicial complex, then its barycentric subdivision is of the form Ch(X), where X is the partially ordered set of non-degenerate simplices in C. Thus N T ∗(X, X) = K∗(C), so that any finitely generated Abelian group arises as N T ∗(X, X). As a consequence, special properties of the pre-additive category N T can only be hidden in its composition. When we identify N T ∗(Y, Z) ∼= KK∗(X; RZ, RY ), then the composition of nat- ural transformations corresponds to the Kasparov composition product. This gets somewhat obscured when we follow the isomorphisms KK∗(X; RZ , RY ) ∼= K∗(cid:0)RY (Z)(cid:1) = K∗(cid:0)S(Y, Z)(cid:1). To describe the composition of natural transformations in terms of K∗(cid:0)S(Y, Z)(cid:1), we must first lift elements of K∗(cid:0)S(Y, Z)(cid:1) back to KK∗(X; RZ, RY ) and then compose them. The lifting requires a formula for the natural isomorphism the unit element in K0(cid:0)RY (Y )(cid:1). (2.3) KK∗(X; RY , A) → K∗(cid:0)A(Y )(cid:1) that occurs in the Representability Theorem. By the Yoneda Lemma, any such natural transformation is of the form f 7→ f∗(ξY ) for a unique ξY ∈ K0(cid:0)RY (Y )(cid:1) = K0(cid:0)S(Y, Y )(cid:1) = K0(cid:0)Ch(Y )(cid:1). The natural transformation in (2.3) is generated by the class of the 1-dimensional trivial vector bundle over the compact space Ch(Y ) or, equivalently, the class of 10 RALF MEYER AND RYSZARD NEST In the examples we consider later, all natural transformations turn out to be products of obvious ones, coming from the K-theory six-term exact sequences (1.4). To check this, we only have to verify that a given element α of KK∗(X; RZ, RY ) lifts a given element of K∗(cid:0)S(Y, Z)(cid:1). The isomorphism (2.3) maps α to [ξZ]⊗RZ (Z) α(Z) in K∗(cid:0)RY (Z)(cid:1) = K∗(cid:0)S(Y, Z)(cid:1), where α(Z) in KK∗(cid:0)RZ (Z), RY (Z)(cid:1) is obtained from α by restriction to Z. This product is easy to compute. To get acquainted with this approach to natural transformations, we compute some important examples. Let Y ∈ LC(X) and U ∈ O(Y ). Since R is a C∗-algebra over X op × X, there is an extension (2.4) of C∗-algebras over X. It contains C∗-algebra extensions RY \U ֌ RY ։ RU RY \U (Z) ֌ RY (Z) ։ RU (Z) for all Z ∈ LC(X). Let Z := Y \ U . The extension (2.4) is semi-split in C∗alg(X) and hence has a class in KK1(X; RU , RZ) and produces an exact triangle (2.5) in KK(X). ΣRU → RZ → RY → RU Lemma 2.11. The maps in the extension triangle (2.5) correspond to the natural transformations FKU [1] ⇐ FKZ ⇐ FKY ⇐ FKU in (1.4). Proof. The natural transformation µY U : FKU ⇒ FKY in (1.4) is induced by the natural ∗-homomorphism j : A(U ) → A(Y ). For A = RU , this map is invertible because S(U, Y ) = S(U, U ) = Ch(U ). Hence j(ξU ) ∈ K0(cid:0)S(U, Y )(cid:1) is again the class of the trivial vector bundle on Ch(U ); this class corresponds to the natural transformation µY U . The restriction map RY ։ RU in (2.4) maps [ξY ] to [ξU ] -- recall that both [ξY ] and [ξU ] are trivial vector bundles. Hence the restriction map RY ։ RU and the natural transformation µZ Y correspond to the same class -- the Y : FKY ⇒ FKZ is induced by the nat- ural ∗-homomorphism p : A(Y ) ։ A(Z). For A = RY , this is the restriction 1-dimensional trivial vector bundle on Ch(U ) -- in K0(cid:0)S(U, Y )(cid:1). ∗-homomorphism C(cid:0)Ch(Y )(cid:1) → C(cid:0)Ch(Z)(cid:1) because S(Y, Y ) = Ch(Y ) and S(Y, Z) = Ch(Z). Since the restriction of a trivial bundle remains trivial, µZ Y corresponds to the trivial 1-dimensional vector bundle on S(Y, Z) = Ch(Z). The embedding RZ ։ RY restricts to an identity map on Z because S(Z, Z) = S(Z, Y ) = Ch(Z). Since this maps [ξZ ] to the trivial bundle, the embedding RZ ։ RY and µZ Y both correspond to the same class -- the 1-dimensional trivial vector bundle on Ch(Z) -- in Similarly, the natural transformation µZ K0(cid:0)S(Y, Z)(cid:1). Finally, we study the boundary map δU Z : FKZ ⇒ FKU [1]. We claim that it corresponds to the class of the extension RZ ֌ RY ։ RU in KK1(X; RU , RZ). To prove this, we use that Ch(Y ) is the join of the spaces Ch(U ) and Ch(Z), so that there is a continuous map f : Ch(Y ) → [0, 1] whose fibres over 0 and 1 are Ch(U ) and Ch(Z), respectively. More precisely, let x0 ≺ x1 ≺ · · · ≺ xn be a strict chain in Y and let ξ be a point of the corresponding simplex with coordinates (t0, . . . , tn) with t0 + · · · + tn = 1, that is, ξ = t0x0 + · · · + tnxn. Then there is j ∈ {0, . . . , n} with x0, . . . , xj ∈ U , xj+1, . . . , xn ∈ Z. We can, therefore, write ξ = tU ξU + tZξZ with ξU = ξZ = t0x0 + · · · + tjxj tU ∈ Ch(U ), tU = t0 + · · · + tj, tj+1xj+1 + · · · + tnxn tZ ∈ Ch(Z), tZ = tj+1 + · · · + tn. C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 11 We define a continuous map f : Ch(Y ) → [0, 1] by ξ 7→ tZ . We have S(U, U ) = Ch(U ) = f −1(0), S(Z, Z) = Ch(Z) = f −1(1) by construction, and hence Now we can compute some boundary maps. The boundary map S(Z, U ) = Ch(Y ) \(cid:0)Ch(U ) ⊔ Ch(Z)(cid:1) = f −1(cid:0)(0, 1)(cid:1). K0(cid:0)S(Z, Z)(cid:1) ∼= K0(cid:0)RZ(Z)(cid:1) → K1(cid:0)RZ(U )(cid:1) ∼= K1(cid:0)S(Z, U )(cid:1) maps the class of the trivial bundle [ξZ ] to f ∗(δ), where δ denotes a generator of Z ∼= K1(cid:0)(0, 1)(cid:1); this follows from the naturality of the boundary map. The boundary map K0(cid:0)S(U, U )(cid:1) ∼= K0(cid:0)RU (U )(cid:1) → K1(cid:0)RZ (U )(cid:1) ∼= K1(cid:0)S(Z, U )(cid:1) for the extension RZ ֌ RY ։ RU maps the class of the trivial bundle [ξU ] to −f ∗(δ), again by naturality of the boundary map. (cid:3) natural transformations FKU ⇒ FKY and FKY ⇒ FKZ are represented by the classes of the trivial vector bundles over the compact spaces S(U, Y ) and S(Y, Z); the natural boundary map FKZ ⇒ FKU [1] is represented by f ∗(δ) for a generator Remark 2.12. The proof also describes the classes in K0(cid:0)S(U, Y )(cid:1), K0(cid:0)S(Y, Z)(cid:1), and K1(cid:0)S(Z, U )(cid:1) that correspond to the natural transformations in (1.4). The of K1(cid:0)(0, 1)(cid:1). KK∗(X; RY , A) → K∗(cid:0)A(Y )(cid:1) induced by ξY is an isomorphism if Y is the min- imal open subset Ux containing some point x ∈ X. The adjointness relation 2.2. Proof of Theorem 2.5. We check first that the natural transformation for all B ∈∈ KK(X) established in [8, Proposition 3.12] yields KK∗(X; ix(A), B) ∼= KK∗(cid:0)A, B(Ux)(cid:1) KK∗(X; ix(C), B) ∼= KK∗(cid:0)C, B(Ux)(cid:1) = FKUx(B), that is, ix(C) represents FKUx. To check that RUx does so as well, we must show that ix(C) and RUx are KK(X)-equivalent. Recall that ix(C) = (C, x), where x denotes the map Prim(C) ∼= {x} ⊆ −→ X, and ix(C)(Z) =(C if x ∈ Z, otherwise 0 for all Z ∈ LC(X). Since Ux = {y ∈ X x (cid:22) y}, the preordered set Ux has a minimal point, namely x. Therefore, the space Ch(Ux) is starlike and hence contractible in a canonical way towards x. The path from a point in ∆I for I ∈ SUx to the base point in ∆x lies in ∆I∪{x}. Since max I ∪ {x} = max I, the contraction preserves the ideals RUx (V ) for V ∈ O(X), so that we get a homotopy equivalence between C(cid:0)Ch(Ux)(cid:1) and ix(C) in C∗alg(X). Thus RUx corepresents FKUx as well. It is easy to see that the natural isomorphism KK∗(X; RUx, ) ∼= FKUx is induced by ξUx . Let Good ⊆ LC(X) be the set of all Z ∈ LC(X) for which the natural trans- formation KK∗(X; RZ, A) → FKZ (A) induced by ξZ is an isomorphism. We must show Good = LC(X). We have just seen that Ux ∈ Good for all x ∈ X. Let Y ∈ LC(X) and U ∈ O(Y ); we claim that all three of U , Y , and Y \ U are good once two of them are. This follows from the Five Lemma because the 12 RALF MEYER AND RYSZARD NEST maps induced by ξZ for Z = U, Y, Y \ U intertwine the maps in the six-term exact sequences (1.4) and KK0(X; RU , A) / KK0(X; RY , A) / KK0(X; RY \U , A) KK1(X; RY \U , A) KK1(X; RY , A) KK1(X; RU , A) for any A ∈∈ KK(X); the latter six-term exact sequence is induced by the semi- split extension (2.5). The commutativity of the relevant diagrams follows from the computations in the proof of Lemma 2.11 (which do not depend on Theorem 2.5). The two-out-of-three property of Good implies: U, V ∈ O(X), U, V, U ∩ V ∈ Good =⇒ U ∪ V ∈ Good because (U ∪ V ) \ U = V \ (U ∩ V ). By induction on the length of U , this implies that all open subsets of X belong to Good. Since any locally closed subset is a difference of two open subsets, we conclude that Good = LC(X). This finishes the proof of Theorem 2.5. 3. An example In this section, we restrict our attention to a special class of spaces, namely, the spaces X = {1, . . . , n} totally ordered by ≤ for n ∈ N. We let [a, b] := {x ∈ X a ≤ x ≤ b}. for a, b ∈ Z. We equip X with the Alexandrov topology, so that the open subsets are [a, n] for all a ∈ X; the closed subsets are [1, b] with b ∈ X, and the locally closed subsets are those of the form [a, b] with a, b ∈ X and a ≤ b. Any locally closed subset of X is connected. 3.1. Computations with the order complex. Since any subset of X is totally ordered, the space Ch([a, b]) is just a closed simplex of dimension b − a for any b ≥ a. We denote the corresponding face of Ch(X) by ∆[a,b]. This is understood to be empty for a > b. From now on, we let Y = [a1, b1], Z = [a2, b2], with 1 ≤ a1 ≤ b1 ≤ n and 1 ≤ a2 ≤ b2 ≤ n. Then eY = [a1, n], e∂Y = [b1 + 1, n], Z = [1, b2], and ∂Z = [1, a2 − 1]. Lemma 2.10 yields S(Y, Z) = ∆[a1,b2] \(cid:0)∆[a1,a2−1] ∪ ∆[b1+1,b2](cid:1). Now we distinguish three cases: Case 1: If a2 ≤ a1 ≤ b2 ≤ b1, then S(Y, Z) = ∆[a1,b2] is a non-empty closed simplex. Case 2: If a2 − 1 ≤ b1, a1 < a2, and b1 < b2, then S(Y, Z) is obtained from a closed simplex by removing two disjoint, non-empty closed faces. Excision yields Hence N T ∗(Y, Z) ∼= K∗(cid:0)S(Y, Z)(cid:1) ∼= Z[0] (this means Z in degree 0). N T ∗(Y, Z) ∼= K∗(cid:0)S(Y, Z)(cid:1) ∼= Z[1] (this means Z in degree 1). that N T ∗(Y, Z) ∼= K∗(cid:0)S(Y, Z)(cid:1) ∼= 0. Case 3: In all other cases, S(Y, Z) is either empty, a difference of two closed sim- plices, or a difference σ \ (τ1 ∪ τ2) for two non-empty closed faces τ1 and τ2 of a simplex σ that intersect. Then τ1 ∪ τ2 and σ are both contractible, so / / O O o o o o C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 13 (3.1) Summing up, we get N T ∗(Y, Z) = Z[0] Z[1] 0 if a2 ≤ a1 ≤ b2 ≤ b1, if a2 − 1 ≤ b1, a1 < a2, and b1 < b2, otherwise. 3.2. Products of natural transformations. Our next task is to identify the natural transformations that correspond to the generators of the groups in (3.1); this also allows us to compute products in N T . First we study the grading preserving transformations that appear in the first case. We introduce a partial order ≥ and a strict partial order ≫ on LC(X) by [a1, b1] ≥ [a2, b2] [a1, b1] ≫ [a2, b2] ⇐⇒ ⇐⇒ a1 ≥ a2 and b1 ≥ b2, a1 > b2. Our computation shows that N T 0(Y, Z) 6= {0} if and only if Y ≥ Z but not Y ≫ Z. This is equivalent to Y ∩ Z being non-empty, closed in Y , and open in Z. Under these assumptions, there is a natural non-zero ∗-homomorphism given by the composition µZ Y : A(Y ) ։ A(Y ∩ Z) ֌ A(Z) because A(Y ∩ Z) is a quotient of A(Y ) and an ideal in A(Z). The natural trans- formation FKY ⇒ FKZ induced by µZ Y maps ξY ∈ FKY,0(RY ), which is the class of the trivial line bundle over S(Y, Y ) = ∆[a1,b1], to the trivial line bundle over S(Y, Z) = ∆[a1,b2]. Since this is the generator of FKZ,0(RY ) = K0(cid:0)S(Y, Z)(cid:1) ∼= Z[0], the natural transformation µZ Y generates N T 0(Y, Z). If Y ≫ Z, then we let µZ Y : A(Y ) → A(Z) be the zero map, which induces the zero transformation FKY ⇒ FKZ. With this convention, we get µZ W for all Y, Z, W ∈ LC(X) with W ≥ Y ≥ Z, also if W ≫ Z; this equation holds on the level of ∗-homomorphisms and, therefore, also for the induced natural transformations. We can sum this up as follows: W = µZ Y ◦ µY Lemma 3.1. The category N T 0 of grading-preserving natural transformations FKY ⇒ FKZ for Y, Z ∈ LC(X) is the pre-additive category generated by natural transformations µZ W = µZ W for W ≥ Y ≥ Z and µZ Y : FKY ⇒ FKZ for all Y ≥ Z with the relations µZ Y ◦ µY Y = 0 for Y ≫ Z. This list of generators is longer than necessary. Clearly, we can write any µZ Y as a product of the transformations µ[a−1,b] for 1 ≤ a < b ≤ n. Moreover, these transformations themselves are indecomposable, that is, they cannot be written themselves as products in a non-trivial way. for 2 ≤ a ≤ b ≤ n and µ[a,b−1] [a,b] [a,b] Now we turn to the natural transformations of degree 1. For any b ∈ X and any C∗-algebra A over X, we have a natural C∗-algebra extension A([b, n]) ֌ A([1, n]) ։ A([1, b − 1]), which generates an odd natural transformation δb : FK[1,b−1] ⇒ FK[b,n]. Composing with the grading preserving natural transformations µ above, we get a natural transformation of degree 1 (3.2) δZ Y : FKY = FK[a1,b1] µ =⇒ FK[1,a2−1] δa2==⇒ FK[a2,n] µ =⇒ FK[a2,b2] = FKZ whenever b1 ≥ a2 − 1. Equation (3.1) predicts that this transformation vanishes if a1 ≥ a2 or b1 ≥ b2. This can be verified as follows. Vanishing for a1 ≥ a2 is clear because then [a1, b1] ≫ [1, a2−1]. By the naturality of the boundary map, the transformation in (3.2) agrees 14 RALF MEYER AND RYSZARD NEST with the composition of µ : FK[a1,b1] ⇒ FK[a1,a2−1] with the boundary map for the extension A([a2, b2]) ֌ A([a1, b2]) ։ A([a1, a2 − 1]). (3.3) If b1 ≥ b2, then µ[a1,a2−1] composite of two maps in a six-term exact sequence vanishes. [a1,b1] factors through the quotient map in (3.3). But the Equation (3.2) produces a natural transformation δZ Y ∈ N T 1(Y, Z) whenever a1 < a2, b1 < b2, and a2 − 1 ≤ b1, that is, whenever (3.1) predicts N T 1(Y, Z) to be non-zero. We claim that δZ Y generates this group. This follows because the natural transformation δZ Y maps the class of the trivial line bundle over S(Y, Y ) to the generator of K1(cid:0)S(Y, Z)(cid:1) ∼= Z. Notice that N T 1([a2, n], Z) = {0} for any Z ∈ LC(X). Since the natural trans- formation (3.2) above factors through FK[a2,n], any product of two odd natural transformations vanishes. Thus the category N T is a split extension of N T 0 by the bimodule N T 1. The bimodule structure on N T 1 is very simple: a product Y ◦ δY µZ W whenever all three natural transformations are defined, and zero otherwise. W is equal to δZ W or δZ Y ◦ µY Example 3.2. To make our constructions more concrete, we now consider the ex- ample n = 2, which corresponds to extensions of C∗-algebras. There are only three non-empty locally closed subsets: 1 = [1, 1], 12 = [1, 2], and 2 = [2, 2]. The order complex is an interval; we label its end points 1 and 2. The map (m, M ) from Ch(X) = [1, 2] to X op × X maps 1 7→ (1, 1), 2 7→ (2, 2), ]1, 2[ 7→ (1, 2). Correspondingly, we have S(1, 1) = {1}, S(1, 2) = ]1, 2[, S(1, 12) = [1, 2[, S(2, 1) = ∅, S(2, 2) = {2}, S(2, 12) = {2}, S(12, 1) = {1}, S(12, 2) = ]1, 2], S(12, 12) = [1, 2]. Taking K-theory, we get N T (1, 1) = Z[0], N T (1, 2) = Z[1], N T (1, 12) = 0, N T (2, 1) = 0, N T (2, 2) = Z[0], N T (2, 12) = Z[0], N T (12, 1) = Z[0], N T (12, 2) = 0, N T (12, 12) = Z[0]. 3.3. Ring-theoretic properties of the natural transformations. We now ob- serve some general ring-theoretic properties of N T for X = {1, . . . , n} with the total order. We exclude the trivial case n = 1. We may replace N T by a Z/2- graded ring by taking the direct sum of N T ∗(Y, Z) for all Y, Z ∈ LC(X)∗ and defining the product as usual for a category ring. Then N T -modules become Z/2- graded modules over this Z/2-graded ring, and ring-theoretic notions such as the Jacobson radical and the balanced tensor product ⊗N T make sense. Definition 3.3. Let N T nil ⊆ N T be the subgroup spanned by the natural trans- formations µZ Y with Y 6= Z and δZ Y with arbitrary Y, Z. Let N T ss ⊆ N T be the subgroup spanned by the natural transformations µY Y with Y ∈ LC(X)∗. Lemma 3.4. The subgroup N T nil is the maximal nilpotent ideal in N T , it is the nilradical and the Jacobson radical of N T . The subgroup N T ss is a semi-simple subring, and N T decomposes as a semi-direct product N T nil ⋊ N T ss. Proof. Since all µY pointwise multiplication. Y are idempotent, N T ss is a subring isomorphic to ZLC(X)∗ It is easy to see that N T nil is an ideal in N T . with It is C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 15 nilpotent, that is, N T k nil = {0} for some k ∈ N, because LC(X)∗ is finite and ≥ is a partial order on it. Since N T = N T nil ⊕ N T ss as Abelian groups, we get the desired semi-direct product decomposition. Since the Jacobson radical of N T ss vanishes, N T nil is both the nilradical and the Jacobson radical of N T . (cid:3) We are going to use Lemma 3.4 to characterise the projective N T -modules. This characterisation involves the following two definitions. Definition 3.5. We call an N T -module M exact if the chain complexes · · · → M (U ) µY U−−→ M (Y ) Y \U µ Y−−−→ M (Y \ U ) δU Y \U−−−→ M (U ) → · · · are exact for all Y ∈ LC(X), U ∈ O(Y ) as in (1.5). Proposition 3.6. Let K ֌ E ։ Q be an extension of N T -modules. If two of the modules K, E, Q are exact, so is the third one. Proof. Given U and Y as above and a module M , let C•(M ) be the chain complex · · · → M (U )[m] → M (Y )[m] → M (Y \ U )[m] → M (U )[m − 1] → · · · . Then C•(K) ֌ C•(E) ։ C•(Q) is an extension of chain complexes. The long exact homology sequence shows that all three of these chain complexes are exact once two of them are exact. (cid:3) Definition 3.7. Given an N T -module M , we let N T nil · M = {x · m x ∈ N T nil, m ∈ M }, Mss := M/N T nil · M. We call Mss the semi-simple part of M . Since the tensor product over N T is right exact, Mss ∼= N T ss ⊗N T M . We need the following more concrete description of Mss or, equivalently, of N T nil · M . Lemma 3.8. Let M be an N T -module and let Y = [a, b] with 1 ≤ a ≤ b ≤ n. Then if b < n, if b = n. (N T nil · M )(Y ) =(ker(cid:0)δ[a+1,b+1] ker(cid:0)µ[1,a] [a,b] : M [a, b] → M [a + 1, b + 1](cid:1) [a,b] : M [a, b] → M [1, a](cid:1) [a+1,b] or µY [a,b+1] if a < b < n, through µY Proof. The first assertion holds because any natural transformation FKZ ⇒ FKY with Z 6= Y factors through µY [a,b+1] if a = b < n, and so on. Here we use that the natural transformations µ[a−1,b] for 2 ≤ a ≤ b ≤ n, µ[a,b−1] [1,a−1] for 2 ≤ a ≤ n already Y or δZ generate N T ∗, that is, all other transformations µZ Y with Y 6= Z can be written as products of these generators. By the way, these natural transformations even form a basis for the subquotient N T nil/N T 2 nil. Now assume that M is exact. If a = b < n, then for 1 ≤ a < b ≤ n, and δ[a,n] [a,b] [a,b] (N T nil · M )[a, a] = range(cid:16)µ[a,a] [a,a+1](cid:17) = ker(cid:16)δ[a+1,a+1] [a,a] (cid:17).  (N T nil · M )(Y ) = If M is exact, then [a+1,b](M [a + 1, b]) + µY µY µY [a,b+1](M [a, b + 1]) [a+1,b](M [a + 1, b]) + δY µY µY [a+1,b](M [a + 1, b]) δY [1,a−1](M [1, a − 1]) [a,b+1](M [a, b + 1]) [1,a−1](M [1, a − 1]) if a < b < n, if a = b < n, if 1 < a < b = n, if 1 = a < b = n, if a = b = n. 16 RALF MEYER AND RYSZARD NEST Similarly, we get [n,n](cid:17), (N T nil · M )[n, n] = ker(cid:16)µ[1,n] [1,n](cid:17). (N T nil · M )[1, n] = ker(cid:16)µ[1,1] Given f1 : A1 → B and f2 : A2 → B and two exact sequences A1 f1−→ B g1−→ C1, A2 g1f2−−−→ C1 g2−→ C2, we have (3.4) range(f1) + range(f2) = ker(g1) + range(f2) = {x ∈ B g1(x) ∈ range(g1 ◦ f2) = ker(g2)} = ker(g2 ◦ g1). [a,b+1] with Y = [a, b]. We get g1 = µ[a,a] If a < b < n, then we apply this to the maps on M induced by f1 = µY [a+1,b] and f2 = µY [a,b+1], and hence g2 = δ[a+1,b+1] . This yields the desired formula for (N T nil · M )[a, b] for a < b < n, using the exactness of M . If a < b = n, then we [1,a−1]. Here we get g1 = µ[a,a] apply the same reasoning to f1 = µY as above, g1 ◦ f2 = δ[a,a] [a,b]. This yields the desired formula for (N T nil · M )[a, b] for a < b = n. (cid:3) [1,a−1], and hence g2 = µ[1,a] [a,a] and g2 ◦ g1 = µ[1,a] and g2 ◦ g1 = δ[a+1,b+1] [a+1,b] and f2 = δY , g1 ◦ f2 = µ[a,a] [a,a] [a,b] Y Y Remark 3.9. The natural transformation δ[a+1,b+1] [a,n] for b = n is the longest natural transformation out of [a, b] in the following sense: it factors through δZ [a,b] whenever the latter is defined and non-zero. Thus Lemma 3.8 identifies N T nil ·M (Y ) with the largest proper subgroup of M (Y ) that is the kernel of some δZ for b < n or µ[1,a] [a,b] or µZ [a,b] [a,b] or µZ [a,b]. The following proposition is a rather trivial variant of the Nakayama Lemma. Unlike in the usual Nakayama Lemma, we do not assume the module to be finitely generated. This is no problem because the relevant ideal N T nil is nilpotent. Proposition 3.10. Let M be an N T -module with Mss = 0. Then M = 0. Proof. By assumption, M = N T nil · M . By induction, this implies M = N T j for all j ∈ N. Since N T k nil = 0 for some k, we get M = 0. nil · M (cid:3) 3.4. Characterisation of free and projective modules. Definition 3.11. For Y ∈ LC(X), the free N T -module on Y is defined by PY (Z) := N T ∗(Y, Z) for all Z ∈ LC(X). An N T -module is called free if it is isomorphic to a direct sum of degree-shifted free modules PY [j], j ∈ Z/2. Theorem 3.12. Let M be an N T -module. Then the following are equivalent: (i) M is a free N T -module. (ii) M is a projective N T -module. (iii) Mss(Y ) = N T ss ⊗N T M (Y ) is a free Abelian group for all Y ∈ LC(X) and TorN T 1 (N T ss, M ) = 0. (iv) M (Y ) is a free Abelian group for all Y ∈ LC(X) and M is exact. 1 Here TorN T denotes the first derived functor of ⊗N T . The first three conditions remain equivalent when we replace N T by any ring that is a nilpotent extension of the ring ZN for some N ∈ N. C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 17 Proof. The Yoneda Lemma asserts that Hom(PY , M ) ∼= M (Y ) for all Y ∈ LC(X) and all N T -modules M . Hence free modules are projective, that is, (1)=⇒(2). A functor of the form M 7→ R ⊗S M for a ring homomorphism S → R always maps free modules to free modules and hence maps projective modules to projective mod- ules. Furthermore, derived functors like TorN T automatically vanish on projective modules. This yields the implication (2)=⇒(3). We are going to prove that (3) implies (1). Since Mss(Y ) is a free Abelian group for all Y , Mss is a free module over N T ss ∼= ZLC(X)∗ . Hence P := N T ⊗N T ss Mss is a free N T -module. The canonical projection M → Mss splits by an N T ss-module homomorphism because Mss is free. This induces an N T -module homomorphism f : P → M because of the adjointness relation 1 HomN T (N T ⊗N T ss X, Y ) ∼= HomN T ss(X, Y ). We claim that f is invertible, so that M ∼= P is a free module as asserted. We have Pss = N T ss ⊗N T N T ⊗N T ss Mss ∼= N T ss ⊗N T ss Mss ∼= Mss. Inspection shows that this isomorphism is induced by f . Since the functor M 7→ Mss is right-exact, this implies coker(f )ss = 0 and hence coker(f ) = 0 by the Nakayama Lemma (Proposition 3.10). That is, f is an epimorphism. Let K := ker(f ), then we get an exact sequence of N T -modules K ֌ P ։ M . The derived functors of N T ss ⊗N T provide a long exact sequence (3.5) 0 → TorN T 1 (N T ss, M ) → Kss → Pss f −→∼= Mss → 0. 1 1 This exact sequence ends at TorN T (N T ss, P ) = 0 because P is projective. Since TorN T (N T ss, M ) = 0 by assumption, we conclude that Kss = 0. Hence another application of the Nakayama Lemma shows that ker(f ) = 0 as well. Thus f is invertible. This finishes the proof of the implication (3)=⇒(1), showing that the first three conditions are equivalent. Furthermore, our argument so far works for any split nilpotent extension of ZN for some N ∈ N because this is the only information about N T that we have used. Nilpotent extensions of the ring ZN always split because we can lift orthogonal idempotents in nilpotent extensions. Free N T -modules are exact, and they consist of free Abelian groups by (3.1). This yields the implication (1)=⇒(4). We are going to prove that (4) implies (3). This will finish the proof of the theorem. Since we will use this once again later, we state half of this argument as a separate lemma: Lemma 3.13. Let M be an exact N T -module. Then TorN T 1 (N T ss, M ) = 0. Proof. Let π : P → M be an epimorphism with a projective N T -module P , and let K := ker π. Since projective modules are exact and K ֌ P ։ M is a module extension, Proposition 3.6 shows that K is exact. We still have an exact sequence as in (3.5). Since K and P are exact, Lemma 3.8 identifies Kss(Y ) and Pss(Y ) in a natural way with subspaces of K(Z) and P (Z) for suitable Z; here we use A/ ker(f ) ∼= range(f ) for a group homomorphism f : A → B. Since the map K(Z) → P (Z) is injective, so is the map Kss(Y ) → Pss(Y ). Hence the map Kss → Pss is a monomorphism, forcing TorN T (cid:3) (N T ss, M ) = 0 by (3.5). 1 To finish the proof of the implication (4)=⇒(3) in Theorem 3.12, it remains to check that Mss(Y ) is free for all Y if M is exact and M (Y ) is free for all Y . We use Lemma 3.8 once again to describe Mss(Y ) as the range of a canonical element in N T ∗(Y, Z) for a suitable Z. Thus Mss(Y ) is isomorphic to a subgroup of M (Z), which is a free group by assumption. Hence Mss(Y ) is free as well. (cid:3) 18 RALF MEYER AND RYSZARD NEST 4. Homological algebra in KK(X) Let X be a sober topological space. We are going to apply to KK(X) the gen- eral machinery for doing homological algebra in triangulated categories discussed in [9]. This theory goes back to the work on relative homological algebra by Samuel Eilenberg and John Coleman Moore ([4]), which was carried over to the setting of triangulated categories by Daniel Christensen [3] and Apostolos Beligiannis [1]. 4.1. An ideal in KK(X). Our starting point is a rough idea of the invariant we want to use. This rough idea is expressed by a homological ideal in the triangulated category. The ideal I in KK(X) relevant for us is defined by (4.1) I(A, B) :=(cid:8)f ∈ KK(X; A, B)(cid:12)(cid:12) It makes no difference if we use LC(X) or LC(X)∗ here. f∗ : K∗(cid:0)A(Y )(cid:1) → K∗(cid:0)B(Y )(cid:1) vanishes for all Y ∈ LC(X)(cid:9). We claim that I is a homological ideal in the triangulated category KK(X); that is, it is the kernel (on morphisms) of a stable homological functor from KK(X) to some stable Abelian category; stability means that the functor intertwines the suspension automorphism on KK(X) with a given suspension automorphism on the target Abelian category. Our starting point is a bare form of filtrated K-theory. Recall the functors FKY : KK(X) → AbZ/2, for Y ∈ LC(X) from Definition 2.1 and let A 7→ K∗(cid:0)A(Y )(cid:1) A 7→(cid:16)K∗(cid:0)A(Y )(cid:1)(cid:17)Y ∈LC(X)∗ F := (FKY )Y ∈LC(X) : KK(X) → YY ∈LC(X)∗ The target categoryQY ∈LC(X)∗ AbZ/2 of F is Abelian and carries an obvious sus- pension functor that shifts the Z/2-grading. The functor F is a stable homological functor, that is, it intertwines the suspension automorphisms and maps exact tri- angles to long exact sequences. By definition, AbZ/2, . (4.2) ker FKY = ker F, I = \Y ∈LC(X)∗ that is, f ∈ I(A, B) if and only if F (f ) = 0. Hence I is a homological ideal with defining functor F . We also have I = ker FK with FK as in Definition 2.4: the two functors F and FK only differ through their target categories. For the time being, we pretend that we do not yet know anything about filtrated K-theory beyond the ideal I it defines. The general machinery will automatically lead us to the functor FK. As explained in [9], the homological ideal I yields various notions of homological algebra. The following descriptions of these notions follow from [9, Lemmas 3.2 and 3.9, Definition 3.21]. • A morphism f ∈ KK∗(X; A, B) is all Y ∈ LC(X); -- I-epic if the induced maps K∗(cid:0)A(Y )(cid:1) → K∗(cid:0)B(Y )(cid:1) are surjective for -- I-monic if the induced maps K∗(cid:0)A(Y )(cid:1) → K∗(cid:0)B(Y )(cid:1) are injective for -- an I-equivalence if the induced maps K∗(cid:0)A(Y )(cid:1) → K∗(cid:0)B(Y )(cid:1) are bijective for all Y ∈ LC(X). all Y ∈ LC(X); • A homological functor F : KK(X) → C to some Abelian category C is I-exact if F (f ) = 0 for all f ∈ I; equivalently, F maps I-epimorphisms to epimorphisms or F maps I-monomorphisms to monomorphisms. C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 19 • An object A ∈∈ KK(X) is -- I-contractible if K∗(cid:0)A(Y )(cid:1) = 0 for all Y ∈ LC(X); -- I-projective if the functor KK∗(X; A, ) is I-exact; equivalently, I(A, B) = 0 for all B ∈∈ KK(X), or: any I-epimorphism B → A splits (see [9] for more equivalent characterisations). • A chain complex · · · → An+1 δn+1−−−→ An δn−→ An−1 δn−1−−−→ An−2 → · · · in KK(X) -- that is, An ∈∈ KK(X) and δn ∈ KK(X; An, An−1) for all n ∈ Z, subject to the condition δn−1 ◦ δn = 0 -- is I-exact (in some degree n) if the induced chain complexes of Z/2-graded Abelian groups · · · → K∗(cid:0)An+1(Y )(cid:1) (δn+1)∗ −−−−−→ K∗(cid:0)An(Y )(cid:1) (δn)∗−−−→ K∗(cid:0)An−1(Y )(cid:1) → · · · are exact (in degree n) for all Y ∈ LC(X). • An I-projective resolution of A ∈∈ KK(X) is an I-exact chain complex · · · → P2 δ2−→ P1 δ1−→ P0 δ0−→ A → 0 → · · · with I-projective entries Pn for all n ∈ N. We shall soon see that there are enough I-projective objects in the sense that any object of KK(X) has an I-projective resolution. Such resolutions are unique up to chain homotopy equivalence once they exist. We use projective resolutions to define derived functors (see [9, Definition 3.27]): just apply the functor to be derived to an I-projective resolution and take homology. In particular, this yields extension groups Extn I(A, B) for all A, B ∈∈ KK(X). Un- like in usual homological algebra, Ext0 I(A, B) may differ from the morphism space in KK(X), compare the exact sequence (4.8) in [6]. 4.2. Enough projective objects. A strategy to find enough projective objects is outlined in [9, §3.6]. The idea is to study the left adjoint functor FK⊢ Y of FKY ; this is defined on P ∈∈ AbZ/2 if there is FK⊢ Y (P ) ∈∈ KK(X) and a natural isomorphism (4.3) for all B ∈∈ KK(X). Notice that FK⊢ Hom(cid:0)P, FKY (B)(cid:1) ∼= KK(X; FK⊢ Y (P ), B) Y need not be defined for all P . Objects of the form FK⊢ Y (P ) are automatically I-projective because the functor KK(X; FK⊢ Y (P ), ) factors through FKY by (4.3) and vanishes on I by (4.2). The simplest case to look for FK⊢ Y (P ) is P = Z[0] (this means Z in degree 0). The defining property of FK⊢ Y (Z[0]) is a natural isomorphism KK(X; FK⊢ Y (Z[0]), B) ∼= Hom(cid:0)Z[0], FKY (B)(cid:1) ∼= FKY,0(B) = K0(cid:0)B(Y )(cid:1). Y (Z[0]) must represent the covariant functor FKY . Theorem 2.5 In other words, FK⊢ provides such representing objects, and yields the following: Proposition 4.1. For any Y ∈ LC(X), the adjoint functor FK⊢ Y is defined on a Z/2-graded Abelian group G = G0 ⊕ G1 if G0 and G1 are free and countable. More precisely, FK⊢ Y Mi∈I Z[εi]! =Mi∈I RY [εi], where I is a countable set and εi ∈ Z/2 for all i ∈ I. Proof. We have just observed that FK⊢ Y (Z[0]) = RY . Since FKY is stable, this implies FK⊢ Y (Z[1]) = RY [1]. It is a general feature of left adjoint functors that they commute with direct sums. Since countable direct sums exist in KK(X), we get the existence of FK⊢ (cid:3) Y on any free countable Z/2-graded Abelian group. 20 RALF MEYER AND RYSZARD NEST Corollary 4.2. There are enough I-projective objects in KK(X), and the class of I-projective objects in KK(X) is generated by the objects RY for Y ∈ LC(X)∗. More precisely, any I-projective objects is a retract of a direct sum of suspensions of these objects. Proof. This follows from Proposition 4.1 and [9, Proposition 3.37]. (cid:3) Often we do not need retracts here, that is, any I-projective object is a direct sum of suspensions of RY for Y ∈ LC(X)∗; for the totally ordered spaces studied in §3, this follows from Theorem 3.12. Since our ideal I is compatible with countable direct sums, the I-contractible objects form a localising subcategory of KK(X), that is, they form a class NI of objects that is closed under countable direct sums, retracts, isomorphism, exact triangles, and suspensions. Furthermore, NI is the complement of the localising subcategory that is generated by the I-projective objects. These two subcategories contain much less information than the ideal itself. Roughly speaking, they will be the same for any reasonable choice of invariant on KK(X) of K-theoretic nature. Proposition 4.3. The localising subcategory that is generated by the I-projective objects is the bootstrap category B(X). It consists of all objects of KK(X) that are KK(X)-equivalent to a tight, nuclear, purely infinite, stable, separable C∗-algebra over X whose simple subquotients belong to the bootstrap category B ⊆ KK. Proof. By definition, B(X) is the localising subcategory of KK(X) that is generated by the objects ix(C) for x ∈ X, see [8]. These generators are I-projective because they represent the functors FKUx, compare the proof of the Representability The- orem 2.5. The proof of this theorem also shows that the representing objects RY belong to the triangulated subcategory of KK(X) generated by RUx for x ∈ X and hence to B(X). Now Corollary 4.2 shows that all I-projective objects belong to B(X). Hence the localising subcategory they generate is contained in the bootstrap class. Conversely, since the generators of the bootstrap class ix(C) are I-projective, the localising subcategory generated by the I-projective objects must contain the whole bootstrap class. This yields the first statement. The second one is contained in [8, Corollary 5.5]. (cid:3) 4.3. The universality of filtrated K-theory. The next step in the general pro- gramme is to determine the universal defining functor for I. This functor is char- acterised by the universal property that it is I-exact and stable homological and that any I-exact homological functor on KK(X) factors through it uniquely (up to natural isomorphism). The advantage of using the universal functor is that it describes I-projective resolutions and the associated I-derived functors in KK(X) by projective resolutions and derived functors in its target Abelian category. This is the crucial step to compute these derived functors. In the presence of enough projective objects, [9, Theorem 3.39] characterises the universal functor by an adjointness property. In our case, this yields: Theorem 4.4. The filtrated K-theory functor FK : KK(X) → Mod(N T )c is the universal I-exact stable homological functor; here Mod(N T )c denotes the category of all countable graded N T -modules. The ring of natural transformations N T comes in automatically at this point. Proof. This is best explained as a special case of a general result on certain homo- logical ideals. Let T be any triangulated category with countable direct sums, and C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 21 let G be an at most countable set of objects of T. Let IG be the stable homological ideal defined by the functor FG : T → YG∈G AbZ, A 7→(cid:0)T∗(G, A)(cid:1)G∈G. We assume that FG(A) is countable for all A ∈∈ T. We are dealing with the case where T = KK(X) and G = {RY Y ∈ LC(X)∗}; Theorem 2.5 identifies T∗(RY , A) = KK∗(X; RY , A) ∼= K∗(cid:0)A(Y )(cid:1) = FKY (A) for all Y ∈ LC(X)∗, so that IG = I with I as in (4.1). Viewing G as a full subcategory of T, it becomes a Z-graded pre-additive cat- egory, so that we get a corresponding category Mod(Gop)c of countable graded right modules. We can enrich the functor FG to a functor G : T → Mod(Gop)c F ′ because the composition in T provides maps T∗(G′, A) ⊗ T∗(G, G′) → T∗(G, A) for all G, G′ ∈ G, A ∈∈ T, which form a right G-module structure on(cid:0)T∗(G, A)(cid:1)G∈G. G is the universal IG-exact functor. We claim that the functor F ′ In the case at hand, our description of the natural transformations FKY ⇒ FKZ in §2.1 means that Mod(Gop)c = Mod(N T )c and F ′ G = FK is filtrated K-theory as defined in Definition 2.4. Hence it suffices to establish the claim above to finish the proof of Theorem 4.4. To do this, we check the conditions in [9, Theorem 3.39]. Idempotent morphisms in KK(X) split because this happens in any triangulated category with countable direct sums (see [10]). Call F ′ G(G) = T( , G) for G ∈ G the free Gop-module on G. Direct sums of free modules are projective, and any object of Mod(Gop)c is a quotient of a countable direct sum of free modules. Hence Mod(Gop)c has enough projective objects. Moreover, HomGop(cid:0)F ′ G(A)(cid:1) ∼= F ′ G(G), F ′ G(A)(G) = T(G, A) shows that the left adjoint F ⊢ of F := F ′ G(G) to G ∈∈ T. Since the domain of F ⊢ is closed under suspensions, countable direct sums, and retracts, the adjoint is defined on all projective modules. Furthermore, F ◦ F ⊢(P ) ∼= P holds for free modules and hence for all projective modules P . Having checked all the hypotheses of [9, Theorem 3.39], we can conclude that F ′ G is indeed universal. (cid:3) G maps F ′ Since FK : KK(X) → Mod(N T )c is universal, [9, Theorem 3.41] now tells us, roughly speaking, that homological algebra in KK(X) with respect to I is equivalent to homological algebra in the Abelian category Mod(N T )c: • An object A of KK(X) is I-projective if and only if FK(A) ∈ Mod(N T )c is projective and KK∗(X; A, B) ∼= HomN T(cid:0)FK(A), FK(B)(cid:1) for all B ∈∈ KK(X). Another equivalent condition is that FK(A) ∈ Mod(N T )c is projective and A belongs to the localising subcategory generated by the I-projective objects; the latter agrees with the bootstrap class by Proposition 4.3. • The functor FK and its partially defined left adjoint FK⊢ restrict to an equivalence of categories between the subcategories of I-projective objects in KK(X) and of projective objects in Mod(N T )c. • For any A ∈∈ KK(X), the functors FK and FK⊢ induce bijections between isomorphism classes of I-projective resolutions of A and isomorphism classes 22 RALF MEYER AND RYSZARD NEST of projective resolutions of FK(A) in Mod(N T )c. That is, a projective res- olution in Mod(N T )c lifts to a unique I-projective resolution in KK(X). This provides the "geometric resolutions" that are used in connection with the usual Universal Coefficient Theorem for KK. • For all n ∈ N, there is a natural isomorphism Extn I(A, B) ∼= Extn N T(cid:0)FK(A), FK(B)(cid:1), where the right hand side denotes extension groups in the Abelian category Mod(N T )c. • For any homological functor G : KK(X) → C, there is a unique right-exact functor ¯G : Mod(N T )c → C with ¯G ◦ FK(P ) = G(P ) for all I-projective P . The left derived functors of G with respect to I are Ln ¯G ◦ FK for n ∈ N, where Ln ¯G : Mod(N T )c → C denotes the nth left derived functor of ¯G. 4.4. The Universal Coefficient Theorem. In the general theory, the next step is to construct a spectral sequence whose E2-term involves the extension groups Extn I(A[m], B); it converges -- in favourable cases -- to KK∗(X; A, B). This spectral sequence is constructed in [3, 6]. Since we aim for an exact sequence, not for a spectral sequence, we only need the special case considered in [9, Theorem 4.4]. This provides the Universal Coefficient Theorem we want under the assumption that FK(A) has a projective resolution of length 1 in Mod(N T )c: Theorem 4.5. Let A, B ∈∈ KK(X). Suppose that FK(A) ∈∈ Mod(N T )c has a projective resolution of length 1 and that A ∈∈ B(X). Then there are natural short exact sequences Ext1 N T(cid:0)FK(A)[j + 1], FK(B)(cid:1) ֌ KKj(X; A, B) ։ HomN T(cid:0)FK(A)[j], FK(B)(cid:1) for j ∈ Z/2, where HomN T and Ext1 in the Abelian category Mod(N T )c and [j] and [j + 1] denote degree shifts. N T denote the morphism and extension groups The bootstrap class appears here because of Proposition 4.3, which identifies it with the localising subcategory generated by the I-projective objects. Corollary 4.6. Let A, B ∈∈ B(X) and suppose that both FK(A) and FK(B) have projective resolutions of length 1 in Mod(N T )c. Then any morphism FK(A) → FK(B) in Mod(N T )c lifts to an element in KK0(X; A, B), and an isomorphism FK(A) ∼= FK(B) lifts to an isomorphism in B(X). Proof. The lifting of a homomorphism follows from Theorem 4.5. Given an iso- morphism f : FK(A) → FK(B), we can lift f and f −1 to elements α and β of KK0(X; A, B) and KK0(X; B, A), respectively. Since β ◦ α lifts the identity map on FK(A), the difference id − β ◦ α belongs to Ext1 latter is a nilpotent ideal in KK(X; A, A) because of the naturality of the exact sequence in Theorem 4.5. Hence (id − βα)2 = 0, so that β ◦ α is invertible. The same argument shows that α ◦ β is invertible, so that α is invertible. (cid:3) N T(cid:0)FK(A)[j + 1], FK(A)(cid:1). The This corollary is what is needed for the classification programme, and it depends on resolutions having length 1. Conversely, if there is A for which FK(A) has no projective resolution of length 1, then it is likely that there exist non-isomorphic B, D ∈∈ B(X) with FK(B) ∼= FK(D). The following theorem provides such a counterexample, but under a stronger assumption. Theorem 4.7. Let I be a homological ideal in a triangulated category T with enough I-projective objects. Let F : T → AIT be a universal I-exact stable homological functor. Suppose that I2 6= 0. Then there exist non-isomorphic objects B, D ∈∈ T for which F (B) ∼= F (D) in AIT. C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 23 Proof. Since I2 6= 0, there is A ∈∈ T with I2(A, ) 6= 0, that is, A is not I2-projective. The ideal I2 has enough projective objects as well, so that there is an exact triangle γ2−→ A2 α2−→ A ι2−→ N2 ΣN2 with ι2 ∈ I2 and an I2-projective object A2 (this is part of the phantom castle constructed in [6], where the same notation is used). Since ι2 ∈ I, this triangle is I-exact and hence provides an extension F (N2)[1] ֌ F ( A2) ։ F (A) in AIT. Even more, this extension splits because ι2 ∈ I2. This follows because the canonical map I(A, N2) → Ext1 I(A, N2[1]) implicitly used above factors through I/I2 and hence annihilates ι2 (see [6, Equa- tion (4.9)]). As a result, F ( A2) ∼= F (A) ⊕ F (N2)[1]. But A2 cannot be isomorphic to A ⊕ N2[1]. If this were the case, then A would be I2-projective, as a retract of the I2-projective object A2. Then I2(A, ) = 0, contradicting our choice of A. Hence A2 6∼= A ⊕ N2[1]. (cid:3) If I2 = 0, then the ABC spectral sequence constructed in [6] degenerates at the third stage, that is, E3 = E∞. But E2 and E3 differ unless projective resolutions have length 1. Hence the vanishing of I2 is probably not sufficient for isomorphisms on the invariant to lift because the boundary map d2 on the second stage of the ABC spectral sequence may provide further obstructions. Whether or not filtrated K-theory gives rise to projective resolutions of length 1 depends on the space in question: we will find positive and negative cases below. Before we turn to examples, we discuss another important issue: does filtrated K-theory exhaust all of Mod(N T )c? This is definitely not the case because of the additional exactness conditions that hold for objects of the form FK(A). The following result is not optimal but sufficient for our purposes. Theorem 4.8. Let G ∈∈ Mod(N T )c have a projective resolution of length 1. Then there is A ∈∈ B(X) with FK(A) ∼= G, and this object is unique up to isomorphism in B(X). Proof. Any projective resolution of length 1 in Mod(N T )c is isomorphic to one of the form · · · → 0 → FK(P1) FK(f ) −−−−→ FK(P0) → G for suitable I-projective objects P1, P0 ∈∈ KK(X) and some f ∈ KK0(X; P1, P0). Here we use that FK restricts to an equivalence of categories between the subcat- egories of I-projective objects of KK(X) and of projective objects of Mod(N T )c by the first paragraph of [9, Theorem 3.41]. We may embed the morphism f in an exact triangle ΣA h−→ P1 f −→ P0 g −→ A. Since FK(f ) is injective, the map f is I-monic; thus g is I-epic and h ∈ I. Therefore, the long exact sequence for FK applied to the above triangle degenerates to a short exact sequence This yields FK(A) ∼= G as desired. The uniqueness of A is already contained in Corollary 4.6. (cid:3) FK(P1) ֌ FK(P0) ։ FK(A). It remains to understand which objects of the category Mod(N T )c have a pro- jective resolution of length 1. 24 RALF MEYER AND RYSZARD NEST 4.5. Resolutions of length 1 in the totally ordered case. We return to the example of the space X = {1, . . . , n} totally ordered by ≤ studied in §3. Let N T be the graded pre-additive category of natural transformations described in §3, and let C = Mod(N T )c be the Abelian category of N T -modules. The following theorem characterises N T -modules with projective resolutions of length 1: Theorem 4.9. Let M ∈∈ C. The following assertions are equivalent: (i) M = FK∗(A) for some A ∈∈ KK(X); (ii) M is exact in the sense of Definition 3.5; (iii) TorN T (iv) M has a free resolution of length 1 in C; (v) M has a projective resolution of length 1 in C; (vi) M has a projective resolution of finite length in C. (N T ss, M ) = 0 for i = 1, 2; i Proof. The exact sequence (1.4) shows that (i) implies (ii). Theorem 4.8 contains the implication (v)=⇒(i), and the implications (iv)=⇒(v)=⇒(vi) are trivial. We will show (ii)=⇒(iii)=⇒(iv) and (vi)=⇒(ii), and this will establish the theorem. First we show that (vi) implies (ii). Let 0 → Pm → · · · → P0 → M be a projective resolution of finite length. By a standard "stabilisation" trick, we can turn this into a free resolution of the same length. Let Zj = ker(Pj → Pj−1) ∼= range(Pj+1 → Pj ). Thus Zm = 0, P0/Z0 ∼= M , and we have exact sequences Zj ֌ Pj ։ Zj−1 because our chain complex is exact. Since Zm = 0, the exactness of the projective modules Pm and Proposition 3.6 show recursively that Zj is exact for j = m − 1, m − 2, . . . , 0, so that M is exact. Thus (vi) implies (ii). Now we prove (ii)=⇒(iii)=⇒(iv). Let P be a countable free module for which there is an epimorphism π : P ։ M , and let K := ker π. We have an extension of N T -modules K ֌ P ։ M . Proposition 3.6 shows that K is exact because P and M are exact. Furthermore, Tori+1(N T ss, M ) ∼= Tori(N T ss, K) for all i ≥ 1 because P is projective. Lemma 3.13 applied to M and K yields Tori(N T ss, M ) = 0 for i = 1, 2 if M is exact, that is, (ii)=⇒(iii). Now assume (iii). The argument above yields Tor1(N T ss, K) = 0. Since P is projective, the Abelian groups P (Y ) are free for all Y ∈ LC(X). The exact sequence in (3.5) yields the same for K(Y ). The criterion in Theorem 3.12.(3) shows that K is projective. (cid:3) Now we combine the existence of projective resolutions of length 1 with The- orem 4.5, which still required this as a hypothesis: Theorem 4.10. Let X be the topological space associated to a totally ordered finite set, and let A and B be C∗-algebras over X. If A ∈∈ B(X), then there is a natural short exact sequence Ext1 N T(cid:0)FK(A)[1], FK(B)(cid:1) ֌ KK∗(X; A, B) ։ HomN T(cid:0)FK(A), FK(B)(cid:1). In particular, any N T -module morphism FK(A) → FK(B) lifts to an element in KK∗(X; A, B). If both A and B belong to the bootstrap class B(X), then an isomorphism FK(A) ∼= FK(B) lifts to a KK-equivalence A ≃ B. Proof. Use Theorem 4.5 and Corollary 4.6 together with the existence of projective resolutions of length 1 ensured by Theorem 4.9. (cid:3) Theorem 4.11. Let X be the topological space associated to a totally ordered finite set, and let A and B be tight, purely infinite, stable, nuclear, separable C∗-algebras over X whose simple subquotients belong to the bootstrap category. Then an iso- morphism FK(A) ∼= FK(B) lifts to an X-equivariant ∗-isomorphism A ∼= B. C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 25 Furthermore, any countable exact N T -modules is the filtrated K-module of some tight, purely infinite, stable, nuclear, separable C∗-algebra over X with simple sub- quotients in the bootstrap category. Proof. A nuclear C∗-algebras over X belongs to the bootstrap category B(X) if and only if its fibres belong to the non-equivariant bootstrap category B (see [8, Corol- lary 4.13]). For a tight C∗-algebra over X, these fibres are the same as the simple It is also shown in [8, Corollary 5.5] that any object of B(X) is subquotients. KK(X)-equivalent to a tight, nuclear, purely infinite, simple, separable C∗-algebra over X whose simple subquotients belong to the bootstrap category B. A deep clas- sification result of Eberhard Kirchberg shows that any KK(X)-equivalence between such objects lifts to an X-equivariant ∗-homomorphism. Now the first assertion fol- lows from Theorem 4.10. The second assertion also uses Theorem 4.8. (cid:3) 5. A counterexample Now we let X := {1, 2, 3, 4} with the partial order 1, 2, 3 < 4 and no relation among 1, 2, 3. Hence the open subsets of X are O(X) =(cid:8)∅, {4}, {1, 4}, {2, 4}, {3, 4}, {1, 2, 4}, {1, 3, 4}, {2, 3, 4}, {1, 2, 3, 4}(cid:9), that is, a non-empty subset is open if and only if it contains 4. The associated directed graph is 4 • 9ssss %❑❑❑❑ • 1 • 2 • 3. We frequently denote subsets of X simply by 124 := {1, 2, 4}, and so on. A C∗-algebra over X is a C∗-algebra A with four distinguished ideals I1 := A(14), I2 := A(24), I3 := A(34), I4 := A(4), such that I1 + I2 + I3 = A and Ii ∩ Ij = I4 for all 1 ≤ i < j ≤ 3 (see [8, Lemma 2.35]). Equivalently, the ideals Ij/I4 for j = 1, 2, 3 decompose A/I4 into a direct sum of three orthogonal ideals. The other distinguished ideals are A(124) = I1 + I2, A(134) = I1 + I3, A(234) = I2 + I3. Any subset of X is locally closed. But a connected locally closed subset is either open or one of the singletons {1}, {2}, and {3}. Hence the set of connected locally closed subsets is LC(X)∗ = {4, 14, 24, 34, 124, 134, 234, 1234, 1, 2, 3}. The order complex Ch(X) is a graph with four vertices 1, 2, 3, 4 and edges joining the first three to the last one: Ch(X) = /.-, ()+1 /.-, ()+2 ▲▲▲▲▲▲ /.-, ()+4 rrrrrr /.-, ()+3 Both maps m, M : Ch(X) → X map the vertices to the corresponding points in X. Whereas M maps the interior of each edge to 4, the map m maps the interior of the edge [j, 4] to j for j = 1, 2, 3. Recall that the space of natural transformations FKY ⇒ FKZ is given by S(Y, Z) := m−1(Y ) ∩ M −1(Z) ⊆ Ch(X). N T ∗(Y, Z) ∼= K∗(cid:0)S(Y, Z)(cid:1), / / % 9 26 RALF MEYER AND RYSZARD NEST Y \Z 4 Z 0 0 0 24 Z 0 Z 0 0 Z[1] 0 0 14 Z Z 0 0 0 0 Z[1] 0 124 34 4 Z Z 0 14 Z 0 24 Z 0 34 Z Z[1] Z 124 Z[1] 0 134 Z[1] 234 Z[1] 0 1234 Z[1]2 Z[1] Z[1] Z[1] 0 Z[1] Z[1] 0 Z[1] 0 0 Z[1] 0 Z[1] Z[1] 0 Z[1] Z[1] Z[1] 1 2 3 134 Z Z 0 Z 0 Z 0 0 0 Z[1] 0 0 3 2 1234 1 234 0 0 0 Z Z Z 0 0 0 Z 0 Z 0 Z Z 0 0 Z Z Z Z Z 0 0 Z Z 0 Z 0 Z 0 Z Z Z Z 0 Z Z Z Z Z 0 Z[1] 0 0 0 Z 0 0 0 0 0 Z Z[1] 0 Table 1. The ring of natural transformations It is straightforward to compute these K-theory groups, and the results are listed in Table 1. Here the rows are labelled by Y , the columns by Z. For instance, the entry Z at (14, 1) means that N T ∗(14, 1) ∼= Z. The trivial 1-dimensional bundle over S(14, 1) generates this group. Hence Remark 2.12 shows that the generator is the natural transformation that we get from the quotient map A(14) ։ A(1). Similar arguments show that all the natural transformations of degree 0 are induced by the familiar restriction and extension ∗-homomorphisms for closed and open subsets. Moreover, the odd natural transformations arise by composing these ∗-homomorphisms with boundary maps in K-theory long exact sequences. All relations that they satisfy are predicted by morphisms of extensions and exactness of the sequences (1.4). The computations in §3 were based on a description of indecomposable morph- isms in the category N T ∗. For the space X in question, these are the maps in the following diagram: i 124 (5.1) 4 i i ;①①①①①①①①① #❋❋❋❋❋❋❋❋❋ i 14 24 34 i i :✉✉✉✉✉✉✉✉✉ $■■■■■■■■■ :✉✉✉✉✉✉✉✉✉ $■■■■■■■■■ i i i 234 134 / 1234 i i %❑❑❑❑❑❑❑❑❑ 9ssssssssss i r :✉✉✉✉✉✉✉✉✉ r $■■■■■■■■■■ r 1 2 3 ◦ δ ◦ δ "❊❊❊❊ <②②②② / 4 ❊❊❊❊ δ ◦ ②②②② Here we write i for the extension transformation for an open subset, r for the restriction transformation for a closed subset, and δ for boundary maps in K-theory long exact sequences. The indecomposable morphisms in (5.1) provide a minimal set of generators for the graded ring N T . To describe N T completely, we list the relations. These are generated by the following: • the cube with vertices 4, 14, . . . , 1234 is a commuting diagram, that is, all the commuting squares involving arrows with label i commute; • the following composite arrows vanish: 124 i−→ 1234 r−→ 3, 134 i−→ 1234 r−→ 2, 234 i−→ 1234 r−→ 1, δ−→ 4 i−→ 14, 1 δ−→ 4 i−→ 24, 2 δ−→ 4 i−→ 34; 3 • the sum of the three maps 1234 → 4 via 1, 2, and 3 vanishes. / / $ % " ; / / # : $ / : / / $ / / / : 9 < C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 27 These relations imply that the diagrams 124 r 1 r − δ ◦ 2 ◦ δ / 4 134 r 1 r − δ ◦ 3 ◦ δ / 4 234 r 3 r − δ ◦ 2 ◦ δ / 4 anti-commute and that the composite of two odd maps vanishes. It is routine to check that the universal pre-additive category with these generators and relations is given by the groups listed in Table 1. Define N T nil and N T ss as in Definition 3.3: N T nil is the linear span of the groups N T ∗(Y, Z) with Y 6= Z and N T ss is spanned by the groups N T ∗(Y, Y ). Then N T nil is a nilpotent ideal in N T and N T ss ∼= ZLC(X)∗ is a semi-simple ring. Thus N T nil is the maximal nilpotent ideal in N T and we have a semi-direct product decomposition N T ∼= N T nil ⋊ N T ss as in Lemma 3.4. The next task is to describe the submodule M ′ := N T nil · M ⊆ M for an exact N T -module M . The following computations are done as in the proof of Lemma 3.8, using (3.4) and that the morphisms in (5.1) generate N T . and symmetrically for 24 and 34; where δ4 and M ′(234). We have 124 denotes a generator of N T 1(124, 4) ∼= Z; symmetry provides M ′(134) M ′(14) = range(cid:0)i14 M ′(124) = range(cid:0)i124 14 : M (14) → M (1)(cid:1), 4 : M (4) → M (14)(cid:1) = ker(cid:0)r1 24 : M (24) → M (124)(cid:1) 14 : M (14) → M (124)(cid:1) +(cid:0)i124 124 : M (124) → M (4)(cid:1), = ker(cid:0)δ4 : M (1) → M (234)(cid:1), 1234 : M (1234) → M (1)(cid:1) = ker(cid:0)δ234 M ′(1) = range(cid:0)r1 : M (4) → M (1234)(cid:1). j : M (j) → M (4)(cid:1) = ker(cid:0)i1234 range(cid:0)δ4 3Xj=1 and symmetrically for 2 and 3, and M ′(4) = But something goes wrong with M ′(1234). Equation (3.4) yields 1 4 range(cid:0)i1234 124 : M (124) → M (1234)(cid:1) +(cid:0)i1234 134 : M (134) → M (1234)(cid:1) = ker(cid:0)δ14 1234 : M (1234) → M (14)(cid:1); to take into account the range of i1234 1234 ◦ i1234 δ14 exact sequence, our method breaks down at this point. 234 as well, we need an exact sequence containing 234 , which is the generator of N T 1(234, 14) ∼= Z. Since there is no such Another symptom but not a cause of problems is that the map δ4 M ′(124) is not the longest map out of 124: that would be δ34 124. 124 that describes As we shall see, the analogues of Theorems 3.12 and 4.9 become false for the space X. First, there is a non-projective exact module M with free Mss; secondly, there is a module that has no projective resolution of length 1; thirdly, there are A, B ∈ B(X) with I2(A, B) 6= 0. Hence Theorem 4.7 provides non-isomorphic objects in the bootstrap class B(X) with isomorphic filtrated K-theory. The con- struction of these counterexamples follows the above pattern: first we find a counter- example to Theorem 3.12, which we use to find one for Theorem 4.9, which is then used to find an example as in Theorem 4.7. We begin with the unexpected non-projective module. Let PY for Y ∈ LC(X)∗ denote the free N T -module on Y , that is, PY (Z) = N T ∗(Y, Z), HomN T (PY , N ) ∼= N (Y ) / / / / / / / / / 28 RALF MEYER AND RYSZARD NEST for any Y, Z ∈ LC(X)∗ and any N T -module N . A natural transformation FKY ⇒ FKZ corresponds to an element in N T ∗(Y, Z) ∼= PY (Z) ∼= HomN T (Pz, PY ) and thus induces a module homomorphism PZ → PY in the opposite direction. Hence the three arrows 124, 134, 234 → 1234 in (5.1) induce a module homomorphism j : P1234 → P 0 := P124 ⊕ P134 ⊕ P234. Table 1 shows that there are no module homomorphisms P 0 → P1234, that is, no non-zero natural transformations from 1234 to 124, 134, or 234. The crucial observation is that j is a monomorphism, so that P1234 becomes a submodule of P 0. Since the longest natural transformations out of 1234 are those to 14, 24 and 34, this follows from the elementary observations that the maps N T ∗(1234, j4) → N T ∗(1234 \ j, j4) for j = 1, 2, 3 are, respectively, the identity map on Z. This follows from the exactness of free modules because N T ∗(j, j4) = 0 by Table 1. We describe the quotient by its values M (Y ) for Y ∈ LC(X)∗ as in (5.1): M := P 0/j(P1234) (5.2) Z[1] i ;✇✇✇✇✇✇✇✇✇ i #●●●●●●●●● i 0 0 0 i i i =③③③③③③③③③ !❉❉❉❉❉❉❉❉❉ =③③③③③③③③③ !❉❉❉❉❉❉❉❉❉ i i i Z Z Z i i "❋❋❋❋❋❋❋❋❋ <①①①①①①①①①① i / Z2 r <①①①①①①①①① r "❋❋❋❋❋❋❋❋❋❋ r Z Z Z ●●●● ●●●● ◦ δ ●●●● ●●●● δ ◦ δ ✇✇✇✇ ✇✇✇✇ ◦✇✇✇✇ ✇✇✇✇ Z[1] The boundary maps δ act by isomorphisms on M because M (j4) = 0 for j = 1, 2, 3. The other maps can be understood by writing M (1234) = Z3/h(1, 1, 1)i and M (j) = Z2/h(1, 1)i for j = 1, 2, 3 as quotients. The three maps Z → Z2 correspond to the three coordinate embeddings Z ֌ Z3, the maps Z2 → Z to the projections Z3 ։ Z2 onto coordinate hyperplanes. The projective resolution 0 → P1234 (5.3) does not split because there exist no non-zero morphisms P 0 → P1234. Hence M is not projective. But Mss is free, and M is exact because the exact modules form an exact category and P1234 and P 0 are exact. Thus M is a counterexample to Theorem 3.12. j −→ P 0 ։ M The module M is directly related to the problem with describing N T nil ·M (1234) encountered above. Since HomN T (PY , N ) ∼= N (Y ) for any N T -module N and any Y ∈ LC(X)∗, the resolution (5.3) provides an exact sequence 0 → HomN T (M, N ) → N (124) ⊕ N (134) ⊕ N (234) → N (1234) → Ext1 N T (M, N ) → 0, so that Ext1 N T (M, N ) ∼= N (1234)/N T nil · N (1234) ∼= Nss(1234). Now we use M to construct a counterexample for Theorem 4.9. Let k ∈ N≥2 and let Mk := M/k · M ; that is, we replace Z by Z/k everywhere in (5.2). This module has a projective resolution of length 2 of the form (5.4) 0 → P1234 (−k,j) −−−−→ P1234 ⊕ P 0 (j,k) −−−→ P 0 ։ Mk, / / ! " ; / / # = ! / < / / " / / = < C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 29 where k denotes multiplication by k. Using this resolution, we compute Ext2(Mk, P1234) ∼= Z/k, Ext1(Mk, P1234) ∼= Hom(Mk, P1234) ∼= 0 because there are no no-zero morphisms P 0 → P1234. Of course, the generator of Ext2(Mk, P1234) is the class of the projective resolution (5.4). Hence Mk admits no projective resolution of length 1 and is a counterexample to Theorem 4.9. Now we claim that Mk is the filtrated K-theory of some C∗-algebra Ak over X in the bootstrap class B(X). To begin with, M is the filtrated K-theory of some such C∗-algebra A by Theorem 4.8. Let Bk be a C∗-algebra in the bootstrap class with K0(Bk) = Z/k and K1(Bk) = 0; for instance, Bk could be the Cuntz algebra Ok+1. Then Ak := A ⊗ Bk has filtrated K-theory Mk by the Kunneth Theorem for the K-theory of tensor products. Theorem 5.1. Let Ak be a C∗-algebra in the bootstrap class with FK(Ak) ∼= Mk as constructed above. Then Ak is not I2-projective. Hence there exist B, D ∈ B(X) that are not KK(X)-equivalent but with the same filtrated K-theory. Proof. The second assertion follows from the first one using Theorem 4.7 applied to the bootstrap class B(X) and the restriction of I to B(X). It remains to prove that Ak cannot be I2-projective. To see this, we lift the resolution (5.4) to an I-projective resolution 0 ◦ /P2 ◦ /P1 ◦ / P0 /Ak in B(X) with boundary maps of degree 1, and embed the latter in a phantom tower (see [6]): ι1 0 Ak N0 Z✺✺✺✺✺✺ π0 P0 / N1 ◦✠✠✠ ✠✠✠ Z✺✺✺✺✺✺ π1 ι2 1 P1 N2 Z✺✺✺✺✺✺ π2 ◦✠✠✠ ✠✠✠ ι3 2 P2 ◦✠✠✠ ✠✠✠ N3 Y✸✸✸✸✸✸ · · · N3 ◦☛☛☛ ☛☛☛ 0 · · · The inductive system (Nj, ιj+1 ) becomes constant at N3 because Pj = 0 for j ≥ 3. Since Ak belongs to the bootstrap class, N3 ∼= 0 (see the proof of [6, Proposition 4.5]). This implies N2 ∼= P2. j The composite map ι2 were I2-projective. Then ι2 sequence would yield that the map ι2 N1 → P0. But 1 ◦ ι1 0 : Ak = N0 → N2 ∼= P2 belongs to I2. Suppose that Ak 0 would vanish, and the long exact homology 1 : N1 → N2 must factor through the map 0 = ι2 KK∗(X; P0, P2) ∼= HomN T(cid:0)FK(P0), FK(P2)(cid:1) = HomN T (P 0, P1234) = 0. Here we have used that filtrated K-theory, by universality, is fully faithful on I-projective objects and that there are no non-zero module homomorphisms P 0 → P1234. Since ι2 1 factors through the zero group, it must be the zero map. But then the map P1 → N1 must be a split surjection, so that N1 is I-projective. Then the I-exact triangle ΣAk → ΣN1 → P0 → Ak provides an I-projective resolution of Ak of length 1, which is impossible because FK(Ak) ∼= Mk has no projective resolution of length 1. As a consequence, Ak is not I2-projective. (cid:3) We can make the two non-equivalent C∗-algebras over X with the same filtrated K-theory more explicit. One of them is Ak ⊕ ΣR1234, the other one is the mapping 0 : Ak = N0 → N2 ∼= R1234 in the phantom tower above. Both cone of the map ι2 have Mk ⊕ P1234[1] as their filtrated K-theory. This counterexample shows that filtrated K-theory does not yet classify purely infinite stable nuclear separable C∗-algebras in the bootstrap class. / / / / / / / / / Z Z o o Z o o Y o o o o 30 RALF MEYER AND RYSZARD NEST Remark 5.2. Refining filtrated K-theory by taking filtrated K-theory with coeffi- cients does not help. This gets rid of the counterexample Ak constructed above, but other objects of B(X) without projective resolution of length 1 remain. An ex- ample is A ⊗ B, where B is a C∗-algebra in the bootstrap class with K∗(B) = Q[0] such as an appropriate UHF-algebra. Its filtrated K-theory is M ⊗ Q. This also has cohomological dimension 2, and this is not affected much by taking K-theory with coefficients because M ⊗ Q is torsion-free. 5.1. A refined invariant. There are at least two ways to identify the source of the problem for the space X. The first point of view is that what is missing is an exact sequence that has the generator α of N T 1(234, 14) as its connecting map. The map α corresponds to a map ΣR14 → R234 between the representing objects, which we also denote by α. In the triangulated category KK(X), we can embed the latter map in an exact triangle (5.5) ΣR14 α−→ R234 → R12344 → R14. The notation R12344 will be explained later. The functors these objects represent sit in a long exact sequence (5.6) · · · → FK14 → FK12344 → FK234 α−→ FK14[1] → · · · which is precisely what we want. The second point of view is that the troublemaker is the non-projective module M . Since M has a projective resolution of length 1, there is a unique object in the bootstrap class with filtrated K-theory M . Actually, this yields the same object as the first point of view: Lemma 5.3. The non-projective module M above agrees with FK(R12344). Proof. The map FKY (α) vanishes for almost all Y ∈ LC(X)∗ simply because the graded groups involved have different parity or one of them vanishes. The only exception is Y = 14. The group FK14(R14) = N T (14, 14) is generated by the identity natural transformation. Since α is the generator of N T 1(234, 14), the map FK14(α) is invertible. Now we apply FK to the long exact sequence for the given exact triangle. Since FK(α) vanishes on most Y and is invertible for Y = 14, we can easily compute the groups FKY (R12344). We get the same groups as for the module M . It remains to check that the isomorphism can be chosen as an N T -module homomorphism. The main step is to check that the map Z2 ∼= FK124(R12344) ⊕ FK134(R12344) → FK1234(R12344) ∼= Z2 is invertible. Together with the known relations between the various natural trans- formations, this implies the assertion. We omit the details of this computation. (cid:3) The representing object R12344 is an algebra of functions on a two-dimensional simplicial complex, which we do not describe here because it is not illuminating. The functor that it represents, however, can be described rather nicely as follows. Let A be a C∗-algebra over X. Pull back the extension A(14) ֌ A(124) ։ A(2) along the quotient map A(234) ։ A(2) to an extension A(14) ֌ A(12344) ։ A(234). The object R12344 represents the functor (5.7) KK∗(X; R12344, A) ∼= K∗(cid:0)A(12344)(cid:1). To see this, two observations are necessary. First, K∗(cid:0)R12344(12344)(cid:1) ∼= Z; the generator of this group yields a natural transformation between the two functors in (5.7). Secondly, this natural transformation is invertible. This follows from the Five Lemma, once we know that it extends the known natural isomorphisms KK∗(X; RY , A) ∼= K∗(cid:0)A(Y )(cid:1) C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 31 for Y = 14 and Y = 234 to a chain map between the long exact sequences that we get from (5.5) and from the extension A(14) ֌ A(12344) ։ A(234). This extension also explains the notation R12344. Now we augment filtrated K-theory by adding the covariant functor B 7→ FK12344(B) := K∗(cid:0)A(12344)(cid:1) ∼= KK∗(X; R12344, B). The new invariant takes values in the category of countable N T ′-modules, where N T ′ is the Z/2-graded category whose object set is LC′ := LC(X)∗ ⊔ {12344} and whose morphisms are the natural transformations between the various filtrated K-groups, including now also FK12344. These natural transformations can be com- puted by the Yoneda Lemma: N T ′ ∗(Y, Z) ∼= KK∗(X; RZ, RY ) ∼= FKZ(RY ) holds for all Y, Z ∈ LC′. The category ring for N T ′ where ∗ is simply the ring KK∗(X; R, R) R := MY ∈LC′ RY . We replace the ideal I in KK(X) studied above by the kernel I′ of the enriched filtrated K-theory functor FK′ : KK(X) → Mod(N T ′)c. The same arguments as above show that there are enough I′-projective objects and that FK′ is the universal I′-exact stable homological functor. The passage from I to I′ has improved the situation because R12344 has now been promoted to an I′-projective object and, therefore, ceases to cause trouble. In principle, something similar can be done in great generality: whenever we have an object of the Abelian approximation that has a projective resolution of length 1, we can lift it uniquely to an object of the triangulated category and refine the ideal by intersecting it with the kernel of the functor this lifted object represents. However, the policy to quieten troublemakers by promotion has the tendency to encourage new troublemakers, so that it is not clear whether this general strategy always resolves all problems after finitely many steps. But in the relatively simple example at hand, this turns out to be the case. To check this, we must describe the category N T ′. If Y, Z ∈ LC(X)∗, then ∗(Y, Z) = N T ∗(Y, Z) is given by the table on page 26. Furthermore, if Z ∈ ∗(12344, Z) ∼= FKZ (R12344) = M (Z) by Lemma 5.3, and this is N T ′ LC(X)∗, then N T ′ described in (5.2). The upshot is: • there are even natural transformations from FK12344 to FK124, FK134, FK234 -- the generators of the respective groups of natural transformations -- such that any natural transformation FK12344 ⇒ FKZ with Z ∈ LC(X)∗ is a sum of natural transformations that factor through one of these three and a natural transformation FKij4 ⇒ FKZ ; • the sum of the three natural transformations FK12344 ⇒ FK1234 via FK124, FK134 and FK234 vanishes, and all other relations follow from these and the already known ones listed after (5.1). The exact triangle (5.5) yields a long exact sequence · · · → N T ′ ∗+1(Y, 234) α−→ N T ′ ∗(Y, 14) → N T ′ ∗(Y, 234) → · · · , which we may use to compute N T ′ ∗(Y, 12344) for all Y ∈ LC′. The map α induces an isomorphism for Y = 234 and the zero map for all other Y because the source and target have opposite parity or one of them vanishes. Thus ∗(Y, 12344) → N T ′ Y N T ′ 4 ∗(Y, 12344) Z2 14, 24, 34 124, 134, 234 1234 1, 2, 3 12344 Z 0 Z[1] Z[1] Z 32 RALF MEYER AND RYSZARD NEST These groups inherit from M their invariance under permutations of 1, 2, 3. Inspect- ing composition with natural transformations in N T , we arrive at the following: • there are even natural transformations FKj4 ⇒ FK12344 for j = 1, 2, 3, such that any natural transformation FKY ⇒ FK12344 with Y ∈ LC(X)∗ factors through one of them; • the sum of the three natural transformations FK4 ⇒ FK12344 vanishes, • the natural transformations FKj4 ⇒ FK1234\j via FK12344 vanish; • all other relations follow from these and the already known ones. As one may expect, the basic natural transformations FK14 ⇒ FK12344 ⇒ FK234 are induced by the maps R234 → R12344 → R14 in the exact triangle (5.5). The indecomposable morphisms of the new category N T ′ are the maps in the following diagram: 4 ?⑦⑦⑦⑦⑦⑦⑦⑦ ❅❅❅❅❅❅❅ 14 24 34 #❋❋❋❋❋❋❋❋ ;①①①①①①①①① / 12344 134 / 1234 ;✈✈✈✈✈✈✈✈✈ #❍❍❍❍❍❍❍❍❍ #❋❋❋❋❋❋❋❋ ;①①①①①①①① 124 234 =③③③③③③③③ !❉❉❉❉❉❉❉❉❉ 1 2 3 ❃❃❃❃ ◦ ❃❃❃❃ / 4 @ ◦ ◦ The category ring of N T ′ again has the by now familiar structure: it is a split spanned by the identity nil that is the subgroup ∼= ZLC′ nilpotent extension of the semisimple algebra N T ′ ss transformations on the objects and a nilpotent ideal N T ′ generated by N T ′(Y, Z) with Y 6= Z. Definition 5.4. A module over N T ′ is exact if it is exact as an N T -module and the three sequences · · · → N∗+1(ij4) → N∗(k4) → N∗(12344) → N∗(ij4) → · · · for {i, j, k} = {1, 2, 3} are exact as well. The range of the invariant FK′ consists of exact N T ′-modules; the three new exact sequences are, in fact, equivalent for symmetry reasons, and the extension · · · → N∗+1(234) → N∗(14) → N∗(12344) → N∗(234) → · · · is built into the definition of FK12344. Let N be an exact N T ′-module and let N ′ := N T ′ nil · N . The description of N ′(14), N ′(1), and N ′(4) is the same as for the category N T , so that these groups remain kernels of certain maps, as needed. Furthermore, N ′(1234) is the kernel of the map N (1234) → N (12344)[1] induced by the generator of N T 1(1234, 12344), so that the problem that appeared for the category N T is cured. The computation of N ′(124) changes because this group is now the range of the arrow N (12344) → N (124). But this is part of a long exact sequence because N is exact, and we get and similarly for N ′(134) and N ′(234). N ′(124) = ker(cid:0)N (124) → N (34)[1](cid:1), Finally, N ′(12344) is the sum of the ranges of the maps N (j4) → N (12344) for j = 1, 2, 3. Using exactness, we identify this in two steps with the kernel of the map N (12344) → N (4)[1] induced by the generator of N T ′ 1(12344, 4). As a result, the submodule N T ′ nil · N is described using kernels of maps N (Y ) → N (Z). By the way, these arrows are the longest arrows starting at Y as in Re- mark 3.9. The same arguments as for totally ordered spaces now show: # # ? / / / / / ; # / = / / ! / ; ; @ C∗-ALGEBRAS OVER TOPOLOGICAL SPACES: FILTRATED K-THEORY 33 Theorem 5.5. An N T ′-module N is free if and only if it is projective, if and only if it is exact and N (Y ) is a free group for all Y ∈ LC′. Theorem 5.6. An N T ′-module N has a projective resolution of length 1 if and only if it is exact. Theorem 5.7. Let A and B be C∗-algebras over the four-point space X under If A belongs to the bootstrap class B(X), then there is a natural consideration. short exact sequence N T ′(cid:0)FK′(A)[1], FK′(B)(cid:1) ֌ KK∗(X; A, B) ։ HomN T ′(cid:0)FK′(A), FK′(B)(cid:1). In particular, morphisms FK′(A) → FK′(B) lift to elements in KK∗(X; A, B). If both A and B belong to the bootstrap class, then an isomorphism FK′(A) ∼= FK′(B) lifts to a KK(X)-equivalence. Corollary 5.8. The map A 7→ FK′(A) is a bijection between the set of isomorphism classes of tight, stable, purely infinite, separable, nuclear C∗-algebras over X with simple subquotients in the bootstrap class and the set of isomorphism classes of countable exact N T ′-modules. Ext1 6. Conclusion We have obtained a Universal Coefficient Theorem that computes KK∗(X; A, B) for A in the bootstrap class and X of a very special form, namely, {1, . . . , n} with the Alexandrov topology from the total order. This Universal Coefficient Theorem can be used to carry over classification results for simple, nuclear, purely infinite C∗-algebras to nuclear, purely infinite C∗-algebras with primitive ideal space X, using filtrated K-theory as the invariant. For general finite topological spaces X, we still get a spectral sequence that computes KK∗(X; A, B) using filtrated K-theory, but this spectral sequence need not degenerate to an exact sequence, so that isomorphisms on filtrated K-theory need not lift to X-equivariant KK-equivalences. In fact, we have found a counter- example. At the same time, we were able to fix the counterexample by refining filtrated K-theory. It is unclear whether such a refinement is available for all finite topological spaces and how it looks like. [1] Apostolos Beligiannis, Relative homological algebra and purity in triangulated categories, J. References Algebra 227 (2000), no. 1, 268 -- 361, doi: 10.1006/jabr.1999.8237. MR 1754234 [2] Alexander Bonkat, Bivariante K-Theorie fur Kategorien projektiver von C ∗-Algebren, Ph.D. Thesis, Westf. Wilhelms-Universitat Munster, http://deposit.ddb.de/cgi-bin/dokserv?idn=967387191 (German). Systeme 2002, [3] J. Daniel Christensen, Ideals in triangulated categories: phantoms, ghosts and skeleta, Adv. Math. 136 (1998), no. 2, 284 -- 339, doi: 10.1006/aima.1998.1735. MR 1626856 [4] Samuel Eilenberg and John Coleman Moore, Foundations of relative homological algebra, Mem. Amer. Math. Soc. No. 55 (1965), 39. MR 0178036 [5] Eberhard Kirchberg, Das nicht-kommutative Michael-Auswahlprinzip und die Klassifikation nicht-einfacher Algebren, C ∗-Algebras (Munster, 1999), Springer, Berlin, 2000, pp. 92 -- 141 (German). MR 1796912 [6] Ralf Meyer, Homological algebra in bivariant K-theory and other triangulated categories. II, Tbil. Math. J. 1 (2008), 165 -- 210. MR 2563811 [7] Ralf Meyer and Ryszard Nest, The Baum -- Connes conjecture via localisation of categories, Topology 45 (2006), no. 2, 209 -- 259, doi: 10.1016/j.top.2005.07.001. MR 2193334 [8] [9] , C ∗-Algebras over topological spaces: the bootstrap class, Munster J. Math. 2 (2009), 215 -- 252. MR 2545613 , Homological algebra in bivariant K-theory and other triangulated categories. I, Tri- angulated categories (Thorsten Holm, Peter Jørgensen, and Raphael Rouqier, eds.), London Math. Soc. Lecture Note Ser., vol. 375, Cambridge Univ. Press, Cambridge, 2010, pp. 236 -- 289. MR 2681710 34 RALF MEYER AND RYSZARD NEST [10] Amnon Neeman, Triangulated categories, Annals of Mathematics Studies, vol. 148, Princeton University Press, Princeton, NJ, 2001. MR 1812507 [11] Gunnar Restorff, Classification of Cuntz -- Krieger algebras up to stable isomorphism, J. Reine Angew. Math. 598 (2006), 185 -- 210, doi: 10.1515/CRELLE.2006.074. MR 2270572 [12] , Classification of Non-Simple C∗-Algebras, Ph.D. Thesis, Københavns Universitet, 2008, http://www.math.ku.dk/~restorff/papers/afhandling_med_ISBN.pdf. [13] Mikael Rørdam, Classification of extensions of certain C ∗-algebras by their six term exact sequences in K-theory, Math. Ann. 308 (1997), no. 1, 93 -- 117, doi: 10.1007/s002080050067. MR 1446202 [14] Steven Vickers, Topology via logic, Cambridge Tracts in Theoretical Computer Science, vol. 5, Cambridge University Press, Cambridge, 1989. MR 1002193 Mathematisches Institut and Courant Research Centre "Higher Order Structures", Georg-August Universitat Gottingen, Bunsenstrasse 3 -- 5, 37073 Gottingen, Germany E-mail address: [email protected] Københavns Universitets Institut for Matematiske Fag, Universitetsparken 5, 2100 København, Denmark E-mail address: [email protected]
1510.05829	4	1510	2016-08-03T12:36:34	Gauge-invriant quasi-free states on the algebra of the anyon commutation relations	[ "math.OA" ]	Let $X=\mathbb R^2$ and let $q\in\mathbb C$, $\|q\|=1$. For $x=(x^1,x^2)$ and $y=(y^1,y^2)$ from $X^2$, we define a function $Q(x,y)$ to be equal to $q$ if $x^1<y^1$, to $\bar q$ if $x^1>y^1$, and to $\Re q$ if $x^1=y^1$. Let $\partial_x^+$, $\partial_x^-$ ($x\in X$) be operator-valued distributions such that $\partial_x^+$ is the adjoint of $\partial_x^-$. We say that $\partial_x^+$, $\partial_x^-$ satisfy the anyon commutation relations (ACR) if $\partial^+_x\partial_y^+=Q(y,x)\partial_y^+\partial_x^+$ for $x\ne y$ and $\partial^-_x\partial_y^+=\delta(x-y)+Q(x,y)\partial_y^+\partial^-_x$ for $(x,y)\in X^2$. In particular, for $q=1$, the ACR become the canonical commutation relations and for $q=-1$, the ACR become the canonical anticommutation relations. We define the ACR algebra as the algebra generated by operator-valued integrals of $\partial_x^+$, $\partial_x^-$. We construct a class of gauge-invariant quasi-free states on the ACR algebra. Each state from this class is completely determined by a positive self-adjoint operator $T$ on the real space $L^2(X,dx)$ which commutes with any operator of multiplication by a bounded function $\psi(x^1)$. In the case $\Re q<0$, the operator $T$ additionally satisfies $0\le T\le -1/\Re q$. Further, for $T=\kappa^2\mathbf 1$ ($\kappa>0$), we discuss the corresponding particle density $\rho(x):=\partial_x^+\partial_x^-$. For $\Re q\in(0,1]$, using a renormalization, we rigorously define a vacuum state on the commutative algebra generated by operator-valued integrals of $\rho(x)$. This state is given by a negative binomial point process. A scaling limit of these states as $\kappa\to\infty$ gives the gamma random measure, depending on parameter $\Re q$.	math.OA	math	Gauge-invariant quasi-free states on the algebra of the anyon commutation relations Eugene Lytvynov Department of Mathematics, Swansea University, Singleton Park, Swansea SA2 8PP, U.K.; e-mail: [email protected] Abstract y ∂− x ∂+ x (x ∈ X) be operator-valued distributions such that ∂+ Let X = R2 and let q ∈ C, q = 1. For x = (x1, x2) and y = (y1, y2) from X 2, we define a function Q(x, y) to be equal to q if x1 < y1, to ¯q if x1 > y1, and to (cid:60)q if x1 = y1. Let x is the adjoint of ∂− x , ∂− ∂+ x . We x , ∂− y ∂+ y = Q(y, x)∂+ x satisfy the anyon commutation relations (ACR) if ∂+ say that ∂+ x for x for (x, y) ∈ X 2. In particular, for q = 1, the y = δ(x − y) + Q(x, y)∂+ x (cid:54)= y and ∂− x ∂+ ACR become the canonical commutation relations and for q = −1, the ACR become the canonical anticommutation relations. We define the ACR algebra as the algebra generated by operator-valued integrals of ∂+ x . We construct a class of gauge-invariant quasi-free states on the ACR algebra. Each state from this class is completely determined by a positive self-adjoint operator T on the real space L2(X, dx) which commutes with any operator of multiplication by a bounded function ψ(x1). In the case (cid:60)q < 0, the operator T additionally satisfies 0 ≤ T ≤ −1/(cid:60)q. Further, for T = κ21 (κ > 0), we discuss the corresponding x . For (cid:60)q ∈ (0, 1], using a renormalization, we rigorously define particle density ρ(x) := ∂+ a vacuum state on the commutative algebra generated by operator-valued integrals of ρ(x). This state is given by a negative binomial point process. A scaling limit of these states as κ → ∞ gives the gamma random measure, depending on parameter (cid:60)q. x , ∂− x ∂− Keywords: Anyon commutation relations; gauge-invariant quasi-free state; parti- cle density; negative binomial point process; gamma random measure. 2010 MSC: 47L10, 47L60, 47L90, 60G55, 60G57, 81R10 1 Preliminaries and introduction The main aim of this paper is to construct a class of gauge-invariant quasi-free states on the algebra of the anyon commutation relations. Let us first recall the definition of the anyon commutation relations and their representation in the anyon Fock space. 1.1 Fock space representation of the anyon commutation re- lations Let X := Rd, let B(X) denote the Borel σ-algebra on X, and let m denote the Lebesgue measure on (X, B(X)). We denote by H := L2(X, m), HC := L2(X → C, m) 1 the L2-space of real-valued, respectively complex-valued functions on X. (The scalar product in HC is supposed to be linear in the first dot and antilinear in the second dot.) Consider a function Q : X 2 → C satisfying Q(x, y) = Q(y, x) and Q(x, y) = 1 for all (x, y) ∈ X 2. In 1995, Liguori and Mintchev [34, 35] introduced the notion of a generalized statistics corresponding to the function Q. Heuristically, this is a family of creation operators ∂+ x is the adjoint of ∂− x at points x ∈ X such that ∂+ x and these operators satisfy the following commutation relations: x and annihilation operators ∂− ∂+ x ∂+ ∂− x ∂− ∂− x ∂+ y = Q(y, x)∂+ y = Q(y, x)∂− y = δ(x − y) + Q(x, y)∂+ y ∂+ x , y ∂− x , y ∂− x . (1) (2) (3) (Formula (2) is, in fact, a consequence of (1).) A rigorous meaning of the operators x and ∂− ∂+ x and the commutation relations (1) -- (3) is given by smearing these relations with functions from the space HC. More precisely, for any h ∈ HC, one defines linear operators a+(h) = m(dx) h(x) ∂+ x , a−(h) = m(dx) h(x) ∂− x (4) (cid:90) X on a dense linear subspace Θ of a complex Hilbert space G such that the adjoint of a+(h) restricted to Θ is a−(h), and these operators satisfy the commutation relations: (cid:90) X y ∂+ x , y ∂− x , (cid:90) (cid:90) (cid:90) X 2 X 2 a+(g)a+(h) = a−(g)a−(h) = m⊗2(dx dy) g(x)h(y)Q(y, x)∂+ m⊗2(dx dy) g(x)h(y) Q(y, x)∂− (cid:90) (5) (6) a−(g)a+(h) = m⊗2(dx dy) g(x) h(y)Q(x, y)∂+ y ∂− x X X 2 g(x) h(x) m(dx) + (7) for any g, h ∈ HC. Of course, the linear operators on the right hand side of formulas (5) -- (7) should be given a rigorous meaning. In the case Q ≡ 1, formulas (5) -- (7) be- come the canonical commutation relations (CCR), describing bosons, while in the case Q ≡ −1, formulas (5) -- (7) become the canonical anticommutation relations (CAR), de- scribing fermions. In the general case, we will call (5) -- (7) the Q-commutation relations (Q-CR). Liguori and Mintchev [34, 35] derived a representation of the Q-CR in the Fock space of Q-symmetric functions. By using also [10], we will now briefly recall this construction. A function f (n) : X n → C is called Q-symmetric if for any i ∈ {1, . . . , n − 1} and (x1, . . . , xn) ∈ X n. f (n)(x1, . . . , xn) = Q(xi, xi+1)f (n)(x1, . . . , xi−1, xi+1, xi, xi+2, . . . , xn). (8) 2 For each n ∈ N, we have H ⊗nC = L2(X n → C, m⊗n). We denote by H (cid:126)nC the subspace of H ⊗nC which consists of all (m⊗n-versions of) Q-symmetric functions from H ⊗nC . We call H (cid:126)nC Consider the group Sn of all permutations of 1, . . . , n. For each π ∈ Sn, we define a function Qπ : X n → C by the n-th Q-symmetric tensor power of HC. Qπ(x1, . . . , xn) := Q(xi, xj). (9) (cid:89) 1≤i<j≤n π(i)>π(j) Note that, in the case Q ≡ 1, we get Qπ ≡ 1, while in the case Q ≡ −1, we get Qπ ≡ (−1)π = sgn π. Here π is the number of inversions of π, i.e., the number of i < j such that π(i) > π(j). For a function f (n) : X n → C, we define its Q-symmetrization by (Pnf (n))(x1, . . . , xn) := 1 n! Qπ(x1, . . . , xn)f (n)(xπ−1(1), . . . , xπ−1(n)). (10) (cid:88) π∈Sn The operator Pn determines the orthogonal projection of H ⊗nC more, for any k, n ∈ N, k < n, we have onto H (cid:126)nC . Further- Pn(Pk ⊗ Pn−k) = Pn. Here P1 denotes the identity operator in HC. For any f (n) ∈ H (cid:126)nC we define the Q-symmetric tensor product of f (n) and g(m) by f (n) (cid:126) g(m) := Pn+m(f (n) ⊗ g(m)). By (11), the tensor product (cid:126) is associative. (11) and g(m) ∈ H (cid:126)mC , For a Hilbert space H and a constant c > 0, we denote by Hc the Hilbert space which coincides with H as a set and the tensor product in Hc is equal to the tensor product in H times c. We define a Q-Fock space over H by ∞(cid:77) F Q(H ) := H (cid:126)nC n! . Here H (cid:126)0C denote by F Q := C. The vector Ω := (1, 0, 0, . . . ) ∈ F Q(H ) is called the vacuum. We fin(H ) the subset of F Q(H ) consisting of all finite sequences n=0 F = (f (0), f (1), . . . , f (n), 0, 0, . . . ) in which f (i) ∈ H (cid:126)iC topology of the topological direct sum of the H (cid:126)nC for i = 0, 1, . . . , n, n ∈ N. This space can be endowed with the fin(H) spaces. Thus, convergence in F Q 3 means uniform finiteness of non-zero components and coordinate-wise convergence in H (cid:126)nC . For each h ∈ HC, we define a creation operator a+(h) and an annihilation operator a−(h) as linear operators acting on F Q fin(H ) that satisfy a+(h)f (n) := h (cid:126) f (n), f (n) ∈ H (cid:126)nC , These operators act continuously on F Q H (cid:126)nC , we have: (a−(h)f (n))(x1, . . . , xn−1) = n a−(h) := (a+(h))∗ (cid:22)F Q fin(H ) . (12) fin(H ). Furthermore, for h ∈ HC and f (n) ∈ (cid:90) h(y) f (n)(y, x1, . . . , xn−1) m(dy). (13) x by formulas (4), we get, for f (n) ∈ x and ∂− X Thus, if we introduce informal operators ∂+ H (cid:126)nC , x f (n) = δx (cid:126) f (n), ∂+ x f (n) = nf (n)(x,·). ∂− where δx is the delta function at x. Now, one can easily give a rigorous meaning to the operators on the right hand side of formulas (5) -- (7) and show that the Q-CR hold. We note that, in the obtained representation of the Q-CR, we only used the values of the function Q m⊗2-almost everywhere. Hence, for this representation, we could assume from the very beginning that there exists a set ∆ ∈ B(X 2) which is symmetric (i.e., if (x, y) ∈ ∆, then (y, x) ∈ ∆) and satisfies m⊗2(∆) = 0, and the function Q is also assume that D ⊂ ∆, where D := {(x, x) x ∈ X} is the diagonal in X 2. only defined on the set (cid:101)X 2 := X 2 \ ∆. Since the measure m is non-atomic, we may In physics, intermediate statistics have been discussed since Leinass and Myrheim [32] conjectured their existence in 1977. The first mathematically rigorous prediction of intermediate statistics was done by Goldin, Menikoff and Sharp [20, 21] in 1980, 1981. The name anyon was given to such statistics by Wilczek [50,51]. Anyon statistics were used, in particular, to describe the quantum Hall effect, see e.g. the review paper [46]. Liguori, Mintchev [34, 35] and Goldin, Sharp [22] showed that anyon statistics can be described by the Q-CR in which X = R2, the set ∆ is chosen as ∆ :=(cid:8)(x, y) ∈ X 2 x1 = y1(cid:9) (14) and q, ¯q, if x1 < y1, if x1 > y1. Q(x, y) = (15) Here, q ∈ C with q = 1, and for x ∈ X we denote by xi the ith coordinate of x. With such a choice of the function Q, formulas (5) -- (7) are called the anyon commutation relations (ACR). We note that Goldin, Sharp [22] realized the ACR by using operators acting on the space of functions of finite configurations in X (or, equivalently, in the symmetric Fock space). (cid:40) 4 Goldin and Majid [19] showed that, in the case where q is a kth root of 1 and q (cid:54)= 1, the corresponding statistics satisfies the natural anyonic exclusion principle, which generalizes Pauli's exclusion principle for fermions: a+(f )k = 0 for each f ∈ HC. (16) For further discussions of anyons in mathematical physics literature (including the discrete setting), see e.g. [13, 15, 17, 19, 33, 36 -- 41] and the references therein. We also refer to the paper [8] which deals with a Fock representation of the commutations relations identified by a sequence of self-adjoint operators in a Hilbert space which have norm ≤ 1 and which satisfy the braid relations. 1.2 Gauge-invariant quasi-free states on the CCR and CAR algebras In the theory of the CCR and CAR algebras, quasi-free states, in particular, gauge- invariant quasi-free states, play a fundamental role. We refer the reader to e.g. Sec- tions 5.2.1 -- 5.2.3 and Notes and Remarks to these sections in [11], and [16, Chapter 17], see also the pioneering book [6, Chapter II] and paper [7]. We note that gauge-invariant quasi-free states describe, in particular, the infinite free Bose gas at finite tempera- ture [3] (see also [14] and Section 5.2.5 in [11]) and the infinite free Fermi gas at both finite and zero temperatures [4] (see also [14] and Section 5.2.4 in [11]). Free analogs of quasi-free states have been discussed in [45], see also [24]. Let us recall that the CCR algebra (or the CAR algebra), A, is a complex alge- bra generated by linear operators a+(h), a−(h) (h ∈ HC) satisfying the CCR (the CAR, respectively). Because of the commutation relations, each element of A can be represented as a finite sum of a constant and operators a(cid:93)1(h1)··· a(cid:93)n(hk), h1, . . . , hk ∈ HC, (cid:93)1, . . . , (cid:93)k ∈ {+,−}, which are in the Wick order . The latter means that there is no i ∈ {1, . . . , k − 1} such that (cid:93)i = − and (cid:93)i+1 = +, i.e., there is no creation operator acting before an annihilation operator. Let τ be a state on the algebra A. One defines n-point functions by S(k,n)(gk, . . . , g1, h1, . . . , hn) := τ(cid:0)a+(gk)··· a+(g1)a−(h1)··· a−(hn)(cid:1), (17) where g1, . . . , gk, h1, . . . , hn ∈ HC and k, n ∈ N. One says that the state τ is gauge- invariant if it is invariant under the group of Bogoliubov transformations a+(h) (cid:55)→ a+(eiθh) = eiθa+(h), a−(h) (cid:55)→ a−(eiθh) = e−iθa−(h), θ ∈ [0, 2π). By (17), τ is gauge-invariant if and only if S(k,n) ≡ 0 for k (cid:54)= n. Thus, a gauge-invariant state is completely determined by S(n,n) (n ∈ N). 5 A state τ is called a gauge-invariant quasi-free state if S(k,n) ≡ 0 for k (cid:54)= n and the n-point functions S(n,n) are completely determined by S(1,1). More precisely, in the case of the CCR algebra, we have (cid:88) S(n,n)(gn, . . . , g1, h1, . . . , hn) = per(cid:2)S(1,1)(gi, hj)(cid:3) = (cid:88) S(n,n)(gn, . . . , g1, h1, . . . , hn) = det(cid:2)S(1,1)(gi, hj)(cid:3) = and in the case of the CAR algebra, we have π∈Sn n(cid:89) n(cid:89) i=1 i=1 S(1,1)(gi, hπ(i)), (18) sgn π S(1,1)(gi, hπ(i)). (19) π∈Sn A gauge-invariant quasi-free state on the CCR algebra is completely identified by a bounded linear operator T in HC, with T ≥ 0, which satisfies S(1,1)(g, h) = (T g, h)HC. (20) Respectively, a gauge-invariant quasi-free state on the CAR algebra is completely iden- tified by a bounded linear operator T in HC, with 0 ≤ T ≤ 1, which satisfies (20). The corresponding representation of the CCR/CAR algebra can be given on the symmetric/antisymetric Fock space over H ⊕H by using the bounded linear operators √ 1 − T in the CAR case, see e.g. T and √ 1 + T in the CCR case and √ T and √ Examples 5.2.18 and 5.2.20 in [11]. 1.3 A brief description of the results While our main interest in this paper will be the ACR, we will actually deal with a slightly more general form of the Q-CR: we will assume that X = Rd with d ≥ 2 and the function Q : (cid:101)X 2 → C satisfies Q(x, y) = Q(x1, y1) (with an obvious abuse of notation). Here (cid:101)X 2 := X 2 \ ∆ with ∆ being given by (14). We saw in the Fock space representation that defining a function Q on (cid:101)X 2 was enough. However, we will see below that, in the general case, this is not enough for relation (3) and we need to specify the values of Q(x, x) for x ∈ X. In the case of the bose and fermi statistics, we take, of course, Q(x, x) ≡ 1 and Q(x, x) ≡ −1, respectively. So from now on we will assume that, for some constant η ∈ R, we have Q(x, y) = η for all (x, y) ∈ ∆, in particular, Q(x, x) = η for all x ∈ X. We will define a Q-CR algebra so that the value η will matter for relation (3), but η will be of no importance to relations (1), (2) as they will still depend on the values of the function Q m⊗2-almost everywhere. We see that, in the anyon case (with q (cid:54)= ±1), the function Q cannot be extended to a continuous function on X 2, so there is a freedom in choosing the value of η. A natural choice for η seems to be η = (cid:60)(q) = (q + ¯q)/2. 6 (cid:90) The form of the Q-CR means that it is not enough to consider a complex algebra generated by the operators (4). Instead, in Section 2, we consider a complex algebra A generated by operator-valued integrals m⊗k(dx1 ··· dxk) ϕ(k)(x1, . . . , xk) ∂(cid:93)1 (21) where the class of functions ϕ(k) : X k → C appearing in the integral (21) will be specified. We show that the anyon exclusion principle (see (16)) holds in the general ACR algebra for q being a root of 1, q (cid:54)= 1. X k xk x1 ··· ∂(cid:93)k , If τ is a state on A, then due to (3), τ is completely determined by the n-point functions S(k,n)(ϕ(k+n)) := τ (cid:18)(cid:90) × ∂+ x1 ··· ∂+ xk X k+n ∂− xk+1 ··· ∂− xk+n (cid:19) m⊗(k+n)(dx1 ··· dxk+n) ϕ(k+n)(x1, . . . , xk+n) . (22) x (cid:55)→ eiθ∂+ x , ∂− x (cid:55)→ e−iθ∂− We say that the state τ is gauge-invariant if it is invariant under the group of x , with θ ∈ [0, 2π), or equivalently Bogoliubov transformations ∂+ if S(k,n) ≡ 0 for k (cid:54)= n. So it is intuitively clear what it should mean that τ is a gauge-invariant quasi-free state: we should have S(k,n) ≡ 0 if k (cid:54)= n and the n-point functions S(n,n) should be completely determined by S(1,1). However, to write down a proper generalization of formulas (18), (19) is not straightforward: instead of the sign of a permutation π we should use the function Qπ (see (9)), and the functions ϕ(k+n) appearing in (22) do not necessarily factorize to separate their variables. We solve this problem in Section 2 by properly introducing certain measures on R2n corresponding to the n-point functions. As a result, the definition of a gauge-invariant quasi-free state for the Q-CR generalizes the available definitions in the CCR and CAR cases. In Section 3 we construct operator-valued integrals (21) in the Q-symmetric Fock space. The presentation in this section is given at a rather general level. In particular, in this section we assume that X is a locally compact Polish space, while m is a non-atomic Radon measure on X. In Section 4, we construct a representation of the Q-CR algebra A and a class of gauge-invariant quasi-free states τ on it. This construction is done in a JQ-symmetric Fock space over H ⊕ H . Here JQ : Z 2 → C is a function on the space Z := X1 (cid:116) X2, the disjoint union of two copies of X. Explicitly, the function JQ is defined through the function Q by formula (54) below. The operator T , being defined analogously to formula (20), satisfies in our setting the following assumptions: • T is a self-adjoint bounded linear operator in the real space H and is extended by linearity to HC; 7 • T commutes with any operator of multiplication by a bounded function ψ(x1); • in the case η ≥ 0, we have T ≥ 0, and in the case η < 0, we have 0 ≤ T ≤ −1/η. For ϕ(2n)(x1, . . . , x2n) = gn(x1)··· g1(xn)h1(xn+1)··· hn(x2n) with g1, . . . , gn, h1, . . . , hn ∈ HC, we obtain S(n,n)(ϕ(2n)) = gi(xi)(T hπ(i))(xi) Qπ(x1, . . . , xn) m⊗n(dx1 ··· dxn), (cid:90) (cid:88) (cid:32) n(cid:89) π∈Sn X n i=1 (cid:33) Finally, compare with (18) -- (20). in Section 5, we discuss the particle density associated with a gauge- invariant quasi-free state on the ACR algebra with η = (cid:60)q. The particle density x for x ∈ X. It follows from the ACR that these is informally defined by ρ(x) := ∂+ operators commute, cf. [19, 22]. Hence, the state on the algebra generated by the X m(dx) f (x)ρ(x) (f running through a space of test commutative operators ρ(f ) :=(cid:82) x ∂− functions on X) should be given by a probability measure µ. In the case of the CCR and CAR algebras, it was shown in [42,44] that, for T being a locally trace class operator, µ is a permanental (determinantal, respectively) point process on X. Note, however, that our assumptions on the operator T exclude locally trace class operators. In this paper, we treat the case where T is a constant operator, T = κ21 with κ > 0. Under this assumption, it is not possible to give a rigorous meaning to ρ(f ) as a self-adjoint linear operator in the JQ-symmetric Fock space over H ⊕ H . So a renormalization is needed. Similarly to the construction of the renormalized square of white noise algebra [1, 2], as renormalization we will use Ivanov's formula [25] which suggests that the square of the delta function, δ2, can be interpreted as cδ, where c is any positive constant. For our calculations, we choose c = 1, so that Ivanov's formula becomes δ2 = δ. We note that a different choice of the constant c would lead to similar results in which the measure m is replaced by cm. So, using this renormalization and the ACR, we rigorously define a functional τ on the algebra generated by commutative operators ρ(f ). However, due to renormaliza- tion, it is not a priori clear whether τ is a state, i.e., whether it is positive definite. We prove that τ is indeed a state if and only if η ∈ [0, 1]. Furthermore, for η = 0, the state τ is given by the Poisson point process on X with intensity measure κ2m, while for η > 0, τ is given by a negative binomial point process on X, which depends on two parameters, η and κ. The latter process takes values in the space Γ(X) of multiple configurations in X, i.e., Radon measures on X which take values in {0, 1, 2, . . . ,∞}. Note also that the negative binomial point process is a measure-valued L´evy process on X whose L´evy measure is finite. Finally, 8 we prove that, for a fixed η > 0, a (scaling) limit of the states depending on κ exists as κ → ∞, and the limiting state is given by the gamma random measure, depending on parameter η. This random measure is known to have many distinguished properties, see e.g. [18, 29 -- 31, 47 -- 49]. We stress that the results of Section 5 are new even in the CCR case. 2 The Q-CR algebra and gauge-invariant quasi-free states on it 2.1 Preliminary definitions In this section, we assume that X = Rd with d ≥ 2, m is the Lebesgue measure on it, Q(x, y) = Q(x1, y1) for (x, y) ∈ (cid:101)X 2, and Q(x, y) = η for (x, y) ∈ ∆. To define the Q-CR algebra, we need an appropriate class of functions ϕ(k) appearing in (21). Let k ∈ N. We denote by Π(k) the set of all partitions of the set {1, . . . , k}, i.e., all collections of mutually disjoint sets whose union is {1, . . . , k}. For each partition the subset of X k which consists of all (x1, . . . , xk) ∈ X k θ ∈ Π(k), we denote by X (k) such that, for all 1 ≤ i < j ≤ k, xi = xj if and only if i and j belong to the same θ with θ ∈ Π(k) form a partition of element of the partition θ. Note that the sets X (k) X k. We denote X (k) := X (k) for the minimal partition θ =(cid:8){1}, {2}, . . . , {k}(cid:9). θ Let θ = {θ1, . . . , θl} ∈ Π(k) and assume that θ min θ1 < min θ2 < ··· < min θl. We have m⊗l(X l\X (l)) = 0, so we can consider m⊗l as a measure on X (l). Consider the mapping Iθ : X (l) → X (k) θ given by Iθ(x1, . . . , xl) = (y1, . . . , yk), where for each i ∈ θ1 we have yi = x1, for each i ∈ θ2 we have yi = x2, etc. We denote by m(k) the pushforward of the measure m⊗l under Iθ. We extend the measure m(k) θ by zero to the whole space X k. Note that m(k) = m⊗k Let us fix any (cid:93)1, . . . , (cid:93)k ∈ {+,−}. We denote by Π(k, (cid:93)1, . . . , (cid:93)k) the subset of Π(k) which consists of all partitions θ = {θ1, . . . , θl} such that each set θi has at most two elements, and if θi = {a, b} has two elements then (cid:93)a (cid:54)= (cid:93)b. We define a measure on X k by for the minimal partition θ =(cid:8){1}, {2}, . . . , {k}(cid:9). θ θ For example, for a measurable function f : X 3 → [0,∞), we have m(k) (cid:93)1,...,(cid:93)k := m(k) θ . θ∈Π(k,(cid:93)1,...,(cid:93)k) (cid:90) X 3 f (x1, x2, x3) m(3)−,−,+(dx1 dx2 dx3) = f (x1, x2, x3) m⊗3(dx1 dx2 dx3) (cid:88) (cid:90) X 3 9 (cid:90) + X 2 f (x1, x2, x1) m⊗2(dx1 dx2) + (cid:90) X 2 f (x1, x2, x2) m⊗2(dx1 dx2). (23) Completely analogously, starting with the Lebesgue measure on R rather than X, on Rk. For example, similarly to (23), for a measurable we define a measure ν(k) function f : R3 → [0,∞), we have (cid:93)1,...,(cid:93)k (cid:90) (cid:90) R3 + R2 f (s1, s2, s3) ν(3)−,−,+(ds1 ds2 ds3) = f (s1, s2, s3) ds1 ds2 ds3 f (s1, s2, s1) ds1 ds2 + f (s1, s2, s2) ds1 ds2. (cid:90) R3 (cid:90) R2 We denote by L0(X k → C, m(k) ) the linear space of classes of complex-valued measurable functions on X k, with any two measurable functions f, g : X k → C being identified if f = g m(k) -a.e. We define a linear mapping (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k Φ(k) (cid:93)1,...,(cid:93)k : H kC × L∞(Rk → C, ν(k) ) → L0(X k → C, m(k) (cid:2)h1, . . . , hk, v(k)(cid:3)(x1, . . . , xk) := h1(x1)··· hk(xk)v(k)(x1 (cid:93)1,...,(cid:93)k 1, . . . , x1 k), ) (24) (cid:93)1,...,(cid:93)k by Φ(k) (cid:93)1,...,(cid:93)k where h1, . . . , hk ∈ HC and v(k) ∈ L∞(Rk → C, ν(k) range of Φ(k) Remark 1. It should be noted that the mapping Φ(k) if h1, . . . , hk ∈ HC, v(k) ∈ L∞(Rk → C, ν(k) , which is a subspace of L0(X k → C, m(k) (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k ). (cid:93)1,...,(cid:93)k (cid:2)g, h2 . . . , hk, v(k)(cid:3) = Φ(k) (cid:93)1,...,(cid:93)k ), and α ∈ L∞(R, ds), we have (cid:2)h1, . . . , hk, w(k)(cid:3), (cid:93)1,...,(cid:93)k Φ(k) (cid:93)1,...,(cid:93)k ). We denote by F(k) (cid:93)1,...,(cid:93)k is not injective. For example, where g(x) := h1(x)α(x1) ∈ HC and w(k)(s1, . . . , sk) := v(k)(s1, . . . , sk)α(s1) ∈ L∞(Rk → C, ν(k) ). (cid:93)1,...,(cid:93)k Below we will deal with linear operators in a complex Hilbert space which will be denoted by operator-valued integrals of the form (21) with (25) the (cid:2)h1, . . . , hk, v(k)(cid:3) ∈ F(k) ϕ(k) = Φ(k) (cid:93)1,...,(cid:93)k (cid:2)h1, . . . , hk, v(k)(cid:3). Our next aim is to give (26) (cid:93)1,...,(cid:93)k . We will also denote these operators by I (k) a rigorous definition of the commutation relations (1) -- (3) satisfied by these operators. Let k ≥ 2, (cid:93)1, . . . , (cid:93)k ∈ {+,−}, and let i ∈ {1, . . . , k − 1}. Let us consider the operator-valued integral (21) with ϕ(k) given by (26). Assume that (cid:93)i = (cid:93)i+1. Then, at least informally, we calculate using either relation (1) or relation (2): (cid:93)1,...,(cid:93)k (cid:90) X k m⊗k(dx1 ··· dxk) ϕ(k)(x1, . . . , xk)∂(cid:93)1 x1 ··· ∂(cid:93)k xk 10 (cid:90) (cid:90) X k × ∂(cid:93)1 m⊗k(dx1 ··· dxk) h1(x1)··· hk(xk)v(k)(x1 x1 ··· ∂(cid:93)i−1 m⊗k(dx1 ··· dxk) h1(x1)··· hi−1(xi−1)hi+1(xi)hi(xi+1)hi+2(xi+2)··· hk(xk) 1, . . . , x1 ··· ∂(cid:93)k i+1, x1 i ) k)Q(x1 ∂(cid:93)i+2 xi+2 ∂(cid:93)i+1 xi+1 ∂(cid:93)i xi xi−1 xk X k × (Ψiv(k))(x1 1, . . . , x1 k)∂(cid:93)1 x1 ··· ∂(cid:93)k xk . (27) = = Here (Ψiv(k))(s1, . . . , sk) := v(k)(s1, . . . , si−1, si+1, si, si+2, . . . , sk)Q(si, si+1) ∈ L∞(Rk → C, ν(k) ). (28) (cid:93)1,...,(cid:93)k Thus, inspired by (27) and (28), we give the following definition: Relation (1) (or relation (2), respectively) means that, for any k ≥ 2, i ∈ {1, . . . , k − 1} and (cid:93)1, . . . , (cid:93)k ∈ {+,−} such that (cid:93)i = (cid:93)i+1 = + (or (cid:93)i = (cid:93)i+1 = −, respectively), we have (cid:2)h1, . . . , hi−1, hi+1, hi, hi+2, . . . , hk, Ψiv(k)(cid:3). I (k) (cid:93)1,...,(cid:93)k (29) Analogously, relation (3) means that for any k ≥ 2, i ∈ {1, . . . , k − 1} and (cid:93)1, . . . , (cid:93)k ∈ {+,−} such that (cid:93)i = − and (cid:93)i+1 = +, we have (cid:93)1,...,(cid:93)k (cid:2)h1, . . . , hk, v(k)(cid:3) = I (k) (cid:2)h1, . . . , hk, v(k)(cid:3) I (k) (cid:93)1,...,(cid:93)k = I (k−2) (cid:93)1,...,(cid:93)i−1,(cid:93)i+2,...,(cid:93)k + I (k) (cid:93)1,...,(cid:93)i−1,(cid:93)i+1,(cid:93)i,(cid:93)i+2,...,(cid:93)k (cid:2)h1, . . . , hi−1, hi+2, . . . , hk, u(k−2)(cid:3) (cid:2)h1, . . . , hi−1, hi+1, hi, hi+2, . . . , hk, Ψ(cid:48) iv(k)(cid:3), (30) where u(k−2)(s1, . . . , sk−2) := hi(x)hi+1(x) (cid:90) X × v(k)(s1, . . . , si−1, x1, x1, si, . . . , sk−2) m(dx) ∈ L∞(Rk−2 → C, ν(k−2) (cid:93)1,...,(cid:93)i−1,(cid:93)i+2,...,(cid:93)k ) (31) and (Ψ(cid:48) iv(k))(s1, . . . , sk) := v(k)(s1, . . . , si−1, si+1, si, si+2, . . . , sk) × Q(si+1, si) ∈ L∞(Rk → C, ν(k) (cid:93)1,...,(cid:93)i−1,(cid:93)i+1,(cid:93)i,(cid:93)i+2,...,(cid:93)k ). (32) Remark 2. In the case k = 2, the second addend on the right hand side of equality (30) is understood as the constant operator u(0), where u(0) := h1(x)h2(x)v(2)(x1, x1) m(dx). (cid:90) X 11 Remark 3. Note that the commutation relations (29), (30) do not depend on the rep- resentation of ϕ(k) ∈ F(k) Remark 4. Note that the commutation relations (29) do not depend on η. Indeed, for (cid:93)i = (cid:93)i+1, we have (cid:0){(s1, . . . , sk) si = si+1}(cid:1) = 0. in the form (26). (33) (cid:93)1,...,(cid:93)k ν(k) (cid:93)1,...,(cid:93)k Hence, in (28), for si = si+1 the value Q(si, si+1) = η plays no role. On the other hand, formula (33) is not true when (cid:93)i = − and (cid:93)i+1 = +. Therefore, for si = si+1 the value Q(si+1, si) = η does matter for (32), hence also for the commutation relation (30). 2.2 Definition of the Q-CR algebra and the anyon exclusion principle We are now in position to define the Q-CR algebra. Let G be a separable, complex Hilbert space. Let Θ be a dense linear subspace of G . We assume that, for any (cid:93)1, . . . , (cid:93)k ∈ {+,−} and any ϕ(k) ∈ F(k) we have a linear operator mapping Θ into Θ. This operator is denoted either as in (21) or by I (k) ϕ(k) is as in (26). These operators will be called operator-valued integrals. (cid:0)h1, . . . , hk, v(k)(cid:1), given that (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k We will assume that the operator-valued integrals satisfy the following axioms. (A1) Consistency condition: For any g1, . . . , gk, h1, . . . , hk ∈ HC, and v(k), w(k) ∈ L∞(Rk → C, ν(k) ), if (cid:93)1,...,(cid:93)k (cid:2)g1, . . . , gk, w(k)(cid:3) = Φ(k) (cid:2)h1, . . . , hk, v(k)(cid:3), (cid:93)1,...,(cid:93)k Φ(k) (cid:93)1,...,(cid:93)k then (g1, . . . gk, w(k)) = I (k) (A2) Linearity: For any (cid:93)1, . . . , (cid:93)k ∈ {+,−}, I (k) I (k) (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k on hi ∈ HC (i = 1, . . . , k) and on v(k) ∈ L∞(Rk → C, ν(k) ). (cid:93)1,...,(cid:93)k (h1, . . . hk, v(k)). (h1, . . . , hk, v(k)) linearly depends (A3) The adjoint operator: The adjoint of any operator I (k) (h1, . . . , hk, v(k)) in the Hilbert space G contains Θ in its domain, and the restriction of this adjoint operator to Θ is equal to the operator I (k)(cid:52)k,...,(cid:52)1 (hk, . . . , h1, v(k)∗), where (cid:93)1,...,(cid:93)k (cid:40) + if (cid:93)i = −, − if (cid:93)i = +, (cid:52)i := v(k)∗(s1, . . . , sk) = v(k)(sk, . . . , s1) ∈ L∞(Rk → C, ν(k)(cid:52)k,...,(cid:52)1 i = 1, . . . , k, ). (A4) Q-commutation relations: The operators ∂+ x satisfy the Q-CR (1) -- (3). A x , ∂− rigorous meaning of these relations is given by formulas (28) -- (32). 12 (A5) Multiplication of operator-valued integrals: For any h1, . . . , hk+n ∈ HC, v(k) ∈ L∞(Rm, ν(k) (cid:93)1,...,(cid:93)k ), w(n) ∈ L∞(Rn, ν(n) (cid:93)k+1,...,(cid:93)k+n ), we have Here I (k) (cid:93)1,...,(cid:93)k (h1, . . . , hk, v(k))I (n) (cid:93)k+1,...,(cid:93)k+n (hk+1, . . . , hk+n, w(n)) = I (k+n) (cid:93)1,...,(cid:93)k+n (h1, . . . , hk+n, v(k) ⊗ w(n)). (v(k) ⊗ w(n))(s1, . . . , sk+n) := v(k)(s1, . . . , sk)w(n)(sk+1, . . . , sk+n) ∈ L∞(Rk+n, ν(k+n) (cid:93)1,...,(cid:93)k+n ). Remark 5. We stress that the value η of the function Q on the diagonal does not matter for the relations (1), (2). Let A denote the complex algebra generated by the operator-valued integrals sat- isfying axioms (A1) -- (A5), with the usual multiplication of operators acting on Θ. We will call A the algebra of Q-commutation relations, or the Q-CR algebra for short. In the case where the function Q is given by (15), we will call A the algebra of anyon commutation relations, or the ACR algebra for short. The following theorem shows that the anyon exclusion principle [19] (see also [10, Proposition 2.9]) holds in the ACR algebra with q being a root of 1. Theorem 6. Let k ∈ N, k ≥ 2. Let q ∈ C be such that q (cid:54)= 1 and qk = 1. Then, in the ACR algebra, we have, for each h ∈ HC, (cid:18)(cid:90) X (cid:19)k m(dx) h(x)∂+ x = 0. +,...,+ = ν(k), the Lebesgue measure on R(k) (R(k) consisting of all Proof. Note that ν(k) (s1, . . . , sk) ∈ Rk with si (cid:54)= sj if i (cid:54)= j). By (29), we get, for any i ∈ {1, . . . , k − 1} and v(k) ∈ L∞(Rk, ν(k)): I (k) +,...,+(h, . . . , h, v(k)) = I (k) +,...,+(h, . . . , h, Ψiv(k)). Here Ψiv(k) is given by formula (28) for (s1, . . . , sk) ∈ R(k). By the proof of Proposi- tion 2.8 in [10], it follows from here that, for each permutation π ∈ Sk, we get I (k) +,...,+(h, . . . , h, v(k)) = I (k) +,...,+(h, . . . , h, Ψπv(k)), (34) where (Ψπv(k))(s1, . . . , sk) = Qπ−1(s1, . . . , sk)v(k)(sπ(1), . . . , sπ(k)). 13 Recall that the function Qπ was defined by (9), and we again used the obvious abuse k) for (x1, . . . , xk) ∈ X k. By (10) and (34), of notation Qπ(x1, . . . , xk) = Qπ(x1 we get 1, . . . , x1 I (k) +,...,+(h, . . . , h, v(k)) = I (k) (35) and the function Pkv(k) is Q-symmetric on R(k). It follows from the proof of Proposi- tion 2.9 in [10] that if we choose v(k) ≡ 1, we get +,...,+(h, . . . , h, Pkv(k)), (Pk1)(s1, . . . , sk) = 1 − qk (1 − q)k! = 0 for all s1 < s2 < ··· < sk. Hence, Pk1 = 0 ν(k)-a.e. Now the statement of the theorem follows by the axioms (A5) and (A2). 2.3 Definition of a gauge-invariant quasi-free state Let τ be a state on the Q-CR algebra A. Because of the Q-CR, τ is completely determined by its n-point functions, which are defined by formula (22) with ϕ(k+n) ∈ , where (cid:93)1 = ··· = (cid:93)k = + and (cid:93)k+1 = ··· = (cid:93)k+n = −. We already discussed F(k+n) (cid:93)1,...,(cid:93)k+n in subsection 1.3 that gauge invariance of τ means that S(k,n) ≡ 0 if k (cid:54)= n. So our aim now is to introduce a proper generalization of formulas (18), (19). We denote the n-point functions by S(n,n)(h1, . . . , h2n, v(2n)) := τ(cid:0)I (2n) (h1, . . . , h2n, v(2n))(cid:1), (cid:93)1,...,(cid:93)2n (cid:93)1,...,(cid:93)2n (36) where (cid:93)1 = ··· = (cid:93)n = + and (cid:93)n+1 = ··· = (cid:93)2n = −. By (A2), the right hand side of (36) identifies a linear functional of v(2n) ∈ L∞(R2n, ν(2n) ). If we assume that this functional continuously depends on v(2n), then, according to the general theory of linear continuous functionals on L∞ spaces, this functional can be identified with a complex- valued, finite-additive measure on R2n that is absolutely continuous with respect to the measure ν(2n) . We will actually assume the following stronger condition to be satisfied. (M) For any h1, . . . , h2n ∈ HC, there exists a (unique) complex-valued measure S(n,n)[h1, . . . , h2n] on R2n which is absolutely continuous with respect to ν(2n) , (cid:93)1 = ··· = (cid:93)n = + and (cid:93)n+1 = ··· = (cid:93)2n = −, and satisfies, for all v(2n) ∈ L∞(R2n, ν(2n) (cid:93)1,...,(cid:93)2n (cid:93)1,...,(cid:93)2n (cid:93)1,...,(cid:93)2n ), S(n,n)(h1, . . . , h2n, v(2n)) = (cid:90) R2n v(2n)(s1, . . . , s2n) S(n,n)[h1, . . . , h2n](ds1 ··· ds2n). (37) 14 We will denote by S(n,n)[h1, . . . , h2n](s1, . . . , s2n) the density of the measure with (cid:93)1 = ··· = We say that τ is a gauge-invariant quasi-free state on the Q-CR algebra A if S(k,n) ≡ S(n,n)[h1, . . . , h2n](ds1 ··· ds2n) with respect to the measure ν(2n) (cid:93)n = + and (cid:93)n+1 = ··· = (cid:93)2n = −. 0 if k (cid:54)= n, and for each n ∈ N and any g1, . . . , gn, h1, . . . , hn ∈ HC, we have (cid:93)1,...,(cid:93)2n S(n,n)[gn, . . . , g1, h1, . . . , hn](sn, . . . , s1, sn+1, . . . , s2n) S(1,1)[gi, hπ(i)](si, sn+π(i)) Qπ(s1, . . . , sn). (38) (cid:19) (cid:88) (cid:18) n(cid:89) π∈Sn i=1 = Remark 7. Note the following slight difference in notations: in formulas (18), (19), the n-point function S(n,n)(gn, . . . , g1, h1, . . . , hn) is linear in each gi and antilinear in each hi, while in our setting the n-point function in (38) depends linearly on both gi and hi. with (cid:93)1 = ··· = (cid:93)n = + and (cid:93)n+1 = ··· = Remark 8. Note that the measure ν(2n) (cid:93)2n = − remains invariant under the transformation (cid:93)1,...,(cid:93)2n R2n (cid:51) (s1, . . . , s2n) (cid:55)→ (sn, . . . , s1, sn+1, . . . , s2n) ∈ R2n. Hence, formulas (37), (38) mean that, for any g1, . . . , gn, h1, . . . , hn ∈ HC and v(2n) ∈ L∞(R2n, ν(2n) ), we have S(n,n)(gn, . . . , g1, h1, . . . , hn, v(2n)) = v(2n)(sn, . . . , s1, sn+1, . . . , s2n) S(1,1)[gi, hπ(i)](si, sn+π(i)) Qπ(s1, . . . , sn) ν(2n) (cid:93)1,...,(cid:93)2n (ds1 ··· ds2n). (39) (cid:90) (cid:19) R2n (cid:93)1,...,(cid:93)2n × (cid:88) π∈Sn (cid:18) n(cid:89) i=1 Below, in Section 4, we will explicitly construct a class of gauge-invariant quasi- free states, but before doing this we wll now construct operator-valued integrals in the Q-Fock space. 3 Operator-valued integrals in the Q-symmetric Fock space In this section, we will assume that X is a locally compact Polish space, B(X) is the Borel σ-algebra on X, and m is a reference measure on (X, B(X)). We assume m to be a Radon measure (i.e., finite on any compact set in X) and non-atomic. Analogously to subsection 1.1, we assume that ∆ is a measurable, symmetric subset of X 2 and satisfies m⊗2(∆) = 0. We also assume that D ⊂ ∆, where D := {(x, x) x ∈ X} is the diagonal in X 2. We denote (cid:101)X 2 := X 2 \ ∆. We fix η ∈ R and consider a function 15 Q(x, y) = η for all (x, y) ∈ ∆. (cid:101)X n :=(cid:8)(x1, . . . , xn) ∈ X n (xi, xj) (cid:54)∈ ∆ for all 1 ≤ i < j ≤ n(cid:9). Q : X 2 → C such that Q(x, y) = 1 and Q(x, y) = Q(y, x) for all (x, y) ∈ (cid:101)X 2, and Analogously to (cid:101)X 2, we define, for each n ≥ 3 Let n ≥ 2. A function f (n) : (cid:101)X n → C is called Q-symmetric if for any i ∈ {1, . . . , n−1} and (x1, . . . , xn) ∈ (cid:101)X n, formula (8) holds. Since m⊗n(X n \ (cid:101)X n) = 0, the function f (n) is defined m⊗n-a.e. on X n. We define the function Qπ : (cid:101)X n → C by formula (9), and the Q-symmetrization of a function f (n) : (cid:101)X n → C by (10). The definitions of H , HC, H (cid:126)nC , Pn, F Q(H ), F Q fin(H ), a+(h), and a−(h) are now similar to subsection 1.1. Let (cid:93)1, . . . , (cid:93)k ∈ {+,−}. Analogously to (24), (25), we define a linear mapping Ξ(k) (cid:93)1,...,(cid:93)k : H kC × L∞(X k → C, m(k) (cid:93)1,...,(cid:93)k ) → L0(X k → C, m(k) ) (cid:93)1,...,(cid:93)k (cid:2)h1, . . . , hk, κ(k)(cid:3)(x1, . . . , xk) := h1(x1)··· hk(xk)κ(k)(x1, . . . , xk), by Ξ(k) (cid:93)1,...,(cid:93)k (40) where h1, . . . , hk ∈ HC and κ(k) ∈ L∞(X k → C, m(k) the range of this mapping. Our aim now is to construct operator-valued integrals of the form (21) with ϕ(k) = Ξ(k) (cid:2)h1, . . . , hk, κ(k)(cid:3). ). We denote by G(k) (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k The following proposition follows immediately from the definition of the creation (cid:93)1,...,(cid:93)k h(y)Q(y, x1)··· Q(y, xi−1) × f (n)(x1, . . . , xi−1, y, xi, . . . , xn−1) m(dy), i=1 X f (n) ∈ H ⊗nC Then, on Ffin(H ), we have a+(h)P = P b+(h), a−(h)P = P b−(cid:0)¯h(cid:1). Here, for f (n) ∈ H ⊗nC , we define P f (n) := Pnf (n). Using the notation (4), we conclude from here the following corollary. 16 (41) (42) Proposition 9. Let F (H ) :=(cid:76)∞ operator, a+(h), and [10, Proposition 3.2]. H ⊗nC n=0 the space Ffin(H ) be defined analogously to F Q continuous operator b+(h)and b−(h) on Ffin(H ) by b+(h)f (n) := h ⊗ f (n), (b−(h)f (n))(x1, . . . , xn−1) := n(cid:88) (cid:90) denote the full Fock space over H , and let fin(H ). For h ∈ HC, we define linear Corollary 10. For any (cid:93)1, . . . , (cid:93)k ∈ {+,−} and any h1, . . . , hk ∈ HC, we have on F Q m⊗k(dx1 ··· dxk) h1(x1)··· hk(xk)∂(cid:93)1 x1 ··· ∂(cid:93)k xk = P b(cid:93)1(h1)··· b(cid:93)k(hk). (43) fin(H ):(cid:90) Here we denoted(cid:90) X k (cid:18)(cid:90) X k := m⊗k(dx1 ··· dxk)h1(x1)··· hk(xk)∂(cid:93)1 x1 . . . , ∂(cid:93)k xk (cid:18)(cid:90) (cid:19) ··· (cid:19) . m(dx1)h(x1)∂(cid:93)1 x1 X m(dxk)h(xk)∂(cid:93)k xk X Let f (n) ∈ H (cid:126)nC . Using Corollary 10, we can write down the action of the operator in formula (43) on f (n) through the function ϕ(k)(x1, . . . , xk) := h1(x1)··· hk(xk). (44) (Note that this function is defined m(k) -a.e.) For example, (cid:93)1,...,(cid:93)k (cid:19) = Pn−1 X 3 i=1 X 2 y1∂− y2∂+ y3f (n) (cid:90) (x1, . . . , xn−1) m⊗2(dy1 dy2) ϕ(3)(y1, y2, y2) m⊗3(dy1 dy2 dy3)ϕ(3)(y1, y2, y3)∂− (cid:26) n(cid:88) ×(cid:2)Q(y1, x1)··· Q(y1, xi−1)f (n)(x1, . . . , xi−1, y1, xi, . . . , xn−1) + Q(y2, y1)Q(y2, x1)··· Q(y2, xi−1)f (n)(x1, . . . , xi−1, y2, xi, . . . , xn−1)(cid:3) (cid:90) n+1(cid:88) n(cid:88) m⊗2(dy1 dy2) ϕ(3)(y1, y2, x1) × Q(y2, x1)··· Q(y2, xi−2)f (n)(x2, . . . , xj−1, y1, xj, . . . , xi−2, y2, xi−1, . . . , xn−1) Q(y2, y1)Q(y1, x1)··· Q(y1, xj−1) Q(y1, x1)··· Q(y1, xj−1)Q(y2, x1)··· Q(y2, xi−1) (cid:20) i−1(cid:88) i=2 X 2 j=2 (cid:18)(cid:90) + + j=i (cid:21)(cid:27) . × f (n)(x2, . . . , xi−1, y2, xi, . . . , xj−1, y1, xj, . . . , xn−1) (cid:2)h1, . . . , hk, κ(k)(cid:3) ∈ G(k) As easily seen, we can replace in the obtained formulas the function ϕ(k) of the form (44) with a function ϕ(k) being given by the right hand side of formula (40). As a result, for each ϕ(k) = Ξ(k) , we have constructed a continuous (cid:93)1,...,(cid:93)k (h1, . . . , hk, κ(k)). linear operator on F Q Note that this operator indeed depends on the values of the function ϕ(k)(x1, . . . , xk) = h1(x1)··· hk(xk)κ(k)(x1, . . . , xk) m(k) fin(H ) which is denoted as in (21) or by I (k) -a.e. (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k 17 The following proposition follows from the construction of the operator-valued inte- grals (compare with [10, Proposition 3.8] regarding the corresponding statement about the Q-CR). (h1, . . . , hk, κ(k)) Proposition 11. The above constructed operator-valued integrals I (k) satisfy the axioms (A1)(cid:48) -- (A5)(cid:48) that are obtained from the axioms (A1) -- (A5) by replac- ing the space L∞(Rk → C, ν(k) ) by L∞(X k → C, m(k) (cid:93)1,...,(cid:93)k ). (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k Let us note that we initially had operator-valued integrals I (k) (h1, . . . , hk, κ(k)) for κ(k) ≡ 1 and then we extended them to an arbitrary function κ(k) ∈ L∞(X k → C, m(k) ). The following lemma shows that, under natural assumptions on the operator-valued integrals, this extension is, in fact, unique. (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k Lemma 12. Assume that the operator-valued integrals I (k) on F Q h1, . . . , hk ∈ HC and any F, G ∈ F Q fin(H ) satisfy the axioms (A1)(cid:48), (A2)(cid:48) and the following assumption: fin(H ) the linear functional (h1, . . . , hk, κ(k)) acting for any (cid:93)1,...,(cid:93)k ) (cid:51) κ(k) →(cid:0) I (k) (cid:93)1,...,(cid:93)k (h1, . . . , hk, κ(k))F, G(cid:1) L∞(X k → C, m(k) (cid:93)1,...,(cid:93)k F Q(H ) ∈ C is continuous and is given by a complex-valued measure on X k. Then the equality I (k) (cid:93)1,...,(cid:93)k (h1, . . . , hk, 1) = I (k) (cid:93)1,...,(cid:93)k (h1, . . . , hk, 1) (45) implies the equality I (k) (cid:93)1,...,(cid:93)k (h1, . . . , hk, κ(k)) = I (k) (cid:93)1,...,(cid:93)k (h1, . . . , hk, κ(k)) for all κ(k) ∈ L∞(X k → C, m(k) (cid:93)1,...,(cid:93)k ). Proof. Denote by m(k) satisfies, for all κ(k) ∈ L∞(X k → C, m(k) (cid:93)1,...,(cid:93)k ), (cid:93)1,...,(cid:93)k [F ; G; h1, . . . , hk] the complex-valued measure on X k that (cid:90) =(cid:0) I (k) X k κ(k)(x1, . . . , xk) m(k) (h1, . . . , hk, κ(k))F, G(cid:1) (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k [F ; G; h1, . . . , hk](dx1 ··· dxk) F Q(H ). (46) As easily seen from the construction of the operator-valued integrals, there exists a complex-valued measure m(k) [F ; G; h1, . . . , hk] on X k that satisfies equality (46) in and I (k) which m(k) , respectively. Hence, it suffices to prove that (cid:93)1,...,(cid:93)k are replaced with m(k) and I (k) (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k m(k) (cid:93)1,...,(cid:93)k [F ; G; h1, . . . , hk] = m(k) (cid:93)1,...,(cid:93)k [F ; G; h1, . . . , hk]. (47) 18 For i ∈ {1, . . . , k}, let Ai ∈ B(X) and denote by χAi the indicator function of the set Ai. We have, by axiom (A1)(cid:48) and (45), [F ; G; h1, . . . , hk](A1 × ··· × Ak) =(cid:0) I (k) F Q(H ) =(cid:0)I (k) (h1χA1, . . . , hkχAk, 1)F, G(cid:1) (cid:93)1,...,(cid:93)k (h1, . . . , hk, χA1 ··· χAk)F, G(cid:1) (h1χA1, . . . , hkχAk, 1)F, G(cid:1) m(k) =(cid:0) I (k) (cid:93)1,...,(cid:93)k F Q(H ) (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k F Q(H ) = m(k) (cid:93)1,...,(cid:93)k [F ; G; h1, . . . , hk](A1 × ··· × Ak), which implies (47). We note that, in this section, we have not yet used the value of the function Q on the diagonal. Proposition 13. The following relation between operators ∂+ obtained representation of the Q-CR algebra in the Q-Fock space F Q(H ): x and ∂− y holds in the y ∂+ x ∂− ∂+ x − ηδ(x, y). y = Q(x, y)∂− X 2 f (2)(x, y)δ(x, y) m⊗2(dx dy) := (cid:82) X f (2)(x, x) m(dx). A rigorous meaning of relation (48) is given analogously to formulas (30) -- (32) (see also formulas (50), (52) below). Here (cid:82) +,−(h1, h2, κ(2)). Since X 2 can be represented as the disjoint union of (cid:101)X 2 and ∆, we Proof. To simplify notation, let us consider the case of an operator-valued integral I (2) have (48) +,−(h1, h2, κ(2)χ(cid:101)X 2) + I (2) +,−(h1, h2, κ(2)) = I (2) I (2) +,−(h1, h2, κ(2)χ∆). +,−(h1, h2, κ(2)χ∆) = I (2) +,−(h1, h2, 0) = 0. Hence, in view of the relation ∂− x ∂+ y = δ(x, y) + Q(x, y)∂+ y ∂− x , But I (2) we get where +,−(h1, h2, κ(2)) = I (2) I (2) +,−(h1, h2, κ(2)χ(cid:101)X 2) = I (2)−,+(h2, h1, (Ψ(cid:48) 1 Ψ(cid:48) 1 κ(2)(x1, x2) = κ(2)(x2, x1)Q(x2, x1). κ(2))χ(cid:101)X 2), (49) (50) On the other hand 1 I (2)−,+(h2, h1, Ψ(cid:48) κ(2)) = I (2)−,+(h2, h1, (Ψ(cid:48) = I (2)−,+(h2, h1, (Ψ(cid:48) 1 1 (cid:90) κ(2))χ(cid:101)X 2) + κ(2))χ(cid:101)X 2) + η (cid:90) X X 19 h1(x)h2(x)κ(2)(x, x)Q(x, x) m(dx) h1(x)h2(x)κ(2)(x, x) m(dx). (51) By (49) and (51), +,−(h1, h2, κ(2)) = I (2)−,+(h2, h1, Ψ(cid:48) I (2) 1 κ(2)) − η (cid:90) X h1(x)h2(x)κ(2)(x, x) m(dx). (52) Remark 14. Assume that X = Rd with d ≥ 2, m is the Lebesgue measure on X and Q(x, y) = Q(x1, y1). Then, we have the inclusion F(k) , where we identify each function v(k)(s1, . . . , sk) ∈ L∞(Rk → C, ν(k) (cid:93)1,...,(cid:93)k ) with the function ⊂ G(k) (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k κ(k)(x1, . . . , xk) := v(k)(x1 1, . . . , x1 k) ∈ L∞(X k → C, m(k) ). (cid:93)1,...,(cid:93)k Thus, the above constructed operator-valued integrals give a representation of the Q- CR algebra in the Fock space F Q(H ). 4 Construction of gauge-invariant quasi-free states In this section we again assume X = Rd with d ≥ 2, m is the Lebesgue measure, Q(x, y) = Q(x1, y1) for (x, y) ∈ (cid:101)X 2 and Q(x, y) = η ∈ R for (x, y) ∈ ∆. 4.1 The operators K1, K2 We fix continuous linear operators K1 and K2 in H . We assume that these operators satisfy the following condition. (C) For a bounded measurable function ψ : R → R, let Mψ denote the continuous linear operator in H given by (Mψf )(x1, . . . , xd) := ψ(x1)f (x1, . . . , xd), f ∈ H . Then, for any bounded measurable function ψ : R → R, both operators K1 and K2 commute with Mψ. 1 and K∗ 2 commute with Mψ. Remark 15. Condition (C) implies that, for any bounded measurable function ψ : R → R, both operators K∗ Remark 16. Condition (C) is satisfied if Ki = 1 ⊗ (cid:101)Ki (i = 1, 2), where (cid:101)K1 and (cid:101)K2 are any continuous linear operators in L2(Rd−1, dx2 ··· dxd). In the general case, the operators Ki have the following structure: (Kif )(x1, x2, . . . , xd) =(cid:0)(cid:101)Ki(x1)f (x1,·)(cid:1)(x2, . . . , xd), i = 1, 2, where, for each x1 ∈ R, Ki(x1) is a continuous linear operator in L2(Rd−1, dx2 ··· dxd) such that Ki is a continuous linear operator in H . 20 Remark 17. The results of this section with a proper modification will also hold for X = R. In this case, condition (C) just means that both K1 and K2 are multiplication operators. In fact, under the latter assumption, we could deal with an arbitrary locally compact Polish space X and a function Q as in Section 3. Remark 18. Note that, for each h ∈ HC, we get Kih = Ki i = 1, 2. We extend the operators K1 and K2 by linearity to HC, the complexification of H . ¯h and similarly for K∗ i , Let i1, . . . , ik ∈ {1, 2} and let us consider the operator Ki1 ⊗···⊗ Kik in H ⊗kC . Let h1, . . . , hk ∈ HC and let v(k) ∈ L∞(Rk, ν(k)). Then ϕ(k)(x1, . . . , xk) := h1(x1)··· hk(xk)v(k)(x1 1, . . . , x1 k) belongs to H ⊗kC . By using condition (C), we get the following equality in H ⊗kC , hence m⊗k-a.e.: (Ki1 ⊗ ··· ⊗ Kik)ϕ(k)(x1, . . . , xk) = (Ki1h1)(x1)··· (Kikhk)(xk)v(k)(x1 1, . . . , x1 k). Hence, we may define a linear operator Ki1 ⊗ ··· ⊗ Kik : F(k) → F(k) (cid:93)1,...,(cid:93)k (cid:93)1,...,(cid:93)k (cid:2)h1, . . . , hk, v(k)(cid:3) := Φ(k) (cid:93)1,...,(cid:93)k (cid:2)Ki1h1, . . . , Kikhk, v(k)(cid:3). (53) by (Ki1 ⊗ ··· ⊗ Kik)Φ(k) (cid:93)1,...,(cid:93)k Indeed, the action of Ki1 ⊗ ··· ⊗ Kik onto ϕ(k) ∈ F(k) sentation (26). (cid:93)1,...,(cid:93)k is independent of the repre- 4.2 The representation of the Q-CR algebra corresponding to the operators K1, K2 Given operators K1, K2 satisfying condition (C), we will now construct a corresponding representation of the Q-CR algebra. Our construction is reminiscent of construction of quasi-free states for the CCR and CAR cases using the representations of Araki, Woods [3] and Araki, Wyss [4], respectively. Let X1 and X2 denote two copies of the space X. Let Z := X1 (cid:116) X2 denote the disjoint union of X1 and X2. Thus, Z = X × {1, 2}. We equip Z with the product topology of the space X and the trivial one on {1, 2}. In particular, Z is a locally compact Polish space. With an abuse of notation, we define a measure m on (Z, B(Z)) so that the restriction of this measure to X1 (or X2, respectively) coincides with the measure m on (X, B(X)). In particular, we get L2(Z → C, m) = L2(X1 → C, m) ⊕ L2(X2 → C, m) = HC ⊕ HC. 21 On some occasions, we will identify a point (x, y) ∈ Z 2 with the corresponding point (x, y) ∈ X 2, i.e., we forget which of the two copies of the space X the points x and y belong to. So, again with an abuse of notation, we define a subset ∆ of Z 2 which consists of those points (x, y) ∈ Z 2 for which (x, y) ∈ ∆, where the latter ∆ is the above introduced subset of X 2. Note that m⊗2(∆) = 0 and ∆ contains the diagonal in Z 2. Similarly, if φ : X 2 → C is a function on X 2 and if (x, y) ∈ Z 2, we will denote by φ(x, y) the value of the function φ at the corresponding point (x, y) ∈ X 2. Let a function JQ : Z 2 → C be defined by JQ(x, y) := Q(x, y), Q(y, x), if x, y ∈ X1 or x, y ∈ X2, if x ∈ X1, y ∈ X2 or x ∈ X2, y ∈ X1. (54) (cid:40) In particular, JQ(x, y) = η for all (x, y) ∈ ∆ and JQ(x, y) = 1, JQ(y, x) = JQ(x, y) for all (x, y) ∈ (cid:101)Z 2 := Z 2 \ ∆. So, according to Section 3, we can define the JQ-Fock space over L2(Z, m), i.e., x,i (i ∈ {1, 2}) the creation and F JQ(L2(Z, m)). For x ∈ X, we denote by ∂+ x,i and ∂− annihilation operators at the point x being identified with the corresponding point of (cid:90) Xi. Thus, analogously to (4), we may write, for h ∈ HC, (cid:90) a−(h, 0) = m(dx) h(x)∂+ a+(h, 0) = (cid:90) (cid:90) x,1, X X m(dx) h(x) ∂− x,1, m(dx) h(x) ∂− x,2. (55) a+(0, h) = m(dx) h(x)∂+ x,2, X a−(0, h) = X We now define (informal) operators D+ x and D− x (x ∈ X) which satisfy, for each (cid:90) (cid:90) x,2, m(dx) (K2h)(x)∂− x,2. m(dx) (K2h)(x)∂+ (57) (56) X h ∈ HC:(cid:90) (cid:90) X X (cid:90) (cid:90) X X m(dx) h(x)D+ x := m(dx) h(x)D− x := m(dx) (K1h)(x)∂− x,1 + m(dx) (K1h)(x)∂+ x,1 + X We will now show that, under the assumption (64) below, the operators D+ x satisfy the Q-CR and lead to a representation of the Q-CR algebra. The latter algebra will be generated by the operator-valued integrals x , D− I(k) (cid:93)1,...,(cid:93)k (h1, . . . , hk, v(k)) (cid:90) := X k m⊗k(dx1 ··· dxk) h1(x1)··· hk(xk)v(k)(x1 1, . . . , x1 k)D(cid:93)1 x1 ··· D(cid:93)k xk , (58) where (cid:93)1, . . . , (cid:93)k ∈ {+,−}, h1, . . . , hk ∈ HC, and v(k) ∈ L∞(Rk, ν(k) (53) -- (57), we can now easily formalize the definition (58). (cid:93)1,...,(cid:93)k ). In view of 22 (cid:90) (cid:90) X X (cid:90) (cid:90) Z Z (cid:90) (cid:90) Z Z (59) (60) (61) We define operators K1 : HC → HC ⊕ HC by K1h := (K1h, 0), K2h := (0, K2h), h ∈ HC. We also denote s(1, +) := −, s(2, +) := +, s(1,−) := +, s(2,−) := −. Using these notations, we can rewrite formulas (56), (57) as follows: m(dx) h(x)D+ x = m(dx) h(x)D− x = m(dx) (K1h)(x)∂s(1,+) x + m(dx) (K1h)(x)∂s(1,−) z + m(dx) (K2h)(x)∂s(2,+) z m(dx) (K2h)(x)∂s(2,−) x , . For g1, . . . , gk ∈ HC ⊕ HC and κ(k) ∈ L∞(Z k → C, m(k) (cid:93)1,...,(cid:93)k ), we denote by I (k) (cid:93)1,...,(cid:93)k (g1, . . . , gk, κ(k)) the corresponding operator-valued integral in the JQ-Fock space F JQ(L2(Z, m)) acting on F JQ fin (L2(Z, m)) as defined in Section 3. We now give a rigorous formulation of the definition (58). For any (cid:93)1, . . . , (cid:93)k ∈ {+,−}, h1, . . . , hk ∈ HC, and v(k) ∈ L∞(Rk, ν(k) ), we define (cid:93)1,...,(cid:93)k (cid:88) I(k) (cid:93)1,...,(cid:93)k := (h1, . . . , hk, v(k)) I (k) s(i1,(cid:93)1),...,s(ik,(cid:93)k)(Ki1h1, . . . , Kikhk, R(k) i1,...,ik v(k)), (62) (i1,...,ik)∈{1,2}k where the function R(k) (cid:0)R(k) i1,...,ik i1,...,ik v(k)(cid:1)(x1, . . . , xk) := (cid:40) v(k) ∈ L∞(Z k → C, m(k) ) is given by (cid:93)1,...,(cid:93)k 1, . . . , x1 k), v(k)(x1 0, if (x1, . . . , xk) ∈ Xi1 × ··· × Xik, otherwise. (63) Theorem 19. Let K1 and K2 be continuous linear operators in H which satisfy con- dition (C) and (64) Let the function JQ : Z → C be defined by (54). Let for any (cid:93)1, . . . , (cid:93)k ∈ {+,−}, h1, . . . , hk ∈ HC, and v(k) ∈ L∞(Rk, ν(k) ), a continuous linear operator 1 K1. (cid:93)1,...,(cid:93)k 2 K2 = 1 + ηK∗ K∗ I(k) (cid:93)1,...,(cid:93)k (h1, . . . , hk, v(k)) on F JQ fin (L2(Z, m)) be defined by (62). These operators satisfy the axioms (A1) -- (A5). Thus, the algebra A generated by these operators gives a representation of the Q-CR algebra. 23 Proof. Axiom (A1) follows from condition (C) and the considerations in the last para- graph of subsection 4.1. Axiom (A2) is trivially satisfied. Axiom (A3) follows from the corresponding property of the operator-valued integrals (61) and and Remark 18. Similarly, axiom (A5) is trivially satisfied. So we only have to check the axiom (A4), i.e., the Q-CR relations. case k = 2 by an easy generalization, which we leave to the interested reader. We will prove (A4) only in the case k = 2. The general case will follow from the So let h1, h2 ∈ HC and v(2) ∈ L∞(R2 → C, ν(2) +,+), we get from (54), (59), (60), (62), (63), Proposition 13, and the JQ-CR satisfied by the operator-valued integrals (61): I(2) +,+(h1, h2, v(2)) = I (2)−,−(K1h1, K1h2, R(2) 1,2 v(2)) + I (2) 1,1 Ψ1v(2)) + I (2) = I (2)−,−(K1h2, K1h1, R(2) + I (2)−,+(K1h1, K2h2, R(2) 1,1 v(2)) + I (2) +,−(K2h1, K1h2, R(2) +,+(K2h2, K2h1, R(2) 2,1 v(2)) 2,2 Ψ1v(2)) +,+(K2h1, K2h2, R(2) 2,2 v(2)) + I (2) +,−(K2h2, K1h1, R(2) 2,1 Ψ1v(2)) + I (2)−,+(K1h2, K2h1, R(2) 1,2 Ψ1v(2)) = I(2) +,+(h2, h1, Ψ1v(2)). By taking the adjoint operators, this also implies that I(2)−,−(h1, h2, v(2)) = I(2) +,+(h2, h1, Ψ1v(2)). Using additionally Remark 15 and (64), we get I(2)−,+(h1, h2, v(2)) = I (2) +,−(K1h1, K1h2, R(2) 1,1 v(2)) + I (2)−,−(K2h1, K1h2, R(2) 2,1 v(2)) (cid:90) (cid:90) 2,2 v(2)) X = I (2)−,+(K1h2, K1h1, R(2) + I (2) +,+(K1h1, K2h2, R(2) 1,2 v(2)) + I (2)−,+(K2h1, K2h2, R(2) 1v(2)) − η 1,1 Ψ(cid:48) 1,2 Ψ(cid:48) 1v(2)) + I (2) + I (2)−,−(K1h2, K2h1, R(2) 2,2 Ψ(cid:48) +,−(K2h2, K2h1, R(2) + I (2) +,−(h2, h1, Ψ(cid:48) +,−(h2, h1, Ψ(cid:48) (cid:0)(−ηK∗ h1(x)h2(x)v(2)(x1, x1) m(dx). 1 K1 + K∗ 1v(2)) + 1v(2)) + 1v(2)) + 2 K2)h1 (cid:90) (cid:90) = I(2) = I(2) X (K1h1)(x)(K1h2)(x)v(2)(x1, x1) m(dx) +,+(K2h2, K1h1, R(2) 2,1 Ψ(cid:48) 1v(2)) (K2h1)(x)(K2h2)(x)v(2)(x1, x1) m(dx) (cid:1)(x)h2(x)v(2)(x1, x1) m(dx) Corollary 20. Let k ∈ N, k ≥ 2. Let q ∈ C be such that q (cid:54)= 1 and qk = 1. Then, for each h ∈ HC, X X (cid:18)(cid:90) (cid:19)k = 0. X m(dx) h(x)D+ x 24 Proof. Immediate by Theorems 6 and 19. Remark 21. In fact, a more general statement can be shown on the space F JQ Let the conditions of Corollary 20 be satisfied. Then, for any h1, h2 ∈ HC, we have fin (L2(Z, m)): (cid:18)(cid:90) X (cid:90) X (cid:19)k m(dx)h1(x)∂− x,1 + m(dx)h2(x)∂+ x,2 = 0. (65) We leave the (nontrivial) proof of formula (65) to the interested reader. 4.3 The associated state Let A be the complex algebra from Theorem 19. We define a state τ on A by τ (a) := (aΩ, Ω)F JQ(L2(Z,m)), a ∈ A. (66) Theorem 22. The state τ defined by (66) is a gauge-invariant quasi-free state on the Q-CR algebra A. The state τ is completely determined by the self-adjoint positive operator T := K∗ +,−), 1 K1 in HC and satisfies, for any g, h ∈ HC and v(2) ∈ L∞(R2, ν(2) S(1,1)(g, h, v(2)) = g(x)(T h)(x)v(2)(x1, x1) m(dx), (67) (cid:90) X (cid:90) Rd−1 or equivalently S(1,1)[g, h](s1, s2) = χD(s1, s2) g(s1, x2, . . . , xd)(T h)(s1, x2, . . . , xd) dx2 ··· dxd. In formula (68), D = {(s1, s2) ∈ R2 s1 = s2}. Remark 23. Note that, by (64), in the case η < 0, the operator T additionally satisfies T ≤ −1/η. Remark 24. Note that the representation of the Q-CR algebra from Theorem 19 de- pends on operators K1 and K2 satisfying (64), while the state τ from Theorem 22 depends only on K1 :=(cid:112)K∗ 1 K1. (68) Proof of Theorem 22. For any g1, . . . , gk, h1, . . . , hn ∈ HC and v(k+n) ∈ L∞(Rk+n, ν(k+n) with (cid:93)1 = ··· = (cid:93)k = +, (cid:93)k+1 = ··· = (cid:93)k+n = −, we have by (59), (60), (62), (63), and (66), (cid:93)1,...,(cid:93)k+n ) S(k,n)(gk, . . . , g1, h1, . . . , hn, v(k+n)) (cid:16) (cid:16) = = (cid:17) (cid:0)gk, . . . , g1, h1, . . . , hn, v(k+n)(cid:1)Ω, Ω (cid:0)K1gk, . . . , K1g1, I(k+n) (cid:93)1,...,(cid:93)k+n I (k+n) (cid:93)k+1,...,(cid:93)k+n,(cid:93)1,...,(cid:93)k F JQ(L2(Z,m)) 25 1,...,1 v(k+n)(cid:1)Ω, Ω K1h1, . . . , K1hn, R(k+n) (cid:17) F JQ(L2(Z,k)) . (69) Hence S(k,n) ≡ 0 if k (cid:54)= n, and for k = n we get from (69): S(n,n)(gn, . . . , g1, h1, . . . , hn, v(2n)) (cid:0)K1gn, . . . , K1g1, K1h1, . . . , K1hn, v(2n)(cid:1)Ω, Ω I (2n) (cid:93)n+1,...,(cid:93)2n,(cid:93)1,...,(cid:93)n (cid:16) = (cid:17) F Q(L2(X,m)) . (70) Note that formulas (67), (68) trivially follow from (70) with n = 1. For the general n ∈ N, analogously to the proof of Lemma 12, it suffices to consider the case where v(2n)(s1, . . . , s2n) = un(s1)··· u1(sn)w1(sn+1)··· wn(s2n), (71) where u1, . . . , un, w1, . . . , wn are indicator functions of sets from B(R). Denote g(cid:48) i(x) := gi(x)ui(x1), h(cid:48) i(x) := hi(x)wi(x1), x ∈ X, i = 1, . . . , n. We get from (A1), Remarks 15, 18, and formulas (10), (70), (71) S(n,n)(gn, . . . , g1, h1, . . . , hn, v(2n)) I (2n) (cid:93)n+1,...,(cid:93)2n,(cid:93)1,...,(cid:93)n n, . . . , K1g(cid:48) 1, K1h(cid:48) 1, . . . , K1h(cid:48) (cid:0)K1g(cid:48) (cid:17) n, 1(cid:1)Ω, Ω (cid:19) n)(cid:1)Ω, Ω n)(cid:1) H ⊗nC F Q(L2(X,m)) F Q(L2(X,m)) F Q(L2(X,m)) n)(cid:1) 1) ⊗ ··· ⊗ (K1g(cid:48) π(1) ⊗ ··· ⊗ h(cid:48) π(n)), g(cid:48) H ⊗nC 1 ⊗ ··· ⊗ g(cid:48) n (cid:1) H ⊗nC m(dxn) (K1g(cid:48) n)(xn)∂− xn 1)(x1)∂− x1 m(dx1) (K1g(cid:48) 1) (cid:126) ··· (cid:126) (K1h(cid:48) n), (K1g(cid:48) 1) ⊗ ··· ⊗ (K1h(cid:48) π(1)) ⊗ ··· ⊗ (K1h(cid:48) n), (K1g(cid:48) 1) ⊗ ··· ⊗ (K1g(cid:48) 1) (cid:126) ··· (cid:126) (K1h(cid:48) n)(cid:1) 1) (cid:126) ··· (cid:126) (K1g(cid:48) (cid:0)(K1h(cid:48) π(n))(cid:3)Qπ, (K1g(cid:48) π(n))(cid:3), g(cid:48) (cid:1) 1 ⊗ ··· ⊗ K∗ 1 )Qπ(K1 ⊗ ··· ⊗ K1)(h(cid:48) π(1)) ⊗ ··· ⊗ (T h(cid:48) 1 ⊗ ··· ⊗ g(cid:48) n H ⊗nC X X = = ··· (cid:16) (cid:18)(cid:90) (cid:90) =(cid:0)(K1h(cid:48) = n!(cid:0)Pn(K1h(cid:48) (cid:88) (cid:0)(cid:2)(K1h(cid:48) (cid:88) (cid:0)(K∗ (cid:88) (cid:2)(T h(cid:48) (cid:0)Qπ (cid:90) (cid:88) (cid:90) (cid:88) π∈Sn × u1(x1 (cid:18) n(cid:89) π∈Sn π∈Sn π∈Sn = = = = X n = π∈Sn X n i=1 Qπ(x1 1, . . . , x1 n)(T hπ(1))(x1)g1(x1)··· (T hπ(n))(xn)gn(xn) 1)··· un(x1 n)wπ(1)(x1 1)··· wπ(n)(x1 n) m⊗n(dx1 ··· dxn) gi(xi)(T hπ(i))(xi) Qπ(x1 1, . . . , x1 n) (cid:19) 26 π−1(n)) m⊗n(dx1 ··· dxn) π−1(1), . . . , x1 gi(si, x2, . . . , xd)(T hπ(i))(si, x2, . . . , xd) dx2 ··· dxd (cid:33) × v(2n)(x1 n, . . . , x1 1, x1 (cid:90) (cid:90) (cid:88) (cid:88) (cid:90) = = = R2n (cid:88) i=1 (cid:32) n(cid:89) (cid:90) (cid:32) n(cid:89) (cid:32) n(cid:89) Rn π∈Sn × Qπ(s1, . . . , sn)v(2n)(sn, . . . , s1, sπ−1(1), . . . , sπ−1(n)) ds1 ··· dsn Rd−1 (cid:90) (cid:33) χD(si, sn+π(i)) gi(si, x2, . . . , xd)(T hπ(i))(si, x2, . . . , xd) dx2 ··· dxd π∈Sn × Qπ(s1, . . . , sn)v(2n)(sn, . . . , s1, sn+1, . . . , s2n) ν(2n) i=1 Rd−1 (cid:93)1,...,(cid:93)2n S(1,1)[gi, hπ(i)](si, sn+π(i)) R2n π∈Sn i=1 × Qπ(s1, . . . , sn)v(2n)(sn, . . . , s1, sn+1, . . . , s2n) ν(2n) (cid:93)1,...,(cid:93)2n (ds1 ··· ds2n) (ds1 ··· ds2n). (72) Hence, by (39), the theorem follows. The following corollary immediately follows from (68) and (72). The reader is advised to compare it with formulas (18) -- (20). Corollary 25. Let v(2n) : R2n → C be identically equal to 1. Then, for any g1, . . . , gn, h1, . . . , hn ∈ HC, (cid:33) (cid:33) S(n,n)(gn, . . . , g1, h1, . . . , hn, 1) (cid:90) (cid:88) = (cid:32) n(cid:89) π∈Sn X n i=1 gi(xi)(T hπ(i))(xi) Qπ(x1, . . . , xn) m⊗n(dx1 ··· dxn). Corollary 26. Let η ≥ 0. Let T be a continuous linear operator in H that is self- adjoint and positive. Assume that, for any bounded measurable function ψ : R → R, the operator T commutes with Mψ (see condition (C)). Extend T by linearity to HC. Then there exists a gauge-invariant quasi-free state τ on the Q-CR algebra A that satisfies (67). 0 ≤ T ≤ −1/η. Proof. Choose K1 := see Remark 16. Now the corollary follows from Theorem 22. √ 1 + ηT . Note that K1 and K2 satisfy condition (C), If η < 0, the latter statement remains true if the operator T additionally satisfies T , K2 := √ 5 Particle density Let operators ∂+ density by x , ∂− x (x ∈ X) satisfy the ACR. We heuristically define the particle ρ(x) := ∂+ x ∂− x , x ∈ X. 27 It follows from the Q-CR that these operators commute, cf. [19, 22]. Indeed, for any x, y ∈ X, ρ(x)ρ(y) = ∂+ x ∂+ x ∂− y ∂− = δ(x − y)∂+ = δ(x − y)∂+ y = δ(x − y)∂+ x ∂− y + Q(x, y)∂+ x ∂− x ∂− y ∂+ x + Q(y, x)∂+ x − δ(x − y)∂+ x ∂− x ∂− x + ∂+ y ∂− y ∂− y ∂+ x y ∂− x ∂− x ∂+ y x ∂− x = ρ(y)ρ(x). K1 = κ1 and K2 =(cid:112)1 + ηκ2 1. We will see below that, in order to properly define (cid:82) In this section, we will study the particle density corresponding to the gauge- invariant quasi-free state from Theorem 22 with T = κ21, where κ > 0 is a con- In the case η < 0, we additionally assume that κ2 < 1/η. Thus, we set stant. X m(dx) ϕ(x)ρ(x) for a test function ϕ : X → R, we will need a certain renormaliza- We will also assume that η = (cid:60)q. Note that, with this choice of η, we get tion. Q(x, x) = 1 2 lim x→y, x1>y1 Q(x, y) + lim x→y, x1<y1 Q(x, y , x ∈ X. We will use this value of Q(x, x) as a 'limiting value' of Q(x, y) (x1 (cid:54)= y1) when per- forming renormalization. 5.1 Renormalization (cid:19) (cid:18) We start with heuristic calculations. By (56) and (57) with the above choice of the operators K1 and K2, we get x = κ∂− D+ D− x = κ∂+ x,1 + x,1 + 1 + ηκ2 ∂+ x,2, 1 + ηκ2 ∂− x,2. (cid:112) (cid:112) (cid:112) Hence, the corresponding particle density is given by ρ(x) = D+ x D− x = κ 1 + ηκ2 R(x), (73) where x,1∂+ x,2∂+ x,2∂− x,1∂− x,1 + β∂+ R(x) := ∂+ x,1 + ∂− x,2 + β−1∂− with β := (cid:112)η + κ−2. We denote by D the space of all real-valued infinitely differ- (cid:82) entiable functions on X with compact support. For each ϕ ∈ D, we denote ρ(ϕ) := X m(dx) ϕ(x)ρ(x). We denote by R the real commutative algebra generated by ρ(ϕ) (ϕ ∈ D) and constants. The involution on this algebra is the indentity mapping. We would like to define a vacuum state τ on R analogously to (66). However, we are not able to intepret ρ(ϕ) as a linear operator in F JQ(L2(Z, m)). So we need a proper renor- malization. For this, as discussed in Introduction, we will use Ivanov's formula [25] in the form δ2 = δ. (74) x,2 28 Below we will denote by (cid:12) symmetric tensor product. Let D(cid:12)an denote the nth symmetric algebraic tensor power of D, i.e., the vector space of finite sums of functions on X n of the form f1(cid:12)···(cid:12) fn, where f1, . . . , fn ∈ D. For each f (n) ∈ D(cid:12)an, we denote W (f (n)) := m⊗n(dx1 ··· dxn)f (n)(x1, . . . , xn)∂+ x1,2∂+ x1,1 ··· ∂+ xn,2∂+ xn,1Ω. (cid:90) X n (cid:18)(cid:90) X (cid:19) (cid:90) X We stress that W (f (n)) is treated as a formal expression. Note that, for f1, . . . , fn ∈ D, W (f1 (cid:12) ··· (cid:12) fn) = m(dx1)f1(x1)∂+ x1,2∂+ x1,1 ··· m(dxn)fn(xn)∂+ xn,2∂+ xn,1 Ω, xj ,1 commute for i (cid:54)= j. We also set D(cid:12)a0 := R where we used that ∂+ and for f (0) ∈ R, we set W (f (0)) := f (0)Ω. We denote by Ffin(D) the real linear space of vectors of the form xi,1 and ∂+ xj ,2∂+ xi,2∂+ F = (f (0), f (1), . . . , f (n), 0, 0, . . . ), where fi ∈ D(cid:12)ai. For such F ∈ Ffin(D), we denote n(cid:88) W (F ) := W (f (i)). i=0 The following proposition will be central for our considerations. Proposition 27. For each ϕ ∈ D, we define a linear operator R(ϕ) : Ffin(D) → Ffin(D) by R(ϕ) := a+(ϕ) + (β + β−1η)a0(ϕ) + a− 1 (ϕ) + ηa− 2 (ϕ) + β−1 (cid:90) X ϕ(x) m(dx). (75) Here a+(ϕ) is the symmetric creation operator: for f (n) ∈ D(cid:12)an we have a+(ϕ)f (n) := ϕ (cid:12) f (n); a0(ϕ) is the neutral operator: for f1, . . . , fn ∈ D n(cid:88) a0(ϕ)f1 (cid:12) ··· (cid:12) fn := f1 (cid:12) ··· (cid:12) (ϕfi) (cid:12) ··· (cid:12) fn; a− 1 (ϕ) is the annihilation operator: 1 (ϕ)f1 (cid:12) ··· (cid:12) fn := a− (cid:90) n(cid:88) i=1 ϕ(x)fi(x) m(dx) f1 (cid:12) ··· (cid:12) fi (cid:12) ··· (cid:12) fn, i=1 X 29 where fi denotes the absence of fi; and a− as follows: 2 (ϕ) is an annihilation operator which acts f1 (cid:12) ··· (cid:12) fmin{i,j}−1 (cid:12) (ϕfifj) (cid:12) ··· (cid:12) fmax{i,j} (cid:12) ··· (cid:12) fn. 2 (ϕ)f1 (cid:12) ··· (cid:12) fn := a− Let i=1 R(ϕ) := n(cid:88) (cid:88) (cid:90) j(cid:54)=i X (cid:16) (cid:112) (cid:17)−1 Then, the JQ-commutation relations satisfied by ∂+ the renormalization formula δ2 = δ imply that, for each F ∈ Ffin(D), x,iΩ = 0, and m(dx)ϕ(x)R(x) = ρ(ϕ). κ x,i, ∂− 1 + ηκ2 x,i, the condition ∂− R(ϕ)W (F ) = W ( R(ϕ)F ). m(dy)ϕ(y)∂+ y,2∂+ y,1 W (F ) = W (a+(ϕ)F ). (76) (77) (cid:90) m⊗(n+1)(dy dx1 ··· dxn)ϕ(y)f (n)(x1, . . . , xn) Proof. We trivially get(cid:90) (cid:90) X For each f (n) ∈ D(cid:12)an, we have m(dy)ϕ(y)∂− × ∂− y,1∂− y,2∂+ y,1∂− x1,2∂+ y,2 W (f (n)) = x1,1 ··· ∂+ xn,2∂+ X n+1 xn,1Ω. X Note that and for i (cid:54)= j Hence, ∂− y,2∂+ xi,2∂+ xi,1 = δ(y − xi)δ+ = δ(y − xi)δ+ xi,2∂− xi,1 + Q(y, xi)∂+ xi,1∂− xi,1 + ∂+ y,2, xi,2∂+ y,2∂+ xi,1 ∂+ xi,1∂+ xj ,2∂+ xj ,1 = ∂+ xj ,2∂+ xj ,1∂+ xi,1. y,1∂− ∂− y,2∂+ n(cid:88) x1,1 ··· ∂+ x1,2∂+ δ(y − xi)∂− = xn,2∂+ xn,1Ω x1,1 ···(cid:0)∂+ y,1∂+ x1,2∂+ xi,2, ∂+ xi,1 (cid:1)··· ∂+ xn,2∂+ xn,1∂+ xi,1Ω. i=1 Here and below (··· ) denotes the absence of the corresponding term. Similarly, we have ∂− y,1∂+ xj ,2∂+ xj ,1 = Q(xj, y)∂+ xj ,2∂− y,1∂+ xj ,1 30 (78) (79) (80) xj ,2δ(y − xj) + ∂+ Thus, using (79) and (81), we continue (80) as follows: = Q(xj, y)∂+ = η∂+ xj ,2δ(y − xj) + ∂+ xj ,2∂+ xj ,1∂− y,1, xj ,2∂+ xj ,1∂− y,1 (cid:88) n(cid:88) n(cid:88) i=1 = η + δ(y − xi)δ(y − xj)∂+ x1,2∂+ j(cid:54)=i δ(y − xi)∂+ x1,2∂+ x1,1 ···(cid:0)∂+ xi,2, ∂+ xi,1 xmax{i,j},2∂+ xmax{i,j},1 xn,2∂+ xn,1∂− y,1∂+ xi,1Ω. x1,1 ···(cid:0)∂+ (cid:1)··· ∂+ (cid:1)··· ∂+ xn,2∂+ xn,1Ω i=1 We also note that δ(y − xi)∂− xi,1Ω = δ2(y − xi)Ω = δ(y − xi)Ω. y,1∂+ Hence, by (80), (82), and (83), y,2 W (f (n)) = W(cid:0)(cid:0)a− 2 (ϕ)(cid:1)f (n)(cid:1) . 1 (ϕ) + ηa− m(dy)ϕ(y)∂− y,1∂− (cid:90) X Similarly, y,2∂− ∂+ y,2∂+ x1,2∂+ n(cid:88) x1,1 ··· ∂+ xn,2∂+ δ(y − xi)∂+ = ∂+ y,2 i=1 δ(y − xi)∂+ x1,2∂+ x1,2∂+ xn,1Ω (cid:1)∂+ x1,1 ···(cid:0)∂+ x1,1 ···(cid:0)∂+ (cid:1)∂+ y,2∂+ xi,2 xi,2 xi,1 ··· ∂+ xn,2∂+ xn,1Ω xi,1 ··· ∂+ xn,2∂+ xn,1Ω δ(y − xi)∂+ x1,2∂+ x1,1 ··· ∂+ xi,2∂+ xi,1 ··· ∂+ xn,2∂+ xn,1Ω, = = which implies i=1 n(cid:88) n(cid:88) (cid:90) i=1 Finally, X m(dy)ϕ(y)∂+ y,2W (f (n)) = W (a0(ϕ)f (n)). y,2∂− ∂− y,1∂+ y,1∂+ = ∂− x1,2∂+ y,1∂+ n(cid:88) = η xn,2∂+ xn,1Ω xn,2∂+ x1,1 ··· ∂+ x1,1 ··· ∂+ x1,2∂+ δ(y − xi)∂+ x1,1 ··· ∂+ x1,2∂+ xn,2∂+ xn,1∂+ y,1Ω x1,1 ··· ∂+ xn,1∂− y,1∂+ xi,2 y,1Ω i=1 + ∂+ x1,2∂+ (cid:0)∂+ xi,1 (cid:1)··· ∂+ xn,2∂+ xn,1∂+ y,1Ω 31 (81) (82) (83) (84) (85) n(cid:88) = η x1,2∂+ x1,1 ··· ∂+ xi,2∂+ xi,1 ··· ∂+ xn,2∂+ xn,1Ω i=1 + ∂+ x1,2∂+ xn,2∂+ xn,1Ω. Here we used that m(dy)ϕ(y)∂− y,1∂+ (cid:90) X δ(y − xi)∂+ x1,1 ··· ∂+ (cid:90) (cid:90) (cid:90) y,1Ω = = X 2 X 2 = y,1∂+ u,1Ω m⊗2(dy du)δ(y − u)ϕ(y)∂− m⊗2(dy du)δ2(y − u)ϕ(y)Ω m⊗2(dy du)δ(y − u)ϕ(y)Ω = (cid:90) X ϕ(y) m(dy)Ω. Thus,(cid:90) X X 2 m(dy) ϕ(y)∂− y,1∂+ y,1W (f (n)) = W (ηa0(ϕ)f (n)) + (cid:90) X ϕ(y) m(dy)W (f (n)). (86) Now, formula (76) follows from (74), (77), (84) -- (86). Corollary 28. We have {rΩ r ∈ R} = {W (F ) F ∈ Ffin(D)}. More precisely, for each F ∈ Ffin(D), there exists a unique r ∈ R such that W (F ) = rΩ. This correspondence is given through formula (76). Vice versa, for each r ∈ R, there exists a unique F ∈ Ffin(D) such that W (F ) = rΩ holds. Proof. By Proposition 27, we have, for ϕ ∈ D, R(ϕ)Ω = W (ϕ) + β−1Ω. Hence W (ϕ) = (R(ϕ) − β−1)Ω. By (75) and (76), for any n ≥ 2 and f1, . . . , fn ∈ D, W (f1 (cid:12) ··· (cid:12) fn) = R(f1)W (f2 (cid:12) ··· (cid:12) fn) 1 (f1) + ηa− (β + β−1η)a0(f1) + a− (cid:18)(cid:20) 2 (f1) + β−1 − W (87) (cid:19) . f2 (cid:12) ··· (cid:12) fn (cid:21) f1(x) m(dx) (cid:90) X (88) Hence, for F ∈ Ffin(D), a unique representation of W (F ) as rΩ (r ∈ R) follows by induction on n and formulas (87), (88). The converse statement follows immediately from Proposition 27 by induction. 32 Corollary 28 implies that, for each r ∈ R, there exists a unique vector F = (f (0), f (1), . . . , f (n), 0, 0, . . . ) ∈ Ffin(D) (89) such that rΩ = f (0)Ω + (cid:90) n(cid:88) i=1 X i m⊗i(dx1 ··· dxi)f (i)(x1, . . . , xi)∂+ x1,2∂+ x1,1 ··· ∂+ xi,2∂+ xi,1Ω. Thus, we can rigorously define a linear mapping τ : R → R by τ (r) := f (0). However, since we used the renormalization, from our construction of τ is not a priori clear whether τ is positive definite, i.e., whether τ (r2) ≥ 0 for each r ∈ R. So, our next aim is to decide whether positive definiteness holds. 5.2 Measure-valued L´evy processes and positive definiteness of τ We start with preliminaries on measure-valued L´evy process. For more detail, see e.g. [23, 26 -- 28, 31]. Let M(X) denote the space of all Radon measures on X. We equip M(X) with the topology of vague convergence. Let B(M(X)) denote the corresponding Borel σ- algebra on M(X). Let (Ω, A , P ) be a provability space. A random measure on X is a measurable mapping γ : Ω → M(X). A competely random measure is a random measure γ such that, for any mutually disjoint sets A1, . . . , An ∈ B0(X), the random variables γ(A1), . . . , γ(An) are independent. Here B0(X) denotes the subset of B(X) that consists of all bounded sets from B(X). A measure-valued L´evy process is a completely random measure γ such that, for any sets A1, A2 ∈ B0(X) with m(A1) = m(A2), the random variables γ(A1) and γ(A2) are identically distributed. and only if there exist a constant c ≥ 0 and a measure ζ on R+ := (0,∞) satisfying It follows from [27] that a random measure γ is a measure-valued L´evy process if such that the Laplace transform of γ is given by (cid:20) (cid:90) E(cid:0)e(cid:104)f,γ(cid:105)(cid:1) = exp support and (cid:104)f, γ(cid:105) :=(cid:82) X c f (x) m(dx)+ f ∈ C0(X), f ≤ 0. (90) Here C0(X) denotes the space of real-valued continuous functions on X with compact X f (x) γ(dx). The measure ζ is called the L´evy measure of the (esf (x)−1) ζ(ds) m(dx) R+ X , measure-valued L´evy process γ. (cid:21) min{s, 1} ζ(ds) < ∞ (cid:90) R+ (cid:90) (cid:90) 33 on X, i.e., Radon measures of the form (cid:80) Let K(X) denote the subset of M(X) consisting of all discrete Radon measures i∈I siδxi, where the set I is either finite or countable, si > 0, and δxi denotes here the Dirac measure with mass at xi. We also assume that xi (cid:54)= xj if i (cid:54)= j. A discrete random measure is a random measure which takes values in K(X) a.s. Each measure-valued L´evy process γ for which c = 0 in formula (90) is a discrete random measure. The space Γ(X) of multiple configurations in X is the subset of K(X) which consists i∈I siδxi with si ∈ N. A point process on X is a of all Radon measures of the form(cid:80) Radon measures of the form(cid:80) i∈I δxi. Each Radon measure(cid:80) with Fourier transform (90) has the property that a.s. γ =(cid:80) The configuration space Γ(X) is defined as the subset of Γ(X) which consists of all i∈I δxi can be identified with the locally finite set {xi i ∈ I} ⊂ X. A simple point process on X is a random measure γ which takes values in Γ(X) a.s. It should be noted that if c = 0 and ζ(R+) < ∞, the measure-valued L´evy process i∈I siδxi, where the set {xi i ∈ I} is locally finite, i.e., a configuration in X. On the other hand, if ζ(R+) = ∞, the set {xi i ∈ I} is a.s. dense in X. random measure γ which takes values in Γ(X) a.s. By choosing c = 0 and the measure ζ in (90) to be zδ1 with z > 0, one obtains the simple point process γ with Laplace transform E(cid:0)e(cid:104)f,γ(cid:105)(cid:1) = exp (cid:20)(cid:90) (cid:21) (ef (x) − 1) zm(dx) . X This γ is called the Poisson point process with intensity measure zm, since for each A ∈ B0(X), the random variable γ(A) has Poisson distribution with parameter zm(A). Theorem 29. The functional τ on the real algebra R is positive definite (i.e., τ (r2) ≥ 0 for each r ∈ R) if and only if η ≥ 0. Furthermore, if η = 0, then τ (ρ(f1)··· ρ(fn)) = E(cid:0)(cid:104)f1, γ(cid:105)···(cid:104)f1, γ(cid:105)(cid:1), f1, . . . , fn ∈ D, where γ is the Poisson point process on X with intensity measure κ2m. (91) If η > 0, then (91) holds with γ being a negative binomial point process. More precisely, γ is the measure-valued L´evy process with Laplace transform (90) in which c = 0 and (cid:18) η ∞(cid:88) (cid:19)k 1 (92) For each A ∈ B0(X), the distribution of the random variable γ(A) is the negative binomial distribution η + κ−2 ζ = δk. k=1 k 1 η (1 + κ2η)−m(A)/η η n! δn. (93) (cid:16) m(A) (cid:17)(n) (cid:19)n (cid:18) η η + κ−2 ∞(cid:88) n=0 34 Here, we used the standard symbol a(n) := a(a + 1)(a + 2)··· (a + n − 1), the so-called rising factorial. The proof of Theorem 29 is based on the property of orthogonal polynomials of a L´evy white noise proved in [43, Theorem 2.1 and Corollaries 2.1, 2.3] and [5], see also [9, Theorem 1.2]. We will now briefly explain this result in the special case of a measure-valued L´evy process. Let γ be a measure-valued L´evy process such that c = 0 in (90) (so that γ is a discrete random measure). We denote ζ(cid:48)(ds) := s2ζ(ds), the so-called Kolmogorov measure of the measure-valued L´evy process γ. We assume that ζ(cid:48) is a probability measure on R+, and furthermore, eεs ζ(cid:48)(ds) < ∞ for some ε > 0. (94) (cid:90) R+ The latter assumption implies that the set of polynomials is dense in L2(R+, ζ(cid:48)). If the support of the measure ζ(cid:48) has infinitely many points, we will denote by (pk)∞ k=0 the system of monic polynomials that are orthogonal with respect to ζ(cid:48). These polynomials satisfy the recurrence relation spk(s) = pk+1(s) + bkpk(s) + akpk−1(s), k = 0, 1, 2, . . . (95) with p−1(s) := 0, ak > 0, and bk ∈ R. Let assume that the σ-algebra A from the probability space (Ω, A , P ) is the min- imal σ-algebra with respect to which γ(A) is measurable for each A ∈ B0(X). We denote by C P the set of continuous polynomials of γ, i.e., the set of random variables of the form f (0) + (cid:104)f (i), γ⊗i(cid:105), (96) where f (i) ∈ D(cid:12)ai and n ∈ N. If f (n) (cid:54)= 0, we call the random variable in (96) a continuous polynomial of γ of degree n. Condition (94) implies that C P is a dense subset of L2(Ω, P ). Let us denote C P n the subset of C P which consists of all polynomials of γ of degree ≤ n. Let M P n denote the closure of C P n in L2(Ω, P ) (measurable polynomials of degree ≤ n). Let OP n := M P n (cid:9) M P n−1, where (cid:9) means the orthogonal difference in L2(Ω, P ) (orthogonal polynomials of degree n). As a result, we get the OP n. For f (n) ∈ D(cid:12)an, we denote by P (f (n)) the orthogonal projection of the monomial (cid:104)f (n), γ⊗n(cid:105) onto OP n. Note that P (f (n)) does not need to belong to C P. For F ∈ Ffin(D) as in (89), we denote i=0 P (f (i)). The set of all such random variables we denote by OC P orthogonal decomposition L2(Ω, P ) = (cid:76)∞ P (F ) := (cid:80)n n=0 (orthogonalized continuous polynomials). n(cid:88) i=1 35 Theorem 30. Let γ be a measure-valued L´evy process that satisfies the above assump- tions. We have OC P ⊂ C P (and, in fact, OC P = C P) if and only is there exist √ constants η ≥ 0 and λ > 0 with λ ≥ 2 η such that: if η = 0 then ζ = λ−2δλ; and if η > 0 then the measure ζ(cid:48) has infinitely many points in its support and the system k=0 of monic polynomials that are orthogonal with respect to ζ(cid:48) satisfies the recur- (pk)∞ rence relation (95) with ak = ηk(k + 1) and bk = λ(k + 1). Furthermore, for any η and λ as above, we have, for ϕ ∈ D and F ∈ Ffin(D), 2 (ϕ) + 2(cid:0)λ + λ2 − 4η(cid:1)−1(cid:105) (cid:112) (cid:17) F , (cid:104)ϕ, γ(cid:105)P (F ) = P a+(ϕ) + λa0(ϕ) + a− (cid:16)(cid:104) where the operators a+(ϕ), a0(ϕ), and a− √ Remark 31. If η > 0 and λ > 2 η, we get from Theorem 30 and e.g. [12, Chapter VI, Section 3] that the L´evy measure of the corresponding measure-valued L´evy process γ is 2 (ϕ) were defined in Proposition 27. 1 (ϕ) + ηa− 1 (ϕ), and a− ζ = 1 η with δ(υ− η υ )k , υ := (λ + λ2 − 4η)/2, and for A ∈ B0(X), the distribution of the random variable γ(A) is the negative binomial distribution(cid:18) υ2 − η (cid:19)m(A)/η ∞(cid:88) (cid:18)m(A) (cid:19)(k) υ2 υ2 k! η k=0 δ(υ− η υ )k . ∞(cid:88) k=1 k υ2 (cid:19)k 1 (cid:18) η (cid:112) (cid:19)k 1 (cid:18) η Thus, γ is a negative binomial random measure. √ η, the L´evy measure of γ is In the case where η > 0 and λ = 2 ζ(ds) = − s√ e η ds, 1 sη and for A ∈ B0(X), the distribution of the random variable γ(A) is the gamma distri- bution (cid:18) m(A) (cid:19)(cid:21)−1 m(A) η 2η Γ η m(∆) η −1 e u − u√ η χR+(u) du. Thus, γ in this case is the gamma random measure, see e.g. [29, 30, 47]. The Laplace transform of γ can also be written in the form E(cid:0)e(cid:104)f,γ(cid:105)(cid:1) = exp − 1 η (cid:90) log(cid:0)1 − √ η f (x)(cid:1) m(dx) (cid:21) , f ∈ C0(X), f ≤ 0. (97) (cid:20) (cid:20) X 36 Proof of Theorem 29. Let η > 0. Set λ = β + β−1η. We have 2(cid:0)λ + (cid:112) λ2 − 4η(cid:1)−1 = β−1. As easily seen, β > √ η. Hence, λ = β + √ > 2 η. η β Let γλ,η be the measure-valued L´evy process on X from Theorem 30 that corresponds to the parameters λ, η. By Proposition 27 and Theorem 30, we get τ(cid:0)R(f1)··· R(fn)(cid:1) = E(cid:0)(cid:104)f1, γλ,η(cid:105)···(cid:104)fn, γλ,η(cid:105)(cid:1), We define the measure-valued L´evy process γ := κ(cid:112)1 + ηκ2 γλ,η. Let ζ and ζλ,η denote ζλ,η under the mapping R+ (cid:51) s (cid:55)→ κ(cid:112)1 + ηκ2 s ∈ R+. By (73) and (98), we get (91). the L´evy measure of γ and γλ,η, respectively. Then ζ is the pushforward of the measure f1, . . . , fn ∈ D. (98) Formulas (92), (93) follow by direct calculations from Remark 31. Note that, since the L´evy measure ζ is concentrated on N, γ is a point process. The proof for η = 0 is analogous. Let f ∈ D. We easily calculate Let us now prove that τ is not positive definite for η < 0. Assume the contrary. (cid:18)(cid:90) τ(cid:0)W (f ⊗ f )2(cid:1) = 2 (cid:19)2 (cid:90) f 2(x) m(dx) + 2η f 4(x) m(dx) ≥ 0. From here we conclude by approximation that, for any cube A in X, X X m(A)2 + ηm(A) ≥ 0, which obviously fails for A small enough. Remark 32. In view of Theorem 29, we can rigorously understand ρ(ϕ) (ϕ ∈ D) as the operator of multiplication by (cid:104)ϕ, γ(cid:105) in the Hilbert space L2(Ω, P ), which maps C P into itself. It follows from [43, Lemma 2.1 and Theorem 2.1] that each ρ(ϕ) is essentially self-adjoint on C P. Corollary 33. Let η > 0. Denote by τκ the state τ on R which corresponds to the operator T = κ21. Let γ be the gamma random measure with Laplace transform (97). Then, for any f1, . . . , fn ∈ D, we have (cid:0)R(f1)··· R(fn)(cid:1) = lim (cid:16) (cid:112) (cid:17)−n = E(cid:0)(cid:104)f1, γ(cid:105)···(cid:104)fn, γ(cid:105)(cid:1). 1 + ηκ2 κ→∞ κ lim κ→∞ τκ (cid:0)ρ(f1)··· ρ(fn)(cid:1) τκ 37 Proof. The statement follows immediately from Proposition 27, Theorem 30, and Re- mark 31 if we note that (cid:0)(η + κ−2)1/2 + (η + κ−2)−1/2η(cid:1) = 2 √ η, lim κ→∞ κ→∞(η + κ−2)−1/2 = η−1/2. lim Acknowledgements The author acknowledges the financial support of the the SFB 701 "Spectral struc- tures and topological methods in mathematics", Bielefeld University, and of the Pol- ish National Science Center, grant no. Dec-2012/05/B/ST1/00626. I am grateful to M. Bozejko, G.A. Goldin, Y. Kondratiev, A. Vershik, and J. Wysocza´nski for useful discussions. I would like to thank the referees for a careful reading of the manuscript and making useful comments and suggestions. References [1] Accardi, L., Franz, U., Skeide, M.: Renormalized squares of white noise and other non-Gaussian noises as L´evy processes on real Lie algebras. Comm. Math. Phys. 228 (2002), 123 -- 150. [2] Accardi, L., Lu, Y.G., Volovich, I.V.: White noise approach to classical and quan- tum stochastic calculi. Centro Vito Volterra, Universita di Roma Tor Vergata, Preprint 275, 1999. [3] Araki, H., Woods, E.: Representations of the C.C.R. for a nonrelativistic infinite free Bose gas, J. Math. Phys. 4 (1963), 637 -- 662. [4] Araki, H., Wyss, W.: Representations of canonical anticommutation relation, Helv. Phys. Acta 37 (1964), 136 -- 159. [5] Berezansky, Yu.M., Lytvynov, E., Mierzejewski, D.A.: The Jacobi field of a L´evy process. Ukrainian Math. J. 55 (2003), 853 -- 858. [6] Berezin, F.A.: The method of second quantization. (In Russian) Nauka, Moscow, 1965. English translation: Academic Press, New York -- London, 1966. [7] Berezin, F.A.: Some remarks on the representations of commutation relations. (In Russian) Uspehi Mat. Nauk 24 (1969), no. 4, 65 -- 88. [8] Bozejko, M., Speicher, R.: Completely positive maps on Coxeter groups, deformed commutation relations, and operator spaces. Math. Ann. 300 (1994), 97 -- 120. 38 [9] Bozejko, M., Lytvynov, E., Rodionova, I.: An extended anyon Fock space and non- commutative Meixner-type orthogonal polynomials in infinite dimensions. Russian Math. Surveys 70 (2015), 857 -- 899. [10] Bozejko, M., Lytvynov, E., Wysocza´nski, J.: Noncommutative L´evy processes for generalized (particularly anyon) statistics. Comm. Math. Phys. 313 (2012), 535 -- 569. [11] Bratteli, O., Robinson, D.W.: Operator algebras and quantum-statistical mechan- ics. II. Equilibrium states. Models in quantum-statistical mechanics. Springer- Verlag, New York-Berlin, 1981. [12] Chihara, T.S.: An introduction to orthogonal polynomials, Gordon and Breach, New York, 1978. [13] Daletskii, A., Kalyuzhnyi, A.: L2 dimensions of spaces of braid-invariant harmonic forms. J. Geom. Phys. 62 (2012), 1309 -- 1322. [14] Dell'Antonio, G.F.: Structure of the algebra of some free systems, Comm. Math. Phys. 9 (1968), 81 -- 117. [15] Dell'Antonio, G., Figari, R., Teta, A.: Statistics in space dimension two. Lett. Math. Phys. 40 (1997), 235 -- 256. [16] Derezi´nski, J., G´erard, C.: Mathematics of quantization and quantum fields. Cam- bridge University Press, Cambridge, 2013. [17] Frappat, L., Sciarrino, A., Sciuto, S., Sorba, P.: Anyonic realizations of the quan- tum affine Lie ailgebra Uq((cid:98)AN−1). Phys. Lett. B 369 (1996), 313 -- 324. [18] Gel'fand, I.M., Graev, M.I., Vershik, A.M.: Models of representations of current groups. Representations of Lie groups and Lie algebras (Budapest, 1971), 121 -- 179, Akad. Kiad´o, Budapest, 1985. [19] Goldin, G.A., Majid, S.: On the Fock space for nonrelativistic anyon fields and braided tensor products. J. Math. Phys. 45, 3770 -- 3787 (2004) [20] Goldin, G.A., Menikoff, R., Sharp, D.H.: Particle statistics from induced repre- sentations of a local current group. J. Math. Phys. 21 (1980), 650 -- 664. [21] Goldin, G.A., Menikoff, R., Sharp, D.H.: Representations of a local current algebra in nonsimply connected space and the Aharonov -- Bohm effect. J. Math. Phys. 22 (1981), 1664 -- 1668. [22] Goldin, G.A., Sharp, D.H.: Diffeomorphism groups, anyon fields, and q commu- tators. Phys. Rev. Lett. 76, 1183 -- 1187 (1996) 39 [23] Hagedorn, D., Kondratiev, Y., Pasurek, T., Rockner, M.: Gibbs states over the cone of discrete measures. J. Funct. Anal. 264 (2013), 2550 -- 2583. [24] Hiai, F.: q-deformed Araki -- Woods algebras. Operator algebras and mathematical physics (Constanta, 2001), 169 -- 202, Theta, Bucharest, 2003. [25] Ivanov V.K.: The algebra of elementary generalized functions. Dokl. Akad. Nauk SSSR 246 (1979), 805 -- 808 (English transl.: Soviet Math. Dokl. 20 (1979), 553 -- 556). [26] Kallenberg, O.: Random measures, 3rd Edition. Akad.-Verl., Berlin, 1983 [27] Kingman, J.F.C.: Completely random measures. Pacific J. Math. 21 (1967), 59 -- 78. [28] Kingman, J.F.C.: Poisson processes. Oxford Studies in Probability, 3. Oxford Science Publications. The Clarendon Press, Oxford University Press, New York, 1993. [29] Kondratiev, Y.G., da Silva, J.L., Streit, L., Us, G.F.: Analysis on Poisson and gamma spaces. Infin. Dimens. Anal. Quantum Probab. Relat. Top. 1 (1998), 91 -- 117. [30] Kondratiev, Y.G., Lytvynov, E.W.: Operators of gamma white noise calculus. Infin. Dimens. Anal. Quantum Probab. Relat. Top. 3 (2000), 303 -- 335. [31] Kondratiev, Y., Lytvynov, E., Vershik, A.: Laplace operators on the cone of Radon measures. J. Funct. Anal. 269 (2015), 2947 -- 2976. [32] Leinass, J.M., Myrheim, J.: On the theory of identical particles. Nuovo Cimento 37 B (1977), 1 -- 23. [33] Lerda, A., Sciuto, S.: Anyons and quantum groups. Nuclear Phys. B 401 (1993), 613 -- 643. [34] Liguori, A., Mintchev, M.: Fock spaces with generalized statistics. Lett. Math. Phys. 33 (1995), 283 -- 295. [35] Liguori, A., Mintchev, M.: Fock representations of quantum fields with generalized statistics. Comm. Math. Phys. 169 (1995), 635 -- 652. [36] Liguori, A., Mintchev, M.: Quantum field theory, bosonization and duality on the half line. Nuclear Phys. B 522 (1998), 345 -- 372. [37] Liguori, A., Mintchev, M., Pilo, L.: Bosonization at finite temperature and anyon condensation. Nuclear Phys. B 569 (2000), 577 -- 605. 40 [38] Liguori, A., Mintchev, M., Rossi, M.: Anyon quantum fields without a Chern- Simons term. Phys. Lett. B 305 (1993), 52 -- 58. [39] Liguori, A., Mintchev, M., Rossi, M.: Unitary group representations in Fock spaces with generalized exchange properties. Lett. Math. Phys. 35 (1995), 163 -- 177. [40] Liguori, A., Mintchev, M., Rossi, M.: Representations of Uq( AN ) in the space of continuous anyons. J. Phys. A 29 (1996), L493 -- L498. [41] Liguori, A., Mintchev, M., Rossi, M.: Fock representations of exchange algebras with involution. J. Math. Phys. 38 (1997), 2888 -- 2898. [42] Lytvynov, E.: Fermion and boson random point processes as particle distributions of infinite free Fermi and Bose gases of finite density. Rev. Math. Phys. 14 (2002), 1073 -- 1098. [43] Lytvynov, E.: Polynomials of Meixner's type in infinite dimensionsJacobi fields and orthogonality measures. J. Funct. Anal. 200 (2003), 118 -- 149. [44] Lytvynov, E., Mei, L.: On the correlation measure of a family of commuting Hermitian operators with applications to particle densities of the quasi-free rep- resentations of the CAR and CCR. J. Funct. Anal. 245 (2007), 62 -- 88. [45] Shlyakhtenko, D.: Free quasi-free states. Pacific J. Math. 177 (1997), 329 -- 368. [46] Stern, A.: Anyons and the quantum Hall effect -- a pedagogical review. Ann. Physics 323, 204 -- 249 (2008) [47] Tsilevich, N, Vershik, A, Yor, M.: An infinite-dimensional analogue of the Lebesgue measure and distinguished properties of the gamma process. J. Funct. Anal. 185 (2001), 274 -- 296. [48] Vershik, A.M., Gel'fand, I.M., Graev, M.I.: Representation of SL(2, R), where R is a ring of functions. Russian Math. Surveys 28 (1973), 83 -- 128. [English translation in "Representation Theory," London Math. Soc. Lecture Note Ser., Vol. 69, pp. 15 -- 60, Cambridge Univ. Press, Cambridge, UK, 1982.] [49] Vershik, A.M., Gel'fand, I.M., Graev, M.I.: Commutative model of the representa- tion of the group of flows SL(2, R)X connected with a unipotent subgroup. Funct. Anal. Appl. 17 (1983), 80 -- 82. [50] Wilczek, F.: Quantum mechanics of fractional-spin particles. Phys. Rev. Lett. 49 (1982), 957 -- 959. [51] Wilczek, F.: Magnetic flux, angular momentum, and statistics. Phys. Rev. Lett. 48 (1982), 1144 -- 1145 41
1207.3038	1	1207	2012-07-12T18:04:27	On C*-Algebras Generated by Isometries with Twisted Commutation Relations	[ "math.OA", "math.KT" ]	In the theory of C-algebras, interesting noncommutative structures arise as deformations of the tensor product. For instance, the rotation algebra may be seen as a scalar twist deformation of the tensor product of the functions on the circle with itself. We deform the tensor product of two Toeplitz algebras in the same way, introducing the universal C-algebra generated by two isometries u and v such that uv=e^{it}vu and uv=e^{-it}vu, for a fixed real parameter t. Since the second relation implies the first one, we also consider the universal C-algebra generated by two isometries u and v with the weaker relation uv=e^{it}vu. Such a "weaker case" does not exist in the case of unitaries, and it turns out to be much more interesting than the twisted "tensor product case" of two Toeplitz algebras. We show that the C-algebra in the "tensor product case" is nuclear, whereas in the "weaker case" it is not even exact. Also, we compute the K-groups and we obtain K_0 = Z and K_1 = 0 for both C-algebras. This answers a question raised by Murphy in 1994 concerning the K-theory of the C-algebra associated to N^2.	math.OA	math	ON C ∗-ALGEBRAS GENERATED BY ISOMETRIES WITH TWISTED COMMUTATION RELATIONS MORITZ WEBER Abstract. In the theory of C ∗-algebras, interesting noncommutative structures arise as deformations of the tensor product. For instance, the rotation algebra Aϑ may be seen as a scalar twist deformation of the tensor product C(S 1)⊗C(S 1). We deform the tensor product of two Toeplitz algebras in the same way, introducing the universal C ∗-algebra T ⊗ϑ T generated by two isometries u and v such that uv = e2πiϑvu and u∗v = e−2πiϑvu∗, for ϑ ∈ R. Since the second relation implies the first one, we also consider the universal C ∗-algebra T ∗ϑ T generated by two isometries u and v with the weaker relation uv = e2πiϑvu. Such a "weaker case" does not exist in the case of unitaries, and it turns out to be much more interesting than the twisted "tensor product case" of two Toeplitz algebras. We show that T ⊗ϑ T is nuclear, whereas T ∗ϑ T is not even exact. Also, we compute the K-groups and we obtain K0(T ∗ϑ T) = Z and K1(T ∗ϑ T) = 0, and the same K-groups for T ⊗ϑ T. This answers a question raised by Murphy in 1994 concerning the K-theory of the C ∗-algebra C ∗(N2), which is T ∗ϑ T with ϑ = 0. Introduction Given two unital C ∗-algebras A and B, a new object may be constructed by forming the (maximal) tensor product A ⊗ B. It is given by the elements of A and B such that all elements of the first C ∗-algebra commute with all elements of the second, i.e. ab = ba for all a ∈ A and b ∈ B. Deformations of these commutation relations (resp. of the tensor product construction) give rise to interesting structures in operator algebras, which are investigated with the aim of understanding settings with noncommutative multiplications. An important example of a noncommutative C ∗-algebra obtained in such a way is the (irrational) rotation algebra Aϑ. It has been studied since the early days of the theory (see for instance [R81], [EH67], [Z68]), and it is one of the guiding lights in the development of several tools and theories (e.g. in Connes' noncommutative geometry, [Co85], [Co94]). Furthermore, it serves as a model for many phenomena in physics (see [GVF01, Ch. 12] for a short overview). Date: September 26, 2018. 2010 Mathematics Subject Classification. 46L05 (Primary); 46L80, 19Kxx, 46L65 (Secondary). Key words and phrases. universal C ∗-algebra, rotation algebra, noncommutative torus, isome- tries, commutation relations, twist. Research supported by the Deutsche Forschungsgemeinschaft and the Graduiertenkolleg "Analytische Topologie und Metageometrie". This work has been done in the context of the author's PhD project at the University of Muenster. 1 2 MORITZ WEBER The rotation algebra Aϑ is the universal C ∗-algebra generated by two unitaries u and v such that uv = λvu, where λ = e2πiϑ is a scalar of absolute value one, for ϑ ∈ R. For ϑ = 0, the rotation algebra may be seen as the tensor product C(S1) ⊗ C(S1) of the algebra C(S1) of continuous functions on the circle with itself. For arbitrary ϑ ∈ R it is the twisted tensor product C(S1) ⊗ϑ C(S1), i.e. it is the universal C ∗-algebra generated by two unitaries with twisted commutation relations. Unitaries are the isomorphisms in the category of Hilbert spaces -- they are iso- metric and surjective. Isometries (i.e. not necessarily surjective unitaries) in turn form a more general class of operators on Hilbert spaces. The universal C ∗-algebra generated by a single isometry is the well-known Toeplitz algebra T, an extension of C(S1) by the C ∗-algebra K of compact operators. In this article, we investigate C ∗-algebras generated by two isometries with twisted commutation relations. Surprisingly, a systematic study of this structure has not yet been done (for partial results by others, see the end of section 1). Furthermore, forming the ∗-algebraic product of isometries reveals problems which do not arise in the case of unitaries. A first step is to twist the tensor product T ⊗ T of two Toeplitz algebras. This C ∗- algebra is generated by two isometries u and v such that uv = vu and u∗v = vu∗. The second relation implies the first one but not conversely (in contrast to the case of two unitaries). Therefore, we introduce two types of twisted commutation relations. First the tensor twist of two isometries T ⊗ϑ T, i.e. the universal C ∗-algebra generated by two isometries u and v such that u∗v = ¯λvu∗ (and hence also uv = λvu), where again λ = e2πiϑ and ϑ ∈ R. Secondly, the free twist of two isometries T ∗ϑT, i.e. the universal C ∗-algebra generated by two isometries u and v such that uv = λvu. For unitaries u and v there is only one way to twist the commutation relation uv = vu whereas for isometries we have these two versions, which behave quite differently. It turns out that the C ∗-algebra T ∗ϑ T is in some sense much more complicated than T ⊗ϑ T. Viewed on the complexity scale of nuclearity and exactness, T ∗ϑ T carries much more "free" structure whereas T ⊗ϑ T behaves more like the tensor product T ⊗ T. This is due to the first main result of this article, namely that T ⊗ϑ T is nuclear whereas T ∗ϑ T is not even exact. In order to show this second, much harder claim, we consider a Wold type decomposition of the isometries u and v. We obtain a good description of the kernel of the canonical map T ∗ϑT → Aϑ and of further important ideals in T ∗ϑ T. The decomposition is inspired by the work of Berger, Coburn and Lebow ([BCL78]) on commuting isometries, which is T ∗ϑ T in the case ϑ = 0. This also leads to the second main result, the computation of the K-groups of T∗ϑT. It turns out that they are the same as those of T⊗ϑ T. While the computations for the C ∗-algebra T ⊗ϑ T are not so hard, we have to make an effort in the case of T ∗ϑ T going beyond tensor product like structures. The computation of the K- groups of T ∗ϑ T answers a question raised by Murphy in 1994 ([M94]) concerning the C ∗-algebra C ∗(N2) which is T ∗ϑ T in the case ϑ = 0. ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 3 Some aspects of T⊗ϑT and T∗ϑT were studied by Proskurin, Kabluchko, Jørgensen, and Samoılenko ([Pr00a], [Pr00b], [Ka01], [JPS05]). See the end of section 1 for an overview on their results and other related work. The article is organized as follows. In section 1, we introduce the C ∗-algebras T ⊗ϑ T and T ∗ϑ T. Section 2 deals with the ideals generated by a defect projection 1 − uu∗ resp. 1 − vv∗ in T ⊗ϑ T (resp. T ⋊ϑ Z). From this we deduce that T ⊗ϑ T is nuclear. In section 3, we prove that the ideal J generated by the two defect projections 1 − uu∗ and 1 − vv∗ in T ∗ϑT contains all non-trivial ideals in T ∗ϑT, for irrational parameters ϑ. This yields the classification of the C ∗-algebras T ⊗ϑ T and T ∗ϑ T with respect to the parameter ϑ, if ϑ is irrational. In section 4, we construct an embedding ι of T ∗ϑT into (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2) by analyzing the Wold decomposition of the isometry uv ∈ T ∗ϑ T. This enables us to give a satisfying picture of the ideal J in T ∗ϑ T, namely it is isomorphic to (C(S1) ∗C C2) ⊗ K. We infer that T ∗ϑ T is not exact. For irrational parameters ϑ, the ideal structure of T ∗ϑT corresponds to that of C(S1) ∗C C2, which we investigate in some concrete cases, see section 5. The K-theory of the C ∗-algebras T ⊗ϑ T and T ∗ϑ T is computed in section 6. At the end of this article, an appendix on examples and constructions of universal C ∗-algebras is attached. 1. The twisted commutation relations of two isometries The rotation algebra Aϑ is an important example of a noncommutative C ∗-algebra (see the introduction of this article). Recall its definition as a universal C-algebra. Definition 1.1. Let ϑ ∈ R be a parameter and put λ = e2πiϑ. The rotation algebra Aϑ is the universal C ∗-algebra generated by two unitaries u and v such that uv = λvu. One of the main features of the rotation algebra is its close relation to a commuta- tive C ∗-algebra, the algebra C(T2) of continuous functions on the 2-torus. This alge- bra may be seen as the universal C ∗-algebra generated by two commuting unitaries, due to the Gelfand isomorphism. For ϑ = 0, the rotation algebra and C(T2) coincide. Thus, the rotation algebra (also called the noncommutative torus) is something of a soft step into the world of noncommutative C ∗-algebras, and many questions about Aϑ find the inspiration for their solutions in the case ϑ = 0. On the other hand, if the parameter ϑ is irrational, the rotation algebra is highly noncommutative. Another perspective is to view the rotation algebra Aϑ as the twisted tensor product of the algebra C(S1) of continuous functions on the circle with itself. Indeed, as C(S1) is the universal C ∗-algebra generated by a single unitary, the tensor product C(S1) ⊗ C(S1) is the universal C ∗-algebra generated by two commuting unitaries u and v (see the appendix for a short overview on universal C ∗-algebras) which is Aϑ in the case ϑ = 0. Hence, the rotation algebra Aϑ is the deformed tensor product C(S1) ⊗ϑ C(S1). 4 MORITZ WEBER Our aim is to study versions of Aϑ involving isometries instead of unitaries, which is a natural generalization. Isometries are isometric transformations of the underly- ing Hilbert space, i.e. they preserve the essential structure of the space, the inner product. Unitaries in turn are surjective isometries -- they are the isomorphisms in the category of Hilbert spaces. These properties may be expressed in an algebraic way, i.e. u is a unitary if and only if uu∗ = u∗u = 1, whereas an isometry v is given by v∗v = 1. The universal C ∗-algebra generated by a single isometry v is the Toeplitz algebra T, a well-known object. The canonical map from T to C(S1), mapping v 7→ u, gives rise to the following short exact sequence. 0 → K → T → C(S1) → 0 Here the C ∗-algebra of compact operators on a separable Hilbert space are denoted by K. In order to twist the tensor product of two Toeplitz algebras, we observe that T ⊗ T is the universal C ∗-algebra generated by two isometries u and v such that uv = vu and u∗v = vu∗. (Note that the relation u∗v = vu∗ is needed to let all ∗-monomials in u commute with all ∗-monomials in v.) If u and v are unitaries, these two relations are equivalent, but in the case of isometries we only have one of the implications. The following statement is also mentioned in an article by Jørgensen, Proskurin, and Samoılenko ([JPS05]). Lemma 1.2. Let u and v be two isometries in a unital C ∗-algebra A and let λ ∈ S1 be a complex number of absolute value one. Then (i) u∗v = ¯λvu∗ ⇒ uv = λvu (ii) u∗v = ¯λvu∗ 6⇐ uv = λvu Proof. (i) We have (uv − λvu)∗(uv − λvu) = 0. (ii) Consider d(λ) ∈ L(ℓ2(N0)), defined by d(λ)en = λnen, and the unilateral shift S ∈ L(ℓ2(N0)), defined by Sen = en+1. Then u′ := d(λ)S and v′ := S are isometries with u′v′ = λv′u′, but u′∗v′ 6= ¯λv′u′∗ since u′∗v′e0 = ¯λe0 whereas v′u′∗e0 = 0. (cid:3) Therefore, we distinguish two cases of twisted commutation relations. Definition 1.3. Let ϑ ∈ R be a parameter and put λ = e2πiϑ ∈ S1. • The tensor twist of two isometries T ⊗ϑ T is defined as the universal C ∗- algebra generated by two isometries u and v such that u∗v = ¯λvu∗. • The free twist of two isometries T ∗ϑT is defined as the universal C ∗-algebra generated by two isometries u and v such that uv = λvu. We immediately see that T ⊗ϑ T = T ⊗ T if ϑ = 0. On the other hand, the (unital) free product of T with itself and the C ∗-algebra T ∗ϑ T do not coincide in the case ϑ = 0. Nevertheless, T ∗ϑT carries much more free structure than T ⊗ϑ T since there are no direct relations between u∗ and v, i.e. the monomial structure of T ∗ϑ T is much more complicated. Also we will see that the free group C ∗-algebra C ∗(F2) embeds into T ∗ϑT. This is a further hint about the features of a free product in the case of T ∗ϑT. In the following we will not distinguish between rational and irrational parameters ϑ if not explicitely stated. Also, we will always write λ := e2πiϑ. ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 5 Let us quickly remark that the deformed tensor product of C(S1)⊗ T is a standard construction. Recall that Aϑ may be constructed as a crossed product of C(S1) by the group of integers Z according to the isomorphism u 7→ λu on C(S1). Likewise T⋊ϑZ is the crossed product of the Toeplitz algebra T by the automorphism v 7→ λv. Remark 1.4. Let ϑ ∈ R be a parameter and put λ = e2πiϑ. Then T ⋊ϑ Z is the universal C ∗-algebra generated by a unitary u and an isometry v such that uv = λvu (or equivalently u∗v = ¯λvu∗). A first note on the difference between T ⊗ϑ T and T ∗ϑ T is the following remark on the range projections of the generating isometries. Remark 1.5. Consider the range projections in T ⊗ϑ T and T ∗ϑ T. (a) The range projections uu∗ and vv∗ in T ⊗ϑ T commute. (b) The range projections uu∗ and vv∗ in T ∗ϑ T do not commute. Proof. A proof of (b) will be given later (see Remark 4.4). (cid:3) At this point, we could ask whether the behavior of the commutator [uu∗, vv∗] is the only difference between T ⊗ϑ T and T ∗ϑ T. Let B denote the quotient of T ∗ϑ T by the ideal generated by the commutator [uu∗, vv∗]. By the representation in the proof of Lemma 1.2(b), B does not equal T ⊗ϑ T, i.e. the canonical homomorphism from T ⊗ϑ T to B is not an isomorphism. Furthermore, we will see later that T ⊗ϑ T is nuclear whereas B is not even exact (see Proposition 5.8). The C ∗-algebras T ⊗ϑ T and T ∗ϑ T have been considered by several authors be- fore, although from a different perspective and not in the homogeneous franework of twisting the commutation relations of isometries. Proskurin ([Pr00a], [Pr00b]) studied the C ∗-algebra T ⊗ϑ T under the name A{0},{λij }, and he proved that it is nuclear. We will give a different proof in section 2. Furthermore he showed, that the defect ideal h1 − uu∗, 1 − vv∗i ⊳ T ⊗ϑ T (which is the kernel of the map T ⊗ϑ T → Aϑ) is the largest ideal in T ⊗ϑ T for irrational parameter ϑ. We extend his approach to the more general case of h1 − uu∗, 1 − vv∗i ⊳ T ∗ϑ T. Proskurin also considered the Fock representation of T ⊗ϑ T and he showed that it is faithful for irrational parameters. Kabluchko ([Ka01]) extended Proskurin's result on the Fock representation to rational parameters. Using this, he gave a concrete description of the defect ideals of T ⊗ϑ T (he denoted T ⊗ϑ T by A{0},Θ) for all parameters ϑ, which can be found in section 2 -- although we prove it in a different way. He showed this using the Fock representation of T ⊗ϑ T; our proof is focused on the algebraic structures. Jørgensen, Proskurin and Samoılenko ([JPS05]) proved in terms of a notion called "∗-wildness", that T ∗ϑ T -- which they call Aq 2 -- is not nuclear. We even go further and show that T ∗ϑT is not exact. They also gave a classification of the C ∗-algebras T∗ϑT, depending on the parameter ϑ. It is directly transferred from the classification of the rotation algebras Aϑ. 6 MORITZ WEBER All of the above articles refer to an article by Bozejko and Speicher ([BS94]), where they introduced the so called qij-relations (or qij-CCR) on elements d1, . . . , dn, namely did∗ j − qijd∗ j di = δij for i, j = 1, . . . , n. Here, the qij are complex scalars with qij ≤ 1 and qij = ¯qji. For n = 2, q11 = q22 = 0 and q12 = ¯λ, these are the relations of T ⊗ϑ T under the correspondence u ↔ d∗ 1 and v ↔ d∗ 2. Thus, T ⊗ϑ T may be seen as the universal C ∗-algebra generated by qij-relations in a very special case. Note, that the di are co-isometries in our case. There is also some connection to Murphy's work on crossed products by semi- groups. In his article Crossed products of C ∗-algebras by semigroups of automor- phisms from 1994 ([M94]), he defines (amongst other things) a C ∗-algebra associ- ated to (unital) semigroups equiped with a 2-cocycle. In [M94, Example 3.3], he ϑ(N2), which is associated to the semigroup N2 and a considers the C ∗-algebra C ∗ 2-cocycle, constructed out of a single complex scalar λ of absolute value one. This is the universal C ∗-algebra generated by two isometries u and v fulfilling the relation uv = λvu; hence C ∗ ϑ(N2) = T ∗ϑ T. In the introduction to this article, he mentions that the K-theory of this object was unknown, even for the case of trivial λ, i.e. for C ∗(N2) = T ∗ϑ T, where ϑ = 0. We may fill in this gap in chapter 6 (see Remark 6.9). According to Murphy, the knowledge of this K-theory would help in the theory of generalized Toeplitz operators (see [M94] or [M96] for references on this). Murphy ([M96]) also investigated the structure of the most important ideal in T ∗ϑ T, namely the defect ideal h1 − uu∗, 1 − vv∗i ⊳ T ∗ϑ T. He did this only for the case ϑ = 0, but not for arbitrary ϑ. Finally, we should mention the considerations by Berger, Coburn and Lebow ([BCL78]) concerning the universal C ∗-algebra generated by commuting isometries u and v, thus the C ∗-algebra T ∗ϑ T in the case of ϑ = 0. Their work plays a crucial role in Section 4. They analyzed the Wold decomposition of the isometry uv ∈ T ∗ϑT for ϑ = 0. We generalize it to arbitrary parameters ϑ which provides our main tool, namely a transparent picture of T ∗ϑ T. 2. Ideals in T ⊗ϑ T and T ⋊ϑ Z In this section, we take a first look at the ideal structure of T ⊗ϑ T and T ⋊ϑ Z. We find a description of their defect ideals generated by 1 − uu∗ resp. by 1 − vv∗. The case of the C-algebra T ∗ϑT is more complicated and will be treated in the next two sections. The main result of this section is to prove that T ⊗ϑ T is nuclear. From the universal property, we infer the existence of the following natural maps, mapping the generators u 7→ u and v 7→ v. T ⊗ϑ T ։ T ⋊ϑ Z ։ Aϑ ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 7 The kernel of the map T ⋊ϑ Z ։ Aϑ is the ideal h1 − vv∗i generated by the defect projection of the isometry v (recall Remark 1.4). In T ⊗ϑ T, we have the ideals h1 − uu∗i and h1 − uu∗, 1 − vv∗i arising as kernels of the according maps to T ⋊ϑ Z resp. Aϑ. The ideal h1 − vv∗i in T ⊗ϑ T is the kernel of the flipped quotient map T ⊗ϑ T → T ⋊−ϑ Z, given by u 7→ v and v 7→ u. In the case of T ⋊ϑ Z and T ⊗ϑ T, we may easily describe the defect ideals h1 −uu∗i and h1 − vv∗i. They are given by a tensor product of C(S1) resp. T with the C ∗-algebra of compact operators. To prove this, we need the following lemma on twisted tensor products with the compacts. Note that the C ∗-algebra K of compact operators may be seen as the universal C ∗-algebra generated by elements xi, i ∈ N0 i xj = δij for all i, j ∈ N0 (see also the appendix). The element with the relations x∗ x0 is a minimal projection, whereas all xi are partial isometries. Definition 2.1. Let A be a C ∗-algebra and α := (αi)i∈N0 ⊆ Aut(A) be automor- phisms αi of A for i ∈ N0 with α0 = id. Let Bα be the universal C ∗-algebra generated by elements a ∈ A (together with the relations of A) and elements xi, i ∈ N0 such that x∗ i xj = δijx0 and axi = xiαi(a) for all a ∈ A, i ∈ N0. We define the twisted tensor product A ⊗α K of A with the compacts to be the ideal generated by all products ab, a ∈ A, b ∈ K in Bα. If A is unital, then A ⊗α K = Bα. If αi = id for all i ∈ N0, then A ⊗α K coincides with the usual tensor product A ⊗ K (see also the appendix). In fact, even for arbitrary automorphisms αi, the twisted tensor product A ⊗α K is isomorphic to the untwisted one, as is shown in the next lemma. Lemma 2.2. Let A be a C ∗-algebra and α := (αi)i∈N0 ⊆ Aut(A) be automorphisms on A, with α0 = id. The twisted tensor product A ⊗α K with the compacts is isomorphic to the untwisted tensor product: A ⊗α K ∼= A ⊗ K via axi ↔ αi(a) ⊗ xi i = xix∗ H = (cid:0)Li≥0 π(xix∗ Proof. We write A ⊗ K as A ⊗id K ⊳ Bid with the automorphisms αi = id. Let π : Bid → L(H) be a faithful representation of Bid and decompose H by i )H(cid:1) ⊕ K for a subspace K ⊆ H (recall that the xix∗ i are mutually orthogonal projections). As A may be embedded into Bid and since i a for all a ∈ A, i ∈ N0, we get representations βi : A → L(π(xix∗ axix∗ i )H), i ) for i ∈ N0; likewise βK : A → L(K), a 7→ π(a)pK, where a 7→ π(αi(a))π(xix∗ pK is the projection onto K ⊆ H. Thus, we get a representation β : A → L(H), βi(a)⊕ βK (a) of A on L(H). It fulfills β(a)π(xi) = π(αi(a)xi) ∈ π(A⊗idK) a 7→ Li≥0 and π(xi) β(αi(a)) = π(xix0) β0(αi(a)x0) = π(αi(a)xi) = β(a)π(xi). From this, we get a representation σ : Bα → L(H), mapping a 7→ β(a) and xi 7→ π(xi). Restricting σ to A ⊗α K, we obtain a map from A ⊗α K to A ⊗id K. Note that A ⊗α K is spanned by elements axix∗ j , hence the image of A ⊗α K under σ is in π(A ⊗id K) since σ(axix∗ j ) = π(αi(a)xi)π(x∗ j ) ∈ π(A ⊗id K). 8 MORITZ WEBER So, we constructed a homomorphism A ⊗α K → A ⊗id K, axi likewise we obtain a homomorphism A ⊗id K → A ⊗α K, axi 7→ α−1 maps are inverse to each other. 7→ αi(a)xi, and i (a)xi. These (cid:3) Recall that C(S1) is the universal C ∗-algebra generated by a unitary u, whereas T is the universal C ∗-algebra generated by an isometry v. Proposition 2.3. The ideals generated by the single defect projections in T ⋊ϑ Z and T ⊗ϑ T are of the following form. (a) C(S1) ⊗ K ∼= h1 − vv∗i ⊳ T ⋊ϑ Z via u ⊗ xi 7→ viu(1 − vv∗) (b) T ⊗ K ∼= h1 − vv∗i ⊳ T ⊗ϑ T via v ⊗ xi 7→ viu(1 − vv∗) T ⊗ K ∼= h1 − uu∗i ⊳ T ⊗ϑ T via v ⊗ xi 7→ uiv(1 − uu∗) Proof. (a) We consider the automorphisms αj on C(S1) for j ∈ N0, given by u 7→ λju, where λ = e2πiϑ. Hence, we can form the according twisted tensor product C(S1) ⊗ϑ K (as an ideal in Bϑ) which is isomorphic to C(S1) ⊗ K by the preceding lemma. Here, Bϑ is the universal C ∗-algebra generated by elements xi, i ∈ N0 and a unitary w such that x∗ i xj = δijx0 and wxi = λixiw for all i ∈ N0. i = λix′ The elements x′ i := vi(1 − vv∗) and u in T ⋊ϑ Z fulfill the relations x′∗ j = δijx′ 0 iu for all i ∈ N0. Thus, there is a homomorphism β : Bϑ → T ⋊ϑ Z and ux′ mapping xi 7→ x′ i and w 7→ u. The ideal C(S1) ⊗ϑ K in Bϑ is spanned by all elements j for k ∈ Z, i, j ∈ N0. Therefore, the restriction of β to C(S1) ⊗ϑ K maps onto wkxix∗ j ) = ukvi(1 − vv∗)(v∗)j for k ∈ Z, the closed linear span S of all elements β(wkxix∗ i, j ∈ N0. The span S is an ideal in T ⋊ϑ Z containing the projection 1 − vv∗, hence β(C(S1) ⊗ϑ K) = S = h1 − vv∗i. i x′ We next show that the restriction of β to C(S1) ⊗ϑ K is injective. Choose a unital, faithful representation π : Bϑ ֒→ L(H) and decompose H = (cid:0)Li≥0 π(xix∗ of ℓ2(N0). There is a Hilbert space isomorphism Li≥0 π(xix∗ i )H(cid:1) ⊕ K, where K ⊆ H is a subspace of H. Write (ei)i≥0 for the standard orthonormal basis i )H ∼= ℓ2(N0) ⊗H0, with i )H = H0, because i xi).) Now, we analyze π(w) and H0 := π(x0)H, via the mapping π(xi)η ↔ ei ⊗ η. (Note, that π(x∗ π(x∗ π(xi) under this decomposition of H. i ) = π((xix0)∗) = π(x0)π(x∗ i ) and π(x0) = π(x∗ The subspace H0 ⊆ H is invariant under π(w), since π(w)η = π(wx0)η = π(x0w)η is in H0 for all η ∈ H0. Even more, π(w)H0 = H0 holds, since π(w∗)H0 ⊆ H0. Therefore w := π(w)H0 ∈ L(H0) is a unitary. Under the Hilbert space isomorphism of K ⊥ and ℓ2(N0)⊗H0, the operator π(w)K ⊥ corresponds to d(λ)⊗ w, where d(λ) ∈ L(ℓ2(N0)) denotes the diagonal operator given by d(λ)en = λnen. This is due to the following computation: π(w)K ⊥(ei ⊗ η) ↔ π(w)π(xi)η = λiπ(xi)π(w)η ↔ (d(λ) ⊗ w)(ei ⊗ η) ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 9 The operator π(xj)K ⊥ correponds to Sj(1 − SS∗) ⊗ 1, where S ∈ L(ℓ2(N0)) is the unilateral shift, given by Sen = en+1: π(xj)K ⊥(ei ⊗ η) ↔ π(xjxi)η = π(xjx∗ 0xi)η = δi0π(xj)η ↔ (Sj(1 − SS∗) ⊗ 1)(ei ⊗ η) On the subspace K ⊆ H, π(w)K is a unitary, since π(w)K = K. Furthermore, the operators π(xi) get annihilated on K for all i ∈ N0. Summarizing, we can write π(w) = (d(λ) ⊗ w) ⊕ π(w)K and π(xi) = (Si(1 − SS∗) ⊗ 1) ⊕ 0. The restriction of π to C(S1) ⊗ϑ K is a representation on the Hilbert space K ⊥, j ) = (d(λ)kSi(1 − SS∗)S∗j ⊗ wk) ⊕ 0. On the other hand, there because π(wkxix∗ is a map σ : T ⋊ϑ Z → L(K ⊥), mapping u 7→ d(λ) ⊗ w and v 7→ S ⊗ 1. Hence, the following diagram commutes and the restriction of β to C(S1) ⊗ϑ K is injective (since π is injective). C(S1) ⊗ϑ K ց π β → T ⋊ϑ Z L(K ⊥) ւ σ We conclude that C(S1) ⊗ϑ K is isomorphic to the ideal h1 − vv∗i in T ⋊ϑ Z. The ideal C(S1) ⊗ϑ K in Bϑ in turn is isomorphic to C(S1) ⊗ K via wxi 7→ λiu ⊗ xi by Lemma 2.2. (b) The proof for the C ∗-algebra T⊗ϑ T is exactly the same. Note that in this case, the operators w ∈ L(H0) and π(w)K ∈ L(K) are not unitaries but isometries. (cid:3) A proof of the fact that T ⊗ K is isomorphic to the defect ideal h1 − vv∗i ⊳ T ⊗ϑ T may be found in an article by Kabluchko ([Ka01, prop. 4]), but he proves it in a quite different way by means of representation theory, via a Fock representation. We attack the problem from a different perspective analyzing the algebraic structure of the span and "untwisting" the twist with the compacts. The result of Proposition 2.3(i) may also be deduced from a result by Williams ([Wi07]). We form the crossed product of the following short exact sequence with the automorphism v 7→ λv on T. 0 −→ K −→ T −→ C(S1) −→ 0 Hence, we obtain the following exact sequence. 0 −→ K ⋊ϑ Z −→ T ⋊ϑ Z −→ C(S1) ⋊ϑ Z −→ 0 The kernel of the map T ⋊ϑ Z → C(S1) ⋊ϑ Z = Aϑ is the defect ideal h1 − vv∗i, on the other hand K ⋊ϑ Z ∼= C(S1) ⊗ K. As an immediate consequence of the preceding proposition, we obtain the following result on T ⊗ϑ T. Theorem 2.4. The C ∗-algebra T ⊗ϑ T is nuclear. 10 MORITZ WEBER Proof. By the previous proposition we have the following short exact sequence: 0 → T ⊗ K → T ⊗ϑ T → T ⋊ϑ Z → 0 Since nuclearity is preserved under taking crossed products with amenable groups, we infer that T⋊ϑ Z is nuclear. Now, T⊗K is nuclear, too, and hence also T⊗ϑ T. (cid:3) We now sketch a more direct proof of the fact that T ⊗ϑ T is nuclear using a result by Rosenberg from 1977 ([Ro77]). Lemma 2.5 ([Ro77, th. 3]). Let A be a unital C ∗-algebra, B ⊆ A be a nuclear C ∗-subalgebra containing the unit of A, s ∈ A be an isometry such that sBs∗ ⊆ B and let A = C ∗(B, s), i.e. let A be generated by B and s. Then A is nuclear. In fact, the isometry s ∈ A induces an endomorphism Ad(s) : B → B, since sBs∗ ⊆ B. Thus, Rosenberg's lemma states that nuclearity is preserved under taking crossed products with the semigroup N, since A = B ⋊Ad(s) N. Now, the C ∗-subalgebra B1 := C ∗(ukvn(v∗)n(u∗)k n, k ∈ N0) ⊆ T ⊗ϑ T is com- mutative and hence nuclear. We apply Rosenberg's lemma to B1, the isometry u, and A1 := C ∗(B1, u) ⊆ T ⊗ϑ T. Thus, A1 is nuclear. We apply it again to B2 := A1 and v ∈ C ∗(B2, v) = T ⊗ϑ T, and we infer that T ⊗ϑ T is nuclear. This proof is a slight modification of a proof by Proskurin ([Pr00a, Prop. 3], [Pr00b]). A glance at the details of the sketched proof shows that the quotient D of T ∗ϑT by the ideal generated by all commutators [ua(u∗)a, vb(v∗)b], a, b ∈ N0 is nuclear. This is because the C ∗-subalgebra B1 ⊆ D -- defined analogously to B1 ⊆ T ⊗ϑ T -- is commutative, exactly because of the relations [ua(u∗)a, vb(v∗)b] = 0 for all a, b ∈ N0. Nevertheless, the canonical map from D to T ⊗ϑ T is not an isomorphism, again due to the representation of Lemma 1.2(b). This explains again that the commutators [ua(u∗)a, vb(v∗)b] are not the only difference between T ⊗ϑ T and T ∗ϑ T, refining the answer to the question after Remark 1.5. The C ∗-algebra T ∗ϑ T in turn is not nuclear. This was shown by Jørgensen, Proskurin and Samoılenko within their concept of "∗-wildness"([JPS05]). A C ∗- algebra is called ∗-wild, if its representation theory may be traced back in a certain way (which we do not specify here) to the representation theory of the C ∗-algebra C ∗(F2) of the free group with two generators. They showed that T ∗ϑT is ∗-wild and that every ∗-wild C ∗-algebra is not nuclear. In section 4 we will show that T ∗ϑ T is not even exact. A satisifying study of the ideals in T ∗ϑ T and also of the ideal h1 − uu∗, 1 − vv∗i in T ⊗ϑ T require further preparation. An explicit description is given in section 4. We end this section by a remark on the ideal h(1 − uu∗)(1 − vv∗)i in T ⊗ϑ T and on the quotient by it. In some sense, it shows how close T ⊗ϑ T and T ⋊ϑ Z are. Again, we view T ⋊ϑ Z as the universal C ∗-algebra generated by a unitary u and an isometry v such that uv = λvu (see Remark 1.4). ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 11 Proposition 2.6. The ideal h(1 − uu∗)(1 − vv∗)i in T ⊗ϑ T is isomorphic to the C ∗-algebra K of compact operators. The quotient D of T ⊗ϑ T by this ideal embeds into (T ⋊ϑ Z) ⊕ (T ⋊ϑ Z) via u 7→ u ⊕ v and v 7→ v ⊕ u∗. Proof. The ideal h(1 − uu∗)(1 − vv∗)i is the closed linear span of all elements of := uavb(1 − uu∗)(1 − vv∗)(v∗)c(u∗)d for a, b, c, d ∈ N0. These the form e(a,b)(d,c) e∗ (a,b)(d,c) = e(d,c)(a,b) and elements fulfill the relations of the matrix units, e(a,b)(d,c)e(f,g)(i,h) = δ(d,c)(f,g)e(a,b)(i,h), where δ(d,c)(f,g) is defined as the Kronecker delta δdf δcg. Thus h(1 − uu∗)(1 − vv∗)i is isomorphic to K (see also the appendix). i.e. By the universal property, there is a homomorphism ϕ from the quotient D to (T ⋊ϑ Z) ⊕ (T ⋊ϑ Z) such that u 7→ u ⊕ v and v 7→ v ⊕ u∗. Furthermore, let π be a unital, faithful representation of D on a Hilbert space H. Then H may be decomposed into H = H1 ⊕ H2 ⊕ H3 such that • π(u) = u1 ⊕ u2 ⊕ u3, where u1 is an isometry and u2, u3 are unitaries, • π(v) = v1 ⊕ v2 ⊕ v3, where v1, v2 are unitaries and v3 is an isometry, • and uivi = λviui for i = 1, 2, 3. Indeed, put ai := (uv)i(1 − uu∗)((uv)∗)i and bj := (uv)j(1 − vv∗)((uv)∗)j in D for i, j ∈ N0. Then the ai and bj are mutually orthogonal projections. We also have uai = ai+1u, ubi = biu, vai = aiv and vbi = bi+1v for all i ∈ N0. Check for instance: uai = λi(uv)iu(vv∗ + (1 − vv∗))(1 − uu∗)((uv)∗)iu∗u = λi(uv)i+1v∗(1 − uu∗)((uv)∗)iu∗u = λiλ−i(uv)i+1(1 − uu∗)((uv)∗)i+1u = ai+1u We put H1 := Li∈N0 π(ai)H, H3 := Li∈N0 π(bi)H and H2 := H ⊖ (H1 ⊕ H3). We infer from the relations on u, ai and bi that π(u)H1 ⊆ H1, π(u)H2 = H2, and π(u)H3 = H3. Similarly we get π(v)H1 = H1, π(v)H2 = H2 and π(v)H3 ⊆ H3. We have a homomorphism σ : (T ⋊ϑ Z) ⊕ (T ⋊ϑ Z) → L(H), mapping u ⊕ 0 7→ 0 ⊕ u2 ⊕ u3 v ⊕ 0 7→ 0 ⊕ v2 ⊕ v3 0 ⊕ u 7→ v∗ 1 ⊕ 0 ⊕ 0 0 ⊕ v 7→ u1 ⊕ 0 ⊕ 0 Since σ ◦ ϕ = π, we conclude that ϕ is injective. (cid:3) 3. The kernel of the map from T ∗ϑ T to Aϑ The kernel of the canonical map from T ∗ϑT to Aϑ mapping u 7→ u and v 7→ v is the ideal h1 − uu∗, 1 − vv∗i generated by the defect projections 1 − uu∗ and 1 − vv∗. It plays an important role in the sequel and it is denoted by J. We prove that it reflects the whole ideal structure of T ∗ϑT, if ϑ is irrational. Also, we give a classification of T ⊗ϑ T and T ∗ϑ T with respect to the parameter ϑ. 12 MORITZ WEBER Since Aϑ is simple for irrational parameters ϑ, we know that J is a maximal ideal in T ∗ϑ T. Even more, we can show that it contains all non-trivial ideals in T ∗ϑ T. We need a technical lemma to prove this statement. Lemma 3.1. Let ϑ be irrational, I ⊳ T ∗ϑ T be an ideal in T ∗ϑ T, ε > 0 and let 1 = w + y + z be a decomposition of the unit such that • w is in the linear span of all elements x(1 − vv∗)x′, where x and x′ are ∗-monomials in u and v, • y ∈ I, • z ∈ T ∗ϑ T with kzk < ε. Then there exists a y′ ∈ I, such that ky′ − 1k < ε. We obtain a similar result, if w is in the linear span of all elements x(1 − uu∗)x′. In fact, this span is dense in the ideal h1 − uu∗i ⊳ T ∗ϑ T. Proof. Put s := uN vN , where N ∈ N is sufficiently large such that s∗w = 0. Put y′ := s∗ys. Then 1 = s∗(w + y + z)s = y′ + s∗zs and thus ky′ − 1k ≤ kzk < ε. (cid:3) Proposition 3.2. If ϑ is irrational, then the ideal J := h1 − uu∗, 1 − vv∗i contains all non-trivial ideals in T ∗ϑ T, i.e. for any ideal T ∗ϑ T 6= I ⊳ T ∗ϑ T we have I ⊆ J. Proof. Let I ⊳ T ∗ϑ T be a non-zero ideal. Since Aϑ is simple, we either have I ⊆ J or I + J = T ∗ϑ T. In the latter case, we have 1 ∈ I + J = I ′ + h1 − vv∗i for I ′ := I + h1 − uu∗i ⊳ T ∗ϑ T. Hence, there is a x ∈ h1 − vv∗i and a y ∈ I ′ such that 1 = x + y. For m ∈ N and εm := 1 m , there are wm and zm in T ∗ϑ T such that 1 = wm + y + zm is a decomposition of the unit in the sense of the preceding lemma, where x = wm + zm. We obtain a sequence of elements y′ m → 1 for m → ∞. Applying the lemma again to 1 ∈ I ′ = I + h1 − uu∗i yields a sequence y′′ m ∈ I such that y′′ m → 1 for m → ∞. We conclude that I = T ∗ϑ T whenever I 6⊆ J. (cid:3) m ∈ I ′ such that y′ The proof of Proposition 3.2 is adapted from the proof for a simpler case, namely for T ⊗ϑ T given by Proskurin ([Pr00a, prop. 6], [Pr00b]). In T ⊗ϑ T, the ideal h1 − uu∗, 1 − vv∗i has a simple form: it is the closed linear span of elements of the form: uavb(1 − uu∗)ε1(1 − vv∗)ε2(v∗)c(u∗)d Here a, b, c, d ∈ N0, ε1, ε2 ∈ {0, 1} and ε1 + ε2 6= 0. In T ∗ϑ T however, the ideal h1 − uu∗, 1 − vv∗i is more complicated. As a corollary from Proposition 3.2, we get that the ideal generated by the two defect projections is the largest ideal in any C ∗-algebra, which is generated by two isometries u and v with uv = λvu, whenever ϑ ∈ R\Q. Corollary 3.3. Let A be a unital, non-zero C ∗-algebra and let u, v ∈ A be two isometries with uv = λvu, where λ = e2πiϑ, ϑ ∈ R\Q. Then h1 − uu∗, 1 − vv∗i ⊳ C ∗(u, v) ⊆ A is the union of all non-trivial ideals in the C ∗-subalgebra generated by u and v. ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 13 Proof. Let ϕ : T ∗ϑ T ։ C ∗(u, v) =: B ⊆ A be the homomorphism with u 7→ u, v 7→ v. Then ϕ(J) = J ′ := h1 − uu∗, 1 − vv∗i ⊳ B. Let B 6= I ′ ⊳ B be an ideal in B. Its preimage I := ϕ−1(I ′) ⊳ T ∗ϑ T is an ideal in T ∗ϑ T, thus I ⊆ J and hence I ′ = ϕ(I) ⊆ ϕ(J) = J ′. (cid:3) Corollary 3.4. Let ϑ be irrational. The ideal J ′ = h1 − uu∗, 1 − vv∗i is the union of all non-trivial ideals in T ⊗ϑ T. The same holds true for h1 − vv∗i in T ⋊ϑ Z. From the fact that J ⊳ T ∗ϑ T is the largest ideal in T ∗ϑ T, we may infer their classification with respect to the parameter ϑ (for irrational ϑ). The following proof is mainly taken from [JPS05]. Proposition 3.5 ([JPS05, prop. 2]). For irrational parameter ϑ, there is a unique, normalized trace on T ∗ϑ T as well as on T ⊗ϑ T, induced by the trace on Aϑ. Proof. On the irrational rotation algebra Aϑ there is a unique, normalized trace τ : Aϑ → C. Thus, τ := τ ◦ ϕ is a normalized trace on T ∗ϑT, where ϕ : T ∗ϑT → Aϑ is the canonical homomorphism. Let σ be another normalized trace on T ∗ϑ T. Then σ(J) = 0, because σ(x(1 − uu∗)x′)2 = σ((1 − uu∗)x′x)2 ≤ σ(1 − uu∗)σ(x∗x′∗x′x) = 0 for all monomials x, x′ in T ∗ϑT by Cauchy-Schwarz; likewise σ(x(1 − vv∗)x′)2 = 0. Then σ : Aϑ → C is a trace on Aϑ, given by σ(b) := σ(a) for an a ∈ T ∗ϑT such that ϕ(a) = b. Therefore we have σ = τ , which implies σ = σ ◦ ϕ = τ ◦ ϕ = τ . (cid:3) Jørgensen, Proskurin and Samoılenko concluded in their article [JPS05] that the classification of the C ∗-algebras T ∗ϑ T is the same as for the rotation algebras Aϑ. We give a slightly different proof, adapted from Proskurin's ([Pr00a], [Pr00b]) classification of the C ∗-algebras T ⊗ϑ T. Proposition 3.6. For irrational parameters ϑ, the classification of the C ∗algebras T ∗ϑ T depends on ϑ in the same way as it does for the rotation algebras Aϑ: T ∗ϑ T ∼= T ∗µ T ⇐⇒ ϑ = ±µ mod Z The same holds true for T ⊗ϑ T and T ⋊ϑ Z. Proof. For ϑ = ±µ mod Z we have either T∗ϑT = T∗µT or T∗ϑT ∼= T∗−ϑT = T∗µT ∼=→ T ∗µ T be an isomorphism via u ↔ v, v ↔ u. For the converse, let α : T ∗ϑ T between T ∗ϑ T and T ∗µ T, and let ϕ : T ∗ϑ T ։ Aϑ and ψ : T ∗µ T ։ Aµ be the canonical surjections onto the rotation algebras. Then α(ker(ϕ)) = ker(ψ) by Proposition 3.2. Thus, α induces an isomorphism from Aϑ to Aµ. Use Corollary 3.4 for the cases of T ⊗ϑ T and T ⋊ϑ Z. (cid:3) 14 MORITZ WEBER 4. A decomposition of the isometries in T ∗ϑ T This section is the heart of this article, since we will develop the main tools for the investigation of T ∗ϑ T. We first present a Wold decomposition of the product uv of the isometries u and v in T ∗ϑ T on some Hilbert space H. This gives rise to a subspace H0 of H, on which uv is a unitary. It turns out that u and v are unitaries on H0, too. A study of u and v on the orthogonal complement of H0 reveals essential parts of their structure. This yields an embedding ι of T ∗ϑT into (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2) from which we may obtain a lot of information in T ∗ϑT. Our approach is adapted from the work of Berger, Coburn and Lebow ([BCL78]), who investigated the C ∗-algebra generated by two commuting isometries; this is the case T ∗ϑ T with ϑ = 0. Using this embedding ι, we show that the ideal J = h1 − uu∗, 1 − vv∗i ⊳ T ∗ϑ T is isomorphic to (C(S1) ∗C C2) ⊗ K. This will be of crucial use for the computation of the K-theory of T ∗ϑT. Also, this proves that T ∗ϑT is not exact, because C(S1) ∗C C2 is not. Finally, we apply the machinery of this section to the C ∗-algebra T ⊗ϑ T. The well-known Wold decomposition states that every isometry v on a Hilbert space H is of the form v = vu ⊕ vs for a decomposition H = Hu ⊕ Hs of the Hilbert space. The operator vu is a unitary and vs is an (amplified) copy of the unilateral shift. Hence, the unilateral shift is "the" isometry and the Wold decomposition reveals the unitary and the shift part of an isometry. In their article from 1978 ([BCL78]), Berger, Coburn and Lebow investigated the representation and index theory for C ∗-algebras generated by commuting isome- tries. They noticed that the Wold decomposition of the product of the commuting isometries yields a subspace for the unitary part, on which the single isometries are unitaries as well. Thereby, they managed to represent the commuting isometries as tensor products of some operators. This approach may be slightly modified to the more general case of twisted commuting isometries -- hence to the C ∗-algebra T ∗ϑT. The point is, that the twist of the multiplication does not affect the relevant subspaces. Let (en)n∈N0 be an orthonormal basis of ℓ2(N0), S ∈ L(ℓ2(N0)) be the unilateral shift and d(λ) ∈ L(ℓ2(N0)) be the rotation operator, given by d(λ)en = λnen for n ∈ N0 and λ ∈ C of absolute value one. Proposition 4.1. Let π : T ∗ϑT → L(H) be a unital representation of T ∗ϑT and let • p := 1 − π(uvv∗u∗) = 1 − π(vuu∗v∗) ∈ L(H) be the defect projection of the isometry π(uv), • H0 be the set of all vectors ξ ∈ H such that for all n > 0 there exists a ξn ∈ H such that ξ = π(uv)nξn, • and let K be the closed linear span of all π(uv)nη, where n ≥ 0 and η ∈ pH. Then the following holds: ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 15 (i) H0 and K are closed linear subspaces of H with H ⊥ 0 = K, and there is a Hilbert space isomorphism K ∼= ℓ2(N0) ⊗ pH via π(uv)nη ↔ en ⊗ η. (ii) The restriction π(uv)H0 ∈ L(H0) is a unitary, and π(uv)K the Hilbert space isomorphism of (i). ∼= S ⊗ 1 under (iii) The restrictions π(u)H0 and π(v)H0 are unitaries. (iv) Put ¯u := u(1 − vv∗) + (1 − uu∗)v∗ and ¯p := v(1 − uu∗)v∗ ∈ T ∗ϑ T. Then ¯u′ := π(¯u) is a unitary on pH and ¯p′ := π(vv∗)p = pπ(vv∗) = π(¯p) ∈ L(pH) is a projection. (v) The restrictions π(u)K and π(v)K are operators on K and they are of the following form, using the Hilbert space isomorphism of (i): π(u)K π(v)K ∼= Sd(λ) ⊗ ¯u′ ¯p′ + d(λ) ⊗ ¯u′(1 − ¯p′) ∼= d(λ)∗S ⊗ (1 − ¯p′)¯u′∗ + d(λ)∗ ⊗ ¯p′ ¯u′∗ Proof. To simplify the notation, we write w := π(uv). (i) To see that H0 is closed, check that (1 − wn(w∗)n)ξ = 0 for all n > 0, whenever ξ ∈ H is a limit of a sequence (ξk)k∈N ⊆ H0. Use pH ⊥ wH to show H0 ⊆ K ⊥, and show inductively ξ = wn((w∗)nξ) for n > 0 and ξ ∈ K ⊥ to obtain H0 ⊇ K ⊥ (use (p(w∗)nξ p(w∗)nξ) = 0 for all n ≥ 0 to show wnww∗(w∗)nξ = wn(w∗)nξ = ξ). If (ηi)i∈I forms an orthonormal basis of pH for some index set I, the ensemble (wnηi)n∈N0,i∈I forms an orthonormal basis of K as well as the elements en ⊗ ηi for ℓ2(N0) ⊗ pH. This yields the Hilbert space isomorphism K ∼= ℓ2(N0) ⊗ pH. (ii) We have wH0 = H0 and hence the isometry wH0 is surjective. Using the Hilbert space isomorphism of (i), we compute: w(en ⊗ η) ↔ w(wnη) = wn+1η ↔ en+1 ⊗ η (iii) Since π(u) and w commute up to a scalar, we get π(u)H0 ⊆ H0. Furthermore, π(u)H0 = H0, because we can write ξ ∈ H0 as ξ = π(u)π(v)ξ1 for some ξ1 ∈ H. Then π(v)ξ1 is in H0 as π(v)ξ1 = wn(λ−nπ(v)ξn+1). (iv) A simple algebraic calculation shows ¯u¯u∗ = ¯u∗¯u = 1 − uvv∗u∗. (v) Since π(u)H0 = H0, we get that π(u)H ⊥ From ¯u¯p = (1 − uu∗)v∗ we conclude ¯u′ ¯p′η = π(v∗)pη = π(v∗)η for any η ∈ pH. We use this and ¯u(1 − ¯p) = ¯u − ¯u¯p = u(1 − vv∗) for the following computation, where n ∈ N0 and η ∈ pH. Under the Hilbert space isomorphism of (i), we have: 0 ; likewise π(v)H ⊥ 0 ⊆ H ⊥ 0 . 0 ⊆ H ⊥ π(u)(en ⊗ η) ↔ π(u)wnη = λnwnπ(u)η = λn(wnπ(u)π(vv∗)η + wnπ(u)(1 − π(vv∗))η) = λn(wn+1π(v∗)η + wnπ(u(1 − vv∗))η) = λn(wn+1¯u′ ¯p′η + wn¯u′(1 − ¯p′)η) ↔ (Sd(λ) ⊗ ¯u′ ¯p′ + d(λ) ⊗ ¯u′(1 − ¯p′))(en ⊗ η) Note, that ¯u′ and ¯p′ are operators on pH, thus ¯u′ ¯p′η, ¯u′(1 − ¯p′)η ∈ pH for η ∈ pH. 16 MORITZ WEBER Analogously, we use ¯p¯u∗ = v(1 − uu∗) and (1 − ¯p′)¯u′∗η = π(u∗)η for any η ∈ pH. This yields: π(v)(en ⊗ η) ↔ π(v)wnη = λ−n(wnπ(v)π(uu∗)η + wnπ(v)(1 − π(uu∗))η) = λ−n(¯λwn+1π(u∗)η + wnπ(v(1 − uu∗))η) = λ−n(¯λwn+1(1 − ¯p′)¯u′∗η + wn ¯p′¯u′∗η) ↔ (¯λSd(λ)∗ ⊗ (1 − ¯p′)¯u′∗ + d(λ)∗ ⊗ ¯p′¯u′∗)(en ⊗ η) Using ¯λSd(λ)∗ = d(λ)∗S, we get the stated result. (cid:3) From Proposition 4.1, we deduce the existence of an embedding ι of T ∗ϑ T into (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2). The unital free product of C(S1) with C2 can be described in several ways. We denote by Z2 the quotient of Z by 2Z. Lemma 4.2. The following C ∗-algebras are isomorphic: (i) The unital free product C(S1) ∗C C2 of C(S1) and C2. (ii) The universal C ∗-algebra generated by a unitary ¯u and a projection ¯p. (iii) The (full) group C ∗-algebra C ∗(Z ∗ Z2). (iv) The crossed product C ∗(F2) ⋊ Z2 under the action that swaps the generators. Proof. C2 is the universal unital C ∗-algebra generated by a projection ¯p under the identification (cid:18) 1 0 (cid:19) ↔ ¯p. As C(S1) can be viewed as the universal C ∗-algebra Furthermore, C ∗(Z) ∼= C(S1) and C ∗(Z2) ∼= C2. Hence the (full) group C ∗-algebra generated by a unitary ¯u, this yields the isomorphism of (i) and (ii). of Z ∗ Z2 may be written as C ∗(Z ∗ Z2) ∼= C ∗(Z) ∗C C ∗(Z2) ∼= C(S1) ∗C C2. From an article by Murphy in 1996 ([M96, proof of th. 6.2]), we get the isomor- phism of (ii) and (iv). The group C ∗-algebra C ∗(F2) may be seen as the universal C ∗-algebra generated by two unitaries a and b without any further relations. The automorphism that swaps a and b is of order two, hence we can form the crossed product with Z2. We write C ∗(F2) ⋊ Z2 as the universal C ∗-algebra generated by two unitaries a and b together with a symmetry z (i.e. a unitary z with z2 = 1 or equivalent z = z∗) such that a = zbz and b = zaz. Since b may be built out of a and z, this is just the universal C ∗-algebra generated by a unitary a and a symmetry z. The homomorphism from C ∗(¯u, ¯p) to C ∗(F2)⋊Z2, mapping ¯u 7→ a and ¯p 7→ (z +1)/2 is inverse to the one in the converse direction, mapping a 7→ ¯u and z 7→ 2¯p − 1. (cid:3) From now on, we will view C(S1) ∗C C2 as the universal C ∗-algebra generated by a unitary ¯u and a projection ¯p. For the next theorem, note that the generators of T ∗ϑ T are denoted by u and v as well as those of T ⋊ϑ Z (recall Remark 1.4). ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 17 Theorem 4.3. There (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2) of the following form. is an embedding of T ∗ϑ T into the C ∗-algebra ι : T ∗ϑ T ֒→ (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2) u 7→ vu ⊗ ¯u¯p + u ⊗ ¯u(1 − ¯p) v 7→ u∗v ⊗ (1 − ¯p)¯u∗ + u∗ ⊗ ¯p¯u∗ Proof. The homomorphism ι exists by the universal property of T ∗ϑT. Further- more, we have a homomorphism ι0 : T ∗ϑ T → Aϑ, mapping u 7→ u and v 7→ v. The direct sum ι0 ⊕ ι is injective. Indeed, let π : T ∗ϑT ֒→ L(H) be a unital, faithful rep- resentation of T ∗ϑT. By Proposition 4.1 we can decompose H = H0 ⊕(ℓ2(N0) ⊗ pH) such that: π(u) = π(u)H0 ⊕ [Sd(λ) ⊗ ¯u′ ¯p′ + d(λ) ⊗ ¯u′(1 − ¯p′)] π(v) = π(v)H0 ⊕ [d(λ)∗S ⊗ (1 − ¯p′)¯u′∗ + d(λ)∗ ⊗ ¯p′¯u′∗] Thus, we have two homomorphism σ0 : Aϑ → L(H0), u 7→ π(u)H0 and v 7→ π(v)H0, and σ : (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2) → L(ℓ2(N0) ⊗ pH), u ⊗ 1 7→ d(λ) ⊗ 1, v ⊗ 1 7→ S ⊗ 1, 1 ⊗ ¯u 7→ 1 ⊗ ¯u′ and 1 ⊗ ¯p 7→ 1 ⊗ ¯p′. Hence π = (σ0 ⊕ σ) ◦ (ι0 ⊕ ι), and ι0 ⊕ ι is injective. Consider now the homomorphism τ : (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2) → Aϑ, given by u ⊗ 1 7→ u, v ⊗ 1 7→ uv, 1 ⊗ ¯u 7→ 1, 1 ⊗ ¯p 7→ 0. Then ι0 = τ ◦ ι. Thus, ι(x) = 0 implies ι0(x) ⊕ ι(x) = 0, which yields x = 0. Therefore ι is injective and the proof is complete. If ϑ is irrational, we can simplify the proof using Proposition 3.2. The restriction πH0 : T ∗ϑ T → L(H0) is isomorphic to the canonical map T ∗ϑ T → Aϑ, since C ∗(πH0(u), πH0(v)) is isomorphic to Aϑ (Aϑ is simple). Thus, the kernel of the is contained in the kernel of the map πH0 by Proposition 3.2. Hence restriction πH ⊥ the restriction πH ⊥ (cid:3) = σ ◦ ι is injective, and therefore also ι. 0 0 Remark 4.4. Direct computations show that we have the following assignment of elements (for i ∈ N0) under the map ι : T ∗ϑ T → (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2). u 7→ vu ⊗ ¯u¯p + u ⊗ ¯u(1 − ¯p) v 7→ u∗v ⊗ (1 − ¯p)¯u∗ + u∗ ⊗ ¯p¯u∗ uv 7→ v ⊗ 1 1 − uu∗ 7→ (1 − vv∗) ⊗ ¯u¯p¯u∗ 1 − vv∗ 7→ (1 − vv∗) ⊗ (1 − ¯p) (uv)i¯u 7→ viu(1 − vv∗) ⊗ ¯u (uv)i¯u∗ 7→ viu∗(1 − vv∗) ⊗ ¯u∗ (uv)i ¯p 7→ vi(1 − vv∗) ⊗ ¯p 18 MORITZ WEBER Here, ¯u := u(1 − vv∗) + (1 − uu∗)v∗ and ¯p := v(1 − uu∗)v∗ in T ∗ϑ T are defined as in Proposition 4.1(iv). From this embedding ι, we immediately see that the range projections uu∗ and vv∗ in T ∗ϑ T do not commute (cf. Remark 1.5). This remark gives us an idea of an explicit picture of the ideal J in T ∗ϑ T. Note ¯p in J are mapped under ι to u(1 − vv∗) ⊗ ¯u resp. that the elements ¯u resp. (1 − vv∗) ⊗ ¯p. This reveals the structure of C(S1) ∗C C2 in the second component, if we "untwist" the multiplication with the unitary u in the image of ¯u. In the first component we have a copy of the algebra of compact operators, if we consider elements (uv)i ¯p(uv)∗j. This yields J ∼= (C(S1) ∗C C2) ⊗ K, which will be verified in the next theorem. Recall that we see the compact operators as the universal C ∗-algebra generated by elements xi, i ∈ N0 such that x∗ Theorem 4.5. The defect ideal J = h1 − uu∗, 1 − vv∗i ⊳ T ∗ϑ T is isomorphic to (C(S1) ∗C C2) ⊗ K via (uv)i¯u ↔ ¯u ⊗ xi and (uv)i ¯p ↔ ¯p ⊗ xi for i ∈ N0. i xj = δijx0. Proof. Like in Proposition 2.3, we first use a twisted version of (C(S1) ∗C C2) ⊗ K and show that it is isomorphic to ι(J) ∼= J. We then apply Lemma 2.2 to get the isomorphism to (C(S1) ∗C C2) ⊗ K. Let αi defined by αi(¯u) := λi ¯u and αi(¯p) := ¯p be automorphisms on C(S1)∗CC2 for i ∈ N0 and form the twisted tensor product with the compacts, (C(S1) ∗C C2) ⊗α K as defined in Definition 2.1. Hence (C(S1) ∗C C2) ⊗α K is an ideal in the universal C ∗-algebra Bα which in turn is generated by a unitary ¯u, a projection ¯p, elements xi for i ∈ N0, and the relations x∗ i xj = δijx0, ¯uxi = λixi ¯u, and ¯pxi = xi ¯p. By the universal property, we obtain the following homomorphism: ϕ : Bα → (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2) ¯u 7→ u ⊗ ¯u, ¯p 7→ 1 ⊗ ¯p, xi 7→ vi(1 − vv∗) ⊗ 1 The ideal (C(S1) ∗C C2) ⊗α K is spanned by all elements ρxix∗ j , where ρ is a ∗- monomial in ¯u and ¯p, and i, j ∈ N0. They are mapped by ϕ to ukvi(1 − vv∗)(v∗)j ⊗ ρ with σ(ρ) = uk, for some k ∈ Z. Here, σ : C(S1) ∗C C2 → C(S1) is given by ¯u 7→ u, ¯p 7→ 1, and we use the canonical embedding of C(S1) ∗C C2 into Bα. The image of J = h1−uu∗, 1−vv∗i⊳T∗ϑT under ι is exactly the closed linear span of these elements ukvi(1−vv∗)(v∗)j ⊗ρ ∈ (T⋊ϑZ)⊗(C(S1)∗CC2). Indeed, let ρ′ ∈ J be the ∗-monomial in ¯u and ¯p ∈ T ∗ϑ T analogous to ρ ∈ C(S1) ∗C C2. Then ι((uv)iρ′((uv)∗)j) equals ukvi(1−vv∗)(v∗)j ⊗ρ up to a scalar, because ι(¯u) = u(1−vv∗)⊗ ¯u, ι(¯p) = (1−vv∗)⊗ ¯p and ι(uv) = v ⊗ 1 (cf. Remark 4.4). On the other hand, the closed linear span by the elements ukvi(1 − vv∗)(v∗)j ⊗ ρ is an ideal in ι(T ∗ϑT) containing ι(1 − uu∗) and ι(1 − vv∗). We conclude that the image of (C(S1) ∗C C2) ⊗α K under ϕ is exactly ι(J). It remains to show, that the restriction of ϕ to (C(S1) ∗C C2) ⊗α K is injective. For this, let π : Bα ֒→ L(H) be a unital, faithful representation of Bα and decompose ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 19 the Hilbert space H = (cid:0)Li∈N0 π(xix∗ H. There is a Hilbert space isomorphism Li∈N0 π(xix∗ i )H(cid:1) ⊕ K, where K ⊆ H is a subspace of i )H ∼= ℓ2(N0) ⊗ H0, where H0 := π(x0)H, via π(xi)η ↔ ei ⊗ η (see the proof of Proposition 2.3). We now compute the form of the elements of Bα on the space K ⊥. Note, that π(¯u)H0 = H0, since ¯ux0 = x0 ¯u and ¯u is a unitary. Hence u := π(¯u)H0 ∈ L(H0) is a unitary on H0. Also p := π(¯p)H0 ∈ L(H0) acts on H0, because ¯px0 = x0 ¯p. On K ⊥, the operators have the following from, using the Hilbert space isomorphism K ⊥ ∼= ℓ2(N0) ⊗ H0. π(¯u)K ⊥(ei ⊗ η) ↔ π(¯uxi)η = λiπ(xi)π(¯u)η ↔ (d(λ) ⊗ u)(ei ⊗ η) π(¯p)K ⊥(ei ⊗ η) ↔ π(¯pxi)η = π(xi)π(¯p)η π(xj)K ⊥(ei ⊗ η) ↔ π(xjxi)η = δi0π(xj)η ↔ (1 ⊗ p)(ei ⊗ η) ↔ (Sj(1 − SS∗) ⊗ 1)(ei ⊗ η) 0xi = δi0xj. Because of π(xj)K = 0 for all j ∈ N0 Here, we used xjxi = xjx∗ and as (C(S1) ∗C C2) ⊗α K is spanned by elements ρxix∗ j , the restriction of π to (C(S1) ∗C C2) ⊗α K is a unital, faithful map to L(K ⊥) mapping ¯u 7→ d(λ) ⊗ u, ¯p 7→ 1 ⊗ p, and xj 7→ (Sj(1 − SS∗)) ⊗ 1. On the space K ⊥ ∼= ℓ2(N0) ⊗H0, the operators d(λ) ⊗1, S ⊗1, 1 ⊗ u and 1 ⊗ p give rise to a representation of (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2). Thus we get a homomorphism: τ : (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2) → L(K ⊥) u ⊗ 1 7→ d(λ) ⊗ 1, 1 ⊗ ¯u 7→ 1 ⊗ u, v ⊗ 1 7→ S ⊗ 1, 1 ⊗ ¯p 7→ 1 ⊗ p We conclude that πK ⊥ = τ ◦ ϕ on (C(S1) ∗C C2) ⊗α K, which proves injectivity for ϕ on (C(S1) ∗C C2) ⊗α K. Therefore, (C(S1) ∗C C2) ⊗α K is isomorphic to ι(J) ∼= J and we may apply Lemma 2.2 to complete the proof. (cid:3) In the case ϑ = 0, Murphy ([M96, proof of th. 6.2]) obtained the same picture for J ⊳ T ∗ϑ T by similar means. But the case of general ϑ remained untouched. This picture of J enables us to compute its K-theory, since the K-groups of free products of C ∗-algebras are completely understood by the work of Cuntz ([Cu82]). We then may derive the K-groups of T ∗ϑ T for all ϑ. This is done in section 6. Furthermore, Theorem 4.5 reveals that T ∗ϑ T is not exact, since C(S1) ∗C C2 is not. The latter may be deduced from Lemma 4.2. Proposition 4.6. The C ∗-algebra C(S1) ∗C C2 is not exact. Proof. By Lemma 4.2, C(S1)∗CC2 is isomorphic to C ∗(F2)⋊Z2. Hence it contains C ∗(F2) as a C ∗-subalgebra. Since C ∗(F2) is not exact due to Simon Wassermann (see [Wa90] or [Wa78] and [Wa76]) and since a C ∗-subalgebra of an exact C ∗-algebra is exact again (a C ∗-algebra is exact if and only if it may be embedded into O2, [KP00]), we conclude that C(S1) ∗C C2 cannot be exact. (cid:3) Theorem 4.7. T ∗ϑ T is not exact. 20 MORITZ WEBER Proof. We have C(S1) ∗C C2 ⊆ (C(S1) ∗C C2) ⊗ K ∼= J ⊆ T ∗ϑ T. (cid:3) Remark 4.8. The following is an overview on some further properties of T ∗ϑ T. (a) As C ∗(F2) embeds into T∗ϑT and as the first C ∗-algebra is not locally reflexive, nor is the second. (cf. [BO08, Cor. 9.1.6 and Lem. 9.2.8]). ⊳ 6= 1 and I ′ (b) Murphy ([M03, Th. 3.3]) showed that the (full) C ∗-algebra of the free prod- uct of a non-trivial, countable, free group F and a non-trivial, countable, amenable group Z is primitive. Hence C(S1) ∗C C2 is primitive by Lemma 4.2. Now let I1, I2 T ∗ϑ T be two non-zero ideals in T ∗ϑ T. For irrational parameter ϑ, we have I1, I2 ⊆ J by Proposition 3.2. Using the isomorphism of Theorem 4.5, we see that these ideals are isomorphic to I ′ 2 ⊗ K 2 in C(S1) ∗C C2. As C(S1) ∗C C2 is not for some non-zero ideals I ′ primitive, I ′ 2 have non-zero intersection, thus the same holds for I1 and I2. We infer that T ∗ϑ T is primitive for irrational ϑ. 1 ⊗ K and I ′ 1 and I ′ (c) The C ∗-algebra C(S1) ∗C C2 is not only a C ∗-subalgebra of T ∗ϑ T but also a hereditary C ∗-subalgebra. Under the isomorphism of Theorem 4.5, we have that C(S1) ∗C C2 is isomorphic to the C ∗-subalgebra of T ∗ϑT generated by the elements ¯u := u(1 − vv∗) + (1 − uu∗)v∗ and ¯p := v(1 − uu∗)v∗ (cf. Remark 4.4 and Theorem 4.5 and using the embedding of C(S1) ∗C C2 into C(S1) ∗C C2 ⊗ K by x 7→ x ⊗ x0). This in turn coincides with the compression pT ∗ϑ Tp of T ∗ϑ T by the projection p := 1 − uvv∗u∗ (which is the unit of C ∗(¯u, ¯p)). Indeed, the projection p corresponds to the minimal projection 1 ⊗ x0 in (C(S1) ∗C C2) ⊗ K, thus: C(S1) ∗C C2 ∼= (1 ⊗ x0)((C(S1) ∗C C2) ⊗ K)(1 ⊗ x0) ∼= pJp = pT ∗ϑ Tp Remark 4.9. The techniques developed in this section may also be applied to T ⊗ϑ T, although this C ∗-algebra is much less complicated. Nevertheless, let us quickly review this case. (a) In analogy to Proposition 4.1, we consider the Wold decomposition of the isometry π(uv), if π is a unital representation of T ⊗ϑ T. In this case, the unitaries ¯u′ and ¯p′ in L(pH) fulfill the relations ¯p′¯u′ = ¯p′¯u′ ¯p′ (with the same notations as in Prop. 4.1). Accordingly, we obtain an embedding ι′ : T ⊗ϑ T ֒→ (T ⋊ϑ Z) ⊗ L. Here, L is the quotient of C(S1) ∗C C2 by the ideal generated by ¯p¯u − ¯p¯u¯p. We deduce that the defect ideal h1 − uu∗, 1 − vv∗i ⊳ T ⊗ϑ T is isomorphic to L ⊗ K. (b) The C ∗-algebra L is much smaller than C(S1) ∗C C2. embedding ψ of L into M2(T) = M2(C) ⊗ T, by In fact, there is an ¯u 7→ (cid:18) v∗ 1 − vv∗ v(cid:19) 0 and ¯p 7→ (cid:18)1 0 0 0(cid:19) . ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 21 (c) The following sequence is exact, hence L is nuclear. 0 → K → L → C(S1) ⊕ C(S1) → 0 Tensoring this sequence with the compacts K (and using item (a)) yields a sequence which Kabluchko proved in 2001 ([Ka01]) to be exact: 0 → K → J ′ → K ⊗ (C(S1) ⊕ C(S1)) → 0 Here J ′ = h1 − uu∗, 1 − vv∗i ⊳ T ⊗ϑ T. For this, he showed that the ideals h1 − uu∗i and h1 − vv∗i ⊳ T ⊗ϑ T are isomorphic to T ⊗ K (cf. Proposition 2.3 of our article) and h1−uu∗i∩h1−vv∗i = K whereas h1−uu∗i∪h1−vv∗i = J ′. Hence J ′/K ∼= (h1 − uu∗i/K) ⊕ (h1 − vv∗i/K) = (C(S1) ⊗ K) ⊕ (C(S1) ⊗ K). Proof. (b) The projections q0 := ¯p − ¯u∗ ¯p¯u and qn := ¯unq0 ¯u−n for n ∈ Z are mutually orthogonal. Furthermore, ¯uqn = qn+1¯u for all n ∈ Z. Thus π(¯u) acts as a bilateral shift on K0 := Ln∈Z π(qn)H ⊆ H, if π : L → L(H) is a representation of L, whereas πK0(¯p) projects onto q−H := Pn≤0 π(qn)H. This yields a representation of M2(T) on L(K0) which proves that ψ is injective, if π is faithful. (c) Consider the homomorphism σ : L → C(S1) ⊕ C(S1), mapping ¯u 7→ (u, u) and ¯p 7→ (1, 0). Its kernel is the ideal h¯u¯p − ¯p¯ui = hq0i, where q0 is the projection given by q0 := ¯p − ¯u∗ ¯p¯u ∈ L. This ideal is spanned by matrix units ¯ekl := ¯ukq0(¯u∗)l for k, l ∈ Z, hence it is isomorphic to the compacts K. (cid:3) 5. The ideal structure of T ∗ϑ T For irrational parameters ϑ, a combination of the results of Proposition 3.2 and Theorem 4.5 gives a good description of the ideal structure of T ∗ϑ T. Since every nontrivial ideal I in T ∗ϑ T is contained in the ideal J = h1 − uu∗, 1 − vv∗i which in turn is isomorphic to (C(S1) ∗C C2) ⊗ K, we conclude that we have a one-to-one correspondence of ideals in T ∗ϑT with ideals in C(S1) ∗C C2, if ϑ is irrational. (Note that every ideal in A ⊗ K is of the form I ′ ⊗ K, if A is unital.) We investigate this correspondence in an explicit way. Furthermore, we prove that not only T ∗ϑT is not exact, but neither is its quotient by the ideal generated by the commutator [uu∗, vv∗]. Recall that the C ∗-subalgebra generated by ¯u := u(1 − vv∗) + (1 − uu∗)v∗ and ¯p := v(1 − uu∗)v∗ in T ∗ϑ T is isomorphic to C(S1) ∗C C2 (see Remark 4.8(c)). Lemma 5.1. Let ϑ be irrational. Let I ⊳ T ∗ϑT and I ′ ⊳ C(S1) ∗C C2 be ideals, and 6= let ϕ resp. β be the according quotient maps. Let ϕ(I ′) = 0 under the identification I ′ ⊆ C(S1) ∗C C2 ∼= C ∗(¯u, ¯p) ⊆ T ∗ϑ T (cf. Remark 4.8(c)). Furthermore, let ψ : T ∗ϑT/I → (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2/I ′) be a map such that the following diagram 22 MORITZ WEBER commutes (where ι is the embedding of Theorem 4.3). T ∗ϑ T ϕ ↓ T ∗ϑ T/I ι ֒→ ψ −→ (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2) ↓ id ⊗ β (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2/I ′) Then I ∼= I ′ ⊗ K. In other words, I corresponds to I ′ under the isomorphism of J and (C(S1) ∗C C2) ⊗ K. Proof. By Proposition 3.2, I is an ideal in J which in turn is isomorphic to (C(S1)∗C C2)⊗K. Thus, I is isomorphic to I0 ⊗K where I0 is an ideal in C(S1)∗C C2. Furthermore, we can identify I ∩ C ∗(¯u, ¯p) in T ∗ϑ T with I0 in C(S1) ∗C C2. Indeed, under the identification of C ∗(¯u, ¯p) ⊆ T ∗ϑ T and C(S1) ∗C C2 via ¯u ↔ ¯u and ¯p ↔ ¯p (by Remark 4.8(c)), we see that the isomorphism of J and (C(S1) ∗C C2) ⊗ K is of the following form (cf. Theorem 4.5): (uv)ix ←→ x ⊗ xi ∀x ∈ C(S1) ∗C C2 ∼= C ∗(¯u, ¯p) ⊆ T ∗ϑ T, i ∈ N0 Now, an element x ∈ I ∩C ∗(¯u, ¯p) is mapped to x⊗x0 ∈ (I0 ⊗K)∩((C(S1)∗C C2)⊗x0) under the isomorphism of J and (C(S1) ∗C C2) ⊗ K. Thus it is mapped to x in I0, since I0 embeds into I0 ⊗ K by a 7→ a ⊗ x0 (recall that x0 is a minimal projection). To prove I ′ ⊆ I0, recall that ϕ(x) = 0 for all x ∈ I ′ by assumption. Hence x is in I ∩ C ∗(¯u, ¯p) = I0. For the converse direction, consider the representation π : T ⋊ϑ Z → L(ℓ2(N0)) defined by π(u) = d(λ) and π(v) = S. Here d(λ) is given by d(λ)en = λnen for all n ∈ N0 and S is the unilateral shift. Consider now γ : C(S1) ∗C C2 → L(ℓ2(N0)) ⊗(C(S1) ∗C C2/I ′), given by γ := (π ⊗id) ◦ ψ ◦ ϕ. (Again we identify C(S1)∗C C2 = C ∗(¯u, ¯p) ⊆ T ∗ϑT.) Then ¯u is mapped by ψ ◦ϕ = (id ⊗β)◦ι to u(1 − vv∗) ⊗ β(¯u) (cf. Remark 4.4). This is mapped to d(λ)(1 − SS∗) ⊗ β(¯u) by (π⊗id). Since d(λ)(1−SS∗) coincides with 1−SS∗ we infer γ(¯u) = (1−SS∗)⊗β(¯u). Similarly, we see γ(¯p) = (1 − SS∗) ⊗ β(¯p), and we conclude γ(x) = (1 − SS∗) ⊗ β(x) for all x ∈ C(S1) ∗C C2 since (1 − SS∗) ⊗ β is a homomorphism. Finally, let x be in I0 = I ∩ C ∗(¯u, ¯p). Then ϕ(x) = 0 which implies (1 − SS∗) ⊗ β(x) = γ(x) = 0, hence β(x) = 0. Thus I0 ⊆ I ′ and we conclude I0 = I ′. (cid:3) We now give a description of some corresponding ideals in the most interesting cases. A quick look at the C ∗-algebra C(S1) ∗C C2 -- again viewed as the universal C ∗-algebra generated by a unitary ¯u and a projection ¯p -- shows that we have the following canonical maps and the according kernels. First, we have the quotient map to the C ∗-algebra L (as defined in Remark 4.9). Secondly, we can consider the quotient of C(S1) ∗C C2 such that ¯u and ¯p commute. This is isomorphic to the direct sum C(S1) ⊕ C(S1) under the assignment ¯u ↔ (u, u) and ¯p ↔ (1, 0). Thirdly, the quotients where ¯p = 1 or ¯p = 0 seem to be natural. Hence we get the following maps and ideals: • C(S1) ∗C C2 → L, mapping ¯u 7→ ¯u and ¯p 7→ ¯p. The kernel is the ideal generated by ¯p¯u− ¯p¯u¯p. Similarly, we consider the ideal generated by ¯u¯p− ¯p¯u¯p. ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 23 • C(S1) ∗C C2 → C(S1) ⊕ C(S1), mapping ¯u 7→ (u, u) and ¯p 7→ (1, 0). The according ideal is generated by ¯u¯p − ¯p¯u. • C(S1) ∗C C2 → C(S1), mapping ¯u 7→ u and ¯p 7→ 1 resp. ¯u 7→ u and ¯p 7→ 0. The ideals are generated by 1 − ¯p resp. ¯p. These ideals are related in an obvious way by: • h¯pi + h1 − ¯pi = C(S1) ∗C C2 • h¯pi ∩ h1 − ¯pi = h¯u¯p − ¯p¯ui (Write the above map C(S1) ∗C C2 → C(S1) ⊕ C(S1) as the direct sum of the two above maps C(S1) ∗C C2 → C(S1).) • h¯p¯u(1 − ¯p)i + h(1 − ¯p)¯u¯pi = h¯u¯p − ¯p¯ui In the next two lemmas we work out the connection between the relations on u and v in T ∗ϑ T on the one side, and ¯u and ¯p in C(S1) ∗C C2 on the other side. Lemma 5.2. Let A be a unital C ∗-algebra and let u and v be isometries in A such that uv = λvu where λ ∈ C is of absolute value one. Put ¯u := u(1−vv∗)+(1−uu∗)v∗, ¯p := v(1 − uu∗)v∗ and p := 1 − vuu∗v∗ as in Proposition 4.1(iv). Then the following equivalences hold. (a) u is unitary ⇐⇒ ¯p = 0 v is unitary ⇐⇒ p(1 − ¯p) = 0 (Note that p is the unit of C ∗(¯u, ¯p) ⊆ A, cf. Remark 4.8(c).) (b) u∗v = ¯λvu∗ ⇐⇒ (1 − uu∗)v∗u(1 − vv∗) = 0 ⇐⇒ ¯p¯u(1 − ¯p) = 0 (1 − uu∗)(1 − vv∗) = 0 ⇐⇒ (1 − ¯p)¯u¯p = 0 (c) u∗v = ¯λvu∗ and (1 − uu∗)(1 − vv∗) = 0 ⇐⇒ ¯u¯p = ¯p¯u Proof. (a) The second equivalence is due to the equality p(1 − ¯p) = 1 − vv∗. (b) If (1 − uu∗)v∗u(1 − vv∗) = 0, then v∗u = ((uu∗) + (1 − uu∗))v∗u((vv∗) + (1 − vv∗)) = λuv∗ Secondly ¯p¯u(1 − ¯p) = v(1 − uu∗)v∗u(1 − vv∗). Thirdly (1 − ¯p)¯u¯p = (1 − vv∗)(1 − uu∗)v∗. If I is an ideal in C(S1)∗CC2, we have a map from T∗ϑT to (T⋊ϑZ)⊗(C(S1)∗CC2/I), analogous to the embedding ι of T ∗ϑ T into (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2) (cf. Theorem 4.3). In the next lemma, we analyze the structure of the elements u and v under this map. (cid:3) Lemma 5.3. Let I ⊳ C(S1) ∗C C2 be an ideal and let β denote its quotient map. Define elements in (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2/I) by u := vu ⊗ β(¯u¯p) + u ⊗ β(¯u(1 − ¯p)) v := u∗v ⊗ β((1 − ¯p)¯u∗) + u∗ ⊗ β(¯p¯u∗) Then u and v are isometries and uv = λvu. Furthermore: (a) If I = h¯pi, then u is even a unitary. If I = h1 − ¯pi, then v is even a unitary. (b) If I = h¯p¯u(1 − ¯p)i, then u∗v = ¯λvu∗. 24 MORITZ WEBER If I = h(1 − ¯p)¯u¯pi, then (1 − uu∗)(1 − vv∗) = 0. Proof. If I = 0, then u and v are isometries with uv = λvu, hence this also holds in any quotient (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2/I). The remaining statements can be checked by a straight forward computation. (Cf. also Remark 4.4.) (cid:3) Proposition 5.4. For irrational parameter ϑ, there are the following correspon- dences between ideals in T ∗ϑ T and in C(S1) ∗C C2 under the isomorphism of J and (C(S1) ∗C C2) ⊗ K of Theorem 4.5. (a) h1 − uu∗i ⊳ T ∗ϑ T corresponds to h¯pi ⊳ C(S1) ∗C C2. h1 − vv∗i ⊳ T ∗ϑ T corresponds to h1 − ¯pi ⊳ C(S1) ∗C C2. (b) hu∗v − ¯λvu∗i ⊳ T ∗ϑ T corresponds to h¯p¯u(1 − ¯p)i ⊳ C(S1) ∗C C2. h(1 − uu∗)(1 − vv∗)i ⊳ T ∗ϑ T corresponds to h(1 − ¯p)¯u¯pi ⊳ C(S1) ∗C C2. (c) hu∗v − ¯λvu∗, (1−uu∗)(1−vv∗)i ⊳ T ∗ϑT corresponds to h¯u¯p− ¯p¯ui ⊳C(S1)∗C C2. (Recall that a description of the quotient of T ∗ϑ T by this ideal is given in Proposition 2.6.) Proof. (a) Under the quotient map ϕ : T∗ϑT → T∗ϑT/h1−uu∗i, we have ϕ(¯p) = 0 by Lemma 5.2 since ϕ(u) is a unitary. Thus, ϕ(h¯pi) = 0 in the sense of Lemma 5.1 (under the identification of C(S1) ∗C C2 and C ∗(¯u, ¯p) ⊆ T ∗ϑ T). By Lemma 5.3, there is a map ψ : T ∗ϑ T/h1 − uu∗i → (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2/h¯pi), mapping ϕ(u) 7→ u ⊗ β(¯u) and ϕ(v) 7→ u∗v ⊗ β(¯u∗), such that the diagram of of Lemma 5.1 commutes. Thus, h1 − uu∗i = ker ϕ ⊳ T ∗ϑ T corresponds to h¯pi ⊳ C(S1) ∗C C2. The other correspondences are obtained in exactly the same way. We conclude that the correspondence of ideals in T ∗ϑT and in C(S1)∗C C2 matches (cid:3) natural ideal to natural ideals. In Remark 1.5, the commutator [uu∗, vv∗] in T ∗ϑ T has been considered. We are now able to show that the quotient of T ∗ϑT by the ideal generated by it is not exact. In some sense, this shows that T ⊗ϑ T and T ∗ϑ T differ much more than just by the structure of their range projections. This answers a question of section 1. Lemma 5.5. Let A be a unital C ∗-algebra and let u and v be isometries in A such that uv = λvu for a given λ ∈ C of absolute value one. With ¯u ∈ A and ¯p ∈ A as in Lemma 5.2, the following relations are equivalent. [¯p, ¯u¯p¯u∗] = 0 ⇐⇒ [uu∗, vv∗] = 0 In fact, even the equality [¯p, ¯u¯p¯u∗] = [uu∗, vv∗] holds. Proof. A simple algebraic computation shows ¯p¯u¯p¯u∗ = vv∗(1 − uu∗). (cid:3) Lemma 5.6. Consider I = h[¯p, ¯u¯p¯u∗]i⊳C(S1)∗CC2 and let u and v be the isometries in (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2/I ′) defined as in Lemma 5.3. Then [uu∗, vv∗] = 0. Proof. Cf. Remark 4.4. (cid:3) ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 25 Proposition 5.7. For irrational ϑ, the ideal h[uu∗, vv∗]i ⊳ T ∗ϑT corresponds to the ideal h[¯p, ¯u¯p¯u∗]i ⊳ C(S1) ∗C C2 under the isomorphism of J and (C(S1) ∗C C2) ⊗ K. Proof. The proof is analogous to that of Proposition 5.4. Next, we show that a quotient of C(S1) ∗C C2 by h[¯p, ¯u¯p¯u∗]i contains C(S1) ∗C C2 as a C ∗-subalgebra, hence it cannot be exact. This transfers directly to the quotient of T ∗ϑ T by the commutator [uu∗, vv∗]. (cid:3) Proposition 5.8. (a) The quotient D of C(S1) ∗C C2 by h[¯p, ¯u¯p¯u∗]i is not exact. (b) The quotient B of T ∗ϑ T by h[uu∗, vv∗]i is not exact, if ϑ is irrational. 1 0(cid:19) and ¯p′ := (cid:18)¯p 0 Proof. (a) Consider the elements ¯u′ := (cid:18)0 ¯u 0 0(cid:19) in the 2 × 2- matrices M2(C(S1) ∗C C2) over C(S1) ∗C C2. Then ¯u′ is a unitary and ¯p′ is a projec- tion. Furthermore, ¯u′ ¯p′¯u′∗ = (cid:18)0 0 0 ¯p(cid:19), thus [¯p′, ¯u′ ¯p′ ¯u′∗] = 0, which yields a surjective homomorphism from D to C ∗(¯u′, ¯p′) ⊆ M2(C(S1) ∗C C2). as ¯u′2 = (cid:18)¯u 0 0 ¯u(cid:19), the homomorphism C(S1) ∗C C2 → C ∗(¯u′2, ¯p′) ⊆ M2(C(S1) ∗C C2) But the C ∗-subalgebra C ∗(¯u′2, ¯p′) of C ∗(¯u′, ¯p′) is isomorphic to C(S1)∗CC2. Indeed, given by ¯u 7→ ¯u′2, ¯p 7→ ¯p′ may be seen as the map id ⊕ρ. Here, the homomorphism ρ : C(S1) ∗C C2 → C(S1) ∗C C2 is given by ¯u 7→ ¯u and ¯p 7→ 0. Thus C ∗(¯u′, ¯p′) ⊆ M2(C(S1)∗CC2) contains the non-exact C ∗-subalgebra C ∗(¯u′2, ¯p′) and therefore is is not exact. Since quotients of exact C ∗-algebras are exact (cf. [BO08, Cor. 9.4.3]), D cannot be exact. (b) Let I be a nontrivial ideal in T ∗ϑ T and let ϕ be the correponding quotient map. Let I ′ be the ideal in C(S1) ∗C C2 corresponding to I. Consider the following diagram of exact sequences: 0 0 → → I = I → → T ∗ϑ T ▽ J → → T ∗ϑ T/I ▽ ϕ(J) → → 0 0 Compare the lower one with the following exact sequence: 0 → I ′ ⊗ K → (C(S1) ∗C C2) ⊗ K → (C(S1) ∗C C2/I ′) ⊗ K → 0 As I ∼= I ′ ⊗ K and J ∼= (C(S1) ∗C C2) ⊗ K, we may conclude that ϕ(J) is isomorphic to (C(S1) ∗C C2/I ′) ⊗ K. With I = h[uu∗, vv∗]i and I ′ = h[¯p, ¯u¯p¯u∗]i, we see that D is a non-exact C ∗- (cid:3) subalgebra of B. 6. The K-groups of T ∗ϑ T and T ⊗ϑ T For the classification of C ∗-algebras the K-theory is the most important ingredient. This section is devoted to the computation of the K-groups of T ⊗ϑ T and T ∗ϑ T. 26 MORITZ WEBER While the case of T⊗ϑ T is not too difficult to treat, the case of T∗ϑT relies essentially on the isomorphism of the ideal J = h1 − uu∗, 1 − vv∗i with (C(S1) ∗C C2) ⊗ K found in section 4. By this we can compute its K-theory using a general result by Cuntz ([Cu82]) on the K-theory of free products of C ∗-algebras. Consider the following two exact sequences: 0 0 → → I = I → → T ∗ϑ T ▽ J → → T ⋊ϑ Z ▽ h1 − vv∗i → → 0 0 By the lower one, we are able to compute the K-groups of the ideal I = h1 − uu∗i in T ∗ϑT (also using the isomorphism of h1 − vv∗i ⊳ T ⋊ϑ Z and T ⊗ K of Proposition 2.3). We then compute the K-groups of T ∗ϑ T using the following exact sequence: 0 → I → T ∗ϑ T → T ⋊ϑ Z → 0 In constrast, a natural first attempt to compute the K-theory of T ∗ϑ T would try to make use of the following exact sequence: 0 → J → T ∗ϑ T → Aϑ → 0 Unfortunately it does not provide enough information in K-theory. Throughout this section, ϑ ∈ R is arbitrary (either rational or irrational). We first recall the K-groups of Aϑ and T ⋊ϑ Z. Remark 6.1. (a) The rotation algebra Aϑ may be written as the crossed prod- uct C(S1) ⋊ϑ Z of C(S1) with the automorphism u 7→ λu on C(S1), where λ = e2πiϑ. We apply the Pimsner-Voiculescu sequence ([PV80]) to ob- tain K0(Aϑ) = Z2, generated by 1 and a Rieffel projection ([R81]), and K1(Aϑ) = Z2, generated by the classes of the unitaries u and v. (If ϑ = 0, then K0(Aϑ) = Z2 is generated by 1 and the Bott projection, cf. also [Y86] for rational parameters ϑ.) (b) The K-groups of T ⋊ϑ Z are K0(T ⋊ϑ Z) = Z, generated by the unit 1, and K1(T ⋊ϑ Z) = Z, generated by the class of the unitary u. Proof. (b) We compute the K-groups of T ⋊ϑ Z in exactly the same way as in the case of the rotation algebra. We use the Pimsner-Voiculescu sequence with respect to the automorphism on T mapping v 7→ λv on T. It induces the identity on the level of K-theory, hence the 6-term exact sequence splits into two parts: K0(T) ↑ 0−→ K0(T) −→ K1(T ⋊ϑ Z) ←− K1(T) 0←− K0(T ⋊ϑ Z) ↓ K1(T) ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 27 Using K0(T) = Z and K1(T) = 0 we get that K0(T ⋊ϑ Z) = K0(T) = Z and K1(T ⋊ϑ Z) = K0(T) = Z. The generator of K0(T) = Z is the unit 1 and thus also of K0(T ⋊ϑ Z) = Z. For the generator of K1(T ⋊ϑ Z) = Z, we take a quick look at Pimsner and Voiculescu's proof of the 6-term exact sequence for A ⋊α Z, where A is a unital C ∗-algebra and α is an automorphism on A. They define T(A, α) to be the C ∗- subalgebra of (A ⋊α Z) ⊗ T generated by all elements a ⊗ 1 for a ∈ A and by u ⊗ v, where u is the adjoint unitary in A ⋊α Z. Then A ⊗ K is isomorphic to the ideal in T(A, α) generated by 1 ⊗ (1 − vv∗), mapping a ⊗ x0 to a ⊗ (1 − vv∗) (x0 being a minimal projection in K). In consequence, the following sequence is exact: 0 → A ⊗ K → T(A, α) → A ⋊α Z → 0 Now, Pimsner and Voiculescu show that Ki(T(A, α)) is isomorphic to Ki(A) and they conclude their picture of the 6-term exact sequence. The isometry u ⊗ v ∈ T(A, α) is mapped to the unitary u ∈ A ⋊α Z. Thus, the connecting map δ from K1(A ⋊α Z) to K0(A ⊗ K) maps the class of u to the class of the defect projection 1 ⊗ (1 − vv∗) of u ⊗ v, which is isomorphic to 1 ⊗ x0 ∈ A ⊗ K. This in turn is mapped to the unit of A under the isomorphism of K0(A ⊗ K) and K0(A). We conclude that the class of the unitary u in T ⋊ϑ Z is mapped to the generator 1 of K0(T), hence it must generate K1(T ⋊ϑ Z) = Z. (cid:3) The K-groups of T ⊗ϑ T are also obtained by a simple investigation of a 6-term exact sequence. Theorem 6.2. The K-groups of T ⊗ϑ T are K0(T ⊗ϑ T) = Z, generated by the unit 1 and K1(T ⊗ϑ T) = 0. Proof. Consider the following exact sequence: 0 → h1 − uu∗i → T ⊗ϑ T → T ⋊ϑ Z → 0 By Proposition 2.3, we know that the ideal h1−uu∗i⊳T⊗ϑT is isomorphic to T⊗K via uiv(1 − uu∗) 7→ v ⊗ xi for i ∈ N0. The K0-group of T ⊗ K is Z, generated by the class of 1 ⊗ x0, thus K0(h1 − uu∗i) = Z is generated by the class of the defect projection 1 − uu∗. But this is mapped to zero under the map K0(h1 − uu∗i) → K0(T ⊗ϑ T). Indeed, the class [1 − uu∗] ∈ K0(h1 − uu∗i) is mapped to [1 − uu∗] ∈ K0(T ⊗ϑ T). But [uu∗] = [u∗u] = 1 in K0(T ⊗ϑ T) which yields 1 = [1 − uu∗] + [uu∗] = [1 − uu∗] + 1. Thus, the generator of K0(h1 − uu∗i) is mapped to zero in K0(T ⊗ϑ T) and the map K0(h1 − uu∗i) → K0(T ⊗ϑ T) is zero. Furthermore K1(h1 − uu∗i) = K1(T) = 0, hence the 6-term exact sequence in K-theory corresponding to the above short exact sequence falls into the parts K0(T ⊗ϑ T) ∼= K0(T ⋊ϑ Z) = Z (generated by the unit) and: 0 → K1(T ⊗ϑ T) → K1(T ⋊ϑ Z) → K0(h1 − uu∗i) → 0 Since K1(T ⋊ϑ Z) = Z and K0(h1 − uu∗i) = Z, we end up with K1(T ⊗ϑ T) = 0. (cid:3) 28 MORITZ WEBER A priori, the situation in the case of T ∗ϑ T is less clear. The defect ideal h1 − uu∗i ⊳ T ∗ϑ T is more complicated than in the case of T ⊗ϑ T and thus the K-groups cannot be computed in the same straightforward manner. But due to the embedding ι of T ∗ϑ T into (T ⋊ϑ Z) ⊗ (C(S1) ∗C C2), we are able to compute the K-groups of the ideal J = h1 − uu∗, 1 − vv∗i in T ∗ϑ T. Recall that J is isomorphic to (C(S1) ∗C C2) ⊗ K (by Theorem 4.5), therefore we first consider the C ∗-algebra C(S1) ∗C C2. In 1982, Cuntz computed the K-groups of the amalgamated free product of C ∗- algebras ([Cu82]). For the unital free product, we have the following result. Lemma 6.3 ([Cu82]). Let A and B be unital C ∗-algebras and let ψ1 : A → C and ψ2 : B → C be unital homomorphisms. Then K0(A ∗C B) is the quotient of K0(A) ⊕ K0(B) by the subgroup generated by ([1A], −[1B]). Secondly, K1(A ∗C B) = K1(A) ⊕ K1(B). Recall, that we view C(S1)∗CC2 as the universal C ∗-algebra generated by a unitary ¯u (corresponding to the unitary of C(S1)) and a projection ¯p (corresponding to the vector (cid:18) 1 0 (cid:19) ∈ C2, cf. Lemma 4.2). We apply the preceding lemma to C(S1) ∗C C2. Proposition 6.4. The K-groups of C(S1) ∗C C2 are K0(C(S1) ∗C C2) = Z2 generated by [¯p] and [1 − ¯p], and K1(C(S1) ∗C C2) = Z generated by [¯u]. Proof. There are unital homomorphisms ψ1 : C(S1) → C, mapping u 7→ 1, and ψ2 : C2 → C the projection onto the first component. By Cuntz' result, K0(C(S1) ∗C C2) is the quotient of K0(C(S1)) ⊕ K0(C2) = Z ⊕ Z2 by the subgroup generated by (cid:18)1, −(cid:18) 1 1 (cid:19)(cid:19). But Z⊕Z2 is generated by exactly this vector together with (cid:18)0,(cid:18) 1 1 (cid:19)(cid:19). So, we end up with K0(C(S1) ∗C C2) = Z2, generated by (cid:18) 1 1 (cid:19) respectively [¯p] and [1 − ¯p]. Furthermore, K1(C(S1) ∗C C2) = K1(C(S1)) ⊕ K1(C2) = K1(C(S1)) = Z. The (cid:3) 0 (cid:19)(cid:19) and (cid:18)0,(cid:18) 0 0 (cid:19) and (cid:18) 0 generator is the class of u ∈ C(S1) respectively ¯u ∈ C(S1) ∗C C2. This yields the K-groups of J in T ∗ϑ T. Corollary 6.5. The K-groups of the ideal J = h1 − uu∗, 1 − vv∗i in T ∗ϑ T are K0(J) = Z2 generated by the classes [1 − vv∗] and [1 − uu∗], and K1(J) = Z generated by [u(1 − vv∗) + (1 − uu∗)v∗ + uvv∗u∗] = [¯u + (1 − ¯u¯u∗)]. Here ¯u is defined by ¯u = u(1 − vv∗) + (1 − uu∗)v∗ ∈ T ∗ϑ T as in Proposition 4.1. Proof. By Proposition 6.4 we know that K0(C(S1)∗C C2) = Z2 is generated by [¯p] and [1− ¯p], thus K0((C(S1)∗CC2)⊗K) = Z2 is generated by [¯p⊗x0] and [(1− ¯p)⊗x0]. By the isomorphism of Theorem 4.5 we infer that ¯p ⊗ x0 ∈ (C(S1) ∗C C2) ⊗ K corresponds to ¯p = v(1 − uu∗)v∗ ∈ J ⊆ T ∗ϑ T, whereas (1 − ¯p) ⊗ x0 corresponds to ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 29 p − ¯p = 1 − vv∗ ∈ J (recall that p is the unit of C ∗(¯u, ¯p) ⊆ J, see Remark 4.8(c)). Because v(1 − uu∗)v∗ is Murray von Neumann equivalent to 1 − uu∗ via v(1 − uu∗), we end up with [1 − vv∗] and [1 − uu∗] as generators of K0(J) = Z2. Now, K1((C(S1) ∗C C2) ⊗K) = Z is generated by [¯u ⊗x0 + (1 ^(C(S 1)∗CC2)⊗K −1 ⊗x0)], where ^(C(S1) ∗C C2) ⊗ K is the unitalization of (C(S1) ∗C C2) ⊗ K. This is, since for any unital C ∗-algebra A, the embedding A ֒→ A ⊗ K, a 7→ a ⊗ x0 lifts to an isomorphism in K-theory. Furthermore, let A and B be unital C ∗-algebras, let ϕ : A → B be a homomorphism and u ∈ Mn(A) a unitary. Then the class [u] ∈ K1(A) is mapped under ϕ∗ to [ϕ(u) + (1n − ϕ(1n))] ∈ K1(B) (see for instance [RLL00, exercise 8.5]). This is the canonical way of turning K1 into a functor. It is "filling up with units". consistent with the embeddings UnA ֒→ Un+1A of unitary matrices u 7→ (cid:18)u By Theorem 4.5, ¯u⊗x0 +(1 ^(C(S 1)∗CC2)⊗K 1(cid:19) by −1⊗x0) ∈ (C(S1)∗C C2)⊗K is isomorphic to ¯u + (1 − ¯u¯u∗) ∈ T ∗ϑ T, where ¯u = u(1 − vv∗) + (1 − uu∗)v∗ ∈ T ∗ϑ T. Note, that 1 − ¯u¯u∗ = uvv∗u∗. (cid:3) Next, we compute the K-groups of the ideal I := h1 − uu∗i in T ∗ϑ T. This will provide enough information to obtain the K-groups of T ∗ϑ T from the following exact sequence: 0 → I → T ∗ϑ T → T ⋊ϑ Z → 0 Proposition 6.6. The K-groups of I = h1 − uu∗i ⊳ T ∗ϑT are K0(I) = Z generated by [1 − uu∗] and K1(I) = 0. Proof. Consider the following exact sequences of ideals in T ∗ϑ T respectively T ⋊ϑ Z (recall the definitions I = h1 − uu∗i and J = h1 − uu∗, 1 − vv∗i): 0 0 → → I = I → → T ∗ϑ T ▽ J → → T ⋊ϑ Z ▽ h1 − vv∗i → → 0 0 As I ′ := h1 − vv∗i ⊳ T ⋊ϑ Z is isomorphic to C(S1) ⊗ K by Proposition 2.3, we know that K0(I ′) = Z is generated by the class of 1 − vv∗, whereas K1(I ′) = Z is generated by the class of u(1 − vv∗) + vv∗. The latter uses the correspondence of − 1 ⊗ x0) ∈ C(S1) ⊗ K with u(1 − vv∗) + (1 − (1 − vv∗)) ∈ T ⋊ϑ Z. u ⊗ x0 + (1 ^C(S 1)⊗K The second above sequence yields the following 6-term exact sequence in K-theory: K0(I) ↑ K1(I ′) −→ ←− K0(J) K1(J) −→ ←− K0(I ′) ↓ K1(I) The map K0(J) = Z2 → K0(I ′) = Z maps one generator [1 − vv∗] of K0(J) to the generator [1 − vv∗] of K0(I ′). Thus it is surjective and hence the connecting 30 MORITZ WEBER map K0(I ′) → K1(I) is zero. Secondly, the map K1(J) = Z → K1(I ′) = Z is an isomorphism, as the generator [u(1−vv∗)+(1−uu∗)v∗ +uvv∗u∗] ∈ K1(J) is mapped to the generator [u(1 − vv∗) + vv∗] ∈ K1(I ′). (Note that the element u ∈ T ⋊ϑ Z is a unitary and uvv∗u∗ = vv∗ in T ⋊ϑ Z.) We conclude K1(I) = 0 and from 0 → K0(I) → K0(J) = Z2 → K0(I ′) = Z → 0 we obtain K0(I) = Z. The generator is [1 − uu∗], which is mapped to the second generator [1 − uu∗] of K0(J) = Z2. (cid:3) The result on I ⊳ T ∗ϑT may also be read in terms of the ideal h¯pi in C(S1) ∗C C2. Remark 6.7. The ideal I = h1 − uu∗i ⊳ T ∗ϑT is isomorphic to the ideal h¯pi ⊗ K in (C(S1) ∗C C2) ⊗ K, mapping v(1 − uu∗)v∗ to ¯p ⊗ x0. Similarly, h1 − vv∗i ⊳ T ∗ϑ T is isomorphic to h1− ¯pi⊗K ⊳(C(S1)∗C C2)⊗K, where 1−vv∗ is mapped to (1− ¯p)⊗x0. The K-groups of h¯pi ⊳C(S1) ∗C C2 are K0(h¯pi) = Z, generated by ¯p and K1(h¯pi) = 0. Similarly, K0(h1 − ¯pi) = Z, generated by 1 − ¯p and K1(h1 − ¯pi) = 0. Proof. Under the isomorphism of J and (C(S1) ∗C C2) ⊗ K of Theorem 4.5, (uv)i ¯p = (uv)iv(1 − uu∗)v∗ ∈ J is mapped to ¯p ⊗ xi ∈ (C(S1) ∗C C2) ⊗ K for i ∈ N0. Thus the image of I = hv(1 − uu∗)v∗i in (C(S1) ∗C C2) ⊗ K is exactly h¯pi ⊗ K (cf. also Proposition 5.4). The ideal h¯pi ⊳ C(S1) ∗C C2 is the kernel of the map C(S1) ∗C C2 → C(S1), which maps ¯u 7→ u and ¯p 7→ 0. Furthermore, the map C(S1) → C(S1) ∗C C2, mapping u 7→ ¯u is a split. Hence, we have two exact sequences in K-theory, for i = 0, 1: 0 → Ki(h¯pi) → Ki(C(S1) ∗C C2) → Ki(C(S1)) → 0 As Ki(C(S1)) = Z, these sequences are split exact which yields the result using Proposition 6.4. (cid:3) We are now able to compute the K-groups of T ∗ϑ T. Theorem 6.8. The K-groups of T ∗ϑ T are K0(T ∗ϑ T) = Z, generated by the class of the unit 1 and K1(T ∗ϑ T) = 0. Proof. Consider the following exact sequence, where I = h1 − uu∗i ⊳ T ∗ϑ T: 0 → I → T ∗ϑ T → T ⋊ϑ Z → 0 Under the map K0(I) → K0(T ∗ϑT), the class of the projection 1 − uu∗ is mapped to zero, hence the map is zero. The proof for this is word-by-word the same as in the case of Theorem 6.2. Hence the 6-term exact sequence in K-theory corresponding to the above short exact sequence falls into the part K0(T ∗ϑ T) ∼= K0(T ⋊ϑ Z) = Z, where [1] ∈ K0(T ∗ϑ T) is mapped to [1] ∈ K0(T ⋊ϑ Z), and the sequence 0 → K1(T ∗ϑ T) → K1(T ⋊ϑ Z) = Z → K0(I) = Z → 0 from which we deduce K1(T ∗ϑ T) = 0. (cid:3) ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 31 Remark 6.9. Murphy ([M94]) considered C ∗-algebras associated to unital semi- ϑ(N2) in the case of groups endowed with a 2-cocycle. He gave the example of C ∗ the semigroup N2 and a 2-cocycle constructed out of a single complex scalar λ of absolute value one. This is exactly the C ∗-algebra T ∗ϑ T. In the introduction to his article, he mentions that the K-theory of this C ∗-algebra was unknown, even in the trivial case of C ∗(N2), which is T ∗ϑ T with ϑ = 0. According to Murphy, the knowledge of this K-theory would help in the theory of generalized Toeplitz operators (see [M94] or [M96] for references on this). Appendix This appendix is supposed to give a brief overview on universal C ∗-algebras. It includes some examples and some constructions of C ∗-algebras viewed as universal C ∗-algebras (cf. [B06, Sect. II.8.3] and [Cu93]). A.1 The Construction of universal C ∗-algebras. Let G = {xi i ∈ I} be a set of generators and denote by P (G) the involutive algebra of non-commutative polynomials in G ∪ G∗. Let R ⊆ P (G) be a set of relations and put J(R) ⊆ P (G) to be the involutive ideal in P (G) generated by the relations R. Then, the quotient A(G, R) of P (G) by J(R) is the universal involutive algebra, generated by G and the relations R. For all x ∈ A(G, R) we define: kxk := sup{p(x) p is a C ∗-seminorm on A(G, R)} A C ∗-seminorm is a seminorm p such that p(xy) ≤ p(x)p(y) and p(x∗x) = p(x)2. If this expression kxk is finite for all x ∈ A(G, R), then k·k is a C ∗-seminorm on A(G, R). We take the quotient of A(G, R) by the null space {x ∈ A(G, R) kxk = 0} and complete it with respect to k·k. We obtain the universal C ∗-algebra generated by G and the relations R, which we denote by C ∗(G R). The most important feature of universal C ∗-algebras is their universal property. This is, for any C ∗-algebra B with elements yi ∈ B, i ∈ I, such that the relations R are fulfilled, there exists a unique ∗-homomorphism from C ∗(G R) to B, mapping xi 7→ yi for all i ∈ I. A.2 The rotation algebra. A classic example is the rotation algebra Aϑ, where ϑ ∈ R is a parameter of rotation. If ϑ is irrational, then Aϑ is called irrational rotation algebra, and rational rotation algebra otherwise. We can define it as the following universal C ∗-algebra: Aϑ = C ∗(cid:0)u, v unitaries(cid:12)(cid:12) uv = e2πiϑvu(cid:1) The algebraic relations for a unitary u are u∗u = uu∗ = 1. (Note that 1 is also one of the generators of Aϑ.) 32 MORITZ WEBER A.3 The Toeplitz algebra. The Toeplitz algebra T may be viewed as the universal C ∗-algebra generated by an isometry, thus: T = C ∗(v v is an isometry, i.e. v∗v = 1) The Toeplitz algebra is isomorphic to the C ∗-subalgebra of L(ℓ2(N0)) generated by the unilateral shift S ∈ L(ℓ2(N0)), which is given by Sen = en+1 for all n ∈ N0. Here (en)n∈N0 is an orthonormal basis of the Hilbert space ℓ2(N0). This isomorphism is due to Coburn's theorem ([Cb67]). A.4 The algebra of complex valued matrices. The C ∗-algebra Mn(C) of complex valued n × n-matrices may be written as a universal C ∗-algebra in the following two ways. (a) Mn(C) is the universal C ∗-algebra generated by elements eij (named matrix ij = eji for units) for i, j ∈ {1, . . . , n} and the relations eijekl = δjkeil and e∗ all i, j, k, l ∈ {1, . . . , n}. (b) Mn(C) is the universal C ∗-algebra generated by elements x1, x2, . . .,xn and the relations x∗ i xj = δijx0 for all i, j ∈ {1, . . . , n}. The first picture of Mn(C) uses the correspondence of eij with the matrix in Mn(C) where all entries are zero except a unit 1 at the i-th row and j-th column. This is an isomorphism since both C ∗-algebras are n2-dimensional. The universal C ∗-algebras of (a) and (b) are isomorphic via eij 7→ xix∗ j . Thus, xi corresponds to the matrix unit ei0. Note, that the element x0 is a projection with xix0 = xi and that the xi are partial isometries. A.5 The algebra of compact operators. The algebra K of compact operators on a separable, infinite-dimensional Hilbert space may be written as a universal C ∗-algebra in the following two ways. (a) K is the universal C ∗-algebra generated by elements eij, where i, j ∈ N0 and the relations eijekl = δjkeil and e∗ ij = eji for all i, j, k, l ∈ N0. (b) K is the universal C ∗-algebra generated by elements xi, i ∈ N0 and the rela- i xj = δijx0 for all i, j ∈ N0. Note, that the element x0 is a projection tions x∗ with xix0 = xi and that the xi are partial isometries for all i ∈ N0. The elements eij correspond to the rank-one-operators matching the j-th basis vector ej to ei (where (ei)i∈N is an orthonormal basis of the Hilbert space). Denote by ϕ the surjection from the universal C ∗-algebra A of item (a) to K obtained by the universal It is an isomorphism, because the C ∗-subalgebra Mn := C ∗(eij i, j = property. 0, . . . n − 1) ⊆ A is isomorphic to Mn(C) (as Mn(C) is simple). Therefore, the ∼= Mn(C). As the restriction of ϕ to Mn is injective, again by simplicity of Mn union of all Mn for n ∈ N0 is dense in A, the map ϕ is isometric on a dense subset of A, thus it is isometric on the whole of A. We conclude that ϕ is an isomorphism. i := vi(1−vv∗) ∈ T i is injective. Now, K is the Recall the connection between K, T and C(S1). The elements x′ fulfill the relations of K. As K is simple, the map xi 7→ x′ ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 33 closed linear span of all elements xix∗ j are the only monomials in K). Hence its image in T is exactly the ideal h1 − vv∗i which is the closed linear span of all elements x′ j = vi(1 − vv∗)(v∗)j. We obtain the following short exact sequence: j , i, j, ∈ N0 (the xix∗ ix′∗ 0 → K → T → C(S1) → 0 A.6 (Full) group C ∗-algebras. We can also formulate some standard con- structions of C ∗-algebras in terms of universal C ∗-algebras. Let G be a discrete group. The (full) group C ∗-algebra C ∗(G) may be viewed as the following universal C ∗-algebra: C ∗(G) = C ∗(ug unitaries for g ∈ G ugh = uguh, ug−1 = u∗ g for all g, h ∈ G) We observe that the unitaries (ug)g∈G encode the structure of the group G. If G = Z, we get C ∗(Z) = C ∗(u u is a unitary) = C(S1) since all unitaries uk 1. If G = Z2, the group of integers Z modulo may be obtained as a power of u1 or u∗ 2Z, then C ∗(Z2) is the universal C ∗-algebra generated by a self-adjoint unitary u such that u2 = 1. Thus the only monomials are of the form 1 and u, and C ∗(Z2) is isomorphic to C2 via u 7→ (cid:18) 1 −1 (cid:19). A.7 Crossed products. Let A be a unital C ∗-algebra, let G be a discrete group, and let α : G → Aut(A) be a group homomorphism. The crossed product of A with G may be seen as the following universal C ∗-algebra: A ⋊α G = C ∗(cid:0)a ∈ A, ug unitaries for g ∈ G(cid:12)(cid:12) the relations (∗) are fulfilled(cid:1) The elements a ∈ A fulfill the relations of A and the relations (∗) are given by: (∗) ugh = uguh, ug−1 = u∗ g, ugau∗ g = αg(a) for all a ∈ A and g, h ∈ G We infer that A ⋊α G is the C ∗-algebra A with adjoint unitaries which implement the automorphisms αg, g ∈ G. If A is not unital, then A ⋊α G is defined as the ideal, generated by all elements a ∈ A in the above universal C ∗-algebra. Consider now a special case. Let A be a unital C ∗-algebra and let α be an automorphism on A. This gives rise to an action of Z on Aut(A), by n 7→ αn. Hence we can form the crossed product of A with Z which is given by: A ⋊α Z = C ∗(cid:0)a ∈ A, u unitary(cid:12)(cid:12) uau∗ = α(a) for all a ∈ A(cid:1) Note that we abuse the notation, because the automorphism as well as the action of Z on Aut(A) are denoted by α. As an example, consider the automorphism v 7→ e2πiϑv of C(S1) where now v denotes the generating unitary of C(S1). Then Aϑ = C(S1) ⋊ϑ Z. Another special case is A = C. We let a discrete group G act trivially on A and we obtain C ⋊ G = C ∗(G). 34 MORITZ WEBER A.8 Tensor products. Let A be a unital, nuclear C ∗-algebra and let B be another unital C ∗-algebra. The tensor product of A and B may be written as: A ⊗ B = C ∗(cid:0)a ∈ A, b ∈ B(cid:12)(cid:12) ab = ba for all a ∈ A, b ∈ B(cid:1) Thus A ⊗ B is the universal C ∗-algebra generated by all elements a ∈ A (with the relations of A) and all b ∈ B (with the relations of B) such that any a ∈ A commutes with any b ∈ B. The units of A and B are identified. If A and B are not necessarily unital, then A ⊗ B is the ideal generated by all products ab in the above C ∗-algebra. A.9 Free products. Let A and B be unital C ∗-algebras. The (unital) free product A ∗C B of A and B is defined as the universal C ∗-algebra generated by all elements a ∈ A (with the relations of A) and all b ∈ B (with the relations of B) under the identification of the units of A and B. There are no further relations between elements from A and those from B. The constructions of the free product of C ∗-algebras and the one for groups fit together in the following way. Let G1 and G2 be discrete groups and denote by G1 ∗ G2 their free product (of groups). Then the full group C ∗-algebra C ∗(G1 ∗ G2) is isomorphic to the unital free product (of C ∗-algebras) of C ∗(G1) and C ∗(G2). (Cf. Blackadar's book on K-theory [B98, chapter V, 10.11.11(f)] for instance.) Acknowledgements My special thanks go to Joachim Cuntz, who encouraged me to work on isometries with twisted commutation relations. The discussions with him were very helpful and I appreciate his ideas and suggestions a lot. I would also like to thank the re- search group in Functional Analysis and Operator Algebras in Munster, in particular Siegfried Echterhoff, Thomas Timmermann, and Christian Voigt. References [B98] Bruce E. Blackadar, K-theory for operator algebras, Second edition, Mathematical Sci- ences Research Institute Publications 5, Cambridge University Press, Cambridge, 1998. [B06] Bruce E. Blackadar, Operator algebras. Theory of C ∗-algebras and von Neumann algebras, Encyclopaedia of Mathematical Sciences 122, Operator Algebras and Non-commutative Ge- ometry III, Springer-Verlag, Berlin, 2006. [BCL78] Charles A. Berger, Lewis A. Coburn, Arnold Lebow, Representation and index theory for C ∗-algebras generated by commuting isometries, J. Functional Analysis 27 (1978), no. 1, 51 -- 99. [BO08] Nathanial P. Brown, Narutaka Ozawa, C ∗-algebras and finite-dimensional approxima- tions, Graduate Studies in Mathematics 88, Amer. Math. Soc., Providence, RI, 2008. [BS94] Marek Bo zejko, Roland Speicher, Completely positive maps on Coxeter groups, deformed commutation relations, and operator spaces, Math. Ann. 300 (1994), no. 1, 97 -- 120. [Cb67] Lewis A. Coburn, The C ∗-algebra generated by an isometry, Bull. Amer. Math. Soc. 73 (1967), 722 -- 726. [Co85] Alain Connes, Noncommutative differential geometry, Inst. Hautes ´Etudes Sci. Publ. Math. No. 62 (1985), 257 -- 360. [Co94] Alain Connes, Noncommutative geometry, Academic Press, Inc., San Diego, CA, 1994. ISOMETRIES WITH TWISTED COMMUTATION RELATIONS 35 [Cu82] Joachim Cuntz, The K-groups for free products of C ∗-algebras, Operator algebras and applications, Part I (Kingston, Ont., 1980), pp. 81 -- 84, Proc. Sympos. Pure Math. 38, Amer. Math. Soc., Providence, RI, 1982. [Cu93] Joachim Cuntz, A survey of some aspects of noncommutative geometry, Jahresber. Deutsch. Math.-Verein. 95 (1993), no. 2, 60 -- 84. [EH67] Edward G. Effros, Frank Hahn, Locally compact transformation groups and C ∗-algebras, Bull. Amer. Math. Soc. 73 (1967), 222 -- 226. [GVF01] Jos´e M. Gracia-Bond´ıa, Joseph C. V´arilly, H´ector Figueroa, Elements of noncom- mutative geometry, Birkhauser Advanced Texts: Basler Lehrbucher, Boston, MA, 2001. [JPS05] Palle E. T. Jørgensen, Daniil P. Proskurin, Yuriı S. Samoilenko, On C ∗-algebras generated by pairs of q-commuting isometries, J. Phys. A 38 (2005), no. 12, 2669 -- 2680. [Ka01] Zakhar A. Kabluchko, On the extension of higher-dimensional noncommutative tori, Methods Funct. Anal. Topology 7 (2001), no. 1, 28 -- 33. [KP00] Eberhard Kirchberg, N. Christopher Phillips, Embedding of exact C ∗-algebras in the Cuntz algebra O2, J. Reine Angew. Math. 525 (2000), 17 -- 53. [M94] Gerard J. Murphy, Crossed products of C ∗-algebras by semigroups of automorphisms, Proc. London Math. Soc. (3) 68 (1994), no. 2, 423 -- 448. [M96] Gerard J. Murphy, C ∗-algebras generated by commuting isometries, Rocky Mountain J. Math. 26 (1996), no. 1, 237 -- 267. [M03] Gerard J. Murphy, Primitivity conditions for full group C ∗-algebras, Bull. London Math. Soc. 35 (2003), no. 5, 697 -- 705. [Pr00a] Daniil P. Proskurin, Stability of a special class of qij -CCR and extensions of irrational rotation algebras, Methods Funct. Anal. Topology 6 (2000), no. 3, 97 -- 104. [Pr00b] Daniil P. Proskurin, Stability of a special class of qij-CCR and extensions of higher- dimensional noncommutative tori, Lett. Math. Phys. 52 (2000), no. 2, 165 -- 175. [PV80] Mihai V. Pimsner, Dan V. Voiculescu, Exact sequences for K-groups and Ext-groups of certain cross-product C ∗-algebras, J. Operator Theory 4 (1980), no. 1, 93 -- 118. [R81] Marc A. Rieffel, C ∗-algebras associated with irrational rotations, Pacific J. Math. 93 (1981), no. 2, 415 -- 429. [RLL00] Mikael Rørdam, Flemming Larsen, Niels J. Laustsen, An introduction to K-theory for C ∗-algebras, London Mathematical Society Student Texts 49, Cambridge University Press, Cambridge, 2000. [Ro77] Jonathan Rosenberg, Amenability of crossed products of C ∗-algebras, Comm. Math. Phys. 57 (1977), no. 2, 187 -- 191. [Wa76] Simon Wassermann, On tensor products of certain group C ∗-algebras, J. Functional Anal- ysis 23 (1976), no. 3, 239 -- 254. [Wa78] Simon Wassermann, A pathology in the ideal space of L(H) ⊗ L(H), Indiana Univ. Math. J. 27 (1978), no. 6, 1011 -- 1020. [Wa90] Simon Wassermann, Tensor products of free-group C ∗-algebras, Bull. London Math. Soc. 22 (1990), no. 4, 375 -- 380. [Wi07] Dana P. Williams, Crossed products of C ∗-algebras, Mathematical Surveys and Mono- graphs 134, Amer. Math. Soc., Providence, RI, 2007. [Y86] Hong Sheng Yin, A simple proof of the classification of rational rotation C ∗-algebras, Proc. Amer. Math. Soc. 98 (1986), no. 3, 469 -- 470. [Z68] Georges Zeller-Meier Produits crois´es d'une C ∗-alg`ebre par un groupe d'automorphismes, J. Math. Pures Appl. (9) 47 (1968), 101 -- 239. Saarland University, Fachrichtung Mathematik, Postfach 151150, 66041 Saarbrucken, Germany E-mail address: [email protected]
1507.03437	3	1507	2016-02-16T03:07:17	On the classification of simple amenable $C*$-algebras with finite decomposition rank, II	[ "math.OA" ]	We prove that every unital stably finite simple amenable $C^*$-algebra $A$ with finite nuclear dimension and with UCT such that every trace is quasi-diagonal has the property that $A\otimes Q$ has generalized tracial rank at most one, where $Q$ is the universal UHF-algebra. Consequently, $A$ is classifiable in the sense of Elliott.	math.OA	math	On the classification of simple amenable C-algebras with finite decomposition rank, II George A. Elliott, Guihua Gong, Huaxin Lin, and Zhuang Niu Abstract We prove that every unital simple separable C-algebra A with finite decomposition rank which satisfies the UCT has the property that A ⊗ Q has generalized tracial rank at most one, where Q is the universal UHF-algebra. Consequently, A is classifiable in the sense of Elliott. 1 Introduction In a recent development in the Elliott program, the program of classification of amenable C- algebras, a certain class of finite unital simple separable amenable C-algebras, denoted by N1, was shown to be classified by the Elliott invariant ([14]). One important feature of this class of C-algebras is that it exhausts all possible values of the Elliott invariant for unital simple separable C-algebras which have finite decomposition rank (a property introduced in [19]; see Definition 2.10 below). The purpose of this note is to show that, in fact, every unital simple separable (non- elementary) C-algebra which has finite decomposition rank and satisfies the Universal Co- efficient Theorem (UCT) is in the class N1. Since every C-algebra in N1 was shown in [14] to be isomorphic to the inductive limit of a sequence of subhomogeneous C-algebras with no dimension growth, the C-algebras in N1 have finite decomposition rank. In other words, the class N1 is precisely the class of all unital simple separable (non-elementary) C-algebras which have finite decomposition rank and satisfy the UCT, and hence we obtain a classification for all of these C-algebras: Theorem 1.1. Let A be a unital simple separable (non-elementary) C-algebra with finite de- composition rank, and assume that A satisfies the UCT. Then A ∈ N1. (See Definition 2.6, below.) Hence, if A and B are two (non-elementary) unital simple separable C-algebras with finite decomposition rank which satisfy the UCT, then A ∼= B if, and only if, (K0(A), K0(A)+, [1A]0, K1(A), T(A), rA) ∼= (K0(B), K0(B)+, [1B]0, K1(B), T(B), rB). In fact, we shall obtain (see Theorem 4.4, below) the formally stronger result that every finite unital simple separable (non-elementary) C-algebra with finite nuclear dimension, which satisfies the UCT, and whose tracial states are all quasidiagonal, is in the class N1. This result, combined with the recent result of [29] (that the quasidiagonality hypothesis is redundant), yields that the class N1 includes all finite unital simple separable C-algebras with finite nuclear dimension which satisfy the UCT-see 4.9 below. (The case of infinite unital simple separable C-algebras with finite nuclear dimension was dealt with fifteen years ago by Kirchberg and Phillips-see 4.10 below.) In a recent paper, [11], Elliott and Niu proved that every unital simple separable (non- elementary) C-algebra A with finite decomposition rank, satisfying the UCT, and such that 1 K0(A) has torsion free rank one, belongs to N1. The present paper is a continuation of [11] with, now, a definitive result. Acknowledgements. During this research, G.A.E. received support from the Natural Sci- ences and Engineering Research Council of Canada (NSERC), and G.G. and H.L. received sup- port from the NSF. Z.N. received support from the Simons Foundation (a Collaboration Grant). G.G., H.L., and Z.N. also received support from The Research Center for Operator Algebras at East China Normal University which is funded by the Science and Technology Commission of Shanghai Municipality (grant No.13dz2260400). 2 Preliminaries Definition 2.1. As usual, let Q denote the field of rational numbers. Let us use the notation Q for the UHF-algebra with K0(Q) = Q and [1Q] = 1. Definition 2.2 (N. Brown [4]). Let A be a unital C-algebra. Denote by T(A) the tracial state space of A, and denote by Tqd(A) the tracial states τ with the following property: For any finite subset F and ǫ > 0, there exists a unital completely positive map ϕ : A → Q such that and τ (a) − tr(ϕ(a)) < ǫ, a ∈ F, kϕ(a)ϕ(b) − ϕ(ab)k < ǫ, a, b ∈ F, where tr is the unique tracial state of Q. Definition 2.3. Let F1 and F2 be two finite dimensional C-algebras and let ψ0, ψ1 : F1 → F2 be two unital homomorphisms. Consider the corresponding mapping torus, C = C(F1, F2, ψ0, ψ1) = {(f, a) ∈ C([0, 1], F2) ⊕ F1 : f (0) = ψ0(a) and f (1) = ψ1(a)}. Denote by C the class of unital C-algebras obtained in this way. C-algebras in the class C are often called Elliott-Thomsen building blocks. They are also called one-dimensional non- commutative CW complexes. Denote by C0 the subclass of C consisting of those C-algebras C ∈ C such that K1(C) = {0}. We shall in fact only work with the Q-stabilizations of these algebras, which can be described just by replacing F1 and F2 with finite direct sums of copies of Q. Definition 2.4 (9.1 of [14]). Let A be a unital simple C-algebra. We shall say that A has generalized tracial rank at most one, if the following property holds: Let ǫ > 0, let a ∈ A+ \{0} and let F ⊂ A be a finite subset. There exist a non-zero projection p ∈ A and a sub-C-algebra C ∈ C with 1C = p such that kxp − pxk < ǫ, x ∈ F, dist(pxp, C) < ǫ, 1 − p . a. x ∈ F, and The last condition means that there exists a partial isometry v ∈ A such that v∗v = 1 − p and vv∗ ∈ aAa. If A has generalized tracial rank at most one, we will write gTR(A) ≤ 1. It was shown in [14] that if gTR(A) ≤ 1, then A is quasidiagonal, and is Z-stable if it is also amenable. 2 Definition 2.5. Let A and B be unital C-algebras and let L : A → B be a contractive completely positive map. Let G be a finite subset of A and δ > 0. Recall that L is said to be G-δ-multiplicative if kL(x)L(y) − L(xy)k < δ for all x, y ∈ G. Given a finite subset P of projections in A, if G is sufficiently large and δ is sufficiently small, then there is a projection q ∈ B such that kL(p) − qk < 1/4. Moreover, for each projection p ∈ G, if δ < 1/4, then the projection q can be chosen such that kL(p) − qk < 2δ. (2.1) Note that if q′ ∈ B is another projection such that kL(p) − q′k < 1/4, then p and q are unitarily equivalent. Recall that [L(p)] often denotes this equivalence class of projections (see e.g. [20]). As usual, when [L(p)] is written, it is understood that G is sufficiently large and δ is sufficiently small that [L(p)] is well defined. Definition 2.6 ([14]). Let A be a unital simple separable C-algebra. Let us say that A has rational generalized tracial rank at most one if gTR(A ⊗ Q) ≤ 1. Let us say that, in addition, A belongs to the class N1 if it satisfies the UCT ([28]), and is Jiang-Su stable, i.e., invariant under tensoring with the Jiang-Su C-algebra ([17]; see also [9]). As pointed out above, it follows from [14] (together with [31]) that, instead of the last property, it is equivalent to assume finite decomposition rank (or even just finite nuclear dimension); see Definition 2.10 below. (This fact is of interest in itself, and is also pertinent in the present context, as it enables the statement of Theorem 1.1-and indeed our proof of it consisting of a reduction to [14]-to be read without reference to the somewhat technical definition of the Jiang-Su algebra.) (Of course, it is an interesting question whether the assumption of finite decomposition rank in Theorem 1.1 might, `a la Toms-Winter (see [15]), be replaced by Jiang-Su stability, together with Murray-von Neumann finiteness.) The following are the main results of [14]: Theorem 2.7. Let A and B be two unital C-algebras in N1. Then A ∼= B if and only if Ell(A) ∼= Ell(B), i.e., A ∼= B if and only if (K0(A), K0(A)+, [1A]0, K1(A), T(A), rA) ∼= (K0(B), K0(B)+, [1B]0, K1(B), T(B), rB). Moreover, any isomorphism between Ell(A) and Ell(B) can be lifted to an isomorphism between A and B. Theorem 2.8. For any non-zero countable weakly unperforated simple ordered group G0 with order unit u, any countable abelian group G1, any non-empty metrizable Choquet simplex T , and any surjective affine map r : T → Su(G0) (Su(G0) is the state space of G0-always non- empty), there exists a (unique) unital simple C-algebra C in N1, which is the inductive limit of a sequence of subhomogeneous C-algebras with two-dimensional spectrum, such that Ell(C) = (G0, (G0)+, u, G1, T, r). Definition 2.9. Let A and B be C-algebras. Recall ([19]) that a completely positive map ϕ : A → B is said to have order zero if ab = 0 ⇒ ϕ(a)ϕ(b) = 0, a, b ∈ A. Definition 2.10 ([19], [33]). A C-algebra A has nuclear dimension at most n, if there exists a net (Fλ, ψλ, ϕλ), λ ∈ Λ, such that the Fλ are finite dimensional C-algebras, and such that ψλ : A → Fλ and ϕλ : Fλ → A are completely positive maps satisfying 3 (1) ϕλ ◦ ψλ → idA pointwise (in norm), (2) kψλk ≤ 1, and (3) for each λ, there is a decomposition Fλ = F (0) is a contractive order zero map. λ ⊕ · · · ⊕ F (n) λ such that each restriction ϕλ F (j) λ Moreover, if the the map ϕλ can be chosen to be contractive itself, then A is said to have decomposition rank at most n. (Recall that finite nuclear dimension immediately implies nuclearity, which by [6] and [16] is equivalent to amenability.) The main theorem of this paper is that the class N1 of C-algebras actually contains (and hence coincides with) the class of all (non-elementary) unital simple separable C-algebras with finite decomposition rank which also satisfy the UCT. In particular it follows (on using both 2.7 and 2.8) that every such C-algebra is the inductive limit of a sequence of subhomogeneous C-algebras (with no dimension growth). 3 Some existence theorems Lemma 3.1. Let A be a unital simple separable amenable quasidiagonal C-algebra satisfying the UCT. Assume that A ∼= A ⊗ Q. Let a finite subset G of A and ǫ1, ǫ2 > 0 be given. Let p1, p2, ..., ps ∈ A be projections such that [1], [p1], [p2], ..., [ps] ∈ K0(A) are Q-linearly independent. (Recall that K0(A) ∼= K0(A ⊗ Q) ∼= K0(A) ⊗ Q.) There are a G-ǫ1-multiplicative completely positive map σ : A → Q with σ(1) a projection satisfying tr(σ(1)) < ǫ2 (where tr denotes the unique trace state on Q), and δ > 0, such that, for any r1, r2, ..., rs ∈ Q with ri < δ, i = 1, 2, ..., s, there is a G-ǫ1-multiplicative completely positive map µ : A → Q, with µ(1) = σ(1), such that [σ(pi)] − [µ(pi)] = ri, i = 1, 2, ..., s. Proof. Let us agree that σ and µ (to be constructed below) are also understood to be required to be sufficiently multiplicative on p1, p2, ..., pk that the classes [σ(pi)] and [µ(pi)] make sense (see 2.5 above). (Similarly for other completely positive approximately multiplicative maps, to be introduced below.) Denote by G0 the subgroup of K0(A) generated by {[1A], [p1], ..., [ps]}. Since [1A], [p1], ..., [ps] are Q-linearly independent, for each i = 1, 2, ..., s, there exists a homomorphism αi : K0(A) → Q ∼= K0(Q) such that αi([pi]) = 1, αi([1A]) = 0, and αi([pj]) = 0, j 6= i. (3.1) We may regard αi as an element of KL(A, Q) (see [8]). Since A is a unital simple amenable It follows from [2] quasidiagonal C-algebra, by [2], A is a unital simple strong NF-algebra. n=1 An, where {An} is an increasing sequence of unital, amenable, residually finite- dimensional C-algebras. It follows from Theorem 5.9 of [20] that there are G-ǫ1-multiplicative completely positive maps σi, µi : A → Q ⊗ K such that σi(1A) and µi(1A) are projections, and that A = S∞ [σi]G0 − [µi]G0 = αiG0, i = 1, 2, ..., s. (3.2) 4 Since αi([1A]) = 0, we have [σi(1A)] = [µi(1A)]. Therefore, without loss of generality, we may assume that σi(1A) = µi(1A) =: Pi, i = 1, 2, ..., s. Consider the projection P := sM i=1 (Pi ⊕ Pi), and the unital G-ǫ1-multiplicative completely positive map sM i=1 (σi ⊕ µi) : A → P (Q ⊗ K)P. Note that P (Q ⊗ K)P ∼= Q. Choose a projection R ∈ Q ⊗ K with 0 < tr(R)≤ min{1, ǫ2} and a rescaling S : Q ⊗ K → Q ⊗ K, P 7→ R. Consider the map σ := S ◦ ( sM i=1 (σi ⊕ µi)) : A → Q ⊗ K and the strictly positive number δ := tr(R) tr(P ) . (Here, tr denotes the tensor product of the traces on Q and K, normalized in the usual way.) Note that since tr(R)≤1, one has tr(σ(1)) = tr(R)≤1, and so we may regard σ as a map from A to Q (rather than Q ⊗ K). Let us show that σ and δ satisfy the condition of the lemma. Let r1, r2, ..., rs ∈ Q be given with ri < δ, i = 1, 2, ..., s. For each i = 1, 2, ..., s, choose a projection Ri ∈ Q ⊗ K with tr(Ri) = ri, and choose a rescaling Si : Q ⊗ K → Q ⊗ K, 1 ⊗ e 7→ Ri, where e is a minimal non-zero projection of K. For each i = 1, 2, ..., s, consider the pair of maps Then, for each i = 1, 2, ..., s, Si ◦ σi, Si ◦ µi : A → Q ⊗ K. [Si ◦ σi(pi)] − [Si ◦ µi(pi)] = ri, [Si ◦ σi(1)] − [Si ◦ µi(1)] = 0, [Si ◦ σi(pj)] − [Si ◦ µi(pj)] = 0, j = 1, 2, ..., s, j 6= i. Consider the direct sum maps σ := (M ri>0 Si ◦ σi) ⊕ (M ri<0 Si ◦ µi) 5 and µ := (M ri>0 Si ◦ µi) ⊕ (M ri<0 Si ◦ σi). It follows from (3.1) and (3.2) that [σ(pi)] − [µ(pi)] = ri, i = 1, 2, ..., s. Note that σ = S ◦ ( sM i=1 (σi ⊕ µi)) = sM i=1 ((S ◦ σi) ⊕ (S ◦ µi)). (3.3) For each i = 1, 2, ..., s, since tr(Si(P )) = tr(P ) · tr(Si(1 ⊗ e)) = tr(P ) · tr(Ri) = tr(P )ri < tr(P )δ = tr(R) = tr(S(P )), there is a rescaling Ti : Q ⊗ K → Q ⊗ K such that S = Si ⊕ Ti. Therefore, by (3.3), σ = (((Si ⊕ Ti) ◦ σi) ⊕ ((Si ⊕ Ti) ◦ µi)) sM i=1 ((Si ◦ σi) ⊕ (Ti ◦ σi)) ⊕ ((Si ◦ µi) ⊕ (Ti ◦ µi)) ((Si ◦ σi) ⊕ (Ti ◦ σi))) ⊕ (M ((Si ◦ µi) ⊕ (Ti ◦ µi))) ⊕ (M ri≤0 ri≥0 ((Si ◦ σi) ⊕ (Ti ◦ σi))) ((Si ◦ µi) ⊕ (Ti ◦ µi))) = i=1 sM sM = (M ⊕(M ri>0 i=1 ri<0 = σ ⊕ γ, where γ = (M (Ti ◦ σi)) ⊕ (M ((Si ◦ σi) ⊕ (Ti ◦ σi))) ⊕ (M (Ti ◦ µi)) ⊕ (M ri>0 ri≤0 ri<0 ri≥0 ((Si ◦ µi) ⊕ (Ti ◦ µi))). Consider the completely positive map µ := µ ⊕ γ. We have [σ(pi)] − [µ(pi)] = [σ(pi)] − [µ(pi)] = ri, i = 1, 2, ..., s, as desired (with µ being regarded as a map from A to Q, as µ(1) = σ(1)). Remark 3.2. The assumption that A is amenable in 3.1 can be removed. In the proof one can apply Theorem 5.5 of [7] in place of Theorem 5.9 of [20]. Let l, r = 1, 2, ... be given. K0(Qr) with Qr) by identifying [1Ql] with (1, 1, ..., 1 In the rest of the paper, we identify K0(Ql) with Ql (and )), where Ql = ) and ([1Qr ] with (1, 1, ..., 1 {z l } {z r } Q ⊕ · · · ⊕ Q and Qr = Q ⊕ · · · ⊕ Q . If ψ : Ql → Qr are unital, then {z l } {z r } (ψ)∗0(1, 1, ..., 1 ) = (1, 1, ..., 1 ), {z l } 6 {z r } and therefore (ψ)∗0(t, t, ..., t {z } l ), t ∈ Q. (3.4) ) = (t, t, ..., t {z } r Lemma 3.3. Let A be a unital simple separable amenable quasidiagonal C-algebra satisfying the UCT. Assume that A ∼= A ⊗ Q. Let G be a finite subset of A, let ǫ1, ǫ2 > 0, and let p1, p2, ..., pk ∈ A be projections such that [1A], [p1], [p2], ..., [ps] ∈ K0(A) are Q-linearly independent. There exists δ > 0 satisfying the following condition. Let ψk : Ql → Qr, k = 0, 1, be unital homomorphisms, where l, r = 1, 2, ... . Set D = {x ∈ Ql : (ψ0)∗0(x) = (ψ1)∗0(x)} ⊂ Ql. There exists a G-ǫ1-multiplicative completely positive map Σ : A → Ql, such that Σ(1A) is a projection, with the following properties: τ (Σ(1A)) < ǫ2, τ ∈ T(Ql), [Σ(1A)], [Σ(pj)] ∈ D, j = 1, 2, ..., s, and, for any r1, r2, ..., rs ∈ Qr satisfying ri,j < δ, where ri = (ri,1, ri,2, ..., ri,r), i = 1, 2, ..., s, there is a G-ǫ1-multiplicative completely positive map µ : A → Qr, with µ(1A) a projection, such that [ψ0 ◦ Σ(pi)] − [µ(pi)] = ri, i = 1, 2, ..., s, and [µ(1A)] = [ψ0 ◦ Σ(1A)]. Proof. Put p0 = 1A and P = {[1A], [p1], ..., [ps]}. Applying Lemma 3.1, we obtain a G-ǫ1- multiplicative σ : A → Q and δ > 0 satisfying the conclusion of Lemma 3.1 with respect to G, ǫ1, ǫ2, and P. Let us show that δ is as desired. For a given integer l = 1, 2, ..., consider the map Σ : A → Ql, the sum of l copies of σ, Σ = σ ⊕ σ ⊕ · · · ⊕ σ. Let us show that Σ has the required properties. Let r = 1, 2, ... and ψk : Ql → Qr, k = 0, 1, be given (as in the statement of the condition on δ to be verified). Since ψ0 and ψ1 are assumed to be unital, [1Qr ] = (1, 1, ..., 1) ∈ D. It then follows that [Σ(pi)] = ([σ(pi)], [σ(pi)], ..., [σ(pi)]) = [σ(pi)](1, 1, ..., 1) ∈ D, i = 0, 1, 2, ..., s, where [σ(pi)] is regarded as a rational number. In other words, [ψ0 ◦ Σ(pi)] = [ψ1 ◦ Σ(pi)], i = 0, 1, 2, ..., s. Since tr([σ(1A)]) < ǫ2, one has τ ([Σ(1A)]) < ǫ2, τ ∈ T(Ql). 7 Let r1, r2, ..., rs ∈ Qr be given such that ri,j < δ, j = 1, 2, ..., r and i = 1, 2, ..., s. Let us show that µ exists as required. Fix j = 1, 2, ..., r and let µj : A → Q (in place of µ) denote the G-ǫ1-multiplicative completely positive map given by Lemma 3.1 for the s-tuple r1,j, r2,j, ..., rs,j . That is, µj(1A) = σ(1A), [σ(pi)] − [µj(pi)] = ri,j ∈ K0(Q), i = 1, 2, ..., s, (3.5) and Define µ : A → Qr by tr(µj(1A))= tr(σ(1A)) < ǫ2. Then, for each i = 1, 2, ..., s, µ(a) = (µ1(a), µ2(a), ..., µr(a) ), a ∈ A. {z r {z l } } [ψ0 ◦ Σ(pi)] − [µ(pi)] = (ψ0)∗0([σ(pi)], [σ(pi)], ..., [σ(pi)] ) − ([µ1(pi)], [µ2(pi)], ..., [µr(pi)] ) = ([σ(pi)], [σ(pi)], ..., [σ(pi)] ) − ([µ1(pi)], [µ2(pi)], ..., [µr(pi)] ) (by (3.4)) {z r } {z r } {z r } = (([σ(pi)] − [µ1(pi)]), ([σ(pi)] − [µ2(pi)]), ..., ([σ(pi)] − [µr(pi)])) = (ri,1, ri,2, ..., ri,r) = ri, (by (3.5)) as desired. A similar computation shows that [ψ0 ◦ Σ(1A)] = [µ(1A)]. Lemma 3.4. Let A be a unital simple separable amenable quasidiagonal C-algebra satisfying the UCT. Assume that A ∼= A ⊗ Q. Let G ⊂ A be a finite subset, let ǫ1, ǫ2 > 0 and let p1, p2, ..., ps ∈ A be projections such that [1A], [p1], [p2], ..., [ps] ∈ K0(A) are Q-linearly independent. Then there exists δ > 0 satisfying the following condition. Let ψk : Ql → Qr, k = 0, 1, be unital homomorphisms, where l, r = 1, 2, .... Set D = {x ∈ Ql : (ψ0)∗0(x) = (ψ1)∗0(x)} ⊂ Ql. There exists a G-ǫ1-multiplicative completely positive map Σ : A → Ql, such that Σ(1A) is a projection, with the following properties: τ (Σ(1A)) < ǫ2, τ ∈ T(Ql), [Σ(1A)], [Σ(pi)] ∈ D, i = 1, 2, ..., s, and, for any r1, r2, ..., rs ∈ Ql satisfying ri,j < δ, where ri = (ri,1, ri,2, ..., ri,l), i = 1, 2, ..., s, there is a G-ǫ1-multiplicative completely positive map µ : A → Ql, with µ(1A) = Σ(1A), such that [Σ(pi)] − [µ(pi)] = ri, i = 1, 2, ..., s. 8 Proof. This is similar to the proof of Lemma 3.3. Put [p0] = [1A] and P = {[p0], [p1], ..., [ps]}. Applying Lemma 3.1, we obtain a G-ǫ1-multiplicative σ : A → Q and δ > 0 satisfying the conclusion of Lemma 3.1 with respect to (G, ǫ1, ǫ2). Let us show that δ is as desired. Consider the map Σ : A → Ql, the sum of l copies of σ, Σ = σ ⊕ σ ⊕ · · · ⊕ σ, for a given l = 1, 2, ... . Then the same argument as that of Lemma 3.1 shows that It is also clear that [Σ(pi)] ∈ D, i = 0, 1, 2, ..., s. τ (Σ(1A)) < ǫ2, τ ∈ T(Ql). Let r1, r2, ..., rk ∈ Ql be given such that ri,j < δ, j = 1, 2, ..., l and i = 1, 2, ..., s. Let us show that µ exists as required. Fix j = 1, 2, ..., r and let µj : A → Q (in place of µ) denote the G-ǫ1-multiplicative completely positive map given by Lemma 3.1 for the s-tuple r1,j, r2,j, ..., rs,j. That is, µj(1A) = σ(1A), and [σ(pi)] − [µj(pi)] = ri,j ∈ K0(Q), i = 1, 2, ..., s. (3.6) Of course, also, Define µ : A → Ql by tr(µj(1A))= tr(σ(1A)) < ǫ2. µ(a) = (µ1(a), µ2(a), ..., µr(a) ), a ∈ A. {z l } Then, for each i = 1, 2, ..., s, [Σ(pi)] − [µ(pi)] = ([σ(pi)], [σ(pi)], ..., [σ(pi)] ) − ([µ1(pi)], [µ2(pi)], ..., [µr(pi)] ) {z l } = (([σ(pi)] − [µ1(pi)]), ([σ(pi)] − [µ2(pi)]), ..., ([σ(pi)] − [µr(pi)])) = (ri,1, ri,2, ..., ri,l) = ri (by (3.6)), {z l } as desired. Moreover, since µj(1A) = σ(1A), we have Σ(1A) = µ(1A). 4 The main result Let us begin by recalling the (stable) uniqueness result used in [11]: Lemma 4.1 (Corollary 2.6 of [11]; see also Lemma 4.14 of [14], and Definition 5.6 and Theorem 5.9 of [21]). Let A be a simple, unital, amenable, separable C-algebra which satisfies the UCT. Assume that A ∼= A ⊗ Q. For any ǫ > 0, any finite subset F of A, there exist δ > 0, a finite subset G ⊂ A, a finite subset P of projections in A, and an integer n ∈ N with the following property. For any three completely positive contractions ϕ, ψ, ξ : A → Q which are G-δ-multiplicative, with ϕ(1) = ψ(1) = 1Q − ξ(1) a projection, [ϕ(p)]0 = [ψ(p)]0 in K0(Q) for all p ∈ P, and tr(ϕ(1)) = tr(ψ(1)) < 1/n, where tr is the unique tracial state of Q, there exists a unitary u ∈ Q such that ku∗(ϕ(a) ⊕ ξ(a))u − ψ(a) ⊕ ξ(a)k < ǫ, a ∈ F. 9 The following two existence results are related to Lemma 16.9 of [14] (and its proof). Lemma 4.2. Let A be a unital simple separable amenable C-algebra which satisfies the UCT. Assume that A ⊗ Q ∼= A. For any ǫ > 0 and any finite subset F of A, there exist δ > 0, a finite subset G of A, and a finite subset P of projections in A with the following property. Let ψ, ϕ : A → Q be two unital G-δ-multiplicative completely positive maps such that [ψ]P = [ϕ]P . Then there is a unitary u ∈ Q and a unital F-ǫ-multiplicative completely positive map L : A → C([0, 1], Q) such that Moreover, if π0 ◦ L = ψ and π1 ◦ L = Ad u ◦ ϕ. tr ◦ ψ(h) − tr ◦ ϕ(h) < ǫ′/2, h ∈ H, for a finite set H ⊂ A and ǫ′ > 0, then L may be chosen such that tr ◦ πt ◦ L(h) − tr ◦ π0 ◦ L(h) < ǫ′, h ∈ H, t ∈ [0, 1]. Here, πt : C([0, 1], Q) → Q is the point evaluation at t ∈ [0, 1]. (4.1) (4.2) (4.3) Proof. This is a direct application of the stable uniqueness theorem (Corollary 2.6) of [11], restated as Lemma 4.1 above. Let F ⊂ A be a finite subset and let ǫ > 0 be given. We may assume that 1A ∈ F and every element of F has norm at most one. Write F1 = {ab : a, b ∈ F}. Note that F ⊆ F1. Let δ, G, P, and n be as assured by Lemma 4.1 for F1 and ǫ/4. We may also assume that F ⊂ G and δ ≤ ǫ/4. Since Q ∼= Q ⊗ Q, we may assume, without loss of generality, that ϕ(a), ψ(a) ∈ Q ⊗ 1. Pick i=0 ei = 1Q. Then, consider the mutually equivalent projections e0, e1, e2, ..., en ∈ Q satisfying Pn maps ϕi, ψi : A → Q ⊗ eiQei, i = 0, 1, ..., n, which are defined by ϕi(a) = ϕ(a) ⊗ ei and ψi(a) = ψ(a) ⊗ ei, a ∈ A, and consider the maps Φn+1 := ϕ = ϕ0 ⊕ ϕ1 ⊕ · · · ⊕ ϕn, Φ0 := ψ = ψ0 ⊕ ψ1 ⊕ · · · ⊕ ψn and Φi := ϕ0 ⊕ · · · ⊕ ϕi−1 ⊕ ψi ⊕ · · · ⊕ ψn, i = 1, 2, ..., n. Since ei is unitarily equivalent to e0 for all i, one has [ϕi]P = [ψj]P , 0 ≤ i, j ≤ n. and in particular, [ϕi]P = [ψi]P , i = 0, 1, .., n. (4.4) Note that, for each i = 0, 1, ..., n, ∼ ψi ⊕ (ϕ0 ⊕ ϕ1 ⊕ · · · ⊕ ϕi−1 ⊕ ψi+1 ⊕ ψi+2 ⊕ · · · ⊕ ψn), Φi Φi+1 ∼ ϕi ⊕ (ϕ0 ⊕ ϕ1 ⊕ · · · ⊕ ϕi−1 ⊕ ψi+1 ⊕ ψi+2 ⊕ · · · ⊕ ψn), 10 where ∼ denotes the relation of unitary equivalence. In view of this, and (4.4), applying Lemma 4.1 we obtain unitaries ui ∈ Q, i = 0, 1, ..., n, such that k Φi+1(a) − Φi(a)k < ǫ/4, a ∈ F1, (4.5) where Φ0 := Φ0 = ψ and Φi+1 := Ad ui ◦ · · · ◦ Ad u1 ◦ Ad u0 ◦ Φi+1, i = 0, 1, ..., n. Put ti = i/(n + 1), i = 0, 1, ..., n + 1, and define L : A → C([0, 1], Q) by πt ◦ L = (n + 1)(ti+1 − t) Φi + (n + 1)(t − ti) Φi+1, t ∈ [ti, ti+1], i = 0, 1, ..., n. By construction, π0 ◦ L = Φ0 = ψ and π1 ◦ L = Φn+1 = Ad un ◦ · · · ◦ Ad u1 ◦ Ad u0 ◦ ϕ. (4.6) Since Φi, i = 0, 1, ..., n, are G-δ-multiplicative (in particular F-ǫ/4-multiplicative), it follows from (4.5) that L is F-ǫ-multiplicative. By (4.6), L satisfies (4.1) with u = un · · · u1u0. Moreover, if there is a finite set H such that (4.2) holds, it is then also straightforward to verify that L satisfies (4.3), as desired. Lemma 4.3. Let A be a unital simple separable amenable C-algebra with T(A) = Tqd(A) which satisfies the UCT. Assume that A ⊗ Q ∼= A. For any σ > 0, ǫ > 0, and any finite subset F of A, there exist a finite set of projections P in A and δ > 0 with the following property. Denote by G ⊆ K0(A) the subgroup generated by P ∪ {1A}. Let κ : G → K0(C) be a positive homomorphism with κ([1A]) = [1C], where C = C([0, 1], Q), and let λ : T(C) → T(A) be a continuous affine map such that τ (κ([p]) − λ(τ )(p) < δ, p ∈ P, τ ∈ T(C). (4.7) (In particular, this entails that T(A) 6= Ø.) Then there is a unital F-ǫ-multiplicative completely positive map L : A → C such that τ ◦ L(a) − λ(τ )(a) < σ, a ∈ F, τ ∈ T(C). (4.8) Proof. Let ǫ, σ and F be given. We may assume that every element of F has norm at most one. Let δ1 (in place of δ), G, and P be as assured by Lemma 4.2 for F and ǫ. We may assume that F ∪ P ⊂ G. Adjoining 1A to P, write P = {1A, p1, p2, ..., ps}. Deleting one or more of p1, p2, ..., ps (but not 1A), we may assume that the set {[1A], [p1], ..., [ps]} is Q-linearly independent. Let δ2 > 0 (in place of δ) be as assured by Lemma 3.1 for ǫ1 = δ1, ǫ2 = σ/4, G, and {p1, p2, ..., ps}. Put δ = min{δ1, δ2/8, 1/4}, and let us show that P and δ are as desired. Let κ and λ be given satisfying (4.7). Let λ∗ : Aff(T(A)) → Aff(T(C)) be defined by λ∗(f )(τ ) = f (λ(τ )) for all f ∈ Aff(T(A)) and τ ∈ T(C). Identify ∂e(T(C)) with [0, 1], and put η = min{δ, σ/12}. Choose a partition 0 = t0 < t1 < t2 < · · · < tn−1 < tn = 1 of the interval [0, 1] such that λ∗(g)(tj) − λ∗(g)(tj−1) < η, g ∈ G, j = 1, 2, ..., n. (4.9) 11 Since T(A) = Tqd(A), there are unital G-δ-multiplicative completely positive maps Ψj : A → Q, j = 0, 1, 2, ..., n, such that tr ◦ Ψj(g) − λ∗(g)(tj) < η, g ∈ G. (4.10) It then follows from (2.1), (4.10), and (4.7) that, for each i = 1, 2, ..., s and each j = 1, 2, ..., n, tr([Ψj(pi)]) − tr([Ψ0(pi)]) < 4δ + 2η + λ∗( pi)(tj) − λ∗( pi)(t0) < 4δ + 2η + 2δ + tr ◦ πtj (κ([pi])) − tr ◦ π0(κ([pi]) = 2η + 6δ ≤ 8δ ≤ δ2. (4.11) (Here, as before, πt is the point evaluation at t ∈ [0, 1].) We also have, by (4.9) and (4.10), that tr(Ψj(g)) − tr(Ψj+1(g)) < 3η, g ∈ G, j = 1, 2, ..., n. (4.12) Consider the differences ri,j := tr([Ψj(pi)]) − tr([Ψ0(pi)]), i = 1, 2, ..., s, j = 1, 2, ..., n. (4.13) By (4.11), ri,j < δ2. Applying Lemma 3.1 we obtain a projection e ∈ Q with tr(e) < σ/4 and G-δ1-multiplicative unital completely positive maps ψ0, ψj : A → eQe, j = 1, 2, ..., n, such that [ψ0(pi)] − [ψj(pi)] = ri,j, i = 1, 2, ..., s, j = 1, 2, ..., n. (4.14) Consider the direct sum maps Φ′ j := ψj ⊕ Ψj : A → (1 ⊕ e)M2(Q)(1 ⊕ e), j = 0, 1, 2, ..., n. Since δ ≤ δ1, these are G-δ1-multiplicative. By (4.13) and (4.14), [Φ′ j(pi)] = [Φ′ 0(pi)], i = 1, 2, ..., s, j = 1, 2, ..., n. (4.15) Define s : Q → Q by s(x) = x/(1 + tr(e)), x ∈ Q. Choose a (unital) isomorphism S : (1 ⊕ e)M2(Q)(1 ⊕ e) → Q such that S∗0 = s. Consider the composed maps, still G-δ1-multiplicative, and now unital, By (4.15), Φj := S ◦ Φ′ j : A → Q, j = 0, 1, 2, ..., n. [Φj]P = [Φj−1]P , j = 1, 2, ..., n, and by (4.12) and the fact that tr(e) < σ/4, tr ◦ Φj(a) − tr ◦ Φj−1(a) < 3η + σ/4 ≤ σ/2, a ∈ F, j = 1, 2, ..., n. (4.16) It follows by Lemma 4.2, applied successively for j = 1, 2, ..., n (to the pairs (Φ0, Φ1), (Ad u1 ◦ Φ1, Ad u1 ◦ Φ2), ..., (Ad un−1 ◦ · · · ◦ Ad u1 ◦ Φn−1, Ad un−1 ◦ · · · ◦ Ad u1 ◦ Φn)), that there are, for each j, a unitary uj ∈ Q and a unital F-ǫ-multiplicative completely positive map Lj : A → C([tj−1, tj], Q) such that π0 ◦ L1 = Φ0, πt1 ◦ L1 = Ad u1 ◦ Φ1, (4.17) 12 and πtj−1 ◦ Lj = πtj−1 ◦ Lj−1, πtj ◦ Lj = Ad uj ◦ · · · ◦ Ad u1 ◦ Φj, j = 2, 3, ..., n. (4.18) Furthermore, in view of (4.16), we may choose the maps Lj such that tr ◦ πt ◦ Lj(a) − λ(tr ◦ πt)(a) < σ, t ∈ [tj−1, tj], a ∈ F, j = 1, 2, ..., n. (4.19) Define L : A → C([0, 1], Q) by πt ◦ L = πt ◦ Lj, t ∈ [tj−1, tj], j = 1, 2, ..., n. Since Lj, j = 1, 2, ..., n, are F-ǫ-multiplicative (use (4.17) and (4.18)), we have that L is a unital F-ǫ-multiplicative completely positive map A → C([0, 1], Q). (Note that the construction of the map L is different from-is based on-the construction of the map L in the proof of Lemma 4.2.) It follows from (4.19) that L satisfies (4.8), as desired. Theorem 4.4. Let A be a unital simple separable C-algebra with finite nuclear dimension. Assume that T(A) = Tqd(A) 6= Ø and that A satisfies the UCT. Then gTR(A ⊗ Q) ≤ 1, and so, (if A is not elementary), A ∈ N1. Proof. Since A is simple, the assumption T(A) = Tqd(A) 6= Ø immediately implies that A is both stably finite and quasidiagonal. We may assume that A ⊗ Q ∼= A. With this assumption, by [27], A has stable rank one. By [9] (see also Corollary 13.47 of [14]), together with the assumption A ∼= A ⊗ Q, there is a unital simple C-algebra C = limn→∞(Cn, ın), where each Cn is the tensor product of a C-algebra in C0 with Q and ın is injective, such that (K0(A), K0(A)+, [1A]0, T(A), rA) ∼= (K0(C), K0(C)+, [1C ]0, T(C), rC ). Choose an isomorphism Γ as above, and write ΓAff for the corresponding map from Aff(T(A)) to Aff(T(C)). Let a finite subset F of A and ǫ > 0 be given. Let the finite set P of projections in A, the finite subset G of A, and δ > 0 be as assured by Lemma 4.2 for F and ǫ. We may assume that 1A ∈ P. Write P = {1A, p1, p2, ..., ps}. Deleting some elements (but not 1A), we may assume that the set {[1A], [p1], [p2], ..., [ps]} ⊂ K0(A) is Q-linearly independent. We may also assume, without loss of generality, that F ∪ P ⊂ G, δ ≤ ǫ, and every element of G has norm at most one. Let σ > 0. Let δ1 > 0 (in place of δ) be as assured by Lemma 3.3 for G, δ (in place of ǫ1), and σ/64 (in place of ǫ2). We may assume that δ1 ≤ 8δ. Let δ3 > 0 (in place of δ) be as assured by Lemma 3.4 for G, δ1/8 (in place of ǫ1) and min{δ1/32, σ/256} (in place of ǫ2). Let P1 (in place of P) and δ2 > 0 (in place of δ) be as assured by Lemma 4.3 for δ1/8 (in place of ǫ), min{δ1/32, σ/256} (in place of σ), and G (in place of F). Replacing P and P1 by their union, we may assume that P = P1. By Lemma 2.7 of [11], there is a unital positive linear map γ : Aff(T(A)) → Aff(T(Cn1)) 13 for some n1 ≥ 1 such that k(ın1,∞)Aff ◦ γ( f ) − ΓAff ( f )k < min{77σ/128, δ2, δ3/2}, f ∈ F ∪ P. (4.20) We may assume, without loss of generality, that there are projections p′ s ∈ Cn1 such that Γ([pi]) = ın1,∞([p′ i]), i = 1, 2, ..., s. To simplify notation, assume that n1 = 1. Let G0 denote the subgroup of K0(A) generated by P. Since G0 is free abelian, there is a homomorphism Γ′ : G0 → K0(C1) such that 1, p′ 2, ..., p′ (ı1,∞)∗0 ◦ Γ′ = ΓG0. We may assume (since P is a basis for G0) that Γ′([pi]) = [p′ i], i = 1, 2, ..., s. (4.21) Since the pair (ΓAff , ΓK0(A)) is compatible, as a consequence of (4.20) and (4.21) we have Write kbp′ i − γ(bpi)k∞ < min{δ2, δ3/2}, i = 1, 2, ...., s. (4.22) C1 = (ψ0, ψ1, Qr, Ql) = {(f, a) ∈ C([0, 1], Qr) ⊕ Ql : f (0) = ψ0(a) and f (1) = ψ1(a)}, where ψ0, ψ1 : Ql → Qr are unital homomorphisms. Denote by πe : C1 → Ql, (f, a) 7→ a the canonical quotient map, and by j : C1 → C([0, 1], Qr) the canonical map j((f, a)) = f, (f, a) ∈ C1 ⊗ Q. Denote by γ∗ : T(C1) → T(A) the continuous affine map dual to γ. Denote by θ1, θ2, ..., θl the extreme tracial states of C1 factoring through πe : C1 → Ql. By the assumption T(A) = Tqd(A), there exists a unital G-min{δ1/8, δ3/8}-multiplicative completely positive map Φ : A → Ql such that trj ◦ Φ(a) − γ∗(θj)(a) < min{13δ1/32, δ3/4, σ/32}, a ∈ G, j = 1, 2, ..., l, (4.23) where trj is the tracial state supported on the jth direct summand of Ql. Moreover, since P ⊂ G, as in (2.1), we also have that trj([Φ(pi)]) − trj(Φ(pi)) < δ3/4, i = 1, 2, ..., s, j = 1, 2, ..., l. (4.24) Set D := (πe)∗0(K0(C1)) = ker((ψ0)∗0 − (ψ1)∗0) ⊆ Ql. It follows from (4.23) that τ (Φ(a)) − (πe)Aff (γ(a))(τ ) < min{13δ1/32, δ3/4, σ/32}, a ∈ G, τ ∈ T(Ql), (4.25) where (πe)Aff : Aff(T(C1)) → Aff(T(Ql)) is the map induced by πe. By (4.25) for a ∈ P ⊂ G, together with (4.24) and (4.22), τ ([Φ(pi)]) − τ ◦ (πe)∗0 ◦ Γ′([pi]) < δ3, τ ∈ T(Ql), i = 1, 2, ..., s. 14 Therefore, applying Lemma 3.4, with ri = [Φ(pi)] − (πe)∗0 ◦ Γ′([pi]), we obtain G-δ1/8- multiplicative completely positive maps Σ1, µ1 : A → Ql, with Σ1(1A) = µ1(1A) a projection, such that τ (Σ1(1A)) < min{δ1/32, σ/256}, τ ∈ T(Ql), [Σ1(P)] ⊆ D, and [Σ1(pi)] − [µ1(pi)] = [Φ(pi)] − (πe)∗0 ◦ Γ′([pi]), i = 1, 2, ..., s. Consider the (unital) direct sum map Φ′ := Φ ⊕ µ1 : A → (1 ⊕ Σ1(1A))M2(Ql)(1 ⊕ Σ1(1A)). (4.26) (4.27) (4.28) (4.29) Note that Φ′, like µ1 and Φ, is G-δ1/8-multiplicative. i = 1, 2, ..., s, It follows from (4.28) that for each [ψ0(Φ′(pi))] = (ψ0)∗0([µ1(pi)] + [Φ(pi)]) = (ψ0)∗0([Σ1(pi)] + (πe)∗0 ◦ Γ′([pi])) (4.30) and [ψ1(Φ′(pi))] = (ψ1)∗0([µ1(pi)] + [Φ(pi)]) = (ψ1)∗0([Σ1(pi)] + (πe)∗0 ◦ Γ′([pi])). (4.31) It follows from (4.30) and (4.31), in view of (4.27) and the fact (use (4.21)) that (πe)∗0◦Γ′([pi])) ∈ (πe)∗0(K0(C1)) = D, that [Φ′(pi)] ∈ D, i = 1, 2, ..., s, i.e., [ψ0(Φ′(pi))] = [ψ1(Φ′(pi))], i = 1, 2, ..., s. (4.32) Set B = C([0, 1], Qr), and (as before) write πt : B → Qr for the point evaluation at t ∈ [0, 1]. Since 1A ∈ P, by (4.27), [Σ1(1A)] ∈ D, and so there is a projection e0 ∈ B such that π0(e0) = ψ0(Σ1(1A)) and π1(e0) = ψ1(Σ1(1A)). It then follows from (4.26) (applied just for τ factoring through ψ0-alternatively, for τ factoring through ψ1) that τ (e0) < min{δ1/32, σ/256}, τ ∈ T(B). (4.33) Let j∗ : T(B) → T(C1) denote the continuous affine map dual to the canonical unital map j : C1 → B. Let γ1 : T(B) → T(A) be defined by γ1 := γ∗ ◦ j∗, and let κ : G0 → K0(B) be defined by κ := j∗0 ◦ Γ′. Then, by (4.21) and (4.22), τ (κ([pi]) − γ1(τ )(pi) = τ (j∗0(Γ′([pi])) − (γ∗ ◦ j∗)(τ )(pi) for all τ ∈ T(B), i = 1, 2, ..., k. = j∗(τ )([p′ i]) − γ(bpi)(j∗(τ )) < δ2 (4.34) The estimate (4.34) ensures that we can apply Lemma 4.3 with κ and γ1 (note that Γ′([1A]) = [1C1 ] and hence κ([1A]) = [1B]), to obtain a unital G-δ1/8-multiplicative completely positive map Ψ′ : A → B such that τ ◦ Ψ′(a) − γ1(τ )(a) < min{δ1/32, σ/256}, a ∈ G, τ ∈ T(B). (4.35) Amplifying Ψ′ slightly (by first identifying Qr with Qr ⊗ Q and then considering H0(f )(t) = f (t) ⊗ (1 + e0(t)) for t ∈ [0, 1]), we obtain a unital G-δ1/8-multiplicative completely positive map Ψ : A → (1 ⊕ e0)M2(B)(1 ⊕ e0) such that (by (4.35) and (4.33)) τ ◦ Ψ(a) − γ1(τ )(a) < 2 min{δ1/32, σ/256} = min{δ1/16, σ/128}. (4.36) 15 Note that for any element a ∈ C1, τ (ψ0(πe(a))) = τ (π0(j(a))) and τ (ψ1(πe(a))) = τ (π1(j(a))), τ ∈ T(Qr). (4.37) (Recall that j : C1 → B is the canonical map.) Therefore (by (4.37)), for any a ∈ G and τ ∈ T(Qr), τ (ψ0(Φ(a))) − γ(a)(τ ◦ π0 ◦ j) = τ (ψ0(Φ(a))) − γ(a)(τ ◦ ψ0 ◦ πe) = τ ◦ ψ0(Φ(a))) − (πe)Aff (γ(a))(τ ◦ ψ0) < min{13δ1/32, σ/32} (by (4.25)). (4.38) The same argument shows that τ (ψ1(Φ(a))) − γ(a)(τ ◦ π1 ◦ j) < min{13δ1/32, σ/32}, a ∈ G, τ ∈ T(Qr). (4.39) Then, for any τ ∈ T(Qr) and any a ∈ G, we have τ ◦ ψ0 ◦ Φ′(a) − τ ◦ π0 ◦ Ψ(a) = τ ◦ ψ0(Φ(a) ⊕ µ1(a)) − τ ◦ π0 ◦ Ψ(a) < τ ◦ ψ0(Φ(a) ⊕ µ1(a)) − γ1(τ ◦ π0)(a) + min{δ1/16, σ/128} < τ ◦ ψ0(Φ(a)) − γ1(τ ◦ π0)(a) + min{3δ1/32, 3σ/256} = τ ◦ ψ0(Φ(a)) − γ(a)(τ ◦ π0 ◦ j) + min{3δ1/32, 3σ/256} < min{13δ1/32, σ/32} + min{3δ1/32, 3σ/256} ≤ 13δ1/32 + 3δ1/32 = δ1/2 (by (4.36)) (by (4.26)) (4.40) (by (4.38)). The same argument, using (4.39) instead of (4.38), shows that τ ◦ ψ1 ◦ Φ′(a) − τ ◦ π1 ◦ Ψ(a) < min{13δ1/32, σ/32} + min{3δ1/32, 3σ/256} (4.41) ≤ δ1/2, τ ∈ T(Qr), a ∈ G. ((4.40) and (4.41)-the σ estimates-will be used later to verify (4.49) and (4.50).) Noting that Ψ and Φ′ are δ1/8-multiplicative on {1A, p1, p2, ..., ps}, by our convention (see (2.1)), we have, for all τ ∈ T(Qr), τ ([Ψ(pi)]) − τ (Ψ(pi)) < δ1/4 and τ ([Φ′(pi)]) − τ (Φ′(pi)) < δ1/4, i = 1, 2, ..., s. Combining these inequalities with (4.40) and (4.41), we have τ ([π0 ◦ Ψ(pi)]) − τ ([ψ0 ◦ Φ′(pi)]) < δ1, i = 1, 2, ..., s, τ ∈ T(Qr). (4.42) Therefore (in view of (4.42)), applying Lemma 3.3 with ri = [π0 ◦ Ψ(pi)] − [ψ0 ◦ Φ′(pi)], we obtain G-δ-multiplicative completely positive maps Σ2 : A → Ql and µ2 : A → Qr, taking 1A into projections, such that [ψk ◦ Σ2(1A)] = [µ2(1A)], k = 0, 1, [Σ2(P)] ⊆ (πe)∗0(K0(C1)) = D, i = 1, 2, ..., s, τ (Σ2(1A)) < σ/64, τ ∈ T(Ql), (4.43) (4.44) (4.45) and, taking into account (4.44), [ψ0 ◦ Σ2(pi)] − [µ2(pi)] = [ψ1 ◦ Σ2(pi)] − [µ2(pi)] = [π0 ◦ Ψ(pi)] − [ψ0 ◦ Φ′(pi)], (4.46) 16 where i = 1, 2, ..., s. Consider the four G-δ-multiplicative direct sum maps (note that Φ′ and Ψ are G-δ1/8- multiplicative, and δ1 ≤ 8δ), from A to M3(Qr), Φ0 := (ψ0 ◦ Φ′) ⊕ (ψ0 ◦ Σ2), Φ1 = (ψ1 ◦ Φ′) ⊕ (ψ1 ◦ Σ2) and Ψ0 := (π0 ◦ Ψ) ⊕ µ2, Ψ1 := (π1 ◦ Ψ) ⊕ µ2. (4.47) We then have that for each i = 1, 2, ..., s, [Ψ0(pi)] − [Φ0(pi)] = ([(π0 ◦ Ψ)(pi)] + [µ2(pi)]) − ([(ψ0 ◦ Φ′)(pi)] + [(ψ0 ◦ Σ2)(pi)]) = ([(π0 ◦ Ψ)(pi)] − [(ψ0 ◦ Φ′)(pi)]) − ([(ψ0 ◦ Σ2)(pi)] − [µ2(pi)]) = 0 (by (4.46)), and [Ψ1(pi)] − [Φ1(pi)] = ([(π1 ◦ Ψ)(pi)] + [µ2(pi)]) − ([(ψ1 ◦ Φ′)(pi)] + [(ψ1 ◦ Σ2)(pi)]) = ([(π1 ◦ Ψ)(pi)] − [(ψ1 ◦ Φ′)(pi)]) − ([(ψ1 ◦ Σ2)(pi)] − [µ2(pi)]) = ([(π0 ◦ Ψ)(pi)] − [(ψ1 ◦ Φ′)(pi)]) − ([(ψ1 ◦ Σ2)(pi)] − [µ2(pi)]) = ([(π0 ◦ Ψ)(pi)] − [(ψ0 ◦ Φ′)(pi)]) − ([(ψ1 ◦ Σ2)(pi)] − [µ2(pi)]) = 0 (by (4.32)) = 0 (by (4.46)). (π0 and π1 are homotopic) Note also that, by construction, Ψi(1A) = Φi(1A) = 1 ⊕ πi(e0), i = 0, 1. Summarizing the calculations in the preceding paragraph, we have On the other hand, for any a ∈ F ⊆ G and any τ ∈ T(Qr), we have [Φi]P = [Ψi]P , i = 0, 1. (4.48) τ (Φ0(a)) − τ (Ψ0(a)) = τ ((ψ0 ◦ Φ′)(a) ⊕ (ψ0 ◦ Σ2)(a)) − τ ((π0 ◦ Ψ)(a) ⊕ µ2(a)) (by (4.45)) < τ ((ψ0 ◦ Φ′)(a)) − τ ((π0 ◦ Ψ)(a)) + σ/32 < min{13δ1/16, σ/32} + min{3δ1/16, 3σ/256} + σ/32 ≤ 5σ/64. (by (4.40)) (4.49) The same argument, using (4.41) instead of (4.40), also shows that τ (Φ1(a)) − τ (Ψ1(a)) < 5σ/64, a ∈ F, τ ∈ T(Qr). (4.50) Since 1A ∈ P, by (4.44), [Σ2(1A)] ∈ D, and so there is a projection e1 ∈ B such that π0(e1) = ψ0(Σ2(1A)) and π1(e1) = ψ1(Σ2(1A)). It then follows from (4.45) (applied just for τ factoring through ψ0-alternatively, for τ factoring through ψ1) that τ (e1) < σ/64, τ ∈ T(B). (4.51) 0 = 1 ⊕ π0(e0) ⊕ π0(e1), E′ 1 = 1 ⊕ π1(e0) ⊕ π1(e1), and D0 = E′ 0M2(Qr)E′ 0, D1 = Set E′ E′ 1M2(Qr)E′ 1. Pick a sufficiently small r′ ∈ (0, 1/4) that kΨ(a)((1 + 2r′)t − r′) − Ψ(a)(t)k < σ/64, a ∈ G, t ∈ [ r′ 1 + 2r′ , 1 + r′ 1 + 2r′ ]. (4.52) 17 L′(a)(t) =   L0(a)(t), Ad u(t) ◦ (πt ◦ Ψ ⊕ µ2)(a), L1(a)(t) t ∈ [−r′, 0), t ∈ [0, 1], t ∈ (1, 1 + r′]. (4.57) It follows from Lemma 4.2 (in view of (4.48), (4.49) and (4.50)) that there exist unitaries u0 ∈ D0 and u1 ∈ D1, and unital F-ǫ-multiplicative completely positive maps L0 : A → C([−r′, 0], D0) and L1 : A → C([1, 1 + r′], D1), such that π−r′ ◦ L0 = Φ0, π0 ◦ L0 = Ad u0 ◦ Ψ0, π1+r′ ◦ L1 = Φ1, π1 ◦ L1 = Ad u1 ◦ Ψ1, τ ◦ πt ◦ L0(a) − τ ◦ π0 ◦ L0(a) < 5σ/32, τ ◦ πt ◦ L1(a) − τ ◦ π1 ◦ L1(a) < 5σ/32, t ∈ [−r′, 0], t ∈ [1, 1 + r′], (4.53) (4.54) (4.55) (4.56) where a ∈ F, τ ∈ T(Qr), and (as before) πt is the point evaluation at t ∈ [−r′, 1 + r′]. Write E3 = 1 ⊕ e0 ⊕ e1 ∈ M3(C([0, 1], Qr)) and B1 = E3(M3(C([0, 1], Qr)))E3. There exists a unitary u ∈ B1 such that u(0) = u0 and u(1) = u1. Consider the projection E4 ∈ M3(C([−r′, 1 + r′], Qr)) defined by E4[−r,0] = E′ 0, E4[0,1] = E3 and E4[1,1+r] = E′ 1. Set B2 = E4(M3(C([−r′, 1 + r′], Qr)))E4. Define a unital F-ǫ-multiplicative (note that F ⊂ G and δ ≤ ǫ) completely positive map L′ : A → B2 by Note that for any a ∈ G, and any τ ∈ T(Qr), by (4.57), if t ∈ [0, 1], then τ (πt(L′(a))) − γ1(π∗ t (τ ))(a) = τ (Ad u(t) ◦ (πt◦Ψ ⊕ µ2)(a)) − γ1(π∗ = τ (πt(Ψ(a))) + τ (µ2(a)) − γ1(π∗ < (π∗ < min{δ1/16, σ/128} + σ/64 ≤ 3σ/128 t (τ ))(a) t (τ ))(a) + σ/64 t (τ ))(Ψ(a))) − γ1(π∗ t (τ ))(a) (by (4.43) and (4.45)) (by (4.36)), (4.58) where π∗ a ∈ F, and any τ ∈ T(Qr), t : T(Qr) → T(B) is the dual of πt : B → Qr. Furthermore, if t ∈ [−r′, 0], then for any τ (πt(L′(a))) − γ1(π∗ 0(τ ))(a) 0(τ ))(a) 0(τ ))(a) + 5σ/32 = τ (L0(a)(t)) − γ1(π∗ < τ (L0(a)(0)) − γ1(π∗ = τ (Ψ0(a)) − γ1(π∗ = τ ((π0 ◦ Ψ)(a) ⊕ µ2(a)) − γ1(π∗ < τ ((π0 ◦ Ψ)(a) − γ1(π∗ 0(τ ))(a) + σ/64 + 5σ/32 < min{δ1/16, σ/128} + σ/64 + 5σ/32 < 23σ/128 0(τ ))(a) + 5σ/32 0(τ ))(a) + 5σ/32 (by (4.55)) (by (4.53)) (by (4.43) and (4.45)) (by (4.36)). (4.59) Again, if t ∈ [1, 1 + r′], then the same argument shows that for any a ∈ F, and any τ ∈ T(Qr), τ (πt(L′(a))) − γ1(π∗ 1(τ ))(a) < 23σ/128. (4.60) Let us modify L′ to a unital map from A to B. First, let us renormalize L′. Consider the isomorphism η : Qr → Qr defined by η(x1, x2, ..., xr) = ( 1 tr2(E3) x2, ..., 1 trr(E3) xr), 1 tr1(E3) x1, 18 for all (x1, x2, ..., xr) ∈ Qr, where (as before) trk is the tracial state supported on the kth direct summand of Qr. Then there is a (unital) isomorphism ϕ : B2 → C([−r′, 1 + r′], Qr) such that ϕ∗0 = η. Let us replace the map L′ by the map ϕ ◦ L′, and still denote it by L′. Note that it follows from (4.58), (4.33), and (4.51) that for any t ∈ [0, 1], any a ∈ F, and any τ ∈ T(Qr), τ (πt(L′(a))) − γ1(π∗ < σ/16 + τ (e0) + τ (e1) < 3σ/128 + min{δ1/16, σ/64} + σ/64 ≤ 7σ/128. t (τ ))(a) The same argument, using (4.59) and (4.60) instead of (4.58), shows that for any a ∈ F, τ (πt(L′(a))) − γ1(π∗ τ (πt(L′(a))) − γ1(π∗ 0(τ ))(a) < 27σ/128, 1(τ ))(a) < 27σ/128, t ∈ [−r′, 0], t ∈ [1, 1 + r′]. Now, put L′′(a)(t) = L′(a)((1 + 2r′)t − r′), t ∈ [0, 1]. (4.61) (4.62) (4.63) (4.64) This perturbation will not change the trace very much, as for any a ∈ F and any τ ∈ T(Qr), if t ∈ [0, r′/(1 + 2r′)], then τ (L′′(a)(t)) − τ (L′(a)(t)) = τ (L′(a)((1 + 2r′)t − r′)) − τ (L′(a)(t)) = τ (L0(a)((1 + 2r′)t − r′)) − (τ (Ψ(a)(t)) + µ2(a)) < τ (L0(a)((1 + 2r′)t − r′)) − τ (Ψ(a)(t)) + σ/64 < τ (L0(a)(0)) − τ (Ψ(a)(t)) + 5σ/32 + σ/64 = τ (Ψ0(a)(0)) − τ (Ψ(a)(t)) + 11σ/64 < σ/64 + 11σ/64 = 3σ/16 (by (4.64)) (by (4.43)) and (4.45) (by (4.55)) (by (4.53)) (by (4.43)) and (4.45). Furthermore, the same argument, now using (4.56) and (4.54), shows that for any a ∈ F, τ ∈ T(Qr), and t ∈ [(1 + r′)/(1 + 2r′), 1], τ (L0(a)((1 + 2r′)t − r′)) − (τ (Ψ(a)(t)) + µ2(a)) < 3σ/16, and, if t ∈ [r′/(1 + 2r′), (1 + r′)/(1 + 2r′)], then τ (L′(a)((1 + 2r′)t − r′)) − τ (L′(a)(t)) = τ (Ψ(a)((1 + 2r′)t − r′) − Ψ(a)(t)) < σ/64 (by (4.52)). Thus, Hence, τ (L′′(a)(t)) − τ (L′(a)(t)) < 3σ/16, a ∈ F, τ ∈ T(Qr), t ∈ [0, 1]. (4.65) τ (πt(L′′(a))) − γ1(π∗ t (τ ))(a) ≤ τ (L′′(a)(t)) − τ (L′(a)(t)) + τ (L′(a)(t)) − γ1(π∗ < 3σ/16 + 27σ/128 = 51σ/128 t (τ ))(a) (by (4.65), (4.61), (4.62), and (4.63)). (4.66) Note that L′′ is a unital map from A to B. It is also F-ǫ-multiplicative, since L′ is. Consider the order isomorphism η′ : Ql → Ql defined by η′(y1, y2, ..., yl) = (a1y1, a2y2, ..., alyl), (y1, y2, ..., yl) ∈ Ql, 19 where aj = 1 trj(1 ⊕ Σ1(1A) ⊕ Σ2(1A)) , j = 1, 2, ..., l, (4.67) and (as before) trj is the tracial state supported on the jth direct summand of Ql. There exists a unital homomorphism ϕ : (1 ⊕ Σ1(1A) ⊕ Σ2(1A))M3(Ql)(1 ⊕ Σ1(1A) ⊕ Σ2(1A)) → Ql such that ϕ∗0 = η′. Therefore, by the constructions of L′′, L′, L0, and L1 ((4.64), (4.57), (4.53), and (4.54)), we may assume that ψ0 ◦ ϕ ◦ (Φ′ ⊕ Σ2) = π0 ◦ L′′ and ψ1 ◦ ϕ ◦ (Φ′ ⊕ Σ2) = π1 ◦ L′′, (4.68) replacing L′′ by Ad w ◦ L′′ for a suitable unitary w if necessary. Define L : A → C1 by L(a) = (L′′(a), ϕ(Φ′(a) ⊕ Σ2(a))), an element of C1 by (4.68). Since L′′ and ϕ ◦ (Φ′ ⊕ Σ2) are unital and F-ǫ-multiplicative (since Φ′ and Σ2 are G-δ-multiplicative, F ⊂ G, and δ ≤ ǫ), so also is L. Moreover, for any a ∈ F, any τ ∈ T(Qr), and any t ∈ (0, 1), it follows from (4.66) that τ (πt(L(a))) − γ1(π∗ t (τ ))(a) < 51σ/128. (4.69) If τ ∈ T(Ql), then, for any a ∈ F, τ (πe(L(a))) − γ∗(π∗ e (τ ))(a) e (τ ))(a) e (τ ))(a) + σ/32 = τ ( ϕ(Φ′(a) ⊕ Σ2(a))) − γ∗(π∗ < τ (Φ′(a) ⊕ Σ2(a)) − γ∗(π∗ < τ (Φ′(a)) − γ∗(π∗ < τ (Φ(a)) − γ∗(π∗ < σ/32 + σ/8 < 51σ/128 e (τ ))(a) + 3σ/64 e (τ ))(a) + σ/8 (by (4.67), (4.33) and (4.45)) (by (4.45)) (by (4.26)) (by (4.25)). Since each extreme trace of C1 factors through either the evaluation map πt or the canonical quotient map πe, by (4.69), τ (L(a)) − γ∗(τ )(a) < 51σ/128, τ ∈ T(C1), a ∈ F. (4.70) Therefore, for any a ∈ F and τ ∈ T(C), we have τ (ı1,∞(L(a))) − ΓAff (a)(τ ) < τ (ı1,∞(L(a))) − γ∗(ı1,∞(τ ))(a) + γ∗(ı1,∞(τ ))(a) − ΓAff (a)(τ ) = τ (ı1,∞(L(a))) − γ∗(ı1,∞(τ ))(a) + (ı1,∞)Aff (a)(τ ) − ΓAff (a)(τ ) < 51σ/128 + 77σ/128 = σ. (by (4.70) and (4.20)) Since F, ǫ, and σ are arbitrary, in this way we obtain a sequence of unital completely positive maps Hn : A → C such that lim n→∞ kHn(ab) − Hn(a)Hn(b)k = 0, a, b ∈ A, and lim n→∞ sup{τ ◦ Hn(a) − ΓAff (a)(τ ) : τ ∈ T(C)} = 0, a ∈ A. It follows from Lemma 3.4 of [22] (which used results obtained in [25] and [32]) that gTR(A⊗Q) ≤ 1. Since A ∼= A ⊗ Q, A is Z-stable and therefore it is in N1. 20 Theorem 1.1 follows from the following corollary. Corollary 4.5. Let A be a unital simple separable C-algebra with finite decomposition rank, satisfying the UCT. Then gTR(A ⊗ Q) ≤ 1. In particular, A is classifiable. Proof. Since A has finite decomposition rank, A is nuclear (see 2.10) and quasidiagonal (5.3 of [19]). It follows from 2.4 of [30] that T(A) 6= Ø. By 8.5 of [3], T(A) = Tqd(A). Now the corollary follows from Theorem 4.4 together with Theorem 2.7. Remark 4.6. It was shown in [12] that any unital simple separable Jiang-Su stable approxi- mately subhomogeneous C-algebra has decomposition rank at most 2. Therefore, it follows from Corollary 4.5 that such a C-algebra is classifiable. This in particular recovers the classification theorem of [10]. Remark 4.7. The special case of Corollary 4.5 for C-algebras with unique trace follows from the results of the earlier papers [24] and [23]. Theorem 4.4 and Corollary 4.5 can also be combined and stated as follows: Theorem 4.8. Let A be a unital simple separable amenable (non-zero) C-algebra which satisfies the UCT. Then the following properties for A ⊗ Q are equivalent: (1) gTR(A ⊗ Q) ≤ 1; (2) A ⊗ Q has finite nuclear dimension and T(A ⊗ Q) = Tqd(A ⊗ Q) 6= Ø; (3) the decomposition rank of A ⊗ Q is finite. Given the fact that every tracial state of a unital simple separable C-algebra with finite decomposition rank is quasidiagonal (8.5 of [3]), it is reasonable to expect that every tracial state of a finite unital simple separable C-algebra with finite nuclear dimension is also quasidiagonal. Indeed, shortly after the present paper was first announced (and posted on arXiv), Tikuisis, White, and Winter proved that, in fact, every tracial state on a unital simple separable amenable C-algebra which satisfies the UCT is quasidiagonal ([29]). Therefore, we have the following statement: Theorem 4.9. Let A be a finite unital simple separable C-algebra with finite nuclear dimension which satisfies the UCT. Then gTR(A ⊗ Q) ≤ 1. In particular, A is classifiable. Remark 4.10. It was established by Kirchberg and Phillips ([18] and [26]) that purely infinite unital simple separable amenable C-algebras which satisfy the UCT are classifiable. It has been shown that these C-algebras have finite nuclear dimension (see [24]). It is also known that every unital simple separable C-algebra with finite nuclear dimension is either finite or purely infinite (see [13] and [31]). Therefore, Theorem 4.9 can now be combined with [18] and [26] to obtain the following overall statement. Corollary 4.11. The class of all unital simple separable (non-elementary) C-algebras with finite nuclear dimension which satisfy the UCT is classifiable by the Elliott invariant. Remark 4.12. It remains open whether every separable amenable C-algebra satisfies the UCT. Bypassing the UCT, we would propose the following questions: (i) Do Theorem 4.8 and Corollary 4.11 hold without assuming the UCT? (ii) Let A be a stably finite unital simple separable amenable C-algebra. Is it true that gTR(A ⊗ Q) ≤ 1? Questions (i) and (ii) are perhaps quite ambitious at the moment. A more realistic question might be the following: (iii) Let A be a stably finite unital simple separable amenable C-algebra. Is it true that if A is quasidiagonal then gTR(A ⊗ Q) ≤ 1? (Note that the converse holds (see 9.8 and 9.9 of [14]).) 21 References [1] B. Blackadar and E. Kirchberg, Generalized inductive limits of finite-dimensional C- algebras, Math. Ann. 307 (1997), 343-380. [2] B. Blackadar and E. Kirchberg, Inner quasidiagonality and strong NF algebras, Pacific J. Math. 198 (2001), 307–329. [3] J. Bosa, N. P. Brown, Y. Sato, A. Tikuisis, S. White, and W. Winter, Covering dimension of C-algebras and 2-colored classification, preprint, arXiv:1506.03974. [4] N. P. Brown, Invariant means and finite representation theory of C ∗-algebras, Mem. Amer. Math. Soc. 184 (2006), no. 865, viii+105 pp. [5] M.-D. Choi and E. G. Effros, Injectivity and operator spaces, J. Funct. Anal. 24 (1977), 156–209. [6] A. Connes, On the cohomology of operator algebras, J. Funct. Anal. 28 (1978), 248–253. [7] M. Dadarlat and S. Eilers. On the classification of nuclear C-algebras, Proc. London Math. Soc. (3) 85 (2002), 168–210. [8] M. Dadarlat and T. Loring, A universal multicoefficient theorem for the Kasparov groups, Duke Math. J. 84 (1996), 355–377. [9] G. A. Elliott, An invariant for simple C-algebras, Canadian Mathematical Society, 1945– 1995, Vol. 3, 61–90, Canad. Math. Soc., Ottawa, 1996. [10] G. A. Elliott, G. Gong, H. Lin, and Z. Niu, The classification of simple separable unital locally ASH algebras, preprint, arXiv:1506.02308. [11] G. A. Elliott and Z. Niu, On the classification of simple amenable C-algebras with finite decomposition rank, in "Operator Algebras and their Applications: A Tribute to Richard V. Kadison", Contemporary Mathematics, Amer. Math. Soc., Providence, R. I., 2016, in press. arXiv:1507.07876. [12] G. A. Elliott, Z. Niu, L. Santiago, and A. Tikuisis, Decomposition rank of approximately subhomogeneous C-algebras, preprint, arXiv:1505.06100. [13] G. Gong, X. Jiang, and H. Su, Obstructions to Z-stability for unital simple C-algebras, Canad. Math. Bull. 43 (2000), 418–426. [14] G. Gong, H. Lin, and Z. Niu, Classification of finite simple amenable Z-stable C ∗-algebras, preprint, arXiv:1501.00135. [15] G. A. Elliott and A. Toms, Regularity properties in the classification program for separable amenable C-algebras, Bull. Amer. Math. Soc. (N.S.) 45 (2008), 229–245. [16] U. Haagerup, All nuclear C-algebras are amenable, Invent. Math. 74 (1983), 305–319. [17] X. Jiang and H. Su, On a simple unital projectionless C ∗-algebra, Amer. J. Math. 121 (1999), 359–413. [18] E. Kirchberg and N. C. Phillips, Embedding of exact C-algebras in the Cuntz algebra O2, J. Reine Angew. Math. 525 (2000), 17–53. 22 [19] E. Kirchberg and W. Winter, Covering dimension and quasidiagonality, Internat. J. Math. 15 (2004), 63–85. [20] H. Lin, Classification of simple tracially AF C-algebras, Canad. J. Math. 53 (2001), 161– 194. [21] H. Lin, An approximate universal coefficient theorem, Trans. Amer. Math. Soc. 357 (2005), 3375–3405. [22] H. Lin, Crossed products and minimal dynamical systems, preprint, arXiv:1502.06658. [23] H. Lin and Z. Niu, Lifting KK-elements, asymptotic unitary equivalence and classification of simple C-algebras, Adv. Math. 219 (2008), 1729–1769. [24] H. Matui and Y. Sato, Decomposition rank of UHF-absorbing C ∗-algebras, Duke Math. J. 163 (2014), 2687–2708. [25] L. Robert, Classification of inductive limits of 1-dimensional NCCW complexes, Adv. Math. 231 (2012), 2802–2836. [26] N. C. Phillips, A classification theorem for nuclear purely infinite simple C-algebras, Doc. Math. 5 (2000), 49–114 (electronic). [27] M. Rørdam, On the structure of simple C-algebras tensored with a UHF-algebra. II, J. Funct. Anal. 107 (1992), 255–269. [28] J. Rosenberg and C. Schochet, The Kunneth theorem and the universal coefficient theorem for Kasparov's generalized KK-functor, Duke Math. J. 55 (1987), 431–474. [29] A Tikuisis, S. White, and W. Winter, Quasidiagonality of nuclear C-algebras, preprint, arXiv:1509.08318. [30] D. Voiculescu, Around quasidiagonal operators, Integr. Equ. Oper. Theory 17 (1993), 137– 149. [31] W. Winter, Nuclear dimension and Z-stability of pure C-algebras, Invent. Math. 187 (2012), 259–342. [32] W. Winter, Classifying crossed product C-algebras, preprint, arXiv:1308.5084. [33] W. Winter and J. Zacharias, The nuclear dimension of C*-algebras, Adv. Math. 224 (2010), 461–498. Department of Mathematics, University of Toronto, Toronto, Ontario, Canada M5S 2E4 E-mail address: [email protected] Department of Mathematics, University of Puerto Rico, San Juan, PR 00936, USA E-mail address: [email protected] Department of Mathematics, University of Oregon, Eugene, OR 97403, USA E-mail address: [email protected] Department of Mathematics, University of Wyoming, Laramie, WY 82071, USA E-mail address: [email protected] 23
1207.2540	1	1207	2012-07-11T04:19:59	The equivalence relations of local homeomorphisms and Fell algebras	[ "math.OA" ]	We study the groupoid C-algebras associated to the equivalence relation induced by a quotient map on a locally compact Hausdorff space. This C-algebra is always a Fell algebra, and if the quotient space is Hausdorff, it is a continuous-trace algebra. We show that the C-algebra of a locally compact, Hausdorff and principal groupoid is a Fell algebra if and only if the groupoid is one of these relations, extending a theorem of Archbold and Somerset about etale groupoids. The C-algebras of these relations are, up to Morita equivalence, precisely the Fell algebras with trivial Dixmier-Douady invariant as recently defined by an Huef, Kumjian and Sims. We use twisted groupoid algebras to provide examples of Fell algebras with non-trivial Dixmier-Douady invariant.	math.OA	math	THE EQUIVALENCE RELATIONS OF LOCAL HOMEOMORPHISMS AND FELL ALGEBRAS LISA ORLOFF CLARK, ASTRID AN HUEF, AND IAIN RAEBURN Abstract. We study the groupoid C∗-algebras C∗(R(ψ)) associated to the equivalence relation R(ψ) induced by a quotient map ψ : Y → X. If Y is Hausdorff then C∗(R(ψ)) is a Fell algebra, and if both Y and X are Hausdorff then C∗(R(ψ)) has continuous trace. We show that the C∗-algebra C∗(G) of a locally compact, Hausdorff and principal groupoid G is a Fell algebra if and only if G is topologically isomorphic to some R(ψ), extending a theorem of Archbold and Somerset. The algebras C∗(R(ψ)) are, up to Morita equivalence, precisely the Fell algebras with trivial Dixmier-Douady invariant as recently defined by an Huef, Kumjian and Sims. We use a twisted analogue of C∗(R(ψ)) to provide examples of Fell algebras with non-trivial Dixmier-Douady invariant. 1. Introduction Important classes of type I C∗-algebras include the continuous-trace algebras and the liminary or CCR algebras, and lately there has been renewed interest in a family of liminary algebras called Fell algebras [12, 5, 1, 13]. A C∗-algebra is a Fell algebra if for each π ∈ A, there exists a positive element a such that ρ(a) is a rank-one projection for all ρ in a neighbourhood of π in A. Fell algebras were named by Archbold and Somerset [2] in respect of Fell's contributions [9], and had been previously studied by Pedersen as "algebras of type I0" [25, §6]. Proposition 4.5.4 of [7] says that a C∗-algebra has continuous trace if and only if it is a Fell algebra and its spectrum is Hausdorff. The Dixmier-Douady class δDD(A) of a continuous-trace algebra A identifies A up to Morita equivalence, and δDD(A) = 0 if and only if A is Morita equivalent to a commutative C∗-algebra [29, Theorem 5.29]. An Huef, Kumjian and Sims have recently developed an analogue of the Dixmier-Douady classification for Fell algebras [13]. Their Dixmier- Douady invariant δ(A) vanishes if and only if A is Morita equivalent to the groupoid C∗-algebra C∗(R(ψ)) of the equivalence relation associated to a local homeomorphism ψ of a Hausdorff space onto the spectrum A. (Since this theorem was not explicitly stated in [13], we prove it here as Theorem 6.1.) So the theory in [13] identifies the C∗-algebras C∗(R(ψ)) as an interesting family of model algebras. Here we investigate the structure of the algebras C∗(R(ψ)) and their twisted analogues, and use them to provide examples of Fell algebras exhibiting certain kinds of behaviour. In particular, we will produce some concrete examples of Fell algebras with non-vanishing Dixmier-Douady invariant. After some background in §2, we discuss in §3 the topological spaces that arise as the spectra of Fell algebras. In §4 we study the locally compact Hausdorff equivalence relation Date: October 15, 2018. 2000 Mathematics Subject Classification. 46L55. Key words and phrases. Fell algebra; continuous-trace algebra; Dixmier-Douady invariant; the C∗- algebra of a local homeomorphism; groupoid C∗-algebra. 1 2 CLARK, AN HUEF, AND RAEBURN R(ψ) associated to a surjection ψ : Y → X defined on a locally compact Hausdorff space Y . We show how extra properties of ψ influence the structure of the groupoid R(ψ). We show in particular that if ψ is a surjective local homeomorphism of Y onto a topological space X, then R(ψ) is ´etale, principal and Cartan, with orbit space naturally homeomorphic to X. We also show that every principal Cartan groupoid has the form R(ψ) for some quotient map ψ. In §5, we show that the twisted groupoid C∗-algebras of the R(ψ) give many examples of Fell algebras. We also show, extending a result of Archbold and Somerset for ´etale groupoids [2], that the C∗-algebra C∗(G) of a principal groupoid G is a Fell algebra if and only if G is topologically isomorphic to the relation R(q) determined by the quotient map q of the unit space G(0) onto G(0)/G. We then illustrate our results with a discussion of the path groupoids of directed graphs, and an example from [11] which fails to be Fell in a particularly delicate way. In Theorem 6.1, we prove that the Dixmier-Douady class δ(A) of a Fell algebra A vanishes if and only if A is Morita equivalent to some C∗(R(ψ)). We then use this to partially resolve a problem left open in [13, Remark 7.10]: when A has continuous trace, how is δ(A) related to the usual Dixmier-Douady invariant δDD(A) of [8, 7, 29]? In Corollary 6.3, we show that when A has continuous trace, δ(A) = 0 if and only if δDD(A) = 0. We also show that if A is Fell and δ(A) = 0, then every ideal I in A with continuous trace has δDD(I) = 0. Since δDD is computable, this allows us to recognise some Fell algebras whose invariant is nonzero. In §7, we describe two examples of Fell algebras which we have found instructive. The first is a Fell algebra whose spectrum fails to be paracompact in any reasonable sense, even though the algebra is separable. The second set of examples are Fell algebras A with nonzero invariant δ(A) and non-Hausdorff spectrum (so that they are not continuous-trace algebras). We close §7 with a brief epilogue on how we found these examples and what we have learned from them. We finish with two short appendices. The first concerns the different twisted groupoid algebras appearing in this paper. We mainly use Renault's algebras associated to a 2- cocycle σ on G from [30], but the proof of Theorem 6.1 uses the twisted groupoid algebra associated to a twist Γ over G from [18], and the proof of Theorem 5.1 uses yet another version from [21]. In Appendix A, we show that when Γ is the twist associated to a continuous cocycle, the three reduced C∗-algebras are isomorphic. In the last appendix, we describe the continuous-trace ideal in an arbitrary C∗-algebra. In the end, we did not need this result, but we think it may be of some general interest: we found it curious that the ideas which work for transformation group algebras in [10, Corollary 18] and [15, Theorem 3.10] work equally well in arbitrary C∗-algebras. 2. Notation and background A groupoid G is a small category in which every morphism is invertible. We write s and r for the domain and range maps in G. The set G(0) of objects in G is called the unit space, and we frequently identify a unit with the identity morphism at that unit. A groupoid is principal if there is at most one morphism between each pair of units. A topological groupoid is a groupoid equipped with a topology on the set of morphisms such that the composition and inverse maps are continuous. A topological groupoid G is FELL ALGEBRAS 3 ´etale if the map r (equivalently, s) is a local homeomorphism. The unit space of an ´etale groupoid is open in G, and the sets s−1(u) and r−1(u) are discrete for every u ∈ G(0). Suppose G is a topological groupoid. Then the orbit of u ∈ G(0) is [u] := r(s−1(u)). For u, v ∈ G(0) we write u ∼ v if [u] = [v], and then ∼ is an equivalence relation on G(0). We write q : G(0) → G(0)/G := G(0)/∼ for the quotient map onto the orbit space. If G is ´etale, then r is open, and then the quotient map is also open because q−1(q(U )) = r(s−1(U )) for U ⊂ G(0). A topological groupoid G is Cartan if every unit u ∈ G(0) has a neighbourhood N in G(0) which is wandering in the sense that s−1(N ) ∩ r−1(N ) has compact closure. Let G be a locally compact Hausdorff groupoid with left Haar system λ = {λu : u ∈ G(0)}. We also need to use the corresponding right Haar system {λu} defined by λu(E) = λu(E−1). A 2-cocycle on G is a function σ : G(2) → T such that σ(α, β)σ(αβ, γ) = σ(β, γ)σ(α, βγ). As in [4], we assume that all our cocycles are continuous and normalised in the sense that σ(r(γ), γ) = 1 = σ(γ, s(γ)), and we write Z 2(G, T) for the set of such cocycles. For such σ, there are both full and reduced twisted groupoid C∗-algebras. Here we work primarily with the reduced version, though in fact the full and reduced groupoid C∗-algebras coincide for the groupoids of interest to us (see Theorem 5.1). Let Cc(G, σ) be Cc(G) with involution and convolution given by f∗(α) = f (α−1)σ(α, α−1) and (f ∗ g)(α) = f (αγ)g(γ−1)σ(αγ, γ−1) dλs(α)(γ); it is shown in [30, Proposition II.1.1] that Cc(G, σ) is a ∗-algebra. The invariance of the Haar system gives (f ∗ g)(α) = f (β)g(β−1α)σ(β, β−1α) dλr(α)(β). (2.1) For u ∈ G(0), there is an induced representation Indσ that, for f ∈ Cc(G, σ) and ξ ∈ L2(s−1(u), λu), G u of Cc(G, σ) on L2(s−1(u), λu) such (cid:90) (cid:90) G (cid:90) G (Indσ u(f )ξ)(α) = f (β)ξ(β−1α)σ(β, β−1α) dλr(α)(β). As in [30, §II.2], the reduced twisted groupoid C∗-algebra C∗ Cc(G, σ) with respect to the reduced norm (cid:107)f(cid:107)r = sup u∈G(0) u(f )(cid:107). (cid:107) Indσ r (G, σ) is the completion of As usual, we write C∗ r (G) for C∗ r (G, 1) and Indu for Ind1 u. If G is ´etale then r−1(r(β)) is discrete, and (2.1), for example, reduces to f ∗ g(α) = f (β)g(β−1α)σ(β, β−1α) = f (β)g(γ)σ(β, γ). (cid:88) (cid:88) r(α)=r(β) α=βγ 3. Topological preliminaries By the standard definition, a topological space X is locally compact if every point of X has a compact neighbourhood. When X is Hausdorff, this is equivalent to asking that every point has a neighbourhood base of compact sets [23, Theorem 29.2]. For general, possibly non-Hausdorff spaces, we say that X is locally locally-compact if every point of X has a neighbourhood basis of compact sets (Lemma 3.1 below explains our choice of 4 CLARK, AN HUEF, AND RAEBURN name). The spectrum of a C∗-algebra is always locally locally-compact (see Lemma 3.2), so this "neighbourhood basis" version of local compactness has attractions for operator algebraists. It also has the advantage, as Munkres points out in [23, page 185], that it is more consistent with other uses of the word "local" in topology. It has been adopted without comment as the definition of local compactness in [3, page 149, problem 29]1 and in [33, Definition 1.16]. However, many topology books, such as [16] and [23], and many real-analysis texts, such as [31] and [24], use the standard "every point has a compact neighbourhood" definition, and we will go along with them. The proof of the following lemma is straightforward. Lemma 3.1. A topological space X is locally locally-compact if and only if every open subset of X is locally compact. A topological space X is locally Hausdorff if every point of X has a Hausdorff neighbour- hood. It is straightforward to verify that a locally Hausdorff space is T1. The following result explains our interest in locally locally-compact and locally Hausdorff spaces. Lemma 3.2. If A is a Fell algebra, then the spectrum of A is locally locally-compact and locally Hausdorff. Proof. Corollary 3.3.8 of [7] implies that A is locally locally-compact, and Corollary 3.4 (cid:3) of [2] that A is locally Hausdorff. Lemma 3.2 has a converse: every locally locally-compact and locally Hausdorff space is the spectrum of some Fell algebra [13, Theorem 6.6(2)]. We will later give a shorter proof of this result (see Corollary 5.5). We warn that a locally locally-compact and locally Hausdorff space may not be para- compact, that compact subsets may not be closed, and that the intersection of two com- pact sets may not be compact (for example, in the spectrum of the Fell C∗-algebra de- scribed in §7.1). So we have found our usual, Hausdorff-based intuition to be distressingly misleading, and we have tried to exercise extreme caution in matters topological. Locally locally-compact and locally Hausdorff spaces have the following purely topo- logical characterisation. Proposition 3.3. (a) Let ψ : Y → X be a local homeomorphism of a locally compact Hausdorff space Y onto a topological space X. Then X is locally locally-compact and locally Hausdorff. If Y is second-countable, so is X. (b) Let X be a locally locally-compact and locally Hausdorff space. Then there are a locally compact Hausdorff space Y and a local homeomorphism ψ of Y onto X. If X is second-countable, then we can take Y to be second-countable. Proof. For (a) suppose that ψ : Y → X is a local homeomorphism. Fix x ∈ X and an open neighbourhood W of x in X. Let y ∈ ψ−1(x). Since ψ−1(W ) is an open neighbourhood of y and ψ is a local homeomorphism, there is a neighbourhood U of y contained in ψ−1(W ) such that ψU is a homeomorphism. Since Y is locally compact and Hausdorff, it is locally locally-compact by [23, Theorem 29.2], and there is a compact neighbourhood K of y contained in U . Then ψ(K) is compact and Hausdorff, and because ψ is open, it is 1Modulo Bourbaki's use of the word "quasi-compact" to mean what we call compact. Dixmier follows Bourbaki, as one should be aware when reading [7]. a neighbourhood of x contained in W . This proves both that X is locally locally-compact and that X is locally Hausdorff. FELL ALGEBRAS 5 Since ψ is continuous and open, the image of a basis for the topology on Y is a basis for the topology on X. Thus X is second-countable if Y is. open cover U of X by Hausdorff sets. Let Y :=(cid:70) For (b), suppose that X is locally locally-compact and locally Hausdorff. Choose an U∈U U , and topologise Y by giving each U the subspace topology from X and making each U open and closed in Y . Then Lemma 3.1 implies that Y is locally compact and Hausdorff, and the inclusion maps U → X combine to give a surjective local homeomorphism ψ : Y → X. If X is second-countable, then we (cid:3) can take the cover to be countable, and Y is also second-countable. 4. The groupoid associated to a local homeomorphism Let ψ be a surjective map from a topological space Y to a set X, and take R(ψ) = Y ×ψ Y := {(y, z) ∈ Y × Y : ψ(y) = ψ(z)}. With the subspace topology and the operations r(y, z) = y, s(y, z) = z and (x, y)(y, z) = (x, z), R(ψ) is a principal topological groupoid with unit space Y . We want to examine the effect of properties of Y and ψ on the structure of R(ψ). We begin by looking at the orbit space. Lemma 4.1. Let ψ be a surjective map from a topological space Y to a set X, and define h : X → Y /R(ψ) by h(x) = ψ−1(x). (a) The function h is a bijection, and h ◦ ψ is the quotient map q : Y → Y /R(ψ). (b) If X is a topological space and ψ is continuous, then h is open. (c) Suppose that ψ : Y → X is a quotient map, in the sense that U is open in X if and only if ψ−1(U ) is open. Then h is a homeomorphism of X onto Y /R(ψ). Proof. (a) If h(x) = h(x(cid:48)), then the surjectivity of ψ implies that there exists at least one z ∈ ψ−1(x) = ψ−1(x(cid:48)), and then x = ψ(z) = x(cid:48). So h is one-to-one. Surjectivity is easy: every orbit ψ−1(ψ(y)) = h(ψ(y)). The same formula h(ψ(y)) = ψ−1(ψ(y)) shows that h ◦ ψ(y) is the orbit q(y) of y. For (b), take U open in X. Then q−1(h(U )) = ψ−1(h−1(h(U ))) = ψ−1(U ) is open because ψ is continuous, and then h(U ) is open by definition of the quotient topology. For (c), take V open in Y /R(ψ). Then ψ−1(h−1(V )) = q−1(V ) is open in Y , and h−1(V ) (cid:3) is open because ψ is a quotient map. Lemma 4.2. Suppose that ψ : Y → X is a quotient map. Then R(ψ) is ´etale if and only if ψ is a local homeomorphism. Proof. Suppose that ψ is a local homeomorphism and (y, z) ∈ R(ψ). There are open neighbourhoods U of y and V of z such that ψU and ψV are homeomorphisms onto open neighbourhoods of ψ(y) = ψ(z). By shrinking if necessary, we may suppose ψ(U ) = ψ(V ). Now W := (U × V ) ∩ R(ψ) is an open neighbourhood of (y, z), and the function w (cid:55)→ (w, (ψV )−1 ◦ ψU (w)) is a continuous inverse for rW . Thus r is a local homeomorphism, and R(ψ) is ´etale. Conversely, suppose that R(ψ) is ´etale and y ∈ Y . Then (y, y) ∈ R(ψ). Since r is a local homeomorphism, there is a neighbourhood W of (y, y) in R(ψ) such that rW is a homeomorphism. By shrinking, we can assume that W = (U × U ) ∩ R(ψ) for some open neighbourhood U of y in Y . We claim that ψ is one-to-one on U . Suppose y1, y2 ∈ U , and 6 CLARK, AN HUEF, AND RAEBURN ψ(y1) = ψ(y2). Then (y1, y1) and (y1, y2) are both in W . Now r(y1, y1) = y1 = r(y1, y2), the injectivity of rW implies that (y1, y1) = (y1, y2), and y1 = y2. Thus ψU is one- to-one, as claimed. Since the orbit map q in an ´etale groupoid is open, and h is a homeomorphism with h ◦ ψ = q, it follows that ψ(U ) = h−1(q(U )) is open. Thus ψ is a (cid:3) local homeomorphism. Proposition 4.3. Suppose that Y and X are topological spaces with Y locally compact Hausdorff, and ψ : Y → X is a surjective local homeomorphism. Then R(ψ) is locally compact, Hausdorff, principal, ´etale and Cartan. Proof. The groupoid R(ψ) is principal because it is an equivalence relation, and is Haus- dorff because Y is Hausdorff. Since ψ is a local homeomorphism, it is a quotient map, and hence Lemma 4.2 implies that R(ψ) is ´etale. Let (y, z) ∈ R(ψ). Then there is an open neighbourhood U of (y, z) such that rU is a homeomorphism onto an open neighbourhood of y. Since locally compact Hausdorff spaces are locally locally-compact, r(U ) contains a compact neighbourhood K of y. Now (rU )−1(K) is a compact neighbourhood of (y, z) in R(ψ), and we have shown that R(ψ) is locally compact. Since locally compact Hausdorff spaces are regular [24, 1.7.9], the following lemma tells us that R(ψ) is Cartan, and hence completes the proof of Proposition 4.3. The extra (cid:3) generality in the lemma will be useful in the proof of Proposition 4.5. Lemma 4.4. Suppose that Y is a regular topological space and ψ : Y → X is a surjection. If R(ψ) is locally compact, then R(ψ) is Cartan. Proof. Let y ∈ Y . We must find a wandering neighbourhood W of y in Y , that is, a neighbourhood W such that (W × W ) ∩ R(ψ) has compact closure in R(ψ). Since R(ψ) is locally compact, there exists a compact neighbourhood K of (y, y) in R(ψ). Since K is a neighbourhood, the interior int K is an open set containing (y, y), and there exists an open set O ⊂ Y × Y such that int K = O ∩ R(ψ). Choose open neighbourhoods U1, U2 of y in Y such that U1×U2 ⊂ O. Since Y is regular, there are open neighbourhoods Vi of y such that V i ⊂ Ui for i = 1, 2. Let C := V 1 ∩ V 2. Then C is a closed neighbourhood of y in Y , and (C × C) ∩ R(ψ) ⊂ (U1 × U2) ∩ R(ψ) ⊂ O ∩ R(ψ) ⊂ K. Since (C × C) ∩ R(ψ) is closed in R(ψ) and K is compact, (C × C) ∩ R(ψ) is compact, (cid:3) and C is the required neighbourhood of y. Proposition 4.3 has an intriguing converse. Suppose that G is a locally compact, Haus- dorff and principal groupoid. We will see that if G is Cartan, then G has the form R(q), where q : G(0) → G(0)/G is the quotient map. The key idea is that, because G is principal, the map r × s : γ (cid:55)→ (r(γ), s(γ)) is a groupoid isomorphism of G onto R(q). The map r × s is also continuous for the product topology on R(q), but it is not necessarily open, and hence is not necessarily an isomorphism of topological groupoids (see Example 5.8 below). But: Proposition 4.5. Suppose G is a locally compact, Hausdorff and principal groupoid which admits a Haar system. Then G is Cartan if and only if r× s is a topological isomorphism of G onto R(q). For the proof we need a technical lemma. FELL ALGEBRAS 7 Lemma 4.6. Let G be a locally compact Hausdorff groupoid which admits a Haar system. If G is Cartan, then r × s is a closed map onto its image r × s(G). Proof. Let C be a closed subset of G and (u, v) be a limit point of r × s(C) in r × s(G). Then there exists γ ∈ G such that r × s(γ) = (u, v) and a net {γi} in C such that r×s(γi) → (u, v). It suffices to show that {γi} has a convergent subnet. Indeed, if γij → γ(cid:48) then γ(cid:48) ∈ C because C is closed, r × s(γ(cid:48)) = (u, v) by continuity, and (u, v) ∈ r × s(C). Since G is Cartan, u has a neighbourhood N in G(0) such that U := (r × s)−1(N × N ) is relatively compact in G. Let V be a relatively compact neighbourhood of γ. We may assume by shrinking V that V ⊂ r−1(N ), and hence that r(V ) ⊂ N . The continuity of multiplication implies that U V is relatively compact. We claim that γi ∈ U V eventually. To see this, we observe that the existence of the Haar system implies that s is open [32, Corollary, page 118], and hence s(V ) is a neighbourhood of v = s(γ). Thus there exists i0 such that s(γi) ∈ s(V ) and r(γi) ∈ N for all i ≥ i0. For each i ≥ i0 there exists βi ∈ V such that s(γi) = s(βi). Now s(γiβ−1 i ) = r(βi) ∈ r(V ) ⊂ N is in U . Thus γi = αiβi ∈ U V for i ≥ i0, as and r(γiβ−1 claimed. Now {γi : i ≥ i0} is a net in a relatively compact set, and hence has a convergent (cid:3) subnet, as required. Proof of Proposition 4.5. Suppose that G is Cartan. The map r × s is always a contin- uous surjection onto R(q), and it is injective because G is principal. Since G is Cartan, Lemma 4.6 implies that r × s is closed as a map onto its image R(q). A bijection is open if and only if it is closed, so r × s is open. Hence r × s is a homeomorphism onto R(q). Conversely, suppose that r × s is a homeomorphism onto R(q). Then r × s is an isomorphism of topological groupoids, and since G is locally compact, so is R(q). Thus (cid:3) R(q) is Cartan by Lemma 4.4, and so is G. i ) = r(γi) ∈ N , so αi := γiβ−1 i Concluding discussion. Lemma 4.4 shows that, if the groupoid R(ψ) associated to a quotient map is locally compact, then R(ψ) is Cartan. On the other hand, Proposition 4.5 says that, if there is a topology on R(ψ) which makes it into a locally compact Cartan groupoid, then that topology has to be the relative topology from the product space Y ×Y . So one is tempted to seek conditions on ψ which ensure that the subset R(ψ) ⊂ Y × Y is locally compact. By Proposition 4.3, it suffices for ψ to be a local homeomorphism. This is not a necessary condition: for example, if (Y, H) is a free Cartan transformation group with H nondiscrete, then the transformation groupoid Y × H is Cartan in our sense, but the quotient map q : Y → Y /H is not locally injective. (In [12] there is a specific example of a free Cartan transformation group which illustrates this.) However, the following example shows that something extra is needed. Example 4.7. Take Y = [0, 1] and X = {a, b} with the topology {X,{a},∅}, and define ψ : Y → X by (cid:40) Then ψ is an open quotient map, but R(ψ) =(cid:0)(0, 1] × (0, 1](cid:1) ∪ {(0, 0)} is not a locally compact subset of [0, 1] × [0, 1] because (0, 0) does not have a compact neighbourhood. ψ(y) = a if t > 0 if t = 0. b 8 CLARK, AN HUEF, AND RAEBURN 5. The groupoids whose C∗-algebras are Fell algebras We begin this section by summarising results from [5, 6, 4, 22] about the twisted groupoid algebras of principal Cartan groupoids. Theorem 5.1. Suppose that G is a second-countable, locally compact, Hausdorff, principal and Cartan groupoid which admits a Haar system. Then the following statements are equivalent: (a) G is Cartan; (b) C∗(G) is a Fell algebra; (c) for all σ ∈ Z 2(G, T), C∗(G, σ) is a Fell algebra. r (G, σ). u induces a homeomorphism of G(0)/G onto If (a) -- (c) these are satisfied, then u (cid:55)→ Indσ C∗(G, σ)∧, and C∗(G, σ) = C∗ Proof. The implication (c) =⇒ (b) is trivial, the equivalence of (a) and (b) is Theorem 7.9 of [5], and the implication (b) =⇒ (c) is part (ii) of [4, Proposition 3.10 (a)]. So it remains to prove the assertions in the last sentence. The orbits in a Cartan groupoid are closed [5, Lemma 7.4], so the orbit space Y /R(ψ) is T1, and it follows from [6, Proposition 3.2] that the map y (cid:55)→ [Lu] described there in- duces a homeomorphism of G(0)/G onto the spectrum of the twisted groupoid C∗-algebra C∗(Gσ; G)MW of Muhly and Williams (see [22] or Appendix A). By [4, Lemma 3.1], C∗(Gσ; G)MW is isomorphic to C∗(G, σ), and by Lemma A.1, this isomorphism carries the equivalence class of Lu to the class of Indσ u induces a homeo- morphism, as claimed. This implies in particular that all the irreducible representations of C∗(G, σ) are induced, so for f ∈ Cc(G, σ) we have u. We deduce that u (cid:55)→ Indσ and C∗(G, σ) = C∗ r (G, σ). (cid:107)f(cid:107) = sup{(cid:107) Indy(f )(cid:107) : y ∈ Y } =: (cid:107)f(cid:107)r, (cid:3) For our first application of Theorem 5.1, we observe that putting the equivalence of (a) and (b) together with Proposition 4.5 gives the following improvement of a result of Archbold and Somerset [2, Corollary 5.9]. (We discuss the precise connection with [2] in Remark 5.3.) Corollary 5.2. Suppose that G is a second-countable, locally compact, Hausdorff and principal groupoid which admits a Haar system λ, and q : G(0) → G(0)/G is the quotient map. Then C∗(G, λ) is a Fell algebra if and only if r× s : G → G(0) × G(0) is a topological isomorphism of G onto R(q). Remark 5.3. The "separated topological equivalence relations" R studied in [2, §5] are the second-countable, locally compact, Hausdorff and principal groupoids that are ´etale. When they say in [2, Corollary 5.9] that "the topologies τp and τ0 coincide," they mean precisely that the map r×s is a homeomorphism for the original topology τ0 on R and the product topology on R(0) × R(0). Theorem 5.1 implies that the full algebra above and the reduced algebra in [2] coincide. So Corollary 5.2 extends [2, Corollary 5.9] from principal `etale groupoids to principal groupoids which admit a Haar system. This is a substantial generalisation since, for example, locally compact transformation groups always admit a Haar system [30, page 17] even though the associated transformation groupoids may not be ´etale. Our proof of Corollary 5.2 seems quite different from the representation-theoretic arguments used in [2]. FELL ALGEBRAS 9 Next we apply Theorem 5.1 to the groupoid associated to a local homeomorphism. Corollary 5.4. Let ψ : Y → X be a local homeomorphism of a second-countable, locally compact and Hausdorff space Y onto a topological space X, and let σ : R(ψ)(2) → T be a continuous normalised 2-cocycle. Then C∗(R(ψ), σ) is a Fell algebra, y (cid:55)→ Indσ y induces a homeomorphism of Y /R(ψ) onto C∗(R(ψ), σ)∧, and C∗(R(ψ), σ) = C∗ Proof. Proposition 4.3 implies that R(ψ) satisfies all the hypotheses of Theorem 5.1, which (cid:3) then gives the result. When X is Hausdorff, we know from [17] that C∗(R(ψ)) has cotinuous trace; the twisted versions we used in [27] to provide examples of continuous-trace algebras with nonzero Dixmier-Douady class. r (R(ψ), σ). We now give our promised shorter proof of the converse of Lemma 3.2. Corollary 5.5. Let X be a second-countable, locally locally-compact and locally Hausdorff topological space. Then there is a separable Fell C∗-algebra with spectrum X. Proof. By Proposition 3.3, there are a second-countable, locally compact and Hausdorff space Y and a surjective local homeomorphism ψ : Y → X. Consider the topological relation R(ψ). Lemma 4.1 gives a homeomorphism h of X onto Y /R(ψ), and Corollary 5.4 says that C∗(R(ψ)) is a separable Fell algebra with spectrum homeomorphic to Y /R(ψ). (cid:3) Next we consider a row-finite directed graph E with no sources, using the conventions of [26]; we also use the more recent convention that, for example, vEnw = {α ∈ En : r(α) = v, s(α) = w}. The infinite-path space E∞ has a locally compact Hausdorff topology with basis the cylinder sets Z(α) = {αx : x ∈ E∞ and r(x) = s(α)}. [20, Corollary 2.2]. The set GE = {(x, k, y) ∈ E∞ × Z × E∞ : there exists n such that xi = yi+k for i ≥ n} is a groupoid with unit space G(0) and ´etale in a topology which has a neighbourhood basis consisting of the sets E = E∞, and this groupoid is locally compact, Hausdorff Z(α, β) = {(αz,β − α, βz) : z ∈ E∞, r(z) = s(α)} parametrised by pairs of finite paths α, β ∈ E∗ with s(α) = s(β) [20, Proposition 2.6]. We know from [11, Proposition 8.1] that GE is principal if and only if E has no cycles (in which case we say E is acyclic). We write x ∼ y to mean (x, k, y) ∈ GE for some k; this equivalence relation on E∞ is called tail equivalence with lag. Proposition 5.6. Suppose that E is an acyclic row-finite directed graph with no sources. Then GE is Cartan if and only if the quotient map q : E∞ → E∞/∼ is a local homeo- morphism. Proof. Suppose that GE is Cartan. The quotient space E∞/∼ is the orbit space of GE. Thus Proposition 4.5 implies that r × s is an isomorphism of topological groupoids of GE onto R(q). Since GE is ´etale, so is R(q). Thus Lemma 4.2 implies that q is a local homeomorphism. 10 CLARK, AN HUEF, AND RAEBURN Conversely, suppose that q is a local homeomorphism. We claim that r×s : GE → R(q) is an isomorphism of topological groupoids. Then, since we know from Proposition 4.3 that R(q) is Cartan, we can deduce that GE is Cartan too. To prove the claim, we observe first that r× s is an isomorphism of algebraic groupoids because GE is principal, and is continuous because r and s are. So it suffices to take α, β ∈ E∗ with s(α) = s(β), and prove that r × s(Z(α, β)) is open. A typical element of r × s(Z(α, β)) has the form (αz, βz) for some z ∈ E∞ with r(z) = s(α). Since q is a local homeomorphism there is an initial segment µ of z such that qZ(µ) is one-to-one. We will prove that (Z(αµ) × Z(βµ)) ∩ R(q) is contained in r × s(Z(α, β)). Let (x, y) ∈ (Z(αµ) × Z(βµ)) ∩ R(q). Then there are x(cid:48), y(cid:48) ∈ E∞ such that x = αµx(cid:48) and y = βµy(cid:48), and q(x) = q(y) implies q(µx(cid:48)) = q(µy(cid:48)). Both µx(cid:48) and µy(cid:48) are in Z(µ), so injectivity of qZ(µ) implies that µx(cid:48) = µy(cid:48), and x(cid:48) = y(cid:48). But now (x, y) = (αµx(cid:48), βµx(cid:48)) = r × s(αµx(cid:48),β − α, βµx(cid:48)) belongs to r × s(Z(α, β)). Thus r × s(Z(α, β)) contains a neighbourhood of (αz, βz), and r × s(Z(α, β)) is open. Now we have proved our claim, and the result follows. (cid:3) The groupoid GE was originally invented as a groupoid whose C∗-algebra is the univer- sal algebra C∗(E) generated by a Cuntz-Krieger E-family [20, Theorem 4.2]. In view of Corollary 5.2, Proposition 5.6 implies that C∗(E) = C∗(GE) is a Fell algebra if and only if q : E∞ → E∞/∼ is a local homeomorphism. So it is natural to ask whether we can identify this property at the level of the graph. We can: s(xn)E∗s(µ) = {µ} for every µ ∈ E∗ with r(µ) = s(xn). Proposition 5.7. Suppose that E is an acyclic row-finite directed graph with no sources. Then the quotient map q : E∞ → E∞/∼ is a local homeomorphism if and only if, for every x ∈ E∞, there exists n such that (5.1) Proof. Suppose first that q is a local homeomorphism, and x ∈ E∞. Then there is an initial segment µ = x1x2 ··· xn of x such that qZ(µ) is injective. We claim that this n satisfies (5.1). Suppose not. Then there exist α, β ∈ s(xn)E∗ such that α (cid:54)= β and s(α) = s(β). Neither can be an initial segment of the other, since this would give a cycle at s(α). So there exists i ≤ min(α,β) such that αi (cid:54)= βi. Then for any y ∈ E∞ with r(y) = s(α), µαy and µβy are distinct paths in Z(µ) with q(µαy) = q(µβy), which contradicts the injectivity of qZ(µ). Conversely, suppose that E has the property described, and let x ∈ E∞. Take n satisfying (5.1), and µ := x1x2 ··· xn. We claim that qZ(µ) is injective. Suppose y = µy(cid:48), z = µz(cid:48) ∈ Z(µ) and q(y) = q(z). Then q(y(cid:48)) = q(z(cid:48)), and there exist paths γ, δ in s(xn)E∗ such that y(cid:48) = γy(cid:48)(cid:48), z(cid:48) = δy(cid:48)(cid:48), say. The existence of y(cid:48)(cid:48) forces s(γ) = s(δ), and (5.1) implies that γ = δ, y(cid:48) = z(cid:48) and y = z. Thus q is locally injective. Since GE is ´etale, (cid:3) the quotient map q is open, and hence q is a local homeomorphism. Example 5.8. Consider the following graph E from [11, Example 8.2]: v1 v2 v3 v4 f (1) 1 f (2) 1 f (1) 2 f (2) 2 f (1) 3 f (2) 3 f (1) 4 f (2) 4 FELL ALGEBRAS 11 Let GE be the path groupoid. Let z be the infinite path with range v1 that passes through each vn, and, for n ≥ 1, let xn be the infinite path with range v1 that includes the edge f (1) n . It is shown in [11, Example 8.2] that the sequence {xn} "converges 2-times in E∞/GE to z", and it then follows from [6, Lemma 5.1] that GE is not Cartan. Applying the criteria of Proposition 5.7 seems to give an easier proof of this: let z be as above. For n+1) = {f (1) n+1}. Thus q : E∞ → E∞/∼ is not a each n, s(zn) = vn+1 and s(zn)E∗s(f (1) local homeomorphism by Proposition 5.7 and hence GE is not Cartan by Proposition 5.6. Proposition 4.5 implies that r × s is not open. n+1, f (2) 6. Fell algebras with trivial Dixmier-Douady invariant If A is a Fell algebra, we write δ(A) for its Dixmier-Douady invariant, as defined in [13, Section 7]. If A is a continuous-trace algebra, then δ(A) makes sense, and A also has a Dixmier-Douady invariant δDD(A) as in [29, §5.3], for example; as pointed out in [13, Remark 7.10], it is not clear whether these invariants are the same. Recall that the properties of being a Fell algebra or having continuous trace are preserved under Morita equivalence by [14, Corollary 14] and [34, Corollary 3.5]. The following is implicitly assumed in [13]. r (R(ψ)). Theorem 6.1. Let A be a separable Fell algebra. Then the Dixmier-Douady invariant δ(A) of A is 0 if and only if there is is a local homeomorphism ψ of a second-countable, locally compact and Hausdorff space onto a topological space such that A is Morita equiv- alent to C∗ Proof. We start by recalling how δ(A) is defined in [13]. By [13, Theorem 5.17], A is Morita equivalent to a C∗-algebra C which has a diagonal C∗-subalgebra D; let h : C → A be an associated Rieffel homeomorphism. By [18, Theorem 3.1], there is a twist Γ → R over an ´etale and principal groupoid R such that C is isomorphic to Kumjian's C∗-algebra C∗(Γ; R)Kum of the twist. Since C is a Fell algebra we may by [13, Proposition 6.3] assume that R = R(ψ), where ψ : D → C is the spectral map, which is a local homeomorphism by [13, Theorem 5.14]. By [19, Remark 2.9], there is an extension Γ → R(ψ) where Γ is the groupoid consisting of germs of continuous local sections of the surjection Γ → R(ψ). Such extensions are called sheaf twists, and the group of their isomorphism classes is denoted by TR(ψ)(S). Let H 2(R(ψ),S) be the second equivariant sheaf cohomology group. The long exact sequence of [19, Theorem 3.7] yields a boundary map ∂1 from TR(ψ)(S) to H 2(R(ψ),S). Finally, set δ(A) = (π∗ h◦ψ)−1(∂1([Γ])) ∈ H 2( A,S) where πh◦ψ : R(h ◦ ψ) → A is given by (y, z) (cid:55)→ h ◦ ψ(y) [13, Definition 7.9]. Quite a bit of the work in [13, §7] is to show that δ(A) is well-defined. Now suppose that δ(A) = 0. Let Γ be a twist associated to A. Then ∂1([Γ]) = 0. Let Λ := T × R(ψ) so that Λ → R(ψ) is the trivial twist. Then the associated sheaf twist Λ is also trivial, whence ∂1([Λ]) = 0. Now 0 = ∂1([Γ]) is sent to 0 = ∂1([Λ]) under a certain natural isomorphism (see [13, Corollary 7.6]), and [13, Lemma 7.12] implies that Γ → R(ψ) and Λ → R(ψ) are equivalent twists. Thus C∗(Γ; R(ψ))Kum and C∗(Λ; R(ψ))Kum are Morita equivalent by [13, Lemma 6.5]. But now A and C∗(Λ; R(ψ))Kum are Morita equivalent. Since Λ → R(ψ) is trivial, C∗(Λ; R(ψ))Kum is isomorphic to C∗ r (R(ψ)) by [13, Lemma A.1]. Thus A and C∗ r (R(ψ)) are Morita equivalent. 12 CLARK, AN HUEF, AND RAEBURN Conversely, suppose that A is Morita equivalent to C∗ of [13], C∗ R(ψ). The associated sheaf twist Λ is also trivial, so ∂1([Λ]) = 0, and δ(C∗ Since A and C∗ phism k : A → C∗ r (R(ψ))∧ such that the induced isomorphism k∗ : H 2(C∗ H 2( A,S) carries 0 = δ(C∗ r (R(ψ)) for some ψ. By Lemma A.1 r (R(ψ)) is isomorphic to the C∗-algebra of the trivial twist Λ := T × R(ψ) → r (R(ψ))) = 0. r (R(ψ)) are Morita equivalent, by Theorem 7.13 of [13] there is a homeomor- r (R(ψ)∧,S) → (cid:3) r (R(ψ))) to δ(A). Thus δ(A) = 0. Proposition 6.2. Let A be a separable Fell algebra. (a) If δ(A) = 0 then δDD(I) = 0 for every ideal I of A with continuous trace. (b) If A has continuous trace and δDD(A) = 0, then δ(A) = 0. Proof. Let B1 and B2 be C∗-algebras with continuous trace and paracompact spectrum X. Propositions 5.32 and 5.33 of [29] together say that δDD(B1) = δDD(B2) if and only if B1 and B2 are Morita equivalent. Below we consider an ideal I in a separable Fell algebra A such that I has continuous trace. Then I is second-countable, locally compact and Hausdorff, and hence is σ-compact. By [24, Proposition 1.7.11], for example, I is paracompact. Thus Propositions 5.32 and 5.33 of [29] apply to continuous-trace C∗- algebras with spectrum homeomorphic to I. (a) Suppose δ(A) = 0, and let I be an ideal of A with continuous trace. By Theorem 6.1, r (R(ψ)) for some local homeomorphism ψ : Y → X. Then A is Morita equivalent to C∗ r (R(ψ)) is also a Fell algebra, and I is Morita equivalent to an ideal J of C∗ C∗ r (R(ψ)) r (R(ψ)) = C∗(R(ψ)). Since R(ψ) is principal with continuous trace. By Corollary 5.4, C∗ and C∗(R(ψ)) is liminary, by [5, Proposition 6.1] there exists an open invariant subset U of the unit space Y of R(ψ) such that J is isomorphic to the C∗-algebra of R(ψ)U := {γ ∈ R(ψ) : r(γ), s(γ) ∈ U}. Since R(ψ)U is principal and C∗(R(ψ)U ) has continuous trace, R(ψ)U is a proper groupoid by [21, Theorem 2.3], and δDD(C∗(R(ψ)U )) = 0 by [21, Proposition 2.2]. Since C∗(R(ψ)U ) and I are Morita equivalent, δDD(I) = 0 by [29, Proposition 5.32]. (b) Suppose δDD(A) = 0. Then δDD(A) = δDD(C0( A)), and A and C0( A) are Morita r (R(ψ)) (cid:3) equivalent by [29, Proposition 5.33]. But C0( A) is isomorphic to C∗(R(ψ)) = C∗ where ψ : A → A is the identity. Thus δ(A) = 0 by Theorem 6.1. The following corollary is immediate from Proposition 6.2. Corollary 6.3. Suppose A is a separable C∗-algebra with continuous trace. Then δ(A) = 0 if and only if δDD(A) = 0. 7. Examples A standard example of a Fell algebra which does not have continuous trace is the algebra A3 = {f ∈ C([0, 1], M2(C)) : f (1) is diagonal} discussed in [29, Example A.25]. Here we describe two variations on this construction. It seems clear to us that our constructions could be made much more general, for example by doubling up along topologically nontrivial subspaces rather than at a single point. FELL ALGEBRAS 13 7.1. A Fell algebra with trivial Dixmier-Douady invariant. We write {ξi}i∈N for the usual orthonormal basis in (cid:96)2(N) and Θij for the rank-one operator Θij : h (cid:55)→ (h ξj)ξi on (cid:96)2(N). We write K = K((cid:96)2(N)), and (7.1) A := {a ∈ C([0, 1],K) : a(1) is diagonal in the sense that (a(1)ξi ξj) = 0 for i (cid:54)= j}. i∈N[0, 1] = (cid:83) We let Y be the disjoint union (cid:70) We write eij for the constant function t (cid:55)→ Θij in C([0, 1],K). i∈N[0, 1] × {i}. Then Y is a locally compact Hausdorff space with the topology in which each [0, 1] × {i} is open, closed and homeomorphic to [0, 1]. There is an equivalence relation ∼ on Y such that (s, i) ∼ (s, j) for all i, j ∈ N and s ∈ [0, 1), and the (1, i) are equivalent only to themselves. We claim that the quotient map ψ : Y → X := Y /∼ is open. To see this, it suffices to take an open set U = W × {i} contained in one level [0, 1] × {i}, and see that ψ(U ) is open. By definition of the quotient topology, we have to show that ψ−1(ψ(U )) is open. If (1, i) is not in U , then (cid:91) j∈N ψ−1(ψ(U )) = {(s, j) : (s, i) ∈ U, j ∈ N} = W × {j}, which is open. If (1, i) ∈ U , then ψ−1(ψ(U )) = (W × {i}) ∪(cid:16)(cid:91) (cid:17) ((W \ {1}) × {j}) , j(cid:54)=i We now consider R(ψ), and write Vij for the subset(cid:0)([0, 1]×{i})×([0, 1]×{j})(cid:1)∩R(ψ). as a finite sum (cid:80){(i,j):supp f∩Vij(cid:54)=∅}(fij ◦ ψij)χVij . By viewing functions in Cc([0, 1)) as which is open. Thus ψ is open, as claimed. Since [0, 1] × {i} is open and ψ[0,1]×{i} is injective, ψ is a surjective local homeomorphism. Then for i (cid:54)= j, the map ψij : ((s, i), (s, j)) (cid:55)→ s is a homeomorphism of Vij onto [0, 1); for i = j, the similarly defined ψii is a homeomorphism of Vii onto [0, 1]. Thus for each f ∈ Cc(R(ψ)), the compact set supp f meets only finitely many Vij. Define fij(s) = f ((s, i), (s, j)). Then fij ∈ Cc([0, 1)) (for i (cid:54)= j) and fii ∈ C([0, 1]), and f can be recovered functions on [0, 1] which vanish at 1, we can define ρ : Cc(R(ψ)) → A by (cid:88) ρ(f ) = fijeij. {(i,j):supp f∩Vij(cid:54)=∅} Proposition 7.1. The function ρ : Cc(R(ψ)) → A extends to an isomorphism of C∗(R(ψ)) onto A. First we check that ρ is a homomorphism on the convolution algebra Cc(R(ψ)). Let f, g ∈ Cc(R(ψ)), and observe that for each s, (7.2) (f ∗ g)((s, i), (s, j)) = (cid:88) f ((s, i), (s, k))g((s, k), (s, j)) = fik(s)gkj(s) k k has only finitely many nonzero terms (and just one if s = 1 and i = j). Since the eij are matrix units in C([0, 1],K), and the operations in A are those of C([0, 1],K), Equation 7.2 implies that ρ is multiplicative. Thus ρ is a ∗-homomorphism. To see that ρ extends to C∗(R(ψ)), we show that ρ is isometric for the reduced norm on Cc(R(ψ)), which by Corollary 5.4 is the same as the enveloping norm. The norm in (cid:88) 14 CLARK, AN HUEF, AND RAEBURN C([0, 1],K) satisfies (cid:107)a(cid:107) = sup{(cid:107)a(t)(cid:107) : t ∈ [0, 1)}, and since [0, 1) is also dense in X, the reduced norm on Cc(R(ψ)) satisfies (cid:107)f(cid:107) = sup{(cid:107) Ind(s,i)(f )(cid:107) : s ∈ [0, 1)}. Thus the following lemma implies that ρ is isometric. Lemma 7.2. For s ∈ [0, 1), we define s : A → B((cid:96)2(N)) by s(f ) = f (s). Then the representation s ◦ ρ of Cc(R(ψ)) is unitarily equivalent to Ind(s,i). Proof. For each i ∈ N we have s−1((s, i)) = {((s, j), (s, i)) : j ∈ N}, so there is a unitary isomorphism Us of (cid:96)2(s−1((s, i))) onto (cid:96)2(N) which carries the point mass e((s,j),(s,i)) into the basis vector ξj. We will prove that Us intertwines s◦ρ and Ind(s,i). Let f ∈ Cc(R(ψ)). On one hand, we have s ◦ ρ(f )(Use((s,j),(s,i))) = fkl(s)Θklξj = fkj(s)ξk. (cid:88) k On the other hand, the induced representation satisfies (cid:0) Ind(s,i)(f )e((s,j),(s,i)) k,l (cid:88) (cid:88) (cid:1)((s, k), (s, i)) = (cid:16)(cid:88) (cid:1) = Us l f ((s, k), (s, l))e((s,j),(s,i))((s, l), (s, i)) and hence Us (cid:0) Ind(s,i)(f )e((s,j),(s,i)) = f ((s, k), (s, j)) = fkj(s), (cid:17) (cid:88) fkj(s)e((s,k),(s,i)) = fkj(s)ξk. (cid:3) k k Indeed, because the qN = (cid:80)N It remains for us to see that ρ is surjective. Since ρ(C∗(R(ψ))) is a C∗-algebra, it suffices to show that ρ(Cc(R(ψ))) is dense in A. i=1 eii form an approximate identity for A, it suffices to take b ∈ qN AqN and show that we can i,j≤N bijeij with bij ∈ C([0, 1]) i=0 biieii, and observe that c(1) = 0. Choose δ ∈ (0, 1) such that supt∈[δ,1) (cid:107)c(t)(cid:107) < , and choose a continuous function h : [0, 1] → [0, 1] i=0 biieii + hc, i=0 dijeij, where dij ∈ Cc([0, 1)) if i (cid:54)= j and dii ∈ C([0, 1]). Set approximate b by some ρ(f ). Fix > 0. We can write b =(cid:80) and bij(1) = 0 for i (cid:54)= j. Set c := b −(cid:80)N such that h = 1 on [0, δ] and h = 0 near 1. Then (cid:107)c− hc(cid:107)∞ < . Set d :=(cid:80)N and then d has the form(cid:80)N f =(cid:80)N (cid:13)(cid:13)(cid:13)∞ = (cid:107)c − hc(cid:107)∞ < . (cid:13)(cid:13)(cid:13)b − n(cid:88) i=0 dijχVij . Then f ∈ Cc(R(ψ)), ρ(f ) = d, and biieii − hc (cid:107)b − ρ(f )(cid:107)∞ = i=1 This completes the proof of Proposition 7.1. Corollary 7.3. The C∗-algebra A in (7.1) is a Fell algebra which does not have continuous trace, and the Dixmier-Douady invariant of A is 0. Proof. By Proposition 7.1, A is isomorphic to C∗(R(ψ)) where ψ is a surjective local homeomorphism. Now Corollary 5.4 implies that A is a Fell algebra with spectrum X, and Theorem 6.1 implies that its Dixmier-Douady invariant vanishes. Because X is not (cid:3) Hausdorff, A does not have continuous trace. FELL ALGEBRAS 15 Remark 7.4. Paracompactness is usually defined only for Hausdorff spaces, and the exam- ple of this section confirms that things can go badly wrong for the sorts of non-Hausdorff spaces of interest to us. For the spectrum X of our algebra A, the sets ψ([0, 1]×{i}) form an open cover of X, but every neighbourhood of the point ψ(1, 1), for example, meets every neighbourhood of every other ψ(1, i), so there cannot be a locally finite refinement. 7.2. A Fell algebra with non-trivial Dixmier-Douady invariant. We describe a Fell algebra A which does not have continuous trace, and has Dixmier-Douady class δ(A) (cid:54)= 0. We do this by combining the construction of [27, §1] (see also [29, Example 5.23]) with that of the algebra A3 at the start of §7. We adopt the notation of [29, Chapter 5]. We start with a compact Hausdorff space S, a finite open cover U = {U1, . . . , Un} of S, and an alternating cocycle λijk : Uijk → T whose class [λijk] in H 2(S,S) is nonzero. By the argument of [28, Lemma 3.4], for example, we may multiply λ by a coboundary and assume that λijk ≡ 1 whenever two of i, j, k coincide. The algebra A(U, λijk) in [29, Example 5.23] has underlying vector space A(U, λijk) = {(fij) ∈ Mn(C(S)) : fij = 0 on S \ Uij}. The product in A(U, λijk) is defined by (fij)(gkl) = (hil), where (cid:40)(cid:80){j:s∈Uijl} λijl(s)fij(s)gjl(s) (7.3) hil(s) = 0 if s ∈ Uijl for some j otherwise, and the involution given by (fij)∗ = (fji). For s ∈ U , we take Is := {i : s ∈ Ui}, and define πi,s : A(U, λijk) → MIs by (7.4) πi,s It is shown in [29, Example 5.23] that A(U, λijk) is a C∗-algebra with (cid:0) (fjk)(cid:1) =(cid:0)λijk(s)fjk(s)(cid:1). (cid:0) (fjk)(cid:1)(cid:13)(cid:13). (cid:13)(cid:13)(fjk)(cid:13)(cid:13) = sup (cid:13)(cid:13)πi,s In fact, and we shall need this later, A(U, λijk) is a continuous-trace algebra with spectrum S and Dixmier-Douady class δDD(A(U, λijk)) = [λijk] (see [29, Proposition 5.40], which simplifies in our case because S is compact and the cover is finite). For our new construction, we fix a point ∗ in U1, and suppose that ∗ is not in any other Ui (we can ensure this is the case by replacing Ui with Ui \ {∗}). We add a copy U0 of U1 to our cover, and set i,s n(cid:71) i=0 n(cid:91) i=0 Y := Ui = (cid:8)(s, i) : 0 ≤ i ≤ n and s ∈ Ui (cid:9). We define a relation ∼ on Y by (s, i) ∼ (s, j) if s (cid:54)= ∗, (∗, 1) ∼ (∗, 1), and (∗, 0) ∼ (∗, 0). This is an equivalence relation, and we define X to be the quotient space and ψ : Y → X to be the quotient map. Thus X consists of a copy of S \{∗} with the subspace topology, and two closed points ψ(∗, 0), ψ(∗, 1) whose open neighbourhoods are the images under ψ of open sets U × {0} and U × {1}, respectively. Lemma 7.5. The function ψ : Y → X is a surjective local homeomorphism. Proof. Quotient maps are always continuous and surjective, and ψ is injective on each Ui × {i}. So it suffices to see that ψ is open, and for this, it suffices to see that for each 16 CLARK, AN HUEF, AND RAEBURN open set W in Ui, ψ(W × {i}) is open in X. By definition of the quotient topology, we need to show that ψ−1(ψ(W × {i})) is open in Y . If ∗ is not in W , then ψ−1(ψ(W × {i})) = (cid:91) ψ−1(ψ(W × {0})) = {W × {0}} ∪(cid:16) (cid:91) {j:Uj∩W(cid:54)=∅} If ∗ ∈ W and i = 0, then (W ∩ Uj) × {j}. ((W \ {∗}) ∩ Uj) × {j}(cid:17) {j:Uj∩W(cid:54)=∅} is open, and similarly for ∗ ∈ W and i = 1. So ψ−1(ψ(W × {i})) is always open, as (cid:3) required. We extend λ to an alternating cocyle on the cover {U0, U1,··· , Un} by setting λ0jk = λ1jk. Then the formula σ(cid:0)((s, i)(s, j)), ((s, j)(s, k))(cid:1) = λijk(s) defines a continuous 2-cocycle σ on R(ψ). Since ψ is a local homeomorphism it follows from Proposition 4.3 that R(ψ) := {(y, z) ∈ Y × Y : ψ(y) = ψ(z)} is a locally compact, Hausdorff and ´etale groupoid, and that C∗(R(ψ), σ) = C∗ r (R(ψ), σ) is a Fell algebra with spectrum homeomorphic to X. The ∗-algebra structure on Cc(R(ψ), σ) is given by f∗((s, i), (s, j)) = f ((s, j), (s, i))λiji(s) = f ((s, j), (s, i)) (7.5) (f ∗ g)((s, i), (s, j)) = f ((s, i), (s, k))g((s, k), (s, j))λikj(s), (cid:88) {k:s∈Uk} and if (s, i) is a unit in R(ψ) then the induced representation Indσ (cid:96)2({(s, j) : s ∈ Uij}) according to the formula (7.6) (cid:88) (s,i)(f )ξ)(s, j) = (Indσ f ((s, j), (s, k))ξ(s, k)λjki(s). (s,i) acts in (cid:96)2(s−1((s, i))) = {k:s∈Uk} Lemma 7.6. There is a homomorphism π0 : C∗(R(ψ), σ) → C such that π0(f ) = f ((∗, 0), (∗, 0)) for f ∈ Cc(R(ψ), σ). Proof. The inverse image s−1((∗, 0)) consists of the single point ((∗, 0), (∗, 0)), so the Hilbert space (cid:96)2(s−1(∗, 0)) is one dimensional, and Indσ (∗,0)(f ) is multiplication by the complex number f ((∗, 0), (∗, 0))λ000(1) = f ((∗, 0), (∗, 0)). In other words, the representation Indσ (∗,0) of Cc(R(ψ), σ) has the property we require of π0. Since the reduced norm is f (cid:55)→ supy∈Y {(cid:107) Indσ (∗,0) is bounded for the (cid:3) reduced norm, and extends to a representation on C∗ Lemma 7.7. Define V0 := U0 \ {∗}, Vi := Ui for i ≥ 1 and V := {V0, V1,··· , Vn}. Then the ideal ker π0 is isomorphic to the C∗-algebra A(V, λijk) of [29, Example 5.23]. Proof. The maps φij : (Vi × {i}) ×ψ (Vj × {j}) → S defined by φij : ((s, i), (s, j)) (cid:55)→ s are homeomorphisms of (Vi × {i}) ×ψ (Vj × {j}) onto Vij. Thus for f ∈ Cc(R(ψ)) such that π0(f ) = 0, we can define fij : Vij → C by fij = f ◦ φ−1 ij . For {i, j} (cid:54)= {0, 1}, the function fij has compact support, and extends uniquely to a continuous function fij on y (f )(cid:107)}, Indσ r (R(ψ), σ) = C∗(R(ψ), σ). FELL ALGEBRAS 17 S with support in Vij; because f ((∗, 0), (∗, 0)) = π0(f ) = 0, the function f01 vanishes on the boundary of V01 = V0, and extends to a continuous function f01 on S which vanishes off V01. At this point, we have constructed a map φ : f (cid:55)→ (fij) of I0 := Cc(R(ψ)) ∩ ker π0 into the underlying set of A(V, λijk). Since fij(s) = f ((s, i), (s, j)), a comparison of (7.5) If s ∈ Vi, then with (7.3) shows that φ is a homomorphism. a comparison of (7.6) with (7.4), and an argument similar to the proof of Lemma 7.2, show that πi,s ◦ φ is unitarily equivalent to the representation Indσ (s,i). Thus φ is isometric for the reduced norm on I0, and since the range is dense in A(V, λijk), φ extends to an isomorphism of the closure ker π0 onto A(V, λijk). (cid:3) Remark 7.8. If we delete the point (∗, 0) from X, we recover the original space S, the groupoid R(ψ) is the one associated to the cover V of S in Remark 3 on page 399 of [27], and the isomorphism Φ of C∗ r (R(ψ), σ) with A(V, λijk) is discussed in that remark. It is also ∗-preserving. Theorem 1 of [27] (or Proposition 5.40 of [29]) implies that A(V, λijk) is a continuous- trace algebra with Dixmier-Douady class δDD(A(V, λijk)) = [λijk] (cid:54)= 0. This implies that the ideal ker π0 in C∗(R(ψ), σ) has δDD(ker π0) (cid:54)= 0. Now Proposition 6.2 implies that δ(C∗(R(ψ), σ)) (cid:54)= 0. 7.3. Epilogue. We started this project looking for a cocycle-based version of the Dixmier- Douady invariant of [13], which would enable us to resolve the issue about compatibility of δ(A) and δDD(A) in [13, Remark 7.10], and to construct concrete families of Fell algebras as in [27]. Since the spectrum X of a Fell algebra is locally locally-compact and locally Hausdorff, it always has covers by open Hausdorff subsets such that the overlaps Uij where cocycles live lie inside large Hausdorff subsets of X. So it seemed reasonable that cocycle-based arguments might work. As we progressed, we realised how crucially the steps by which one refines covers, as in the proof of [29, Proposition 5.24], for example, depend on the existence of locally finite refinements. In the example of §7.1, this local finiteness fails spectacularly. So even though we know that that the algebra in §7.1 is a Fell algebra, and even though we know it must have vanishing Dixmier-Douady invariant, it is hard to see how a cocycle-based theory could accommodate it. The second part of our project has worked to some extent, in that we can see how to build lots of Fell algebras from ordinary Cech cocycles. However, we can also see that the possibilities are almost limitless, and at this stage there seems little hope of finding a computable invariant. Appendix A. Twisted groupoid C∗-algebras There are several different ways of twisting the construction of a groupoid C∗-algebra. They include: (a) Renault's C∗(G, σ) associated to a 2-cocycle σ : G(2) → T on a groupoid G from (b) Kumjian's C∗(Γ; G)Kum associated to a twist Γ over a principal, ´etale groupoid G (c) Muhly and Williams' C∗(E; G)MW associated to an extension E of a groupoid G [30, II.1] (which we discuss in §2 and use in §5 and §7); in [18, §2] (which we use in §6); by T in [21, §2] (which we use in the proof of Theorem 5.1); 18 CLARK, AN HUEF, AND RAEBURN (d) the reduced C∗-algebras C∗ r (G, σ) and C∗ r (E; G)MW corresponding to (a) and (c), respectively. Here we only consider second-countable, locally compact, Hausdorff and principal group- oids G with a left Haar system λ = {λu}. Let σ : G(2) → T be a continuous normalised 2-cocycle on G. Following [30, page 73] we denote by Gσ the associated extension of G by T: thus Gσ is the groupoid T× G with the product topology, with range and source maps r(z, α) = (1, r(α)) and s(z, α) = (1, s(α)), multiplication (w, α)(z, β) = (wzσ(α, β), αβ) and inverse (z, α)−1 = (z−1σ(α, α−1)−1, α−1). Then Gσ is a locally compact Hausdorff groupoid with left Haar system σ = {σu}, where σu is the product of the normalised Haar measure on T and λu. A twist Γ over a principal ´etale groupoid G has an underlying principal T-bundle over G, and in [18, page 985] Kumjian observes that the twists whose underlying bundle is trivial are in one-to-one correspondence with continuous 2-cocycles σ. Set Cc(Gσ; G) := {f ∈ Cc(Gσ) : f (z · γ) = zf (γ) for z ∈ T, γ ∈ Gσ}. It is easy to check that Cc(Gσ; G) is a ∗-subalgebra of the usual convolution algebra Cc(Gσ). The C∗-algebra C∗(Gσ; G)MW is by definition the completion of Cc(Gσ; G) in the supremum norm (cid:107)f(cid:107) = sup{(cid:107)L(f )(cid:107)}, where L ranges over a collection of appropriately continuous ∗-representations of Cc(Gσ; G). By [4, Proposition 3.7], C∗(Gσ; G)MW is a direct summand of C∗(Gσ). We know from [4, Lemma 3.1] that the map ρ : Cc(Gσ; G) → Cc(G, σ) defined by ρ(f )(α) = f (1, α) is a ∗-isomorphism of Cc(Gσ; G) onto Renault's twisted convolution algebra Cc(G, σ) (the algebra Cc(Gσ; G) is denoted Cc(Gσ,−1) in [4]), and that ρ extends to an isomorphism of C∗(Gσ; G)MW onto C∗(G, σ). Let u ∈ G(0). For the proof of Theorem 5.1 we need to know that the isomorphism ρ sends the class of certain irreducible representations Lu of C∗(Gσ; G)MW defined in [21, §3] to the class of the representation Indσ u of C∗(G, σ). The Hilbert space of Lu is the u := {g ∈ Cc(Gσ; G) : supp g ⊂ T × s−1(u)} with respect to the completion Hu of H 0 u. Then Lu(f )g = f ∗ g, where the convolution takes place in Cc(Gσ; G). We compute using the formulas in [4, Remark 2.3] and the cocycle identity for the triple (β, β−1, α): inner product (f g) = (cid:82) G f (1, α)g(1, α) dλu(α). Let f ∈ Cc(Gσ; G) and g ∈ H 0 (cid:90) (cid:90) (cid:90) (cid:90) f (1, β)g(cid:0)(1, β)−1(1, α)(cid:1) dλr(α)(β) f (1, β)g(cid:0)σ(β, β−1)σ(β−1, α), β−1α(cid:1) dλr(α)(β) f (1, β)g(cid:0)σ(β, β−1α)σ(ββ−1, α), β−1α(cid:1) dλr(α)(β) f (1, β)g(1, β−1α)σ(β, β−1α) dλr(α)(β). = = G G G = (Lu(f )g)(1, α) = (A.1) G Let U : Cc(Gσ; G) → L2(s−1(u), λu) be the map defined by (U g)(α) = g(1, α). Then U ex- tends to a unitary U from the Hilbert space Hu of Lu onto the Hilbert space L2(s−1(u), λu) FELL ALGEBRAS 19 of Indσ u. We have (Indσ u(ρ(f ))U (g))(α) = = (cid:90) (cid:90) G ρ(f )(β)U (g)(β−1α)σ(β, β−1α) dλr(α)(β) f (1, β)g(1, β−1α)σ(β, β−1α) dλr(α)(β), G which is (U (Lu(f )g))(α) using (A.1). We have proved: Lemma A.1. Let G be a principal groupoid with Haar system and σ : G(2) → T a u ◦ρ is unitarily equivalent to continuous normalised 2-cocycle. Then for each unit u, Indσ Lu. Now let G be a principal and ´etale groupoid. It is shown in [18, pages 977 -- 8] that there is a positive-definite C0(G(0))-valued inner product on Cc(Gσ; G), and that the action of Cc(Gσ; G) by left multiplication on itself extends to an action by adjointable operators on the Hilbert-module completion H(Gσ). Then the C∗-algebra C∗(Gσ; G)Kum is by definition the completion of Cc(Gσ; G) in L(H(Gσ)). Proposition A.2. Let G be an ´etale and principal groupoid, and σ : G(2) → T a contin- uous normalised 2-cocycle. The homomorphism ρ of Cc(Gσ; G) onto Cc(G, σ) extends to an isomorphism of C∗(Gσ; G)Kum onto the reduced crossed product C∗ r (G, σ). In view of what we already know about ρ from [4], it suffices to check that ρ is iso- metric for the given norm on Cc(Gσ) ⊂ L(H(Gσ)) and the reduced norm on Cc(G, σ). The general theory of Hilbert bimodules says that, if π is a faithful representation of C0(G(0)), then the induced representation H(Gσ)-Ind π is faithful on L(H(Gσ)). We can u∈G(0) u, and then H(Gσ)-Ind π = in particular take π to be the atomic representation(cid:76) (cid:76) u∈G(0) H(Gσ)-Ind u. So Proposition A.2 follows from the following lemma. Lemma A.3. Let G be an ´etale and principal groupoid, and σ : G(2) → T a continuous normalised 2-cocycle. For each u ∈ G(0), the representation H(Gσ)-Ind u is unitarily equivalent to the representation (Indσ Proof. Since G is ´etale, the representation Indσ formula u of C∗(G, σ) acts on (cid:96)2(s−1(u)) by the u) ◦ ρ of Cc(Gσ; G). (cid:88) h(β)ξ(β−1α)σ(β, β−1α) (Indσ u(h)ξ)(α) = (A.2) r(β)=r(α) for h ∈ Cc(G), ξ ∈ (cid:96)2(s−1(u)), α ∈ G. The representation H(Gσ)-Ind u acts on (the completion of) Cc(Gσ; G) ⊗C0(G(0)) C, which is Cc(Gσ; G) with the inner product (f g) = (g∗ ∗ f )(u). The Haar system on Gσ is the product of the normalised Haar measure on T (cid:16) (cid:88) and the counting measure on s−1(u). Let f, g ∈ Cc(Gσ; G). We have (f g) = (g∗ ∗ f )(u) = (cid:16) (cid:88) (cid:90) (cid:16) (cid:88) (cid:88) g∗(z, α−1)f (z, α) (cid:17) (cid:17) g(z, α)f (z, α) (cid:90) (cid:90) (cid:17) dz = s(α)=u dz T s(α)=u T = zg(1, α)zf (1, α) dz = f (1, α)g(1, α), T s(α)=u s(α)=u and it follows that the Hilbert space of H(Gσ)-Ind u is the space Hu. The action of H(Gσ)-Ind u(f ) on g is by multiplication, so H(Gσ)-Ind u(f ) = Lu(f ). Now the result (cid:3) follows from Lemma A.1. 20 CLARK, AN HUEF, AND RAEBURN Corollary A.4. Suppose that G is a second-countable, locally compact, Hausdorff, ´etale, principal and Cartan groupoid, and that σ is a continuous normalised 2-cocycle on G. Then (A.3) C∗(Gσ; G)MW ∼= C∗(G, σ) = C∗ r (G, σ) ∼= C∗(Gσ; G)Kum. Proof. The first isomorphism is from [4, Lemma 3.1], the equality in the middle is from (cid:3) Theorem 5.1, and the second isomorphism is from Proposition A.2. Remark A.5. We can relax the hypothesis "G is Cartan" in Corollary A.4. That hypoth- esis was used in the proof of Theorem 5.1 to see that u (cid:55)→ [Lu] is a homeomorphism of G(0)/G onto (C∗(Gσ; G)MW)∧; then Lemma A.1 implies that every irreducible representa- tion of C∗(G, σ) is induced, and we have equality in the middle of (A.3). However, if we merely know that "G(0)/G is T0", then we can use [6, Theorem 3.4] in place of [6, Propo- sition 3.2] in the proof of Theorem 5.1, still deduce that u (cid:55)→ [Lu] is a homeomorphism, and follow the rest of the argument to get C∗(G, σ) = C∗ r (G, σ). Appendix B. The ideal of continuous-trace elements If A is a C∗-algebra, we write m(A) for the ideal spanned by the positive elements a such that π (cid:55)→ tr(π(a)) is continuous on A, as in [7, §4.5.2]. The closure of m(A) is an ideal in A, which we call the ideal of continuous-trace elements. Proposition B.1. Suppose that A is a C∗-algebra. Then U := {π ∈ A : π has a closed Hausdorff neighbourhood} is an open Hausdorff subset of A. Proof. Let x ∈ U , and choose a closed Hausdorff neighbourhood N of x. Then for each point y in the interior int N , N is a closed Hausdorff neighbourhood of y, and hence y ∈ U . Thus U is open. To see that U is Hausdorff, let x, y ∈ U , and choose closed Hausdorff neighbourhoods Nx of x and Ny of y. If x ∈ int Ny, then since Ny is Hausdorff we can choose open sets Vx and Vy in Ny such that x ∈ Vx, y ∈ Vy and Vx ∩ Vy = ∅; then Wx := Vx ∩ (int Ny) and Wy = Vy ∩ (int Ny) are open subsets of int Ny with the same property, and they are open in A. A similar argument works if y ∈ Nx. It remains to deal with the case where x /∈ int Ny and y /∈ int Nx. Since A is locally locally-compact, int Ny contains a compact neighbourhood K of y. Since Ny is Hausdorff, K is closed in Ny, and since Ny is closed in A, K is closed in A. Now (int Nx)\ K is open in A, and contains x because x /∈ int Ny. Thus (int Nx) \ K and int K are disjoint open neighbourhoods of x and y. (cid:3) Corollary B.2. Suppose that A is a C∗-algebra. Then V := {π ∈ A : π is a Fell point and has a closed Hausdorff neighbourhood} is an open Hausdorff subset of A. Proof. The set of Fell points is an open subset of A, so V is the intersection of two open sets. It is Hausdorff because it is a subset of the Hausdorff set U of Proposition B.1. (cid:3) Corollary B.3. Suppose that A is a C∗-algebra. Then the set V of Corollary B.2 is the spectrum of the ideal of continuous-trace elements of A. FELL ALGEBRAS 21 Proof. We show that π ∈ V if and only if π satisfies conditions (i) and (ii) of [15, Lemma 2.2], and then our corollary follows from that lemma. First suppose that π ∈ V . Then since V is open it is the spectrum of an ideal J of A. Since π is a Fell point of A it is also a Fell point of J. To see this, let a ∈ A+ and W an open neighbourhood of π in A such that ρ(a) is a rank-one projection for all ρ ∈ W . Let f ∈ Cc( J)+ such that f is identically one on a neighbourhood W (cid:48) of π in J. Using the Dauns-Hofmann theorem, f · a is in J +, and ρ(f · a) = f (ρ)ρ(a) = ρ(a) is a rank-one projection for ρ ∈ W ∩ W (cid:48). Thus J is a continuous-trace algebra by [7, Proposition 4.5.4], and π satisfies (i). To verify (ii), note that π has a closed Hausdorff neighbourhood N , and a neighbour- hood base of compact sets [7, Corollary 3.3.8]. Since π ∈ int N , the compact neighbour- hoods K with K ⊂ int N also form a neighbourhood base. Since N is Hausdorff and closed, such K are also closed in A. Next suppose that π satisfies (i) and (ii). Then π belongs to the spectrum of a continuous-trace ideal J, and is a Fell point in J; since J is open in A, it trivially follows that π is a Fell point in A and that J is an open Hausdorff neighbourhood of π. Now (ii) implies that J contains a closed neighbourhood, and π ∈ V . (cid:3) References C∗-dynamical system, Proc. London Math. Soc. 96 (2008), 545 -- 581. [1] R.J. Archbold and A. an Huef, Strength of convergence and multiplicities in the spectrum of a [2] R.J. Archbold and D.W.B. Somerset, Transition probabilities and trace functions for C∗-algebras, Math. Scand. 73 (1993), 81 -- 111. Math. Soc., to appear; arXiv:1105.0718. 251 -- 266. [3] N. Bourbaki, General Topology, Hermann, Paris, 1966. [4] J.H. Brown and A. an Huef, Decomposing the C∗-algebras of groupoid extensions, Proc. Amer. [5] L.O. Clark, Classifying the types of principal groupoid C∗-algebras, J. Operator Theory 57 (2007), [6] L.O. Clark and A. an Huef, The representation theory of C∗-algebras associated to groupoids, Math. [7] J. Dixmier, C∗-Algebras, North-Holland, Amsterdam, 1977. [8] J. Dixmier and A. Douady, Champs continus d'espaces hilbertiens et de C∗-alg`ebres, Bull. Soc. Proc. Cambridge Philos. Soc. 153 (2012), 167 -- 191. Math. France 91 (1963), 227 -- 284. [9] J.M.G. Fell, The structure of algebras of operator fields, Acta Math. 106 (1961), 233 -- 280. [10] P. Green, C∗-algebras of transformation groups with smooth orbit space, Pacific J. Math. 72 (1977), 71 -- 97. [11] R. Hazlewood and A. an Huef, Strength of convergence in the orbit space of a groupoid, J. Math. [12] A. an Huef, The transformation groups whose C∗-algebras are Fell algebras, Bull. London Math. Anal. Appl. 383 (2011), 1 -- 24. Soc. 33 (2001), 73 -- 76. [13] A. an Huef, A. Kumjian and A. Sims, A Dixmier-Douady theorem for Fell algebras, J. Funct. Anal. 260 (2011), 1543 -- 1581. [14] A. an Huef, I. Raeburn and D.P. Williams, Properties preserved under Morita equivalence, Proc. [15] A. an Huef and D.P. Williams, Ideals in transformation-group C∗-algebras, J. Operator Theory 48 Amer. Math. Soc. 135 (2007), 1495 -- 1503. (2002), 535 -- 548. [16] J.L. Kelley, General Topology, Van Nostrand, New York, 1955. [17] A. Kumjian, Preliminary algebras arising from local homeomorphisms, Math. Scand. 52 (1983), [18] A. Kumjian, On C∗-diagonals, Canad. J. Math. 38 (1986), 969 -- 1008. 269 -- 278. 22 [19] A. Kumjian, On equivariant sheaf cohomology and elementary C∗-bundles, J. Operator Theory 20 CLARK, AN HUEF, AND RAEBURN (1988), 207 -- 240. Funct. Anal. 144 (1997), 505 -- 541. [20] A. Kumjian, D. Pask, I. Raeburn and J. Renault, Graphs, groupoids and Cuntz-Krieger algebras, J. [21] P.S. Muhly and D.P. Williams, Continuous trace groupoid C∗-algebras, Math. Scand. 66 (1990), [22] P.S. Muhly and D.P. Williams, Continuous trace groupoid C∗-algebras II, Math. Scand. 70 (1992), 231 -- 241. 127 -- 145. [23] J.R. Munkres, Topology, second edition, Prentice Hall, Upper Saddle River, 2000. [24] G.K. Pedersen, Analysis Now, Graduate Texts in Math., vol. 118, Springer-Verlag, New York, 1989. [25] G.K. Pedersen, C∗-Algebras and their Automorphism Groups, London Math. Soc. Monographs, vol. 14, Academic Press, London, 1979. [26] I. Raeburn, Graph Algebras, CBMS Regional Conference Series in Mathematics, vol. 103, Amer. [27] I. Raeburn and J.L. Taylor, Continuous trace C∗-algebras with given Dixmier-Douady class, J. Math. Soc., Providence, 2005. Austral. Math. Soc (Series A) 38 (1985), 394 -- 407. C∗-algebras, J. Funct. Anal. 116 (1993), 245 -- 276. veys and Monographs, vol. 60, Amer. Math. Soc., Providence, 1998. [28] I. Raeburn and D.P. Williams, Topological invariants associated with the spectrum of crossed product [29] I. Raeburn and D.P. Williams, Morita Equivalence and Continuous-Trace C∗-Algebras, Math. Sur- [30] J. Renault, A Groupoid Approach to C∗-Algebras, Springer-Verlag, Berlin, 1980. [31] W. Rudin, Real and Complex Analysis, second edition, McGraw Hill, Boston, 1991. [32] A.K. Seda, On the continuity of Haar measure on topological groupoids, Proc. Amer. Math. Soc. 96 [33] D.P. Williams, Crossed Products of C∗-Algebras, Math. Surveys and Monographs, vol. 134, Amer. [34] H.H. Zettl, Ideals in Hilbert modules and invariants under strong Morita equivalence of C∗-algebras, (1986), 115 -- 120. Math. Soc., Providence, 2007. Arch. Math. (Basel) 39 (1982), 69 -- 77. 9054, New Zealand. Department of Mathematics and Statistics, University of Otago, PO Box 56, Dunedin E-mail address: {lclark, astrid, iraeburn}@maths.otago.ac.nz
1207.5090	1	1207	2012-07-21T01:53:22	A rotational approach to triple point obstructions	[ "math.OA" ]	Subfactors where the initial branching point of the principal graph is 3-valent are subject to strong constraints called triple point obstructions. Since more complicated initial branches increase the index of the subfactor, triple point obstructions play a key role in the classification of small index subfactors. There are two strong triple point obstructions called the triple-single obstruction and the quadratic tangles obstruction. Although these obstructions are very closely related, neither is strictly stronger. In this paper we give a more general triple-point obstruction which subsumes both. The techniques are a mix of planar algebraic and connection-theoretic techniques with the key role played by the rotation operator.	math.OA	math	A ROTATIONAL APPROACH TO TRIPLE POINT OBSTRUCTIONS NOAH SNYDER Abstract. Subfactors where the initial branching point of the principal graph is 3-valent are subject to strong constraints called triple point ob- structions. Since more complicated initial branches increase the index of the subfactor, triple point obstructions play a key role in the classification of small index subfactors. There are two strong triple point obstructions called the triple-single obstruction and the quadratic tangles obstruction. Although these obstructions are very closely related, neither is strictly stronger. In this paper we give a more general triple-point obstruction which subsumes both. The techniques are a mix of planar algebraic and connection-theoretic techniques with the key role played by the rotation operator. 1. Introduction The principal graph of a subfactor begins with a type A string and then hits an initial branch point (unless the graph is Ak or A∞). It is natural to stratify subfactors based on how complex this initial branch point is. Furthermore, complex initial branches increase the norm of the graph and thus the index of the subfactor. This means that small index subfactors can only have simple initial branches. The simplest possibility is an initial triple point (in this case the dual graph also begins with a triple point). Subfactors beginning with an initial triple point are subject to strong constraints known as triple point obstructions. For example, a triple point obstruction due to Ocneanu shows that as long as the index is at least 4 the initial triple point must be at odd depth. These triple point obstructions play a crucial role in the classification of small index subfactors [Haa94, MS10, MPPS10, IJMS11, PT10]. The current state of the art of triple point obstructions is given in [MPPS10], but the status is somewhat unsatisfatory as there are two main results neither of which is strictly stronger than the other. One result applies more generally and proves a certain inequality, while the other (due to Jones [Jon03]) has stricter assumptions but replaces the inequality with a finite list of values. The former is proved using connections and the latter using planar algebras. The main result of this paper is a mutual generalization of these two triple point obstructions which proves the stronger conclusion using only the weaker assumptions. As one might expect, this paper uses a mix of connections and planar algebras following [IJMS11]. Furthermore, one can think of this argu- ment as giving an alternate proof of the triple point obstruction from [Jon03]. Date: November 1, 2018. 1 2 NOAH SNYDER Before stating the three relevant results, we fix some notation which we will use throughout the paper. Suppose that N ⊂ M is an n − 1 supertransitive finite index subfactor, let Γ and Γ(cid:48) denote the principal and dual principal graphs. Let [k] denote the quantum number (νk − ν−k)/(ν − ν−1) where ν is a number such that the index is [2]2. Let β and β(cid:48) denote the initial triple points at depth n− 1 (which is necessarily odd by Ocneanu's obstruction), let α1 and γ1 be the vertices at depth n − 2, let α2 and α3 be the two vertices at depth n on Γ, and let γ2 and γ3 be the two vertices at depth n on Γ(cid:48). We will conflate vertices with the corresponding simple bimodules and the corresponding simple projections in the planar algebra. Assume without loss of generality that dim α2 ≥ dim α3 and dim γ2 ≥ dim γ3. Theorem 1 (Triple-single obstruction). [MPPS10, Thm 3.5] If γ3 is 1-valent, then dim(α2) − dim(α3) ≤ 1. Theorem 2 (Quadratic tangles obstruction). [Jon03] Suppose that γ3 is 1- valent and that γ2 is 3-valent, then r + 1 r = λ + λ−1 + 2 [n][n + 2] + 2 where λ is the scalar by which rotation acts on the 1-dimensional perpendicular complement of Temperley-Lieb at depth n and r = dim(α2) dim(α3). Since λ is a root of unity, we know that −2 ≤ λ + λ−1 ≤ 2. Hence the QT obstruction gives an inequality, and (as observed by Zhengwei Liu) this inequality turns out to be precisely the one in the triple-single obstruction [MPPS10, Lemma 3.3]. Thus the QT obstruction is stronger (replacing an interval of possibilities with a finite list) when both apply, but the triple- single obstruction has a weaker assumption. The main result of this paper is the following mutual generalization of Theorems 1 and 2. Theorem 3. Suppose that γ3 is 1-valent, then λ + λ−1 + 2 [n][n + 2] r + 1 r = + 2. The main ideas in this paper came out of joint work with Scott Morrison, and I would like to thank him for many helpful conversations. I would also like to thank Vaughan Jones, Dave Penneys, and Emily Peters. This work was supported by an NSF Postdoctoral Fellowship at Columbia University and DARPA grant HR0011-11-1-0001. 2. Background We quickly summarize the key idea of [IJMS11, §5.2], which is that the action of rotation on the planar algebra can be read off from the connection. Since rotational eigenvalues must be roots of unity this gives highly nontrivial constraints on candidate connections. We assume that the reader is familiar with both planar algebras and connections, referring the reader to [IJMS11] for more detail. A ROTATIONAL APPROACH TO TRIPLE POINT OBSTRUCTIONS Given a subfactor N ⊂ M we get a certain collection of matrices called a connection. This connection depends on a choice of certain intertwiners, and thus is only well-defined up to gauge automorphisms. Let the branch matrix U denote the 3-by-3 matrix coming from the connection at the initial branch vertex of Γ. The key idea from [IJMS11, §5.2] is that there is a canonical gauge choice for U , called the diagrammatic branch matrix, coming from the planar algebra. This choice is both easy to recognize and has nice properties, as captured by the following two results. 3 Lemma 4. [IJMS11, Lemma 5.6] When n is odd the diagrammatic branch matrix is characterized within its gauge class by the property that all the entries in the first row and column are positive real numbers. Proposition 5. Let U be the diagrammatic branch matrix for a subfactor with an initial triple point. Suppose that x is an n-box in the perpendicu- lar complement of Temperley-Lieb, and write x = a2 . Let (c1, c2, c3) = U (0, a2, a3). Then c1 = 0 and c2 dim α3 2 (x). dim α2 γ3√ is ρ 1 + a3 α2√ α3√ + c3 γ2√ dim γ2 dim γ3 Proof. This is a restatement of [IJMS11, Corollary 5.3] in our special case. (cid:3) See [IJMS11, pp. 18 -- 19] for a worked example. In order to apply the previous proposition we will want an explicit formula for vectors in the perpendicular complement to Temperley-Lieb in the n-box space and the action of rotation there. Recall that the rotation ρ preserves shading and thus is an endomorphism of each box space, while ρ 1 2 changes rotation and thus is a map from one box space to a different box space. We will use λ to denote the scalar by which ρ acts on the 1-dimensional perpendicular complement to Temperley-Lieb in the n-box space. Note that this is an nth root of unity. Lemma 6. Let r = dim α2 dim α3 T = 1√ Lieb. rα3 and rγ3 are each in the perpendicular complement of Temperley- √ λ is some square root of the rota- tional eigenvalue for the action of rotation on the perpendicular compliment of Temperley-Lieb. √ λ T , where r α2 − √ and r = dim γ2 dim γ3 . Then T = 1√ Furthermore ρ 1 γ2 − √ 2 (T ) = r Proof. These calculations (with slightly different conventions) were done in an early version of [Jon03]. Seeing that T and T are perpendicular to Temperley- Lieb is straightforward (you only need to work out their inner product with two specific Jones-Wenzl projections). Since half-click rotation preserves Temperley- Lieb and is an isometry, it also preserves the perpendicular complement of 2 (T ) is some scalar multiple of T . To work out which Temperley-Lieb. Thus ρ 1 scalar multiple this is you compute their norms. This tells you that the square (cid:3) of this scalar is λ. √ √ λ all Remark 7. There are many square roots in this paper. Other than square roots are positive square roots of positive numbers. λ will always be chosen such that the previous lemma works. In the final statement of the main theorem no λ appears, so this subtlety is not very important. √ 4 NOAH SNYDER Combining the previous two results we have the following concrete state- ment, which will supply the main ingredient of our proof of Theorem 3. Corollary 8. The diagrammatic branch matrix U sends (0,(cid:112)dim(α3),−(cid:112)dim(α2)) (cid:55)→ λ(0,(cid:112)dim(γ3),−(cid:112)dim(γ2)). √ 3. Proof of Theorem 3 The idea of this argument is that having a 1-valent vertex allows us to solve for the branch matrix, and thus we can read off the rotational eigenvalue (since the diagrammatic branch matrix acts on the appropriate vectors by rotation). This gives an identity between the dimensions of objects and the rotational eigenvalue. We begin with a quick calculation of the branch matrix following the proof of the triple-single obstruction [MPPS10, Thm 3.1]. Since α1, γ1, β, and β(cid:48) are in the initial string there dimensions are [n − 1], [n − 1], [n], and [n], respectively. Since γ3 is 1-valent, we have dim γ2 = [n+2] [2] . Us- ing the 1-valence of γ3 the normalization condition on connections determines the magnitude of several of the entries in the branch matrix. Furthermore, unitarity of U allows us to work out several more of the entries. In particular, the branch matrix is gauge equivalent to the matrix below, where p = dim(α2) and q = dim(α3), where σ and τ are unknown phases, and where ? denotes unknown entries which will play no role in the calculation. [2] and dim γ3 = [n]  1 [n] (cid:113) [n−1] (cid:113) [n−1][n+2] [2][n] [2][n]2 (cid:112)[n − 1]p (cid:112)[n − 1]q (cid:113) q (cid:113) p σ [2][n] ? τ [2][n] ?  U = The first row and column of this matrix are clearly positive, so by Lemma 4 we see that U is the diagrammatic branch matrix. Remark 9. This matrix is the transpose of the matrix found in [MPPS10] because the calculation there is done for Γ(cid:48) instead of Γ. As shown in [IJMS11], the diagrammatic branch matrices of Γ and Γ(cid:48) are always transposes. We would like to solve for σ and τ . Orthogonality of the first two rows of U tells us that 1 + σp + τ q = 0. Although 1 + σp + τ q = 0 is one equation in two unknowns, it actually deter- mines σ and τ since they are phases: σ = −1 + τ q p 1 = σ¯σ = 1 + τ q 1 + ¯τ q 1 + (τ + ¯τ )q + q2 p τ + ¯τ = = p p2 − q2 − 1 p2 . q q2−p2−1 . A ROTATIONAL APPROACH TO TRIPLE POINT OBSTRUCTIONS 5 This determines the real part of τ , and thus τ itself. Similarly, σ + ¯σ = p Now that we have a very explicit understanding of U we apply it to a rotational eigenvector. Corollary 8 tells us that U sends √ q,−√ p) (cid:55)→ (0, √ λ (cid:115) (cid:33) . [n + 2] [2] ,− [n] [2] (cid:115) (cid:32) (cid:115) 0, σ − τ = √ λ [n + 2][n] . pq Looking at the middle coordinate of that identity we see that (cid:115) Comparing the real parts of both sides yields the following formula √ λ + ( 1√ λ ) [n + 2][n] pq = (σ + ¯σ) − (τ + ¯τ ) q2 − p2 − 1 − p2 − q2 − 1 = (q − p)((p + q)2 − 1) p q (q − p)([n + 1]2 − 1) pq (q − p)([n][n + 2]) = = pq Squaring both sides and rearranging proves the theorem. pq . = √ (cid:16)(cid:113) [n+2][n] (cid:17) Remark 10. You might guess that σ − τ = would give a second condition coming from the imaginary parts. In fact there's no new information there, because the two sides automatically have the same norm. λ pq References [Haa94] √ Uffe Haagerup. Principal graphs of subfactors in the index range 4 < [M : N ] < 3+ 2. In Subfactors (Kyuzeso, 1993), pages 1 -- 38. World Sci. Publ., River Edge, NJ, 1994. MR1317352. [IJMS11] M. Izumi, V. F. R. Jones, S. Morrison, and N. Snyder. Subfactors of index less than 5, part 3: quadruple points. Communications in Mathematical Physics, September 2011. arXiv:1109.3190. Vaughan F. R. Jones. Quadratic tangles in planar algebras, 2003. arXiv:1007. 1158. [Jon03] [MS10] [MPPS10] Scott Morrison, David Penneys, Emily Peters, and Noah Snyder. Classification of subfactors of index less than 5, part 2: triple points. International Journal of Mathematics, 2010. arXiv:1007.2240, accepted June 28 2011. Scott Morrison and Noah Snyder. Subfactors of index less than 5, part 1: the principal graph odometer. Communications in Mathematical Physics, 2010. arXiv:1007.1730, accepted June 28 2011. David Penneys and James Tener. Classification of subfactors of index less than 5, part 4: cyclotomicity. International Journal of Mathematics, 2010. arXiv:1010. 3797, accepted June 28 2011. [PT10]
1204.6431	1	1204	2012-04-28T20:56:37	Two families of Exel-Larsen crossed products	[ "math.OA" ]	Larsen has recently extended Exel's construction of crossed products from single endomorphisms to abelian semigroups of endomorphisms, and here we study two families of her crossed products. First, we look at the natural action of the multiplicative semigroup $\mathbb{N}^\times$ on a compact abelian group $\Gamma$, and the induced action on $C(\Gamma)$. We prove a uniqueness theorem for the crossed product, and we find a class of connected compact abelian groups $\Gamma$ for which the crossed product is purely infinite simple. Second, we consider some natural actions of the additive semigroup $\mathbb{N}^2$ on the UHF cores in 2-graph algebras, as introduced by Yang, and confirm that these actions have properties similar to those of single endomorphisms of the core in Cuntz algebras.	math.OA	math	TWO FAMILIES OF EXEL-LARSEN CROSSED PRODUCTS NATHAN BROWNLOWE AND IAIN RAEBURN Abstract. Larsen has recently extended Exel's construction of crossed products from single endomorphisms to abelian semigroups of endomorphisms, and here we study two families of her crossed products. First, we look at the natural action of the multiplica- tive semigroup N× on a compact abelian group Γ, and the induced action on C(Γ). We prove a uniqueness theorem for the crossed product, and we find a class of con- nected compact abelian groups Γ for which the crossed product is purely infinite simple. Second, we consider some natural actions of the additive semigroup N2 on the UHF cores in 2-graph algebras, as introduced by Yang, and confirm that these actions have properties similar to those of single endomorphisms of the core in Cuntz algebras. 1. Introduction In [14], Exel considered dynamical systems (A, α, L) consisting of an endomorphism α of a unital C ∗-algebra A and a transfer operator L for α, which is a positive linear map L : A → A satisfying L(α(a)b) = aL(b) for a, b ∈ A. To each (A, α, L), he associated a crossed product C ∗-algebra A ⋊α,L N. Exel showed how to realise the Cuntz-Krieger algebras as Exel crossed products [14], and the analogous construction for nonunital A includes more general graph algebras [8]. Larsen [26] has recently extended Exel's construction to actions of abelian semigroups, and we will call her algebras Exel-Larsen crossed products. They are by definition the Cuntz-Pimsner algebras of product systems of Hilbert bimodules constructed using transfer operators. Larsen considers in particular the natural action of the multiplicative semigroup N× = {a ∈ Z : a > 0} on a compact abelian group Γ, and the associated action α : N× → End C(Γ) given by αa(f )(g) = f (ga). She shows that under certain conditions (see (G1 -- 3) in §3 below), there is a compatible action L of N× by transfer operators obtained by averaging over the solutions h of ha = g. When Γ is the group Z of integral ad`eles, for example, her construction yields the Hecke C ∗-algebra of Bost and Connes [3] (see [26, Remark 5.10]); when Γ is the circle group T, C(T) ⋊α,L N× is the purely infinite simple C ∗-algebra QN of Cuntz [10] (see [5, Theorem 5.2] and [19]); and Brownlowe has shown that the C ∗-algebras of higher-rank graphs can be obtained via a modification of her construction [4]. Here we investigate two families of Exel-Larsen crossed products. The first family are the C(Γ) ⋊α,L N× from [26, §5]. We find general conditions on Γ which ensure that C(Γ)⋊α,L N× is purely infinite and simple, which apply in particular when Γ is a solenoid or a higher-dimensional torus. Our main tools are Yamashita's work on the C ∗-algebras of topological higher-rank graphs [33], which are also Cuntz-Pimsner algebras of product Date: January 12, 2021. This research has been supported by the Australian Research Council and the University of Otago. The authors thank Astrid an Huef for helpful comments and suggestions. 1 2 NATHAN BROWNLOWE AND IAIN RAEBURN systems, and the structure theory for compact groups, including some surprisingly new results about coverings of solenoids [17]. The second family of interest to us involves actions of the additive semigroup N2 on the core in the C ∗-algebras of rank-2 graphs with a single vertex. These graphs Λθ were introduced in the context of non-selfadjoint operator algebras in [21], and their C ∗-algebras have been studied in [12, 35, 36]; the actions of N2 of interest to us were discussed in [35]. The papers [35, 36] suggest that, when the algebra C ∗(Λθ) is simple, it behaves very much like a Cuntz algebra, which can be viewed as the graph algebra of a rank-1 graph with a single vertex. We compute the Exel crossed products for these actions, in direct parallel with a result for the Cuntz algebras in [20, Theorem 6.6]. We begin with a quick review of product systems in §2. Then in §3 we discuss general properties of Exel-Larsen crossed products arising from compact abelian groups. Our main general results are a uniqueness theorem and a theorem which identifies a family of compact groups whose crossed products are purely infinite and simple. In §4, we discuss examples, including higher-dimensional tori and solenoids. The last section contains our work on Exel-Larsen crossed products associated to 2-graphs. We need to use a theorem of Davidson and Yang about the periodicity of 2-graphs with a single vertex, and in an Appendix we give a short graph-theoretic proof of their theorem. 2. Product systems of Hilbert bimodules We are interested in two families of C ∗-algebras, both of which are by definition the Cuntz-Pimsner algebras of product systems of Hilbert bimodules over a fixed C ∗-algebra. So we begin by setting out our conventions. A Hilbert bimodule over a C ∗-algebra A is a right Hilbert A-module M with a left action of A by adjointable operators, implemented by a homomorphism φ of A into the C ∗-algebra L(M) of adjointable operators. (In some papers, such M are described as "right-Hilbert A -- A bimodules", or "correspondences over A.") Suppose that P is an abelian semigroup with an identity e; in this paper, P is either the multiplicative semigroup N× of positive integers or the additive semigroup N2. A product system over P is a collection {Mp : p ∈ P } of Hilbert bimodules over the same C ∗-algebra A together with an associative multiplication on Fp∈P Mp such that (m, n) 7→ mn induces an isomorphism of Mp ⊗A Mq onto Mpq; we always assume that the bimodule Me is AAA, and that the products from Me × Mp → Mp and Mp × Me → Mp are given by the module actions of A on Mp. A representation ψ = {ψp : p ∈ P } of a product system in a C ∗-algebra B consists of linear maps ψp : Mp → B such that each (ψp, ψe) is a representation of Mp, and ψ is Cuntz-Pimsner covariant if each (ψp, ψe) is Cuntz-Pimsner covariant. The Cuntz- Pimsner algebra1 O(M) is generated by a universal Cuntz-Pimsner covariant represen- tation kM = {kM,p}. These algebras were introduced by Fowler [16]; we follow [31] in writing Πψ for the representation of O(M) satisfying (Πψ) ◦ kM,p = ψp. Next we suppose that (A, P, α, L) is an Exel-Larsen system as in [26]: thus α : P → End A is an action of P by unital endomorphisms of a unital C ∗-algebra A, and L is 1The potentially different Cuntz-Pimsner algebras O(M ) of [16] and N O(M ) of [31] coincide in our case, because the left actions φp : A → L(Mp) all have range in K(Mp) and our semigroups N× and N2 are lattice-ordered in the sense that every pair has a least upper bound (see [31, §5.1]). EXEL-LARSEN CROSSED PRODUCTS 3 an action of P by transfer operators Lp for αp. We put a bimodule structure on each ALp = A by a · x · b = axαp(b); give it a pre-inner product defined by hx, yiLp = Lp(x∗y); mod out by the vectors of length 0 and complete to get a right Hilbert A-module Mp (denoted MLp in [7]). The left action of A on ALp gives a homomorphism φp of A into the algebra L(Mp) of adjointable operators, and we write qM p (x) or just qp(x) for the image of x ∈ A = ALp in Mp. Larsen proves2 in [26, §3.2] that M :=F Mp is a product system of Hilbert bimodules over A with (2.1) p (x)qM qM r (y) = qM pr (xαp(y)). The Exel-Larsen crossed product A ⋊α,L P is by definition the Cuntz-Pimsner algebra O(M). In the families of Exel-Larsen systems of interest to us, we can make some simplifi- cations. The endomorphisms αp are unital, and the transfer operators Lp are faithful, in the sense that Lp(a∗a) = 0 =⇒ a = 0. The Hilbert bimodules are finitely generated as right modules, so no modding out or completing is required. In fact our bimodules Mp usually have finite orthonormal bases {ei : 1 ≤ i ≤ Np}, and then the identity i=1 Θa·ei,ei implies in particular that φp has range in the algebra K(Mp) of φp(a) = PNp compact operators. 3. Exel-Larsen systems from compact abelian groups Let Γ be a compact abelian group, and let ω be the action of N× by endomorphisms of Γ given by ωa(g) = ga for a ∈ N×. Let α denote the action of N× by endomorphisms of C(Γ) given by αa(f )(g) = f (ωa(g)) = f (ga). We will require that (G1) ωa(Γ) has finite index in Γ for all a ∈ N×; (G2) ker ωa is finite for all a ∈ N×; and (G3) ker ωab = ker ωa · ker ωb for all a, b ∈ N×. In [26, Proposition 5.1] Larsen proved that under the conditions (G1 -- 3), the formula La(f )(g) :=( 1 0 ker ωaPωa(h)=g f (h) if g ∈ ωa(Γ) otherwise (3.1) gives an action L of N× by transfer operators for α. So (C(Γ), N×, α, L) is an Exel-Larsen system. We can now run the system (C(Γ), N×, α, L) through Larsen's construction. We observe as in [28, Lemma 3.3] that C(Γ)La is complete in the norm defined by the inner product. The argument of [2, Example 2.1(b)] shows that each Ma := C(Γ)La has an orthonormal basis with Na := Γ/ωa(Γ) elements, and hence each φa has range in K(Ma). Then C(Γ) ⋊α,L N× := O(M) is generated by the universal Cuntz-Pimsner covariant representation kM of the product system M :=Fa∈N× Ma. We want to use a description of C(Γ) ⋊α,L N× as a topological-graph algebra. The topological graph in question is a topological ∞-graph in the sense of [33], where the 2We prefer the conventions of [7] to those of [26], where Larsen completes Aαp(1) rather than A. For us, αp(1) = 1, so we can just ignore the αp(1). However, one can also ignore it in general: the completion process involves modding out by vectors of norm 0, and every m − mαp(1) has norm 0, so qp(x) = qp(xαp(1)). 4 NATHAN BROWNLOWE AND IAIN RAEBURN rank-∞ case is explicitly allowed. This is relevant to our situation because prime fac- torisation gives an isomorphism of our multiplicative semigroup N× onto the additive semigroup N∞. We choose to suppress this isomorphism by retaining multiplicative notation, and call our graph a topological N×-graph. However, the existence of the iso- morphism N× ∼= N∞ means that the results of [33] apply (as was previously done in [33, Remark 5.6], for example). To help give a feel for our conventions, we describe a standard example. Note straight- away that our multiplicative notation means that the set of vertices in Λ is denoted Λ1 rather than Λ0. Example 3.1. The N×-graph ΩN× has vertex set Ω1 N× = N×, and the morphisms are pairs (a, b) in N× × N× such that a divides b. The range of (a, b) is a, the source of (a, b) is b, and the degree of (a, b) is a−1b. We now describe the topological N×-graph associated to our system (C(Γ), N×, α, L). Proposition 3.2. For a ∈ N× let Γa := {(a, g) : g ∈ Γ}, and give ΛΓ := Sa∈N× Γa the topology in which each Γa is compact and open. Define r, s : ΛΓ → Λ1 := Γ and d : ΛΓ → N× by r(a, g) = g, s(a, g) = ga and d(a, g) = a. Then Λ = (Γ, ΛΓ, r, s, d) is a topological N×-graph with composition (a, g)(b, ga) := (ab, g); Λ is row-finite and has no sources. The only nontrivial part of this assertion is that the source map s is a local homeo- morphism, and this follows from the next lemma. Lemma 3.3. If Γ is a compact abelian group satisfying (G1) and (G2), then each ωa is a proper local homeomorphism. Proof. The homomorphism ωa factors through a continuous isomorphism of Γ/ ker ωa onto ωa(Γ), which is an open subgroup of Γ because it has finite index in Γ. Since the domain of this isomorphism is compact and the range is Hausdorff, it is a homeomor- phism; since the quotient map from Γ to Γ/ ker ωa is open, so is ωa. The map ωa is proper because it is a continuous map from a compact space into a Hausdorff space. Since ωa is open, to see that it is a local homeomorphism it suffices to find an open neighbourhood U of e such that ωaU is one-to-one. Since ker ωa is finite, we can find an open neighbourhood U of e such that gU ∩hU = ∅ for h 6= g in ker ωa. Suppose h1, h2 ∈ U 2 ∈ ker ωa. Since h1 ∈ U and h1 = h1h−1 and ωa(h1) = ωa(h2). Then h1h−1 2 U, we have U ∩h1h−1 2 U 6= ∅. So h1h−1 2 = e, and hence h1 = h2. Thus ωaU is one-to-one. (cid:3) 2 h2 ∈ h1h−1 Proof of Proposition 3.2. Routine calculations show that (Γ, ΛΓ, r, s) is a category. The map r is trivially continuous, Lemma 3.3 implies that s is a local homeomorphism, and d is locally constant, hence continuous. The continuity of multiplication in Γ implies that composition is continuous, and since the product of open sets in Γ is open, composition in Λ is open. To see the factorisation property, suppose that (a, g)(b, ga) can be rewritten as (a, k)(b, l); then taking the ranges of both sides gives g = k, and composability on the right forces l = ka = ga. So Λ is a topological N×-graph. The maps rΓa are bijections, so Λ is row-finite and has no sources. (cid:3) EXEL-LARSEN CROSSED PRODUCTS 5 We can now apply the construction of [33, page 305] to the topological N×-graph Λ, and thus obtain a product system X over N× of Hilbert bimodules over the algebra C(Λ1) = C(Γ). Yamashita's topological-graph algebra O(Λ) is then the Cuntz-Pimsner algebra O(X) of this product system. We want to show that our C(Γ) ⋊α,L N× is naturally isomorphic to O(Λ), and to do this we need to compare the product systems M and X. When we use the obvious identification of C({a} × Γ) with C(Γ) to view Xa as a copy {qX a (x) : x ∈ C(Γ)} of C(Γ), the bimodule Xa is very similar to Ma: indeed, the underlying spaces are the same, the left and right actions are the same, and the only difference is in the inner products, where the one on Xa is given by hqX a (x), qX a (y)i(g) =(Pha=g x(g)y(g) 0 if g ∈ ωa(Γ) otherwise for x, y ∈ C(Γa) = C(Γ); because of the way the source maps are defined in Λ, the product system structure is given by qX We now define ψa : Xa → Ma by ψa(qX a (x). Then for each fixed a ∈ N×, ψa is an isomorphism of Hilbert bimodules. Further, property (G3) implies that a (x)qX b (y) = qX ab(xαa(y)). a (x)) = ker ωa1/2qM ψab(qX a (x)qX b (y)) = ψab(qX ab(xαa(y))) = ker ωab1/2qM ab (xαa(y)) b (y) = ψa(qX = ker ωa1/2 ker ωb1/2qM a (x)qM a (x))ψb(qX b (y)). Thus the {ψa : a ∈ N×} constitute an isomorphism of product systems of Hilbert bimodules, and we have: Proposition 3.4. Let jX = {jX,a} and kM be the canonical Cuntz-Pimsner covariant representations of X and M in O(Λ) := O(X) and C(Γ) ⋊α,L N× := O(M). Then there is an isomorphism Ψ of O(Λ) onto C(Γ) ⋊α,L N× such that Ψ ◦ jX,a = ker ωa1/2kM,a for every a ∈ N×. We now use this topological-graph realisation to prove a uniqueness theorem for the crossed products C(Γ) ⋊α,L N×. Theorem 3.5. Suppose that Γ is a compact abelian group satisfying (G1 -- 3), and that the torsion subgroup Tor Γ has empty interior. If π : C(Γ) ⋊α,L N× → B is a unital homomorphism into a C ∗-algebra B and π ◦ kM,1 is injective, then π is injective. We will deduce Theorem 3.5 from Yamashita's uniqueness theorem for topological k-graphs. We recall that an infinite path in a topological N×-graph Λ with range v is a degree-preserving continuous functor µ : ΩN× → Λ with µ(1) = v. For m ∈ N×, the translate τ m(µ) is the infinite path such that τ m(µ)(a, b) := µ(ma, mb). Following Yamashita, we say that Λ is aperiodic if for each open set V in Λ1, there exists an infinite path µ ∈ Λ∞ with range in V such that m 6= n implies τ m(µ) 6= τ n(µ). Yamashita's uniqueness theorem is about aperiodic graphs, so the next lemma explains why we need the condition on the torsion subgroup in Theorem 3.5. Lemma 3.6. Let Γ be a compact abelian group satisfying (G1 -- 3), and let Λ be the topological N×-graph in Proposition 3.2. 6 NATHAN BROWNLOWE AND IAIN RAEBURN (a) For g ∈ Γ we define µg : ΩN× → Λ by µg(a, b) = (a−1b, ga). Then µg ∈ Λ∞, and for every β ∈ Λ∞ we have β = µr(β). (b) Λ is aperiodic if and only if int(Tor Γ) = ∅. Proof. (a) Fix g ∈ Γ. Then for composable morphisms (a, b), (b, c) in ΩN× we have µg(a, b)µg(b, c) = (a−1b, ga)(b−1c, gb) = (a−1b, ga)(b−1c, (ga)a−1b) = (a−1c, ga) = µg(a, c) = µg((a, b)(b, c)), so µg is a functor, and µ is trivially degree-preserving. So µg is an infinite path in Λ. To see that they all have this form, let β ∈ Λ∞ and take g = r(β). Then for (a, b) ∈ ΩN×, we have β(1, b) = β((1, a)(a, b)) = β(1, a)β(a, b). Since r(β) = g, we have β(1, b) = (b, g), and hence the factorisation property implies that β(1, a) = (a, g) and β(a, b) = (a−1b, ga) = µg(a, b). Hence β = µg. (b) Since each path in Λ∞ = {µg} is determined by its range, we have τ a(µg) 6= τ b(µg) ⇐⇒ r(τ a(µg)) 6= r(τ b(µg)) ⇐⇒ ga 6= gb. Thus Λ is aperiodic if and only if each open set V in Λ1 = Γ contains some g = r(µg) with infinite order. But this is equivalent to saying that the torsion subgroup has empty interior. (cid:3) Proof of Theorem 3.5. The homomorphism π is the integrated form Πρ of a Cuntz- Pimsner covariant representation {ρM,a : a ∈ N×} in B, and then ρM,1 = π ◦ kM,1 is injective. Composing with the isomorphism Ψ of Proposition 3.4 gives a Cuntz-Pimsner covariant representation {ρM,a ◦ Ψ} of X with (ρ ◦ Ψ)1 = ρM,1 injective. Since Tor Γ has empty interior, Lemma 3.6 says that Λ is aperiodic. So Yamashita's uniqueness theorem [33, Theorem 3.11] implies that Π(ρ ◦ Ψ) is injective. Hence Πρ = Π(ρ ◦ Ψ) ◦ Ψ−1 is injective too. (cid:3) Lemma 3.7. Let Γ be a compact abelian group satisfying (G1 -- 3). Then the topological graph Λ of Proposition 3.2 is minimal in the sense of [33, 4.1] if and only if {h ∈ Γ : ha = gb for some a, b ∈ N×} = Γ for every g ∈ Γ. (3.2) Proof. The equivalence of (ii) and (ii) in [33, Theorem 4.7] says that the graph Λ is minimal if and only if Orb(g, µ) := [b∈N× [a∈N× rΓa(s−1 Γa (µ(b))) (3.3) is dense in Λ1 = Γ for every g ∈ Γ and every µ ∈ Λ∞ with r(µ) = g. Lemma 3.6 (a) implies that for each g there is exactly one such path µ, namely the path µg. But for this path, the bth vertex µg(b) is gb, and Orb(g, µg) is the set whose closure appears in (3.2). (cid:3) If G also satisfies int(Tor Γ) = ∅, then Λ is aperiodic, and it follows from [33, Theo- rem 4.7] that (3.2) is equivalent to simplicity of C ∗(Λ) = C(Γ) ⋊α,L N×. For connected groups, (3.2) always holds, and we get the following. Theorem 3.8. Suppose that Γ is a connected compact abelian group satisfying (G2) and int(Tor Γ) = ∅. Then Γ also satisfies (G1) and (G3), and C(Γ) ⋊α,L N× is purely infinite and simple. EXEL-LARSEN CROSSED PRODUCTS 7 The following proof uses the following characterisations of connectedness for a com- pact abelian group Γ (the second holds also for locally compact Γ): (i) [18, Theorem 24.25]: Γ is connected if and only if Γ is divisible, in the sense that for every g ∈ Γ and a ∈ N×, there exists h ∈ Γ with ha = g, (ii) [18, Corollary 7.9]: Γ is connected if and only if for every neighbourhood U of the identity in Γ, we have S∞ n=1 U n = Γ. Proof of Theorem 3.8. We again take Λ as in Proposition 3.2. From (i) above we know that Γ is divisible, which means each ωa is surjective, and we have (G1). As pointed out in [26, Section 5.1], if ωa is surjective, then for each b ∈ N× the map g 7→ ωa(g) is a homomorphism of ker ωab onto ker ωb with kernel equal to ker ωa, and this implies (G3). Since O(Λ) ∼= C(Γ) ⋊α,L N×, it suffices to show that Λ satisfies the conditions of [33, Theorem 4.13]. Since int(Tor Γ) = ∅, Lemma 3.6 says that ΛΓ is aperiodic. Thus we need to show that ΛΓ is minimal and contracting, in the sense of [33, Theorem 4.7(iii)] and [33, Definition 4.12], respectively. We let g ∈ G, and claim that {h ∈ Γ : ha = g for some a ∈ N×} is dense in Γ. Suppose h ∈ Γ and U is an open neighbourhood of the identity e ∈ Γ; we need to find n=1 U n. Since k ∈ hU and a ∈ N× such that ka = g. We know from (ii) that Γ = S∞ Γ is compact, there exists a finite subset F of N× with Γ = Sa∈F U a. For b, c ∈ N× with bc the map u 7→ ueb−1c is an injection of U b in U c. Thus the {U b} are nested, and there exists a ∈ N× with Γ = U a. Now choose h0 ∈ U such that ha 0 = gh−a, and 0 = hagh−a = g, proving the claim. Now then k := hh0 satisfies k ∈ hU and ka = haha Lemma 3.7 implies that Λ is minimal To see that Λ is contracting, consider any open neighbourhood U of e. Then as in the previous paragraph there exists a ∈ N× with Γ = U a. This implies that U is a contracting open set, as defined in [33, 4.11], and hence that Λ is contracting at e, as in [33, 4.12]. The result now follows from [33, Theorem 4.13]. (cid:3) 4. Examples We now apply the general results of the previous section to some specific compact abelian groups Γ. 4.1. Tori. For each positive integer l the torus Tl is a connected compact abelian group. Each ωa satisfies ker ωa = al < ∞, so (G2) holds. We now claim that int(Tor Tl) = ∅, or equivalently that the elements of infinite order are dense in Tl. To see this, let U be an open set in Tl, and choose w ∈ U with first coordinate of the form eiθ for some rational θ; then the exponent in the first coordinate of wa is never an integer multiple of 2πi, and w has infinite order. So Theorem 3.8 gives the following result. Proposition 4.1. For each positive integer l, the crossed product C(Tl) ⋊α,L N× is purely infinite and simple. Remark 4.2. This result is known for l = 1, because C(T) ⋊α,L N× is isomorphic to Cuntz's algebra QN (by [5, Theorem 5.2]), and Cuntz proved in [10] that QN is purely infinite simple. 8 NATHAN BROWNLOWE AND IAIN RAEBURN 4.2. Solenoids. Let P = (p1, p2, . . . ) be a sequence of primes. The P -adic solenoid is the inverse limit ΣP := lim←−(T, z 7→ zpi), which is a compact abelian group. We will use results of Charatonik-Covarrubias [9] and Gumerov [17] to study C(ΣP ) ⋊α,L N×. The solenoid ΣP is connected because it is the inverse limit of connected spaces [13, Theorem 6.1.20]. The maps ωa : ΣP → ΣP are studied in [17] (there denoted ha P ). Proposition 3 of [17] says that ωa is a homeomorphism if each prime divisor of a occurs infinitely often in P , and Theorem 1 of [17] says that ker ωa = a if and only if no prime divisor of a occurs infinitely often in P . Let P1 be the set of primes which occur infinitely often in P . Then each a has a unique factorisation a = bc where b is coprime to every p ∈ P1, and c has all its prime factors in P1, and we have ker ωa = ker(ωb ◦ ωc) = ker ωb = b. Hence ΣP satisfies (G2). We claim that the elements of infinite order are dense in ΣP . To see this we consider the canonical continuous homomorphism πn of the inverse limit lim←−(T, z 7→ zpi) onto the nth copy of T. Then the sets {π−1 n (U) : n ≥ 1, U is open in T} form a basis for the topology on ΣP . If g ∈ ΣP has some πn(g) = eiθ with θ rational, then g has infinite order, and since the πn are surjective, there are such points in every π−1 n (U). So the elements of infinite order are dense, as claimed, and int(Tor ΣP ) = ∅. Thus Theorem 3.8 gives: Proposition 4.3. The crossed product C(ΣP ) ⋊α,L N× is purely infinite and simple. 4.3. The p-adic integers. For each prime p, the p-adic integers are the elements of the inverse limit Zp := lim←− Z/pkZ, where each map Z/pk+1Z → Z/pkZ is reduction modulo pk. Each Z/pkZ is a finite ring, and hence the inverse limit is a compact ring; we are interested in the additive group (Zp, +), which is a compact abelian group but is not connected. For a ∈ N×, the map ωa takes x to ax, and is injective. We can write a = pnb, where b is coprime to p, and then Zp : ωa(Zp) = pn. So Zp satisfies (G1 -- 3). The transfer operator La on C(Zp) is given by La(f )(x) =(f (a−1x) 0 if x ∈ aZp = range ωa, otherwise and a 7→ La is an action by endomorphisms of C(Zp) such that αa ◦La = id. So we are in the situation of [26, Theorem 5.7], and [26, Corollary 5.8] implies that C(Zp) ⋊α,L N× is isomorphic to the Stacey crossed product C(Zp) ×St L N× (as in [32, 23, 20], for example). We can say more about this Stacey crossed product using the analysis of [27, 6]. It follows from [25, Theorem 4.3] that the ideal C0(Zp\{0}) in C(Zp) is extendibly invariant in the sense of Adji [1], and hence by [25, Theorem 1.7] gives an ideal C0(Zp\{0})×St L N× in C(Zp) ×St L N× isomorphic to C ∗(N×) = C(T∞). L N× with quotient C ×St To analyse the structure of the ideal, set P := {b ∈ N× : b is coprime to p}, and note that (n, b) 7→ pnb is an isomorphism of N × P onto N×. Similarly, with U(Zp) the group of units in the ring Zp, the map (n, x) 7→ pnx is a homeomorphism of N × U(Zp) onto Zp\{0}. This homeomorphism gives a tensor-product decomposition C0(Zp\{0}) = c0(N) ⊗ C(U(Zp)), and in this decomposition the endomorphism αpna decomposes as τn ⊗ σa, where τn(f )(m) =(f (m − n) 0 if m ≥ n if m < n, EXEL-LARSEN CROSSED PRODUCTS 9 and σa(g)(x) = g(ax) (see [27, page 175]). Now Theorem 2.6 of [25] implies that (cid:0)c0(N) ⊗ C(U(Zp))(cid:1) ×St α N× ∼=(cid:0)c0(N) ×St τ N(cid:1) ⊗(cid:0)C(U(Zp)) ×St σ P(cid:1). As in the proof of [27, Lemma 2.5], c0(N) ×St τ N is isomorphic to K(ℓ2(N)). The endo- morphisms σa are in fact automorphisms, and hence σ extends to an action σ of the subgroup P P −1 of Q∗. Theorem 4.1 of [6] implies that C(U(Zp)) ×St σ P = C(U(Zp)) ×σ (P P −1) is a simple AT-algebra with real rank zero and a unique tracial state. 4.4. The integral ad`eles. The integral ad`eles are the elements of the ring Z := Qp prime Zp, and the additive group (Z, +) satisfies (G1 -- 3). As in the preceding ex- ample, [26, Corollary 5.8] gives isomorphisms C(Z) ⋊α,L N× ∼= C(Z) ⋊L,α N× ∼= C(Z) ⋊St L N×, (4.1) and since Z is the Pontryagin dual of Q/Z, [23, Corollary 2.10] implies that the Stacey crossed product is isomorphic to the Bost-Connes algebra (as observed in [26, Exam- ple 5.9]). The statements of [26, Lemma 4.5 and Theorem 5.7] show that the isomor- phisms in (4.1) preserve the copies of C(Z), so Theorem 3.5 gives [23, Theorem 3.7]. In view of the isomorphism (4.1), the results of [24] show that the Exel-Larsen crossed product C(Z) ⋊α,L N× has a complicated ideal structure, and is in particular not simple. Remark 4.4. Of course this proof of [23, Theorem 3.7] goes through Yamashita's unique- ness theorem for row-finite graphs with no sources from [33]. So it is mildly curious that he was not able to do this in [33], because the topological-graph model he was using has sources (see [33, Remark 5.6]). He has since proved a more general uniqueness theorem which does apply to this example (see [34], especially Example 5.17). 5. Exel-Larsen systems from higher-rank graphs Let E be a finite directed graph with one vertex, and partition the edge set E1 = B ⊔ E1 E1 R into blue and red edges. We suppose that there are N1 blue edges and N2 red edges, and we use multiindex notation: for n = (n1, n2) ∈ N2 we write N n := N n1 1 N n2 2 . We write EBB, EBR, ERB and ERR for the sets of blue-blue, blue-red, red-blue and red- red paths of length 2, respectively. We suppose that we have a bijection θ : E2 RB, and we write3 θ1(ef ) ∈ E1 B for the edges in θ(ef ), so that θ(ef ) = θ1(ef )θ2(ef ). R and θ2(ef ) ∈ E1 BR → E2 θ = E1 B, Λ(0,1) A theorem of Kumjian and Pask [22, §6] says that there is a unique 2-graph Λθ such that Λ(1,0) R, and the paths of degree (1, 1) have factorisations ef and θ(ef ). The C ∗-algebra Oθ := C ∗(Λθ) has been studied by Davidson and Yang [12, 35, 36]. The C ∗-algebra Oθ is generated by two Cuntz families {se : e ∈ E1 B} and {sf : f ∈ E1 R} such that sesf = sθ1(ef )sθ2(ef ), and is universal for such families. As usual, we have = E1 θ Oθ = span{sλs∗ µ : λ, µ ∈ Λθ}. 3This is different from the notation in [35, 36], where the sets E 1 B and E 1 R are listed as e1, · · · , em and f1, · · · , fn, and θ is a map on sets of labels. Yang writes "eifj = fj ′ ei′ where θ(i, j) = (i′, j ′)". 10 NATHAN BROWNLOWE AND IAIN RAEBURN There is a gauge action γ of T2 characterised by γz(sλs∗ Oγ θ is the fixed-point algebra µ) = zd(λ)−d(µ)sλs∗ µ, and the core Oγ θ = span{sλs∗ µ : d(λ) = d(µ)}, which is a UHF algebra of type (N1N2)∞. Yang observes in [35, §3.1] that there is an action α : N2 → End Oθ such that αn(a) = Xλ∈Λn θ sλas∗ λ for a ∈ Oθ. This is the analogue for the rank-2 graph algebra Oθ of the canonical unital endo- In [35, §3], Yang investigates other unital endomorphisms of the Oθ, and finds striking parallels with the known properties of endomorphisms of the Cuntz algebras. morphism α(a) = Pn i of the Cuntz algebra On = C ∗(si). 1=1 sias∗ An Huef and Raeburn have recently considered an Exel system (Oγ n, α, L) involving the restriction of the canonical endomorphism α to the core, and showed that Oγ n ⋊α,L N is the Cuntz algebra On2. (To recover this from [20, Theorem 6.6], take N = n in [20, §6], so that the projection p appearing there is the identity of Mn(C).) Here we show that the analogous Exel-Larsen system for Yang's endomorphic action has similar properties: the crossed product is another Oη associated to a larger graph Λη. We first need an action by transfer operators. Proposition 5.1. For each n ∈ N2, the function Ln : Oθ → Oθ defined by Ln(a) = 1 N n Xλ∈Λn θ s∗ λasλ is a transfer operator for αn such that Ln(1) = 1, and we then have Lm ◦ Ln = Lm+n. Proof. The map L is clearly positive and linear, and Ln(1) = 1 because each sλ is an isometry and there are N n summands. For a, b ∈ Oθ we have Ln(αn(a)b) = 1 N n Xλ∈Λn θ s∗ λ(cid:16) Xµ∈Λn θ sµas∗ µ(cid:17)bsλ. Since λ and µ have the same degree, s∗ has just one nonzero term, λsµ vanishes unless µ = λ. Thus the inside sum Ln(αn(a)b) = 1 N n Xλ∈Λn θ as∗ λbsλ, and factoring out a gives Ln(αn(a)b) = aLn(b). So Ln is a transfer operator for αn. To see that Lm ◦ Ln = Lm+n, we take a ∈ Oθ and calculate: 1 1 Lm ◦ Ln(a) = N m Xµ∈Λm θ s∗ µ(cid:16) 1 N n Xλ∈Λn θ s∗ λasλ(cid:17)sµ = N m+n Xµ∈Λm θ λ∈Λn θ s∗ µλasµλ. (5.1) The factorisation property implies that {µλ : µ ∈ Λm (5.1) implies that Lm ◦ Ln(a) = Lm+n(a). θ , λ ∈ Λn θ } is just Λm+n θ , and hence (cid:3) EXEL-LARSEN CROSSED PRODUCTS 11 Both αn and Ln map the core Oγ θ , N2, α, L) is an Exel- Larsen system. We now describe the corresponding crossed product Oγ θ ⋊α,L N2. Recall that we are writing kM = {kM,n : n ∈ N2} for the canonical Cuntz-Pimsner covariant representation of M in O(M) = Oγ θ into itself, and hence (Oγ θ ⋊α,L N2. Theorem 5.2. Consider the coloured directed graph F = (F 0, F 1 = F 1 F 0 = E0, F 1 RR, and define η : F 2 BB and F 1 BR → F 2 B = E2 R = E2 RB by B ⊔ F 1 R) with η((ef )(gh)) = (θ1(eg)θ1(f h))(θ2(eg)θ2(f h)). Then η is a bijection. If {tef : ef ∈ F 1 which generate Oη, then there is an isomorphism π of Oη onto Oγ B} and {tgh : gh ∈ F 1 π(tef ) = kM,(1,0)(N 1/2 1 q(1,0)(ses∗ B and gh ∈ F 1 R. for ef ∈ F 1 f )) and π(tgh) = kM,(0,1)(N 1/2 (5.2) R} denote the Cuntz families θ ⋊α,L N2 such that 2 q(0,1)(sgs∗ h)), We first show that η is a bijection. If η((ef )(gh)) = η((e′f ′)(g′h′)), then repeated applications of the factorisation property in Λθ show that e = e′, f = f ′, g = g′ and h = h′, whence ef = e′f ′ and gh = g′h′. So η is an injection of F 2 RR into RB = E2 F 2 In view of [15, Lemma 2.6] and [20, Lemma 6.3], it is not surprising that the right BB. Since both sets have N 2 2 elements, η is a bijection. BR = E2 BBE2 RRE2 1 N 2 Hilbert Oγ θ -modules Mn have orthonormal bases. Proposition 5.3. For every n ∈ N2, (cid:8)mn thonormal basis for the right Hilbert Oγ µνmn mm µν := N n/2qn(sµs∗ ν) : µ, ν ∈ Λn θ -module Mn. The multiplication in M satisfies αβ = mm+n (5.3) (µα)(νβ). θ(cid:9) is an or- Proof. Fix µ, ν ∈ Λm satisfies θ and α, β ∈ Λn θ , and recall from (2.1) that the multiplication in M We have qm(sµs∗ ν)qn(sαs∗ β) = qm+n(sµs∗ ναm(sαs∗ β)). sµs∗ ναm(sαs∗ β) = sµs∗ sλsαs∗ βs∗ sµs∗ νsλsαs∗ βs∗ λ; ν(cid:16) Xd(λ)=m λ(cid:17) = Xd(λ)=m since d(ν) = m = d(λ) and the {sλ : d(λ) = m} are isometries, only the term with λ = ν survives, and we have sµs∗ ναm(sαs∗ β) = sµsαs∗ βs∗ This immediately gives the formula (5.3), and since β = qm(sµs∗ θ } generates Mm as a right Hilbert Oγ θ , and compute µν} is orthonormal, we take µ, ν, α, β ∈ Λn qm(sµs∗ ν) · sαs∗ µν : µ, ν ∈ Λm (5.4) also implies that {mm To check that {mn ν = sµαs∗ ναm(sαs∗ β)), νβ. (5.4) θ -module. hmn µν, mn αβi = N nLn((sµs∗ ν)∗(sαs∗ s∗ λsνs∗ µsαs∗ βsλ. β)) = Xd(λ)=n Since all the paths have the same degree n, all the terms vanish unless µ = α. When µ = α, only the term with λ = ν survives, and it is zero unless β = ν. So the inner product is zero unless (µ, ν) = (α, β), and then is 1 because all the sµ are isometries. (cid:3) 12 NATHAN BROWNLOWE AND IAIN RAEBURN As in [15] and [20], the orthonormal bases for Mn give rise to Cuntz families in O(M): Proposition 5.4. For ef ∈ F 1 ef B = E2 Tef := kM,(1,0)(m(1,0) BB and gh ∈ F 1 RR, we set ) and Tgh := kM,(0,1)(m(0,1) R = E2 gh ). Then {Tef : ef ∈ F 1 B} and {Tgh : gh ∈ F 1 R} are Cuntz families, and they satisfy Tef Tgh = Tη1((ef )(gh))Tη2((ef )(gh)). (5.5) Proof. Since ef Tef = kM,(1,0)(m(1,0) T ∗ ef )∗kM,(1,0)(m(1,0) ef ) = kM,0(hm(1,0) ef , m(1,0) ef i) = kM,0(1), and since kM,0 is unital [8, Corollary 3.5], each Tef is an isometry. The reconstruction formula for the orthonormal basis {m(1,0) ef } implies that φ(1,0)(1) = 1 = Xef ∈F 1 B Θm(1,0) ef ,m(1,0) ef , and hence the Cuntz-Pimsner covariance of (kM,(1,0), kM,0) implies that kM,0(1) = kM,0(φ(1,0)(1)) = Xef ∈F 1 B kM,(1,0)(m(1,0) ef )kM,(1,0)(m(1,0) ef Tef T ∗ ef . )∗ = Xef ∈F 1 B Thus {Tef } is a Cuntz family. The same arguments work for {Tgh}. To establish (5.5), we calculate using (5.3): hs∗ ef (cid:1)kM,(0,1)(cid:0)m(0,1) Tef Tgh = kM,(1,0)(cid:0)m(1,0) gh (cid:1) = kM,(1,1)(cid:0)m(1,1) (eg)(f h)(cid:1) = N (1,1)kM,(1,1)(cid:0)q(1,1)(sesgs∗ f )(cid:1) = N (1,1)kM,(1,1)(cid:0)q(1,1)(sθ1(eg)sθ2(eg)s∗ = N (1,1)kM,(1,1)(cid:0)q(1,1)(sθ1(eg)θ2(eg)s∗ = kM,(1,1)(cid:0)m(1,1) θ1(eg)θ2(eg),θ1(f h)θ2(f h)(cid:1) = kM,(0,1)(cid:0)m(0,1) θ1(eg)θ1(f h)(cid:1)kM,(1,0)(cid:0)m(1,0) = Tη1((ef )(gh))Tη2((ef )(gh)). θ2(f h)s∗ θ1(f h))(cid:1) θ1(f h)θ2(f h))(cid:1) θ2(eg)θ2(f h)(cid:1) (cid:3) Proposition 5.4 and the universal property of Oη imply that there is a homomorphism π := πT : Oη → Oγ θ ⋊α,L N2 such that π(tef ) = Tef = kM,(1,0)(N 1/2 π(tgh) = Tgh = kM,(0,1)(N 1/2 1 q(1,0)(ses∗ 2 q(0,1)(sgs∗ f )) h)) for ef ∈ F 1 B, and for gh ∈ F 1 R. We complete the proof of Theorem 5.2 by proving the following Proposition. Proposition 5.5. The homomorphism π is an isomorphism of Oη onto Oγ θ ⋊α,L N2. EXEL-LARSEN CROSSED PRODUCTS 13 Proof. The homomorphism π is equivariant for the gauge action of T2 on Oη = C ∗(Λη) and the gauge action of T2 on Oγ θ ⋊α,L N2 = O(M). Because each Tef is an isometry, the homomorphism π is unital, and hence the image of the only vertex projection pv = 1C ∗(Λη) is nonzero. Thus the gauge-invariant uniqueness theorem [22, Theorem 3.4] implies that π is injective. The formula (5.3) implies that the image of π contains the image kM,n(mn µν for Mn. Since these elements generate kM,n(Mn) as a kM,0(Oγ µν) of every basis element mn θ )- module, we have kM,n(Mn) ⊂ π(Oη) for each n ∈ N2. So to see that π is surjective it suffices to show that the range of π contains the image kM,0(Oγ θ ) of the coefficient algebra; since the range is a C ∗-subalgebra, it suffices to show that the range contains each kM,0(sµs∗ ν) with d(µ) = d(ν). So fix such an element with d(µ) = d(ν) = n, say. We claim that φn(sµs∗ ν) ∈ L(Mn) satisfies φn(sµs∗ ν) = Xd(λ)=n Θmn µλ,mn νλ . (5.6) Both sides of (5.6) are adjointable operators on the right Hilbert Oγ θ -module Mn, and θ -module homomorphisms, so it suffices to check that they agree in particular are right Oγ on a typical basis element mn αβ. We have αβ) = N n/2sµs∗ νqn(sαs∗ φn(sµs∗ ν)(mn =(0 N n/2qn(sµs∗ β) = mn µβ νsαs∗ β) β) = N n/2qn(sµs∗ if ν 6= α if ν = α. On the other hand, we have Xd(λ)=n Θmn µλ,mn νλ (mn mn µλ · hmn νλ, mn αβi. αβ) = Xd(λ)=n The inner product vanishes unless (ν, λ) = (α, β), and then is the identity in Oγ θ . So every summand vanishes unless ν = α, and then only the λ = β summand survives, so Xd(λ)=n Θmn µλ,mn νλ(cid:0)mn αβ(cid:1) = mn µβ · 1 = mn µβ. Thus we have (5.6), as claimed. In view of (5.6), the Cuntz-Pimsner covariance of kM implies that kM,0(sµs∗ ν) = (kM,n, kM,0)(1)(φM n (sµs∗ Since we have already seen that each kM,n(cid:0)mn that kM,0(sµs∗ ν) belongs to the range of π, and π is surjective. kM,n(cid:0)mn νλ(cid:1)∗ µλ(cid:1)kM,n(cid:0)mn ν)) = Xd(λ)=n µλ(cid:1) belongs to the range of π, we deduce (cid:3) . Davidson and Yang have found a necessary and sufficient condition for aperiodicity of O(Λθ) [12, Theorem 3.1] (and we give a new proof of their result in an appendix). It is interesting to note that their condition is not easy to verify, and they discuss at some length an algorithm for determining whether a given Λθ is periodic. However, their result gives the following criteria for simplicity of our crossed products. 14 NATHAN BROWNLOWE AND IAIN RAEBURN Corollary 5.6. Let Λθ be a 2-graph with one vertex such that the corresponding directed graph has N1 ≥ 2 blue edges and N2 ≥ 2 red edges, and let η be the bijection described in (5.2). The Exel-Larsen crossed product Oγ θ ⋊α,L N2 is simple if and only if, for every pair of integers a and b with N a , there exists µ ∈ Λ(a,0) θ ⋊α,L N2 is simple, then it is purely infinite. 2 and every bijection ρ : Λ(a,0) with µλ 6= ρ(µ)ρ−1(λ). If Oγ and λ ∈ Λ(0,b) η → Λ(0,b) 1 = N b η η η 1 = N b 2 if and only if (N 2 Proof. First note that N a 2 )b, so we can apply [12, Theorem 3.4] (or Theorem A.1 below) to the graph Oη, and deduce that Oη is aperiodic if and only if the condition on the bijections Λ(a,0) holds. Since graphs with a single vertex are trivially cofinal, [29, Theorem 3.1 and Lemma 3.2] imply that Oη is simple if and only if the condition in the theorem holds. Since the hypothesis of [30, Proposition 8.8] trivially holds in our graphs, we know from that result that Oη is purely infinite. Thus the result follows from Theorem 5.2. (cid:3) η → Λ(0,b) 1 )a = (N 2 η Remark 5.7. If ln N1 and ln N2 are not rationally related, then there are no such bijec- tions ρ, our condition is trivially satisfied, and Corollary 5.6 says that Oγ θ ⋊α,L N2 is purely infinite and simple. Appendix A. On Davidson and Yang's characterisation of periodicity In [12, Theorem 3.1], Davidson and Yang find a necessary and sufficient condition for a 2-graph with one vertex to be periodic. In their proof, they view the graph as a semigroup F+ θ which they and Power had previously constructed in [11]. Here we give a short proof of their theorem that argues directly in terms of the graph. We use the finite-path formulation of periodicity from [29]. θ , and use some concrete isometric representations of F+ Theorem A.1 (Davidson and Yang). Let Λ be a finite 2-graph with a single vertex and at least two edges of each colour. Then Λ is periodic if and only if there are positive integers a, b and a bijection γ of Λ(a,0) onto Λ(0,b) such that, for each (µ, ν) ∈ Λ(a,0)×Λ(0,b), the path µν in Λ(a,b) has (0, b) + (a, 0) factorisation γ(µ)γ−1(ν). Proof. Suppose that Λ is periodic. Then there exist m, n ∈ N2 such that m 6= n and µ(m, m+d(µ)−(m∨n)) = µ(n, n+d(µ)−(m∨n)) for all µ with d(µ) ≥ m ∨ n. (A.1) We claim that we cannot have m ≤ n. To see this, suppose that m ≤ n and λ is any path with d(λ) = n. Then our hypotheses imply that there are at least 2 paths of degree n − m, and in particular there is one ν which is not equal to λ(m, n). Then µ := λν does not satisfy (A.1). So we cannot have one of m or n larger than the other, and (possibly after swapping m and n) there exist a and b positive such that m = (m ∧ n) + (a, 0) and n = (m ∧ n) + (0, b). Now taking a path ν with d(ν) = m ∧ n and applying (A.1) to µ = νλ shows that λ((a, 0), d(λ) − (0, b)) = λ((0, b), d(λ) − (a, 0)) for all λ with d(λ) ≥ (a, b). (A.2) We next claim that for each µ ∈ Λ(a,0), there is a unique γ(µ) ∈ Λ(0,b) such that (αµ)((a, 0), (a, b)) = γ(µ) = (µβ)((0, 0), (0, b)) for all α, β ∈ Λ(0,b). (A.3) EXEL-LARSEN CROSSED PRODUCTS 15 The uniqueness in the factorisation property implies that there is at most one such γ(µ). To see there is one, take α, β ∈ Λ(0,b), and consider λ := αµβ ∈ Λ(a,2b). Then (αµ)((a, 0), (a, b)) = (αµβ)((a, 0), (a, b)) = (αµβ)((0, b), (0, 2b)) by (A.2) = (µβ)((0, 0), (0, b)). Since the first term is independent of β and the last term is independent of α, the claim follows. Swapping the roles of blue and red paths in the argument of the preceding paragraph shows that for each ν ∈ Λ(0,b), there is a unique δ(ν) ∈ Λ(a,0) such that (τ ν)((0, b), (a, b)) = δ(ν) = (νσ)((0, 0), (a, 0)) (A.4) We claim that δ is an inverse for γ. Take µ ∈ Λ(a,0). To compute γ(µ), we pick any α ∈ Λ(0,b), and then (A.3) gives γ(µ) = (αµ)((a, 0), (a, b)). To compute δ(γ(µ)), we take τ = (αµ)((0, 0), (a, 0)). Then τ γ(µ) = αµ, and (A.4) gives for all τ, σ ∈ Λ(a,0). δ(γ(µ)) = (τ γ(µ))((0, b), (a, b)) = (αµ)((0, b), (a, b)) = µ. A similar argument shows that γ ◦ δ is the identity. We have now proved that γ is a bijection with γ−1 = δ. We next need to show that µν = γ(µ)δ(ν). Take α ∈ Λ(0,b) and σ ∈ Λ(a,0), and consider the path λ := αµνσ: the diagram in Figure 1 should help see what is happening: the • o σo • o ν µ • • α • • • • • Figure 1. The path λ of degree (2a, 2b). unbroken arrows represent blue paths of length b and the dashed arrows red paths of length a. Periodicity in the form of (A.2) tells us that the top left-hand corner in Figure 1, which is µν = λ((0, b), (a, 2b)), is the same as the bottom right-hand corner λ((a, 0), (2a, b)) = (αµνσ)((a, 0), (a, b))(αµνσ)((a, b)(2a, b)) = (αµ)((a, 0), (a, b))(νσ)((0, 0), (a, 0)), which by (A.3) and (A.4) is precisely γ(µ)δ(ν). Thus γ has all the required properties. Now suppose that we have a, b and γ as in the theorem. Because γ is a bijection, we also have αβ = γ−1(α)γ(β) (A.5) We will prove that Λ is periodic by showing that every λ ∈ Λ with d(λ) ≥ (a, b) satisfies (A.2). Take such a λ. Because it suffices to prove (A.2) for a longer path λµ, we may suppose that d(λ) = (Ma, Nb) for some positive integers M, N, and then it suffices to prove that for α ∈ Λ(0,b), β ∈ Λ(a,0). λ(cid:0)(ma, (n + 1)b), ((m + 1)a, (n + 2)b)(cid:1) = λ(cid:0)((m + 1)a, nb), ((m + 2)a, (n + 1)b)(cid:1) (A.6) o o ✤ ✤ ✤ ✤ ✤ ✤ ✤ ✤ ✤ ✤ ✤ ✤ ✤ ✤ ✤ ✤ ✤ ✤ o o o o o o o o o o o 16 NATHAN BROWNLOWE AND IAIN RAEBURN for every pair m, n satisfying 0 ≤ m ≤ M − 2 and 0 ≤ n ≤ N − 2. Factor λ(cid:0)(ma, nb), ((m + 2)a, (n + 2)b)(cid:1) = αµνσ with µ, σ ∈ Λ(a,0) and α, ν ∈ Λ(0,b), and then we are back in the situation of Figure 1. The top left corner of Figure 1 has factorisations µν = γ(µ)γ−1(ν). The relation (A.5) implies that the top right and bottom left corners have factorisations νσ = γ−1(ν)γ(σ) and αµ = γ−1(α)γ(µ). Together, these imply that the bottom right-hand corner is γ(µ)γ−1(ν), which is the same as the top left corner. This says that λ satisfies (A.6), as required. (cid:3) References [1] S. Adji, Invariant ideals of crossed products by semigroups of endomorphisms, Functional Analysis and Global Analysis (T. Sunada and P. W. Sy, Eds.), Springer-Verlag, Singapore, 1997, pages 1 -- 8. [2] L.W. Baggett, N.S. Larsen, J.A. Packer, I. Raeburn and A. Ramsay, Direct limits, multiresolution analyses, and wavelets, J. Funct. Anal. 258 (2010), 2714 -- 2738. [3] J.-B. Bost and A. Connes, Hecke algebras, Type III factors and phase transitions with spontaneous symmetry breaking in number theory, Selecta Math. (N.S.) 1 (1995), 411 -- 457. [4] N. Brownlowe, Realising the C ∗-algebra of a higher-rank graph as an Exel crossed product, J. Operator Theory, to appear; arXiv:0912.4635. [5] N. Brownlowe, A. an Huef, M. Laca and I. Raeburn, Boundary quotients of the Toeplitz algebra of the affine semigroup over the natural numbers, Ergodic Theory Dynam. Systems 32 (2012), 35 -- 62. [6] N. Brownlowe, N.S. Larsen, I.F. Putnam and I. Raeburn, Subquotients of Hecke C ∗-algebras, Ergodic Theory Dynam. Systems 25 (2005), 1503 -- 1520. [7] N. Brownlowe and I. Raeburn, Exel's crossed product and relative Cuntz-Pimsner algebras, Math. Proc. Cambridge Philos. Soc. 141 (2006), 497 -- 508. [8] N. Brownlowe, I. Raeburn and S.T. Vittadello, Exel's crossed product for non-unital C ∗-algebras, Math. Proc. Cambridge Philos. Soc. 149 (2010), 423 -- 444. [9] J.J. Charatonik and P.P. Covarrubias, On covering mappings on solenoids, Proc. Amer. Math. Soc. 130 (2002), 2145 -- 2154. [10] J. Cuntz, C ∗-algebras associated with the ax + b-semigroup over N, in: K-Theory and Noncommu- tative Geometry (Valladolid, 2006), European Math. Soc., 2008, pages 201 -- 215. [11] K.R. Davidson, S.C. Power and D. Yang, Dilation theory for rank 2 graph algebras, J. Operator Theory 63 (2010), 245 -- 270. [12] K.R. Davidson and D. Yang, Periodicity in rank 2 graph algebras, Canad. J. Math. 61 (2009), 1239 -- 1261. [13] R. Engelking, General Topology, Sigma Series in Pure Math., vol.6, second ed., Heldermann Verlag, Berlin, 1989. [14] R. Exel, A new look at the crossed-product of a C ∗-algebra by an endomorphism, Ergodic Theory Dynam. Systems 23 (2003), 1 -- 18. [15] R. Exel, A. an Huef and I. Raeburn, Purely infinite simple C ∗-algebras associated to integer dilation matrices, Indiana Univ. Math. J., to appear; arXiv:1003.2097. [16] N.J. Fowler, Discrete product systems of Hilbert bimodules, Pacific J. Math. 204 (2002), 335 -- 375. [17] R.N. Gumerov, On finite-sheeted covering mappings onto solenoids, Proc. Amer. Math. Soc. 133 (2005), 2771 -- 2778. [18] E. Hewitt and K.A. Ross, Abstract Harmonic Analysis, Vol. I. Springer-Verlag, Berlin, 1963. [19] J.H. Hong, N.S. Larsen and W. Szyma´nski, The Cuntz algebra QN and C ∗-algebras of product systems, Proceedings of the EU-Network Noncommutative Geometry Fourth Annual Meeting, Bucharest, 2011, to appear; arXiv:1108.0790. [20] A. an Huef and I. Raeburn, Stacey crossed products associated to Exel systems, Integral Equations Operator Theory 72 (2012), 537 -- 561. [21] D.W. Kribs and S.C. Power, The analytic algebras of higher rank graphs, Math. Proc. Royal Irish Acad. 106A (2006), 199 -- 218. EXEL-LARSEN CROSSED PRODUCTS 17 [22] A. Kumjian and D. Pask, Higher rank graph C ∗-algebras, New York J. Math. 6 (2000), 1 -- 20. [23] M. Laca and I. Raeburn, A semigroup crossed product arising in number theory, J. London Math. Soc. 59 (1999), 330 -- 344. [24] M. Laca and I. Raeburn, The ideal structure of the Hecke C ∗-algebra of Bost and Connes, Math. Ann. 318 (2000), 433 -- 451. [25] N.S. Larsen, Nonunital semigroup crossed products, Math. Proc. Royal Irish Acad. 100A (2000), 205 -- 218. [26] N.S. Larsen, Crossed products by abelian semigroups via transfer operators, Ergodic Theory Dynam. Systems 30 (2010), 1147 -- 1164. [27] N.S. Larsen, I.F. Putnam and I. Raeburn, The two-prime analogue of the Hecke C ∗-algebra of Bost and Connes, Indiana Univ. Math. J. 51 (2002), 171 -- 186. [28] N.S. Larsen and I. Raeburn, Projective multi-resolution analyses arising from direct limits of Hilbert modules, Math. Scand. 100 (2007), 317 -- 360. [29] D.I. Robertson and A. Sims, Simplicity of C ∗-algebras associated to higher-rank graphs, Bull. London Math. Soc. 39 (2007), 337 -- 344. [30] A. Sims, Gauge-invariant ideals in the C ∗-algebras of finitely aligned higher-rank graphs, Canad. J. Math. 58 (2006), 1268 -- 1290. [31] A. Sims and T. Yeend, C ∗-algebras associated to product systems of Hilbert bimodules, J. Operator Theory 64 (2010), 349 -- 376. [32] P.J. Stacey, Crossed products of C ∗-algebras by ∗-endomorphisms, J. Austral. Math. Soc. (Ser. A) 54 (1993), 204 -- 212. [33] S. Yamashita, Cuntz's ax + b-semigroup C ∗-algebra over N and product system C ∗-algebras, J. Ramanujan Math. Soc. 24 (2009), 299 -- 322. [34] S. Yamashita, Cuntz-Krieger type uniqueness theorem for topological higher-rank graph C ∗-algebras, arXiv:0911.2978. [35] D. Yang, Endomorphisms and modular theory of 2-graph C ∗-algebras, Indiana Univ. Math. J. 59 (2010), 495 -- 520. [36] D. Yang, Type III von Neumann algebras associated with Oθ, Bull. London Math. Soc., available online at the journal; arXiv:1104.5697. School of Mathematics and Applied Statistics, University of Wollongong, NSW 2522, Australia E-mail address: [email protected] Department of Mathematics and Statistics, University of Otago, PO Box 56, Dunedin 9054, New Zealand E-mail address: [email protected]
1108.4200	3	1108	2013-07-18T16:52:29	On the structural theory of II_1 factors of negatively curved groups, II: Actions by product groups	[ "math.OA", "math.GR" ]	This paper includes a series of structural results for von Neumann algebras arising from measure preserving actions by product groups on probability spaces. Expanding upon the methods used earlier by the first two authors \cite{CS}, we obtain new examples of strongly solid factors as well as von Neumann algebras with unique or no Cartan subalgebra. For instance we show that every II$_1$ factor associated with a weakly amenable group in the class $\mathcal S$ of Ozawa is strongly solid, \cite{OzSolid}. There is also the following product version of this result: any maximal abelian $\star$-subalgebra of any II$_1$ factor associated with a finite product of weakly amenable groups in the class $\mathcal S$ of Ozawa has an amenable normalizing algebra. Finally, pairing some of these results with cocycle superrigidity results from \cite{IoaCSR}, it follows that compact actions by finite products of lattices in $Sp(n, 1)$, $n \geq2$, are virtually $W^*$-superrigid.	math.OA	math	ON THE STRUCTURAL THEORY OF II1 FACTORS OF NEGATIVELY CURVED GROUPS, II. ACTIONS BY PRODUCT GROUPS IONUT CHIFAN, THOMAS SINCLAIR, AND BOGDAN UDREA Abstract. This paper contains a series of structural results for von Neumann algebras arising from measure preserving actions by product groups on prob- ability spaces. Expanding upon the methods used earlier by the first two authors, we obtain new examples of strongly solid factors as well as von Neu- mann algebras with unique or no Cartan subalgebra. For instance, we show that every II1 factor associated with a weakly amenable group in the class S of Ozawa is strongly solid. There is also the following product version of this result: any maximal abelian ⋆-subalgebra of any II1 factor associated with a finite product of weakly amenable groups in the class S of Ozawa has an amenable normalizing algebra. Finally, pairing some of these results with a cocycle superrigidity result of Ioana, it follows that compact actions by finite products of lattices in Sp(n, 1), n ≥ 2, are virtually W ∗-superrigid. Contents Introduction Statement of results 1. Popa's Intertwining Techniques 2. Relative Arrays and Relative Quasi-cocycles 2.1. Relative arrays 2.2. Relative quasi-cocycles 3. Weak Amenability for Groups and von Neumann Algebras 4. The Gaussian Construction, Bimodules and Weak Containment 5. A Path of Automorphisms of the Extended Roe Algebra Associated with the Products of Gaussian Actions 6. Proofs of the Main Results Acknowledgements References 1 2 4 5 5 7 12 13 15 16 28 28 Introduction An important motivation in the study of II1 factors-in fact, one of von Neu- mann's original motivations in inventing the subject-is that they provide an ana- lytical and algebraic framework for the representation theory of groups and ergodic theory. The usefulness of this observation lies in the fact that classification questions Date: February 8, 2018. 2010 Mathematics Subject Classification. 46L07; 43A22. 1 2 I. CHIFAN, T. SINCLAIR, AND B. UDREA in ergodic theory or representation theory can often be reformulated as questions on the algebraic structure of certain von Neumann algebras, and that such ques- tions may be approached with strategies and techniques beyond those which are available in the standard ergodic or representation-theoretic toolkits. A notable example of the translation of problems from ergodic theory to the theory of von Neumann algebras is the fundamental result of Singer [64] which states that the or- bit equivalence class of a free, ergodic, p.m.p. action of a countable discrete group is in one-to-one correspondence with the group of automorphisms of a canonical associated II1 factor which preserves a canonical subalgebra. Thus, the problem of characterizing all group actions orbit equivalent to a given one is reduced to the calculation of the symmetry group of some algebraic object. With such applications in mind, Popa developed in the first half of the last decade a powerful theory for the classification of algebraic structure in II1 factors which he termed deformation/rigidity [50, 51, 52]. Popa's techniques rapidly led to the settling of several long-standing problems in the theory of II1 factors [50] as well as far-reaching classification results in the orbit equivalence theory of ergodic actions, notably, Popa's cocycle superrigidity theorems [53, 55]. Following Popa's seminal work, the classification of II1 factors has witnessed a rebirth. To list some of the major accomplishments which have occurred in the last several years: the cocycle superrigidity theorems of Popa [53, 55] and Ioana [25]; work on the classification of Cartan subalgebras by Ozawa and Popa [45, 46]; the discovery W∗-superrigid groups and actions with substantial contributions by Ioana, Peterson, Popa, and Vaes, among others, [47, 57, 26, 28, 8, 66]; and the study of various structural properties for von Neumann algebras such as strong solidity initiated by Ozawa and Popa [45, 46] and continued by others [20, 21, 63, 9]. This paper is the continuation of an article [9] by the first two authors. The broad theme of that article was the application of geometric techniques in the context of Popa's deformation/rigidity theory to obtain structural results for II1 factors associated to Gromov hyperbolic groups and their actions on measure spaces. This was accomplished in part through the reinterpretation of Ozawa's C∗-algebraic structural theory of group factors [39, 40] in terms of Peterson's cohomological approach [48] to Popa's deformation/rigidity theory. However, partly for reasons of clarity, there are aspects of Ozawa's theory which were not touched upon in the previous paper-specifically, the use of "small" families of subgroups to unify various structural theorems [5, 61]. The aim of this paper is to incorporate these techniques into the deformation/rigidity approach in [9]. The main applications which will be addressed are to p.m.p. actions of countable discrete groups Γ which fall into two basic cases: (1) Γ is generated by a pair of subgroups (G1, G2) which are rather "free" with respect to each other (precisely, Γ is relatively hyperbolic to {G1, G2}), or (2) Γ is generated by a pair of "negatively curved" groups {G1, G2} with a high degree of commutation. Aside from this we will also be able the sharply generalize most of the results in the previous paper to cover a more general class of groups. Statement of results. The main result of this paper will be the following theorem which improves Theorem B/Theorem 4.1 of [9] in two ways. First, we are able to extend the theorem to the more general class of exact groups which admit proper arrays into weakly ℓ2 representations (i.e., bi-exact groups) rather than just proper quasi-cocycles. Secondly, we are able to deal with groups which are "negatively the generating set S =SΣ∈G Σ (i.e., RQ(Γ,G,Hπ) 6= ∅: see §2). II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 3 curved" with respect to a collection of "small" subgroups, which includes, primarily, the widely studied class of relatively hyperbolic groups [3, 15]. The result and its proof are inspired by Ozawa's general semi-solidity theorem (Theorem 15.1.5 in [5]) and Ozawa and Popa's Theorem B in [46], viewed through the framework developed by the first two authors in [9]. Theorem 0.1 (Theorem 6.1). Let Γ be an exact group, let π : Γ → U(Hπ) be a weakly-ℓ2 representation and assume that one of the following holds: (1) Γ admits a proper array into Hπ (i.e., RA(Γ,{e},Hπ) 6= ∅: see §2); or, (2) there exists G, a family of subgroups of Γ such that Γ admits a quasi- cocycle which is metrically proper relative to the length metric coming from Also let Γ y X be a free, ergodic p.m.p. action, denote by M = L∞(X) ⋊ Γ the corresponding crossed-product von Neumann algebra, and let P ⊆ M be any weakly compact embedding with P diffuse. Then the following holds: • if Γ satisfies condition (1) above then either the normalizing algebra NM (P )′′ • if Γ satisfies condition (2) above then either the normalizing algebra NM (P )′′ is amenable or P (cid:22)M L∞(X). is amenable or there exists a group Σ ∈ G such that P (cid:22)M L∞(X) ⋊ Σ. As a consequence any free ergodic weakly compact action[45] of any weakly amenable group Γ in the class S of Ozawa [40] gives rise to a von Neumann algebra with unique Cartan subalgebra. Moreover for all these groups as well as all Γ that are hyperbolic relative to a collection of subgroups which are, in some sense, peripheral (cf. §1.2 and §4 in [15]), then LΓ is strongly solid, as the following corollary demonstrates. For instance this will be the case when Γ is any group in the measure equivalence class of an arbitrary limit group in the sense of Sela. These groups should be con- sidered as generalizations of non-uniform lattices in rank one Lie groups, which may admit finitely many cusp subgroups. Corollary 0.2 (Corollary 6.8). Let Γ be a weakly amenable group and let π : Γ → U(Hπ) be an weakly-ℓ2 representation such that one of the following holds: either RA(Γ,{e},Hπ) 6= ∅, or there exists G, a family of amenable, malnormal subgroups of Γ such that RQ(Γ,G,Hπ) 6= ∅. If Λ is any M E-subgroup of Γ then LΛ is strongly solid i.e., given any diffuse amenable subalgebra A ⊆ LΛ its normalizing algebra NLΛ(A)′′ is still amenable. In particular, every amenable subgroup of Λ has amenable normalizer. Our techniques also allow us to obtain structural results for normalizers in di- rect products of negatively curved groups. Such groups are interesting in that they provide highly tractable examples of groups which exhibit higher-rank (rigid) phe- nomena (cf. [8, 55, 35, 36]). On the other hand, the next result will show that the structure of their group factors may be reduced to the study of their rank one components (this "rank one" decomposition is algebraically unique by [44]; see also Theorem C in [9]). The result is optimal, though more intricate to state than the previous, since one needs to account for the presence of commutation between the factors. Theorem 0.3 (Theorem 6.5). For every i = 1, 2 let Γi be an exact group such that RA(Γi,{e}, ℓ2(Γi)) 6= 0. Let (Γ1 × Γ2) y X be a free, ergodic, p.m.p. action and denote by M = L∞(X)⋊ (Γ1× Γ2) the corresponding crossed-product von Neumann 4 I. CHIFAN, T. SINCLAIR, AND B. UDREA algebra. If P ⊆ M is any weakly compact embedding with P diffuse, then one can find projections p0, p1, p2, p3 ∈ Z(NM (P )′ ∩ M ) with p0 + p1 + p2 + p3 = 1 such that: (1) NM (P )′′p0 is amenable; (2) P p1 (cid:22)M L∞(X) ⋊ Γ1; (3) P p2 (cid:22)M L∞(X) ⋊ Γ2; (4) P p3 (cid:22)M L∞(X). The next corollary, for the special case of tensor products of free group factors, is an unpublished result of Ioana and the first author [7]. It would be interesting to know whether the result also holds true for generic higher rank lattices, e.g. SL(3, Z). Corollary 0.4. Let Γ1, Γ2 be i.c.c. hyperbolic groups and denote by M = LΓ1 ¯⊗LΓ2. If A ⊆ M is an amenable subalgebra such that A′ ∩ M is amenable (e.g. when A is either a m.a.s.a. or an irreducible, amenable subfactor of M ) then its normalizing algebra NM (A)′′ is amenable. The following corollary is complementary to Corollary 6.2 in [8], which holds for product actions of rigid groups which are sufficiently mixing. Interestingly, for actions between these two extremes, the result is known to fail (Example 2.22 in [36]). Corollary 0.5. If Γ1, Γ2 are hyperbolic groups with property (T) (e.g. Γi lattices in Sp(n, 1) n ≥ 2), then any free, ergodic, profinite (or more generally compact) action (Γ1 × Γ2) y X is virtually W ∗-superrigid. 1. Popa's Intertwining Techniques We will briefly review the concept of intertwining two subalgebras inside a von Neumann algebra, along with the main technical tools developed by Popa in [50, 51]. Given N a finite von Neumann algebra, let P ⊂ f N f , Q ⊂ N be diffuse subalgebras for some projection f ∈ N . We say that a corner of P can be intertwined into Q inside N if there exist two non-zero projections p ∈ P , q ∈ Q, a non-zero partial isometry v ∈ pN q, and a ⋆-homomorphism ψ : pP p → qQq such that vψ(x) = xv for all x ∈ pP p. Throughout this paper we denote by P (cid:22)N Q whenever this property holds, and by P (cid:14)N Q its negation. The partial isometry v is called an intertwiner between P and Q. Popa established an efficient criterion for the existence of such intertwiners (The- orems 2.1-2.3 in [51]). Particularly useful in concrete applications is the following analytic description of absence of intertwiners. Theorem 1.1 (Corollary 2.3 in [51]). Let N be a von Neumann algebra and let P ⊂ f N f , Q ⊂ N be diffuse subalgebras for some projection f ∈ N . Then the following are equivalent: (1) P (cid:14)M Q; (2) For every finite set F ⊂ N f and every ǫ > 0 there exists a unitary v ∈ U(P ) such that Xx,y∈F kEQ(xvy∗)k2 2 ≤ ǫ. II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 5 Notice that in the intertwining concept presented above we a priori have no con- trol over the image ψ(pP p) inside qQq. When trying to get unitary conjugacy this often becomes a significant issue and additional analysis regarding the position of ψ(pP p) inside qQq is required. Sometimes the ⋆-homomorphism ψ can be suitably modified to automatically preserve certain properties from the inclusion P ⊂ N to the inclusion ψ(pP p) ⊆ qQq. For instance, Ioana showed in Lemma 1.5 of [27] that if P ⊂ N is a m.a.s.a. then ψ can be chosen so that ψ(pP p) ⊆ qQq is again a m.a.s.a. Applying his argument one can show that ψ can be chosen to also preserve the irreducibility of the inclusion P ⊂ N . The precise technical result which will be of essential use to derive some of our main applications is the following: Proposition 1.2. Let N be a von Neumann algebra together with subalgebras P, Q ⊆ N such that P ′∩N = C1. If we assume that P (cid:22)N Q then one can find pro- jections p ∈ P , q ∈ Q, a ⋆-homomorphism φ : pP p → qQq and a non-zero partial isometry v ∈ qN p such that φ(x)v = vx, for all x ∈ pP p, and φ(pP p)′ ∩ qQq = Cq. The proof of this result follows the same strategy as the proof of Lemma 1.5 of [27], so it will be omitted. We end this section by recalling two important intertwining results from the work of Popa [50, 51]. These results play a very important role in deriving some of our main applications. The first result describes an inclusion of von Neumann algebras where we have complete control over general intertwiners of subalgebras. To properly introduce the statement we need a definition. Given an inclusion of countable groups Σ < Γ we say that Σ is malnormal in Γ if and only if for every γ ∈ Γ \ Σ we have γΣγ−1 ∩ Σ is finite. Proposition 1.3 (Theorem 3.1 in [51]). Let Σ < Γ be a malnormal group, let Γ y A be a trace preserving action and denote by M = A ⋊ Γ the corresponding crossed product von Neumann algebra. Also let p ∈ A ⋊ Σ be a projection and suppose that P ⊆ p(A ⋊ Σ)p is a diffuse subalgebra such that P (cid:14)A⋊Σ A. If there exist elements x, x1, x2, . . . , xn ∈ M such that P x ⊆Pi xiP then x ∈ A ⋊ Σ. The second result which will be needed in the sequel is Popa's unitary conjugacy criterion for Cartan subalgebras. Theorem 1.4 (Appendix 1 in [50]). Let N be a II1 factor and A, B ⊂ N two semiregular m.a.s.a. (i.e., their normalizing algebras NN (A)′′ and NN (B)′′ are sub- factors of N ). If B0 ⊂ B is a von Neumann subalgebra such that B′0 ∩ N = B, and B0 (cid:22)N A, then there exists a unitary u ∈ N such that uAu∗ = B. 2. Relative Arrays and Relative Quasi-cocycles In this section we consider relative versions of the notions of arrays [9] and quasi- cocycles [34, 30, 31] for groups. This will allow us to generalize, from the viewpoint of deformation/rigidity theory, the structural results obtained in [9]. After intro- ducing the definitions, we summarize a few useful properties, relating these with other concepts extant in the literature. In the last part of the section we will present several examples, some of them arising naturally from geometric group theory. 2.1. Relative arrays. Assume that Γ is a countable, discrete group together with : i ∈ I}, a family of subgroups of Γ and π : Γ → U(H), a unitary G = {Σi representation. 6 I. CHIFAN, T. SINCLAIR, AND B. UDREA Definition 2.1. We say that a group Γ admits a proper array relative to G into H if there exists a map r : Γ → H which satisfies the following conditions: (1) πγ(r(γ−1)) = ±r(γ) for all γ ∈ Γ, i.e., (anti-)symmetry; (2) for every γ ∈ Γ we have sup δ∈Γ kr(γδ) − πγ(r(δ))k = C(γ) < ∞; (3) the map γ → kr(γ)k is proper with respect to G, i.e. for every C > 0 there exist finite subsets F ⊂ G and H, K ⊂ Γ such that {γ ∈ Γ : kr(γ)k ≤ C} ⊆ ∪Σ∈F HΣK. Notation 2.2. Given a map φ : Γ → R and ℓ ∈ R, we say that lim φ(γ) = ℓ γ→∞/G if for every ǫ > 0 there exist finite sets H, K ⊂ Γ,F ⊂ G such that φ(γ) − ℓ < ǫ for all γ 6∈ HF K. From now on the set of all such relative arrays will be denoted by RA(Γ,G,Hπ). Notice that when G consists of the trivial subgroup only, one recovers the notion of proper, (anti-)symmetric arrays as defined in [9]. For further discussion on arrays the reader may consult section 1 in [9]. When considering exact groups, the above notion of relative array into the left regular representation is closely related with the notion of bi-exactness introduced by Ozawa (Definition 15.1.2 in [5]). We are indebted to Narutaka Ozawa for kindly demonstrating to us the direct implication in the following result. Proposition 2.3. Let Γ be an exact group together with G a family of subgroups. Then RA(Γ,G, ℓ2(Γ)) 6= ∅ if and only if Γ is bi-exact with respect to G. Remark 2.4. A recent result of Popa and Vaes [58] establishes the same result under the weaker assumption that RA(Γ,G,Hπ) 6= ∅ for some weakly-ℓ2 represen- tation π. Proof. The reverse implication can be shown using the same method as in [9] and therefore we only prove the direct implication. So let r : Γ → ℓ2(Γ) an array relative to the family G and denote by π : Γ → U(ℓ2(Γ)) the left regular representation. Let Prob(Γ) be the set of positive Borel probability measures on Γ. For any f ∈ ℓ∞(Γ) we let φ(f ) the natural "diagonal" operator acting by pointwise multiplication. Then we define the map µ : Γ → Prob(Γ) by letting 1 hµ(γ), fi = kr(γ)k2hφ(f )r(γ), r(γ)i, for all γ ∈ Γ and f ∈ ℓ∞(Γ). Also if we fix s, t ∈ Γ then denoting by Cs = supγ∈Γ kr(sγ)− πs(r(γ))k and using the triangle inequality together with the (anti- )symmetry of the array r we have that (1) kr(sγt) − πs(rγ)k ≤ kr(sγt) − r(sγ)k + kr(sγ) − πs(r(γ))k = kπsγt(r(t−1γ−1s−1)) − πsγ(r(γ−1s−1))k + Cs = kπt(r(t−1γ−1s−1)) − r(γ−1s−1)k + Cs ≤ Ct + Cs, II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 7 for all γ ∈ Γ. have In the remaining part we will use this estimate to show that for all s, t ∈ Γ we (2) lim γ→∞/G kµ(sγt) − s · µ(γ)k = 0, which in turn will give the desired conclusion. To see this we fix s, t, γ ∈ Γ and f ∈ ℓ∞(Γ). Then applying the triangle inequality in combination with (1) and the Cauchy-Schwarz inequality we have ≤ ≤ 1 hµ(sγt), fi − hs · µ(γ), fi kr(sγt)k2 hφ(f )r(sγt), r(sγt) − πs(r(γ))i + 1 Cs + Ct 1 + 1 kr(γ)k2(cid:19)hφ(f )r(sγt), πs(r(γ))i(cid:12)(cid:12)(cid:12)(cid:12) + kr(sγt)k2 − +(cid:12)(cid:12)(cid:12)(cid:12)(cid:18) kr(sγt)k2 kφ(f )r(sγt)k +(cid:12)(cid:12)(cid:12)(cid:12) ≤ 2(Cs + Ct)kfk(cid:18) kr(γ)k2 hφ(f )r(sγt) − πs(r(γ)), r(γ)i kr(sγt)k2 − kr(γ)k(cid:19) . kr(γ)kkφ(f )r(sγt) − πs(r(γ))k kr(sγt)k 1 + 1 1 1 + 1 kr(γ)k2(cid:12)(cid:12)(cid:12)(cid:12)kφ(f )r(sγt)kkr(γ)k + Since r is assumed to be proper with respect to the set G then limγ→∞/G kr(sγt)k = ∞, limγ→∞/G kr(γ)k = ∞ and thus taking the limit in the previous inequality we get (2). (cid:3) 2.2. Relative quasi-cocycles. In the same spirit, if Γ is a group together with a family of subgroups G = {Σi : i ∈ I} and a unitary representation π : Γ → U(H), we say that pair (Γ,G) admits a relative quasi-cocycle into H if there exists a map r : Γ → H satisfying the following condition: (1) there exists a constant C > 0 such that sup γ,δ∈Γkr(γδ) − πγ(r(δ)) − r(γ)k ≤ C. (2) the map γ → kr(γ)k is proper relative to G. From now on, the set of all such relative quasi-cocycles we will denoted by RQ(Γ,G,Hπ). Using the terminology from [65], it is clear that RQ(Γ,G,Hπ) is a subset of QH 1(Γ,Hπ) which is stable under scalar multiplication and translation by uniformly bounded maps, without being in general a vector subspace. It is also straightforward that every relative quasi-cocycle is a relative array, i.e., we always have RQ(Γ,G,Hπ) ⊆ RA(Γ,G,Hπ). The next proposition summarizes a few basic properties which follow directly from definitions. Proposition 2.5. For each n ∈ N let Gn be a family of subgroups of Γ together with πn : Γ → U(Hn) a unitary representation. Then we have the following: (1) If G1 ⊂ G2 then RA(Γ,G1,Hπ1 ) ⊆ RA(Γ,G2,Hπ1); 8 I. CHIFAN, T. SINCLAIR, AND B. UDREA (2) If r ∈ RA(Γ,G1,H1) and c : Γ → H1 is a uniformly bounded map then r + c ∈ RA(Γ,G,H); (3) If Gn = G1 and πn = π1 for all n and there exists a sequence rn ∈ RA(Γ,G1,Hπ1) with uniformly bounded defects such that rn converges to r uniformly then r ∈ RA(Γ,G1,Hπ1 ); : Σ1 ∈ G1, Σj ∈ Gj , sj ∈ Γ}. If for every n ∈ N there exists cn > 0 and rn ∈ RA(Γ,G,Hπn ) satisfying (4) Denote by ∧nGn = {Σ1 ∩Tj6=1 sjΣjs−1 Pn c2 nkrn(γ)k2 < ∞ for all γ ∈ Γ, then RA(Γ,∧nGn,⊕Hπn) 6= ∅. j Cocycles, quasi-cocycles, and arrays combine both geometric and representation- theoretical data in a way that can be used to efficiently extract information about a group's internal structure. For instance, by the same proof as in Proposition 1.5.3 of [9] we can locate centralizers of certain subgroups and, in some cases, even normalizers. This property, generically called the "spectral gap rigidity principle", is the main intuition for the von Neumann algebraic structural results obtained in the subsequent sections. Proposition 2.6. Let Γ be a countable group, G be a family of subgroups, and π : Γ → U(Hπ) a representation such that RA(Γ,G,Hπ) 6= ∅. If Λ < Γ is a subgroup such that 1 ⊀ π ↾Λ then there exists h ∈ Γ and Σ ∈ G such that its centralizer CΓ(Λ) satisfies [CΓ(Λ) : hΣh−1 ∩ CΓ(Λ)] < ∞. Proof. Let q : Γ → Hπ be an array. Since 1 ⊀ π ↾Λ there exists a finite, symmetric subset S ⊂ Λ and K′ > 0 such that (3) kπs(ξ) − ξk, for all ξ ∈ Hπ. kξk ≤ K′Xs∈S Since q is an array there exists K′′ > 0 such that kq(sγ) − πs(q(γ))k ≤ K′′ for all s ∈ S and γ ∈ Γ. Set K = max{K′, K′′}. Then, for every γ ∈ CΓ(Λ) we have that kq(γ)k ≤ KXs∈S ≤ KXs∈S ≤ KXs∈S ≤ KXs∈S = KXs∈S ≤ 2K 2S kπs(q(γ)) − q(γ)k kq(sγ)) − q(γ)k + K 2S kq(γs)) − q(γ)k + K 2S kπγsq(s−1γ−1)) − πγ(q(γ−1))k + K 2S kπsq(s−1γ−1)) − q(γ−1)k + K 2S This shows that q is bounded on CΓ(Λ) and since q is proper with respect to the family G. It follows that CΓ(Λ) is small with respect to the family G which means that there exists a finite collection of groups Σi ∈ G and a finite set of Ωi = hiΣih−1 elements hi, ki ∈ Γ such that CΓ(Λ) ⊆ Si hiΣiki. Therefore, if we denote by , there exists a finite set of elements ℓi ∈ Γ such that CΓ(Λ) ⊆Si Ωiℓi. i II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 9 In particular this implies that CΓ(Λ) =Si (Ωiℓi ∩ CΓ(Λ)). After dropping all the empty intersections we can assume that Ωiℓi ∩ CΓ(Λ) 6= ∅, for all i. Hence there exists si ∈ Ωi such that ri = siℓi ∈ CΓ(Λ) and we obviously have that CΓ(Λ) =[i (Ωiri ∩ CΓ(Λ)) =[i (Ωi ∩ CΓ(Λ)) ri. Finally, by Lemma 4.1 in [37], the previous relation implies that Ωi ∩ CΓ(Λ) have finite index in CΓ(Λ) and we are done. (cid:3) Here are two concrete situations when this happens: Λ has property (T) and the restriction π ↾Λ has no invariant vectors; Λ is not co-amenable with respect to a subgroup Σ < Γ and π is the left semi-regular representation ℓ2(Γ/Σ). Moreover, if Γ is weakly amenable (cf. §3), Λ is amenable, and the normalizing group satisfies 1 ⊀ π ↾NΓ(Λ) then Λ is small with respect to G. This is an easy consequence of Corollary 3.2 in the sequel but in the in this form can be shown by direct arguments similar to the above proof. Examples 2.7. There are many examples of groups that admit relative quasi- cocycles (arrays) into various representations. First we analyze a few examples arising from canonical group constructions: A. Exact sequences. Let L, K, Γ be groups such that 0 → L → K → Γ → 0 is a short exact sequence. If RA(Γ,{e}, ℓ2(Γ)) 6= ∅ then we have RA(K,{L}, ℓ2(K/L)) 6= ∅. B. Product groups. Let Γ1, Γ2, . . . , Γn be a collection of groups, and denote by Γ = Γ1 × Γ2 × ··· × Γn. For every 1 ≤ i ≤ n we denote by Γi the subgroup of the direct product Γ which consists of all elements whose ith coordinate is trivial. Assume that Gi is family of subgroups of Γi and denote by G =Si{Λ× Γi : Λ ∈ Gi}. If RA(Γi,Gi,Hi) 6= ∅ for all 1 ≤ i ≤ n, then RA(Γ,G,⊗iHi) 6= ∅. For the proof of this fact see Proposition 1.10 in [9]. In particular Γ1 × Γ2 admits an array into a weakly-ℓ2 representation that is proper with respect to {Γ1, Γ2} whenever Γ1 and Γ2 admit proper arrays into weakly-ℓ2 representations. C. Semidirect products. Let Γ and A be countable discrete groups together with G a family of subgroups of Γ and assume that ρ : Γ → Aut(A) is an action by group automorphisms. Let π : Γ → U(Hπ) be a unitary representation and define π : A ⋊ρ Γ → U(Hπ) by letting πaγ(ξ) = πγ(ξ) for every a ∈ A, γ ∈ Γ and ξ ∈ Hπ. If c ∈ RA(Γ,G,Hπ) then the formula c(aγ) = c(γ) defines an array which belongs to RA(A ⋊ρ Γ,{A ⋊ρ Σ : Σ ∈ G},Hπ). We now look at semidirect products by finite groups. So let Γ be a countable discrete group together with a family of subgroups G, Λ be a finite group, and ρ : Λ → Aut(Γ) be an action by automorphisms. It is an exercise for the reader to check that for any r ∈ RA(Γ,G, ℓ2(Γ)), the map r′(γα) = λδ(r(ρδ−1 (γ))) 2 Xδ∈Λ 1 Λ 1 defines an array belonging to RA(Γ ⋊ρ Λ,G, ℓ2(Γ ⋊ρ Λ)), where γ ∈ Γ, α ∈ Λ and λ is the left regular representation on ℓ2(Γ ⋊ρ Λ). D. Free products. Let {Γn}1≤i≤n be a finite collection of groups. Denote by Γ = ⋆iΓi their free product, and let π : Γ → U(Hπ) a unitary representation. If for every 1 ≤ i ≤ n we have RQ(Γi,{e},Hπ) 6= ∅, then the proof of Lemma 5.1 and Theorem 10 I. CHIFAN, T. SINCLAIR, AND B. UDREA 5.3 in [65] show that RQ(Γ,{e},H⊕n π ) 6= ∅. Note that the when considering quasi- cocycles proper with respect to families of subgroups, it is not clear whether the resulting quasi-cocycle is proper to any canonical family of subgroups rather than just finite length subsets over the families of subgroups we started with. However, if we assume that Σ⊳Γi is a common normal subgroup, Γ = ⋆ΣΓi is the amalgamated free product over Σ, and for every 1 ≤ i ≤ n we have RQ(Γi/Σ,{e}, ℓ2(Γi/Σ)) 6= ∅ then RQ(Γ,{Σ}, ℓ2(Γ/Σ)) 6= ∅. In connection to this notice that if Σ is an amenable (non necessarily normal) subgroup then it follows from [5] that RA(Γ,{Gi : 1 ≤ i ≤ n}, ℓ2(Γ)∞) 6= ∅. E. HNN-extensions. Denote by Γ = (H, L, θ) the HNN-extension associated with a given inclusion groups L < H and a monomorphism θ : L → H. We also assume that K ⊳ H is a normal subgroup which contains L and θ(L) and from now on we will denote by L1 = L, L−1 = θ(L). The group Γ may be presented as {H, t : θ(ℓ) = tℓt−1, ℓ ∈ L}. By Britton's Lemma, every element γ ∈ Γ has a canonical reduced form γ = γ0tε1 γ1tε2 . . . γn−1tεn γn, where γi ∈ H, εi ∈ {−1, 1} and whenever εi 6= εi+1 we have that γi /∈ Lεi , for all 1 ≤ i ≤ n − 1. Assume that q : H/K → ℓ2(H/K) is a quasi-cocycle. By the construction in the first example there exists a quasi-cocycle c : H → ℓ2(H/K) which vanishes on K and moreover c ∈ RA(H,{K}, ℓ2(H/K)) whenever q is proper. We can define a map r : Γ → ℓ2(Γ/K) in the following way: for every γ = γ0tε1 γ1tε2 . . . γn−1tεn γn and s = 0, 1 we let q(γ) = λγ0tε1 γ1tε2 ...γn−1tεn c(γn) + λγ0tε1 γ1tε2 ...γn−1ds(tεn ) + rs +λγ0tε1 γ1tε2 ...γn−1tεn−1 c(γn−1) + λγ0tε1 γ1tε2 γn−2ds(tεn−1 ) + + . . . + λγ0 ds(tε1 ) + c(γ0), where d1(tε) = δtεK and d0(tε) = 0 for all ε ∈ {−1, 1}. Here λ denotes the left semi-regular representation ℓ2(H/K). It is a straightforward exercise to see that this map is well defined and it satisfies the quasi-cocycle relation. Moreover, when q = 0, the map is actually a 1-cocycle. Therefore, applying part (4) in Proposition 2.5 we have that r1 0 ⊕ r0 0 ⊕ r0 0 ⊕ r0 q ∈ RQ(Γ,{K}, ℓ2(Γ/K) ⊕ ℓ2(Γ/K)). q is a quasi- cocycle into ℓ2(Γ/K) ⊕ ℓ2(Γ/K). If q is assumed proper it follows that r1 q is proper with respect to various subsets of Γ, e.g. sets of words with finite length over t's whose letters from H are "small" over K. However, to have properness with respect to subgroups we need to impose additional assumptions on K. For instance, one may assume that L and θ(L) have finite index in K, in which case we would have r1 F. Inductive limits. Let Γn ր Γ be an inductive limit of groups and for each n ∈ N let Gn be a family of subgroups of Γn such that Gn ⊆ Gn+1. Assume that for each n, there exists rn ∈ RA(Γn,Gn, ℓ2(Γn)) so that: (1) supn,m supγ∈Γmin(n,m) krn(γ) − rm(γ)k < ∞; (2) supn∈N kCn(γ)k < ∞, for every γ ∈ Γ; (3) for every C > 0 there exists nC ∈ N such that for all n ≥ nC we have For every γ ∈ Γ we define a map r : Γ → ℓ2(Γ) by letting r(γ) = rn(γ), where n is chosen to be the smallest natural number such that γ ∈ Γn. The above properties then imply that r ∈ RQ(Γ,∪nGn, ℓ2(Γ)). {γ ∈ Γn+1 : krn+1(γ)k ≤ C} ⊂ Γn. II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 11 The reader may verify that this above construction together with Proposition 1.10 in [9] shows that if there exists a sequence rn ∈ RA(Γn,{e}, ℓ2(Γn)) with uniform bounded equivariance then r ∈ RA(⊕nΓn,{e}, ℓ2(⊕Γn)). In particular we have that if RA(Γ,{e}, ℓ2(Γ)) 6= ∅ then RA(Γ⊕∞,{e}, ℓ2(Γ⊕∞)) 6= ∅. As expected, to obtain relative quasi-cocycles we have to impose stronger as- sumptions. For example, if there exist relative quasi-cocycles rn ∈ RQ(Γn,Gn, ℓ2(Γn)) satisfying supm,n supγ∈Γmin(n,m) krn(γ) − rm(γ)k < ∞, supn∈N Dn < ∞, and condi- tion (3), the same construction as before shows that RQ(Γ,∪nGn, ℓ2(Γ)) 6= ∅. We also notice that by a basic rescaling procedure the same conclusion follows if we completely drop the uniform boundedness on the defects Dn, keep condition (3), and replace the first condition by the following: there exists a sequence Kn ≥ Dn such that sup m,n sup γ∈Γmin(n,m) k 1 Kn rn(γ) − 1 Km rm(γ)k < ∞. The examples presented above arise more or less from canonical algebraic con- structions. More interestingly, relative quasi-cocycles on groups can be constructed naturally from purely geometric considerations. Below we single out a class of such examples which are intensely studied in geometric group theory. G. Relatively hyperbolic groups. The results in [34, 30, 31] imply that every Gromov hyperbolic group Γ admits a proper quasi-cocycle into a multiple of ℓ2(Γ) (Lemma 4.2 in [65]). Using similar reasoning we will show a relative version of this result for groups which are relatively hyperbolic in the sense of Bowditch [3]. Briefly, given a group Γ together with a family of subgroups G, we say that Γ is hyperbolic relative to G if there exists a graph K on which Γ acts such that the following conditions are satisfied: a) Γ and every Σ ∈ G are finitely generated, b) K is fine (see (1) in Definition 2 from [3]) and has thin triangles, c) there are finitely many orbits and each edge stabilizer is finite, and d) the infinite vertex stabilizers are precisely the elements of G and their conjugates. Here are some examples of relatively hyperbolic groups: a free product is rela- tively hyperbolic with respect to its factors; if Γ is hyperbolic relative to a family of subgroups G and α : Σ1 → Σ2 is a monomorphism with Σi ∈ G, then the HNN extension Γ⋆α is hyperbolic with respect to G \ {Σ1, Σ2} [12]; geometrically finite Kleinian groups are hyperbolic with respect to their cusp subgroups [15]; the fun- damental group of a complete hyperbolic manifold of finite volume is hyperbolic relative to its cusp subgroups [15]; Sela's limit groups are hyperbolic relative to their maximal noncyclic abelian subgroups [12]. Mineyev and Yaman [32] showed that whenever Γ is hyperbolic relative to a finite set G of subgroups, there exists an ideal hyperbolic tuple (Γ,G, X, ν′) (Definition 42 in [32]). Furthermore, using this in combination with the machinery developed in [30], they constructed a homological Q-bicombing in X which is Γ-equivariant, anti- symmetric, quasi-geodesic, and has bounded area (Theorem 47 in [32]). Therefore, applying the same arguments as in the proof of Theorem 7.13 of [35], we see that this bicombing gives rise naturally to relative quasi-cocycles for Γ into a multiple of the left semi-regular representations with respect to some conjugates of elements in G. In effect, the bounded area together with anti-symmetry will imply the quasi- cocycle relation and being quasi-geodesic will imply properness with respect to the family G. 12 I. CHIFAN, T. SINCLAIR, AND B. UDREA Proposition 2.8. If a group Γ is hyperbolic relative to a finite family of subgroups G, then we have that RQ(Γ,G,⊕i,jℓ2(Γ/γjΣiγ−1 )) 6= ∅, for some γj ∈ Γ and Σi ∈ G. j Finally, we mention that from the work of Ozawa it is known that for every group Γ that is relatively hyperbolic with respect to a family of amenable subgroups we have that RA(Γ,{e}, ℓ2(Γ)∞) 6= ∅. 3. Weak Amenability for Groups and von Neumann Algebras The notion of weak amenability for groups was introduced by Cowling and Haagerup in [10]. There are several equivalent definitions ([4, 10]) and for the reader's convenience we recall the following: Definition 3.1. A countable discrete group Γ is said to be weakly amenable with constant C if there exists a sequence of finitely supported functions φn : Γ → C (completely bounded) norm of the Schur multiplier on B(ℓ2(Γ)) associated with such that φn → 1 pointwise and lim supn kbφnkcb ≤ C, where kbφnkcb denotes the the kernel bφn : Γ × Γ → C given by bφn(γ, δ) = φn(γ−1δ). The Cowling-Haagerup constant Λcb(Γ) is defined to be the infimum of all C for which such a sequence (φn) exists. If Γ is not weakly amenable then we write Λcb(Γ) = ∞. specifying their Cowling-Haagerup constants: Below we summarize some families of groups known to be weakly amenable, also (1) all amenable groups (Λcb(Γ) = 1); (2) all lattices in SO(n, 1) and SU (n, 1) (Λcb(Γ) = 1) or lattices in Sp(n, 1) (Λcb(Γ) = 2n − 1), [10]; (3) Coxeter groups (Λcb(Γ) = 1) [29]; (4) more generally, all groups which act properly on finite dimensional CAT(0)- cube complexes (Λcb(Γ) = 1), [18, 33]; (5) all hyperbolic groups (in this case no explicit constants were computed), [41]. (6) all limit groups in the sense of Sela (Λcb(Γ) = 1); this is an observation due to Ozawa based on a result from [19]. Groups which are not weakly amenable include Z2 ⋊ SL2(Z), [23] (see also, [13]), lattices in higher-rank simple Lie groups, and any non-amenable wreath products of the form Z ≀ Σ, [43]. The class of weakly amenable groups is closed under taking subgroups, cartesian products, co-amenable extensions, measure equivalence [43], and inductive limits of groups with uniformly bounded Cowling-Haagerup constants. However, it is not known whether weak amenability is closed under taking a free product of two groups except in the case that the Cowling–Haagerup constants of both groups are one [60]. By analogy with the group case discussed above, one can define a similar ap- proximation property for von Neumann algebras. The precise formulation is the following. Definition 3.2. A von Neumann algebra M is said to have the weak* completely bounded approximation property, abbreviated WCBAP, if there is a sequence of II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 13 ultraweakly-continuous finite-rank maps (φn) on M such that φn → idM in the point-ultraweak topology and lim supn kφnkcb < ∞. In [45] Ozawa and Popa discovered that the presence of this finite-dimensional approximation (with constant one) on a group imposes a certain type of "rigidity" on its internal structure. More precisely, they showed that if Λcb(Γ) = 1 then for any amenable subgroup Ω < Γ with non-amenable normalizing group NΓ(Ω) there exists an Ω ⋊ NΓ(Ω) invariant state on ℓ∞(Ω), where the semidirect product Ω ⋊ NΓ(Ω) acts on Ω by (γ, a) · x = γaxγ−1. In other words, the natural action of the normalizer NΓ(Ω) on Ω is fairly "small"; for instance, it cannot be of Bernoulli type. Later, Ozawa showed that in fact all weakly amenable groups satisfy this property, [43]. In fact, this rigidity even manifests in the von Neumann–algebraic context, as follows: Theorem 3.3 (Ozawa and Popa [45], Ozawa [43]). Let M be a von Neumann algebra which has WCBAP and let P ⊂ M be a diffuse amenable subalgebra. Then the natural action by conjugation of the normalizer NM (P ) y P is weakly compact, i.e., there exists a net of positive unit vectors (ηn)n∈N in L2(M ) ¯⊗L2( ¯M ) such that: (A) kηn − (v ⊗ ¯v)ηnk → 0, for all v ∈ U(P ); (B) k[u ⊗ ¯u, ηn]k → 0, for all u ∈ NM (P ); (C) h(x ⊗ 1)ηn, ηni = τ (x) = h(1 ⊗ ¯x)ηn, ηni, for all x ∈ M . In combination with deformation techniques, weak compactness turned out to be an powerful tool for obtaining many important structural results for group–measure space factors [45, 46, 21, 63]. 4. The Gaussian Construction, Bimodules and Weak Containment Given an orthogonal representation π : Γ → O(HR) of a countable, discrete group there exists a way of associating to it a p.m.p. action of Γ on a measure space such that the induced Koopman representation is unitarily equivalent to the infinite direct sum of the symmetric tensor powers of π (see the proof of Lemma 3.5 in [66]). This is called the Gaussian construction associated to (Γ, π,HR). We briefly describe this construction here, indicating how it can be extended to measure preserving actions by product groups. If π : Γ → O(HR) is an orthogonal representation, the Gausssian construction as described in [49] or [63] provides a probability measure space (Yπ, ν) and a family ω(ξ)ξ∈H of unitaries in L∞(Yπ) such that L∞(Yπ) is generated as a von Neumann algebra by the ω(ξ)'s and the following relations hold: (1) ω(0) = 1, ω(ξ1 + ξ2) = ω(ξ1)ω(ξ2), ω(ξ)∗ = ω(−ξ) for all ξ, ξ1, ξ2 ∈ HR (2) τ (ω(ξ)) = exp(−kξk2) where τ is the trace on L∞(Yπ) given by integration. The action σ of Γ on L∞(Yπ) is given by σg(ω(ξ)) = ω(πg(ξ)), for all ξ ∈ HR. Suppose now that Γ1 × Γ2 acts in a trace preserving manner on an abelian von Neumann algebra (A, τ ) and denote by M = A ⋊ (Γ1 × Γ2) the corresponding crossed product von Neumann algebra. For each i = 1, 2 let πi : Γi → O(Hi) be an orthogonal representation which is weakly contained in the (real) left regular representation of Γi. Let L2(Yπi)0 = L2(Yπi ) ⊖ C1 be the Koopman representation of the Gaussian action corresponding to πi which, by the assumptions, it is also weakly contained in the left regular representation. Consider the Hilbert space K = 14 I. CHIFAN, T. SINCLAIR, AND B. UDREA L2(A) ¯⊗L2(Yπ1)0 ¯⊗L2(Yπ2 )0 ¯⊗ℓ2(Γ1 × Γ2) with the M -bimodular structure defined as (aug) · (ξ ⊗ ξ1 ⊗ ξ2 ⊗ δk) · (buh) = (aσg(ξ)σgk(b)) ⊗ (πg(ξ1 ⊗ ξ2)) ⊗ (δgkh), for every a, b ∈ A, ξ ∈ L2(A), ξ1 ∈ L2(Yπ1 )0, ξ2 ∈ L2(Yπ2 )0, and g, k, h ∈ Γ1 × Γ2. Here π = π1 ⊗ π2. One of the key ingredients needed in the proof of Theorem 6.5 is that whenever A is amenable the above M -bimodule is weakly contained in the coarse M -bimodule. Lemma 4.1 (Fell's absorption principle). As an M -bimodule, K is isomorphic with a multiple of L2(hM, Ai, T r). In particular, when A is amenable, it follows that K is weakly contained in the coarse bimodule, L2(M ) ¯⊗L2(M ). Proof. First we notice that when πi is weakly contained in ρi then the bimodule associated to the pair (π1, π2) is weakly contained in the bimodule associated with the pair (ρ1, ρ2). It is therefore enough to prove the statement in the case when πi is the (real) left regular representation of Γi. Throughout the proof, we denote by Γ = Γ1×Γ2. Since K is canonically identified with L2(A) ¯⊗ℓ2(Γ) ¯⊗ℓ2(Γ), we will obtain the desired conclusion by showing that the map L2(A) ¯⊗ℓ2(Γ) ¯⊗ℓ2(Γ) ∋ ξ ⊗ δg ⊗ δh → ξugeAug−1h ∈ L2(hM, Ai, T r) implements an isomorphism between the two bimodules. To this purpose it suffices to show that (4) h(aus) · (ξ ⊗ δg ⊗ δh) · (but), (a′us′ ) · (ξ′ ⊗ δg′ ⊗ δh′) · (b′ut′)i = haus(ξugeAug−1h)but, a′us′(ξ′ug′ eAug′−1h′ )b′ut′iT r, for all a, a′, b, b′ ∈ A, ξ, ξ′ ∈ L2(A), and s, t, g, h, s′, t′, g′, h′ ∈ Γ. to On the one hand, by definitions, the left side in the previous equation is equal h(aus) · (ξ ⊗ δg ⊗ δh) · (but), (a′us′) · (ξ′ ⊗ δg′ ⊗ δh′) · (b′ut′)i = h(aσs(ξ)σsh(b)) ⊗ δsg ⊗ δsht, (a′σs′ (ξ′)σs′h′ (b′)) ⊗ δs′g′ ⊗ δs′h′t′i = δsg,s′g′ δsht,s′h′t′h(aσs(ξ)σsh(b)), (a′σs′ (ξ′)σs′h′(b′))i = δsg,s′g′ δsht,s′h′t′ τ (σs′h′(b′∗)σs′ (ξ′∗)a′∗aσs(ξ)σsh(b)). On the other hand, using basic computations and τ (σs′g′ (x)) = τ (x) for all x ∈ A we see that the right side of (4) is equal to haus(ξugeAug−1h)but, a′us′(ξ′ug′ eAug′−1h′ )b′ut′iT r = T r(ut′−1 b′∗uh′−1g′ eAug′−1ξ′∗us′−1 a′∗ausξugeAug−1hbut) = T r(eAug′−1 ξ′∗us′−1 a′∗ausξugeAug−1hbutt′−1b′∗uh′−1g′ eA) = τ (EA(ug′−1 ξ′∗us′−1 a′∗ausξug)EA(ug−1hbutt′−1b′∗uh′−1g′ )) = τ (EA(σg′−1 (ξ′∗)σg′−1s′−1 (a′∗aσs(ξ))ug′−1s′−1sg)EA(σg−1h(b)σg−1htt′−1 (b′∗)ug−1htt′−1h′−1g′ )) = δg′−1s′−1sg,eδg−1htt′−1h′−1g′,eτ (σg′−1 (ξ′∗)σg′−1s′−1 (a′∗aσs(ξ))σg−1h(b)σg−1htt′−1 (b′∗) = δsg,s′g′ δg−1ht,g′−1h′t′ τ (σ(s′g′)−1 (σs′ (ξ′∗)a′∗aσs(ξ)σs′g′g−1h(b)σs′g′g−1htt′−1 (b′∗)) = δsg,s′g′ δsht,s′h′t′τ (σs′ (ξ′∗)a′∗aσs(ξ)σsh(b)σs′h′(b′∗)). This establishes (4) and hence the conclusion of the lemma. (cid:3) II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 15 5. A Path of Automorphisms of the Extended Roe Algebra Associated with the Products of Gaussian Actions Let Γ = Γ1 × Γ2 yσ X be a measure preserving action of Γ on a measure : Γ1 → O(Hi) . space X. Assume we are given orthogonal representations πi As shown in the previous section, to these representations we can associate the Gaussian actions Γi yπi (Yπi , νi) (in a slight abuse of notation we will denote the Gaussian action by the same letter). Next we consider the product action Γ yπ1⊗π2 (Yπ1 × Yπ2 , ν1 × ν2) and the diagonal action of Γ on (X × Yπ1 × Yπ2, µ× ν1 × ν2). To this action, following [9], we can associate the extended Roe algebra C∗u(Γ y Z) (where Z = X × Yπ1 × Yπ2 ). Additionally, given any pair of quasi-cocycles qi : Γi → Hi for the respective representations πi, i = 1, 2, we can construct a one-parameter family (αt)t∈R of ∗-automorphisms of C∗u(Γ y Z), by exponentiating the qi's. This traces back to the construction of a malleable deformation of LΓ from a cocycle b as carried out in §3 of [63]. Moreover, this family will be pointwise continuous with respect to the uniform norm as t → 0 (Theorem 5.3). Given the quasi-cocycles qi : Γi → Hi, one can construct, following section §1.2 of [63], two one-parameter families of maps υi t : Γi → U(L∞(Yπi , νi)) defined by the t(γi)(x) = exp(√−1tqi(γi)(x)), where γi ∈ Γi, x ∈ Yπi , respectively. To formula υi understand this formula, the reader must think about Hi as being identified with a subspace of L2(Yπi, νi), viewing the elements qi(γi) as functions on Yπi. The same computations as in [49, 63] show the following: Proposition 5.1. Assuming the same notations as above, we have that: (1) If the representation πi is weakly-ℓ2, i = 1, 2, then the (tensor) product of (1) VtVs = Vt+s, VtV ∗t = V ∗t Vt = 1 (2) If the array is anti-symmetric we have JVtJ = Vt and if it is symmetric we have JVtJ = V−t. Here we denoted by J : L2(Z) ¯⊗ℓ2(Γ) → L2(Z) ¯⊗ℓ2(Γ) is Tomita's conjugation. The unitary Vt implements an inner ⋆-automorphism αt on B(L2(Z) ¯⊗ℓ2(Γ)) by for all x ∈ B(L2(Z) ¯⊗ℓ2(Γ)). The αt then restricts to a family letting αt(x) = VtxV ∗t of inner automorphisms of the uniform Roe algebra. Moreover, when restricting to the uniform Roe algebra one can recover from αt the multipliers introduced in section 2 of [9] by the formula mt([xγ,δ]) = ([κt(γ, δ)xγ,δ]). Precisely, we have EM ◦ αt(x) = 1X ⊗ mt(x) for all x ∈ C∗u(Γ). The same computations as in [9] can be used to show that, αt, when restricted to the Roe algebra, is a C∗-deformation, i.e., it is pointwise-k · k∞ continuous. t(γi)(y)υi 0(Yπ2 ) is also weakly-ℓ2; 0(Yπ1 )⊗L2 t(γi, δi), i = 1, 2, and γi, δi ∈ Γi. Koopman representations π1 ⊗ π2L2 t(δi)∗(y)dµπi (y) = κi Y υi t(γi, δi) = exp(−tkqi(γi) − qi(δi)k). (2) R πi Here, κi With the help of these maps we can construct a path of unitary operators Vt ∈ B(L2(Z) ¯⊗ℓ2(Γ)) = B(L2(Yπ1 ) ¯⊗L2(Yπ2 ) ¯⊗L2(X) ¯⊗ℓ2(Γ)) by letting Vt(ξ1 ⊗ ξ2 ⊗ η⊗ t (γ2)ξ2 ⊗ η ⊗ δ(γ1,γ2) for every η ∈ L2(X), ξi ∈ L2(Yπi ), δ(γ1,γ2)) = υ1 t (γ1)ξ1 ⊗ υ2 and γi ∈ Γi, where i = 1, 2. The computations in [9] show that the Vt enjoy the following basic properties. Proposition 5.2. For every t, s ∈ R we have that: 16 I. CHIFAN, T. SINCLAIR, AND B. UDREA Theorem 5.3 (Lemma 2.6 in [9]). Let q be any symmetric or anti-symmetric array. Assuming the notations above, for every x ∈ L∞(X) ⋊σ,r Γ we have (5) (6) where k · k∞ denotes the operatorial norm in B(L2(X) ¯⊗ℓ2(Γ)); here e denotes the orthogonal projection from L2(Z) ¯⊗ℓ2(Γ) onto L2(X) ¯⊗ℓ2(Γ). The proof of the following result is straightforward and we leave it to the reader. k(αt(x) − x) ◦ ek∞ → 0 as t → 0; k(αt(JxJ) − JxJ) ◦ ek∞ → 0 as t → 0, Proposition 5.4. Let q be any symmetric or anti-symmetric array. Assuming the notations above, for every x ∈ L∞(X) ⋊σ,r Γ, v ∈ U(L∞(X) ⋊σ Γ), and every t ∈ R we have (7) k(αt ⊗ 1(x ⊗ v) − x ⊗ v) ◦ (e ⊗ 1)k∞ = k(αt(x) − x) ◦ ek∞; k(αt ⊗ 1( J(x ⊗ v) J) − J(x ⊗ v) J ) ◦ (e ⊗ 1)k∞ = k(αt(JxJ) − JxJ) ◦ ek∞, (8) where on the left hand side of the above formulas k · k∞ denotes the operatorial norm in B([L2(X) ¯⊗ℓ2(Γ)] ¯⊗[L2(X) ¯⊗ℓ2(Γ)]), J is Tomita's conjugation operator on [L2(Z) ¯⊗ℓ2(Γ)] ⊗ [L2(X) ¯⊗ℓ2(Γ)], and e ⊗ 1 is the orthogonal projection from [L2(Z) ¯⊗ℓ2(Γ)] ¯⊗[L2(X) ¯⊗ℓ2(Γ)] onto [L2(X) ¯⊗ℓ2(Γ)] ¯⊗[L2(X) ¯⊗ℓ2(Γ)]. Combining the two previous results we obtain the following Corollary 5.5. Let q be any symmetric or anti-symmetric array. Assuming the notations above, for every x ∈ L∞(X) ⋊σ,r Γ and v ∈ U(L∞(X) ⋊σ Γ) we have sup η∈([L2(X) ¯⊗ℓ2(Γ)] ¯⊗[L2(X) ¯⊗ℓ2(Γ)])1 kVt ⊗ 1(η(x ⊗ v)) − (Vt ⊗ 1(η))(x ⊗ v)k → 0 and η∈([L2(X) ¯⊗ℓ2(Γ)] ¯⊗[L2(X) ¯⊗ℓ2(Γ)])1 kVt ⊗ 1((x ⊗ v)η) − (x ⊗ v)Vt ⊗ 1(η)k → 0, sup as t → 0. 6. Proofs of the Main Results We start by proving the main technical result of the paper which involves product of groups. Specifically, we obtain a result describing all weakly compact embeddings in the crossed product von Neumann algebras arising from actions of products of hyperbolic groups (Theorem 6.5). Our approach follows the general outline of the proof of Theorem B in [46] and Theorem B in [9]. However, it is based on a new ingredient which allows us to treat the more general case of arrays rather than just quasi-cocycles as proved in [9]. This was influenced by the approach taken in [7]. Theorem 6.1. Let Γ be an exact group together with a family of subgroups G = {Σi}, and let π : Γ → U(Hπ) be a weakly-ℓ2 representation. Also, let Γ y X be a free, ergodic action and denote by M = L∞(X) ⋊ Γ the corresponding crossed- product von Neumann algebra. 1. If RA(Γ,G,Hπ) 6= ∅ and P ⊆ M is a diffuse subalgebra, then there exist and P pi (cid:22)M L∞(X) ⋊ Σi, for all i. In particular, either P ′ ∩ M is amenable or there exists a group Σ ∈ G such that P (cid:22)M L∞(X) ⋊ Σ. 2. If RQ(Γ,G,Hπ) 6= ∅ and P ⊆ M is a weakly compact embedding with P projections p0, pi ∈ Z(P ′ ∩ M ) such that p0 +Wi pi = 1, (P ′ ∩ M )p0 is amenable diffuse, then there exist projections p0, pi ∈ Z(NM (P )′∩M ) such that p0+Wi pi = 1, II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 17 p0, pi ∈ Z(NM (P )′ ∩ M ) such that p0 +Wi pi = 1, NM (P )′′p0 is amenable and NM (P )′′p0 is amenable and P pi (cid:22)M L∞(X) ⋊ Σi, for all i. In particular, either NM (P )′′ is amenable or there exists a group Σ ∈ G such that P (cid:22)M L∞(X) ⋊ Σ. 3. Assume that G is a family of normal subgroups of Γ. If RA(Γ,G,Hπ) 6= ∅ and P ⊆ M is a weakly compact embedding with P diffuse, then there exist projections P pi (cid:22)M L∞(X) ⋊ Σi, for all i. In particular, either NM (P )′′ is amenable or there exists a group Σ ∈ G such that P (cid:22)M L∞(X) ⋊ Σ. Proof. As stated, the first part is Theorem 3.2 in [9] while the second part follows exactly as in the proof of Theorem 4.1 in [9]. Indeed the only ingredient needed for this is to adapt Proposition 2.6 in [9] to the case of quasi-cocycles that are proper with respect to a family of subgroups. This is a straightforward exercise, and we leave it to the reader. So we only prove the third part. We establish the following notations: Let {Σi : i ∈ I} be an enumeration of G, Mi = L∞(X) ⋊ Σi for each i ∈ I, N = NM (P )′′, and Z = Z(N′ ∩ M ). Let p0 ∈ Z be the maximal projection such that N p0 is amenable. For each i, let pi ∈ (P ′ ∩ M )(1 − p0) be a maximal projection satisfying P pi (cid:22)M L∞(X) ⋊ Σi (obtained via a standard maximality argument). By maximality, we must have that pi ∈ Z(P ′ ∩ M ). Moreover, we have that pi ∈ Z. Indeed, if u ∈ NM (P ), let pi = upiu∗(1 − p0 − pi). Then pi + pi satisfies the same condition and by the maximality of pi, we get that pi = 0. Thus pi = upiu∗, for every u ∈ NM (P ), hence for any i, by maximality. Also due to the maximality, N p has no amenable direct summand. pi ∈ Z. Therefore, to prove the theorem, we only need to show that p0 +W pi = 1. By contradiction, assume that p := 1 − (p0 +Wi pi) 6= 0. Note that P p (cid:14)M Mi, vectors (ηn)n∈N in L2(M ) ⊗ L2( ¯M ) such that: (A) kηn − (v ⊗ ¯v)ηnk → 0, for all v ∈ U(P ); (B) k[u ⊗ ¯u, ηn]k → 0, for all u ∈ NM (P ); and (C) h(x ⊗ 1)ηn, ηni = τ (x) = h(1 ⊗ ¯x)ηn, ηni, for all x ∈ M . By assumption P ⊂ M is weakly compact, so there exists a net of positive unit 0(Y π)⊗L2(X)⊗ℓ2(Γ) which as we remarked before is weakly contained So let H = L2 as an M -bimodule in the coarse bimodule. Fixing t > 0 we consider the unitary Vt associated with an array q as defined in the previous sections. Next denote by ηn,t = (Vt ⊗ 1)(p ⊗ 1)ηn, ζn,t = (e ⊗ 1)ηn,t = (e ◦ Vt ⊗ 1)(p ⊗ 1)ηn, and ξn,t = ηn,t − ζn,t = (e⊥ ⊗ 1)ηn,t ∈ H ⊗ L2(M ). Using these notations we show next the following inequality: Lemma 6.2. where Lim is an ultralimit on N. n kξn,tk ≥ Lim 1 2kpk2, Proof. We argue by contradiction, so by passing to a subsequence we can assume that (9) kξn,tk < 1 2kpk2 for all n. Denoting by ζn = (p ⊗ 1)ηn we have kηn,tk = kζnk = kpk2 and using the identity k(e ⊗ 1)(ηn,t)k2 + k(e⊥ ⊗ 1)(ηn,t)k2 = kηn,tk2 = kpk2 2 in combination with (9) we 18 have (10) I. CHIFAN, T. SINCLAIR, AND B. UDREA k(e ⊗ 1)(ηn,t)k > √3 2 kpk2, for all n. The main strategy is to prove that relation (10) together with the assumption P p (cid:14)M Mi, for all i will enable us to show that, after passing to an infinite sub- sequence of ζn, one can construct an infinite sequence F1,F2,F3, . . . of mutually disjoint finite subsets of Γ satisfying the following property: there exists i ∈ I and 1 > D > 0 such that for every k ∈ N we can find nk ∈ N such that for all j ≤ k we have (11) kPΣiFj (ζnk )k ≥ D √j kpk2. Here and throughout the proof PΣiFj stands for the orthogonal projection from L2(M ) ¯⊗L2( ¯M ) onto the closed linear span of the set {Miug ⊗ L2( ¯M ) : g ∈ Fj}. First, we briefly explain how this claim leads to a contradiction, thus finishing the proof of the lemma. Since the sets Fj are disjoint, relation (11) implies kpk2 2 = kζnkk2 ≥ impossible when letting k be sufficiently large. ), for all k ∈ N. This is obviously kPΣiFj (ζnk )k2 2 ≥ D2kpk2 2( kXj=1 kXj=1 1 j So we are left to prove (11). To show this we will proceed by induction on k. We begin by establishing the base case k = 1. Since ζn ∈ L2(M ) ¯⊗L2( ¯M ), we write ζn =Pg∈Γ ζn g ∈ L2(X) ¯⊗L2( ¯M ). Then, using the definition of Vt, a straightforward computation shows that g δg, where ζn ke ⊗ 1(ηn,t)k2 =Xg∈Γ exp(−2t2kq(g)k2)kζg nk2, for all n. When combined with (10) this formula implies that, for all n we have exp(−2t2kq(g)k2)kζg nk2 > 3 4kpk2 2. Xg∈Γ Since the map g → kq(g)k is proper relative to the family G, then the set {g ∈ Γ : exp(−t2kq(g)k2) ≥ 1 2} is contained in F′RF′ for some finite subsets F′ ⊂ Γ and R ⊂ G and, using the inequality (12), we further deduce that 3 4kpk2 2 < 1 4 Xg∈Γ\F ′R′F ′ kζg nk2 + Xg∈F ′R′F ′ kζg nk2, for all n, where we have denoted by R′ = ∪Σ∈RΣ ⊂ Γ. By basic algebraic manipulations, the above inequality gives that Xg∈F ′R′F ′ kζg 2 3kpk2 2 for all n which implies nk2 > (12) (13) kPF ′R′F ′(ζn)k >r 2 3kpk2. Also note that since G is a family of normal subgroups then there exists finite subset F ⊂ Γ such that F′R′F′ = R′F and hence by (13) we have (14) kPR′F (ζn)k >r 2 3kpk2. (17) KΣi ∩ ΣiS = ∅; kv′ − vpk2 6 D2 6k(k + 1)Fkpk2. II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 19 Notice that R′F = ∪Σ∈RΣF and if we denote by s = R then using the inclusion- exclusion principle together with the triangle inequality in (14) we see that after passing to an infinite subsequence ζn there exists Σi ∈ R such that for all n (15) So if we set D = n1 = 1. 1√22s−13 1 kPΣiF (ζn)k > then case k = 1 follows form (15) by letting F1 = F and √22s−13kpk2. To conclude we now show the inductive step, i.e., assuming that we have con- structed the sets F1,F2, . . . ,Fk ⊂ Γ, we indicate how to construct Fk+1 ⊂ Γ and j=1(FjF−1) ⊂ Γ and notice nk+1 ∈ N satisfying (11). Consider the finite set S = ∪k that since Σi is quasi-normal then the set ΣiSΣi is small over G. Thus, since S is finite and P p (cid:14)M Mi for any i, by Popa's intertwining techniques, there exist a unitary v ∈ U(P ), a finite set K ⊂ Γ, and an element v′ in the linear span of {uh : h ∈ K} such that (16) Next, we show that for n ∈ N and z ∈ M we have (18) Fix n and denote by P the orthogonal projection onto L2(Mi) ¯⊗L2( ¯M ), i.e., P = PΣi . We have PΣiF (ζn) =Ph∈F P (ζn(u∗h⊗ 1))(uh⊗ 1) and by the Cauchy-Schwarz inequality we deduce k(z ⊗ 1)PΣiF (ζn)k ≤ Fkzk2. (19) k(z ⊗ 1)PΣiF (ζn)k2 6 FXh∈F k(z ⊗ 1)P (ζn(u∗h ⊗ 1))k2 Now we let EMi to be the conditional expectation from M onto Mi and we denote by a = EMi (z∗z) 2 . Using the formulas h(x ⊗ 1)P (ζ), P (ζ)i = h(EMi (x) ⊗ 1)P (ζ), P (ζ)i and k(x ⊗ 1)ηnk = kxk2, for all ζ ∈ L2(M ) ¯⊗L2( ¯M ) and x ∈ M , we obtain the following: 1 k(z ⊗ 1)P (ζn(u∗h ⊗ 1))k2 = h(z∗z ⊗ 1)P (ζn(u∗h ⊗ 1)), P (ζn(u∗h ⊗ 1))i = h(EMi (z∗z) ⊗ 1)P (ζn(u∗h ⊗ 1)), P (ζn(u∗h ⊗ 1))i = k(a ⊗ 1)P (ζn(u∗h ⊗ 1))k2 ≤ k(a ⊗ 1)ζnk2 = kapk2 2 = kzk2 2. 2 ≤ kak2 2 = kP ((a ⊗ 1)ζn(u∗h ⊗ 1))k2 It is clear that the last inequalities combined with (19) give (18). Applying the triangle inequality, for all v ∈ U(P ) and all n ∈ N, we have (20) kζn − (v′ ⊗ ¯v)PΣiF (ζn)k ≤ kζn − (vp ⊗ ¯v)ζnk + kζn − PΣiF (ζn)k + k((v′ − vp) ⊗ 1)PΣiF (ζn)k 20 I. CHIFAN, T. SINCLAIR, AND B. UDREA Since p and v commute, we have ζn − (vp ⊗ ¯v)ζn = (p ⊗ 1)(ηn − (v ⊗ ¯v)ηn). Thus, since limn→∞ kηn − (v ⊗ ¯v)ηnk2 = 0, we can find nk+1 ≥ nk such that for all n ≥ nk+1 we have (21) D2 kζn − (vp ⊗ ¯v)ζnk ≤ 6k(k + 1)kpk2. Using (18) for z = vp − v′ in combination with (17) for all n we have (22) D2 k((v′ − vp) ⊗ 1)PF (ζn)k ≤ 6k(k + 1)kpk2. Altogether, (20), (21), (22), and (15) show that that for all n ≥ nk+1 we have (23) kζn − (v′ ⊗ ¯v)PΣiF (ζn)k ≤ 3k(k + 1)kpk2 + k(ζn − PΣiF (ζn))k D2 < D2 3k(k + 1) +(cid:18)1 − D2 k (cid:19) 1 2!kpk2. Finally, we let Fk+1 = KF and because Σi is normal in Γ and it follows from (16) that ΣiFk+1 is disjoint from ΣiF , ΣiF2, . . . , ΣiFk. Moreover, we have that (v′ ⊗ v)PΣiF (ζn) belongs to the closed linear span of {Miuh ⊗ ¯M : h ∈ Fk+1}. Thus, PΣiFk+1((v′ ⊗ ¯v)PΣiF (ζn)) = (v′ ⊗ ¯v)PΣiF (ζn) and by (23) we have that 2!kpk2, for all n ≥ nk+1. kζn − PΣiFk+1(ζn)k < D2 3k(k + 1) +(cid:18)1 − k (cid:19) 1 D2 Using this in combination with D < 1 then basic calculations show that, for all n ≥ nk+1, we have kPΣiFk+1(ζn)k >1 − = D2 ≥(cid:18) D2 k − which ends the proof of (11). k − D2 3k2(k + 1)2 +(cid:18)1 − 1 2 D2 2!2 k (cid:19) 1 3k2(k + 1)2(cid:18)1 − 2D2 D4 3k4(k + 1)4 − 2 D2 k(k + 1)(cid:19) 1 = D √k + 1 , kpk2 D2 2 2! 1 k (cid:19) 1 kpk2 (cid:3) Lemma 6.3 (Lemma 4.4 in [9]). For every ε > 0 and every finite set K ⊂ L∞(X) ⋊σ,r Γ with distk·k2(y, (N )1) ≤ ε for all y ∈ K one can find tε > 0 and a finite set LK,ε ⊂ NM (P ) (in this set we allow the situation when the same element is repeated finitely many times) such that (24) h((yx − xy) ⊗ 1)ξn,t, ξn,ti ≤ 10ε + Xv∈LK,ε k[v ⊗ ¯v, ηn]k, for all y ∈ K, kxk∞ ≤ 1, tε > t > 0, and n. II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 21 Note. After publication we discovered that the proof of Lemma 4.4 as given in [9] contains a minor gap which does not affect the substance of the argument. We take the opportunity to provide a corrected proof. Proof. Fix ε > 0 and y ∈ K. Since N = NM (P )′′ by the Kaplansky density theorem there exists a finite set Sy = {vi} ⊂ NM (P ) and scalars µi such that kPi µivik∞ ≤ 1 and (25) ky −Xi µivik2 ≤ 0.5ε. Also by the Kaplansky density theorem there exists contractions wi ∈ L∞(X)⋊σ,r Γ such that for all i and all y ∈ K we have (26) =: ε′. ε kw∗i − v∗i k2 = kwi − vik2 ≤ Since the elements wi, y ∈ L∞(X)σ,r ⋊ Γ, then using Corollary 5.5 one can find a positive number tε > 0 such that, for all tε > t ≥ 0 and all i we have the following seven inequalities: Pi µi (27) n kVt ⊗ 1((p ⊗ 1)ηn(y∗ ⊗ 1)) − (Vt ⊗ 1((p ⊗ 1)ηn))(y∗ ⊗ 1)k ≤ ε; sup n kVt ⊗ 1((y∗ ⊗ 1)(p ⊗ 1)ηn) − (y∗ ⊗ 1)(Vt ⊗ 1((p ⊗ 1)ηn))k ≤ ε; sup n kVt ⊗ 1((w∗i ⊗ ¯v∗i )(p ⊗ 1)ηn) − (w∗i ⊗ ¯v∗i )(Vt ⊗ 1((p ⊗ 1)ηn))k ≤ ε′; sup n kVt ⊗ 1((p ⊗ 1)ηn(w∗i ⊗ ¯v∗i )) − (Vt ⊗ 1((p ⊗ 1)ηn))(w∗i ⊗ ¯v∗i )k ≤ 0.5ε′; sup n kVt ⊗ 1((p ⊗ 1)ηn(wi ⊗ ¯vi)) − (Vt ⊗ 1((p ⊗ 1)ηn))(wi ⊗ ¯vi)k ≤ 0.5ε′; sup n kVt ⊗ 1((wi ⊗ ¯vi)(p ⊗ 1)ηn) − (wi ⊗ ¯vi)(Vt ⊗ 1((p ⊗ 1)ηn))k ≤ ε′; sup n kVt ⊗ 1((Xi sup µiwi − y) ⊗ 1)(p ⊗ 1)ηn) − (Xi µiwi − y) ⊗ 1)(Vt ⊗ 1((p ⊗ 1)ηn))k ≤ ε. Next we will proceed in several steps to show inequality (24). First we fix tε > t > 0. Then, using the triangle inequality in combination with kxk∞ ≤ 1, the second inequality in (27), and the M -bimodularity of e⊥ = 1 − e, we see that h(x ⊗ 1)ξn,t, (y∗ ⊗ 1)ξn,ti − h(xy ⊗ 1)ξn,t, ξn,ti n kVt ⊗ 1((y∗ ⊗ 1)(p ⊗ 1)ηn) − (y∗ ⊗ 1)(Vt ⊗ 1((p ⊗ 1)ηn))k+ ≤ sup + h(x ⊗ 1)ξn,t, (e⊥Vty∗p ⊗ 1)ηni − h(xy ⊗ 1)ξn,t, ξn,ti ≤ ε + h(x ⊗ 1)ξn,t, (e⊥Vty∗p ⊗ 1)ηni − h(xy ⊗ 1)ξn,t, ξn,ti Further, the Cauchy-Schwarz inequality together with statement (C) from the def- inition of weak compactness and (25) enable us to see that the last quantity above is smaller than ≤ ε + k((y∗p −Xi ≤ 1.5ε + Xi ¯µiv∗i )p ⊗ 1)ηnk + Xi µih(x ⊗ ¯v∗i )ξn,t, (e⊥Vtpv∗i ⊗ ¯v∗i )ηni − h(xy ⊗ 1)ξn,t, ξn,ti µih(x ⊗ 1)ξn,t, (e⊥Vtv∗i p ⊗ 1)ηni − h(xy ⊗ 1)ξn,t, ξn,ti 22 I. CHIFAN, T. SINCLAIR, AND B. UDREA Since vi is a unitary then by applying the triangle inequality several times the last quantity above is smaller than (28) ≤ 1.5ε +Xi ≤ 1.5ε +Xi + Xi ≤ 1.5ε +Xi + Xi µik[v∗i ⊗ ¯v∗i , ηn]k + Xi µi (k[v∗i ⊗ ¯v∗i , ηn]k + k(p ⊗ 1)ηn((w∗i − v∗i ) ⊗ ¯v∗i )k) + µih(x ⊗ ¯v∗i )ξn,t, (e⊥Vtp ⊗ 1)(ηnv∗i ⊗ ¯v∗i )i − h(xy ⊗ 1)ξn,t, ξn,ti µih(x ⊗ ¯v∗i )ξn,t, (e⊥Vtp ⊗ 1)(ηnw∗i ⊗ ¯v∗i )i − h(xy ⊗ 1)ξn,t, ξn,ti µi (k[v∗i ⊗ ¯v∗i , ηn]k + kηn((w∗i − v∗i ) ⊗ 1)k) + µih(x ⊗ ¯v∗i )ξn,t, (e⊥Vtp ⊗ 1)(ηnw∗i ⊗ ¯v∗i )i − h(xy ⊗ 1)ξn,t, ξn,ti We notice that, since ηn is a positive vector and J is an isometry then for all z ∈ M we have kηn(z ⊗ 1)k = kJ(z∗ ⊗ 1)Jηnk = k(z∗ ⊗ 1)ηnk = kz∗k2 = kzk2. Using this identity (right traciality) in combination with (26), the triangle inequality, and the fourth respectively the fifth inequality in (27) imply that the last quantity in (28) is smaller than n kVt ⊗ 1((p ⊗ 1)ηn(w∗i ⊗ ¯v∗i ))− µih(x ⊗ ¯v∗i )ξn,t, (e⊥Vtp ⊗ 1(ηn))(w∗i ⊗ ¯v∗i )i − h(xy ⊗ 1)ξn,t, ξn,ti µik[v∗i ⊗ ¯v∗i , ηn]k + Xi µik[v∗i ⊗ ¯v∗i , ηn]k+ µih(x ⊗ ¯v∗i )(e⊥Vtp ⊗ 1(ηn))(wi ⊗ ¯vi), ξn,ti − h(xy ⊗ 1)ξn,t, ξn,ti ≤ 2.5ε +Xi µi(k[v∗i ⊗ ¯v∗i , ηn]k + sup − (Vt ⊗ 1((p ⊗ 1)ηn))(w∗i ⊗ ¯v∗i )k)+ + Xi ≤ 3.5ε +Xi = 3.5ε +Xi + Xi ≤ 3.5ε +Xi + Xi µi(k[v∗i ⊗ ¯v∗i , ηn]k + sup n kVt ⊗ 1((p ⊗ 1)ηn(wi ⊗ ¯vi)) − (Vt ⊗ 1(p ⊗ 1ηn))(wi ⊗ ¯vi)k) µih(x ⊗ ¯v∗i )(e⊥Vt ⊗ 1((p ⊗ 1)ηn(wi ⊗ ¯vi)), ξn,ti − h(xy ⊗ 1)ξn,t, ξn,ti µih(x ⊗ ¯v∗i )ξn,t(wi ⊗ ¯vi), (e⊥Vtp ⊗ 1(ηn))i − h(xy ⊗ 1)ξn,t, ξn,ti Using the right traciality of ηn, (26), the fifth inequality in (27), and vi being a unitary which commutes with p we see that the last quantity above is smaller than II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 23 µi (k[v∗i ⊗ ¯v∗i , ηn]k + k(p ⊗ 1)ηn((wi − vi) ⊗ ¯vi)k) µih(x ⊗ ¯v∗i )(e⊥Vt ⊗ 1((p ⊗ 1)ηn(vi ⊗ ¯vi)), ξn,ti − h(xy ⊗ 1)ξn,t, ξn,ti µi (k[v∗i ⊗ ¯v∗i , ηn]k + k[vi ⊗ ¯vi, ηn]k) µih(x ⊗ ¯v∗i )(e⊥Vt ⊗ 1((vi ⊗ ¯vi)(p ⊗ 1)ηn), ξn,ti − h(xy ⊗ 1)ξn,t, ξn,ti ≤ 4.5ε +Xi + Xi ≤ 5.5ε +Xi + Xi Thus, by the triangle inequality in combination with (26), the (left) traciality of ηn, and the sixth inequality in (27) if we continue above we obtain that ≤ 5.5ε +Xi + Xi ≤ 6.5ε +Xi + h(x(Xi ≤ 7.5ε +Xi + sup µi (k[v∗i ⊗ ¯v∗i , ηn]k + k[vi ⊗ ¯vi, ηn]k + k(((wi − vi)p)vi) ⊗ ¯vi)ηnk) + µih(x ⊗ ¯v∗i )(e⊥Vt ⊗ 1((wi ⊗ ¯vi)(p ⊗ 1)ηn), ξn,ti − h(xy ⊗ 1)ξn,t, ξn,ti µi(k[v∗i ⊗ ¯v∗i , ηn]k + k[vi ⊗ ¯vi, ηn]k+ n kVt ⊗ 1((wi ⊗ ¯vi)(p ⊗ 1)ηn) − (wi ⊗ ¯vi)(Vt ⊗ 1((p ⊗ 1)ηn))k)+ µiwi) ⊗ 1)ξn,t, ξn,ti − h(xy ⊗ 1)ξn,t, ξn,ti µi (k[v∗i ⊗ ¯v∗i , ηn]k + k[vi ⊗ ¯vi, ηn]k) + h(x(Xi µiwi − y) ⊗ 1)ξn,t, ξn,ti. Inequality (25) together with (C), the Cauchy-Schwarz inequality, kxk∞ ≤ 1, and the seventh inequality in (27) show that the last quantity above is smaller than ≤ 7.5ε +Xi ≤ 7.5ε +Xi ≤ 8.5ε +Xi = 8.5ε +Xi ≤ 10ε +Xi µi (k[v∗i ⊗ ¯v∗i , ηn]k + k[vi ⊗ ¯vi, ηn]k) + k((Xi µi (k[v∗i ⊗ ¯v∗i , ηn]k + k[vi ⊗ ¯vi, ηn]k) + k(Xi µi (k[v∗i ⊗ ¯v∗i , ηn]k + k[vi ⊗ ¯vi, ηn]k) + k((Xi µi (k[v∗i ⊗ ¯v∗i , ηn]k + k[vi ⊗ ¯vi, ηn]k) + k(Xi µi (k[v∗i ⊗ ¯v∗i , ηn]k + k[vi ⊗ ¯vi, ηn]k) . µiwi − y) ⊗ 1)ξn,tk µiwi − y) ⊗ 1)(Vt ⊗ 1((p ⊗ 1)ηn))k µiwi − y)p) ⊗ 1)ηnk µiwi − y)pk2 In conclusion, (24) follows from the previous inequalities if we let LK,ε be the collection of all elements vi, v∗i for all the i's and all y ∈ K where each vi (resp. v∗i ) is repeated [µi] + 1 times. Lemma 6.4 (Lemma 4.5 in [9]). For every ε > 0 and any finite set F0 ⊂ U(N ) there exist a finite set F0 ⊂ F ⊂ M , a c.c.p. map ϕF,ε : span(F ) → L∞(X) ⋊σ,r Γ, (cid:3) 24 I. CHIFAN, T. SINCLAIR, AND B. UDREA and tε > 0 such that (29) for all u ∈ F0 and kxk∞ ≤ 1. ψtε(ϕF,ε(up)∗xϕF,ε(up)) − ψtε(x) ≤ 116ε, Next we briefly explain how to use the previous lemmas in order to get the proof of the theorem. First notice that, as an M -bimodule, H is weakly contained in the coarse bimodule. Then following the same argument as in Theorem B of [46] we define a state ψt on N = B(H) ∩ ρ(M op)′. Explicitly, if we denote by ξn,t = e ⊗ 1(ηn,t) we let ψt(x) = Limn To get the proof, from here on, one can proceed exactly as explained in Theorem 4.1 in [9]. Namely we use the same Lemmas 4.3 and 4.4 from [9] and the final part in the proof of Theorem B in [46] to conclude that N p is amenable. We leave the details to the reader. kξn,tk2 h(x ⊗ 1)ξn,t, ξn,ti for every x ∈ N . 1 Finally, we note that very recently Popa and Vaes extended the third part to arbitrary families of subgroups; even more impressively, when Γ is weakly amenable, they showed the result holds for arbitrary subalgebras P , without the weak compact embedding assumption, [58]. (cid:3) 1} in Γ1 and G2 = {Σk Theorem 6.5. Let Γ1, Γ2 be exact groups each having a family of quasi-normal subgroups G1 = {Σj 2} in Γ2. and for each i let πi : Γi → U(Hπi) be a weakly-ℓ2 representation such that RA(Γi,Gi,Hπi ) 6= ∅. Let Γ1 × Γ2 y X be a measure-preserving action on a probability space and denote by M = L∞(X) ⋊ (Γ1 × Γ2). If P ⊂ M is a weakly compact embedding (cf. [45]), then one can find projections p0, pk 2 = 1 such that the following hold: 2 ∈ Z(NM (P )′ ∩ M ) with p0 +W pk 1 +W pj 1, pj (1) NM (P )′′p0 is amenable; 1 (cid:22)M L∞(X) ⋊ (Γ1 × Σk (2) P pk 2 (cid:22)M L∞(X) ⋊ (Σj (3) P pj 2), for all k; 1 × Γ2), for all j. 2), M j 1 = L∞(X) ⋊ (Γ1 × Σk 1, pj 1 = upk 1, we get that pk Proof. We establish the following notations: M k 2 = L∞(X)⋊(Σj 1×Γ2), N = NM (P )′′ and Z = Z(N′∩M ). Let p0 ∈ Z be the maximal projection such that N p0 is amenable. For each k, let pk 1 ∈ (P ′ ∩ M )(1 − p0) be a maximal projection satisfying the condition in (2)(obtained via a standard maximality argument). Similarly, for each j, let pj 1) be a maximal projection satisfying the condition in (3). By maximality, we must have that pk 2 ∈ Z. Indeed, if u ∈ NM (P ), let pk 1 also satisfies (2) and by the maximality 1u∗, for every u ∈ U(P ), hence pk of pk 2 ∈ Z(P ′∩M ). Moreover, we have that pk 1u∗(1 − p0 − pk 2 ∈ (P ′ ∩ M )(1 − p0 −W pk Therefore, to prove the theorem, we only need to show that p0 +W pk 2 = 1. By contradiction, assume that p := 1 − (p0 +W pk 2) 6= 0. Note that P p (cid:14)M M k 2 , for any j and k, by maximality. Also, note that N p has no amenable direct summand, for the same reason. Now use the remarks in Examples 2.7. B for Γ = Γ1 × Γ2. It follows that RA(Γ,G,H1 ¯⊗H2) 6= ∅, where G = {Σ1 × Γ2 : Σ1 ∈ G1}∪{Γ1 × Σ2 : Σ2 ∈ G2}. So we can apply the third part of Theorem 6.1 for the amenable subalgebra P p ⊂ L∞(X) ⋊ Γ. Thus either NM (P p)′′ is amenable, which is impossible (because it contains N p which is non-amenable), 1 +W pj 1 +W pj 1 and P p (cid:14)M M j 1 ∈ Z. 1). Then pk 1 + pk 1 = upk 1 = 0. Thus pk 1, pj II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 25 or there is a Λ ∈ G such that P p (cid:22)M L∞(X) ⋊ Λ. Suppose Λ = Σj j. But this means P p (cid:22)M M j ends the proof. 1 × Γ2, for some 2 for some j, which is again a contradiction, and this (cid:3) Proof of Corollary 0.4. Applying the previous theorem for A = C1 and Σi = e there exist p0, p1, p2 ∈ Z with p0 + p1 + p2 = 1 such that p0NM (P )′′ is amenable, p1B (cid:22)M LΓ1, and p2B (cid:22)M LΓ2. Therefore, the conclusion follows if we show that p0 = 1. Assuming this is not the case one can find p1 6= 0 such that p1B (cid:22)M LΓ1. Then Remark 3.8 in [67] implies that LΓ′1 ∩ M (cid:22)M p1B′ ∩ M and since LΓ′1 ∩ M = LΓ2 then we have LΓ2 (cid:22)M p1B′∩M . This is, however, a contradiction because LΓ2 is a non-amenable factor while p1B′∩ M is assumed to be an amenable algebra. (cid:3) Proof of Corollary 0.5. Assume that Λ y Y is a free, ergodic action which is W ∗- equivalent to Γ1 × Γ2 y X. This amounts to the existence of an ⋆-isomorphism ψ : L∞(Y )⋊ Λ → L∞(X)⋊ (Γ1× Γ2). We will denote by A = L∞(X), B = L∞(Y ), M = A ⋊ (Γ1 × Γ2), M1 = A ⋊ Γ1, and M2 = A ⋊ Γ2. Below we will prove that there exists a unitary x ∈ U(M ) such that xψ(B)x∗ = A. Notice that since C = ψ(B) Cartan in M its normalizing algebra is non- amenable so by Theorem 6.5 we can assume that C (cid:22)M M1. Therefore one can find nonzero projections p ∈ C, q ∈ M1, a partial isometry v ∈ M , and a ⋆- homomorphism φ : Cp → qM1q such that for all x ∈ Cp we have (30) φ(x)v = vx. Since C is a maximal abelian subalgebra of M then by Lemma 1.5 in [27] we can assume that φ(Cp) ⊂ qM1q is also a maximal abelian subalgebra. Fixing u ∈ NpMp(Cp) we can easily see that for all x ∈ Cp we have (31) vuv∗φ(x) = vuxv∗ = vuxu∗uv∗ = φ(uxu∗)vuv∗. Notice that vuv∗vu∗v∗ = φ(uv∗vu∗)vv∗ is a projection and hence vuv∗ is a partial isometry. Also, applying the conditional expectation EqM1q to equation (31), we obtain that for all x ∈ Cp we have EqM1q(vuv∗)φ(x) = φ(uxu∗)EqM1q(vuv∗). Taking the polar decomposition EqM1q(vuv∗) = wuEqM1q(vuv∗), the previous equation entails that EqM1q(vuv∗) ∈ φ(Cp)′ ∩ qM1q = φ(Cp) and for all x ∈ Cp we have wuφ(x) = φ(uxu∗)wu. This implies in particular that wuw∗u, w∗uwu ∈ φ(Cp)′∩qM1q = φ(Cp) and therefore wu ∈ GN qM1q(φ(Cp)), the normalizing groupoid of φ(Cp) in qM1q. Altogether, we have shown that EqM1q(vuv∗) ⊆ GN qM1q(φ(Cp))′′. By [14], we have that GN qM1q(φ(Cp))′′ = NqM1q(φ(Cp))′′ and since the above containment holds for every u ∈ NpMp(Cp)′′ and NpMp(Cp)′′ = pM p we have that EqM1q(vM v∗) ⊆ NqM1q(φ(Cp))′′, and hence vv∗M1vv∗ ⊆ NqM1q(φ(Cp))′′. This shows in particular that NqM1q(φ(Cp))′′ is non-amenable; therefore, by Theorem B in [9] we have that φ(Cp) (cid:22)M1 A. By Remark 3.8 in [67] this further implies that C (cid:22)M A. Finally, by Theorem 1.4, one can find a unitary x ∈ U(M ) such that xφ(B)x∗ = xCx∗ = A. 26 I. CHIFAN, T. SINCLAIR, AND B. UDREA In particular, our claim shows that the actions Γ1×Γ2 y X and Λ y Y are orbit equivalent. Note that, since Γ1 and Γ2 have property (T) then so is the product Γ1 × Γ2, so it follows from Ioana's Cocycle Superrigidity Theorem [25] that the actions Γ1 × Γ2 y X and Λ y Y are virtually conjugate. (cid:3) Corollary 6.6. Let Γi be weakly amenable groups and let π : Γi → U(Hπ) be weakly-ℓ2 representations such that RA(Γ,{e},Hπ) 6= ∅ (e.g. Γi are hyperbolic). If Γ1 × Γ2 y X and Λ y Y are any p.m.p. actions such that Λ admits an infinite amenable normal subgroup Σ < Γ for which the restriction Σ y Y is still ergodic then Γ1 × Γ2 y X ≇OE Λ y Y . Proof. We will assume that Γ1×Γ2 y X ∼=OE Λ y Y and then show that this leads to a contradiction. Thus there exists a ⋆-isomorphism ψ : L∞(Y ) ⋊ Λ → L∞(X) ⋊ (Γ1 × Γ2). We will also denote by A = L∞(X), B = L∞(Y ), P = ψ(L∞(Y ) ⋊ Σ), M = A ⋊ (Γ1 × Γ2), M1 = A ⋊ Γ1, M2 = A ⋊ Γ2, and notice that ψ(B) = A. Since the Cowling-Haagerup constant is an ME -invariant [11] it follows that ψ(LΣ) is a weakly compact embedding in M . Since Σ is normal in Λ, then applying Theorem 6.5, we can assume that ψ(LΣ) (cid:22)M M1 and since ψ(B) = A we conclude that P (cid:22)M M1. Therefore, one can find nonzero projections p ∈ P , q ∈ M1, a partial isometry v ∈ M , and a ⋆-homomorphism φ : pP p → qM1q such that for all x ∈ pP p we have (32) φ(x)v = vx. Since P is an irreducible subfactor of M , by Proposition 1.2 we can assume that φ(pP p) ⊂ qM1q is also a irreducible subfactor. Fixing u ∈ NpMp(pP p) we can easily see that for all x ∈ pP p we have (33) vuv∗φ(x) = vuxv∗ = vuxu∗uv∗ = φ(uxu∗)vuv∗. Notice that vuv∗vu∗v∗ = φ(uv∗vu∗)vv∗ is a projection and hence vuv∗ is a partial isometry. Also, applying the conditional expectation EqM1q to equation (33), we obtain that for all x ∈ pP p we have EqM1q(vuv∗)φ(x) = φ(uxu∗)EqM1q(vuv∗). Taking the polar decomposition EqM1q(vuv∗) = wuEqM1q(vuv∗), the previous equation entails that EqM1q(vuv∗) ∈ φ(pP p)′ ∩ qM1q = Cq and for all x ∈ pP p we have wuφ(x) = φ(uxu∗)wu. This implies in particular that wuw∗u, w∗uwu ∈ φ(pP p)′ ∩ qM1q = Cq and therefore wu is a scalar multiple of a normalizing unitary in ∈ NqM1q(φ(pP p)). Altogether, we have shown that EqM1q(vuv∗) ⊆ NqM1q(φ(pP p))′′. Since the above containment holds for every u ∈ NpMp(pP p) and NpMp(pP p)′′ = pM p we have that EqM1q(vM v∗) ⊆ NqM1q(φ(pP p))′′, and hence vv∗M1vv∗ ⊆ NqM1q(φ(pP p))′′. This shows in particular that NqM1q(φ(pP p))′′ is non-amenable; therefore, by Theorem B in [9] we have that φ(pP p) (cid:22)M1 A. By Remark 3.8 in [67] this would imply that P (cid:22)M A, which is an obvious contradic- tion. (cid:3) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) n\i=1 i < ∞. (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 27 In the remaining part of the section we explain how the second and third part of Theorem 6.1 can be successfully exploited to produce new examples of von Neumann algebras with either unique Cartan subalgebra or no Cartan subalgebras. With this purpose in mind, we introduce the following definition. Definition 6.7. A subgroup Σ < Γ is called weakly malnormal if there exist finitely many elements γ1, γ2, . . . , γn ∈ Γ such that γiΣγ−1 Therefore, when the second and the third part in the intertwining theorem above is combined with Corollary 7 from [22] we immediately obtain the following uniqueness (absence) of Cartan subalgebra statement. Corollary 6.8. Let Γ be a weakly amenable group and let π : Γ → U(Hπ) be a weakly-ℓ2 representation such that one of the following cases holds: (1) RQ(Γ,G,Hπ) 6= ∅ for a family of weakly malnormal subgroups G of Γ, or (2) RA(Γ,{e},Hπ) 6= ∅. Also let Γ y X be a weakly compact, free action. If Γ is as in the first case (1) above we assume in addition that the restrictions Σ y X are ergodic for all Σ ∈ G. Then L∞(X) ⋊ Γ has unique Cartan subalgebra. If in addition Γ is i.c.c. and G is a family of malnormal groups then LΓ has no Cartan subalgebra. Proof. Let B be a Cartan subalgebra of M = L∞(X) ⋊ Γ. Since Γ is weakly amenable and the action is weakly compact, it follows that M = L∞(X) ⋊ Γ is weakly amenable, so the inclusion B ⊂ M is weakly compact, by [45]. In the first case we apply (2) in Theorem 6.1 above and see that, since NM (B)′′ = M is non- amenable, B (cid:22)M L∞(X) ⋊ Σ, for some Σ ∈ G. Using Corollary 7 from [22] we obtain that in fact B (cid:22)M L∞(X), from which it follows, by using Appendix 1 in [50], that B and L∞(X) are unitarily conjugated. In the second case we apply (3) in the same Theorem 6.1 above to see that again B (cid:22)M L∞(X), and the conclusion follows. (cid:3) Employing the same strategy as in the proof of Corollary B.2 from [9] and using the fact that the class of weakly amenable groups is closed under taking ME- subgroups, we obtain new structural results for measure equivalence of groups. Corollary 6.9. Let Γ be a weakly amenable group and let π : Γ → U(Hπ) be a weakly-ℓ2 representation such that one of the following holds: either RQ(Γ,G,Hπ) 6= ∅ for a family of amenable, malnormal subgroups G, or RQ(Γ,{e},Hπ) 6= ∅. If Λ is any M E-subgroup of Γ then LΛ is strongly solid i.e., given any diffuse amenable subalgebra A ⊆ LΛ its normalizing algebra NLΛ(A)′′ is still amenable. In particu- lar, every amenable subgroup of Λ has amenable normalizer. The following are examples of groups that satisfy the conditions required in the above corollary: any weakly amenable group that is in the class S of Ozawa [40]–in particular any weakly amenable group Γ that is hyperbolic relative to a family of amenable subgroups (e.g. Sela's limit groups which are weakly amenable and hyper- bolic with respect to their noncyclic maximal abelian subgroups [12]); any weakly amenable HNN extension Γ⋆α of a group Γ, where α : Σ1 → Σ2 is a monomorphism with Σi ∈ G; any infinite free product ⋆n∈NΓn where Γn is hyperbolic relative to a 28 I. CHIFAN, T. SINCLAIR, AND B. UDREA finite family Gn of malnormal groups, Λcb(Γn) = 1 and Gn = {e} for all but finitely many n's – in this case we choose G = ∪nGn. Acknowledgements The authors would like to thank Adrian Ioana, Narutaka Ozawa, and Jesse Peterson for useful discussions. They are also grateful to the anonymous referees for many useful suggestions. References 1. Scot Adams: Indecomposability of equivalence relations generated by word hyperbolic groups, Topology 33 (1994), no. 4, 785–798. 2. Bachir Bekka, Pierre de la Harpe, and Alain Valette: Kazhdan's property (T), New Mathe- matical Monographs, vol. 11, CUP, Cambridge, 2008. xiv+472 pp. 3. B. H. Bowditch: Relatively hyperbolic groups. Internat. J. Algebra and Computation 22 (2012), 1250016, 66pp. 4. Marek Bozejko and Massimo Picardello: Weakly amenable groups and amalgamated free prod- ucts, Proceedings of the American Mathematical Society 117 (1993), no. 4, 1039–1046. 5. Nathanial P. Brown and Narutaka Ozawa: C∗-algebras and finite-dimensional approxima- tions, Graduate Studies in Mathematics, vol. 88, AMS, Providence, RI. 6. Marc Burger and Nicolas Monod: Continuous bounded cohomology and applications to rigidity theory, Geometric and Functional Analysis 12 (2002), 219–280. 7. Ionut Chifan and Adrian Ioana: Strong solidity in tensor products of free group factors, manuscript (2009), unpublished. 8. Ionut Chifan and Jesse Peterson: Some unique group-measure space decomposition results, Duke Math. J., to appear. 9. Ionut Chifan and Thomas Sinclair: On the structural theory of II1 factors of negatively curved groups, Annales Scientifiques de l'ENS 46 (2013), 1–33. 10. Michael Cowling and Uffe Haagerup: Completely bounded multipliers of the Fourier algebra of a simple Lie group of real rank one, Inventiones Mathematicae 96 (1989), no. 3, 507–549. 11. Michael Cowling and Robert J. Zimmer: Actions of lattices in Sp(1, n), Ergodic Theory Dynam. Systems 9 (1989), no. 2, 221–237. 12. Francois Dahmani: Combination of convergence groups, Geometry and Topology 7 (2003), 933–963. 13. Brian Dorofaeff: The Fourier algebra of SL(2, R) ⋊ Rn, n ≥ 2, has no multiplier bounded approximate unit, Mathematische Annalen 297 (1993), no. 4, 707–724. 14. Henry A. Dye: On groups of measure preserving transformations. II. American Journal of Mathematics, 85 (1963), 551–576. 15. Benson Farb: Relatively hyperbolic groups, Geometric and Functional Analysis 8 (1998), 810– 840. 16. Alex Furman: Orbit equivalence rigidity, Annals of Mathematics (2) 150(1999), no. 3,1083– 1108. 17. Daniel Groves: Limit groups for relatively hyperbolic groups, I: The basic tools, Algebraic and Geometric Topology, 9 (2009), 1423–1466. 18. Erik Guentner and Nigel Higson: Weak amenability of CAT(0)-cubical groups, Geometriae Dedicata, 143 (2009), no. 1, 137–156. 19. Erik Guentner, Nigel Higson and Shmuel Weinberger: The Novikov Conjecture for Linear Groups, Publications Math´ematiques de l'IH´ES 101 (2005) 243–268. 20. Cyril Houdayer: Strongly solid group factors which are not interpolated free group factors, Mathematische Annalen 346 (2010), 969–989. 21. Cyril Houdayer and Dimitri Shlyakhtenko: Strongly solid II1 factors with an exotic MASA, International Mathematics Research Notices 6 (2011), 1352–1380. 22. Cyril Houdayer, Sorin Popa, and Stefaan Vaes: A class of groups for which every action is W ∗-superrigid, to appear in Groups, Geometry, and Dynamics. 23. Uffe Haagerup: Injectivity and decomposition of completely bounded maps, Operator algebras and their connections with topology and ergodic theory (Bu¸steni, 1983), Lecture Notes in Mathematics, vol. 1132, Springer, Berlin, 1985, pp. 170–222. II1 FACTORS OF NEGATIVELY CURVED GROUPS, II 29 24. Uffe Haagerup: Group C∗-algebras without the completely bounded approximation property, Preprint (1986). 25. Adrian Ioana: Cocycle superrigidity for profinite actions of property (T) groups, Duke Math- ematical Journal, 157 (2011), no. 2, 337–367. 26. Adrian Ioana: W-superrigidity for Bernoulli actions of property (T) groups, J. Amer. Math. Soc. 24 (2011), 1175–1226. 27. Adrian Ioana: Uniqueness of the group measure space decomposition for Popa's HT factors, Geom. Funct. Anal. 22 (2012), 699–732. 28. Adrian Ioana, Sorin Popa, and Stefaan Vaes: A class of superrigid group von Neumann algebras, Ann. Math. 178 (2013), 231–286. 29. Tadeusz Januszkiewicz: For Coxeter groups zg is a coefficient of a uniformly bounded rep- resentation, Fundamenta Mathematicae 174 (2002), no. 1, 79–86. 30. Igor Mineyev: Straightening and bounded cohomology of hyperbolic groups, Geometric and Functional Analysis 11 (2001), no. 4, 807–839. 31. Igor Mineyev, Nicolas Monod, and Yehuda Shalom: Ideal bicombings for hyperbolic groups and applications, Topology 43 (2004), no. 6, 1319–1344. 32. Igor Mineyev and Asli Yaman: Relative hyperbolicity and bounded cohomology, Preprint, 2009. 33. Naokazu Mizuta: A Bozejko–Picardello type inequality for finite-dimensional CAT(0) cube complexes, Journal of Functional Analysis 254 (2008), no. 3, 760–772. 34. Nicolas Monod: Continuous bounded cohomology of locally compact groups, Lecture Notes in Mathematics, vol. 1758, Springer, Berlin, 2001. 35. Nicolas Monod and Yehuda Shalom: Cocycle superrigidity and bounded cohomology for neg- atively curved spaces, Journal of Differential Geometry 67 (2004), no. 3, 395–455. 36. Nicolas Monod and Yehuda Shalom: Orbit equivalence rigidity and bounded cohomology, Annals of Mathematics (2) 164 (2006), no. 3, 825–878. 37. B. H. Neumann: Groups covered by permutable sets. Journal of London Mathematical Society (2) 29 (1954), no. 2, 236–248. 38. Narutaka Ozawa: Amenable actions and exactness for discrete groups, C. R. Acad. Sci. Paris S´er. I Math. 330 (2000), no. 8, 691–695. 39. Narutaka Ozawa: Solid von Neumann algebras, Acta Mathematica, 192 (2004), 111–117. 40. Narutaka Ozawa: A Kurosh type theorem for type II1 factors, International Mathematics Research Notices (2006) Art. ID 97560. 41. Narutaka Ozawa: Weak amenability of hyperbolic groups, Groups, Geometry, and Dynamics 2 (2008), no. 2, 271–280. 42. Narutaka Ozawa: An example of a solid von Neumann algebra, Hokkaido Mathematical Jour- nal 38 (2009), 567–571. 43. Narutaka Ozawa: Examples of groups which are not weakly amenable, Kyoto J. Math. 52 (2012), 333–344. 44. Narutaka Ozawa and Sorin Popa: Some prime factorization results for type II1 factors , Inventiones Mathematicae, 156 (2004), 223–234. 45. Narutaka Ozawa and Sorin Popa: On a class of II1 factors with at most one Cartan subalgebra, Annals of Mathematics (2) 172 (2010), 713–749. 46. Narutaka Ozawa and Sorin Popa: On a class of II1 factors with at most one Cartan subalgebra II, American Journal of Mathematics, 132 (2010), 841–866. 47. Jesse Peterson: Examples of group actions which are virtually W-superrigid, Preprint, 2009. 48. Jesse Peterson: L2-rigidity in von Neumann algebras, Inventiones Mathematicae, 175 (2009), 417–433. 49. Jesse Peterson and Thomas Sinclair: On cocycle superrigidity for Gaussian actions, Ergodic Theory and Dynamical Systems 32 (2012), 249–272. 50. Sorin Popa: On a class of type II1 factors with Betti numbers invariants, Annals of Mathe- matics (2) 163 (2006), 809–899. 51. Sorin Popa: Strong rigidity of II1 factors arising from malleable actions of w-rigid groups I, Inventiones Mathematicae 165 (2006), no. 2, 369–408. 52. Sorin Popa: Deformation and rigidity for group actions and von Neumann algebras, Interna- tional Congress of Mathematicians, Vol. I, 445–477, Eur. Math. Soc., Zurich, 2007. 53. Sorin Popa: Cocycle and orbit equivalence superrigidity for malleable actions of w-rigid groups, Inventiones Mathematicae 170 (2007), 243-295. 30 I. CHIFAN, T. SINCLAIR, AND B. UDREA 54. Sorin Popa: On Ozawa's property for free group factors, International Mathematics Research Notices (2007), no. 11, 10pp. 55. Sorin Popa: On the superrigidity of malleable actions with spectral gap, Journal of the Amer- ican Mathematical Society 21 (2008), 981–1000. 56. Sorin Popa and Stefaan Vaes: Strong rigidity of generalized Bernoulli actions and computa- tions of their symmetry groups, Advances in Mathematics 217 (2008), 833-872. 57. Sorin Popa and Stefaan Vaes: Group measure space decomposition of factors and W*- superrigidity, Inventiones Mathematicae 182 (2010), 371-417. 58. Sorin Popa and Stefaan Vaes: Unique Cartan decomposition for II1 factors arising from arbitrary actions of hyperbolic groups, J. Reine Angew. Math., to appear. 59. John Roe: Lectures on coarse geometry, University Lecture Series, vol. 31, AMS, Providence, RI, 2003. vii+175 pp. 60. ´Eric Ricard and Quanhua Xu: Khintchine type inequalities for reduced free products and applications, Journal fur die reine und angewandte Mathematik 599 (2006), 27–59. 61. Hiroki Sako: Measure equivalence rigidity and bi-exactness of groups, Journal of Functional Analysis 257 (2009), no. 10, 3167–3202. 62. Yehuda Shalom: Rigidity, unitary representations of semisimple groups, and fundamental groups of manifolds with rank one transformation groups, Annals of Mathematics (2) 152 (2000) no. 1, 113–182. 63. Thomas Sinclair: Strong solidity of group factors from lattices in SO(n,1) and SU(n,1), Journal of Functional Analysis 260 (2011), no. 11, 3209–3221. 64. Isadore M. Singer: Automorphisms of finite factors, American Journal of Mathematics 77 (1955), 117-133. 65. Andreas Thom: Low degree bounded cohomology invariants and negatively curved groups, Groups, Geometry, and Dynamics 3 (2009), no. 2, 343–358. 66. Stefaan Vaes: One-cohomology and the uniqueness of the group measure space decomposition of a II1 factor, Math. Ann. 355 (2013), 661–696. 67. Stefaan Vaes: Explicit computations of all finite index bimodules for a family of II1 factors, Annales Scientifiques de l'´Ecole Normale Sup´erieure. (4) 41 (2008), no. 5, 743–788. 68. Dan-Virgil Voiculescu: The analogues of entropy and of Fisher's information measure in free probability theory: the absence of Cartan subalgebras, Geometric and Functional Analysis 6 (1996), no.1, 172–199. Ionut Chifan, University of Iowa,14 McLean Hall, Iowa City, IA, USA 52242 and IMAR, Bucharest, Romania E-mail address: [email protected] Thomas Sinclair, Department of Mathematics, UCLA, Los Angeles, CA, USA 90095– 1555 E-mail address: [email protected] Bogdan Udrea, University of Iowa, 14 McLean Hall, Iowa City, IA, USA 52242 and IMAR, Bucharest, Romania E-mail address: [email protected]
1211.0484	1	1211	2012-11-02T16:23:07	On examples of intermediate subfactors from conformal field theory	[ "math.OA", "math.QA" ]	Motivated by our subfactor generalization of Wall's conjecture, in this paper we determine all intermediate subfactors for conformal subnets corresponding to four infinite series of conformal inclusions, and as a consequence we verify that these series of subfactors verify our conjecture. Our results can be stated in the framework of Vertex Operator Algebras. We also verify our conjecture for Jones-Wassermann subfactors from representations of Loop groups extending our earlier results.	math.OA	math	On examples of intermediate subfactors from conformal field theory Feng Xu∗ Department of Mathematics University of California at Riverside Riverside, CA 92521 E-mail: [email protected] 2 1 0 2 v o N 2 ] . A O h t a m [ 1 v 4 8 4 0 . 1 1 2 1 : v i X r a Abstract Motivated by our subfactor generalization of Wall's conjecture, in this paper we determine all intermediate subfactors for conformal subnets corresponding to four infinite series of conformal inclusions, and as a consequence we verify that these series of subfactors verify our conjecture. Our results can be stated in the framework of Vertex Operator Algebras. We also verify our conjecture for Jones-Wassermann subfactors from representations of Loop groups extending our earlier results. ∗Supported in part by NSF grant and an academic senate grant from UCR. 1 1 Introduction Let M be a factor represented on a Hilbert space and N a subfactor of M which is irreducible, i.e.,N ′ ∩ M = C. Let K be an intermediate von Neumann subalgebra for the inclusion N ⊂ M. Note that K ′ ∩ K ⊂ N ′ ∩ M = C, K is automatically a factor. Hence the set of all intermediate subfactors for N ⊂ M forms a lattice under two natural operations ∧ and ∨ defined by: K1 ∧ K2 = K1 ∩ K2, K1 ∨ K2 = (K1 ∪ K2)′′. The commutant map K → K ′ maps an intermediate subfactor N ⊂ K ⊂ M to M ′ ⊂ K ′ ⊂ N ′. This map exchanges the two natural operations defined above. Let M ⊂ M1 be the Jones basic construction of N ⊂ M. Then M ⊂ M1 is canonically isomorphic to M ′ ⊂ N ′, and the lattice of intermediate subfactors for N ⊂ M is related to the lattice of intermediate subfactors for M ⊂ M1 by the commutant map defined as above. Let G1 be a group and G2 be a subgroup of G1. An interval sublattice [G1/G2] is the lattice formed by all intermediate subgroups K, G2 ⊆ K ⊆ G1. By cross product construction and Galois correspondence, every interval sublattice of finite groups can be realized as intermediate subfactor lattice of finite index. Hence the study of intermediate subfactor lattice of finite index is a natural generalization of the study of interval sublattice of finite groups. The study of intermediate subfactors has been very active in recent years(cf. [4],[12], [27], [25], [22], and [39] for a partial list). In 1961 G. E. Wall conjectured that the number of maximal subgroups of a finite group G is less than G, the order of G (cf. [31]). In the same paper he proved his conjecture when G is solvable. See [28] for more recent result on Wall's conjecture. Wall's conjecture can be naturally generalized to a conjecture about maximal elements in the lattice of intermediate subfactors. What we mean by maximal elements are those subfactors K 6= M, N with the property that if K1 is an intermediate subfactor and K ⊂ K1, then K1 = M or K. Minimal elements are defined similarly where N is not considered as an minimal element. When M is the cross product of N by a finite group G, the maximal elements correspond to maximal subgroups of G, and the order of G is the dimension of second higher relative commutant. Hence a natural generalization of Wall's conjecture as proposed in [37] is the following: Conjecture 1.1. Let N ⊂ M be an irreducible subfactor with finite index. Then the number of maximal intermediate subfactors is less than dimension of N ′ ∩ M1 (the dimension of second higher relative commutant of N ⊂ M). We note that since maximal intermediate subfactors in N ⊂ M correspond to min- imal intermediate subfactors in M ⊂ M1, and the dimension of second higher relative commutant remains the same, the conjecture is equivalent to a similar conjecture as above with maximal replaced by minimal. In [37],[14], Conjecture 1.1 is verified for subfactors coming from certain confor- mal field theories and subfactors which are more closely related to groups and more 2 generally Hopf algebras. In this paper we investigate Conjecture 1.1 for conformal subnets A ⊂ B (cf. Definition 2.1) with finite index. Then Conjecture 1.1 in this case states: Conjecture 1.2. Suppose that conformal subnets A ⊂ B (cf. Definition 2.1) has finite index. Then the number of minimal (resp. maximal) subnets between A and B is less than the dimension of the space of bounded maps from the vacuum representation of B to itself which commutes with the action of A. In the above conjecture we have included both maximal and minimal cases since the dual of conformal subnet A ⊂ B is not conformal subnet. It is also straightforward to phrase the above conjecture in terms of Vertex Operator Algebras (VOAs) and its sub-VOAs. Note that any finite group G is embedded in a finite symmetric group Sn, and using the theory of permutation orbifolds as in [38] we can always find a completely rational net B such that G acts properly on B and with fixed point subnet A. In this case the intermediate subnets between A and B are in one to one correspondence with subgroups of G. So in this orbifold case the minimal version of Conjecture 1.2 is equivalent to Wall's conjecture. Hence Conjecture 1.2 is highly nontrivial even if we assume that B is completely rational. Though the orbifold case of Conjecture 1.2 in general is out of reach at present, there are very interesting other examples of subnets coming from conformal field theory (CFT). A large class of such examples come from conformal inclusions (cf. §2.5), and they provide a large class of subfactors which are not related to groups. In view of Conjecture 1.2 it is a natural question to investigate intermediate subnets of such examples, and this is the main goal of our paper. Our results Th. 3.8, Th. 3.11 give a complete list of intermediate conformal subnets in subnets coming from four infinite series of maximal conformal inclusions, and as consequence, we are able to verify Conjecture 1.2 in these examples. Our results show that the intermediate subnets in these examples are very rare. The key idea behind the proof of Th. 3.8 is the property of induced adjoint representation: Prop. 3.6 shows that such induced representation is always irreducible when the intermediate subnet does not have additional weight 1 element. By locality consideration in Lemma 2.8 this forces the intermediate subnet to be simply simple current extensions when it has no additional weight 1 element. In the case when the intermediate net has additional weight 1 element, we use smeared vertex operators as in [34] and maximality of conformal inclusions to show that the intermediate subnet is in fact the largest net. The proof makes use of the analogue of statement in VOA theory that weight 1 element of a VOA forms a Lie algebra. The proof of Th. 3.11 is much simpler and make use of normal inclusions as in §4.2 of [35]. By using properties of smeared vertex operators in §3.1, we can translate our results in Th. 3.8, Th. 3.11 into statements about intermediate VOAs (cf. Th. 3.14). We think it is an interesting question to find a VOA proof of Th. 3.14. In §4 we extend our earlier results in §5 of [39] on Jones-Wassermann subfactors and we verify that these subfactors verify Conjecture 1.1. 3 In addition to what are already described as above, we have included a preliminary section §2 where we introduce the basic notion of conformal nets, subnets , conformal inclusions, and induction to describe the background of our results in §3 and §4. 2 Preliminaries 2.1 Preliminaries on sectors Given an infinite factor M, the sectors of M are given by Sect(M) = End(M)/Inn(M), namely Sect(M) is the quotient of the semigroup of the endomorphisms of M modulo the equivalence relation: ρ, ρ′ ∈ End(M), ρ ∼ ρ′ iff there is a unitary u ∈ M such that ρ′(x) = uρ(x)u∗ for all x ∈ M. Sect(M) is a ∗-semiring (there are an addition, a product and an involution ρ → ¯ρ) equivalent to the Connes correspondences (bimodules) on M up to unitary equiva- lence. If ρ is an element of End(M) we shall denote by [ρ] its class in Sect(M). We define Hom(ρ, ρ′) between the objects ρ, ρ′ ∈ End(M) by Hom(ρ, ρ′) ≡ {a ∈ M : aρ(x) = ρ′(x)a ∀x ∈ M}. We use hλ, µi to denote the dimension of Hom(λ, µ); it can be ∞, but it is finite if λ, µ have finite index. See [26] for the definition of index for type II1 case which initiated the subject and [29] for the definition of index in general. Also see §2.3 [18] for expositions. hλ, µi depends only on [λ] and [µ]. Moreover we have if ν has finite index, then hνλ, µi = hλ, ¯νµi, hλν, µi = hλ, µ¯νi which follows from Frobenius duality. µ is a subsector of λ if there is an isometry v ∈ M such that µ(x) = v∗λ(x)v,∀x ∈ M. We will also use the following notation: if µ is a subsector of λ, we will write as µ ≺ λ or λ ≻ µ. A sector is said to be irreducible if it has only one subsector. 2.2 Local nets By an interval of the circle we mean an open connected non-empty subset I of S1 such that the interior of its complement I ′ is not empty. We denote by I the family of all intervals of S1. A net A of von Neumann algebras on S1 is a map I ∈ I → A(I) ⊂ B(H) from I to von Neumann algebras on a fixed separable Hilbert space H that satisfies: A. Isotony. If I1 ⊂ I2 belong to I, then A(I1) ⊂ A(I2). 4 If E ⊂ S1 is any region, we shall put A(E) ≡ WE⊃I∈I A(I) with A(E) = C if E has empty interior (the symbol ∨ denotes the von Neumann algebra generated). The net A is called local if it satisfies: B. Locality. If I1, I2 ∈ I and I1 ∩ I2 = ∅ then where brackets denote the commutator. [A(I1),A(I2)] = {0}, The net A is called Mobius covariant if in addition satisfies the following properties C,D,E,F: C. Mobius covariance. There exists a non-trivial strongly continuous unitary rep- resentation U of the Mobius group Mob (isomorphic to P SU(1, 1)) on H such that U(g)A(I)U(g)∗ = A(gI), g ∈ Mob, I ∈ I. D. Positivity of the energy. The generator of the one-parameter rotation subgroup of U (conformal Hamiltonian), denoted by L0 in the following, is positive. E. Existence of the vacuum. There exists a unit U-invariant vector Ω ∈ H (vacuum vector), and Ω is cyclic for the von Neumann algebra WI∈I A(I). By the Reeh-Schlieder theorem Ω is cyclic and separating for every fixed A(I). The modular objects associated with (A(I), Ω) have a geometric meaning ∆it I = U(ΛI(2πt)), JI = U(rI) . Here ΛI is a canonical one-parameter subgroup of Mob and U(rI) is a antiunitary acting geometrically on A as a reflection rI on S1. This implies Haag duality: A(I)′ = A(I ′), I ∈ I , where I ′ is the interior of S1 \ I. F. Irreducibility. WI∈I A(I) = B(H). Indeed A is irreducible iff Ω is the unique U-invariant vector (up to scalar multiples). Also A is irreducible iff the local von Neumann algebras A(I) are factors. In this case they are either C or III1-factors with separable predual in Connes classification of type III factors. By a conformal net (or diffeomorphism covariant net) A we shall mean a Mobius covariant net such that the following holds: G. Conformal covariance. There exists a projective unitary representation U of Diff(S1) on H extending the unitary representation of Mob such that for all I ∈ I we have U(φ)A(I)U(φ)∗ = A(φ.I), φ ∈ Diff(S1), U(φ)xU(φ)∗ = x, x ∈ A(I), φ ∈ Diff(I ′), 5 where Diff(S1) denotes the group of smooth, positively oriented diffeomorphism of S1 and Diff(I) the subgroup of diffeomorphisms g such that φ(z) = z for all z ∈ I ′. associates to each I a normal representation of A(I) on B(H) such that A (DHR) representation π of A on a Hilbert space H is a map I ∈ I 7→ πI that π I ↾A(I) = πI , I ⊂ I, I, I ⊂ I . π is said to be Mobius (resp. diffeomorphism) covariant if there is a projective unitary representation Uπ of Mob (resp. Diff(S1)) on H such that πgI(U(g)xU(g)∗) = Uπ(g)πI(x)Uπ(g)∗ for all I ∈ I, x ∈ A(I) and g ∈ Mob (resp. g ∈ Diff(S1)). of itself and we will call this representation the vacuum representation. By definition the irreducible conformal net is in fact an irreducible representation Let G be a simply connected compact Lie group. By Th. 3.2 of [7], the vacuum positive energy representation of the loop group LG (cf. [30]) at level k gives rise to an irreducible conformal net denoted by AGk. By Th. 3.3 of [7], every irreducible positive energy representation of the loop group LG at level k gives rise to an irreducible covariant representation of AGk. Given an interval I and a representation π of A, there is an endomorphism of A localized in I equivalent to π; namely ρ is a representation of A on the vacuum Hilbert space H, unitarily equivalent to π, such that ρI ′ = id ↾ A(I ′). We now define the statistics. Given the endomorphism ρ of A localized in I ∈ I, choose an equivalent endomorphism ρ0 localized in an interval I0 ∈ I with ¯I0 ∩ ¯I = ∅ and let u be a local intertwiner in Hom(ρ, ρ0) , namely u ∈ Hom(ρ I, ρ0, I) with I0 following clockwise I inside I which is an interval containing both I and I0. The statistics operator ǫ(ρ, ρ) := u∗ρ(u) = u∗ρ I(u) belongs to Hom(ρ2 I). We will call ǫ(ρ, ρ) the positive or right braiding and ǫ(ρ, ρ) := ǫ(ρ, ρ)∗ the negative or left braiding. I, ρ2 Let B be a conformal net. By a conformal subnet (cf. [22]) we shall mean a map I ∈ I → A(I) ⊂ B(I) that associates to each interval I ∈ I a von Neumann subalgebra A(I) of B(I), which is isotonic and conformal covariant with respect to the representation U, namely A(I1) ⊂ A(I2), I1 ⊂ I2, U(g)A(I)U(g)∗ = A(g.I) for all g ∈ Diff(S1) and I ∈ I. Note that by Lemma 13 of [22] for each I ∈ I there exists a conditional expectation EI : B(I) → A(I) such that EI preserves the vector state given by the vacuum of A. 6 Definition 2.1. Let A be a conformal net. A conformal net B on a Hilbert space H is an extension of A or A is a subnet of B if there is a DHR representation π of A on H such that π(A) ⊂ B is a conformal subnet. The extension is irreducible if π(A(I))′ ∩ B(I) = C for some (and hence all) interval I, and is of finite index if π(A(I)) ⊂ B(I) has finite index for some (and hence all) interval I. The index will be called the index of the inclusion π(A) ⊂ B and is denoted by [B : A]. If π as representation of A decomposes as [π] = Pλ mλ[λ] where mλ are non-negative integers and λ are irreducible DHR representations of A, we say that [π] = Pλ mλ[λ] is the spectrum of the extension. For simplicity we will write π(A) ⊂ B simply as A ⊂ B. Lemma 2.2. If A ⊂ B is a conformal subnet with finite index, then A ⊂ B is irreducible. Proof. This is proved in Cor. 3.6 of [2], without assumption of conformal covariance of A but under the additional assumption that A is strongly additive to ensure the equivalence of local and global intertwiners, but for conformal net A the equivalence of local and global intertwiners for finite index representations are proved in §2 of [13], thus the proof of Cor. 3.6 of [2] applies verbatim. (cid:4) j j ) = δij1/dλ, E(T¯λiT¯λ∗ Lemma 2.3. Suppose that A ⊂ B has finite index , and let [π] = Pλ mλ[λ] be as in Definition above. Fix an interval I and suppose that λ, ¯λ is localized on I. (1) Let Kλ := {T ∈ B(I)T a = aλ(a)T,∀a ∈ A(I)}. Then Kλ is a vector space of dimension mλ ≤ dλ. One can find isometries Tλi ∈ Kλ, Tλi ∈ K¯λ, 1 ≤ i ≤ mλ such λi ∈ A(I)T¯λi; that Tλia = λ(a)Tλi,∀a ∈ A, E(TλiTλ∗ Every b ∈ B(I) can be written as b = Pλi (2) Let Lλ ⊂ Kλ be subspaces with the following properties:(a) LλLµ ⊂ Pν A(I)Lν; λ ⊂ A(I)L¯λ. Then there is an intermediate subnet A ⊂ C ⊂ B such that C(I) = Pλ A(I)Lλ. Conversely every intermediate subnet arises this way; (3) If Ω is the vacuum vector of B, and denote by AΩ = H0, T ∗ λiAΩ = Hλi, then as Hilbert space H = Lλi,1≤i≤mλ λi : H0 → Hλi is a unitary intertwiner between the action of λ(A(I)) on H0 and A(I) on Hλi. Proof. (1) and(2) follow from §3 of [25] and §2 of [22]. For (3), only unitarity has to be checked. We have Hλi, and the map √dλT ∗ ) = δij1/dλ, T ∗ (b) L∗ dλT ∗ λiE(Tλib); hT ∗ λia1Ω, T ∗ λia2Ωi = ha∗ and the proof is complete. 2E(TλiT ∗ λi)a1Ω, Ωi = 1/dλha1Ω, a2Ωi,∀a1, a2 ∈ A(I), (cid:4) Induced endomorphisms 2.3 Suppose a conformal net A and a representation λ is given. Fix an open interval I of the circle and Let M := A(I) be a fixed type III1 factor. Then λ give rises to an endomorphism still denoted by λ of M. Suppose {[λ]} is a finite set of all equivalence classes of irreducible, covariant, finite-index representations of an irreducible local 7 conformal net A. We will use ∆A to denote all finite index representations of net A and will use the same notation ∆A to denote the corresponding sectors of M. We will denote the conjugate of [λ] by [¯λ] and identity sector (corresponding to the vacuum representation) by [1] if no confusion arises, and let N ν λµ = h[λ][µ], [ν]i. Here hµ, νi denotes the dimension of the space of intertwiners from µ to ν (denoted by Hom(µ, ν)). The univalence of λ and the statistical dimension of (cf. §2 of [13]) will be denoted by ωλ and d(λ) (or dλ)) respectively. Suppose that ρ ∈ End(M) has the property that γ = ρ¯ρ ∈ ∆A. By §2.7 of [23], we can find two isometries v1 ∈ Hom(γ, γ2), w1 ∈ Hom(1, γ) such that ¯ρ(M) and v1 generate M and 1γ(w1) = d−1 ρ v∗ 1w1 = v∗ v1v1 = γ(v1)v1 By Thm. 4.9 of [23], we shall say that ρ is local if 1γ(w1) = d−1 ρ v∗ 1w1 = v∗ v1v1 = γ(v1)v1 ¯ρ(ǫ(γ, γ))v1 = v1 (1) (2) (3) Note that if ρ is local, then (4) For each (not necessarily irreducible) λ ∈ ∆A, let ε(λ, γ) : λγ → γλ (resp. ε(λ, γ)), be the positive (resp. negative) braiding operator as defined in Section 1.4 of [33]. Denote by λε ∈ End(M) which is defined by ωµ = 1,∀µ ≺ ρ¯ρ λε(x) : = ad(ε(λ, γ))λ(x) = ε(λ, γ)λ(x)ε(λ, γ)∗ λε(x) : = ad(ε(λ, γ))λ(x) = ε(λ, γ)∗λ(x)ε(λ, γ)∗,∀x ∈ M. By (1) of Theorem 3.1 of [33], λερ(M) ⊂ ρ(M), λερ(M) ⊂ ρ(M), hence the following definition makes sense: Definition 2.4. If λ ∈ ∆A define two elements of End(M) by aρ λ(m) := ρ−1(λερ(m)), aρ λ(m) := ρ−1(λερ(m)),∀m ∈ M. λ) will be referred to as positive (resp. negative) induction of λ with respect aρ λ (resp. aρ to ρ. Remark 2.5. For simplicity we will use aλ, aλ to denote aρ inductions are with respect to the same ρ. λ, aρ λ when it is clear that The endomorphisms aλ are called braided endomorphisms in [33] due to its braid- ing properties (cf. (2) of Corollary 3.4 in [33]), and enjoy an interesting set of prop- erties (cf. Section 3 of [33]). We summarize a few properties from [33] which will be used in this paper: (cf. Th. 3.1 , Co. 3.2 and Th. 3.3 of [33] ): 8 Proposition 2.6. (1). The maps [λ] → [aλ], [λ] → [aλ] are ring homomorphisms; (2) aλ ¯ρ = aλ ¯ρ = ¯ρλ; (3) When ρ¯ρ is local, haλ, aµi = haλ, aµi = haλ ¯ρ, aµ ¯ρi = haλ ¯ρ, aµ ¯ρi; (4) (3) remains valid if aλ, aµ (resp. aλ, aµ) are replaced by their subsectors. In particular we have haλ, σi = hλ, ρσ ¯ρi if σ ≺ aµ. The following is Porp. 2.24 of [39]: Proposition 2.7. Suppose that ρ¯ρ ∈ ∆. Then: (1) ρ is local iff h1, aµi = hρ¯ρ, µi,∀µ ∈ ∆A; (2) ρ = ρ′ρ′′ = ρ′ ρ′′ where ρ′, ρ′′, ρ′, ρ′′ ∈ End(M), and ρ′, ρ′ are local which verifies hρ′ ¯ρ′, µi = h1, aµi = h1, aρ′ µ i hρ′ ρ′, µi = h1, aµi = h1, aρ′ µ i ∀µ ∈ ∆A. We refer to ρ′ (resp. ρ′′) as the left (resp.right) local support of ρ. The following Lemma is Prop. 3.23 of [2] (The proof was also implicitly contained in the proof of Lemma 3.2 of [33]): Lemma 2.8. If ρ¯ρ is local, then [aλ] = [aλ] iff ε(λ, ρ¯ρ)ε(ρ¯ρ, λ) = 1 iff ε(λ, µ)ε(µ, λ) = 1,∀µ ∈ ρ¯ρ. We shall make use of the following notation in §4: Definition 2.9. For λ, µ ∈ ∆A, Z ρ λµ := haλ, aµi. 2.4 Jones-Wassermann subfactors from representation of Loop groups Let G = SU(n). We denote LG the group of smooth maps f : S1 7→ G under pointwise multiplication. The diffeomorphism group of the circle DiffS1 is naturally a subgroup of Aut(LG) with the action given by reparametrization. In particular the group of rotations RotS1 ≃ U(1) acts on LG. We will be interested in the projective unitary representation π : LG → U(H) that are both irreducible and have positive energy. This means that π should extend to LG ⋊ Rot S1 so that H = ⊕n≥0H(n), where the H(n) are the eigenspace for the action of RotS1, i.e., rθξ = exp(inθ) for θ ∈ H(n) and dim H(n) < ∞ with H(0) 6= 0. It follows from [30] that for fixed level k which is a positive integer, there are only finite number of such irreducible representations indexed by the finite set P k ++ = (cid:26)λ ∈ P λ = Xi=1,··· ,n−1 λiΛi, λi ≥ 0 , Xi=1,··· ,n−1 λi ≤ k(cid:27) 9 where P is the weight lattice of SU(n) and Λi are the fundamental weights. We will write λ = (λ1, ..., λn−1), λ0 = k −P1≤i≤n−1 λi and refer to λ0, ..., λn−1 as components of λ. We will use Λ0 or simply 1 to denote the trivial representation of SU(n). For is given by the µ S(δ∗) /S(δ Λ0) where S(δ) λ S(δ) λ S(δ) ν λ, µ, ν ∈ P k Kac-Peterson formula: ++, define N ν ++ λµ = Pδ∈P k S(δ) λ = c Xw∈Sn εw exp(iw(δ) · λ2π/n) is an orthonormal system. It is shown in [17] P. 288 that N ν where εw = det(w) and c is a normalization constant fixed by the requirement that S(δ) λµ are non-negative µ integers. Moreover, define Gr(Ck) to be the ring whose basis are elements of P k ++ with structure constants N ν ++ is defined by λ 7→ λ∗ = the conjugate of λ as representation of SU(n). Note that λ → Sλµ gives a representation of Gr(Ck). λµ. The natural involution ∗ on P k S1µ We shall also denote S(Λ) S-matrix of LSU(n) at level k. Λ0 by S(Λ) 1 . Define dλ = S(λ) S(Λ0) 1 1 . We shall call (S(δ) ν ) the We shall encounter the Zn group of automorphisms of this set of weights, generated by σ : λ = (λ1, λ2,· · · , λn−1) → σ(λ) = (k − 1 − λ1 − · · · λn−1, λ1,· · · , λn−2). Define col(λ) = Σi(λi − 1)i. col(λ) will be referred to as the color of λ. The central n of SU(n) acts on representation of SU(n) labeled by λ as exp( 2πicol(λ) element exp 2πi ). The irreducible positive energy representations of LSU(n) at level k give rise to an irreducible conformal net A (cf. [18]) and its covariant representations. We will use λ = (λ1, ...λn−1) to denote irreducible representations of A and also the corresponding endomorphism of M = A(I). All the sectors [λ] with λ irreducible generate the fusion ring of A. For λ irreducible, the univalence ωλ is given by an explicit formula (cf. 9.4 of [PS]). Let us first define hλ = c2(λ) k+n where c2(λ) is the value of Casimir operator on representation of SU(n) labeled by dominant weight λ. hλ is usually called the conformal dimension. Then we have: ωλ = exp(2πihλ). The conformal dimension of λ = (λ1, ..., λn−1) is given by n hλ = 1 2n(k + n) X1≤i≤n−1 i(n−i)λ2 i + 1 n(k + n) X1≤j≤i≤n−1 j(n−i)λjλi+ 1 2(k + n) X1≤j≤n−1 j(n−j)λj (5) The following result is proved in [32] (See Corollary 1 of Chapter V in [32]). Theorem 2.10. Each λ ∈ P (k) generated by all λ ∈ P (k) ++ has finite index with index value d2 ++ is isomorphic to Gr(Ck). λ. The fusion ring 10 Remark 2.11. The subfactors in the above theorem are called Jones-Wassermann subfactors after the authors who first studied them (cf. [15],[32]). Definition 2.12. v := (1, 0, ..., 0), v0 := (1, 0, ..., 0, 1), ωi = kΛi, 0 ≤ i ≤ n − 1. v (resp. v0) will be referred to as vector (resp. adjoint) representation. The following is observed in [11]: Lemma 2.13. Let (0, ..., 0, 1, 0, ...0) be the i-th (1 ≤ i ≤ n − 1) fundamental weight. Then [(0, ..., 0, 1, 0, ...0)λ] are determined as follows: µ ≺ (0, ..., 0, 1, 0, ...0)λ iff when the Young diagram of µ can be obtained from Young diagram of λ by adding i boxes on i different rows of λ, and such µ appears in [(0, ..., 0, 1, 0, ...0)λ] only once. Lemma 2.14. (1) If [λ] 6= ωi for some 0 ≤ i ≤ n − 1, then v0 ≺ λ¯λ; (2) If λ1λ2 is irreducible, then either λ1 or λ2 = ωi for some 0 ≤ i ≤ n − 1; (3) Suppose that λ has color 0modn. Then λ ≺ vm 0 for some m ∈ N. Proof. (1), (2) is lemma 2.30 of [39]. By the lemma above λ ≺ vl for some l ∈ N, and since col(λ) = 0 modn, we have l = nl1, l1 ∈ N. Since 1 ≺ vn, 1 ≺ ¯vn, we have [vn] ≺ [vn¯vn] = ([v0] + [1])n, and (3) follows. (cid:4) 2.5 Subnets from conformal inclusions Let G ⊂ H be inclusions of compact simply connected Lie groups. LG ⊂ LH is called a conformal inclusion if the level 1 projective positive energy representations of LH decompose as a finite number of irreducible projective representations of LG. LG ⊂ LH is called a maximal conformal inclusion if there is no proper subgroup G′ of H containing G such that LG ⊂ LG′ is also a conformal inclusion. A list of maximal conformal inclusions can be found in [24]. Let H 0 be the vacuum representation of LH, i.e., the representation of LH asso- ciated with the trivial representation of H. Then H 0 decomposes as a direct sum of irreducible projective representation of LG at level K. K is called the Dynkin index of the conformal inclusion. We shall write the conformal inclusion as GK ⊂ H1. Note that it follows from the definition that AH1 is an extension of AGK . We shall limit our consideration to the following conformal inclusions so we can use the results of [33]: SU (n)n−2 ⊂ SU (cid:18) n(n − 1) SU (n)n+2 ⊂ SU (cid:18) n(n + 1) SU (n)n ⊂ Spin(n2 − 1)1, N ≥ 2; (cid:19)1 (cid:19)1 ; SU (n)m × SU (m)n ⊂ SU(mn)1. , N ≥ 4; 2 2 (6) (7) (8) (9) Note that except equation (9), the above cover all the infinite series of maximal conformal inclusions of the form SU(N) ⊂ H with H being a simple group. 11 3 Intermediate subnets in confonmal subnets as- sociated with conformal inclusions Let A ⊂ B be conformal subnets associated with conformal inclusions in §2.5, i.e., A = AGk ⊂ B = AH1. Our goal in this section is to list all intermediate subnets A ⊂ C ⊂ B. The spectrum [π] = Pλ mλλ of A ⊂ B is given by [1] and [21]. One interesting feature is that all mλ = 1. We write HB = ⊕λHλ with H0 the vacuum representation of A, and HB (resp. HC) the vacuum representation space of B (resp. HC). Fix an interval I and let M = A(I) ⊂ C(I), ρ ∈ End(M), ρ¯ρ = HC ∈ ∆A where we use HC to denote the restriction of the vacuum representation of C to A. For λ ∈ ∆A, we will write aC λ := aρ λ. 3.1 Smeared Vertex Operators Let g (resp. h) be the Lie algebra of G (resp. H). Choose a basis eα, e−α, hα in hC := h ⊗ C with α ranging over the set of roots as in §2.5 of [30]. Let Xα := i(eα + e−α), Yα := (eα − e−α). Denote by h the affine Kac-Moody algebra (cf. P. 163 of [20]) associated to hC. Note h = hC ⊗ C[t, t−1] ⊕ Cc, where Cc is the 1-dimensional center of h. For X ∈ h, Define X(n) := X ⊗ tn, X(z) := Pn X(n)z−n−1 as on Page 312 of [19]. Let π0 be the vacuum representation of LH1 on HB with vacuum vector Ω. Let D be the generator of the action of the rotation group on HB. H 0 B will denote the finite linear sum of the eigenvectors of D. For ξ ∈ HB, we define xs = (1+D)sx, s ∈ R. H s := {x ∈ H 0 B xs < ∞} and H(∞) = ∩s∈RH s. Note that when s ≥ 0, H s is a complete space under the norm .s. Clearly H 0 B (resp. H(∞)) will be called finite energy vectors (resp. smooth vectors). The eigenvalue of D is sometimes referred to as energy or weight. Let us recall a few elementary facts about vertex operators which will be used. B) to be See [8] or [16] for an introduction on vertex operator algebras. Define End(H 0 the space of all linear operators (not necessarily bounded) from H 0 B ⊂ H(∞). The elements of H 0 B to H 0 B and set End(H 0 B)[[z, z−1]] := {Xn∈Z vnznvn ∈ End(H 0 B)}. By the statement on P. 154 of [9] which follows from Th. 2.4.1 of [9] there exists a linear map ψ ∈ H 0 B → V (ξ, z) = Xm∈Z ψ(m)z−m−1 ∈ End(H 0 B)[[z, z−1]] with the following properties: (1) ψ(−1)Ω = ψ; (2) If ψ = Xi1(−1)...Xit(−1)Ω, 12 then V (ψ, z) =: Xi1(z)...Xit(z) : where :, : are normal ordered products (cf. (2.38), (2.39) of [5]). V (ψ, z) is called a vertex operator of ψ. Let f = Pm f (m)zm be a smooth test function. Define (1 + m)sf (m). fs = Xn∈Z The smeared vertex operator V (ψ, f ) is defined to be: V (ψ, f ) = 1 2πi ZS1 V (ψ, z)f dz = Xm f (m)ψ(m). V (ψ, f ) is a well defined operator on H 0 V (ψ, f ) on H 0 B. It is defined by the equation B. Let V (ψ, f )F A be the formal adjoint of hV (ψ, f )x, yi = hx, V (ψ, f )F Ayi,∀x, y ∈ H 0 B where h,i is the inner product on Hilbert space H 0 B. Lemma 3.1. The subspace spanned by V (ψ, f )Ω,∀ψ ∈ H 0 suppf ∈ I, is dense in Hλ. Proof. The proof is essentially the same as the proof of Reeh-Schlieder Th. Let ξ ∈ Hλ be a vector which is orthogonal to the subspace spanned by V (ψ, f )Ω,∀ψ ∈ λ, suppf ∈ I. Suppose J is an open interval such that ¯J ⊂ I, and f is a smooth H 0 function with support in J. Consider the function λ = Hλ ∩ H 0 B,∀f smooth, F (z) = hexp(izD)ξ, V (ψ, f )Ωi. Since the spectrum of D on Hλ is a subset of non-negative integers, it follows that F (z) is holomorphic on the upper half plane, continues on the real line, and vanishes on an open interval on the real line. It follows by Schwartz reflection principle that F (z) is identically zero, and we have hexp(itD)ξ, V (ψ, f )Ωi = hξ, exp(−itD)V (ψ, f ) exp(itD)Ωi = 0,∀t ∈ R. On the other hand exp(−itD)V (ψ, f ) exp(itD)Ω = V (ψ, Rt(f ))Ω, where Rt(f )(z) = f (exp(it)z). Choose a covering of S1 by intervals RtiI, 1 ≤ n ≤ n and smooth functions fi with support in RtiI, 1 ≤ n ≤ n such that P1≤i≤n fi = 1 z , then 0 = hξ, X1≤i≤n V (ψ, fi)Ωi = hξ, ψi,∀ψ ∈ H 0 λ. Since H 0 λ is dense in Hλ, we conclude that ξ = 0 and the lemma is proved. (cid:4) 13 Recall HB = ⊕λHα as representations of LGk or AGk. The lowest energy space of Hα, denoted by Hλ(0) is a highest weight module of G with weight λ. The vertex operator is a primary vertex operator for g with highest weight λ(cf. [19] and [9]). By a slightly V (ψ, z) : Hλ(0) → End(H 0 B)[[z, z−1]] H 0 abuse of notations we write such operator as V (λ) = Pm V (λ)mz−m−1. Definition 3.2. We define V (λ)V (µ)H0 to be the linear span of V (λ)mV (µ)nH 0 0 ,∀n, m. Note that by definition V (λ)V (µ)H0 is a g submodule of HB, and if V (λ)V (µ)H0 ⊃ 0 , then λ = ¯µ. The weight 1 element in HB is a Lie algebra isomorphic to hC and will be identified with hC. It has a subspace isomorphic to gC. The vertex operator associated with h, V (h, z) is usually written as h(z), similarly we write V (h, f ) as h(f ). h(f ) are skew adjoint unbounded operators if f = f ∗. The orthogonal complement of gC in hC, denoted by hC ⊖ gC is a direct sum of Hλ(0) with hλ = 1. Lemma 3.3. (1) Let Tλ, Tµ be as in Lemma 2.3. Then T ∗ λ T ∗ µ H0 ⊂ V (λ)V (µ)H0; (2) If E(TνT ∗ λ T ∗ µ ) 6= 0, then Hν ⊂ V (λ)V (µ)H0; Proof. Ad (1): We choose interval I1 which is disjoint from I. By Lemma 3.1 it is sufficient to check that for all smooth f with support in I1 and ψ ∈ H 0 µ, T ∗ λ V (ψ, f )Ω ∈ V (λ)V (µ)H0. By choosing H trivial in Prop. 2.3 of [34], we know that V (ψ, f ) is affiliated with B(I1), and by locality we have T ∗ λ Ω. Since V (µ)V (λ)H0 is a g module, the orthogonal complement of V (µ)V (λ)H0 is a direct of irreducible g module. λ V (ψ, f )Ω = V (ψ, f )T ∗ Now suppose that ξ ∈ H 0 B is orthogonal to V (µ)V (λ)H0. Choose χn ∈ H 0 λ such λ Ω in norm. Then by Lemma 1 of [34] that χn → T ∗ hV (ψ, f )χn, ξi = 0 = hχn, V (ψ, f )∗ξi. Now let n go to infinity we have hT ∗ λ Ω, V (ψ, f )∗ξi = 0 = hV (ψ, f )T ∗ λ Ω, ξi, and (1) is proved. λ T ∗ ν E(TνT ∗ Ad (2): Since T ∗ µ ), it follows that from (1) if E(TνT ∗ λ T ∗ µ = Pν T ∗ µ ) 6= λ T ∗ µ )∗ ∈ A(I) is an isometry up to non-zero constant, and so Hν = λ T ∗ µA(I)Ω ⊂ V (λ)V (µ)H0. 0, then E(TνT ∗ ν A(I)Ω ⊂ T ∗ T ∗ Lemma 3.4. Suppose that Hλ ∈ HC with hλ = 1. Let ψ ∈ Hλ ∩ (h ⊖ g), and f = f ∗ a smooth function with support in I. Then exp(V (ψ, f )) ∈ C. λ T ∗ (cid:4) 14 Proof. Let EC : B → C be the conditional expectation which is implemented by the projection PC on HB with range HC. We first show that PCV (ψ, f ) exp(tV (ψ, f ))Ω = V (ψ, f )PC exp(tV (ψ, f ))Ω. For any b ∈ B(I ′) we have hPCV (ψ, f ) exp(tV (ψ, f ))Ω, bΩi = hV (ψ, f ) exp(tV (ψ, f ))Ω, EC(b)Ωi = −hexp(tV (ψ, f ))Ω, V (ψ, f )EC(b)Ωi. Since by Prop. 2.3 of [34] V (ψ, f ) is skew self adjoint and is affiliated with B(I), and note that V (ψ, f )Ω ∈ Hλ ⊂ HC, it follows that hexp(tV (ψ, f ))Ω, V (ψ, f )EC(b)Ωi = hexp(tV (ψ, f ))Ω, EC(b)V (ψ, f )Ωi = hexp(tV (ψ, f ))Ω, PCbV (ψ, f )Ωi = hPC exp(tV (ψ, f ))Ω, V (ψ, f )bΩi. By (2) of Lemma 4 in [34] PC exp(tV (ψ, f ))Ω ∈ H(∞), and H(∞) is in the domain of skew self adjoint operator V (ψ, f ). It follows that hPC exp(tV (ψ, f ))Ω, V (ψ, f )bΩi = −hV (ψ, f )PC exp(tV (ψ, f ))Ω, bΩi, and we have shown that hPCV (ψ, f ) exp(tV (ψ, f ))Ω, bΩi = hV (ψ, f )PC exp(tV (ψ, f ))Ω, bΩi. By Reeh-Schleder Th. we have shown that PCV (ψ, f ) exp(tV (ψ, f ))Ω = V (ψ, f )PC exp(tV (ψ, f ))Ω. Set F (t) := hPC exp(tV (ψ, f ))Ω, exp(tV (ψ, f ))Ωi. Then F (0) = 1 and F ′(t) = hV (ψ, f ) exp(tV (ψ, f ))Ω, PC exp(tV (ψ, f ))Ωi+ hexp(tV (ψ, f ))Ω, PCV (ψ, f ) exp(tV (ψ, f ))Ωi = 0 where we have used PCV (ψ, f ) exp(tV (ψ, f ))Ω = V (ψ, f )PC exp(tV (ψ, f ))Ω and PC exp(tV (ψ, f ))Ω ∈ H(∞), and H(∞) is in the domain of skew self adjoint operator V (ψ, f ). It follows that F (t) = 1 and we conclude that exp(tV (ψ, f ))Ω ∈ HC which proves our lemma. (cid:4) The following uses an analogue of VOA statement that weight 1 space has a Lie algebra structure. 15 Lemma 3.5. If Hλ ⊂ HC with hλ = 1, then C = B. Proof. By Lemma 3.4 for any ψ ∈ Hλ∩h, and f = f ∗ a smooth function with support in I, we have exp(V (ψ, f )) ∈ C. Since the conformal inclusions are maximal, it follows that the Lie algebra generated by g and ψ is in fact Lie algebra h. By Lie's formula, if exp(iV (ψj, fj)) ∈ C(I), j = 1, 2 then ((exp(V (ψ1, f1)/n) exp(V (ψ2, f2)/n)(exp(−V (ψ1, f1)/n) exp(−V (ψ2, f2)/n))n2 converges strongly to On the other hand exp([V (ψ1, f1), V (ψ1, f1)]). [V (ψ1, f1), V (ψ2, f1)] = V ([ψ1, ψ2], f g) + hψ1, ψ2iZS1 f1f2dz/z. It follows that for any ψ ∈ h, and smooth functions f = f ∗ with support in I we have that exp(V (ψ, f )) ∈ C(I). Since B(I) is generated as a von Neumann algebra by such elements, we have shown that C = B. (cid:4) 3.2 Induction of the adjoint representation The following is a key observation in this section, and is already implicitly contained in (3) of Lemma 2.33 in [39]. Proposition 3.6. Suppose that C contains no weight 1 element except those in A, then aC v0 is irreducible. Proof. By Lemma 2.13 we have [v0 2] = [1]+2[v0]+[(2, 0, ..., 0, 2)]+[(0, 1, 0, ..., 1, 0)]+[(0, 1, 0, ..., 0, 2)]+[(2, 0, ..., 0, 1, 0)] By computing the conformal dimensions of the descendants of v0 we have 2 using equation (5) h(2,0,...,0,2) = 2 + 2n k + n , h(0,1,...,0,2) = h(2,0,...,1,0) = 2n k + n , h(0,1,...,1,0) = 2n − 2 k + n Hence if C contains no weight 1 element except those in A, then haC v0, aC v0i = hHC, v0v0i = 1 where recall that we use HC to denote the restriction of the vacuum representation of C to A, and the proposition is proved. (cid:4) Lemma 3.7. Suppose that ǫ(λ, v0)ǫ(v0, λ) = 1, then λ = ωi for some 0 ≤ i ≤ n. 16 Proof. By definition we have Sv0λ S1λ = dv0. From [v][¯v] = [1] + [v0] we have Svλ S1λ S¯vλ S1λ = dv0 + 1 ≤ dvd¯v = dv0 + 1 S1λ = dv. For any positive integer k, suppose that It follows that we must have Svλ [vk] = Pµ mµ[µ], then we have Sµλ Sµλ dn mµ S1λ ≤ Xµ S1λ ≤ Xµ v = Xµ S1λ = dµ,∀µ ≺ vk. Since every irrep of A occurs in some vk, it follows S1λ = dµ,∀µ. Square both sides and sum over µ, we have proved (cid:4) It follows that Sµλ that we must have Sµλ that dλ = 1, and hence the Lemma. mµdµ = dn v . mµ 3.3 List of intermediate subnets from conformal inclusions Theorem 3.8. (1) For the subnet A ⊂ B corresponding to conformal inclusions in 8, when n is odd (resp. even) the intermediate subnet C are in one to one correspondence with the abelian subgroup Zn (resp. Zn/2) generated by ω, (resp. ω2) i.e., if ωi, ik = n (resp. ω2i, 2ik = n) is a generator of this subgroup, then the spectrum of C is HC = P1≤j≤k Hωij (resp. HC = P1≤j≤k Hω2ij ); (2): For the subnet A ⊂ B corresponding to conformal inclusions in 6, 7, when n is odd there is no intermediate subnet. When n = 2m is even, the only nontrivial intermediate subnet C is a Z2 extension of A by the simple current ωm, i.e., the spectrum is HC = H0 + Hωm; v0] = [aC v0 is irreducible, it follows that aC Proof. Ad (1): By Lemma 3.5 we can assume that C has no weight 1 elements besides those of A. By Prop. 3.6 we know that aC v0 is irreducible. In the case of conformal inclusions in 8, since the vector representation of LH, when restricting to A, contains the adjoint representation, it follows from (4) of 2.6 that aC v0 must contain a DHR representation of C. Since aC v0 is a DHR representation of C, i.e., [aC v0]. By Lemma 2.8 we must have for any λ ∈ HC, ǫ(λ, v0)ǫ(v0, λ) = 1. By Lemma 3.7 we conclude that λ = ωi for some 0 ≤ i ≤ n. (1) now follows easily by inspection of the spectrum of A ⊂ B in [1]. Ad (2): In the case of conformal inclusions in 6 (resp. 7) , we note that the vector representation of LH, when restricting to A, contains the antisymmetric representa- tion (0, 1, 0, ..., 0) (resp. symmetric representation (2, 0, 0, ...0) of A).) Since v¯v, aC (0,2,0,...,0)] + [aC (2,0,...,0)] + [aC (0,1,0,...,0)], [aC (1,1,0,...,0)]i = haC v¯vi = 2, v2, aC haC v2i = h[aC 17 (2,0,...,0) and aC it follows that both aC for the case of conformal inclusions in 6 (resp. 7), aC representations of C. It follows that if λ ∈ HC, then ǫ(λ, (0, 1, 0, ..., 0))ǫ((0, 1, 0, ..., 0), λ) = 1 (resp.ǫ(λ, (2, 0, ..., 0))ǫ((2, 0, ..., 0), λ) = 1). Similarly ǫ(λ, (0, 0, ..., 0, 1, 0))ǫ((0, 0, ..., 1, 0), λ) = 1 (resp.ǫ(λ, (0, 0, ..., 2))ǫ((0, 0, ..., 2), λ) = 1). (0,1,0,...,0) are irreducible. Hence as in the proof of (1), (2,0,...,0)) are DHR (0,1,0,...,0) (resp. aC Since by Lemma 2.14 v0 appears in the product of (0, 1, 0, ..., 0) (resp. (2, 0, ..., 0))and its conjugate, it follows that ǫ(λ, v0)ǫ(v0, λ) = 1. By Lemma 3.7 we conclude that λ = ωi, 1 ≤ i ≤ n, and (2) follows by inspection of the spectrum of A ⊂ B as given in [21]. (cid:4) We note that the same idea in the proof of Theorem above gives a proof of the following: Corollary 3.9. Suppose ASU (n)k ⊂ C, n 6= n, n± 2, and there is a representation of C, when restricting to ASU (n)k, contains v0. Then C is an extension by simple currents. Remark 3.10. Since conformal inclusion SU(2)10 ⊂ Spin(5)1 is not a simple current extension, this example shows that the condition in the above corollary is necessary. In fact in this case the adjoint representation of ASU (2)10 does not appear in the restriction of any irreps of ASpin(5)1. Theorem 3.11. For the subnet A ⊂ B associated with 9, let (n, m) = p, n = n1p, m = m1p. Then the intermediate subnets C are in one to one correspondence with the sub- group of Zp, i.e.,each such C has spectrum HC = P0≤l≤k2 H(ωn1 k1 l, ωm1 k1l) with k1k2 = p, where we use λ to denote the highest weights of SU(m)n. Proof. Since ASU (n)m ⊂ B is normal (cf. §4 of [35]), it follows that for each (λ, λ) ∈ HC, λ is a ring isomorphism. So if (λi, λi) ∈ HC, i = we must have [aC 1, 2, and ], it follows there must be a λ3 such that λ3 ≺ λ1λ2, then [aC λ3] = [aC λ3 It follows that C are in one to one correspondence with the set RC of λ with color zero mod n which are closed under conjugation and fusion product. If RC contains any λ with dλ 6= 1, by Lemma 2.14 we have v0 ∈ R, and it follows that R contains all λ with color zero mod n, in which case C = B. Now assume that dλ = 1 if λ ∈ RC. The RC must be a subgroup generated by ωn1k1, k1k2 = p. Since the color of elements in RC is zero mod n, our Theorem follows. (cid:4) ], and λ → aC λ3] ≺ [aC ] and (λ3, λ3) ∈ HC. λ] = [aC λ [aC λ1λ2] = [aC λ1 λ2 By checking the list of intermediate subnets from Th. 3.8 and Th. 3.11 we immediately have: Corollary 3.12. Conjecture 1.2 is true for AGK ⊂ AH1 where GK ⊂ H1 are confor- mal inclusions in 6,7,8 and 9. For the conformal inclusions Gk ⊂ H1, we write VA (resp. VB) the VOA (cf. [9]) associate with affine g at level k (resp. affine h at level 1) We have natural inclusion VA ⊂ VB. We are interested in VOA VC such that VA ⊂ VC ⊂ VB. We say a VOA is 18 simple it is irreducible as a representation over itself. Note that VC will be direct sum of g modules, and we can write VC = ⊕Hλ and we refer to those λ which appear in VC as the spectrum of VC. Proposition 3.13. For any simple VOA VC such that VA ⊂ VC ⊂ VB, there corre- sponds a unique intermediate subnet C such the spectrum of A ⊂ C is the same as the spectrum of VA ⊂ VC. Proof. Fix an interval I. For each λ in the spectrum of VA ⊂ VC, denote by Tλ be as in Lemma 2.3. By Lemma 3.3, it follows that if E(TνT ∗ µ ) 6= 0 where λ, µ are in the spectrum of VA ⊂ VC, then ν is also in the spectrum of VA ⊂ VC. If λ is in the spectrum of VA ⊂ VC but ¯λ is not, then by the remark after Definition 3.2 the action of VC on Hλ will span an invariant subspace of HC which does not contain H0, contradicting our assumption that C is simple. By (2) of Lemma 2.3, A, Tλ where λ is in the spectrum of VA ⊂ VC generate an intermediate subnet C with its spectrum the same as the spectrum of VA ⊂ VC. (cid:4) λ T ∗ The above proposition immediately implies the following theorem: Theorem 3.14. The set of intermediate simple VOA VC in VA ⊂ VB for conformal inclusions 6,7,8 and 9 are in one to one correspondence with the set of intermediate subnets C of A ⊂ B with the same spectrum as given in Th. 3.8 and Th. 3.11. Remark 3.15. We note that the simple intermediate VOAs in the above theorem are simple current extensions of affine VOAs, and they are well understood in VOA literature (cf.[6]). 4 Verifying Conjecture 1.1 for Jones-Wassermann subfactors In this section we extend the results in Cor. 5.23 of [39]. Let λ be an irreducible representation of ASU (n)k localized on I, M := ASU (n)k(I). Suppose λ = c1c2 where ci ∈ End(M), i = 1, 2, c1(M) is an intermediate subfactor of λ(M) ⊂ M. We note that c1¯c1 ≺ λ¯λ. We say the intermediate subfactor c1(M) is of abelian type if [c1¯c1] = P1≤i≤j[ωij1], jj1 = n. The following Lemma appears as Lemma 5.22 in the correction of proof of [39] and we include its proof: Lemma 4.1. Assume that Z c1 1µ = δ1µ,∀µ where Z c1 is defined as in Definition 2.9. Then hc1c2, c1c2i = hc1¯c1, ¯c2c2i. Proof By §2 of [10] we have Z c1 µ1µ2 = δµ1τ (µ2) where µ → τ (µ) is an order two automorphism of fusion algebra. It follows that [aµ] = [aτ (µ)], and by [3] irreducible sectors of ¯c1νc1 are of the form aµ,∀µ. Since hc2¯c2, aµi = hc2, aµc2i = hc2, c2µi = h¯c2c2, aµi = ha¯c2c2, aµi, 19 we conclude that [c2¯c2] = [a¯c2c2], and hc1¯c1, ¯c2c2i = hc1, ¯c2c2c1i = hc1, c1a¯c2c2i = hc1, c1c2¯c2i = hc1c2, c1c2i (cid:4) Theorem 4.2. (1) Suppose that k 6= n − 2, n + 2, n. then λ is maximal iff there is no 1 ≤ i ≤ n − 1 such that [ωiλ] = [λ]; (2) When λ is not maximal, the maximal intermediate subfactor is either abelian type or at most one given by c1(M) with λ = c1c2, [¯c2][c2] = [1] + [ωm], n = 2m. Proof Ad (1): (1) is Cor. 5.23 in [39]. We include its proof which will be modified in our proof of (2). When k = 1 the Cor. is obvious. By Lemma 2.33 of [39] we can assume that k ≥ 2 and dv0 > 1. As in the proof of Cor. 5.21 in [39], λ is maximal implies that there is no 1 ≤ i ≤ n − 1 such that [ωiλ] = [λ]. Now suppose that there is no 1 ≤ i ≤ n − 1 such that [ωiλ] = [λ]. If Svλ 6= 0, then λ is maximal by Cor. 5.20 of [39]. If k = 2, the S matrix elements are equal to that of S matrix elements for SU(2)n up to phase factors, and it follows easily that Svλ 6= 0 if there is no 1 ≤ i ≤ n − 1 such that [ωiλ] = [λ]. Suppose that k ≥ 3, Svλ = 0. Since [v¯v] = [1] + [v0] we have Sv0λ = −S1λ 6= 0. Assume that M1 is an intermediate subfactor between λ(M) and M, and λ = c1c2 with c1(M) = M1 and c1 = c′ 1 as in Prop. 2.7. Apply Lemma 2.20 of [39] we v0, ac′ have hac′ v0] and by Lemma 2.36 of [39] [c′ 1.] Since λ = c′ 1] = [1]. By Prop. 2.7 we must have Z c1 µ1µ2 = δµ1τ (µ2) where τ (µ) = ωmcol(µ)µ or τ (µ) = ωmcol(µ) ¯µ, m ≥ 0. We claim that in fact [ωm] = [1] and τ (µ) = µ. First we show that τ (µ) = ωmcol(µ)µ. If instead τ (µ) = ωmcol(µ) ¯µ, since k ≥ 3, τ ((0, 1, 0, ..., 0)) 6= (0, 1, 0, ..., 0), by Lemma 2.20 of [39] we must have Sλ(0,1,0,...,0) = 0. From the fusion rule v0i ≥ 1. By Lemma 2.33 of [39] we must have [ac′ 1¯c′ 1] = P1≤j≤n/j1 1c2, [ωj1λ] = [λ], and by assumption j1 = n and [c′ 1c′′ µ1 = δµ1,∀µ. By §2 of [10] we have Z c1 [ωjj1]. By Frobenius reciprocity we have [ωj1c′ 1] = [c′ 1 v0] = [ac′ 1 1¯c′ 1 1 1c′′ [(0, 1, 0, ..., 0)(0, 0, ..., 0, 2)] = [(0, 1, 0, ..., 0, 2)] + [v0] we must have Sλ(0,1,0,...,0,2) 6= 0. By Lemma 2.20 of [39] we must have τ ((0, 1, 0, ..., 0, 2)) = (0, 1, 0, ..., 0, 2) = (2, 0, 0, ..., 1, 0), a contradiction. So we conclude that τ (µ) = ωmcol(µ)µ,∀µ. It follows that [aµ] = [aωmcol(µ)aµ], and in particular [av] = [aωmav]. So we have [ωmvc1] = [c1av] = [c1av] = [vc1], and similarly [c2ω−m¯v] = [c2¯v]. ωm 6≺ c1¯c1, ωm 6≺ ¯c2c2. On the other hand we have If [ωm] 6= [1], by our assumption on λ we have h¯vωmv, c1¯c1i ≥ 1,h¯vωmv, ¯c2c2i ≥ 1 It follows that ωmv0 ≺ c1¯c1, ωmv0 ≺ ¯c2c2, and hc1¯c1, ¯c2c2i ≥ 2. By Lemma 4.1 we conclude that λ = c1c2 is not irreducible, contradicting our assumption. Hence [ωm] = 20 [1] and Zµ1µ2 = δµ1µ2. The rest of the proof now follows in exactly the same way as in the proof of Prop. 5.20 of [39]. maximal intermediate subfactor, and c1 = c′ c1 = c′ Ad (2): As in the proof of (1) we assume that λ = c1c2 with c1(M) a nontrivial 1. By our assumption we must have 1] 6= [1]. In this case as in the proof of (1) above we must have [c1¯c1] = [ωjj1]. µγ = δµ,ωµcol(γ)γ. By Corollary 1] = [1]. Then as above we have Z c1 Now suppose that [c′ 1 if [c′ P1≤j≤n/j1 3.14 of [39] we can find c ≺ µc1 for some µ such that ωli, pl = n. 1c′′ [c¯c] = X1≤i≤p Since Z c = Z c1, it follows the left local support of c is trivial. If [ωm] = [1], then we are as in the end of proof of (1), and in that case [c1] = [1], contradicting our assumption that c1(M) a nontrivial maximal intermediate subfactor. So [ωm] 6= [1]. If ωm ≺ c1¯c1, then by maximality of c(M) we have [c1¯c1] = Pi[ωqi], i.e., c1(M) comes from abelian part of λ. Now assume that ωm 6≺ c1¯c1, then as in (1) we must have ωm ≺ ¯c2c2. Since c ≺ µc1 for some µ, we have c¯c ≺ µc1¯c1 ¯µ ≺ µλ¯λ¯µ, so col(ωli) = 0 mod n, so we have nli. Similarly since ωm ≺ ¯c2c2, col(ωm) = 0 mod n. On the other hand since the map µ → ωmcol(µ)µ has order two, it follows that ω2m = 1. So we must have n = 2m. From hωli = kli n n − li , 2 it follows that hωli ∈ Z if i is even. These ωli with i even will generate local simple currents (cf Definition 2.3 and Prop. 2.15 of [36]), and it follows that the left local support of c is nontrivial if li 6= 0modn for some even i. So we conclude that 2l = 0modn, and [c¯c] = [1] + [ωm]. Note that there are λ1, λ2 such that c1 ≺ λ1c, ¯c2 ∈ λ2c. From hc1, λ1ci = hc1¯c, λ1i ≥ 1, we have dc1√2 ≥ dλ1, and similarly d¯c2√2 ≥ dλ2. Since [ωmc1] 6= [c1], we have [λ1c] ≻ [c1] + [ωmc1], and by computing statistical Similarly if λ2c is not irreducible, we must have [¯c2¯c] = [λ2]. But we have dimension [λ1c] = [c1] + [ωmc1], and [c1¯c] = [λ1]. h¯c2¯c, ¯c2¯ci = h¯c2¯ccc2, 1i = h¯c2ac¯cc2, 1i = hc¯c¯c2c2, 1i ≥ 2, where we have used [¯cc] = [ac¯c] and ¯c2c2 ≻ [1] + [ωm]. From this we conclude that [¯c2] = [λ2c]. We have [λ] = [c1c2] = [c1¯c¯λ2] = [λ1¯λ2]. By Lemma 2.14 , we must have ¯λ2 = ωi, [λ] = [c1¯cωi]. Now we check that the intermediate subfactor c1(M) is uniquely fixed. Suppose there is an intermediate subfactor f1(M) such that [λ] = [f1f2] and [ ¯f2f2] = [1] + [ωm]. Then ¯c ¯f2 has statistical dimension two and decompose into sum of two irreducible en- domorphisms, it follows that there is an automorphism α such that [f2] = [α¯cωi], and 21 [f1α] = [c1β] for some automorphism β ≺ ¯cc. By Cor. 2.4 of of [37] the intermediate subfactor f1(M) is determined by equivalence class [f1, f2] with equivalence relation [f1, f2] ∼ [f1ρ, ρ−1f2] where ρ is any automorphism. We have [f1, f2] = [c1βα−1, α¯cωi] ∼ [c1β, ¯cωi] ∼ [c1, β−1¯cωi] = [c1, ¯cωi]. Corollary 4.3. Suppose that k 6= n− 2, n + 2, n. Then each irreducible representation λ of ASU (n)k verifies both maximal and minimal version of Conjecture 1.1. Proof. Since the dual of λ is ¯λ, it is sufficient to verify the maximal version of Con- jecture 1.1. We may assume that dλ > 1. By Lemma 2.14 [λ][¯λ] = X1≤i≤p [ωiq] + [v0] + ... where pq = n, and ... are possible additional irreps. We note that the set of maximal intermediate subfactors for λ coming the abelian part are bounded by p − 1, and by Th. 4.2, there is at most one more maximal intermediate subfactor, and our corollary follows. (cid:4) Remark 4.4. It will be interesting to remove the condition k 6= n, n ± 2 in Th.4.2 and Cor. 4.3. This condition is used in the proof of Th. 4.2 to ensure that ac1 v0 is irreducible. One can remove this condition if one can find a different way of proving that ac1 v0 is irreducible. References [1] D. Altschuler, M. Bauer and C. Itzykson, The branching rules of conformal embeddings, Comm. Math. Phys., 132 (1990), 349-364. [2] J. Bockenhauer, D. E. Evans, Modular invariants, graphs and α-induction for nets of subfactors. I., Comm.Math.Phys., 197, 361-386, 1998. [3] J. Bockenhauer, D. E. Evans, Y. Kawahigashi, Chiral structure of modular invariants for sub- factors, Comm.Math.Phys., 210, 733-784, 2000. [4] D. Bisch and V. F. R. Jones Algebras associated to intermediate subfactors, Invent. Math. 128 (1997), no. 1, 89 -- 157. [5] C. Dong, Introduction to vertex operator algebras I, S ¯urikaisekikenky ¯usho K ¯oky ¯uroku, No. 904 (1995), 1-25. Also see q-alg/9504017. [6] C. Dong and J. Lepowsky, Generalized vertex algebras and relative vertex operators, Progress in Mathematics, 112 (1993). [7] J. Frohlich and F. Gabbiani, Operator algebras and Conformal field theory, Comm. Math. Phys., 155, 569-640 (1993). [8] I. B. Frenkel, J. Lepowsky and J. Ries, Vertex operator algebras and the Monster, Academic, New York, 1988. [9] I. Frenkel and Y. Zhu, Vertex operator algebras associated to representations of affine and Virasoro algebras, Duke Math. Journal (1992), Vol. 66, No. 1, 123-168. 22 [10] T. Gannon, P. Ruelle, and M. A. Walton, Automorphism modular invariants of current algebras, Comm. Math. Phys. 179 (1996), no. 1, 121 -- 156. [11] F. Goodman and H. Wenzl, Littlewood-Richardson coefficients for Hecke algebras at roots of unity, Adv. Math. 82 (1990), no. 2, 244 -- 265. [12] P. Grossman and Vaughan F. R. Jones, Intermediate subfactors with no extra structure, J. Amer. Math. Soc. 20 (2007), no. 1, 219 -- 265. [13] D. Guido & R. Longo, The conformal spin and statistics theorem, Commun. Math. Phys. 181 (1996) 11 -- 35. [14] R. Guralnick and F. Xu, On a subfactor generalization of Wall's conjecture. J. Algebra 332 (2011), 457468 [15] V. F. R. Jones, Fusion en algbres de von Neumann et groupes de lacets (d'aprs A. Wassermann). (French) [Fusion in von Neumann algebras and loop groups (after A. Wassermann)], Sminaire Bourbaki, Vol. 1994/95. Astrisque No. 237 (1996), Exp. No. 800, 5, 251 -- 273. [16] V. G. Kac, Vertex algebras for beginners, AMS, 1997. [17] V. G. Kac, "Infinite Dimensional Lie Algebras", 3rd Edition, Cambridge University Press, 1990. [18] V. G. Kac, R. Longo and F. Xu, Solitons in affine and permutation orbifolds, Comm. Math. Phys. 253 (2005), no. 3, 723 -- 764. [19] A. Tsuchiya and Y. Kanie, Vertex Operators in conformal field theory on P 1 and monodromy representations of braid group, Adv. Studies in Pure Math. 16 (88), 297-372. [20] V. G. Kac and M. Wakimoto, Modular and conformal invariance constraints in representation theory of affine algebras, Advances in Math., 70, 156-234 (1988). [21] F. Levstein and J. I. Liberati, Branching rules for conformal embeddings, Commun. Math. Phys. 173 (1995), 1 -- 16. [22] R. Longo, Conformal subnets and intermediate subfactors, Commun. Math. Phys. 237 n. 1-2 (2003), 7 -- 30. [23] R. Longo & K.-H. Rehren, Nets of subfactors, Rev. Math. Phys. 7 (1995) 567 -- 597. [24] P. Goddard, W. Nahm and D. Olive, Symmetric spaces, Sugawara's energy momentum tensor in two dimensions and free fermions, Phys. Lett. B 160 (1985), no. 1-3, 111 -- 116 [25] M. Izumi, R. Longo & S. Popa, A Galois correspondence for compact groups of automorphisms of von Neumann Algebras with a generalization to Kac algebras, J. Funct. Analysis, 155, 25-63 (1998). [26] V. F. R. Jones, Index for subfactors, Invent. Math. 72 (1983) 1 -- 25. [27] Vaughan F. R. Jones and F. Xu, Intersections of finite families of finite index subfactors, Internat. J. Math. 15 (2004), no. 7, 717 -- 733. [28] M. W. Liebeck, L. Pyber and A. Shalev, On a conjecture of G. E. Wall, J. Algebra 317 (2007), no. 1, 184 -- 197. [29] M. Pimsner, & S. Popa, Entropy and index for subfactors, Ann. Scient. Ec. Norm. Sup. 19 (1986), 57 -- 106. [30] A. Pressley and G. Segal, "Loop Groups" Oxford University Press 1986. [31] G. E. Wall, Some applications of the Eulerian functions of a finite group, J. Austral. Math. Soc. 2 1961/1962 35 -- 59. 23 [32] A. Wassermann, Operator algebras and Conformal field theories III, Invent. Math. 133 (1998), 467-538. [33] F. Xu, New braided endomorphisms from conformal inclusions, Commun. Math. Phys. 192 (1998) 347 -- 403. [34] F. Xu, Algebraic coset conformal field theories. Comm. Math. Phys. 211 (2000), no. 1, 1-43. [35] F. Xu, Mirror extensions of local nets. Comm. Math. Phys. 270 (2007), no. 3, 835-847. [36] F. Xu, An application of mirror extensions. Comm. Math. Phys. 290 (2009), no. 1, 83-103. [37] F. Xu, On intermediate subfactors of Goodman-de la Harpe-Jones subfactors, Commun. Math. Phys. 298 (2010), no.3, 707 -- 739. [38] F. Xu, Some computations in the cyclic permutations of completely rational nets, Comm. Math. Phys. 267 (2006), no. 3, 757-782. [39] F. Xu, On representing some lattices as lattices of intermediate subfactors of finite index , Adv. Math. 220 (2009), no. 5, 1317-1356. Corrections in the proof of Cor. 5.23 in arXiv:math/0703248. 24
1208.0096	1	1208	2012-08-01T04:22:46	Automatic continuity of derivations on C-algebras and JB-triples	[ "math.OA" ]	We introduce the notion of a Jordan triple module and determine the precise conditions under which every derivation from a JB-triple E into a Banach (Jordan) triple E-module is continuous. In particular, every derivation from a real or complex JB-triple into its dual space is automatically continuous. Among the consequences, we prove that every triple derivation from a C*-algebra A to a Banach triple A-module is continuous. In particular, every Jordan derivation from A to a Banach A-bimodule is a derivation, a result which complements a classical theorem due to B.E. Johnson and solves a problem which has remained open for over ten years.	math.OA	math	AUTOMATIC CONTINUITY OF DERIVATIONS ON C∗-ALGEBRAS AND JB∗-TRIPLES ANTONIO M. PERALTA AND BERNARD RUSSO Abstract. We introduce the notion of a Jordan triple module and determine the precise conditions under which every derivation from a JB∗-triple E into a Banach (Jordan) triple E-module is continuous. In particular, every derivation from a real or complex JB∗-triple into its dual space is automatically continuous. Among the consequences, we prove that every triple derivation from a C∗-algebra A to a Banach triple A-module is continuous. In particular, every Jordan derivation from A to a Banach A-bimodule is a derivation, a result which complements a classical theorem due to B.E. Johnson and solves a problem which has remained open for over ten years. 1. Introduction Results on automatic continuity of linear operators defined on Banach algebras comprise a fruitful area of research intensively developed during the last sixty years. The monographs [48], [14] and [16] review most of the main achievements obtained during the last fifty years. In the words of A.M. Sinclair (see [48, Introduction]), "the continuity of a multiplicative linear functional on a unital Banach algebra is the seed from which these results on the automatic continuity of homomorphisms grew". A linear mapping D from a Banach algebra A to a Banach A-bimodule is said to be a derivation if D(ab) = D(a)b + aD(b), for every a, b in A. The pioneering work of W. G. Bade and P. C. Curtis (see [2]) studies the automatic continuity of a module homomorphism between bi-modules over C(K)-spaces. Some techniques developed in the just quoted paper were exploited by J.R. Ringrose to prove that every (associative) derivation from a C∗-algebra A to a Banach A-bimodule M is continuous (compare [43]). The case in which M = A was previously treated by S. Sakai [45] by way of spectral theory in A (= M ). A Jordan derivation from a Banach algebra A into a Banach A-module is a linear map D satisfying D(a2) = aD(a) + D(a)a, (a ∈ A), or equivalently, D(ab+ba) = aD(b)+D(b)a+D(a)b+bD(a), (a, b ∈ A). Sinclair proved that a bounded Jordan derivation from a semisimple Banach algebra to itself is a derivation, although this result fails for derivations of semisimple Banach First author partially supported by D.G.I. project no. MTM2008-02186, and Junta de Andaluc´ıa grants FQM0199 and FQM3737. 2 PERALTA AND RUSSO algebras into a Banach bi-module [46, Theorem 3.3]. Nevertheless, a cele- brated result of B.E. Johnson states that every bounded Jordan derivation from a C∗-algebra A to a Banach A-bimodule is an associative derivation (cf. [31]). In view of the intense interest in automatic continuity problems in the past half century, it is natural to ask if the assumption of boundedness is needed in Johnson's result. It is therefore somewhat surprising that the following problem has remained open for fifteen years. Problem 1. Is every Jordan derivation from a C∗-algebra A to a Banach A-bimodule automatically continuous? This problem was already posed in [51, Question 14.i]. According to [6, §5], Problem 1 "is an intriguing open question". In 2004, J. Alaminos, M. Bresar and A.R. Villena gave a positive answer to the above problem for some classes of C∗-algebras including the class of von Neumann algebras and the class of abelian C∗-algebras (cf. In the setting of general C∗-algebras the question has remained open. [1]). Problem 1 has a natural generalization to the setting of Banach Jordan algebras. In the category of JB∗-algebras, S. Hejazian and A. Niknam es- tablished in [25] that every Jordan derivation from a JB∗-algebra J into J or into J ∗ is automatically continuous. We recall that a linear mapping D from a JB∗-algebra J to a Jordan Banach J-bimodule is said to be a Jordan derivation if D(a ◦ b) = D(a) ◦ b + a ◦ D(b), for every a, b in J, where ◦ de- notes the Jordan product in J and the action of J on the Jordan J-module (defined below). The above quoted paper actually contains a theorem which provides nec- essary and sufficient conditions to guarantee that a Jordan derivation from a JB∗-algebra J into a Jordan Banach J-module is continuous (cf. [25, The- orem 2.2]). The same authors show the existence of discontinuous Jordan derivations from JB∗-algebras into Jordan Banach modules (compare [25, §3]). When the domain JB∗-algebra is a commutative or a compact C∗- algebra A, the same authors proved that every Jordan derivation from A into a Jordan Banach A-module is continuous (cf. [25, Theorem 2.4 and Corollary 2.7]). In the setting of general C∗-algebras, however, the following question remains open (also for fifteen years). Problem 2. Is every Jordan derivation from a C∗-algebra A to a Jordan Banach module automatically continuous? Prior to the writing of this paper, it apparently had escaped the atten- tion of functional analysts that combining a theorem of Cuntz ([13], see Lemma 19 below) with the theorems just quoted from [1] and [25] concern- ing commutative C∗-algebras yields positive answers to both Problems 1 and 2. We therefore can now state the following theorem. Theorem 3. Every Jordan derivation from a C∗-algebra A to a Banach A-module or to a Jordan Banach module is continuous. AUTOMATIC CONTINUITY OF DERIVATIONS 3 As a consequence of our main results, we are able to treat both Problems 1 and 2 from a new and more general point of view. We introduce the class of Banach (Jordan) triple modules, a class which includes, besides Banach modules over Banach algebras and Banach Jordan modules over Banach Jordan algebras, the dual space of every real or complex JB∗-triple. In this setting, a conjugate linear (resp., linear) mapping δ from a complex (resp., real) Jordan triple E to a triple E-module is called a derivation if (1) δ{a, b, c} = {δ(a), b, c} + {a, δ(b), c} + {a, b, δ(c)} , for every a, b, c ∈ E. We determine (in Theorem 13) the precise conditions in order that a derivation from a complex JB∗-triple, E, into a Banach (Jordan) triple E- module is continuous. We subsequently show that every derivation from a real or complex JB∗-triple into its dual space is automatically continuous, a fact which has significance for the forthcoming study by the authors of weak amenability. From one point of view (another is through infinite dimensional holomor- phy) the theory of JB∗-triples may be viewed as parallel to the theory of C∗-algebras. The analog of the theorem of Sakai mentioned above, namely, the automatic continuity of a derivation from a JB∗-triple into itself, that is, a linear map satisfying the derivation property (1), was proved by T. J. Barton and Y. Friedman [3] in the complex case and extended to the real case in [27]. Among the consequences of our main results, we obtain a com- pletely different proof for the automatic continuity results obtained in the just quoted papers [3] and [27]. We shall see that there exist examples of triple derivations from a JB∗- triple E to a Banach triple E-module which are not continuous (see Remark 16). In our last results we show that these examples cannot appear when the domain is a C∗-algebra. More concretely, in Theorem 20 and Corollaries 21, 22, and 23 we prove that every triple (resp., Jordan) derivation from a C∗-algebra A to a Banach triple A-module (resp., to a Jordan Banach A-module) is automatically continuous, which constitute the solutions to Problems 1 and 2 and a completely different proof of the automatic conti- nuity result of Ringrose quoted above. In section 2 of this paper we recall the definition and basic properties of Jordan triples, define Jordan triple modules and submodules, and intro- duce and study a basic tool in our paper: the quadratic annihilator of a submodule. In section 3 we prove the automatic continuity results by relat- ing triple derivations to triple module homomorphisms and using the well known technique of separating spaces. The final section contains an analysis of the automatic continuity of every triple derivation from a C∗-algebra A to a Banach triple A-module, which leads to a unified solution to Problems 1 and 2. 4 PERALTA AND RUSSO Our definition of Jordan triple module is motivated by the theory of mod- ules over a Jordan algebra due to Jacobson [30], together with the definition in the special case of the dual of a Banach Jordan triple, which was suggested by Tom Barton some time ago. Subsequently, we noticed that Jordan triple modules were defined in [37] in a form more suitable to a purely algebraic setting. Our definition is more suitable for the applications to C∗-algebras. All of our results, excepting Theorem 13, are valid for real or complex JB∗-triples. It should be noted however that the key to the solutions to Problems 1 and 2 is that Theorem 13 is valid for the self-adjoint part of a C∗-algebra, considered as a (reduced) real JB∗-triple (see Proposition 17). 2. Jordan triple Modules 2.1. Jordan triples. A complex (resp., real) Jordan triple is a complex (resp., real) vector space E equipped with a triple product E × E × E → E (xyz) 7→ {x, y, z} which is bilinear and symmetric in the outer variables and conjugate linear (resp., linear) in the middle one satisfying the so-called "Jordan Identity": L(a, b)L(x, y) − L(x, y)L(a, b) = L(L(a, b)x, y) − L(x, L(b, a)y), for all a, b, x, y in E, where L(x, y)z := {x, y, z}. When E is a normed space and the triple product of E is continuous, we say that E is a normed Jordan triple. If a normed Jordan triple E is complete with respect to the norm (i.e. if E is a Banach space), then it is called a Jordan-Banach triple. Unless otherwise specified, the term "normed Jordan triple" (resp., "Jordan-Banach triple") will always mean a real or complex normed Jordan triple (resp., a real or complex Jordan-Banach triple). A summary of the basic facts about the important subclass of JB∗-triples (defined below), some of which are recalled here, can be found in [44] and some of the references therein, such as [34],[21],[22],[49] and [50]. A subspace F of a Jordan triple E is said to be a subtriple if {F, F, F } ⊆ F . We recall that a subspace J of E is said to be a triple ideal if {E, E, J} + {E, J, E} ⊆ J. When {J, E, J} ⊂ J we say that J is an inner ideal of E. We recall that a real (resp., complex) Jordan algebra is a (not-necessarily associative) algebra over the real (resp., complex) field whose product is abelian and satisfies (a ◦ b) ◦ a2 = a ◦ (b ◦ a2). A normed Jordan algebra is a Jordan algebra A equipped with a norm, k.k, satisfying ka ◦ bk ≤ kak kbk, a, b ∈ A. A Jordan Banach algebra is a normed Jordan algebra whose norm is complete. Every Jordan algebra is a Jordan triple with respect to {a, b, c} := (a ◦ b) ◦ c + (c ◦ b) ◦ a − (a ◦ c) ◦ b. AUTOMATIC CONTINUITY OF DERIVATIONS 5 Every real or complex associative Banach algebra (resp., Jordan Banach algebra) is a real Jordan-Banach triple with respect to the product {a, b, c} = 1 2 (abc + cba) (resp., {a, b, c} = (a ◦ b) ◦ c + (c ◦ b) ◦ a − (a ◦ c) ◦ b). An element e in a Jordan triple E is called tripotent if {e, e, e} = e. Each tripotent e in E induces two decomposition of E (called Peirce decomposi- tions) in the form: E = E0(e) ⊕ E1(e) ⊕ E2(e) = E1(e) ⊕ E−1(e) ⊕ E0(e) where Ek(e) = {x ∈ E : L(e, e)x = k 2 x} for k = 0, 1, 2 and Ek(e) is the k- eigenspace of the operator Q(e)x = {e, x, e} for k = 1, −1, 0. The projection onto Ek(e), which is contractive, is denoted by Pk(e) for k = 0, 1, 2. The following Peirce rules are satisfied: (a) E2(e) = E1(e) ⊕ E−1(e) and E0(e) = E1(e) ⊕ E0(e), (b) {Ei(e), Ej (e), Ek(e)} ⊆ Eijk(e) if ijk 6= 0, (c) {Ei(e), Ej (e), Ek (e)} ⊆ Ei−j+k(e), where i, j, k = 0, 1, 2 and El(e) = 0 for l 6= 0, 1, 2, (d) {E0(e), E2(e), E} = {E2(e), E0(e), E} = 0. A JB∗-algebra is a complex Jordan Banach algebra A equipped with an (Recall that algebra involution ∗ satisfying k {a, a∗, a} k = kak3, a ∈ A. {a, a∗, a} = 2(a ◦ a∗) ◦ a − a2 ◦ a∗). A (complex) JB∗-triple is a complex Jordan Banach triple E satisfying the following axioms: (JB∗1) For each a in E the map L(a, a) is an hermitian operator on E with non negative spectrum. (JB∗2) k{a, a, a}k = kak3 for all a in A. Every C∗-algebra (resp., every JB∗-algebra) is a JB∗-triple with respect 2 (ab∗c + cb∗a) (resp., {a, b, c} := (a ◦ b∗) ◦ c + (c ◦ to the product {a, b, c} = 1 b∗) ◦ a − (a ◦ c) ◦ b∗). We recall that a real JB∗-triple is a norm-closed real subtriple of a complex JB∗-triple (compare [29]). The class of real JB∗-triples includes all complex JB∗-triples, all real and complex C∗- and JB∗-algebras and all JB-algebras. A complex (resp., real) JBW∗-triple is a complex (resp., real) JB∗-triple which is also a dual Banach space (with a unique isometric predual [4, 39]). It is known that the triple product of a JBW∗-triple is separately weak∗ continuous (c.f. [4] and [39]). The second dual of a JB∗-triple E is a JBW∗- triple with a product extending the product of E [17, 29]. It is also known that, for each tripotent e in a complex JB∗-triple E, E2(e) is a JB∗-algebra with product and involution given by x ◦e y := {x, e, y} and x♯e := {e, x, e}, respectively. In the case of E being a real JB∗-triple E1(e) is a JB-algebra with respect to the product given in the above lines (JB- algebras are precisely the self adjoint parts of JB∗-algebras [52]). 6 PERALTA AND RUSSO A tripotent e in a real or complex JB∗-triple E is called minimal if E1(e) = Re. In the complex setting this is equivalent to say that E2(e) = Ce, because E−1(e) = iE1(e), whereas in the real situation the dimensions of E1(e) and E−1(e) need not be correlated. When E is a JB∗-triple or a real JB∗-triple, a subtriple I of E is a triple ideal if and only if {E, E, I} ⊆ I or {E, I, E} ⊆ I or {E, I, I} ⊆ I (compare [7]). 2.2. Jordan triple modules. Let A be an associative algebra. Let us recall that an A-bimodule is a vector space X, equipped with two bilinear products (a, x) 7→ ax and (a, x) 7→ xa from A × X to X satisfying the following axioms: a(bx) = (ab)x, a(xb) = (ax)b, and, (xa)b = x(ab), for every a, b ∈ A and x ∈ X. Let J be a Jordan algebra. A Jordan J-module is a vector space X, equipped with two bilinear products (a, x) 7→ a ◦ x and (x, a) 7→ x ◦ a from J × X to X, satisfying: a ◦ x = x ◦ a, a2 ◦ (x ◦ a) = (a2 ◦ x) ◦ a, and, 2((x ◦ a) ◦ b) ◦ a + x ◦ (a2 ◦ b) = 2(x ◦ a) ◦ (a ◦ b) + (x ◦ b) ◦ a2, for every a, b ∈ J and x ∈ X (see [30, §II.5,p.82]). Let E be a complex (resp. real) Jordan triple. A Jordan triple E-module (also called triple E-module) is a vector space X equipped with three map- pings {., ., .}1 : X × E × E → X, {., ., .}2 : E × X × E → X and {., ., .}3 : E × E × X → X satisfying the following axioms: (JT M 1) {x, a, b}1 is linear in a and x and conjugate linear in b (resp., trilinear), {abx}3 is linear in b and x and conjugate linear in a (resp., trilinear) and {a, x, b}2 is conjugate linear in a, b, x (resp., trilinear) (JT M 2) {x, b, a}1 = {a, b, x}3, and {a, x, b}2 = {b, x, a}2 for every a, b ∈ E and x ∈ X. (JT M 3) Denoting by {., ., .} any of the products {., ., .}1, {., ., .}2 and {., ., .}3, the identity {a, b, {c, d, e}} = {{a, b, c} , d, e} − {c, {b, a, d} , e} + {c, d, {a, b, e}} , holds whenever one of the elements a, b, c, d, e is in X and the rest are in E. When E is a Jordan Banach triple and X is a triple E-module which is also a Banach space and, for each a, b in E, the mappings x 7→ {a, b, x}3 and x 7→ {a, x, b}2 are continuous, we shall say that X is a triple E-module with continuous module operations. When the products {., ., .}1, {., ., .}2 and {., ., .}3 are (jointly) continuous we shall say that X is a Banach (Jordan) triple E-module. AUTOMATIC CONTINUITY OF DERIVATIONS 7 It is obvious that every real or complex Jordan triple E is a real triple E-module. Actually, every triple ideal J of E is a (real) triple E-module. It is problematical whether every complex Jordan triple E is a complex triple E-module for a suitable triple product. We shall see later that triple modules have a priori a different behavior than bi-modules over associative algebras and Jordan modules (see Remark 16). Every real or complex associative algebra A (resp., Jordan algebra J) is a real Jordan triple with respect to {a, b, c} := 1 2 (abc + cba), a, b, c ∈ A (resp., {a, b, c} = (a ◦ b) ◦ c + (c ◦ b) ◦ a − (a ◦ c) ◦ b) , a, b, c ∈ J). It is not hard to see that every A-bimodule X is a real triple A-module with respect to the products {a, b, x}3 := 1 2 (axb + bxa), and that every Jordan module X over a Jordan algebra J is a real triple J-module with respect to the products {a, b, x}3 := (a ◦ b) ◦ x + (x ◦ b) ◦ a − (a ◦ x) ◦ b and {a, x, b}2 = (a ◦ x) ◦ b + (b ◦ x) ◦ a − (a ◦ b) ◦ x. 2 (abx + xba) and {a, x, b}2 = 1 Hereafter, the triple products {·, ·, ·}j, j = 1, 2, 3, which occur in the definition of Jordan triple module will be denoted simply by {·, ·, ·} whenever the meaning is clear from the context. It is a little bit more laborious to check that the dual space, E∗, of a complex (resp., real) Jordan Banach triple E is a complex (resp., real) triple E-module with respect to the products: (2) and (3) {a, b, ϕ} (x) = {ϕ, b, a} (x) := ϕ {b, a, x} {a, ϕ, b} (x) := ϕ {a, x, b}, ∀ϕ ∈ E∗, a, b, x ∈ E. Given a triple E-module X over a Jordan triple E, the space E ⊕ X can be equipped with a structure of real Jordan triple with respect to the prod- uct {a1 + x1, a2 + x2, a3 + x3} = {a1, a2, a3} + {x1, a2, a3} + {a1, x2, a3} + {a1, a2, x3}. Consistent with the terminology in [30, §II.5], E ⊕ X will be called the triple split null extension of E and X. A subspace S of a triple E-module X is said to be a Jordan triple submod- ule or a triple submodule if and only if {E, E, S} ⊆ S and {E, S, E} ⊆ S. Every triple ideal J of E is a Jordan triple E-submodule of E. 2.3. Quadratic annihilator. Given an element a in a Jordan triple E, we shall denote by Q(a) the conjugate linear operator on E defined by Q(a)(b) := {a, b, a} . The following formula is always satisfied Q(a)Q(b)Q(a) = Q(Q(a)b), (a, b ∈ E). and remains true for Q(·) acting on a triple E-module X: (4) {a, {b, {a, x, a} , b} , a} = {{a, b, a} , x, {a, b, a}} , x ∈ X. For each submodule S of a triple E-module X, we define its quadratic annihilator, AnnE(S), as the set {a ∈ E : Q(a)(S) = {a, S, a} = 0}. Since 8 PERALTA AND RUSSO S is triple submodule of X, it follows by (4) that (5) and (6) {a, E, a} ⊂ AnnE(S), ∀a ∈ AnnE(S) {b, AnnE(S), b} ⊆ AnnE(S), ∀b ∈ E. Consequently, AnnE(S) is an inner ideal of E whenever it is a linear subspace of E. Further, AnnE(S) is a triple ideal of E whenever E is a JB∗-triple and AnnE(S) is a linear subspace of E, since as noted earlier, for JB∗-triples, (6) implies {E, AnnE(S), E} ⊂ AnnE(S). Let E be a Jordan triple. Two elements a and b in E are said to be orthogonal (written a ⊥ b) if L(a, b) = L(b, a) = 0. A direct application of the Jordan identity yields that, for each c in E, (7) a ⊥ {b, c, b} whenever a ⊥ b. Given an element a in a Jordan triple E, we denote a[1] = a, a[3] = {a, a, a} The element a is called nilpotent if a[2n+1] = 0 for some n. Jordan triples and a[2n+1] := (cid:8)a, a[2n−1], a(cid:9) (∀n ∈ N). The Jordan identity implies that a[5] = (cid:8)a, a, a[3](cid:9) , and by induction, a[2n+1] = L(a, a)n(a) for all n ∈ N. are power associative, that is,(cid:8)a[k], a[l], a[m](cid:9) = a[k+l+m]. A Jordan triple E for which the vanishing of {a, a, a} implies that a itself vanishes is said to be anisotropic. It is easy to check that E is anisotropic if and only if zero is the unique nilpotent element in E. Let a and b be two elements in a Jordan triple E. If L(a, b) = 0, then, for each c in E, the Jordan identity implies that {L(b, a)c, L(b, a)c, L(b, a)c} = 0. Therefore, in an anisotropic Jordan triple, a ⊥ b if and only if L(a, b) = 0. Let a be an element in a real (resp., complex) JB∗-triple E. Denoting by Ea the JB∗-subtriple generated by the element a, it is known that Ea is JB∗-triple isomorphic (and hence isometric) to C0(L) = C0(L, R) (resp., C0(L) = C0(L, C)) for some locally compact Hausdorff space L ⊆ (0, kak], such that L ∪ {0} is compact. It is also known that denoting by Ψ the triple isomorphism from Ea onto C0(L), then Ψ(a)(t) = t (t ∈ L) (compare [34, Lemma 1.14], [35, Proposition 3.5] or [11, Page 14]). The set L is called the triple spectrum of a. It should be noticed here that, in the setting of real or complex JB∗-triples orthogonality is a "local concept"(compare Lemma 1 in [10], whose proof remains valid for real JB∗-triples). Indeed, two elements a and b in a real JB∗-triple E are orthogonal if and only if one of the following equivalent statements holds: (a) {a, a, b} = 0, (b) Ea ⊥ Eb, (c) {b, b, a} = 0, (d) a ⊥ b in a subtriple of E containing both elements. AUTOMATIC CONTINUITY OF DERIVATIONS 9 Let E be a (real or complex) Jordan Banach triple. We have already mentioned that E∗ is a triple module with respect to the products given in (2) and (3). The triple module structure of E∗ satisfies the following additional property: given a and b in E with a ⊥ b (in E), we have {a, b, ϕ} = {ϕ, b, a} = 0 for every ϕ ∈ E∗. That is, a ⊥ b in the Jordan triple E ⊕ E∗. Orthogonal elements in E lift to orthogonal elements in the split null extension E ⊕ E∗. Let X be a triple module over a Jordan triple E. We shall say that X has the property of lifting orthogonality (LOP in short) if {a, b, x} = 0, for every x ∈ X, a, b ∈ E with a ⊥ b. We have just remarked that for every Jordan Banach triple E, E∗ is a triple E-module satisfying LOP. When a Jordan triple E is regarded as a real triple E-module with its natural products, then E also has LOP. However, not every triple module has this property. Let A be a C∗-algebra regarded as a complex JB∗-triple with respect to {a, b, c} := 1 2 (ab∗c + cb∗a). As noted earlier, the vector space X = A is a real triple A-module with respect to the products {a, b, x}3 := 1 2 (axb + bxa). Two elements a and b in a real or complex C∗-algebra A are orthogonal if and only if ab∗ = b∗a = 0 or equivalently, in the triple sense, aa∗b + ba∗a = 0 or bb∗a + ab∗b = 0 (compare [10, Lemma 1]). It is not hard to find a C∗- algebra A containing two orthogonal elements a, b with {a, b, x}3 6= 0 for some x ∈ A. 2 (abx + xba) and {a, x, b}2 := 1 Let J be a norm-closed subspace of a JB∗-triple (resp., a real JB∗-triple) E. Clearly, J is a triple ideal of E if and only if J is a triple E-submodule of E. Let J be a triple ideal of E regarded as a Jordan triple E-submodule. We clearly have AnnE (J) := {a ∈ E : Q(a)(J) = 0} ⊇ J ⊥ := {a ∈ E : a ⊥ J}. Suppose now that a ∈ AnnE (J). Replacing J with its weak∗-closure in E∗∗, we may assume that E is a JBW∗-triple, J is a weak∗-closed triple ideal and Q(a)(J) = 0. By [28, Theorem 4.2 (4)], there exists a weak∗- closed triple ideal K in E such that E = J ⊕ K and J ⊥ K. Writing a = a1 + a2 with a1 ∈ J and a2 ∈ K, we deduce, by orthogonality, that a[3] 1 = Q(a)(a1) ∈ Q(a)(J) = 0, and hence a = a2 ⊥ J. We state this as a Lemma. Lemma 4. Let E be a JB∗-triple (resp., a real JB∗-triple). For each triple ideal J in E we have AnnE (J) = J ⊥ is a norm closed triple ideal of E. (cid:3) Let E be a JB∗-triple (resp., a real JB∗-triple). For each x in E, E(x) will denote the norm-closure of {x, E, x} in E. It is known that E(x) coincides with the norm-closed inner ideal of E generated by x and Ex ⊆ E(x) (see [9]). By [9, Proposition 2.1], E(x) is a JB∗-subalgebra of the JBW∗-algebra E(x)∗∗ = E(x) 2 (r(x)), where r(x) is the (so called) range tripotent of x in E∗∗. It is also known that x ∈ E(x)+. = E∗∗ w∗ 10 PERALTA AND RUSSO For each functional ϕ ∈ E∗, there exists a unique tripotent s = s(ϕ) in E∗∗ satisfying that ϕ = ϕP2(s) and ϕE ∗∗ 2 (s) is a faithful normal positive functional on E∗∗ 2 (s) (compare [21, Proposition 2] and [39, Lemma 2.9] and [40, Lemma 2.7], respectively). The tripotent s(ϕ) is called the support tripotent of ϕ in E∗∗. Proposition 5. Let E be a JB∗-triple (resp., a real JB∗-triple). For each triple submodule S ⊂ E∗, (a) the quadratic annihilator AnnE (S) is a norm closed triple ideal of E, (b) AnnE (S) = E ∩(cid:16)Tϕ∈S E∗∗ 0 (s(ϕ))(cid:17), (c) {AnnE (S), AnnE (S), S} = 0 in the triple split null extension E ⊕E∗. Proof. We prove (b) first. For each a ∈ AnnE (S) and each ϕ ∈ S, we have by definition, {a, ϕ, a} = 0 and hence ϕQ(a)(E) = 0. It follows that E(a) ⊆ ker(ϕ) for every ϕ ∈ S, a ∈ AnnE (S). In particular, ϕ(a) = 0. Since S is a triple submodule, for every b ∈ E, {ϕ, b, a} ∈ S, so {ϕ, b, a}(a) = 0, that is, ϕ {a, a, b} = 0. Fix ϕ ∈ S. We have already seen that ϕ {a, a, b} = 0 for every b ∈ E. Since E is weak∗-dense in E∗∗ and ϕ {a, a, .} is weak∗-continuous on E∗∗, we deduce that ϕ {a, a, b} = 0, for every b ∈ E∗∗. Thus, (8) where s = s(ϕ) ∈ E∗∗ denotes the support tripotent of ϕ in E∗∗. ϕ {a, a, s(ϕ)} = 0, Proposition 2 and Lemma 1.5 in [21] together with the Peirce arithmetic imply that the mapping (x, y) 7→ ϕ {x, y, s} = ϕ {P2(s)x, P2(s)y, s} + ϕ {P1(s)x, P1(s)y, s} is faithful and positive on E∗∗ x ∈ E∗∗ 2 (s) ⊕ E∗∗ 1 (s) and ϕ {x, x, s} = 0 if and only if x = 0. By (8), 0 = ϕ {a, a, s(ϕ)} = ϕ {P2(s)a + P1(s)a, P2(s)a + P1(s)a, s} , 2 (s) ⊕ E∗∗ 1 (s), that is, ϕ {x, x, s} ≥ 0 for every which implies that P2(s)a = P1(s)a = 0. We have shown that AnnE (S) ⊆ E ∩ E∗∗ 0 (s(ϕ)), for every ϕ ∈ S. This assures that (9) AnnE (S) ⊆ E ∩\ϕ∈S E∗∗ 0 (s(ϕ)) . To prove the reverse inclusion, let b belong to the right side of (9), let ϕ ∈ S and let c ∈ E have Peirce decomposition c = c2 + c1 + c0 with respect to s(ϕ). From Peirce arithmetic, {b, ϕ, b} (c) = ϕ {b, c, b} = ϕ {b, c0, b} = 0, proving equality in (9) and establishing (b). To prove (c), let b, c ∈ AnnE (S) and ϕ ∈ S. Then for x = x2 + x1 + x0 ∈ E (with respect to s(ϕ)), by Peirce rules and properties of the support AUTOMATIC CONTINUITY OF DERIVATIONS 11 tripotent, {b, c, ϕ} (x) = ϕ {c, b, x} = ϕ {c, b, x2}+ϕ {c, b, x1}+ϕ {c, b, x0} = 0, which proves c). Because of (5) and (6), to prove (a) it remains to show that AnnE (S) is a linear subspace of E. Take a, b ∈ AnnE (S). Since, by Peirce arithmetic, Q(a, b)(E) ⊆ E ∩ E∗∗ 1 (s(ϕ))) , for every ϕ ∈ S, it follows that {a, ϕ, b} = 0, and {a, b, ϕ} = 0, for every ϕ ∈ S. Therefore (using only the first of these two facts), 0 (s(ϕ)), and L(a, b)(E) ⊆ E ∩ (E∗∗ 0 (s(ϕ)) ⊕ E∗∗ Q(a + b)ϕ = Q(a)ϕ + Q(b)ϕ + 2Q(a, b)ϕ = 0, for every a, b ∈ AnnE (S) and ϕ ∈ S, which implies that AnnE (S) is a linear subspace of E and completes the proof. (cid:3) Remark 6. Let E be a real or complex JB∗-triple regarded as a real Banach triple E-module. It can be easily seen that norm-closed triple E-submodules and norm-closed triple ideals of E coincide. The conclusions in the above Proposition 5 remain true for any norm-closed triple E-submodule (i.e. norm-closed triple ideal) of E. Indeed, let S = J be a norm-closed triple ideal of E. By Lemma 4, AnnE(J) = J ⊥, which implies that {AnnE (J), AnnE (J), J} = 0, in the triple split null extension E ⊕ E. 3. Triple derivations and triple module homomorphisms 3.1. Triple derivations. Separating spaces have been revealed as a use- ful tool in results of automatic continuity. This tool has been applied by many authors in the study of automatic continuity of binary and ternary homomorphims, derivations and module homomorphisms (see, for example, [41, 2, 53, 32, 33, 47, 48, 14] and [15], among others). These spaces also play an important role in the subsequent generalisations of Kaplansky's theorem (compare [12, 26] and [18]). Let T : X → Y be a linear mapping between two normed spaces. Follow- ing [42, Page 70], the separating space, σY (T ), of T in Y is defined as the set of all z in Y for which there exists a sequence (xn) ⊆ X with xn → 0 and T (xn) → z. The separating space, σX (T ), of T in X is defined by σX (T ) := T −1(σY (T )). A straightforward application of the closed graph theorem shows that a linear mapping T between two Banach spaces X and Y is continuous if and only if σY (T ) = {0} (c.f. [12, Proposition 4.5]). It is known that σX (T ) and σY (T ) are closed linear subspaces of X and Y, respectively. A useful property of the separating space σY (T ) asserts that for every bounded linear operator R from Y to another Banach space Z, the com- position RT is continuous if and only if σY (T ) ⊆ ker(R). Further, there exists a constant M > 0 (which does not depend on R nor Z) such that kRT k ≤ M kRk, whenever RT is continuous (compare [48, Lemma 1.3]). 12 PERALTA AND RUSSO Let E be a complex (resp., real) Jordan triple and let X be a triple E- module. We recall that a conjugate linear (resp., linear) mapping δ : E → X is said to be a derivation if δ{a, b, c} = {δ(a), b, c} + {a, δ(b), c} + {a, b, δ(c)} . Note that derivations on complex JB∗-triples to themselves are linear mappings but that a derivation from a complex JB∗-triple into a complex triple module is conjugate linear by this definition. This is not inconsistent, since as we have noted earlier, it is not clear that a complex JB∗-triple E can be made into a complex triple E-module. Lemma 7. Let δ : E → X be a triple derivation from a Jordan Banach triple to a Banach (Jordan) triple E-module. Then σX (δ) is a norm-closed triple E-submodule of X and σE (δ) is a norm-closed subtriple of E. Proof. Given a, b in E and x ∈ σX (δ), there exists a sequence (cn) in E with (cn) → 0 and δ(cn) → x in norm. The sequence ({a, b, cn}) (resp., ({a, cn, b})) tends to zero in norm and δ {a, b, cn} = {δa, b, cn} + {a, δb, cn} + {a, b, δ(cn)} → {a, b, x} (resp., δ {a, cn, b} → {a, x, b}), which proves the first statement. If a, b, c ∈ σE(δ), then δa, δb, δc ∈ σX (δ) and by the first statement (cid:3) δ {a, b, c} ∈ σX(δ), as required. Let δ : E → X be a triple derivation from a Jordan Banach triple E to a Banach triple E-module. Since σX (δ) is a norm closed triple E-submodule of X, AnnE(σX (δ)) is a norm closed inner ideal of E whenever it is a linear subspace of E (actually, in such a case, it is a triple ideal when E is a real or complex JB∗-triple). Let us take a in E. Since δ is in particular a linear mapping, from the useful property mentioned above, σX (δ) ⊆ ker(Q(a)) if and only if Q(a)δ is a continuous linear mapping from E to X, and we deduce that AnnE(σX (δ)) = {a ∈ E : Q(a)δ is continuous}. Moreover, for each a in E, δQ(a) = Q(a)δ + 2Q(a, δa), and it follows that Q(a)δ is continuous if and only if δQ(a) is. 3.2. Triple module homomorphisms. Let X and Y be two triple E- modules over a real or complex Jordan triple E. A linear mapping T : X → Y is said to be a triple E-module homomorphism if the identi- ties T {a, b, x} = {a, b, T (x)} and T {a, x, b} = {a, T (x), b} , hold for every a, b ∈ E and x ∈ X. As above, AnnE(σY (T )) = {a ∈ E : Q(a)T is continuous} AUTOMATIC CONTINUITY OF DERIVATIONS 13 and since a triple module E-homomorphism T : X → Y commutes with Q(a) (acting on X), we have AnnE(σY (T )) = {a ∈ E : T Q(a) is continuous}, where Q(a) acts on Y . The argument applied in the proof of Lemma 7 is also valid to prove the following result. Lemma 8. Let E be a Jordan Banach triple and let T : X → Y be a triple E-module homomorphism between two Banach space which are triple E-modules with continuous module operations. Then σY (T ) and σX (T ) are norm closed triple E-submodules of Y and X, respectively. (cid:3) The following lemma provides a key tool needed in our main result. Lemma 9. Let E be a Jordan Banach triple, X a Banach triple E-module satisfying LOP, Y a Banach space which is a triple E-module with contin- uous module operations and T : X → Y a triple module homomorphism. Then for every sequence (an) of mutually orthogonal non-zero elements in E, we have: (a) Q(an)2T is continuous for all but a finite number of n; (b) a[3] (c) the set n belongs to AnnE(σY (T )) for all but a finite number of n; ( kQ(a[3] kank6 n )T is continuous) n )T k : Q(a[3] is bounded. Proof. Suppose that the statement (a) of the lemma is false. Passing to a subsequence, we may assume that Q(an)2T is an unbounded operator for every natural n. In this case we can find a sequence (xn) in X sat- isfying kxnk ≤ 2−nkank−2, and kQ(an)2T (xn)k > n Kn, where Kn is the norm of the bounded conjugate linear operator Q(an) : Y → Y , Q(an)y = {an, y, an}. Since Q(an)2T is discontinuous Kn = kQ(an)k 6= 0, for every n. (Note that kQ(a)k ≤ M kak2 for some constant M .) module X. For n 6= k, the LOP and the identity k=1 Q(ak)(xk) defines an element z in the Banach triple The series P∞ {x, an, {ak, an, ak}} + {ak, {an, x, an} , ak} = {{x, an, ak} , an, ak} + {ak, an, {x, an, ak}} shows that {ak, {an, x, an} , ak} = 0. That is, Q(ak)Q(an) = 0 for k 6= n and the same argument shows that for any b ∈ E, (10) Q(ak, b)Q(an) = 0 for n 6= k. Hence, for each natural n, we have KnkT (z)k ≥ kQ(an)T (z)k = kT Q(an)(z)k = kT Q(an)2(xn)k = kQ(an)2T (xn)k > Kn n, 14 PERALTA AND RUSSO which is impossible. This proves (a). Since Q(an)2T is continuous for all but a finite number of n and the module operations are continuous on Y , it follows that Q(an)Q(an)2T = Q(an)3T = Q(a[3] n ∈ AnnE(σY (T ))) for all but a finite number of n. This proves (b). n )T is continuous (and hence, a[3] [3] n )T k kank6 In order to prove (c) we may assume that Q(an)2T is continuous for every natural n. Arguing by reduction to the absurd, we assume that suming that kank = 1, for every n. By the Cantor diagonal process we may find a doubly indexed subsequence (ap,q)p,q∈N of (an) and a doubly indexed (cid:26) kQ(a sequence (xp,q) in the unit sphere of X such that(cid:13)(cid:13)(cid:13)Q(a[3] Let bp := P+∞ : n ∈ N(cid:27) is unbounded. There is no loss of generality in as- p,q) T (xp,q)(cid:13)(cid:13)(cid:13) > 42q q p. q=1 2−q ap,q ∈ E. We observe that ap,q ⊥ al,m for every (p, q) 6= (l, m). It is therefore clear that (bp) is a sequence of mutually orthogonal elements in E. Having in mind that X satisfies LOP, we deduce from (4) and (10) that Q(bp)2Q(ap,q)(x) = 4−2q Q(a[3] p,q)(x), for every x in X. Thus, (cid:13)(cid:13)Q(bp)2T Q(ap,q)(xp,q)(cid:13)(cid:13) =(cid:13)(cid:13)T Q(bp)2Q(ap,q)(xp,q)(cid:13)(cid:13) p,q)T (xp,q)(cid:13)(cid:13)(cid:13) > q p, = 4−2q(cid:13)(cid:13)(cid:13)T Q(a[3] p,q)(xp,q)(cid:13)(cid:13)(cid:13) = 4−2q(cid:13)(cid:13)(cid:13)Q(a[3] for every p, q in N, which shows that Q(bp)2T is unbounded for every p ∈ N. This contradicts the first statement of the lemma and proves (c). (cid:3) Let E be a complex (resp., real) Jordan triple and let X be a triple E- It is not hard to see that for every derivation δ : E → X the module. mapping Θδ : E → E ⊕ X a 7→ a + δ(a) is a real linear Jordan triple monomorphism between from the real Jordan triple E to the triple split null extension E ⊕ X. (We observe that, in this case, E is regarded as a real Jordan triple whenever it is a complex Jordan triple). When X is a Jordan Banach triple E-module over a real or complex JB∗- triple E, we define a norm, k.k0, on the triple split null extension of E and X by the assignment a + x 7→ ka + xk0 := kak + kxk. The real Jordan triple E ⊕X becomes a real Jordan Banach triple. It is not hard to see that, in this setting, a derivation δ is continuous if and only if the triple monomorphism Θδ is. Moreover, the separating spaces σX (δ) and σE⊕X (Θδ) are linked by the the following identity (11) Moreover, σE⊕X (Θδ) = {0} × σX(δ). AnnE(σE⊕X(Θδ)) = AnnE(σX (δ)). AUTOMATIC CONTINUITY OF DERIVATIONS 15 The linear space Θδ(E) is a subtriple of E ⊕ X and is made into a triple E-module for the products {a, b, Θδ(c)} = Θδ({a, b, c}) = {Θδ(a), Θδ(b), Θδ(c)} = {a, Θδ(b), c} , (a, b, c ∈ E). These products can be extended to the k.k0-closure, Θδ(E), of Θδ(E). Under this point of view, the mapping Θδ : E → Θδ(E) is a triple E-module homomorphism. The following result derives from the previous Lemma 9, since Q(a)Θδ = Q(a) ⊕ Q(a)δ. Corollary 10. Let E be a complex (resp., real) JB∗-triple, X a Banach space which is a triple E-module with continuous module operations and let δ : E → X be a triple derivation. Then for every sequence (an) of mutually orthogonal non-zero elements in E, Q(an)2δ is continuous for all but a finite number of n. It follows that a[3] n belongs to AnnE(σX (δ)) for all but a finite number of n. Moreover, the set ( kQ(a[3] kank6 n )δ is continuous) n )δk : Q(a[3] is bounded. (cid:3) Let E be a real or complex JB∗-triple. We shall say that E is algebraic if all singly-generated subtriples of E are finite-dimensional. If in fact there exists m ∈ N such that all single-generated subtriples of X have dimension ≤ m, then E is said to be of bounded degree, and the minimum such an m will be called the degree of E. Corollary 11. Let E be a complex (resp., real) JB∗-triple, X a Banach triple E-module and let δ : E → X be a triple derivation. Suppose that AnnE(σX (δ)) is a norm closed triple ideal of E. Then every element in E/AnnE(σX (δ)) has finite triple spectrum, in other words, the JB∗-triple E/AnnE(σX (δ)) is isomorphic to a Hilbert space or, equivalently, algebraic of bounded degree. Proof. Let a be an element in the JB∗-triple F = E/AnnE(σX (δ)). Let Ia denote the intersection of Ea with AnnE(σX (δ)). It is clear that Ia is a norm closed triple ideal of Ea. Moreover, the subtriple Fa is JB∗-triple isomorphic to the quotient of Ea with Ia. Ea is JB∗-triple isomorphic (and hence isometric) to C0(L) = C0(L, C) (resp., C0(L) = C0(L, R)) for some locally compact Hausdorff space L ⊆ (0, kak] (called the triple spectrum of a) such that L ∪ {0} is compact. We shall identify Ea with C0(L). It is known (compare [20, Proposition 3.10]) that Ea/Ia ∼= C0(Λ) where Λ = {t ∈ L : b(t) = 0, for every b ∈ Ia}. We claim that the set Λ is finite. Otherwise, there exists an infinite se- quence (tn) in Λ. We find a sequence (fn) of mutually orthogonal elements 16 PERALTA AND RUSSO in C0(L) such that fn(tn) 6= 0 and hence fn 6∈ Ia and f [3] 6∈ Ia. Since orthogonality is a "local" concept, (fn) is a sequence of mutually orthog- onal elements in E and (f [3] n ) 6∈ AnnE(σX (δ)), we have a contradiction to Corollary 10. n It follows that Ea/Ia ∼= Fa is finite dimensional. The final statement (cid:3) follows from [8, §4] and [5, §3, Theorems 3.1 and 3.8]. 3.3. Automatic continuity results. Our main result (Theorem 13) will be proved in two steps, the first being the following proposition. Proposition 12. Let E be a complex (resp., real) JB∗-triple, X a Banach triple E-module, and let δ : E → X be a triple derivation. Assume that AnnE(σX (δ)) is a (norm-closed) linear subspace of E and that in the triple split null extension E ⊕ X, {AnnE (σX (δ)), AnnE (σX (δ)), σX (δ)} = 0. (12) Then δAnnE (σX (δ)) : AnnE(σX (δ)) → X is continuous. Proof. By Lemma 7, σX (δ) is a triple E-submodule of X. Since we are assuming that AnnE(σX (δ)) is a norm-closed subspace of E, as we have seen, it is a norm-closed triple ideal of E. Fix two arbitrary elements a, b in AnnE(σX (δ)). Since a+b ∈ AnnE(σX (δ)), for every x in σX (δ), we have 2 {a, x, b} = {a + b, x, a + b} − {a, x, a} − {b, x, b} = 0, Hence, in addition to our assumption (12), we also have {a, x, b} = 0, for every x ∈ σX (δ), a, b ∈ AnnE(σX (δ)), that is, (13) {AnnE (σX (δ)), σX (δ), AnnE (σX (δ)} = 0. Considering L(a, b) and Q(a, b) as linear mappings from X to X de- fined by L(a, b)(x) = {a, b, x} and Q(a, b)(x) = {a, x, b} (x ∈ X), we deduce from (12), (13) that σX(δ) ⊂ ker L(a, b) ∩ ker Q(a, b) and there- fore that L(a, b)δ, Q(a, b)δ : E → X are continuous operators for every a, b ∈ AnnE(σX (δ)). When L(a, b) and Q(a, b) as considered as (real) linear operators from E to E, the compositions δL(a, b) and δQ(a, b) satisfy the identities δL(a, b)(c) = {δ(a), b, c} + {a, δ(b), c} + {a, b, δ(c)} = L(δa, b)(c) + L(a, δb)(c) + L(a, b)δ(c) and δQ(a, b)(c) = {δ(a), c, b} + {a, δ(c), b} + {a, c, δ(b)} = Q(δa, b)(c) + Q(a, b)δ(c) + Q(a, δb)(c). for an arbitrary c ∈ E. Since X is a Banach triple E-module, the continuity of L(a, b)δ and Q(a, b)δ as operators from E to X implies that the mappings AUTOMATIC CONTINUITY OF DERIVATIONS 17 c 7→ δ({a, b, c}) and c 7→ δ({a, c, b}) are continuous linear operators from E to X. Let W : AnnE(σX (δ)) × AnnE(σX (δ)) × AnnE(σX (δ)) → X be the real trilinear mapping defined by W (a, b, c) := δ({a, b, c}). We have already seen that W is separately continuous whenever we fix two of the variables in (a, b, c) ∈ AnnE(σX (δ)) × AnnE(σX (δ)) × AnnE(σX (δ)). By repeated application of the uniform boundedness principle, W is (jointly) continu- ous. Therefore, there exists a positive constant M such that kδ {a, b, c} k ≤ M kak kbk kck, for every a, b, c ∈ AnnE(σX (δ)). Finally, since AnnE(σX (δ)) is a JB∗-subtriple of E, for each a in In AnnE(σX (δ)), there exists b in AnnE(σX (δ)) satisfying that b[3] = a. this case kδ(a)k = kδ {b, b, b} k ≤ M kbk3 = M k {b, b, b} k = M kak, which shows that the restriction of δ to AnnE(σX (δ)) is continuous. (cid:3) We can state now the main results of the paper. Theorem 13. Let E be a complex JB∗-triple, X a Banach triple E-module, and let δ : E → X be a triple derivation. Then δ is continuous if and only if AnnE(σX (δ)) is a (norm-closed) linear subspace of E and {AnnE (σX (δ)), AnnE (σX (δ)), σX (δ)} = 0, in the triple split null extension E ⊕ X. Proof. If δ is continuous AnnE(σX (δ)) = AnnE({0}) = E is a linear sub- space of E and {E, E, 0} = 0. Conversely, let us suppose that E is a complex JB∗-triple and that AnnE(σX (δ)) is a norm-closed subspace of E and hence a norm-closed triple ideal of E. In order to simplify notation, we denote J = AnnE(σX (δ)), while the projection of E onto E/J will be denoted by a 7→ π(a) = a. By Corollary 11, E/J is algebraic of bounded degree m. Thus, for each element a in E/J there exist mutually orthogonal minimal tripotents e1, . . . , ek in E/J and 0 < λ1 ≤ . . . ≤ λk with k ≤ m such that a = j=1 λjej. We shall show in the next two paragraphs that e1, . . . , ek ∈ J, and hence, a ∈ J. This will show that E = J and application of Proposi- tion 12 will complete the proof. Pk Suppose that e is a minimal tripotent in E/J, where e ∈ E is a repre- sentative in the class e. In this case (E/J)2 (e) = Ce. Take an arbitrary sequence (an) converging to 0 in E. For each natural n, there exists a scalar µn ∈ C such that π(Q(e)(an)) = Q(e)(π(an)) = Q(e)(an) = µne = π(µne). The continuity of π and the Peirce 2 projection P2(e) assure that µn → 0. It follows that the sequence Q(e)(an) − µne lies in J and tends to zero in norm. 18 PERALTA AND RUSSO By Proposition 12, δJ is continuous. Therefore, δ(Q(e)(an)) = δ(Q(e)(an) − µne) + µnδ(e) → 0. Since (an) is an arbitrary norm null sequence in E, the linear mapping δQ(e) : E → X is continuous, and hence e ∈ AnnE(σX (δ)) = J, or equiva- lently, e = 0. (cid:3) Let E be a real JB∗-triple. By [29, Proposition 2.2], there exists a unique complex JB∗-triple structure on the complexification bE = E ⊕ i E, and a unique conjugation (i.e., conjugate-linear isometry of period 2) τ on bE such that E = bEτ := {x ∈ bE : τ (x) = x}, that is, E is a real form of a complex JB∗-triple. Let us consider defined by τ ♯ : bE∗ → bE∗ τ ♯(f )(z) = f (τ (z)). The mapping τ ♯ is a conjugation on bE∗. Furthermore the map (bE∗)τ ♯ −→ (bEτ )∗ (= E∗) f 7→ f E is an isometric bijection, where (bE∗)τ ♯ [29, Page 316]). := {f ∈ bE∗ : τ ♯(f ) = f } (compare Remark 14. Let δ : E → E∗ be a triple derivation from a real JB∗-triple to its dual. It is not hard (but tedious) to see that, under the identifications the latter is seen as a triple E-module. Actually, although the calculations are tedious, the triple products of every real triple E-module, X, can be appropriately extended to its algebraic given in the above paragraph, the mapping bδ : bE → bE∗, bδ(x + iy) := δ(x) − iδ(y) is conjugate-linear and a triple derivation from bE to bE∗, when complexification bX = X ⊕ iX to make the latter a complex triple bE-module. a (conjugate linear) triple derivationbδ : bE → bX. Further, every (real linear) triple derivation δ : E → X can be extended to Corollary 15. Let E be a real or complex JB∗-triple. (a) Every derivation δ : E → E is continuous. (b) Every derivation δ : E → E∗ is continuous. Proof. The proof in the complex case follows now from Proposition 5 and Theorem 13. (In Theorem 13, we consider E as a real triple and as a real E-module, and δ as a real-linear map.) The statements in the real setting are, by Remark 14, direct consequences of the corresponding results in the complex case. (cid:3) AUTOMATIC CONTINUITY OF DERIVATIONS 19 The first statement of the above corollary was already established in [3, Corollary 2.2] and [27, Remark 1]. The proof given here is completely in- dependent. The second statement is new and will be important for a forth- coming study by the authors of weak amenability for JB∗-triples. Recall that every derivation of a complex C∗-algebra A into a Banach A-bimodule is automatically continuous [43]. The class of Banach triple modules over real or complex JB∗-triples is strictly wider than the class of Banach bimodules over C∗-algebras. Our next remark shows that, in the more general setting of triple derivations from real or complex JB∗-triples to Banach triple modules the continuity is not, in general, automatic. Remark 16. Let H be a real Hilbert space with inner product denoted by (., .). Suppose that dim(H) ≥ 2. Let J denote the Banach space C1 ⊕ℓ1 H. It is known that J is a JB-algebra with respect to the product (λ11 + a1) ◦ (λ21 + a2) := λ1a2 + λ2a1 + (λ1λ2 + (a1, a2))1. The JB-algebra (J, ◦) is called a spin factor (see [24]). It follows that J is a real JB∗-triple via {a, b, c} := (a ◦ b) ◦ c + (c ◦ b) ◦ a − (a ◦ c) ◦ b, (a, b, c ∈ J). It was already noticed by Hejazian and Niknan (see [25, Definition 3.2]) that every Banach space X can be considered as a (degenerate) Jordan J-module with respect to the products (λ11 + a1) ◦ x = x ◦ (λ11 + a1) = λ1x, (x ∈ X, λ1 ∈ R, a1 ∈ H). Since every linear mapping D : J → X with D(1) = 0 is a Jordan derivation (i.e. D(a ◦ b) = D(a) ◦ b + a ◦ D(b), ∀a, b ∈ J), for every infinite dimensional spin factor J, there exists a discontinuous derivation from J to a degenerate Jordan J-module. Each degenerate Banach Jordan J-module X is a Banach triple J-module with respect to {a, b, x} := (a ◦ b) ◦ x + (x ◦ b) ◦ a − (a ◦ x) ◦ b and {a, x, b} = (a ◦ x) ◦ b + (b ◦ x) ◦ a − (a ◦ b) ◦ x (a, b ∈ J, x ∈ X), and each linear mapping δ : J → X with δ(1) = 0 is a triple derivation. Thus, for each infinite dimensional spin factor J there exists a discontinuous triple derivation from J to a Banach triple J-module. 4. Derivations on a C∗-algebra A celebrated result due to J.R. Ringrose establishes that every (associa- tive) derivation from a C∗-algebra A to a Banach A-bimodule is continuous (cf. [43]). We have already commented that S. Hejazian and A. Niknam gave in [25, §3] an example of a discontinuous Jordan derivation from a JB∗-algebra to a Jordan Banach module. Based on this example, we have already shown the existence of a discontinuous triple derivation from a JB∗- triple to a Banach triple module (see Remark 16). The aim of this section is to show that these two counterexamples cannot be found when the domain is a C∗-algebra, thereby providing positive answers to Problems 1 and 2 20 PERALTA AND RUSSO (Corollaries 20 and 21). We shall also see that Ringrose's Theorem derives as a consequence of our results (Corollary 22). We shall actually prove a stronger result: every triple derivation from a C∗-algebra to a Banach triple module is automatically continuous (Theorem 19), which will imply these three corollaries. We shall need a technical reformulation of Theorem 13 above. Theorem 13 has been established only for complex JB∗-triples. The proof given in Section 3 is not valid for real JB∗-triples. The obstacles appearing in the real setting concern the structure of the Peirce-2 subspace associated with a minimal tripotent. We have already commented that, in case of E being a complex JB∗-triple, the identity E−1(e) = iE1(e) holds for every tripotent e in E, whereas in the real situation the dimensions of E1(e) and E−1(e) are not, in general, correlated. For example, every infinite dimensional rank- one real Cartan factor C contains a minimal tripotent e satisfying that C 1(e) = Re and dim(C −1(e)) = +∞ (compare [19, Remark 2.6]). Following [38, 11.9], we shall say that a real JB∗-triple E is reduced when- ever E2(e) = Re (equivalently, E−1(e) = 0) for every minimal tripotent e ∈ E. Reduced real Cartan factors were studied and classified in [38, 11.9] and in [36, Table 1]. Reduced real JB∗-triples played an important role in the study of the surjective isometries between real JB∗-triples developed in [19]. Having the above comments in mind, it is not hard to check that, in the particular subclass of reduced real JB∗-triples the proof of Theorem 13 remains valid line by line. We therefore have: Proposition 17. Let E be a reduced real JB∗-triple, X a Banach triple E-module, and let δ : E → X be a triple derivation. Then δ is continuous if and only if AnnE(σX (δ)) is a (norm-closed) linear subspace of E and {AnnE (σX (δ)), AnnE (σX (δ)), σX (δ)} = 0, in the triple split null extension E ⊕ X. (cid:3) Every closed ideal of a reduced real JB∗-triple is a reduced real JB∗-triple. It is also true that the self-adjoint part, Asa, of a C∗-algebra, A, is a reduced real JB∗-triple with respect to the product (14) {a, b, c} := abc + cba 2 (a, b, c ∈ Asa), or equivalently, (15) {a, b, c} := (a ◦ b) ◦ c + (c ◦ b) ◦ a − (a ◦ c) ◦ b, (a, b, c ∈ Asa). Indeed, writing e = p − q for a minimal partial isometry e ∈ Asa with p and q orthogonal projections, it is easy to check that e = p or e = −q and it follows that if exe = −x, then x = 0. In particular, for each closed triple ideal J of Asa, the quotient Asa/J is a reduced real JB∗-triple. AUTOMATIC CONTINUITY OF DERIVATIONS 21 Our next result is a consequence of the previous proposition. Note that the fact that Asa is a reduced JB∗-triple is only needed in the case that A is an abelian C∗-algebra. Proposition 18. Let A be an abelian C∗-algebra whose self adjoint part is denoted by Asa. Then, every triple derivation from Asa to a real Jordan- Banach triple Asa-module is continuous. In particular, every triple deriva- tion from A into a real Jordan-Banach triple A-module is continuous. Proof. Let δ : Asa → X be a triple derivation from Asa into a real Jor- dan triple Asa-module. The statement of the proposition will follow from Proposition 17 as soon as we prove that Ann(σX (δ)) = AnnAsa(σX (δ)) is a (norm-closed) linear subspace of Asa and {Ann(σX (δ)), Ann(σX (δ)), σX (δ)} = 0. Let us take a ∈ Ann(σX (δ)). Having in mind that a ∈ Ann(σX (δ)) if, and only if, Q(a)δ (or equivalently, δQ(a)) is a continuous operator from Asa to X (see the comments after Lemma 7), we observe that δQ(a) is a continuous mapping from Asa to X. Obviously, for each b in Asa, the operator Lb : Asa → Asa, c 7→ cb = bc is continuous. Since A is abelian we have L(a2, b) = Q(a)Lb = LbQ(a). Therefore δL(a2, b) = δQ(a)Lb is a continuous operator from Asa to X. The identity δL(a2, b) = L(δ(a2), b) + L(a2, δ(b)) + L(a2, b)δ shows that L(a2, b)δ is continuous. It is easy to check, from the definition of σX (δ), that(cid:8)a2, b, x(cid:9) = 0, for every x ∈ σX (δ). It follows that (cid:8)a2, b, x(cid:9) = 0, for every a ∈ Ann(σX (δ)), b ∈ Asa and x ∈ σX (δ). (16) It is known that a can be written in the form a = a1 − a2, where a1 and a2 are two orthogonal positive elements in Asa. It is also known that Q(a)(Asa) ∈ Ann(σX (δ)). Therefore, a3 1 = Q(a)(a1) ∈ Ann(σX (δ)) and hence a6 1)(Asa) ⊆ Ann(σX (δ)). This implies that the ideal of Asa generated by a6 1 lies in Ann(σX (δ)), which guarantees that a1 ∈ Ann(σX (δ)). We can similarly show that a2 belongs to Ann(σX (δ)). A similar argument shows that a 2 ∈ Ann(σX (δ)). Now, we deduce from (16) that 1Asa = Q(a3 1 , a 1 2 1 2 (17) {a, b, x} = {a1, b, x} − {a2, b, x} = 0, for every a ∈ Ann(σX (δ)), b ∈ Asa and x ∈ σX (δ), or equivalently, δL(a, b) and L(a, b)δ are continuous operators for every a ∈ Ann(σX (δ)) and b ∈ Asa. Since A is abelian, L(a, b) = Q(a, b) in Asa, it follows from (17), that δQ(a, b) and Q(a, b)δ are continuous operators from Asa to X for every a ∈ Ann(σX (δ)) and b ∈ Asa. This implies that (18) {a, x, b} = 0, for every a ∈ Ann(σX (δ)), b ∈ Asa and x ∈ σX (δ). Finally, given a, c in Ann(σX (δ)), we deduce from (18) that Q(a + c)(σX (δ)) = Q(a)(σX (δ)) + Q(c)(σX (δ)) + 2Q(a, c)(σX (δ)) = 0, 22 PERALTA AND RUSSO which shows that a + c ∈ Ann(σX (δ)), and hence the latter is a linear subspace of Asa. (cid:3) Given any element x in a C∗-algebra A, we shall denote by C(x) the C∗-subalgebra of A generated by x. The following theorem, due to J. Cuntz (see [13]) will be required later. Lemma 19. [13, Theorem 1.3] Let A be a C∗-algebra and f a linear func- tional on A. If f is continuous on C(h) for all h = h∗ in A, then A is continuous on A. By the uniform boundedness theorem, a linear mapping T from A to a normed space X is continuous if and only if it restriction to C(h) is continuous for all h = h∗ in A. (cid:3) Let δ : A → X be a triple derivation from a C∗-algebra to a Banach triple A-module. For each self-adjoint element h in A, the Banach space X can be regarded as a Jordan Banach C(h)-module by restricting the module operation from A to C(h). Since δC(h) : C(h) → X is a triple derivation from an abelian C∗-algebra into a Banach triple C(h)-module, Proposition 18 assures that δC(h) is continuous. Combining this argument with the above Cuntz's theorem we have: Theorem 20. Let A be a C∗-algebra. Then every triple derivation from A (respectively, from Asa) into a complex (respectively, real) Jordan Banach triple A-module is continuous. (cid:3) Since every Jordan derivation is a triple derivation, and every Jordan module is a Jordan triple module, we have: Corollary 21 (Solution to Problem 2). Let A be a C∗-algebra. Then every Jordan derivation from A into a Jordan-Banach A-module X is continuous.(cid:3) It is due to B.E. Johnson that every continuous Jordan derivation from a C∗-algebra A to a Banach A-bimodule is a derivation (cf. [31, Theorem 6.2]). As we have just seen, the hypothesis of continuity can be omitted in the just quoted theorem. Thus: Corollary 22 (Solution to Problem 1). Let A be a C∗-algebra. Then every Jordan derivation from A into a Banach A-bimodule X is continuous. In particular, every Jordan derivation from A to X is a derivation, by Johnson's theorem. (cid:3) Let D : A → X be an associative (resp., Jordan) derivation from a C∗- algebra to a Banach A-bimodule. The space X, regarded as a real Banach space, is a real Banach triple Asa-module with respect to the product defined in (15), where, in this case, one element in (a, b, c) is taken in X and the other two in Asa. The restriction of D to Asa, δ = DAsa : Asa → X is a (real linear) triple derivation. Hence, Theorem 20 implies that δ (and hence D) is continuous. Thus: AUTOMATIC CONTINUITY OF DERIVATIONS 23 Corollary 23 (Ringrose). Let A be a C∗-algebra. Then every derivation from A into a Banach A-bimodule X is continuous. (cid:3) In [23], U. Haagerup and N.J. Laustsen presented a new proof of Johnson's Theorem. Applying a result of automatic continuity in [25, Corollary 2.3], the just quoted authors proved that every Jordan derivation from a C∗- algebra A to A∗ is bounded and hence an inner derivation (cf. [23, Corollary 2.5]). In [6], M. Bresar studied a more general class of Jordan derivations from a C∗-algebra A to an A-bimodule X. An additive mapping d : A → X satisfying d(a ◦ b) = d(a) ◦ b + a ◦ d(b), for every a, b ∈ A, is called an additive Jordan derivation. An additive Jordan derivation is said to be proper when it is not an associative derivation. Every (linear) Jordan derivation D : A → X is an additive Jordan derivation. However, the reciprocal implication is, in general, false. Actually, from [6, Theorem 5.1], for each unital C∗-algebra A, then there exists a proper additive Jordan derivation from A into some unital A-bimodule if, and only if, A contains an ideal of codimension one. References [1] J. Alaminos, M. Bresar, A.R. Villena, The strong degree of von Neumann algebras and the structure of Lie and Jordan derivations, Math. Proc. Cambridge Philos. Soc. 137, no. 2, 441-463 (2004). [2] W.G. Bade, P.C. Curtis, Homomorphisms of commutative Banach algebras, Amer. J. Math. 82, 589-608 (1960). [3] T.J. Barton, Y. Friedman, Bounded derivations of JB-triples, Quart. J. Math. Oxford 41, 255-268 (1990). [4] T.J. Barton, R.M. Timoney, Weak∗-continuity of Jordan triple products and its ap- plications, Math. Scand. 59, 177-191 (1986). [5] J. Becerra-Guerrero, G. L´opez, A. M. Peralta and A. Rodr´ıguez-Palacios, Relatively weakly open sets in closed unit balls of Banach spaces, and real JB∗-triples of finite rank, Math. Ann. 330, 45-58 (2004). [6] M. Bresar, Jordan derivations revisited, Math. Proc. Cambridge Philos. Soc. 139, no. 3, 411-425 (2005). [7] L. J. Bunce, Structure of representations and ideals of homogeneous type in Jordan algebras, Quart. J. Math. Oxford Ser. (2) 37, no. 145, 1-10 (1986). [8] L.J. Bunce and C.-H. Chu, Compact operations, multipliers and Radon-Nikodym property in J B ∗-triples, Pacific J. Math. 153, 249-265 (1992). [9] L.J. Bunce, Ch.-H. Chu, B. Zalar, Structure spaces and decomposition in JB∗-triples. Math. Scand. 86, 17-35 (2000). [10] M. Burgos, F.J. Fern´andez-Polo, J. Garc´es, J. Mart´ınez, A.M. Peralta, Orthogonality preservers in C∗-algebras, JB-algebras and JB-triples, J. Math. Anal. Appl. 348, 220-233 (2008). [11] M. Burgos, A. M. Peralta, M. Ram´ırez, and M. E. Ruiz Morillas, von Neumann regularity in Jordan-Banach triples, in Proceedings of Jordan structures in algebra and analysis meeting (Almer´ıa, 2009), Eds: J. Carmona et al., 65-88, Punto Rojo, Almer´ıa, 2010. [12] S. B. Cleveland, Homomorphisms of non-commutative -algebras, Pacific J. Math. 13, 1097-1109 (1963). [13] J. Cuntz, On the continuity of semi-norms on operator algebras, Math. Ann. 220, no. 2, 171-183 (1976). 24 PERALTA AND RUSSO [14] H.G. Dales, Automatic continuity: a survey, Bull. London Math. Soc. 10, no. 2, 129-183 (1978). [15] H.G. Dales, On norms on algebras, Proc. Centre Math. Anal. Austral. Nat. Univ. 21, 61-69 (1989). [16] H. G. Dales, Banach algebras and automatic continuity. London Mathematical Society Monographs. New Series, 24. Oxford Science Publications. The Clarendon Press, Oxford University Press, New York, 2000. [17] S. Dineen, The second dual of a JB∗-triple system, in: J. Mujica (Ed.), Complex analysis, Functional Analysis and Approximation Theory, North-Holland, Amster- dam, 1986. [18] F.J. Fern´andez-Polo, J. Garc´es, A.M. Peralta, A Kaplansky theorem for JB∗-triples, to appear in Proc. Amer. Math. Soc. [19] F.J. Fern´andez-Polo, J. Mart´ınez Moreno, and A.M. Peralta, Surjective isometries between real JB∗-triples, Math. Proc. Cambridge Phil. Soc., 137, 709-723 (2004). [20] Y. Friedman, B. Russo, Function representation of commutative operator triple sys- tems, J. London Math. Soc. (2) 27, no. 3, 513-524 (1983). [21] Y. Friedman, B. Russo, Structure of the predual of a JBW∗-triple. J. Reine Angew. Math. 356, 67-89 (1985). [22] Y. Friedman and B. Russo, A Gelfand-Naimark theorem for JB∗-triples, Duke Math. J. 53, 139-148 (1986). [23] U. Haagerup, N.J. Laustsen, Weak amenability of C∗-algebras and a theorem of Goldstein, in Banach algebras '97 (Blaubeuren), 223-243, de Gruyter, Berlin, 1998. [24] H. Hanche-Olsen, E. Størmer, Jordan operator algebras, Monographs and Studies in Mathematics 21, Pitman, London-Boston-Melbourne 1984. [25] S. Hejazian, A. Niknam, Modules Annihilators and module derivations of JB∗- algebras, Indian J. pure appl. Math., 27, 129-140 (1996). [26] S. Hejazian, A. Niknam, A Kaplansky theorem for JB∗-algebras, Rocky Mountain J. Math. 28, no. 3, 977-982 (1998). [27] T. Ho, J. Martinez-Moreno, A.M. Peralta, B. Russo, Derivations on real and complex JB∗-triples, J. London Math. Soc. (2) 65, no. 1, 85-102 (2002). [28] G. Horn, Characterization of the predual and ideal structure of a JBW∗-triple, Math. Scand. 61, no. 1, 117-133 (1987). [29] J.M. Isidro, W. Kaup, A. Rodr´ıguez, On real forms of JB∗-triples, Manuscripta Math. 86, 311-335 (1995). [30] N. Jacobson, Structure and representation of Jordan algebras, Amer. Math. Soc. Col- loq. Publications, vol 39, 1968. [31] B.E. Johnson, Symmetric amenability and the nonexistence of Lie and Jordan deriva- tions, Math. Proc. Cambridge Philos. Soc. 120, no. 3, 455-473 (1996). [32] B.E. Johnson, A.M. Sinclair, Continuity of derivations and a problem of Kaplansky, Amer. J. Math. 90, 1067-1073 (1968). [33] B.E. Johnson, A.M. Sinclair, Continuity of linear operators commuting with contin- uous linear operators. II., Trans. Amer. Math. Soc. 146, 533-540 (1969). [34] W. Kaup, A Riemann Mapping Theorem for bounded symmentric domains in com- plex Banach spaces, Math. Z. 183, 503-529 (1983). [35] W. Kaup, On spectral and singular values in JB-triples, Proc. Roy. Irish Acad. Sect. A 96, no. 1, 95-103 (1996). [36] W. Kaup, On real Cartan factors, Manuscripta Math. 92, 191-222 (1997). [37] O. Loos, Representations of Jordan triples, Trans. Amer. Math. Soc. 185, 199-211 (1973). [38] O. Loos, Bounded symmetric domains and Jordan pairs, Math. Lectures, University of California, Irvine 1977. [39] J. Mart´ınez, A. M. Peralta, Separate weak-continuity of the triple product in dual real JB∗-triples, Math. Z. 234, 635-646 (2000). AUTOMATIC CONTINUITY OF DERIVATIONS 25 [40] A.M. Peralta, L.L. Stach´o, Atomic decomposition of real JBW∗-triples, Quart. J. Math. Oxford 52, no. 1, 79-87 (2001). [41] C. Rickart, The uniqueness of norm problem in Banach algebras, Ann. of Math, 51, 615-628 (1950). [42] C. Rickart, Genenal Theory of Banach Algebras, Van Nostrand, New York, 1960. [43] J. R. Ringrose, Automatic continuity of derivations of operator algebras, J. London Math. Soc. (2) 5 , 432-438 (1972). [44] B. Russo, Structure of JB∗-triples, In: Jordan Algebras, Proceedings of the Ober- wolfach Conference 1992, Eds: W. Kaup, K. McCrimmon, H. Petersson, 209-280, de Gruyter, Berlin, 1994. [45] S. Sakai, On a conjecture of Kaplansky, Tohoku Math. J., 12, 31-33 (1960). [46] A.M. Sinclair, Jordan homomorphisms and derivations on semisimple Banach alge- bras, Proc. Amer. Math. Soc. (3) 24, 209-214 (1970). [47] A.M. Sinclair, Homomorphisms from C∗-algebras, Proc. London Math. Soc. (3) 29, 435-452 (1975). [48] A.M. Sinclair, Automatic continuity of linear operators, London Mathematical Soci- ety Lecture Note Series, No. 21, Cambridge University Press, Cambridge-New York- Melbourne, 1976. [49] H. Upmeier, Symmetric Banach manifolds and Jordan C*-algebras, North-Holland, Amsterdam, 1985 [50] H. Upmeier, Jordan algebras in analysis, operator theory, and quantum mechanics, CBMS, Regional conference, No. 67 (1987). [51] A. R. Villena, Automatic continuity in associative and nonassociative context, Irish Math. Soc. Bull. No. 46, 43-76 (2001). [52] J.D. Maitland Wright, Jordan C∗-algebras, Michigan Math. J. 24, no. 3, 291-302 (1977). [53] B. Yood, Homomorphisms on normed algebras, Pacific J. Math. 8, 373-381 (1958). E-mail address: [email protected] Departamento de An´alisis Matem´atico, Facultad de Ciencias, Universidad de Granada, 18071 Granada, Spain. E-mail address: [email protected] Department of Mathematics, UC Irvine, Irvine CA, USA
1402.1048	5	1402	2015-02-23T09:35:06	Random walk questions for linear quantum groups	[ "math.OA", "math.QA" ]	We study the discrete quantum groups $\Gamma$ whose group algebra has an inner faithful representation of type $\pi:C^*(\Gamma)\to M_K(\mathbb C)$. Such a representation can be thought of as coming from an embedding $\Gamma\subset U_K$. Our main result, concerning a certain class of examples of such quantum groups, is an asymptotic convergence theorem for the random walk on $\Gamma$. The proof uses various algebraic and probabilistic techniques.	math.OA	math	RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS TEODOR BANICA AND JULIEN BICHON Abstract. We study the discrete quantum groups Γ whose group algebra has an inner faithful representation of type π : C ∗(Γ) → MK(C). Such a representation can be thought of as coming from an embedding Γ ⊂ UK. Our main result, concerning a certain class of examples of such quantum groups, is an asymptotic convergence theorem for the random walk on Γ. The proof uses various algebraic and probabilistic techniques. Introduction A discrete quantum group Γ is dual to a compact quantum group G, and vice versa. Of particular interest is the case where Γ is finitely generated, which corresponds to the case where G is a matrix quantum group. The associated unital Hopf C ∗-algebras A = C ∗(Γ) = C(G) were axiomatized by Woronowicz in [29], [30]. The algebra A has a Haar functional,R : A → C, and possesses a certain distinguished distribution of χ ∈ (A,R ). There are two motivations for this problem: element χ = T r(u) ∈ A, and one interesting problem is that of computing the probabilistic (1) Random walks. For a discrete group Γ =< g1, . . . , gN > we have χ = g1 + . . . + gN , whose moments are the numbers cp = #{i1, . . . , ipgi1 . . . gip = 1}. whose moments are the numbers cp = dim(F ix(u⊗p)). (2) Representation theory. For a compact group G ⊂u UN we have χ(g) = T r(u(g)), In general, understanding the structure of A, and computing the law of χ = T r(u), are non-trivial questions. The available methods here fall into two main classes: are known, by definition or computation, Tannakian methods apply. (1) Category theory. When the relations between the standard coordinates uij ∈ A (2) Matrix models. Here the idea is to search for models for the variables uij ∈ A. The first idea is well-known, going back to old work of Brauer [12], when G is classical, and to old work of Kesten [19], when Γ is classical. In the quantum group context, this idea has been heavily developed, starting with [30]. See [8], [17], [23]. Once such a model found, matrix analysis gives us information about χ. The second idea, while having a big potential as well, is more of an "underground" one. Only some general algebraic theory is available here ([1], [3], [5], [7], [13], [16]), and there is still a lot of work to be done, in order for this idea to really "take off". 2000 Mathematics Subject Classification. 46L65 (46L54). Key words and phrases. Quantum group, Random walk, Freeness. 1 2 TEODOR BANICA AND JULIEN BICHON We will do here some work in this direction. First of all, a matrix model for A will be by definition an inner faithful representation of type π : A → MK(C). The existence of such a representation is a linearity type condition on Γ, because in the classical case, the representation must come from a group embedding Γ ⊂ UK. The simplest models are those coming from the Fourier representations Z ⊂ UZ of the finite abelian groups Z. Assuming now that we are in a product situation, Z = X×Y , the corresponding matrix model can be twisted by a parameter belonging to a torus, Q ∈ TZ, and produces in this way a certain quantum group GQ. We will study here GQ, and the random walk on the dual quantum group ΓQ, our main result being: Theorem. When Q ∈ TZ is generic, with X = αK,Y = βK, K → ∞ we have law(χ) =(cid:18)1 − 1 αβK 2(cid:19) δ0 + 1 αβK 2 D 1 βK (πα/β) where πt is the Marchenko-Pastur (or free Poisson) law of parameter t. The proof uses various algebraic and probabilistic techniques, notably Hopf algebra methods from [2], [5], [11], [24] and free probability theory from [10], [20], [22], [26]. Generally speaking, the situation that we have, with quantum groups GQ which are undeformed (S2 = id) and which depend critically on the arithmetics of Q ∈ TZ, is of course quite beautiful, and reminds a bit the theory of the algebras Uq(g) at q = 1, coming from [15], [18]. At Z = 4 the arithmetic specializations can be computed by using methods from [4]. In general, this remains to be explored. From a more applied point of view now, the main raison d'etre of the compact quantum groups is that of acting on noncommutative manifolds coming from quantum physics, cf. [9], [14]. There are many interesting questions here, regarding the potential applications of the matrix model techniques. One problem is that of unifying the present results with those in [6], and then trying to investigate more specialized models. The paper is organized as follows: 1 is a preliminary section, in 2-3-4 we develop a number of algebraic methods for computing the quantum groups associated to the above matrix models, and in 5 we state and prove the random walk results. Acknowledgements. The work of TB was partly supported by the "Harmonia" NCN grant 2012/06/M/ST1/00169. 1. Matrix models We fix a Hopf algebra A = C ∗(Γ) = C(G), satisfying Woronowicz's axioms in [29], [30]. We assume in addition that the square of the antipode is the identity, S2 = id. By [29], this is the same as assuming that the Haar integration functionalR : A → C has the trace property R ab = R ba. generated groups, Γ =< g1, . . . , gN >, and the compact Lie groups, G ⊂ UN . In short, we use the "minimal" framework covering the finitely The axioms are in fact very simple, as follows: RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 3 Definition 1.1. A unitary Hopf algebra is a pair (A, u) formed by a C ∗-algebra A, and a unitary matrix u ∈ MN (A) whose transpose ut is unitary too, such that: (1) The formula ∆(uij) =Pk uik ⊗ ukj defines a morphism ∆ : A → A ⊗ A. (2) The formula ε(uij) = δij defines a morphism ε : A → C. (3) The formula S(uij) = u∗ ji defines a morphism S : A → Aop. We write A = C ∗(Γ) = C(G), and call Γ, G the underlying quantum groups. At the level of basic examples, given a finitely generated group Γ =< g1, . . . , gN >, we have the group algebra A = C ∗(Γ), with u = diag(g1, . . . , gN ). Also, given a compact Lie group G ⊂ UN , we have the algebra A = C(G), with uij(g) = gij. See [29], [30]. We refer to the recent book [21] for a detailed presentation of the theory. Let us explain now what we mean by matrix model for A. We can of course consider embeddings of type A ⊂ MK(C), with K < ∞, but this is not very interesting, because it can only cover the finite dimensional case. Observe that such an embedding always exists at K = ∞, by the GNS theorem, but this is just a theoretical result. The answer comes from the notion of inner faithfulness, introduced in [5]: Definition 1.2. Let π : A → R be a C ∗-algebra representation. (1) The Hopf image of π is the smallest quotient Hopf algebra A → A′ producing a (2) When A = A′, we say that π is inner faithful. That is, we call π : A → R inner factorization of type π : A → A′ → R. faithful when there is no factorization π : A → A′ → R. Here the existence of A′ as in (1) comes from standard Hopf algebra theory. See [5]. As a basic example, when Γ is a classical group, the representation π must come from a unitary group representation Γ → UR, and the factorization in (1) is simply the one obtained by taking the image, Γ → Γ′ ⊂ UR. Thus π is inner faithful when Γ ⊂ UR. Also, given a compact group G, and elements g1, . . . , gK ∈ G, we can consider the representation π = ⊕ievgi : C(G) → CK. The minimal factorization of π is then via C(G′), with G′ = < g1, . . . , gK >. Thus π is inner faithful when G = < g1, . . . , gK >. Finally, observe that the representation A′ → R constructed in Definition 1.2 (1) is Now back to the matrix model problematics, we can formulate: inner faithful. Thus, we have many other potential examples. See [5], [13]. Definition 1.3. A matrix model for A is a C ∗-algebra representation π : A → MK(C) which is inner faithful in the sense of Definition 1.2. When the underlying discrete quantum group Γ is classical, such a model must come from a group embedding Γ ⊂ UK. At the group dual level, given a compact group G, and elements g1, . . . , gK ∈ G, we can consider the representation π : C(G) → MK(C) given by π(ϕ) = diag(ϕ(gi)). By the above, π is a matrix model when G = < g1, . . . , gK >. 4 TEODOR BANICA AND JULIEN BICHON Further examples include the fibers of the Pauli matrix representation of A = C(S+ 4 ), studied in [4], [6], [16]. When dropping the assumption S2 = id, examples appear as well from certain q-deformations of enveloping Lie algebras, with q 6= 1, see [1]. Let us also mention that, given an abstract algebra A satisfying the axioms in Definition 1.1, deciding whether A has or not a matrix model is a subtle analytic problem. See [13]. Let us record, for future reference, the group and group dual statements: Proposition 1.4. Given g1, . . . , gK ∈ UN , consider the discrete group Γ =< g1, . . . , gK >, and the compact group G = Γ. We have then matrix models, as follows: (1) π : C ∗(Γ) → MN (C), mapping g → g. (2) ν : C(G) → MK(C), mapping ϕ → diag(ϕ(gi)). Proof. Both the assertions are elementary, and follow from the above discussion. For full details here, we refer to our previous paper [5]. (cid:3) As explained in the introduction, we are interested here in using matrix models for solving some concrete questions, regarding the random walk on Γ. In order to discuss such questions, we must first study the Haar functional of A. We recall that such a Haar functional exists, thanks to the general results of Woronowicz in [29]. We use in what follows multi-indices of exponents, ε = (ε1, . . . , εp) ∈ {1,∗}p. Given a p,p+1, using the square matrix w ∈ Mn(R), we define w⊗ε ∈ Mnp(R) by wε = wε1 leg-numbering notation, and the standard identification Mnp(R) ≃ Mn(C)⊗p ⊗ R. The general available results on the matrix models can be summarized as follows: 1,p+1 . . . wεp (1) We have F ix(u⊗ε) = F ix(U ⊗ε), where F ix(W ) = {ξW ξ = ξ}. Proposition 1.5. Let π : A → MK(C) be a matrix model, mapping uij → Uij. apbp)a1...ap,b1...bp is the orthogonal projection on F ix(U ⊗ε). 1 G = (tr ◦ π)∗r, with φ ∗ ψ = (φ ⊗ ψ)∆. i1j1 . . . U εp apbp = (T r ε )a1...ap,b1...bp, where (Tε)i1...ip,j1...jp = tr(U ε1 a1b1 . . . uεp G, whereR r r=1R r kPk (2) (RG uε1 (3) RG = limk→∞ (4) R r a1b1 . . . uεp G uε1 ipjp). Proof. These results are known from [5], [7], the proof being as follows: (1) This follows from Tannakian duality [30], see [5]. (2) This follows from (1) and from the Peter-Weyl theory in [29], see [5]. (3) This follows by using idempotent state methods, see [7]. (4) This formula, useful in conjunction with (3), is elementary, see [3], [7]. (cid:3) Let us try now to compute the Kesten type measure µ = law(T r(u)). As a first observation, in the real case, u = ¯u, the character χ = T r(u) is self-adjoint, and by unitarity of u, it satisfies χ ≤ N. Thus in this case µ is a probability measure, supported on [−N, N]. It is well-known that Γ is amenable when N ∈ supp(µ). In general, µ is a ∗-distribution, in the sense of noncommutative probability theory. Such a ∗-distribution is uniquely determined by its ∗-moments. See [22], [26]. We have the following result, coming from Proposition 1.5 above: RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 5 Proposition 1.6. Let µr be the law of χ = T r(u) with respect toR r (1) We have the convergence formula µ = limk→∞ (2) The ∗-moments of µr are cr kPk ε = T r(T r 1 ε ), where (Tε)i1...ip,j1...jp = tr(U ε1 i1j1 . . . U εp ipjp). G = (tr ◦ π)∗r. r=0 µr, in moments. Proof. These results are basically known since [3], the proof being as follows: (1) This follows from the limiting formula in Proposition 1.5 (3). (2) This follows from Proposition 1.5 (4), by summing over ai = bi. The above discussion regarding µ applies to each µr. More precisely, when u = ¯u each (cid:3) Let us prove now that, under suitable assumptions, µr is the law of a certain explicit µr is a probability measure on [−N, N]. In general, µr is a ∗-distribution. matrix. In order to do so, we will need a certain duality operation, as follows: Definition 1.7. Let π : A → MK(C) be a matrix model, mapping uij → Uij. K ) → MN (C) by vkl → U ′ kl. (1) We set (U ′ (2) We perform the Hopf image construction, as to get a model ρ : A′ → MN (C). Here the quantum group U + kl)ij = (Uij)kl, and define eρ : C(U + K is the free analogue of the unitary group UK, constructed K ) is by definition the universal one by Wang in [27]. More precisely, the algebra C(U + generated by the entries of a K × K biunitary matrix v = (vij). See [27]. Thus this matrix is indeed biunitary, and produces a representation ρ as in (1). Observe that the matrix constructed in (1) is given by U ′ = ΣU, where Σ is the flip. The operation A → A′ is a duality, in the sense that we have A′′ = A. Let us first the finite quantum groups, for which we refer to [29], and the basic group/group dual matrix model constructions, from Proposition 1.4 above: discuss a few basic examples, following [3], [5]. We use the duality operation G → bG for Proposition 1.8. With A = C(G), A′ = C(G′), we have: (1) G < ∞ =⇒ G′ = bG. (2) G = < g1, . . . , gK > ⇐⇒ G′ = \< g1, . . . , gK >. Proof. These results are known since [3], [5], the proof being as follows: (1) Assume that (C(G), u) is as in Definition 1.1 above, with G < ∞, and that we have a matrix model π : C(G) → MK(C). We can then construct a fundamental corepresentation for C(G)∗, by the formula wkl(x) = (π(x))kl, and a ∗-representation ρ : C(G)∗ → MN (C), by the formula ρ(ϕ) = (ϕ(ukl))kl. We have: ρ(wab) = (wab(vkl))kl = (ρ(vkl)ab)kl = ((Ukl)ab)kl = ((U ′ ab)kl)kl = U ′ ab examples of matrix models, described in Proposition 1.4 above, are as follows: (2) Given unitaries g1, . . . , gk ∈ UN , set Γ =< g1, . . . , gK > and G = Γ. The standard -- We have a model π : C ∗(Γ) → MN (C), given by g → g. Since we have Uij = δijgi in Thus we have G′ ⊂ bG, and by interchanging G,bG, we obtain the result. See [3]. this case, the associated biunitary matrix is diagonal, U =Pl ell ⊗ gl. 6 TEODOR BANICA AND JULIEN BICHON -- We have a model ν : C(G) → MK(C), given by ν(ϕ) = diag(ϕ(gl)). Here we have Uij = diag((gl)ij) =Pl ell(gl)ij, and so U =Pijl eij ⊗ ell(gl)ij =Pl gl ⊗ ell. Summarizing, for these two models the respective biunitary matrices arePl ell⊗ gl and Pl gl ⊗ ell, related indeed by the flip operation Σ. See [5] for details. distributions, given by the formula Dr(law(X)) = law(rX). We denote by D the dilation operation for probability measures, or for general ∗- We have the following result, extending previous findings from [3]: (cid:3) Theorem 1.9. Consider the rescaled measure ηr = D1/N (µr). r (A′). p/N p of ηr satisfy γr (1) The moments γr (2) ηr has the same moments as the matrix T ′ (3) In the real case u = ¯u we have ηr = law(T ′ p(A) = γp r = Tr(A′). r). p = cr Proof. All results follow from Proposition 1.6 (2), as follows: (1) We have the following computation: cr p(A) = T r(T r (Tp)i1 1...i1 p,i2 1...i2 p . . . . . . (Tp)ir 1...ir p,i1 1...i1 p 1 . . . Ui1 p) . . . . . . tr(Uir pi2 1 . . . Uir p) pi1 1i1 p ) =Xi 1i2 tr(Ui1 = Xi N rXi Xj = 1 (Ui1 1i2 1 )j 1 1 j 1 2 . . . (Ui1 p)j 1 pi2 pj 1 1 . . . . . . (Uir 1i1 1 )jr 2 . . . (Uir 1 jr p)jr pjr 1 pi1 In terms of the matrix (U ′ kl)ij = (Uij)kl, then by permuting the terms in the product on the right, and finally with the changes ib b , we obtain: b , jb a ↔ ia )i1 a ↔ ja p . . . . . . (U ′ pi2 jr 1 jr 2 )i1 1i2 1 . . . (U ′ pj 1 j 1 1 )ir 1i1 1 . . . (U ′ pjr jr 1 )ir pi1 p cr p(A) = = = 1 1 N rXi Xj N rXi Xj N rXi Xj 1 (U ′ j 1 1 j 1 2 (U ′ j 1 1 j 1 2 (U ′ 1 j 2 j 1 1 )i1 1i2 1 )i1 1i1 2 . . . (U ′ jr 1 jr 2 . . . (U ′ j 1 r j 2 r )ir 1i1 1 . . . . . . (U ′ pj 1 j 1 1 )i1 p . . . (U ′ pi2 pjr jr 1 )ir pi1 p )i1 ri1 1 . . . . . . (U ′ jp 1 j 1 1 )ip 2 . . . (U ′ 1ip jp r j 1 r )ip r ip 1 On the other hand, if we use again the above formula of cr matrix U ′, and with the changes r ↔ p and i ↔ j, we obtain: . . . . . . (U ′ jp 1 j 1 . . . (U ′ j 1 r j 2 r cp r(A′) = (U ′ j 1 1 j 2 )i1 )i1 ri1 1i1 1 2 1 1 1 N pXi Xj p(A), but this time for the )ip 2 . . . (U ′ 1ip jp r j 1 r )ip r ip 1 Now by comparing this with the previous formula, we obtain N rcr p(A) = N pcp r(A′). Thus we have cr p(A)/N p = cp r(A′)/N r, and this gives the result. RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 7 (2) By using (1) and the formula in Proposition 1.6 (2), we obtain: cr p(A) N p = cp r(A′) N r = r)p) T r((T ′ N r = tr((T ′ r)p) But this gives the equality of moments in the statement. (3) This follows from the moment equality in (2), and from the standard fact that for (cid:3) self-adjoint variables, the moments uniquely determine the distribution. 2. Projective models In general, the use of the above methods is quite limited. We restrict now attention to a certain special class of models, for which more general theory can be developed. We recall that a square matrix u = (uij) is called "magic" if its entries are projections (p = p2 = p∗), which sum up to 1 on each row and column. The basic example is provided by the matrix coordinates uij : SN ⊂ ON → R, given by uij(σ) = δiσ(j). The following key definition is due to Wang [28]: Definition 2.1. C(S+ N ) is the universal C ∗-algebra generated by the entries of a N × N This algebra satisfies the axioms in Definition 1.1, so the underlying space S+ magic matrix u, with ∆(uij) =Pk uik ⊗ ukj, ε(uij) = δij, S(uij) = uji. N is a compact quantum group, called quantum permutation group. The canonical embedding SN ⊂ S+ in pairs, π : A → MK(C), uij → Uij and π′ : A′ → MN (C), u′ formula being (U ′ from now on to restrict the attention to the case K = N. We have: Definition 2.2. A matrix model π : A → MN (C), with dual model π′ : A′ → MN (C), is called projective if both π, π′ appear from representations of C(S+ Now back to the matrix models, we recall from Definition 1.7 that such models come ij, the connecting ij)kl = (Ukl)ij, or, equivalently, U ′ = ΣU, where Σ is the flip. We agree N is an isomorphism at N = 1, 2, 3, but not at N ≥ 4. See [28]. ij → U ′ N ). In other words, with A = C(G), A′ = C(G′), the projectivity condition states that kl, the N . Equivalently, with U =Pij eij ⊗ Uij and U ′ =Pkl ekl ⊗ U ′ we have G, G′ ⊂ S+ projectivity condition states that both matrices (Uij) and (U ′ The basic examples of such models are those coming from the complex Hadamard matrices. We recall that such a matrix, H ∈ MN (C), has by definition its entries on the unit circle, and its rows are pairwise orthogonal. At the level of examples, the Fourier matrix FG of any finite abelian group G is Hadamard, of size N = G. See [25]. The models associated to the Hadamard matrices are constructed as follows: ij) must be magic. Proposition 2.3. If H ∈ MN (C) is Hadamard, with rows H1, . . . , HN ∈ TN , then produces a projective model. In addition we have U ′(H) = U(H t). Uij = P roj(cid:18) Hi Hj(cid:19) = 1 N(cid:18) HikHjl HilHjk(cid:19)kl 8 TEODOR BANICA AND JULIEN BICHON Proof. The vectors H1, . . . , HN being pairwise orthogonal, we obtain: D Hi Hj , Hi HkE =Xr Hir Hjr · Hkr Hir =Xr Hkr Hjr =< Hk, Hj >= Nδjk A similar computation gives < Hi/Hj, Hk/Hj >= Nδik, so the matrix of rank one projections Uij = P roj(Hi/Hj) is magic. Moreover, by using the formula P roj(ξ) = 1 ξ2 (ξiξj)ij, the projections Uij are indeed given by the formula in the statement. Regarding now the last assertion, this follows from: (U ′ ij)kl = (Ukl)ij = 1 N · HkiHlj HkjHli = 1 N · (H t)ik(H t)jl (H t)jk(H t)il In particular the matrix U ′ is magic as well, and this finishes the proof. (cid:3) We can deform the tensor products of projective models, as follows: Proposition 2.4. Given two projective models π : A → MM (C), ν : B → MN (C), mapping uij → Uij, vij → Vij, the matrix W = U ⊗Q V given by (Wia,jb)kc,ld = QicQjd QidQjc (Uij)kl(Vab)cd produces a projective model, for any choice of the parameter matrix Q ∈ MM ×N (T). Proof. Let us first check that the elements Wia,jb are self-adjoint. We have indeed: (W ∗ ia,jb)kc,ld = (Wia,jb)ld,kc = ¯Qid ¯Qjc ¯Qic ¯Qjd (Uij)lk · (Vab)dc = QicQjd QidQjc (Uij)kl(Vab)cd = (Wia,jb)kc,ld We verify now the magic condition. First, we have: (Wia,jbWia,me)kc,ld = Xnf = Xnf = Xf (Wia,jb)kc,nf (Wia,me)nf,ld QicQjf Qif Qjc (Uij)kn(Vab)cf Qif Qmd QidQmf QicQjf Qmd QjcQidQmf (Vab)cf (Vae)f dXn (Uim)nl(Vae)f d (Uij)kn(Uim)nl RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 9 The last sum on the right being (UijUim)kl = δjm(Uij)kl, we obtain: (Wia,jbWia,me)kc,ld = δjmXf QicQjd QjcQid (Vab)cf (Vae)f d(Uij)kl = δjm QicQjd QjcQid (Uij)klXf (Vab)cf (Vae)f d The last sum on the right being (VabVae)cd = δbe(Vab)cd, we obtain: (Wia,jbWia,me)kc,ld = δjmδbe QicQjd QjcQid (Uij)kl(Vab)cd = δjmδbe(Wia,jb)kc,ld Thus the elements Wia,jb are indeed projections, which are pairwise orthogonal on rows. In order to conclude that W is magic, we check that the sum on the columns is 1: Xia (Wia,jb)kc,ld =Xi QicQjd QidQjc (Uij)klXa (Vab)cd = δcdXi (Uij)kl = δcdδkl It remains to prove that W ′ is a projective model too. Since W is a model, so is W ′, so it suffices to prove that the entries of W ′ are projections. We have: (W ′ ia,jb)kc,ld = (Wkc,ld)ia,jb = QkaQlb QkbQla (Ukl)ij(Vcd)ab = QkaQlb QkbQla Now since both U ′, V ′ are projective models, we obtain: (U ′ ij)kl(V ′ ab)cd (W ′∗ ia,jb)kc,ld = (W ′ ia,jb)ld,kc = ¯Qla ¯Qkb ¯Qlb ¯Qka (U ′ ij)lk · (V ′ ab)dc = QlbQka QlaQkb (U ′ ij)kl(V ′ ab)cd = (W ′ ia,jb)kc,ld Finally, the ckeck of the idempotent condition goes as follows: ia,jb)kc,nf (W ′ ia,jb)nf,ld (U ′ ij)kn(V ′ ab)cf QnaQlb QnbQla (U ′ ij)nl(V ′ ab)f d (W ′2 (W ′ ia,jb)kc,ld = Xnf = Xnf QkbQlaXf QkaQnb QkbQna QkaQlb = (V ′ ab)cf (V ′ (U ′ ij)kn(U ′ ij)nl ab)f dXn = QkaQlb QkbQla (V ′ ab)cd(U ′ ij)kl = (W ′ ia,jb)kc,ld We conclude that W ′ is magic too, and this finishes the proof. (cid:3) 10 TEODOR BANICA AND JULIEN BICHON As an example, assume that we are given two Hadamard matrices, H ∈ MM (C) and K ∈ MN (C), and let us form the deformed tensor product H ⊗Q K = (QibHijKab)ia,jb, which is Hadamard as well. See [25]. The model associated to this matrix is: (Uia,jb)kc,ld = 1 MN · (QicHikKac)(QjdHjlKbd) (QidHilKad)(QjcHjkKbc) = QicQjd QidQjc (U H ij )kl(U K ab )cd Thus, the ⊗Q operations for models and for Hadamard matrices are compatible. Another theoretical remark is that, in the projective model framework, the operation constructed in Proposition 2.4 has a dual counterpart, constructed as follows: Proposition 2.5. With U, V as in Proposition 2.4, the matrix W ◦ = U Q⊗ V given by (W ◦ ia,jb)kc,ld = (Uij)kl(Vab)cd QkaQlb QkbQla produces a projective model. We have W ′ = U ′ Q⊗ V ′ and W ◦′ = U ′ ⊗Q V ′. Proof. We use the following formula, already met in the proof of Proposition 2.4: (W ′ ia,jb)kc,ld = (Wkc,ld)ia,jb = QkaQlb QkbQla (Ukl)ij(Ucd)ab = QkaQlb QkbQla (U ′ ij)kl(V ′ ab)cd we have W ′ = U ′ With the convention in the statement for the products Q⊗, this means precisely that Q⊗ V ′. The last assertion is proved similarly, because we have: ab)cd (Ukl)ij(Vcd)ab = kc,ld)ia,jb = ij)kl(V ′ (U ′ (W ◦′ ia,jb)kc,ld = (W ◦ QicQjd QidQjc QicQjd QidQjc Finally, the fact that the matrix W ◦ produces indeed a projective model follows from (cid:3) Proposition 2.4, by using the two connecting formulae that we just proved. Once again, we have here a compatibility with the known complex Hadamard ma- trix constructions, and namely with the operation H Q⊗ K = (QjaHijKab)ia,jb from [25]. Indeed, the projective model associated to such a matrix is: (Uia,jb)kc,ld = 1 MN · (QkaHikKac)(QlbHjlKbd) (QlaHilKad)(QkbHjkKbc) = QkaQlb QlaQkb (U H ij )kl(U K ab )cd As a last theoretical result about the deformed tensor products, constructed in Propo- sition 2.4 and Proposition 2.5 above, here is an alternative definition for them: Proposition 2.6. We have fW = eU13QδeV24 andgW ◦ = eV24QδeU13, where and where the U → eU operation is defined by (eUij)kl = (Ujl)ik. Qδ = diag(cid:18)QicQjd QidQjc(cid:19)icjd RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 11 Proof. According to the definition of W in Proposition 2.4, we have: fW = XiajbXkcld = XiajbXkcld = eU13 Xjalc eia,jb ⊗ ekc,ld QjaQlc QjcQla (Ujl)ik(Vbd)ac eij ⊗ eab ⊗ ekl ⊗ ecd ejj ⊗ eaa ⊗ ell ⊗ ecc QjaQlc QjcQla (eUij)kl(eVab)cd QjcQla!eV24 QjaQlc We recognize in the middle the diagonal matrix Qδ in the statement, and we are there- fore done with the proof of the first formula. Similarly, we have: gW ◦ = XiajbXkcld = XiajbXkcld = eV24 Xibkd eia,jb ⊗ ekc,ld QibQkd QkbQid (Ujl)ik(Vbd)ac eij ⊗ eab ⊗ ekl ⊗ ecd eii ⊗ ebb ⊗ ekk ⊗ edd QibQkd QkbQid (eUij)kl(eVab)cd QkbQid!eU13 QibQkd But this gives the second formula in the statement, and we are done. (cid:3) We should mention the above description of the deformed tensor products is in fact map models to models, or projective models to projective models, even in the most simple cases. As an example here, for a Fourier matrix model, (Uij)kl = Fi−j,k−l, where F = FX not very enlightening, because the operation U → eU given by (eUij)kl = (Ujl)ik does not is the Fourier matrix of a finite abelian group X, we have (eUij)kl = Fj−l,i−k. Now since we have (eU ∗ ij)kl = Fj−k,l−i, we see that the matrices eUij are not self-adjoint. Let us study now the truncated moments. First, we have: Lemma 2.7. The truncated moments for W = U ⊗Q V are given by 1Qi1 2Qi1 ∆U (i)∆V ′(bt) pQi2 Qi2 Qi1 Qi1 Qi1 Qi1 Qi2 Qi2 Qir Qir . . . . . . cr p = . . . pb1 pb1 1br 1br 1b1 2 1b1 1 1 pb1 p 1b1 2 1b1 1 pb1 1 1 (MN)rXib 1br 2 1br 1 . . . Qir Qir pQi1 1Qi1 pbr pbr pbr 1 pbr p where ∆U (i) = M r · (T U p )i1 1...i1 p,i2 1...i2 p . . . . . . (T U p )ir 1...ir p,i1 1...i1 p, for i ∈ Mr×p(1, . . . , M). 12 TEODOR BANICA AND JULIEN BICHON Proof. We will use several times, in forward and in backwards form, the following com- putation, which already appeared in the proof of Theorem 1.9 (1) above: cr 1 1i2 p,i2 1...i1 1...i2 (T U p )i1 tr(Ui1 . . . Ui1 p(U) = Xi = Xi M rXi Xj (MN)rXia Xjb cr p(W ) = (Ui1 = 1i2 1 1 1 In double index notation, with Uij replaced by Wia,jb, the formula is: p . . . . . . (T U p )ir 1...ir p,i1 1...i1 p p) . . . . . . tr(Uir pi2 1i1 1 )j 1 1 j 1 2 . . . (Ui1 p)j 1 pi2 pj 1 1 . . . Uir p) pi1 . . . . . . (Uir 1i1 1 )jr 2 . . . (Uir 1 jr p)jr pjr 1 pi1 (Wi1 1a1 1,i2 1a2 1 )j 1 1 b1 1,j 1 2b1 2 . . . . . . (Wi1 pa1 p,i2 pa2 p)j 1 pb1 p,j 1 1b1 1 . . . . . . (Wir 1ar 1,i1 1a1 1 )jr 1 br 1,jr 2 br 2 . . . . . . (Wir par p,i1 pa1 p)jr pbr p,jr 1 br 1 Now with Wia,jb being as in Proposition 2.4 above, we obtain: cr p(W ) = 1 (MN)rXib Xj Xa Qi1 Qi1 1b1 1 1b1 2 Qi2 Qi2 1b1 2 1b1 1 . . . Qi1 Qi1 pb1 pb1 1 pQi2 Qi2 pb1 1 pb1 p . . . . . . Qir Qir 1Qi1 2Qi1 1br 1br 1br 2 1br 1 . . . Qir Qir pQi1 1Qi1 pbr pbr pbr 1 pbr p (Ui1 1i2 1 )j 1 1 j 1 2 . . . (Ui1 p)j 1 pi2 pj 1 1 . . . . . . (Uir 1i1 1 )jr 2 . . . (Uir 1 jr p)jr pjr 1 pi1 (Va1 1)b1 1b1 2 . . . (Va1 1a2 p)b1 pa2 pb1 1 . . . . . . (Var 1)br 2 . . . (Var 1br 1a1 p)br pbr 1 pa1 The middle sum can be compacted by using the computation in the beginning of this proof. The last sum can be compacted too, by using a similar computation, after switching indices by using (Vab)cd = (V ′ cr p(W ) = 1 (MN)rXib 1 2 1 1b1 1b1 1b1 pb1 pb1 . . . Qi1 Qi1 cd)ab. We obtain the following formula: Qi2 Qi1 1Qi1 2Qi1 Qi1 Qi2 M r · (T U p )i1 N p · (T V ′ pQi2 Qi2 p . . . . . . (T U p )ir 2 . . . . . . (T V ′ Qir Qir . . . . . . r )b1 r )b1 p...br 1...br 2...br 1...ir 1...i1 1...i2 p,b1 p,i2 p,i1 1,b1 pb1 pb1 p 1br 1br 1b1 2 1 1 1...i1 p 1...br 1 1br 2 1br 1 . . . Qir Qir pQi1 1Qi1 pbr pbr pbr 1 pbr p But this gives the formula in the statement, and we are done. (cid:3) In order to further advance, we use the following notion: Definition 2.8. A model π : A → MN (C), mapping uij → Uij, is called positive if for any p ∈ N, and any choice of the indices i1, . . . , ip and j1, . . . , jp. tr(Ui1j1 . . . Uipjp) ≥ 0 RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 13 In other words, the model is called positive if the functionalR 1 G = tr◦π from Proposition 1.5 (3) is positive on all the products of standard coordinates ui1j1 . . . uipjp. Equivalently, the matrix T U p in Proposition 1.5 (4) must have positive entries, for any p ∈ N. Once again, the basic examples here come from the Hadamard matrices. In the context of Proposition 2.3 above, with the notations there, we have: tr(Ui1j1 . . . Uipjp) = = = = 1 1 NXk N p+1Xk N p+1Xk1 N p+1(cid:28) Hi1 1 1 Hj1 N < Ha Hb (Ui1j1)k1k2 . . . . . . (Uipjp)kpk1 Hi1k1Hj1k2 Hi1k2Hj1k1 . . . . . . HipkpHjpk1 Hipk1Hjpkp Hi1k1Hjpk1 Hj1k1Hipk1 . . . . . .Xkp Hjp(cid:29) . . . . . .(cid:28) Hip Hjp Hip , HipkpHjp−1kp HjpkpHip−1kp , Hip−1 Hjp−1(cid:29) In particular, if the quantities Cabcd = 1 > are all positive, then the positivity condition is satisfied. Observe that this is the case for the Fourier matrix FX of a finite abelian group X, where we have Cabcd = δa−b,c−d, with all indices taken in X. , Hc Hd Now back to the general case, we have the following result, that we believe of interest, and which is the best one that we could find at the abstract level: Theorem 2.9. If U, V ′ come from positive projective models, with W = U ⊗Q V we have: cr p(W ) ≤ cr p(U)cr p(V ) Thus, the moments of µr Proof. By using Qij = 1 for any i, j, the formula in Lemma 2.7 gives: W are bounded by those of the usual tensor product. cr p(W ) ≤ 1 (MN)rXib ∆U (i)∆V ′(bt) Now observe that the computation in the beginning of the proof of Lemma 2.7 reads p(U). Thus, assuming that we have positivity, the ∆ quantities on the right are both positive, we can remove the absolute value sign, and we obtain: Pi ∆U (i) = M rcr cr p(W ) ≤ 1 (MN)rXib = N p−rcr = cr p(U)cr p(U)cp p(V ) ∆U (i)∆V ′(bt) = r(V ′) = N p−rcr 1 (MN)r · M rcr p(V ) p(U)N r−pcr p(U) · N pcp r(V ′) Here we have used Theorem 1.9 (1). Now since this formula tells us that the moments (cid:3) V , this gives the last assertion as well. W are bounded by those of µr of µr U × µr 14 TEODOR BANICA AND JULIEN BICHON 3. Abelian groups In this section we further restrict the attention, to a very special class of projective models. Generally speaking, the problem is that the complex Hadamard matrices, which are the main source of projective models, are quite complicated objects, and the only elementary example is the Fourier matrix FX of a finite abelian group X. See [25]. Let us first recall the construction of this matrix: Proposition 3.1. Let X = ZN1 × . . . × ZNk be a finite abelian group, and consider the matrix FX = FN1 ⊗ . . . ⊗ FNk , where FN = (wij), with w = e2πi/N . (1) In the cyclic group case, X = ZN , we have FX = FN . (2) In general, FX is the matrix of the Fourier transform over X. (3) With F = FX we have Fi+j,k = FikFjk, Fi,j+k = FijFik, F−i,j = Fi,−j = ¯Fij. Proof. All these results are well-known: (1) This is clear from definitions. (2) This is well-known in the cyclic group case, and in general, it follows by using the compatibility between the product of groups × and the tensor product of matrices ⊗. (3) This is clear in the cyclic group case, and then in general as well. (cid:3) Observe that each FN , and hence each FX , is a complex Hadamard matrix. Now let us go back to Proposition 2.3 above. By using the formulae in (3) above, we see that the matrix constructed there, with H = FX, is given by: (Uij)kl = 1 N · FikFjl FilFjk = 1 N (FikFi,−l)(F−j,kF−j,−l) = 1 N Fi,k−lF−j,k−l = 1 N Fi−j,k−l Thus, the projective models associated to the Fourier matrices, coming from Proposition 2.3 above, can be in fact introduced directly, as follows: Definition 3.2. Associated to a finite abelian group X is the projective model coming from the matrix (Uij)kl = 1 π : C(X) → MX(C) N Fi−j,k−l, where F = FX. Observe that the models U, U ′ fall into the general framework of Proposition 1.8 (2) above, but with both U, U ′ being twisted by the Fourier transform. Now let X, Y be finite abelian groups, and let us try to understand the projective model constructed by deforming the tensor product of the corresponded Fourier models: RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 15 Definition 3.3. Given two finite abelian groups X, Y , we consider the corresponding Fourier models U, V , we construct W = U ⊗Q V as in Proposition 2.5, and we factorize C(S+ X×Y ) πQ MX×Y (C) %❑❑❑❑❑❑❑❑❑❑ 9rrrrrrrrrr π C(GQ) with C(GQ) being the Hopf image of πQ, as in Definition 1.2. Explicitely computing the compact quantum group GQ, as function of the parameter matrix Q ∈ MX×Y (T), and understanding the random walk on the corresponding group dual ΓQ = bGQ, will be our main purpose, in the reminder of this paper. In order to do so, we use the following notion, from [11]: Definition 3.4. Let C(S+ damental corepresentations denoted u, v. We let M ) → A and C(S+ N ) → B be Hopf algebra quotients, with fun- A ∗w B = A∗N ∗ B/ < [u(i) with the Hopf algebra structure making wia,jb = u(i) ab , vij] = 0 > ab vij a corepresentation. The fact that we have indeed a Hopf algebra follows from the fact that w is magic. In terms of quantum groups, if A = C(G), B = C(H), we write A ∗w B = C(G ≀∗ H): C(G) ∗w C(H) = C(G ≀∗ H) The ≀∗ operation is then the free analogue of ≀, the usual wreath product. See [11]. We will need as well the following elementary lemma: Lemma 3.5. If X is a finite abelian group then C(X) = C(S+ X)/ < uij = ukl∀i − j = k − l > with all the indices taken inside X. Proof. Observe first that C(Y ) = C(S+ X )/ < uij = ukl∀i − j = k − l > is commutative, because uijukl = uijui,l−k+i = δj,l−k+iuij and ukluij = ui,l−k+iuij = δj,l−k+iuij. Thus we have Y ⊂ SX , and since uij(σ) = δiσ(j) for any σ ∈ Y , we obtain: i − j = k − l =⇒ (σ(j) = i ⇐⇒ σ(l) = k) But this condition tells us precisely that σ(i) − i must be independent on i, and so σ(i) = i + x for some x ∈ X, and so σ ∈ X, as desired. (cid:3) We can now factorize representation πQ in Definition 3.3, as follows: / / % 9 16 TEODOR BANICA AND JULIEN BICHON Proposition 3.6. We have a factorization C(S+ X×Y ) πQ MX×Y (C) &◆◆◆◆◆◆◆◆◆◆ 8♣♣♣♣♣♣♣♣♣♣♣ π C(Y ≀∗ X) ab =Pj Wia,jb and by Vij =Pa Wia,jb, independently of b. given by U (i) Proof. With K = FX , L = FY and M = X, N = Y , the formula of the magic matrix W ∈ MX×Y (MX×Y (C)) associated to H = K ⊗Q L is: = (Wia,jb)kc,ld = QicQjd QidQjc · Ki−j,k−lLa−b,c−d 1 MN · QicQjd QidQjc · KikKjl KilKjk · LacLbd LadLbc 1 MN · Our claim that the representation πQ constructed in Definition 3.3 can be factorized in three steps, up to the factorization in the statement, as follows: C(S+ X×Y ) πQ MX×Y (C) C(S+ Y ≀∗ S+ X) C(S+ Y ≀∗ X) C(Y ≀∗ X) Indeed, the construction of the map on the left is standard, see [11], and this produces the first factorization. Regarding the second factorization, this comes from the fact that since the elements Vij depend on i − j, they satisfy the defining relations for the quotient algebra C(S+ X ) → C(X), coming from Lemma 3.5. Finally, regarding the third factoriza- tion, observe that the above matrix Wia,jb depends only on a − b. By summing over j we obtain that U (i) ab depends only on a − b, and by using Lemma 3.5, we are done. (cid:3) In order to further factorize the representation in Proposition 3.6, we use: Definition 3.7. If H y Γ is a finite group acting by automorphisms on a discrete group, the corresponding crossed coproduct Hopf algebra is C ∗(Γ) ⋊ C(H) = C ∗(Γ) ⊗ C(H) Observe that C(H) is a subcoalgebra, and that C ∗(Γ) is not a subcoalgebra. The with comultiplication ∆(r ⊗ δk) =Ph∈H(r ⊗ δh) ⊗ (h−1 · r ⊗ δh−1k), for r ∈ Γ, k ∈ H. quantum group corresponding to C ∗(Γ) ⋊ C(H) is denotedbΓ ⋊ H. Now back to the factorization in Proposition 3.6, the point is that we have: Lemma 3.8. With L = FY , N = Y we have an isomorphism given by vij → 1 ⊗ vij and u(i) C(Y ≀∗ X) ≃ C ∗(Y )∗X ⋊ C(X) ab = 1 NPc Lb−a,cc(i) ⊗ 1. / / & 8 / / / / 3 3 / / 7 7 O O RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 17 Proof. We know that C(Y ≀∗ X) is the quotient of C(Y )∗X ∗ C(X) by the relations [u(i) ab , vij] = 0. Now since vij depends only on j − i, we obtain [u(i) ab , vi,l−k+i] = 0, and so we are in a usual tensor product situation, and we have: ab , vkl] = [u(i) C(Y ≀∗ X) = C(Y )∗X ⊗ C(X) Let us compose now this identification with Φ∗X ⊗ id, where Φ : C(Y ) → C ∗(Y ) is the Fourier transform. We obtain an isomorphism as in the statement, and since Φ(uab) = 1 (cid:3) ab is indeed the one in the statement. Here is now our key lemma, which will lead to further factorizations: NPc Lb−a,cc, the formula for the image of u(i) Lemma 3.9. With c(i) =Pa Lacu(i) a0 and εke =Pi Kikeie we have: Qi,e−cQi−k,e QieQi−k,e−c εk,e−c π(c(i))(εke) = In particular if c1 + . . . + cs = 0 then π(c(i1) 1 . . . c(is) s ) is diagonal, for any i1, . . . , is. Proof. We have the following formula: π(c(i)) =Xa Lacπ(u(i) a0) =Xaj LacWia,j0 On the other hand, in terms of the basis in the statement, we have: Wia,jb(εke) = 1 N δi−j,kXd QidQje QieQjd La−b,d−eεkd We therefore obtain, as desired: π(c(i))(εke) = 1 N Xad = Xd Lac QidQi−k,e QieQi−k,d La,d−eεkd = 1 NXd QidQi−k,e QieQi−k,d εkdXa La,d−e+c QidQi−k,e QieQi−k,d εkdδd,e−c = Qi,e−cQi−k,e QieQi−k,e−c εk,e−c r Regarding now the last assertion, this follows from the fact that each matrix of type π(c(ir) ) acts on the standard basis elements εke by preserving the left index k, and by rotating by cr the right index e. Thus when we assume c1 + . . . + cs = 0 all these rotations compose up to the identity, and we obtain indeed a diagonal matrix. (cid:3) We have now all needed ingredients for refining Proposition 3.6: 18 TEODOR BANICA AND JULIEN BICHON Theorem 3.10. We have a factorization as follows, C(S+ X×Y ) πQ (◗◗◗◗◗◗◗◗◗◗◗◗ C ∗(ΓX,Y ) ⋊ C(X) MX×Y (C) ρ 6♠♠♠♠♠♠♠♠♠♠♠♠♠ ] = 1Pr cr =Pr dr = 0 >. . . . c(is) where ΓX,Y = Y ∗X / < [c(i1) Proof. Assume that we have a representation π : C ∗(Γ) ⋊ C(X) → ML(C), let Λ be a X-stable normal subgroup of Γ, so that X acts on Γ/Λ and that we can form the crossed coproduct C ∗(Γ/Λ) ⋊ C(X), and assume that π is trivial on Λ. Then π factorizes as: . . . d(js) , d(j1) 1 s s 1 C ∗(Γ) ⋊ C(X) π ML(C) )❙❙❙❙❙❙❙❙❙❙❙❙❙❙ 6♥♥♥♥♥♥♥♥♥♥♥♥ ρ C ∗(Γ/Λ) ⋊ C(X) With Γ = Y ∗X , and by using Lemma 3.8 and Lemma 3.9, this gives the result. (cid:3) 4. Formal deformations In general, further factorizing the representation found in Theorem 3.10 above is a quite complicated task. In this section we restrict attention to the case where the parameter matrix Q is generic, in the sense that its entries are as algebrically independent as possible, and we prove that the representation in Theorem 3.10 is the minimal one. Our starting point is the group ΓX,Y found above: Definition 4.1. Associated to two finite abelian groups X, Y is the discrete group ΓX,Y = Y ∗X.*[c(i1) 1 . . . c(is) s , d(j1) 1 . . . d(js) s ] = 1(cid:12)(cid:12)(cid:12)Xr cr =Xr dr = 0+ where the superscripts refer to the X copies of Y , inside the free product. We will need a more convenient description of this group. The idea here is that the above commutation relations can be realized inside a suitable semidirect product. Given a group acting on another group, H y G, we denote as usual by G ⋊ H the semidirect product of G by H, i.e. the set G × H, with multiplication (a, s)(b, t) = (as(b), st). Now given a group G, and a finite abelian group Y , we can make Y act on GY , and form the product GY ⋊ Y . Since the elements of type (g, . . . , g) are invariant, we can form as well the product (GY /G) ⋊ Y , and by identifying GY /G ≃ GY −1 via the map (1, g1, . . . , gY −1) → (g1, . . . , gY −1), we obtain a product GY −1 ⋊ Y . With these notations, we have the following result: / / ( 6 / / ) 6 RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 19 Proposition 4.2. The group ΓX,Y has the following properties: (1) ΓX,Y ≃ Z(X−1)(Y −1) ⋊ Y . (2) ΓX,Y ⊂ Z(X−1)Y ⋊ Y via c(0) → (0, c) and c(i) → (bi0 − bic, c) for i 6= 0, where bic are the standard generators of Z(X−1)Y . Proof. We prove these assertions at the same time. We must prove that we have group morphisms, given by the formulae in the statement, as follows: ΓX,Y ≃ Z(X−1)(Y −1) ⋊ Y ⊂ Z(X−1)Y ⋊ Y 1 . . . C (is) Our first claim is that the formula in (2) defines a morphism ΓX,Y → Z(X−1)Y ⋊ Y . Indeed, the elements (0, c) produce a copy of Y , and since we have a group embedding Y ⊂ ZY ⋊ Y given by c → (b0 − bc, c), the elements C (i) = (bi0 − bic, c) produce a copy of Y , for any i 6= 0. In order to check now the commutation relations, observe that: C (i1) s = bi10 − bi1c1 + bi2c1 − bi2,c1+c2 + . . . + bis,c1+...+cs−1 − bis,c1+...+cs,Xr cr! Thus Pr cr = 0 implies C (i1) abelian group, we have the commutation relations, and our claim is proved. ∈ Z(X−1)Y , and since we are now inside an Using the considerations before the statement of the proposition, it is routine to con- struct an embedding Z(X−1)(Y −1) ⋊ Y ⊂ Z(X−1)Y ⋊ Y such that we have group mor- phisms whose composition is the group morphism just constructed, as follows: . . . C (is) 1 s s . . . c(is) It remains to prove that the map on the left is injective. For this purpose, consider the morphism ΓX,Y → Y given by c(i) → c, whose kernel T is formed by the elements of type c(i1) 1 ΓX,Y → Z(X−1)(Y −1) ⋊ Y ⊂ Z(X−1)Y ⋊ Y , withPr cr = 0. We get an exact sequence, as follows: 1 → T → ΓX,Y → Y → 1 This sequence splits by c → c(0), so we have ΓX,Y ≃ T ⋊ Y . Now by the definition of ΓX,Y , the subgroup T constructed above is abelian, and is moreover generated by the elements (−c)(0)c(i), i, c 6= 0. Finally, the fact that T is freely generated by these elements follows from the computation in the proof of Lemma 4.4 below. (cid:3) Let us specify now what our genericity assumptions are: Definition 4.3. We use the following notions: pr1 1 . . . prm (1) We call p1, . . . , pm ∈ T root independent if for any r1, . . . , rm ∈ Z we have (2) A matrix Q ∈ MX×Y (T), taken to be dephased (Q0c = Qi0 = 1), is called generic m = 1 =⇒ r1 = . . . = rm = 0. if the elements Qic, with i, c 6= 0, are root independent. We will need the following lemma: 20 TEODOR BANICA AND JULIEN BICHON Lemma 4.4. Assume that Q ∈ MX×Y (T) is generic, and put θke ic = Qi,e−cQi−k,e QieQi−k,e−c For every k ∈ X, we have a representation πk : ΓX,Y → UY given by πk(c(i))ǫe = θke ic ǫe−c. The family of representations (πk)k∈X is projectively faithful in the sense that if for some t ∈ ΓX,Y , we have that πk(t) is a scalar matrix for any k, then t = 1. Proof. The representations πk arise from Lemma 3.9. With ΓX,Y = T ⋊ Y , as in the proof of Proposition 4.2, we see that for t ∈ ΓX,Y such that πk(t) is a scalar matrix for any k, then t ∈ T , since the elements of T are the only ones having their image by πk formed by the proof of Proposition 4.2, for Ric ∈ Z, and consider the quantities: diagonal matrices. Now write t =Qi6=0,c6=0((−c)(0)(c)(i))Ric with the generators of T as in A(k, e) = Yi6=0Yc6=0 = Yi6=0Yc6=0 = Yj6=0Yc6=0 jc )Rjc ·Yc6=0Yj6=0 ic )Ric ·Yc6=0 jc )Rjc+Pi6=0 Ric (θke (θke ic )Ric(θke )Ric =Yi6=0Yc6=0 0c )−Pi6=0 Ric =Yj6=0Yc6=0 (θke jc )Pi6=0 Ric (θke (θke ic (θke 0c ) −1 (θke 0c )−Ric (θke We have πk(t)(ǫe) = A(k, e)ǫe for any k, e. Our assumption is that for any k, we have A(k, e) = A(k, f ) for any e, f . Using the root independence of the elements Qic, i, c 6= 0, we see that this implies Ric = 0 for any i, c, and this proves our assertion. (cid:3) We will need as well the following lemma: Lemma 4.5. Let π : C ∗(Γ) ⋊ C(H) → L be a surjective Hopf algebra map, such that πC(H) is injective, and such that for r ∈ Γ and f ∈ C(H), we have: Then π is an isomorphism. π(r ⊗ 1) = π(1 ⊗ f ) =⇒ r = 1 Proof. We use here various tools from [2], [24]. Put A = C ∗(Γ) ⋊ C(H). We start with the following Hopf algebra exact sequence, where i(f ) = 1 ⊗ f and p = ε ⊗ 1: C → C(H) i→ A p → C ∗(Γ) → C Since π ◦ i is injective, and Hopf subalgebra π ◦ i(C(H)) is central in L, we can form the quotient Hopf algebra L = L/(π ◦ i(C(H))+L, and we get another exact sequence: C → C(H) π◦i−−→ L q → L → C RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 21 Note that this sequence is indeed exact, e.g. by centrality. So we get the following diagram with exact rows, with the Hopf algebra map on the right surjective: C C / C(H) i / A π / C(H) π◦i / L p q / C ∗(Γ) C / L / C Since a quotient of a group algebra is still a group algebra, we get a commutative diagram with exact rows as follows: C C / C(H) i / A π / C(H) π◦i / L p q′ / C ∗(Γ) C / C ∗(Γ) / C Here the Hopf algebra map on the right is induced by a surjective morphism u : Γ → Γ, g 7→ g. By the five lemma we just have to show that u is injective. So, let g ∈ Γ be such that u(g) = 1. Then q′π(g ⊗ 1) = up(g ⊗ 1) = u(g) = g = 1. For g ∈ Γ, put: gA = {a ∈ A p(a1) ⊗ a2 = g ⊗ a} gL = {l ∈ L q′(l1) ⊗ l2 = g ⊗ l} The commutativity of the right square ensures that π(gA) ⊂ gL. Then with the previous g, we have π(g ⊗ 1) ∈ 1L = πi(C(H)) (exactness of the sequence), so π(g ⊗ 1) = π(1 ⊗ f ) for some f ∈ C(H). We conclude by our assumption that g = 1. We have now all ingredients for proving our first main result: (cid:3) Theorem 4.6. When Q is generic, the minimal factorization for πQ is C(S+ X×Y ) πQ MX×Y (C) (◗◗◗◗◗◗◗◗◗◗◗◗ 6♠♠♠♠♠♠♠♠♠♠♠♠♠ π C ∗(ΓX,Y ) ⋊ C(X) where ΓX,Y ≃ Z(X−1)(Y −1) ⋊ Y is the discrete group constructed above. Proof. We want to apply Lemma 4.5 to the morphism θ : C ∗(ΓX,Y ) ⋊ C(X) → L arising from the factorization in Theorem 3.10, where L denotes the Hopf image of πQ, which / / / / / / / / / / / / / / / / / / / / ( 6 22 TEODOR BANICA AND JULIEN BICHON produces the following commutative diagram (see [5]): C(S+ X×Y ) (❘❘❘❘❘❘❘❘❘❘❘❘❘❘❘ MX×Y (C) 6❧❧❧❧❧❧❧❧❧❧❧❧❧❧❧❧ π πQ L θ C ∗(ΓX,Y ) ⋊ C(X) The first observation is that the injectivity assumption on C(X) holds by construction, and that for f ∈ C(X), the matrix π(f ) is "block scalar", the blocks corresponding to the indices k in the basis εke in the basis from Lemma 3.9. Now for r ∈ ΓX,Y with θ(r ⊗ 1) = θ(1 ⊗ f ) for some f ∈ C(X), we see, using the commutative diagram, that we will have that π(r ⊗ 1) is block scalar. By Lemma 4.4, the family of representations (πk) of ΓX,Y , corresponding to the blocks k, is projectively faithful, so r = 1. We can apply indeed Lemma 4.5, and we are done. (cid:3) 5. Random walks In this section we compute the Kesten type measure µ = law(χ) for the quantum group G = GQ found in Theorem 4.6. Our results here will be a combinatorial moment formula, a geometric interpretation of it, and an asymptotic convergence result. The moment formula is as follows: Proposition 5.1. We have the moment formula ZG χp = 1 X · Y #(cid:26) i1, . . . , ip ∈ X d1, . . . , dp ∈ Y(cid:12)(cid:12)(cid:12) [(i1, d1), (i2, d2), . . . , (ip, dp)] = [(i1, dp), (i2, d1), . . . , (ip, dp−1)](cid:27) where the sets between square brackets are by definition sets with repetition. Proof. According to the various formulae in sections 2 and 3 above, the factorization found in Theorem 4.6 is, at the level of standard generators, as follows: X×Y ) → C ∗(ΓX,Y ) ⊗ C(X) → MX×Y (C) uia,jb → 1 C(S+ Y Pc Fb−a,cc(i) ⊗ vij → Wia,jb c(i)! ⊗ δ1 c(i) ⊗ vii =Xic c(i) ⊗ vii = Xic χ = 1 Y Xiac Thus, the main character is given by: / / ( % % 6 O O ✤ ✤ ✤ @ @ RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 23 Now since the Haar functional of C ∗(Γ) ⋊ C(H) is the tensor product of the Haar functionals of C ∗(Γ), C(H), this gives the following formula, valid for any p ≥ 1: ZG χp = 1 XZbΓX,Y Xic c(i)!p Let Si = Pc c(i). By using the embedding in Proposition 4.2 (2), with the notations there we have Si =Pc(bi0 − bic, c), and these elements multiply as follows: Si1 . . . Sip = Xc1...cp bi10 − bi1c1 + bi2c1 − bi2,c1+c2 +bi3,c1+c2 − bi3,c1+c2+c3 + . . . . . . . . . . . . + bip,c1+...+cp−1 − bip,c1+...+cp , c1 + . . . + cp In terms of the new indices dr = c1 + . . . + cr, this formula becomes:   Si1 . . . Sip = Xd1...dp bi10 − bi1d1 + bi2d1 − bi2d2 +bi3d2 − bi3d3 + . . . . . . . . . . . . + bipdp−1 − bipdp , dp   Now by integrating, we must have dp = 0 on one hand, and on the other hand: [(i1, 0), (i2, d1), . . . , (ip, dp−1)] = [(i1, d1), (i2, d2), . . . , (ip, dp)] Equivalently, we must have dp = 0 on one hand, and on the other hand: [(i1, dp), (i2, d1), . . . , (ip, dp−1)] = [(i1, d1), (i2, d2), . . . , (ip, dp)] Thus, by translation invariance with respect to dp, we obtain: It follows that we have the following moment formula: Si1 . . . Sip = ZbΓX,Y ZbΓX,Y Xi 1 Y #(cid:26)d1, . . . , dp ∈ Y(cid:12)(cid:12)(cid:12) #(cid:26) i1, . . . , ip ∈ X d1, . . . , dp ∈ Y(cid:12)(cid:12)(cid:12) Si!p = 1 Y [(i1, d1), (i2, d2), . . . , (ip, dp)] = [(i1, dp), (i2, d1), . . . , (ip, dp−1)](cid:27) = [(i1, dp), (i2, d1), . . . , (ip, dp−1)](cid:27) [(i1, d1), (i2, d2), . . . , (ip, dp)] Now by dividing by X, we obtain the formula in the statement. The formula in Proposition 5.1 can be interpreted as follows: (cid:3) Proposition 5.2. With M = X, N = Y we have the formula law(χ) =(cid:18)1 − 1 N(cid:19) δ0 + 1 N law(A) where A ∈ C(TM N , MM (C)) is given by A(q) = Gram matrix of the rows of q. 24 TEODOR BANICA AND JULIEN BICHON Proof. According to Proposition 5.1, we have the following formula: ZG χp = = = 1 1 MN Xi1...ip Xd1...dp MN ZTM N Xi1...ip Xd1...dp MN ZTM N Xi1...ip Xd1 1 δ[i1d1,...,ipdp],[i1dp,...,ipdp−1] qi1d1 . . . qipdp qi1dp . . . qipdp−1 dq qi1d1 qi2d1! Xd2 qi2d2 qi3d2! . . .Xdp qipdp qi1dp dq Consider now the Gram matrix in the statement, A(q)ij =< Ri, Rj >, where R1, . . . , RM are the rows of q ∈ TM N ≃ MM ×N (T). We have then: ZG χp = = = 1 1 MN ZTM N MN ZTM N MN ZTM N 1 < Ri1, Ri2 >< Ri2, Ri3 > . . . < Rip, Ri1 > A(q)i1i2A(q)i2i3 . . . A(q)ipi1 T r(A(q)p)dq = 1 NZTM N tr(A(q)p)dq But this gives the formula in the statement, and we are done. (cid:3) where the measure in Proposition 5.1 converges, after some suitable manipulations. The problem now is that of finding the good regime, M = f (K), N = g(K), K → ∞, We denote by NC(p) the set of noncrossing partitions of {1, . . . , p}, and for π ∈ P (p) we denote by π ∈ {1, . . . , p} the number of blocks. See [22]. We will need: Lemma 5.3. With M = αK, N = βK, K → ∞ we have: cp K p−1 ≃ pXr=1 #nπ ∈ NC(p)(cid:12)(cid:12)(cid:12)π = ro αr−1βp−r p+1(cid:0)2p p(cid:1)(αK)p−1. In particular, with α = β we have cp ≃ 1 Proof. We use the combinatorial formula in Proposition 5.1 above. Our claim is that, with π = ker(i1, . . . , ip), for π ∈ NC(p) the contribution to cp is Cπ ≃ απ−1βp−πK p−1, and for π /∈ NC(p), the contribution is Cπ = O(K p−2). As a first observation, since there are M(M − 1) . . . (M − π + 1) ≃ M π choices for a multi-index (i1, . . . , ip) ∈ X p satisfying ker i = π, we have: Cπ ≃ M π−1N −1#nd1, . . . , dp ∈ Y(cid:12)(cid:12)(cid:12)[dαα ∈ b] = [dα−1α ∈ b],∀b ∈ πo RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 25 Consider now the partition σ = ker d. The contribution of σ to the above quantity Cπ is then given by ∆(π, σ)N(N − 1) . . . (N − σ + 1) ≃ ∆(π, σ)N σ, where: ∆(π, σ) =(1 if b ∩ c = (b − 1) ∩ c,∀b ∈ π,∀c ∈ σ 0 otherwise We use now the fact, coming from [10], that for π, σ ∈ P (p) satisfying ∆(π, σ) = 1 we have π + σ ≤ p + 1, with equality when π, σ ∈ NC(p) are inverse to each other, via Kreweras complementation. This shows that for π /∈ NC(p) we have Cπ = O(K p−2), and that for π ∈ NC(p) we have Cπ ≃ M π−1N −1N p−π−1 = απ−1βp−πK p−1, as claimed. (cid:3) We denote by πt the free Poisson (or Marchenko-Pastur) law of parameter t > 0. It is known that the p-th moment of πt is given by Pπ∈N C(p) tπ, and in particular that the Also, we denote by D the dilation operation, Dr(law(X)) = law(rX). p(cid:1). See [20], [22], [26]. Theorem 5.4. With M = αK, N = βK, K → ∞ we have: 1 p-th moment of π1 is the Catalan number p+1(cid:0)2p αβK 2(cid:19) δ0 + In particular with α = β we have µ =(cid:0)1 − 1 µ =(cid:18)1 − 1 Proof. At α = β, this follows from Lemma 5.3. In general now, we have: 1 αβK 2 D 1 βK (πα/β) α2K 2 D 1 αK (π1). α2K 2(cid:1) δ0 + 1 α Xπ∈N C(p)(cid:18)α β(cid:19)π βp = βp α Z xpdπα/β(x) βK(cid:19) dx cp K p−1 ≃ Xπ∈N C(p) απ−1βp−π = When α ≥ β, where dπα/β(x) = ϕα/β(x)dx is continuous, we obtain: cp = 1 αKZ (βKx)pϕα/β(x)dx = 1 αβK 2Z xpϕα/β(cid:18) x But this gives the formula in the statement. When α ≤ β the computation is similar, with a Dirac mass as 0 dissapearing and reappearing, and gives the same result. (cid:3) As a first comment, when interchanging α, β we obtain D 1 (πβ/α), which is a consequence of the well-known formula πt−1 = Dt(πt). This latter formula is best understood by using Kreweras complementation (see [22]), which gives indeed: (πα/β) = D 1 αK βK Z xpdπt(x) = Xπ∈N C(p) tπ = tp+1 Xπ∈N C(p) t−π = tZ (tx)pdπt−1(x) Let us state as well an explicit result, regarding densities: 26 TEODOR BANICA AND JULIEN BICHON Proposition 5.5. With M = αK, N = βK, K → ∞ we have: µ =(cid:18)1 − 1 αβK 2(cid:19) δ0 + 1 αβK 2 ·p4αβK 2 − (x − αK − βK)2 2πx dx In particular with α = β we have µ =(cid:0)1 − 1 α2K 2(cid:1) δ0 + 1 α2K 2 · Proof. According to the well-known formula for the density of the free Poisson law (see [20], [22]), the density of the continuous part D 1 (πα/β) is indeed given by: √ 4αK x −1 2π . βK q4 α β − ( x βK − 1 − α β )2 2π · x βK = p4αβK 2 − (x − αK − βK)2 2πx With α = β now, we obtain the second formula in the statement, and we are done. (cid:3) Observe that at α = β = 1, where M = N = K → ∞, the measure in Theorem 5.4, (π1), is supported by [0, 4K]. On the other hand, since the groups ΓM,N are all amenable, the corresponding measures are supported on [0, MN], and so on [0, K 2] in the M = N = K situation. The fact that we don't have a convergence of supports is not surprising, because our convergence is in moments. namely µ =(cid:0)1 − 1 K 2(cid:1) δ0 + 1 K 2 D 1 K References [1] N. Andruskiewitsch and J. Bichon, Examples of inner linear Hopf algebras, Rev. Un. Mat. Argentina 51 (2010), 7 -- 18. [2] N. Andruskiewitsch and J. Devoto, Extensions of Hopf algebras, St. Petersburg Math. J. 7 (1996), 17 -- 52. [3] T. Banica, Truncation and duality results for Hopf image algebras, Bull. Pol. Acad. Sci. Math. 62 (2014), 161 -- 179. [4] T. Banica and J. Bichon, Quantum groups acting on 4 points, J. Reine Angew. Math. 626 (2009), 74 -- 114. [5] T. Banica and J. Bichon, Hopf images and inner faithful representations, Glasg. Math. J. 52 (2010), 677 -- 703. [6] T. Banica and B. Collins, Integration over the Pauli quantum group, J. Geom. Phys. 58 (2008), 942 -- 961. [7] T. Banica, U. Franz and A. Skalski, Idempotent states and the inner linearity property, Bull. Pol. Acad. Sci. Math. 60 (2012), 123 -- 132. [8] T. Banica and R. Speicher, Liberation of orthogonal Lie groups, Adv. Math. 222 (2009), 1461 -- 1501. [9] J. Bhowmick, F. D'Andrea and L. Dabrowski, Quantum isometries of the finite noncommutative geometry of the standard model, Comm. Math. Phys. 307 (2011), 101 -- 131. [10] P. Biane, Some properties of crossings and partitions, Discrete Math. 175 (1997), 41 -- 53. [11] J. Bichon, Free wreath product by the quantum permutation group, Alg. Rep. Theory 7 (2004), 343 -- 362. [12] R. Brauer, On algebras which are connected with the semisimple continuous groups, Ann. of Math. 38 (1937), 857 -- 872. [13] A. Chirvasitu, Residually finite quantum group algebras, preprint 2014. RANDOM WALK QUESTIONS FOR LINEAR QUANTUM GROUPS 27 [14] A. Connes, A unitary invariant in Riemannian geometry, Int. J. Geom. Methods Mod. Phys. 5 (2008), 1215 -- 1242. [15] V.G. Drinfeld, Quantum groups, Proc. ICM Berkeley (1986), 798 -- 820. [16] D. Enders, A characterization of semiprojectivity for subhomogeneous C ∗-algebras, preprint 2014. [17] A. Freslon, On the partition approach to Schur-Weyl duality and free quantum groups, preprint 2014. [18] M. Jimbo, A q-difference analog of U (g) and the Yang-Baxter equation, Lett. Math. Phys. 10 (1985), 63 -- 69. [19] H. Kesten, Symmetric random walks on groups, Trans. Amer. Math. Soc. 92 (1959), 336 -- 354. [20] V.A. Marchenko and L.A. Pastur, Distribution of eigenvalues in certain sets of random matrices, Mat. Sb. 72 (1967), 507 -- 536. [21] S. Neshveyev and L. Tuset, Compact quantum groups and their representation categories, SMF (2013). [22] A. Nica and R. Speicher, Lectures on the combinatorics of free probability, Cambridge Univ. Press (2006). [23] S. Raum and M. Weber, The full classification of orthogonal easy quantum groups, preprint 2013. [24] H.-J. Schneider, Some remarks on exact sequences of quantum groups, Comm. Algebra 21 (1993), 3337 -- 3357. [25] W. Tadej and K. Zyczkowski, A concise guide to complex Hadamard matrices, Open Syst. Inf. Dyn. 13 (2006), 133 -- 177. [26] D.V. Voiculescu, K.J. Dykema and A. Nica, Free random variables, AMS (1992). [27] S. Wang, Free products of compact quantum groups, Comm. Math. Phys. 167 (1995), 671 -- 692. [28] S. Wang, Quantum symmetry groups of finite spaces, Comm. Math. Phys. 195 (1998), 195 -- 211. [29] S.L. Woronowicz, Compact matrix pseudogroups, Comm. Math. Phys. 111 (1987), 613 -- 665. [30] S.L. Woronowicz, Tannaka-Krein duality for compact matrix pseudogroups. Twisted SU(N) groups, Invent. Math. 93 (1988), 35 -- 76. T.B.: Department of Mathematics, Cergy-Pontoise University, 95000 Cergy-Pontoise, France. [email protected] J.B.: Department of Mathematics, Clermont-Ferrand University, F-63177 Aubiere, France. [email protected]
0810.3906	2	0810	2011-01-11T19:59:01	The Radial Masa in a Free Group Factor is Maximal Injective	[ "math.OA" ]	The radial (or Laplacian) masa in a free group factor is the abelian von Neumann algebra generated by the sum of the generators (of the free group) and their inverses. The main result of this paper is that the radial masa is a maximal injective von Neumann subalgebra of a free group factor. We also investigate tensor products of maximal injective algebras. Given two inclusions $B_i\subset M_i$ of type $\mathrm{I}$ von Neumann algebras in finite von Neumann algebras such that each $B_i$ is maximal injective in $M_i$, we show that the tensor product $B_1\ \bar{\otimes}\ B_2$ is maximal injective in $M_1\ \bar{\otimes}\ M_2$ provided at least one of the inclusions satisfies the asymptotic orthogonality property we establish for the radial masa. In particular it follows that finite tensor products of generator and radial masas will be maximal injective in the corresponding tensor product of free group factors.	math.OA	math	The radial masa in a free group factor is maximal injective Jan Cameron Junsheng Fang∗ Mohan Ravichandran [email protected] [email protected]. [email protected] Stuart White [email protected] November 8, 2018 1 1 0 2 n a J 1 1 ] . A O h t a m [ 2 v 6 0 9 3 . 0 1 8 0 : v i X r a Abstract The radial (or Laplacian) masa in a free group factor is the abelian von Neumann algebra generated by the sum of the generators (of the free group) and their inverses. The main result of this paper is that the radial masa is a maximal injective von Neumann subalgebra of a free group factor. We also investigate tensor products of maximal injective algebras. Given two inclusions Bi ⊂ Mi of type I von Neumann algebras in finite von Neumann algebras such that each Bi is maximal injective in Mi, we show that the tensor product B1 ⊗ B2 is maximal injective in M1 ⊗ M2 provided at least one of the inclusions satisfies the asymptotic orthogonality property we establish for the radial masa. In particular it follows that finite tensor products of generator and radial masas will be maximal injective in the corresponding tensor product of free group factors. 1 Introduction In [15], Popa showed that the von Neumann algebra generated by a single generator a1 of a free group FK = ha1, . . . , aKi is a maximal injective von Neumann subalgebra of the factor LFK associated to FK. This resolved a problem of Kadison by demonstrating that self-adjoint elements in a II1 factor need not contained in a hyperfinite subfactor. The subalgebra generated by a1 is known as a generator masa (maximal abelian subalgebra) in LFK. There is another naturally occurring masa in the free group factor LFK, the von i ). This was shown to be maximal abelian by Pytlik in [18] and is known as the radial masa in LFK. The main objective of this paper is to show that this radial masa gives another example of an abelian maximal injective von Neumann subalgebra of LFK. Neumann subalgebra generated by PK i=1(ai + a−1 The radial masa shares many properties with generator masas. Dixmier defined a masa A in a II1 factor M to be singular if every unitary u ∈ M with uAu∗ = A lies in A, and ∗Partially supported by the fundamental research funds for the central universities of China. 1 showed in [6] that a generator masa in a free group factor is singular. Singularity of the radial masa was established by Rădulescu by means of an intricate calculation in [19], which provides a central ingredient in this paper. An alternative combinatorial proof was given by Sinclair and Smith in [21]. To establish singularity of the radial masa, Rădulescu computed its Pukánszky invariant, historically the most sucessful invariant for singular masas (see [23] for a discussion of this invariant). The Pukánszky invariant of a generator masa is easily computed using the group-subgroup methods of [22] and both the radial and generator masas have Pukánszky invariant {∞}. It is natural to ask whether these masas are conjugate in LFK. This seems unlikely, but we have been unable to find a proof. As noted in Proposition 3.1 below, a straightforward deduction from [16] shows that the radial masa is not inner conjugate to a generator masa in LFK. Question 1.1. Does there exist an automorphism of LFK which maps the radial masa onto a generator masa? The critical ingredient in Popa's proof of the maximal injectivity of a generator masa B in a free group factor LFK is an asymptotic orthogonality property at the Hilbert space level. Lemma 2.1 of [15] shows that if x1, x2 are elements of an ultrapower (LFK)ω with EBω (x1) = EBω (x2) = 0 and y1, y2 are elements of LFK with EB(y1) = EB(y2) = 0, then x1y1 is orthogonal to y2x2 in the Hilbert space L2((LFK)ω). To do this, Popa first shows that those elements x ∈ LFK which approximately commute with the generator a1 and are orthogonal to B must be essentially supported on the collection of words which begin and end with large powers of a1 or a−1 1 . The asymptotic orthogonality follows when y1, y2 are group elements in FK \ {an 1 : n ∈ Z} and a density argument then gives the result for general y1 and y2 (see also [23, Section 14.2]). Most of this paper is taken up with establishing an asymptotic orthogonality condition for the radial masa (Theorem 6.2). The natural orthonormal basis of ℓ2(FK) does not behave well in calculations involving the radial masa, however to prove that the radial masa is singular Rădulescu introduced in [19] a collection of vectors which form a basis (albeit not quite an orthonormal basis) for ℓ2(FK) which can be thought of as spanning double cosets coming from the radial masa. In Section 3, we set up notation for discussing the radial masa and describe the properties of Rădulescu's basis. In Section 4, we establish the essential location of the support of an element x in LFK which is both orthogonal to the radial masa and approximately commutes with it. We do this in Lemma 4.3, the proof of which is contained in the lemmas preceding it. Section 5 contains the second step of [15, Lemma 2.1] for the radial masa. In this section, we use techniques from [21] to count certain words in FK which we use in Section 6 to establish the asymptotic orthogonality result from the results of Section 4. All of the calculations described above are performed in LFK ⊗ N, where N is an arbitrary finite von Neumann algebra. The only additional difficulties this introduces are notational, and it enables us to show that certain tensor products involving the radial masa are also maximal injective, as described below. In [15, 4.5 (1)], Popa asked how maximal injective von Neumann algebras behave under tensor products. That is, if Bi ⊂ Mi (i = 1, 2) are two inclusions of von Neumann algebars with Bi a maximal injective von Neumann subalgebra of Mi, must B1 ⊗ B2 be a maximal injective von Neumann subalgebra of M1 ⊗ M2? Various authors have subsequently worked on this question. The first progress was made by Ge and Kadison, who in [10] gave a 2 positive answer when B1 = M1 is an injective factor. This result was subsequently improved by Strătilă and Zsidó who removed the factor assumption to show ([25, Theorem 6.7]) that if M1 is an injective von Neumann algebra and M2 is a von Neumann algebra with a separable predual, then M1 ⊗ B2 is a maximal injective von Neumann subalgebra of M1 ⊗ M2 whenever B2 is a maximal injective von Neumann subalgebra of M2. In [20], Shen investigated the tensor product of copies of a generator masa inside the a free group factor, answering Popa's question positively in this case. He was also able to obtain a positive result for an infinite tensor product of generator masas and so obtain the first example of an abelian subalgebra which is maximal injective in a McDuff II1 factor. In [20], it was necessary to develop an additional orthogonality condition styled upon [15, Lemma 2.1] for tensor products of generator masas. Recent progress was made by the second author, who in [8] gave a positive answer to Popa's question in the case that Z(B1) is atomic and M1 has a separable predual. In Section 2, we examine how asymptotic orthogonality properties based on [15, Lemma 2.1] imply maximal injectivity. Using Popa's intertwining lemma and ingredients from [8, 20], we are able to give a technical improvement of the argument which deduces the maximal injectivity of the generator masa from the asymptotic orthogonality condition. This argu- ment doesn't require any further assumptions on the masa beyond singularity and is also applicable to the tensor product of maximal injective algebras, provided we can establish our asymptotic orthogonality condition after tensoring by a further finite von Neumann algebra (see Definition 2.1 below). In particular, it follows that if A ⊂ M is a singular masa of a II1 factor with the asymptotic orthogonality property, such as a generator masa or the radial masa in a free group factor, and B is a type I maximal injective von Neumann subalgebra of a II1 factor N with separable predual, then A ⊗ B is maximal injective in M ⊗ N (Theorem 2.8, taking A to be a masa). This result strengthens Shen's finite tensor product results from [20] and enlarges the class of positive answers to Popa's question. The paper concludes with Section 7, which contains some final remarks. We observe that the additional ingredients Popa uses with asymptotic orthogonality to show a generator masa is maximal injective in [15] can be deduced from maximal injectivity and Ozawa's solidity of the free group factors introduced in [12]. In this way, we see that these properties of the generator masas are also satisfied by the radial masa. We end with a brief discussion of maximal nuclearity of the generator and radial masas in the reduced group C∗-algebras of free groups. Acknowledgements The work in this paper originated during a visit of the fourth named author to the University of New Hampshire in April 2008. He would like to thank the faculty and students of the Department of Mathematics, and in particular Don Hadwin, for their hospitality during this visit. Section 2 was undertaken at the Workshop in Analysis and Probability at Texas A&M University in August 2008. It is a pleasure to record our gratitude to the workshop organisers and NSF for providing financial support to the workshop. The authors would like to thank Simon Wassermann for simulating conversations regard- ing maximal nuclear C∗-subalgebras of C∗-algebras. Finally, the authors would like to thank the referee for their helpful suggestions, which have improved the exposition of the paper. 3 2 Asymptotic Orthogonality and Maximal Injectivity In this section we examine how asymptotic orthogonality properties give rise to maximal injective von Neumann algebras. Definition 2.1. Let A be a type I von Neumann subalgebra of a type II1 von Neumann algebra M with a fixed faithful normal trace τM . Let N be a finite von Neumann algebra with a fixed faithful normal trace τN . Say that A ⊂ M has the asymptotic orthogonality property after tensoring by N if there is a non-principal ultrafilter ω ∈ βN\N such that x(1)y1 ⊥ y2x(2) in L2((M ⊗ N)ω, (τM ⊗ τN )ω), whenever x(1), x(2) are elements of (A ⊗ C1)′ ∩ (M ⊗ N)ω with E(A⊗N )ω (x(i)) = 0 for i = 1, 2, and y1, y2 ∈ M ⊗ N with EA⊗N (yi) = 0 for i = 1, 2. Say that A has the asymptotic orthogonality property if it has this property when N = C1. Note that the ultraproduct (M ⊗ N)ω used in this definition is constructed with respect to the product trace τM ⊗ τN . In section 2 of [15], Popa shows that the generator masa inside a free group factor has the asymptotic orthogonality property. The arguments given in [15] (see also [23, Section 14.2]) also show that the generator masa has the asymptotic orthogonality property after tensoring by any N without further work. As noted in [15], given any separable diffuse type I von Neumann algebra B 6= C1, one can form the free product B ∗ R (where R is the hyperfinite II1 factor). The methods of [15] show that B is maximal injective in B ∗ R. Indeed, these inclusions B ⊂ B ∗ R have the asymptotic orthogonality property after tensoring by any N. In this section we show how the maximal injectivity of a singular masa satisfying the asymptotic orthogonality property can be established without further hypotheses (Corollary 2.3). We also examine tensor products, showing how the asymptotic orthogonality property after tensoring by N can be used to show that certain tensor products A⊗B of maximal injective subalgebras are again maximal injective. In subsequent sections we shall show that the radial masa in a free group factor also has the asymptotic orthogonality property after tensoring by any N. It will follow that finite tensor products of generator and radial masas are maximal injective. We need some technical observations, the first of which easily follows from Popa's intertwining lemma. Lemma 2.2. Let L be an injective type II1 von Neumann algebra with a separable predual equipped with a fixed faithful normal trace. Let A be a type I von Neumann subalgebra of L. Then there exists a unitary u ∈ L′ ∩ Lω with EAω (u) = 0. Proof. Since L is injective with separable predual, we can find a sequence (Ln) of finite dimensional subalgebras of L whose union is weakly dense in L. It suffices to show that for all ε > 0 and n ∈ N there exists a unitary u ∈ L′ n ∩ L with kEA(u)k2 < ε. However, if there exists some n and ε > 0 for which no such unitary could be found, then Popa's intertwining lemma [17] (see also [3, Theorem F.12, 4 ⇒ 1] for the exact statement we are using) shows that a corner of L′ n ∩ L embeds into A inside L. This can not happen, since L′ n ∩ L is necessarily type II1 and A is type Ifin. If A is a maximal injective von Neumann subalgebra in a finite von Neumann algebra M, then A is singular in M (see [13, Lemma 3.6] for example). Any singular masa A in M necessarily has A′ ∩ M ⊆ A so that any intermediate subalgebra A ⊂ L ⊂ M has Z(L) ⊆ L′ ∩ M ⊆ A′ ∩ M ⊆ A. 4 Corollary 2.3. Let A be a singular masa in a II1 von Neumann algebra M with a separa- ble predual. Suppose that A has the asymptotic orthogonality property, then A is maximal injective. Proof. Let L be an injective von Neumann algebra with A ⊂ L ⊂ M. Let p be the maximal central projection in L so that Lp is type II1. As noted above, p ∈ A. If p 6= 0, use Lemma 2.2 to find a unitary u in (Lp)′ ∩ (Lp)ω with EAω (u) = 0. Choosing some non-zero v ∈ Lp with EA(v) = 0, the asymptotic orthogonality property gives uv ⊥ vu. On the other hand uv = vu, so uv = 0 and v = 0, giving a contradiction. Hence p = 0 and L is a finite type I von Neumann algebra. By [11], A is regular in L (see also [20, Lemma 2.3]), so by singularity of A in M it follows that L = A. The proof of our second technical observation, which allows us to handle tensor products, is contained in [20] modulo a theorem from [25]. We include the details for completeness. Lemma 2.4. Let A ⊂ M and B ⊂ N be two inclusions of finite von Neumann algebras with separable preduals. Suppose that A is type I, A′ ∩ M ⊂ A and that B is a maximal injective von Neumann subalgebra of N. Let A ⊗ B ⊂ L ⊂ M ⊗ N be an injective intermediate von Neumann subalgebra. Then EA⊗N (L) = A ⊗ B. Proof. Since A is finite and type I, it follows that (A′∩M)′ ∩M = A. Indeed if A is of type In, we can express A as Z ⊗ Mn for some abelian Z and construct a corrresponding factorisation M = M ⊗ Mn. Then (A′ ∩ M) = (Z ′ ∩ M) ⊗ C1 so the condition that A′ ∩ M ⊂ A implies that Z is a masa in M . Hence (A′ ∩ M)′ ∩ M = (Z ⊗ C1)′ ∩ ( M ⊗ Mn) = Z ⊗ Mn = A. The general case follows by taking a direct sum. Given x ∈ L, EA⊗N (x) is the unique element of minimal k·k2 in co2{uxu∗ : u ∈ U((A′ ∩ M) ⊗ Z(N))} as ((A′ ∩ M) ⊗ Z(N))′ ∩ (M ⊗ N) = A ⊗ N. Since (A′ ∩ M) ⊗ Z(N) ⊂ A ⊗ B ⊂ L, it follows that EA⊗N (L) ⊂ L. If x ∈ L has EA⊗N (x) /∈ A ⊗ B, then we have A ⊗ B ( (A ⊗ B, EA⊗N (x))′′ ⊆ A ⊗ N, where (A ⊗ B, EA ⊗ N (x))′′ is a subalgebra of L and hence injective. This contradicts [25, Theorem 6.7], which shows that A ⊗ B is maximal injective in A ⊗ N, whenever A is injective and B is maximal injective in a von Neumann algebra N with separable predual. These ingredients enable us to gain control over certain intermediate injective subalge- bras. Lemma 2.5. Let M, N be finite von Neumann algebras with separable preduals. Let A be a type I von Neumann subalgebra of M with A′ ∩ M ⊆ A and the asymptotic orthogonality property after tensoring by N. Let B be a type I von Neumann subalgebra of N which is a maximal injective von Neumann subalgebra of N. If L is an intermediate injective von Neumann algebra between A ⊗ B and M ⊗ N, then L is necessarily type I. Proof. Let p be the maximal central projection of L such that Lp is type II1. Note that p ∈ L′ ∩ (M ⊗ N) ⊂ A ⊗ B. If p 6= 0, then Lemma 2.2 gives us a unitary u ∈ (Lp)′ ∩ (Lp)ω, with E((A⊗B)p)ω (u) = 0, where the ultrapower (Lp)ω is constructed from the product trace on 5 M ⊗ N. Since (A ⊗ B)p ( Lp, there is some non-zero v ∈ Lp with EA⊗B(v) = 0. Regarding u ∈ (M ⊗ N)ω and v ∈ M ⊗ N, we have EA⊗N (v) = EA⊗BEA⊗N (v) = EA⊗N EA⊗B(v) = 0, where the first equality uses Lemma 2.4 to see that EA⊗N (v) ∈ A ⊗ B. By writing u = (un) for some un ∈ L with EA⊗B(un) = 0 for all n, it follows that EA⊗N (un) = 0 for each n just as above and so E(A⊗N )ω (u) = 0. As A ⊂ M has the asymptotic orthogonality property after tensoring by N, we have uv ⊥ vu. Since u ∈ (Lp)′ ∩ (Lp)ω and v ∈ Lp, u and v commute. Therefore uv = 0 and hence v = 0, which is a contradiction. Therefore p = 0. With the notation of the previous lemma, if we additionally assume that A is a singular masa with the asymptotic orthogonality property and B is a masa which is maximal injective, we can then deduce the maximal injectivity of A ⊗ B in M ⊗ N as follows. Suppose L is an injective von Neumann algebra between A ⊗ B and M ⊗ N. By Lemma 2.5, L is a finite type I von Neumann algebra. Since B is singular in M, A ⊗ B is singular in M ⊗ N by [24, Corollary 2.4]. On the other hand, every masa in a finite type I von Neumann algebra is regular by [11] (see also [20, Lemma 2.3]). Ergo L = A ⊗ B. Returning to the more general situation of type I algebras A and B, the argument of the previous paragraph breaks down as there need not be a non-trivial normalising unitary of A ⊗ B in L. Indeed, take P to be M2(C) ⊕ C1, naturally embedded inside Q = M3(C), then P ′ ∩ Q ⊆ P but every normalising unitary of P in Q lies in P . To circumvent this difficulty, we need to use groupoid normalisers. Recall that if P ⊂ Q is an inclusion of von Neumann algebras with P ′ ∩ Q ⊆ P , then a partial isometry v ∈ Q is a groupoid normaliser of P if vP v∗ ⊆ P and v∗P v ⊆ P . We write GN Q(P ) for the collection of all groupoid normalisers. Recall too that if A is a singular masa in a finite von Neumann algebra M, then GN M (A) ⊂ A (see [23, Lemma 6.2.3(iv)]). To use the argument of the preceding paragraph in a groupoid normaliser context, we need the groupoid normaliser analogue of Kadison's fact that masas in finite type I von Neumann algebras are regular. Lemma 2.6. Let P ⊂ Q be an inclusion of finite type I von Neumann algebras with separable preduals and P ′ ∩ Q ⊆ P . Then GN Q(P )′′ = Q. Proof. As P ′ ∩ Q ⊆ P , it follows that Z(Q) ⊆ Z(P ). Hence we can assume that Q is type In for some n ∈ N, as the general case follows by taking a direct sum. By [14, Theorem 3.3], P contains a masa of Q, say B. By [11, Theorem 3.19], we can write Q = A ⊗ Mn, where A is the centre of Q and B = A ⊗ Dn, where Dn are the diagonal matrices of Mn. An easy calculation shows that p ⊗ ei,j is a groupoid normaliser of P for every projection p ∈ A and the (ei,j)n i,j=1 are the standard matrix units of Mn. These elements evidentally generate Q. In [7], the second author introduced the notation of complete singularity of an inclusion P ⊆ Q, which implies that GN Q(P ) ⊂ P . Since it was shown in [7, Proposition 3.2] that a maximal injective von Neumann subalgebra is completely singular, the next lemma follows immediately. Alternatively, one can establish the lemma directly from Connes' characteri- sation of amenability in [5] by showing that if P is an injective von Neumann algebra and v is a groupoid normaliser of P , then (P ∪ {v})′′ is also injective. 6 Lemma 2.7. Let P be a maximal injective von Neumann subalgebra of a von Neumann algebra Q. Then GN Q(P ) ⊂ P . We are now in a position to prove the main result of this section. Theorem 2.8. Let M, N be finite von Neumann algebras with separable preduals. Let A be a type I von Neumann subalgebra of M with GN M (A) ⊂ A and the asymptotic orthogonality property after tensoring by N. Let B be a type I von Neumann subalgebra of N which is a maximal injective von Neumann subalgebra of N. Then A ⊗ B is maximal injective in M ⊗ N. Proof. Let L be an injective von Neumann algebra with A ⊗ B ( L ⊆ M ⊗ N. By Lemma 2.5, L is type I and so A ⊗ B ⊂ L is an inclusion of finite type I von Neumann algebras with (A ⊗ B)′ ∩ L ⊂ A ⊗ B. By Lemma 2.6, there is a groupoid normaliser v of A ⊗ B in L with v /∈ A ⊗ B. On the other hand GN M (A) ⊂ A by hypothesis and GN N (B) ⊂ B by Lemma 2.7. Corollary 5.6 of [9] shows that GN M ⊗N (A ⊗ B)′′ = GN M (A)′′ ⊗ GN N (B)′′ = A ⊗ B, and this gives a contradiction. Hence A ⊗ B is maximal injective in M ⊗ N. Taking B = N = C in the previous theorem immediately yields the next corollary, which we could also establish directly in exactly the same way as Corollary 2.3. Corollary 2.9. Let A ⊂ M be an inclusion of a type I von Neumann algebra inside a finite von Neumann algebra with separable predual such that GN M (A) ⊂ A and A has the asymptotic orthogonality property. Then A is maximal injective in M. Finally we note that the asymptotic orthogonality property does not pass to tensor products. Indeed, no inclusion A ⊗ B ⊂ M ⊗ N with A and B masas in type II1 factors can have the asymptotic orthogonality property. To see this, take a unitary u in A′ ∩ M ω with EAω (u) = 0 by Lemma 2.2 and some non-zero v ∈ N with EB(v) = 0. Then (u ⊗ 1) lies in (A ⊗ B)′ ∩ (M ⊗N)ω with E(A⊗B)ω (u ⊗ 1) = 0 and 1 ⊗ v ∈ M ⊗ N with EA⊗ B(1 ⊗ v) = 0 and uv = vu. 3 The radial masa in a free group factor For K ≥ 2, let FK denote the free group on K generators. Write a1 . . . , aK for the generators of FK. We regard FK as a subset of the free group factor LFK. The unique faithful normal tracial state τ on LFK induces a pre-Hilbert space norm kxk2 = τ (x∗x)1/2 on LFK and completing LFK in this norm yields the Hilbert space ℓ2(FK). Therefore, we can regard LFK as a subset of ℓ2(FK) as well as an von Neumann algebra acting on ℓ2(FK). Given g ∈ FK, write g for the length of g. For n ≥ 0, let wn ∈ LFK be the sum of all words of length n in FK. The recurrence relations w1wn = wnw1 =  wn+1 + (2K − 1)wn−1, n > 1; n = 1; w2 + 2Kw0, w1, n = 0; 7 (3.1) from [4, Theorem 1], show that w1 generates an abelian von Neumann subalgebra of LFK, which we denote by A. This is the radial or Laplacian subalgebra of LFK which was shown to be masa in LFK by Pytlik, [18]. Write L2(A) for the closure of A in ℓ2(FK) and note that (wn/ kwnk2)∞ n=0 forms an orthonormal basis for L2(A). As promised in the introduction, we note that the radial and generator masas are not inner conjugate in LFK. Proposition 3.1. There is no unitary u ∈ LFK with uAu∗ = {a1}′′. Proof. Let Φ be the automorphism of LFK induced by the automorphism of FK which interchanges the generators a1 and a2 and fixes the other generators. Note that Φ(w1) = w1 so that Φ fixes the radial masa pointwise. If such a unitary u existed, then Φ(u)AΦ(u)∗ = Φ(uAu∗) = Φ({a1}′′) = {a2}′′ and then θ = ad uΦ(u)∗ gives an inner automorphism of LFK with θ({a2}′′) = {a1}′′ and this contradicts [16, Corollary 4.3]. To show that the radial masa is singular in LFK, Rădulescu decomposed ℓ2(F2) into a direct sum of orthogonal A − A-bimodules, [19]. Before setting out this decomposition, we need some preliminary notation. For l ≥ 0, let Wl denote the subspace of ℓ2(FK) spanned by the words of length l in FK and let ql be the orthogonal projection from ℓ2(FK) onto Wl. Given a vector ξ ∈ Wl for l ≥ 1, and integers r, s ≥ 0, define ξr,s = qr+s+l(wrξws) (2K − 1)(r+s)/2 . (3.2) Rădulescu's definition does not have the scaling factor (2K − 1)(r+s)/2 on the denominator, which we introduce simplify our subsequent calculations. Given a word g of length l ≥ 1, qr+s+l(wrgws) is precisely the sum of all reduced words of length r + s + l which contain g from the (r + 1)-th letter to the (r + l)-th letter. Since there are precisely (2K − 1)(r+s) such words, the chosen scaling factor ensures that kgr,sk2 = kgk2. This observation also shows that gr,s ⊥ hr,s for two distinct words g, h of length l. As such our scaling factor ensures that kξr,sk2 = kξk2, for all l ≥ 1, ξ ∈ Wl and r, s ≥ 0. Thus ξ 7→ ξr,s extends to an isometry ℓ2(FK) ⊖ L2(A) → ℓ2(FK) ⊖ L2(A). Define ξr,s = 0 when either r < 0 or s < 0. In [19], Rădulescu constructed a basis for ℓ2(FK) ⊖ L2(A) which lends itself to calculations involving the radial masa and as such plays a key role in this paper. The next lemma sets out the properties we need. Since our normalisation conventions differ from [19], we indicate how to derive these properties. Lemma 3.2. There is a sequence of orthonormal vectors (ξi)∞ properties. i=1 in ℓ2(FK) with the following 1. Each ξi lies in Wl(i) for some l(i) ≥ 1 and has (ξi)∗ = ±ξi. 2. The subspaces Span(AξiA) 2 are pairwise orthogonal in ℓ2(FK) and ⊕Span(AξiA) 2 = ℓ2(FK) ⊖ L2(A). 3. For those i with l(i) > 1, the spaces Span(AξiA) 2 have orthonormal bases (ξi r,s)r,s≥0. 8 4. For each i, j > 0, the map Ti,j which sends ξi onto Span(AξjA) operator from Span(AξiA) C0 > 1 (which depends on K but not i or j) such that r,s to ξj 2 2 r,s extends to a bounded invertible . Furthermore, there exists a constant kTi,jk ,(cid:13)(cid:13)T −1 i,j (cid:13)(cid:13) ≤ C0, i, j > 0. (3.3) Proof. For each l ≥ 1, let Sl denote the linear subspace of Wl spanned by {ql(w1g), ql(gw1) : g ∈ FK, g ≤ l − 1}. As in Theorem 7 of [19], for each l ≥ 1 we choose an orthonormal basis (ξj,l)j for Wl ⊖ Sl. As the subspaces Wl ⊖ Sl are self-adjoint, we can additionally insist that the elements of this basis satisfy (ξj,l)∗ = ±ξj,l (as Rădulescu does when l = 1 in [19, Theorem 7]). Re-indexing these elements as (ξi)i, gives a sequence (ξi)i, where ξi ∈ Wl(i) which satisfies condition (i). Condition (ii) follows from [19, Lemma 4] (the first statement is part (b) of this lemma, while part (a) gives the second statement). Lemma 3(a) of [19] ensures that for those i with l(i) ≥ 2, the elements (ξi r,s)r,s≥0 are pairwise orthogonal. Since our normalisation conventions above ensure that kξi r,sk2 = kξik2 for all such r, s condition (iii) follows. When l(i), l(j) > 1 the maps Ti,j in condition (iv) are unitary operators by (iii). It then suffices to establish (iv) for some constant C ′ 0 when l(i) = 1 and some fixed j with l(j) > 1 as all other cases follow by composition. In this case our map Ti,j is the map T ′ 0 appearing in the proof of [19, Lemma 6] and which is noted to be bounded and invertible there. One can explicitly estimate the bound on these maps from the results in [19], however this is not necessary as there are only finitely many i with l(i) = 1, so there is certainly a uniform bound C ′ 0 on the norm of Ti,j and T −1 i,j . Note that the set {ξi r,s : i ≥ 1, r, s ≥ 0} does not give an orthonormal basis for ℓ2(FK) ⊖ L2(A), due to the presence of some indices i with l(i) = 1. However, the last two facts above show that this set is at least a basis for ℓ2(FK) ⊖ L2(A) and the 2-norm it induces is equivalent to the norm k·k2. We shall subsequently use this in a tensor product setting. Proposition 3.3. Let H be a Hilbert space and let η ∈ (ℓ2(FK) ⊗ H) ⊖ (L2(A) ⊗ H). Then there exist vectors (αi r,s. Furthermore ξi r,s ⊗ αi r,s)r,s≥0,i>0 in H such that η = Xi≥1 H(cid:17)1/2 0 kηkℓ2(FK )⊗H ≤(cid:16) Xi≥1 r,sk2 r≥0,s≥0 kαi r≥0,s≥0 C −1 ≤ C0 kηkℓ2(FK )⊗H , where C0 > 1 is the constant appearing in equation (3.3). Remark 3.4. When the Hilbert space H is of the form L2(N) for some finite von Neumann algebra N, it makes sense to refer to the conjugate x∗ of an element x ∈ ℓ2(FK) ⊗ L2(N) (indeed, this is defined by extending the conjugation operation from LFK ⊗ N to the Hilbert space). We can expand both x and x∗ in terms of the ξi r,s ⊗ αi r,s s,r, it follows r,s. Since Lemma 3.2 (i) and (3.2) give ξi r,s, writing x =Pi≥1, r,s≥0 ξi r,s = ±ξi r,s ⊗ βi and x∗ = Pi≥1, r,s≥0 ξi r,s = ±αi that βi s,r for each i, r and s. 9 A critical component of Rădulescu's computations in [19] are recurrence relations analo- gous to (3.1) showing how the ξi r,s behave under multiplication with w1. We will need these relations subsequently so we recall them here. Again it is slightly more convenient for our purposes to use a suitable normalisation, so we define w1 = w1/(2K − 1)1/2. Applying our normalisation conventions to [19, Lemma 1], for i ≥ 1 and r, s ≥ 0, there exist values of σ(i) = ±1 so that and r−1,s, ξi r+1,s + ξi ξi 1,s, 1,s + σ(i) ξi 2K−1ξi 0,s−1, r,s−1, ξi r,s+1 + ξi ξi r,1, r,1 + σ(i) ξi 2K−1ξi r−1,0, w1ξi r,s =  r,s w1 =  ξi r ≥ 1; r = 0, l(i) ≥ 2; r = 0, l(i) = 1; s ≥ 1; s = 0, l(i) ≥ 2; s = 0, l(i) = 1. (3.4) (3.5) −1,s and ξi Since ξi identical to the first cases. r,−1 are defined to be zero, the second cases of the relations above are 4 Locating the support of elements of A′ ∩ (LFK)ω In this section we establish the first half of the technical estimates required to show that the radial masa has the asymptotic orthogonality property in LFK. Our objective, realised in Lemma 4.3, is to control the support of elements of A′ ∩ (LFK)ω which are orthogonal to Aω. We find that the essential support of elements in LFK orthogonal to A and approxiately commuting with A must be contained in the closed linear span of the ξi r,s for sufficiently large r and s. We perform these calculations after tensoring by an arbitrary finite von Neumann algebra N as our objective is the asymptotic orthogonality property after tensoring by such an N. For each i, define λ(i) =(1, 1 − σ(i) 2K−1, l(i) ≥ 2; l(i) = 1. Lemma 4.1. Let N be a finite von Neumann algebra and suppose that x ∈ ℓ2(FK) ⊗ L2(N) has x ⊥ L2(A ⊗ N) and kxk2 = 1. Write with convergence in ℓ2(FK) ⊗ L2(N), for some αi r,s ∈ L2(N). Then, for each s′ ≥ s ≥ 1 x = Xr,s≥0, i≥1 ξi r,s ⊗ αi r,s, kαi r−s,0 + λ(i)αi r−s+2,0 + · · · + λ(i)αi r+s−2,0 + λ(i)αi r+s,0k2 kαi r,sk2 i≥1 (cid:12)(cid:12)(cid:12)(cid:16)Xr≥s′ ≤(cid:16)Xr≥s i≥1 2(cid:17)1/2 i≥1 −(cid:16)Xr≥s′ 2(cid:17)1/2 r+s,0)k2 2(cid:17)1/2(cid:12)(cid:12)(cid:12) kαi r,s − (αi r−s,0 + λ(i)αi r−s+2,0 + · · · + λ(i)αi r+s−2,0 + λ(i)αi ≤ 3s−1C0 kx( w1 ⊗ 1) − ( w1 ⊗ 1)xk2 , (4.1) 10 where C0 is the constant of Proposition 3.3. Proof. Note that the first inequality in the Lemma is immediate from the triangle inequality, so it suffices to prove the second inequality. Write ε = kx( w1 ⊗ 1) − ( w1 ⊗ 1)xk2 and define r,s = 0 whenever r < 0 or s < 0. Recalling the convention that ξi αi r,s is zero whenever r < 0 or s < 0, the recurrence relations (3.4) and (3.5) show that [ w1, ξi r,s] = These relations give   ξi r+1,s + ξi ξi r+1,s + ξi ξi 1,s − ξi ξi r+1,0 + λ(i)ξi ξi 1,0 − ξi 0,1, r−1,s − ξi r−1,s − ξi r,s+1 − ξi r,s+1 − ξi r,s−1, r,s−1, 0,s+1 − λ(i)ξi 0,s−1, r,1, r−1,0 − ξi l(i) ≥ 2, r, s ≥ 0 l(i) = 1, r, s ≥ 1 l(i) = 1, r = 0, s > 0 l(i) = 1, r > 0, s = 0 l(i) = 1, r = s = 0. (ξr r+1,s + ξi r−1,s − ξi r,s+1 − ξi r,s−1) ⊗ αi r,s (4.2) (ξi r+1,s + ξi r−1,s − ξi r,s+1 − ξi r,s−1) ⊗ αi r,s [ w1 ⊗ 1, x] = Xr,s≥0, l(i)≥2 + Xr,s≥1, l(i)=1 + Xs≥1, l(i)=1 + Xr≥1, l(i)1=1 + Xl(i)=1 (ξi (ξi 1,s − ξi 0,s+1 − λ(i)ξi 0,s−1) ⊗ αi 0,s (ξi r+1,0 + λ(i)ξi r−1,0 − ξi r,1) ⊗ αi r,0 1,0 − ξi 0,1) ⊗ αi 0,0, r,s ⊗ βi r,s for some βi Pr,s≥0, i≥1 ξi with convergence in ℓ2(FK) ⊗ L2(N). We then rearrange this sum to write it in the form r,s ∈ L2(N). For l(i) ≥ 2 and r, s ≥ 0 contributions to βi r,s arise from the (r + 1, s), (r − 1, s), (r, s + 1) and (r, s − 1) terms in the first sum on the right of (4.2) (the cases when either r = 0 or s = 0 do not cause extra difficulty, as our notational conventions ensure that ξi r,−1 = 0). Thus r,s = αi βi r,s when l(i) = 1 can also be computed from (4.2). For example we can only obtain a contribution to βi 0,0 with l(i) = 1 from the s = 1 term in line 3 of (4.2) and the r = 1 term of line 4, giving −1,s = αi r,s+1 when l(i) = 2. The values of βi r,−1 = 0 and αi r+1,s − αi r−1,s + αi r,s−1 − αi −1,s = ξi 11 βi 0,0 = λ(i)(αi 1,0 − αi 0,1) in this case. Continuing in this fashion, we obtain [ w1 ⊗ 1, x] = Xl(i)≥2, r,s≥0 + Xl(i)=1, r,s≥1 + Xl(i)=1, r≥1 + Xl(i)=1, s≥1 + Xl(i)=1 ξi r,s ⊗ (αi r−1,s + αi r+1,s − αi r,s−1 − αi r,s+1) ξi r,s ⊗ (αi r−1,s + αi r+1,s − αi r,s−1 − αi r,s+1) ξi r,0 ⊗ (λ(i)αi r+1,0 + αi r−1,0 − αi r,1) 0,s ⊗ (αi ξi 1,s − αi 0,s−1 − λ(i)αi 0,s+1) λ(i)ξi 0,0 ⊗ (αi 1,0 − αi 0,1). Proposition 3.3 gives C 2 kαi r−1,s + αi r+1,s − αi r,s−1 − αi r,s+1k2 2 (4.3) 0 ε2 ≥ Xl(i)≥2, r,s≥0 + Xl(i)=1, r,s≥1 + Xl(i)=1, r≥1 + Xl(i)=1, s≥1 + Xl(i)=1 kαi r−1,s + αi r+1,s − αi r,s−1 − αi r,s+1k2 2 kλ(i)αi r+1,0 + αi r−1,0 − αi r,1k2 2 kαi 1,s − αi 0,s−1 − λ(i)αi 0,s+1k2 2 kλ(i)(αi 1,0 − αi 0,1)k2 2, where C0 is the constant appearing in Proposition 3.3. Considering the terms with s = 0 from the first line of right hand side of (4.3) and those terms from the third line gives Xr≥1 l(i)≥2 kαi r−1,0 + αi r+1,0 − αi r,1k2 2 + Xr≥1 l(i)=1 kλ(i)αi r+1,0 + αi r−1,0 − αi r,1k2 2 ≤ C 2 0 ε2, since αi r,−1 is defined to be zero. As λ(i) = 1 whenever l(i) ≥ 2, we have kαi r−1,0 + λ(i)αi r+1,0 − αi r,1k2 2 ≤ C 2 0 ε2. Xr≥1 i≥1 This gives the s = 1 case of (4.1). For the s = 2 case, take those terms on the right-hand side of (4.3) with s = 1 and r ≥ 2 to obtain Xr≥2 i≥1 kαi r−1,1 + αi r+1,1 − αi r,0 − αi r,2k2 2 ≤ C 2 0 ε2. (4.4) 12 Now r,2 + αi αi r,0 − αi r+1,1 − αi r,2 − (αi r−2,0 + λ(i)αi r,0 + λ(i)αi (4.5) r−1,1 =(cid:16)αi −(cid:16)αi −(cid:16)αi r+1,1 − (αi r,0 + λ(i)αi r−1,1 − (αi r−2,0 + λ(i)αi r+2,0)(cid:17) r,0)(cid:17), r+2,0)(cid:17) so, (4.4) and the s = 1 case of the lemma combine to give kαi (cid:16) Xr≥2, i≥1 ≤C0ε +(cid:16) Xr≥2, i≥1 r,2 − (αi r−2,0 + λ(i)αi r,0 + λ(i)αi r+2,0)k2 kαi r+1,1 − (αi r,0 + λ(i)αi r+2,0)k2 2(cid:17)1/2 +(cid:16) Xr≥2, i≥1 2(cid:17)1/2 kαi r−1,1 − (αi r−2,0 + λ(i)αi r,0k2 2(cid:17)1/2 (4.6) ≤C0ε + C0ε + C0ε = 3C0ε. We need to sum over r ≥ 2 above so that the sum in the second term in (4.6) is over those r + 1 ≥ 1 which allows the s = 1 case of the lemma to be used. Suppose inductively that the lemma holds for 1 ≤ s ≤ s0 with s0 ≥ 2. Taking those terms in (4.3) with s = s0 and r ≥ s0 gives Xr≥s0 i≥1 (cid:13)(cid:13)αi r−1,s0 + αi r+1,s0 − αi r,s0−1 − αi 2 2 ≤ C 2 0 ε2. r,s0+1(cid:13)(cid:13) The identity αi r,s0−1 + αi =αi r,s0+1 − αi r−1,s0 − αi r−(s0+1),0 + λ(i)αi r+1,s0 r−(s0−2),0 + · · · + λ(i)αi r−(s0−1),0 + λ(i)αi r+1−s0),0 + λ(i)αi r−1−s0,0 + λ(i)αi r−(s0−4),0 + · · · + λ(i)αi r+1−(s0−3),0 + · · · + λ(i)αi r−1−(s0−3),0 + · · · + λ(i)αi r+s0−2,0 + λ(i)αi r+s0−4,0 + λ(i)αi r+(s0+1),0(cid:1) r+(s0−1),0(cid:1) r+1+(s0−3),0 + λ(i)αi r−1+(s0−3),0 + λ(i)αi + αi r,s0+1 −(cid:0)αi r,s0−1 −(cid:0)αi r+1,s0 −(cid:0)αi −(cid:0)αi r−1,s0 −(cid:0)αi −(cid:0)αi the inductive hypothesis, (4.7) and the triangle inequality imply that (4.7) (4.8) r+1+s0),0(cid:1)(cid:1) r−1+s0,0(cid:1)(cid:1) , 2(cid:17)1/2 kαi r,s0+1 − (αi r−(s0+1),0 + λ(i)αi r−(s0−2),0 + · · · + λ(i)αi r+s0−2,0 + λ(i)αi r+(s0+1),0)k2 (cid:16) Xr≥s0+1 i≥1 ≤C0ε + 3(s0−2)C0ε + 3(s0−1)C0ε + 3(s0−1)C 2 0 ε ≤ 3 · 3(s0−1)C0ε = 3s0C0ε. By induction, the lemma holds for all s. The crux of Lemma 4.3 is contained in the next Lemma. 13 Lemma 4.2. Let N be a finite von Neumann algebra. Let x = (xn) be an element of (A ⊗ r,s ⊗ αn,i r,s C1)′ ∩ (LFK ⊗ N)ω with kxnk2 = 1 and EA⊗N (xn) = 0 for all n. Write xn =Pi,r,s ξi r,s ∈ L2(N) and convergence in ℓ2(FK) ⊗ L2(N). Then for some αn,i lim n→ω Xi≥1, r≥0 kαn,i r,0k2 2 = lim kαn,i 0,sk2 2 = 0. n→ω Xi≥1, s≥0 Proof. Let εn = kxn( w1 ⊗ 1) − ( w1 ⊗ 1)xnk2. For t ≥ 1, r+2,0 + · · · + λ(i)αn,i r,0 + λ(i)αn,i kαn,i r+2t,0k2 kαn,i (r+t)−t,0 + λ(i)αn,i r+t−(t−2),0 + · · · + λ(i)αn,i (r+t)+t,0k2 2(cid:1)1/2 2(cid:1)1/2 (cid:0) Xr≥0, i≥1 =(cid:0) Xr+t≥t, i≥1 ≤(cid:0) Xr+t≥t, i≥1 kαn,i r+t,tk2 2(cid:1)1/2 + 3t−1C0εn =(cid:0) Xr≥t, i≥1 2(cid:1)1/2 kαn,i r,t k2 + 3t−1C0εn, (4.9) where the inequality comes from Lemma 4.1. Another use of this lemma with s′ = t + 1 and s = t − 1 gives kαn,i (cid:0) Xr≥0, i≥1 =(cid:0) Xr+t+1≥t+1, i≥1 ≤(cid:0) Xr+t+1≥t+1, i≥1 Using r+2,0 + λ(i)αn,i r+4,0 + · · · + λ(i)αn,i r+2t,0k2 2(cid:1)1/2 kαn,i (r+t+1)−(t−1),0 + λ(i)αn,i (r+t+1)−(t−3),0 + · · · + λ(i)αn,i (r+t+1)+(t−1),0k2 2(cid:1)1/2 kαn,i r+t+1,t−1k2 2(cid:1)1/2 + 3t−2C0εn =(cid:0) Xr≥t+1, i≥1 kαn,i r,t−1k2 2(cid:1)1/2 + 3t−2C0εn. (4.10) 1 2 kβ − γk2 2 ≤ kβk2 r+2,0 + · · · + λ(i)αn,i with β = αn,i the inequalities (4.9) and (4.10) give r,0 + λ(i)αn,i 2 + kγk2 2, β, γ ∈ L2(N), r+2t,0 and γ = αn,i r+2,0 + λ(i)αn,i r+4,0 + · · · + λ(i)αn,i r+2t,0, 1 r,t k2 r,t k2 kαn,i kαn,i kαn,i 2 Xr≥0, i≥1 ≤(cid:16)(cid:0) Xr≥t, i≥1 2 + Xr≥t+1 ≤Xr≥t +(cid:16)2 · 3t−1C0(Xr≥t 2 + Xr≥t+1 =Xr≥t kαn,i r,t k2 i≥1 i≥1 i≥1 i≥1 i≥1 r,0 − (1 − λ(i))αn,i r+2,0k2 2 2(cid:1)1/2 + 3t−1C0εn(cid:17)2 kαn,i r,t−1k2 2 (4.11) +(cid:16)(cid:0) Xr≥t+1, i≥1 kαn,i r,t−1k2 2(cid:1)1/2 + 3t−2C0εn(cid:17)2 kαn,i r,t k2 2)1/2 + 2 · 3t−2C0( Xr≥t+1 i≥1 kαn,i r,t−1k2 2)1/2(cid:17)εn + (32t−2 + 32t−4)C 2 0 ε2 n kαn,i r,t−1k2 2 + C 2 0(cid:16)2 · 3t−1 + 2 · 3t−2(cid:17)εn + C 2 0 (32t−2 + 32t−4)ε2 n, 14 Summing (4.11) over t = 1, . . . , t0, and crudely estimating the first two sums on the r,sk2 2 ≤ C 2 0 kxnk2 2 = C 2 0 comes from Proposition for each n, t ∈ N, where the estimate Pr,s,i kαn,i 3.3. right-hand side gives kαn,i r,0 − (1 − λ(i))αn,i r+2,0k2 2 t0 2 Xr≥0, i≥1 ≤ 2 Xr≥0,t≥0 i≥1 kαn,i r,t k2 2 + C 2 0 (2εn(3t−1 + 3t−2) + (32t−2 + 32t−4)ε2 n) t0 Xt=1 Using Pr,t,i kαn,i r,t k2 2 ≤ C 2 0 kxnk2 2 = C 2 0 again, we have Xr≥0, i≥1 ≤ 4C 2 0 t0 kαn,i r,0 − (1 − λ(i))αn,i r+2,0k2 2 + 2C 2 0 t0 t0 Xt=1 (cid:0)2εn(3t−1 + 3t−2) + (32t−2 + 32t−4)ε2 n(cid:1) . (4.12) The first term of (4.12) can be made arbitrarily small by fixing a suitably large value for t0. Thus kαn,i r,0 − (1 − λ(i))αn,i r+2,0k2 = 0, (4.13) lim n→ω Xr≥0, i≥1 as limn→ω εn = 0. When l(i) > 1, we have λ(i) = 1 so (4.13) implies lim n→ω Xr≥0, l(i)>1 kαn,i r,0k2 2 = 0. For those i with l(i) = 1, we have (1 − λ(i)) = 1/(2K − 1). Since (cid:18)1 − 1 2K − 1(cid:19) kαn,i r,0k2 2 − 1 2K − 1 kαn,i r+2,0k2 2 2 hαn,i r,0, αn,i r+2,0i ≤ kαn,i ≤ kαn,i 2 − r,0k2 r,0 + (1 − λ(i))αn,i 2K − 1 r+2,0k2 2, (4.13) also implies that (cid:18)1 − 2 2K − 1(cid:19) Xr≥0, l(i)=1 kαn,i r,0k2 2 ≤ Xr≥0, l(i)=1 Combining the cases l(i) = 1 and l(i) > 1 gives kαn,i r,0 − (1 − λ(i))αn,i r+2,0k2 2 → 0, n → ω. lim n→ω Xr≥0, i≥1 kαn,i r,0k2 2 = 0. 15 Finally, replacing x by x∗ interchanges the roles of r and s (see Remark 3.4). Therefore lim n→ω Xs≥0, i≥1 kαn,i 0,sk2 2 = 0, which is the second limit required for the lemma. We can now deduce the main result of this section, showing that the elements of LFK which are orthogonal to A and approximately commute with w1 are essentially supported on the span of the ξi r,s for large r and s. Lemma 4.3. Let N be a finite von Neumann algebra. Let x = (xn) be an element of (A ⊗ r,s ⊗ αn,i r,s C1)′ ∩ (LFK ⊗ N)ω with kxnk2 = 1 and EA⊗N (xn) = 0 for all n. Write xn =Pi,r,s ξi r,s ∈ L2(N) and convergence in ℓ2(FK) ⊗ L2(N). For each m ∈ N, for some αn,i lim n→ω Xi≥1, r≤m or s≤m kαn,i r,sk2 2 = 0. Proof. Lemma 4.2 gives lim n→ωXi≥1 r≥0 kαn,i r,0k2 2 = 0. (4.14) For fixed s0 > 0, we can combine (4.14) with Lemma 4.1 and the triangle inequality to obtain kαn,i r,s0k2 2 = 0. lim n→ωXi≥1 r≥s0 (4.15) By replacing x with x∗ (using Remark 3.4) we interchange the roles of r and s and hence, for each r0 ≥ 0 (4.16) kαn,i r0,sk2 2 = 0. lim n→ωXi≥1 s≥r0 Combining the limits (4.15) and (4.16) establishes the lemma. 5 Counting Words in FK This section contains the combinatorial ingredients required to show that the radial masa has the asymptotic orthogonality property. Let g, h ∈ FK. Say that there are exactly i ≥ 0 cancelations in the product gh if gh = g + h − 2i. Given non-empty subsets σ, τ of {a±1 K } and n ≥ 0, let wn(σ, τ ) denote the sum of all words in FK of length n which begin with an element of σ and end with an element of τ . Write νn(σ, τ ) for the number of words in this sum so that νn(σ, τ ) = kwn(σ, τ )k2 2. We abuse notation when sets of singletons are involved, writing νn(a1, τ ) instead of νn({a1}, τ ), for example. The values νn(σ, τ ) can be explicitly computed by solving certain difference equations, see Lemma 4.3 and Remark 4.4 of [21], but for our purposes the following estimate of Sinclair and Smith will suffice. 1 , . . . , a±1 16 Proposition 5.1 (([21, Corollary 4.5])). There exists a constant C1 > 0, which depends only on K, such that νn(σ1, τ1) − νn(σ2, τ2) ≤ C1, for all n ≥ 0 and all subsets σ1, σ2, τ1, τ2 ⊂ {a±1 1 , . . . , a±1 K } with σ1 = σ2 and τ1 = τ2. Recall that for a vector ξ ∈ Wl, and m ≥ 0 we defined (ξ)m,m = (2K −1)−mq2m+l(wmξwm) in (3.2). In this section, our objective is to estimate where g1, g2, h, k1, k2 ∈ FK have g1 = g2, k1 6= e, k2 6= e and m is sufficiently large. Fix such g1, g2, h, k1 and k2 and let m > 2 max(g1, h). (cid:12)(cid:12)(cid:12) h(k1)m,m(g1 − g2), h(k2)m,mi(cid:12)(cid:12)(cid:12) , Write Sm(k1, g1) for the collection of all words of the form xk1yg1, where x, y have length m and there are no cancelations in the products xk1 and k1y. Write T m(k2, h) for the collec- tion of all words of the form hxk2y, where x, y have length m and there are no cancelations in the products xk2 and k2y. These definitions are constructed so that (k1)m,mg1 = 1 (2K − 1)m Xs∈Sm(k1,g1) s, h(k2)m,m = 1 (2K − 1)m Xt∈T m(k2,h) t, as (ki)m,m is the sum of all words xkiy with x = y = m and no cancelations in the products xki and kiy normalised by a factor of (2K − 1)−m. Therefore h(k1)m,mg1, h(k2)m,mi = 1 (2K − 1)2m T m(k2, h) ∩ Sm(k1, g1). (5.1) For 0 ≤ i ≤ g1, write Sm i (k1, g1) for those words xk1yg1 in Sm(k1, g1) which have exactly i Lemma 5.2. With the notation above, cancelations in the product yg1 so that Sm(k1, g1) =Sg1 T m(k2, h) ∩ Sm(k1, g1) − T m(k2, h) ∩ Sm(k1, g2)(cid:12)(cid:12)(cid:12) where C1 is the constant of Proposition 5.1. (cid:12)(cid:12)(cid:12) i=0 Sm i (k1, g1). ≤ C1(2K − 1)m+h, Proof. Let l = g1 = g2 and express g1 and g2 as reduced words g1 = g1,1 . . . g1,l, g2 = g2,1 . . . g2,l. There are (2K − 1)m reduced words x ∈ FK of length m with no cancelations in the product xk2. Given such an x, write Ip(x) =(cid:12)(cid:12)(cid:12) {y ∈ FK : hxk2y ∈ Sm(k1, gp), y = m, k2y = m + k2}(cid:12)(cid:12)(cid:12) and it then suffices to show that , p = 1, 2 I1(x) − I2(x) ≤ C1(2K − 1)h, (5.2) for all such x. Fix such an x, and let j be the number of cancelations in the product hx. By comparing the lengths of words, I1(x) = I2(x) = 0 unless there is some integer i with 0 ≤ i ≤ g1 and k1 + g1 − 2i = k2 + h − 2j, 17 for some j with 0 ≤ j ≤ h. Furthermore, the words hxk2y lying in the sets Sm(k1, g1) and Sm(k1, g2) appearing in (5.2) must actually lie in Sm i (k1, g2) respectively. We now consider three cases individually. i (k1, g1) and Sm Firstly suppose that hxk2 ≥ m+k1. The words of Sm(k1, g1) and Sm(k1, g2) all contain a copy of k1 from the (m + 1)-th letter to the (m + k1)-th letter. Therefore we can assume that k1 is contained in hxk2 from the (m + 1)-th letter to the (m + k1)-th letter, otherwise both I1(x) = 0 and I2(x) = 0 in which case (5.2) is immediate. For hxk2y to lie in Sm i (k1, g1), the last l − i letters of y must be g1,i+1 . . . g1,l. That is, writing y = y1 . . . ym, we must have ym−l+i+1 = g1,i+1, ym−l+i+2 = g1,i+2, . . . , ym = g1,l. Hence hxk2y must have the form hxk2y1 . . . ym−l+ig1,i+1 . . . g1,l, where there is no cancelation in the product k2y1, ym−l+i 6= y−1 1,i+1 (this condition gives no restriction if i = l) and ym−l+i 6= g1,i (this condition gives no restriction if i = 0). This last condition is necessary as we must have m−l+i+1 = g−1 hxk2y1 . . . ym−l+i(g−1 1,i . . . g−1 1,1g1,1 . . . g1,i)g1,i+1 . . . g1,l, (5.3) with no cancelation between ym−l+i and g−1 1,i as otherwise ym−l+i would cancel with g1,i+1 in (5.3). Furthermore every such y1 . . . ym−l+i gives rise to some y with hxk2y ∈ Sm(k1, g1). Therefore I1(x) = νm−l+i({a±1 1 , . . . , a±1 K } \ {last letter of k2}, {a±1 1 , . . . , a±1 K } \ {g−1 1,i+1, g1,i}) with the appropriate adjustments if i = 0 or i = l. Similarly I2(x) = νm−l+i({a±1 1 , . . . , a±1 K } \ {last letter of k2}, {a±1 1 , . . . , a±1 K } \ {g−1 2,i+1, g2,i}), with the appropriate adjustments if i = 0 or i = l. Proposition 5.1 implies that I1(x) − I2(x) ≤ C1, and so (5.2) follows. Now consider the case that m+1 ≤ hxk2 < m+k1. Let r = hxk2−m. We can assume that the first r letters of k1 make up the last r letters of hxk2, otherwise both I1(x) and I2(x) are zero. Any y for which hxk2y ∈ Sm i (k1, g1) must commence with the last (k1 − r)-letters of k1, that is we may assume y1, . . . , yk1−r are given. Arguing exactly as in the previous paragraph we see that (making adjustments if i = 0 or i = l) I1(x) = νm−l−(r+1)+i({a±1 1 , . . . , a±1 K } \ {y−1 k1−r}, {a±1 1 , . . . , a±1 K } \ {g−1 1,i+1, g1,i}) and I2(x) = νm−l−(r+1)+i({a±1 1 , . . . , a±1 K } \ {y−1 k1−r}, {a±1 1 , . . . , a±1 K } \ {g−1 2,i+1, g2,i}). Again (5.2) follows from Proposition 5.1. Finally, consider the case that hxk2 ≤ m and let r = m − hxk2 and note that r = In this case any y for m − (m + k2 + h − 2j) ≤ h (recalling that 0 ≤ j ≤ h). 18 which hxk2y lies in Sm i (k1, g1) must contain a copy of k1 from the (r + 1)-th position to the (r + k1)-th position. Any choices of y1, . . . , yr which lead to no cancelations in k2y and yrk1 are permissible and for each such choice, arguing just as before (and making appropriate adjustments when i = 0 or i = l), there are νm−l−(r+k1)+i({a±1 1 , . . . , a±1 N } \ {y−1 r+k1}, {a±1 1 , . . . , a±1 N } \ {g−1 1,i+1, g1,i}) choices of yr+k1+1 . . . ym giving rise to an element hxk1y in Sm (2K − 1)r choices of y1, . . . , yr, so Proposition 5.1 gives i (k1, g1). There are at most I1(x) − I2(x) ≤ C1(2K − 1)r ≤ C1(2K − 1)h. Thus (5.2) also holds in this case. We need one final combinatorial ingredient for our proof of the maximal injectivity of the radial masa. Lemma 5.3. Let g1, g2, h ∈ FK have g1 = g2. Write I for the set of all pairs (k1, k2) in FK with k1 6= e and k2 6= e such that h(k1)m,m(g1 − g2), h(k2)m,mi 6= 0 for some m > 2 max(g1, h). Then there exists a constant C2, depending only on K, g1 and h so that, for k1 fixed, there are at most C2 values of k2 with (k1, k2) ∈ I and for k2 fixed, there are at most C2 values of k1 with (k1, k2) ∈ I. Proof. Fix k1 ∈ FK. For m > 2 max(g1, h), the inner product h(k1)m,m(g1 − g2), h(k2)m,mi is zero when both the intersections Sm(k1, g1) ∩ T m(k2, h) and Sm(k1, g2) ∩ T m(k2, h) are empty. These intersections are both empty unless there are i and j with 0 ≤ i ≤ g1 and 0 ≤ j ≤ h and a word of length 2m+k1 +g1 −2i = 2m+k2 +h −2j containing k1 from the (m + 1)-letter to the (m + k1)-th letter and containing k2 from the (m + h − 2j + 1)-th letter to the (m + k2 + h − 2j)-th letter. Since the possible values of i and j depend only on g1 and h, it suffices to find C ′ 2, depending only on g1, h and K (and not m), such that for each i and j, there are at most C ′ 2 values of k2 satisfying these conditions. For fixed i and j, we have the freedom to choose the first max(0, 2j − h)-letters of k2 and the last max(0, g1 − 2i)-letters of k2 (subject to their being no cancelations) the remaining letters are determined by k1. Estimating crudely, there are at most (2K)h+g1 such choices and so we can take C ′ 2 = (2K)h+g1. A similar argument shows that the value of C2 obtained also satisfies the second statement of the lemma. 6 Maximal Injectivity of the Radial Masa In this section, we combine the results of the previous two sections to show that the radial masa has the asymptotic orthogonality property. The next lemma converts the combinatorial arguments of the previous section into a suitable form for use in Theorem 6.2. Lemma 6.1. Let g1, g2, h ∈ FK have g1 = g2. There exists a constant C3, which depends only on g1, g2, h and K, such that for all finite von Neumann algebras N, all 19 m > 2 max(g1, h) and all vectors η1, η2 ∈ ℓ2(FK) ⊗ L2(N) which lie in the closed lin- ear span of ξi r+m,s+m ⊗ L2(N) for i ≥ 1 and r, s ≥ 1, we have (cid:12)(cid:12)(cid:12) hη1((g1 − g2) ⊗ z1,2), (h ⊗ z2,2)η2i(cid:12)(cid:12)(cid:12) for all z1,2, z2,2 ∈ N. ≤ C3(2K − 1)−m kη1k2 kη2k2 kz1,2k kz2,2k , r+m,s+m = (ξi r,s)m,m. Each ξi Proof. Note that ξi r,s lies in Wl(i)+r+s, the span of the words of length l(i) + r + s and the map ζ 7→ ζm,m is an isometry Wl(i)+r+s → Wl(i)+r+s+2m. Thus ξi r+m,s+m lies in the span of the elements km,m for k ∈ Fr with k = l(i) + r + s. Note too that the elements (km,m)k∈Fr\{e} form an orthonormal set in ℓ2(FK). This follows as km,m consists of the normalised sum of all words of length k + 2m which contain k beginning in the (m + 1)-th position. All the words in the sum for k1 are then orthogonal to the words in the sum for k2 unless k1 = k2. The hypotheses of the lemma and the preceding paragraph allow us to write for some αk, βk ∈ L2(N) with Pk kαkk2 2 = kη2k2 2 and so km,m ⊗ βk, km,m ⊗ αk, 2 = kη1k2 η1 =Xk6=e η2 =Xk6=e 2 and Pk kβkk2 hη1((g1 − g2) ⊗ z1,2), (h ⊗ z2,2)η2i(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) h(k1)m,m(g1 − g2), h(k2)m,mihαk1z1,2, z2,2βk2i(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Xk1,k26=e kαk1k2kβk2k2(cid:12)(cid:12)(cid:12) ≤ kz1,2k kz2,2k Xk1,k26=e h(k1)m,m(g1 − g2), h(k2)m,mi(cid:12)(cid:12)(cid:12) . (6.1) Let I be the set of all pairs (k1, k2) such that the inner product h(k1)m,m(g1 − g2), h(k2)m,mi is non-zero for some m > 2 max(g1, h). By Lemma 5.2 and (5.1) h(k1)m,m(g1 − g2), h(k2)m,mi ≤ C1(2K − 1)h−m, m > 2 max(g1, h), where C1 is the constant (depending only on K) from Proposition 5.1. Applying the Cauchy- Schwartz inequality to (6.1)) gives hη1((g1 − g2) ⊗ z1,2), (h ⊗ z2,2)η2i(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) ≤C1(2K − 1)h−m(cid:0) X(k1,k2)∈I kαk1k2 2(cid:1)1/2(cid:0) X(k1,k2)∈I kβk2k2 2(cid:1)1/2 kz1,2k kz2,2k . Lemma 5.3 gives us a constant C2 (depending only on g1, h and K) such that for each k1, there are at most C2 words of k2 with (k1, k2) ∈ I and for each k2, there are at most C2 words k1 with (k1, k2) ∈ I. Therefore X(k1,k2)∈I kαk1k2 2 ≤ C2Xk1 kαk1k2 2 = C2kη1k2 2, and X(k1,k2)∈I kβk2k2 2 ≤ C2kη2k2 2. 20 Thus (cid:12)(cid:12)(cid:12) hη1((g1 − g2) ⊗ z1,2), (h ⊗ z2,2)η2i(cid:12)(cid:12)(cid:12) ≤ C1C2(2K − 1)h−m kη1k2 kη2k2 kz1,2k kz2,2k , and the result follows, where the constant C3 is given by C3 = C1C2(2K − 1)h. We are now in a position to prove that the radial masa in the free group factor LFK has the asymptotic orthogonality property (after tensoring by any finite von Neumann algebra). The maximal injectivity of the radial masa then follows from the results of Section 2. In fact we are able to show slightly more; in the theorem which follows we do not require that EA⊗N (y2) = 0. This is to be expected, as a similar result for the generator masa can be obtained from Popa's calculations in [15]. Theorem 6.2. Let N be a finite von Neumann algebra with a fixed faithful normal, nor- malised trace. Let x(1), x(2) ∈ (A ⊗ C1)′ ∩ (LFK ⊗ N)ω and y1, y2 ∈ LFK ⊗ N satisfy E(A⊗N )ω (x(1)) = E(A⊗N )ω (x(2)) = 0 and y1, y2 ∈ LFK ⊗ N with EA⊗N (y1) = 0. Then x(1)y1 ⊥ y2x(2). In particular, the radial masa A in the free group factor LFK has the asymptotic orthogonality property after tensoring by N. Proof. By linearity and density, we may first assume that y2 is of the form h⊗z2,2 for a single group element h ∈ FK and z2,2 ∈ N. We may also assume that y1 is an elementary tensor z1,1 ⊗ z1,2, where z1,1 is an element of CFK with EA(z1,1) = 0, and this space decomposes as Ll≥1 Wl ⊖Cwl. Each Wl⊖Cwl is spanned by elements of the form g1−g2 with g1 = g2 = 1. To see this take some non-zero r = Pg=l βgg ∈ Wl ⊖ Cwl and note that Pg=l βg = 0 as r ⊥ wl. The support of r consists of those g with βg 6= 0. If this support has precisely two elements then r is already a multiple of some g1 − g2. Otherwise, choose some g1, g2 in this support and define r′ = r − βg1(g1 − g2). This still lies in Wl ⊖ Cwl and has a strictly smaller support. Proceeding by induction, it follows that the elements g1 − g2 with g1 = g2 = l span Wl ⊖ Cwl. Thus we may assume that y1 = (g1 −g2)⊗z1,2 for some g1 = g2. Choose representatives n ), (x(2) n k2 = n ) of x(1), x(2) with EA⊗N (x(1) n k2 = kx(1)k2, kx(2) n ) = EA⊗N (x(2) n ) = 0 and kx(1) (x(1) kx(2)k2 for all n. r,s ⊗ αj,n,i r,s , for αj,n,i r,s ∈ L2(N). For each m > 0, Lemma 4.3 We can write x(j) shows that n = Pi,r,s ξi lim n→ω Xi≥1 r≤m or s≤m kαj,n,i r,s k2 2 = 0, j = 1, 2. In particular, if we define elements in ℓ2(FK) ⊗ L2(N), then η(j) n = Xi≥1 r,s>m r,s ⊗ αj,n,i ξi r,s (cid:10)x(1)((g1 − g2) ⊗ z1,2), (h ⊗ z2,2)x(2)(cid:11) = lim = lim n ((g1 − g2) ⊗ z1,2), (h ⊗ z2,2)x(2) n ((g1 − g2) ⊗ z1,2), (h ⊗ z2,2)η(2) n (cid:11) n (cid:11) . n→ω(cid:10)x(1) n→ω(cid:10)η(1) 21 Lemma 6.1 gives us a constant C3, which depends only on g1, h and N so that Since limn→ω kη(j) n k2 = kx(j)k2, we obtain n ((g1 − g2) ⊗ z1,2), (h ⊗ z2,2)η(2) (cid:12)(cid:12)(cid:12)(cid:10)η(1) n (cid:11)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:10)x(1)((g1 − g2) ⊗ z1,2), (h ⊗ z2,2)x(2)(cid:11)(cid:12)(cid:12)(cid:12) n (cid:13)(cid:13)2(cid:13)(cid:13)η(2) ≤ C3(2K − 1)−m(cid:13)(cid:13)η(1) n (cid:13)(cid:13)2 kz1,2k kz2,2k . ≤ C3(2K − 1)−m(cid:13)(cid:13)x(1)(cid:13)(cid:13)2(cid:13)(cid:13)x(2)(cid:13)(cid:13)2 kz1,2k kz2,2k , for all m ∈ N. Hence x(1)((g1 − g2) ⊗ z1,2) ⊥ (h ⊗ z2,2)x(2), as required. As the radial masa is singular in LFK by Rădulescu's original computation [19] and so has GN LFK (A) ⊂ A (see [23, Lemma 6.2.3(iv)]), the corollaries below follow from Corollary 2.3 and Theorem 2.8. Corollary 6.3. For each K ≥ 2, the radial masa is a maximal injective von Neumann subalgebra of LFK. Corollary 6.4. Let A be the radial masa in LFK. Let B ⊂ N be an inclusion of a type I von Neumann algebra in a finite von Neumann algebra such that B is maximal injective in N. Then A ⊗ B is maximal injective in LFK ⊗ N. Corollary 6.5. Fix n ∈ N. For 1 ≤ i ≤ n, let Ai be either a generator or radial masa in LFKi for some Ki ≥ 2. Then the tensor product A1 ⊗ . . . ⊗ An is maximal injective in LFK1 ⊗ . . . ⊗ LFKn. 7 Concluding Remarks Let B denote a generator masa in LFK. As a preliminary step in his proof that B is maximal injective in LFK, Popa uses the fact (see [16, Proposition 4.1]) that if B0 is a diffuse subalgebra of B, then B′ 0 ∩ LFK = B. It follows that if L is any intermediate von Neumann algebra between B and LFK, then there is a family of orthogonal central projections (pi)i≥0 in L which sum to 1 such that Lp0 = Bp0 and for i ≥ 1, each Lpi is a II1 factor. Popa then uses the critical asymptotic orthogonality calculation [15, Lemma 2.1] to show that these factors Lpi can not have property Γ and the maximal injectivity of B follows (see Corollary 3.3 of [15]). The argument of section 2 allows us to deduce the maximal injectivity of B directly from Popa's asymptotic orthogonality calculation. Using Ozawa's solidity of LFK [12], the additional properties of the generator masa found in Popa's proof can then be recovered. This works for any maximal injective von Neumann subalgebra of a solid II1 factor. In particular, the radial masa shares all the properties of the generator masas developed in [15]. Recall that a von Neumann algebra M is solid if the relative commutant of every diffuse von Neumann subalgebra of M is injective. Proposition 7.1. Let B be a maximal injective von Neumann subalgebra of a solid II1 factor M. If L is an intermediate von Neumann subalgebra between B and M, then there is a family of orthogonal central projections (pi)i≥0 in L which sum to 1 such that Lp0 = Bp0 and for i ≥ 1, each Lpi is a II1 factor without property Γ. 22 Proof. Note that B′ ∩M ⊆ B by maximal injectivity, so that L′ ∩M = Z(L) ⊆ Z(B). Let p0 be the maximal central projection of L such that Lp0 is diffuse. Then Lp0 ⊆ (Z(L)p0 ⊕B(1− p0))′ ∩ M and this last algebra is injective by solidity of M. Since (Z(L)p0 ⊕ B(1 − p0))′ ∩ M contains B, we have Lp0 ⊆ B by maximal injectivity of B. Let (pi)i>0 denote the minimal projections of Z(L)(1 − p0) so that for i > 0, each Lpi is a non injective factor of type II1. These factors can not have property Γ, by Popa's observation [12, Proposition 7]. C∗-algebra. Write B0 for the C∗-subalgebra of C ∗ A0 for the C∗-subaglebra of C ∗ subalgebras of C ∗ The generator and radial algebras also give rise to masas at the level of the reduced r (FK) generated by the first generator and r (FK) generated by w1. Both A0 and B0 are maximal abelian r (FK). The later claim is well known, while the former can be found in [18]. r (FK). r (FK) which 0 is an injective von Neumann subalgebra of LFK containing the 0 = B by maximal injectivity 0 and hence L0 is abelian. Since A0 and B0 are maximal Proposition 7.2. The algebras A0 and B0 are maximal nuclear C∗-subalgebras of C ∗ Proof. Represent C ∗ contains A0 or B0. Then L′′ radial masa A or the generator masa B so that L′′ of these algebras. In particular L′′ abelian subalgebras of C ∗ r (FK) on ℓ2(FK) and let L0 be a nuclear C∗-subalgebra of C ∗ r (FK) it follows that L0 = A0 or L0 = B0. 0 = A or L′′ An identical argument shows that any finite tensor product of n copies of the algebras A0 or B0 is maximal nuclear in the spatial tensor product of n copies of C ∗ r (FK), since [26, Theorem 4] the tensor product of masas is maximal abelian in the spatial tensor product. As the algebra B0 has the extension property ([2, Example (i)]), it follows that B0 ⊗ B0 is r (FK) for any C∗-norm α, [28]. However, B0 ⊗ B0 need not be a a masa in C ∗ maximal nuclear C∗-algebra in C ∗ r (FK) when α is larger than the minimal norm. We would like to thank Simon Wassermann bringing this to our attention and for allowing us to include the following argument here. r (FK) ⊗α C ∗ r (FK) ⊗α C ∗ r (FK) ⊙ C ∗ r (FK) ⊙ C ∗ r (FK) → C ∗ r (FK) ⊗ C ∗ Let α denote the conjugacy norm on C ∗ r (FK), given by the representation r (FK) on B(ℓ2(FK)), where λ and ρ are the left and right regular λ × ρ of C ∗ representations respectively. The kernel of the canonical ∗-homomorphism π : C ∗ r (FK) ⊗α C ∗ r (FK) consists precisely of the compact operators K(ℓ2(FK)) and this forms the only non-trivial ideal in this tensor product [1] (see also [27]). Consequently the C∗-subalgebra D of C ∗ r (FK) generated by B0 ⊗ B0 and K(ℓ2(FK)) is a nuclear extension of a nuclear C∗-algebra so nuclear. Furthermore, this algebra D is a maximal nuclear C∗-subalgebra of C ∗ r (FK). Indeed, if D1 is a nuclear C∗-subalgebra of C ∗ r (FK) containing D, then π(D1) is a nuclear C∗-subalgebra containing B0 ⊗ B0 in C ∗ r (FK), we have π(D1) = B0 ⊗ B0. Thus D1 = D. r (FK). By maximal nuclearity of B0 ⊗ B0 in C ∗ r (FK) ⊗α C ∗ r (FK) ⊗α C ∗ r (FK) ⊗α C ∗ r (FK) ⊗ C ∗ r (FK) ⊗ C ∗ References [1] C. A. Akemann and P. A. Ostrand. On a tensor product C ∗-algebra associated with the free group on two generators. J. Math. Soc. Japan, 27(4):589 -- 599, 1975. [2] R. J. Archbold. Extensions of states of C ∗-algebras. J. London Math. Soc. (2), 21(2):351 -- 354, 1980. 23 [3] N. P. Brown and N. Ozawa. C ∗-algebras and finite-dimensional approximations, vol- ume 88 of Graduate Studies in Mathematics. American Mathematical Society, Provi- dence, RI, 2008. [4] J. M. Cohen. Operator norms on free groups. Boll. Un. Mat. Ital. B (6), 1(3):1055 -- 1065, 1982. [5] A. Connes. Classification of injective factors. Cases II1, II∞, IIIλ, λ 6= 1. Ann. of Math. (2), 104(1):73 -- 115, 1976. [6] J. Dixmier. Sous-anneaux abéliens maximaux dans les facteurs de type fini. Ann. of Math. (2), 59:279 -- 286, 1954. [7] J. Fang. On completely singular von Neumann subalgebras. Proc. Edinb. Math. Soc. (2), 52, 607 -- 618, 2009. [8] J. Fang. On maximal injective subalgebras of tensor products of von Neumann algebras. J. Funct. Anal., 244(1):277 -- 288, 2007. [9] J. Fang, R. R. Smith, S. A. White, and A. D. Wiggins. Groupoid normalizers and tensor products. J. Funct. Anal., 258, 20 -- 49, 2010. [10] L. Ge and R. Kadison. On tensor products for von Neumann algebras. Invent. Math., 123(3):453 -- 466, 1996. [11] R. V. Kadison. Diagonalizing matrices. Amer. J. Math., 106(6):1451 -- 1468, 1984. [12] N. Ozawa. Solid von Neumann algebras. Acta Math., 192(1):111 -- 117, 2004. [13] N. Ozawa and S. Popa. On a class of II1 factors with at most one cartan subalgebra. Ann. of Math. (2), 172, 713-749, 2010. [14] S. Popa. On a problem of R. V. Kadison on maximal abelian ∗-subalgebras in factors. Invent. Math., 65(2):269 -- 281, 1981/82. [15] S. Popa. Maximal injective subalgebras in factors associated with free groups. Adv. in Math., 50(1):27 -- 48, 1983. [16] S. Popa. Orthogonal pairs of ∗-subalgebras in finite von Neumann algebras. J. Operator Theory, 9(2):253 -- 268, 1983. [17] S. Popa. Strong rigidity of II1 factors arising from malleable actions of w-rigid groups, I. Invent. Math., 165:369 -- 408, 2006. [18] T. Pytlik. Radial functions on free groups and a decomposition of the regular represen- tation into irreducible components. J. Reine Angew. Math., 326:124 -- 135, 1981. [19] F. Rădulescu. Singularity of the radial subalgebra of L(FN ) and the Pukánszky invari- ant. Pacific J. Math., 151(2):297 -- 306, 1991. 24 [20] J. Shen. Maximal injective subalgebras of tensor products of free group factors. J. Funct. Anal., 240(2):334 -- 348, 2006. [21] A. M. Sinclair and R. R. Smith. The Laplacian MASA in a free group factor. Trans. Amer. Math. Soc., 355(2):465 -- 475 (electronic), 2003. [22] A. M. Sinclair and R. R. Smith. The Pukánszky invariant for masas in group von Neumann factors. Illinois J. Math., 49(2):325 -- 343, 2005. [23] A. M. Sinclair and R. R. Smith. Finite von Neumann algebras and masas, volume 351 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 2008. [24] A. M. Sinclair, R. R. Smith, S. A. White, and A. Wiggins. Strong singularity of singular masas in II1 factors. Illinois J. Math., 51(4):1077 -- 1083, 2007. [25] Ş. Strătilă and L. Zsidó. The commutation theorem for tensor products over von Neu- mann algebras. J. Funct. Anal., 165(2):293 -- 346, 1999. [26] S. Wassermann. The slice map problem for C ∗-algebras. Proc. London Math. Soc. (3), 32(3):537 -- 559, 1976. [27] S. Wassermann. Tensor products of free-group C ∗-algebras. Bull. London Math. Soc., 22(4):375 -- 380, 1990. [28] S. Wassermann. Tensor products of maximal abelian subalgebras of C ∗-algebras. Glasg. Math. J., 50(2):209 -- 216, 2008. 25
1305.6004	1	1305	2013-05-26T08:35:46	Semigroup Operator Algebras and Quantum Semigroups	[ "math.OA", "math.QA" ]	A detailed study of the semigroup $C^\ast$-algebra is presented. This $C^\ast$-algebra appears as a "deformation" of the continuous functions algebra on a compact abelian group. Considering semigroup $C^\ast$-algebras in this framework we construct a compact quantum semigroups category. Then the initial group is a compact subgroup of the new compact quantum semigroup, the natural action of this group is described. The dual space of such $C^\ast$-algebra is endowed with Banach -algebra structure, which contains the algebra of measures on a compact group. The dense weak Hopf -algebra is given. It is shown that the constructed category of compact quantum semigroups can be embedded to the category of abelian semigroups.	math.OA	math	SEMIGROUP OPERATOR ALGEBRAS AND QUANTUM SEMIGROUPS AUKHADIEV MARAT, GRIGORYAN SUREN, AND LIPACHEVA EKATERINA Abstract. A detailed study of the semigroup C ∗-algebra is presented. This C ∗-algebra appears as a "deformation" of the continuous functions algebra on a compact abelian group. Considering semigroup C ∗-algebras in this framework we construct a compact quantum semigroups category. Then the initial group is a compact subgroup of the new compact quantum semigroup, the natural action of this group is described. The dual space of such C ∗-algebra is endowed with Banach -algebra structure, which contains the algebra of measures on a compact group. The dense weak Hopf -algebra is given. It is shown that the constructed category of compact quantum semigroups can be embedded to the category of abelian semigroups. Contents red(S) red(S) red(S) Introduction Ideals in the algebra C∗ 1. 2. A C∗-algebra C∗ 2.1. Preliminaries 2.2. Algebra C∗ 2.3. Dynamical system, generated by the regular representation 2.4. 3. Quantization of semigroup algebras 3.1. Compact quantum semigroup 3.2. Quantization of the semigroup S 3.3. Weak Hopf algebra 3.4. Compact quantum subgroup 3.5. Dual algebra and Haar functional 3.6. Category of quantum semigroups 3.7. Different comultiplications on the Toeplitz algebra References 1 3 3 4 6 8 10 10 12 14 15 17 18 19 20 1. Introduction There exist two different approaches to quantization -- algebraic and topologi- cal. The first one is based on deformation of universal enveloping algebras. This approach takes its roots in the early works by V.G. Drinfeld [Dri86]. The the- ory of compact quantum groups and semigroups is a part of the second approach, Date: March, 2013. 1991 Mathematics Subject Classification. Primary 46L05, 16W30; Secondary 46L87, 46L65. Key words and phrases. C ∗-algebra, compact quantum semigroup, Haar functional, Toeplitz algebra, isometric representation, inverse representation, Banach algebra, weak Hopf algebra. 1 2 AUKHADIEV MARAT, GRIGORYAN SUREN, AND LIPACHEVA EKATERINA which started with S.L. Woronowicz's works [Wor87b] and L.L. Vaksman and Ya.S. Soibel'man. It was shown later in [RTF90] that these two approaches are equivalent. Let us explain in short the process of quantization in the C∗-algebras framework. Let P be a compact semigroup, i.e. a Hausdorff compact space with continuous associative operation (x, y) → xy. Denote by C(P ) the algebra of continuous functions on P . Then C(P ) is a commutative unital C∗-algebra, which encodes all the topological information about the space P . We identify C(P × P ) with In what follows symbol ⊗ denotes minimal C∗-tensor product. C(P ) ⊗ C(P ). Define a map ∆ : C(P ) → C(P ) ⊗ C(P ): ∆(f )(x, y) = f (xy). Obviously, ∆ is a continuous unital -homomorphism. The associativity of multi- plication in P is encoded in the following property of ∆, called coassociativity: (∆ ⊗ id)∆ = (id ⊗ ∆)∆ Thus, all the information about the compact group P is encoded in a pair (C(P ), ∆). Now, let A be a commutative unital C∗-algebra. Then, by Gelfand-Naimark theorem, A is isomorphic to the algebra of all continuous functions C(P ) on some compact Hausdorff space P . If A is endowed with a -homomorphism ∆ : A → A ⊗ A, satisfying the coassociativity condition, then equation f (xy) = ∆(f )(x, y) endowes a semigroup structure on P . Quantization means a passage from the commutative algebra C(P ) to a non- commutative unital C∗-algebra A. Due to noncommutativity this algebra does not corresppond to any semigroup. Nevertheless, algebra A can be regarded as an al- gebra of continuous functions on some virtual compact geometrical object. This object is called a quantum space. In other words, quantum space is an object of a category, dual to the category of C∗-algebras (see [Wor91] for details). Usually quantum space correspoding to C∗-algebra A is denoted by QS(A) [So l09]. A unital -homomorphism ∆ : A → A ⊗ A satisfying coassociativity condition is called comultiplication. Analogously to the classical case, ∆ turns quantum space QS(A) into a quantum semigroup. Then the algebra A with its comultiplication represents an algebra of functions on a quantum semigroup [MT04]. The pair (A, ∆) is called a compact quantum semigroup [A.M98]. Compact quantum semigroup (A, ∆) is called commutative, if σ∆ = ∆, where σ : A ⊗ A → A ⊗ A is a flip: σ(a ⊗ b) = b ⊗ a. By definition of S.L. Woronowicz [Wor98], the compact quantum semigroup (A, ∆) is a compact quantum group if each of the sets (1 ⊗ A)∆(A), (A ⊗ 1)∆(A) is linear dense in A ⊗ A. In the classical case (when A is commutative) (A, ∆) corresponds to an ordinary group, the density conditions of the above-mentioned sets corresponds to existense of inverse elements in P . At the moment the quantum groups theory is widely represented (see [Wor87a], [VVD01] [A.M98], [Pod95]). Most of non-trivial examples are constructed by defor- mation of Hopf -algebras generated by classical groups, and then taking a closure. Futher in each case it should be proved that the comultiplication on the Hopf algebra extends to C∗-algebra. SEMIGROUP OPERATOR ALGEBRAS AND QUANTUM SEMIGROUPS 3 On the other hand, problem of constructing a compact quantum semigroup on a concrete C∗-algebra remains open. This problem is unsolved even for a well-known Toeplitz algebra. There are not so many works devoted to the theory of quantum semigroups. The problem of giving a nontrivial structure of a compact quantum semigroups on a concrete C∗-algebra has been actively studied only the last years. For example, in paper [Kaw08] the comultiplication is defined an a direct sum of Cuntz algebras. This work is devoted to the construction of compact quantum semigroups on semigroup C∗-algebras, which are generated by the "deformed" representations of commutative C∗-algebra of continuous functions on a compact abelian group G by multiplicators on L2(G). Such algebras form a class of noncommutative C∗- algebras, close in their characteristics to C(G). The paper consists of two sections. The first is devoted to properties of a C∗- algebra C∗ red(S). Some basic facts of this section can be found in [GS09]. The second section is devoted to the cunstruction and investigation of properties of compact quantum semigroup on the C∗-algebra C∗ red(S). 2. A C∗-algebra C∗ red(S) 2.1. Preliminaries. Let Γ be an additive discrete torsion-free group. follows S denotes a semigroup, which generates group Γ, i.e. In what Γ = {a − b a, b ∈ S}. Unless otherwise stated, we suppose that S contains 0 -- a neutral element of the group Γ. We endow S with a quasi-order: a ≺ b, if b = a + c, where c ∈ S. This quasi-order is called natural. Quasi-order is called an order, if a ≺ b and b ≺ a imply a = b. Quasi-oder is an order if and only if S does not contain any non-trivial group. The order on S is called total, if for any elements a and b, a ≺ b or b ≺ a. In case the natural order on S is total, Γ = SS(−S), where −S is a set of oppposite elements to the elements of S, and ST(−S) = {0}. Let Is(H) be a semigroup of all isometric operators on a Hilbert space H. A semigroup homomorphism π : S → Is(Hπ) is called an isometric representa- tion of semigroup S, where Hπ is a Hilbert space. In what follows the image if an element a ∈ S in Is(Hπ) is denoted by Tπ(a). Let Cπ(S) be a closed C∗-subalgebra in B(Hπ), generated by operators Tπ(a) and Tπ(a)∗. Representation π : S → Is(Hπ) is called irreducible, if algebra Cπ(S) is irre- ducible on Hπ. Two representations π1 : S → Is(Hπ1 ) and π2 : S → Is(Hπ2) generate canonically isomorphic C∗-algebras if there exists an -isomorphism τ : Cπ1 (S) → Cπ2 (S) such that the following diagram is commutative. (2.1) π1−→ Cπ1 (S) S ↓ π2 ւ τ Cπ2 (S) Two representations π1 : S → Is(Hπ1 ) and π2 : S → Is(Hπ2 ) are unitarily equiv- alent, if there exists a unitary operator U : Hπ1 → Hπ2 such that π1 = U π2U ∗. A representation of semigroup S on l2(S) by a shift operator a → Ta is called a regular representation. A C∗-algebra generated by this representation is denoted C∗ red(S). 4 AUKHADIEV MARAT, GRIGORYAN SUREN, AND LIPACHEVA EKATERINA Coburn [Cob67] proved that all isometric non-unitary representations of semi- group Z+ generate canonically isomorphic C∗-algebras. Later Douglas [Dou72] and Murphy [Mur87] generalized this result for totally ordered semigroups. Then the following question seems natural: which conditions on the semigroup S are equivalent to the property that any isometric representations generate canonically isomorphic C∗-algebras? Theorem 1. [AT12] The following conditions are equivalent. (1) all isometric non-unitary representations of S generate canonically isomor- phic C∗-algebras, (2) the natural order is total on S. We call operators of type Tπ(a) and Tπ(a)∗ elementary monomials, and finite product of elementary monomials is called a monomial. A set of all monomials forms a semigroup, which we denote by S∗ π. A representation π : S → Is(Hπ) is called an inverse representation of the semigroup S, if S∗ π is an inverse semigroup with respect to -operation. Recall that a semigroup is called inverse (or a generalized group), if it contains with each element x a unique x∗, called inverse to x, such that Theorem 2. The following conditions are equivalent. xx∗x = x, x∗xx∗ = x∗. (1) The natural order is total on S, (2) For any non-unitary representations π1 : S → Is(Hπ1 ) and π2 : S → Is(Hπ2 ) there exist -isomorphisms between semigroups S∗ π1 and S∗ π2. In [AT12] it was shown that each regular representation is inverse. Therefore, there exists at least one inverse representation of S. 2.2. Algebra C∗ red(S). Recall that l2(S) = {f : S → CXa∈S f (a) 2< ∞} is a Hilbert space with respect to scalar product (f, g) =Xa∈S f (a)g(a). A set of functions {ea}a∈S, ea(b) = δa,b forms an orthonormal basis in l2(S). We define a representation of S by a map a → Ta, where Taeb = ea+b for any b ∈ S. C∗-algebra, generated by operators Ta and T ∗ red(S) and called a reduced semigroup C∗-algebra of the semigroup S. We denote by S∗ red a semigroup of monomials, generated by this representation, and by P (S) an algebra of finite linear combinations of such monomials. b for all a, b ∈ S is denoted by C∗ Semigroup S is a net with respect to the natural order defined in previous section: for any a and b in S there exists c such that a ≺ c or b ≺ c. Lemma 1. Let V be a monomial in S∗ Then there exist a, b ∈ S such that π, where π : S → Is(Hπ) is a representation. Tπ(c)∗V Tπ(c) = Tπ(a)∗Tπ(b). lim c∈S SEMIGROUP OPERATOR ALGEBRAS AND QUANTUM SEMIGROUPS 5 Proof. Suppose V is a multiple of {Tπ(ai)∗}n j=1. Take c = a + b, bj. Then relations Tπ(d)∗Tπ(d) = I and Tπ(d)Tπ(k) = i=1 and {Tπ(bj)}m where a = ai, b = Tπ(d + k) imply nPi=1 mPj=1 Tπ(c)∗V Tπ(c) = Tπ(a)∗Tπ(b). Now let c ≺ d, i.e. d = c + k for some k ∈ S. Then Tπ(d)V Tπ(d) = Tπ(k)∗(Tπ(c)∗V Tπ(c))Tπ(k) = Tπ(k)∗Tπ(a)∗Tπ(b)Tπ(k) = Tπ(a)∗Tπ(b). And we have the required relation. (cid:3) Using this result we introduce the following notion. In conditions of Lemma 1, we call b − a ∈ Γ an index of monomial V , denoted by indV = b − a. Note that this definition of index coincides with the same notion for a regular representation introduced in [GS09]. The following statement follows from Lemma 1. Lemma 2. ind(V1 · V2) = indV1 + indV2. Lemma 3. Let indV = a−b for a monomial V in S∗ or V ed = 0. Proof. Since Taeb = ea+b, and T ∗ have red. Then either V ed = ed+(a−b) a eb = eb−a if a ≺ b or T ∗ a eb = 0 in other case, we if V ed 6= 0, then T ∗ c V Tced = V ed. Passing to the limit in S we obtain T ∗ a Tbed = V ed. Therefore, V ed = ed+a−b. (cid:3) Remark 1. One can easily verify that due to Lemma 3, monomials with zero index are projections, with eigenvectors in a basis ea, a ∈ S. Therefore S∗ red is an inverse semigroup, where passing to inverse element is realized by involution. Denote by Ac a closed linear space in C∗ red(S), generated by linear combinations of monomials with index c. Lemma 4. (1) Ac · Ab ⊆ Ac+b, (2) A0 is a commutative C∗-algebra Proof. 1) Take monomials V and W with corresponding indices c and b. Then ind(V · W ) = c + b. Consequently, V · W ∈ Ac+b. Statement 2) is obvious. (cid:3) Note that Ac = T ∗ b A0Ta, if c = a − b. Indeed, for V in Ac we have V = T ∗ b TbV T ∗ a Ta. Therefore, c = indV = indT ∗ b + ind(TbV T ∗ a ) + indTa = = a − b + indTbV T ∗ a . Denote by Pa : l2(S) → l2(S) an orthogonal projection on space Cea, Paf = (f, ea)ea. Define a conditional expectation Q : B(l2(S)) → B(l2(S)), Q(A) = Consequently, TbV T ∗ a ∈ A0. PaAPa, where A ∈ B(l2(S)). La∈S Lemma 5. A restriction of Q to C∗ to A0. red(S) is a conditional expectation from C∗ red(S) 6 AUKHADIEV MARAT, GRIGORYAN SUREN, AND LIPACHEVA EKATERINA Proof. Since ea, a ∈ S are eigenvectors for any monomial with zero index, we have Q(V ) = V, if indV = 0. Let indV = c. Then, by Lemma 3 indPaV Pa = 0 for any a ∈ S. Therefore, indQ(V ) = 0 and finally, Q(C∗ (cid:3) red(S)) = A0. Fix an element c in group Γ, and define a linear map Qc : B(l2(S)) → B(l2(S)) by the following relation. Qc(A) =Xa∈S Pa+cAPa. Here Pa+c = 0 if a + c is not contained in S. Lemma 6. Operator Qc maps the algebra C∗ red(S) on space Ac. Proof. Take monomial V with indV = c. We have V Pa = Pa+cV for any a ∈ S. red(S)) = Hence, Qc(V ) = V . Ac. (cid:3) If indV 6= c, then Qc(V ) = 0. Therefore, Qc(C∗ For any A in C∗ red(S) we call Ac = Qc(A) a c-coefficient of A. Lemma 7. Following conditions are equivalent: (1) A = 0, (2) Qc(A) = 0 for all c ∈ Γ. Proof. Condition Ac = 0 implies that PaAPb = 0 for any a, b ∈ S. Equivalently, (Aeb, ea) = 0 for any a and b in S. Since {ea}a∈S forms an orthonormal basis in l2(S), we get A = 0. The reverse follows immediately from Lemma 6. (cid:3) The following statement is an obvious consequence. Theorem 3. C∗-algebra C∗ resented in the following way red(S) is a graded C∗-algebra and can be formally rep- C∗ red(S) =Mc∈Γ Ac, and Ac = T ∗ a A0Tb, if c = b − a, b, a ∈ S. 2.3. Dynamical system, generated by the regular representation. A space C(G, C∗ red(S), is a C∗-algebra with pointwise multiplication, pointwise involution and uniform norm: red(S)) of all continuous functions on a group G taking values in C∗ f ∗(α) = (f (α))∗, kf k∞ = sup α∈G red(S)) is a function of the following type kf (α)k, f ∈ C(G, C∗ red(S)). A generalized polynomial in C(G, C∗ P (α) = nXi=1 χai(α)Ai, where Ai ∈ C∗ a dense involutive subalgebra in the algebra C(G, C∗ red(S), ai ∈ Γ. A set of generalized polynomials P (G, C∗ red(S)). Obviously, red(S)) forms Ai =Z χ−ai(α)P (α)dµ(α), where µ is a normed Haar measure of group G. G SEMIGROUP OPERATOR ALGEBRAS AND QUANTUM SEMIGROUPS 7 If {Pn(α)}∞ then the limit n=1 is a Cauchy sequence, which converges to A(α) in C(G, C∗ red(S)), n→∞Z lim G Pn(α)χ−a(α)dµ(α) exists and equals a-coefficient of A(α): Aa =Z G χ−a(α)A(α)dµ(α). Thus, each function A(α) in C(G, C∗ Fourier series red(S)) is uniquely represented in a formal with Fourier coefficients χa(α)Aa A(α) ≃Xa∈Γ Aa =Z χ−a(α)A(α)dµ(α). Denote by C∗(G, S) a C∗-subalgebra of C(G, C∗ red(S)), generated bby elements G for all a in S. cTa(α) = χa(α)Ta and cTa ∗ (α) = χ−a(α)Ta, Lemma 8. Each function A(α) ∈ C∗(G, S) is formally represented by A(α) =Ma∈Γ χaAa, where Aa lies in space Aa. sented in a form Proof. A finite product bV (α) of functions Ta(α) and T ∗ where V is a monomial, i.e. bV without χ, and c = indV . Hence, each generalized polynomial P (α) in P (G, S) = P (G, C∗(G, S))T C∗(G, S) is represented in a form bV (α) = χc(α)V, b (α), a, b ∈ S can be repre- P (α) = χai(α)Aai , nXi=1 where Aai ∈ Aai. (cid:3) Theorem 4. There exists a -isomorphism between C∗-algebras C∗(G, S) and C∗ red(S). Proof. Let e be a neutral element of G. Define a map be : C(G, S) → C∗ be(A(α)) = A(e). Obviously, be is a surjective -homomorphism. Suppose A(α) ∈ Kerbe and A 6= 0. Then red(S) by χa(α)Aa, A(α) =Xa∈Γ where Aa ∈ Aa. Since A(e) = 0, by virtue of Lemma 7, Aa = 0 for any a ∈ Γ. (cid:3) Therefore, A(α) = 0, which shows that be is a -isomorphism. 8 AUKHADIEV MARAT, GRIGORYAN SUREN, AND LIPACHEVA EKATERINA red(S)) → C(G, C∗ Consider a shift bα : C(G, C∗ red(S)), bα(f ) = fα, where fα(β) = f (αβ). Obviously, it is a -automorphism on algebra C(G, C∗(S)) for all α ∈ G. Then one can define a representation τ : G → Aut(C(G, C∗ red(S))) of G into - automorphism group of C∗-algebra C(G, C∗ red(S)). Due to G-invariance of algebra P (G, S), its closure C∗(G, S) is invariant as well. This implies that representa- tion τ is also a reresentation of G into group Aut(C∗(G, S)). Hence, a triple (C∗(G, S), G, τ ) is a C∗-dynamical system. Finally, we obtain the following state- ment. Theorem 5. Mapping τ is a representation of group G in Aut(C∗ red(S)). Thus, there exists a non-trivial C∗-dynamical system (C∗ red(S), G, τ ). 2.4. Ideals in the algebra C∗ Hilbert space L2(G, dµ): red(S). Consider a representation of C(G) on a f 7→ Tf , Tf g = f · g. This map realizes a -isomorphism between C(G) and a commutative subalgebra in the algebra B(L2(G, dµ)) of all bounded linear operators on L2(G, dµ). One can easily verify that kTf k = kf k∞. Denote by H 2(S) subspace in L2(G, dµ) of functions with spectrum in S, i.e. f a χa. The Fourier transform maps H 2(S) on lies in H 2(S) if and only if f = Pa∈S l2(S). Let PS : L2(G, dµ) → H 2(S) be an orthonormal projection on H 2(S). C f Lemma 9. For any f ∈ C(G) we have kPSTf PSk = kf k∞. Proof. Assume that f is a polynomial: f = nXai∈Γ,i=1 C f ai χai. Since kTf k = kf k∞ for any ε > 0 there exists g = mXj=1 C g bj χbj in L2(G, dµ) such that kgk2 = 1 and kTf gk2 > kf k∞ − ε. Take a in S such that a+bj and a+ai +bj lie in S for any 1 ≤ i ≤ n and 1 ≤ j ≤ m. Then functions χag and χaf g lie in H 2(S). Obviously, kχagk2 = kgk2 = 1 and Hence, PSTf PSχag = χaf g. kPSTf PSχagk2 = kf gk2 > kf k∞ − ε. (cid:3) Note that Fourier transform maps PSTχcPS to Tc in H 2(S), and PSTχ−cPS to a in l2(S) for any c ∈ Γ. If c = b − a, where a, b ∈ S, then T ∗ a Tb, PSTχcPS = PSTχ−aPSTχbPS = T ∗ (PSTχaPS)∗ = PSTχ−a PS. SEMIGROUP OPERATOR ALGEBRAS AND QUANTUM SEMIGROUPS 9 In what follows for c = b − a we denote by Tc operator T ∗ Denote by Tf a Fourier transform of PSTf PS, and call f a symbol of operator a Tb. Tf . Corollary 1. Assume that f ∈ C(G) is represented as f ≃Xc∈Γ C f c χc. Then the Fourier transform of PSTf PS is represented as Tf ≃Xc∈Γ C f c Tc with kTf k = kf k∞. Corollary 2. The C∗-subalgebra in B(l2(S)), generated by operators Tf , f ∈ C(G) coincides with C∗ red(S). Thus, mapping f → Tf is an isometric embedding of C(G) into C∗ Note that Tf · Tg = Tf ·g if and only if the spectrum sp(f ) and sp(g) of functions red(S). f and g is contained in S. A commutator ideal K of algebra C∗ red(S) is an ideal, generated by commutators TcTd − TdTc, where d, c ∈ Γ. Note that K is a non-unital C∗-algebra. A restriction of representation τ : G → Aut(C∗ red(S)) onto K generates represen- tation τK : G → AutK: τ (α)(TcTd − TdTc) = χc+d(α)(TcTd − TdTc). Consequently, dynamical system (C∗ red(S)/K, G, τ ), where C∗ red(S), G, τ ) implies two other dynamical red(S)/K is a quotient algebra. systems (K, G, τ ) and (C∗ Lemma 10. A ∈ C∗ red(S). f ∈ C(G), Tf = lim −→ c∈S T ∗ c ATc. Proof. C∗ red(S). A = nXi=1 C A i Vi, C A i ∈ C, Vi ∈ S∗ red. T ∗ c ATc = lim −→ c∈S nXi=1 CiTdi, di = indVi, i = 1, 2, ...n. 1, nXi=1 CiTdi = Tg, g = nXi=1 Ciχdi. Lemma 11. The quotient algebra C∗ red(S)/K is isomorphic to C(G). Proof. Denote by [Tc] equivalence class of monomial Tc by ideal K. Obviously, Hence, for a monomial V with c = indV , we have [Tc] = [V ]. It implies that for [Tc][Td] = [Tc+d]. (cid:3) A = nXi=1 C A i Vi, 10 AUKHADIEV MARAT, GRIGORYAN SUREN, AND LIPACHEVA EKATERINA we have the following relation for equivalence classes [A] = nXi=1 C A i [Tdi], where di = indVi. Therefore, [A] = [Tf ], f = C A i χdi. Finite linear combinations of monomials are dense in C∗ C(G) such that [A] = [Tf ]. red(S). Thus, for any A ∈ C∗ red(S) there exists f ∈ nPi=1 Note that if [A] = [Tf ], [B] = [Tg], [A] + [B] = [Tf +g] and [A · B] = [Tf ·g] with k[Tf ]k ≤ kTf k = kf k∞. Consequently, mapping f → [Tf ] is a -isomorphism between C∗-algebras C(G) and C∗ (cid:3) red(S)/K. Lemma 12. An element A ∈ C∗ red(S) lies in K if and only if T ∗ c ATc = 0. lim −→ c∈S Proof. Suppose lim −→ c∈S T ∗ c ATc = 0 herewith A /∈ K. Then [A] = Tf 6= 0 for some function f with the following relation [T ∗ c ATc] = [T ∗ c Tf Tc] = [Tf ]. The last relation follows from Lemma 10. This causes a contradiction. The reverse statement is obvious. (cid:3) Theorem 6. A short exact sequence 0 → K id−→ C∗ red(S) ξ −→ C(G) → 0, where id is an embedding and ξ is a quotient map, splits by means of involution- preserving map ρ : C(G) → C∗ red(S) such that ξ ◦ ρ = id. Proof. Define an isometric map ρ : C(G) → C∗ A ∈ K we have red(S), which maps f → Tf . For any kTf + Ak ≥ kTf k. Indeed, Therefore, C∗ kTf + Ak ≥ lim −→ c∈S red(S) = ρ(C(G))L K. kT ∗ c (Tf + A)Tck = kTf k. (cid:3) 3. Quantization of semigroup algebras 3.1. Compact quantum semigroup. Let K be a compact semigroup and denote by C(K) an algebra of continuous functions on K. It was shown in introduction that this algebra is endowed with a comultiplication ∆ such that (C(K), ∆) reflects the structure of a compact semigroup K. But not every pair (C(K), ∆) with a coassociative -homomorphism ∆ : C(K) → C(K) ⊗ C(K) encodes the structure of K. One can see this for comultiplication defined by ∆(f ) = f ⊗1. This motivates us to search for a quantization of semigroup which reflects the structure and properties of initial semigroup. The main aim of this section is to put in correspondence a compact quantum semigroup for any abelian cancellative semigroup. For this purpose we shall use the theory of C∗-algebras C∗ red(S) and results of the previous section. But in order to construct compact quantum semigroups we introduce a bit different definition. SEMIGROUP OPERATOR ALGEBRAS AND QUANTUM SEMIGROUPS 11 Let A be a C∗-, and A ⊙ A algebraic tensor product. Faithful representations π1 : A → B(H1) and π2 : A → B(H2) generate representation π = π1 ⊗π2 : A⊙A → B(H1) ⊗ B(H2). Closure of π(A ⊙ A) in operator norm is a tensor product A ⊗π A, generated by representation π. Consider a dual space (A ⊗π A)∗ to the C∗-algebra A ⊗π A, i.e. a space of all linear continuous functionals on A ⊗π A. For any ξ, η ∈ (A)∗ there exists a unique functional ξ ⊗ η on A ⊗π A, such that (ξ ⊗ η)(A ⊗ B) = ξ(A)η(B), A, B ∈ A. Therefore one can define (A)∗ ⊗π (A)∗ as a closure of (A)∗ ⊙ (A)∗ in Banach space (A ⊗π A)∗. Note that (A)∗ ⊗π (A)∗ = (A ⊗π A)∗ if and only if A is a finite- dimensional C∗-algebra. Every embedding ∆ : A → A ⊗π A induces a Banach algebra structure on A∗: product of ξ, η ∈ A∗ is a functional ξ × η ∈ A∗ such that Note that (ξ × η)(A) = (ξ ⊗ η)∆(A). kξ × ηk ≤ kξk · kηk. Generally, Banach algebra A∗ need not to be associative. If ∆ : A → A ⊗π A is not only an embedding but also a -homomorphism and A∗ is an associative Banach algebra, then we call ∆ a π1 ⊗ π2-comultiplication, and pair (A, ∆) is called a compact quantum semigroup. If any functionals commute, ξ × η = η × ξ, then quantum semigroup (A, ∆) is called cocommutative. Now the coassociativity condition for ∆ makes no sense: (I ⊗ ∆)∆ = (∆ ⊗ I)∆. Proposition 1. Let ∆ : A → A ⊗π A be a comultiplication. If π′ -representations of algebra A, then there exists comultiplication ∆′ : A → A ⊗′ where π′ = π′ 2 are faithful π A, 1, π′ 1 ⊗ π′ 2. Proof. An identity map id on A induces a canonical isomorphism I ⊗ I : A ⊗π A → π A. Define ∆′ = (I ⊗ I)∆. It is sufficient to show (ξ ⊗ η)∆′ = (ξ ⊗ η)∆ for A ⊗′ any ξ, η ∈ A∗. Indeed, we have (ξ ⊗ η)∆′(A) = (ξ ⊗ η)(I ⊗ I)∆(A) = (ξ ⊗ η)∆(A). (cid:3) This proposition implies the following statement, which shows the equivalence of two definitions. Corollary 3. Following conditions are equivalent: (1) There exists a (standard) comultiplication ∆ : A → A ⊗min A (2) For any faithful -representations π1, π2 of the C∗-algebra A there exists a comultiplication ∆′ : A → A ⊗π A, where π = π1 ⊗ π2. We provide this definition with following examples. Let C(S1) denote the C∗-algebra of all continuous functions on a unit circle S1, and M (S1) = (C(S1))∗ -- a space of all regular Borel measures on S1. 12 AUKHADIEV MARAT, GRIGORYAN SUREN, AND LIPACHEVA EKATERINA Example 1. Define ∆ : C(S1) → C(S1) ⊗ C(S1) by setting ∆(einθ) = einθ ⊗ einτ , n ∈ Z. By virtue of Grothendieck Theorem, we have C(S1) ⊗ C(S1) = C(S1 × S1). Due to this equation, the map ∆ has the following form: ∆(f )(eiθ, eiτ ) = f (eiθ · eiτ ). In this case a pair (C(S1), ∆) is a compact quantum group and M (S1) is an asso- ciative Banach algebra with respect to convolution. Example 2. Define ∆2 : C(S1) → C(S1)⊗C(S1), which maps eiθ → ei2θ⊗ei2τ , i.e. ∆(f )(eiθ,eiτ ) = f (ei2θ ·ei2τ ). Then (C(S1), ∆) is not a compact quantum semigroup. The space M (S1) is a non-associative Banach algebra with a zero divisor in this case. Recall that a compact quantum semigroup (A, ∆) is called a compact quantum group, if linear spaces (A ⊗min I)∆(A) and (I ⊗min A)∆(A) are dense in A ⊗min A. 3.2. Quantization of the semigroup S. In this section we give a comutiplication ∆ : C∗ red(S) and show some properties of a compact quantum semigroup (C∗ red(S) → C∗ red(S) ⊗ C∗ red(S), ∆). Cartesian product Γ2 = Γ × Γ of the group Γ is endowed with a natural group structure: (a, b) + (c, d) = (a + c, b + d). Obviously, semigroup S2 = S × S gen- erates group Γ. Let C∗ red(S2) be a C∗-algebra on l2(S2), generated by a regular representation of S2. Lemma 13. The following equations hold l2(S2) = l2(S) ⊗ l2(S), C∗ red(S2) = C∗ red(S) ⊗ C∗ red(S). Proof. Mapping e(a,b) → ea ⊗ eb extends to a linear isometric map U : l2(S2) → l2(S) ⊗ l2(S), taking the basis {e(a,b)}(a,b)∈S 2 of the space l2(S2) to the basis {e(a,b)}(a,b)∈S 2 in l2(S) ⊗ l2(S). This induces an operator T(a,b) → Ta ⊗ Tb, (a, b) ∈ S2. Obviously, T(a,b) = U ∗Ta ⊗ TbU. Therefore the map T(a,b) → Ta ⊗ Tb extends to an -isomorphism C∗ C∗ red(S2) → (cid:3) red)∗ = S∗ red × S∗ red, where (S2 red)∗ is an inverse semigroup generated red(S) ⊗ C∗ Note that (S2 red(S). by monomials in C∗ red(S2) (see section 2.2), and P (S2) = P (S) ⊙ P (S). We start a construction of quantum semigroup on C∗ red(S) with its subalgebra P (S). Define ∆ : P (S) → P (S) ⊗ P (S) by relation ∆( nXi=1 λiVi) = nXi=1 λiVi ⊗ Vi, where Vi is a monomial. Now our aim is to show that this comultiplication extends to an embedding red(S). To this end, divide the basis {ec ⊗ ed}c,d∈S of ∆ : C∗ the Hilbert space l2(S) ⊗ l2(S) to equivalence classes in the following way. red(S) → C∗ red(S) ⊗ C∗ We say that elements ec ⊗ ed and el ⊗ ek of the basis in l2(S) ⊗ l2(S) equivalent (ec ⊗ ed ∼ el ⊗ ek), if ec+a ⊗ ed+a = el+b ⊗ ek+b for some a, b ∈ S. One can easily see that ec⊗ed ∼ el⊗ek and el⊗ek ∼ em⊗en imply ec⊗ed ∼ em⊗en. Note that ec⊗ed ∼ SEMIGROUP OPERATOR ALGEBRAS AND QUANTUM SEMIGROUPS 13 el ⊗ ek if and only if d − c = k − l. Denote by Hc a Hilbert subspace in l2(S) ⊗ l2(S), Hc. generated by family {el ⊗ ekk − l = c, k, l ∈ S}. Obviously, l2(S) ⊗ l2(S) = Lc∈Γ Each of the spaces Hc is invariant under C∗-algebra ∆(P (S)). Indeed, let V be a monomial, el ⊗ ek ∈ Hc and (V ⊗ V )el ⊗ ek 6= 0. Due to Lemma 3, where c = indV . (V ⊗ V )(el ⊗ ek) = el+c ⊗ ek+c, Lemma 14. Operator ∆ : P (S) → P (S)⊗P (S) extends to an embedding ∆ : C∗ C∗ red(S) ⊗ C∗ red(S). red(S) → Proof. Fixing a ∈ S, Q ∈ P (S), we show the commutativity of the following dia- gram: Ha ρa−−−−→ l2(S) (3.1) ρa−−−−→ l2(S) where ρa is a projection onto the first component: Ha yPaQ nXi=1 λieai ⊗ ebi ) = λieai, ∆a(Q)=∆(Q)Hay nXi=1 ρa( nPi=1 Pa : l2(S) → ρa(Ha) ⊂ l2(S) is an orthogonal projection. Obviously, ρa is an isometric embedding of Ha into l2(S). It is sufficient to check 3.1 for basic elements ec ⊗ ed in Ha. Suppose Q = λiVi. Then ∆(Q)(ec ⊗ ed) = nXi=1 λ′ iec+ai ⊗ ed+ai, where ai = indVi and i =(cid:26) λ′ λi, if c + ai, d + ai ∈ S 0, if either c + ai /∈ S, or d + ai /∈ S. Consequently, On the other hand, ρa(∆)(Q)(ec ⊗ ed) = ((PaQ)ρa)(ec ⊗ ed) = Pa(Q(ec)) = Pa( λ′ iec+ai. nXi=1 nXi=1 λiec+ai) = nXi=1 λ′ iec+ai. And commutativity of diagram (3.1) is proved. Further, kρa∆a(Q)k = kPaQρak. Due to the fact that ρa is embedding and Pa is projection, we have k∆a(Q)k ≤ kQk for all a ∈ Γ. Note that the space H0 is generated by ea ⊗ ea for all a ∈ S. Operator ρ0 : H0 → l2(S) is a bijection and P0 is an identity. Therefore, k∆0(Q)k = kQk for all Q ∈ P (S). Thus, operator ∆ : P (S) → P (S) ⊗ P (S) extends to a continuous embedding ∆ : C∗ red(S) → C∗ red(S) ⊗ C∗ red(S). (cid:3) 14 AUKHADIEV MARAT, GRIGORYAN SUREN, AND LIPACHEVA EKATERINA Corollary 4. Suppose A ∈ C∗ red(S) and Aea = λieai. Then ∞Pi=1 ∆(A)(ea ⊗ ea) = ∞Xi=1 λi(eai ⊗ eai). Proof. By virtue of definition of ∆ on the subalgebra P (S) of algebra C∗ have red(S), we Further, ∆( λiVi) = λiVi ⊗ Vi, λiVi ∈ P (S). nXi=1 nXi=1 nXi=1 nXi=1 λiViea = nXi=1 λ′ ieai, where λ′ i = λi in case a + indVi ∈ S, and λ′ i = 0 otherwise. This implies that ( nXi=1 λiVi ⊗ Vi)(ea ⊗ ea) = nXi=1 λ′ ieai ⊗ eai. Since P (S) is dense in C∗ red(S), we get the required statement. (cid:3) Theorem 7. The pair (C∗ red(S), ∆) is a compact quantum semigroup. Proof. Let us show that multiplication induced by comultiplication ∆ by Lemma red(S))∗ is accosiative. Indeed, for any monomial V , we 14 on a Banach algebra (C∗ have the following (ξ × (η × ζ))(V ) = (ξ ⊗ (η × ζ))∆(V ) = = ξ(V ) · (η × ζ)(V ) = ξ(V ) · η(V ) · ζ(V ) = ((ξ × η) × ζ)(V ). (cid:3) 3.3. Weak Hopf algebra. In 1998 F.Li [Li98] defined a notion, naturally general- izing a Hopf algebra. This generalization is based on weakening an antipode condi- tion, and called a weak Hopf algebra. Let A be an algebra with a comuliplication ∆. Denote by m a three-argument multiplication map in algebra A, m(a ⊗ b ⊗ c) = abc, the same as a product of two elements. Suppose there exists a linear map T : A → A such that (3.2) (3.3) m(id ⊗ T ⊗ id)(∆ ⊗ id)∆ = id m(T ⊗ id ⊗ T )(∆ ⊗ id)∆ = T Then A is called a weak Hopf algebra, and T is a weak antipode. Note that weak Hopf algebra is not bialgebra, since a counit is not required. Obviously, T (1) = 1. Every Hopf algebra admits a unique antipode. However, this is not true for a weak Hopf algebra. Other weak Hopf algebra properties are shown in [Li99]. The analogous connection between a Hopf algebra and a group takes place here as well. Weak Hopf algebras are closely related to inverse semigroups. Consider an inverse semigroup M and semigroup algebra C[M ]. Algebra C[M ] admits comultiplication and weak antipode: ∆(x) = x ⊗ x, ∆(Xi λixi) =Xi λixi ⊗ xi, T (x) = x∗. SEMIGROUP OPERATOR ALGEBRAS AND QUANTUM SEMIGROUPS 15 Clearly, equations (3.2, 3.3) hold and C[M ] is a weak Hopf algebra, called an inverse semigroup bialgebra. Let A be a weak Hopf algebra with a weak antipode T . Then a set of group-like elements in A forms a regular semigroup, where inverse of any x is T (x). Thus, every inverse semigroup is associated with a classical weak Hopf algebra, and every weak Hopf algebra contains an inverse semigroup. Another example is algebra Mn(C), generated by matrix units Ei,j. Comultipli- cation and weak antipode are defined on generators: ∆(Ei,j ) = Ei,j ⊗ Ei,j , T (Ei,j) = Ej,i. Matrix units form an inverse semigroup. Therefore, this weak Hopf algebra turns out to be an inverse semigroup bialgebra. A.M. Vershik in [Ver07] proves result, which shows the connection between in- verse semigroup bialgebras and finite-dimensional bialgebras: Theorem 8 ( [Ver07]). Semigroup bialgebra of a finite inverse semigroup S with a unit, convolution multiplication and pointwise comultiplication is a semi-simple cocommutative involutive bialgebra. Conversely, every finite semi-simple cocommu- tative involutive bialgebra is isomorphic (as an involutive bialgebra) to a semigroup algebra of some finite inverse semigroup with unit. Recall that algebra C∗ red(S) can be regarded as a C∗-algebra generated by iso- metric operators Ta, T ∗ b for all a, b ∈ S. Since the regular representation is inverse, semigroup Σ generated by such operators is also inverse (see [AT12] for details). The inverse for Ta in this semigroup is T ∗ a . Thus, we obtain the following. Theorem 9. Compact quantum semigroup (C∗ Hopf algebra (C[Σ], ∆,∗ ). red(S), ∆) contains a dense weak This result motivates to search for a certain category of quantum semigroups which admits an analogue of the well-known fact that every compact quantum group contains a dense ∗-Hopf algebra [Wor98]. 3.4. Compact quantum subgroup. In this section we extend a notion of com- pact quantum subgroup from quantum groups [So l09] to quantum semigroups. We start with classical (commutative) case, the algebra of functions on a semigroup. Let P ′ ⊂ P be a compact subsemigroup in a compact semigroup P . Then there exists a natural ∗-epimorphism π : C(P ) → C(P ′), which is a restriction of functions from semigroup P to P ′. Denote corresponding comultiplications on this algebras by ∆ and ∆′. We get the following relation: (3.4) ∆′π = (π ⊗ π)∆. Now take compact quantum semigroups (A, ∆), (A′, ∆′). Following the classical case, we call (A′, ∆′) a compact quantum subsemigroup in (A, ∆), if there exists a ∗-epimorphism π : A → A′, which verifies (3.4). Actually, coassociativity of ∆′ follows from relation (3.4) and coassociativity of ∆. It is sufficient to require that ∆′ is a unital ∗-homomorphism. If moreover (A, ∆) is a compact quantum group, then (A′, ∆′) is a compact quantum subgroup. 16 AUKHADIEV MARAT, GRIGORYAN SUREN, AND LIPACHEVA EKATERINA Closed ideal J ⊂ A is called a coideal in a compact quantum semigroup (A, ∆) if ∆(J) ⊆ J ⊗ A + A ⊗ J. Theorem 10. Let J be a coideal in (A, ∆). There exists comultiplication ∆J , turning A/J into a compact quantum subsemigroup in (A, ∆) by means of quotient map π : A → A/J. Any compact quantum subsemigroup has this form. Proof. One can find the proof for the quantum group case in [Pin07]. Observing that the density conditions are not necessary in semigroup case, one can repeat the same here. (cid:3) An action of a compact quantum semigroup (A, ∆) on a unital C∗-algebra B is a unital -homomorphism δ : B → B ⊗ A such that (δ ⊗ id)δ = (id ⊗ ∆)δ. Action δ is called ergodic, if the fixed point algebra Bδ = {b ∈ Bδ(b) = b ⊗ I} coincides with C. Obviously, any compact quantum subsemigroup (A′, ∆′) in a compact quantum semigroup (A, ∆) naturally acts on the algebra A: δ = (id ⊗ π)∆. Then the fixed point algebra Aδ is called a quantum left coset space. Space Aδ with ∆δ A : Aδ → A ⊗ Aδ is a quantum left quotient space (A, ∆)/(A′, ∆′). Let us show that G is a compact (quantum) subgroup in (C∗ red(S), ∆). Further K denotes the commutator ideal in algebra C∗ red(S). Theorem 11. Ideal K is a coideal in C∗ (C∗ red(S), ∆) contains a classical compact quantum subgroup (C(G), ∆G). red(S). Compact quantum semigroup Proof. By Lemma 11, quotient map π : C∗ red(S)/K is a -epimorphism on algebra C(G). Here we have π(Ta) = χa, and χa is a character of group G. For any a ∈ S we have red(S) → C∗ (π ⊗ π)∆(Ta) = (π ⊗ π)(Ta ⊗ Ta) = χa ⊗ χa. On the other and, classical comultiplication on C(G) acts on χa in the following way: ∆G(χa)(α, β) = χa(αβ) = χa(α)χa(β) = (χa ⊗ χa)(α ⊗ β). Since chaacters generate algebra C(G), we obtain (π ⊗ π)∆ = ∆Gπ. It is sufficient to show that K is a coideal. To this end, consider monomials V, W . Then we have ∆(V W − W V ) = V W ⊗ V W − W V ⊗ W V = = (V W − W V ) ⊗ V W + W V ⊗ (V W − W V ) ∈ K ⊗ C∗ = (V W − W V ) ⊗ V W + W V ⊗ V W − W V ⊗ W V = red(S) + C∗ red(S) ⊗ K. Since K is a closed ideal and ∆ is a -homomorphism, the required injection holds for all elements. (cid:3) SEMIGROUP OPERATOR ALGEBRAS AND QUANTUM SEMIGROUPS 17 Thus, (C(G), ∆G) is a compact quantum subsemigroup in (C∗ red(S), ∆). For- red(S)) is endowed with an action of mally, G ⊂ QS(C∗ G. red(S)). Moreover, QS(C∗ Proposition 2. The natural action of compact quantum subgroup (C(G), ∆G) on the algebra C∗ red(S) is non-ergodic and coincides with representation τ : G → AutC∗ red(S), defined in section 2.3. The left coset space equals A0. Compact quan- tum quotient space (C∗ red(S), ∆)/(C(G), ∆G) is a compact (classical) quantum semi- group (A0, ∆A0 ). Proof. Take a ∈ S. We have τ (α)(Ta) = χa(α)Ta. This means that Therefore, τ (·)(Ta) ∈ C(G, C∗ red(S)) = C∗ red(S) ⊗ C(G). τ (·)(·) : C∗ red(S) → C∗ red(S) ⊗ C(G), herewith τ (·)(Ta) = Ta ⊗ χa. On the other hand, δ(Ta) = (id ⊗ π)(Ta ⊗ Ta) = Ta ⊗ χa. As noted above, τ (α)(V ) = V for V ∈ A0. Therefore the space A0 is a left coset space with respect to the action δ. (cid:3) 3.5. Dual algebra and Haar functional. In the previous section we endowed the dual space (C∗ red(S))∗ with an associative Banach algebra structure. By cocom- mutativity of ∆, this algebra is commutative. Futher recall that Haar functional is a state h ∈ A∗, subject to relation for all φ ∈ A∗ herewith λφ ∈ C. h × φ = φ × h = λφ · h Theorem 12. Haar functional in (C∗ red(S))∗ is defined by relation h(A) = (Ae0, e0) for all A ∈ C∗ red(S). Proof. One can easily verify that for any monomial V in C∗ is not equal to zero only in case V = I. Therefore, for any φ ∈ A∗ we have red(S), the scalar (V e0, e0) h × φ(V ) = h(V ) · φ(V ) =(cid:26) 0, V 6= I, φ(I), V = I It follows that h × φ = φ(I) · h. (cid:3) Equation C∗ red(S) = KL (C(G)) (see Theorem 6) implies red(S))∗ = K ⊥M (C(G))⊥. (C∗ Define mappings Φc : A → A, A 7→ T ∗ c : A∗ → A∗, (Φ∗ Φ∗ c ATc, c ∈ S c ξ)(A) = ξ(Φc(A)). Obviously, these maps are continuous. 18 AUKHADIEV MARAT, GRIGORYAN SUREN, AND LIPACHEVA EKATERINA Lemma 15. For any c ∈ S operator Φ∗ ((C(G)))⊥. c is identical on K ⊥, and Φ∗ c : ((C(G)))⊥ → Proof. Since K is ideal, Φc : K → K. Any monomial V in (C(G)) has a form T ∗ a Tb for some a, b ∈ S. Therefore, c T ∗ Consequently, ξ ∈ K ⊥ if and only if Φ∗ Φc(V ) = T ∗ a TbTc = T ∗ c (ξ) = ξ. a Tb = V. (cid:3) Obviously, ((C(G)))⊥ is an ideal in A∗ with respect to mulitplication ×. Denote by M (G) the space of regular Borel measures on group G, dual of Γ. Theorem 13. K ⊥ is a Banach algebra isomorphic to M (G). The following exact sequence splits. 0 → ((C(G)))⊥ → A∗ → K ⊥ = M (G) → 0 3.6. Category of quantum semigroups. Consider a category Sab of discrete abelian cancellative semigroups, with arrows (morphisms) being semigroup mor- phisms. Denote by QS red a category of compact quantum semigroups (C∗ red(S), ∆) constructed above, for all semigroups S ∈ Obj(Sab), herewith unital compact quantum semigroup morphisms. Recall that a *-homomorphism π : C∗ red(S1) → C∗ red(S2) is called a unital morphism of the corresponding compact quantum semi- groups, if it is unital and satisfies following diagram. C∗ red(S1) π−−−−→ C∗ red(S2) y∆2 (3.5) ∆1y C∗ red(S1) ⊗ C∗ red(S1) π⊗π−−−−→ C∗ red(S2) ⊗ C∗ red(S2) Thus, morphisms in category QS red are compact quantum semigroup morphisms. Lemma 16. Suppose that A is an isometry in C∗ Then A = Tc, where c ∈ S. red(S), such that ∆(A) = A ⊗ A. Proof. Let us show that the set {ea}a∈S is A-invariant. , Aea = ∞Xi=1 λieai. Then we have (A ⊗ A)(ea ⊗ ea) = Aea ⊗ Aea = ∞Xi,j=1 λiλj eai ⊗ eaj . On the other hand, by Corollary 4, ∆(A)(ea ⊗ ea) = ∞Xi=1 λi(eai ⊗ eai). The family {ea ⊗ eb}a,b∈S forms an orthonormal basis in l2(S) ⊗ l2(S). Therefore, λiλj = 0 for i 6= j. Consequently, all λj = 0 except for one. further, Aea = λeb. By the same reasoning, we obtain λ = λ2. It follows that λ = 1. Assume that Aea = eb with b = c − a. Let us verify Aed = ed+c for any d ∈ S. Suppose that Aed = ed+k, k 6= c. Element ea ⊗ ed lies in Hilbert space Hl, where SEMIGROUP OPERATOR ALGEBRAS AND QUANTUM SEMIGROUPS 19 l = d − a, i.e. the space generated by family {eai ⊗ ebi }∞ Since ∆(A)(ea ⊗ ed) lies in Hl, we obtain i=1 such that bi − ai = l. ∆(A)(ea ⊗ ed) = ea+c ⊗ ed+k, with (d + k) − (a + c) = l. Consequently, k = c. Since A is an isometry, c ∈ S and A = Tc. (cid:3) Theorem 14. There exists an injective functor which maps QS red to the category Sab. Proof. By commutativity of (3.5), we have ∆2(π(Ta)) = π(Ta) ⊗ π(Ta). Operators Ta and π(Ta) are isometric. By Lemma 16, π(Ta) = Ta′ for some a′ ∈ S. Thus, π generates a homomorphism φ : S1 → S2, a 7→ a′. (cid:3) An automorphism σ of algebra C∗ red(S) is called a compact quantum semigroup red(S), ∆), if (σ ⊗ σ)∆ = ∆σ. red(S), ∆)) the automorphism group of the compact quantum automorphism of (C∗ Denote by Aut((C∗ semigroup (C∗ red(S), ∆). Corollary 5. Aut((C∗ red(S), ∆)) is isomorphic to automorphism group Aut(S). Note that for non-negative real numbers R+ and non-negative rational num- bers Q+ semigroups every morphism is an automorphism, and Aut(R+) = R, Aut(Q+) = Q. For non-negative integers Z+ we have that group Aut(Z+) is trivial, and the morphism semigroup matches Z+. 3.7. Different comultiplications on the Toeplitz algebra. The structure of a compact quantum semigroup on the Toeplitz algebra T is given in [AGL11], denoted by (T , ∆). One can ask if there exists another comultiplication on T . The Toeplitz algebra is generated by the regular isometric representation of semigroup Z+, so T = C∗ red(Z+) constructed in section 3.2 matches the comultiplication on the Toeplitz algebra. However, Z+ is not unique as an abelian semigroup which regular representation generates algebra T . red(Z+). Obviously, the comultiplication on C∗ Consider semigroup S = {0, 2, 3, 4, ...}. Obviously, S generates group Z: Then C∗ mials represents a right-shift operator on l2(S): red(S) is the Toeplitz agebra. Indeed, the following combination of mono- Z = S − S. T = (I − T ∗ 3 T2T ∗ 2 T3)T2 + T ∗ 2 T3. Futher, (C∗ 3.2. But this comultiplication ∆S differs from comultiplication ∆: red(S), ∆S) is a compact quantum semigroup with ∆S defined in Section ∆(T ) = T ⊗ T, 3 T2T ∗ ∆S(T ) = T2 ⊗ T2 − T ∗ 2 T3T2 ⊗ T ∗ Thus, the compact quantum semigroups (C∗ 3 T2T ∗ red(S), ∆S) and (T , ∆) do not match. And the results of this paper show, that generally, the Toeplitz algebra can be given at least as many structures of compact quantum semigroups, as many semigroups generating Z exist. 2 T3T2 + T ∗ 2 T3 ⊗ T ∗ 2 T3. Lemma 17. A morphism taking (C∗ red(S), ∆S) to (T , ∆) does not exist. 20 AUKHADIEV MARAT, GRIGORYAN SUREN, AND LIPACHEVA EKATERINA Proof. Suppose, π is a morphism of the corresponding compact quantum semi- groups. Then, by theorem 14, there exists a morphism φ : S → Z+. Take m = φ(2), k = φ(3). We have m + m + m = φ(2 + 2 + 2) = φ(3 + 3) = k + k. Therefore, m is even and k > m. Consequently, π(T ∗ 2 T3) is an isometric operator, which is not true for T ∗ 2 T3) = 0, and the commutator ideal (the idea of compact operators) is a kernel of homomorphism π. (cid:3) 2 T3. This means that π(I − T ∗ 3 T2T ∗ This research was partially supported by RFBR grant no 12-01-97016. References [AGL11] M. A. Aukhadiev, S. A. Grigoryan, and E. V. Lipacheva, Infinite-dimensional compact quantum semigroup, Lobachevskii J. Math. 32 (2011), no. 4, 304 -- 316. [A.M98] A.Van Daele A.Maes, Notes on Compact Quantum Groups, Nieuw Arch. Wisk. 4 (1998), no. 16, 73 -- 112. [AT12] M. A. Aukhadiev and V. H. Tepoyan, Isometric representations of totally ordered semi- groups, Lobachevskii J. Math. 33 (2012), no. 3, 239 -- 243. [Cob67] L. A. Coburn, The C ∗-algebra generated by an isometry, Bull. Amer. Math. Soc. 73 (1967), 722 -- 726. [Dou72] R. G. Douglas, On the C ∗-algebra of a one-parameter semigroup of isometries, Acta Math. 128 (1972), no. 3-4, 143 -- 151. [Dri86] V.G. Drinfeld, Quantum Groups, Proceedings ICM Berkeley (1986), 798 -- 820. [GS09] S. A. Grigoryan and A. F. Salakhutdinov, C ∗-algebras generated by semigroups, Izv. Vyssh. Uchebn. Zaved. Mat. (2009), no. 10, 68 -- 71. [Kaw08] Katsunori Kawamura, C ∗-bialgebra defined by the direct sum of Cuntz algebras, J. [Li98] [Li99] Algebra 319 (2008), no. 9, 3935 -- 3959. Fang Li, Weak Hopf algebras and some new solutions of the quantum Yang-Baxter equation, J. Algebra 208 (1998), no. 1, 72 -- 100. Fang Li, Weak Hopf algebras and regular monoids, J. Math. Res. Exposition 19 (1999), no. 2, 325 -- 331. [MT04] G. J. Murphy and L. Tuset, Aspects of compact quantum group theory, Proc. Amer. Math. Soc. 132 (2004), no. 10, 3055 -- 3067 (electronic). [Mur87] G. J. Murphy, Ordered groups and Toeplitz algebras, J. Operator Theory 18 (1987), no. 2, 303 -- 326. [Pin07] Claudia Pinzari, Embedding ergodic actions of compact quantum groups on C ∗-algebras into quotient spaces, Internat. J. Math. 18 (2007), no. 2, 137 -- 164. [Pod95] Piotr Podle´s, Symmetries of quantum spaces. Subgroups and quotient spaces of quantum SU(2) and SO(3) groups, Comm. Math. Phys. 170 (1995), no. 1, 1 -- 20. [RTF90] N.Yu. Reshetikhin, L.A. Takhtadzhyan, and L.D. Faddeev, Quantization of Lie groups and Lie algebras., Leningr. Math. J. 1 (1990), no. 1, 193 -- 225 (English. Russian original). Piotr Miko laj So ltan, On quantum semigroup actions on finite quantum spaces, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 12 (2009), no. 3, 503 -- 509. [So l09] [Ver07] A.M. Vershik, Krein duality, positive 2-algebras, and the dilation of comultiplications., Funct. Anal. Appl. 41 (2007), no. 2, 99 -- 114 (English. Russian original). [VVD01] Stefaan Vaes and Alfons Van Daele, Hopf C ∗-algebras, Proc. London Math. Soc. (3) 82 (2001), no. 2, 337 -- 384. [Wor87a] S. L. Woronowicz, Compact matrix pseudogroups, Comm. Math. Phys. 111 (1987), no. 4, 613 -- 665. [Wor87b] S.L. Woronowicz, Twisted SU(2)-group: an example of non-commutative differential calculus, Publ.RIMS (1987), no. 23, 171 -- 181. [Wor91] S.L. Woronowicz, Unbounded elements affiliated with C ∗-algebras and non-compact quantum groups, Commun. Math. Phys. (1991), no. 136, 399 -- 432. SEMIGROUP OPERATOR ALGEBRAS AND QUANTUM SEMIGROUPS 21 [Wor98] S.L. Woronowicz, Compact quantum groups, Symetries quantiques, North-Holland, Am- sterdam (1998), 845 -- 884. E-mail address, Aukhadiev Marat: [email protected] E-mail address, Grigoryan Suren: [email protected] E-mail address, Lipacheva Ekaterina: [email protected] (Aukhadiev M.A., Grigoryan S.A. and Lipacheva E.A.) 420066, Russia, Kazan, Krasnoselskaya str., 51 Kazan State Power Engineering University
1807.07381	2	1807	2019-09-17T15:49:43	Subalgebras of simple AF-algebras	[ "math.OA" ]	It is shown that if A is a separable, exact C-algebra which satisfies the Universal Coefficient Theorem (UCT) and has a faithful, amenable trace, then A admits a trace-preserving embedding into a simple, unital AF-algebra with unique trace. Modulo the UCT, this provides an abstract characterization of C-subalgebras of simple, unital AF-algebras. As a consequence, for a countable, discrete, amenable group G acting on a second countable, locally compact, Hausdorff space X, C_0(X) \rtimes_r G embeds into a simple, unital AF-algebra if, and only if, X admits a faithful, invariant, Borel, probability measure. Also, for any countable, discrete, amenable group G, the reduced group C-algebra C_r(G) admits a trace-preserving embedding into the universal UHF-algebra.	math.OA	math	SUBALGEBRAS OF SIMPLE AF-ALGEBRAS CHRISTOPHER SCHAFHAUSER Abstract. It is shown that if A is a separable, exact C∗-algebra which satisfies the Universal Coefficient Theorem (UCT) and has a faithful, amenable trace, then A admits a trace-preserving embedding into a simple, unital AF-algebra with a unique trace. Modulo the UCT, this provides an abstract characterization of C∗-subalgebras of simple, unital AF-algebras. As a consequence, for a countable, discrete, amenable group G acting on a second countable, locally compact, Hausdorff space X, C0(X) ⋊r G embeds into a simple, unital AF-algebra if, and only if, X admits a faithful, invariant, Borel, probability measure. Also, for any countable, discrete, amenable group G, the reduced group C∗-algebra C∗ r (G) admits a trace-preserving embedding into the universal UHF-algebra. Introduction It follows from the work of Murray and von Neumann in [57] that there is a unique sep- arably acting, hyperfinite II1-factor R and that any separably acting, hyperfinite, tracial von Neumann algebra admits a trace-preserving embedding into R. A fundamental result of Connes in [12] characterizes hyperfinite von Neumann algebras abstractly by showing the equivalence of hyperfiniteness and injectivity among von Neumann algebras. Together, these results show that a separably acting, finite von Neumann algebra embeds into R if, and only if, it is hyperfinite, a result which has an instrumental role in the theory of subfactors initiated by Jones in [40]. A C∗-algebraic analogue of the Murray-von Neumann classification theorem is given by Elliott's celebrated result in [26] classifying approximately finite-dimensional (AF) C∗- algebras and ∗-homomorphisms between such algebras in terms of the non-stable K0-group. The problem of finding abstract characterizations of AF-algebras and C∗-subalgebras of AF-algebras was posed by Effros in [24] with the latter problem motivated in part by the AF-embedding of the irrational rotation algebras of Pimsner and Voiculescu in [64] which led to the classification of such C∗-algebras in terms of the angle of rotation. Although an abstract characterization of AF-algebras among all C∗-algebras seems out of reach, such a characterization is possible among simple, unital C∗-algebras due to the remarkable success of Elliott's classification programme for simple, nuclear C∗-algebras over the last several decades with the final steps taken in [35], [29], and [82]. The problem of characterizing C∗-subalgebras of AF-algebras has received much atten- tion over the last several decades with the standing conjecture being that a C∗-algebra Date: September 18, 2019. 2010 Mathematics Subject Classification. Primary: 46L05. Key words and phrases. AF-Embedding, Amenable trace. 1 2 CHRISTOPHER SCHAFHAUSER embeds into an AF-algebra if, and only if, it is separable, exact, and quasidiagonal (see Sec- tion 7 of [4]). Shortly after the Pimsner-Voiculescu AF-embedding result of [64], Pimsner characterized in [63] which C∗-algebras of the form C(X) ⋊ Z can be embedded into an AF-algebra in terms of the underlying action of Z on X. Similar results were obtained by Brown for crossed products of AF-algebras by Z in [7] and for crossed products of UHF-algebras by Zk in [8]. Many other partial results along these lines have appeared in [41], [51], and [53] for example. The latter result in [53] also has an important role in the work of Ozawa, Rørdam, and Sato in [62] showing that the C∗-algebra of an elemen- tary amenable group embeds into an AF-algebra, a result which was later extended to all countable, discrete, amenable groups in [82]. Aside from crossed products, it is known that all residually stably finite, type I C∗- algebras embed into AF-algebras [81], separable, exact, residually finite-dimensional C∗- algebras satisfying the UCT embed into AF-algebras [18], and the cone over any separable, exact C∗-algebra embeds into an AF-algebra [60]. Also, combining Ozawa's result in [60] with the techniques introduced by Spielberg in [81], Dadarlat has obtained AF-embeddings of continuous fields of C∗-algebras in [19] provided the base space is sufficiently connected and at least one fibre is AF-embeddable. See Chapter 8 of [10] for a well-written survey of the AF-embedding problem for C∗-algebras. Despite the remarkable work on the AF-embedding problem given in the results above, in each of these results, the methods used are very specific to the class of C∗-algebras under consideration and shed very little light on the abstract AF-embedding problem. This paper introduces a systematic method for producing embeddings into certain simple, unital AF-algebras. It is well known that any C∗-subalgebra of a simple, unital AF-algebra must be separable and exact and must admit a faithful, amenable trace. Modulo the Universal Coefficient Theorem (UCT) of [73], the present paper shows these are the only obstructions. There are no known counterexamples to the UCT among separable, exact C∗-algebras with a faithful, amenable trace, and it is an important open problem whether all separable, nuclear C∗-algebras satisfy the UCT. Theorem A. If A is a separable, exact C ∗-algebra which satisfies the UCT and admits a faithful, amenable trace, then there is a simple, unital AF-algebra B with a unique trace and a trace-preserving embedding A ֒→ B. Amenable traces were introduced by Connes in [12] in the von Neumann algebra setting and are characterized by the existence of almost multiplicative, almost trace-preserving, completely positive, contractive maps into Mn, the algebra of n × n matrices over C, where "almost" is measured in the 2-norm defined by the normalized trace on Mn. A key step in Connes's proof that injective II1-factors are hyperfinite consists of showing that the trace on an injective II1-factor is amenable. In fact, the amenability of the trace characterizes injectivity of II1-factors. Amenable traces were introduced in the C∗-algebraic setting by Kirchberg in [43] and extensively developed by Brown in [9]. There is a very close connection between the theory of amenable traces on C∗-algebras and von Neumann algebras. For example, if A is an exact C∗-algebra, a trace τA on A is amenable if, and only if, πτA(A)′′ is hyperfinite. Furthermore, if τA is faithful, then the GNS representation πτA : A → πτ (A)′′ is faithful, and hence any exact C∗-algebra A SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 3 with a faithful, amenable trace admits a trace-preserving embedding into a hyperfinite von Neumann algebra. This observation, which can be viewed as a weak∗-version of Theorem A, is the starting point of the proof of the quasidiagonality theorem of Tikuisis, White, and Winter in [82] and Gabe in [33] which, in turn, has an important role in the proof of Theorem A. Building on the work of Ozawa, Rørdam, and Sato in [62], it was shown in [82] that if G is a countable, discrete, amenable group, then C∗ r(G) embeds into an AF-algebra, although these results give no control over the codomain AF-algebra. Combining the techniques introduced in the present paper with the work of Higson and Kasparov [39], Luck [52], and Tu [85] on the Baum-Connes conjecture yields a much sharper AF-embedding result showing group C∗-algebras embed into the universal UHF-algebra Q = N∞ For a discrete group G, the group von Neumann algebra L(G) and the reduced group n=1 Mn. C∗-algebra C∗ where e ∈ G is the neutral element. r(G) are equipped with the usual faithful trace given by Pg∈G cg · g 7→ ce Theorem B. For a countable, discrete group G, the following are equivalent: (1) G is amenable; (2) L(G) admits a trace-preserving embedding into R; (3) C∗ r(G) admits a trace-preserving embedding into Q. In fact, the methods introduced here also yield an AF-embedding result for crossed products of abelian C∗-algebras by amenable groups in the spirit of Pimsner's result for crossed products C(X) ⋊ Z in [63] and Lin's result for crossed products C(X) ⋊ Zk in [51]. Theorem C. If X is a second countable, locally compact, Hausdorff space and G is a countable, discrete, amenable group acting on X, then C0(X) ⋊ G embeds into a simple, unital AF-algebra if, and only if, X admits a faithful, G-invariant, Borel, probability measure. The new technical tool facilitating these results is a classification theorem for faithful ∗-homomorphisms into certain AF-algebras. The precise statement of Theorem D is given in Corollary 5.4. Theorems A, B, and C will be deduced from Theorem D at the beginning of Section 6. Theorem D. If A is a separable, unital, exact C ∗-algebra satisfying the UCT with a faithful, amenable trace and B is a simple, unital AF-algebra with a unique trace and divisible K0-group, then the unital, trace-preserving embeddings A → B are classified up to approximate unitary equivalence by their behaviour on the K0-group. A more general classification result is given in Theorem 5.3 which does not require any inductive limit structure or nuclearity assumption on the codomain C∗-algebra. Along with the applications to AF-embeddability listed above, this technical refinement of Theorem D also leads to a self-contained proof of an abstract characterization of AF-algebras among simple, unital C∗-algebras with a unique trace and divisible K0-group in Corollary 6.7. These results also lead to new examples of MF algebras arising as reduced crossed products by free groups (see Corollary 6.5) building on [42, 66, 67, 79] and give AF-embedding results for k-graph algebras extending those of [77] and [11]. 4 CHRISTOPHER SCHAFHAUSER Special cases of Theorem D have appeared in several places in the literature. When A is an AF-algebra or an AT-algebra of real rank zero, Theorem D is a special case of classical results of Elliott in [26] and [28], respectively. For a commutative or, more generally, approximately homogeneous (AH) C∗-algebra A, classification results for embeddings from A into an AF-algebra form a crucial part of the AF-embedding results for crossed products obtained by Lin in [51]. Also, the classification results for simple, nuclear C∗-algebras with tracial approximation structure, initiated by Lin in [50] and culminating in the remarkable work of Gong, Lin, and Niu in [35], depends heavily on classification results for embeddings between such algebras. The power of Theorem D and what distinguishes Theorem D from existing classification results is the general hypothesis on the domain C∗-algebra. The proof is motivated by the classical classification results for tracially AF-algebras in [17] and [48], but it avoids making any use of the internal structure of the domain C∗-algebra and makes no use of inductive limit models such as the AH-algebras used in [48]. The proof of Theorem D, and the more general Theorem 5.3 below, serve as proof of concept for an abstract approach to Elliott's classification programme which does not depend on tracial approximations. A more general result of this form will appear in forthcoming work of the author with Carri´on, Gabe, Tikuisis, and White. On the proof of Theorem D. The proof of Theorem D is heavily motivated by the classification theorem for separable, simple, unital, nuclear, tracially AF-algebras satisfy- ing the UCT due to Lin in [50] and the classification of ∗-homomorphisms between such algebras due to Dadarlat in [17], and, on a non-technical level, there are very close analo- gies between the proofs of these results. The class of tracially AF-algebras was introduced by Lin in [48] motivated in part by an approximation condition proved by Popa in [65] for simple, unital, quasidiagonal C∗-algebras of real rank zero. Roughly, a tracially AF- algebra is a C∗-algebra B which admits the following approximation condition in the spirit of Egoroff's Theorem: there is an approximately central projection in B with large trace such that the corner pBp is locally approximated by a finite-dimensional C∗-subalgebra where the approximations are in operator norm. The basic strategy for the classification results in [17] and [50] is to use the tracial finite- dimensional approximations for A to produce approximately multiplicative maps from A into a suitably well-behaved AH-algebra which approximately preserve the tracial data. From here, one then perturbs these maps on tracially small corners of A to adjust the behaviour of the maps on the infinitesimal elements of the K-theory groups. Then, the structure of AH-algebras allows one to construct an embedding of the given AH-algebra into B which implements an isomorphism on the invariant so that, upon composing this embedding with the approximately multiplicative maps into the AH-algebra, one obtains approximately multiplicative maps A → B with prescribed behaviour on K-theory and traces. Together with a uniqueness result for approximately multiplicative maps from A to B relying on a Weyl-von Neumann type theorem for Hilbert module representations as in [49] or [21] and using the tracial approximation structure of B, an intertwining argument allows one to construct ∗-homomorphisms from A to B with prescribed behaviour on K-theory and traces. SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 5 Changing perspective slightly, the almost multiplicative, almost trace-preserving maps A → B can be encoded as a trace-preserving ∗-homomorphism ψ : A → Bω where Bω denotes the norm ultrapower of B. Adjusting the behaviour of these approximations on the tracially small corner of A can be viewed as classifying, up to unitary equivalence, all ∗-homomorphisms ϕ : A → Bω which are trace-zero perturbations of ψ. It should be noted that this formalism is also explicitly used in the classical tracially AF-algebra classification results (usually with sequence algebras in place of ultrapowers, but this is mostly a matter of taste). In the abstract setting of Theorem D, the existence of ψ will follow directly from the quasidiagonality theorem of [82] and [33] (see Theorem 1.2 below). The new step is an abstract method for describing the trace-zero perturbations of ψ up to unitary equivalence. Consider the trace-kernel extension 0 JB jB Bω qB Bω 0 associated to B where Bω is the 2-norm ultrapower of B and JB = {b ∈ Bω : τBω (b∗b) = 0}. Roughly, Bω and JB will play the roles of the tracially large and small corners of B, respectively. When B has a unique trace and no non-zero, finite-dimensional representations, Bω is a II1-factor, and hence all trace-preserving ∗-homomorphisms A → Bω factor through the von Neumann algebra πτA(A)′′ associated to A and τA. When A is exact and τA is amenable, πτA(A)′′ is a hyperfinite von Neumann algebra, and hence the classifica- tion of normal ∗-homomorphisms from hyperfinite von Neumann algebras to II1-factors yields a classification result for trace-preserving ∗-homomorphisms A → Bω up to unitary equivalence. It is through exploiting this observation that both the tracial approximation assumptions on A and factoring through a model AH-algebra are avoided. Having classification of trace-preserving ∗-homomorphisms A → Bω ∼= Bω/JB, the goal becomes to lift this classification along qB to obtain classification of ∗-homomorphisms A → Bω up to unitary equivalence. Using extension theoretic methods, it is shown in [78] that (at least when B ∼= Q), modulo a certain KK-obstruction, a faithful, nuclear ∗-homomorphism A → Bω can be lifted to a nuclear ∗-homomorphism A → Bω. The strategy behind Theorem D is to use extension theoretic methods to show that the nuclear ∗-homomorphisms A → Bω lifting a given faithful, nuclear ∗-homomorphism A → Bω are parametrized up to unitary equivalence by the Kasparov group KKnuc(A, JB).1 Now, using the UCT and the K-theoretic assumptions on B, the group KKnuc(A, JB) is computed in terms of the K0-groups of A and Bω. Together with an intertwining argument given in Section 5, this will prove Theorem D. In order to illustrate the lifting process suggested above, let us consider the uniqueness result (Proposition 4.3) in more detail; the corresponding existence result (Proposition 4.2) has a similar flavour. Let A and B be as in Theorem D, and suppose ϕ, ψ : A → Bω are unital, full, nuclear ∗-homomorphisms with τBω ϕ = τBω ψ and K0(ϕ) = K0(ψ). 1The C∗-algebra JB is not σ-unital. In the actual proof, all computations will be done in a sufficiently large, separable C∗-subalgebra of JB. 6 CHRISTOPHER SCHAFHAUSER Since ϕ and ψ agree on the trace on Bω, qBϕ, qBψ : A → Bω agree on the trace on Bω. As Bω is a II1-factor and ϕ and ψ are nuclear, qBϕ and qBψ are unitarily equivalent (see Proposition 1.1 below). Let u ∈ Bω denote a unitary conjugating qBψ to qBϕ. If F ∈ Bω with qB(F ) = u, then F ∗F − 1Bω , F F ∗ − 1Bω , ϕ(a) − F ψ(a)F ∗ ∈ JB for all a ∈ A.2 Hence viewing Bω as a C∗-subalgebra of the multiplier algebra M (JB) of JB, the triple (ϕ, ψ, F ) defines an element in KKnuc(A, JB). Under the hypotheses of Theorem D, since K0(ϕ) = K0(ψ), this KK-class vanishes, so there are a ∗-homomorphism π : A → M (JB ⊗ K) and a unitary V ∈ M (JB ⊗ K) such that V − F ⊕ 1M (JB ⊗K) ∈ JB ⊗ K and ϕ ⊕ π = ad(V )(ψ ⊕ π). In Section 3, following the techniques introduced in [78] for the case B = Q, it is shown that, modulo separability issues, JB is stable and has the corona factorization property. From here, the fullness of ϕ and ψ together with a Weyl-von Neumann type absorption theorem due to Elliott and Kucerovsky in [30] (Theorem 2.3 below) is used to remove the summand π and show there is a unitary U ∈ M (JB) with U − F ∈ JB and ϕ = ad(U )ψ. But now, as F ∈ Bω and U − F ∈ JB ⊆ Bω, we have U ∈ Bω. Since U conjugates ψ to ϕ by construction, this shows the uniqueness result. The paper is organized as follows. In Section 1, some preliminary results on amenable traces are collected along with some general machinery for reducing problems to separable C∗-algebras. Section 2 records some KK-theoretic prerequisites and obtains the non- stable KK-theoretic results needed in Section 4. Section 3 is devoted to the trace-kernel extension and proves certain extension-theoretic regularity conditions for the trace-kernel ideal. The main classification results are given in Sections 4 and 5; the former section proves classification results for embeddings into ultrapower C∗-algebras, and the latter section restates these results in terms of approximate morphisms and, from here, obtains Theorem D via an intertwining argument. Finally, Theorems A, B, and C along with some other consequences of Theorem D are proved in Section 6. Acknowledgements. I would like to thank Jos´e Carri´on, Jamie Gabe, Aaron Tikuisis, and Stuart White for several helpful conversations regarding this work and for their com- ments on early drafts of this paper. I am also grateful to Matt Kennedy for pointing out the Choquet theoretic result used in the proof of Corollary 6.3. Finally, I thank the anonymous referees for several helpful comments on this paper. 1. Preliminaries 1.1. Amenable Traces. Throughout, the word trace is reserved for a tracial state on a C∗-algebra. Given a C∗-algebra A and a trace τA on A, let τA : K0(A) → R denote the induced state on K0(A). In the case A has a unit, τA([p]) = (τ ⊗ TrK)(p) for a projection 2In fact, as the unitary group of B ω is path-connected, u lifts to a unitary. For technical reasons, it will be helpful to take a unitary lift of u and replace ψ with a unitary conjugate of ψ to arrange F = 1Bω . SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 7 p ∈ A ⊗ K where TrK is the usual tracial weight on the C∗-algebra K of compact operators on a separable, infinite-dimensional Hilbert space. For n ≥ 1, let τMn denote the unique trace on the C∗-algebra Mn of n × n matrices over C and define the 2-norm k · k2 on Mn by kak2 = τMn(a∗a)1/2 for all a ∈ Mn. A trace τA on a C∗-algebra A is called amenable if there is a net ϕi : A → Mn(i) of completely positive, contractive maps with kϕi(aa′) − ϕi(a)ϕi(a′)k2 → 0 and τMn(i)(ϕi(a)) → τA(a) for all a, a′ ∈ A. See [9] or Chapter 6 of [10] for a detailed treatment of amenable traces. Note that by Theorem 4.2.1 of [9], all traces on nuclear C∗-algebras are amenable. Exploiting the connections between amenable traces and hyperfinite von Neumann al- gebras given in Theorem 3.2.2 of [9] leads to the following uniqueness result which is well known in the case when A is nuclear. Proposition 1.1. If A is a C ∗-algebra, M is a finite factor, and ϕ, ψ : A → M are weakly nuclear ∗-homomorphisms such that τMϕ = τMψ, then there is a net of unitaries (ui) in M such that kϕ(a) − uiψ(a)u∗ i k2 → 0 for all a ∈ A. Proof. Let τA = τMϕ and note that ϕ and ψ induce normal ∗-homomorphisms ¯ϕ, ¯ψ : πτA(A)′′ → M such that ¯ϕ(πτA(a)) = ϕ(a) and ¯ψ(πτA(a)) = ψ(a) for all a ∈ A. As ¯ϕ is faithful and normal, Lemma 1.5.11 of [10] implies that there is a normal, completely positive map θ : M → πτA(A)′′ such that πτ (a) = θ(ϕ(a)) for all a ∈ A. Since ϕ is weakly nuclear, πτ (A)′′ is hyperfinite by the equivalence of (5) and (6) in Theorem 3.2.2 of [9]. The result follows from the classification of normal ∗-homomorphisms from hyperfinite von Neumann algebras into finite factors; see the proof of Proposition 2.1 in [13] for example. (cid:3) The following fundamental result was proved for nuclear C∗-algebras by Tikuisis, White, and Winter in [82] and was extended to exact C∗-algebras by Gabe in [33] (see also [78] for a short proof). This result is the starting point for the existence result in Theorem D. Let Q = Nn≥1 Mn denote the universal UHF-algebra and let Qω denote the norm ultrapower of Q with respect to a fixed free ultrafilter ω on the natural numbers. Recall that a ∗-homomorphism ϕ : A → B between C∗-algebras A and B is full if for every non-zero a ∈ A, ϕ(a) generates B as an ideal. Theorem 1.2. If A is a separable, unital, exact C ∗-algebra satisfying the UCT and τA is a faithful, amenable trace on A, then there is a unital, full, nuclear ∗-homomorphism ϕ : A → Qω such that τQω ϕ = τA. The result is not quite stated this way in the references given above. The existence of a unital, nuclear, trace-preserving ∗-homomorphism A → Qω follows from Theorem 3.8 and Proposition 3.4(ii) in [33], and this ∗-homomorphism is necessarily full by Lemma 2.2 in [82] and the faithfulness of τA. 8 CHRISTOPHER SCHAFHAUSER 1.2. Separability Issues. Throughout the paper, several non-separable C∗-algebras such as ultraproducts and their trace-kernel ideals (as defined in Section 3) will be considered. The lack of separability causes technical issues in certain arguments; this is especially the case with KK-theoretic considerations where all C∗-algebras are typically required to be separable or, at the very least, σ-unital. This section collects some general methods for reducing problems to the separable setting. Definition 1.3 (Blackadar, Section II.8.5 of [3]). A property (P ) of C∗-algebras is called separably inheritable if (1) whenever A is a C∗-algebra satisfying (P ) and A0 is a separable C∗-subalgebra of A, there is a separable C∗-subalgebra A of A which satisfies (P ) and contains A0, and (2) whenever A1 ֒→ A2 ֒→ A3 ֒→ · · · is an inductive system of separable C∗-algebras An satisfies (P ). with injective connecting maps, if each An satisfies (P ), then lim −→ Many important properties of C∗-algebras are separably inheritable such as exactness, nuclearity, simplicity, real rank zero, and stable rank one, to name a few. Also, the meet of countably many separably inheritable properties is separably inheritable. See Section II.8.5 of [3] for proofs of these facts and for many more examples of separably inheritable properties. The following slight variation of separable inheritability will be useful. Definition 1.4. Let (P ) be a property of separable C∗-algebras. A C∗-algebra A separably satisfies (P ) if whenever A0 is a separable C∗-subalgebra of A, there is a separable C∗- subalgebra A of A which satisfies (P ) and contains A0. Note that if (P ) is a separably inheritable property of C∗-algebras and A is a C∗- algebra satisfying (P ), then A separably satisfies (P ). Note also that if (P ) is a property of separable C∗-algebras preserved under sequential inductive limits with injective connecting maps, then separably (P ) is a separably inheritable property. The following is analogous to II.8.5.3 in [3], and the same proof holds here. Proposition 1.5. Let (Pi) be a countable family of properties of separable C ∗-algebras preserved under sequential inductive limits with injective connecting maps. If A is a C ∗- algebra separably satisfying (Pi) for each i, then A separably satisfies the meet of the (Pi). The following result will be crucial for KK-theoretic considerations related to the trace- kernel extension defined in Section 3. Proposition 1.6. Consider an extension 0 j I E q D 0 of C ∗-algebras and suppose for each X ∈ {I, E, D}, (PX ) is a property of separable C ∗- algebras preserved under sequential inductive limits with injective connecting maps and X separably satisfies (PX ). If for each X ∈ {I, E, D}, a separable C ∗-subalgebra X0 of X is given, then for each X ∈ {I, E, D}, there is a separable C ∗-subalgebra X of X which SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 9 satisfies (PX ) and contains X0 and such that there is a homomorphism 0 0 I I j E E D D q 0 0 of extensions where the vertical arrows are the inclusion maps. Proof. For X ∈ {I, E, D}, we construct an increasing sequence of separable C∗-subalgebras (Xn)∞ n=1 of X containing X0 such that for each n ≥ 1, Xn satisfies (PX ), and In−1 ⊆ j−1(En) ⊆ In. q(En−1) ⊆ Dn ⊆ q(En), Assuming this has been done, for each X ∈ {I, E, D}, let X be the closed union of the Xn and note that X is a separable C∗-subalgebra of X. As each Xn satisfies (PX ), so does X by hypothesis. By construction, q( E) = D, and j−1( E) = I, so the result follows. We construct the desired C∗-subalgebras In, En, and Dn inductively starting from the given C∗-subalgebras I0, D0, and E0. Assume n ≥ 1 and In−1, Dn−1, and En−1 have been constructed. The C∗-subalgebra of D generated by Dn−1 and q(En−1) is separable, and hence there is a separable C∗-subalgebra Dn of D which satisfies (PD) and contains both Dn−1 and q(En−1). Fix a countable, dense set Tn ⊆ Dn and let Sn ⊆ E be a countable set with q(Sn) = Tn. Then the C∗-subalgebra of E generated by j(In−1), En−1, and Sn is separable, and hence there is a separable C∗-subalgebra En of E which satisfies (PE) and contains j(In−1), En−1, and Sn. Then Tn = q(Sn) ⊆ q(En) since Sn ⊆ En, and as Tn is dense in Dn and the ∗-homomorphism qEn has closed range, Dn ⊆ q(En). Also, as j(In−1) ⊆ En and j is injective, In−1 ⊆ j−1(En). Finally, as j−1(En) is a separable C∗-subalgebra of I, there is a separable C∗-subalgebra In of I which satisfies (PI ) and contains j−1(En). This completes the construction. (cid:3) Corollary 1.7. Let (P ) be a property of separable C ∗-algebras preserved under sequential inductive limits with injective connecting maps. If (P ) is preserved by ideals, quotients, or extensions of separable C ∗-algebras, then separably (P ) has the same permanence property among all C ∗-algebras. Proof. We only consider the case of extensions as the other two results are similar. Let (P ) be a property of separable C∗-algebras preserved by extensions. Suppose I is an ideal of a C∗-algebra A such that I and A/I both separably satisfy (P ) and suppose A0 is a separable C∗-subalgebra of A. By Proposition 1.6, there are separable C∗-subalgebras I, A, and B of I, A, and A/I, respectively, such that I and B satisfy (P ), A contains A0, and there is a homomorphism 0 0 I I A A B A/I 0 0 of extensions where the vertical arrows are the inclusion maps. Now A is a separable C∗-subalgebra of A satisfying (P ) and containing A0. (cid:3) 10 CHRISTOPHER SCHAFHAUSER The method for reducing to separable C∗-algebras given here also behaves well with hereditary subalgebras. Proposition 1.8. If (P ) is a property of separable C ∗-algebras preserved by hereditary subalgebras, then the property separably (P ) is preserved by hereditary subalgebras. Proof. Suppose A is a C∗-algebra separably satisfying (P ) and B ⊆ A is a hereditary subalgebra. Let B0 be a separable C∗-subalgebra of B. There is a separable C∗-subalgebra A of A which satisfies (P ) and contains B0. Now, B := B0 AB0 is a separable C∗-subalgebra of B containing B0. Moreover, B is a hereditary subalgebra of A whence satisfies (P ). (cid:3) The following result will be used heavily in Section 4. Proposition 1.9. Suppose A and B are C ∗-algebras such that A is separable and B is If ϕ : A → B is a full, nuclear ∗-homomorphism, there is a separable, unital unital. C ∗-subalgebra B0 of B such that ϕ(A) ⊆ B0 and the corestriction of ϕ to B0 is full and nuclear. Proof. As A is separable and ϕ is nuclear, for each integer n ≥ 1, there are an integer d(n) ≥ 1 and completely positive maps θn : A → Md(n) and ρn : Md(n) → B such that kρn(θn(a)) − ϕ(a)k → 0 for all a ∈ A. By Proposition II.8.5.7 of [3], there is a sequence (an)∞ n=1 ⊆ A \ {0} such that for every non-zero ideal I ⊆ A, there is an n ≥ 1 with an ∈ I. As ϕ is full and B is unital, for each n ≥ 1, there are an integer k(n) ≥ 1 and elements bn,i, b′ n,i ∈ B for i = 1, . . . , k(n) such n,i = 1B. i=1 bn,iϕ(an)b′ that Pk(n) Let B0 denote the C∗-subalgebra of B generated by ϕ(A), ρn(Md(n)), bn,i, and b′ n,i for i = 1, . . . , k(n) and n ≥ 1. Then B0 is separable, and if ϕ0 is the corestriction of ϕ to B0, then ϕ0 is nuclear. Suppose a ∈ A \ {0} and let I denote the ideal of B0 generated by ϕ(a). Then ϕ−1(I) is an ideal in A which is non-zero as a ∈ ϕ−1(I), and hence there is an integer n ≥ 1 such that an ∈ ϕ−1(I). Now, ϕ(an) ∈ I, and since ϕ(an) is full in B0 by construction, I = B0. This shows that ϕ0 is full. (cid:3) A version of the following result appeared in an early version of [78]. Proposition 1.10. Suppose G is a countable, abelian group and A is a C ∗-algebra. For i = 0, 1, the natural group homomorphisms lim −→ HomZ(G, Ki(A0)) −→ HomZ(G, Ki(A)) and lim −→ Ext1 Z(G, Ki(A0)) −→ Ext1 Z(G, Ki(A)) are isomorphisms where the limit is taken over all separable C ∗-subalgebras A0 of A. Proof. As G is a countable, abelian group, there is an extension 0 ZX ZY G 0 SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 11 for countable sets X and Y where ZX and ZY denote the free abelian groups generated by X and Y , respectively. For every C∗-algebra B and for i = 0, 1, there is a natural exact sequence HomZ(G, Ki(B)) HomZ(ZY, Ki(B)) HomZ(ZX, Ki(B)) Ext1 Z(G, Ki(B)). As inductive limits preserve exact sequences, it is enough to prove that the natural map lim −→ HomZ(ZX, Ki(A0)) −→ HomZ(ZX, Ki(A)) is an isomorphism for each countable set X. Adding a unit to A if necessary, we may assume A is unital. We only prove the case when i = 0 as the case i = 1 then follows by Bott periodicity (or by a similar argument). To show surjectivity, let f : ZX → K0(A) be given. For each x ∈ X, there are an integer n(x) ≥ 1 and projections px and qx in Mn(x)(A) with f (x) = [px] − [qx]. If A0 denotes the unital C∗-subalgebra of A generated by the entries of the projections px and qx, then there is a group homomorphism f0 : ZX → K0(A0) given by f0(x) = [px] − [qx]. If ι0 : A0 → A denotes the inclusion, then K0(ι0)f0 = f . To show injectivity, let A0 ⊆ A be a separable, unital C∗-subalgebra, let ι0 : A0 → A denote the inclusion, and suppose f, g : ZX → K0(A0) are such that K0(ι0)f = K0(ι0)g. For each x ∈ X, there are an integer n(x) ≥ 1 and projections px, qx, p′ x ∈ Mn(x)(A0) such that f (x) = [px] − [qx] and g(x) = [p′ x]. For each x ∈ X, there are an integer k(x) ≥ 1 and a partial isometry vx ∈ M2n(x)+k(x)(A) with x] − [q′ x, q′ v∗ xvx = px ⊕ q′ x ⊕ 1⊕k(x) A and vxv∗ x = p′ x ⊕ qx ⊕ 1⊕k(x) A Let A1 denote the C∗-subalgebra of A generated by A0 and the entries of each vx for x ∈ X. If ι1,0 : A0 ֒→ A1 denotes the inclusion, then K0(ι1,0)f = K0(ι1,0)g. (cid:3) 1.3. Tensorial Absorption and Separability. Recall from [84] that a separable, unital, infinite-dimensional C∗-algebra D is strongly self-absorbing if there is an isomorphism D → D ⊗ D approximately unitarily equivalent to the first factor embedding. For a strongly self-absorbing C∗-algebra D, a separable C∗-algebra A is D-stable if A ⊗ D ∼= A. The only strongly self-absorbing C∗-algebra needed here is the universal UHF-algebra Q. A local condition characterizing D-stability for separable, unital C∗-algebras can be extracted from Theorem 2.2 of [84] which shows a separable, unital C∗-algebra A is D- stable if, and only if, there is a unital embedding of D into the central sequence algebra A∞ ∩ A′ of A where A∞ := ℓ∞(A)/c0(A). This local characterization extends to the non-separable setting with D-stability replaced by separable D-stability. Lemma 1.11. For a strongly self-absorbing C ∗-algebra D, a unital C ∗-algebra A is sepa- rably D-stable if, and only if, for every finite set F ⊆ A, for every finite set G ⊆ D, and for every ε > 0, there is a unital, completely positive map ϕ : D → A such that kϕ(dd′) − ϕ(d)ϕ(d′)k < ε and kaϕ(d) − ϕ(d)ak < ε for all a ∈ F and d, d′ ∈ G. Proof. Assume first that A is separably D-stable and fix a finite set F ⊆ A, a finite set G ⊆ D, and ε > 0. There is a separable, unital, D-stable C∗-subalgebra A of A containing 12 CHRISTOPHER SCHAFHAUSER F. By Theorem 2.2 of [84], there is a unital embedding ϕ∞ : D → A∞ ∩ A′. As D is nuclear, the Choi-Effros lifting theorem implies that there are unital, completely positive maps ϕn : D → A representing ϕ∞. Hence lim n→∞ kϕn(dd′) − ϕn(d)ϕn(d′)k = lim n→∞ kaϕn(d) − ϕn(d)ak = 0 for all a ∈ A and d, d′ ∈ D. Take ϕ = ϕn for some sufficiently large n. Conversely, suppose the approximation condition holds and let A0 be a separable, unital C∗-subalgebra of A. Let F0,n be an increasing sequence of finite subsets of A0 with dense union and let Gn be an increasing sequence of finite subsets of D with dense union. There are unital, completely positive maps ϕ0,n : D → A such that kϕ0,n(dd′) − ϕ0,n(d)ϕ0,n(d′)k < 1 n and kaϕ0,n(d) − ϕ0,n(d)ak < 1 n for all a ∈ F0,n and d, d′ ∈ Gn. Let A1 denote the C∗-subalgebra of A generated by A0 and ϕ0,n(D) for each n ≥ 1 and note that A1 is separable. Iterating this argument, there are an increasing sequence of separable C∗-subalgebras Ak of A and sequences of unital, completely positive maps ϕk,n : D → Ak+1 with lim n→∞ kϕk,n(dd′) − ϕk,n(d)ϕk,n(d′)k = lim n→∞ kaϕk,n(d) − ϕk,n(d)ak = 0 for all a ∈ Ak, d, d′ ∈ D, and k ≥ 0. Let A ⊆ A denote the closed union of the Ak. A reindexing argument produces a sequence of unital, completely positive maps ψn : D → A such that lim n→∞ kψn(dd′) − ψn(d)ψn(d′)k = lim n→∞ kaψn(d) − ψn(d)ak = 0 for all a ∈ A and d, d′ ∈ D. The sequence ψn induces a unital ∗-homomorphism D → A∞ ∩ A′, and since A is separable, A is D-stable by Theorem 2.2 of [84]. As A contains A0 by construction, this shows A is separably D-stable. (cid:3) The next proposition collects some permanence properties of separable D-stability. Proposition 1.12. For a strongly self-absorbing C ∗-algebra D, hereditary subalgebras, quotients, and extensions of separably D-stable C ∗-algebras are separably D-stable, and ℓ∞-products and ultraproducts of unital, separably D-stable C ∗-algebras are separably D- stable. Proof. By Corollary 3.4 in [84], D-stability is preserved by sequential inductive limits of separable C∗-algebras and by Corollaries 3.1 and 3.3 and Theorem 4.3 of [84], hereditary subalgebras, quotients, and extensions of separable, D-stable C∗-algebras are D-stable. Now, separable D-stability passes to hereditary subalgebras by Proposition 1.8 and is preserved by quotients and extensions by Corollary 1.7. For ℓ∞-products, the result follows from Lemma 1.11 by choosing approximately central approximate morphisms into each factor of the product and taking the product of these maps. For ultraproducts, the result follows from the result for ℓ∞-products and quotients. (cid:3) SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 13 2. Some KK-theory This section contains a brief overview of KK-theory and collects the results on absorbing representations which will be needed in the classification results in Section 4. With one exception in the proof of Proposition 3.3, we will work exclusively with the nuclear KK- bifunctor KKnuc(−, −) introduced by Skandalis in [80]. 2.1. Basics of KK-theory. Let A be a separable C∗-algebra and let B be a σ-unital C∗-algebra. The word representation will refer to a ∗-homomorphism A → M (B ⊗ K). A representation ϕ : A → M (B ⊗ K) is called weakly nuclear if the completely positive map A → B ⊗ K given by a 7→ b∗ϕ(a)b is nuclear for all b ∈ B ⊗ K. Let Enuc(A, B) denote the set of pairs (ϕ, ψ) such that ϕ, ψ : A → M (B ⊗ K) are weakly nuclear representations with ϕ(a) − ψ(a) ∈ B ⊗ K for all a ∈ A. Such a pair (ϕ, ψ) is called a (weakly nuclear ) Cuntz pair. A homotopy between Cuntz pairs (ϕ0, ψ0) and (ϕ1, ψ1) in Enuc(A, B) is a Cuntz pair (Φ, Ψ) ∈ Enuc(A, C([0, 1], B)) such that for t ∈ {0, 1}, composing Φ and Ψ with the evaluation map M (C([0, 1], B) ⊗ K) ։ M (B ⊗ K) at t produces ϕt and ψt, respectively. Let KKnuc(A, B) denote the set of homotopy classes of Cuntz pairs in Enuc(A, B) and let [ϕ, ψ] denote the class of a Cuntz pair (ϕ, ψ) in KKnuc(A, B). 1 + t2t∗ 2 = 1, u = t1s∗ 1 + t2s∗ 1 + s2s∗ Let s1, s2 ∈ M (B ⊗ K) be isometries with s1s∗ 2 = 1. Given representations θ, ρ : A → M (B ⊗ K), define a representation θ ⊕s1,s2 ρ = s1θ(·)s∗ 2 called the Cuntz sum with respect to s1 and s2. For another choice of isometries t1, t2 ∈ M (B ⊗ K) with t1t∗ 2 is a unitary with ad(u)(θ ⊕s1,s2 ρ) = θ ⊕t1,t2 ρ, and hence, up to unitary equivalence, the Cuntz sum is independent of the choice of s1 and s2. If (ϕ1, ψ1) and (ϕ2, ψ2) are Cuntz pairs in Enuc(A, B), then (ϕ1 ⊕s1,s2 ϕ2, ψ1 ⊕s1,s2 ψ2) is also a Cuntz pair in Enuc(A, B). Since the unitary group of M (B ⊗ K) is path-connected in the operator norm topology by the main result of [14] (see also [56] for the case when B is unital), the class [ϕ1 ⊕s1,s2 ϕ2, ψ1 ⊕s1,s2 ψ2] in KKnuc(A, B) is independent of the choice of isometries s1 and s2; abusing notation, this element will be written as [ϕ1 ⊕ ϕ2, ψ1 ⊕ ψ2]. The set KKnuc(A, B) is an abelian group with addition given by Cuntz sum. 1 + s2ρ(·)s∗ If ϕ : A → B is a nuclear ∗-homomorphism and p ∈ K is a rank one projection, define a representation ϕp : A → M (B ⊗ K) by ϕp(a) = ϕ(a) ⊗ p for a ∈ A. Then (ϕp, 0) defines a Cuntz pair in Enuc(A, B) and the corresponding element of KKnuc(A, B) is denoted by [ϕ]. The element [ϕ] is independent of the choice of the rank one projection p. Given a separable C∗-algebra A and a ∗-homomorphism θ : B → D between σ-unital C∗-algebras B and D, there is an induced group homomorphism θ∗ : KKnuc(A, B) → KKnuc(A, D). In this way, KKnuc(A, −) becomes a covariant functor from the category of σ-unital C∗- algebras to the category of abelian groups; see [80] for the details. We will only need an explicit computation of θ∗ in the following special case. Proposition 2.1. Suppose A and E are separable C ∗-algebras, I ⊆ E is an ideal in E with I ⊗ K ∼= I, and ϕ, ψ : A → E are nuclear ∗-homomorphisms with ϕ(a) − ψ(a) ∈ I for all a ∈ A. Let λ : E → M (I) denote the canonical ∗-homomorphism and note that (λϕ, λψ) ∈ Enuc(A, I). If j : I → E denotes the inclusion map, then j∗[λϕ, λψ] = [ϕ] − [ψ] in KKnuc(A, E). 14 CHRISTOPHER SCHAFHAUSER Proof. As in [80] (see also Chapter 17 of [2]), KKnuc(A, I) can be realized as homotopy classes of Kasparov modules (θ0, θ1, F ) where θi : A → B(Ki) is a weakly nuclear repre- sentation of A on a countably generated Hilbert I-module Ki and F : K0 → K1 is an adjointable operator which, modulo the compacts, is a unitary intertwining of θ0 and θ1. Viewing I has a Hilbert module over itself, the Cuntz pair (λϕ, λψ) defines the Kasparov module (λϕ, λψ, 1M (I)) which defines the element [λϕ, λψ] ∈ KKnuc(A, I). If H0 = I viewed as a Hilbert E-module, then there is a natural isomorphism M (I) → B(H0). Let ϕ0, ψ0 : A → B(H0) denote the representations given by composing λϕ and λψ with this isomorphism. Then j∗[λϕ, λψ, 1M (I)] = [ϕ0, ψ0, 1B(H0)] in KKnuc(A, E) by the construction of j∗ given in [80] (see also Section 17.8 of [2]). Consider H1 = E as a Hilbert E-module. If ϕ1, ψ1 : A → B(H1) denote the represen- tations induced by ϕ and ψ, respectively, then [ϕ1, ψ1, 1B(H1)] = [ϕ] − [ψ] in KKnuc(A, I). It suffices to show (ϕ0, ψ0, 1B(H0)) and (ϕ1, ψ1, 1B(H1)) are homotopic. To this end, define H = {f ∈ C([0, 1], E) : f (0) ∈ I} and view H as a Hilbert C([0, 1], E)-module in the natural way. Define Φ, Ψ : A → B(H) by Φ(a)(f )(t) = ϕ(a)f (t) and Ψ(a)(f )(t) = ψ(a)f (t) for all a ∈ A, f ∈ C([0, 1], E), and t ∈ [0, 1]. Then (Φ, Ψ, 1B(H)) is a homotopy from (ϕ0, ψ0, 1B(H0)) to (ϕ1, ψ1, 1B(H1)). (cid:3) For a separable C∗-algebras A, a σ-unital C∗-algebras, B, and a representation ψ : A → n=1 be a M (B ⊗ K), define the infinite repeat ψ∞ : A → M (B ⊗ K) as follows. Let (sn)∞ sequence of isometries in M (B ⊗ K) such that P∞ snψ(a)s∗ n = 1 and let n=1 sns∗ ψ∞(a) = n ∈ M (B ⊗ K) ∞ Xn=1 for a ∈ A where convergence is in the strict topology. Then ψ∞ is a ∗-homomorphisms, and up to unitary equivalence, ψ∞ is independent of the choice of the sequence (sn)∞ n=1; in particular, ψ ⊕ ψ∞ and ψ are unitarily equivalent. Note also that if ψ is weakly nuclear, then so is ψ∞. 2.2. Absorbing Representations. Let A and B be C∗-algebras such that A is separable and B is σ-unital. Given two representations ϕ, ψ : A → M (B ⊗ K), write ϕ ∼ ψ if there is a sequence of unitaries (un)∞ n=1 ⊆ M (B ⊗ K) such that (1) kϕ(a) − unψ(a)u∗ (2) ϕ(a) − unψ(a)u∗ nk → 0 as n → ∞ and n ∈ B ⊗ K for all n ≥ 1 for all a ∈ A. Definition 2.2. Suppose A and B are C∗-algebras such that A is separable and B is σ-unital. A representation ϕ : A → M (B ⊗ K) is called (unitally) nuclearly absorbing if for all (unital) weakly nuclear representations θ : A → M (B ⊗ K), ϕ ⊕ θ ∼ ϕ. Consider the special case when B = C. All representations A → M (K) ∼= B(ℓ2(N)) are weakly nuclear as K is nuclear. Now, Voiculescu's representation theorem, as stated in Theorem II.5.8 of [22] for example, is the statement that a unital representation ϕ : A → SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 15 M (K) is unitally nuclearly absorbing if, and only if, ϕ is faithful and ϕ(A) ∩ K = 0. There is a far reaching generalization of this result for nuclearly absorbing representations due to the work of Elliott and Kucerovsky in [30] (see also [46] and [31]). A σ-unital C∗-algebra B has the corona factorization property if for all projections p ∈ M (B ⊗ K), 1 - p ⊕ p implies 1 - p (see [46]). The corona factorization property is a very weak regularity property of C∗-algebras. See [45, 58, 59] and the references within for several examples and connections to other regularity properties. In this paper, the examples of interest will be separable C∗-subalgebras of the trace-kernel ideal JB associated to an appropriate C∗-algebra B as discussed in the introduction and introduced formally in Section 3. When B is unital, a ∗-homomorphism ϕ : A → B is called unitizably full if the uni- tization ϕ† : A† → B is full. Note that when A and B are unital C∗-algebras, a ∗- homomorphism ϕ : A → B is unitizably full if, and only if, ϕ is full and 1B − ϕ(1A) is full. Theorem 2.3. If A is a separable C ∗-algebra and B is a σ-unital C ∗-algebra with the corona factorization property, then every unitizably full representation A → M (B ⊗ K) is nuclearly absorbing. Proof. After adding units, it is enough to show that every unital, full representation ϕ : A → M (B ⊗ K) is unitally nuclearly absorbing. Let q : M (B ⊗ K) → M (B ⊗ K)/(B ⊗ K) denote the quotient map. By [46], qϕ is a purely large extension, and hence, by the main result of [30], qϕ is unitally nuclearly absorbing as an extension. For any unital, weakly nuclear representation ψ : A → M (B ⊗ K), there is a unitary u ∈ M (B ⊗ K) such that u(ϕ(a) ⊕ ψ∞(a))u∗ − ψ∞(a) ∈ B ⊗ K for all a ∈ A. By the equivalence of (iv) and (v) in Theorem 3.4 of [34], ϕ ⊕ ψ∞ ∼ ϕ. As ψ∞ ⊕ ψ and ψ∞ are unitarily equivalent, ϕ ⊕ ψ ∼ ϕ ⊕ ψ∞ ⊕ ψ ∼ ϕ ⊕ ψ∞ ∼ ϕ, so ϕ is unitally nuclearly absorbing. (cid:3) Given two representations ϕ, ψ : A → M (B ⊗ K), write ϕ ∼asymp ψ if there is a norm continuous family (ut)t≥0 ⊆ M (B ⊗ K) of unitaries such that (1) kϕ(a) − utψ(a)u∗ (2) ϕ(a) − utψ(a)u∗ t k → 0 as t → ∞, and t ∈ B ⊗ K for all t ≥ 0 for all a ∈ A. The following folklore result shows nuclearly absorbing representations also satisfy a stronger asymptotic absorption condition. Proposition 2.4. Suppose A and B are C ∗-algebras such that A is separable and B is σ-unital. If ϕ : A → M (B ⊗ K) is a nuclearly absorbing representation and ψ : A → M (B ⊗ K) is a weakly nuclear representation, then ϕ ⊕ ψ ∼asymp ϕ. Proof. As ϕ is nuclearly absorbing, ϕ ⊕ ψ∞ ∼ ϕ, and hence ϕ ⊕ ψ∞ ∼asymp ϕ by the equivalence of (v) and (vi) in Theorem 3.4 of [34]. Now, ϕ ⊕ ψ ∼asymp ϕ ⊕ ψ∞ ⊕ ψ ∼asymp ϕ ⊕ ψ∞ ∼asymp ϕ. 16 CHRISTOPHER SCHAFHAUSER (cid:3) 2.3. Destabilizing KK-theory. Following Dadarlat and Eilers in [20], two representa- tions ϕ, ψ : A → M (B ⊗K) are properly asymptotically unitarily equivalent, written ϕ ≅ ψ, if there is a norm continuous family of unitaries (ut)t≥0 in B ⊗ K + C1M (B⊗K) ⊆ M (B ⊗ K) such that (1) kϕ(a) − utψ(a)u∗ (2) ϕ(a) − utψ(a)u∗ t k → 0 as t → ∞ and t ∈ B ⊗ K for all t ≥ 0 for all a ∈ A. The word proper reflects that the path of unitaries is taken from the minimal unitization of B ⊗ K instead of the multiplier algebra. This is a subtle, but very critical, difference with the relation ∼asymp above. The following result, which is essentially due to Dadarlat and Eilers in [20], shows the relevance of proper asymptotic equivalence in KK-theory. This is in stark contrast with Proposition 2.4 above, which shows the relation ∼asymp is a rather weak equivalence relation on representations. Theorem 2.5. Suppose A is a separable C ∗-algebra, B is a σ-unital C ∗-algebra, and (ϕ, ψ) ∈ Enuc(A, B) is a Cuntz pair. The following are equivalent: (1) [ϕ, ψ] = 0 ∈ KKnuc(A, B); (2) there is a weakly nuclear representation θ : A → M (B ⊗K) such that ϕ⊕θ ≅ ψ ⊕θ; (3) for any weakly nuclear, nuclearly absorbing representation θ : A → M (B ⊗ K), ϕ ⊕ θ ≅ ψ ⊕ θ. Proof. By Theorem 3.10 of [20], (1) and (2) are equivalent to (3') for any weakly nuclear, nuclearly absorbing representation θ : A → M (B ⊗ K), ϕ ⊕ θ∞ ≅ ψ ⊕ θ∞. Hence it suffices to show (3) is equivalent to (3'). If θ is weakly nuclear and nuclearly absorbing, then θ∞ is weakly nuclear, and hence θ ⊕ θ∞ ∼asymp θ by Proposition 2.4. So θ∞ ∼asymp θ, and the result follows from Lemma 3.4 in [20]. (cid:3) The next two results in this section will allow us to control the stabilizations appearing in the previous theorem when all representations considered are assumed to be nuclearly absorbing. These results form a critical part of the classification results in Section 4 and play a role analogous to that of the stable uniqueness theorem used in earlier classification results. Proposition 2.6. Suppose A and B are C ∗-algebras such that A is separable and B is σ-unital. If x ∈ KKnuc(A, B) and ψ : A → M (B ⊗ K) is a weakly nuclear, nuclearly absorbing representation, then there is a weakly nuclear, nuclearly absorbing representation ϕ : A → M (B ⊗ K) such that (ϕ, ψ) is a Cuntz pair in Enuc(A, B) and x = [ϕ, ψ]. Proof. Let x ∈ KKnuc(A, B) and ψ be given. There is a Cuntz pair (θ, ρ) in Enuc(A, B) such that x = [θ, ρ]. By Proposition 2.4, there is a norm continuous family (ut)t≥0 of unitaries in M (B ⊗ K) such that (1) kψ(a) − ut(ρ ⊕ ψ)(a)u∗ (2) ψ(a) − ut(ρ ⊕ ψ)(a)u∗ t k → 0 as t → ∞ and t ∈ B ⊗ K for all t ≥ 0 SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 17 for all a ∈ A. For each t ≥ 0 and a ∈ A, ut(ρ ⊕ ψ)(a)u∗ 0 differ by an element of B ⊗ K as both elements differ from ψ(a) by an element of B ⊗ K. As (θ, ρ) is a Cuntz pair, we have t and u0(ρ ⊕ ψ)(a)u∗ u0((θ ⊕ ψ)(a) − (ρ ⊕ ψ)(a))u∗ 0 ∈ B ⊗ K for all a ∈ A. Hence (ad(u0)(θ⊕ψ), ad(ut)(ρ⊕ψ)) is a Cuntz pair for all t ≥ 0 and defines a homotopy3 between the Cuntz pairs (ad(u0)(θ ⊕ ψ), ad(u0)(ρ ⊕ ψ)) and (ad(u0)(θ ⊕ ψ), ψ). Now, x = [θ ⊕ ψ, ρ ⊕ ψ] = [ad(u0)(θ ⊕ ψ), ad(u0)(ρ ⊕ ψ)] = [ad(u0)(θ ⊕ ψ), ψ]. To complete the proof, define ϕ = ad(u0)(θ ⊕ ψ). (cid:3) In the uniqueness portion of Theorem D, we will need a way to relate two weakly nuclear representations ϕ, ψ : A → M (B ⊗ K) when (ϕ, ψ) forms a Cuntz pair and [ϕ, ψ] = 0 in KKnuc(A, B). Ideally, if ϕ and ψ are nuclearly absorbing, then [ϕ, ψ] = 0 would imply ϕ ≅ ψ. This is known to be the case when B = K (see Theorem 3.12 in [20]) but is not known in general. The following technical variation will be sufficient for our purposes. A stronger result of this form will appear in forthcoming work of the author with Carri´on, Gabe, Tikuisis, and White. Proposition 2.7. Suppose A is a separable C ∗-algebra, E is a separable, unital, Q-stable C ∗-algebra, and I ⊆ E is an ideal such that I ⊗ K ∼= I. Suppose ϕ, ψ : A → E are nuclear ∗-homomorphisms such that ϕ(a) − ψ(a) ∈ I for all a ∈ A. Let λ : E → M (I) be the canonical ∗-homomorphism and note that (λϕ, λψ) ∈ Enuc(A, I). If λϕ and λψ are nuclearly absorbing representations and [λϕ, λψ] = 0 ∈ KKnuc(A, I), then there is a sequence of unitaries (un)∞ n=1 ⊆ E such that for all a ∈ A. kϕ(a) − unψ(a)u∗ nk → 0 Proof. Since [λϕ, λψ] = 0 ∈ KKnuc(A, I), λϕ is nuclearly absorbing, and λψ is weakly nuclear, we have λϕ ⊕ λϕ ≅ λψ ⊕ λϕ by Theorem 2.5. Reversing the roles of ϕ and ψ, we have λϕ ⊕ λψ ≅ λψ ⊕ λψ. In particular, there is a sequence (un)∞ n=1 ⊆ M2(I + C1M (I)) such that k(λϕ ⊕ λϕ)(a) − un(λψ ⊕ λψ)(a)u∗ nk → 0 for all a ∈ A. As the restriction of λ to M2(I + C1E) is injective, it follows that ϕ ⊕ ϕ and ψ ⊕ ψ are approximately unitarily equivalent as ∗-homomorphisms A → M2(E). As there is a unital embedding M2 ֒→ Q, ϕ⊗ 1Q and ψ ⊗ 1Q are approximately unitarily equivalent as ∗-homomorphisms A → E ⊗ Q. As E is Q-stable and Q is strongly self- absorbing, there is, by Remark 2.7 in [84], a sequence θn : E ⊗ Q → E of unital ∗- homomorphisms such that θn(x ⊗ 1Q) → x for all x ∈ E. An ε/3 argument implies that ϕ and ψ are approximately unitarily equivalent as ∗-homomorphisms A → E. (cid:3) 3Here the homotopy is defined on [0, ∞] in place of [0, 1]. 18 CHRISTOPHER SCHAFHAUSER 3. The Trace-Kernel Extension Let B be a simple, unital C∗-algebra with a unique trace τB and define the 2-norm on B by kbk2 = τB(b∗b)1/2 for all b ∈ B. Let ℓ∞(B) denote the C∗-algebra of all bounded sequences in B and, for a free ultrafilter ω on the natural numbers, define Bω := ℓ∞(B)/{b ∈ ℓ∞(B) : lim n→ω Bω := ℓ∞(B)/{b ∈ ℓ∞(B) : lim n→ω kbnk = 0} and kbnk2 = 0}. Since τB is contractive, kbk2 ≤ kbk for all b ∈ B, and hence there is a natural extension 0 JB jB Bω qB Bω 0 where qB is the canonical quotient map, JB = ker(qB), and jB is the inclusion map. The C∗-algebra JB is referred to as the trace-kernel ideal associated to B, and the extension is called the trace-kernel extension associated to B. The trace-kernel extension has been extensively studied in connection with the Toms- Winter conjecture (with a modified definition when B has more than one trace); see [5, 44, 54, 55, 75, 76, 83] for example. In the case B = Q, this extension also has a crucial role in the proof of the quasidiagonality theorem (Theorem 1.2 above), where the basic strategy is to lift a trace-preserving ∗-homomorphism into Qω ∼= Rω along the quotient map qQ to obtain a trace-preserving ∗-homomorphism into Qω. This was made more explicit in [78] where the lift was obtained through extension theoretic methods. Many of the properties of JQ proved in [78] also hold for JB for much more general C∗-algebras B (Proposition 3.2 below). The following definition is taken from [78]. Definition 3.1. A C∗-algebra I is called an admissible kernel if I has real rank zero and stable rank one, K0(I) is divisible, K1(I) = 0, the Murray-von Neumann semigroup V (I) is almost unperforated, and every projection in I ⊗ K is Murray-von Neumann equivalent to a projection in I. The following result collects the key properties of the trace-kernel extension needed in the next section. Proposition 3.2. If B is a simple, unital, Q-stable C ∗-algebra with a unique trace τB such that every quasitrace on B is a trace and K1(B) = 0, then (1) Bω is a II1-factor, (2) Bω has real rank zero and stable rank one, has a unique trace τBω , has strict comparison of positive elements with respect to the trace, is separably Q-stable, and has trivial K1-group, and (3) JB is an admissible kernel. Proof. For (1), as B has a unique trace τB, πτB (B)′′ is a II1-factor (see Theorem 6.7.4 and Corollary 6.8.5 in [23]), and hence so is the tracial ultrapower (πτB (B)′′)ω. By Remark 4.7 in [44], Bω ∼= Bω/JB ∼= (πτB (B)′′)ω. For (2), the C∗-algebra B has real rank zero by Theorem 7.2 of [71], stable rank one by Corollary 6.6 in [70], and strict comparison by Theorem 5.2 of [71]. All three properties are known to be preserved by ultraproducts; for strict comparison, this follows from Lemma SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 19 1.23 in [5], and for real rank zero and stable rank one, see the proof of Proposition 3.2 in [78] for example. As B is Q-stable, Bω is separably Q-stable by Proposition 1.12. By Theorem 8 in [61], the unique trace on Bω is the trace τBω induced by τB. As Bω has stable rank one, to show K1(Bω) = 0, it suffices by Theorem 2.10 of [69] to show that the unitary group of Bω is path connected. Let u be a unitary in Bω and fix a sequence of unitaries (un)∞ n=1 in B lifting u. As K1(B) = 0 and B has stable rank one, another application of Theorem 2.10 in [69] shows un is in the path component of the identity for each n ≥ 1. Since B has real rank zero, Theorem 5 of [47] implies there is a unitary vn ∈ B with finite spectrum such that kun − vnk < 1/n. Write vn = eihn for a self-adjoint hn ∈ B with khnk ≤ π and let h denote the self-adjoint element of Bω determined by the sequence (hn)∞ n=1. Then u = eih, and hence u is in the path component of the identity in the unitary group of Bω. For (3), as Bω has real rank zero and stable rank one, so does JB as both properties pass of ideals by Corollary 2.8 in [6] and Theorem 4.3 in [68], respectively. If d ≥ 1 is an integer and p ∈ Md(JB) is a projection, then (τBω ⊗ TrMd)(p) = 0 < 1 = (τBω ⊗ TrMd)(1Bω ⊕ 0⊕(d−1) Bω ) where TrMd is the tracial functional on Md normalized at a rank one projection. As Bω has strict comparison, there is a partial isometry v ∈ Md(Bω) such that v∗v = p and vv∗ ≤ 1Bω ⊕ 0⊕(d−1) . Since v∗v ∈ Md(JB), we have vv∗ ∈ Md(JB), and hence vv∗ = q ⊕ 0⊕(d−1) for some projection q ∈ JB. So every projection in JB ⊗ K is Murray- von Neumann equivalent to a projection in JB. Bω Bω Since Bω is separably Q-stable, so is JB by Proposition 1.12. Hence there is an increasing net Ji of a separable, Q-stable C∗-subalgebras of JB with dense union. It is well known that, since Ji is Q-stable, V (Ji) is unperforated and K0(Ji) is divisible; this can be shown using the continuity of V (−) and K0(−) and writing Ji ⊗ Q as the inductive limit of the ∗-homomorphisms Mk(Ji) → Mℓ(Ji) where k divides ℓ and the maps are given by the diagonal embeddings with multiplicity ℓ/k. Using the continuity of the functors V (−) and K0(−) again, it follows that V (JB) is almost unperforated and K0(JB) is divisible as both properties are preserved by inductive limits. It remains to show K1(JB) = 0. Consider the exact sequence K0(Bω) K0(qB ) K0(Bω) ∂0 K1(JB) K1(jB ) K1(Bω) induced by the trace-kernel extension. As K1(Bω) = 0, it suffices to show K0(qB) is surjective. Suppose t ∈ [0, 1] and write t = limn→ω tn with tn ∈ Q ∩ [0, 1]. As B is unital and Q-stable, there is a unital embedding Q ֒→ B. Hence there is a sequence of projections (pn)∞ n=1 ⊆ B with τB(pn) = tn. If p denotes the projection in Bω defined by the sequence pn, then τBω (p) = t. Hence the group homomorphism τBω : K0(Bω) → R is surjective, and since Bω is a II1-factor, the group homomorphism τBω : K0(Bω) → R is an isomorphism. Now, τBω = τBω K0(qB), so K0(qB) is surjective, and hence K1(JB) = 0. (cid:3) The relevant properties of admissible kernels are collected in the following result. Proposition 3.3. (1) The property of being an admissible kernel is separably inheritable. 20 CHRISTOPHER SCHAFHAUSER (2) If I is an admissible kernel, then Mn(I) is an admissible kernel for all n ≥ 1. (3) If I is a separable admissible kernel, then I is stable and has the corona factoriza- tion property. (4) If A is a separable C ∗-algebra satisfying the UCT and I is a separable admissible kernel, then the canonical homomorphism KKnuc(A, I) → HomZ(K0(A), K0(I)) is an isomorphism. Proof. (1) is Proposition 4.1 in [78]. In (2), the only non-trivial claim is that real rank zero and stable rank one are preserved by taking matrix algebras; this follows from Theorem 2.10 in [6] and Theorem 3.3 in [68], respectively. The first two paragraphs of the proof of Theorem 2.1 in [78] show (3). Since K1(I) = 0 and K0(I) is divisible, it is an immediate consequence of the UCT that the natural map KK(A, I) → HomZ(K0(A), K0(I)) is an isomorphism. By Theorem 4.1 and Proposition 7.1 in [73], A is KK-equivalent to a commutative C∗-algebra, and hence, by Propositions 3.2 and 3.3 in [80], the canonical map KKnuc(A, I) → KK(A, I) is an isomorphism. (cid:3) 4. Classification of ∗-homomorphisms into ultrapowers The goal of this section is to produce classification results for unital, full, nuclear ∗- homomorphisms from a separable, exact C∗-algebra A satisfying the UCT into an ultra- power of a suitably well-behaved codomain B as outlined in the introduction. The fol- lowing pullback lemma, shown to me by Jamie Gabe, will be used to control the range of a representation A → M (JB) in the proof of Proposition 4.2. This lemma is implicitly contained in the proof Theorem 5.1 in [78] and was explicitly stated in a slightly different form in an earlier version available on the arXiv. As the result does not appear in the published version of [78], the statement and proof are reproduced here. Lemma 4.1. Consider a commuting diagram 0 0 I I j1 j2 α1 β2 P α2 B2 B1 β1 D 0 0 of C ∗-algebras with exact rows. If A is a C ∗-algebra and ϕi : A → Bi are ∗-homomorphisms for i = 1, 2 such that β1ϕ1 = β2ϕ2, then there is a unique ∗-homomorphism ϕ : A → P such that αiϕ = ϕi for i = 1, 2. Moreover, ϕ is nuclear if, and only if, ϕ1 and ϕ2 are nuclear. Proof. Consider the pullback C∗-algebra Q = {(b1, b2) ∈ B1 ⊕ B2 : β1(b1) = β2(b2)} and define π : P → Q by π(p) = (α1(p), α2(p)). A diagram chase shows π is an isomor- phism. For existence, the ∗-homomorphism ϕ is given by ϕ(a) = π−1(ϕ1(a), ϕ2(a)), and uniqueness reduces to the statement ker(α1) ∩ ker(α2) = ker(π) = 0. If ϕ is nuclear, then ϕi = αiϕ is nuclear for i = 1, 2. Conversely, suppose ϕ1 and ϕ2 are nuclear. Fix a C∗-algebra C and let ρ : A ⊗max C → A ⊗min C be the canoni- cal ∗-homomorphism. As maximal tensor products preserve exact sequences, there is a SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 21 commuting diagram 0 0 I ⊗max C j1⊗id P ⊗max C α1⊗id B1 ⊗max C I ⊗max C j2⊗id B2 ⊗max C β2⊗id D ⊗max C α2⊗id β1⊗id 0 0 with exact rows. As ϕi is nuclear, there is a ∗-homomorphism ψi : A ⊗min C → Bi ⊗max C such that ϕi ⊗max idC = ψiρ for each i = 1, 2 by Corollary 3.8.8 of [10]. Applying the first part of the lemma to the maps ψ1 and ψ2, there is a ∗-homomorphism ψ : A ⊗min C → P ⊗max C such that (αi ⊗max idC)ψ = ψi. Since (αi ⊗max idC )ψρ = ϕi ⊗max idC = (αi ⊗max idC)(ϕ ⊗max idC), the uniqueness portion of the first part of the lemma implies ψρ = ϕ ⊗max idC. So ϕ ⊗max idC factors through A ⊗min C. As C was arbitrary, ϕ is nuclear by Corollary 3.8.8 of [10]. (cid:3) The next two propositions provide a first approximation to the existence and uniqueness results in Theorem D and its technical refinement given in Theorem 5.3. For any C∗-algebra C, ι2 : C → M2(C) denotes the inclusion into the (1, 1)-corner, and for any ∗-homomorphism f : C1 → C2 between C∗-algebras C1 and C2, the induced ∗-homomorphism M2(C1) → M2(C2) is still denoted by f . Proposition 4.2. Suppose A is a separable, unital, exact C ∗-algebra satisfying the UCT and B is a simple, unital, Q-stable C ∗-algebra with a unique trace τB such that every quasitrace on B is a trace and K1(B) = 0. If τA is a faithful, amenable trace on A and σ : K0(A) → K0(Bω) is a group homo- morphism such that τBω σ = τA and σ([1A]) = [1Bω ], then there is a unital, full, nuclear ∗-homomorphism ϕ : A → Bω such that K0(ϕ) = σ and τBω ϕ = τA. Proof. As B is Q-stable, there is a unital embedding Q → B which is necessarily trace- preserving by the uniqueness of the trace on Q. This induces a unital, trace-preserving embedding Qω → Bω. Composing this embedding with the ∗-homomorphism A → Qω given by Theorem 1.2 yields a unital, full, nuclear ∗-homomorphism ψ : A → Bω such that τBω ψ = τA. Note that τBω K0(qBψ) = τBω K0(qB)σ. As Bω is a II1-factor by Proposition 3.3, τBω is an isomorphism, and hence K0(qBψ) = K0(qB)σ. So the image of σ − K0(ψ) is contained in ker(K0(qB)) = im(K0(jB)). Using again that Bω is a II1-factor, K1(Bω) = 0, and hence K0(jB) is injective. So there is a group homomorphism κ : K0(A) → K0(JB) such that K0(jB)κ = σ − K0(ψ). By Proposition 1.9, there is a separable C∗-subalgebra E0 of Bω containing ψ(A) such that the corestriction of ψ to E0 is full and nuclear. By Proposition 1.10, there is a separable C∗-subalgebra I0 of JB such that κ factors as the composition of a group homo- morphism κ0 : K0(A) → K0(I0) and the group homomorphism K0(I0) → K0(JB) induced by the inclusion I0 ֒→ JB. As JB is an admissible kernel by Proposition 3.2 and being an admissible kernel is separably inheritable by Proposition 3.3, Proposition 1.6 implies there are a separable admissible kernel I ⊆ JB containing I0, a separable C∗-subalgebra E ⊆ Bω 22 CHRISTOPHER SCHAFHAUSER containing E0, and a separable C∗-subalgebra D ⊆ Bω such that there is a homomorphism 0 0  jB I ιI JB q qB E ιE Bω D ιD Bω 0 0 of extensions where the vertical maps are the inclusions. Let ψ : A → E denote the corestriction of ψ to E and let κ : K0(A) → K0(I) be the composition of κ0 with the group homomorphism K0(I0) → K0(I) induced by the inclusion I0 ֒→ I. As ψ : A → E is full and nuclear, ι2 ψ : A → M2(E) is unitizably full and nuclear. Let λ : M2(E) → M (M2(I)) be the canonical ∗-homomorphism. By Proposition 3.3, M2(I) is a separable admissible kernel whence is stable and has the corona factorization property. Note that λι2 ψ is unitizably full since λ is unital and ι2 ψ is unitizably full. Now by Theorem 2.3, λι2 ψ is nuclearly absorbing. As M2(I) is a separable admissible kernel and A satisfies the UCT, the canonical group homomorphism KKnuc(A, M2(I)) → HomZ(K0(A), K0(M2(I))) is an isomorphism by Proposition 3.3. Let x ∈ KKnuc(A, M2(I)) be a lift of K0(ι2)κ. By Proposition 2.6, there is a weakly nuclear representation θ : A → M (M2(I)) such that (θ, λι2 ψ) is a Cuntz pair in Enuc(A, M2(I)) and [θ, λι2 ψ] = x. As A is exact and θ is weakly nuclear, Proposition 3.2 of [32] implies that θ is nuclear. Lemma 4.1 applied to the diagram 0 0 M2(I)  M2(E) q M2(D) λ M2(I) M (M2(I)) M (M2(I))/M2(I) 0 0 implies there is a nuclear ∗-homomorphism ϕ2 : A → M2(E) such that λ ϕ2 = θ and q ϕ2 = qι2 ψ. Now in KKnuc(A, M2(E)), we have by Proposition 2.1. In particular, as x induces the group homomorphism K0(ι2)κ, ∗(x) = ∗[λ ϕ2, λι2 ψ] = [ ϕ2] − [ι2 ψ] K0( ϕ2) − K0(ι2 ψ) = K0(ι2)κ. Let ϕ2 = ιE ϕ2 : A → M2(Bω). Then K0(ϕ2) − K0(ι2ψ) = K0(ι2jB)κ = K0(ι2)σ − K0(ι2ψ) by the choice of κ, and hence K0(ϕ2) = K0(ι2)σ. In particular, [ϕ2(1A)] = K0(ι2)(σ([1A])) = [1Bω ] ∈ K0(Bω). As Bω has stable rank one by Proposition 3.2, Bω has cancellation of projections by Proposition 6.5.1 of [2], and there is a unitary u ∈ M2(Bω) with uϕ2(1A)u∗ = 1Bω ⊕ 0Bω . Now, there is a unital ∗-homomorphism ϕ : A → Bω such that ι2ϕ = ad(u)ϕ2. SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 23 We claim ϕ is the desired ∗-homomorphism. Note that K0(ι2ϕ) = K0(ϕ2) = K0(ι2)σ by the unitary invariance of K0. By the stability of K0, the map K0(ι2) is an isomorphism, and hence K0(ϕ) = σ. By construction, qBϕ2 = qBι2ψ, so if τM2(Bω ) is the trace on M2(Bω) induced by τBω , then τA = τBω ψ = 2τM2(Bω )ϕ2 = 2τM2(Bω)ι2ϕ = τBω ϕ. As ϕ is a compression of ϕ2 and ϕ2 is nuclear, ϕ is also nuclear. For each a ∈ A+ \ {0}, τBω (ϕ(a)) = τA(a) > 0 since τA is faithful. Since Bω has strict comparison by Proposition 3.2, it follows from Lemma 2.2 in [82] that ϕ is full. (cid:3) Proposition 4.3. Suppose A is a separable, unital, exact C ∗-algebra satisfying the UCT and B is a simple, unital, Q-stable C ∗-algebra with a unique trace τB such that every quasitrace on B is a trace and K1(B) = 0. If ϕ, ψ : A → Bω are unital, full, nuclear ∗-homomorphisms such that K0(ϕ) = K0(ψ) and τBω ϕ = τBω ψ, then there is a unitary u ∈ Bω such that ϕ = ad(u)ψ. Proof. Note that τBω qBϕ = τBω qBψ. As Bω is a II1-factor by Proposition 3.2 and qBϕ and qBψ are nuclear, qBϕ and qBψ are unitarily equivalent by Proposition 1.1 and a reindexing argument. Let ¯v be a unitary in Bω with qBϕ = ad(¯v)qBψ. Again as Bω is a II1-factor, the unitary group of Bω is path connected. Hence there is a unitary v ∈ Bω such that qB(v) = ¯v. Replacing ψ with ad(v)ψ, we may assume qBϕ = qBψ. By Propositions 1.9 and 1.10, there is a separable C∗-subalgebra E0 of Bω containing ϕ(A) and ψ(A) such that the corestrictions of ϕ and ψ to E0 are full and nuclear and agree on K0. By Proposition 3.2, JB is an admissible kernel, Bω is separably Q-stable, and Bω is a II1-factor and, therefore, has trivial K1-group. As having trivial K1-group and being an admissible kernel are separably inheritable properties, Proposition 1.6 implies there are separable C∗-subalgebras I ⊆ JB, E ⊆ Bω, and D ⊆ Bω such that I is an admissible kernel, E contains E0 and is Q-stable, K1(D) = 0, and there is a homomorphism 0 0  jB I ιI JB q qB E ιE Bω D ιD Bω 0 0 of extensions where the vertical arrows are the inclusion maps. Let ϕ, ψ : A → E be the corestrictions of ϕ and ψ to E, respectively. Then q ϕ = q ψ, ϕ and ψ are unital, full, and nuclear, and K0( ϕ) = K0( ψ). Let λ : M2(E) → M (M2(I)) be the canonical ∗-homomorphism. Note that the image of λι2ϕ − λι2ψ is contained in M2(I), so (λι2ϕ, λι2ψ) is a Cuntz pair in Enuc(A, M2(I)). If κ : K0(A) → K0(M2(I)) is the group homomorphism induced by this Cuntz pair, then K0()κ = K0(ι2 ϕ) − K0(ι2 ψ) = 0 by Proposition 2.1. As K1(D) = 0, K0() is injective, and hence κ = 0. Proposition 3.3 implies M2(I) is an admissible kernel and the canonical group homomorphism KKnuc(A, M2(I)) → HomZ(K0(A), K0(M2(I))) 24 CHRISTOPHER SCHAFHAUSER is an isomorphism, so [λι2 ϕ, λι2 ψ] = 0. As ϕ and ψ are full and nuclear and λ is unital, we have λι2 ϕ and λι2 ψ are unitizably full and nuclear. As M2(I) is a separable admissible kernel, Proposition 3.3 implies M2(I) It follows that λι2 ϕ and λι2 ψ is stable and satisfies the corona factorization property. are nuclearly absorbing by Theorem 2.3, and Proposition 2.7 implies ι2 ϕ and ι2 ψ are approximately unitarily equivalent. Now, ι2ϕ and ι2ψ are approximately unitarily equivalent, and hence, by a reindexing argument, there is a unitary w ∈ M2(Bω) with ι2ϕ = ad(w)ι2ψ. As ϕ and ψ are unital, w commutes with ι2ϕ(1A) = ι2ψ(1A) = 1Bω ⊕ 0Bω . Therefore, w = u ⊕ u′ for unitaries u and u′ in Bω, and ϕ = ad(u)ψ. (cid:3) 5. The classification theorem The goal of this section is to show the classification results for ∗-homomorphisms from A to Bω given in Section 4 imply the analogous results for ∗-homomorphisms from A to B under the same hypothesis. The ideas used here go back at least as far as the work of Lin in [49] and Dadarlat and Eilers in [21] and have been used heavily since then. The uniqueness result for ∗-homomorphisms from A to B follows immediately from the uniqueness result for ∗-homomorphisms from A to Bω. The difficulty lies in the existence result where only approximately multiplicative maps from A to B can be produced directly from a ∗-homomorphism from A to Bω. To make up for this, a technical uniqueness result for approximately multiplicative maps is needed. Let A and B be C∗-algebras, let G ⊆ A be a finite set, and let δ > 0 be given. A linear, self-adjoint function ϕ : A → B is called (G, δ)-multiplicative if kϕ(aa′) − ϕ(a)ϕ(a′)k < δ for all a, a′ ∈ G. Following Section 3.3 of [21], a K0-triple for a unital C∗-algebra A is a triple (G, δ, P) where G ⊆ A is a finite set, δ > 0, and P ⊆ P∞(A) is a finite set of projections in matrices over A such that whenever B is a unital C∗-algebra and ϕ : A → B is a linear, self-adjoint, (G, δ)-multiplicative map, kϕ(p2) − ϕ(p)2k < 1/4 for all p ∈ P. Note that if P ⊆ P∞(A) is any finite set, then for all sufficiently large finite sets G ⊆ A and sufficiently small δ > 0, the triple (G, δ, P) is a K0-triple for A. Let χ denote the characteristic function of [1/2, ∞) defined on the real numbers. Note that if (G, δ, P) is a K0-triple for A and ϕ : A → B is a linear, self-adjoint, (G, δ)- multiplicative map, then 1/2 is not in the spectrum of ϕ(p) for all p ∈ P. Hence for p ∈ P, we may define ϕ#(p) := [χ(ϕ(p))] ∈ K0(B). In this way, every linear, self-adjoint, (G, δ)-multiplicative map ϕ : A → B defines a function ϕ# : P → K0(B). Lemma 5.1. Suppose A is a separable, unital, exact C ∗-algebra satisfying the UCT and B is a simple, unital, Q-stable C ∗-algebra with a unique trace τB such that every quasitrace on B is a trace and K1(B) = 0. If τA is a faithful, amenable trace on A and σ : K0(A) → K0(B) is a group homomor- phism such that τBσ = τA and σ([1A]) = [1B], then for any K0-triple (G, δ, P) for A, there is a unital, completely positive, nuclear, (G, δ)-multiplicative map ϕ : A → B such that ϕ#(p) = σ([p]) for all p ∈ P and τB(ϕ(a)) − τA(a) < δ for all a ∈ G. SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 25 Proof. Suppose σ and τA are given as in the statement and let (G, δ, P) be a K0-triple for A. Let ιB : B → Bω denote the diagonal embedding. By Proposition 4.2, there is a unital, nuclear ∗-homomorphism ϕω : A → Bω such that K0(ϕω) = K0(ιB)σ and τBω ϕω = τA. By the Choi-Effros lifting theorem, there is a sequence of unital, completely positive, nuclear maps ϕn : A → B representing ϕω. Let S1 = \a,a′∈G {n ≥ 1 : kϕn(aa′) − ϕn(a)ϕn(a′)k < δ} and note that S1 ∈ ω and, for each n ∈ S1, ϕn is (G, δ)-multiplicative. Fix p ∈ P and let d, k ≥ 1 be integers such that p ∈ Md(A) and there are projections e, f ∈ Mk(B) with σ([p]) = [e] − [f ]. Then [ϕω(p)] = [ιB(e)] − [ιB(f )]. Now, there are an integer ℓ ≥ 1 and a partial isometry v ∈ Md+k+ℓ(Bω) such that Let (vn)∞ v∗v = ϕω(p) ⊕ ιB(f ) ⊕ 1⊕ℓ Bω Bω ⊕ ιB(e) ⊕ 1⊕ℓ Bω . n=1 be a bounded sequence in Md+k+ℓ(B) lifting v and note that B ⊕ e ⊕ 1⊕ℓ nvn − χ(ϕn(p)) ⊕ f ⊕ 1⊕ℓ and vv∗ = 0⊕d kv∗ lim n→ω B k = lim n→ω kvnv∗ n − 0⊕d B k = 0. As p ∈ P was arbitrary and P is finite, S := \p∈P {n ∈ S1 : (ϕn)#(p) = [σ(p)]} ∈ ω. Let T = Ta∈G{n ≥ 1 : τB(ϕ(a)) − τA(a) < δ} and note that T ∈ ω since τBω ϕ = τA. Now, S ∩ T ∈ ω, and in particular, S ∩ T 6= ∅. Fix n ∈ S ∩ T and define ϕ = ϕn. Lemma 5.2. Suppose A is a separable, unital, exact C ∗-algebra satisfying the UCT and B is a simple, unital, Q-stable C ∗-algebra with a unique trace τB such that every quasitrace on B is a trace and K1(B) = 0. (cid:3) For any faithful trace τA on A, finite set F ⊆ A, and ε > 0, there is a K0-triple (G, δ, P) for A such that if ϕ, ψ : A → B are unital, completely positive, nuclear, (G, δ)- multiplicative maps with ϕ#(p) = ψ#(p), τB(ϕ(a)) − τA(a) < δ, and τB(ψ(a)) − τA(a) < δ for all a ∈ G and p ∈ P, then there is a unitary u ∈ B such that for all a ∈ F. kϕ(a) − uψ(a)u∗k < ε Proof. Assume the result is false and fix a faithful trace τA on A, a finite set F ⊆ A, and ε > 0 where the result fails. Let (Gn)∞ n=1 be an increasing sequence of finite subsets of A with dense union, let (δn)∞ n=1 be a decreasing sequence of positive real numbers converging n=1 be an increasing sequence of finite subsets of P∞(A) with dense4 to zero, and let (Pn)∞ union such that (Gn, δn, Pn) is a K0-triple for each n ≥ 1. For each n ≥ 1, there are unital, completely positive, nuclear, (Gn, δn)-multiplicative maps ϕn, ψn : A → B such that (ϕn)#(p) = (ψn)#(p), τB(ϕn(a)) − τA(a) < δn, and τB(ψn(a)) − τA(a) < δn for all a ∈ Gn and p ∈ Pn but such that for each unitary un ∈ B, there is an a ∈ F with 4Here P∞(A) is equipped with the metric induced by the norm on A ⊗ K. kϕn(a) − unψn(a)u∗ nk ≥ ε. 26 CHRISTOPHER SCHAFHAUSER Let ϕω, ψω : A → Bω denote the functions induced by the sequences (ϕn)∞ n=1 and (ψn)∞ n=1, respectively, and note that ϕω and ψω are unital ∗-homomorphisms. Since A is exact, ϕω and ψω are nuclear by Proposition 3.3 in [15]. Also, τBω ϕω = τBω ψω = τA. Since τA is faithful and Bω has strict comparison with respect to τBω by Proposition 3.2, ϕω and ψω are full by Lemma 2.2 in [82]. Fix p ∈ Pn and let d ≥ 1 be an integer with p ∈ Md(A). As (ϕk)#(p) = (ψk)#(p) for all k ≥ n, [χ(ϕk(p))] = [χ(ψk(p))] in K0(B) for all k ≥ n. Note that B has stable rank one by Corollary 6.6 in [70], and hence B has cancellation of projections by Proposition 6.5.1 of [2]. Now, there is a partial isometry vk ∈ Md(B) with v∗ kvk = χ(ϕk(p)) and vkv∗ k=1 defines a partial isometry v in Md(Bω) with v∗v = χ(ϕω(p)) = ϕω(p) and vv∗ = χ(ψω(p)) = ψω(p). Hence [ϕω(p)] = [ψω(p)] in K0(Bω). This shows K0(ϕω) = K0(ψω). k = χ(ψk(p)) for all k ≥ n. The sequence (vk)∞ Proposition 4.3 now shows there is a unitary u ∈ Bω with ϕω(a) = uψω(a)u∗ for all a ∈ A. If (un)∞ n=1 ⊆ B is a sequence of unitaries lifting u, then for each a ∈ A, lim n→ω kϕn(a) − unψn(a)u∗ nk = 0, and in particular, for some integer n ≥ 1, kϕn(a) − unψn(a)u∗ nk < ε for all a ∈ F which is a contradiction. (cid:3) Intertwining the previous two lemmas produces the following classification theorem which is the main technical result of the paper. Theorem 5.3. Suppose A is a separable, unital, exact C ∗-algebra satisfying the UCT and B is a simple, unital, Q-stable C ∗-algebra with a unique trace τB such that every quasitrace on B is a trace and K1(B) = 0. (1) If τA is a faithful, amenable trace on A and σ : K0(A) → K0(B) is a group homomorphism such that τBσ = τA and σ([1A]) = [1B], then there is a unital, faithful, nuclear ∗-homomorphism ϕ : A → B such that K0(ϕ) = σ and τBϕ = τA. (2) If ϕ, ψ : A → B are unital, faithful, nuclear ∗-homomorphisms with K0(ϕ) = n=1 ⊆ B such K0(ψ) and τBϕ = τBψ, then there is a sequence of unitaries (un)∞ that lim n→∞ kϕ(a) − unψ(a)u∗ nk = 0 for all a ∈ A. Proof. Note that (2) follows immediately from Lemma 5.2 with τA := τBϕ. For (1), fix an increasing sequence of finite sets Fn ⊆ A with dense union and a sequence εn > 0 with Pn εn < ∞. Let (Gn, δn, Pn) be K0-triples for A satisfying Lemma 5.2 for the trace τA and the pair (Fn, εn). Enlarging Gn and Pn and decreasing δn, we may assume the Gn are increasing with union dense in A, the δn are decreasing and converging to zero, and the Pn are increasing with union dense in P∞(A). By Lemma 5.1, for each n ≥ 1, there is a unital, completely positive, nuclear, (Gn, δn)- multiplicative map ψn : A → B with (ψn)#(p) = σ([p]) for all p ∈ Pn and τB(ψn(a)) − SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 27 τA(a) < δn for all a ∈ Gn. By Lemma 5.2, for each n ≥ 1, there is a unitary un+1 ∈ B such that kψn(a) − un+1ψn+1(a)u∗ n+1k < εn for all a ∈ Fn. Define ϕ1 = ψ1 and define ϕn = ad(u2u3 · · · un)ψn for n ≥ 2. Then for all n ≥ 1 and a ∈ Fn, kϕn(a) − ϕn+1(a)k = kψn(a) − un+1ψn+1(a)u∗ n+1k < εn. This implies (ϕn(a))∞ n=1 Fn whence, by an ε/3 argument, is Cauchy for all a ∈ A. The desired map ϕ : A → B is given by ϕ(a) = limn→∞ ϕn(a). (cid:3) n=1 is Cauchy for all a ∈ S∞ The following result provides the rigorous statement of Theorem D. Corollary 5.4 (Theorem D). Suppose A is a separable, unital, exact C ∗-algebra satisfying the UCT and B is a simple, unital AF-algebra with a unique trace τB and divisible K0- group. (1) If τA is a faithful, amenable trace on A and σ : K0(A) → K0(B) is a group homomorphism such that τBσ = τA and σ([1A]) = [1B], then there is a unital, faithful ∗-homomorphism ϕ : A → B such that K0(ϕ) = σ and τBϕ = τA. (2) If ϕ, ψ : A → B are unital, faithful ∗-homomorphisms with K0(ϕ) = K0(ψ) and τBϕ = τBψ, then there is a sequence of unitaries (un)∞ n=1 ⊆ B such that lim n→∞ kϕ(a) − unψ(a)u∗ nk = 0 for all a ∈ A. Proof. As B is an AF-algebra, we have that K0(B) is torsion-free, K1(B) = 0, and every quasitrace on B is a trace. As K0(B) is divisible and torsion-free, the homomorphism K0(B) → K0(B) ⊗Z Q given by [p] 7→ [p] ⊗ 1Q for all projections p ∈ B ⊗ K is an isomorphism. Also, viewing B ⊗ Q as the inductive limit of the diagonal embeddings Mk(B) → Mℓ(B) for integers k, ℓ ≥ 1 with k dividing ℓ, the group homomorphisms K0(Mk(B)) → K0(B) ⊗Z Q : [p] 7→ [p] ⊗ 1 k induce an isomorphism K0(B ⊗ Q) → K0(B) ⊗Z Q. These isomorphisms induce an iso- morphism K0(B) ∼= K0(B ⊗ Q) of ordered abelian groups, and it follows from Elliott's classification of AF-algebras that B ∼= B ⊗ Q. As B is nuclear, all ∗-homomorphisms A → B are nuclear, so the result follows from Theorem 5.3. (cid:3) 6. Applications Proof of Theorem A. We may assume A is unital. Let τA be a faithful, amenable trace on A and let G0 = spanQ{(τA ⊗ TrK)(p) : p ∈ A ⊗ K is a projection} ⊆ R. By [27], if G0 is equipped with the order and unit inherited from R, then G0 is a simple dimension group with a unique state given by the inclusion G0 ֒→ R. Let B be a unital AF- algebra with K0(B) ∼= G0 and note that B is simple and has a unique trace. Composing the map K0(A) → G0 induced by the trace with an isomorphism G0 → K0(B) produces a group homomorphism σ : K0(A) → K0(B) compatible with the unit and the trace. Corollary 5.4 now implies the existence of a unital, trace-preserving embedding A → B. (cid:3) 28 CHRISTOPHER SCHAFHAUSER Proof of Theorem B. The implications (1) ⇔ (2) ⇐ (3) are well known. To show (1) implies (3), let G be a countable, discrete, amenable group. A result of Higson and Kasparov in [39] shows G satisfies the Baum-Connes conjecture, and hence a result of Luck in [52] shows that if τC∗ r(G), then for all projections p ∈ C∗ r(G) produces a group homomorphism K0(C∗ r(G)) → K0(Q) compatible with the unit and trace. A result of Tu in [85] shows C∗ r(G) satisfies the UCT, and the result follows from Corollary 5.4. (cid:3) r (G) ⊗ TrK)(p) ∈ Q. As K0(Q) ∼= Q, the trace on C∗ r (G) denote the canonical trace on C∗ r(G) ⊗ K, (τC∗ Proof of Theorem C. If C0(X) ⋊r G embeds into a simple, unital AF-algebra B, then any trace on B induces a faithful, G-invariant, Borel measure on X with mass at most 1, and hence after rescaling, X admits a probability measure of the desired form. Conversely, suppose G is amenable and suppose µ is a faithful, G-invariant, Borel, If E : C0(X) ⋊ G → C0(X) is the canonical conditional probability measure on X. expectation, then the map C0(X) ⋊ G → C given by a 7→ ZX E(a) dµ is a faithful trace on C0(X) ⋊ G which is amenable as C0(X) ⋊r G is nuclear. A result of Tu in [85] shows C0(X) ⋊r G satisfies the UCT, so the result follows from Theorem A. (cid:3) In [51], Lin has shown that if A is an AH-algebra and α is an action of Zk on A such that A admits a faithful, α-invariant trace, then the crossed product A ⋊r Zk embeds into an AF-algebra. Theorem A implies the following generalization of this result. Corollary 6.1. Suppose A is a separable, exact C ∗-algebra satisfying the UCT, G is a countable, discrete, torsion-free, abelian group, and α is an action of G on A. If A admits a faithful, α-invariant, amenable trace, then A ⋊α G embeds into an AF-algebra. Proof. Let Gn be an increasing sequence of finitely generated subgroups of G with union G. If αn is the restriction of α to an action of Gn on A, then the canonical ∗-homomorphism lim −→ A ⋊αn Gn −→ A ⋊α G is an isomorphism. As each Gn is finitely generated, torsion-free, and abelian, Gn ∼= Zd(n) for some integer d(n) ≥ 1. It follows that A ⋊α G satisfies the UCT. Also, A is separable and exact, and if E : A ⋊α G → A is the canonical conditional expectation and τA is a faithful, α-invariant, amenable trace on A, then τAE is a faithful trace on A ⋊α G. As τA is amenable and A is exact, πτA(A)′′ is injective. Therefore, πτAE(A ⋊α G)′′ ∼= πτA(A)′′ ¯⋊G, is injective by Proposition 6.8 of [12], and hence τAE is amenable. The result now follows from Theorem A. (cid:3) The AF-embedding problem for C∗-algebras of countable 1-graphs was solved in [77]. For countable, cofinal, row-finite 2-graphs with no sources, a similar result was obtained by Clark, an Huef, and Sims in [11]. Using their techniques together with Theorem A, the main result of [11] extends to k-graphs. SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 29 Corollary 6.2. Let k ≥ 1 be an integer, let Λ be a countable, cofinal, row-finite k-graph with no sources, and let A1, . . . , Ak denote the coordinate matrices of Λ. The following are equivalent: (1) C∗(Λ) embeds into an AF-algebra; (2) C∗(Λ) is quasidiagonal; (3) C∗(Λ) is stably finite; (4) (cid:16)Pk (5) Λ admits a faithful graph trace. i=1 im(1 − At i)(cid:17) ∩ Z+Λ0 = {0}; Proof. The equivalence of (2) through (5) is Theorem 1.1 in [11], and it is well known that (1) implies (2). The same proof given in Lemma 3.7 of [11] shows (5) implies (1) by appealing to Theorem A above in place of Corollary B of [82]. (cid:3) A C∗-algebra A is matricial field (MF ) if there is a net ϕi : A → Mn(i) of linear, self-adjoint functions such that kϕi(aa′) − ϕi(a)ϕi(a′)k → 0 and kϕi(a)k → kak for all a, a′ ∈ A. Similarly, a trace τA on a C∗-algebra A is called matricial field (MF ) if there is a net ϕi : A → Mn(i) of linear, self-adjoint functions such that kϕi(aa′) − ϕi(a)ϕi(a′)k → 0 and τMn(i)(ϕi(a)) → τA(a) for all a, a′ ∈ A. The class of MF algebras was introduced by Blackadar and Kirchberg in [4]. Haagerup and Thorbjørnsen have shown in [37] that C∗ r(F2) is MF where F2 is the free group on two generators. From here, it follows as well that C∗ r(F ) is MF for all free groups F (one first reduces to the case where F is countable and then considers an embedding F ֒→ F2). Using this result, many reduced crossed products of C∗-algebras by free groups have been shown to be MF in [42, 66, 67, 79]. The results above allow for substantial generalizations of these results. Corollary 6.3. Suppose A is a separable, exact C ∗-algebra satisfying the UCT and α is an action of a free group F on A. If τ is a trace on A ⋊r F such that τ A is faithful and amenable, then τ is MF. Proof. Adding a unit to A if necessary, we may assume A is unital. By Theorem 4.8 in [67], there is a group homomorphism σ : K0(A) → K0(Qω) such that σ([1A]) = [1Qω ], τQω σ = τ A, and σK0(αs) = σ for each s ∈ F . By Proposition 4.2, there is a unital, full, nuclear ∗-homomorphism ϕ : A → Qω such that K0(ϕ) = σ and τQω ϕ = τ A. For each s in a free generating set for F , we have K0(ϕαs) = K0(ϕ) and τQω ϕαs = τQω ϕ, and hence, by Proposition 4.3, there is a unitary us ∈ Qω such that ϕαs = ad(us)ϕ. As F is free, the function s 7→ us extends to a unitary representation of F on Qω with ϕαs = ad(us)ϕ for all s ∈ F . This shows the trace τ A is α-MF in the sense of [67], and hence, if E : A ⋊r F → A denotes the conditional expectation, then the trace (τ A)E is MF. When F 6∼= Z, τ = (τ A)E by [38], and therefore, τ is MF. 30 CHRISTOPHER SCHAFHAUSER Now assume F = Z. Since A is exact and τ A is amenable, πτ A(A)′′ is injective. Since π(τ A)E(A ⋊ Z)′′ ∼= πτ A(A)′′ ¯⋊Z is injective by Proposition 6.8 of [12], (τ A)E is amenable. Now, A ⋊ Z is separable and exact, satisfies the UCT, and admits a faithful, amenable trace (τ A)E. By Theorem 3.8 in [33], (τ A)E is quasidiagonal, and hence A ⋊ Z is quasidiagonal. By Theorem 4.1 in [33], every amenable trace on A ⋊ Z is quasidiagonal and, in particular, is MF. So it suffices to show τ is amenable. Let T (A) denote the set of traces on A. Since A is unital, T (A) is a Choquet simplex by Theorem 3.1.18 in [74]. It was shown by Kirchberg in Lemma 3.4 of [43] that the set of amenable traces on A is a weak∗-closed face in T (A), so by Corollary II.5.20 of [1], there is a continuous, affine function f : T (A) → [0, 1] such that for ρ ∈ T (A), f (ρ) = 0 if, and only if, ρ is amenable. Let α be the action of T on A ⋊ Z dual to α. Then (τ A)E = ZT τ αz dz. As f ((τ A)E) = 0 and the map z 7→ f (τ αz) is positive and continuous on T, we have f (τ αz) = 0 for all z ∈ T by the faithfulness of the Lebesgue measure on T. Therefore, τ αz is an amenable trace for all z ∈ T, and in particular, τ = τ α1 is an amenable trace. (cid:3) Corollary 6.4. If A is a locally type I C ∗-algebra and α is an action of a free group F on A, then every trace on A ⋊r F is MF. Proof. We may assume A is separable. Suppose τ is a trace on A ⋊r F . Then J := {a ∈ A : τ (a∗a)} = 0 is an α-invariant ideal of A and τ vanishes on the ideal J ⋊r F of A ⋊r F . Hence, as F is exact, τ factors through (A/J)⋊r F . As A is locally type I, so is A/J. After replacing A with A/J, we may assume τ A is faithful. By Theorem 1.1 in [16], locally type I algebras satisfy the UCT. Since locally type I algebras are nuclear, the result follows from Corollary 6.3. (cid:3) The following is a substantial generalization of a result of Kerr and Nowak (Theorem 5.2 in [42]) which gives the same statement in the case when A is commutative. An action of a group G on a C∗-algebra A is called minimal if A admits no non-trivial G-invariant ideals. Corollary 6.5. Suppose A is a separable, unital, nuclear C ∗-algebra satisfying the UCT and F is a free group acting minimally on A. Then the following are equivalent: (1) A ⋊r F is MF; (2) A ⋊r F is stably finite; (3) A admits an invariant trace. Proof. All MF C∗-algebras are stably finite, so (1) implies (2). For (2) implies (3), if A ⋊r F is stably finite, then since A ⋊r F is unital and exact, A ⋊r F admits a trace τ by Corollary 5.12 in [36], and then τ A is an invariant trace on A. Finally, for (3) implies (1), suppose A admits an invariant trace τA. Since {a ∈ A : τA(a∗a) = 0} is an invariant ideal in A and α is minimal, τA is faithful. As A is nuclear, τA is amenable, and hence τAE is MF by Corollary 6.3. Since τA and E are faithful, τAE is faithful. So A ⋊r F has a faithful, MF trace whence is MF. (cid:3) SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 31 Theorem D also allows for a description of closed unitary orbits of normal operators in simple, unital AF-algebras with a unique trace and divisible K0-group in terms of spectral data. Corollary 6.6. Suppose B is a simple, unital AF-algebra with a unique trace and divisible K0-group. Two normal operators a and b in B are approximately unitarily equivalent if, and only if, σ(a) = σ(b), τB(f (a)) = τB(f (b)) for all f ∈ C(σ(a)), and for every compact, open set U ⊆ σ(a), the projections χU (a) and χU (b) are unitarily equivalent. Proof. Let X = σ(a) = σ(b) and let a and b be the ∗-homomorphisms C(X) → B given by the functional calculus. If X = σ(a) = σ(b), then K0(C(X)) is canonically isomorphic to C(X, Z) and is generated as a group by the elements [χU ] for compact, open sets U ⊆ X. Hence K0(a) = K0(b) if, and only if, χU (a) and χU (b) are unitarily equivalent for each compact, open set U ⊆ X. Now apply the uniqueness portion of Theorem 5.3 to the ∗-homomorphisms a and b. (cid:3) We end with an abstract characterization of AF-algebras among the class of simple, unital C∗-algebras with a unique trace and divisible K0-group. The following result is known and can be deduced from the classification of separable, simple, unital, nuclear, Z-stable C∗-algebras which satisfy the UCT and have a unique trace. The minimal route through the literature seems to be the quasidiagonality theorem of [82], the results of Matui-Sato to show A is tracially AF [55], and the classification of separable, simple, unital, nuclear, tracially AF-algebras satisfying the UCT due to Lin in [50]. The latter result makes heavy use of deep classification and structural results for approximately homogeneous C∗-algebras. It is worth emphasizing that the proof given here does not depend on tracial approximations of any kind and does not depend on any inductive limit structure beyond that for AF-algebras. Corollary 6.7. Suppose A is a simple, unital, C ∗-algebra with a unique trace and divisible K0-group. Then A is an AF-algebra if, and only if, A is separable, nuclear, and Q-stable, A satisfies the UCT, and K1(A) = 0. Proof. It is well known that AF-algebras are separable, nuclear, and satisfy the UCT. As K0(A) is divisible, K0(A ⊗ Q) ∼= K0(A) and hence A is Q-stable by the classification of AF-algebras. Conversely, suppose A is a separable, simple, unital, nuclear, Q-stable C∗-algebra with a unique trace, A satisfies the UCT, and K1(A) = 0. As A is simple, the trace on A is necessarily faithful, so A is stably finite. Hence (K0(A), K + 0 (A), [1A]) is an ordered group by Proposition 6.3.3 in [2]. As A is Q-stable and has a unique trace, A has real rank zero by Theorem 7.2 of [71]. Also, A has stable rank one by Corollary 6.6 in [70] and, therefore, has cancellation of projections by Proposition 6.5.1 of [2]. Therefore, K0(A) has Riesz interpolation by Corollary 1.6 in [86]. Also, as A is Q-stable, K0(A) is unperforated. By the Effros- Handelman-Shen Theorem [25], there is a unital AF-algebra B with K0(A) ∼= K0(B) as ordered groups. Necessarily, B is simple and has a unique trace. Let σ : K0(A) → K0(B) be a unital order isomorphism. As K0(B) has unique state, we have τBσ = τA. Applying Theorem 5.3, there is a unital, nuclear ∗-homomorphism ϕ : A → B with K0(ϕ) = σ and τBϕ = τA. 32 CHRISTOPHER SCHAFHAUSER As B is an AF-algebra and A has stable rank one, there is a ∗-homomorphism ψ : B → A such that K0(ψ) = σ−1 and τAψ = τB by the classification of ∗-homomorphisms out of AF- algebras (or by Theorem 5.3). Using Theorem 5.3, ψϕ is approximately unitarily equivalent to the identity on A and ϕψ is approximately unitarily equivalent to the identity on B. By Elliott's intertwining argument (see Corollary 2.3.4 of [72]), A ∼= B. (cid:3) References [1] E. Alfsen, Compact convex sets and boundary integrals, Ergebnisse der Mathematik und ihrer Grenz- gebiete, Band 57, Springer-Verlag, New York-Heidelberg, 1971. [2] B. Blackadar, K-Theory for Operator Algebras, second edition. Mathematical Sciences Research Institute Publications 5. Berkeley, CA, 1998. [3] B. Blackadar, Operator Algebras: Theory of C ∗-algebras and von Neumann algebras, Encyclopedia of Mathematical Sciences, Vol 122, Springer-Verlag, Berlin Heidelberg New York, 2006. [4] B. Blackadar, E. Kirchberg, Gereralized inductive limits of finite-dimensional C∗-algebras, Math. Ann., 307 (1997), 343 -- 380. [5] J. Bosa, N. Brown, Y. Sato, A. Tikuisis, S. White, W. Winter, Covering dimension of C∗-algebras and 2-coloured classification, to appear in Mem. Amer. Math. Soc. (arXiv:1506.03974 [math.OA]) [6] L. Brown, G. Pedersen, C∗-algebras of real rank zero, J. Funct. Anal., 99 (1991), 131 -- 149. [7] N. Brown, AF embeddability of crossed products of AF algebras by the integers, J. Funct. Anal., 160 (1998), 150 -- 175. [8] N. Brown, Crossed products of UHF algebras by some amenable groups, Hokkaido Math. J., 29 (2000), 201 -- 211. [9] N. Brown, Invariant means and finite representation theory of C∗-algebras, Mem. Amer. Math. Soc., 184 (2006). [10] N. Brown, N. Ozawa, C ∗-Algebras and Finite-Dimensional Approximations, Graduate Studies in Mathematics, 88, Amer. Math. Soc., Providence, RI, 2008. [11] L. Clark, A. an Huef, A. Sims, AF-embeddability of 2-graph algebras and quasidiagonality of k-graph algebras, J. Funct. Anal., 271 (2016), 958 -- 991. [12] A. Connes, Classification of injective factors: cases II1, II∞, IIIλ, λ 6= 1, Ann. of Math. (2), 104 (1976), 73 -- 115. [13] A. Cuiperca, T. Giordano, P. Ng, Z. Niu, Amenability and uniqueness, Adv. Math., 240 (2013), 325 -- 345. [14] J. Cuntz, N. Higson, Kuiper's theorem for Hilbert modules, Operator algebras and mathematical physics (Iowa City, Iowa, 1985), 429 -- 435, Contemp. Math., 62, Amer. Math. Soc., Providence, RI, 1987. [15] M. Dadarlat, Quasidiagonal morphisms and homotopy, J. Funct. Anal., 151 (1997), 213 -- 233. [16] M. Dadarlat, Some remarks on the Universal Coefficient Theorem in KK-theory, Operator algebras and mathematical physics (Constant¸a, 2001), 65 -- 74, Theta, Bucharest, 2003. [17] M. Dadarlat, Morphisms of simple tracially AF algebras, Internat. J. Math., 15 (2004), 919 -- 957. [18] M. Dadarlat, On the topology of the Kasparov groups and its applications, J. Funct. Anal., 228 (2005), 394 -- 418. [19] M. Dadarlat, On AF embeddability of continuous fields, C. R. Math. Acad. Sci. Soc. R. Can., 31 (2009), 16 -- 19. [20] M. Dadarlat, S. Eilers, Asymptotic unitary equivalence in KK-theory, K-Theory, 23 (2001), 305 -- 322. [21] M. Dadarlat, S. Eilers, On the classification of nuclear C∗-algebras, Proc. Lond. Math. Soc. (3), 85 (2002), 168 -- 210. [22] K. Davidson, C ∗-Algebras by Example, Fields Institute Monographs 6, American Mathematical So- ciety, Providence, RI, 1996. [23] J. Dixmier, C ∗-Algebras, North-Holland Publishing Company, Amersterdam, 1977. SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 33 [24] E. Effros, On the structure theory of C∗-algebras: Some old and new problems, Proceedings of Symposia in Pure Mathematics, 38 (1982), 19 -- 34. [25] E. Effros, D. Handelman, C. Shen, Dimension groups and their affine representations, Amer. J. Math., 102 (1980), 385 -- 407. [26] G. Elliott, On the classification of inductive limits of sequences of semisimple finite-dimensional algebras, J. Algebra, 38 (1976), 29 -- 44. [27] G. Elliott, On totally ordered groups, and K0, pp. 1 -- 49 in Ring Theory (Waterloo, 1978), edited by D. Handelman and J. Lawrence, Lecture Notes in Math. 734, Springer, Berlin, 1979. [28] G. Elliott, On the classification of C∗-algebras of real rank zero, J. Reine Angew. Math., 443 (1993), 179 -- 219. [29] G. Elliott, G. Gong, H. Lin, Z. Niu, On the classification of simple amenable C∗-algebras with finite decomposition rank, II, preprint (2015). (arXiv:1507.03437 [math.OA]). [30] G. Elliott, D. Kucerovsky, An abstract Voiculescu-Brown-Douglas-Fillmore absorption theorem, Pacific J. Math., 198 (2001), 385 -- 409. [31] J. Gabe, A note on non-unital absorbing extensions, Pacific J. Math., 284 (2016), 383 -- 393. [32] J. Gabe, Lifting theorems for completely positive maps, preprint (2016). (arXiv:1508.00389v3 [math.OA]) [33] J. Gabe, Quasidiagonal traces on exact C∗-algebras, J. Funct. Anal., 272 (2017), 1104 -- 1120. [34] J. Gabe, E. Ruiz, Absorbing representations with respect to closed operator convex cones, preprint (2015). (arXiv:1503.07799 [math.OA]) [35] G. Gong, H. Lin, Z. Niu, Classification of simple amenable Z-stable C∗-algebras, preprint (2015). (arXiv:1501.00135 [math.OA]). [36] U. Haagerup, Quasitraces on exact C∗-algebras are traces, C. R. Math. Acad. Sci. Soc. R. Can., 36 (2014), 67 -- 92. [37] U. Haagerup, S. Thorbjørnsen, A new application of random matrices: Ext(C∗ red(F2)) is not a group, Ann. of Math. (2), 162 (2005), 711 -- 775. [38] P. de la Harpe, G. Skandalis, Powers' property and simple C∗-algebras, Math. Ann., 273 (1986), 241 -- 250. [39] N. Higson, G. Kasparov, Operator K-theory for groups which act properly and isometrically on Hilbert space, Invent. Math., 144 (2001), 23 -- 74. [40] V. Jones, Index for subfactors, Invent. Math., 72 (1983), 1 -- 25. [41] T. Katsura, AF-embeddability of crossed products of Cuntz algebras, J. Funct. Anal., 196 (2002), 427 -- 442. [42] D. Kerr, P. Nowak, Residually finite actions and crossed products, Ergodic Theory Dynam. Systems, 32 (2012), 1585 -- 1614. [43] E. Kirchberg, Discrete groups with Kazhdan's property T and the factorization property are resid- ually finite, Math. Ann., 299 (1994), 551 -- 563. [44] E. Kirchberg, M. Rørdam, Central sequence algebras and tensorial absorption of the Jiang-Su algebra, J. Reine Angew. Math., 695 (2014), 175 -- 214. [45] E. Kirchberg, M. Rørdam, When central sequence C∗-algebras have characters, Internat. J. Math., 26 (2015), 32 pages. [46] D. Kucerovsky, P. Ng, The corona factorization property and approximate unitary equivalence, Houston J. Math., 32 (2006), 531 -- 550. [47] H. Lin, Exponential length of C∗-algebras with real rank zero and the Brown-Pedersen conjectures, J. Funct. Anal., 114 (1993), 1 -- 11. [48] H. Lin, Tracially AF C∗-algebras, Trans. Amer. Math. Soc., 353 (2001), 693 -- 722. [49] H. Lin, Stable approximate unitary equivalence of homomorphisms, J. Operator Theory, 47 (2002), 343 -- 478. [50] H. Lin, Classification of simple C∗-algebras of tracial topological rank zero, Duke Math. J., 125 (2004), 91 -- 119. [51] H. Lin, AF-embedding of the crossed products of AH-aglebras by finitely generated abelian groups, Int. Math. Res. Pap. IMRP 2008, rpn007, 67 pages. 34 CHRISTOPHER SCHAFHAUSER [52] W. Luck, The relation between the Baum-Connes Conjecture and the Trace Conjecture, Invent. Math., 149 (2002), 123 -- 152. [53] H. Matui, AF embeddability of crossed products of AT algebras by the integers and its applications, J. Funct. Anal., 192 (2002), 562 -- 580. [54] H. Matui, Y. Sato, Strict comparison and Z-absorption of nuclear C∗-algebras, Acta Math., 209 (2012), 179 -- 196. [55] H. Matui, Y. Sato, Decomposition rank of UHF-absorbing C∗-algebras, Duke Math. J., 163 (2014), 1441 -- 1496. [56] J. Mingo, K-theory and multipliers of stable C∗-algebras, Trans. Amer. Math. Soc., 299 (1987), 397 -- 411. [57] F. Murray, J. von Neumann, On rings of operators, IV, Ann. of Math. (2), 44 (1943), 716 -- 808. [58] E. Ortega, F. Perera, M. Rørdam, The corona factorization property and refinement monoids, Trans. Amer. Math. Soc., 363 (2011), 4505 -- 4525. [59] E. Ortega, F. Perera, M. Rørdam, The corona factorization property, stability, and the Cuntz semigroup of a C∗-algebra, Int. Math. Res. Not. IMRN, 1 (2012), 34 -- 66. [60] N. Ozawa, Homotopy invariance of AF-embeddability, Geom. Funct. Anal., 13 (2003), 216 -- 223. [61] N. Ozawa, Dixmier approximation and symmetric amenability for C∗-algebras, J. Math. Sci. Univ. Tokyo, 20 (2013), 349 -- 374. [62] N. Ozawa, M. Rørdam, Y. Sato, Elementary amenable groups are quasidiagonal, Geom. Funct. Anal., 25 (2015), 307 -- 316. [63] M. Pimsner, Embedding some transformation group C∗-algebras into AF-algebras, Ergodic Theory Dynam. Systems, 3 (1983), 613 -- 626. [64] M. Pimsner, D. Voiculescu, Imbedding the irrational rotation C∗-algebra into an AF-algebra, J. Operator Theory, 4 (2008), 201 -- 210. [65] S. Popa, On local finite-dimensional approximation of C∗-algebras, Pacific J. Math., 181 (1997), 141 -- 158. [66] T. Rainone, MF actions and K-theoretic dynamics, J. Funct. Anal., 267 (2014), 542 -- 578. [67] T. Rainone, C. Schafhauser, Crossed products of nuclear C∗-algebras by free groups and their traces, to appear in Adv. Math. (arXiv:1601.06090v2 [math.OA]) [68] M. Rieffel, Dimension and stable rank in the K-theory of C∗-algebras, Proc. Lond. Math. Soc. (3), 46 (1983), 301 -- 333. [69] M. Rieffel, The homotopy groups of the unitary groups of non-commutative tori, J. Operator Theory, 17 (1987), 237 -- 254. [70] M. Rørdam, On the structure of simple C∗-algebras tensored with a UHF-algebra, J. Funct. Anal., 100 (1991), 1 -- 17. [71] M. Rørdam, On the structure of simple C∗-algebras tensored with a UHF-algebra, II, J. Funct. Anal., 107 (1992), 255 -- 269. [72] M. Rørdam, Classification of nuclear, simple C∗-algebras, in Classification of nuclear C ∗-algebras. Entropy in operator algebras, volume 126 of Encyclopaedia Math. Sci., pages 1 -- 145, Springer, Berlin, 2002. [73] J. Rosenberg, C. Schochet, The Kunneth theorem and the universal coefficient theorem for Kas- parov's generalized K-functor, Duke Math. J., 55 (1987), 431 -- 474. [74] S. Sakai, C ∗-Algebras and W ∗-Algebras, Springer-Verlag, New York/Berlin, 1971. [75] Y. Sato, Trace spaces of simple nuclear C∗-algebras with finite-dimensional extreme boundary, preprint (2012). (arXiv:1209.3000v1 [math.OA]). [76] Y. Sato, S. White, W. Winter, Nuclear dimension and Z-stability, Invent. Math., 202 (2015), 893 -- 921. [77] C. Schafhauser, AF-embeddings of graph C∗-algebras, J. Operator Theory, 74 (2015), 177 -- 182. [78] C. Schafhauser, A new proof of the Tikuisis-White-Winter theorem, to appear in J. Reine Angew. Math. (arXiv:1701.02180 [math.OA]) [79] C. Schafhauser, MF traces and the Cuntz semigroup, preprint (2017). (arXiv:1705.06555 [math.OA]) [80] G. Skandalis, Une notion de nucl´earit´e en K-th´eorie (d'apr`es J. Cuntz), K-Theory, 1 (1988), 549 -- 573. SUBALGEBRAS OF SIMPLE AF-ALGEBRAS 35 [81] J. Spielberg, Embedding C∗-algebra extensions into AF algebras, J. Funct. Anal., 81 (1988), 325 -- 344. [82] A. Tikuisis, S. White, W. Winter, Quasidiagonality of nuclear C∗-algebras, Ann. of Math. (2), 185 (2017), 229 -- 284. [83] A. Toms, S. White, W. Winter, Z-stability and finite dimensional tracial boundaries, Int. Math. Res. Not. IMRN, 10 (2015), 2702 -- 2727. [84] A. Toms, W. Winter, Strongly self-absorbing C∗-algebras, Trans. Amer. Math. Soc., 359 (2007), 3999 -- 4029. [85] J. Tu, La conjecture de Baum-Connes pour les feuilletages moyennables, K-Theory, 17 (1999), 215 -- 264. [86] S. Zhang, A Riesz decomposition property and ideal structure of multiplier algebras, J. Operator Theory, 24 (1990), 209 -- 225. Department of Mathematics, University of Nebraska -- Lincoln, Lincoln, NE 68588 E-mail address: [email protected]
1106.4535	1	1106	2011-06-22T19:23:39	The tiling C*-algebra viewed as a tight inverse semigroup algebra	[ "math.OA", "math.RA" ]	We realize Kellendonk'?s C-algebra of an aperiodic tiling as the tight C-algebra of the inverse semigroup associated to the tiling, thus providing further evidence that the tight C*-algebra is a good candidate to be the natural associative algebra to go along with an inverse semigroup.	math.OA	math	The tiling C-algebra viewed as a tight inverse semigroup algebra Ruy Exel∗, Daniel Gon¸calves†, Charles Starling‡ Abstract We realize Kellendonk´s C-algebra of an aperiodic tiling as the tight C-algebra of the inverse semigroup associated to the tiling, thus providing further evidence that the tight C-algebra is a good candidate to be the natural associative algebra to go along with an inverse semigroup. 1 Introduction The notion of an inverse semigroup was introduced independently by Wagner [15] and Pre- ston [11] in the early 50's to give an abstract formulation for the algebraic structure formed by all partialy defined bijective maps on a given set, under the operation of composition in the largest domain where it makes sense. Perhaps the best way to illustrate that inverse semigroups do yield such an abstraction is via what is now known as the Wagner -- Preston Theorem, according to which every inverse semigroup is isomorphic to a semigroup of par- tially defined bijections. One of the main themes behind the present work is the question of what is the natural associative algebra to go along with a given inverse semigroup, generalizing the notion of the group algebra of a given group. Since inverse semigroups are special examples of ∗Departamento de Matem´atica, Universidade Federal de Santa Catarina, 88040-900 Florian´opolis SC, Brazil. Email: [email protected]. Partially supported by CNPq. †Departamento de Matem´atica, Universidade Federal de Santa Catarina, 88040-900 Florian´opolis SC, Brazil. E-mail: [email protected]. ‡Department of Mathematics, University of Ottawa, 585 King Edward, Ottawa ON, K1N 6N5. Email: [email protected]. 1 semigroups, they may of course be studied simply as semigroups. In particular one may construct their semigroup algebra, as first studied by Munn [10], or their l1-algebra as discussed by Hewitt and Zuckerman [4]. The l1-algebra of an inverse semigroup has been considered by Barnes [1], where the main problem treated in [4] was solved in a very natural way. The l1-algebra of an inverse semigroup over the field of the complex numbers has a natural structure of a -algebra and hence it is natural to consider its envelopping C- algebra, as studied by Duncan and Paterson [2]. The extensive literature on inverse semigroup algebras (be it the purely algebraic object, the l1-algebra, or the envelopping C-algebra) could be considered definitive were it not for the fact that it has some limitations. The first hint that the established notion fails to deliver in some situatons has already been noticed more than thirty years ago in Renault's thesis [13, 2.8]. For example, if one lets On be the "Cuntz inverse semigroup" [13, 2.2], it is perhaps not unreasonable to expect that the C-algebra of On be isomorphic to the Cuntz algebra On. Unfortunately, this is not so: Duncan and Paterson's construction applied to On gives the Toeplitz extension of On, rather than On itself. i = 1" involves sums, besides products, and hence it is not obvious how to express it in (multiplicative) semigroup terms. The trouble is that the famous Cuntz relation "Pn i=1 SiS ∗ Addressing this question the first named author has introduced an alternative construc- tion of a C-algebra from an inverse semigroup, based on the idea of "tight representations", which, when applied to the Cuntz inverse semigroup, does produce the Cuntz algebra. The purpose of the present work is to apply these ideas to inverse semigroups which naturally occur in the study of tilings, the study of which traces back to Wang [16]. A tiling of Rn gives rise to an inverse semigroup, which was originally introduced and studied by Kellendonk [5, 8]: Given a tiling T of Rn, one considers the set of all triples (t1, P, t2), where P is a patch in T (see below for the precise definitions), and t1 and t2 are tiles in P . The equivalence class of such a triple modulo the action of the group of translations of Rn is often called a "doubly-pointed pattern class". The set S of all doubly-pointed pattern classes, to which is added an ad-hoc "zero element", forms an inverse semigroup under the multiplication defined by [t1, P, t2][t′ 2], 1, P ′, t′ 2] = [t1, P ∪ P ′, t′ provided t2 = t1 and P ∪ P ′ is a patch in T . When these conditions are not met the product is simply set to be zero. In a close parallel to the above question of whether or not one obtains the Cuntz algebra as the C-algebra of the Cuntz inverse semigroup, one may ask if the C-algebra of an aperiodic tiling T , as defined in [7], may be obtained from the inverse semigroup S described above. The C-algebra associated to a tiling is the C-algebra of an ´etale equivalence relation Rpunc which is translational equivalence on a suitable space of tilings. Under suitable conditions on the tilings considered, the algebra C ∗(Rpunc) contains a generating set which form an inverse semigroup with respect to the C-algebra product which is isomorphic to the inverse semigroup associated to a tiling above. The answer to whether C ∗(Rpunc) can be obtained from S is again negative for Duncan and Paterson's construction. Kellendonk addresses this issue for the specific example of a tiling in [8] and for general 0-E-unitary inverse semigroups in Lawson's book [9, 9.2]. As our main result shows, the answer also becomes affirmative if we use the "tight C-algebra" of S instead. This may be viewed as further evidence that the tight C-algebra of an inverse semigroup is a good candidate to be the natural associative algebra to go along with an inverse semigroup. 2 Inverse Semgroups Recall that a semigroup S is said to be an inverse semigroup if for every s ∈ S there is an element s∗ ∈ S such that We denote by E the set of idempotent elements in S. s∗ss∗ = s∗ and ss∗s = s. Let X be a set. Denote by I(X) the inverse semigroup of all bijections between subsets of X, with the operation of composition of functions in the largest domain on which the composition makes sense. Let S be an inverse semigroup and let X be a locally compact Hausdorff space. An action of S on X is a semigroup homomorphism θ : S → I(X) such that for every s ∈ S the map θs is continuous and its domain is open in X, and the union of the domains of the θs is all of X. The object we use to study actions of inverse semigroups is the groupoid of germs associated the action. Let S be an inverse semigroup acting on a locally compact Hausdorff space X by an action θ. We denote by Y the subset of S × X given by Y = {(s, x) ∈ S × X : x ∈ Ds∗s}, where Ds∗s denotes the domain of θs∗s, and for every (s, x) and (t, y) in Y we say that (s, x) ∼ (t, y), if x = y and there exists an idempotent e such that x ∈ De and se = te. We denote the equivalence class of (s, x) by [s, x] and call it the germ of s at x. The groupoid of germs is then defined as G = Y / ∼, with the set of composible pairs given by G(2) = {([s, x], [t, y]) ∈ G × G : x = θt(y)} and operations defined by [s, x] · [t, y] = [st, y], [s, x]−1 = [s∗, θs(x)]. A basis for the topology of G consists or the collection of all Θ(s, U), where Θ(s, U) = {[s, x] ∈ G : x ∈ U}, for any s ∈ S and any open subset U ⊆ Ds∗s. 3 Tilings A tile is a subset of Rn homeomorphic to the closed unit ball. A partial tiling is a collection of tiles whose interiors are pairwise disjoint. A finite partial tiling will be called a patch. The support of a partial tiling is the union of its tiles. We define a tiling to be a partial tiling whose support is Rn. Given U ⊂ Rn and a partial tiling T , T (U) is all the tiles that intersect U, that is, T (U) = {t ∈ T t ∩ U 6= ∅}. For x ∈ Rn, T ({x}) is frequently abbreviated as T (x). Two partial tilings T and T ′ are said to agree on U if T (U) = T ′(U), and are said to be compatible if T and T ′ agree on Int(supp(T ) ∩ supp(T ′)). Given a vector x ∈ Rn we can take any subset U ⊂ Rn and form its translate by x, namely U + x = {u + x u ∈ U}. Thus, given a tiling T we can form another tiling by translating every tile by x. We denote the new tiling by T + x = {t + x t ∈ T } A set of tiles P = {p1, p2, . . . , pN } is called a set of prototiles for T if t ∈ T ⇒ t = pi +x for some pi ∈ P and x ∈ Rn. Prototiles may carry labels to distinguish translationally equivalent tiles. We also insist that the pi, when viewed as subsets of Rn, have the origin in their interior. This allows us to define a designated point in each tile. If t is a tile in a tiling and t = p + x(t) for some p ∈ P we say that the puncture of t is x(t). A tiling for which T + x = T for some non-zero x ∈ Rn is called periodic. A tiling for which no such non-zero vector exists is called aperiodic. Given a set T of tilings, we would define a metric on T . This definition is standard, for instance see [6]. Definition 1. Suppose that T is a set of tilings and that T, T ′ ∈ T . We define the distance between T and T ′ to be d(T, T ′) = inf{1, ǫ ∃ x, x′ ∈ Rn ∋ x , x′ < ǫ, (T − x)(B(0, 1/ǫ)) = (T ′ − x′)(B(0, 1/ǫ))}. This is called the tiling metric. Two tilings are close in this metric if they agree on a large ball around the origin up to a small translation. If T is a tiling, then we complete the set T + Rn = {T + x x ∈ Rn} in this norm and obtain a metric space ΩT . This space is called the continuous hull of T (also called the tiling space associated with T ). One can verify that given a Cauchy sequence {Ti}i∈N in T + Rn one can find a tiling T ′ with Tn → T ′, and that if P is a set of prototiles for T then P is a set of prototiles for T ′. The space ΩT can also be characterized as the set of all tilings T ′ such that every patch in T ′ appears as some translate of a patch in T . Although the inverse semigroup of a tiling defined below can be defined in a more general setting, the C-algebra of a tiling has most interest under some additional assumptions on the tiling. Definition 2. (Assumption 1) A tiling T is said to have finite local complexity if for every r > 0 the set {T (B(x, r)) x ∈ Rn} contains only finitely many different patches modulo translation. If T has finite local complexity, then ΩT is compact; see [12], Lemma 2. Definition 3. (Assumption 2) A tiling T is said to have repetitivity (or is repetitive) for any patch P ⊂ T there exists R > 0 such that for every x ∈ Rn there exists y ∈ Rn such that P + y ⊂ T (BR(x)). The tiling T is repetitive if and only if for every T ′ ∈ ΩT we have ΩT ′ = ΩT ; see [14], Lemma 1.2. Definition 4. (Assumption 3) A tiling T is called strongly aperiodic if ΩT contains no periodic tilings. Suppose that P is a set of prototiles for T . We let Ωpunc ⊂ ΩT be the set of tilings in ΩT that have a puncture at the origin. In other words, T ∈ Ωpunc if and only if T (0) ∈ P. The space Ωpunc is sometimes called the punctured hull of the tiling. If T satisifies the three assumptions above, then its punctured hull is homeomorphic to a Cantor set (see for example [6], p. 187). The space Ωpunc has a neighbourhood base consisting of sets of the following form: for a patch P and tile t ∈ P , define U(P, t) = {T ∈ Ωpunc P − xt ⊂ T }. Each of these sets is clopen and the set of all such U(P, t) forms a neighbourhood base for Ωpunc. Notice that for x ∈ Rn, we have U(P, t) = U(P + x, t + x). We now can define the inverse semigroup associated to a tiling T . We let M be the set of patches P such that the P appears somewhere in T and the support of P has connected interior. Let P, Q ∈ M, t1, t2 ∈ P and s1, s2 ∈ Q. We say the triples (t1, P, t2) and (s1, Q, s2) are equivalent, and write (t1, P, t2) ∼ (s1, Q, s2), if there exists x ∈ Rn such that t1 + x = s1, t2 + x = s2 and P + x = Q. The equivalence class of (t1, P, t2) is denoted [t1, P, t2] and is called a "doubly pointed pattern class". Let S = {[t1, P, t2] P ∈ M, t1, t2 ∈ P } ∪ {0}. We give S the structure of an inverse semigroup as follows. Let [t1, P, t2] and [t′ 2] be two doubly pointed pattern classes. Suppose we can find x, x′ ∈ Rn such that P + x and P ′ + x′ are patches in T and such that t2 + x = t′ 1 + x′. Then we define the product of these two classes to be 1, P ′, t′ [t1, P, t2][t′ 1, P ′, t′ 2] = [t1 + x, (P + x) ∪ (P ′ + x′), t′ 2 + x′] This product is well-defined because (P + x) ∪ (P ′ + x′) is a patch in T . We notice that 1) 6= ∅. If no such the product [t1, P, t2][t′ translations exist, we define the product to be zero. We also define all products involving 0 to be 0, and we define 2] is non-zero if and only if U(P, t2) ∩ U(P ′, t′ 1, P ′, t′ [t1, P, t2]∗ = [t2, P, t1]. This product and involution give S the structure of an inverse semigroup; see for example [9, 9.5], [8] or [5]. Notice when forming the product [t1, P, t2][t′ 2] we may pick repre- 1 and P ∪ P ′ is a patch. In this case we obtain the sentatives in each class so that t2 = t′ nicer formula 1, P ′, t′ [t1, P, t2][t′ 1, P ′, t′ 2] = [t1, P ∪ P ′, t′ 2]. We notice that there is an action θΩ of S on the space Ωpunc. Let s = [t1, P, t2] and define xs as x(t1) − x(t2). Then s can be seen as a partial bijection θΩ s : U(P, t2) → U(P, t1) θΩ s (T ) = T − xs. When we restrict to Ωpunc, we no longer have an Rn action. If T is a punctured tiling, then T + x need not be a a punctured tiling. We may still consider translational equivalence restricted to this subspace. Define Rpunc = {(T, T + x) T, T + x ∈ Ωpunc} If we give this the topology from Ωpunc × Rn, then Rpunc becomes a principal r-discrete groupoid. The unit space of Rpunc is clearly Ωpunc. Our next goal is to show that the groupoid of germs associated to the action of S on Ωpunc is isomorphic to Rpunc as a topological groupoid. In our case, Y = {(s, T ) ∈ S × Ωpunc : T ∈ Ds∗s} = {([t1, P, t2], T ) ∈ S × Ωpunc : T ∈ U(P, t2)}, and the equivalence relation on Y can be characterized as follows: Lemma 1. ([t1, P, t2], T ) ∼ ([t′ exists a patch P ′′ and a tile t ∈ P ′′, such that P ′′ ⊇ P ∪ P ′ and T ∈ U(P ′′, t). 2], T ′) if and only if T = T ′, t1 = t′ 1, P ′, t′ 1, t2 = t′ 2 and there 1, P ′, t′ 2] and suppose (s, T ) ∼ (s′, T ′). Then T = T ′ Proof. Let s = [t1, P, t2], s′ = [t′ and there exists s′′ = [t, P ′′, t] such that T ∈ U(P ′′, t) and ss′′ = s′s′′. Also, since (s, T ) and (s′, T ′) are in Y we have that T ∈ U(P, t2) ∩ U(P ′, t′ 2) ∩ U(P ′′, t), which implies that 1, P ′ ∪ P ′′, t] and hence T (0) = t2 = t′ ss′′ = s′s′′ iff P ′′ ⊇ P ∪ P ′ and t1 = t′ 1. 2 = t. It follows that ss′′ = [t1, P ∪ P ′′, t] and s′s′′ = [t′ For the converse it is enough to take the idempotent as e = [t, P ′′, t] in the definition of ∼. (cid:4) Note that if (s, T ) and (s′, T ) are equivalent, then xs = xs′, and so both pairs represent translating T by the same vector. This leads to the following. Theorem 1. The groupoid of germs associated to the action of S on Ωpunc and Rpunc are isomorphic as topological groupoids. Proof. Let s = [t1, P, t2], s′ = [t′ 1, P ′, t′ 2] ∈ S. We will show that the map α : G → Rpunc [s,T] 7→ (T − xs, T ) , where xs is the vector that takes the puncture of t2 to the puncture of t1, is a homeomor- phism. That α is well defined follows directly from lemma 1. To see that α is a bijection note that its inverse is given by α−1 ((T + x, T )) = [(T (0), P, T (x)), T ], where P is a patch containing T (0) ∪ T (x). Next we would like to show that the groupoid operations are preserved by the map α. Recall that G(2) = {([s, T ], [s′, T ′]) ∈ G × G : T = θs′(T ′)}, that is G(2) = {([s, T ], [s′, T ′]) ∈ G × G : T = T ′ − xs′} and notice that since T ∈ U(P, t2) and T ′ ∈ U(P ′, t′ 2), T = punc = {((T, T + x), (T ′, T ′ + x′)) ∈ Rpunc × T ′ − xs′ implies that t2 = t′ Rpunc : T + x = T ′}. Now it is not hard to check that ([s, T ], [s′, T ′]) ∈ G(2) if and only if (α([s, T ]), α([s′, T ′])) = ((T − xs, T ), (T ′ − xs′, T ′)) ∈ R(2) 1. Also recall that R(2) punc and we have that: α([s, T ][s′, T ′]) = α([ss′, T ′]) = α([[t1, P ∪ P ′, t′ 2], T ′]) = (T ′, T ′ − xr), where xr the the vector from the puncture of t′ 2 to the puncture of t1. On the other hand α([s, T ])α([s′, T ′]) = (T − xs, T )(T ′ − xs′, T ′) = (T ′ − xs − xs′, T ′) = (T ′ − xr, T ′) and hence α is multiplicative. To see that α preserves the inverse operation note that α([s, T ]−1) = α([s∗, θs(T )]) = α([[t2, P, t1], T − xs]) = (T, T − xs) = α([s, T ])−1 For the topological part, notice that a basis for the topology of Rpunc is given by the collection of graphs of the maps s : U(P, t2) → U(P, t1) given by s(T ) = T − xs and a basis for the topology of G consists or the collection of all Θ(s, U) = {[s, T ] ∈ G : T ∈ U}, where s ∈ S and U is any open subset of U(P, t2). So it is clear that a basis element of Rpunc is taken by α to an open set in G. Now, let Θ(s, U) be a basis element of the the topology of G. Then α(Θ(s, U)) = {(T − xs, T ) : T ∈ U ⊆ U(P, t2)}. We show that α(Θ(s, U)) is open. Let T1 ∈ U. Since U is open there exists a patch P ′ such that T1 ∈ U(P ′, t2) ⊆ U ⊆ U(P, t2) and this implies that {(T − xs, T ) : T ∈ U(P ′, t2)} is an open neighboorhood of (T1 − xs, T1) (since it is the graph of (t2, P ′, t1)) contained in α(Θ(s, U)). It follows that α is bicontinuous and we conclude that α is a homeomorphism as desired. (cid:4) 4 Tight Representations of S Recall that E = {e ∈ S e∗ = e2 = e}, the set of idempotents in S. In our case, they are elements of the form [t, P, t]. There is a natural partial order on E given by e ≤ f ⇔ e = ef . In our case, [t, P, t] ≤ [t′, P ′, t′] ⇔ P ′ ⊂ P, t = t′ for suitible choices of representatives of each class. We also notice that [t, P, t] ≤ [t′, P ′, t′] if and only if t = t′ and U(P, t) ⊂ U(P ′, t′). Definition 5. Let X be a partially ordered set with minimum element 0. A filter is a non-empty subset ξ ⊂ X such that 1. 0 /∈ ξ 2. if x ∈ ξ and y ≥ x, then y ∈ ξ. 3. if x, y ∈ ξ, ∃z ∈ ξ such that x, y ≥ z. An ultra-filter is a filter that is not properly contained in another. We notice that filters in E are closed under the product. Let T be a partial tiling contained in some tiling in Ωpunc and whose support contains the origin, and let t = T (0). Then define ξT = {[t, P, t] ∈ E t ∈ P ⊂ T }. In words, ξT is the set of all the [t, P, t] such that P is a patch around the origin in T . Lemma 2. Let T be a partial tiling contained in some tiling in Ωpunc and whose support contains the origin. Then the set ξT is a filter in E. If T is a tiling, then ξT is an ultra- filter. Furthermore, if ξ ⊂ E is a non-empty filter, then there is a partial tiling T such that ξ = ξT , and a tiling T ′ ∈ Ωpunc such that T ⊂ T ′. If ξ is an ultra-filter, T is neccessarily a tiling. Proof. First we prove that ξT is a filter: The zero element is clearly not in ξT . If we have [t, P, t] ≥ [t, P ′, t], and [t, P ′, t] ∈ ξT , then P ⊂ P ′, and so P ⊂ T , hence [t, P, t] ∈ ξT . If [t, P, t], [t, P ′, t] ∈ ξT , then P and P ′ are compatible so [t, P, t][t, P ′, t] = [t, P ∪ P ′, t] is in ξT , and this product is less than both. To see that if T is a tiling then ξT is an ultra-filter, suppose that we had a filter ξ with ξT ( ξ. Take [t′, P ′, t′] ∈ ξ \ ξT , and find M > 0 such that supp(P ) ⊂ BM (0) (this is the ball of radius M in Rn). Then P = T (BM (0)) is a patch in T around the origin, so [t, P, t] ∈ ξT ⊂ ξ. Filters are closed under products and 0 is not in a filter so [t, P, t][t′, P ′, t′] must be non-zero. Hence t = t′ and [t, P, t] ≤ [t′, P ′, t′] implies that [t′, P ′, t′] ∈ ξT , a contradiction. We conclude that ξT is an ultra-filter. Now, suppose that ξ is a non-empty filter. Then we can find [t, P, t] ∈ ξ. Filters in E are closed under products, so if we have another element in ξ it is necessarily of the form [t, P ′, t] for some patch P ′ that's compatible with P and such that U(P, t) ∩ U(P ′, t) 6= ∅. So let T = [[t,P ′,t]∈ξ P ′. As defined T is a partial tiling because all such patches must be compatible since filters are closed under products and filters do not contain 0. It is clear from the definition of ξT that ξ ⊂ ξT . Suppose that [t, P ′, t] ∈ ξT . For each tile q ∈ P ′, there is a patch Pq such that q ∈ Pq and [t, Pq, t] ∈ ξ because T is the union of tiles in such patches. From this, we have that Pq. P ′ ⊂ [q∈P ′ Since ξ is closed under products, ht,Sq∈P ′ Pq, ti ∈ ξ, and because of the above, [t, P ′, t] ≥"t, [q∈P ′ Pq, t# and hence [t, P ′, t] ∈ ξ. It follows that ξ = ξT as required. That T is contained in some tiling T ′ ∈ Ωpunc is obtained by noticing that the collection of closed sets {U(P ′, t)}[t,P ′,t]∈ξ has the property that any finite subcollection has non-empty intersection since all finite products are non-zero. Since Ωpunc is compact, this implies that U(P ′, t) 6= ∅. \[t,P ′,t]∈ξ Now, if we take any T ′ in the above intersection, then T ′ agrees with T everywhere, hence T ⊂ T ′. Now suppose that ξ is an ultra-filter. From above, ξ = ξT for some partial tiling T , and T ⊂ T ′ for some tiling T ′. It is clear that ξT ⊂ ξT ′ and that the containment is proper unless T is a tiling. (cid:4) A character is a non-zero map φ : E → {0, 1} such that φ(ef ) = φ(e)φ(f ) and which takes 0 to 0. Let E0 be the set of all characters on E. We give E0 the topology of pointwise convergence. If φ is a character, then the set ξφ = {x ∈ E φ(x) = 1} is a filter. Likewise, if ξ is a filter, then the map φξ(x) =( 1, if x ∈ ξ 0, otherwise is a character. Characters on E and filters in E are in one-to-one correspondence. Let E∞ be the set of all characters coming from ultra-filters, and denote its closure in E0 as Etight. Denote the character associated with ξT as φT . That is to say, Then we have the following Theorem 2. The map φT ([t, P, t]) =( 1, if T ∈ U(P, t) 0, otherwise . Ψ : Ωpunc → E∞ T 7→ φT is a homeomorphism. Furthermore, compactness of Ωpunc means that Ωpunc is homeomorphic to Etight as well. Proof. Injectivity, at least intuitively, makes a lot of sense. The ultra-filter ξT is the set of all patches around the origin in T . The union of all such patches is precisely T . We let T 6= T ′ and find patches P, P ′ ⊂ T ′ and U ⊂ Rn around the origin such that P and P ′ do not overlap. Hence [T (0), P, T (0)][T ′(0), P ′, T ′(0)] = 0. Since filters are closed under the product, we must have that ξT 6= ξT ′. Filters and characters are in one-to-one correspondence, so we must have that φT 6= φT ′. Surjectivity is given by Lemma 2. Now, the topology on E0 has a sub-basis consisting of sets of the form b([t, P, t], U) = {f ∈ E0 f ([t, P, t]) ∈ U}, where U is a subset of {0, 1}. Consider the basic open set U(P, t) ⊂ Ωpunc. Then Ψ(U(P, t)) = {φT T ∈ U(P, t)} = b([t, P, t], {1}) ∩ E∞ Hence Ψ is an open map. Now Ψ−1(b([t, P, t], {0}) ∩ E∞) = U(P, t)c which is open due to U(P, t) being closed. Hence Ψ is bicontinuous and thus a homeomor- ∼= E∞ = E∞ = Etight. (cid:4) phism. Since Ωpunc is compact, E∞ must be as well and so Ωpunc Following [3] we know that there is an intrinsic action, called θ, on any inverse semi- group S on Etight, such that for every idempotent e ∈ E, the domain of θe is De = {φ ∈ Etight : φ(e) = 1} and θs : Ds∗s → Dss∗ is given by θs(φ)(e) = φ(s∗es). The groupoid of germs arising from this action is called the tight groupoid of S and is denoted Gtight. Its associated C-algebra, C ∗(Gtight) is called the tight C-algebra of S. Recall that S also acts on Ωpunc by the action θΩ, that is, for s = [t1, P, t2] ∈ S, θΩ s : U(P, t2) → U(P, t1) is defined via θΩ s (T ) = T − xs, where xs is the vector from the puncture of t2 to the puncture of t1. Next we show that these two actions commute via the homeomorphism Ψ. Proposition 1. For all s = [t1, P, t2] ∈ S the diagram below commutes. U(P, t2) Ψ Ds∗s θΩ s θs U(P, t1) Ψ / Dss∗ Proof. First we show that Ψ(U(P, t2)) = {φ ∈ E∞ : φ[t2, P, t2] = 1}: Let φ ∈ E∞ be such that φ[t2, P, t2] = 1. By lemma 2, φ = φξT for some T ∈ Ωpunc and since φξT [t2, P, t2] = 1 iff T ∈ U(P, t2), we have that φ = Ψ(T ). The other inclusion is trivial. Now, since ψ is a homeomorphism, ψ(U(P, t2)) is closed and hence equal to the closure of {φ ∈ E∞ : φ[t2, P, t2] = 1}, which in turn is equal to {φ ∈ Etight : φ[t2, P, t2] = 1}, since if φ ∈ Etight and φ[t2, P, t2] = 1 then there exists a sequence (φi) in E∞ converging to φ, which implies that for i large enough φi[t2, P, t2] = φ[t2, P, t2] = 1. It follows that Ψ(U(P, t2)) = {φ ∈ Etight : φ[t2, P, t2] = 1} = Ds∗s. Analogously one can show that Ψ(U(P, t1)) = Dss∗. Now let T ∈ U(P, t2). Then θs(Ψ(T ))[t, P ′, t] = θs(φt)[t, P ′, t] = φT ([t2, P, t1][t, P ′, t][t1, P, t2]) and the element in which φT is applied is non-zero iff t = t1 and U(P, t1) ∩ U(P ′, t1) 6= ∅, which happens if and only if t = t1 and P and P ′ are compatible, in which case θs(Ψ(T ))[t, P ′, t] = φT ([t2, P ∪ P ′, t2]) . It follows that θs(Ψ(T ))[t, P ′, t] =  1, if t = t1 and P and P ′ are compatible. 0, otherwise . / / / On the other hand, Ψ(θΩ s (T ))[t, P ′, t] = Ψ(T − xs)[t, P ′, t] = φT −xs[t, P ′, t] =  and since T ∈ U(P, t2) we have that T −xs ∈ U(P ′, t) iff t = t1 and P and P ′ are compatible. It follows that θs ◦ Ψ = Ψ ◦ θΩ (cid:4) s as desired. 1, 0, if T − xs ∈ U(P ′, t) otherwise We combine these to obtain our main result. Theorem 3. Let T be a tiling which is strongly aperiodic, repetitive, and has finite local complexity and let S be its associated inverse semigroup. Then the groupoids Rpunc and tight groupoid associated with S, Gtight are isomorphic as topological groupoids. In particular, the tight C-algebra associated with S is isomorphic to the usual C-algebra associated to a tiling C ∗(Rpunc). References [1] B. A. Barnes, "Representations of the l1-algebra of an inverse semigroup", Trans. Amer. Math. Soc. 218 (1976), 361 -- 396. [2] J. Duncan and A. L. T. Paterson, "C-algebras of inverse semigroups", Proc. Edin- burgh Math. Soc. (2) 28 (1985), 41 -- 58. [3] R. Exel, "Inverse semigroups and combinatorial C-algebras", Bull. Braz. Math. Soc. (N.S.) 39 (2008), 191 -- 313. [4] E. Hewitt and H. S. Zuckerman, "The l1-algebra of a commutative semigroup", Trans. Amer. Math. Soc. 83 (1956), 70 -- 97. [5] J. Kellendonk, "The local structure of tilings and their integer group of coinvariants", Commun. Math. Phys. 187 (1997), 115-157. [6] J. Kellendonk and I.F. Putnam, "Tilings, C∗-algebras, and K-Theory", CRM Mono- graph Series, vol 13 (2000), 177 -- 206. [7] J. Kellendonk, "Noncommutative Geometry of Tilings and Gap Labelling", Rev. Math. Phys. 7 (1995), 1133 -- 1180. [8] J. Kellendonk, "Topological equivalence of tilings", J. Math. Phys. 38 (1997), no. 4, 1823-1842. [9] M. V. Lawson, "Inverse Semigroups: The Theory of Partial Symmetries," World- Scientific, Singapore, 1998. [10] W. D. Munn, "On semigroup algebras", Proc. Cambridge Philos. Soc. 51 (1955), 1 -- 15. [11] G. B. Preston, "Inverse semi-groups", Journal of the London Mathematical Society 29 (1954), 396 -- 403. [12] C. Radin and M. Wolff, "Space Tilings and Local Isomorphism", Geom. Dedicata, 42 (1992), 355-360. [13] J. Renault, "A groupoid approach to C-algebras", Lecture Notes in Mathematics vol. 793, Springer, 1980. [14] B. Solomyak, "Nonperiodicity implies unique composition for self-similar translation- ally finite tilings", Discrete Comput. Geom. 20 (1998), 265-279. [15] V. V. Wagner, "Generalised groups", Proceedings of the USSR Academy of Sciences 84 (1952), 1119 -- 1122. [16] H. Wang, "Proving theorems by pattern recognition", II, Bell Systems Tech. J. 40 (1961), 1-41
1007.0363	1	1007	2010-07-02T13:47:12	Isometric coactions of compact quantum groups on compact quantum metric spaces	[ "math.OA" ]	We propose a notion of isometric coaction of a compact quantum group on a compact quantum metric space in the framework of Rieffel where the metric structure is given by a Lipnorm. We prove the existence of a quantum isometry group for finite metric quantum spaces, preserving a given state.	math.OA	math	Isometric coactions of compact quantum groups on compact quantum metric spaces by Johan Quaegebeur(1) and Marie Sabbe(1) Abstract We propose a notion of isometric coaction of a compact quantum group on a compact quantum metric space in the framework of Rieffel where the metric structure is given by a Lipnorm. We prove the exis- tence of a quantum isometry group for finite metric quantum spaces, preserving a given state. 1 Introduction This paper pertains to the study quantum symmetries of a (classical or quan- tum) space. This problem can be formulated and studied in various settings. In [16], Wang considers the case where the space is finite and doesn't carry any extra structure. The quantum symmetry groups he obtains, can thus be interpreted as quantum permutation groups. The spaces we are interested in in this paper, are metric spaces, both classical and quantum. Symmetries of a (quantum) metric space should then preserve the extra metric structure of the space, i.e. they should be 'isometric' in a sense to be made precise. For finite metric spaces, the 'quantum isometry group' is therefore expected to be a quantum subgroup of the quantum permutation group of Wang. For finite classical metric spaces, this problem was studied by Banica [1]. He has given a definition for a quantum symmetry of a classical finite metric space. With this definition, he was able to construct a quantum isometry group as a subgroup of Wang's quantum permutation group. The framework of Banica is still a half-classical framework since the metric spaces he considers, are classical. The concept of a metric space also has a quantum version. One approach of quantizing metric spaces is by spectral triples ([7]). These triples are used in non-commutative geometry to quantize the differential structure of a manifold as well as the metrical structure of the space. The quantum isometries of spectral triplets have been studied in a number of papers (e.g. [8], [3], [9], [4], [5], [2], [6]). There exists another framework to describe quantum metric spaces. In [12], Rieffel introduces a framework in which he only quantizes the metri- cal information of the space, disregarding the differential structure. Hence 1Department of Mathematics; K.U.Leuven; Celestijnenlaan 200B; B–3001 Leuven (Bel- gium). E-mails: [email protected] and [email protected] 1 Rieffel considers quantum spaces that come only with a quantum metric (encoded by a so called Lipnorm) and carry no further structure. This is the framework we will be working in. In fact, this paper wants to formulate a suitable answer to Rieffel's suggestion at the end of section 6 of [13]: "It would be interesting to develop and study the notion of a 'quantum isome- try group' for quantum metric spaces as quantum subgroups of the quantum symmetry groups studied by Wang [16]". The quantum metric spaces considered by Rieffel, are compact. Classi- cally, the isometry group of a compact space is a compact group. Hence, if we want to define a 'quantum isometry group' in Rieffel's framework, we might expect it to be compact. Therefore, we gather some information on compact quantum groups and compact quantum metric spaces together with other preliminary notions in the second section. Next, we introduce and motivate our notion of isometric actions in the third section. Actually, we will need two notions: 1-isometric coactions and full-isometric coactions. The former notion is the more natural one from the intuitive point of view. The latter is technically stronger and will allow us to construct a quantum isometry group in some cases. In the fourth section, we prove that both notions coincide with the existing definition of Banica [1] for quantum isometries of finite classical metric spaces. In the following sections, we further explore the notion(s) of isometric coactions we introduced. In the fifth section we mainly prove Theorem 5.2 which is a quantum version of the classical result that a group acting on a metric space has a largest subgroup acting isometrically. Using that theo- rem, we can prove in section 6 the main result of this paper: the existence of a quantum isometry group for finite quantum metric spaces preserving a given state on the space, as subgroups of the quantum symmetry groups of Wang. 2 Preliminaries First we need to recall some introductory definitions. We define compact quantum groups, compact quantum metric spaces and coactions. 2.1 Compact quantum groups Definition 2.1. A compact quantum group (CQG) is a pair (A, ∆) where 1. A is a unital C-algebra 2. the 'comultiplication' ∆ : A → A ⊗ A is a -homomorphism such that • ∆ is 'coassociative': (ι ⊗ ∆) ∆ = (∆ ⊗ ι) ∆ • the sets ∆(A)(1 ⊗ A) and ∆(A)(A ⊗ 1) are dense in A ⊗ A. 2 The notion of a compact quantum group generalizes the notion of a 'classical' compact group. Indeed, any compact group G can be seen as a CGQ: take A = C(G) (the commutative C-algebra of continuous complex valued functions on G) and define the comultiplication ∆ : C(G) → C(G) ⊗ C(G) ∼= C(G × G) : f 7→ ∆f by (∆f )(s, t) = f (st) for s, t ∈ G. (1) Conversely, if (A, ∆) is a CQG for which the C-algebra A happens to be commutative, then there is a compact group G such that A can be identified with C(G) and such that, under this identification, ∆ is given by (1). Definition 2.2. Let (A, ∆) and ( A, ∆) be two compact quantum groups. We say that ϕ : A → A is a morphism of CQGs from (A, ∆) to ( A, ∆) if ϕ is a -homomorphism such that (ϕ ⊗ ϕ)∆ = ∆ϕ. For more information on compact quantum groups we refer to [11],[18]. 2.2 Compact quantum metric spaces Definition 2.3. A compact quantum metric space (CQMS) is a pair (B, L) where (i) B is a unital C-algebra (ii) L is a Lipnorm on B, i.e. L : B → [0, +∞] is a seminorm such that – L(b) = L(b∗) for every b ∈ B – ∀b ∈ B : L(b) = 0 ⇔ b ∈ C1 – L is lower semicontinuous – the ρL-topology coincides with the weak -topology on S(B), where ρL is the metric on the state space S(B) defined by ρL(µ, ν) = sup(cid:8)µ(b) − ν(b) (cid:12)(cid:12) b ∈ B, L(b) ≤ 1(cid:9) for every µ, ν ∈ S(B). The classical notion of a compact metric space fits into this framework. Indeed, let (X, d) be a compact metric space. Put B = C(X) and consider the Lipschitz seminorm Ld : C(X) → [0, +∞] : f 7→ sup(cid:26) f (x) − f (y) d(x, y) x, y ∈ X, x 6= y(cid:27) . (cid:12)(cid:12)(cid:12)(cid:12) 3 One can prove that this seminorm is a Lipnorm. Conversely, if (B, L) is a CQMS for a commutative C-algebra B, then B = C(X) for some compact space X and one can construct a metric d on X by restricting the metric ρL on the state space to the set of pure states µx : C(X) → C : f 7→ f (x) (where x ∈ X). Rieffel has proven in [12] that L is now equal to the Lipschitz seminorm Ld. Another source of examples of compact quantum metric spaces is given by some spectral triples. A spectral triple (B, H, D) is a -subalgebra B of the bounded operators on a Hilbert space H, together with a Dirac operator D on H. This is a self-adjoint (possibly unbounded) operator on H such that [D, b] has a bounded extension for every b ∈ B. In some cases, the normclosure of B in B(H) will be a CQMS for the seminorm L determined by L(b) = [D, b] for b ∈ B. For more information and examples of compact quantum metric spaces, we refer to the work of Rieffel [12], [14]. Note that in [14] and later articles, the definition of a Lipnorm is without the requirement of lower semicontinu- ity. But Rieffel shows that, starting from any, possibly not lower semicon- tinuous, Lipnorm L on B , one can always construct a lower semicontinuous Lipnorm L such that L and L induce the same metric on the state space of B. Therefore, we prefer to have this requirement in the definition. 2.3 Coactions Definition 2.4. A coaction of a CQG (A, ∆) on a unital C-algebra B is a unital -homomorphism α : B → B ⊗ A such that (i) (ι ⊗ ∆) α = (α ⊗ ι) α, (ii) the set α(B) (1 ⊗ A) is norm dense in B ⊗ A. We say that α is faithful if the set {(ψ ⊗ ι)α(b) b ∈ B, ψ ∈ B∗} generates A as a C-algebra. A classical action of a compact group G on a space X fits in this frame- work by taking α : C(X) → C(X) ⊗ C(G) ∼= C(X × G) : f 7→ α(f ) determined by α(f )(x, s) = f (s · x) for x ∈ X, s ∈ G. Definition 2.5. If B is a C-algebra, (A, ∆) is a compact quantum group and α : B → B ⊗ A is a coaction, we call the triple (A, ∆, α) a quantum transformation group (QTG) of B. We say that the QTG (A, ∆, α) is faithful if the coaction α is faithful. If ψ is a functional on B, we say that α preserves ψ if (ψ ⊗ ι)α(b) = ψ(b)1A 4 for all b ∈ B. In that case, we say that (A, ∆, α) is a quantum transfor- mation group of the pair (B, ψ). If (A, ∆, α) and ( A, ∆, α) are two quantum transformation groups of B or (B, ψ), we say that ϕ : A → A is a morphism of QTGs from (A, ∆, α) to ( A, ∆, α) if ϕ is a morphism of compact quantum groups such that (ι ⊗ ϕ)α = α. 3 Isometric coactions Now the natural question arises: if in the Definition 2.4 B happens to have the extra structure of a compact quantum metric space as in Definition 2.3, when would it make sense to call the action α isometric? Before we answer this question in general, we take a look at a specific case. Consider the situation where the CQG is actually a classical group, i.e. the situation where A is C(G) for some compact group G. Let α : B → B ⊗ C(G) be a coaction of the group G on a compact quantum metric space (B, L). To simplify notations, we identify B ⊗ C(G) with C(G, B), by b ⊗ f : G → B : s 7→ f (s)b for b ∈ B and f ∈ C(G). isometric coaction should be: we call α isometric if In this context, it is rather evident what an ∀b ∈ B, ∀s ∈ G : L(α(b)(s)) = L(b). (2) Notice that, in case of a classical CQMS (B, L), i.e. when B = C(X) and L = Ld for some compact metric space (X, d), condition (2) expresses that the action of G on X is isometric in the classical sense. Now we want to extend this definition to the double-quantum case, where a CQG is acting on a CQMS. Therefore we must get rid of everything that refers to the individual group elements. To start with, we write α(b)(s) as (ι ⊗ ωs)α(b) where ωs is the pure state on C(G) given by evaluating in s: ωs : C(G) → C : f 7→ f (s). (3) To get rid of the s in this formula, we want to replace ωs by an arbitrary state ω on A. Let's first take a look at what happens for a convex combination ω = i=1 λiωsi for some group elements s1, · · · , sk ∈ G and some positive num- i=1 λi = 1. Then, if the equality L((ι ⊗ ωs)α(b)) = Pk bers λ1, · · · , λk with Pk L(b) holds for every s ∈ G and every b ∈ B, we have that L((ι ⊗ ω)α(b)) ≤ k Xi=1 λiL((ι ⊗ ωsi)α(b)) = L(b) 5 for all b ∈ B. Using the lower semicontinuity of L, one can show that the above inequality also holds for arbitrary states, which will be done in Lemma 3.4. These ideas motivate the following definition for a coaction to be isometric. For later purposes, we formulate the definition in a more general setting. Definition 3.1. For a unital C-algebra A and a CQMS (B, L), a -homomorphism α : B → B ⊗ A is called 1-isometric if and only if ∀ω ∈ S(A), ∀b ∈ B : L((ι ⊗ ω)α(b)) ≤ L(b). (4) When writing this paper, we discovered that this notion already ap- peared in a recent paper of H. Li ([10], definition 8.8). He calls a coaction invariant if the above inequality holds. However, he did not further explore this notion in the direction we are studying (the problem of the existence of a quantum isometry group). Let us explain why we used the terminology '1-isometric', and why the definition was formulated for -homomorphisms instead of coactions. This has to do with the following problem: suppose we let two CQGs (A1, ∆1) and (A2, ∆2) act 1-isometrically on a CQMS (B, L) (denote the coactions by α1 and α2). In order to have a good definition of isometric coactions, we want the 'coaction' obtained by applying both coactions successively (i.e. the - homomorphism (α1 ⊗ ι)α2), to be isometric. Firstly, the -homomorphism (α1 ⊗ ι)α2 is not necessarily a coaction for the comultiplication one can define on A1 ⊗ A2. That is why we will also be using the terminology '1- isometric' for -homomorphisms α from a CQMS (B, L) to B ⊗ A where A is a C-algebra. Moreover, we cannot prove why condition (4) would hold for (α1 ⊗ ι)α2. To solve this problem, we strengthen the notion of 1-isometric to 'full isometric'. In order to formulate the definition, we introduce some notations. Notation 3.2. Let B be a C-algebra and (Ci, ∆i, βi) be QTGs of B for i = 1, · · · , n. Then we denote β1 ∗ β2 for the -homomorphism (β1 ⊗ ι)β2 : B → B ⊗ C1 ⊗ C2. Inductively we can define β := β1 ∗ · · · ∗ βn : B → B ⊗ C1 ⊗ · · · ⊗ Cn. Definition 3.3 (Full-isometric coaction). A coaction α of a CQG (A, ∆) on a CQMS (B, L) is called full-isometric if and only if for all faithful QTGs (Ci, ∆i, βi) (i = 1, · · · , n) such that β := β1 ∗ · · · ∗ βn is 1-isometric, we have that α ∗ β is 1-isometric, or ∀ω ∈ S(A ⊗ C1 ⊗ · · · ⊗ Cn), ∀b ∈ B : L((ι ⊗ ω)(α ⊗ ι)β(b)) ≤ L(b). As the terminology suggests, being full-isometric implies being 1-isometric. The inequality (4) is not always very practical to work with. The follow- ing lemma states that it is enough to have inequality (4) for some specific states, for example the set of all pure states. 6 Lemma 3.4. Suppose we have two unital C-algebras C and B, and a - homomorphism β : B → B ⊗ C. Assume that S ⊆ S(C) is a set of states on C such that the convex hull of S is weakly-* dense in the state space S(C). Then, for any lower semicontinuous seminorm L on B, the following are equivalent: (i) for every ω ∈ S and every b ∈ B, we have L((ι ⊗ ω)β(b)) ≤ L(b) (ii) for every ω ∈ S(C) and every b ∈ B, we have L((ι ⊗ ω)β(b)) ≤ L(b). Proof. Suppose (i) holds. First we check that the inequality in (ii) holds if i=1 λiωi for i=1 λi = 1. Then, ω is taken from the convex hull of S. Indeed, suppose ω = Pk some states ωi in S and some positive numbers λi with Pk for any b ∈ B, L((ι ⊗ ω)β(b)) ≤ k Xi=1 λiL((ι ⊗ ωi)β(b)) ≤ k Xi=1 λiL(b) = L(b). Now, if ω is an arbitrary state on C, then ω is the weak-* limit of some net of states ωi in the convex hull of S. This implies that, for any b ∈ B, the net (ι ⊗ ωi)β(b) converges weakly to (ι ⊗ ω)β(b). Hence (ι ⊗ ω)β(b) belongs to the weak closure of the convex set Y = {(ι ⊗ ψ)β(b) ψ in the convex hull of S}. But the weak closure of a convex set equals its norm closure. Therefore we can take a net of states ψn in the convex hull of S, such that the net (ι ⊗ ψn)β(b) norm-converges to (ι ⊗ ω)β(b). Using the lower semi-continuity of L, we find L((ι ⊗ ω)β(b)) = L(lim i (ι ⊗ ψi)β(b)) ≤ lim i L((ι ⊗ ψi)β(b)) ≤ L(b). (cid:4) There are two half classical cases that have been studied before. We prove that in these settings, the defined notions of 1-isometries and full- isometries both coincide with the existing notion of isometries. First we study the case where A is commutative, and thus of the form C(G) for a compact group G. Theorem 3.5. Let G be a compact group, (B, L) a CQMS and α : B → B ⊗ C(G) a coaction. Then the following are equivalent: (i) α is isometric in the sense of (2) (ii) α is 1-isometric (Definition 3.1) (iii) α is full-isometric (Definition 3.3). 7 Proof. • (iii) implies (ii): straightforward. • (ii) implies (i). For s ∈ G, we write ωs for the state in C(G) defined in (3). Then we use αs to denote the mapping αs : B → B : b 7→ α(b)(s) = (ι ⊗ ωs)α(b). We fix elements b ∈ B and s ∈ G. By (ii) we know that L(αs(b)) ≤ L(b). On the other hand we also have L(b) = L(αe(b)) = L(αs−1(αs(b))) ≤ L(αs(b)). Both inequalities together prove (i). • (i) implies (iii). Choose faithful QTGs (Ci, ∆i, βi) for i = 1, · · · , n such that β = β1 ∗ · · · ∗ βn is 1-isometric. We will write C for C1 ⊗ · · · ⊗ Cn. To prove that α ∗ β is 1-isometric, we fix an element b ∈ B and a pure state ω on C(G) ⊗ C. Since C(G) is commutative, we can use theorem 4.14 of [15] to write ω = ωs ⊗ ϕ for an element s ∈ G and a pure state ϕ on C. But then, by (i), we have L((ι ⊗ ω)(α ∗ β)(b)) = L(αs((ι ⊗ ϕ)β(b))) = L((ι ⊗ ϕ)β(b)) ≤ L(b). Lemma 3.4 shows that this inequality holds for all states on C(G) ⊗ C since it holds for all pure states. (cid:4) More specifically, this theorem implies that in the double-classical case, where A = C(G) for a compact group G and B = C(X) for a metric space (X, d), the definitions 3.1 and 3.3 are equivalent to the usual notion of an isometric group action. There is another special case of our setting that has been studied before, namely the case of a compact quantum group acting on a finite classical metric space. This has been studied by Banica in [1]. The setting is as follows. Let α : C(X) → C(X) ⊗ A be a coaction of a CQG (A, ∆) on a finite metric space (X, d) with n points. A coaction on any finite space X with n points is completely described by an (n × n) matrix a = (aij)i,j=1,...,n over A. Indeed, consider the standard basis {δ1, . . . , δn} of C(X), where δi is the function that is one in the i-th point of X, and zero elsewhere. The coaction α is determined by the elements α(δj ) for j ∈ {1, . . . , n}. As the element α(δj ) belongs to C(X) ⊗ A we can write α(δj ) = n Xi=1 δi ⊗ aij for some elements aij in A. The properties of α being a coaction translate into properties of the elements aij (see [16]): the matrix a = (aij) has to 8 be a so called magic biunitary, which means that its rows and columns are partitions of the unity of A with projections, or, explicitly, for all i, j ∈ {1, · · · , n}, we have: ij = aij = a2 a∗ ij, n Xk=1 aik = 1, n Xk=1 akj = 1. (5) The comultiplication on the elements aij is given by Note that (5) implies that ∆(aij) = n Xk=1 aik ⊗ akj. aijaik = 0 = ajiaki for all i, j, k ∈ {1, · · · , n} with j 6= k. Moreover, if the coaction α is faithful, i.e. if A is generated by {aij i, j = 1, . . . , n}, then (A, ∆) is of Kac type with bounded antipode S given by S(aij) = aji. Now suppose d is a metric on X. Consider the (n×n)-matrix (d(i, j))i,j=1,...,n which we also denote by d. In [1] Banica calls the coaction α of A on X isometric if the matrices a and d commute: ad = da. (6) Does Banica's notion of isometric coactions coincide with the notions in- troduced in definitions 3.1 and 3.3 for the general 'double-quantum' setting? This question is answered in the following theorem, which we will prove in the next section. Theorem 3.6. Let (X, d) be a finite metric space with n points and (A, ∆) a CQG acting faithfully on X by the coaction α : C(X) → C(X) ⊗ A. i=1 δi ⊗ aij where δj is 1 on the j-th point of X and 0 elsewhere. Write Take elements aij in A such that α(δj ) = Pn • Ld(f ) = supn f (x)−f (y) • a for the (n × n)-matrix (aij)i,j=1···n (cid:12)(cid:12)(cid:12) d(x,y) x, y ∈ X, x 6= yo for f ∈ C(X) • d for the (n × n)-matrix (d(i, j))i,j=1···n. Then the following are equivalent: (a) ad = da (b) α is 1-isometric (Definition 3.1) (c) α is full-isometric (Definition 3.3). 9 4 Proof of Theorem 3.6 4.1 Part 1: (b) implies (a) Before we can start the proof of the first implication of Theorem 3.6, we want to express the commutation of a and d in a different, for our purposes more practical way. This is done in the following lemma. We will need this lemma in the second part of the proof too. Lemma 4.1. Using the notations of Theorem 3.6, the following are equiv- alent: (A) ad = da (B) aijakl = 0 for every i, j, k, l ∈ X with d(i, k) 6= d(j, l). Proof. Suppose (A) holds, so a and d commute. We want to prove (B). Take points i, j, k, l in X such that d(i, k) 6= d(j, l). Because a and d commute, we have n n aixd(x, l) = Xx=1 d(i, y)ayl Xy=1 If we multiply this equality on the left by aij, we get aijd(j, l) = n Xy=1 d(i, y)aij ayl because aixaiy is zero whenever x 6= y. Similarly, by multiplying this by akl on the right, we get aijakld(j, l) = d(i, k)aij akl. Because d(j, l) 6= d(i, k) this equality is only possible if aijakl is zero which is what we wanted to prove. Conversely, suppose (B) holds. We want to prove (A), so we need a and d to commute. But for every i, j in X, we have (ad−da)ij = n Xx=1 aixd(x, j)−d(i, x)axj = n Xx,y=1 d(x, j)aixayj−d(i, y)aixayj = 0 The last equality expresses the given fact that aixayj is zero whenever d(x, j) is different from d(i, y). We also used the fact that the rows and columns of a sum up to the unit element. (cid:4) Now we can start proving that (b) implies (a) in Theorem 3.6. Using the notations of that theorem, we suppose Ld((ι ⊗ ω)α(f )) ≤ Ld(f ) for every f ∈ C(X) and every state ω on A. We want to prove ad = da. Before we give the actual proof, we need another lemma. The formula- tion of that lemma needs some notations. Since we are working in a finite 10 metric space, the set D = {d(i, j) i, j ∈ X} of all distances is a finite set. So we can rearrange the numbers in D, writing D = {d0, d1, · · · , dN } where dk < dl if k < l. For every distance dγ ∈ D, and every point i ∈ X, we can look at the set of all points that are at a distance dγ from i. We denote this set by V γ i = {j ∈ X d(i, j) = dγ}. Lemma 4.2. We use the notations of Theorem 3.6 and assume the inequal- ity in (b). If a state ω on A is one on an element aij, then it is zero on every akl with d(i, k) 6= d(j, l). In other words: for all γ ∈ {0, · · · , N }, all i, j ∈ X and all ω ∈ S(A) with ω(aij) = 1, we have ∀k ∈ V γ ∀k 6∈ V γ i , ∀l 6∈ V γ i , ∀l ∈ V γ j j : ω(akl) = 0, : ω(akl) = 0. Proof. We prove this lemma by induction on γ. Step 1: First of all, we want to check the lemma for γ = 0. We fix points i, j ∈ X and a state ω on A such that ω(aij) = 1. The value γ = 0 corresponds with the smallest distance in the metric space, which is zero. First, we take k ∈ V 0 also take some l 6∈ V 0 i . This means that d(i, k) = d0 = 0, so k = i. We j . (hence l 6= j) and we want to show that ω(akl) is zero. But ω(akl) = ω(ail) ≤ Px6=j ω(aix), because all aix are positive and ω is a positive map. Moreover, Px∈X aix = 1 and since ω(1) = 1, we have ω(akl) ≤ 1 − ω(aij) = 0. Hence ω(akl) = 0, which is the first item we had to prove. Secondly, we take k 6∈ V 0 i , so k 6= i, and l ∈ V 0 j , so l = j. Then we have, by a similar argument as before, that ω(akl) = ω(akj) ≤ Px6=i ω(axj) = 1 − ω(aij) = 0, so ω(akl) = 0. Step 2: We proceed by induction. We suppose that the lemma holds for every value in {0, . . . , γ} for a certain γ. We want to prove that it also holds for γ + 1. Again, we fix points i and j in X, and a state ω on A such that ω(aij) = 1. We take k ∈ V γ+1 , and l 6∈ V γ+1 . i j We are going to use the inequality Ld((ι ⊗ ω)α(f )) ≤ Ld(f ) (7) for the chosen state ω and for f = Dj, where Dj : X → C : x 7→ d(x, j). Us- ing the triangle inequality, it is easy to check that Ld(Dj) = 1. So inequality (7) implies (ι ⊗ ω)α(Dj )(k) − (ι ⊗ ω)α(Dj )(i) d(k, i) ≤ 1. Notice that Dj = Px∈X d(j, x)δx, so, using the formula from Theorem 3.6 for the α(δx), we get Xx∈X d(j, x)ω(akx) − Xx∈X (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) d(j, x)ω(aix)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 11 ≤ d(i, k) = dγ+1 In the first step, we have proven that ω(aix) = 0 whenever x 6= j. We also know that d(j, j) = 0. Those two facts give Xx∈X (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) d(j, x)ω(akx)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ≤ dγ+1. (8) i j j . On the other hand, since k ∈ V γ+1 Because of the induction hypothesis, we have that ω(akx) = 0 if d(j, x) < dγ+1. Indeed, in that case, we can find a value c ∈ {0, · · · , γ} such that d(j, x) = dc and hence x ∈ V c , we know that k 6∈ V c i . As c ≤ γ, it follows from the induction hypothesis that ω(akx) = 0. This means that the left hand side of inequality (8) is a convex combination of those distances d(j, x) that are at least dγ+1, although the convex combination itself is smaller than dγ+1. It is clear that is only possible if the coefficients of the distances d(j, x) > dγ+1 are zero. So, ω(akx) = 0 unless d(j, x) = dγ+1. In particular, since l 6∈ V γ+1 , this implies that ω(akl) = 0. j i We still have to prove the second case, for k 6∈ V γ+1 and l ∈ V γ+1 . In this case, we define a new state ω′ : A → C : a 7→ ω(S(a)) where S is the antipode of (A, ∆). Since ω′(aji) = ω(S(aji)) = ω(aij) = 1, we can use the first case of this lemma on the distance γ + 1, the points j and i in X and the state ω′. Since l ∈ V γ+1 and k 6∈ V γ+1 , the first case states that 0 = ω′(alk) = ω(akl) and this concludes the proof. (cid:4) j i Using the previous lemma, we now need to show the commutation of a and d to complete the proof of the first implication of Theorem 3.6. Choose any i, j, k, l ∈ X with d(i, k) 6= d(j, l). By Lemma 4.1 it suffices to show that aijakl = 0. Suppose that the product aijakl is nonzero. Then also aijaklaij = (aijakl)(aij akl)∗ is nonzero. Hence we can find a state ω on A such that ω(aijaklaij) is nonzero. But then also ω(aij) is nonzero, because of the Cauchy-Schwarz inequality: 0 < ω(aij(aklaij))2 ≤ ω(aij)ω(aijaklaij). Using this state ω, we define a new state ωij : A → C : x 7→ ω(aijxaij) ω(aij) . One can easily check that this is indeed a state and that ωij(aij) = 1. So we can use the first case of the Lemma 4.2 on the number γ for which d(i, k) = dγ, on the points i and j, and on the state ωij, and conclude that ωij(akl) = 0. But this clearly contradicts the fact that ωij(akl) = ω(aij aklaij ) is nonzero because of the choice of ω. Therefore aijakl = 0, and ω(aij ) this concludes the proof of this first implication. 12 4.2 Part 2: (a) implies (b) Since this part of the theorem will not use anything specific about the CQG- structure we have on A (except for the fact that the matrix a is a magic biunitary), we can reformulate this part in a more general setting. Theorem 4.3. Let (X, d) be a finite metric space with n points, A a unital C-algebra and α : C(X) → C(X) ⊗ A a -homomorphism. Take elements aij in A such that α(δj) = Pn i=1 δi ⊗ aij. We use the same notations as in Theorem 3.6 for δi, Ld, a and d. Suppose a is a magic biunitary matrix. Then, if ad = da, it follows that α is 1-isometric. We use the notations from the theorem, and we suppose that the magic biunitary matrix a commutes with d. We want to prove that Ld((ι ⊗ ω)α(f )) ≤ Ld(f ) for every f ∈ C(X) and every state ω on A. First note that it is sufficient to prove the following lemma: Lemma 4.4. Using the notations and assumptions of Theorem 4.3, we fix elements x, y ∈ X and a state ω on A. Then there exist positive numbers λij (i, j ∈ X) such that for all i, j ∈ X, we have Pj∈X λij = ω(axi) Pi∈X λij = ω(ayj) λij = 0 if d(i, j) 6= d(x, y)   Suppose the lemma holds, then we can prove Theorem 4.3. Indeed, we fix a state ω ∈ S(A) and a function f ∈ C(X). Now we can rewrite (ι ⊗ ω)α(f ) = Xi∈X f (i)(ι ⊗ ω)α(δi) = Xi,j∈X f (i)ω(aji)δj Choose elements x, y ∈ X with x 6= y. Using Lemma 4.4, we can write (ι ⊗ ω)α(f )(x) − (ι ⊗ ω)α(f )(y) d(x, y) d(x, y) d(x, y) = (cid:12)(cid:12)Pi∈X f (i)(ω(axi) − ω(ayi))(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)Pi∈X f (i)(cid:16)Pj∈X λij −Pj∈X λji(cid:17)(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)Pi,j∈X λij(f (i) − f (j))(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) f (i) − f (j) λij Ld(f ) d(x, y) λij d(i, j) Xi,j∈X ≤ Xi,j∈X = Ld(f ) 13 Because this holds for all elements x and y in X with x 6= y, we find that Ld((ι ⊗ ω)α(f )) ≤ Ld(f ) for every f ∈ C(X), which would conclude the proof of Theorem 4.3. Of course, we still have to prove Lemma 4.4. This lemma is combinato- rial, and its proof will use a famous theorem in graph theory. To formulate this theorem, we need to introduce some notations. The graph theory we need concerns flow networks. A flow network is a set V of vertices, and a set E ⊆ V × V of directed edges. Every edge (u, v) in E has a positive capacity c(u, v). We also fix two vertices s and t, called the source and the sink. An s-t-flow in such a network is a mapping from the set of edges to the positive numbers, which maps every edge (v, w) ∈ E on f (v, w), the 'flow from v to w'. The conditions are that the flow f (v, w) through an edge is always smaller than the capacity c(v, w) of the edge. Secondly, the flow can not stay inside a vertex (except for the source and the sink). So the flow entering a vertex must equal the flow leaving the vertex, or Pu∈V f (u, v) = Pw∈V f (v, w) for every vertex v ∈ V \{s, t}. In this mathematical notation we suppose that the capacity c(v, w) (and hence also the flow f (v, w)) is zero when there is no edge (v, w) in E. The value F of an s-t-flow is then the sum of all the flow leaving the source Pv∈V f (s, v). This equals the flow entering the sink. An s-t-cut is a partition of the vertices. We cut the graph into two pieces, partitioning the set of vertices in two sets, S and T in such a way that S contains the source s and T contains the sink t. The capacity of such a cut is then the sum of the capacities of all edges crossing the cut from S to T . Mathematically, this is Pv∈S,w∈T c(v, w). Now we can formulate the famous max-flow-min-cut theorem briefly. Theorem 4.5 (Max-flow-min-cut theorem). The maximal value of an s-t- flow is equal to the minimal capacity of an s-t-cut. This theorem will be used to prove the following combinatorial lemma. Lemma 4.6. Suppose we have positive numbers α1, α2, · · · , αn and i=1 βi = 1. Next, suppose that for every i ∈ {1, · · · , n}, there exists a set Vi ⊆ {1, · · · , n} such that for every subset Z of {1, · · · , n} β1, β2, · · · , βn such that Pn i=1 αi = Pn Xi∈Z αi ≤ Xj∈Si∈Z Vi βj. (9) Then, for i, j ∈ {1, · · · , n}, we can take positive numbers λij such that 14 for all i, j ∈ {1, · · · , n} j=1 λij = αi i=1 λij = βj Pn Pn λij = 0 if j 6∈ Vi   Proof. We construct a flow network. We have n vertices on the left hand side, which we label l1, l2, · · · , ln, and n vertices on the right hand side, labeled r1, r2, · · · , rn. For short, we denote L for {l1, · · · , ln} and R for {r1, · · · , rn}. For every i ∈ {1, · · · , n}, we have an edge from the source s to vertex li with capacity αi, and an edge from vertex ri to the sink t with capacity βi. In the center, we have an edge from vertex li to every vertex rj with j ∈ Vi. We give those edges capacity 1. We denote the set of vertices by V and the set of edges by E. From now on, we will see Vi as the subset of all vertices rj with j ∈ Vi. An example for n = 5 can be found on figure 1. s α1 α2 α3 α4 α5 l1 l2 l3 l4 l5 1 1 1 1 r1 r2 r3 r4 r5 β1 β2 β3 β4 β5 t Figure 1: Example of a flow network when n = 5 V1 = {2}, V2 = {1}, V3 = ∅, V4 = {5}, V5 = {4} We want to use the max-flow-min-cut theorem to find a maximal flow in this flow network. For this end, we investigate the minimal capacity of an s-t-cut. An example of an s-t-cut with capacity one is S = {s}, T = V \{s}. We claim that this capacity is minimal. So we want to prove that every other s-t-cut has a capacity of at least one. Let us consider all possible cases. • First we suppose S contains non of the vertices on the left hand side: L ⊆ T . Then the capacity of the cut is Xv∈S w∈T c(v, w) ≥ Xw∈L c(s, w) = n Xi=1 c(s, li) = n Xi=1 αi = 1. 15 • If one of the central edges is crossing the cut from S to T , then the capacity of the cut is bigger than one. This is, if S ∩ L 6= ∅ and T ∩Sli∈S Vi 6= ∅, then we have a vertex rj ∈ T in one of the Vi, with li ∈ S. Then li belongs to S and rj to T , so the capacity of the cut is bigger than c(li, rj), which is 1 because rj ∈ Vi. • The last case is the case where S ∩ L 6= ∅ but T ∩Sli∈S Vi = ∅. This means that Vi ⊆ S when li ∈ S. If we now use the given inequality (9) on the set S ∩ L, we get Xli∈S αi ≤ Xrj ∈Sli∈S Vi βj For the total capacity of the cut, this gives Xv∈S w∈T c(v, w) ≥ Xw∈(T ∩L) c(s, w) + Xv∈(S∩R) c(v, t) = Xli∈T αi + Xrj ∈Sli∈S Vi βj ≥ Xli∈T αi + Xli∈S αi = n Xi=1 αi = 1 We have proven that the minimal capacity of an s-t-cut in the flow network we consider is one. Because of the max-flow-min-cut theorem, this implies that we can have a maximal flow f of value one through this network. If we denote the flow in the edge (li, rj) by λij, we have found the positive numbers we were looking for: • It is obvious that all λij = f (li, rj) are positive. • For the first condition, we have to calculate Pn j=1 f (li, rj). This is the total amount of flow leaving the vertex li, so this has to equal the amount of flow entering li. The only edge through which the flow can enter li, is the edge (s, li), which has capacity αi. In order to have a total flow of value one, the flow in every edge (s, lk) must k=1 αk = 1). Hence, in particular, the flow through the edge (s, li) equals its capacity αi. So we have proven j=1 λij = Pn reach its capacity αk (since Pn that Pn • For the second condition, we calculate Pn j=1 λij = f (s, li) = αi. i=1 f (li, rj) simi- larly, we see that this is the total amount of flow entering the vertex rj. This is equal to the amount of flow leaving rj, which is the flow through the edge (rj, t). With a similar reasoning as for the first condition, we see that the flow f (rj, t) equals the capacity βj of the edge. i=1 λij = Pn • If j 6∈ Vi, there is no edge from li to rj. Clearly, there cannot be any (cid:4) flow going from li to rj, so λij = f (li, rj) is zero. 16 We want to use this purely combinatorial lemma to prove Lemma 4.4. Proof of Lemma 4.4: Fix elements x and y in X and a state ω on A. It is obvious that we want to use Lemma 4.6 for the numbers αi = ω(axi) and βi = ω(ayi). To make the third condition of Lemma 4.6 match with the third condition of Lemma 4.4, we take Vi to be the set {j ∈ X d(i, j) = d(x, y)}. For example, if the metric space looks like figure 2(a) (where the numbers represent the distances between the points), and we chose x to be c1 and y to be c2, then the flow network used in the proof of Lemma 4.6 will look like the network in figure 2(b) (which we also used as an example on figure 1 in the proof of Lemma 4.6). c3 1 c4 3 1 1 2 2 3 c2 2 c5 2 2 (a) l1 ω(a11) l2 ω(a12) c1 s ω(a13) l3 ω(a14) ω(a14) l4 l5 r1 r2 ω(a21) ω(a22) r3 ω(a23) t ω(a24) r4 ω(a25) r5 1 1 1 1 (b) Figure 2: An example of a metric space and the corresponding flow network In order to use Lemma 4.6, we need to check the inequality ω(axi) ≤ Xi∈Z Xj∈X : ∃i∈Z: d(i,j)=d(x,y) ω(ayj) for every subset Z ⊆ X. But this follows from the fact that a and d commute. Indeed, by using the commutation relation in the way we obtained in Lemma 4.1, we get for Z ⊆ X Xi∈Z axi = Xi∈Z axi Xj∈X ayj = Xi∈Z axi Xj∈X : ∃i∈Z: d(i,j)=d(x,y) ayj because axi ayj = 0 whenever d(i, j) 6= d(x, y). Because both factors in the previous product are projections, we have axi ≤ Xi∈Z Xj∈X : ∃i∈Z: d(i,j)=d(x,y) ayj. Since ω is a positive map, we immediately get the desired inequality. Hence we can apply Lemma 4.6 to obtain the desired positive numbers λij. (cid:4) 17 4.3 Part 3: Equivalence of (b) with (c) Theorem 3.6 also states that, in case of a finite classical space, the notions 1-isometric (Definition 3.1) and full-isometric (Definition 3.3) coincide. We already know that (c) implies (b). To prove that (b) implies (c), we need another lemma. Lemma 4.7. Let (B, L) be a CQMS and suppose we have CQGs (Ci, ∆i) with bounded counits ǫi and coactions βi : B → B ⊗ Ci (i = 1, · · · , m). If β := β1 ∗ · · · ∗ βm : B → B ⊗ C1 ⊗ · · · ⊗ Cm is 1-isometric, then βi is 1-isometric for every i ∈ {1, · · · , m}. Proof. We fix i ∈ {1, · · · , m}. Since (ι⊗ǫj)βj = idB for every j ∈ {1, · · · , m}, one can compute that (ι ⊗ ǫ1 ⊗ · · · ⊗ ǫi−1 ⊗ ι ⊗ ǫi+1 ⊗ · · · ⊗ ǫn)β = βi. Now we can easily prove that βi is 1-isometric if β is 1-isometric: we first choose b ∈ B and ω ∈ S(Ci). Then, because of the previous formula for βi, we have L((ι ⊗ ω)βi(b)) = L((ι ⊗ ǫ1 ⊗ · · · ⊗ ǫi−1 ⊗ ω ⊗ ǫi+1 ⊗ · · · ⊗ ǫm)β(b)) ≤ L(b) since ǫ1 ⊗ · · · ⊗ ǫi−1 ⊗ ω ⊗ ǫi+1 ⊗ · · · ⊗ ǫn is a state on C1 ⊗ · · · ⊗ Cm. (cid:4) Proof of (b) implies (c) in Theorem 3.6. We suppose that the coaction α of the CQG (A, ∆) on (C(X), Ld) is 1-isometric, which means that the matrix a = (aij) commutes with the distance matrix. To prove that α is full- isometric, we first choose faithful QTGs (Ci, ∆i, βi) for i = 1, · · · , m such that β = β1 ∗ · · · ∗ βm is 1-isometric. We want to prove that α ∗ β is 1- isometric. We denote X for A ⊗ C1 ⊗ · · · ⊗ Cn, and we use xij for the matrix elements in X such that, for every j ∈ {1, · · · , n} we have (α ∗ β)(δj ) = n Xk=1 δi ⊗ xij. Then, by Theorem 4.3, it is sufficient to prove that the matrix x = (xij)ij is a magic biunitary that commutes with the distance-matrix d. We need to compute the elements xij. For every k ∈ {1, · · · , m} we can take elements c(k) for every ij j ∈ {1, · · · , n}. We can also take elements aij in A such that for every j ∈ {1, · · · , n} we have α(δj ) = Pn i=1 δi ⊗ aij. Then one can check that for in Ck such that βk(δj ) = Pn every i, j ∈ {1, · · · , n}, we have i=1 δi ⊗ c(k) ij xij = n Xk1,··· ,km=1 aik1 ⊗ c(1) k1k2 ⊗ c(2) k2k3 ⊗ · · · ⊗ c(m−1) km−1km ⊗ c(m) kmj (10) 18 Now we can check whether the elements xij satisfy all the conditions needed to apply Theorem 4.3. Firstly, since α and all the βk are coactions, we are magic know that the matrix a = (aij)ij and all matrices c(k) = (cid:16)c(k) ij (cid:17)ij biunitary matrices. From this, one can easily see that also x will be a magic biunitary. Secondly, since α is given to be 1-isometric, we know that a commutes with d. We also supposed β to be 1-isometric, which by Lemma 4.7 implies that every coaction βk is 1-isometric. Notice that the conditions of Lemma 4.7 are met here since all Ck have bounded counits as they are faithfully acting on C(X) [16]. Hence all matrices c(k) commute with d. Then, using (10), it is straightforward to verify that d commutes with x as well. (cid:4) 5 Subgroup acting isometrically Classically, when a group G is acting on a metric space (X, d), the subset H consisting of those elements g of G for which the action X → X : x 7→ gx is isometric forms a subgroup of G. We show in Theorem 5.2 that this remains true in the general quantum context for full-isometric coactions. Before stating the theorem, we need to introduce a definition. Definition 5.1. Let (A, ∆) be a compact quantum group and I ⊆ A a subset of A. We say that I is a Woronowicz Hopf C-ideal if I is a two-sided closed -ideal in A such that (p ⊗ p)∆(I) = {0} where p is the canonical projection A → A/I : a 7→ a + I. In the above situation, A/I becomes a compact quantum group for the comultiplication defined by ∆I : A/I → A/I ⊗ A/I : p(a) 7→ (p ⊗ p)∆(a). The CQG (A/I, ∆I ) can be considered as a compact quantum subgroup of (A, ∆). Theorem 5.2. If α is a faithful coaction of a CQG (A, ∆) with bounded counit ǫ on a CQMS (B, L), then there exists a proper Woronowicz Hopf C-ideal I in A such that αI : B → B ⊗ A/I : b 7→ (ι ⊗ pI )α(b) is a faithful and full-isometric coaction, where pI is the canonical projection of A onto A/I. The ideal I can be chosen to be the smallest one possible, in the sense that for any other Woronowicz Hopf C-ideal J with αJ = (ι⊗pJ )α a faithful and full-isometric coaction, we have that I ⊆ J. 19 We need a couple of lemmas before we can start the proof of this theorem. Lemma 5.3. Let S be a set of states on a unital C-algebra A such that (i) S separates the points of A (i.e. ∀a ∈ A\{0}, ∃ω ∈ S : ω(a) 6= 0) (ii) For all a ∈ A and all ω ∈ S with ω(a∗a) 6= 0, the state ωa belongs to S, where ωa is defined by ωa : A → C : x 7→ ω(a∗xa) ω(a∗a) . Then for every element a of A, a ≥ 0 is equivalent to ω(a) ≥ 0 for all ω in S. Proof. We fix an element a ∈ A. First of all, it is clear that, if a is positive, ω(a) will also be positive for all ω in S, since every state is a positive mapping. Next, we suppose that ω(a) ≥ 0 for all ω in S, and we want to prove that a is positive. First we check that a is self-adjoint. Suppose not, then a − a∗ is nonzero. Since the states in S separate the points of A, there exists an ω ∈ S such that ω(a − a∗) 6= 0. But then ω(a) 6= ω(a∗) = ω(a) which contradicts the fact that ω(a) is a positive (and hence real) number. So we may suppose that a is self-adjoint. Then we can write a as a+ −a− for some positive elements a+ and a− in A with a+a− = 0. To prove that a is positive, we have to show that a− is zero. But if a− were non-zero, then also (a−)3 would be non-zero. Hence we could find a state ω in S such that ω((a−)3) 6= 0. But then also ω((a−)2) 6= 0, since by the Cauchy-Schwarz inequality we have 0 < ω((a−)3)2 ≤ ω((a−)2)ω((a−)4). By the second property of the set S, we now know that ωa− belongs to S, which implies that ωa−(a) ≥ 0. But ωa−(a) = ω(a−aa−) ω((a−)2) = ω(a−a+a− − (a−)3) ω((a−)2) = − ω((a−)3) ω((a−)2) and this would be strictly negative, since both ω((a−)3) and ω((a−)2) are supposed to be positive and non-zero. We get an obvious contradiction and conclude that a = a+ is positive. (cid:4) Lemma 5.4. Let S be a set of states on a unital C-algebra A such that for every element a ∈ A we have a ≥ 0 iff ω(a) ≥ 0 for all ω ∈ S. Then the convex hull of S is weakly -dense in the state space S(A). Proof. Suppose the lemma does not hold. Then we can take a state ω in S(A)\conv(S). As conv(S) is weakly -compact, there exists, by the Hahn-Banach separation theorem, a linear, weakly -continuous mapping ψ : A∗ → C and a real number λ such that Re(ψ(ϕ)) ≤ λ < Re(ψ(ω)) 20 for all states ϕ in conv(S). But since ψ is linear and weakly -continuous, there must exist an element a in A such that ψ(ϕ) = ϕ(a) for all ϕ in A∗. Now the above equality reads Re(ϕ(a)) ≤ λ < Re(ω(a)) for all ϕ in conv(S). Since ω and all ϕ ∈ conv(S) are self-adjoint, we can use b to denote Re(a) and then we get ϕ(b) ≤ λ < ω(b) for all ϕ in conv(S). So, for the element λ1 − b (where 1 is the unit element of A), we get ϕ(λ1 − b) = λ − ϕ(b) ≥ 0 for all ϕ ∈ conv(S), and hence in particular for all ϕ in S. By the assumption in the lemma, this would mean that λ1−b is a positive element of A. But since ω is a state, this would imply that ω(λ1 − b) is positive, which gives a contradiction with ω(b) > λ. (cid:4) Lemma 5.5. Let A be a unital C-algebra and Π be a set of -homomorphisms π : A → Cπ where the Cπ are unital C-algebras. We suppose Π separates the points of A, so if a ∈ A is non-zero, then there exists an element π ∈ Π such that π(a) 6= 0. Then the convex hull of the set S = {ω ◦ π π ∈ Π, ω ∈ S(Cπ)} is weakly -dense in S(A). Proof. We will use the above two lemma's to prove this result. We verify both conditions for Lemma 5.3. Firstly, we choose a non-zero element a ∈ A. Since Π separates the points of A, there is a π ∈ Π : A → Cπ such that π(a) is non-zero. But then it is clear that there also exists a state ω on Cπ such that ω(π(a)) 6= 0. Hence we have found a state ω ◦ π in S that is non-zero on a, which shows that S separates the points of A. Secondly, we take an element a ∈ A and an arbitrary state ϕ in S. Then there exists a π ∈ Π : A → Cπ and a state ω on Cπ such that ϕ = ω◦π. Since π is a -homomorphism, one can compute that ϕa = ωπ(a) ◦ π if ϕ(a∗a) 6= 0. This clearly implies that ϕa belongs to S, since ωπ(a) is still a state on Cπ. So, from Lemma 5.3, we may conclude that an element a ∈ A is positive if and only if ϕ(a) ≥ 0 for all ϕ ∈ S. But then Lemma 5.4 tells us that the convex hull of S is weakly -dense in the state space S(A). (cid:4) Now we have all the results needed to prove the main Theorem 5.2 of this section. To be able to construct the desired Woronowicz Hopf C-ideal, we introduce some notations. Definition 5.6. Let B be a unital C-algebra and take QTGs (A, ∆, α) and (Ci, ∆i, αi) for i = 1, · · · , n. We say that π : A → C1 ⊗ · · · ⊗ Cn is a multi- morphism of QTGs if π is of the form π1 ∗ · · · ∗ πn for some morphisms of 21 QTGs πi : A → Ci (i = 1, · · · , n). We used the notation π1 ∗ π2 to denote the -homomorphism (π1 ⊗ π2)∆ : A → C1 ⊗ C2, and inductively one can define π1 ∗ · · · ∗ πn. Definition 5.7. Let (B, L) be a CQMS and (A, ∆, α) a QTG of B. Let πi : (A, ∆, α) → (Cπi, ∆πi, απi) be morphisms of QTGs for i = 1, · · · , n. We say that the multi-morphism of QTGs π = π1 ∗ · · · ∗ πn is admissible if every coaction απi is faithful and full-isometric (i = 1, · · · , n). We denote the set of all admissible multi-morphisms of QTGs on (A, ∆, α) by ΠA. Proof of Theorem 5.2: We claim that the desired Woronowicz Hopf C- ideal will be I = \π∈ΠA ker(π). For the first part of the theorem, we have to prove that the set I is a Woronowicz Hopf C-ideal. It is easy to see that I is a closed two-sided - ideal, since every kernel of a -homomorphism is a closed -ideal. We want to show that (pI ⊗ pI)∆(I) = {0}. To do so, it is sufficient to prove that (π1⊗π2)∆(I) = {0} for every π1, π2 ∈ ΠA. So we choose π1 = π(1) 1 ∗· · ·∗π(1) n : A → C (1) 1 ⊗ · · · ⊗ C (1) m in ΠA. This means that (C (j) i ) are faithful QTGs such that the coactions α(j) are morphisms of QTGs. i n and π2 = π(2) 1 ⊗ · · · ⊗ C (2) 1 ∗ · · · ∗ π(2) m : A → C (2) , ∆(j) , α(j) i i are full-isometric and π(j) But then obviously i π1 ∗ π2 = π(1) 1 ∗ · · · ∗ π(1) n ∗ π(2) 1 ∗ · · · ∗ π(2) m is still a multi-morphism of QTGs, and since all α(j) are faithful and full- isometric, we know that π1 ∗ π2 is admissible. Hence (π1 ⊗ π2)∆(I) = (π1 ∗ π2)(I) = {0}, so I is a Worononwicz Hopf C-ideal. i Notice that the counit ǫ : A → C is an admissible morphism of QTGs. Therefore I is a proper Woronowicz Hopf C-ideal. Hence it makes sense to consider the CQG A/I with comultiplication ∆I : A/I → A/I ⊗ A/I : a + I 7→ (pI ⊗ pI )∆(a) and the coaction αI of (A/I, ∆I ) on B, as defined in the theorem. Since α is faithful, it is clear that also the induced coaction αI = (ι⊗pI)α is faithful. To prove that αI is full-isometric, we first choose faithful QTGs (Ci, ∆i, βi) for i = 1, · · · , n such that β = β1 ∗ · · · ∗ βn is 1-isometric. Denote C1 ⊗ · · · ⊗ Cn by C. We want to prove that αI ∗ β : B → B ⊗ A/I ⊗ C is 1-isometric, using Lemma 3.4. We choose a multi-morphism π ∈ ΠA. Then we can write π = π1∗· · ·∗πm : (A, ∆, α) → (Cπi, ∆πi, απi) of QTGs. We use for some morphisms πi 22 Cπ to denote Cπ1 ⊗ · · · ⊗ Cπm. We introduce some more notations: for i ∈ {1, · · · , m} we write ¯πi for the unique ∗-homomorphism from A/I to Cπi such that ¯πi ◦ pI = πi. Remark that this is well-defined since πi(I) = {0}. We denote ¯π1 ∗ · · · ∗ ¯πm by ¯π, where the ∗-product is defined using the comultiplication ∆I. We will also write α(j) for α ∗ · · · ∗ α, the ∗-product of j factors α (hence α(j) is a mapping from B to B ⊗ A⊗j). The symbol ∆(j) denotes (∆ ⊗ ι ⊗ · · · ⊗ ι) · · · (∆ ⊗ ι)∆, where we have j factors in the composition (hence ∆(j) is a mapping from A to A⊗(j+1)). Notice that for any state ω on Cπ ⊗ C, the mapping ω(¯π ⊗ ι) is a state on A/I ⊗ C. Moreover, for any b ∈ B, we have L((ι ⊗ ω(¯π ⊗ ι))(αI ∗ β)(b)) = L((ι ⊗ ω)(ι ⊗ ¯π ⊗ ι)(ι ⊗ pI ⊗ ι)(α ⊗ ι)β(b)) = L((ι ⊗ ω)(ι ⊗ π ⊗ ι)(α ⊗ ι)β(b)) = L((ι ⊗ ω)(ι ⊗ π1 ⊗ · · · ⊗ πm ⊗ ι)(ι ⊗ ∆(m−1) ⊗ ι)(α ⊗ ι)β(b)) = L((ι ⊗ ω)(ι ⊗ π1 ⊗ · · · ⊗ πm ⊗ ι)(α(m) ⊗ ι)β(b)) = L((ι ⊗ ω)(απ1 ∗ · · · ∗ απm ⊗ ι)β(b)) = L((ι ⊗ ω)(απ1 ∗ · · · ∗ απm ∗ β)(b)) Now, since π is admissible, we know that every coaction απi is faithful and full-isometric. In particular απm is full-isometric, so, because of the choice of β, we now know that απm ∗ β is 1-isometric. Since απm is faithful and απm−1 is full-isometric, this implies that απm−1 ∗ απm ∗ β is 1-isometric. We can continue this argument to conclude that απ1 ∗ · · · ∗ απm ∗ β is 1- isometric. Together with the above calculation, it follows that L((ι ⊗ ω(¯π ⊗ ι))(αI ∗ β)(b)) ≤ L(b) for any element b ∈ B, any π : A → Cπ in ΠA and any state ω on Cπ ⊗ C. By Lemma 3.4, we need to show that the convex hull of the set S = {ω(¯π ⊗ ι) π ∈ ΠA : A → Cπ, ω ∈ S(Cπ ⊗ C)} is weakly -dense in the set of states on A/I ⊗ C, to conclude that αI ∗ β is 1-isometric. This will follow from applying Lemma 5.5 to the C-algebra A/I ⊗ C and the set Π of -homomorphisms ¯π ⊗ ι : A/I ⊗ C → Cπ ⊗ C with π ∈ ΠA. It is sufficient to prove that this set of -homomorphisms separates the points of A/I ⊗ C. We choose an element x in A/I ⊗ C and suppose x is nonzero. Then, since the product states separate the points of the minimal tensor product, there exists a state ψ on C such that (ι ⊗ ψ)(x) is non-zero. We can take an element y in A such that pI (y) = (ι ⊗ ψ)(x). Since this element is non-zero in A/I, we know that y 6∈ I. Hence, there exists a morphism π in ΠA such that π(y) is non- zero. But then ¯π(ι ⊗ ψ)(x) = ¯π(pI (y)) = π(y) is non-zero in Cπ. So we may conclude that (¯π ⊗ ι)(x) is non-zero, which proves that Π separates the points of A/I ⊗C. This finishes the proof of the fact that αI is full-isometric. 23 For the third part of the theorem, let J be a second Woronowicz Hopf C-ideal in A, such that the coaction αJ : B → B ⊗ A/J : b 7→ (ι ⊗ pJ )α(b) is faithful and full-isometric. We denoted the canonical projection of A onto A/J by pJ . We want to prove that I ⊆ J. Since J is a Woronowicz Hopf C-ideal, we know that A/J is a compact quantum group with comultiplication ∆J : A/J → A/J ⊗ A/J : pJ (a) 7→ (pJ ⊗ pJ )∆(a). Hence, by definition, the mapping pJ is a morphism of QTGs from (A, ∆, α) to (A/J, ∆J , αJ ). Moreover, since αJ is a faithful and full-isometric coaction, we know that pJ is an admissible morphism of QTGs, and hence pJ belongs to ΠA. But then, by construction, I is a subset of the kernel of pJ , which is exactly J. (cid:4) 6 Existence of isometry groups? The ultimate goal of this theory would be to find something like a quantum isometry group of a compact quantum metric space (B, L). This quantum isometry group would have to be a universal object in the category of all faithful quantum transformation groups of B that preserve the extra struc- ture L. We already know in what way the structure should be preserved: Definition 6.1. Let (B, L) be a compact quantum metric space. We call a quantum transformation group (A, ∆, α) of B full-isometric if α is a full-isometric coaction. For a CQMS (B, L), we will be considering the category C(B, L) of all faithful and full-isometric transformation groups (A, ∆, α) of B. The mor- phisms are morphisms of transformation groups, as defined in Definition 2.5 It is also possible that B has some additional structure, on top of the Lipnorm. For example, we can consider a functional ψ on B that should be preserved by the coaction. To this purpose, we introduce a second category. For a CQMS (B, L) and a functional ψ on B, we consider the category C(B, L, ψ) of all faithful and full-isometric transformation groups of the pair (B, ψ) as defined in Definition 2.5. The morphisms are still the morphisms of transformation groups. Definition 6.2. The quantum isometry group of a CQMS (B, L) is the universal object in the category C(B, L), if such an object exists. Hence it is an object (A, ∆, α) in C(B, L) such that, for every object ( A, ∆, α) in 24 C(B, L), there is a unique morphism of transformation groups from (A, ∆, α) to ( A, ∆, α). If ψ is a functional on B, we define the quantum isometry group of (B, L, ψ) to be the universal object in C(B, L, ψ) if such an object exists. Obviously it is not immediately clear under what conditions such a quan- tum isometry group exists, but Theorem 5.2 is certainly a step in the right direction. This theorem allows us to define the quantum isometry group if we have a 'quantum permutation group' at our disposition. For example, for finite compact quantum metric spaces, we can define a quantum isometry group fixing the trace (or another functional), since S. Wang defined the quantum automorphism group [16],[17]. Theorem 6.3. Let (B, L) be a finite compact quantum metric space and let ψ be a functional on B. Then there exists a quantum isometry group of (B, L, ψ). Proof. Denote by (A, ∆, α) the quantum automorphism group of (B, ψ), according to Wang's definition [16]. This means that (A, ∆, α) is a universal object in the category C(B, ψ) of all faithful transformation groups of the pair (B, ψ). Note that (A, ∆) is of Kac type, and hence has a bounded counit. Then we claim that (A/I, ∆I , αI ) is the desired quantum isometry group of (B, ψ), where I is the Woronowicz Hopf C-ideal as in (the proof of) Theorem 5.2. That is, I is the intersection of all kernels of admissible multi- morphisms of QTGs on (A, ∆, α). As before, we used ∆I to denote the induced comultiplication of A/I, and αI to denote the induced coaction (ι ⊗ pI)α. From Theorem 5.2 we already know that (A/I, ∆I , αI ) is a faithful and full-isometric quantum transformation group. To prove the universality, let ( A, ∆, α) be a second faithful and full-isometric quantum transformation group of (B, ψ). We want to show that there exists a unique morphism of quantum transformation groups from (A/I, ∆I , αI ) to ( A, ∆, α). Because of the universality of the object (A, ∆, α) in the category C(B, ψ), we know that there is a unique morphism of QTGs θ : A → A from (A, ∆, α) to ( A, ∆, α). Since θ is a morphism of QTGs, we know that αθ = (ι ⊗ θ)α equals α, hence αθ is full-isometric. But then, by definition, θ is admissible. It is now clear that θ(I) = {0}, by the construction of I. We have verified that the mapping ¯θ : A/I → A : (a + I) 7→ θ(a). (11) is well defined. Remark that ¯θpI = θ. We claim that ¯θ is the desired mor- phism of quantum transformation groups from (A/I, ∆I , αI ) to ( A, ∆, α). 25 • It is clear that ¯θ is a unital -homomorphism since θ is a unital - homomorphism. • We have (¯θ ⊗ ¯θ)∆I pI = (¯θ ⊗ ¯θ)(pI ⊗ pI)∆ = (θ ⊗ θ)∆ = ∆θ = ∆¯θpI which means that the mappings (¯θ ⊗ ¯θ)∆I and ∆¯θ are equal on A/I. • We also have (ι ⊗ ¯θ)αI = (ι ⊗ ¯θ)(ι ⊗ pI )α = (ι ⊗ θ)α = α. To conclude the proof, we have to check the uniqueness of ¯θ. But if we take any morphism of quantum transformation groups σ from (A/I, ∆I , αI ) to ( A, ∆, α), then σpI would be a morphism of quantum transformation groups from (A, ∆, α) to ( A, ∆, α). Because of the uniqueness of θ, this implies that σpI = θ. But then of course σ = ¯θ, which proves the uniqueness of ¯θ. (cid:4) The previous theorem shows that every finite quantum metric space had an isometry group (fixing a state). We compute this isometry group in a 'small' example, where the isometry group is non-classical: the quantum space M2(C) ⊕ C ⊕ C. One can check that the quantum symmetry group of Wang (preserving the trace) reduces to the C-algebra A generated by four generators x, y, z, p with relations x2 = −yz 2xx∗ + yy∗ + zz∗ = 1 p∗ = p = p2   and such that the -algebra generated by x, y, z is commutative. The comultiplication is defined by ∆(x) = (zz∗ − yy∗) ⊗ x + x ⊗ z + x∗ ⊗ y ∆(y) = (yx∗ − xz∗) ⊗ 2x + y ⊗ z + z∗ ⊗ y ∆(z) = (xy∗ − zx∗) ⊗ 2x + z ⊗ z + y∗ ⊗ y ∆(p) = p ⊗ p + (1 − p) ⊗ (1 − p). We will write the elements of M2(C) ⊕ C ⊕ C as triples. We use the notation eij (i, j = 1, 2) for the matrix units in M2(C). Then the coaction of A on M2(C) ⊕ C ⊕ C is the unital -homomorphism defined by α(e12, 0, 0) = (e11 − e22, 0, 0) ⊗ x + (e12, 0, 0) ⊗ z + (e21, 0, 0) ⊗ y α(0, 1, 0) = (0, 1, 0) ⊗ p + (0, 0, 1) ⊗ (1 − p). There are several Lipnorms that make this space into a CQMS. The Lipnorm L we will consider here will be the following: L((cid:18) a b c d (cid:19) , e, f ) = a − d + b + c + a − e + a − f for all a, b, c, d, e, f ∈ C. 26 For this Lipnorm, one can prove that a representation π of A is admissible if and only if π(x) = π(y) = 0. Hence, to find the quantum isometry group, we take the quotient of A by the ideal generated by x and y. We find that the quantum isometry group of M2(C) ⊕ C ⊕ C is the universal C-algebra generated by a unitary element z and a projection p. The comultiplication is given by ∆(z) = z ⊗ z ∆(p) = p ⊗ p + (1 − p) ⊗ (1 − p) and the coaction α is given by α(e12, 0, 0) = (e12, 0, 0) ⊗ z α(0, 1, 0) = (0, 1, 0) ⊗ p + (0, 0, 1) ⊗ (1 − p). References [1] Teodor Banica. Quantum automorphism groups of small metric spaces. Pacific J. Math., 219:27–51, 2005. [2] Teodor Banica and Debashish Goswami. Quantum isometries and non- commutative spheres. Comm. Math. Phys., to appear, 2009. [3] Jyotishman Bhowmick and Debashish Goswami. Quantum group J. Func. Anal., of orientation-preserving Riemannian isometries. 257(8):2530–2572, October 2009. [4] Jyotishman Bhowmick and Debashish Goswami. Quantum isome- try groups: Examples and computations. Commun. Math. Phys., 285(2):421 – 444, January 2009. [5] Jyotishman Bhowmick and Debashish Goswami. Quantum isometry groups of the podles spheres. J. Funct. Anal., 258(9):2937 – 2960, May 2010. [6] Jyotishman Bhowmick and Adam Skalski. Quantum isometry groups of noncommutative manifolds associated to group c-algebras. J. Geom. Phys., to appear, 2010. [7] Alain Connes. 87.compact metric spaces, fredholm modules, and hy- perfiniteness. Ergodic Theory Dynamical Systems, 9(2):207–220, 1989. [8] Debashish Goswami. Quantum group of isometries in classical and non- commutative geometry. Commun. Math. Phys., 285(1):141–160, Jan- uary 2009. [9] Debashish Goswami. Quantum isometry group for spectral triples with real structure. SIGMA, 6(007), 2010. 27 [10] Hanfeng Li. Compact quantum metric spaces and ergodic actions of compact quantum groups. J. Funct. Anal., 256(10):3368–3408, May 2009. [11] Ann Maes and Alfons Van Daele. Notes on compact quantum groups. Nieuw Arch. Wisk. (4), 16(1-2):73–112, 1998. [12] Marc A. Rieffel. Metrics on state spaces. Doc. Math., 4:559–600, 1999. [13] Marc A. Rieffel. Gromov-hausdorff distance for quantum metric spaces. Mem. Amer. Math. Soc., 168(796):1–65, March 2004. [14] Marc A. Rieffel. Compact quantum metric spaces. In R. S. Doran and R. V. Kadison, editors, Operator algebras, quantization, and noncom- mutative geometry, pages 315–330. Contemp. Math., 365, Amer. Math. Soc., Providence, RI, 2004. [15] Masamichi Takesaki. Theory of Operator Algebras I, volume 124 of Encyclopaedia of Mathematical Sciences, Operator Algebras and Non- Commutative Geometry. Springer, 2nd edition, 2002. [16] Shuzhou Wang. Quantum symmetry groups of finite spaces. Commun. Math. Phys., 195(1):195–211, July 1998. [17] Shuzhou Wang. Ergodic actions of universal quantum groups on oper- ator algebras. Commun. Math. Phys., 203(2):481–498, June 1999. [18] S.L. Woronowicz. Compact quantum groups. In Sym´etries Quantiques (Les Houches, 1995), pages 845–884, Amsterdam, 1998. North-Holland. 28
1006.4420	1	1006	2010-06-23T05:36:59	Connes-Landi Deformation of Spectral Triples	[ "math.OA" ]	We describe a way to deform spectral triples with a 2-torus action and a real deformation parameter, motivated by deformation of manifolds after Connes-Landi. Such deformations are shown to have naturally isomorphic $K$-theoretic invariants independent of the deformation parameter.	math.OA	math	CONNES-LANDI DEFORMATION OF SPECTRAL TRIPLES MAKOTO YAMASHITA Abstract. We describe a way to deform spectral triples with a 2-torus action and a real deformation parameter, motivated by deformation of manifolds after Connes-Landi. Such deformations are shown to have naturally isomorphic K- theoretic invariants independent of the deformation parameter. 1. Introduction Let M be a compact smooth Riemannian manifold endowed with a smooth action of 2-torus T2 = (R/Z)2. Connes and Landi [CL] defined an isospectral deformation Mθ of M for a deformation parameter θ ∈ R/Z. The "smooth function algebra" C∞(Mθ) of Mθ is, as a linear space given by the smooth function algebra C∞(M ) of M , but endowed with a deformed product f ∗θ g = eπiθ(mn′−m′n)f g when f is an T2-eigenvector of weight (m, n) ∈ Z2 in C∞(M ) and g is a one of weight (m′, n′). In the case where M = T2 and the action is the translation, one obtains the noncommutative torus T2 θ whose function algebra is generated by the two unitaries u and v subject to the relation uv = e2πiθvu. They also showed that Dirac type elliptic operator over M determined by the metric structure on M such as the signature operator or the spin Dirac operator continue to make sense over Mθ as unbounded self adjoint operators of compact resolvent over Hilbert spaces on which C∞(Mθ) acts. Li [Li] showed that the spectral triple defined this way qualifies as a "compact quantum metric space" of Rieffel [Rie4]. This construction can be easily generalized to a spectral triple (A, H, D) over a noncommutative algebra endowed with a smooth action of T2 when the "Dirac operotor" D is equivariant with respect to this action. In such a case obtain a new spectral triple (Aθ, H, D) over the deformed algebra exactly as in the smooth Riemannian manifold case. The C∗-algebraic closure of Aθ is a particular case of the deformation considered in [Rie3], where such a deformation is shown to have the same K-group as the original algebra. One question which arises after such deformation is to compute the pairing of the Chern-Connes character of thus obtained spectral triple with the K-group of the deformed algebra. In the case of noncommutative torus T2 θ, Pimsner and Voiculescu showed that it has the same K-theory as the ordinary 2-torus in [PV], then Connes [Con1] [Con2] made the computations of the periodic cyclic cohomology group and the pairing with the K0-group in terms of the connections of projective modules as classified by Rieffel [Rie1]. In this paper we show that 1) Connes-Landi deformation gives isomorphic HP- group as the original algebra, compatible with the K-theory isomorphism (Corollary 13), 2) deformation of a spectral triple give the pairing of the same image (Theorem 3), somewhat generalizing the above computations for T2 θ to the general Mθ. In the course of the proof of the latter invariance of Chern-Connes characters we describe 2000 Mathematics Subject Classification. 58B34;46L87. Key words and phrases. noncommutative geometry, spectral triple, K-theory. 1 2 MAKOTO YAMASHITA the image of the invariant cyclic cocycles under the former isomorphism. In doing so we obtain a simple description of the phenomenon like hτ (θ), K0C(T2 θ)i = Z + θZ for the gauge invariant tracial state τ (θ) on C(T2 θ) which is sensitive to the change of deformation parameter as opposed to Chern-Connes character of equivariant spectral triples. 2. Preliminaries In this section we give a basic definition of Connes-Landi deformation of spectral triples with a 2-torus action and related constructions. Throughout this paper we consider regular spectral triple ([GBVF], [Ren]) with an additional assumption on the smoothness of the torus action. Let (A, H, D) be an even spectral triple (in other words, a K-cycle over A); thus, H = H 0 ⊕ H 1 is a graded Hilbert space, A is a ∗-subalgebra of the algebra of the bounded even operators B(H 0) ⊕ B(H 1) on H and D is an odd unbounded self adjoint on H which has bounded commutator with the elements A: [D, a] is bounded and a(1 + D2)−1/2 is compact for any a ∈ A. Let δ denote the densely defined closed derivation T 7→ [D , T ] on B(H). Recall k=1 dom δk. Let A denote that (A, H, D) is said to be regular when A + [D, A] ⊂ ∩∞ the operator norm closure of A in B(H). The completion Aδ of A, with respect to the seminorms(cid:13)(cid:13)δk(−)(cid:13)(cid:13) and(cid:13)(cid:13)δk([D, −])(cid:13)(cid:13) for k ∈ N = {0, 1, . . .}, is stable under holomorphic functional calculus inside A ([Ren], Proposition 16). Let σ be an action of T2 on A by ∗-automorphisms. In the following we assume that it is strongly continuous with respect to the Fr´echet topology on A, i.e. for any a ∈ A the map t 7→ σt(a) is continuous with respect to the seminorms on A mentioned above. Now, suppose that σt is spatially implemented on H by a strongly continuous even unitary representation Ut : T2 → U(H 0) × U(H 1) on H satisfying (1) σt(a) = AdUt (a), AdUt (D) = D for t ∈ T2 and a ∈ A. This condition implies that σ is isometric with respect to the seminorms on A mentioned above: (2) (cid:13)(cid:13)δkσt(a)(cid:13)(cid:13) =(cid:13)(cid:13)δk(a)(cid:13)(cid:13) , t = U(t,0) and U (2) Put U (1) (cid:13)(cid:13)δk([D, σt(a)])(cid:13)(cid:13) =(cid:13)(cid:13)δk([D, a])(cid:13)(cid:13) . t = U(0,t). For i = 1, 2, let hi denote the generator hiξ = lim t→0 U (1) t ξ − ξ t A. In addition, we write [h, a](α) for the iterated derivation and put σ(i) t = AdU (i) t [h1, · · · , [h1 , [h2, · · · , [h2 , a] · · · ]] · · · ]. α1× {z } α2× {z } Since the derivations [hi, −] are closable and have dense domains on A, σ extends to a strongly continuous action of T2 on A by Theorems 1.4.9 and 1.5.4 of [Bra]. Let A∞ be the subalgebra of Aδ consisting of the elements a such that the map t 7→ σt(a) admit arbitrary order of derivatives in Aδ. The algebra A∞ is stable under holomorphic functional calculus inside A. On the other hand, by the strong continuity of σ and (2), the element σf (a) =ZT2 f (t)σt(a)dt is contained in Aδ for any smooth function f on T2 and any a ∈ Aδ. Thus A∞ is a dense subalgebra of A which contains A and is closed under the holomorphic CONNES-LANDI DEFORMATION OF SPECTRAL TRIPLES 3 functional calculus. Moreover it can be characterized as the joint domain of the following densely defined seminorms on A; (3) νk,α(a) =(cid:13)(cid:13)(cid:13)δk([h, a](α))(cid:13)(cid:13)(cid:13) +(cid:13)(cid:13)(cid:13)δk([D, [h, a](α)])(cid:13)(cid:13)(cid:13) for k ∈ N and α = (α1, α2) ∈ N2. Remark 1. When we define the algebra A∞, the higher derivatives of the functions t 7→ σt(a) for a ∈ A∞ are only required to have their derivatives of arbitrary order in Aδ with respect to the operator norm topology. Then for any aA∞ there exist derivatives of σt(a) with respect to the seminorms(cid:13)(cid:13)δk(a)(cid:13)(cid:13) +(cid:13)(cid:13)δk([D, a])(cid:13)(cid:13) for k ∈ N. Indeed, for i = 1, 2, we know that ∂1σt(a) exists with respect to the operator norm, equals [hi, σt(a)], and it is an element of Aδ. For any k the operator δk([hi, σt(a)]) is bounded and equals [hi, σt(δk(a))]. Now, one has δk(σt(a)) = δk(a) +Z t1 0 δk([h1, σ(1) r (a)])dr +Z t2 0 δk([h2, σ(2) s σ(1) t1 (a)])ds, which shows ∂2δk(σt(a)) = δk([h2, σt(a)]). This shows that the map σt(a) is differ- is [h2, σt(a)]. With an analogous argument one has ∂1σt(a) = [h1, σt(a)] with re- spect to this seminorm. Next, by induction on the order α of differentiation, one obtains that the higher order derivatives of σt(a) exists with respect to the semi- entiable by ∂2 with respet to the seminorm(cid:13)(cid:13)δk(a)(cid:13)(cid:13) and that the partial derivative norm (cid:13)(cid:13)δk(a)(cid:13)(cid:13) and agrees with the ones with respect to the operator norm case. The case for the seminorms(cid:13)(cid:13)δk([D, a])(cid:13)(cid:13) fro k ∈ N is similar. By the stability under the holomorphic functional calculus, the change of algebras from Aδ to A∞ does not affect the K0-group (which is isomorphic to K0(A)) and they have the same K-cycle given by H and D. In the rest of the paper we assume that A = A∞. The Hilbert spaces H i for i = 0, 1 decomposes into a direct sum of eigenspaces (m,n) of weights T2 ≃ Z2 characterized by H i ξ ∈ H i (m,n) ⇔ Utξ = e2πi(mt1+nt2)ξ for t = (t1, t2) ∈ T2. Similarly, for any bounded operator T on H, put (4) e−2πi(mt1+nt2) AdUt (T )dµ(t) T(m,n) =ZT2 where µ is the normalized Haar measure on T2. Then T(m,n) satisfies AdUt (T(m,n)) = e2πi(mt1+nt2)T(m,n) and T can be expressed as a sumP(m,n) T(m,n) which converges Let B(H)fin denote the subspace of B(H) consisting of the operators T where in the strong operator topology. T(m,n) = 0 except for finitely many (m, n). Put Afin = A ∩ B(H)fin. Definition 1. Let θ be an arbitrary real number. Given a bounded operator T on H which is an eigenvector of weight (m, n) ∈ Z2 for the action AdUt of T2, we define a new bounded operator T (θ) on H by T (θ)ξ = eπiθ(mn′−m′n)T ξ for ξ ∈ H(m′,n′). We extend this to the operators in B(H)fin by putting T (θ) = (m,n). P T (θ) Note that we have (T (θ))∗ = (T ∗)(θ). When T is homogeneous of weight (m, n) and S is a one of weight (m′, n′), we have T (θ)S(θ) = eπi(mn′−m′n)θ(T S)(θ). 4 MAKOTO YAMASHITA Remark 2. We adopted a presentation of T (θ) which is slightly different from the one given in [CL]. Let V be the unitary operator on H characterized by V ξ = eπiθm′n′ ξ for ξ ∈ H(m′,n′), and φ be the linear transformation of B(H)fin characterized by φ(T ) = e−πiθmnT when T ∈ B(H)(m,n). Then we have which agrees with the deformation given by Connes and Landi. V φ(T )(θ)V ∗ξ = e2πiθmn′ T ξ, Lemma 1. Let T be a bounded operator on H which admits derivatives of the map t 7→ AdUt (T ) up to the fourth degree in B(H) in the weak operator topology. Then (m,n) is absolutely convergent to a bounded operator T (θ) in the the sumP(m,n)∈Z2 T (θ) operator norm topology. Proof. For i = 1, 2, let ∂i denote the partial differentiation in the direction of ti of functions F (t1, t2) defined on T2. Let (m, n) be any element of Z2. We claim that the bounded operator S = ∂1 AdUt (T )t=(0,0) satisfies Sm,n = mTm,n. Indeed, when ξ ∈ Hm′,n′ and η ∈ Hm′′,n′′, we have hAdU (1) s (T )ξ, ηi − hT ξ, ηi s = e2πi(m′′−m′)s − 1 s hT ξ, ηi. The left hand side converges to hSξ, ηi as s → 0, while the right hand side converges to 2πi(m′′ − m′)hT ξ, ηi. Hence we have that e−2πi(mt1+nt2)hAd Ut(S)ξ, ηi = 2πi(m′′ − m′)e−2πi( mt1+nt2)hT ξ, ηi, where one puts m = −m − m′ + m′′ and n = −n − n′ + n′′. This leads to hSm,nξ, ηi =(2πi(m′′ − m′)hT ξ, ηi 0 (m = m′′ − m′, n = n′′ − n′) (otherwise). This agrees with the value of 2πimhTm,nξ, ηi. Hence the bounded operators Sm,n and 2πimTm,n agree on the linear spans of the Hm′,n′ for (m′, n′) in H. Since this subspace is dense, we have established the claim. Iterating the above argument, we obtain ( (∂2 1 + ∂2 for any (m, n) ∈ Z2. Since the correspondence T 7→ T(m,n) is a contraction, we have 2 )2 AdUt (T )(cid:12)(cid:12)t=(0,0))m,n = 16π4(m2 + n2)2T(m,n) 2 )2 AdUt (T )(cid:12)(cid:12)t=(0,0)(cid:13)(cid:13)(cid:13) . 16π4(m2 + n2)(cid:13)(cid:13)(cid:13) (∂2 (m2 + n2)−2 is summable on Z2 and(cid:13)(cid:13)(cid:13)T (θ) (m,n)(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)T(m,n)(cid:13)(cid:13). (cid:13)(cid:13)T(m,n)(cid:13)(cid:13) ≤ for (m, n) ∈ Z2 \ {(0, 0)} Now, the assertion of Lemma follows from the fact that (cid:3) Lemma 2. The subspace Aθ ⊂ B(H) of the operators a(θ) for a ∈ A is closed under multiplication. 1 1 + ∂2 Proof. Let a and b be arbitrary elements of A. We than have (5) a(θ)b(θ) = X(m,n),(m′,n′)∈Z2 eπiθ(mn′−m′n)am,nbm′,n′ (θ) . In order to show that the series in the the right hand side defines an element of A = A∞, it is enough to show that it is uniformly convergent with respect to any of the seminorms νk,α for k ∈ N and α ∈ N2. CONNES-LANDI DEFORMATION OF SPECTRAL TRIPLES 5 By induction on k and α, we know that there exist numbers Ck,α l,β indexed by k, l ∈ N and α, β ∈ N2 satisfying δk([h, a′b′](α)) = Xl≤k,β≤α Ck,α l,β δl([h, a′](β))δk−l([h, b′](α−β)) for any elements a′, b′ ∈ A. Let us fix k ∈ N and α ∈ N2 now. Then, for any l ≤ k and β ≤ α, the functions σt(δl([h, a](β))) and σt(δl([h, b](β))) admit bounded derivatives in t ∈ T2 of order 4. Hence we have the absolute convergence of X(m,n)∈Z2 δl([h, am,n](β)) by Lemma 1. Hence the infinite series X(m,n),(m′,n′)∈Z2 δl([h, am,n](β))δk−l([h, bm′,n′](α−β)) is also absolutely convergent, which implies the convergence of X(m,n),(m′,n′)∈Z2 eπiθ(mn′−m′n)δl([h, am,n](β))δk−l([h, bm′,n′ ](α−β)). Combining this for all the indices l ≤ k and β ≤ α, one obtains the convergence of the right hand side of (5) with respect to the seminorm δk([h, a′](α)) for a′ ∈ A. With a similar argument we also have the convergence of that series with respect to the seminorm δk([D, [h, a′](α)]), hence also with respect to νk,α. This proves the assertion of Lemma. (cid:3) Definition 2. Let A, H, D, σ and U be as above and θ be an arbitrary real number. The algebra of the operators a(θ) for a ∈ A is called the Connes-Landi deformation Aθ of A. The operator norm closure of Aθ inside B(H) is called the Connes-Landi deformation Aθ of A. Remark 3. Let σ be an action of Rd on a C∗-algebra A and J a skew symmetric matrix of size d. Rieffel [Rie2] defined a deformed product (6) a ×J b =Z σJu(a)σv(b)dudv on the σ-smooth part A∞ of A by means of oscillatory integral. In our setting, the action σ of the 2-torus induces an action of R2. For the 2 × 2 skew symmetric matrix J, consider the following matrix (cf. loc. cit. Example 10.2) 0 (cid:19) . J =(cid:18) 0 − θ θ 2 2 Then the deformed C∗-algebra (A∞, ×J ), which is obtained as the C∗-algebraic closure of (A∞, ×J ), is isomorphic to Aθ. Hence the presentation of the latter on H as in Definition 2 gives a representation of (A∞, ×J ) on the graded Hilbert space H as even operators. Note that the equation (6) defines a product on A under our assumption. The correspondence a 7→ a(θ) gives a representation of (A, ×J ) on H. In the remaining of the section we show that Aθ satisfies the conditions assumed for A. Lemma 3. For any a ∈ A and any θ ∈ R, the operators δk([h, a(θ)](α)) and δk([D, [h, a(θ)](α)]) are bounded for k ∈ N and any α ∈ N2. 6 MAKOTO YAMASHITA of Pm,n δk([h, a](α))(θ) Proof. Applying Lemma 1 to T = δk([h, a](α)), one has the absolute convergence (m,n)](α)) and T 7→ δk([h, T ](α)) is closed on B(H), one has a(θ) ∈ dom δk([h, −](α)) and δk([h, a(θ)](α)) = δk([h, a](α))(θ). (cid:3) (m,n). Since δk([h, a](α))(θ) (m,n) = δk([h, a(θ) Lemma 4. For any a ∈ A and θ ∈ R, the operator a(θ)(1 + D2)−1/2 on H is compact. Proof. The operator a(m,n)(1 + D2)−1/2, and a(θ) (m,n)(1 + D2)−1/2 which is similar to the former, are compact for any pair (m, n) ∈ Z2. Applying Lemma 1 to the operator a(1 + D2)−1/2, we know that a(θ)(1 + D2)−1/2 = lim N→∞ Xm,n<N a(θ) m,n(1 + D2)−1/2 is also compact. (cid:3) Proposition 5. The triple (Aθ, H, D) is an even regular spectral triple. The action AdUt on Aθ by T2 is smooth and Aθ is complete with respect to the seminorms νk,α for k ∈ N and α ∈ N. Proof. The fact that (Aθ, H, D) is a regular spectral and the smoothness of AdUt on Aθ follows from Lemmas 3 and 4. Let (a(θ) k )k∈N be a Cauchy sequence with respect to the seminorms νk,α in Aθ, convergent to a bounded operator S on H. Then the map t 7→ AdUt (S) is smooth on T2. Hence we have a bounded operator S(−θ) on H. It remains to show that S(−θ) ∈ Aδ, since one would have S = (S(−θ))(θ) ∈ Aθ then. Let m ∈ N and α ∈ N2. As in the proof of Lemma 1, there is a universal constant C such that k − a(θ) (cid:13)(cid:13)(cid:13)δm([h, (a(θ) k′ )(−θ)](α))(cid:13)(cid:13)(cid:13) ≤ C(cid:13)(cid:13)(cid:13)AdUt (δm([h, a(θ) k − a(θ) k′ )](α))(cid:13)(cid:13)(cid:13)C 4(T2) for any k, k′. Thus we obtain (cid:13)(cid:13)(cid:13)δm([h, ak − ak′ ](α))(cid:13)(cid:13)(cid:13) < C max α′≤4 νm,α+α′ (a(θ) k − a(θ) k′ ). There is also a similar estimate for kδm([D, ak − ak′ ])k. Hence the sequence (ak)k in A is a Cauchy sequence for the seminorms νk,α, which is convergent to S(−θ). This shows S(−θ) ∈ Aδ. (cid:3) Remark 4. The newly obtained spectral triple (Aθ, H, D) again satisfies Aθ = A∞ θ . Hence one can form (Aθ)θ ′ for yet another deformation parameter θ′, which is identified to Aθ+θ ′. We also have A0 = A. Example 1. Let A be the smooth function algebra C∞(T2) over the 2-torus, H = L2(T2)⊕2 and /D =(cid:20) 0 i∂1 + ∂2 i∂1 − ∂2 0 (cid:21) . Consider the action of T2 on A given by the translation. Then the spectral triple (A, H, /D) is regular and satisfies A = A∞. Given a real parameter θ, the corre- sponding regular Connes-Landi deformation Aθ is precisely the algebra of Laurent seriesP(m,n)∈Z2 a(m,n)umvn of rapid decay coefficients over the two unitaries u and v satisfying uv = e2πiθvu. CONNES-LANDI DEFORMATION OF SPECTRAL TRIPLES 7 3. K-theory of Connes-Landi deformation We keep the notations A, A, σ, Aθ and Aθ of the previous section. We extend σ to A. As remarked in the previous section, as an abstract C∗-algebra Aθ is the de- formation algebra studied by Rieffel in [Rie2], [Rie3] and shown to have isomorphic K-group as A. In this section we elaborate a similar crossed product construction in order to show that Aθ and A have isomorphic periodic cyclic cohomology groups which is compatible with the isomorphism between the K-groups. Note that there is an action γ of T2 on C(T2 θ) called the gauge action, given by γt(uavb) = e2πi(t1a+t2b)uavb. We obtain the diagonal product action σ ⊗ γ of T2 on the (minimal) tensor product A ⊗min C(T2 θ). Proposition 6. The algebra Aθ is isomorphic to the subalgebra of A ⊗min C(T2 θ) consisting of the fixed elements under the diagonal action σ ⊗ γ of T2. Proof. When a ∈ A is a T2-eigenvector of weight (m, n), a ⊗ u−mv−n is in the fixed point algebra (A ⊗min C(T2 θ))σ⊗γ . The correspondence a(θ) 7→ e−πimnθa ⊗ u−mv−n defines a multiplicative map f from Aθ ∩ B(H)fin into (A ⊗min C(T2 θ))σ⊗γ. absolute convergence of Pm,n am,nu−mv−n inside (A ⊗min C(T2 Let a be an element of A. By the estimate given by Lemma 1, one has the θ))σ⊗γ . Hence f extends to the subalgebra Aθ of Aθ. Since Aθ is stable under holomorphic functional calculus, the spectral radius of a(θ) in A is the same as that in Aθ. It means that f is a contraction and extends to a ∗-homomorphism of Aθ into (A ⊗min C(T2 θ))σ⊗γ. We first prove that f is injective by contradiction. Suppose that there was a nonzero positive element x in the ideal ker f of Aθ. On one hand we have f (x0,0) = x0,0 ⊗ 1 6= 0. On the other hand, since f is equivariant for the action AdUt on Aθ θ))σ⊗γ. Hence the ideal ker f is invariant under AdUt and and σ ⊗ 1 on (A ⊗min C(T2 x(0,0) ∈ ker f . This is a contradiction, hence f is injective. Next we prove the surjectivity of f . By construction its image contains the linear span E of the elements of the form a ⊗ u−mv−n for where (m, n) ∈ Z2 and a ∈ A is a T2-eigenvector of weight (m, n). Since any homomorphism between C∗-algebras has a closed image, it is enough to show that the closure of E agrees with (A ⊗min C(T2 First, note that if a sequence (ak)k∈N in A converges to a T2-eigenvector a in A of weight (m, n) ∈ Z2, so does the sequence ((ak)m,n)k∈N. Hence for any T2- eigenvector a ∈ A of weight (m, n), the element a ⊗ u−mv−n lies in the closure of E. θ))σ⊗γ . Next, given any element x in (A ⊗min C(T2 θ))σ⊗γ , eigenspace decomposition with respect to the action σ ⊗ 1 gives us the T2-eigenvectors (am,n)(m,n)∈Z2 satisfying Let c(k) m,n be the sequence finitely supported coefficients defined by am,n ⊗ u−mv−n = e2πi(−mt1−nt2)ZT2 m,n = (cid:12)(cid:12)([m, k + m] × [n, k + n]) ∩ ([0, k] × [0, k]) ∩ Z2(cid:12)(cid:12) (σt ⊗ 1)(x). k2 c(k) . By the Fej´er kernel argument as in [BC], the sequence X(m,n)∈Z2 c(k) m,nam,n ⊗ u−mv−n, converges to x in norm. This shows that E is dense in (A ⊗min C(T2 θ))σ⊗γ. (cid:3) 8 MAKOTO YAMASHITA Let L1(T2, A; σ) denote the algebra of A-valued measurable integrable functions on For i = 1, 2, let dσ(i) denote the automorphism of T2 ⋉σ A which is dual to σ(i). T2 endowed with the σ-convolution product (f ∗ g)t =R fsσs(gt−s)ds. Then the of the normed ∗-algebra of the Laurent polynomials Pk fkvk with coefficients in T2 ⋉σ A by σ(2) can be described as the C∗-algebraic closure L1(T2, A; σ), whose product is given by crossed product Z⋉ σ(2) The algebra Z ⋉ σ(2) T2 ⋉σ A admits two automorphisms of interest, the first being and the second being 2 (gn)vm+n. X fmvm ∗X gnvn =X fm ∗ σm dσ(1)(X fkvk) =Xdσ(1)(fk)vk αθ(X fkvk) =X e2πikθfkvk. Let (dσ(1), αθ) denote an automorphism of Z ⋉ σ(2) of these two automorphisms. Lemma 7. The algebra Aθ is strongly Morita equivalent to Z⋉ T2 ⋉σ A given by the composition dσ(1),αθ Z⋉ σ(2) T2⋉σA. θ). Hence Aθ, which is isomorphic to (A ⊗min C(T2 Proof. Since the action of T2 on C(T2 θ) has full spectrum, so does the one on A ⊗ C(T2 by Proposition 6, is strongly Morita equivalent to T2 ⋉ (A ⊗min C(T2 θ)) (this follows from [Ng], although it should have been known to experts beforehand for this particular case of compact abelian group action). The latter algebra is generated by A and the operatorsR dtftUt of unitaries Ut for t ∈ T2 integrated with coefficient functions f in L1(T2), u and v satisfying θ))T2 AdUt (a) = σt(a) [a, umvn] = 0 AdUt (umvn) = e2πi(mt1+nt2)umvn uvu∗ = e2πiθv for any t ∈ T2 and (m, n) ∈ Z2. This can be identified to Z ⋉ A. dσ(1),αθ Z ⋉ σ(2) T2 ⋉σ (cid:3) By abuse of notation, let σ(2) θ denote the automorphism of T⋉ σ(1) A characterized by σ(2) θ (f )t = σ(2) θ (ft) (f ∈ L1(T, A), t ∈ T). Then we denote by (dσ(1), σ(2) θ ) the composition of dσ(1) and σ(2) Lemma 8. The algebra Aθ is strongly Morita equivalent to Z ⋉ θ . dσ(1),σ(2) θ T ⋉ σ(1) A. Proof. Let V denote the unitary represented by U (2) Z ⋉ σ(2) T2 ⋉σ A. Thus, V is characterized by θ in the multiplier algebra of V.f vk = σ(2) θ (g)vk, fkvk.V = e2πkiθgvk invariant and the crossed product of Z⋉ σ(2) for f ∈ L1(T2, A), where g denotes the function f (t1, t2 − θ). Then V is (dσ(1), αθ) (dσ(1), αθ) and the one by AdV ◦(dσ(1), αθ) define isomorphic algebras. T2⋉σ A by Z given by the automorphism Now, one computes AdV ◦dσ(1) ◦ αθ(f vk) = AdV (e2πikθhvk) = σ(2) θ (h)vk, CONNES-LANDI DEFORMATION OF SPECTRAL TRIPLES 9 where h(t1, t2) = e2πit1 f (t1, t2). hence the automorphism AdV ◦(dσ(1), αθ) is equal to the composition of dσ(1) and σ(2) θ . Thus we have an isomorphism between the crossed products as Z ⋉ ( dσ(1),αθ) T ⋉ σ(1) Z ⋉ T ⋉ σ(2) A ≃ Z ⋉ dσ(2) dσ(1),σ(2) θ T ⋉ σ(1) Z ⋉ T ⋉ σ(2) A. dσ(2) The right hand side can be also expressed as Z ⋉ σ(2) T ⋉ σ(2) Z ⋉ dσ(1),σ(2) θ T ⋉ σ(1) A. By Takesaki-Takai duality [Tak], this algebra is isomorphic to the tensor product of Z ⋉ σ(1) A with the compact operator algebra K, which proves our assertion. (cid:3) dσ(1),σ(2) T ⋉ θ Remark 5. The strong Morita equivalence of Lemma 8 can be described in terms of a bimodule E0 as follows. As a linear space, take E0 as the tensor product ℓ2Z ⊗ A. Define left actions of A, C∗(T) and Z on E0 by a(m,n).δk ⊗ b(m′,n′) = δk ⊗ ab, zl.δk ⊗ b(m,n) =(δk ⊗ b 0 (l = −k + m), (otherwise), u.δk ⊗ b(m,n) = e2πiθnδk+1 ⊗ b, where a(m,n), b(m,n) denotes any element of A with weight (m, n) for σ, zl is the function t 7→ e2πikt in the convolution algebra C∗T and u is the generating unitary of C∗Z. These form a covariant representation of Z, T and A with respect to the action σ(1) of T on A and (σ1, σ(2) θ ) of Z on T ⋉ Next, consider a right action of Aθ on E0 by σ(1) A. δk ⊗ b(m′,n′).a(θ) (m,n) = e2πiθknδk+m ⊗min ba. Then E0 has an Aθ-valued inner product hδka(m,n), δlb(m′,n′)i = δk−m,l−m′ (a∗b)(θ). The completion E of E0 with respect to this inner product becomes a bimodule over Z ⋉ σ(1) A and Aθ. Moreover the operators coming from Z ⋉ T ⋉ A are T ⋉ σ1,σ(2) θ precisely the Aθ-compact operators. Hence we obtain Z ⋉ σ1,σ(2) θ T ⋉ σ(1) A ≃ End0 Aθ (E) ≃ K ⊗ Aθ. Let (dσ(1), σ(2) θt ) denote the action of R on R ⋉ σ(1) A given by θt )t0 f )s = e2πt0sσ(0,θt0)(fs). ((dσ(1), σ(2) Proposition 9. The algebra Aθ is strongly Morita equivalent to R⋉ A. ( dσ(1),σ(2) θt ) R⋉ σ(1) Proof. For any automorphism α of A, the mapping cone MαA of α is defined to be the algebra of functions from R to A satisfying f (t + 1) = α(f (t)) for any t ∈ R. It admits a natural action α, called the suspension flow, of R by (αsf )t = f (t + s) Regarding this flow we have the strong Morita equivalence between R ⋉ α MαA and R ⋉α A. The crossed product R ⋉ σ(1) A is isomorphic to the mapping cone of dσ(1) on σ(1) A, and the suspension flow is identified to dσ(1) : R y R ⋉ We claim that the mapping cone of the automorphism(dσ(1), σ(2) σ(1) A Indeed, there is the 'untwisting isomorphism' f 7→ σ(1) A which is compatible (cid:3) is isomorphic to R ⋉ βf, (βf )t = σ(2) with the suspension flow. σ(1) A. θ ) on T ⋉ θt (ft) from M dσ(1) σ(1) A to M ( dσ(1),σ(2) σ(1) A. T ⋉ T ⋉ T ⋉ θ 10 MAKOTO YAMASHITA Corollary 10 ([Rie3]). The K-groups of A and Aθ are naturally isomorphic. Proof. By Proposition 9, the K-group of Aθ can be identified to that of R ⋉ R ⋉ to the K-group of R ⋉ depend on θ. dσ(1),σ(2) σ(1) A. By Connes-Thom isomorphism [Con], the latter is naturally isomorphic σ(1) A with parity exchange in the degree. This does not (cid:3) θt 3.1. Smooth KK-equivalence of CL deformations. Now we turn to the smooth analogue of the above KK-equivalence between the Connes-Landi deformations. The strong Morita equivalence of Proposition 9 restricts to an isomorphism between the rapid decay infinite matrix algebra with coefficients in Aθ and the "smooth" crossed product Z ⋉ T ⋉ A. Let T ⋉ σ(1) A denote the algebra given by the convolution product on the linear space C∞(T; A) with convolution product twisted by σ(1). This algebra has semi- norms given by the seminorms νk,α for k ∈ N and α ∈ N2 on A and the derivations with respect to the variable on T. Next, let Z ⋉ σ(1) A denote the algebra of sequences (fn)n∈Z in T ⋉ σ(1) A which are rapid decay with respect to the aforementioned seminorms, endowed with the convolution product twisted by σ(1)σ(2)(θ). We have the inclusions T ⋉ σ(1) A ⊂ Z ⋉ T ⋉ A into the corresponding C∗-algebras. σ(1) A ⊂ T ⋉ A and Z ⋉ σ(1)σ(2)(θ) σ(1)σ(2)(θ) T ⋉ T ⋉ Let K∞ denote the algebra of the rapid decay matrices on Z ([Phi], Section 2): the elements of K∞ are the matrices (ai,j)i,j∈Z indexed by Z satisfying for k ∈ N. The algebra K∞ becomes a Fr´echet algebra with respect to the family of (semi-)norms p1 + i2 + j2 k ai,j → 0 (i, j → ∞) µk((ai,j)i,j∈Z) = max i,j∈Zp1 + i2 + j2 k ai,j, for k ∈ N. The projective tensor product K∞ ⊗ Aθ is identified to the algebra of the Aθ-valued matrices which are rapid decay with respect to the seminorms on Aθ (loc. cit. Corollary 2.4). Proposition 11. The Fr´echet algebras Z ⋉ isomorphic via the correspondences dσ(1),σ(2) θ T ⋉ σ(1) A and K∞ ⊗ Aθ are T ⋉ σ(1) A ∋ unf (t) 7→ kunf (t)(l, m)l,m∈Z (7) Z ⋉ dσ(1),σ(2) θ and =(cid:18)ZT2 (a)Xm∈Z ZT σ(1) t e2πi(mt−(n+m−l)s)σ(s−t,(n−l)θ)(f (t))dtds(cid:19)(θ) e−2πimtumdt ←[ a(θ) ∈ Aθ, ul0−m0 e−2πim0s ←[ el0,m0 = δ(l0,m0)(l, m) ∈ K∞. Here the terms in the left hand side denote elements in Z ⋉ the ones in the right hand side denote Aθ-valued matrices in K∞ ⊗ Aθ. dσ(1),σ(2) T ⋉ θ σ(1) A and Proof. These formulae determine well-defined linear maps between the two algebras by the assumption on the smoothness of σ and the description of the smooth crossed products above. On the other hand, the multiplicativity in the assertion follows from the bimodule described in Remark 5. We may take a "basis" (δk ⊗1)k∈Z of E0 over the right action of Aθ. We indicate the case for (7) in the following. Suppose that f ∈ C∞(T; A) CONNES-LANDI DEFORMATION OF SPECTRAL TRIPLES 11 is a function of the form t 7→ e2πilta for some integer l and a ∈ A of weight (m, n′) for σ. First we have unf (t).δk ⊗ 1 =(e2πiθnn′ 0 δk+n ⊗ a (k = m − l) (otherwise). On the other hand we have e2πiθnn′ δk+n ⊗ a = (δk−m+n ⊗ 1).(e2πiθ{(m−k−n)n′+nn′}a)(θ). Hence unf (t) represents a operator which moves the (m − l)-th base to the (n − l)- th base multiplied by e2πiθln′ a(θ). Thus we need to show that the formula (7) applied to our choice of f gives e2πiθln′ a(θ)δn−l,m−l(l′, m′). Now, when l′ and m′ are arbitrary integers, kunf (l′, m′) =ZT2 =ZT2 =(e2πiθ(n−l′)n′ 0 e2πi(m′t−(n+m′−l′)s)σ(s−t,(n−l′)θ)(f (t))dtds e2πi{m′t−(n+m′−l′)s+(s−t)m+(n−l′)θn′+lt}adtds a (l + m′ − m = 0, m − n − m′ + l′ = 0) (otherwise). In the nontrivial case of l + m′ − m = 0 and m − n − m′ + l′ = 0 one has l′ = n − l and (n − l′)n′ = ln, which is the desired relation. The rest is proved in a similar way. (cid:3) There is also a continuous analogue of the above argument. Let K∞ R denote the algebra of the compact operators on L2(R) whose integral kernel belong to the R ⊗ Aθ is identified to the Schwartz class on R2. The projective tensor product K∞ algebra of the Aθ-valued integral kernels I(x, y) which satisfy max x,y (cid:13)(cid:13)(cid:13)δk((xm + yn)(∂m′ x + ∂n′ y )I(x, y))(cid:13)(cid:13)(cid:13) < ∞ for any m, n, m′ and n′ in N. Let S∗(R; A, σ(1)) be the convolution algebra of the A-valued Schwartz functions θt ) restricts to a smooth action on with respect to the action σ. The action (dσ(1), σ(2) this algebra, hence we may take the convolution algebra S∗(R; S∗(R; A, σ(1)), (dσ(1), σ(2) θt )) of the S∗(R; A, σ(1))-valued Schwartz functions. Let R ⋉ σ(1) A denote this algebra. We write f (t, τ ) for a typical element of this algebra, where for any fixed τ , the function t 7→ f (t, τ ) is in S∗(R; A, σ(1)). Then one has the following analogue of Proposition 11: dσ(1),σ(2) R ⋉ θt σ(1) A and K∞ Proposition 12. The Fr´echet algebras R ⋉ isomorphic. The correspondence between these two algebras are given by dσ(1),σ(2) R ⋉ θt R ⊗ Aθ are R ⋉ dσ(1),σ(2) θt R ⋉ σ(1) A ∋ ft,τ 7→ kft,τ (λ, µ) = (cid:18)ZR2 e2πi(µt−(τ +µ−λ)s)σ(s−t,(τ −λ)θ)(ft,τ )dtdsdτ(cid:19)(θ) , (k(λ, µ))e−2πi(µ−λ+τ )t+µsdtdλdµ ←[ k(λ, µ) ∈ K∞ R ⊗ Aθ. and f k s,τ =ZR σ(1) t 12 MAKOTO YAMASHITA Corollary 13. There is a natural isomorphism Φθ : HP∗(A) → HP∗(Aθ) between the periodic cyclic cohomology groups which is compatible with the identification of the K-groups of Corollary 10. Proof. On the one hand, by Elliott-Natsume-Nest [ENN] Theorem 6.2, one has natural isomorphisms HP∗(R ⋉ σ(1) A) ≃ HP∗+1(R ⋉ σ(1) A) ≃ HP∗(A) R ⋉ dσ(1),σ(2) θt compatible with the Connes-Thom isomorphism K∗(R⋉ σ(1) A) ≃ K∗(A). On the other hand, by Proposition 12, one has HP∗(R ⋉ σ(1) A) ≃ HP∗(Aθ) (by the stability of HP∗ for K∞ R ⊗ −, loc. cit. Theorem 4.3). Combining these and remembering that the isomorphisms involved are transpose to the ones between K-groups of the C∗-algebraic completions, we have the assertion. (cid:3) dσ(1),σ(2) dσ(1),σ(2) R ⋉ R⋉ θt θt Remark 6. Although we need the above continuous crossed product presentation R ⊗ Aθ later in order to investigate the deformation of cyclic cocycles, the of K∞ periodic cyclic cohomology isomorphism of Corollary 13 itself can be deduced from the algebra isomorphism of Proposition 11 and the result of [Nes]. 4. Preservation of dimension spectrum under deformation We keep the notations A, H, D, σ and U as in the previous sections. In general the regularity of the spectral triple is not guaranteed to be preserved under a deformation as in the case of Podle´s sphere [NT]. We show that in the case of Connes-Landi deformation, the regularity is well preserved. As in the previous section consider a spectral triple (A, H, D) endowed with an action of T2 satisfying the condition (1) and we assume that A = A∞. Let B be the algebra generated by δk(a) for a ∈ A + [D, A] and k ∈ N. Similarly, let Bθ denote the algebra generated by δk(Aθ + [D, Aθ]) for k ∈ N. Now suppose that D is n-summable. Given a bounded operator T on H, the function ζT (s) = Tr(T D−s) is called the zeta function associated to T . This is a priori holomorphic in the region {z ∈ C ℜ(z) > n}. If the functions of the form ζT for T ∈ B admit meromorphic extension to the whole plane C, the dimension spectrum of (A, H, D) is defined to be the collection of the polls of the analytic continuation of the functions of the form ζT (s) for T ∈ B [CM]. Theorem 1. Let (A, H, D) be a regular spectral triple whose zeta functions ζT for T ∈ B admit meromorphic extensions to C. Suppose that there is a smooth action σ by T2 which is spatially implemented by a strongly continuous unitary representation U∗ on H and satisfies A = A∞. The dimension spectrum of the spectral triple (Aθ, H, D) is equal to that of (A, H, D). Proof. The elements of the algebra B satisfy the assumption of Lemma 1. Moreover when a1, . . . , aj are elements of A and k1, . . . , kj are positive integers, one has (cid:0)δk1 (a1) · · · δkj (aj)(cid:1)(θ) = δk1(a(θ) 1 ) · · · δkj (a(θ) j ). There is also an analogous identity involving [D, a] for a ∈ A. Hence the algebra Bθ is the collection of the operators of the form T (θ) for T ∈ B. Since D commutes with Ut, the function ζT (s) only depends on T(0,0) for any bounded operator T on H. Since we have the equality T (θ) (0,0) = T(0,0), which leads to ζT (θ) (s) = ζT (s). This shows that the dimension spectrum of (Aθ, H, D) is the same as that of (A, H, D). (cid:3) CONNES-LANDI DEFORMATION OF SPECTRAL TRIPLES 13 5. Deformation of invariant cyclic cocycles In the rest of the paper we investigate the pairing of the character of the spectral triple (Aθ, H, D) with K∗(Aθ) for varying θ ∈ R under the natural isomorphisms given by Corollaries 10 and 13. We start with a review of Elliott-Natsume-Nest's isomorphism HCn(A) → HCn+1(R ⋉σ A), φ 7→ #σφ [ENN]. Let A be a Fr´echet *-algebra which is dense and close under holomorphic func- tional calculus inside a C∗-algebra A. Let σ be a smooth action of R on A. Recall that close graded traces on graded differential algebras containing A in the degree 0 give cyclic cocycles on A, and conversely any cyclic n-cocycle can be represented as a closed graded trace on the universal differential graded algebra Ω(A) = ⊕Ω(A)n (where Ω(A)0 = A and Ω(A)n = A ⊗ n ⊕ A ⊗ n+1 for n > 0). Suppose that a cyclic n-cocycle φ on A is given by a closed graded trace of degree n, which is also denoted by φ by abuse of notation, on a graded differential algebra Ω containing A in degree 0. Let S∗R be the convolution algebra of the Schwartz class functions on R, E the direct sum Ω(S∗R)0 ⊕ Ω(S∗R)1. We construct a differential graded algebra structure on Ω ⊗ E containing R ⋉σ A in the degree 0 as follows. The space Ω ⊗ E0 is naturally identified to R ⋉ Ω = S∗(R; Ω). There is a deriva- tion d : Ω ⊗ E0 → Ω ⊗ E given by d(ω ⊗ f ) = (dω) ⊗ f + (−1)deg ωω ⊗ df for any homogeneous element ω ∈ Ω. On Ω ⊗ E1, the differential is defined by d(ω ⊗ f dg) = (dω) ⊗ f dg. The 1-forms in Ω(S∗R)1 act on Ω ⊗ E by df (ω ⊗ g) = d(f (ω ⊗ g)) − f d(ω ⊗ g). These, together with the original product structure of Ω, defines a structure of a graded differential algebra on Ω ⊗ E. Then the closed graded trace #σφ on Ωn ⊗ E1 = Ω ⊗ S∗(R) ⊗ S∗(R) correspond- ing to the cyclic (n + 1)-cocycle #σφ over R ⋉σ A is defined as follows: When φ is invariant under σ, (8) reduces to #σφ(f ) = 2πiZ ∞ −∞Z t 2πiZ ∞ −∞ 0 (8) (9) φ(σsf (−t, t))dsdt. tφ(f (−t, t))dt. This correspondence of φ to #σφ is compatible with the Connes-Thom isomor- phism Φσ : K∗A → K∗+1R ⋉σ A ([ENN], Theorem 6.2). Let δ denote the generator d dt of σ. When φ is a σ-invariant cyclic n-cocycle δ(a) = lim t→0 σt(a) φ(f 0, . . . , f n) = φ(σt(f 0), . . . , σt(f n)) (∀t ∈ T2) over A, we obtain a new n + 1-cocycle iδφ ([Con3] Chapter 3, Section 6.β) by iδφ(a0da1 · · · dan+1) = (−1)jφ(a0da1 · · · δ(aj) · · · dan+1). n+1Xj=1 On the other hand, noting that the convolution algebra of the A-valued rapid decay functions S∗(R; A) is identified to R⋉σ A, one obtains the dual cocycle φ over R⋉σA 14 by MAKOTO YAMASHITA φ(f 0, . . . , f n) =ZPn j=0 tj =0 φ(f 0 t0 , σt0 (f 1 t1 ), . . . , σPj<n tj (f 0 tn )) for f j ∈ S∗(R; A). We also have the dual action σ : R y R ⋉σ A and the iso- R ⊗ A, which implies HCk(R ⋉σ R ⋉σ A) ≃ HCk(A). morphism R ⋉σ R ⋉σ A ≃ K∞ Regarding these constructions one has the following generalization of [ENN], Propo- sition 3.11. Proposition 14. Let φ be a σ-invariant n-cocycle over A. The class of the cyclic (n + 1)-cocycle iδφ in HCn+1(A) agrees with that of #σ φ. Proof. Let b0, . . . , bn+1 be elements of R⋉σ A and f0, . . . , fn+1 be functions in S∗R. For each 0 ≤ j ≤ n + 1, bj ⊗ fj represents the element t 7→ bj fj(t) ∈ S(R; A). We consider the element (10) ω = (b0 ⊗ f0)d(b1 ⊗ f1) · · · d(bn+1 ⊗ fn+1) of Ω ⊗ E. Each term in the above can be expanded as d(bj ⊗ fj) = dbj ⊗ fj + bj ⊗ dfj. Since the terms containing more than one dfj's vanish and the only terms contributing to the pairing of ω with #σ φ are the elements in Ωn ⊗ E1. Hence #σ φ(ω) can be expressed as 2πiZsn+1=0 n+1Xj=1 (11) where tj φ(ηj(t0, . . . , tn+1))ξj(t0, . . . , tn+1)dt0 · · · dtn, ηj(t0, . . . , tn+1) = b0σs0 (db1) · · · σsj−1 (bj) · · · σsn (dbn+1), ξj(t0, . . . , tn+1) = f0(t0) · · · fj · · · fn+1(tn+1)dfj(tj), sk = t0 + · · · + tk. The derivation δf (t) = tf (t) on S∗(R; R ⋉ A) is the generator of the double dual action σ : R y R ⋉σ R ⋉σ A. The j-th term of (11) is equal to φ(b0 ⊗ f0)d(b1 ⊗ f1) · · · δ(bj ⊗ fj) · · · d(bn+1 ⊗ fn+1). φ(ω) = iδ(φ ⊗ Tr)(ω). Collecting these terms we obtain #σ φ(ω) = iδ There is a one-parameter unitary ut such that the double dual action σ is conju- R ⊗ A ≃ R ⋉σ R ⋉σ A by the formula σt = Adut ◦(IdK σt). By gated to IdK ⊗σ on K∞ Connes's 2 × 2-matrix trick, we obtain an action Φ(σ, u∗) of R on M2(R ⋉σ R ⋉σ A) by The generator of this action can be written as x21 x22 (cid:21)(cid:19) =(cid:20) σt(x11) Φ(σ, u∗)t(cid:18)(cid:20) x11 x12 x21 x22 (cid:21)(cid:19) =(cid:20) δσ,u∗(cid:18)(cid:20) x11 x12 σt(x12)u∗ t utσt(x21) Adut σt(x22) (cid:21) . (cid:21) , δ(x12) − x12h δ(x22) δ(x11) hx21 + δ(x21) (12) where h is the generator h = lim t→0 ut − 1 t of ut which is in the multiplier algebra of R ⋉σ R ⋉σ A, so that we have δ(x) = δ(x) + [h, x]. CONNES-LANDI DEFORMATION OF SPECTRAL TRIPLES 15 Now, the derivation δσ,u∗ of (12) determines a cyclic n + 1-cocycle ψ = iδσ,u∗ (φ ⊗ Tr) on M2(R ⋉σ R ⋉σ A). There are two embeddings of R ⋉σ R ⋉σ A into M2(R ⋉σ R ⋉σ A), given by Ψ1(x) =(cid:20) x 0 0 (cid:21) , 0 Ψ2(x) =(cid:20) 0 0 x (cid:21) . 0 The pullback HC∗(M2(R ⋉σ R ⋉σ A)) → HC∗(R ⋉σ R ⋉σ A) by these two homo- morphisms in cyclic cohomology become the same map. But the pullback of ψ by Ψ1 is iδφ ⊗ Tr while the one by Ψ2 is iδ(φ ⊗ Tr), hence these two cocycles deter- mine the same class in the cyclic cohomology group, which implies iδφ = #σ φ in HC∗(R ⋉σ R ⋉σ A). (cid:3) Now we consider an even regular spectral triple (A, H, D) with an action of T2 satisfying A∞ = A as in the previous sections. Note that we have the estimates νl,α( dk dtk σ(i) t (a)) = νl,α′ (a) for k ∈ N and α ∈ N2, where α′ = (α1 + k, α2) or α′ = (α1, α2 + k) corresponding to the cases i = 1, 2. Hence the actions σ(i) are smooth action on A∞ = A in the sense above. the dual cocycle φ over the crossed product T ⋉ Let φ be a cyclic n-cocycle on A which is invariant under σ. Then we obtain σ(1) A. Then again, φ is invariant σ(1) ⋉ A, σ(1) A, which φ induces a one φ(θ) on Aθ via the embedding under the action (dσ(1), σ(2) and with a similar process another cyclic cocycle on R ⋉ θ ). Then we obtain a cocycle φ. The cocycle φ on Z ⋉ dσ(1),σ(2) R ⋉ T dσ(1),σ(2) θt θ we still denote by given into a corner of K∞ ⊗ Aθ ≃ Z ⋉ dσ(1),σ(2) θ T σ(1) ⋉ A by Proposition 11. We record the formula for φ(θ) for a ∈ A: (13) φ(θ)(a(θ) 0 , . . . , a(θ) n ) = X φ((a0)m0,n0 , b(1,m1,l1), . . . , b(n,mn,ln)), m0+···+mn=0,l0+···+ln=0 where b(k,mk,lk) = e2πiθ(Pj<k mj )lk (ak)mk,lk . In the particular case of n = 0, φ is given by a trace on A and φ(θ) is given by the corresponding trace φ(θ)(a(θ)) = φ(a(0,0)) = φ(a). Theorem 2. Let φ be a σ-invariant cyclic n-cocycle on A. Then the cyclic n- cocycle φ(θ) on Aθ corresponds to the nonhomogeneous cyclic cocycle (14) φ + θiδ(1) iδ(2) φ on A under the natural isomorphism of Corollary 13. σ(1) A Proof. It is enough to check that the cyclic n-cocycle corresponds to the one of (14) on A via the natural isomorphism in HP∗. The action (σ1, σ(2) σ(1) A is generated by the derivation θt ) of R on R ⋉ dσ(1),σ(2) θt R ⋉ φ on R ⋉ δ(θ)(f )s = sfs + θ d dt2 σ(0,t2)(fs). Let δ (1) denote the generator of the dual action dσ(1) and δ(2) that of the action σ(2). The equation above shows that δ(θ) = dσ(1) + θδ(2). Applying Proposition 14, 16 MAKOTO YAMASHITA the cocycle φ corresponds to the cyclic cocycle iδ(θ) φ = iδ (1) φ + θiδ(2) φ on R ⋉ σ(1) A. The first term iδ (1) φ in the right hand side of the above is the cocycle which corresponds to φ under the isomorphism HP∗(A) ≃ HP∗+1(R ⋉ σ(1) A). On the other hand, one has iδ(2) φ = [iδ(2) φ for the second term. This n + 1-cocycle on R ⋉ σ(1) A corresponds to the n + 2-cocycle iδ(1) iδ(2) φ on A again by Proposition 14. Combining these two, one obtains the assertion of the Proposition. (cid:3) Example 2. Let M be a compact smooth manifold, endowed with a smooth action σ of T2. Then M admits a Riemannian metric which is invariant under σ. Then the algebra of the smooth functions on M , the de Rham complex Ω∗(M ) graded by the degree of forms and the operator d + d∗ densely defined on H = L2(Ω∗(M )) and the induced representation of T2 on H satisfies the assumption of this paper. The volume form dv on M is invariant under σ and it defines an invariant trace i = 1, 2. By the preceding proposition, the map K0(CMθ) → C induced by the trace τ (θ) on CMθ corresponds to the map K 0(M ) → C induced by the current τ : f 7→RM f dv on C(M ). Let Xi denote the vector fields on M generating σi for f + g0dg1 ∧ dg2 7→ZM ZM When E is a vector bundle over M , the above map on the class of E in K 0(M ) gives the number ch(E) ∧ (dv + θiX1 iX2 dv) = vol(M ) rk(E) + θZM f + θ(g0hdg1, X1ihdg2, X2i − hdg1, X2ihdg2, X1i)dv. c1(E) ∧ iX1 iX2 dv. 5.1. Chern-Connes character of deformed triple. The algebra R ⋉ σ(1) A acts on the Hilbert space L2(R) ¯⊗ H in the following way: a function f in S∗(R, A) transforms vectors in L2(R; H) ≃ L2(R) ¯⊗ H by (π(f )ξ)t =ZR fsU(s,0)ξt−sds. Next we define an action of the algebra R ⋉ dσ(1),σ(2) θt R ⋉ σ(1) A on L2(R; H) by (π(f )ξ)t =ZR σ(1) A). for f ∈ S∗(R; R ⋉ e2πist(π(f s)U(0,θs)ξ)tds Lemma 15. The correspondence f 7→ π(f ) defines a representation of R ⋉ R ⋉ dσ(1),σ(2) θt σ(1) A on L2(R) ¯⊗ H. Similarly, we have a representation of Z ⋉ dσ(1),σ(2) θ T ⋉ σ(1) A on ℓ2Z ¯⊗ H by zlaδk ⊗ ξ =(δk ⊗ aξ 0 (aξ ∈ H(m,n), l + k − m = 0) (otherwise) and wδk ⊗ ξ = e2πinθδk+1 ⊗ ξ for any ξ ∈ H(m,n). In the following we make a more detailed analysis of the strong Morita equiv- alence given by Proposition 9 in relation to the unbounded selfadjoint operator D. Proposition 16. There is a unitary operator U0 on ℓ2(Z) ¯⊗ H satisfying U ∗ D)U0 = 1 ⊗ D and 0 (1 ⊗ U ∗ 0 (Z ⋉ dσ(1),σ(2) θ T ⋉ σ(1) A)U0 = K(ℓ2(Z)) ⊗ Aθ. CONNES-LANDI DEFORMATION OF SPECTRAL TRIPLES 17 Proof. Let U0 denote the unitary operator on ℓ2Z ¯⊗ H given by U0δk⊗ξ = e2πiθnkδk+m⊗ ξ for any k ∈ Z and ξ ∈ H(m,n). When a ∈ A(p,q) and ξ ∈ Hm,n, we have U ∗ 0 zlaU0δk ⊗ ξ = U0zlae2πiθknδk+m ⊗ ξ =(e2πiθq(p−k)+pnδk−p ⊗ aξ 0 (l + k − p = 0) (otherwise). This is the effect of e2πiθqle−l,−(l+p) ⊗ a(θ) ∈ K(ℓ2Z) ⊗min Aθ on δk ⊗ ξ, where em,n denotes the matrix element δk 7→ δn,kδm on ℓ2Z for any (m, n) ∈ Z2. Similarly, one has U ∗ on ℓ2Z. Hence we have U ∗ 0 similarly show that U0(K(ℓ2Z) ⊗ Aθ)U ∗ dσ(1),σ(2) Z ⋉ 0 wU0 = v ⊗ 1 where v is the unitary operator δk 7→ δk+1 σ(1) AU0 ⊂ K(ℓ2(Z)) ⊗min Aθ. One can (cid:3) T ⋉ T ⋉ 0 ⊂ Z ⋉ σ(1) A. θ dσ(1),σ(2) θ As a consequence of Proposition 16 the image under π of the smooth subalgebra σ(1) A have bounded commutators with 1 ⊗ D. dσ(1),σ(2) σ(1) A of Z ⋉ dσ(1),σ(2) T ⋉ T ⋉ θ θ Z ⋉ In particular, (15) (Z ⋉ dσ(1),σ(2) θ T ⋉ σ(1) A, ℓ2Z ¯⊗ H, 1 ⊗ D) is a spectral triple over Z ⋉ dσ(1),σ(2) θ T ⋉ σ(1) A. Corollary 17. The character of the spectral triple (15) corresponds to the map induced by the character of D over Aθ via the strong Morita equivalence of Lemma 8. R ⊗ Aθ given by AdU is Proof. The isomorphism of R ⋉ equal to the strong Morita equivalence of Lemma 8. By the U -invariance of 1 ⊗ D, R ⊗ Aθ. The its character over R ⋉ latter induces the same class as the character of D in the cyclic cohomology group via the isomorphism HC∗(K∞ (cid:3) σ(1) A is identified to that over K∞ R ⊗ Aθ) ≃ HC∗(Aθ). σ(1) A onto K∞ dσ(1),σ(2) dσ(1),σ(2) R ⋉ R ⋉ θt θt Suppose that (A, H, D) is n-summable (0 < n ∈ 2N). Then the character chD(f0, . . . , fn) = Trs(γf0[F, f1] · · · [F, fn]) (F = D D−1) of this triple for is a cyclic n-cocycle which is invariant σ. Lemma 18. The character of the spectral triple (15) over Z ⋉ equal to chD. dσ(1),σ(2) θ T ⋉ σ(1) A is Corollary 19. Under the strong Morita equivalence of Proposition 9, the character of D over Aθ corresponds to chD over R ⋉ R ⋉ σ(1) A. dσ(1),σ(2) θt Remark 7. As a consequence of Corollary 19, or directly from the formula (13), the cocycle ch(θ) (A,H,D) over Aθ agrees with ch(Aθ ,H,D). Theorem 3. Let (A, H, D) be an even regular spectral triple endowed with a smooth action of T2 satisfying A = A∞. Then the Chern-Connes characters of (Aθ, H, D) and (A, H, D) induce the same maps on K0(Aθ) via the isomorphism Λθ of Corol- lary 10. Proof. Put φ = chD. By Corollary 19, the character of D over Aθ corresponds to σ(1) A. As in the proof of Proposition 2, this cocycle the cocycle φ + θiδ(2) φ on R ⋉ A. corresponds to the cocycle iδ (1) φ on R ⋉ dσ(1),σ(2) R ⋉ θt 18 MAKOTO YAMASHITA Given an action α on a C∗-algebra B, let Φα denote the Connes-Thom isomor- phism [Con] K∗B → K∗+1(R ⋉α B). Let y be any element of K1(R ⋉ A). Then one has φ, Φ−1 h dσ(1) σ(2) θt (y)i = hiδ (1) φ + θiδ(2) φ, yi = hiδ (1) φ, yi + θhiδ(2) φ, yi. φ is equal to a character of a spectral triple, its pairing with any element in φ, yi + θhiδ(2) φ, yi must Since K0(R ⋉ stay inside Z regardless of the value of θ, which implies hiδ(2) φ, yi = 0. σ(1) A) must be an integer. Hence hiδ (1) dσ(1),σ(2) R ⋉ θt R ⋉ σ(1) A) be the isomorphism given by Propo- Let Ψ : K0(Aθ) → K0(R ⋉ dσ(1),σ(2) θt sition 9. For any element x ∈ K0(Aθ), one has hchD, xi = h φ, Ψ(x)i = hiδ (1) φ, Φ dσ(1)σ(2) θt Ψ(x)i = hchD, Φσ(1) Φ dσ(1)σ(2) θt Ψ(x)i. Since x 7→ Φσ(1) Φ dσ(1)σ(2) one obtains the assertion of the theorem. Ψ(x) is the natural isomorphism Λθ : K0(Aθ) → K0(A), (cid:3) θt Remark 8. In the proof of Theorem 3 we saw that the cyclic (n + 2)-cocycle iδ(1) iδ(2) chD on A pairs trivially with K0(A). It is very likely that this cocycle gives the trivial class of HP0(A). Note that this phenomenon is specific to the cases where one has a unitary representation Ut on H satisfying (1). Otherwise, iδ(1) iδ(2) chD can pair nontrivially with K0(A). For example, the trace τ on T2 given by the Haar integral is a character of a 2-summable even spectral triple over C∞(T2) and one has hiδ(1) iδ(2) τ, K 0T2i = Z. Remark 9. After a major portion of the results in this paper was obtained the author has learned from N. Higson that the invariance of the index pairing as stated in Theorem 3 can be explained in a purely C∗-algebraic framework as follows: let R act on the C∗-algebra R ⋉ θt ) at the fiber of θ ∈ [0, 1]. Then the resulting crossed product algebra R ⋉ (R ⋉ σ(1) A ⊗min C[0, 1]) act on the Hilbert space L2(R) ¯⊗ H ¯⊗ L2([0, 1]) determined by the action of R ⋉ σ(1) A as described at the beginning of Section 5.1 and the pointwise multiplication representation of C[0, 1] on L2[0, 1]. Then the R ⋉ (R ⋉ σ(1) A ⊗min C[0, 1])-C[0, 1]- module L2(R) ¯⊗ H ¯⊗ L2([0, 1]) together with the unbounded C[0, 1]-endomorphism 1L2(R) ⊗D ⊗1L2([0,1]) define an element α of KK(R⋉(R⋉ σ(1) A ⊗min C[0, 1], C[0, 1]). The evaluation at θ ∈ [0, 1] gives us KK-equivalences σ(1) A ⊗min C[0, 1]) ≃KK R ⋉ σ(1) A ⊗min C[0, 1] by (dσ(1), σ(2) evθ : R ⋉ (R ⋉ dσ(1),σ(2) σ(1) A, R ⋉ R ⋉ θt dσ(1),σ(2) θt evθ : C[0, 1] ≃KK C. These KK-equivalences intertwine α and the element of K 0(Aθ) given by D, while the composition ev0 ◦ ev−1 gives the natural isomorphism between K0(Aθ) and K0(A). θ Acknowledgement. The author would like to thank Y. Kawahigashi for his contin- uing support. He is also grateful to N. Higson, E. Blanchard, G. Skandalis, R. Tomatsu, N. Ozawa, R. Ponge, T. Natsume who gave many insightful suggestions during the various stages of research. He also benefitted from conversations with C. Oikonomides, T. Fukaya, S. Oguni and many others. References [BC] [Bra] E. B´edos and R. Conti. On twisted Fourier analysis and convergence of Fourier series on discrete groups. J. Fourier Anal. Appl. 15(2009), 336 -- 365. O. Bratteli. Derivations, dissipations and group actions on C ∗-algebras, volume 1229 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1986. CONNES-LANDI DEFORMATION OF SPECTRAL TRIPLES 19 [Con] [CM] A. Connes. An analogue of the Thom isomorphism for crossed products of a C ∗-algebra by an action of R. Adv. in Math. 39(1981), 31 -- 55. A. Connes and H. Moscovici. The local index formula in noncommutative geometry. Geom. Funct. Anal. 5(1995), 174 -- 243. [Con1] A. Connes. C ∗ alg`ebres et g´eom´etrie diff´erentielle. C. R. Acad. Sci. Paris S´er. A-B 290(1980), A599 -- A604. [Con2] A. Connes. Noncommutative differential geometry. Inst. Hautes ´Etudes Sci. Publ. Math. 62(1985), 257 -- 360. [Con3] A. Connes. Noncommutative geometry. Academic Press Inc., San Diego, CA, 1994. [CL] A. Connes and G. Landi. Noncommutative manifolds, the instanton algebra and isospec- tral deformations. Comm. Math. Phys. 221(2001), 141 -- 159. [ENN] G. A. Elliott, T. Natsume, and R. Nest. Cyclic cohomology for one-parameter smooth crossed products. Acta Math. 160(1988), 285 -- 305. [NT] [Nes] [Ng] [Li] [GBVF] J. M. Gracia-Bond´ıa, J. C. V´arilly, and H. Figueroa. Elements of noncommutative ge- ometry. Birkhauser Advanced Texts: Basler Lehrbucher. [Birkhauser Advanced Texts: Basel Textbooks]. Birkhauser Boston Inc., Boston, MA, 2001. H. Li. θ-deformations as compact quantum metric spaces. Comm. Math. Phys. 256(2005), 213 -- 238. S. Neshveyev and L. Tuset. A local index formula for the quantum sphere. Comm. Math. Phys. 254(2005), 323 -- 341. R. Nest. Cyclic cohomology of crossed products with Z. J. Funct. Anal. 80(1988), 235 -- 283. C.-K. Ng. Morita equivalences between fixed point algebras and crossed products. Math. Proc. Cambridge Philos. Soc. 125(1999), 43 -- 52. N. C. Phillips. K-theory for Fr´echet algebras. Internat. J. Math. 2(1991), 77 -- 129. M. Pimsner and D. Voiculescu. Exact sequences for K-groups and Ext-groups of certain cross-product C ∗-algebras. J. Operator Theory 4(1980), 93 -- 118. A. Rennie. Smoothness and locality for nonunital spectral triples. K-Theory 28(2003), 127 -- 165. [Phi] [PV] [Ren] [Rie1] M. A. Rieffel. Projective modules over higher-dimensional noncommutative tori. Canad. J. Math. 40(1988), 257 -- 338. [Rie2] M. A. Rieffel. Deformation quantization for actions of Rd. Mem. Amer. Math. Soc. 106(1993), x+93. [Rie3] M. A. Rieffel. K-groups of C ∗-algebras deformed by actions of Rd. J. Funct. Anal. 116(1993), 199 -- 214. [Rie4] M. A. Rieffel. Metrics on states from actions of compact groups. Doc. Math. 3(1998), [Tak] 215 -- 229 (electronic). H. Takai. On a duality for crossed products of C ∗-algebras. J. Functional Analysis 19(1975), 25 -- 39. Graduate School of Mathematical Sciences, the University of Tokyo, 3-8-1 Komaba Meguro-ku, Tokyo, JAPAN E-mail address: [email protected]
1311.0030	1	1311	2013-10-31T20:38:59	Local derivations on Rings containing a von Neumann algebra and a question of Kadison	[ "math.OA" ]	We prove that if M is a von Neumann algebra whose abelian summand is discrete, then every local derivation on the algebra of all measurable operators affilated with M is a derivation. This answers a question of Richard Kadison.	math.OA	math	Local derivations on Rings containing a von Neumann algebra and a question of Kadison. Don Hadwin, Jiankui Li, Qihui Li, and Xiujuan Ma Abstract. We prove that if M is a von Neumann algebra whose abelian sum- mand is discrete, then every local derivation on the algebra of all measurable operators affilated with M is a derivation. This answers a question of Richard Kadison. At a conference held in Texas A & M University of 2012, Richard Kadison gave a talk about his joint work with Zhe Liu [10, 11] in which they proved that the only derivation that maps the algebra S(M) of closed densely defined operators affiliated with a factor von Neumann algebra M of type II1 into that von Neumann algebra is zero. Kadison asked whether every local derivation on S(M) is a derivation. In this note we prove a general ring-theoretic result which implies that for all von Neumann algebras M whose abelian summand is discrete, every local derivation on S(M) of all measurable operators is a derivation. If R is a ring (resp. algebra) and δ : R → R is an additive (resp. linear) mapping, we say that δ is a derivation if, for all a, b ∈ R, we have δ (ab) = δ (a) b + aδ (b) . We say that an additive mapping δ is a local derivation if, for every x ∈ R there is a derivation ρx on R such that δ (x) = ρx (x) . To prove our main result, we need two lemmas. The first is a result of [3, Corollary 4.5]. Suppose R is a ring with identity 1 and n ≥ 2 is an integer. We say that a subset {Eij : 1 ≤ i, j ≤ n} ⊂ R is a system of n × n matrix units for R if and only if Pn i=1 Eii = 1 and, for 1 ≤ i, j, s, t ≤ n, EijEst = 0 if j 6= s and EijEst = Eit if j = s. In [3, p.11], Bresar shows that for any unital ring B, the ring Mn(B) is generated by the set of all idempotents in B, where 2 ≤ n. The following Lemma 1 is a special case of [3, Corollary 4.5]. Lemma 1. If R is a ring with identity 1 that possesses a system of n × n matrix units for some n ≥ 2, then every local derivation on R is a derivation. 2000 Mathematics Subject Classification. Primary 46L57, 13N15; Secondary 16W25. Key words and phrases. local derivation, derivation, affiliated algebra, von Neumann algebra. 1 2 DON HADWIN, JIANKUI LI, QIHUI LI, AND XIUJUAN MA Note that, in general, derivations need not leave ideals invariant, e.g., differenti- ation on the polynomials. However, if p a central idempotent, then every derivation (hence, every local derivation) δ leaves pR and (1 − p) R invariant, since δ (pa) = δ ((pa) p) = paδ (p) + δ (pa) p. Moreover, pR is isomorphic to R/ (1 − p) R. This yields the following corollary to the preceding lemma. A family P of central idempotents for a ring R is separating if and only if, for each nonzero x ∈ R, there is a p ∈ P such that px 6= 0. Lemma 2. Suppose R is a ring with identity and P is a separating family of central idempotents such that, for each p ∈ P, every local derivation on pR is a derivation. Then every local derivation on R is a derivation. Proof. Suppose δ : R → R is a local derivation, a ∈ R and p ∈ P. Then there is a derivation ρ on R such that δ (pa) = ρ (pa) = ρ (p (pa)) = pρ (pa)+ρ (p) pa = p [ρ (pa) + ρ (p) a] = pgr (pa) = pδ (pa) . It follows that δ (pR) ⊂ pR and that δpR is a local derivation. Hence δpR is a derivation. Thus, for every a, b ∈ R and every p ∈ P, we have pδ (ab) = δ (pab) = δ ((pa) (pb)) = paδ (pb) + δ (pa) pb = p [aδ (b) + δ (a) b] . Hence p [δ (ab) − [aδ (b) + δ (a) b]] = 0. Since P is separating, we see that δ is a derivation on R. (cid:3) An abelian von Neumann algebra M is discrete if it is generated by its minimal nonzero projections; equivalently, if the identity operator in M is the sum of the minimal projections in M. Since M is abelian, it follows that QM = CQ for every minimal projection Q in M. Every von Neumann algebra on a Hilbert space H is the direct sum of algebras Mn with 1 ≤ n < ∞ (the finite type In summands) and a von Neumann algebra M∞ which is the direct sum of algebras of type I∞, II, and III (not all summands need be present). Call the corresponding central projections Pn with 1 ≤ n ≤ ∞. Theorem 1. Suppose M is a von Neumann algebra on a Hilbert space H. Then (1) If M1 = 0 and R is a ring containing M with the same identity as M such that P = {Pn : 2 ≤ n ≤ ∞} is a separating family of central idempotents for R, then every local derivation on R is a derivation. (2) If M1 is discrete, then every local derivation on the algebra of closed densely defined operators affiliated with M is a derivation. Proof. (1) . By [9, Theorem 6.6.5], we know that PnM contains an n × n system of matrix units for 2 ≤ n < ∞, and it follows from [9, Lemma 6.5.6] that P∞M contains a 2 × 2 system of matrix units. Since PnM ⊂ PnR is a unital embedding and P is separating, it follows from the two lemmas above that every local derivation on R is a derivation. LOCAL DERIVATIONS 3 (2) . Let R be the algebra of all measurable operators affiliated with M. How- ever, since M1 is discrete, P1 is the orthogonal sum of a family {Qλ : λ ∈ Λ} of minimal projections and that, for each λ ∈ Λ, QλM = CQλ, which means that QλM′ = B (QλH) Qλ, which, in turn, implies that QλR = CQλ, so ev- ery local derivation on QλR is a derivation. Since Since the elements of R are densely defined operators on H, and Pλ∈Λ Qλ + P2≤n≤∞ Pn = 1, it follows that {Qλ : λ ∈ Λ} ∪ {Pn : 2 ≤ n ≤ ∞} is a separating family of central idempotents for R. Arguing as in the proof of (1) we can apply the lemmas to see that every local derivation on R is a derivation. (cid:3) In [1, Theorem 3.8], the authors give necessary and sufficient conditions on a commutative von Neumann algebra M for the existence of local derivations which are not derivations on the algebra S(R) of measurable operators affiliated with M. For a von Neumann algebra M, we can define the set S(M) of all measurable operators affiliated with M and the set LS(M) of all local measurable operators affiliated with M In [12], Muratov and Chilin show that LS(M) is a unital ∗-algebra when equipped with algebraic operations of the strong addition, multiplication, and tak- ing the adjoint of an operator and S(S) is a unital ∗-subalgebra of LS(M). Suppose that M is a von Neumann algebra with a faithful normal semi-finite trace τ . Let S(M, τ ) denote the algebra of all τ -measurable operators affiliated with M. It is clear that M ⊆ S(M) ⊆ LS(M) ( see for details [1, 2, 13]). In Theorem 1, if we choose that R is M, S(M) or LS(M), we can obtain [1, Theorem 2.5 and Proposition 2.7]. By Theorem 1 and [1, Theorem 3.8], we can completely answer [15, Conjecture 48]. Acknowledgement The first author is partially supported by a travel grant from the Simons Foun- dation and the second author is supported by NSF of China. References [1] S. Albeverio, Sh. A. Ayupov, K. K Kudaybergenov and B. O. Nurjanov, Local derivations on algebras of measurable operators, Commun. in Contemp. Math. 13(4) (2011), 643 -- 657. [2] Sh. A. Ayupov and K. K Kudaybergenov, Derivations on algebras of measurable operators, Infin. Dimens. Ana. Quantum Probab. Relat. Top. 13(2) (2010), 305 -- 337. [3] M. Bresar, Characterizing homomorphisms, derivations and multipliers in rings with idem- potents, Proc. Roy. Soc. Edinburgh Sect.A 137 (2007), 9 -- 21. [4] D. Hadwin and J. Li, Local derivations and local automorphisms, J. Math. Anal. Appl. 290 (2003), 702 -- 714. [5] D. Hadwin and J. Li, Local derivations and local automorphisms on some algebras, J. Oper. Theory 60 (2008), 29 -- 44. [6] B. Johnson, Local derivations on C∗-algebras are derivations, Trans. Amer. Math. Soc. 353 (2001), 313 -- 325. [7] B. Johnson, Symmetric amenability and the nonexistence of Lie and Jordan derivations, Math. Proc. Cambridge Philos. Soc. 120 (1996), 455 -- 473. [8] R. Kadison, Local derivations, J. Algebra 130 (1990), 494 -- 509. [9] R. Kadison and J. R. Ringrose, Fundamentals of the Theory of Operator Algebras II, Grad- uate Studies in Math. 16. [10] R. Kadison and Zhe Liu, Derivations of Murray-von Neumann algebras, Math. Scand. To appear. 4 DON HADWIN, JIANKUI LI, QIHUI LI, AND XIUJUAN MA [11] R. Kadison and Zhe Liu, A note on derivations of Murray-von Neumann algebras, To ap- pear. [12] M. A. Muratov and V. I. Chilin, Algebras of Measurable and Local Measurable Operators, (Ukrainian Academy of Sciences, 2007). [13] M. A. Muratov and V. I. Chilin, ∗-algebras of unbounded operators affiliated with a von Neumann algebra, J. Math. Sci. 140 (2007), 445 -- 451. [14] I. E. Segal, A non-commutative extension of abstract integration, Ann. Math. 57 (1953), 401 -- 457. [15] http://www.math.tamu.edu/kerr/workshop/problems 2012.pdf Department of Mathematics, University of New Hampshire, Durham, NH 03824, USA E-mail address: [email protected] URL: http://euclid.unh.edu/~don Department of Mathematics, East China University of Science and Technology, Shanghai 200237, China E-mail address: [email protected] Department of Mathematics, East China University of Science and Technology, Shanghai 200237, China E-mail address: [email protected] Department of Mathematics, Hebei University of Technology, Tianjing, 300130, China E-mail address: [email protected]
1905.12447	1	1905	2019-05-28T14:07:44	On the Decomposition Theorems for C*-algebras	[ "math.OA" ]	Elliott dimension drop interval algebra is an important class among all $C^*$-algebras in the classification theory. Especially, they are building stones of $\mathcal{AHD}$ algebra and the latter contains all $AH$ algebras with the ideal property of no dimension growth. In this paper, we will show two decomposition theorems related to the Elliott dimension drop interval algebra. Our results are key steps in classifying all $AH$ algebras with the ideal property of no dimension growth.	math.OA	math	ON THE DECOMPOSITION THEOREMS FOR C ∗-ALGEBRAS Chunlan Jiang1 Liangqing Li2, and Kun Wang3∗ Abstract Elliott dimension drop interval algebra is an important class among all C ∗-algebras in the classification theory. Especially, they are building stones of AHD algebra and the latter contains all AH algebras with the ideal property of no dimension growth. In this paper, we will show two decomposition theorems related to the Elliott dimension drop interval algebra. Our results are key steps in classifying all AH algebras with the ideal property of no dimension growth. Keywords: C ∗-algebra, Elliott dimension drop interval algebra, decomposition theorem, spectral distribution property AMS subject classification: Primary: 46L35, 46L80. 1 Introduction Classification theorems have been obtained for AH algebras -- the inductive limits of cut downs of matrix algebras over compact metric spaces by projections -- and AD algebras -- the inductive limits of Elliott dimension drop interval algebras in two special cases: 1. Real rank zero case: all such AH algebras with no dimension growth and such AD algebras (See [4], [12], [7], [8], [13], [1], [14]- [17], [3], and [2]); 2. Simple case: all such AH algebras with no dimension growth (which includes all simple AD algebras by [11]) (See [5], [6], [33], [42], [43], [26]- [29], [18], and [9]). In [9], the authors pointed out two important possible next steps after the com- pletion of classification of simple AH algebras (with no dimension growth). One of these is the classification of simple ASH algebras -- the simple inductive limits of 1College of Mathematics and Information Science, Hebei Normal University, Shijiazhuang, 050024, China. E-mail: [email protected] 2Department of Mathematics, University of Puerto Rico at Rio Piedras, PR 00936, USA. E-mail: [email protected] 3Department of Mathematics, University of Puerto Rico at Rio Piedras, PR 00936, USA. E-mail: [email protected] ∗Corresponding author. 1 subhomogeneous algebras (with no dimension growth). The other is to generalize and unify the above-mentioned classification theorems for simple AH algebras and real rank zero AH algebras by classifying AH algebras with the ideal property. In this article, we have achieved several key results for the second goal by providing two decomposition theorems. As in [8], let TII,k be the 2-dimensional connected simplicial complex with H 1(TII,k) = 0 and H 2(TII,k) = Z/kZ, and let Ik be the subalgebra of Mk(C[0, 1]) defined by Ik = {f ∈ Mk(C[0, 1]) : f (0) ∈ C · 1k and f (1) ∈ C · 1k}. This algebra is called an Elliott dimension drop interval algebra. Denote by HD the class of algebras consisting of direct sums of building blocks of the forms Ml(Ik) and P Mn(C(X))P , with X being one of the spaces {pt}, [0, 1], S1, and TII,k, and with P ∈ Mn(C(X)) being a projection. (In [2], this class is denoted by SH(2), and in [23], this class is denoted by B). We will call a C ∗-algebra an AHD algebra, if it is an inductive limit of algebras in HD. In [20], [21], [29], and [24], it is proved that all AH algebras with the ideal property of no dimension growth are inductive limits of algebras in the class HD -- that is, they are AHD algebras. By this reduction theorem, to classify AH algebras with the ideal property, we must study the properties of homomorphisms between those basic building blocks. In the local uniqueness theorem for classification, it requires the homomorphisms involved to satisfy a certain spectral distribution property, called the sdp property (more specifically, sdp(η, δ) property introduced in [18] and [9] for some positive real numbers η and δ). This property automatically holds for the homomorphisms φn,m (provided that m is large enough) giving rise to a simple inductive limit procedure. But for the case of general inductive limit C ∗-algebras with the ideal property, to obtain this sdp property, we must pass to certain good quotient algebras which cor- responding to simplicial sub-complexes of the original spaces; a uniform uniqueness theorem, that does not depend on the choice of simplicial sub-complexes involved, is required. For the case of an interval, whose simplicial sub-complexes are finite unions of subintervals and points, such a uniform uniqueness theorem is proved in [22] (see [27] and [5] also). But for the general case, there are no uniqueness theorem for the general case involving arbitrary finite subsets of Mn(C(TII,k)) (or Ml(Ik)). In this paper, we prove decomposition theorems between such building blocks or between a building block of this kind and a homogeneous building block. And we will compare the decompositions of two different homomorphisms in the last part of chapter 4. Such decomposition and comparison results will be used in the proof of the uniqueness theorem for AH algebras with the ideal property in [19] by Gong, Jiang and Li. 2 2 Notation and terminology In this section, we will introduce some notation and terminology. Definition 2.1. Let X be a compact metric space and ψ : C(X) → P Mk1(C(Y ))P (with rank(P ) = k) be a unital homomorphism. For any point y ∈ Y , there are k mutually orthogonal rank 1 projections p1, p2, · · · , pk with {x1(y), x2(y), · · · , xk(y)} ⊂ X (may be repeat) such that pi = P (y) and kPi=1 ψ(f )(y) = kXi=1 f (xi(y))pi, ∀f ∈ C(X). We denote the set {x1(y), x2(y), · · · , xk(y)} (counting multiplicities), by Spψy. We shall call Spψy the spectrum of ψ at the point y. 2.2. For any f ∈ Ik ⊂ Mk(C[0, 1]) = C([0, 1], Mk(C)) as in 3.2 of [13], let function f : [0, 1] −→ C ⊔ Mk(C) (disjoint union) be defined by That is, f (t) is the value of irreducible representation of f corresponding to the point t. Similarly, for f ∈ Ml(Ik), we can define f : [0, 1] −→ Ml(C) ⊔ Mlk(C), by if t = 0 and f (0) = λ1k if t = 1 and f (1) = µ1k if 0 < t < 1 . if t = 0 and f (0) = a ⊗ 1k if t = 1 and f (1) = b ⊗ 1k if 0 < t < 1 . λ, µ, f (t), f (t) =  f (t) =  P (y) = u(cid:18) 1rank(P ) 0 a, b, f (t), 0 2.3. Suppose that φ : Ik −→ P Mn(C(Y ))P is a unital homomorphism. Let r = rank(P ). For each y ∈ Y , there are t1, t2, · · · , tm ∈ [0, 1] and a unitary u ∈ Mn(C) such that 0 (cid:19) u∗   0n−r and φ(f )(y) = u for all f ∈ Ik. f (t1) f (t2) . . .   f (tm) 3 u∗ ∈ P (y)Mn(C)P (y) (1) 2.4. Let φ be the homomorphism defined by the equation (2.1) above with t1, t2, · · · , tm as appeared in the diagonal of the matrix. We define the set Spφy to be the points t1, t2, · · · , tm with possible fraction multiplicity. If ti = 0 or 1, we will assume that the multiplicity of ti is 1 k ; if 0 < ti < 1, we will assume that the multiplicity of ti is 1. For example if we assume t1 = t2 = t3 = 0 < t4 ≤ t5 ≤ · · · ≤ tm−2 < 1 = tm−1 = tm, then Spφy = {0∼ 1 k , 0∼ 1 k , 0∼ 1 k , t4, t5, · · · , tm−2, 1∼ 1 k , 1∼ 1 k }, which can also be written as Spφy = {0∼ 3 k , t4, t5, · · · , tm−2, 1∼ 2 k }. Here we emphasize that, for t ∈ (0, 1), we do not allow the multiplicity of t to be non-integral. Also for 0 or 1, the multiplicity must be multiple of 1 k (other fraction numbers are not allowed). Let ψ : C[0, 1] −→ P Mn(C(Y ))P be defined by the following composition ψ : C[0, 1] ֒→ Ik φ −→ P Mn(C(X))P, where the first map is the canonical inclusion. Then we have Spψy = {Spφy}∼k -- that is, for each element t ∈ (0, 1), its multiplicity in Spψy is exactly k times of the multiplicity in φy. 2.5. (a) we use ♯(.) to denote the cardinal number of a set. Very often, the sets under consideration will be sets with multiplicity, in which case we shall also count multiplicity when we use the notation ♯. The set may also contain fractional point. For example, ♯{01, 12, 0, 0, 1} = 5. (b) We shall use a∼k to denote a, a, · · · a . For example {a∼3, b∼2} = {a, a, a, b, b}. (c) For any metric space X, any x0 ∈ X and c > 0, let Bc(x0) , {x ∈ Xd(x, x0) < c}, the open ball with radius c and center x0. (d) Suppose that A is a C ∗-algebra, B ⊂ A a subset (often a subalgebra), F ⊂ A is a finite subset and ε > 0. If for each element f ∈ F , there is an element g ∈ B such that kf − gk < ε, then we shall say that F is approximately contained in B to within ε, and denote this by F ⊂ε B. 4 k {z } (e) Let X be a compact metric space. For any δ > 0, a finite set {x1, x2, · · · , xn} is said to be δ-dense in X if for any x ∈ X, there is xi ∈ {x1, x2, ..., xn} such that dist(x, xi) < δ. (f) We shall use • or •• to denote any possible positive integers. (g) For any two projections p, q ∈ A, by [p] ≤ [q] we mean that p is unitarily equivalent to a sub-projection of q. And we use p ∼ q to denote that p is unitarily equivalent to q. 2.6. Let A = Ml(Ik). Then every point t ∈ (0, 1) corresponds to an irreducible representation πt, defined by πt(f ) = f (t). The representations π0 and π1 defined by π0 = f (0) and π1 = f (1) are no longer irreducible. We use 0 and 1 to denote the corresponding points for the irreducible representations. That is, π0(f ) = f (0), and π1(f ) = f (1). Or we can also write f (0) , f (0) and f (1) , f (1). Then the equation (∗) could be written as φ(f )(y) = u f (t1)   f (t2) . . . f (tm) 0n−r   u∗, where some of ti may be 0 or 1. In this notation, up to unitary equivalence, f (0) is equal to diag(f (0), f (0), · · · , f (0) ) . k {z } Under this notation, we can also write 0∼ 1 k as 0. Then the example of Spφy in 2.4 can be written as Spφy = {0∼ 1 k , 0∼ 1 k , 0∼ 1 k , t4, t5, · · · , tm−2, 1∼ 1 k , 1∼ 1 k } = {0, 0, 0, t4, t5, · · · , tm−2, 1, 1}. 2.7. For a homomorphism φ : A −→ Mn(Ik), where A = Ik or C(X), and for any t ∈ [0, 1], define Spφt = Spψt, where ψ is defined by the composition ψ : A φ−→ Mn(Il) → Mnl(C[0, 1]). Also Spφ0 = Sp(π0 ◦ φ). Hence, Spφ0 = {Spφ0}∼k. 5 2.8. Let φ : Mn(A) −→ B be a unital homomorphism. It is well known (see 1.34 and 2.6 of [8]) that there is an identification of B with (φ(e11)Bφ(e11)) ⊗ Mn(C) such that φ = φ1 ⊗ idn : Mn(A) = A ⊗ Mn(C) −→ (φ(e11)Bφ(e11)) ⊗ Mn(C) = B, where e11 is the matrix unit of upper left corner of Mn(A) and φ1 = φe11Mn(A)e11 : A −→ φ(e11)Bφ(e11). If we further assume that A = Ik or C(X) (with X being a connected CW complex) and B is either QMn(C(Y ))Q or Ml(Ik1), then for any y ∈ SpB, define Spφy , Sp(φ1)y. Here, we use the standard notation that if B = P Mm(C(Y ))P then SpB = Y ; and if B = Ml(Ik), then Sp(B) = [0, 1]. 2.9. Let A and B be either of form P Mn(C(X))P (with X path connected) or of form Ml(Ik). Let φ : A −→ B be a unital homomorphism, we say that φ has property sdp(η, δ) (spectral distribution property with respect to η and δ) if for any η-ball Bη(x) = {x′ ∈ X dist(x′, x) < η)} ⊂ X(= Sp(A)) and any point y ∈ Sp(B), ♯(Spφy ∩ Bη(x)) ≥ δ · ♯Spφy, counting multiplicity. If φ is not unital, we say that φ has sdp(η, δ) if the correspond- ing unital homomorphism φ : A −→ φ(1A)Bφ(1A) has property sdp(η, δ). 2.10. Set P nX = X × X × · · · × X / ∼, where the equivalence relation ∼ is defined n {z } by (x1, x2, · · · , xn) ∼ (x′ 1, x′ 2, · · · , x′ n) if there is a permutation σ of {1, 2, · · · , n} such that xi = x′ A metric d on X can be extended to a metric on P nX by σ(i) for each 1 ≤ i ≤ n. d([x1, x2, · · · , xn], [x′ 1, x′ 2, · · · , x′ n]) = min max 1≤i≤n d(xi, x′ σ(i)), σ where σ is taken from the set of all permutations, and [x1, x2, · · · , xn] denote the equivalence class of (x1, x2, · · · , xk) in P kX. 2.11. Let X be a metric space with metric d. Two k-tuple of (possible repeating) points {x1, x2, · · · , xn} ⊂ X and {x′ 1, x′ n} ⊂ X are said to be paired within η if there is a permutation σ such that 2, · · · , x′ d(xi, x′ σ(i)) < η, i = 1, 2, · · · , k. 6 This is equivalent to the following statement. [x′ n] as points in P nX, then 2, · · · , x′ 1, x′ If one regards [x1, x2, · · · , xn] and d([x1, x2, · · · , xn], [x′ 1, x′ 2, · · · , x′ n]) < η. 2.12. For X = [0, 1], let P (n,k)X, where n, k ∈ Z+\{0}, denote the set of n k elements from X, in which only 0 or 1 may appear fractional times. That is, each element in X is of the form {0∼ n0 k , t1, t2, · · · , tm, 1∼ n1 k } with 0 < t1 ≤ t2 ≤ · · · ≤ tm < 1 and n0 k + m + n1 k = n k . An element in P (n,k)X can always be written as k , t1, t2, · · · , ti, 1∼ k1 {0∼ k0 k }, (2) (3) where 0 ≤ k0 < k, 0 ≤ k1 < k, 0 ≤ t1 ≤ t2 ≤ · · · ≤ ti ≤ 1 and k0 k + i + k1 ti could be 0 or 1.) In the above representations 2 and 3, we know that k = n k . (Here Let and k0 ≡ n0 (mod k) and k1 ≡ n1 (mod k). y = [0∼ k0 k , t1, t2, · · · , ti, 1∼ k1 k ] ∈ P (n,k)X with k0, k1, k′ 0, k′ k′ 0 k , t′ y′ = [0∼ 1, t′ 1 ∈ {0, 1, · · · , k − 1}. 2, · · · , t′ i, 1∼ k′ 1 k ] ∈ P (n,k)X, We define dist(y, y′) as the following: if k0 6= k′ 0 or k1 6= k′ 1, then dist(y, y′) = 1; if k0 = k′ 0 and k1 = k′ 1 (consequently i = i′), then dist(y, y′) = max 1≤j≤i tj − t′ j, as we order the {tj} and {t′ i, respectively. Note that P (n,1)X = P nX with the same metric. Let φ, ϕ : Ik −→ Mn(C) be two unital homomorphisms. Then Spφ and Spψ define two elements in P (n,k)[0, 1]. We say that Spφ and Spψ can be paired within η, if dist(Spφ, Spψ) < η. j} as t1 ≤ t2 ≤ · · · ≤ ti and t′ 2 ≤ · · · ≤ t′ 1 ≤ t′ Note that if dist(Spφ, Spψ) < 1, then KK(φ) = KK(ψ). 2.13. Let A = P Mk(C(X))P , or Ml(Ik) and X1 ⊂ Sp(A) be a closed subset -- that is, X1 is a closed subset of X or of [0, 1]. We define AX1 to be the quotient algebra A/I, where I = {f ∈ A, f X1 = 0}. Evidently Sp(AX1) = X1. If B = QMk(C(Y ))Q, φ : A −→ B is a homomorphism, and Y1 ⊂ Sp(B)(= Y or [0, 1]) is a closed subset, then we use φY1 to denote the composition. φY1 : A φ −→ B → BY1. 7 If Sp(φY1) ⊂ X1 ∪ X2 ∪ · · · ∪ Xk, where X1, X2, · · · , Xk are mutually disjoint closed subsets of X, then the homomorphism φY1 factors as A −→ AX1∪X2∪···∪Xn = nMi=1 AXi −→ BY1. Y1 to denote the part of φY1 corresponding to the map AXi −→ BY1. We will use φXi Hence φY1 =Li φXi Y1 . 3 Decomposition Theorem I In this section, we will prove the following theorem. Theorem 3.1. Let F ⊂ Ik be a finite set, ε > 0. There is an η > 0, satisfying that if φ : Ik → P M•(C(X))P (dim(X) ≤ 2) is a unital homomorphism such that for any x ∈ X, ♯(Spφ′ x ∩ [0, η 4 ]) ≥ k and ♯(Spφ′ x ∩ [1 − η 4 ], 1]) ≥ k, where φ′ : C[0, 1] ı−→ Ik φ−→ P M(C(X))P, then there are three mutually orthogonal projections with Q0, Q1, P1 ∈ P M•(C(X))P Q0 + Q1 + P1 = P and a unital homomorphism ψ1 : Mk(C[0, 1]) → P1M•(C(X))P1 such that (1) write ψ(f ) = f (0)Q0 + f (1)Q1 + (ψ1 ◦ ı)(f ), then k φ(f ) − ψ(f ) k< ε for all f ∈ F ⊂ Ik ⊂ Mk(C[0, 1]), and (2) rank(Q0) ≤ k and rank(Q1) ≤ k. We will divide the proof into several steps. 8 3.2. Let η > 0 (and η < 1) be such that if t − t′ < η, then k f (t) − f (t′) k< ε 6 for all f ∈ F . We will prove that this η is as desired. Let a unital homomorphism φ : Ik → P M•C(X)P satisfy that ♯(Spφx ∩ [0, η 4 , 1]) ≥ k for each x ∈ X, we will prove such φ has the decomposition as desired. 3.3. Let rank(P ) = n. And let ei,j ∈ Mn(C) be the matrix units. For any closed set Y ⊂ [0, 1], define hY ∈ C[0, 1] ⊂ Ik (considering C[0,1] as in the center of Ik) as 4 ]) ≥ k and ♯(Spφx ∩ [1 − η hY (t) =   1, 1 − 0, 12n η if t ∈ Y dist(t, x), if dist(t, x) ≤ if dist(t, x) ≥ η 12n η 12n . 12n , 1 − η Define H ′ = {hY Y is closed} ∪ {hY eij Y ⊂ [ η 12n ] is closed}. Note that for a closed set Y ⊂ [ η hY (0) = hY (1) = 0, and therefore hY eij ∈ Ik. Note also that the family H ′ is equally continuous. There is a finite set H ⊂ H ′ satisfying that for any h′ ∈ H ′, ∃h ∈ H such that 12n , 1 − η 12n ], k h − h′ k≤ ε 12(n + 1)2 . For finite set H ∪ F , ε > 0, and φ : Ik → P M•(C(X))P , there is a τ > 0 such that the following are true: (a) For x, x′ ∈ X with dist(x′, x) < τ , Spφx and Spφx′ can be paired within η 24n2 . This is equivalent to the condition that Spφ′x can be paired with Spφ′x′ to within 24n2 (since KK(φx) = KK(φx′)), where φ′ = φ ◦ ı is as the above. (b) For x, x′ ∈ X with dist(x′, x) < τ , η kφ(h)(x) − φ(h)(x′)k ≤ ε 12(n + 1)2 , regarding φ(h)(x) ∈ P (x)M•(C)P (x) ⊂ M•(C) and φ(h)(x′) ∈ P (x′)M•(C)P (x′) ⊂ M•(C). In particular, kP (x) − P (x′)k < 3.4. Choose any simplicial decomposition on X such that for any simplex ∆ ⊂ X, the set 12(n+1)2 since 1 ∈ H. ε Star(△) = ∪{ ◦ ∆′ ∆′ is a simplex of X with ∆′ ∩ ∆ 6= ∅} has diameter at most τ 3.5. We will construct the homomorphism ψ : Ik → P M•(C(X))P which is of the form ◦ ∆′ is the interior of the simplex ∆′. 2 , where ψ(f ) = f (0)Q0 + f (1)Q1 + ψ1(f ) 9 as described in the theorem. Our construction will be carried out simplex by simplex. First, define the restriction of map ψ to P M•(C(X))P v = P (v)M•(C)P (v) for each vertex v ∈ X. The homomorphism is denoted by ψ{v} : Ik → P (v)M•(C)P (v). (Here and below, we refer the reader to 2.13 for the notation ψX1 for a subset X1 ⊂ X.) Next, we will define, for each 1-simplex [a, b] ⊂ X, the homomorphisms ψ[a,b] : Ik → P [a,b]M•(C([a, b]))P [a,b] which will give the same maps as the previously defined maps ψ{a} and ψ{b} on the boundary {a, b}. Finally, we will define, for each 2-simplex ∆ ⊂ X, the homomor- phism ψ∆ : Ik → P ∆M•(C(∆))P ∆ such that ψ∂△ should be the same as what previously defined. 3.6. For each simplex ∆ of any dimension, let C∆ denote the center of the simplex. That is, if ∆ is a vertex v, then C∆ = v; if ∆ is a 1-simplex identified with [a,b], then C∆ = a+b 2 ; and if ∆ is a 2-simplex identified with a triangle in R2 with vertices {a, b, c} ⊆ R2, then C∆ = a+b+c 3.7. According to each simplex ∆ (of possible dimensions 0, 1, or 2), we will divide φ the set Spφ′∆ ⊂ [0, 1] into pieces, where φ′ : C[0, 1] ֒→ Ik −→ P M•C(X)P . (Recall Spφ′x = {Spφx}∼k, and Spφ′x has no fractional multiplicity). So for each x ∈ X, 3 ∈ R2 which is barycenter of ∆. Spφ′x = n = rank(P ) (counting multiplicity). If we order Spφ′x as 0 ≤ λ1(x) ≤ λ2(x) ≤ · · · ≤ λn(x) ≤ 1, then all functions λi are continuous functions. By path connectedness of simplex ∆, the set Spφ∆ can be written as Spφ∆ = [a0, b0] ∪ [a1, b1] ∪ · · · ∪ [ak′−1, bk′−1] ∪ [ak′, bk′] with 0 ≤ a0 ≤ b0 < a1 ≤ b1 < a2 ≤ b2 < · · · < ak′−1 ≤ bk′−1 < ak′ ≤ bk′ ≤ 1. (Note that, if ai = bi, then [ai, bi] = {ai} is a degenerated interval.) We will group the above intervals into groups T0 ∪ T1 ∪ · · · ∪ Tlast such that Spφ∆ = ∪Tj, with the condition that for any λ ∈ Tj, µ ∈ Tj+1, we have λ < µ, 10 4 + η according to the following procedure: (i) Spφ∆ ∩ [0, η should be in the group T0; and Spφ∆ ∩ [1 − ( η 4 + η with bi ≥ 1 − ( η (ii) If ai − bi−1 ≤ η (iii) If ai − bi−1 > η [ai, bi] are in different groups, say, Tj and Tj+1. 4 + η 12n ] ⊂ T0, that is, all the above intervals [ai, bi] with ai ≤ η 4 + η 12n 12n ), 1] ⊂ Tlast, that is all [ai, bi] 4 + η 12n ) will be grouped into the last group Tlast; 12n , then [ai−1, bi−1] and [ai, bi] are in the same group, say Tj; 12n , ai > η 12n and bi−1 < 1 − ( η 12n ), then [ai−1, bi−1] and 4 + η Denote Tlast by Tl∆ (i.e., l∆ = last) -- if there is no confusion, we will call Tl∆ by Tl. Let t0 = 0, s0 = max{ η 4 , min Tl}, sl = 1; and for 1 < i < l, let ti = min Ti, and si = max Ti. Then Ti ⊂ [ti, si]. With the above notation, we have the following lemma. 4 , max T0}, tl = min{1 − η 4 + η 6 ; Lemma 3.8. With the above notation, we have the following (a) length [t0, s0] ≤ η (b) length [tl, sl] ≤ η/4 + η/6; (c) length [ti, si] ≤ η/6 for i ∈ {1, 2, · · · , l − 1}; (d) ti+1 − si > η Proof. From (ii) of 3.7, we know that min Ti+1 − max Ti > η know that min T1 > η 12n for i ∈ {0, 1, 2, · · · , l − 1}. 12n and max Tl−1 < 1 − ( η 4 + η 4 + η 12n ; and from (i), we 12n ). Hence (d) holds. The following fact is well known. Fact: For any two sequences 0 ≤ λ1 ≤ λ2 ≤ · · · ≤ λn ≤ 1 and 0 ≤ µ1 ≤ µ2 ≤ i=1 can be paired within σ if and only if λi − µi < σ i=1 and {µi}n · · · ≤ µn ≤ 1, {λi}n for all i ∈ {1, 2, · · · , n}. Note that ∆ is path connected and Spφ∆ = i 6= j. We conclude that for any z, z′ ∈ ∆ and i, [ai, bi] with [ai, bi] ∩ [aj, bj] = ∅ if k′Si=1 ♯(Spφz ∩ [ai, bi]) = ♯(Spφz′ ∩ [ai, bi]) counting multiplicity. In our construction, we know that Spφz and Spφz′ can be paired within η/24n2, using the above mentioned fact. We know also that and can be paired within η/24n2. Consequently Spφz ∩ [ai, bi] Spφz′ ∩ [ai, bi] [ai, bi] ⊂η/24n2 [ai, bi] ∩ SpφC∆, where C∆ is the center of simplex ∆. Note that SpφC∆ ∩ [ai, bi] is a finite set with at most n points in [0, 1] and η/24n2-neighborhood of each point is a closed interval of length at most (η/24n2) · 2 = η/12n2. Hence we have length[ai, bi] ≤ (η/12n2) · n = η/12n. 11 Furthermore, each Tj contains at most n intervals [ai, bi]. And for each consecutive pair of intervals in Tj (0 < j < l), we have [ai, bi] ∪ [ai+1, bi+1] ⊂(cid:0)η 4 + η 12n , 1 − ( η 4 + η 12n )(cid:1) and the distance between them ai+1 − bi ≤ η/12n. That is, the gap between them is at most η/12n. Hence for each i ∈ {1, 2, · · · , l − 1}, the length of [ti, si] is at most n · η 12n + (n − 1) · η 12n < η/6 (at most n possible intervals and n − 1 gaps). Also, and length[t0, s0] < η 4 + η 6 length[tl, sl] < η 4 + η 6 . 3.9. For each simplex ∆ with face ∆′ ⊂ ∆, we use Ti(∆) and Tj(∆′) to denote the sets [ti(∆), si(∆)] or [tj(∆′), sj(∆′)] as in 3.7, corresponding to ∆ and ∆′. Then evidently, the decomposition Spφ∆′ =[j (Tj(∆′) ∩ Spφ∆′), is a refinement of the decomposition Spφ∆ = ∪(Ti(∆) ∩ Spφ∆) -- that is, if two elements λ, µ ∈ Spφ∆′ are in the set Tj(∆′) for a same index j, then they are in the set Ti(∆) for a same index i. 3.10. For each simplex ∆, consider the homomorphism φ : Ik → P M•(C(∆))P = A∆ . Since Spφ∆ ⊂Sl j=0 Tj(∆) = [tj, sj], φ factors through as lSj=1 Ik → ⊕l j=0Ik[tj,sj] ⊕φj−−→ P M•(C(∆))P. Let Pj(x) = φj(1k[tj,sj])(x) for each x ∈ ∆. Then Pj(x) are mutually orthogonal projections satisfying lXj=0 Pj(x) = P (x). By the assumption of Theorem 3.1, we have rank(P0) ≥ k and rank(Pl) ≥ k. 12 3.11. Now we define ψ : Ik → A∆ simplex by simplex, starting with vertices -- the zero dimensional simplices. Let v ∈ X be a vertex. As in 3.7, we write Spφ{v} = l[i=0 [ti, si]\ Spφ{v}, where 0 = t0 < s0 < t1 ≤ s1 < · · · < tl−1 ≤ sl−1 < tl < sl = 1, with [0, η/4] ⊂ [t0, s0] ⊂ [0, η/2], [1 − η/4, 1] ⊂ [tl, sl] ⊂ [1 − η/2, 1], 0 ≤ si − ti < η/6 for each i ∈ {1, 2, · · · , l − 1}, and ti+1 − si > η/12n for each i ∈ {0, 1, 2, · · · , l − 1}. Recall that φ{v} : Ik → P (v)M•(C)P (v) (as in 3.10) can be written as φ{v} = diag(φ0, φ1, · · · , φl) : Ik → lMi=0 PiM•(C)Pi ⊂ P (v)M•(C)P (v), where φi = φ[ti,si] {v} below, we refer the reader to 2.13 for the notation φZj : Ik → Ik[ti,si] → PiM•(C)Pi and P (v) = Pl i=0 Pi. (Here and X1 (X1 ⊂ X), which makes sense, provided that Sp(φX1) ⊂ Sj Zj, where {Zj} are mutually disjoint closed subsets of the spectrum of the domain algebra of φ.) From now on, we will use diag0≤i≤l (φi) to denote diag(φ0, φ1, · · · , φl). Define ψi : Ik[ti,si] → PiM•(C)Pi by ψi = φi if 1 ≤ i ≤ l − 1, (That is, we do not modify φi for 1 ≤ i ≤ l − 1.) For i = 0 (the case i = l is similar) we do the following modification. There is a unitary u ∈ M•(C) such that 13 f (0) f (0) . . .   f (0)  j×j f (ξ1) φ0(f )(v) = u   f (ξ2) . . . f (ξ••) 0 . . . 0   u∗, where ξi ∈ (0, s1], 0 < ξ1 ≤ ξ2 ≤ · · · ≤ ξ•• ≤ s1. Or write it as φ0(f )(v) = udiag(f (0)∼j, f (ξ1), f (ξ2), · · · , f (ξ••), 0, · · · , 0)u∗. If 0 < j ≤ k then we do not do any modification and just let ψ0 = φ0. If j > k, then write j = kk′ + j′ with 0 < j′ ≤ k, choose ξ′ ∈ (0, ξ1), and define ψ0(f )(v) = udiag(f (0)∼j′ , f (ξ′)∼k′ , f (ξ1), f (ξ2), · · · , f (ξ••), 0, · · · , 0)u∗. That is, change kk′ terms of f (0) in the diagonal of the definition of φ0 to k′ terms of the form f (ξ′). If j = 0, then we change ξ1 to 0, that is, ψ0(f )(v) = udiag(f (0)∼k, f (ξ2), · · · , f (ξ••), 0, · · · , 0)u∗. Since ξ′ − 0 < η (see 3.2). We modify φl in a similar way to define ψl. Let 2 and ξ1 − 0 < η 2 , we have kφ0(f ) − ψ0(f )k < ε 6, for all f ∈ F ψ{v} = diag(ψ0, ψ1, · · · , ψl) : Ik → P (v)M•(C)P (v), {v} . Then kφ(f ) − ψ(f )k < ε where ψi = ψ[ti,si] Remark 3.12. Let us emphasize that the homomorphisms ψi are the same as φi for i ∈ {1, 2, · · · , l{v} − 1}. But we do modify φ0 and φl (l = l{v}) to get ψ0 and ψl. 6 for all f ∈ F . Also, we have Sp(ψ0) ⊂ [0, s0] and Sp(ψl{v}) ⊂ [tl{v}, 1]. Furthermore, ψi(1) = φi(1) for any i, and consequently ψ(1) = φ(1). 14 3.13. Now consider 1-simplex ∆ = [a, b] ⊂ X. We need to define ψ∆ = ψ[a,b] from previously defined ψ{a} and ψ{b}. According to 3.7, write Spφ∆ = Spφ∆∩Tj(∆) l∆Sj=1 with T0(∆) = [0, s0(∆)] and Tl∆(∆) = [tl∆(∆), 1]. Recall that in the definition of ψ{a}, ψ{b}, we use the decomposition and φ{a} = diag1≤j≤l{a}(φTj({a}) {a} ) φ{b} = diag1≤j≤l{b}(φTj({b}) {b} ) and only modified φ0 = φ[0,s0{a}] {a} (or φ[0,s0{b}] {b} ) and φl{a} = φ For ∆ = [a, b], let us consider the decomposition ({a}),1] [tl{a} {a} (or φ [tl{b} {b} ({b}),1] ). φ∆ = l∆Mj=1 φ[tj(∆),sj (∆)] ∆ . From the above, we know that for any 0 < j < l∆, the definition of ψ[tj(∆),sj (∆)] the same as (φ[tj(∆),sj (∆)] ){a}, since the decomposition {a} ∆ is Spφ{a} = l{a}[j=1 Tj({a}) ∩ Spφ{a} is finer than the decomposition Spφ{a} = l∆[j=1 Tj(∆) ∩ Spφ{a} (see 3.9) and only partial maps involving [0, s1{a}] (⊂ [0, s1(∆)]) and [tl{a}({a}), 1] (⊂ [tl∆(∆), 1]) are modified. The same is true for φ{b} and ψ{b}. Therefore, we can define the partial maps ψ[tj(∆),sj (∆)] ∆ = φ[tj(∆),sj (∆)] ∆ for 0 < j < l∆. The only parts need to be modified are φ[0,s0(∆)] 3.14. Now denote φ[0,s0(∆)] tl(∆)(∆) by tl. Now we have two unital homomorphisms ∆ (∆ = [a, b]) by φ0 and φ[tl(∆),1] ∆ ∆ and φ[tl(∆),1] ∆ . by φl, and s0(∆) by s0, φ0 : Ik[0,s0] → P0M•C(∆)P0 15 and φl : Ik[tl,1] → PlM•C(∆)Pl, where P0, Pl are defined as in 3.10. We will do the modification of φ0 to get ψ0 (the one for φl is completely the same). We already have the definitions of ψ0{a} and ψ0{b}. Note that P0 ∈ M•(C(∆)) can be written as φ(h[0,s0]), where h[0,s0] is the test function appeared in 3.3, which is equal to 1 on [0, s0] and 0 on [s0 + η 12n , 1]. (Note that φ(h[0,s0]) is a projection since Spφ ⊂ [0, s0] ∪ [t1, 1] and t1 > s0 + η 12n .) Consequently, kP0(x) − P0(y)k < ε 12(n + 1)2 for all x, y ∈ [a, b] = ∆ (see (b) of 3.3). There exists a unitary W ∈ M•(C(∆)) such that P0(x) = W (x)   1   . . . 1  rank(P0)×rank(P0) 0 . . . 0 ε 6(n + 1)2 .   W ∗(x), for all x ∈ ∆ and kW (x) − W (y)k < To define it suffices to define since Note that where ψ0 : Ik[0,s0] → P0M•(C(∆))P0, AdW ◦ ψ0 : Ik[0,s0] → Mrank(P0)(C(∆)), W ∗P0W =(cid:18) 1rank(P0) 0 ♯(Speψ0{a} ∩ {0}) = rank(P0) 0 0 (cid:19) . (mod k), eψ0 : C[0, s0] ֒→ Ik[0,s0] ψ0−→ P0({a})M•(C)P0({a}). (This is true since the multiplicities of all the spectra other than 0 are multiples of k ). Similarly, ♯(Speψ0{b} ∩ {0}) = rank(P0) 16 (mod k). Also, from the definition of ψ on the vertices (namely on {a} and {b}) from 3.11, we know that ♯(Speψ0{b} ∩ {0}) = ♯(Speψ0{a} ∩ {0}) , k′ ≤ k. Lemma 3.15. Suppose that two unital homomorphisms α′, α′′ : Ik[0,s0] → Mrank(P0)(C) satisfy that counting multiplicity, where eα′ (or eα′′) is the composition α′ −→ Mrank(P0)(C)) (or C[0, s0] ֒→ Ik[0,s0] 0 < ♯(Speα′ ∩ {0}) = ♯(Speα′′ ∩ {0}) ≤ k C[0, s0] ֒→ Ik[0,s0] α′′ −→ Mrank(P0)(C)), then there is a homomorphism α : Ik[0,s0] → Mrank(P0)(C[a, b]), such that 0 < ♯(Speαt ∩ {0}) ≤ k, for all t ∈ [a, b] and α{a} = α′, α{b} = α′′, where again eα is the composition α−→ Mrank(P0)(C[a, b]). C[0, s0] ֒→ Ik[0,s0] Proof. We can regard [a, b] = [0, 1]. There are two unitaries u, v ∈ Mrank(P0)(C), a number k′ ∈ {1, 2, . . . , k}, and two finite sequences of numbers: such that and α′(f ) = u α′′(f ) = v 0 < ξ1 ≤ ξ2 ≤ · · · ≤ ξ• ≤ s0 0 < ξ′ ≤ ξ′ • ≤ s0 2 ≤ · · · ≤ ξ′  k′×k′ f (ξ1)  k′×k′ f (ξ′ 1)     f (0) . . . f (0) f (0) . . . f (0) 17 . . . f (ξ•) . . . f (ξ′ •)         u∗ v∗. Let u(t), 0 ≤ t ≤ 1 follows. For 0 ≤ t ≤ 1 2, 2 be any unitary path with u(0) = u, u( 1 2) = v. Define α as f (0) . . .   f (0)   k′×k′ f (ξ1)   f (ξ2) . . . f (ξ•)   u∗(t); α(f )(t) = u(t) and for 1 2 ≤ t ≤ 1, α(f )(t) = v   f (0) . . .   f (0)   k′×k′ f ((2 − 2t)ξ1 + (2t − 1)ξ′ 1) . . . Then α is a desired homomorphism. 3.16. Applying the above lemma, we can define   v∗. f ((2 − 2t)ξ• + (2t − 1)ξ′ •) such that and α : Ik[0,s0] → Mrank(P0)(C[a, b]) ı ◦ α{a} = AdW (a) ◦ ψ0{a} ı ◦ α{b} = AdW (b) ◦ ψ0{b}, where ı : Mrank(P0)(C) → M•(C) is defined by ı(A) =(cid:18) A 0 0 0 (cid:19) . Define ψ0 : Ik[0,s0] → P0M•(C(∆))P0 by ψ0 = AdW ∗ ◦ (ı ◦ α) -- that is, for any t ∈ [a, b] = ∆, ψ0(f )(t) = W (t)(cid:18) α(f )(t) 0 0 (cid:19) W ∗(t). 0 18 [a,b] and φl = φ[tl,1] As mentioned in 3.13, when we modify φ[a,b] to obtain ψ[a,b], we only need to [a,b] . The modifications of φl to ψl are the same as modify φ0 = φ[0,s0] the one from φ0 to ψ0. Thus we have the definition of ψ[a,b] = diag0≤i≤l(ψi). 3.17. Let us estimate the difference of φ[a,b] and ψ[a,b] on the finite set F ⊂ Ik. Note that φ[a,b] = diag0≤i≤l(φi), ψ[a,b] = diag0≤i≤l(ψi) and φi = ψi, for 0 < i < l. So we only need to estimate kφ0(f ) − ψ0(f )k and kφl(f ) − ψl(f )k. Note that φ0 and ψ0 are from Ik[0,s0] to P0M•(C[a, b])P0, where P0 is as in 3.14. And both AdW ◦ φ0 and AdW ◦ ψ0 can be regarded as ı ◦ φ′ and ı ◦ ψ′ for where is given by φ′, ψ′ : Ik[0,s0] → Mrank(P0)(C[a, b]), ı : Mrank(P0)(C[a, b]) → M•(C[a, b]) ı(A) =(cid:18) A 0 0 0 (cid:19) . Claim: Let α : Ik[0,s0] → Mrank(P0)(C[a, b]) be any unital homomorphism. Then we have α(f ) − (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  f (0) . . . f (0)  rank(P0) ≤ sup 0<ξ≤s0 kf (ξ) − f (0)k. (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  k′×k′ f (ξ1) In fact, for each x ∈ [a, b], there exist ux ∈ U(Mrank(P0)(C)) and k′ ∈ {1, 2, · · · , k} and 0 ≤ ξ1 ≤ ξ2 ≤ · · · ≤ ξ•• ≤ s0 such that α(f )(x) = ux f (0) . . . f (0)     It follows that   u∗ x. . . . f (ξ••) kα(f )(x) − f (0) · 1rank(P0)k 19 f (0)   . . . f (0)   −   f (ξ••) f (0)   . . . f (0)   k′×k′ f (ξ1) . . . 0 . . . 0  k′×k′       0 = = ux  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  f (ξ1) − f (0) . . . f (ξ••) − f (0) ≤ sup 0≤ξ≤s0 kf (ξ) − f (0) k .     u∗ x (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) f (0)   k′×k′ f (0) . . .  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  Thus, the claim is true. It follows from the claim that kφ(f )(t) − ψ(f )(t)k ≤ 2 max(cid:0) sup 0≤ξ≤s0 kf (ξ) − f (0)k, sup tl≤ξ≤1 kf (ξ) − f (1)k(cid:1) ≤ 2 · ε 6 for all t ∈ [a, b], and f ∈ F , as s0 − 0 < η definition of ψ on the 1-skeleton X (1) ⊂ X satisfying 2 and tl − 1 < η 2 . Hence we have the kφ(f )(t) − ψ(f )(t)k < ε 3 for all t ∈ X (1) and f ∈ F . 3.18. Now fix a 2-simplex ∆ ⊂ X. We will define ψ∆ : Ik → P M•(C(∆))P based on the previous definition of Again, write where ψ∂∆ : Ik → P M•C(∂∆)P. φ∆ = diag0≤i≤l(∆)(φi), φi = φ[ti(∆),si(∆)] ∆ = Ik[ti,si] → PiM•(C(∆))Pi 20 and Pi are projections defined on ∆ with l(∆)Xi=0 Pi(x) = P (x), ∀x ∈ ∆. For each face ∆′ ⊂ ∂∆, we know that the decomposition Spφ∆′ = l∆′[j=0 Tj(∆′)\ Spφ∆′ = l∆′[j=0 [tj(∆′), sj(∆′)]\ Spφ∆′ is finer than the decomposition Spφ∆′ = l∆[j=0 Tj(∆)\ Spφ∆′ = l∆[j=0 [tj(∆), sj(∆)]\ Spφ∆′. Consequently, [0, s0(∆′)] ⊂ [0, s0(∆)] and [tl(∆′), 1] ⊂ [tl(∆), 1]. Note that when we define ψ∆′ by modifying φ∆′, we only modify the parts of φ[0,s0(∆′)] ∆′ and φ [tl(∆′),1] ∆′ -- that is, [s0(∆′)+δ,tl(∆′)(∆′)−δ] ∆′ φ = ψ [s0(∆′)+δ,tl(∆′)(∆′)−δ] ∆′ , where δ ∈ (0, η 12n ). Hence φ [t1(∆),sl(∆)−1(∆)] ∆′ = ψ [t1(∆),sl(∆)−1(∆)] ∆′ since and t1(∆) > s0(∆) + η 12n η ≥ s0(∆′) + δ 12n Because ∆′ ⊂ ∂∆ is an arbitrary face, we have sl(∆)−1 < tl(∆)(∆) − < tl(∆′)(∆′) − δ. φ [t1(∆),sl(∆)−1(∆)] ∂∆ = ψ [t1(∆),sl(∆)−1(∆)] ∂∆ . Therefore similar to what we did on 1-simplexes, define ψ[tj(∆),sj (∆)] ∆ = φ[tj(∆),sj (∆)] ∆ for j ∈ {1, 2, · · · , l(∆) − 1}. Then we only need to modify φ[0,s0(∆)] φ = φl. We will only do it for φ0. ∆ [tl(∆),1] ∆ = φ0 and 21 3.19. We have the definition of unital homomorphism ψ0∂∆ : Ik[0,s0] → P0M•(C(∂∆))P0 such that ♯(Speψ0x ∩ {0}) = k′ ∈ {1, 2, · · · , k} for any x ∈ ∂∆, where ∆ is a 2-simplex and eψ0 is defined as the composition ψ0−→ P0M•(C(∂∆))P0. C[0, s0] ֒→ Ik(0,s0] We need to extend it to a homomorphism ψ0∆ : Ik[0,s0] → P0M•(C(∆))P0 such that ♯(Speψ0∆ ∩ {0}) = k′ for all x ∈ ∆. Once this extension is obtained, as in 3.17, we can use the claim in 3.17 to prove that φ[0,s0] are approximately equal to within ε 3 for all f ∈ F . (Note that in the argument of 3.17, the estimation is true which do not depend on the choice of the extension. It only uses s0 − 0 < η/2 < η, and kf (t) − f (t′)k < ε and ψ[0,s0] 6 whenever t − t′ < η.) ∆ ∆ There is a W ∈ U(M•(C(∆))) such that P0(x) = W (x)(cid:18) 1rank(P0) 0 0 (cid:19) W ∗(x) 0 for all x ∈ ∆. Again, if we can extend (AdW ◦ ψ0)∂∆ : Ik[0,s0] →(cid:18) 1rank(P0) 0 0 0 (cid:19) M•(C(∆))(cid:18) 1rank(P0) 0 0 (cid:19) , 0 (cid:19) M•(C(∆))(cid:18) 1rank(P0) 0 0 0 0 (cid:19) , to α∆ : Ik[0,s0] →(cid:18) 1rank(P0) 0 0 then we can set ψ0∆ = AdW ∗ ◦ α∆ to obtain our extension. But (AdW ◦ ψ0)∂∆ (or α∆) should be regarded as a homomorphism from Ik[0,s0] to Mrank(P0)(C(∂∆)) (or to Mrank(P0)(C(∆)). Hence the construction of ψ0∆ follows from the following lemma. Lemma 3.20. Let β : Ik[0,s0] → Mn′(C(S1)) be a unital homomorphism such that for any x ∈ S1, ♯(Sp(β ◦ ı)x ∩ {0}) = k′ ∈ {1, 2, · · · , k} for some fixed k′ (not depending on x), where ı : C[0, s0] → Ik[0,s0]. Then there is a homomorphism β : Ik[0,s0] → Mn′(C(D)), 22 (where D is the disk with boundary S1) such that ♯Sp(β ◦ ı)x\{0} = k′ for all x ∈ D and π ◦ β = β, where π : Mn′(C(D)) → Mn′(C(S1)) is the restriction. Proof. Let h(t) = t · 1k be the function in the center of Ik[0,s0]. Then β(h) is a self adjoint element in Mn′(C(S1)). For each z ∈ S1, write the eigenvalue of β(h)(z) in increasing order 0 = λ1(z) ≤ λ2(z) ≤ · · · ≤ λn′(z) ≤ s0. Then λ1, λ2, · · · , λn′ are continuous functions from S1 to [0, s0]. From the assump- tion, we know that λ1(z) = λ2(z) = · · · = λk′(z) = 0 and for all j > k′, λj(z) > 0. (Note that each λj(j > k′) repeats some multiple of k times.) Consequently, there is ξ ∈ (0, s0] such that λj(z) ≥ ξ for all j > k′. Hence β factors through as where and with Ik[0,s0] → Ik{0} ⊕ Ik[ξ,s0] diag(β0,β1) −−−−−−→ Mn′(C(S1)), β0 : Ik{0}(= C) → Q0Mn′(C(S1))Q0 β1 : Ik[ξ,s0](= Mk(C[ξ, s0])) → Q1Mn′(C(S1))Q1 Q0 + Q1 = 1n′ ∈ Mn′(C(S1)). Note that rank(Q0) = k′, and rank(Q1) = n′ − k′, which is a multiple of k. Write rank(Q1) = n′ − k′ = kk′′. There is a unitary u ∈ Mn(C(S1)) such that uQ0u∗ =(cid:18) 1k′ 0 0 (cid:19) 0 and uQ1u∗ =(cid:18) 0 0 1n′−k′ (cid:19) . 0 Hence with and Adu∗ ◦ β = diag(β′ 0, β′ 1) 0 : Ik{0}(= C) → Mk′(C(S1)) β′ 1 : Ik[ξ,s0](= Mk(C[ξ, s0])) → Mkk′′(C(S1)). β′ 23 Evidently, c β′ 0(c) =  . . .   = c · 1k′ ∈ Mk′(C(S1)), c ∀c ∈ C. 1, there exist β′′ : C[ξ, s0] → Mk′′(C(S1)) and a unitary V ∈ Mkk′′(C(S1)) such For β′ that V β′ 1(f )V ∗ = β′′ ⊗ idk(f ), ∀f ∈ Mk(C[ξ, s0]). Let Then W =(cid:18) 1k′   f (0) . . . (AdW ∗ ◦ β)(f ) = 0 V (cid:19) · u. 0   . f (0) β′′ ⊗ idk(f ) Let m be the winding number of the map S1 ∋ z 7−→ det(W (z)) ∈ T ⊆ C. Then W ∈ U(Mn′(C(S1))) is homotopic to W ′ ∈ Mn′(C(S1)) defined by W ′(z) = zm 1 1     . . . 1 , ∀z ∈ S1 = T. Let {wr} 1 2 ≤r≤1 be a unitary path in Mn′(C(S1)) with (z) = W ′(z) w 1 2 and w1(z) = W (z), ∀z ∈ S1. Evidently the homomorphism β′′ : C[ξ, s0] → Mk′′(C(S1)) is homotopic to the homomorphism β′′′ : C[ξ, s0] → Mk′′(C(S1)) 24 defined by β′′′(f )(e2πiθ) = f (ξ)1k′′ -- that is, β′′′(f )(eiθ) is the constant matrix f (ξ)1k′′ (which does not depend on θ). There is a path {βr}0≤r≤ 1 of homomorphisms 2 βr : C([ξ, s0]) → Mk′′(C(S1)) = β′′ and β0 = β′′′. such that β 1 2 Finally, regard D = {reiθ, 0 ≤ r ≤ 1}, and define β : Ik[0,s0] → Mn′(C(D)) by f (0) . . . f (0) f (0) . . . f (0)  k′×k′  k′×k′       w∗ r (eiθ)     This homomorphism is as desired.     β(f )(reiθ) = (β′′ ⊗ idk)(f )(eiθ) wr(eiθ), if 1 2 ≤ r ≤ 1 , if 0 ≤ r ≤ 1 2 . (βr ⊗ idk)(f )(eiθ) 3.21. Proof of Theorem 3.1 From 3.3 -- 3.20, we have constructed with the property ψ : Ik → P M•(C(X))P kφ(f ) − ψ(f )k < ε 3 for all f ∈ F . And importantly, for each x ∈ X, ♯(Speψx ∩ {0}) is a constant k′ ∈ {1, 2, · · · , k} and ♯(Speψx ∩ {1}) is also a constant k′ 1 ∈ {1, 2, · · · , k}, where eψ is the composition C[0, 1] ֒→ Ik ψ−→ P M•(C(X))P. Let h(t) = t · 1k ∈ Ik be the canonical function in the center of Ik. Then ψ(h) ∈ P M•(C(X))P is a self adjoint element. For each x ∈ X, denote the eigenvalues of ψ(h)(x) by 0 ≤ λ1(x) ≤ λ2(x) ≤ · · · ≤ λrank(P )(x) ≤ 1. Then all λi(x) are continuous functions from X to [0, 1]. Furthermore, λ1(x) = λ2(x) = · · · = λk′(x) = 0, 25 0 < λk′+1(x) ≤ λk′+2(x) ≤ · · · ≤ λrank(P )−k′ 1(x) < 1, λrank(P )−k′ 1+1(x) = λrank(P )−k′ 1+2(x) = · · · = λrank(P )(x) = 1. ξ1 = min x∈X λk′+1(x) > 0 and ξ2 = max λrank(P )−k′ 1(x) < 1. and Let Then That is, ψ factors through as Spψ ⊂ {0} ∪ [ξ1, ξ2] ∪ {1}. Ik → C ⊕ Mk(C[ξ1, ξ2]) ⊕ C diag(α0,ψ1,α1) −−−−−−−−→ P M•(C(X))P, where we identify Ik{0} = C and Ik{1} = C. Let Q0 = α0(1), Q1 = α1(1) and P1 = ψ1(1Mk(C([ξ1,ξ2]))). Finally, regarding ψ1 as Mk(C[0, 1]) restriction we finish the proof of Theorem 3.1. −−−−−−→ Mk(C([ξ1, ξ2])) ψ1−→ P1M•(C(X))P1, ✷ 3.22. From the definition of ψ in the above procedure, for every x ∈ X, the map ψx : Ik ψ −→ P M•(C(X))P evaluate at x −−−−−−−−→ P (x)M•(C)P (x) is defined when the construction of ψ∆ : Ik → P M•(C(∆))P ◦ is carried out for the unique simplex ∆ such that x ∈ ∆ (the interior of ∆). And when we define ψ∆ by modifying φ∆, the only modifications are made on the two parts φ[0,s0(∆)] and φ[tl(∆),1] . Consequently, ∆ ∆ Spφx ∩ (s0(∆), tl(∆)) = Spψx ∩ (s0(∆), tl(∆)) as sets with multiplicity. On the other hand for any simplex ∆, s0(∆) < η tl(∆)(∆) > 1 − η 2 . Hence 2 and Spφx ∩ [ η 2 , 1 − η 2 ] = Spψx ∩ [ η 2 , 1 − η 2 ]. If we further assume that φ has property sdp(η/4, δ), then ψ has property sdp(η, δ). As a consequence, we can use the decomposition theorem for ψ1 : Mk(C[0, 1]) → P1M•(C(X))P1 to study the homomorphisms φ, ψ : Ik → P M•(C(X))P. Note that the homomor- phisms f 7→ f (0)Q0 and f 7→ f (1)Q1 factor through the C ∗-algebra C. 26 3.23. Lemma 3.20 is not true for the case k′ = 0. In fact, there exists a unital homo- morphism α : Mk(C) → Mk(C(S1)), which can not be extended to a homomorphism α : Mk(C) → Mk(C(D)). Let πs0 : Ik[0,s0] → Mk(C) be the map defined by evaluat- ing at the point s0. Then β = α ◦ πs0 : Ik[0,s0] → Mk(C(S1)) can not be extended to β : Ik[0,s0] → Mk(C(D)) such that ♯Sp(β ◦ ı)xT{0} = k′ = 0 for all x ∈ D, where ı is the canonical map from Mk(C) to Ik[ζ,s0] for some 0 < ζ < s0. 4 Decomposition Theorem II Our next task is to study the possible decomposition of φ : C(X) → Ml(Ik2) for X being [0, 1], S1 or TII,k. The cases of [0, 1], S1 are more or less known (see [4] and [11]). Let us assume X is a 2-dimensional connected simplicial complex. The following lemma is essentially due to H. Su (See [40]). The case of X = graph was stated in [26]. Lemma 4.1. For any connected simplicial complex X, a finite set F ⊂ C(X) which generates C(X), η > 0 and a positive interger n > 0, there is a δ > 0, such that for any two unital homomorphisms φ, ψ : C(X) → Mn(C), if kφ(f ) − ψ(f )k < δ for all f ∈ F, then Sp(φ) and Sp(ψ) can be paired within η. This is a consequence of Lemma 2.2 and Lemma 2.3 of [40]; also see the argument 2.1.3 in [26]. For the case of graphs, it was stated in 2.1.9 of [26]. Lemma 4.2. For any connected simplicial complex X, a finite generating set F ⊂ C(X), ε > 0 and positive integer n > 0, there is δ > 0 with the following property: If x1, x2, · · · , xn ∈ X are n points (possibly repeating), u, v ∈ Mn(C) are two unitaries such that u f (x1)   f (x2) . . . f (xn) f (x1)   u∗ − v   f (x2) . . . f (xn)   v∗(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)   u∗ < δ, t′ < ε (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) for all f ∈ F , then there is a path of unitaries ut ∈ Mn(C) connecting u and v (i.e, u0 = u, u1 = v) with the property that ut f (x1)   f (x2) . . . f (xn)   u∗ t − ut′   f (x1) f (x2) . . . f (xn) for all f ∈ F and t, t′ ∈ [0, 1] (of course δ depends on both ε and n). 27 This was proved in step 2 and step 3 of the proof of Theorem 3.1 of [40]. The following lemma reduces the study of φ : C(X) → Ml(Ik) to the study of homomorphism φ1 : C(Γ) → Ml(Ik), where Γ ⊂ X is 1-skeleton of X under a certain simplicial decomposition. Since Γ is a graph, then we will apply the technique in [26] and [27] to obtain the decomposition of φ1. Lemma 4.3. Let X be a 2-dimensional simplicial complex. For any F ⊂ C(X), ε > 0, η > 0, and any unital homomorphism φ : C(X) → Ml(Ik), there is a simpicial decomposition of X with 1-skeleton X (1) = Γ and a homomorphism φ1 : C(Γ) → Ml(Ik) such that: 1. kφ(f ) − φ1 ◦ π(f )k < ε, where π : C(X) → C(Γ) is given by π(f ) = f Γ; 2. For any t ∈ [0, 1], Spφt and Sp(φ1 ◦ π)t can be paired within η. Proof. By Lemma 4.1, we only need to prove that there exists a homomorphism φ1 to satisfy condition (1). Without loss of generality, we assume that F generates C(X). By 4.2, there is an ε′ > 0 such that for any x1, x2, · · · , xkl ∈ X and unitaries u, v ∈ Mkl(C), if f (x1) f (x2) . . . f (xkl) f (x1) f (x2) . . . f (xkl) then there is a continuous path ut with u0 = u, u1 = v satisfying that f (x1) f (x2) . . . f (xkl) f (x1) f (x2) . . . f (xkl) Recall for the simplicial complex, a continuous path {x(t)}0≤t≤1 is called piecewise linear if there are a sequence of points 0 = t0 < t1 < · · · < tn = 1 such that {x(t)}ti≤t≤ti+1 fall in the same simplex of X and are linear there. Note that the property of piecewise linear is preserved under any subdivision of the simplicial complex. For the simplicial complex X, we endow the standard metric on X, briefly described as below (see [18, 1.4.1] for detail). Identify each n-simplex with an n- simplex in Rn whose edges are of length 1, preserving affine structure of the simplexes. Such identifications give rise to a unique metric on the simplex ∆. For any two points x, y ∈ X, d(x, y) is defined to be the length of the shortest path connecting 28 u ut     (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)   u∗ − v   u∗ t − ut′       v∗(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)   u∗ t′ < ε′, < ε 3 . x and y. (The length is measured in individual simplex, by breaking the path into small pieces). With this metric, if x0, x1 ∈ X with d(x0, x1) = d, then there is a piecewise linear path x(t) with length d such that x(0) = x0, x(1) = x1. Furthermore, d(x(t), x(t)′) ≤ d for all t, t′ ∈ [0, 1]. In fact, we can choose x(t), such that d(x(t), x(t′)) = t′ − t · d. There is an η′ < η 2η′, 4 such that the following is true: For any x, x′ ∈ X with d(x, x′) < f (x) − f (x)′ < ε′ 3 . Let δ > 0, such that if t − t′ ≤ δ, then kφ(f )(t) − φ(f )(t′)k < ε′ 3 , ∀f ∈ F, and Spφt and Spφt′ can be paired within η′. Dividing the interval [0, 1] into pieces 0 = t0 < t1 < t2 < · · · < t• = 1, with ti+1 − ti < δ. We first define ψ : C(X) → Ml(Ik) such that ψ is close to φ on F to within ε 3, Spφt and Spψt can be paired within η′, and with extra property that on each interval [ti, ti+1]; Spψt = {α1(t), α2(t), · · · , αlk(t)} with all αj : [ti, ti+1] → X being piecewise linear. Set ψ{ti} = φ{ti}, for each ti (i = 0, 1, 2, · · · , •) -- that is, ψ(f )(ti) = φ(f )(ti) for i = 0, 1, 2, · · · , •. And we will define ψ{t} for t ∈ (ti, ti+1) by interpolating the definitions between ψ{ti} and ψ{ti+1}. (Note that we do not change the definitions of φ{0} and φ{1}, hence ψ is a homomorphism into Ml(Ik) instead of Mlk(C[0, 1]).) Let Spψ{ti} = {α1, α2, · · · , αlk} ⊂ X Spψ{ti+1} = {β1, β2, · · · , βlk} ⊂ X. Since Spψ{ti} and Spψ{ti+1} can be paired within η′, we can assume dist(αi, βi) < η′. There exist two unitaries u, v ∈ Mlk(C) such that   u∗ and ψ(f )(ti+1) = v    v∗ − v  3 for each j, we have f (β1) . . . f (βkl) f (αkl) 29   v∗. f (β1) . . . f (βkl) < ε′ 3 .  v∗(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  ψ(f )(ti) = u  f (α1) . . . f (αkl) Note that kf (αj) − f (βj)k < ε′ f (α1) . . . (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) v  Combining with kψ(f )(ti) − ψ(f )(ti+1)k < ε′ 3 , we get f (α1) . . .   u∗ − v  f (αkl) f (α1) . . . f (αkl) Since ε′ is the number δ in Lemma 4.2 for ε path u(t), ti ≤ t ≤ ti+ti+1 with u(ti) = u, 2 3 , applying Lemma 4.2, there is a unitary u( ti+ti+1 2 ) = v such that . 2ε′ 3 <  v∗(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)   u∗(t′)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)  < ε 3 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) u  u(t)  (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) f (α1) . . .   u∗(t) − u(t′)  f (α1) . . . f (αkl) ]. f (αkl) for all t, t′ ∈ [ti, ti+ti+1 2 that There are piecewise linear paths ri(t) with ri( ti+ti+1 ) = αi and ri(ti+1) = βi such 2 Define ψ(f ) as follows: For t ∈ [ti, ti+ti+1 2 ], d(ri(t), ri(t′)) ≤ dist(αi, βi) < η′. ψ(f )(t) = u(t)  f (α1) . . .   u∗(t); f (αkl) for t ∈ [ ti+ti+1 2 , ti+1], ψ(f )(t) = v f (r1(t)) f (r2(t)) . . .     v∗. f (rkl(t)) 2 Then {Spψt, t ∈ [ti, ti+1]} is a collection of kl piecewise linear maps from [ti, ti+1] to X. (Note that for t ∈ [ti, ti+ti+1 ], we use constant maps which are linear.) Now subdivid the simplicial complex X so that each simplex of the subdivision has diameter at most η′, and so that all the points in Spφ{0} = Spψ{0} and Spφ{1} = Spψ{1} are vertices. With this simplicial decomposition we have, Spψ ∩ ∆ $ ∆, for every 2-simplex ∆. This is true because Spψ[ti,ti+1] is the union of the collection of images of kl piecewise linear maps from [ti, ti+1] to X, and a finite union of line segments must be 1-dimensional. Hence for each simplex ∆ of dimension 2, we can choose a point x∆ ∈ ◦ ∆, such that x∆ 6∈ Spψ. 30 There is a σ > 0 such that Spψ has no intersection with Bσ(x∆) = {x ∈ X, dist(x, x∆) ≤ σ} for all ∆. Let Y = X\(cid:0) ∪ {Bσ(x∆) ∆ is 2-simplex}(cid:1). Then Spψ ⊂ Y . That is, ψ factors through C(Y ) as ψ : C(X) restriction −−−−−−→ C(Y ) ψ1−→ Ml(Ik). Let α : Y → X (1) be the standard retraction defined as a map sending ∆\{x∆} to ∂∆ for each simplex ∆. Then d(x, α(x)) < η′. Let φ1 : C(X (1)) → Ml(Ik) be defined by ψ1 ◦ α∗ : C(X (1)) α∗ −→ C(Y ) ψ1−→ Ml(Ik). Evidently φ1 is as desired. Corollary 4.4. Suppose that φ : C(X) → Ml(Ik) is a unital homomorphism. For any finite set F ⊂ C(X), ε > 0, and η > 0, there is a unital homomorphism ψ : C(X) → Ml(Ik) such that (1) φ(f )(0) = ψ(f )(0), φ(f )(1) = ψ(f )(1) for all f ∈ C(X); (2) kφ(f ) − ψ(f )k < ε for all f ∈ F ; (3) Spφt and Spψt can be paired to within η; (4) For each t ∈ (0, 1), the maximal multiplicity of Spψt is one -- that is, ψ{t} has distinct spectra. Proof. Applying Lemma 4.3, we reduce the case of C(X) to the case of C(X (1)), where X (1) is a 1-dimensional simplicial complex. The corollary of this case is almost the same as the special case of [26, Theorem 2.1.6] (where we let Y = [0, 1]). Note that from the proof of Theorem 2.1.6 in [26], if we do not require the homomorphism ψ to have distinct spectrum at the end points 0 and 1, then we do not need to modify the original homomorphism φ at these two end points. The proof goes the same way as the proof there with some small modifications. We briefly describe them as below. One divides the interval Y = [0, 1] into small pieces [0, 1] = ∪m−1 i=0 [yi, yi+1] with y0 = 0 < y1 < y2 · · · < ym = 1, as in the proof of [26, Theorem 2.1.6]. Define ψyi with 1 ≤ i ≤ m − 1, by slightly modifying φyi so that ψyi has distinct spectra; but define ψ0 = φ0 and ψ1 = φ1 (no modification are made at the ending points). Therefore, in our case, ψ0 and ψ1 do not have distinct spectra -- this is the only difference from [26, Theorem 2.1.6]. For all intervals [yi, yi+1] with 1 ≤ i ≤ m−2, the constructions of ψ[yi,yi+1] are the same as in the proof of [26, Theorem 2.1.6]. For the constructions of ψ[0,y1] and ψ[ym−1,1], we need to modify [26, Lemma 2.1.1] and [27, Lemma 2.1.2] accordingly, in an obvious way, and then apply these modifications. For example, [26, Lemma 2.1.1] should be modified to the following case: among two 31 l-element sets X 0 = {x0 is distinct. That is, the following statement is true with the same proof: l } and X 1 = {x1 2, · · · , x1 2, · · · , x0 1, x0 1, x1 l } -- only one of them Let X = X1 ∨ X2 ∨ · · · ∨ Xk be a bunch of k intervals Xi = [0, 1] (1 ≤ i ≤ k) and Y = [0, 1]. Suppose that X 0 = {x0 1, x0 2, · · · , x0 l } ⊂ X and X 1 = {x1 1, x1 2, · · · , x1 l } ⊂ X j if i 6= j. Then there are l continuous functions f1, f2, · · · , fl : Y → X i 6= x1 with x1 such that (1) as sets with multiplicity, we have {f1(0), f2(0), · · · , fl(0)} = X 0, and {f1(1), f2(1), · · · , fl(1)} = X 1, (2) for each t ∈ (0, 1] ⊂ Y and i 6= j, we have fi(t) 6= fj(t). Remark 4.5. In Corollary 4.4, we can further assume that Spψ{0} and Spψ{1} have eigenvalue multiplicity just k as homomorphisms from C(X) to Mlk(C[0, 1]), or equiv- alently, both maps C(X) ψ−→ Ml(Ik) evaluate at 0 −−−−−−−→ Ml(C) and C(X) ψ−→ Ml(Ik) evaluate at 1 −−−−−−−→ Ml(C) have distinct spectrum. To do this, we first extend the definition of the original φ to a slightly larger interval [−δ, 1 + δ] as below. Find u ∈ Ml(C) and x1, x2, · · · , xl ∈ X such that φ(f )(0) = u f (x1)   f (x2) . . .   u∗ ⊗ 1k. f (xl) Since X is path connected and X 6= {pt}, there are functions αi : [−δ, 0] → X such that {αi(−δ)}l i=1 is a set of distinct l points, αi(0) = xi, and dist(αi(t), αi(0)) are as small as we want. Define φ(f )(t) = u  f (α1(t)) . . . f (αl(t))   u∗ ⊗ 1k, for t ∈ [−δ, 0]. Similarly, we can define φ(f )(t) for t ∈ [1, 1 + δ], so that φ1+δ as a homomorphism from C(X) to Mkl(C) has multiplicity exactly k and φ(f )(1 + δ) ∈ Ml(C) ⊗ 1k. One 32 can reparemetrize [−δ, 1 + δ] to [0, 1] so that φ0 and φ1 as homomorphisms from C(X) to Mkl(C) have multiplicity exactly k. Then we apply the corollary to perturb φ to ψ without changing the definition at the end points. Remark 4.6. The same argument can be used to prove the following result. Let X 6= {pt} be a connected finite simplicial complex of any dimension. Let Y be a 1-dimensional simplicial complex. Then any homomorphism φ : C(X) → Mn(C(Y )) can be approximated arbitrarily well by a homomorphism ψ with distinct spectrum. This is a strengthened form of [18, Theorem 2.1] for the case dim(Y ) = 1. The following Theorem for X = gragh, is a slight modification of [28, Theorem 2.7]. Theorem 4.7. Let X be a connected simplicial complex of dimension at most 2, and G ⊂ C(X) be a finite set which generates C(X). For any ε > 0, there is an η > 0 such that the following statement is true. Suppose that φ : C(X) → Ml1l2+r(Ik) is a unital homomorphism satisfying the following condition: There are l1 continuous maps a1, a2, · · · , al1 : [0, 1](= Sp(Ik)) → X such that for every y ∈ [0, 1], Spφy (considered as a homomorphism from C(X) to M(l1l2+r)k(C[0, 1])) and Θ(y) can be paired within η, where Θ(y) = {a1(y)∼l2k, a2(y)∼l2k, · · · , al1−1(y)∼l2k, al1(y)∼(l2+r)k}. It follows that there are l1 mutually orthogonal projections p1, p2, · · · , pl1 ∈ Ml1l2+r(Ik) such that (i) for all g ∈ G and y ∈ Y kφ(g)(y) − p0φ(g)(y)p0 ⊕ l1Xk=1 g(ak(y))pkk < ε, where p0 = 1 −Pl1 i=1 pi; (ii) rank(pi) = (l2 − 3)k for 1 ≤ i < l1, rank(pl1) = (l2 + r − 3)k (as projections in M(l1l2+r)k(C[0, 1])) and rank(p0) = 3l1k. Proof. We will apply [28, Theorem 2.7] (using map ai to replace map b ◦ ai as in [28, Remark 2.8]) and its proof (see 2.9-2.16 of [28]) for the case Y in [28, Theorem 2.7] being [0,1]. As a matter of fact, in the proof of [28, Theorem 2.7], Li does use that X to be graph, for only one property that any homomorphism from C(X) to MnC(Y ) (Y graph) can be approximated arbitrarily well by homomorphisms with distinct spectra. By Remark 4.6, [28, Theorem 2.7] holds for the case X 6= {pt} being any connected simplicial complex and Y , a graph. 33 For finite set G ⊂ C(X), and ε > 0, choose η > 0 such that dist(x1, x2) ≤ η implies g(x1) − g(x2) < ε 4 for all g ∈ G, as in [28, 2.16]. Without lose of generality, we can assume that the Spφt is distinct for any t ∈ (0, 1) and Spφ0 and Spφ1 have multiplicities exact k as in Corollary 4.4 and Remark 4.5 above. When we go through Li's proof in [28], we need to make the projections pi to satisfy the extra condition: pi(0), pi(1) ∈ (Ml1l2+r(C)) ⊗ 1k ⊆ M(l1l2+r)k(C). We will repeat part of the proof of [28, Theorem 2.7] and point out how to modify it. As in the proof of [28, Theorem 2.7], we can choose an open cover U0, U1, · · · , U• of [0, 1] with U0 = [0, b0), U1 = (a1, b1), U2 = (a2, b2), · · · , U•−1 = (a•−1, b•−1), U• = (a•, 1], 0 < a1 < b0 < a2 < b1 < a3 < b2 < · · · < a• < b•−1 < 1. We will define P i U (i = 1, 2, · · · , l1) as same as in [28, 2.12] for U = Ui(0 < i < •) -- note that Spφy, for y ∈ (a1, b•−1) ⊂ (0, 1), are distinct. For U0 and U•, a special care is needed as follows. We will only do it for U0 (it is the same for U•). Write Spφ0 = {λ∼k q } with q = l1l2 + r. Then {λ1, λ2, · · · , λq} can be paired with {a0(0)∼l2, a2(0)∼l2, · · · , al1−1(0)∼l2, al1(0)∼(l2+r)} (note ∼ l2k is changed to ∼ l2 here) to within η. We can divide {λ1, λ2, · · · , λq} into groups {λ1, λ2, · · · , λq} = j=1 E′j (where E′j = l2 if 1 ≤ j ≤ l1 − 1, and E′j = l2 + r if j = l1) such that 2 , · · · , λ∼k 1 , λ∼k Sl1 dist(λi, aj(0)) < η, for all λi ∈ Ej. Let σ′ satisfy the following conditions: (1) σ′ < min{dist(λi, λj), (2) σ′ < η − max{dist(λi, aj(0)), λi ∈ Ej}. i 6= j}; We can choose b1 (> b0 > a1 > 0) being so small that for any y ∈ [0, b1], Spφy and 2 . Then for each y ∈ [0, b1], 2 and dist(aj(y), aj(0)) < σ′ Spφ0 can be paired to within σ′ Spφy can be written as a set of {λ1 1(y), λ2 1(y), · · · , λk 1(y), λ1 2(y), λ2 2(y), · · · , λk 2(y), · · · , λ1 q(y), · · · , λk q (y)} i (0) = λi. Then let Ej(y) be the set {λi′ i (y); λi ∈ E′j}. In this way we have, with λj if λi′ i ∈ Ej, then dist(λi′ i (y), aj(y)) < η. U0(y) and P j Let both P j U1(y) (defined on U0 = [0, b0) and U1 = (a1, b1)) be the spectral projections corresponding to Ej(y). In particular, P j U0(0) ∈ Ml1l2+r(C) ⊗ 1k. We can define pj(y) as a subprojection of P j U (y) (for U ∋ y) as in 2.9-2.16 of [28] for each y ∈ [b0, a•] but with rank (pj(y)) = (l2 − 3)k (instead of l2 − 3 in [28]) for 34 1 ≤ j ≤ l1 − 1 and rank(pl1(y)) = (l2 + r − 3)k (instead of l2 + r − 3 in [28]). Also we can choose an arbitrary sub projection pj(0) < P j U0(0) ∈ Ml1l2+r(C) ⊗ 1k of form pj(0) = p′ j(0)) = l2 − 3 for 1 ≤ j ≤ l1 − 1, and rank(p′ j(0) ⊗ 1k ∈ Ml1l2+r(C) ⊗ 1k with rank(p′ l1(0)) = l2 + r − 3. Consequently, rank(pj(0)) = (l2 − 3)k and rank(pl1(0)) = (l2 + r − 3)k. Finally, connect pj(0) and pj(b0) by pj(y) for y ∈ [0, b0] inside P j U0(y). As one can see from 2.16 of [28], if the projections pj(y) are subprojections of P j U (y), then all the estimations in that proof hold. After we do similar modifications for P j U•(y) and pj(y) near point 1, we will get pj(y) ∈ Ml1l2+r(Ik) instead of M(l1l2+r)k(C[0, 1]). (This method was also used in the proof of [13, Theorem 3.10].) The following result is a generalization of [18, Proposition 4.42]. Theorem 4.8. Let X be a connected finite simplicial complex of dimension at most 2, ε > 0 and F ⊂ C(X), a finite set of generators. Suppose that η ∈ (0, ε) satisfies that if dist(x, x′) ≤ 2η, then kf (x) − f (x′)k < ε 4 for all f ∈ F . For any δ > 0 and positive integer J > 0, there exist an integer L > 0 and a finite set H ⊆ Af f T C(X)(= CR(X)) such that the following holds. If φ, ψ : C(X) → B = MK(Ik) (or B = P M•(C(Y ))P ) are unital homomorphisms 4 , for all h ∈ H, with the properties: (a) φ has sdp(η/32, δ); (b) K ≥ L (or rank(P ) ≥ L); (c) kAf f T φ(h) − Af f T ψ(h)k < δ then there are three orthogonal projections Q0, Q1, Q2 ∈ B, two homomorphisms φ1 ∈ Hom(C(X), Q1BQ1)1 and φ2 ∈ Hom(C(X), Q2BQ2)1, and a unitary u ∈ B such that (1) 1B = Q0 + Q1 + Q2; (2) kφ(f ) −(cid:0)Q0φ(f )Q0 + φ1(f ) + φ2(f )(cid:1)k < ε k(Adu ◦ ψ)(f ) −(cid:0)Q0(Adu ◦ ψ)(f )Q0 + φ1(f ) + φ2(f )(cid:1)k < ε, for all f ∈ F ; (3) φ2 factors through C[0, 1]; (4) Q1 = p1 + · · ·+ pn with (rank(Q0) + 2)J < rank(pi) (i = 1, 2, · · · , n), where rank: K0(B) → Z is the map induced on K0 by the evaluation map at 0 or 1. (which is rank pi(0) for B = MK(Ik), where rank pi(0) is regarded as projections in MK(C) not MK(Mk(C))), and φ1 is defined by and φ1(f ) = nXi=1 f (xi)pi, ∀f ∈ C(X), 35 where p1, p2, · · · , pn are mutually orthogonal projections and {x1, x2, · · · , xn} ⊂ X is an ε-dense subset of X. Proof. For the case B = P M•(C(Y ))P , this is [18, Proposition 4.42]. The proof for the case B = MK(Ik) is almost the same as the proof of [18, Proposition 4.42], replacing [18, Theorem 4.1] by Theorem 4.7 above. The only thing one should notice is that, in [18, Lemma 4.33], rankφ(1) = K, the K should be corresponding to K in our theorem (not Kk) and Θ(y) should be defined as Θ(y) =(cid:8)α ◦ β1(y)∼L2k, α ◦ β2(y)∼L2k, · · · , α ◦ βL−1(y)∼L2k, α ◦ βL(y)∼(L2+L1)k(cid:9) . (Note in the above, we use ∼ L2k and ∼ (L2 + L1)k to replace ∼ L2 and ∼ (L2 + L1) in [18].) In the proof of this version of [18, Lemma 4.33], one can choose the homomorphism ψ′ : C(X) → Mk(C[0, 1]) (not to MKk(C[0, 1])) as the map ψ there, with kAf f T φ(f ) − Af f T ψ′(f )k < ∀f ∈ H(η, δ, x) δ 4 as in [18, Lemma 4.33]. Then let ψ = ψ′ ⊗ ık, where ık : C → Mk(C) is defined by ık(λ) = λ · 1k. With this modification, we have Spψ′ y being Θ′(y) = {α ◦ β1(y)∼L2, α ◦ β2(y)∼L2, · · · , α ◦ βL−1(y)∼L2, α ◦ βL(y)∼(L2+L1)} and Spψy being Θ(y) = {α ◦ β1(y)∼L2k, α ◦ β2(y)∼L2k, · · · , α ◦ βL−1(y)∼L2k, α ◦ βL(y)∼(L2+L1)k} as desired. All other parts of the proof are exactly the same. For the proof of uinqueness theorem in [19], it is important to have a simultaneous decomposition for two homomorphisms as below. Theorem 4.9. Let X be a connected finite simplicial complex of dimension at most 2, ε > 0 and F ⊂ C(X), a finite set of generators. Suppose that η ∈ (0, ε) satisfies that if dist(x, x′) ≤ 2η, then kf (x) − f (x′)k < ε 4 for all f ∈ F . Let κ be a fixed simplicial structure of X. For any δ > 0 and positive integer J > 0, there exist an integer L > 0 and a finite set H ⊆ Af f T C(X)(= CR(X)) such that the following holds. If X1 is a connected sub-complex of (X, κ), and if φ, ψ : C(X1) → B = MK(Ik) (or B = P M•(C(Y ))P ) are unital homomorphisms with the following properties: (a) φ has sdp(η/32, δ); (b) K ≥ L (or rank(P ) ≥ L); (c) kAf f T φ(hX1) − Af f T ψ(hX1)k < δ 4 , for all h ∈ H, 36 then there are three orthogonal projections Q0, Q1, Q2 ∈ B, two homomorphisms φ1 ∈ Hom(C(X1), Q1BQ1)1 and φ2 ∈ Hom(C(X1), Q2BQ2)1, and a unitary u ∈ B such that (1) 1B = Q0 + Q1 + Q2; (2) kφ(f X1) −(cid:0)Q0φ(f X1)Q0 + φ1(f X1) + φ2(f X1)(cid:1)k < ε k(Adu ◦ ψ)(f X1) −(cid:0)Q0(Adu ◦ ψ)(f X1)Q0 + φ1(f X1) + φ2(f X1)(cid:1)k < ε for all f ∈ F ; (3) φ2 factors through C[0, 1]; (4) Q1 = p1 + · · · + pn with (rank(Q0) + 2)J < rank(pi) (i = 1, 2, · · · , n), and φ1 is defined by and φ1(f ) = f (xi)pi ∀f ∈ C(X), nXi=1 where p1, p2, · · · , pn are mutually orthogonal projections and {x1, x2, · · · , xn} ⊂ X1 is an ε-dense subset of X1. Proof. Suppose that {Xi}i are all connected sub-complexes of (X, κ) (there are finitely many of them for a fixed simplicial structure of a finite complex). Apply Theorem 4.7 to each Xi to obtain Li and Hi ⊆ Af f T (C(Xi)) as in the theorem. By Tietze Extension Theorem, there are finite sets Hi ⊆ Af f T (C(X)) such that Hi ⊆ {hXi h ∈ Hi}. Evidently L = maxi{Li} and H = ∪i Hi are as desired. Acknowledgement The authors would like to express our special thanks of gratitude to Professor Guihua Gong who suggested us to do this interesting problem. We also benefit a lot from discussions with him. References [1] M. Dadarlat, Reduction to dimension three of local spectra of Real rank zero C ∗-algebras, J. Reine Angew. Math. 460(1995) 189-212. [2] M. Dadarlat and G. Gong, A classification result for approximately homogeneous C ∗-algebras of real rank zero, Geometric and Functional Analysis, 7(1997) 646- 711. [3] S. Eilers, A complete invariant for AD algebras with bounded torsion in K1, J. Funct. Anal. 139(1996), 325-348. [4] G. A. Elliott, On the classification of C ∗-algebras of real rank zero, J. Reine Angew. Math. 443(1993) 263-290. 37 [5] G. A. Elliott, A classification of certain simple C ∗-algebras, Quantum and Non- Commutative Analysis, Kluwer, Dordrecht, (1993), 373-388. [6] G. A. Elliott, A classification of certain simple C ∗-algebras, II, J. Ramaunjan Math. Soc., 12 (1997), 97-134. [7] G. A. Elliott and G. Gong, On the inductive limits of matrix algebras over two-tori, American. J. Math 118(1996) 263-290. [8] G. A. Elliott and G. Gong, On the classification of C ∗-algebras of real rank zero, II, Ann. of Math, 144(1996) 497-610. [9] G. A. Elliott, G. Gong and L. Li, On the classification of simple inductive limit C ∗-algebras, II: The isomorphism Theorem, Invent. Math. 168(2)(2007) 249-320. [10] G. A. Elliott, G. Gong and L. Li, Injectivity of the connecting maps in AH inductive limit systems, Canand. Math. Bull., 26(2004) 4-10. [11] G. A. Elliott, G. Gong, X. Jiang, H. Su: Aclassification of simple limits of dimension drop C ∗−algebras, Fields Inst. Commun., 13, 125-143 (1997). [12] G. A. Elliott, G. Gong, H. Lin, C.Pasnicu: Abelian C ∗−subslgebras of C ∗−algebras of real rank zero and inductive limit C ∗−algebras, Duke Math. J., 83, 511-554 (1996). [13] G. A. Elliott, G. Gong, H. Su: On the classification of C ∗−algebras of real rank zero, IV: Reduction to local spectrum of dimension two, Fields Inst. Commun., 20, 73-95 (1998). [14] G. Gong, Approximation by dimension drop C ∗-algebras and classification, C. R. Math. Rep. Acad. Sci Can., 16(1994) 40-44. [15] G. Gong, Classification of C ∗-algebras of real rank zero and unsuspended E- equivalent types, J. Funct. Anal. 152(1998) 281-329. [16] G. Gong, On inductive limit of matrix algebras over higher dimension spaces, Part I, Math Scand., 80(1997) 45-60. [17] G. Gong, On inductive limit of matrix algebras over higher dimension spaces, Part II, Math Scand., 80(1997) 61-100. [18] G. Gong, On the classification of simple inductive limit C ∗-algebras, I: Reduction Theorems, Doc. Math., 7(2002) 255-461. [19] G. Gong, C. Jiang, L. Li, A classification of inductive limit C ∗-algebras with ideal property, preprint. 38 [20] G. Gong, C. Jiang, L. Li, C. Pasnicu, AT structure of AH algebras with ideal property and torsion free K-theory, J. Func. Anal. 58(2010) 2119-2143. [21] G. Gong, C. Jiang, L. Li, C. Pasnicu, A Reduction theorem for AH algebras with the ideal property, Int. Math. Res. Not. IMRN, 2018, no. 24, 7606-7641. [22] K. Ji and C. Jiang, A complete classification of AI algebra with ideal property, Canadian. J. Math, 63(2), (2011), 381-412. [23] C. Jiang, A classification of non simple C ∗-algebras of tracial rank one: In- ductive limit of finite direct sums of simple TAI C ∗-algebras, J. Topol. Anal. 3 No.3(2011), 385-404. [24] C. Jiang, Reduction to dimension two of local spectrum for AH algebras with ideal property, Canad. Math. Bull., 60 (2017), no. 4, 791-806. [25] C. Jiang, K. Wang, A complete classification of limits of splitting interval al- gebras with the ideal property, J. Ramanujan Math. Soc., 27, No. 3 (2012) 305-354. [26] L. Li, On the classification of simple C ∗-algebras: Inductive limit of matrix algebras over trees, Mem Amer. Math, Soc., 127(605) 1997. [27] L. Li, Simple inductive limit C ∗-algebras: Spectra and approximation by interval algebras, J. Reine Angew Math ,507(1999) 57-79. [28] L. Li, Classification of simple C ∗-algebras: Inductive limit of matrix algebras over 1-dimensional spaces, J. Func. Anal., 192(2002) 1-51. [29] L. Li, Reduction to dimension two of local spectrum for simple AH algebras, J. of Ramanujian Math. Soc., 21 No.4(2006) 365-390. [30] H. Lin, Tracially AF C ∗−algebras, Trans. Amer. Math. Soc., 353 (2001) No. 2, 693-722. [31] H. Lin, Simple nuclear C ∗−algrbras of tracial topological rank one, J. Funct. Anal., 251(2007), No. 2, 601-679. [32] H. Lin, Crossed products and minimal dynamical systems, J. Topol. Anal., 10 (2018), no. 2, 447-469. [33] K. E. Nielsen and K. Thomsen, Limits of circle algebras. Expo. Math., 14 (1996), 17-56. [34] C. Pasnicn, Shape equiralence, nonstable K-theory and AH algebras, Pacific J. Math, 192(2000) 159-182. 39 [35] C. Pasnicu, The ideal property in crossed products, Proc. Amer. Math. Soc., 131 (7)(2003) 2103-2108. [36] C. Pasnicu, Extension of AH algebras with the ideal property, Proc. Edinb. Math. Soc., (2) 42 (1)(1999) 65-76. [37] C. Pasnicu, On the AH algebras with the ideal property, J. Operator Theory, 43 (2)(2000) 389-407. [38] C. Pasnicu, Ideals generated by projections and inductive limit C ∗−algebras, Rocky Mountain J. Math., 31 (3)(2001) 1083-1095. [39] M. Rørdam, Classification of certain infinite simple C ∗−algebras, J. Funct. Anal., 131(1995), 415-458. [40] H. Su, On the classification of C ∗-algebras of real rank zero: inductive limits of matrix algebras over non-Hausdorff graphs, Memoirs of the American Mathe- matical Society, Vol 114, No. 547, 1995. [41] K. Thomsen, Insuctive limit of interval algebras, American J. of Math, 116, 605 -- 620 (1994). [42] K. Thomsen, Limits of certain subhomogeneous C ∗-algebras, Mem. Soc. Math. Fr. (N.S.), 71(1999). [43] K. Thomsen, Inductive limit of interval algebras: the simple case. In: Arak; H etal. (eds), Quantum and non-commutative analysis, Kluwer Dordrecht (1993) 399-404. [44] J. Villadsen,: The range of the Elliott invariant, J. Reine Angew. Math., 462, (1995), 31-35. [45] K. Wang, On invariants of C ∗-algebras with the ideal property, Journal of Non- commutative Geometry, Volume 12, Issue 3, 2018, 1199-1225. 40
1603.06979	1	1603	2016-03-22T21:02:35	Wavelets and spectral triples for fractal representations of Cuntz algebras	[ "math.OA" ]	In this article we provide an identification between the wavelet decompositions of certain fractal representations of $C^*-$algebras of directed graphs of M. Marcolli and A. Paolucci, and the eigenspaces of Laplacians associated to spectral triples constructed from Cantor fractal sets that are the infinite path spaces of Bratteli diagrams associated to the representations, with a particular emphasis on wavelets for representations of $\mathcal{O}_D$. In particular, in this setting we use results of J. Pearson and J. Bellissard, and A. Julien and J. Savinien, to construct first the spectral triple and then the Laplace Beltrami operator on the associated Cantor set. We then prove that in certain cases, the orthogonal wavelet decomposition and the decomposition via orthogonal eigenspaces match up precisely. We give several explicit examples, including an example related to a Sierpinski fractal, and compute in detail all the eigenvalues and corresponding eigenspaces of the Laplace Beltrami operators for the equal weight case for representations of Cuntz algebras, and in the uneven weight case for certain representations of $\mathcal{O}_2$, and show how the eigenspaces and wavelet subspaces at different levels are related.	math.OA	math	Wavelets and spectral triples for fractal representations of Cuntz algebras C. Farsi, E. Gillaspy, A. Julien, S. Kang, and J. Packer October 7, 2018 Abstract In this article we provide an identification between the wavelet decompositions of certain fractal representations of C ∗ algebras of directed graphs of M. Marcolli and A. Paolucci [19], and the eigenspaces of Laplacians associated to spectral triples constructed from Cantor fractal sets that are the infinite path spaces of Bratteli diagrams associated to the representations, with a particular emphasis on wavelets for representations of Cuntz C ∗-algebras OD. In particular, in this setting we use results of J. Pearson and J. Bellissard [20], and A. Julien and J. Savinien [15], to construct first the spectral triple and then the Laplace -- Beltrami operator on the associated Cantor set. We then prove that in certain cases, the orthogonal wavelet decomposition and the decomposition via orthogonal eigenspaces match up precisely. We give several explicit examples, including an example related to a Sierpinski fractal, and compute in detail all the eigenvalues and corresponding eigenspaces of the Laplace -- Beltrami operators for the equal weight case for representations of OD, and in the uneven weight case for certain representations of O2, and show how the eigenspaces and wavelet subspaces at different levels first constructed in [8] are related. 2010 Mathematics Subject Classification: 46L05. Key words and phrases: Weighted Bratteli diagrams; Ultrametric Cantor set; Spectral triples; Laplace Beltrami operators; Wavelets. Contents 1 Introduction 2 Cantor sets associated to directed graphs 2.1 Directed graphs and Bratteli diagrams . . . . . . . . . . . . . . . . . . . . 2.2 Cuntz algebras and representations on fractal spaces . . . . . . . . . . . . . 2 4 4 8 3 The action of OD on L2(SA, H) 10 3.1 The Sierpinski fractal representation for OD . . . . . . . . . . . . . . . . . 10 3.2 The measure-preserving isomorphism . . . . . . . . . . . . . . . . . . . . . 12 1 4 Spectral triples and Laplacians for Cuntz algebras 15 4.1 The Cuntz algebra OD and its Sierpinski spectral triple . . . . . . . . . . . 15 . . . . . . . . . . . . . . . . . . . . . . . . 20 4.2 The Laplace -- Beltrami operator 5 Wavelets and eigenfunctions for OD 21 5.1 Wavelets on SA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 6 Spectral triples and Laplacians for the Cuntz algebra OD: the uneven 25 weight case 6.1 The spectral triple . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 6.2 The Laplace -- Beltrami Operator . . . . . . . . . . . . . . . . . . . . . . . . 28 6.3 Eigenvalues and eigenfunctions for the O2 case . . . . . . . . . . . . . . . . 29 1 Introduction In the 2011 paper [19], M. Marcolli and A. Paolucci, motivated by work of A. Jonnson [13] and R. Strichartz [23], studied representations of Cuntz -- Krieger C ∗-algebras on Hilbert spaces associated to certain fractals, and constructed what they termed "wavelets" in these Hilbert spaces. These wavelets were so called because they provided an orthogonal decomposition of the Hilbert space, and the partial isometries associated to the C ∗-algebra in question gave "scaling and translation" operators taking one orthogonal subspace to another. The results of Marcolli and Paolucci were generalized first to certain fractal representations of C ∗-algebras associated to directed graphs and then to representations of higher-rank graph C ∗-algebras C ∗(Λ) by some of the authors of this article in [7] and [8]. The k-graph C ∗-algebras C ∗(Λ) of Robertson and Steger [22] are particular examples of these higher-rank graph algebras, and it was shown in [7] that for these Robertson -- Steger C ∗-algebras there is a faithful representation of C ∗(Λ) on L2(X, µ), where X is a fractal space with Hausdorff measure µ. Moreover, this Hilbert space also admits a wavelet decomposition -- that is, an orthogonal decomposition such that the representation of C ∗(Λ) is generated by "scaling and translation" operators that move between the orthogonal subspaces. As in Marcolli and Paolucci's original construction, the wavelets in [7] and [8] had a characteristic structure, in that they were chosen to be orthogonal to a specific type of function in the path space that could be easily recognized. Earlier, the theory of spectral triples and Fredholm modules of A. Connes had gen- erated great interest [5], and such objects had been constructed for dense subalgebras of several different classes of C ∗-algebras, including the construction of spectral triples by E. Christensen and C. Ivan on the C ∗-algebras of Cantor sets [3], which in turn motivated the work of J. Pearson and J. Bellissard, who constructed spectral triples and related Laplacians on ultrametric Cantor sets [20]. Expanding on the work of Pearson and Bel- lisard, A. Julien and J. Savinien studied similarly constructed Laplacians on fractal sets constructed from substitution tilings [15]. In both the papers of Pearson and Bellissard and of Julien and Savinien, after the Laplacian operators were described, spanning sets of functions for the eigenspaces of the Laplacian were explicitly described in terms of 2 differences of characteristic functions. It became apparent to the authors of the current paper that certain components of the wavelet system as described in [7] and the explicit eigenfunctions given by Julien and Savinien in [15] seemed related, and one of the aims of this paper is to analyze this similarity in the case of C ∗-algebras of directed graphs as represented on their infinite path spaces. Indeed, we will show that under appropriate hypotheses, each orthogonal subspace described in the wavelet decomposition of [7] can be expressed as a union of certain of the eigenspaces of the Laplace -- Beltrami operator from [15]. We suspect that the hypotheses required for this result can be substantially weakened from their statement in Theorem 5.1 below, and plan to explore this question in future work [9]. More broadly, the goal of this paper is to elucidate the connections between graph C ∗-algebras, wavelets on fractals, and spectral triples. We focus here on the case of one particular directed graph, namely the graph ΛD which has D vertices and, for each pair (v, w) of vertices, a unique edge e with source w and range v. Again, many of the results presented here will hold in greater generality; see the forthcoming paper [9] for details. In this paper we introduce the graph C ∗-algebra (also known as a Cuntz algebra) associated to ΛD; discuss the associated representations on fractal spaces as in [19, 7]; and present the associated spectral triples and Laplace -- Beltrami operators associated to (fractal) ultrametric Cantor sets as adapted from recent work by Julien-Savinien, Pearson and Bellissard, Christensen et al., see e.g. [15, 20, 4]. In particular we show in Theorem 5.1 that when one constructs the Laplace -- Beltrami operator of [15] associated to the infinite path space of ΛD (which is an ultrametric Cantor set), the wavelets in [19] are exactly the eigenfunctions of the Laplacian. We then compute in detail all the eigenfunctions and eigenvalues of the Laplace -- Beltrami operator associated to a representation of the Cuntz algebra OD on a Sierpinski type fractal set (see [19] Section 2.6 and Section 3.1 below) for the definition of this representation). For several different choices of a measure on the infinite path space of ΛD, we also compute all the the eigenfunctions and eigenvalues of the associated Laplace -- Beltrami operator; in the case when D = 2 and this measure arises from assigning the two vertices of ΛD the weights r and 1 − r for some r ∈ [0, 1], we compare these results to wavelets associated to certain representations of O2 analyzed in Section 3 of [8]. In further work [9] we will generalize these constructions to more general directed graphs and to higher-rank graphs, and also explain how to generalize certain other spectral triples associated to directed graphs, such as those described in [2], [3], [10], [24], and [14], to higher-rank graphs. The structure of the paper is as follows. In Section 2, we review the definition of directed graphs, with an emphasis on finite graphs and the construction of both the in- finite path space and Bratteli diagrams associated to finite directed graphs, the first as described in [19] among other places, and the second as described in [21]. When the incidence matrices for our graphs are {0, 1} matrices, the infinite path space can defined in terms of both edges and vertices, and we describe this correspondence, together with the identification of the infinite path space Λ∞ with the associated infinite path space of the Bratteli diagram ∂BΛ for a finite directed graph Λ. In so doing, we note that these spaces are Cantor sets. We also review the semibranching function systems of K. 3 Kawamura [17] and Marcolli and Paolucci [19] in this section, with an emphasis on those systems giving rise to representations of the Cuntz algebras OD. In Section 3, we review representations of OD on the L2-spaces of Sierpinski fractals first constructed by Marcolli and Paolucci in [19], and show that these representations are equivalent to the standard positive monic representations of OD defined by D. Dutkay and P. Jorgensen in [6]. In Section 4, we review the construction of spectral triples associated to weighted Bratteli diagrams, described by Pearson and Bellissard in [20] and Julien and Savinien in [15], and provide explicit details of their construction for a variety of weights on the Bratteli diagram ∂BD associated to the graph ΛD. We describe in Theorem 4.9 the conditions under which the measure on ∂BD agrees with the measure introduced by Marcolli and Paolucci, which we describe in Section 2. We also introduce the Laplace -- Beltrami oper- ator of Pearson and Bellissard [20] in this setting and review the specific formulas for its eigenvalues and associated eigenspaces. In Section 5 we review the construction of Mar- colli and Paolucci's wavelets associated to representations of Cuntz -- Krieger C ∗-algebras on the L2-spaces of certain fractal spaces, with the notation for these subspaces provided in earlier papers [7, 8] with an emphasis on representations of the Cuntz C ∗-algebra OD, and prove our main theorem (Theorem 5.1), which is that in all cases that we consider, the wavelet subspaces for Marcolli and Paolucci's representations can be identified with the eigenspaces of the Laplace -- Beltrami operator associated to the related Bratteli dia- gram. In Section 6, we examine certain representations of OD where the weights involved are unevenly distributed among the vertices of ΛD, and specializing to the study of un- even weights associated to representations of O2, we compute explicitly the associated eigenvalues and eigenspaces for the Laplace -- Beltrami operatore in this case, and pro- vide the correspondence between these eigenspaces and certain wavelet spaces for monic representations of O2 first computed in [8]. to Judith Packer). This work was partially supported by a grant from the Simons Foundation (#316981 2 Cantor sets associated to directed graphs We begin with a word about conventions. Throughout this paper, N consists of the positive integers, N = {1, 2, 3, . . .}; we use N0 to denote {0, 1, 2, 3 . . .}. The symbol ZN indicates the set {0, . . . , N − 1}. The Bratteli diagrams we discuss below do not have a root vertex; indeed, we think of the edges in a Bratteli diagram as pointing towards the zeroth level of the diagram. See Remark 2.5 for more details. 2.1 Directed graphs and Bratteli diagrams Definition 2.1. A directed graph Λ consists of a set of vertices Λ0 and a set of edges Λ1 and range and source maps r, s : Λ1 → Λ0. We say that Λ is finite if Λn = {e1e2 . . . en : ei ∈ Λ1, r(ei) = s(ei−1) ∀ i} 4 is finite for all n ∈ N. If γ = e1 · · · en, we define r(γ) = r(e1) and s(γ) = s(en), and we write γ = n. By convention, a path of length 0 consists of a single vertex (no edge): if γ = 0 then γ = (v) for some vertex v. We say that Λ has no sources if vΛn = {γ ∈ Λn : r(γ) = v} 6= ∅ for all v ∈ Λ0 and all n ∈ N. We say that Λ is strongly connected if vΛw = [n∈N{γ ∈ vΛn : s(γ) = w} 6= ∅ for all v, w ∈ Λ0. In a slight abuse of notation, if Λn denotes the set of finite paths of length n, we denote by Λ = ∪n∈N0Λn the set of all finite paths, and by Λ∞ the set of infinite paths of a finite directed graph Λ: Λ∞ =(cid:26)(ei)i∈N ∈ ∞ Yi=1 Λ1 : s(ei) = r(ei+1) ∀ i ∈ N(cid:27). For γ ∈ Λ, we write [γ] ⊆ Λ∞ for the set of infinite paths with initial segment γ: [e1 . . . en] =(cid:8)(fi)i ∈ Λ∞ : fi = ei ∀ 1 ≤ i ≤ n(cid:9). We say that a path γ = e1 . . . en has length n and write γ = n. If γ = (v) is a path of length 0, then [γ] = [v] = {(fi)i ∈ Λ∞ : r(f1) = v}. Given a finite directed graph Λ, the vertex matrix A of Λ is an Λ0 × Λ0 matrix with entry A(v, w) = vΛ1w counting the number of edges with range v and source w in Λ. Remark 2.2. As shown in [18] Corollary 2.2, if Λ is finite and source-free, the cylinder sets {[γ] : γ ∈ Λ} form a compact open basis for a locally compact, totally disconnected, Hausdorff topology on Λ∞.1 If Λ is finite, Λ∞ is also compact. (1) According to [11] Proposition 8.1, a strongly connected finite directed graph Λ has a distinguished Borel measure M on the infinite path space Λ∞ which is given in terms of the spectral radius ρ(A) of the vertex matrix A; M([γ]) = ρ(A)−γPs(γ), (2) where (Pv)v∈Λ0 is the unimodular Perron -- Frobenius eigenvector of the vertex matrix A. (See section 2 of [7] for details). Definition 2.3. Let Λ be a finite directed graph with no sources. The Bratteli diagram associated to Λ is an infinite directed graph BΛ, with the set of vertices V = ⊔n≥0Vn and the set of edges E = ⊔n≥1En such that (a) For each n ∈ N0, Vn ∼= Λ0 and En+1 ∼= Λ1. (b) There are a range map and a source map r, s : E → V such that r(En) ⊆ Vn−1 and s(En) ⊆ Vn for all n ∈ N. 1Note that if Λ is finite, it is also row-finite, according to the definition given in Section 2 of [18]. 5 A path γ of length n ∈ N in BΛ is an element e1e2 . . . en = (e1, e2, . . . , en) ∈ En n Yi=1 which satisfies ei = 1 ∀i, and s(ei) = r(ei+1) for all 1 ≤ i ≤ n − 1. We denote by FBΛ the set of all finite paths in the Bratteli diagram BΛ, and by F nBΛ the set of all finite paths in the Bratteli diagram BΛ of length n. We denote by ∂BΛ the set of infinite paths in the Bratteli diagram BΛ; ∂BΛ = {e1e2 · · · = (e1, e2, . . . ) ∈ ∞ Yn=1 En : ei = 1, s(ei) = r(ei+1) ∀ i ∈ N}. Given a (finite or infinite) path γ = e1e2 . . . in BΛ and m ∈ N, we write If m = 0 we write γ[0, 0] = r(γ). γ[0, m] = e1e2 · · · em. Remark 2.4. Any finite path γ of a length n in a directed graph (or a Bratteli diagram) is given by a string of n edges e1e2 . . . en, which can be written uniquely as a string of vertices v0v1 . . . vn such that r(ei) = vi−1 and s(ei) = vi for 1 ≤ i ≤ n. Conversely, if the vertex matrix A has all entries either 0 or 1 (as will be the case in all of our examples), a given string of vertices v0v1 . . . vn with vi ∈ Vn for all n ∈ N0 corresponds to at most one string of edges, and hence at most one finite path γ. Thus even though our formal definition of a path is given as a string of edges, sometimes we use the notation of a string of vertices for a path. Remark 2.5. Note that our description of a Bratteli diagram is different from the one in [15] and [1]. First, the edges in En in [15] and in [1] have source in Vn and range in Vn+1; in other words, they point in the opposite direction from our edges. More substantially, though, in [15] and [1] every finite (or infinite) path in a Bratteli diagram starts from a vertex called a root vertex, ◦, and any finite path that ends in Vn is given by ǫr(e1)e1e2 . . . en, where for each vertex v ∈ V0, there is a unique edge ǫv connecting ◦ and v. This implies that a finite path that ends in Vn consists of n + 1 edges in their Bratteli diagram. However, our description of a Bratteli diagram in Definition 2.3 does not include a root vertex, and a finite path that ends in Vn consists of n edges. Thus, when we discuss Theorem 4.3 of [15] in Sections 4.2 and 6 below, we will need to introduce a single path, the "empty path" of length -1, which we will denote by γ[0,−1] for any and all paths γ ∈ FBΛ. The cylinder set of this path is [◦] = ∂BΛ when we translate Theorem 4.3 of [15] to our setting. Remark 2.6. As is suggested by the notation, a finite directed graph and its associated Bratteli diagram encode the same information in their sets of finite and infinite paths. We wish to emphasize this correspondence in this paper, to illuminate the way tools from a variety of disciplines combine to give us information about wavelets on fractals. 6 Remark 2.7. If Λ is a strongly connected finite directed graph, then Λ has no sources by Lemma 2.1 of [11]. Hence every vertex of the associated Bratteli diagram BΛ also receives an edge. Example 2.8. Consider a directed graph Λ with two vertices v, w and four edges f1, f2, f3 and f4 given as follows: f1 v f2 f3 w f4 Note that Λ is finite and strongly connected, and (consequently) has no sources. The vertex matrix A is given by and the associated Bratteli diagram BΛ is A =(cid:18)1 1 1 1(cid:19) , v0 v1 v2 .... v3 w0 w1 w2 .... w3 Proposition 2.9. Let Λ be a finite directed graph. If every vertex v in the directed graph Λ receives two distinct infinite paths, then Λ∞ (equivalently, ∂BΛ) has no isolated points and hence it is a Cantor set. Proof. Recall that a Cantor set is a totally disconnected, compact, perfect topological space. Moreover, Λ∞ is always compact Hausdorff and totally disconnected by Corollary 2.2 of [18], so it will suffice to show that Λ∞ has no isolated points. Suppose Λ∞ has an isolated point (ei)i∈N. Since the cylinder sets form a basis for the topology on Λ∞, this implies that there exists n ∈ N such that [e1 · · · en] only contains (ei)i∈N. In other words, for each m ≥ n, there is only one infinite path with range s(em), contradicting the hypothesis of the proposition. Corollary 2.10. If Λ is a finite directed graph with {0, 1} vertex matrix A and every row sum of A is at least 2, then Λ∞ (equivalently, ∂BΛ) is a Cantor set. Proof. Note that the sum of the vth row of A represents the number of edges in Λ with range v. If every vertex receives at least two edges, then any cylinder set [γ] will contain infinitely many elements, so Λ∞ has no isolated points. Corollary 2.10 tells us that the infinite path space of Example 2.8 is a Cantor set. 7 2.2 Cuntz algebras and representations on fractal spaces Definition 2.11 ([6, Definition 2.1]). Fix an integer D > 1. The Cuntz algebra OD is the universal C ∗-algebra generated by isometries {Ti}D−1 i=0 satisfying the Cuntz relations and T ∗ j Ti = δijI, D−1 TiT ∗ i = I. Xi=0 (3) (4) The above definition of OD is equivalent to the definition of OAD in the beginning of section 2 of [19] associated to the matrix AD that is a D× D matrix with 1 in every entry: AD = . (5) 1 1 1 ... 1 1 1 1 ... 1 ... ... 1 1 1 ... 1 1 1 1 ... 1 ... ... ...     As had been done previously by K. Kawamura [17], Marcolli and Paolucci constructed representations of OD (and more generally, the Cuntz -- Krieger algebras OA associated to a matrix A) by employing the method of "semibranching function systems." We note for completeness that the semibranching function systems of Kawamura [17] were for the most part defined on finite Euclidean spaces, e.g. the unit interval [0, 1], whereas the semibranching function systems used by Marcolli and Paolucci [19] were mainly defined on Cantor sets. Definition 2.12 (cf. [17], [19, Definition 2.1], [1, Definition 2.16] ). Let (X, µ) be a measure space, fix an integer D > 1 and let {σi : X → X}i∈ZD be a collection of µ- measurable maps. The family of maps {σi}i∈ZD is called a semibranching function system on (X, µ) with coding map σ : X → X if the following conditions hold: 1. For i ∈ ZD, set R[i] = σi(X). Then we have µ(X\ ∪i∈ZD R[i]) = 0 and µ(R[i] ∩ R[j]) = 0 for i 6= j. 2. For i ∈ ZD, we have µ ◦ σi ≪ µ and the Radon -- Nikodym derivative satisfies 3. For all i ∈ ZD, we have d(µ ◦ σi) dµ > 0, µ-a.e. σ ◦ σi(x) = x, µ-a.e. (6) Kawamura and then Marcolli and Paolucci observed the following relationship between semibranching function systems and representations of OD : 8 Proposition 2.13 (cf. [19, Proposition 2.4], [6, Theorem 2.22]). Let (X, µ) be a measure space, and let {σi : X → X}i∈ZD be a semibranching function system on (X, µ) with coding map σ : X → X. For each i ∈ ZD define Si : L2(X, µ) → L2(X, µ) by Si(ξ)(x) = χR[i](x)(cid:16) dµ ◦ σi dµ (σ(x))(cid:17)− 1 2 ξ(σ(x)) for ξ ∈ L2(X, µ) and x ∈ X. Then the family {Si}i∈ZD satisfies the Cuntz relations Equations (3) and (4), and therefore generates a representation of the Cuntz algebra OD. Example 2.14. Let ΛD be the directed graph associated to the vertex matrix AD. We can define a semibranching function system {(σi)i∈ZD , σ} on the Cantor set (Λ∞ D , M) by thinking of elements of Λ∞ D as sequences of vertices (vi)i∈N0 with vj ∈ ZD ∀ j. With this convention, we set σi(v0v1v2 . . .) = (iv0v1v2 . . .) and σ(v0v1 . . .) = (v1v2 . . .). Then the Radon -- Nikodym derivative d(M ◦σi) dM is given by d(M ◦ σi) dM = 1 D since the cylinder set R[i] has measure 1 given by D for all i, and the associated operators Si are Si(ξ)(v0v1v2 . . .) =(cid:26) √Dξ(v1v2 . . .) 0 if v0 = i else. This representation of OD is faithful by Theorem 3.6 of [7], since every cycle in ΛD has an entrance. Example 2.15 (cf. [19, Proposition 2.6]). Take an integer D > 1, and let KD =Q∞ j=1[ZD]j, which is called the Cantor group on D letters in Definition 2.3 of [6]. As described in Section 2 of [8], KD has a Cantor set topology which is generated by cylinder sets According to Section 3 of [6], there is a measure νD on KD given by [n] = {(ij)∞ j=1 ∈ KD : i1 = n}. νD([n1n2 . . . nm]) = 1 D = 1 Dm . m Yj=1 Note that νD is a Borel measure on KD with respect to the cylinder-set Cantor topology. For each j ∈ ZD, define σj on KD by σj ((i1i2 · · · ik · · · )) = (ji1i2 · · · ik · · · ). Then R[j] = σj(KD) = {(ji1i2 · · · ik · · · ) : (i1i2 · · · ik · · · ) ∈ KD} = [j], 9 P =(cid:18) 1 D , 1 D , . . . , 1 D(cid:19), and consequently M([e1 . . . en]) = 1 and, denoting by σ the one-sided shift on KD, σ ((i1i2 · · · ik · · · )) = (i2i3 · · · ik+1 · · · ), we have that σ ◦ σj(x) = x for all x ∈ KD and j ∈ ZD. Marcolli and Paolucci show in Section 2.1 of [19] that this data gives a semibranching function system. Moreover, since the measure of each set R[i] is 1 D , the Radon -- Nikodym derivative d(νD◦σi) satisfies dνD d(νD ◦ σi) dνD = 1 D . Thus, Proposition 2.13 implies that there is a family of operators {Si}i∈ZD ⊆ B(L2(KD, νD)) that generates a representation of the Cuntz algebra OD. Moreover, this representation is faithful by Theorem 3.6 of [7]. To see this, let ΛD denote the directed graph with vertex matrix AD, and note that labeling the vertices of ΛD by {0, 1, . . . , D− 1} allows us to identify an infinite path (ei)i∈N ∈ ∂BD with the sequence (r(ei))i∈N ∈ KD. Moreover, in this case the Perron -- Frobenius eigenvector associated to AD is Dn+1 = νD(cid:0)[r(e1)r(e2)· · · r(en)s(en)](cid:1). Since the cylinder sets generate the topology on both KD and on ∂BD, this identification is measure-preserving. Thus, the representation {Si}i∈ZD of OD on L2(KD, νD) is equivalent to the infinite path representation of Example 2.14. We can apply Theorem 3.6 of [7] to this latter representation to conclude that it is faithful, since every cycle in the graph ΛD associated to AD has an entry. In other words, C ∗(cid:0){Si}i∈ZD(cid:1) ∼= OD. 3 The action of OD on L2(SA, H) As mentioned in the Introduction, we wish to show that when we represent OD on a 2-dimensional Sierpinski fractal SA, this representation of OD also gives rise to wavelets. We will then compare these wavelets with the eigenfunctions of the Laplace -- Beltrami operator ∆s of [15] that is associated to AD, the D × D matrix of all 1's (that is, the matrix associated to the Cuntz algebra OD). To compare these functions, we will establish a measure-preserving isomorphism between SA and the infinite path space of the directed graph (equivalently, Bratteli diagram) associated to OD in this section. (See Theorem 3.1 below). 3.1 The Sierpinski fractal representation for OD Let N and D be positive integers with N ≥ 2, and let A be a N × N{0, 1}-matrix with exactly D entries consisting of the number 1. Suppose that the nonzero entries of A are in 10 positions {(aj, bj)}D−1 j=0 , where aj, bj ∈ {0, 1 · · · , N − 1} and in a lexicographic ordering we have (a0, b0) < (a1, b1) < · · · < (aD−1, bD−1). Here we say (a, b) < (a′, b′) if either a < a′ or if a = a′ and b < b′. In Section 2.6 of [19], Marcolli and Paolucci defined the Sierpinski fractal associated to A, SA ⊂ [0, 1]2, as follows: SA = (cid:26)(x, y) =(cid:18) ∞ Xi=1 For each j ∈ ZD, we define xi N i , ∞ Xi=1 yi N i(cid:19) : xi, yi ∈ ZN , Axi,yi = 1, ∀i ∈ N(cid:27). bj + , + N y N aj N τj(x, y) =(cid:18) x Lemma 2.23 of [19] tells us that the operators {τj}j∈ZD form a semibranching function system with coding map τ , and hence determine a representation of the Cuntz algebra OD associated to AD given in (5), on the Hilbert space L2(SA, H). Here H is the Hausdorff measure on the fractal SA. N(cid:19) and τ (x, y) =(cid:18)N(cid:16)x − N(cid:17), N(cid:16)y − N(cid:17)(cid:19). (7) x1 y1 According to the work of Hutchinson [12], we have D Moreover, the work of [12] shows that the Hausdorff measure H on SA is the unique SA = τj(SA). [i=1 Borel probability measure on SA satisfying the self-similarity equation H = D−1 Xi=0 H(τj(SA)) = 1 D 1 D In other words, It follows that, since (τj)∗(H). (8) H(SA)) = 1 D . ∞ yi τj(SA)) =(cid:26)(cid:18) ∞ N i(cid:19) : (x1, y1) = (aj, bj)(cid:27), xi Xi=1 N i , H(cid:18)(cid:26) ∞ N i ) ∈ SA : (x1, y1) = (aj, bj)(cid:27)(cid:19) = Xi=1 Xi=1 Xi=1 xi N i , yi ∞ 1 D . By repeatedly applying the measure-similitude equation (8) we obtain ∞ H(cid:18)(cid:26)(cid:16) Xi=1 xi N i , ∞ Xi=1 yi N i(cid:17) ∈ SA : ∀1 ≤ i ≤ M, (xi, yi) = (aji, bji)(cid:27)(cid:19) = H(τj1 ◦ τj2 ◦ · · · ◦ τjM (S)A) =(cid:16) 1 D(cid:17)M . (9) 11 3.2 The measure-preserving isomorphism In this section, we discuss in more detail the relationship between the representation of OD on L2(SA, H) and the infinite path representation of OD on L2(∂BD, M) described in Example 2.14. First, we note that the Hausdorff dimension of the Sierpinski fractal SA introduced above is ln D ln N , as established in Hutchinson's paper [12].2 In particular, in the classical case of the Sier- pinski triangle corresponding to the 2 × 2 matrix A =(cid:18)1 0 1 1(cid:19) , the Hausdorff dimension of SA is ln 3 ln 2. The main goal of this section is to prove the following: Theorem 3.1. Let A be the N × N matrix with entries consisting of only 0's and 1's with D incidences of 1's in the entries (a0, b0) < (a1, b1) < · · · < (aD−1, bD−1), where aj, bj ∈ ZN . Consider the Sierpinski gasket fractal SA = (cid:26)(cid:16)Xi∈N xi N i ,Xi∈N yi N i(cid:17) : A(xi, yi) = 1, ∀i ∈ N(cid:27). Then there is a measure-theoretic isomorphism Υ = Φ ◦ Θ : (∂BD, M) → (SA, H), where (∂BD, M) is the infinite path space of the Bratteli diagram associated to the D × D matrix with all ones, and M is the measure given by Equation (2): M[γ] = D−γ−1. Moreover, if {Si}i∈ZD denotes the infinite path representation of OD on (∂BΛ, M), and {Ti}i∈ZD denotes the representation of OD on (SA, H) associated to the semibranching function system (7), then for all i ∈ ZD, Ti = Si ◦ Υ. Proof. Let SA denote the D-element symbol space of pairs from ZN with 1's in the cor- responding entry of A : SA = {(a0, b0), (a1, b1), (a2, b2),· · · , (aD−1, bD−1)} ⊂ ZN × ZN , 2This formula is not in line with [19, Equation (2.64)], which gives ln D/(2 ln N ) for the Hausdorff dimension. However, said equation appears to be a typo: the dimension should be 2 when D = N 2 (i.e. when SA is the unit square). 12 and let XA be the infinite product space XA = Q∞ i=1 SA. Giving SA the discrete topology and XA the product topology, we see that XA is a Cantor set, by the arguments of Section 2 of [8]. For every i ∈ N, let µi,A be the normalized counting measure on SA; that is, for S ⊂ SA, #(S) , µi,A(S) = D and let µA denote the infinite product measure µA =Q∞ (cid:2)(aj1, bj1)(aj2, bj2)· · · (ajM , bjM )(cid:3) denote the cylinder set i=1(µi,A). Note if we let (10) [(aj1, bj1)(aj2, bj2)· · · (ajM , bjM )] = {((xi, yi))∞ i=1 ∈ XA : (xi, yi) = (aji, bji) ∀ 1 ≤ i ≤ M}, then µA([(aj1, bj1)(aj2, bj2)· · · (ajM , bjM )]) = 1 DM . Define now a map Φ : XA → SA by Φ(cid:0)((xi, yi))∞ i=1(cid:1) =(cid:16) ∞ Xi=1 xi N i , ∞ Xi=1 yi N i(cid:17). i, y′ The map Φ is continuous from the product topology on XA to the topology on SA inherited from the Euclidean topology on [0, 1] × [0, 1]. The map Φ is not one-to-one, but if we let E ⊂ XA denote the set of points on which Φ is not injective, µA(E) = 0. Indeed, let's examine the set of points of XA where Φ may not be one-to-one: non-injectivity can come from pairs of sequence of the forms (xi, yi)i, (x′ i)i where xi is eventually N − 1 and x′ i is eventually 0, and similarly exchanging x and y. Notice also that if A has no ones either on the first or on the last row, there will be no such pairs for which xi is eventually N − 1 and x′ i is eventually 0. Therefore, since A has D total entries equaling 1, if two such pairs (xi, yi)i and (x′ i)i are going to have the same image under Φ, there need to be at most D − 1 ones on the first row, and the same on the last row. Therefore, the measure of the set of pairs (xi, yi)i for which xi is eventually N − 1 is smaller than [(D − 1)/D]n for all n: it has zero measure. We reason similarly for the set of pairs (xi, yi) for which xi is eventually 0, for which yi is eventually 0 and for which yi is eventually N − 1. In conclusion, the set of points in XA on which Φ has a risk of not being one-to-one has measure zero. i, y′ We also note that since Φ is continuous, it is a Borel measurable map, and that for any Borel subset B of SA, µA ◦ [Φ]∗(B) = H(B). This is the case because a length-M cylinder set in SA (that is, any cylinder set (cid:2)(x1, y1), . . . , (xM , yM )(cid:3) consisting of all points in SA whose first M pairs of N-adic digits 13 are fixed) has H-measure cylinder sets of the form 1 DM , whereas when one pulls such sets back via Φ, we obtain which also have measure D−M . Since these sets generate the Borel σ-algebras for SA and XA respectively, we get the desired equality of the measures. (cid:2)(aj1, bj1)(aj2, bj2)· · · (ajM , bjM )(cid:3) ⊆ XA Now let BD be the Bratteli diagram with D vertices at each level, associated to the matrix AD given in (5) (and, hence, to the directed graph ΛD with D vertices and all possible edges). We equip the infinite path space ∂BD with the measure of Equation (2), which in this case is M([γ]) = D−γ−1. Label the vertices of Λ0 by ZD = {0, 1, . . . , D−1}, and define Θ : ∂BD → XA by Θ((ei)i≥1) =(cid:0)(ar(e1), br(e1)), (ar(e2), br(e2)), (ar(e3), br(e3)), . . .)(cid:1); in other words, Θ takes an infinite path (written in terms of edges) (ei)i∈N to the sequence of vertices (r(ei))i∈N it passes through, and then maps this sequence of vertices to the corresponding element of XA. The map Θ is bijective, since each pair of vertices has exactly one edge between them. In addition, both Θ and Θ−1 are continuous, since both the topology on ∂BD and the topology on XA are generated by cylinder sets. In other words, Θ is a homeomorphism, and M = µA ◦ [Θ]∗. We thus have shown that Υ = Φ◦Θ is a Borel measure-theoretic isomorphism between the measure spaces (∂BD, M) and (SA, H). A routine computation, using the fact that N i ! , will show that for any i ∈ ZD, Ti = Si ◦ Υ to finish the proof. Υ((ei)i∈N) = Xi∈N ,Xi∈N ar(ei) N i br(ei) We now recall the definition of Dutkay and Jorgensen [6] of a monic representation of OD : Definition 3.2 (cf. [6, Definition 2.6]). Let D ∈ N, and let KD be the infinite product Cantor group defined earlier. Let σi : KD → KD, 0 ≤ i ≤ D − 1 be as in Example 2.15. A nonnegative monic system is a pair (µ, (fi)i∈ZD ) where µ is a Borel probability measure on KD and (fi)i∈ZD are nonnegative Borel measurable functions in L2(KD, µ) such that µ ◦ σ−1 i ≪ µ, and such that for all i ∈ ZD d(µ ◦ σ−1 i ) dµ = (fi)2 with the property that fi(x) 6= 0, µ a.e. on σi(KD), ∀i ∈ ZD. By Equation (2.9) of [6], there is a natural representation of OD on L2(KD, µ) associ- ated to a monic system (µ, (fi)i∈ZD ) given by Sif = fi(f ◦ σ), (i ∈ ZD, f ∈ L2(KD, µ)). If µ = νD and we set fi = √Dχσi(KD), the corresponding monic system is called the standard positive monic system for OD. 14 Corollary 3.3. The representation of OD on L2(SA, H) described in Section 3.1 above is equivalent to the monic representation of OD corresponding to the standard positive monic system on L2(KD, νD). Proof. Theorem 3.1, combined with the measure-theoretic identification of (KD, νD) and (∂BD, M) established in Example 2.15, implies that we have a measure-theoretic iso- morphism between (KD, νD) and (SA, H). Thus, to show that the corresponding rep- resentations of OD are unitarily equivalent, it only remains to check that the opera- tors Si = fi(f ◦ σ) associated to the standard positive monic system, and the operators {Ti}i∈ZD , match up correctly. To that end, observe that Si(ξ)(v0v1 . . .) = fi(v0v1 . . .)ξ(v1v2 . . .) =(√Dξ(v1v2 . . .) 0 if v0 = i else. = Si(ξ)(v0v1 . . .). Since Theorem 3.1 established that the operators Si and Ti are unitarily equivalent, the Corollary follows. 4 Spectral triples and Laplacians for Cuntz algebras Let AD be the D × D matrix with 1 in every entry and consider the Bratteli diagram BD associated to AD. If D ≥ 2, then every row sum of AD is at least 2 by construction, and hence the associated infinite path space of the Bratteli diagram, ∂BD, is a Cantor set. In this section, by using the methods in [15], we will construct a spectral triple on ∂BD. This spectral triple gives rise to a Laplace -- Beltrami operator ∆s on L2(∂BD, µD), where µD is the measure induced from the Dixmier trace of the spectral triple as in Theorem 4.9 below. We also compute explicitly the orthogonal decomposition of L2(∂BD, µD) in terms of the eigenfunctions of the Laplace -- Beltrami operator ∆s (cf. [15, Theorem 4.3]). 4.1 The Cuntz algebra OD and its Sierpinski spectral triple Definition 4.1. Let Λ be a finite directed graph; let F (BΛ)◦ be the set of all finite paths on the associated Bratteli diagram, including the empty path whose length we set to −1 by convention.. A weight on BΛ (equivalently, on Λ) is a function w : F (BΛ)◦ → (0,∞) satisfying (a) w(◦) = 1 (b) lim n→∞ sup{w(η) : η ∈ Λn = F nBΛ} = 0, where we denoted by Λn = F nBΛ the set of finite paths of length n on Λ (equiva- lently, BΛ). (c) For any finite paths η, ν with s(η) = r(ν), we have w(ην) < w(η). 15 A Bratteli diagram BΛ with a weight w is called a weighted Bratteli diagram. Remark 4.2. Observe that a weight that satisfies Definition 2.9 of [15] on the vertices of a Bratteli diagram BΛ induces a weight on the finite paths of the Bratteli diagram as in Definition 4.1 above. In fact in [15] and [20] the authors define a weight on FBΛ by defining the weight first on vertices, and then extending it to finite paths via the formula w(η) = w(s(η)), for η ∈ FBΛ. space ∂BΛ ∼= Λ∞; see Theorem 4.9 below for details. We will show below that a weight on BΛ induces in turn a measure on the infinite path Definition 4.3. An ultrametric d on a topological space X is a metric satisfying the strong triangle inequality: d(x, y) ≤ max{d(x, z), d(y, z)} for all x, y, z ∈ X. Proposition 4.4 ([15, Proposition 2.10]). Let BΛ be a weighted Bratteli diagram with weight w. We define a function dw on ∂BΛ × ∂BΛ by dw(x, y) =(w(x ∧ y) 0 if x 6= y otherwise , where x∧ y is the longest common initial segment of x and y. (If r(x) 6= r(y) then we say x ∧ y is the empty path ◦, and w(◦) = 1.) Then dw is an ultrametric on ∂BΛ. Note that the ultrametric dw induces the same topology on ∂BΛ as the cylinder sets in (1); thus, (∂BΛ, dw) is called an ultrametric Cantor set. Definition 4.5. Let AD be a D × D matrix with 1 in every entry and let BD be the associated Bratteli diagram. Fix λ > 1, and set d = ln D/ ln λ. We define a weight wλ D on the Bratteli diagram BD by setting (a) wλ D(◦) = 1. (b) For any level 0 vertex v ∈ V0 of BD, wλ (c) For any finite path γ ∈ F nBD of length n, D(v) = 1 D . wD(γ) = λ−n 1 D . According to [15], after choosing a weight on BD, we can build a spectral triple asso- ciated to it as in the following Theorem. Note that this result is a special case of Section 3 of [15]. 16 Theorem 4.6. Fix an integer D > 1 and λ > 1. Let (BD, wλ D) be the weighted Bratteli diagram with the choice of weight wλ w) be the associated ultrametric Cantor set. Then there is an even spectral triple (CLip(∂BD),H, πτ , /D, Γ), where D as in Definition 4.5. Let (∂BD, dλ • CLip(∂BD) is the pre-C ∗-algebra of Lipschitz continuous functions on (∂BD, dλ w), • for each choice function τ : FBD → ∂BD × ∂BD,3 a faithful representation πτ of CLip(∂BD) is given by bounded operators on the Hilbert space H = ℓ2(FBD) ⊗ C2 as πτ (f ) = Mγ∈F (BD)◦(cid:18)f (τ+(γ)) 0 0 f (τ−(γ))(cid:19) ; • the Dirac operator /D on H is given by /D = Mγ∈F (BD)◦ 1 0(cid:19) ; D(γ)(cid:18)0 1 • the grading operator is given by Γ = 1ℓ2(F (BD)◦) ⊗(cid:18)1 0 −1(cid:19). 1 wλ 0 Definition 4.7 (cf. [15, Theorem 3.8]). The ζ-function associated to the spectral triple of Theorem 4.6 is given by ζ λ D(s) := 1 2 Tr( /D−s) = Xγ∈F (BD)◦(cid:0)wλ D(γ)(cid:1)s . (11) Proposition 4.8 (cf. [15, Theorem 3.8]). The ζ-function in Equation (11) has abscissa of convergence equal to d = ln D/ ln λ. Proof. By a straightforward calculation we get (if we denote by F q(BD)◦ the set of paths of length q): Xγ∈F (BD)◦(cid:16)wλ D(γ)(cid:17)s = D−s Xq≥−1 Card(F q(BD)◦)λ−qs = D−s Xq≥−1 Dq+1λ−qs, where Card(S) denotes the cardinality of the set S. It is clear that this sum converges precisely when D/λs is smaller than 1, that is whenever s > ln D ln λ . 3A choice function τ : FBD → ∂BD × ∂BD is a function that satisfies τ (γ) =: (τ+(γ), τ−(γ)) where dw(τ+(γ), τ−(γ)) = wλ D(γ). 17 D associated to the ultrametric dλ mension of ∂BD ∼= Λ∞ It is known that the abscissa of convergence coincides with the upper Minkowski di- w [20, Theorem 2]. In the self-similar cases (when the weight is given as in Definition 4.5), the upper Minkowski dimension turns out to coincide with the Hausdorff dimension [16, Theorem 2.12]. In particular, when the scaling factor λ is just N, the Hausdorff dimensions of (Λ∞ ) and SA coincide, where we equip SA with the metric induced by the Euclidean metric on [0, 1]2. D , dwN D The Dixmier trace µλ D(f ) of a function f ∈ CLip(∂BD) is given by the expression below; see Theorem 3.9 of [15] for details. µλ D(f ) = lim s↓d Tr( /D−sπτ (f )) Tr( /D−s) = lim s↓d Tr( /D−sπτ (f )) 2ζ λ D(s) . (12) In particular the limit given in (12) induces a measure µλ D on ∂BD characterized as follows. If f = χ[γ] is the characteristic function of a cylinder set [γ], and if FγBD = {η ∈ FγBD : η = γη′} denotes the set of all finite paths with initial segment γ, we have µλ D([γ]) = µλ D(χ[γ]) = lim s↓d Pη∈Fγ (BD)◦(cid:0)wλ Pη∈F (BD )◦(cid:0)wλ D(η)(cid:1)s D(η)(cid:1)s . (13) D on ∂BD It actually turns out, as we prove in Theorem 4.9 below, that the measure µλ is independent of λ; so we will also write, with notation as above µD([γ]) = µD(χ[γ]) = µλ D([γ]) = µλ D(χ[γ]) Moreover, by combining Theorem 3.1 with Theorem 4.9 below, we see that µD agrees with the Hausdorff measure of SA. Theorem 4.9. For any choice of scaling factor λ > 1, the measure µλ D on ∂BD induced by the Dixmier trace agrees with the measure M associated to the infinite path representation of OD. Namely, for any finite path γ ∈ FBD, we have µD([γ]) = 1 Dγ+1 = M([γ]). (14) Proof. Note that, although the proof of this Theorem is very long for the more general case of Cuntz -- Krieger algebras (cf. [15, Theorem 3.9]), it considerably simplifies for the case of Cuntz algebras covered here. First note that for the choice of the empty path γ = ◦ (whose cylinder set corresponds to the whole space), we have f (s) = Pη∈F (BD)◦(wλ Pη∈F (BD)◦(wλ D, for a finite path γ 6= ◦ of length n in F nBD. Define, according D(η))s D(η))s = 1 = µλ D ) = M(Λ∞ D(Λ∞ (15) D ). Now we will compute µλ to Equation (13), f (s) = Pη∈Fγ BD 1 +Pη∈F BD 18 (wλ D(η))s D(η))s . (wλ (16) Note that in the above expression we isolated the term corresponding to the empty = 1s = 1. Moreover, since γ is not the empty path, then η = ◦ D(η)s only depends on D(η) = D−1λ−(n+q). For path, for which (cid:16)wλ does not occur in the sum in the numerator. If η ∈ FγBD, then wλ the length of η, say η = n + q for some q ∈ N0, and hence wλ q ∈ N0, let D(◦)(cid:17)s F qBD = {η ∈ FBD : η = q}, F q γBD = {η ∈ FγBD : η = n + q}. Then we can write f (s) = Card(F q D−sPq∈N0 1 + D−sPq∈N0 Card(F q γBD)(cid:0)λ−(n+q)(cid:1)s γBD)(cid:0)λ−q(cid:1)s . Since the vertex matrix AD of the Bratteli diagram BD has 1 in every entry, every edge in BD has D possible edges that could follow it. Also note that η ∈ F qBD has its range in V0 and its source in Vq, and hence we get Card(F qBD) = Dq+1. But any finite path η ∈ F q paths η ∈ F q fixed, we get γBD can be written as η = γη′. Since γ is fixed, the number of γBD is the same as the number of possible paths η′. Since r(η′) = s(γ) is also By multiplying both numerator and denominator of f (s) by Ds, we obtain Card(F q γBD) = Dq. 1 λns Pq∈N0 Ds +Pq∈N0 Dqλ−qs Dq+1λ−qs f (s) = = 1 λns Dq(cid:0)λ−(n+q)(cid:1)s D−sPq∈N0 Dq+1(cid:0)λ−q(cid:1)s = 1 + D−sPq∈N0 λs(cid:17)q Pq∈N0(cid:16) D λs(cid:17)q(cid:17) (cid:16)Ds + DPq∈N0(cid:16) D λs(cid:17)q λs < 1, thus Pq∈N0(cid:16) D . ln λ , we have D Since s > ln D (again multiplying numerator and denominator of f (s) by 1 − D λs ), converges and is equal to 1 1− D λs . Thus f (s) = 1 λns 1 1− D λs = 1 λns 1 λs )Ds + D(cid:17) Now take the limit s ↓ d and recall that λd = D. So we have (1 − D (cid:16)(1 − D (Ds + D 1 1− D λs ) λs ) → 0 and hence f (s) = lim s↓d 1 λnd 1 D = 1 Dn 1 D = 1 Dn+1 , which is the desired result by Equation (13). 19 4.2 The Laplace -- Beltrami operator In Section 4 of [15], the authors use the spectral triple associated to a weighted Bratteli diagram to construct a non-positive definite self-adjoint operator with discrete spectrum (which they fully describe) defined on the infinite path space of the given Bratteli dia- gram. Moreover, they show in Theorem 4.3 of [15] that the eigenfunctions of ∆s form an orthogonal decomposition of the L2-space of the boundary. Therefore, by applying the results of Section 4 of [15] to the spectral triples of Section 4.1 above, we obtain, after we choose a weight wλ D on BD as in Definition 4.5, a non- positive definite self-adjoint operator ∆s on L2(∂BD, µD) for any s ∈ R, where µD is the measure on ∂BD given in (13). (Recall that µD does not depend on λ). Namely, for any s ∈ R, the Laplace -- Beltrami operator ∆s on L2(∂BD, µD) is given by hf, ∆s(g)i = Qs(f, g) = 1 2ZE Tr( /D −s[ /D, πτ (f )]∗ [ /D, πτ (g)] dµD(τ ), (17) where Dom Qs = span{χ[γ] : γ ∈ FBD} and Qs is a closable Dirichlet form, and µD(τ ) is the measure induced by the Dixmier trace on the set E of choice functions. Moreover, the eigenfunctions of ∆s form an orthogonal decomposition of L2(∂BD, µD). In the remainder of this section we give the details of this decomposition and formulas for the eigenvalues. In Section 5 below, we describe the relationship between this orthogonal decomposition and the wavelet decomposition of L2(Λ∞, M) computed in [7]. Theorem 4.10. [15, Theorem 4.3] Let (BD, wD D) be the weighted Bratteli diagram as in Theorem 4.6. (Note that we made here the choice λ = D for simplicity.) Let ∆s be the Laplace -- Beltrami operator on L2(∂BD, µD) given by (17). Then the eigenvalues of ∆s are 0, associated to the constant function 1, and the eigenvalues {λη}η∈(F BD)◦ with corresponding eigenspaces {Eη}η∈(F BD)◦ of ∆s are given by λ◦ =(cid:0)Gs(◦)(cid:1)−1 = λη = −2 − 2D3−s 1 − D(3−s)η 1 − D3−s − ; 2D D − 1 2D3η+4 (D − 1)Ds(η+1) , η ∈ F (BD) E◦ = spannD−1(cid:0)χv − χv′(cid:1)o : v 6= v′ ∈ V0o, with Eη = spann Dη+2(cid:0)χ[ηe] − χ[ηe′](cid:1) : η ∈ F (BD), e 6= e′, e = e′ = 1, r(e) = r(e′) = s(η)o. Proof. This follows from evaluating the formulas given in Theorem 4.3 of [15], using Theorem 4.9 above to calculate the measures of the cylinder sets, and recalling that the diameter diam[γ] of a cylinder set is given by the weight of γ. 20 To be precise, since there are D edges with a given range v, the size of the set (cid:8)(e, e′) ∈ Λ1 × Λ1 : r(e) = r(e′) = v, e 6= e′(cid:9) is D(D − 1) for any vertex v. Therefore, for any path η ∈ Λ, the constant Gs(η) from Theorem 4.3 of [15] is given by Gs(η) = D(D − 1)D−2(η+2) 2wD(η)s−2 = (D − 1)Ds(η+1) 2D4η+5 . Observe that, in the notation of [15], a path of "length 0" corresponds to the empty path ◦, that is, whose cylinder set gives entire infinite path space, and a path of "length 1" corresponds to a vertex. In general, the length of a path in [15] corresponds to the number of vertices that this path traverses; hence a path of length n for them is a path of length n − 1 for us. In order to compute the eigenvalues λη described in Theorem 4.3 of [15], then, we also need to calculate Gs(◦) = Gs(Λ∞). Since the infinite path space has diameter 1 by Proposition 2.10 of [15], we obtain Gs(◦) = Gs(Λ∞) = D(D − 1) 2D2 = D − 1 2D . Now, if we denote the empty path ◦ by a path of "length −1," we can rewrite the formula (4.3) from [15] for the eigenvalue λη associated to a path η as 1 Dk+2 − 1 Gs(η[0, k]) − Dk+1 1 Dη+1Gs(η) λη = = η−1 Xk=−1 1 − D D D−1 2D + η−1 Xk=0 1 − D Dk+2 2D4k+5 (D − 1)Ds(k+1) − 2D3η+4 (D − 1)Ds(η+1) = −2 − 2D3−s 1 − D(3−s)η 1 − D3−s − 2D3η+4 (D − 1)Ds(η+1) , using the notation of Definition 2.3, and the fact that η−1 Xk=0 2D3k+3 Ds(k+1) = 2D3−s η−1 Xk=0 D3k Dsk = 2D3−s 1 − D(3−s)η 1 − D3−s . 5 Wavelets and eigenfunctions for OD In this section, we connect the eigenspaces Eγ of Theorem 4.10 with the orthogonal decomposition of L2(∂BD, M) associated to the wavelets constructed in [19] Section 3 (see 21 also Section 4 of [7]). We begin by describing the wavelet decomposition of L2(∂BD, M), which is a special case of the wavelets of [19] and [7]. To be precise, the wavelets we discuss here are those associated to the D × D matrix AD consisting of all 1's, but the wavelets described in [19] are defined for any matrix A with entries from {0, 1}. Let ΛD denote the directed graph with vertex matrix AD. In what follows, we will assume that we have labeled the D vertices of Λ0 D by ZD = {0, 1, . . . , D − 1}, and we will write infinite paths in Λ∞ D = ∂BD as strings of vertices (i1i2i3 . . .) where ij ∈ ZD for all j. Denote by V0 the (finite-dimensional) subspace of L2(∂BD, M) given by Define an inner product on CD by V0 = span{χσi(∂BD) : i ∈ ZD}. (cid:10)(xj), (yj)(cid:11)P F = 1 D D−1 Xj=0 xjyj. (18) We now define a set of D linearly independent vectors {cj : 0 ≤ j ≤ D − 1} ⊂ CD, where cj = (cj 0, . . . , cj D−1), by c0 ℓ = 1 ∀ ℓ ∈ ZD, and {cj : 1 ≤ j ≤ D − 1} an orthonormal basis for the subspace {(1, 1,· · · , 1)}⊥, with ⊥ taken with respect to the inner product h·,·iP F . We now note that we can write each set R[k] = σk(∂BD) as a disjoint union: R[k] = R[kj], D−1 Gj=0 where Thus in terms of characteristic functions, R[kj] = (cid:8)(i1i2 · · · in · · · ) ∈ ∂BD : i1 = k and i2 = j(cid:9). D−1 Now, define functions {f j,k}D−1 χR[k] = Xj=0 j,k=0 on ∂BD by f j,k(x) = √D χR[kj] for k ∈ ZD. cj ℓχR[kℓ](x). D−1 Xℓ=0 Moreover, since c0 ℓ = 1 for all ℓ, we have f 0,k = √D D−1 Xℓ=0 ℓ χR[kℓ] = √D c0 D−1 Xℓ=0 χR[kℓ] = √DχR[k]. 22 It follows that span(cid:8)f 0,k(cid:9)D−1 k=0 = span(cid:8)χR[k](cid:9)D−1 k=0 = V0. Now, we can use the functions f j,k to construct a wavelet basis of L2(∂BD, M). First, a definition: for any word w = w1w2 · · · wn ∈ (ZD)n, write Sw = Sw1Sw2 · · · Swn, where Swi ∈ L2(∂BD, M) is the operator defined in Proposition 2.13. Theorem 5.1 ([19, Theorem 3.2]; [7, Theorem 4.2]). Fix an integer D > 1. Let {Sk}k∈ZD be the operators on L2(∂BD, M) described in Proposition 2.13. Let {f j,k : j, k ∈ ZD} be the functions on ∂BD defined in the above paragraphs. Define W0 = span{f j,k : j, k ∈ ZD, j 6= 0}; Wn = span{Sw(f j,k) : j, k ∈ ZD, j 6= 0, and w ∈ (ZD)n}. Then the subspaces V0 and {Wn}∞ n=0 are mutually pairwise orthogonal in L2(∂BD, M) and To calculate the functions Sw(f j,k), we first observe that L2(∂BD, M) = span V0 ⊕h SiχR[k] = √DχR[ik]; ∞ Mn=0 Wni! . consequently, if w = w1w2 · · · wn, Sw(f j,k) = D(n+1)/2 D−1 Xℓ=0 cj ℓχ[w1w2···wnkℓ]. (19) If we instead write the finite path w as γ, and observe that the edges in En+1 with range k are in bijection with the pairs (kℓ)ℓ∈ZD , we see that for any path γ ∈ FBD with γ = n−1, (20) cj eχ[γke]. Sγ(f j,k) = D(n+1)/2 Xe∈En+1 A few more calculations lead us to the following Theorem 5.2. Let ΛD be the directed graph whose D × D adjacency matrix consists of all 1's. For each γ ∈ Λ, let Eγ be the eigenspace of the Laplace -- Beltrami operator ∆s described in Theorem 4.10. Then for all n ≥ 0 we can write In particular, Wn = Mγ∈Λn Eγ. L2(Λ∞, µ) = V−1 ⊕ W−1 ⊕"Mn≥0 Mγ∈Λn Eγ# = V0 ⊕"Mn≥0 Mγ∈Λn Eγ#. 23 Moreover, for all i ∈ ZD and all γ ∈ FBD,, the isometry Si given by if v1 = i, else. Sif ((v1v2 . . .)) =(D1/2f ((v2v3 . . .)) 0 maps Eγ to Eiγ unitarily. Proof. Let γ ∈ FBD be a path of length n. Recall that the subspaces Eγ are spanned by functions of the form χ[γe] − χ[γe′], where e 6= e′ are edges in En+1. In other words, if we write a spanning function ξe,e′ = χ[γe]−χ[γe′] of Eγ as a linear combination of characteristic functions of cylinder sets, we have ξe,e′ = Xf ∈En+1 df χ[γf ] where de = 1, de′ = −1, df = 0 ∀ f 6= e, e′. In other words, the vector (df )r(f )=s(γ),f ∈En+1 is in the subspace (1, 1, . . . , 1)⊥ of CD which is orthogonal to (1, 1, . . . , 1) in the inner product (18). It follows that Eγ ⊆ Wn whenever γ = n. there are Dn+1 paths γ of length n, and Eγ ⊥ Eη for all γ, η with γ = η. Therefore, Now, Theorem 4.3 of [15] tells us that each space Eγ has dimension D − 1. Moreover, dim  [γ=n Eγ  = Dn+1(D − 1). Similarly, dimWn = Card(F n−1BD)Card({f j,k}j6=0) = Dn · D(D − 1). This equality of dimensions thus implies that Wn = [γ=n Eγ ∀ n ∈ N0. For the last assertion, we simply observe that Si is an isometry with SiS∗ i = idEi. 5.1 Wavelets on SA Let A be an N × N {0, 1}-matrix with precisely D nonzero entries. In this section we will describe wavelets on SA associated to the Cuntz algebra OD using the measure-preserving isomorphism between (SA, H) and (∂BD, M) described in Theorem 3.1. Since all edges in ΛD can be preceded (or followed) by any other edge, this infinite path space corresponds simply to [0, 1] by thinking of points in [0, 1] as infinite sequences in {0, . . . , D − 1}N and using the D-adic expansion. The natural correspondence between SA and points from [0, 1] in their D-adic expan- sions is given by labeling the nonzero entries in A by the elements of {0, 1, . . . , D − 1}, 24 and then identifying a cylinder set [(x1, y1), (x2, y2), . . . , (xn, yn)] in SA with the cylinder [d1 . . . dn], where di ∈ {0, . . . , D − 1} is the integer corresponding to Axi,yi. Thus, we obtain wavelets on SA by using this identification to transfer the wavelets associated to the infinite path representation of OD into functions on SA. These wavelets will agree with the eigenfunctions Eγ of the Laplace -- Beltrami operator associated to the Bratteli diagram for OD, by Theorem 5.2 above. Theorem 4.10 as a wavelet decomposition of L2(∂BD, M), with χ[γe′](cid:27). To be more precise, Theorem 5.2 implies that we can interpret the eigenfunctions of Eγ = span(cid:26) 1 χ[γe] − M[γe′] M[γe] 1 Here γ is a finite path in the graph ΛD associated to OD; writing γ as a string of vertices, equivalently, γ = d0d1d2 · · · dn for di ∈ {0, . . . , D − 1}. Thus, if di ∈ ZD corresponds to the pair (xi, yi) ∈ SA, and e, e′ ∈ {0, . . . , D − 1} correspond to the pairs (z, w), (z′, w′) in the symbol set SA, the wavelet on L2(SA, H) associated to M [γe′] χ[γe′] is 1 M [γe]χ[γe] − 1 1 H([(x1, y1), (x2, y2), . . . , (xn, yn), (z, w)]) χ[(x1,y1),(x2,y2),...,(xn,yn),(z,w)] H([(x1, y1), (x2, y2), . . . , (xn, yn), (z′, w′)]) 1 − χ[(x1,y1),(x2,y2),...,(xn,yn),(z ′,w′)] 1 = Dn+2 (cid:0)χ[(x1,y1),(x2,y2),...,(xn,yn),(z,w)] − χ[(x1,y1),(x2,y2),...,(xn,yn),(z ′,w′)](cid:1) . This correspondence allows us to transfer the spaces Eγ from L2(∂BD, M) to L2(SA, H), giving us an orthogonal decomposition of the latter. Moreover, the "scaling and trans- lation" operators Si of Theorem 5.2 from the infinite path representation of OD transfer (via the same correspondence between pairs (x, y) with A(x, y) 6= 0 and elements of {0, . . . , D − 1}) to the operators Ti on L2(SA, H) introduced in Theorem 3.1. In other words, these operators Ti allow us to move between the orthogonal subspaces of L2(SA, H), enabling us to view this as a wavelet decomposition. 6 Spectral triples and Laplacians for the Cuntz alge- bra OD: the uneven weight case 6.1 The spectral triple D) is different from the Perron -- Frobenius weights wλ We are going to work in the general framework of Section 4 with the difference that the weight (which we call wr D we previously defined in Definition 4.5. For this section, we require that our weight is defined on finite paths as in Definition 4.1, rather than on vertices as in Definition 4.5. In particular, the weight wr D(γ) will not depend only on the length and the source of γ, but also on the precise sequence of edges making up γ. D will not be self-similar in the sense that wr 25 Definition 6.1. Fix a vector r = (r1, . . . , rD) of positive numbers satisfying Pi ri = 1. (We also note that this condition is not essential, although it makes a nice normalization.) The weight wr D on the graph ΛD with D vertices v1, . . . , vD (equivalently, the Bratteli diagram BD) associated to the matrix AD is defined as follows. 1. Whenever γ is the trivial (empty) path ◦, we set wr 2. Associate to each vertex vi the weight ri: D(◦) = 1. wr D(v) = rv, ∀v ∈ Λ0. 3. Given a path γ = (e1 . . . en) with ej = 1, s(ei) = vji, and r(e1) = vj0, we set the weight of γ to be wr D(γ) = rji. n Yi=0 4. The diameter diam[η] of a cylinder set [η] is defined to be equal to its weight, diam[η] = wr D(η). choice of our normalization. Note in particular that [◦] = Λ∞ D and so diam[◦] = 1, which is consistent with the The set of finite paths on a graph has a natural tree structure. In fact, if we denote by (e1 . . . en) a string of composable edges (thus requiring s(ei−1) = r(ei), ∀i) then the "parent" of (e1 . . . en) is (e1 . . . en−1); the root is the path ◦ of length −1 which corresponds to Λ∞ D(γ) decreases to 0 as the length of γ, γ, increases to infinity. Therefore, the Pearson -- Bellissard construction from [20] applies, and there is a spectral triple associated to the set of infinite paths as in Theorem 4.6 (see also [15, 9]). To be more precise, we have: D . In addition, the weight wr Proposition 6.2. Let BD be the Bratteli diagram associated to the matrix AD. Let (BD, wr w) be the as- sociated ultrametric Cantor set. Then there is an even spectral triple (CLip(∂BD),H′, π′ D) be the weighted Bratteli diagram given in Definition 6.1. Let (∂BD, dr τ , /D′, Γ′). The ζ-function associated to the spectral triple of Theorem 4.6 is given by ζ r D(s) = 1 2 Tr( /D′−s) = Xλ∈F (BD)◦(cid:0)wr D(λ)(cid:1)s. We now want to compute the abscissa of convergence sr of the above ζ-function. Proposition 6.3. The abscissa of convergence sr of the ζ-function ζ r the spectral triple in Proposition 6.2 is 1. D(s) associated to 26 Proof. The formula for the ζ-function can be written as follows: ζ r D(s) = 1 2 Tr( /D′−s) = ∞ Xn=−1 Xλ∈Λn(cid:16)wr D(λ)(cid:17)s , (21) with the convention that a path of length −1 is the empty path ◦ with associated cylinder set Λ∞. In order to enumerate how many paths of which weight there are in F (B)◦, we will use the following argument. Consider the following formal polynomial in D variables X1, . . . , XD with integer coefficients: d . Xi=1 P (X1, . . . , XD) =(cid:16) Xi(cid:17)n+1 After expanding, each monomial is of the form cQi X αi i where c is a constant. The constant c counts how many partitions of {0, . . . , n} into D (possibly empty) subsets there are, of cardinality respectively α1, . . . , αD. The set of such partitions for all possible choices of α1, . . . , αD is in bijection with F nBD: given γ = (e1 . . . en) (with ei = 1 and s(ei−1) = r(ei), ∀i), let Ui = {j ∈ {0, . . . , n} : s(ej) = vi}. One sees that {Ui}D i=1 defines a partition of {0, . . . , n}, and the map from F nBD, the set of finite paths of BD of length n, to the set of such partitions is a bijection. Indeed, Now, we see that the sum in Equation (21) can be rewritten as wr D(γ) =Yi rαi i . ∞ ∞ D ζ r D(s) = Xn=−1 P (r1, . . . , rD)s(n+1) = i(cid:17)n+1 Xi=1 This is a geometric series, which converges if and only if Pi rs i < 1. The function s 7→ Pi rs is a decreasing function on R+ (since all the ri are less than 1), and Pi ri = 1. Remark 6.4. Note that this guarantees that the upper Minkowski dimension of (∂BD, dwr) is 1, see [20, Theorem 2]. Therefore, the abscissa of convergence is exactly sr = 1. Xn=−1(cid:16) rs . i D(γ). Theorem 6.5. The measure µr D([γ]) = wr µr Proof. Note first that for the case γ = ◦, the result follows immediately from the defini- tions of µr D on ∂BD induced by the Dixmier trace is defined by D and wr D. Given a cylinder set [γ] 6= Λ∞, we have D(γη)(cid:17)s r Pη : r(η)=s(γ)(cid:16)wr D([γ]) = lim s→s+ ζ r D(s) µr . 27 One remark is in order: if γ is a path of length n and 0 < m < n, then n wr D(γ) = wr D(cid:0)s(ei)(cid:1) D(cid:0)s(ei)(cid:1)(cid:19)(cid:18)wr D(cid:0)s(ei)(cid:1)(cid:19)(cid:18)wr wr m wr wr Yi=1 D(cid:0)r(e1)(cid:1) =(cid:18)wr Yi=1 D(cid:0)r(e1)(cid:1) =(cid:18)wr Yi=1 D(cid:0)r(e1)(cid:1) D(e1e2 · · · em)wr D(γ)wr m = wr D(cid:0)s(em+1)(cid:1) D(cid:0)r(em+2)(cid:1) n n Yi=m+2 Yi=m+2 w(cid:0)s(ei)(cid:1)(cid:19) D(cid:0)s(ei)(cid:1)(cid:19) wr D(γη) is not wr D(em+2 · · · en). In particular, wr D(η). Indeed, any path of the form γη with s(γ) = r(η) can be written uniquely as γeη′ where e is the unique edge with r(e) = s(γ) and s(e) = r(η′). By the computation above, wr D(η′). Moreover, since ΛD has precisely one edge connecting any pair of vertices, every finite path η′ in Λ gives rise to exactly one e such that s(e) = r(η′) and r(e) = s(γ). Therefore, D(γeη′) = wr D(γ)wr Xη : r(η)=s(γ)(cid:0)wr where α(s) =Pη′∈Λ(cid:16)wr D(γη)(cid:1)s D(η′)(cid:17)s µr D([γ]) = lim α(s), D(η′)(cid:17)s =(cid:0)wr D(γ)(cid:1)s D(γ)(cid:1)s(cid:16)wr = Xη′∈Λ(cid:0)wr . Moreover, since lims→1+ α(s) = +∞, we have s→1+(cid:16)wr D(γ)(cid:17)s 1 + α(s) D(γ). = wr α(s) In particular, we do not have µr D = µD = M. This should not be completely surprising, however. The Perron -- Frobenius measure M = µD is the unique measure on Λ∞ under the following assumptions: the measure is a probability measure, and µD[γ] only depends on γ and s(γ). The second assumption is not satisfied for µr D. D does not define a self-similar ultrametric Cantor set in the sense of [16, Definition 2.6], since again, the diameter of [γ] does not just depend on γ and s(γ) but also on the specific sequence of edges. Note also that the choice of weight wr 6.2 The Laplace -- Beltrami Operator As in Section 4.2, the Dixmier trace associated to the spectral triple of Proposition 6.2 induces the probability measure µr D(τ ) on the set of choice functions; thus, by the classical theory of Dirichlet forms we can define a Laplace -- Beltrami operator ∆r s on L2(∂BD, µr D) as in Proposition 4.1 of [15] by hf, ∆r s(g)i = Qs(f, g) = 1 2ZE Tr( /D −s[ /D, πτ (f )]∗ [ /D, πτ (g)] dµr D(τ ), (22) 28 As before, ∆r where Dom Qs = span{χγ : γ ∈ FB2} is a closable Dirichlet form. spectrum of ∆r the eigenfunctions of ∆r s is self-adjoint and has pure point spectrum, and we can describe the s explicitly. For our case we can additionally compute the eigenvalues and s as follows. Theorem 6.6. [15, Theorem 4.3] Let ∆r given by (22). Then the eigenvalues {λr given by, for η ∈ F ∂BD, Xk=−1 λr η = η−1 1 Gs(η[0, k])(cid:16)µr D[ηe] − µr χ[ηe′] µr D[ηe′] Er η = span(cid:26) χ[ηe] s be the Laplace -- Beltrami operator on L2(∂B2, µr D) η} and corresponding eigenspaces {Er s are η} of ∆r D[η[0, k + 1]] − µr D[η[0, k])](cid:17) − µr D[η] Gs(η) , : e 6= e′, e = e′ = 1, r(e) = r(e′)(cid:27), where η[0,−1] = ◦ and χ[◦] = ∂BD, Gs(η[0,−1]) = 1 FBD, D(ξ)2−s Xe6=e′∈r−1(s(ξ)) Gs(ξ) = wr 1 2 2Pv6=w∈Λ0 µr D[ξe] µr µr D[ξe′]. D[v] µr D[w], and for ξ ∈ In addition, 0 is an eigenvalue for the constant function 1, and λ◦ = (Gs(◦))−1 is an eigenvalue with eigenspace Er ◦ = span(cid:26) χ[v] D[v] − µr χ[v′] µr D[v′] : v 6= v′, v, v′ ∈ V0(cid:27). Proof. Although Theorem 4.3 of [15] is stated only for the case when the weight function w(γ) only depends on the length and the source of the path γ, as in Definition 4.5, a careful examination of the proof of that Theorem will show that the same proof works verbatim in the case of the weight wr D. 6.3 Eigenvalues and eigenfunctions for the O2 case We are going to explicitly compute here the eigenvalues for the Laplace -- Beltrami operator ∆r s in the D = 2 case. Theorem 6.6 specializes in the case of O2 to give The formulas in Proposition 6.8 below allow us to compute in principle the eigenvalue associated to any finite path. However it seems difficult to get an explicit formula that covers all the cases as the calculations in full generality are difficult to manage because of challenging bookkeeping. Lemma 6.7. With notation as above, for a finite path ξ(p, q) ∈ F (∂B2) having p vertices equal to v1 and q vertices equal to v2 we have we have More generally, if ξ is any path, one can write Gs(ξ(p, q)) = r4p+1−ps(1 − r)4q+1−qs Gs(ξ) =(cid:0)µr 2[ξ](cid:1)4−sr(1 − r). 29 Proof. We start with the second point. If ξ is coded by its vertices, ξ = (v0, . . . , vξ) and e 6= e′ are vertices such that r(e) = r(e′) = s(ξ), then wr D(ξ))2r(1 − r). Since µr D(ξe′) = (wr D(ξe)wr D(ξ), we have D[ξ] = wr 1 Gs(ξ) = 2(cid:0)µr D[ξ](cid:1)2−s2(cid:0)µr D[ξ](cid:1)2r(1 − r) and the result follows. (Note that the factor 2 appears because (v1, v2) and (v2, v1) are the two pairs in the index of the sum defining Gs(ξ).) For the first point, we compute Gs(ξ(p, q)) = (1/2)(cid:2)rp(1 − r)q(cid:3)2−s2(cid:0)(rp(1 − r)qr)(rp(1 − r)q(1 − r)(cid:1). Proposition 6.8. Let ∆r for the choice of weight induced by s be the Laplace -- Beltrami operator on L2(∂B2, µr 2) given by (22) wr 2(v1) = r, wr 2(v2) = (1 − r), where r ∈ [0, 1] is fixed. (Note that the notation used above is slightly different from the notation we used in Theorem 6.6). Let η ∈ F ∂B2 of length n be determined by the string of vertices (v0, . . . , vn); also we write (v0, . . . , vk) for η[0, k], for any k ≤ n. Then we have λr η = wr 2(v0) − 1 r(1 − r) + n−1 Xk=0 (µr 2[v0, . . . , vk])s−3 r(1 − r) (wr 2(vk+1) − 1) − 2[η])s−3 (µr r(1 − r) . Proof. We will use the fact that if one codes η by its vertices η = (v0, . . . , vη), then µr 2[v0, . . . , vk] = µr 2[vi+1, . . . , vk], as was established in the proof of Theo- rem 6.5. Consequently, we can factor the term (µr 2[v0, . . . , vi]µr 2[η[0, k]]) as follows: 2[η[0, k + 1]] − µr µr 2[η[0, k + 1]] − µr 2[η[0, k]] = µr = µr 2[v0, . . . , vk+1] − µr 2[v0, . . . , vk+1](µr 2[v0, . . . , vk] 2[vk] − 1). We therefore compute λr η = η−1 Xk=−1 1 Gs(η[0, k])(cid:16)µr 2[η[0, k + 1]] − µr 2[η[0, k]](cid:17) − µr 2[η] Gs(η) , that is (using point 2 of Lemma 6.7) λr η = + − n−1 µr 2[v0] − 1 Gs(◦) Xk=0 2[η] r(1 − r)(µr µr r(1 − r)(µr 2[η])4−s . 1 2[v0, . . . , vk])4−s µr 2[v0, . . . , vk](wr 2(vk+1) − 1) The result follows from algebraic simplifications. Note in particular that Gs(◦) = r(1 − r). 30 One can also construct representations and wavelet spaces of O2 associated to the weighted Bratteli diagram (∂B2, wr 2); see Theorem 3.8 of [8]. This is the analogue of Theorem 5.1 above for the uneven weight case. We now compute the eigenspaces corre- sponding to the eigenvalues of Proposition 6.8 above, and show that they coincide with the wavelet spaces described in [8] Theorem 3.8. In other words, we will show that if Wk are the orthogonal subspaces of L2(∂B2, wr 2) described in Theorem 3.8 of [8], Wk = Mη:η=k Er η. What is done below is similar to the result of Theorem 5.2, but we allow unequal weights in what follows. By evaluating the formulas given in Theorem 6.6 we obtain: Proposition 6.9. Let ∆r 2) given by (22) for the choice of weight induced by the choice on the vertices v1 and v2 of the associated graph as s be the Laplace -- Beltrami operator on L2(∂B2, µr wr 2(v1) = r1, wr 2(v2) = r2 = 1 − r1, where r = r1 ∈ [0, 1] is fixed. If we let ◦ denote the empty path, then the eigenspace Er with eigenvalue λr 2 , hence has dimension 1 and ◦ is given by Λ∞ ◦ Er ◦ = spann χ[v1] 2([v1]) − µr χ[v2] 2([v2])o = span(cid:26) χ[v1] r − µr χ[v2] 1 − r(cid:27) . Given a finite non-empty path η = vj0vj1 . . . vjn ∈ F ∂B2 with n + 1 vertices, where ji ∈ {1, 2} ∀ i, the eigenspace Er η described in Proposition 6.8 is given by η with corresponding eigenvalue λr Er η = spann χ[ηe] 2[ηe] − µr = span(cid:26) (Qn 1 i=0 rji)r χ[ηe′] µr 2[ηe′] χ[vj0 vj1 ...vjn v1] − : e 6= e′, e = e′ = 1, r(e) = r(e′) = s(η)o χ[vj0 vj1 ...vjn v2](cid:27) . 1 (Qn i=0 rji)(1 − r) We now show how the scaling functions generating V0 in Theorem 3.8 of [8] fit into the eigenspace picture described above. Lemma 6.10. Let r ∈ [0, 1] be given, and let µr 2 be the Markov probability measure on the infinite path space Λ∞ 2 corresponding to the weight assigning r to the vertex v1 and 1 − r to the vertex v2. Let V−1 denote the space of constant functions on Λ∞ 2 . Then the scaling space V0 described in Theorem 3.8 of [8] as the span of {χ[v1], χ[v2]}, the characteristic functions of cylinder sets corresponding to the vertices, can be written as where Er ◦ is the eigenspace corresponding to the empty path. V0 = V−1 ⊕ Er ◦, 31 Proof. We note that V0, being generated by the orthogonal functions χ[v1] and χ[v2], has dimension 2. On the other hand, the space V−1 of constant functions on Λ∞ 2 has dimension 1 and Er ◦ = spann χ[v1] 2([v1]) − µr χ[v2] 2([v2])o µr also has dimension 1 and is orthogonal to V−1. It follows by a dimension count that as desired. V0 = V−1 ⊕ Er ◦, Proposition 6.11. Let µr 2 be the Markov probability measure on the infinite path space Λ∞ 2 corresponding to the weight assigning r to the vertex v1 and 1 − r to the vertex v2. Then for the corresponding representation of O2 on L2(Λ∞ 2) defined in Theorem 3.8 of [8], we have 2 , µr W0 = spanη:η=0{Er η}, where Er η are the eigenspaces of the Laplace-Beltrami operator defined in Proposition 6.9. Proof. As in Theorem 3.8 of [8] and Section 5 above, we have an inner product on C2 defined by 2 h(xj), (yj)i = xj · yj · rj, Xj=1 and a fixed vector c0 = c0,k = (1, 1). For k = 1, 2, we find an orthonormal basis for {c0,k}⊥ denoted by {c1,k}, where c1,k = (c1,k But here, a straightfoward calculation shows that we can take ℓ )ℓ∈{1,2}. c1,k =pr(1 − r)(cid:16)1 r ,− 1 1 − r(cid:17), k = 1, 2. Therefore the wavelet ψ1,k of Theorem 3.8 of [8] is given by χ[vkv2] ψ1,k = pr(1 − r) √rk (cid:20)χ[vkv1] r − =pr(1 − r)rk(cid:20) χ[vkv1] 1 − r(cid:21) 2([vkv1]) − µr µr χ[vkv2] 2([vkv2])(cid:21). χ[vkv2](cid:27) 1 Recall that Er vk = span(cid:26) 1 µr 2([vkv1]) is a one-dimensional subspace of L2(∂B2, µr scalar multiple of the single spanning vector from Er the span of the two vectors from Er v1 and Er since W0 is defined to be the span of the vectors ψ1,k, the result follows. χ[vkv1] − 2). Moreover, each vector ψ1,k is evidently a η for η = vk a path of length 0. Taking v2 gives exactly the span of the ψ1,k for k = 1, 2; 2([vkv2]) µr 32 We now relate higher dimensional wavelet subspaces to the corresponding eigenspaces for the Laplacian: 2 be the Markov probability measure on the infinite path space Λ∞ Lemma 6.12. Let µr 2 corresponding to the weight assigning r to the vertex v1 and 1 − r to the vertex v2. Then for the corresponding representation of O2 on L2(Λ∞ 2) defined in Theorem 3.8 of [8], we have Wk = spanη:η=k{Er η}, 2 , µr where Er η are the eigenspaces of the Laplacian defined in Proposition 6.9. Proof. We prove the result by induction. We have proved the result for k = 0 directly. We now suppose that for k = n we have shown where, as defined in Theorem 3.8 of [8], Wn = spanη:η=n(cid:8)Er η(cid:9), Wn = span(cid:8)Sw(ψ1,k) : k = 1, 2, w is a word of length n(cid:9), for ψ1,1 and ψ1,2 the wavelets of Lemma 6.11, and for w = w1w2 · · · wn a word of length n, where wi ∈ Z2, Sw = Sw1Sw2 · · · Swn, where (writing an infinite path x as a sequence of vertices) and 0 S0f (x) =(r−1/2f (u2u3 . . .) S1f (x) =((1 − r)−1/2f (u2u3 . . .) 0 if x = (v1u2u3 . . .), else. if x = (v2u2u3 . . .), else. From this and the induction hypothesis, it follows that where a typical element of Er Wn+1 = span(cid:8)S0( Wn), S1( Wn)(cid:9) = spanη:η=n(cid:8)S0(Er η), S1(Er η looks like η)(cid:9), χ[ηe] 2([ηe]) − µr χηe′ µr 2([ηe′]) . Now if η = u0u1 · · · un is a path of length n whose n + 1 vertices are given in order by u0u1u2 · · · un, we compute directly that and S0χ[η] = 1 √r χ[v1η], S1χ[η′] = χ[v2η′]. 1 √1 − r 33 Therefore we can write, for η of length n and e and e′ of length 1 with e 6= e′, S0(cid:18) χ[ηe] 2([ηe]) − µr χ[ηe′] 2([ηe′])(cid:19) = µr 1 √r(cid:20) χ[v1ηe] 2([ηe]) − µr = √r(cid:20) χ[v1ηe] 2([v1ηe]) − µr χ[v1ηe′] µr 2([ηe′])(cid:21) 2([v1ηe′])(cid:21) χ[v1ηe′] µr which is a constant multiple of χ[v1ηe] 2([v1ηe]) − µr χ[v1ηe′] µr 2([v1ηe′]) which is a spanning function for the one-dimensional subspace Er χ[ηe′] 2([ηe′])) is a constant multiple of µr v1η. Similarly, S1( χ[ηe] µr 2([ηe])− χ[v2ηe] 2([v2ηe′]) − µr χ[v2ηe′] µr 2([v2η′e′]) , which spans Er length n and some vertex vi, with i = 1, 2, it then follows that v2η. Since all paths of length n + 1 are of the form viη for some path η of spanη:η=n{S0(Er η), S1(Er η)} = spanη′:η′=n+1(Er η′). But this shows that Wn+1 = spanη′:η′=n+1(Er η′), and the induction step of the proof is complete. The above results have established the following: 2 be the Markov probability measure on the infinite path space Λ∞ Theorem 6.13. Let µr 2 corresponding to the weight assigning r to the vertex v1 and 1 − r to the vertex v2. Then for the corresponding representation of O2 on L2(Λ∞ 2) defined in Theorem 3.8 of [8], we have that the kth-order wavelets defined there are all constant multiples of functions of the form 2 , µr χ[ηe] 2([ηe]) − µr χ[ηe′] µr 2([ηe′]) , η = k, e = e′ = 1, r(e) = r(e′) = s(η). As in the case of Lemma 6.11, the constant coefficient needed to transform the wavelet function Sηψ1,k into the spanning function of Eηvk can be be computed to be where j is the number of v1's appearing as vertices in the path ηvk. pr(1 − r)[√r]j[√1 − r]k+1−j, 34 References [1] S. Bezuglyi and P. E. T. Jorgensen, Representations of Cuntz -- Krieger algebras, dy- namics on Bratteli diagrams, and path-space measures, in "Trends in Harmonic Anal- ysis and Its Applications", 57 -- 88, Contemp. Math. 650, Amer. Math. Soc., Provi- dence, RI, 2015. [2] A. L. Carey, J. Phillips, and A. Rennie, Semifinite spectral triples associated with graph C ∗-algebras, Traces in number theory, geometry and quantum fields, 35-36 , Aspects Math., E38, Friedr. Vieweg, Wiesbaden, 2008. [3] E. Christensen and C. Ivan, Spectral triples for AF C ∗-algebras and metrics on the Cantor set, J. Operator Theory 56 (2006), 17 -- 46. [4] E. Christensen, C. Ivan, M. L. Lapidus, and E. Schrohe, Spectral triples and the geometry of fractals, J. Noncommut. Geom. 6 (2012), 249 -- 274. [5] A. Connes, Compact metric spaces, Fredholm modules, and hyperfiniteness, Ergodic Theory Dynamical Systems 9 (1989), 207 -- 220. [6] D.E. Dutkay and P.E.T. Jorgensen, Monic representations of the Cuntz algebra and Markov measure, J. Funct. Anal. 267 (2014), 1011 -- 1034. [7] C. Farsi, E. Gillaspy, S. Kang and J. Packer, Separable representations, KMS states, and wavelets for higher-rank graphs, J. Math. Anal. Appl. 434 (2016) 241 -- 270. [8] C. Farsi, E. Gillaspy, S. Kang and J. Packer, Wavelets and graph C ∗-algebras, arXiv:1601.0006. [9] C. Farsi, E. Gillaspy, S. Kang, A. Julien, and J. Packer, Spectral triples on k-graphs, in preparation. [10] M. Goffeng and B. Mesland, Spectral triples and finite summability on Cuntz-Krieger algebras, Doc. Math. 20 (2015), 89170. [11] A. an Huef, M. Laca, I. Raeburn and A. Sims, KMS states on the C ∗-algebra of a higher-rank graph and periodicity in the path space, J. Funct. Anal. 268 (2015) 1840 -- 1875. [12] J. E. Hutchinson, Fractals and self-similarity, Indiana Univ. Math. J. 30 (1981), 713 -- 747. [13] A. Jonsson, Wavelets on fractals and Besov spaces, J. Fourier Anal. Appl. 4(3) (1998), 329 -- 340. [14] A. Julien and I. Putnam, Spectral triples for subshifts, J. Funct. Anal. 270 (2016), 1031 -- 1063. 35 [15] A. Julien and J. Savinien, Transverse Laplacians for substitution tilings, Comm. Math. Phys. 301 (2011), 285 -- 318. [16] A. Julien and J. Savinien, Embeddings of self-similar ultrametric Cantor sets, Topol- ogy Appl. 158 (2011), 2148 -- 2157. [17] K. Kawamura, The Perron-Frobenius operators, invariant measures and representa- tions of the Cuntz -- Krieger algebras, J. Math. Phys. 46 (2005), 083514, 6 pp. [18] A. Kumjian, D. Pask, I. Raeburn, and J. Renault, Graphs, groupoids, and Cuntz -- Krieger algebras, J. Funct. Anal. 144 (1997), 505 -- 541. [19] M. Marcolli and A.M. Paolucci, Cuntz -- Krieger algebras and wavelets on fractals, Complex Anal. Oper. Theory 5 (2011), 41 -- 81. [20] J. C. Pearson, J. V. Bellissard, Noncommutative Riemannian geometry and diffusion on ultrametric Cantor sets, J. Noncommut. Geom. 3 (2009), 447 -- 481. [21] I. Raeburn Graph Algebras, CBMS Reg. Conf. Series in Math., Vol, 103, American Mathematical Society, Providence, RI, 2005, vii+113 pp. [22] G. Robertson and T. Steger, Affine buildings, tiling systems and higher rank Cuntz- Krieger algebras, Journal fur die reine und angewandte Mathematik 513 (1999), 115 -- 144. [23] R. Strichartz, Piecewise linear wavelets on Sierpinski gasket type fractals, J. Fourier Analysis and Applications 3 (1997), 387 -- 416. [24] M.F. Whittaker, Spectral triples for hyperbolic dynamical systems, J. Noncommut. Geom. 7 (2013), no. 2, 563 -- 582. Carla Farsi, Elizabeth Gillaspy, Sooran Kang, Judith Packer : Department of Mathematics, University of Colorado at Boulder, Boulder, Colorado, 80309- 0395 E-mail address: [email protected], [email protected], [email protected], [email protected] Antoine Julien : Department of Mathematical Sciences, Norwegian Univer- sity of Science and Technology 7491 Trondheim, Norway. E-mail address, [email protected] 36
1605.05766	1	1605	2016-05-18T21:55:07	Traces arising from regular inclusions	[ "math.OA" ]	We study the problem of extending a state on an abelian $C^$- subalgebra to a tracial state on the ambient $C^$-algebra. We propose an approach that is well-suited to the case of regular inclusions, in which there is a large supply of normalizers of the subalgebra. Conditional expectations onto the subalgebra give natural extensions of a state to the ambient $C^$-algebra; we prove that these extensions are tracial states if and only if certain invariance properties of both the state and conditional expectations are satisfied. In the example of a groupoid $C^$-algebra, these invariance properties correspond to invariance of associated measures on the unit space under the action of bisections. Using our framework, we are able to completely describe the tracial state space of a Cuntz-Krieger graph algebra. Along the way we introduce certain operations called graph tightenings, which both streamline our description and provides connections to related finiteness questions in graph $C^*$-algebras. Our investigation has close connections with the so-called unique state extension property and its variants.	math.OA	math	TRACES ARISING FROM REGULAR INCLUSIONS DANNY CRYTSER, GABRIEL NAGY Abstract. We study the problem of extending a state on an abelian C ∗- subalgebra to a tracial state on the ambient C ∗-algebra. We propose an ap- proach that is well-suited to the case of regular inclusions, in which there is a large supply of normalizers of the subalgebra. Conditional expectations onto the subalgebra give natural extensions of a state to the ambient C ∗-algebra; we prove that these extensions are tracial states if and only if certain invari- ance properties of both the state and conditional expectations are satisfied. In the example of a groupoid C ∗-algebra, these invariance properties corre- spond to invariance of associated measures on the unit space under the action of bisections. Using our framework, we are able to completely describe the tracial state space of a Cuntz-Krieger graph algebra. Along the way we intro- duce certain operations called graph tightenings, which both streamline our description and provides connections to related finiteness questions in graph C ∗-algebras. Our investigation has close connections with the so-called unique state extension property and its variants. Introduction A trace on an complex algebra A is a linear functional φ : A → C satisfying If A is a C∗-algebra, and the trace φ is also a φ(xy) = φ(yx) for all x, y ∈ A. state, it is simply called a tracial state. In this paper we study tracial states on C∗-algebras A by reconstructing them from their restrictions to abelian subalgebras B ⊂ A. The material is organized as follows In Section 1 our approach focuses on the case when a conditional expectation E : A → B exists and the "candidate" tracial state on A is φ ◦ E, where φ ∈ S(B). In other words, we focus on states on A that factor through E; equivalently, states that vanish on ker E. In orther to characterize such states, we identify a certain invariance condition on φ, coupled with a a suitable normalization condition on E (both conditions employ normalizers of B). Section 2 specializes our investigation to the case of ´etale groupoid C∗-algebras, where the natural abelian C∗-algebra to consider is C0(G(0)) -- the C∗-algebra of continuous functions that vanish at ∞ on the unit space G(0). In this framework, the invariance conditions treated in Section 1 become measure theoretical in nature. In Section 3 we explore the link between the invariance and normalization con- ditions from Section 1 and certain state extension properties. When the so-called extension property holds, the tracial state space of A can be completely described by its restrictions to B. The paper concludes with Section 4, where the case of graph C∗-algebras is fully investigated, using the results proved in the previous sections. Given some directed graph E, our main goal is the complete parametrization of the tracial state space of the associated C∗-algebra C∗(E), solely in graph theoretical language. Earlier work in this direction ([17], [12]) identified the notion of graph traces as a major 1 2 DANNY CRYTSER, GABRIEL NAGY ingredient. In many instances, graph traces are not sufficient for exhausting all tracial states, and our analysis shows exactly what additional structure is necessary: cyclical tags on graph traces. The usage of cyclical tags alone, although necessary, is still insufficient for describing all tracial states on C∗(E); however, this deficiency can be fixed using graph operations called tightenings. 1. Invariant states on abelian C-subalgebras Following [8] and [15], given a C∗-algebra inclusion B ⊂ A, an element n ∈ A is said to normalize B if nBn∗ ∪ n∗Bn ⊂ B. The collection of such normalizers is denoted by NA(B), or simply N (B) when there is no danger of confusion. Clearly N (B) is closed under products and adjoints, and contains B. A C∗-inclusion B ⊂ A is said to be regular, if N (B) generates A as a C∗-algebra. (Equivalently, if the span of N (B) is dense in A.) Most of the C∗-algebra inclusions B ⊂ A we are going to deal with in this paper are non-degenerate, in the sense that B contains an approximate unit for A. (Of course, if A is unital, then non-degeneracy of B is equivalent to the fact that B contains the unit of A.) Note that, if B ⊂ A is a non-degenerate C∗-subalgebra, then n∗n, nn∗ ∈ B for any n ∈ N (B). Definition 1.1. Assume B ⊂ A is a non-degenerate and let φ be a state on B ⊂ A. (1) Given n ∈ N (B), we say that φ is n-invariant if (1) ∀ b ∈ B : φ(nbn∗) = φ(n∗nb). (2) Given N0 ⊂ N (B), we say that φ is N0-invariant if φ is n-invariant for all n ∈ Σ. (3) Lastly, if φ is N (B)-invariant, then we simply say that φ is fully invariant. The collection of fully invariant states on B ⊂ A is denoted by Sinv(B). Comment. The restriction τ B of any tracial state τ ∈ T (A) is clearly a fully invariant state on B, so we have an affine w∗-continuous map (2) T (A) ∋ τ 7−→ τ B ∈ Sinv(B). This paper aims at understanding when the map (2) is either surjective, or injective, or both. The most important features of normalizers and invariant states are collected in Proposition 1.3 below. Both in its proof and elsewhere in the paper, we are going to employ the following well known technical results and notations. Fact 1.2. Assume x is an element in some C∗-algebra A. (i) For any function f ∈ C ([0, ∞)), the elements f (xx∗), f (x∗x) ∈ A, given by continuous functional calculus, satisfy the equality (3) (4) xf (x∗x) = f (xx∗)x. (ii) When specializing to the kth root functions f (t) = t1/k, we also have the equalities lim k→∞ (xx∗)1/kx = lim k→∞ x(x∗x)1/k = x. TRACES ARISING FROM REGULAR INCLUSIONS 3 (iii) If we fix a double sequence (f ℓ k)∞ k,ℓ=1 of polynomials in one variable, such that (5) ∀ k ∈ N : lim ℓ→∞ tf ℓ k(t) = t1/k, uniformly on compact K ⊂ [0, ∞) (this is possible by the Stone-Weierstrass Theorem), then: (6) (7) lim k→∞ lim k→∞ lim ℓ→∞ lim ℓ→∞ xf ℓ k(x∗x)x∗x = lim k→∞ f ℓ k(xx∗)xx∗x = lim k→∞ lim ℓ→∞ lim ℓ→∞ xx∗xf ℓ k(x∗x) = x, xx∗f ℓ l (xx∗)x = x. Proposition 1.3. Let B ⊂ A be a non-degenerate abelian C∗-subalgebra of a C∗- algebra A. (i) nB = Bn for all n ∈ N (B). (ii) All states φ ∈ S(B) are B-invariant. (iii) If φ ∈ S(B) is n-invariant for some n ∈ N (B), then φ is also n∗-invariant. (iv) If φ ∈ S(B) is both n1-invariant and n2-invariant, for some n1, n2 ∈ N (B), then φ is also n1n2-invariant. (v) If N0 ⊂ N (B) is a sub-∗-semigroup, generated as a ∗-semigroup by some subset W ⊂ N (B), and φ ∈ S(B) is W -invariant, then φ is N0-invariant. (vi) A state φ ∈ S(B) is fully invariant if and only if ∀ n ∈ N (B) : φ(nn∗) = φ(n∗n). (8) Proof. (i) It suffices to show that for any n ∈ N (B) and any b ∈ B, we have nb ∈ Bn and bn ∈ nB. If we fix n and b, then using the f ℓ k's from Fact 1.2, combined with the commutativity of B, we have (9) nb = lim k→∞ lim ℓ→∞ f ℓ k(nn∗)nn∗nb = lim k→∞ lim ℓ→∞ f ℓ k(nn∗)nbn∗n. k(nn∗)nbn∗ all Since n normalizes B, we know that nbn∗ ∈ B, so the elements bℓ belong to B, and then (9), which now simply states that nb = limk→∞ limℓ→∞ bℓ kn, clearly proves that nb ∈ Bn. The fact that bn ∈ nB is proved exactly the same way. k = f ℓ (ii) This is obvious, since B is abelian. (iii) Take a sequence {bk} ⊂ B such that bn = limk nbk. Then φ(n∗bn) = lim k φ(n∗nbk) = lim k φ(nbkn∗) = φ(bnn∗) = φ(nn∗b). (iv) Suppose that b ∈ B. Take a sequence {ck} ⊂ B such that (n∗ 1n1)n2 = limk n2ck. Then φ(n1n2bn∗ 2) = lim k φ(n2ckbn∗ 2) = lim k φ(n∗ 2n2ckb) = 2n∗ 1) = φ(n∗ = φ(n∗ 1n1n2bn∗ 2n∗ 1n1n2b), so that φ is n1n2-invariant. Part (v) follows immediately from (iii) and (iv). (vi) The "if" implication (for which it suffices to prove (1) only for positive b) follows from the observation, that for any n ∈ N (B) and any b ∈ B+, the element x = nb1/2 is again in N (B), so applying condition (8) to x will clearly imply φ(nbn∗) = φ(b1/2n∗nb1/2) = φ(n∗nb). Conversely, if φ is fully invariant, then ∀ n ∈ N (B) : φ(nn∗) = lim λ φ(nuλn∗) = lim λ φ(n∗nuλ) = φ(n∗n), 4 DANNY CRYTSER, GABRIEL NAGY where (uλ) ⊂ B be an approximate identity for A. (cid:3) Besides the notion of invariance for states on a C∗-subalgebra, we will also use the following two additional variants. Definition 1.4. Given a state ψ ∈ S(A), we say that an element x ∈ A centralizes ψ if ψ(xa) = ψ(ax) for all a ∈ A. It is easy to see that the set Zψ = {x ∈ A : x centralizes ψ} is a C∗-subalgebra of A. (Obviously, ψ is always tracial when restricted to Zψ. In particular, ψ is tracial on A, if and only if its centralizer Zψ contains a set that generates A as a C∗-algebra.) Definition 1.5. If B ⊂ A is a C∗-subalgebra and n ∈ N (B), we will say that a map Φ : A → B is normalized by n if Φ(nan∗) = nΦ(a)n∗ for all a ∈ A. Lemma 1.6. Let B ⊂ A be a non-degenerate abelian C-subalgebra with a condi- tional expectation E : A → B, which is normalized by some n ∈ N (B). For a state φ ∈ S(B), the following are equivalent: (i) φ is n-invariant state on B; (ii) φ ◦ E ∈ S(A) is a state on A, which is centralized by n. Proof. The implication (ii) ⇒ (i) is pretty obvious, and holds even without the assumption that E is normalized by n. Indeed, if b ∈ B, then nbn∗ = E(nbn∗) and bn∗n = E(bn∗n), so if φ ◦ E is centralized by n, then: φ(nbn∗) = (φ ◦ E)(cid:0)n(bn∗)(cid:1) = (φ ◦ E)(cid:0)(bn∗)n(cid:1) = φ(bn∗n) = φ(n∗nb). For the proof of (i) ⇒ (ii), we fix a ∈ A and we show that φ(cid:0)E(an)(cid:1) = φ(cid:0)E(na)(cid:1). k) as in Fact 1.2(iii). Since E is a conditional expectation, it Fix polynomials (f ℓ follows that (10) E(an) = lim k→∞ By the n-invariance of φ, we have k(n∗n)(cid:1) n∗n. Since E is a conditional expectation onto an abelian C-subalgebra, we have: k(nn∗)nn∗ = k(nn∗)nn∗na), so when we return to (12) and we also use (7), we finally get: k(nn∗)nn∗E(na) = E(f ℓ k(nn∗)nn∗) = E(na)f ℓ E(naf ℓ = f ℓ φ(cid:0)E(an)(cid:1) = lim k→∞ lim ℓ→∞ φ(cid:0)E(cid:0)f ℓ k(nn∗)nn∗na(cid:1)(cid:1) = φ(cid:0)E(na)(cid:1). Theorem 1.7. Let B ⊂ A be a non-degenerate abelian C-subalgebra with a con- ditional expectation E : A → B, which is normalized by some set N0 ⊂ N (B). For a state φ ∈ S(B), the following are equivalent: (cid:3) lim ℓ→∞ E(cid:0)anf ℓ φ(cid:0)E(an)(cid:1) = lim k→∞ = lim k→∞ φ(cid:0)E(an)(cid:1) = lim k→∞ = lim k→∞ k→∞ lim ℓ→∞ lim ℓ→∞ lim ℓ→∞ k(n∗n)n∗n(cid:1) = lim φ(cid:0)E(cid:0)anf ℓ φ(cid:0)nE(cid:0)anf ℓ φ(cid:0)E(nanf ℓ φ(cid:0)E(cid:0)naf ℓ E(cid:0)anf ℓ k(n∗n)(cid:1) n∗n(cid:1) = k(n∗n)(cid:1) n∗(cid:1) . k (n∗n)n∗)(cid:1) = k(nn∗)nn∗(cid:1)(cid:1) . lim ℓ→∞ lim ℓ→∞ (11) (12) Because E is normalized by n, with the help of (3) our computation continues as: TRACES ARISING FROM REGULAR INCLUSIONS 5 (i) φ is N0-invariant; (ii) φ ◦ E is centralized by all elements of the C∗-subalgebra C∗(B ∪ N0) ⊂ A; (iii) the restriction (φ ◦ E)C ∗(B∪N0) is a tracial state on C∗(B ∪ N0). Proof. (i) ⇒ (ii). Assume φ is N0-invariant. By Lemma 1.6, we clearly have the inclusion N0 ⊂ Zφ◦E, so (using the fact that Zφ◦E is a C∗-subalgebra of A) in order to prove statement (ii), it suffices to show that φ ◦ E is also centralized by B, which is pretty clear, since B is abelian. The implication (ii) ⇒ (iii) is trivial, since any state becomes tracial when restricted to its centralizer. (iii) ⇒ (i). Assume (φ ◦ E)C ∗(B∪N0) is a tracial. In particular, N0 centralizes this restriction, so by Lemma 1.6 (applied to C∗(B ∪ N0) in place of A), it again follows that φ is N0-invariant. (cid:3) 2. Invariant states in the ´etale groupoid framework The invariance conditions from Section 1 can be neatly described in the context of ´etale groupoid C∗-algebras, which we briefly recall here. A groupoid is a set G along with a subset G(2) ⊂ G × G of composable pairs and two functions: composition G(2) ∋ (α, β) 7−→ αβ ∈ G and an involution G ∋ γ 7−→ γ−1 ∈ G (the inversion), such that the following hold: (i) γ(ηζ) = (γη)ζ whenever (γ, η), (η, ζ) ∈ G(2); (ii) (γ, γ−1) ∈ G(2) for all γ ∈ G, and γ−1(γη) = η and (γη)η−1 = γ for (γ, η) ∈ G(2). Elements satisfying u = u2 ∈ G are called units of G and the set of all such units is denoted G(0) ⊂ G and called the unit space of G. There are maps r, s : G → G(0) defined by r(γ) = γγ−1 s(γ) = γ−1γ that are called, respectively, the range and source maps. If A, B ⊂ G, then AB = {γ ∈ G : ∃α ∈ A, β ∈ B, such that αβ = γ}. It is not difficult to show that (α, β) ∈ G(2) if and only if s(α) = r(β). For a given unit u ∈ G(0) there is an associated group G(u) = {γ ∈ G : r(γ) = s(γ) = u}; this is called the isotropy or stabilizer group of u. The union of all isotropy groups in G forms a subgroupoid of G called Iso(G), the isotropy bundle of G. A groupoid is called principal (or an equivalence relation) if Iso(G) = G(0); that is, if no unit has non-trivial stabilizer group. Throughout this present paper a groupoid G will be called ´etale, if it is endowed with a Hausdorff, locally compact and second countable topology so that (a) the composition and inversion operations are continuous (the domain of ◦ is equipped with the relative product topology), and furthermore, (b) the range and source maps are local homeomorphisms. sX←−−− X By condition (b), for each γ ∈ G, there exists an open set γ ∈ X ⊂ G, such that rX−−−→ r(X) are homeomorphisms onto open sets in G; such the maps s(X) an X is called a bisection. Note that in the ´etale case, the unit space G(0) is in fact clopen in G, and all range and source fibers r−1(u), s−1(u), u ∈ G(0), are discrete in the relative topology; hence compact subsets of G intersect any given range (or source) fiber at most finitely many times. 6 DANNY CRYTSER, GABRIEL NAGY In order to define a C∗-algebra from an ´etale groupoid G, it is necessary to specify a ∗-algebra structure on Cc(G). This is given by (f × g)(γ) = X(α,β)∈G(2):αβ=γ f ∗(γ) = f (γ−1). f (α)g(β); (Compactness of supports ensures that the sum involved in the definition of the product gives a well-defined element of Cc(G).) As G(0) is open in G, we have an inclusion Cc(G(0)) ⊂ Cc(G), which turns Cc(G(0)) into a ∗-subalgebra. However, the ∗-algebra operations on Cc(G(0)) inherited from Cc(G) coincide with the usual (pointwise!) operations: h∗ = ¯h and h × k = hk, ∀ h, k ∈ Cc(G(0)). In fact, some- thing similar can be said concerning the left and right Cc(G(0))-module structure of Cc(G): for all f ∈ Cc(G), h ∈ Cc(G(0)) we have (13) (14) (f × h)(γ) = f (γ)h(cid:0)s(γ)(cid:1); (h × f )(γ) = h(cid:0)r(γ)(cid:1)f (γ). Following Renault ([14]), for an ´etale groupoid G, the full C∗-norm on Cc(G) is given as kf k = sup(cid:8)(cid:13)(cid:13)π(f )(cid:13)(cid:13) : π non-degenerate ∗-representation of Cc(G)(cid:9) , and the full groupoid C∗-algebra C∗(G) is defined to be the completion of Cc(G) in the full C∗-norm. When restricted to Cc(G(0)), the full C∗-norm agrees with the usual sup-norm k · k∞, so by completion, the embedding Cc(G(0)) ⊂ Cc(G) gives rise to a non-degenerate inclusion C0(G(0)) ⊂ C∗(G). At the same time, one can also consider the restriction map, which ends up being a contractive map contractive linear map E : C∗(G) → C0(G(0)), which is in fact a conditional expec- tation. We refer to E as the natural expectation. Using the KSGNS construction as- (Cc(G), k · k) ∋ f 7−→ f G(0) ∈ (cid:0)Cc(G(0)), k · k∞(cid:1), so by completion one obtains a sociated with E ([9]) we obtain a ∗-representation πE : C∗(G) → L(cid:0)L2 (C∗(G), E)(cid:1), where L2 (C∗(G), E) is the Hilbert C0(G(0))-module obtained by completing C∗(G) in the norm given by the inner product habiC0(G(0)) = E(a∗b). With this rep- resentation in mind, the quotient C∗(G)/ker πE is the so-called reduced groupoid C∗-algebra, denoted by C∗ red(G). An alternative description of the ideal ker πE is to employ the usual GNS-representations πevu◦E, associated with the states evu : C0(G(0)) ∋ h 7−→ h(u) ∈ C, u ∈ G(0). With these (honest) representations in evu ◦ E ∈ S(cid:0)C∗(G)(cid:1) that are obtained by composing E with evaluation maps mind, we have ker πE =Tu∈G(0) ker πevu◦E. As was the case with the full groupoid C∗-algebra, after composing with the quotient map πred : C∗(G) → C∗ red(G), we still have an embedding Cc(G) ⊂ C∗ red(G) as the com- pletion of the convolution ∗-algebra Cc(G) with respect to a (smaller) C∗-norm, denoted k · kred. As before, when restricted to Cc(G(0)), the norm k · kred agrees with k · k∞, so C0(G(0)) still embeds in C∗ red(G), and furthermore, since the nat- ural expectation E vanishes on ker πE, we will have a reduced version of natural expectation, denoted by Ered : C∗ red(G) → C0(G(0)), which satisfies Ered ◦ πred = E.. As pointed out for instance in [15], a large supply of normalizers for C0(G(0)) are those elements of the groupoid C∗-algebra represented by functions f ∈ Cc(G) red(G), so we can also view C∗ TRACES ARISING FROM REGULAR INCLUSIONS 7 supported in bisections. We shall refer to such elements as elementary normal- malizers, along with 0, red(G) -- as a C∗- algebra. Using the embedding of Cc(G) in the groupoid (full or reduced) C∗-algebra, izers of C0(G(0)). Note that the collection Nelem(cid:0)C0(G(0))(cid:1) of elementary nor- is a ∗-subsemigroup of N(cid:0)C0(G(0))(cid:1), and furthermore Nelem(cid:0)C0(G(0))(cid:1) generate the ambient algebra -- C∗(G) or C∗ we interpret Nelem(cid:0)C0(G(0))(cid:1) as a subset in Cc(G), namely: (15) Nelem(cid:0)C0(G(0))(cid:1) = [X bisection Cc(X) ⊂ Cc(G). Comment. In order to avoid any unnecessary notational complications or duplica- tions, the results and definitions in the remainder of this section are stated only us- red(G) as the ambient C∗-algebra. However, with only ing the reduced C∗-algebra C∗ a few explicitly noted exceptions, by composing with the quotient ∗-homomorphism red(G), the same results will hold if we use the full C∗-algebra πred : C∗(G) → C∗ C∗(G) instead; we leave it to the reader to write down the missing statements corresponding to the full case (by simply erasing the subscript "red" from the statements). The ´etale groupoid framework is particularly convenient because one of the hy- potheses in Lemma 1.6 above is automatically satisfied. Proposition 2.1. The natural conditional expectation Ered : C∗ red(G) → C0(G(0)) is normalized by all elementary normalizers. In particular, for a state φ on C0(G(0)), the following are equivalent: (i) φ is an Nelem(cid:0)C0(G(0))(cid:1)-invariant state on C0(G(0)); (ii) φ ◦ Ered is a tracial state on C∗ red(G). Proof. Assume n ∈ Cc(X), for some bisection X ⊂ G. In order to prove the first assertion, we must show that Ered(n × f × n∗) = n × Ered(f ) × n∗, for all f ∈ Cc(G). Fix f , as well as x ∈ G(0). Then Ered(n × f × n∗)(u) =(n(γ)2f (s(γ)) 0 if ∃γ ∈ X ∩ r−1(u) ∩ s−1(supp f ) else . It is straightforward to verify that this is the same as (n × Ered(f ) × n∗) (u). The second statement is a direct consequence of Theorem 1.7, combined with (cid:3) the fact that Nelem(C0(G(0))) generates C∗ red(G) as a C∗-algebra. We want to characterize the Nelem(C0(G(0)))-invariant states on C0(G(0)) -- here- after referred to as elementary invariant states -- completely in measure-theoretical terms on G. We introduce the following terminology in parallel with Definition 1.1. Definition 2.2. Let G be an ´etale topological groupoid with unit space G(0), and let µ be a positive Radon measure on G(0). (1) Given an open bisection X ⊂ G, we say that µ is X-balanced if µ(XBX −1) = µ(s(X) ∩ B) for any Borel set B ⊂ G(0). (2) If X is a family of open bisections, then we say that µ is X -balanced if µ is X-balanced for all X ∈ X . (3) If µ is X-balanced for every open bisection X, then we say that µ is totally balanced. 8 DANNY CRYTSER, GABRIEL NAGY Notations. Given a proper continuous function between locally compact spaces h : X → Y , and a Radon measure µ on X, we denote its h-pushforward by h∗µ. This is a Radon measure on Y , given by (h∗µ)(A) = µ(h−1(A)), for any Borel set A ⊂ Y . Note that the pushforward construction is covariant: (g ◦ f )∗µ = g∗(f∗µ). By Riesz's Theorem, we have a bijective correspondence (16) Prob(X) ∋ µ 7−→ φµ ∈ S(cid:0)C0(X)(cid:1) between the space of Radon probability measures on X and the state space of C0(X), defined as follows. For each µ ∈ Prob(X), the associated state φµ ∈ S(cid:0)C0(X)(cid:1) is: φµ(f ) =ZX f (x) dµ(x), f ∈ C0(X). On the level of positive linear functionals, the pushforward construction corresponds to composition: (h∗φ)(f ) = φ(cid:0)f ◦ h), f ∈ C0(Y ), h : X → Y. Lemma 2.3. With G as above, let X ⊂ G be an open bisection. For a finite Radon measure µ on G(0), the following are equivalent: (i) µs(X) =(cid:0)s ◦ (rX )−1(cid:1)∗(µr(X)); (ii) µ(cid:0)s(B)(cid:1) = µ(cid:0)r(B)(cid:1), for all Borel subsets B ⊂ X; (iii) µ(cid:0)s(K)(cid:1) = µ(cid:0)r(K)(cid:1), for all compact subsets K ⊂ X; (iv) µ is X-balanced. (In condition (i) we use the restriction notation for measures: if µ is a finite Radon measure on G(0) -- thought as a function µ : Bor(G(0)) → [0, ∞), and D ⊂ G(0) is some open subset, then µD is the Radon measure on D obtained by restricting µ to Bor(D).) Proof. The equivalence (i) ⇔ (ii) is trivial, because the maps s(X) r(X) are homeomorphisms onto open sets. sX←−−− X rX−−−→ The equivalence (ii) ⇔ (iv) follows from the observation that, for any Borel set B ⊂ G(0), the set B′ = X ∩ s−1(B) ⊂ X is Borel, and furthermore, the sets that appear in the definition of X-invariance are precisely XBX −1 = r(B′) and s(X) ∩ B = s(B′). Lastly, the equivalence (ii) ⇔ (iii) follows from regularity and finiteness of µ. (cid:3) We are interested in balanced measures because they are tied up with elementary invariance. Lemma 2.4. Let G be an ´etale groupoid with unit space G(0), let µ be a Radon probability measure on G(0), and let φµ be the state on the C∗-subalgebra C0(G(0)) ⊂ C∗ red(G) given by (16). For an open bisection X ⊂ G, the following conditions are equivalent: (i) µ is X-balanced; (ii) φµ is Cc(X)-invariant. (As in (15), Cc(X) ⊂ NC ∗ Proof. The entire argument will be based on the following red(G)(cid:0)C0(G(0))(cid:1).) TRACES ARISING FROM REGULAR INCLUSIONS 9 Claim. For any n ∈ Cc(X) and any b ∈ Cc(G(0)), one has the equalities: (17) (18) (19) 2 φµ(n∗ ×n×b) =Zs(X)(cid:12)(cid:12)(cid:0)n ◦ (sX )−1(cid:1) (u)(cid:12)(cid:12) φµ(n×b×n∗) =Zr(X)(cid:12)(cid:12)(cid:0)n ◦ (rX )−1(cid:1) (u)(cid:12)(cid:12) φµ(n×b×n∗) =Zs(X)(cid:12)(cid:12)(cid:0)n ◦ (sX )−1(cid:1) (u)(cid:12)(cid:12) b(u) d(cid:0)µs(X)(cid:1) (u); 2(cid:0)b ◦ s ◦ (rX )−1(cid:1) (u) d(cid:0)µr(X)(cid:1) (u); b(u) d(cid:0)s ◦ (rX )−1(cid:1)∗(cid:0)µr(X)(cid:1) (u). 2 The equality (17) follows from the definition of the convolution multiplication and ∗-involution, which yields so we can multiply the functions n∗n and b to obtain: 0 (n∗ × n)(u) =((cid:12)(cid:12)n(cid:0)(sX )−1(u)(cid:1)(cid:12)(cid:12) (n∗ × n × b)(u) =((cid:12)(cid:12)n(cid:0)(sX )−1(u)(cid:1)(cid:12)(cid:12) 0 2 2 u ∈ s(X) u 6∈ s(X) b(u) u ∈ s(X) u 6∈ s(X). Likewise, the equality in (18) follows from (n × b × n∗)(u) =((cid:12)(cid:12)n(cid:0)(rX )−1(u)(cid:1)(cid:12)(cid:12) 0 2 · b(cid:0)s(cid:0)(rX )−1(u)(cid:1)(cid:1) u ∈ r(X) u 6∈ r(X) which implies that the support of n × b × n∗ is contained in X(supp b)X −1 ⊂ r(X). Lastly, the equality between the right-hand sides of (18) and (19) follows immediately by applying the definition of the pushforward (20) Zs(X) f d(cid:0)s ◦ (rX )−1(cid:1)∗(cid:0)µr(X)(cid:1) =Zr(X)(cid:0)f ◦ s ◦ (rX )−1(cid:1) d(cid:0)µr(X)(cid:1) , 2 to functions f ∈ Cc(cid:0)s(X)(cid:1) of the form: f (u) =(cid:12)(cid:12)n ◦(cid:0)(sX )−1(cid:1) (u)(cid:12)(cid:12) Having proved the Claim, the implication (i) ⇒ (ii) follows from Lemma 2.3, which yields: b(u). (21) ∀ n ∈ Cc(X), b ∈ Cc(G(0)) : φµ(n∗ ×n×b) = φµ(n×b×n∗). By density, (21) holds for all n ∈ Cc(X), b ∈ C0(G(0)), thus φµ is n-invariant for all n ∈ Cc(X). As for the implication (ii) ⇒ (i), all we have to observe is that, if φµ is Cc(X)- invariant, then (21) is valid, which by the identities (17) and (19), simply state that the equality f d(cid:0)s ◦ (rX )−1(cid:1)∗(cid:0)µr(X)(cid:1) =Zs(X) f d(cid:0)µs(X)(cid:1) , 2 b(u), n ∈ Cc(X), b ∈ Cc(G(0)). (22) (23) holds for all functions of the form: Zs(X) f (u) =(cid:12)(cid:12)(cid:0)n ◦ (sX )−1(cid:1) (u)(cid:12)(cid:12) Since (using a partition of unity argument) the functions of the above form linearly span all functions in Cc (s(X)), the equality (22) simply states that so by Lemma 2.3, it follows that µ is indeed X-balanced. (cid:0)s ◦ (rX )−1(cid:1)∗(cid:0)µr(X)(cid:1) = µs(X), (cid:3) 10 DANNY CRYTSER, GABRIEL NAGY Combining Proposition 2.1 with Lemma 2.4, we now reach the following conclu- sion. Theorem 2.5. Let G be an ´etale groupoid with unit space G(0), let µ be a probability Radon measure on G(0), and let φµ be the state on the C∗-subalgebra C0(G(0)) ⊂ C∗ red(G) given by (16). The following conditions are equivalent: (i) µ is totally balanced; (ii) φµ is elementary invariant; (iii) φµ is fully invariant; (iv) φµ ◦ Ered is a tracial state on C∗ red(G). (cid:3) In concrete situations, one would like to check condition (i) from the above Theorem in an "economical" way. To be more precise, assuming that a given measure µ ∈ Prob(G(0)) is X -balanced, for some collection of bisections X , we seek a natural subalgebra on which φµ◦Ered is tracial (as in Theorem 1.7), and furthermore find criteria on X which ensure that our subalgebra is in fact all of C∗ red(G). Parts of the Lemma below mimic corresponding statements from Proposition 1.3. (Each one of statements (i) -- (iii) has an implicit statement built-in: the new sets, such as X ′, X −1 and X1X2 are always bisections.) Proposition 2.6. Let G be an ´etale groupoid with unit space G(0) and let µ be a Radon probability measure on G(0), (i) If µ is X-balanced, for some bisection X, then µ is X ′-balanced, for any open subset X ′ ⊂ X. (ii) If µ is X-balanced, for some bisection X, then µ is X −1-balanced. (iii) If µ is both X1- and X2-balanced, for two bisections X1, X2, then µ is X1X2-balanced. (iv) Assume X is an open set, written as a union X = Sj∈J Xj of bisections, such that sX , rX : X → G(0) are injective. Then X is a bisection, and if µ is Xj-balanced for all j ∈ J, then µ is X-balanced. Proof. Statements (i) and (ii) are trivial from Lemma 2.3. Before we prove (iii), we need some clarifications. First of all, the set X1X2 is obtained as the image of the open set X1 ◦ X2 = {(α, β) ∈ X1 × X2 : s(α) = r(β)} = X1 × X2 ∩ G(2) ⊂ G(2). p1←−− X1 ◦ X2 p1←−− X1 × X2 under composition map m : G(2) → G. Secondly, by the bisection property, the p2−−→ X2 give rise to two restrictions of the coordinate maps X1 p2−−→ p2(X1 ◦ X2) onto open subsets of homeomorphisms p1(X1 ◦ X2) X1 and X2 respectively, and furthermore the compositions s ◦ p1 and r ◦ p2 agree on X1 ◦ X2, and the resulting map, denoted here by t : X1 ◦ X2 →⊂ G(0) is a homeomorphism onto an open subset D ⊂ G(0). (This open set is simply D = t(X1 ◦ X2) = s(X1) ∩ r(X2). By construction, X1X2 = ∅ ⇔ s(X1) ∩ r(X2) = ∅.) Furthermore, again by the bisection property, mX1◦X2 : X1 ◦ X2 → X1X2 is also a homeomorphism onto an open set, so composing its inverse with the coordinate maps, we obtain two homeomorphisms qk = pk ◦ (mX1◦X2 )−1 : X1X2 → Xk, k = 1, 2, which satisfy sX1X2 = s ◦ q1 and rX1X2 = r ◦ q2. Using all these three homeomorphisms, the fact that X1X2 is a bisection is obvious. Not only are the maps s(X1X2) map r ◦ q2 = s ◦ q1 = t ◦ (mX1◦X2 )−1 : X1X2 → D. sX1 X2←−−−−− X1X2 rX1X2 −−−−−→ r(X1X2) homeomorphisms, but so is the TRACES ARISING FROM REGULAR INCLUSIONS 11 After all these preparations, statement (iii) follows from the observation that the X1- and X2-balancing features imply that, for any Borel set B ⊂ X1X2 we have µ(cid:0)s(B)(cid:1) = µ(cid:0)s(cid:0)q2(B)(cid:1)(cid:1) = µ(cid:0)r(cid:0)q2(B)(cid:1)(cid:1) = = µ(cid:0)s(cid:0)q1(B)(cid:1)(cid:1) = µ(cid:0)r(cid:0)q1(B)(cid:1)(cid:1) = µ(cid:0)r(B)(cid:1), so the desired conclusion follows from Lemma 2.3. (iv). Since we have the equalities s(X) =Sj∈J s(Xj) and r(X) =Sj∈J s(Xj), it follows that s(X) and r(X) are open. The fact that both s(X) are homeomorphisms follows by local compactness. rX−−−→ r(X) sX←−−− X Finally, to prove that µ is X-balanced, we apply criterion (iii) from Lemma 2.3. Start with some compact set K ⊂ X, and using compactness write it as a finite k=1 Bjk , where Bjk ⊂ Xjk , k = 1, . . . , n are Borel sets. Using disjoint union K =Sn the fact that µ is Xj-balanced for all j, we know that µ(cid:0)s(Bjk )(cid:1) = µ(cid:0)r(Bjk )(cid:1), for all k, so using that s and r are homeomorphisms, we also have s(K) =Sn r(K) =Sn k=1 s(Bjk ) and k=1 r(Bjk ) (disjoint unions of Borel sets in s(X) and r(X) respectively), so we have n n Xk=1 [k=1 µ(cid:0)s(K)(cid:1) = µ(cid:0) µ(cid:0)s(Bjk )(cid:1) = s(Bjk )(cid:1) = Xk=1 [k=1 µ(cid:0)r(Bjk )(cid:1) = µ(cid:0) r(Bjk )(cid:1) = µ(cid:0)r(K)(cid:1). = n n (cid:3) Using the above result, combined with Lemma 2.4, we immediately obtain the following measure-theoretic groupoid analogue of Theorem 1.7. Theorem 2.7. Assume W is a collection of bisections in the ´etale groupoid G, and let X be the inverse semigroup generated by W. For a measure µ ∈ Prob(G(0)), the following are equivalent: (i) µ is W-balanced; (ii) µ is X -balanced; (iii) the state φµ ◦ Ered is tracial when restricted to the subalgebra C∗(cid:18)C0(G(0)) ∪ [W ∈W Cc(W )(cid:19) = span(cid:18)C0(G(0)) ∪ [X∈X Cc(X)(cid:19). (cid:3) Remark 2.8. A sufficient condition for a collection X of bisections of G to satisfy the equality span(cid:18)C0(G(0)) ∪ [X∈X Cc(X)(cid:19) = C∗ red(G) is that X covers G r G(0). This follows using a standard partition of unity ar- gument, which implies the equality Cc(G) = span(cid:18)C0(G(0)) ∪SX∈X Cc(X)(cid:19). As a consequence, the desired "economical" criterion for traciality of φµ ◦ Ered is as follows. Corollary 2.9. Assume G, W and X are as in Theorem 2.7. If µ ∈ Prob(G(0)) is W-balanced, and X covers G r G(0), then φµ ◦ Ered is tracial on C∗ (cid:3) red(G). 12 DANNY CRYTSER, GABRIEL NAGY 3. Tracial states via extension properties So far, assuming that an non-degenerate abelian C∗-subalgebra B ⊂ A is the range of a conditional expectation E : A → B, we have examined certain conditions both for a state φ ∈ S(B) and for E, that ensure that φ ◦ E is a trace. In the groupoid framework, the natural conditional expectation E exhibited nice behavior (elementary invariance), so the focus was solely placed on φ. In this section we provide another framework, in which again the conditional expectation in question will also be normalized by all n ∈ N (B). (As a side issue one should also be concerned with the uniqueness of conditional expectation.) A natural class of subalgebras to which this analysis can be carried on nicely are Renault's Cartan subalgebras ([15]; see also the Comment following Corollary 3.3 below). As it turns out, very little from the Cartan subalgebra machinery is needed for our purposes: the almost extension property ([11]), which requires that the set P1(B ↑ A) = {ω ∈ B : ω has a unique extension to a state on A} is weak-∗ dense in B -- the Gelfand spectrum of B. (A slight strengthening of the above condition will be introduced in the Comment following Lemma 3.2 below.) The utility of the almost extension property is exhibited by Lemma 3.2 below, in preparation of which we need the following simple fact. Fact 3.1. Let ω be a state on B ⊂ A with extension θ ∈ S(A), so that θB = ω. If x, y ∈ A and satisfy either (1) y∗y ∈ B and ω(y∗y) = 0, or (2) xx∗ ∈ B and ω(xx∗) = 0, then θ(xy) = 0. In particular, if b ∈ B satisfies 0 ≤ b ≤ 1 and ω(b) = 1, then ∀ a ∈ A : θ(a) = θ(ab) = θ(ba) = θ(bab). Proof. Apply the Cauchy-Schwarz inequality for the sesquilinear form: haa′i = θ(a∗a′). The second statement follows from the first one applied with y = 1 − b. (cid:3) Lemma 3.2 (compare to [8, Lemma 6]). Let B ⊂ A be a non-degenerate abelian C∗-subalgebra with the almost extension property, and let E : A → B be a condi- tional expectation. Then E is normalized by all n ∈ N (B). Comment. As noted in [11], the almost extension property implies that at most one conditional expectation E : A → B can exist. In the case such an expectation does exist and the almost extension property holds, we say that the inclusion B ⊂ A has the conditional almost extension property. Proof of Lemma 3.2. Fix some normalizer n ∈ N (B), and let us prove that (24) E(nan∗) = nE(a)n∗, for all a ∈ A. Fix polynomials (f ℓ k) as in Fact 1.2(iii), so we have (25) E(nan∗) = lim k→∞ lim ℓ→∞ E(nn∗nf ℓ k(n∗n)af ℓ k(n∗n)n∗nn∗). TRACES ARISING FROM REGULAR INCLUSIONS 13 Likewise, and using also the fact that E is a conditional expectation, we also have (26) nE(a)n∗ = lim k→∞ = lim k→∞ lim ℓ→∞ lim ℓ→∞ nn∗nf ℓ k(n∗n)E(a)f ℓ k(n∗n)n∗nn∗ = nE(n∗nf ℓ k(n∗n)af ℓ k(n∗n)n∗n)n∗, Inspecting (25) and (26), we now see that it suffices to prove (24) for elements of the form a = n∗a1n; in other words, instead of (24), it suffices to prove (27) ∀ a ∈ A : E(nn∗ann∗) = nE(n∗an)n∗, As both sides of this equation belong to B, we only need show that () ω(E(nn∗ann∗)) = ω(nE(n∗an)n∗) for all ω ∈ P1(B ↑ A). Suppose that ω(nn∗) = 0. In this case, we have by Fact 3.1 that both sides of () are zero. Suppose that ω(nn∗) > 0 and define two states ψω and θω on A by ψω(a) = (ω ◦ E)(nn∗ann∗) ω(nn∗)2 and θω(a) = ω(nE(n∗an)n∗) ω(nn∗)2 , so () is equivalent to the equality ψω = θω (of states on A). Note that, if b ∈ B, then ψω(b) = θω(b) = ω(b), so that both states ψω and θω are extensions of ω ∈ P1(B ↑ A), so by uniqueness we have ψω = θω, and () is established. (cid:3) In the context of the conditional almost extension property, Theorem 1.7 has the following consequences. Corollary 3.3. Let B ⊂ A be a non-degenerate abelian C-subalgebra with the conditional almost extension property, let E : A → B be its (unique) conditional expectation, and let φ be a state on B. (a) For a subset N0 ⊂ N (B) the following are equivalent: (i) φ is N0-invariant; (ii) φ ◦ E is centralized by all elements of C∗(B ∪ N0) ⊂ A; (iii) the restriction (φ ◦ E)C ∗(B∪N0) is a tracial state on C∗(B ∪ N0). (b) In particular, if B is regular, then φ ◦ E is a trace on A if and only φ is (cid:3) fully invariant. (Of course, statement (b) can be slightly relaxed, by requiring that φ is only N0- invariant for a subset N0 ⊂ N (B) which together with B generates A as a C∗- algebra.) Comment. A natural class exhibiting the conditional almost extension property are Cartan subalgebras, as defined by Renault in [15]. They are regular non-degenerate inclusions B ⊂ A, in which • B is maximal abelian (masa) in A, and • there exists a faithful conditional expectation E : A → B (which is neces- sarily unique). As pointed out for instance in [3], Cartan subalgebras do have the the conditional almost extension property, but there are many examples of regular non-degenerate abelian C∗-subalgebra inclusions B ⊂ A with the conditional almost extension property, which are non-Cartan. In fact, for ´etale groupoids, the equivalent condi- tion to the almost extension property is topological principalness: the set of units u ∈ G(0) with trivial isotropy G(u) is dense in G(0). For topologically princi- red(G) and C0(G(0)) ⊂ C∗(G) have pal groupoids, both inclusions C0(G(0)) ⊂ C∗ 14 DANNY CRYTSER, GABRIEL NAGY the conditional almost extension property. However, since the (full) conditional expectation E : C∗(G) → C0(G(0)) is not faithful in general, C0(G(0)) is gener- ally not Cartan in C∗(G). On the other hand, since the (reduced) expectation Ered : C∗ red(G) → C0(G(0)) is faithful, C0(G(0)) is Cartan in C∗ red(G). Up to this point, we have seen that for regular non-degenerate abelian C∗- subalgebras B ⊂ A with the conditional almost extension property, Corollary 3.3(b) provides us with an injective w∗-continuous affine map (28) Sinv(B) ∋ φ 7−→ φ ◦ E ∈ T (A), which is a right inverse of the restriction map (2); in particular, it follows that for such inclusions, the map (2) is surjective. Question. If B ⊂ A is a regular non-degenerate abelian C∗-subalgebra with the conditional almost extension property, under what additional circumstances is the map (28) also surjective? (If this is the case, this would imply that the restriction map (2) is in fact an affine w∗-homeomorphism.) As the Example below suggests, even in the case of Cartan inclusions, the map (28) may fail to be surjective. Example 3.4. Let B = C(D) ⊂ A = C(D) ⋊α Z = C∗(C(D), u), where α is rotation of D by an irrational multiple of π and u is the unitary that implements the automorphism in the crossed product. Then B is a Cartan subalgebra as can be directly verified. The conditional expectation is given on the dense set of Laurent polynomials in u by E(X fnun) = f0. (It is obvious that E(un) = 0 for all n 6= 0.) As 0 is a fixed point under the rotation α, we have that (ev0(·)1, id) is a covariant representation of (C(D), α) in C∗(Z) ∼= C(T), thus it induces a ∗-homomorphism ρ : A → C(T). Any state ψ on C(T) defines a state ψ ◦ ρ on A, which is clearly tracial since C(T) is abelian and ρ is a ∗-homomorphism. A tracial state of this form factors through E if and only if it maps {un}n6=0 to 0, so taking for instance ψ = evz to be a point evaluation at z ∈ T, then clearly (evz ◦ ρ)(u) = z 6= 0, so the trace τ = evz ◦ ρ ∈ T (A) does not belong to the range of the map (28). Remark 3.5. In connection with the above example, the reason that the map φ → φ ◦ E fails to be surjective is the fact that the state ev0 on C(D) does not have a unique extension to a state on C(D) ⋊ Z. Such an obstruction can be avoided if we consider inclusions with the (honest) extension property, which are those non- degenerate abelian C∗-subalgebra inclusion B ⊂ A for which every pure state on B has a unique extension to a state on A. As shown in [7] and [1], the extension property implies the following: • B is maximal abelian; • there exists a unique conditional expectation E : A → B • ker E = [A, B] (the closed linear span of the set of elements of the form ab − ba, a ∈ A, b ∈ B). From the last two properties it follows immediately that any tracial state τ ∈ T (A) vanishes on ker E. Thus, any tracial state factors through E, and is completely determined by its restriction to B. Since restrictions of the form τ(cid:12)(cid:12)B, τ ∈ T (A) are always fully invariant, Corollary 3.3 has the following immediate consequence. TRACES ARISING FROM REGULAR INCLUSIONS 15 Corollary 3.6. If B ⊂ A is a regular abelian C∗-subalgebra algebra inclusion with the extension property, and E : A → B is its associated conditional expectation, then the map Sinv(B) ∋ φ 7−→ φ ◦ E ∈ T (A) is an affine w∗-homeomorphism, with inverse τ → τ B. (cid:3) Example 3.7. For an ´etale groupoid G, the inclusions of C0(G(0)) into either the full or reduced C∗-algebra of G have the extension property if and only if G is principal : all units in G have trivial isotropy group. In the case when G is a principal groupoid, the above combined with Theorem 2.5 (in both its reduced and full versions) establishes a bijection between the set of totally balanced measures on G(0) and the tracial state spaces of both C∗(G) and C∗ In particular, if Γ is a discrete group acting freely on X, then the tracial state spaces of both crossed-product C∗-algebras C0(X) ⋊ Γ and C0(X) ⋊red Γ are naturally identified with the Γ-invariant Radon probability measures on X. red(G). The condition that the groupoid be principal (or for crossed products, that the action be free) cannot be relaxed, especially in the non-amenable case, as the following example shows. Let F2 -- the free group on two generators -- act on by translation on its Alexandrov compactification F2 ∪ {∞} (by keeping ∞ fixed), so that the associated action of F2 on the unitized on c0(F2)∼ is given by αg(f +c1) = λg(f ) + c1, where λ is the left-shift action on c0(F2). It is not hard to show that c0(F2)∼ ⋊red F2 has a unique tracial state. On the other hand, the full crossed product c0(F2)∼ ⋊ F2 has the full group C∗-algebra C∗(F2) as quotient, and so it must have infinitely many tracial states. 4. Graph C-algebras In this section we provide a method for parametrizing tracial state spaces on graph C∗-algebras. Our approach complements the treatment in [18] by giving an explicit parametrization of the tracial state space of a graph C∗-algebra. We begin with a quick review of graph terminology and notation, most of which are borrowed from [13]. A directed graph E = (E0, E1, r, s) consists of two countable sets E0, E1 as well as range and source maps r, s : E1 → E0. A vertex is regular if r−1(v) is finite and non-empty. A vertex which is not regular is called singular ; a singular vertex is either a source (r−1(v) = ∅) or an infinite receiver (r−1(v) infinite). is denoted by En; the collection S∞ A finite path in E is a sequence λ = e1 . . . en of edges satisfying s(ek) = r(ek+1) for k = 1, . . . , n − 1. (Note that we are using the right-to-left convention.) The length λ = e1 . . . en is defined to be λ = n, and the set of paths of length n in E n=0 En of all finite paths in E is denoted E∗. (The vertices E0 are included in E∗ as the paths of length zero.) An infinite path in E is an infinite sequence e1e2 . . . of edges in E satisfying s(ek) = r(ek+1) for all k; the set of these paths is denoted by E∞. If λ = e1 . . . en is a finite path then we define its range r(λ) to be r(e1), and its source s(λ) to be s(en). The range of an infinite path is defined the same way. In order to avoid any confusion, for any vertex v ∈ E0, and any n ∈ N ∪ {∞}, the set {λ ∈ En : r(λ) = v, λ = n} will be denoted by r−n(v). If λ is a finite path and ν is a finite (or infinite) path with s(λ) = r(ν), then we can concatenate the paths to form λν. Whenever a (finite or infinite) path σ can 16 DANNY CRYTSER, GABRIEL NAGY be decomposed as σ = λν, we write λ ≺ σ (or σ ≻ λ) and we denote ν by σ ⊖ λ. A cycle is a finite path λ of positive length with r(λ) = s(λ). Given a cycle λ = e1 . . . en ∈ E∗, an entry to λ is a path f1f2 . . . fj, j > 0, with r(f1) = r(ek) and f1 6= ek, for some k. If no entry to λ exists, we say that λ is entry-less. It fairly easy to see that every entry-less cycle λ can be written uniquely as a repeated concatenation λ = νm, of a simple entry-less cycle ν, i.e. the number of vertices in ν equals ν. An infinite path x is called periodic if there exist α, λ ∈ E∗, with s(α) = r(λ) = s(λ), such that x = αλ∞ (that is, x is obtained by following α and then repeating the cycle λ forever). If x = αλ∞, and λ has minimal length among any cycle in such a decomposition, then the period of x is defined to be λ and is denoted per(x). Definition 4.1. If B is a C∗-algebra then a Cuntz-Krieger E-family in B is a set {Se, Pv}e∈E1,v∈E0 , where the Se are partial isometries with mutually orthogonal range projections and the Pv are mutually orthogonal projections which also satisfy: (i) S∗ (ii) Se S∗ e Se = Ps(e); e ≤ Pr(e); (iii) if v is regular, then Pv =Pr(e)=v Se S∗ e . The C∗-subalgebra of B generated by {Se, Pv}e∈E1,v∈E0 is denoted C∗(S, P ). The graph algebra C∗(E) is the universal C∗-algebra generated by a Cuntz-Krieger E-family, C∗(E) = C∗(s, p), where {se, pv} are the universal generators. For any Cuntz-Krieger E-family {Se, Pv}e∈E1,v∈E0 there is a unique ∗-homomorphism πS,P : C∗(E) → C∗(S, P ) satisfying πS,P (se) = Se and πS,P (pv) = Pv. For an E-family {S, P } and a finite path λ = e1 . . . en in E∗, there is an associated partial isometry Sλ = Se1 Se2 . . . Sen in C∗(S, P ). (If λ = 0, so λ reduces to a vertex v ∈ E0, then Sλ = Pv.) When specializing to C∗(E), we have partial isometries denoted sλ, λ ∈ E∗. By construction, all sλ ∈ C∗(E), λ ∈ E∗ are partial isometries: the source λ will be denoted from λsλ = ps(λ); the range projection sλs∗ projection of sλ is s∗ now on by pλ. As it turns out, one has the equality (29) C∗(E) = span{sαs∗ β : α, β ∈ E∗, s(α) = s(β)}. The products sαs∗ β listed in the right-hand side of (29) are referred to as the span- ning monomials, and the set of all these elements is denoted by G(E). The equality (29) is due to the fact that G(E) ∪ {0} is a ∗-semigroup, which is a consequence of the following product rule: (30) (sαs∗ β)(sλs∗ sαs∗ ν(β⊖λ), sα(λ⊖β)s∗ ν, 0, if λ ≺ β if β ≺ λ otherwise ν) =  Since all projections pv, v ∈ E0 are mutually orthogonal, for any finite set net (qV )V ∈Pfin(E0) forms an approximate unit for C∗(E), hereafter referred to as V ⊂ E0, the sum qV = Pv∈V pv will be again a projection, and furthermore, the the canonical approximate unit. The ∗-subalgebraSV ∈Pfin(E0) qV C∗(E)qV will be denoted by C∗(E)fin. TRACES ARISING FROM REGULAR INCLUSIONS 17 Passing from a graph to a sub-graph does not always produce a meaningful link between the associated C∗-algebras. The best suited objects that allow such links are the identified as follows: given some graph E, a subset H ⊂ E0 is called • hereditary, if r(e) ∈ H implies s(e) ∈ H • saturated, if whenever v ∈ E0 is regular and {s(e) : e ∈ r−1(v)} ⊂ H, it follows that v ∈ H. Any subset H ⊂ E0 is contained in a minimal saturated set H called its saturation, which is the union H =S∞ (31) k=0 Hk, where H0 = H and, for k > 1, Hk = Hk−1 ∪ {v ∈ E0 : v regular and s(r−1(v)) ⊂ Hk−1}. Clearly, the saturation of a hereditary set is again hereditary. The main point about considering such sets is the fact (see [13]) that, whenever H ⊂ E0 is saturated and hereditary, and we form the sub-graph E \ H = (E0 r H, s−1(E0 r H), r, s), then we have a natural surjective ∗-homomorphism ρH : C∗(E) → C∗(E \ H), defined on the generators as ρH (pv) =(pv, 0, if v ∈ E0 r H; otherwise; ρH (se) =(se, 0, if s(e) ∈ E0 r H; otherwise. (A sub-graph of this form will be called canonical.) The ideal ker ρH is simply the closed two-sided ideal generated by {pv}v∈H; alternatively, it is also described as: ker ρH = span{sαs∗ β : α, β ∈ E∗, s(α) = s(β) ∈ H}. The gauge action on C∗(E) is the point-norm continuous group homomorphism β) = zα−βsαs∗ γ : T ∋ z 7−→ γz ∈ Aut(cid:0)C∗(E)(cid:1), given on the generators by γz(pv) = pv, v ∈ E0 and γz(se) = zse, e ∈ E1. On the spanning monomials listed above, the automorphisms γz, z ∈ T, act as γz(sαs∗ β. The gauge invariant uniqueness theorem of an Huef and Raeburn (see [5]) states that, given some C∗-algebra A equipped with a group homomorphism θ : T ∋ z 7−→ θz ∈ Aut(A), and a gauge invariant ∗-homomorphism π : C∗(E) → A (that is, such that θz (π(x)) = π (γz(x)), ∀ x ∈ C∗(E), z ∈ T), the condition that π is injective is equivalent to the condition that π(pv) 6= 0, for all v ∈ E0. There are two distinguished abelian C∗-subalgebras of C∗(E) which we use to define states on C∗(E), the first of which is defined as follows. Definition 4.2. Let E be a directed graph. Then the diagonal D ⊂ C∗(E) is the C∗-subalgebra of C∗(E) generated by the set GD(E) = {pα}α∈E ∗ . (We sometimes use the notation D(E), when specifying the graph is necessary.) Remark 4.3. As it turns out, GD(E) ∪ {0} is an abelian semigroup of projections; more specifically, by (30), the product rule for GD(E) is: (32) pαpβ = pβpα =  pα, pβ, 0, if β ≺ α if α ≺ β otherwise Using the semigroup property, span GD(E). We can also write D(E) = (cid:2)Pv∈E0 D(E)pv(cid:3)− it follows that we can in fact present D(E) = , with each summand 18 DANNY CRYTSER, GABRIEL NAGY presented as D(E)pv = span{pα : α ∈ E∗, pα ≤ pv} = span{pα : α ∈ E∗, r(α) = v}. As it turns out, each corner D(E)pv is in fact a unital abelian AF-subalgebra, with unit pv, so D itself is an abelian AF-algebra, which contains the canonical approximate unit (qV )V ∈Pfin(E0). As explained for instance in [10], the Gelfand spectrum \D(E) of the diagonal C∗-subalgebra D(E) can be identified with the set E≤∞ = E∞ ∪ {x ∈ E∗ : s(x) is singular } with evaluation maps defined by evD x (pα) = 1 if α ≺ x, and 0 otherwise. In other words, for each α ∈ E∗, when we view pα ∈ D(E) as a continuous function on \D(E) ≃ E≤∞, this function will be the indicator function of the compact-open set Z(α) = {x ∈ E≤∞ : α ≺ x}. Furthermore, the sets Z(α), α ∈ E∗ form a basis for the topology, so clearly \D(E) is a totally disconnected. When identifying D(E) ≃ C0(cid:0)\D(E)(cid:1), the algebraic sum (without closure!) D(E)fin =Pv∈E0 D(E)pv gets naturally identified with Cc(cid:0)\D(E)(cid:1), the algebra of continuous functions with compact support. Remark 4.4. Cylinder sets can be used to analyze path (in)comparability. To be more precise, given two paths, α, β ∈ E∗, the following statements hold. I. (Comparability Rule) The inequality α ≺ β is equivalent to the reverse inclusion Z(α) ⊃ Z(β). II. (Orthogonality Rule) Conditions (i) -- (iv) below are equivalent: αsβ = 0; (i) s∗ (ii) the projections pα and pβ are orthogonal, i.e. pαpβ = 0; (iii) α and β are incomparable, i.e. α 6≺ β and β 6≺ α; (iv) Z(α) ∩ Z(β) = ∅. Remark 4.5. Among all paths x ∈ E≤∞, the ones of interest to us will be those that represent isolated points in the spectrum \D(E). On the one hand, if E has sources (i.e. vertices v ∈ E0 with r−1(v) = ∅), then all finite paths that start at sources are determine isolated points in \D(E). On the other hand, the infinite paths x = e1e2 · · · ∈ E∞ that produce isolated points in \D(E) are precisely those with the property that there exists k such that r−1(r(en)) = {en}, for all n ≥ k. If this is the case, if we form α = e1e2 . . . ek−1, then {x} = Z(α). Among those paths, the periodic ones will play an important role in our discussion. Definition 4.6. A finite path α = e1e2 . . . en ∈ E∗ (possibly of length zero) is called a ray if there is a a simple entry-less cycle ν, such that s(α) = s(ν), and furthermore, no edge ek from α is appears in ν. (Note: In [10], rays were called distinguished paths.) In this case, the cycle ν (which is uniquely determined by α) is referred to as the seed of α. We caution the reader that zero-length rays are permitted: they are what we will call cyclic vertices. For reasons explained in the second paragraph below, the (possibly empty) set of all rays in E will be denoted by E∗ ip. By definition, any two distinct rays α1 6= α2 are incomparable, so by the Or- thogonality Rule (Remark 4.4) they satisfy: s∗ α1 sα2 = s∗ α2 sα1 = 0. TRACES ARISING FROM REGULAR INCLUSIONS 19 Clearly, rays parametrize the set E∞ ip of infinite periodic paths that yield isolated points in \D(E): any such path can be uniquely presented as x = αν∞, with α ray and ν the seed of α, and its period (as a function from N to E1) is per(x) = ν. When it would be necessary to emphasize the sole dependence on α, we also denote the infinite path αν∞ simply by ξα. When we collect the corresponding points in \D(E), we obtain a countable open set Σip = {evD x : x ∈ E∞ ip } ⊂ \D(E). Remark 4.7. Associated with the space E≤∞ we have the path representation πpath : C∗(E) → B(ℓ2(E≤∞)) given on generators by (see [13] for details): πpath(se)δx =(δex 0 r(x) = s(e) otherwise; πpath(pv)δx =(δx 0 r(x) = v otherwise. In general, πpath is not faithful; however, it is always faithful on the diagonal subalgebra D(E). This embedding gives us a explicit form of the identification \D(E) = E≤∞ as follows: for x ∈ E≤∞, the associated character on D(E) is simply evD x (a) = hδxπpath(a)δxi. For future use, we denote the subalgebras πpath(D(E)) and πpath(C∗(E)) of B(ℓ2(E≤∞)) by Dpath(E) and Apath(E), respectively. Notation. As shown in [10, Prop. 3.1], a spanning monomial b = sαs∗ normal if and only if one of the following holds: β ∈ C∗(E) is (a) α = β, so w = sαs∗ (b) α ≺ β and β ⊖ α is an entry-less cycle; (c) β ≺ α and α ⊖ β is an entry-less cycle. α ∈ GD(E); The set of normal spanning monomials in C∗(E) is denoted by GM(E). Definition 4.8. The abelian core M(E) is the C∗-subalgebra of C∗(E) generated by the set GM(E) of normal spanning monomials. Notations. If b ∈ GM(E) r GD(E) (i.e. b is of either type (b) or (c) above), then b is a normal partial isometry, so its adjoint b∗ also acts as its pseudo-inverse. For this reason, we will denote b∗ simply by b−1. More generally, we will allow arbitrary negative integer exponents, by letting b−m be an alternative notation for b∗m. We will also allow zero exponents, by agreeing that b0 = bb∗ = b∗b, a monomial which in fact belongs to GD(E). (Equivalently, for any b ∈ GM(E) r GD(E), the C∗- subalgebra C∗(b) ⊂ C∗(E) generated by b is a unital abelian C∗-algebra, and b is a unitary element in C∗(b).) Remark 4.9. In general, for a monomial b ∈ GM(E) r GD(E), there might be multiple ways to present it as sαs∗ β, with α and β as in (b) or (c) above, but after careful inspection, one can show that b can be uniquely presented as b = sαsm α = α)m, where α ∈ E∗ is a ray with seed ν and m is some non-zero integer, so if (sαsνs∗ we let bα = sαsνs∗ α (recall that ν is uniquely determined by α), then we can present ν s∗ GM(E) r GD(E) = {bm α : α ray, m non-zero integer}. Clearly, using our exponent conventions, GM(E) r GD(E) is closed under taking adjoints, because (bm α . As it turns out, GM(E) ∪ {0} is an abelian ∗- semigroup; besides the product rules (32) for GD(E), the remaining rules which α )∗ = b−m 20 DANNY CRYTSER, GABRIEL NAGY involve the monomials in GM(E) r GD(E) are: (33) (34) (35) b0 α = pα, for all rays α; bm α pβ = pβbm α1 bm2 bm1 α2 = bm2 α2 bm1 if β ≺ ξα otherwise α , 0, α =(bm α1 =(bm1+m2 α1 0, , if α1 = α2 otherwise By the above ∗-semigroup property, M(E) ⊂ C∗(E) is an abelian C∗-subalgebra which contains D(E), and it can also be described as M(E) = span GM(E). Fur- thermore, the images of D(E) and M(E) under the path representation agree; that is, πpath(M(E)) = Dpath(E). In general, M(E) is much larger than D(E); in fact, M(E) = D(E)′, the commutant of D(E) in C∗(E). As was the case with the diagonal, we have M(E) = (cid:2)Pv∈E0 M(E)pv(cid:3)− the summand M(E)pv now presented as span ({bm α : m ∈ Z, α ∈ E∗ ip, r(α) = v} ∪ {pα : α ∈ E∗, r(α) = v}) , , with continuous functions with compact support. so, upon identifying M(E) ≃ C0(cid:0)\M(E)(cid:1), the (non-norm-closed) algebraic sum M(E)fin = Pv∈E0 M(E)pv is naturally identified with Cc(cid:0)\M(E)(cid:1), the algebra of define the twisted representation Θ : C∗(E) → C(cid:0)T, Apath(E)(cid:1) by Definition 4.10. (Twisted path representation.) With the notation as above, Θ(a)(z) = πpath(γz(a)) z ∈ T, a ∈ C∗(E). For any pair (z, x) ∈ T × E≤∞, we define the state ωz,x on C∗(E) by ωz,x(a) = hδxΘ(a)(z)δxi. Remark 4.11. As πpath is injective on D(E), the gauge-invariant uniqueness the- orem implies that Θ is injective. (The gauge action on the codomain is by trans- lation: (λz(f ))(w) = f (z−1w).) In particular, Θ yields an injection of M(E) into C(T, Dpath(E)). Therefore the spectrum of M(E) can be recovered as a quotient of the spectrum of C(T, Dpath(E)) (that is, T × E≤∞), by the natural equivalence relation implemented by Θ. Specifically, if (z, x) ∈ T × E≤∞, then the restriction ωz,xM(E) is a pure state on M(E). The equivalence relation ∼ on T × E≤∞ is simply given by: (36) (z1, x1) ∼ (z2, x2) ⇔ ωz1,x1M(E) = ωz2,x2M(E). Since the restrictions of these states on the diagonal act as ωz,xD(E) = evD x , it is fairly obvious that (z1, x1) ∼ (z2, x2) implies x1 = x2. The precise description of the equivalence classes (z, x)∼ = {(z1, x1) ∈ T × E≤∞ : (z1, x1) ∼ (z, x)} goes as follows. (37) (z, x)∼ =(zUper(x) × {x}, T × {x}, if x ∈ E∞ ip if x ∈ E≤∞ r E∞ ip (For any integer n ≥ 1, the symbol Un denotes the group of nth roots of unity.) Lemma 4.12. Let E be a directed graph. TRACES ARISING FROM REGULAR INCLUSIONS 21 (i) When we equip the quotient space T × E≤∞/ ∼ with the quotient topology, the map (z, x)∼ 7−→ ωz,xM(E) is a homeomorphism of onto the spectrum of M(E). (ii) For every ray α, if we regard pα as a continuous function on \M(E), then pα is the characteristic function of a compact-open subset Tα, which is homeomorphic to T. Specifically, if ν is the seed of α, and x = αν∞ ∈ E∞ ip is the associated periodic path, then Tα = {(z, x)∼}z∈T and the map T/Uν ∋ zUν 7−→ (z, x)∼ ∈ Tα is a homeomorphism. Alternatively, Tα is naturally identified with the spectrum -- computed in the unital C∗-algebra C∗(bα) -- of the normal partial isometry bα = sαsν s∗ α. (iii) The compact-open sets (Tα)α∈E ∗ ip are mutually disjoint. When we consider Tα, and fix a positive Radon measure µ on \M(E) with cor- Ωip = Sα∈E ∗ ip responding positive linear functional φµ on M(E)fin = Cc(\M(E)), then (38) ZΩip f dµ = Xα∈E ∗ ip φ(f pα) for all f ∈ M(E)fin = Cc(\M(E)). (cid:3) Proof. Parts (i) and (ii) are established in [10] and [2]. For part (iii) we only need to justify the first statement, because the rest follows from the Lebesgue dominated convergence theorem. This follows immediately from the observation that any two distinct rays α1, α2 are incomparable, so by (30) the projections pα1 and pα2 are orthogonal, thus the sets {Tα}α ray form a countable disjoint compact-open cover of Ωip. (cid:3) Remark 4.13. Both D(E) and M(E) are abelian regular C∗-subalgebras in C∗(E), since all generators pv, v ∈ E0 and se, e ∈ E1, normalize both of them. It is shown in [10] that M(E) is in fact a Cartan subalgebra of C∗(E), with its (unique) conditional expectation acting on generators as (39) EM(sαs∗ β) =(sαs∗ 0, β, β ∈ GM(E) if sαs∗ otherwise Within this framework, Theorem 1.7 has the following consequence. Corollary 4.14. For a state φ on M(E), the following conditions are equivalent: (i) The composition φ ◦ EM is a tracial state on C∗(E). (ii) φ is se-invariant for all e ∈ E1. (iii) φ is fully invariant. (cid:3) Remark 4.15. In general, D(E) is not Cartan, and there may exist more than one conditional expectation onto it. One expectation -- hereafter referred to as the Haar expectation -- always exists, defined as ED(a) =ZT γz (EM(a)) dm(z) =ZT EM (γz(a)) dm(z). (Here m denotes the normalized Lebesgue measure on T; the second equality follows from (39), which clearly implies that EM is gauge invariant.) The Haar expectation 22 DANNY CRYTSER, GABRIEL NAGY acts on the spanning monomials as: (40) ED(sαs∗ β) =(pα, 0, if α = β otherwise that ED is faithful. Since the integration mapRT γz(a) dm(z) is always a faithful positive map, it follows Using formulas (40) it is easy to see that ED is also normalized by all pv, v ∈ E0, and se, s∗ e, e ∈ E1, so we also have the following analogue of Corollary 4.14. Corollary 4.16. For a state ψ on D(E), the following conditions are equivalent: (i) The composition ψ ◦ ED is a tracial state on C∗(E). (ii) φ is se-invariant for all e ∈ E1. (iii) φ is fully invariant. (cid:3) Remark 4.17. Either using Corollary 4.16 or directly from the definition, it follows that any fully invariant state ψ on D(E) satisfies (41) ∀ α ∈ E∗ : ψ(pα) = ψ(ps(α)). In particular, a fully invariant state on D(E) is completely determined by its values on the projections pv, v ∈ E0. Definition 4.18. Let E be a directed graph. A graph trace on E is a function g : E0 → [0, ∞) such that: (a) for any v ∈ E0, g(v) ≥Pe:r(e)=v g(s(e)); (b) for any regular v, we have equality in (a). Note that, for any graph trace g, its null space Ng = {v ∈ E0 : g(v) = 0} is a saturated hereditary set. if kgk1 < ∞, or infinite, otherwise. Depending on the quantity kgk1 =Pv∈E0 g(v), a graph trace g is declared finite, We denote the set of all graph traces on E by T (E), and the set of finite graph traces on E by Tfin(E). Lastly, we define the set T1(E) = {g ∈ T (E) : kgk1 = 1}, the elements of which are termed normalized graph traces. Theorem 4.19. A map g : E0 → [0, ∞) is a graph trace on E, if and only if, every finite tuple Ξ = (ξi, λi)n i∈I ⊂ R × E∗ satisfies (42) Pi∈I ξipλi ≥ 0 ⇒ Pi∈I ξig(s(λi)) ≥ 0. Proof. To prove the "if" implication, assume g satisfies condition (42) and let us verify conditions (a) and (b) from Definition 4.18. To check condition (a), start off by fixing some v ∈ E0, and notice that, since for every finite set F ⊂ r−1(v), we In order to check (b), simply notice that, if v is regular (so r−1(v) is both finite and non-empty), the by the Cuntz-Krieger relations, we have an equality pv = have pv ≥ Pe∈F pe (by the Cuntz-Krieger relations), then by (42), it follows that g(v) ≥ Pe∈F g(s(e)); this clearly implies the inequality g(v) ≥ Pe∈r−1(v) g(s(e)). Pe∈r−1(v) pe, so applying (42) both ways (writing the equality as two inequalities), we clearly get g(v) =Pe∈r−1(v) g(s(e)). To prove the "only if " implication, we fix a graph trace g and we prove the implication (42). As a matter of terminology, if a tuple Ξ satisfies the inequality (43) Pi∈I ξipλi ≥ 0, TRACES ARISING FROM REGULAR INCLUSIONS 23 we will call Ξ admissible. Our proof will use induction on the number hΞi = I +Pi∈I λi. If hΞi = 1, then I = 1, thus I is a singleton {i0} and λi0 is a path of length 0, i.e. a vertex v ∈ E0; in this case, (42) is same as the implication "ξpv ≥ 0 ⇒ ξg(v) ≥ 0," which is trivial, since g takes non-negative values. Assume (42) holds whenever hΞi < N , for some N > 1, and show that (42) holds when hΞi = N . Fix an admissible tuple ξ with hΞi = N (so (43) is satisfied), and let us prove the inequality If we consider the set W = {r(λi) : i ∈ I}, then we can split (disjointly) I = (44) Pi∈I ξj g(s(λi)) ≥ 0, Sv∈W Iv, where Iv = {i : r(λi) = v} and we will have Pi∈I ξjg(s(λi)) =Pv∈W Pi∈Iv ξig(s(λi)), with each tuple Ξv = (ξi, λi)i∈Iv admissible. (This is obtained by multiplying the inequality (43) by pv.) In the case when W has at least two vertices, we have hΞvi < hΞi, ∀ v ∈ W , so the inductive hypothesis can be used, and the desired conclusion follows. Based on the above argument, for the remainder of the proof we can assume that W is a singleton, so we have a vertex v ∈ E0, such that r(λi) = v, ∀ i ∈ I. Split I = I 0 ∪ I +, where I 0 = {i ∈ I : λi = 0} and I + = {i ∈ I : λi > 0}. Since W is a singleton, the set I 0 consists of all I for which λi = v. The case when I + = ∅ is trivial, because that would mean that all λi will be equal to v, so for the remainder of the proof we are going to assume that I + 6= ∅. With this set-up the hypothesis (43) reads and the desired conclusion (44) reads: ≥ 0, (cid:0)Pi∈I 0 ξi(cid:1)pv +Pi∈I + ξipλi (cid:0)Pi∈I 0 ξi(cid:1)g(v) +Pi∈I + ξig(s(λi)) ≥ 0. Since I + is non-empty (and finite), we can find a finite non-empty set F ⊂ E1 (In the case when I 0 = ∅, we let Pi∈I 0 ξi = 0.) which allows us to split I + as a disjoint union of non-empty sets I + = Se∈F Ie, the element q =Pe∈E ses∗ where Ie = {i ∈ I : λi ≻ e}. Using the Cuntz-Krieger relations, it follows that e ∈ D is a projection satisfying q ≤ pv, so the difference In either case, it follows that = 0, ∀ i ∈ I +, so when we multiply (45) by q′ we obtain: s∗ λi q′ = pv − q is also a (possibly zero) projection. q′sλi (47) Likewise multiplying (45) by each ses∗ (cid:0)Pi∈I 0 ξi(cid:1)q′ ≥ 0. e we obtain (45) (46) (48) so if we multiply on the left by s∗ e and on the right by se, we obtain: For each e ∈ F , we can form the tuple Ξe = (ξi, λi)i∈I 0∪Ie by letting (cid:0)Pi∈I 0 ξi(cid:1)ses∗ e +Pi∈Ie (cid:0)Pi∈I 0 ξi(cid:1)ps(e) +Pi∈Ie ξj sλi s∗ λi ≥ 0, ξisλi⊖es∗ λi⊖e ≥ 0. λi =(s(e), λi ⊖ e, if i ∈ I 0 if j ∈ Ie 24 DANNY CRYTSER, GABRIEL NAGY and then (48) shows that all Ξe are admissible. Since we obviously have hΞei < hΞi, ξig(s(λi ⊖ e)) ≥ 0, by the inductive hypothesis we obtain(cid:0)Pi∈I 0 ξj(cid:1)g(s(e))+Pi∈Ie which combined with the obvious equality s(λi ⊖ e) = s(λi) yields: We we sum all these inequalities (over e ∈ E), we obtain: (cid:0)Pi∈I 0 ξj(cid:1)g(s(e)) +Pi∈Ie ξj g(s(λi)) ≥ 0. (49) (50) Comparing this inequality with the desired conclusion (46), we see that it suffices to show that (cid:0)Pi∈I 0 ξi(cid:1)(cid:0)Pe∈F g(s(e))(cid:1) +Pi∈I + ξig(s(λi)) ≥ 0. (cid:0)Pi∈I 0 ξi(cid:1)g(v) ≥(cid:0)Pi∈I 0 ξi(cid:1)(cid:0)Pe∈F g(s(e))(cid:1). (51) The case when I 0 = ∅ is trivial, since both sides will equal zero, so for the remain- e, it follows that v is regular and F = r−1(v), so by condition (ii) in the graph trace der, we can assume I 0 6= ∅. In the case when q′ = 0, that is, when pv =Pe∈F ses∗ definition, it follows that g(v) =Pe∈F g(s(e)) and again (51) becomes an equality. which yields g(v) ≥Pe∈F g(s(e)); this means that desired inequality would follow once we prove thatPi∈I 0 ξi ≥ 0, an inequality which is now (under the assumption Lastly, in the case when q′ 6= 0, we use condition (i) in the graph trace definition, that q′ is a non-zero projection) a consequence of (47). (cid:3) In preparation Proposition 4.22 below, which contains two easy applications of Theorem 4.19, we introduce the following terminology. Definition 4.20. A vertex v ∈ E0 is said to be essentially left infinite, if there exists an infinite set X ⊂ E∗ of mutually incomparable paths such that s(α) = v for all α ∈ X. Remark 4.21. One particular class of essentially left infinite vertices are those that emit entries into cycles, i.e. vertices v that have some path α = e1e2 . . . em of positive length, with s(α) = v, such that e1 is an entry to a cycle. Indeed, if e1 enters a cycle ν, then all paths νnα, n ∈ N, are mutually incomparable. Another class of essentially left infinite vertices are those that emit paths to infinitely many vertices. (In [19], such vertices are called left infinite.) The following result generalizes [12, Lemma 3.3(i)] and part of the proof of [19, Theorem 3.2]. Proposition 4.22. Let E be a directed graph, g be a graph trace on E, and v ∈ E0 be some vertex. Assume either one of the hypotheses below is satisfied (a) v emits an entry to a cycle; or (b) g is finite and v is essentially left infinite. Then g(v) = 0. Proof. The main ingredient in the proof is the observation that, for any finite set F of mutually incomparable paths starting at v, one has the inequality (52) g(w) ≥ F · g(v). Xw∈r(F ) Indeed, if we list F as {α1, . . . , αn} (with all α's distinct, i.e. n = F ), then by j=1 pαj , and then mutual incomparability, we have the inequality Pw∈r(F ) pw ≥Pn (52) follows immediately from Theorem 4.19. TRACES ARISING FROM REGULAR INCLUSIONS 25 By assumption, in either case, we can find an infinite set Y ⊂ E∗ of mutually incomparable paths starting at v, such that the sum M = Pw∈r(Y ) g(w) is finite. (In case (a), as seen in the preceding remark, we can ensure that r(Y ) is a sin- gleton; case (b) is trivial, by finiteness of g.) The desired conclusion now follows immediately from (52), which implies M ≥ n · g(v) for arbitrarily large n. (cid:3) Comment. As we will see shortly, graph traces on E correspond to certain maps on the "compactly supported" diagonal subalgebra D(E)fin =SV ∈Pfin(E0) D(E)qV , which will eventually yield tracial positive functionals on the dense ∗-subalgebra C∗(E)fin ⊂ C∗(E). Although neither D(E)fin, nor M(E)fin, nor C∗(E)fin, are C∗-algebras, they are nevertheless unions of increasing nets of unital C∗-algebras: D(E)fin = SV ∈Pfin(E0) D(E)qV , M(E)fin = SV ∈Pfin(E0) M(E)qV , and C∗(E)fin = SV ∈Pfin(E0) qV C∗(E)qV . (Recall that, for any finite subset V ⊂ E0, the projection qV is defined to be Pv∈V pv.) It is clear that the conditional expectations EM and ED map C∗(E)fin onto M(E)fin and D(E)fin, respectively, so Corollaries 4.14 and 4.16 have suitable statements applicable to C∗(E)fin, with the word "state" replaced by "positive linear functional." By definition, positivity for linear func- tionals defined on each one of these ∗-algebras is equivalent to the positivity of their restrictions to each of the cut-off algebras corresponding to V ∈ Pfin(E0). Upon identifying D(E)fin = Cc(\D(E)) and M(E)fin = Cc(\M(E)), the positive cones D(E)+ fin correspond precisely to the non-negative continuous compactly supported functions. fin and M(E)+ With this set-up in mind, Theorem 4.19 has the following consequence. Theorem 4.23. For any graph trace g on E, there exists a unique positive linear functional η = ηg : D(E)fin → C, such that (53) When restricted to the unital C∗-algebras D(E)qV , V ∈ Pfin(E0), the positive linear functionals ηg, g ∈ T (E), have norms: ηg(pλ) = g(s(λ)), ∀ λ ∈ E∗. (cid:13)(cid:13)(cid:13) ηgD(E)qV(cid:13)(cid:13)(cid:13) = Xv∈V g(v). In particular, for g ∈ T (E), the functional ηg is norm-continuous, if and only if g is finite, and in this case, one has kηgk = kgk1. Proof. Let A be the complex span of {pλ}λ∈E ∗, and let Ah be its Hermitean part, which is the same as the real span of {pλ}λ∈E ∗. An application of The- orem 4.19 shows that there is a unique R-linear functional θ : Ah → R with θ(pλ) = g(s(λ)) for all λ ∈ E∗. If we fix V ∈ Pfin(E0) and x ∈ AhqV , another application of Theorem 4.19 to the inequality −xqV ≤ x ≤ xqV shows that θ(x) ≤ θ(qV )x. Thus for each V ∈ Pfin(E0), there is a unique C-linear her- mitean functional ηV : D(E)qV → C with ηV = ηV (qV ), so that ηV is in fact positive with norm equal to Pv∈V g(v). Clearly if V ⊂ W are both finite subsets of E0, then ηW D(E)qV = ηV ; thus, by density, there exists a unique positive linear functional ηg defined on all of D(E) such that ηgD(E)qV = ηV if V ∈ Pfin(E0). (cid:3) Comment. As a ∗-subalgebra in C∗(E)fin, both D(E)fin and M(E)fin are non- degenerate (since they both contain {qV }V ∈Pfin(E), as well as regular, because they 26 DANNY CRYTSER, GABRIEL NAGY are normalized by all se, e ∈ E1 and all pv, v ∈ E0. Given a positive linear functional η on either one of these algebras, it then makes sense to define what it means for it to be se-invariant. Remark 4.24. The map g 7−→ ηg establishes a affine bijective correspondence between T (E) and the space of positive linear functionals on D(E)fin that are se- invariant for all e ∈ E1. The inverse of this correspondence is obtained as follows. Given a linear positive functional θ on D(E)fin which is se-invariant, for all e ∈ E1, the associated graph trace is simply the map (54) gθ : E0 ∋ v 7−→ θ(pv) ∈ [0, ∞). When we specialize to the case of interest to us, Theorem 4.23 yields the following statement. Theorem 4.25. For any normalized graph trace g, there exists a unique state ψg ∈ S(D(E)) satisfying (55) ψg(pλ) = g(s(λ)), ∀ λ ∈ E∗. All states ψg, g ∈ T1(E) are fully invariant, and furthermore, the correspondence (56) T1(E) ∋ g 7−→ ψg ∈ Sinv(D(E)) is an affine bijection, which has as its inverse the correspondence (57) Sinv(D(E)) ∋ θ 7−→ gθ ∈ T1(E) defined as in (54). (cid:3) Comment. Using Corollary 4.16, it follows that for any g ∈ T1(E), the composition χg = ψg ◦ ED defines a tracial state on C∗(E); this way we obtain an injective correspondence (58) T1(E) ∋ g 7−→ χg ∈ T (C∗(E)). Of course, any tracial state τ ∈ T (C∗(E)) becomes invariant, when restricted to D(E), so using (57) we obtain a correspondence (59) T (C∗(E)) ∋ τ 7−→ gτ ∈ T1(E). Theorem 4.25 shows that this map is surjective, because the correspondence (58) is clearly an affine right inverse for (59). The surjectivity of (59) is also proved in [17], by completely different means. Remark 4.26. Using formulas (40), given a normalized graph trace g ∈ T1(E), the associated tracial state χg = ψg ◦ ED -- hereafter referred to as the Haar trace induced by g -- acts on the spanning monomials as: (60) χg(sαs∗ β) =(g(s(α)), 0, if α = β otherwise Among other things, the above formulas prove that χg is in fact gauge invariant, i.e. χg ◦ γz = χg, for all z ∈ T. Conversely, every gauge invariant tracial state τ ∈ T (C∗(E)) arises this way. Indeed, if τ is such a trace, then by gauge invariance it follows that, whenever TRACES ARISING FROM REGULAR INCLUSIONS 27 α, β ∈ E∗ are such that α 6= β, we must have τ (sαs∗ α = β, then β) = 0; furthermore, if τ (sαs∗ β) = τ (s∗ βsα) =(τ (0) = 0, τ (s∗ αsα) = τ (ps(α)), otherwise if α 6= β so in all cases we get τ (sαs∗ β) = χgτ (sαs∗ β). To summarize: • the range of the injective correspondence (58) is the set T (C∗(E))T of gauge invariant tracial states; • when restricting the correpondence (59) to T (C∗(E))T, one obtains an affine isomorphism (61) T (C∗(E))T ∋ τ 7−→ gτ ∈ T1(E). When searching for an analogue of Theorem 4.25, with D(E) replaced by M(E), it is obvious that the space T (E) is not sufficient, so additional structure needs to be added to it. Definition 4.27. The cyclic support of a function g : E0 → C is defined to be the set suppcg = {v ∈ E0 : v ∈ E0 cyclic, g(v) 6= 0}. (Recall that a cyclic vertex v is one visited by a simple entry-less cycle. Equivalently, v is a ray of length zero.) A cyclically tagged graph trace consists of a pair (g, µ), where g is a graph trace and a map µ : suppcg ∋ v 7−→ µv ∈ Prob(T) -- hereafter referred to as the tag. Note that our definition includes the possibility of an empty tag in the case when suppcg = ∅. (More on this in Theorem 4.41 below.) The space of all such pairs will be denoted by T ct(E). The adjective "finite," "infinite," or "normalized," is attached to (g, µ) precisely when it applies to g. Using this terminology, one has the following extension of Theorem 4.23. Theorem 4.28. For any cyclically tagged graph trace (g, µ) on E, there exists a unique positive linear functional η = η(g,µ) : M(E)fin → C, such that (i) η(g,µ)(pλ) = g(s(λ)), for every finite path λ ∈ E∗; (ii) for any ray α and any integer m 6= 0, η(g,µ)(bm α ) =(g(s(α))RT zm dµs(α)(z), 0, if g(s(α)) 6= 0, otherwise When restricted to the unital C∗-algebras M(E)qV , V ∈ Pfin(E0), the positive linear functionals η(g,µ), (g, µ) ∈ T ct(E), have norms: (cid:13)(cid:13)(cid:13) η(g,µ)M(E)qV(cid:13)(cid:13)(cid:13) = Xv∈V g(v). In particular, for any (g, µ) ∈ T ct(E), the functional η(g,µ) is norm-bounded if and only if g is finite, and in this case, one has kη(g,µ)k = kgk1. Proof. Assume (g, µ) ∈ T ct(E) is fixed throughout the entire proof. Fix for the moment some a ray α with g(s(α)) 6= 0, and consider the C∗-subalgebra C∗(bα) ⊂ M(E). (Recall that, if ν is the seed of the ray α, then bα is the normal partial isometry sαsνs∗ α.) As pointed out in Lemma 4.12, using the fact that the projection 28 DANNY CRYTSER, GABRIEL NAGY α = pα is the characteristic function of the compact-open set Tα ⊂ \M(E), we have b0 of course the equality M(E)pα = C∗(bα), so using the surjective ∗-homomorphism πα : M(E) ∋ a 7−→ apα ∈ C∗(bα) ∼−−→ C(T), we can define a state ωα on M(E) by ωα(a) =ZT πα(a) dµs(α). Specifically, if we write the compression apα as a f (bα), for some f ∈ C(T), then ωα(a) = RT f (z) dµs(α)(z). Using the product rules (32), (34) and (35), it follows that on the generator set GM(E), the state ωα acts as (62) ωα(pλ) =(1, if λ ≺ ξα; 0, otherwise; ωα(bm α1 ) =(RT zm dµs(α)(z), 0, if α1 = α; otherwise. Define now the functional θ : M(E)fin → C by (63) θ(a) = Xα∈E ∗ ip g(s(α))6=0 g(s(α))ωα(a), a ∈ M(E)fin. Concerning the point-wise convergence of the sum in (63), as well as its positivity, they are a consequence of the following fact. Claim. For any vertex v ∈ E0, one has the inequality (64) In particular, the sum (65) g(s(α)) ≤ g(v). Xα∈E ∗ ip r(α)=v θv = Xα∈E ∗ ip r(α)=v g(s(α))ωαM(E)pv is a norm-convergent sum, thus θv is a positive linear functional on M(E)pv with norm (66) kθvk = Xα∈E ∗ ip r(α)=v g(s(α)). The inequality (64) follows from the observation that, for any finite set F of rays with range v, the projections {pα}α∈F satisfy the inequality Pα∈F pα ≤ pv, which by Theorem 4.19 impliesPα∈F g(s(α)) ≤ g(v). The equality (66) is now clear from the positivity of θv, which combined with (62) yields: kθvk = θv(pv) = Xα∈E ∗ ip r(α)=v g(s(α))ωα(pv) = Xα∈E ∗ ip r(α)=v g(s(α)). Using the Claim, we see that θ given in (63) is indeed correctly defined, positive and it can alternatively be presented as θ(a) =Pv∈E0 θv(a) (a sum which has only TRACES ARISING FROM REGULAR INCLUSIONS 29 finitely many non-zero terms for each a ∈ M(E)fin). By construction, θ acts on the generator set GM(E) as: (67) (68) ip λ≺ξα g(s(α)), λ ∈ E∗ θ(pλ) = Xα∈E ∗ α ) =(g(s(α))RT zm dµs(α)(z), θ(bm 0, if α ∈ E∗ otherwise ip and g(s(α)) 6= 0 Next we consider the positive linear functional ηg : D(E)fin → C associated to g, as constructed in Theorem 4.23, and the linear positive functional ηg ◦ED : M(E)fin → (Here we use the fact that ED maps C∗(E)fin onto D(E)fin.) Using Riesz' C. Theorem, there is a positive Radon measure υ on \M(E), such that ηg (ED(f )) = f dυ, for all f ∈ Cc(\M(E)) = M(E)fin. Using this measure, we now define R\M(E) the desired positive linear functional η on Cc(\M(E)) = M(E)fin by: (69) (70) η(f ) = θ(f ) +Z\M(E)rΩip f dυ = = θ(f ) + ηg (ED(f )) − Xα∈E ∗ = θ(f ) + ηg (ED(f )) − Xα∈E ∗ ip ip ηg (ED(f pα)) = ηg (ED(f )pα) . (The equality (69) follows from Lemma 4.12.) To check condition (i), start with some λ ∈ E∗ and observe that, for all rays α, we have the equalities pλpα =(pα, 0, if λ ≺ ξα otherwise which by (67) imply that Xα∈E ∗ ip ηg (ED(pλpα)) = Xα∈E ∗ ip λ≺ξα ηg(pα) = Xα∈E ∗ ip λ≺ξα g(s(α)) = θ(pλ), so by (69) we obtain the desired property η(pλ) = ηg(pλ) = g(s(λ)). In order to check condition (ii), we simply verify that, for any ray α and any integer m, we have the equality (71) η(bm α ) = θ(bm α ). α) = η(pα) = g(s(α)) = θ(b0 The case when m = 0 we have b0 η(b0 ED vanishes on G(E) r GD(E) -- by (40) -- we have ED(bm trivial using (70). α = pα, so by condition (i) and (68), we have α). In the case when m 6= 0, we notice that since α ) = 0, and then (71) is The remaining statements in the Theorem (including the uniqueness of η) are pretty clear, since any positive linear functional η satisfying conditions (i) and (ii) must satisfy ηD(E)fin = ηg, from which the continuity of the restrictions ηM(E)qV follows immediately. (cid:3) 30 DANNY CRYTSER, GABRIEL NAGY One aspect not addressed so far is invariance of the states η. For this purpose, the following definition is well-suited. Definition 4.29. Two cyclic vertices are said to be equivalent if they are visited by the same entry-less cycle. A cyclically tagged graph trace (g, µ) is said to be consistent if µv = µv′ whenever v and w are equivalent. (Note that if two cyclic vertices v, w are equivalent, then g(v) = g(w).) The space of all consistent cyclically tagged traces on E is denoted by T cct(E). As agreed earlier, the adjective "finite," "infinite," or "normalized," is attached to an element (g, µ) ∈ T cct(E), precisely when it applies to g. In particular, the space of normalized consistent cyclically tagged graph traces on E is denoted by T cct (E). 1 Proposition 4.30. A cyclically tagged graph trace (g, µ) is consistent if and only if the associated positive functional η(g,µ) : M(E)fin → C constructed in Theorem 4.28 is se-invariant for all e ∈ E1. Proof. Assume (g, µ) is consistent, and let us show the invariance of η(g,µ), which amounts to checking, that for each e ∈ E1, we have: (i) η(g,µ)(sepλs∗ α s∗ (ii) η(g,µ)(sebm e) = η(g,µ)(pepλ), ∀ λ ∈ E∗; α ), ∀ α ∈ E∗ e) = η(g,µ)(pebm ip, m ∈ Z. Property (i) is obvious, since η(g,µ) agrees with the se-invariant functional ηg on D(E)fin. As for condition (ii), we only need to verify it if s(e) = r(α) (other- wise both sides are zero). Also notice that if α > 0, then eα is also a ray with s(eα) = s(α), which satisfies sebm eα, so by condition (ii) in The- orem 4.28, we have η(g,µ)(sebm e) = η(g,µ)(bm α ). In the remaining case, α = 0, so α reduces to a vertex v = r(ν), for some simple entry-less cycle ν. If e is not an edge in ν, then it is a ray, thus the preceding argument still applies (we will have sebm e ). If e is an edge on ν, then sebm r(e), with r(e) obviously equivalent to v, and the desired equality -- which now reads η(g,µ)(bm v ) -- follows from the equalities g(v) = g(r(e)) and µv = µr(e). eα) = g(s(eα))RT zm dµs(eα)(z) = g(s(α))RT zm dµs(α)(z) = η(g,µ)(bm r(e)) = η(g,µ)(bm e = bm e = bm v s∗ v s∗ e = bm α s∗ α s∗ Conversely, notice first that, if η(g,µ) is se-invariant, for all e ∈ E1, then it will also satisfy the identity (72) η(g,µ)(sλas∗ λ) = η(g,µ)(pλa), ∀ λ ∈ E∗, a ∈ M(E)fin. Secondly, observe that, if v, v′ are equivalent cyclic vertices, presented as v = s(ν) and v′ = s(ν′) for two simple entry-less cycles, then we can write ν = αβ and ν′ = βα for two suitably chosen paths α, β ∈ E∗. This clearly implies that bv′ = sβbvs∗ β, which also yields bm v′ = sβbm v s∗ β, ∀ m ∈ Z. Combining these two observations with condition (ii) from Theorem 4.28, it follows that, if η(g,µ) is invariant, then for any two equivalent cyclic vertices v and v′ we have (with α, β as above): ZT zm dµv′ (z) = η(g,µ)(bm v′ ) = η(g,µ)(sβbm v s∗ β) = η(g,µ)(pβbm v ) = = η(g,µ)(bm v ) =ZT zm dµv(z), ∀ m ∈ Z, which clearly implies µv′ = µv. (cid:3) TRACES ARISING FROM REGULAR INCLUSIONS 31 Remark 4.31. The map (g, µ) 7−→ η(g,µ) establishes a affine bijective correspon- dence between T cct(E) and the space of positive linear functionals on M(E)fin that are se-invariant for all e ∈ E1. The inverse of this correspondence is the map θ 7−→ (gθ, µθ) defined as follows. Given a linear positive functional θ on M(E)fin which is se-invariant, for all e ∈ E1, the graph trace gθ is given by (54), and the tag µθ = (µθ v)v∈suppcgθ is given (implicitly) by (73) f (z) dµθ v(z) = θ(f (bv)) gθ(v) ZT , ∀ v ∈ suppcgθ, f ∈ C(T). When we specialize to states, we now have the following extension of Theo- rem 4.25. 1 Theorem 4.32. For any normalized consistent cyclically tagged graph trace (g, µ) ∈ T cct (E), there exists a unique state φ(g,µ) ∈ S(M(E)) satisfying (i) φ(g,µ)(pλ) = g(s(λ)), for every finite path λ ∈ E∗; (ii) for any ray α and any integer m: φ(g,µ)(bm α ) =(g(s(α))RT zm dµs(α)(z), 0, if g(s(α)) 6= 0, otherwise All states φ(g,µ), (g, µ) ∈ T cct spondence 1 (E) are fully invariant, and furthermore, the corre- (74) T cct 1 (E) ∋ (g, µ) 7−→ φ(g,µ) ∈ Sinv(M(E)) is an affine bijection, which has as its inverse the correspondence (75) Sinv(M(E)) ∋ θ 7−→ (gθ, µθ) ∈ T cct 1 (E) defined as in (54) and (73). (cid:3) Comment. Using Corollary 4.14, it follows that for any (g, µ) ∈ T cct (E), the com- position τ(g,µ) = φ(g,µ) ◦ EM defines a tracial state on C∗(E); this way we obtain an injective correspondence 1 (76) T cct 1 (E) ∋ (g, µ) 7−→ τ(g,µ) ∈ T (C∗(E)). Of course, any tracial state τ ∈ T (C∗(E)) becomes invariant, when restricted to M(E), so using (75) we obtain a correspondence (77) T (C∗(E)) ∋ τ 7−→ (gτ , µτ ) ∈ T cct 1 (E). Theorem 4.32 shows that this map is surjective, because the correspondence (76) is clearly an affine right inverse for (77). Remark 4.33. The range of (76) clearly contains the range of (58), which equals T (C∗(E))T. After all, any trace g ∈ T1(E) can be tagged using the constant map µ : suppcg → Prob(T) that takes µv to be the Haar measure for every v, and it is straightforward to verify that for this particular tagging one, has τ(g,µ) = χg. Concerning the range of (76), one legitimate question is whether it equals the whole tracial state space T (C∗(E)). Using the bijection (74), this question is equiv- alent to the surjectivy of the map (78) Sinv(M(E)) ∋ φ 7−→ φ ◦ EM ∈ T (C∗(E)). 32 DANNY CRYTSER, GABRIEL NAGY As we have seen in Corollary 3.6, a sufficient condition for the surjectivity of (78) is the condition that the inclusion M(E) ⊂ C∗(E) has the (honest) extension property. As it turns out, this issue can be neatly described using the graph. Theorem 4.34. The inclusion M(E) ⊂ C∗(E) has the extension property, if and only if no cycle in E has an entry. Proof. To prove the "if" implication, assume that no cycle in E has an entry, fix a pure state ω on M(E), and let φ be an extension of ω to C∗(E). In order to prove uniqueness of φ, it suffices to show that the value of φ on a standard generator sαs∗ β is independent of the choice of φ. By assumption, there is a x ∈ E≤∞ and z ∈ T such that ω = ωz,x as in Lemma 4.12. On the one hand, by Fact 3.1 and the observation that ω(pγ) = 1 for all γ ≺ x, it follows that β) = φ(pγ sαs∗ φ(sαs∗ βpγ). (79) ∀ γ ≺ x : On the other hand, using the results from [10, Section 3], it follows that there is γ ≺ x such that pγsαs∗ βpγ belongs to M(E). (In the language of [10], x must be essentially aperiodic by our assumption on E.) Using (79) it follows that φ(sαs∗ β) = ω(pγsαs∗ βpγ), and the desired conclusion follows. For the "only if" direction, we show that if there is a cycle ν ∈ E∗ that has an entry, then we can construct a pure state on M(E) which has multiple extensions to states on C∗(E). Consider the path x = ν∞ ∈ E∞ formed by following ν infinitely many times. For each z ∈ T consider the state ωz,x ∈ S(C∗(E)) introduced in Definition 4.10, given by As explained in Remark 4.11, since x 6∈ E∞ ip , it follows that: ωz,x(a) = hδxπpath(γz(a))δxi. (z, x) ∼ (1, x), ∀ z ∈ T, which by Lemma 4.12 means that all restrictions ωz,xM(E), z ∈ T, coincide, so they are all equal to the pure state ϑ ∈ \M(E) corresponding to the equivalence class (1, x)∼ = T × {x}. However, as states on C∗(E), the functionals ωz,x, z ∈ T cannot all be equal, since for example we have ωz,x(ν) = zν, ∀ z ∈ T. (cid:3) Definition 4.35. A graph E is tight, if every cycle is entry-less. Combining Theorem 4.34 with Corollary 3.6 and Theorem 4.32 we now obtain the following statement. Theorem 4.36. If E is tight, then the correspondence (76) is an affine isomor- phism between the space T cct (cid:3) (E) and the tracial state space T (C∗(E)). 1 Remark 4.37. Tight graphs are interesting in other respects: they are the only graphs that yield finite, stably finite, quasi-diagonal, or AF-embeddable C∗-algebras ([16]), as well as the only graphs that yield graph algebras with stable rank one ([6]). A graph which yields a C∗-algebra with Hausdorff spectrum must be tight, although this is not sufficient [4, Ex. 10]. In the remainder of this paper we aim to parametrize the entire tracial state space T (C∗(E)) for arbitrary graphs by employing Theorem 4.36 in conjunction with certain procedures that replace the graph E with a tight sub-graph E′, in such a way that the tracial state spaces T (C∗(E)) and T (C∗(E′)) coincide. Since TRACES ARISING FROM REGULAR INCLUSIONS 33 the sub-graphs that are best suited for analyzing how the trace spaces change are the canonical ones, the following terminology is all we need. Definition 4.38. If E is a directed graph, a tightening of E is a canonical sub- graph, i.e. one that can be presented as E \ H, for some saturated hereditary subset H ⊂ E0, in such a way that (a) E \ H is tight, and (b) the canonical ∗-homomorphism ρH : C∗(E) → C∗(E \ H) implements a bijective correspondence: T (C∗(E \ H)) ∋ τ 7−→ τ ◦ ρH ∈ T (C∗(E)) Since ρH is always surjective, the correspondence from (b) is always injective, so the only requirement in our definition is its surjectivity. When it comes to parametrizing tracial states on graph C∗-algebras, the most useful and natural tightening is as follows. Example 4.39. Let E be a graph, and let C = CE be the set of vertices which emit entrances into cycles. The set C is obviously hereditary, but not saturated in general, so we need to take its saturation C. As it turns out, E \ C constitutes a tightening of E. First of all, since passing from E to E \ C clearly removes all entries into the cycles in E, it is clear that E \ C is tight. Secondly, in order to justify the surjectivity of T (C∗(E \ C)) ∋ τ 7−→ τ ◦ ρH ∈ T (C∗(E)), (80) all we must show is the fact that all tracial states on C∗(E) vanish on ker ρC , for which it suffices to prove the inclusion H ⊂ Ng, which in itself is a consequence of Proposition 4.22. The sub-graph constructed in the above Example is called the minimal tighten- ing, and is denoted by Etight. The canonical ∗-homomorphism will be denoted by ρtight : C∗(E) → C∗(Etight). Combining this construction with Theorem 4.36 we now obtain. Theorem 4.40. For any directed graph E, the map T cct 1 (Etight) ∋ (g, µ) 7−→ τ(g,µ) ◦ ρtight ∈ T (C∗(E)) is an affine isomorphism. (cid:3) The final result in this paper deals with a graph-theoretic characterization of automatic gauge invariance for tracial states, which as pointed out in Remark 4.26 is equivalent to the surjectivity of the map (58). In [19], it is shown that this feature is implied by condition (K). However, as Theorem 4.41 below shown, this is not necessary. Theorem 4.41. For a directed graph E, the following conditions are equivalent: (i) all tracial states on C∗(E) are gauge invariant; (ii) the source of each cycle in E is essentially left infinite. Proof. (i) ⇒ (ii): Suppose that λ = e1 . . . em is a cycle such that v = s(λ) = r(e1) is not essentially left infinite; we show how to construct a tracial state on C∗(E) which is not gauge-invariant. Note that as v is not essentially infinite, in particular it does not emit an entrance to any cycle; therefore, none of the edges in λ will be removed when forming Etight, and so we can assume that E is tight. (Since the canonical quotient π : C∗(E) → C∗(Etight) is equivariant for the respective 34 DANNY CRYTSER, GABRIEL NAGY gauge actions, a non-gauge invariant tracial state on C∗(E)tight) will give rise to a non-gauge invariant trace on C∗(E).) Say that a path µ ∈ E∗ is acyclic if it cannot be written as µ = ανβ for α, β ∈ E∗ and ν a cycle. Let A denote the set of all acyclic paths with source v; note that any two paths in A are incomparable, and so A must be finite because v is not essentially left infinite. For w ∈ E0 let g(w) = A ∩ r−1(w); it is straightforward to verify that g is a finite graph trace with g(v) = 1 which we can normalize to obtain g′ ∈ T1(E). Note that the cyclic support of g′ is precisely r({e1, . . . , em}) (as v is not essentially left infinite, it emits no entrances to cycles). Now we can take any z ∈ T \ Uλ and let µs(ei) = δz for all i = 1, . . . , m. The affiliated tracial state τ(g,µ) ∈ T (C∗(E)) will satisfy τ(g,µ)(bλ) = g(s(λ))zλ 6= 0 so that in particular τ(g,µ) is not gauge-invariant. (ii) ⇒ (i): Suppose that the source of each cycle is essentially left infinite. Any finite graph trace must vanish on an essentially left infinite vertex as in Proposition 4.22; hence if every source of every cycle is essentially left infinite, then there are no vertices in the cyclic support of any graph trace, and so there are no taggings to consider. Thus every tracial state on C∗(Etight) is gauge-invariant, which shows that every tracial state on C∗(E) is gauge-invariant. (cid:3) Comment. Besidese the minimal tightening Etight introduced in this paper, other tightenings could naturally be considered. The same arguments as those used in Example 4.39 can be used with C replaced by another hereditary subset H ⊂ E0, as long as: (a) the canonical sub-graph E \ H is tight, and (b) one has the inclusion H ⊂ Ng, for all g ∈ T1(E). One way to ensure (a) is to take H to contain CE. As far as condition (b) is concerned, we could use Proposition 4.22 as a guide. In particular, we can consider the set L = LE of all essentially left infinite vertices. Since LE is potentially much larger than CE, the resulting subgraph E \ LE will potentially be considerably smaller than Etight (and thus easier to analyze regarding graph traces). References 1. R.J. Archbold, J.W. Bunce, and K.D. Gregson. Extensions of states of C ∗-algebras, ii. Proc. Edinb. Math. Soc., 92A:113 -- 122, 1982. 2. J. Brown, G. Nagy, and S. Reznikoff. A generalized cuntz -- krieger uniqueness theorem for higher-rank graphs. J. Funct. Anal., 266:2590 -- 2609, 2013. 3. J. Brown, G. Nagy, S. Reznikoff, A. Sims, and D. Williams. Cartan Subalgebras in C ∗-Algebras of Hausdorff Etale Groupoids. arxiv:1503.03521. 4. G. Goehle. Groupoid C ∗-algebras with Hausdorff spectrum. Bull. Aust. Math. Soc., 88:232 -- 242, 2013. 5. A. an Huef and I. Raeburn. The ideal structure of cuntzkrieger algebras. Ergodic Theory Dynam. Systems, 17:611624, 1997. 6. J. Jeong, G. Park, and D. Shin. Stable rank and real rank of graph C ∗-algebras. Pac. J. Math, 200(2):331 -- 343, 2001. 7. R. Kadison and I. Singer. Extensions of pure states. Amer. Jour. Math, 81(2):383 -- 400, 1960. 8. A. Kumjian. On C ∗-Diagonals. Can. J. Math., 38(4):969 -- 1008, 1986. 9. E. C. Lance. Hilbert C ∗-modules: A toolkit for operator algebraists, volume 210 of London Math Soc. Lecture Note Series. Cambridge Univ. Press, 1994. TRACES ARISING FROM REGULAR INCLUSIONS 35 10. G. Nagy and S. Reznikoff. Abelian core of graph algebras. J. London Math. Society, 3:889 -- 908, 2012. 11. G. Nagy and S. Reznikoff. Pseudo-diagonals and uniqueness theorems. Proc. Amer. Math. Soc., 142(1):263 -- 275, January 2014. 12. D. Pask and A. Rennie. The noncommutative geometry of graph C ∗-algebras I: The index theorem. J. Funct. Anal., 233:92 -- 134, 2006. 13. I. Raeburn. Graph Algebras. CBMS Lecture Notes. American Mathematical Society, 2005. 14. J. Renault. A Groupoid Approach to C ∗-Algebras. Lecture Notes in Mathematics. Springer Verlag, 1980. 15. J. Renault. Cartan Subalgebras in C ∗-Algebras. Irish Math. Soc. Bulletin, 61:29 -- 63, 2008. 16. C. Schafhauser. Af-embeddings of Graph Algebras. arXiv:1405.7757. 17. M. Tomforde. The ordered K0-group of a graph C ∗-algebra. C.R. Math. Acad. Sci. Soc., 25:19 -- 25, 2003. 18. M. Tomforde. A unified approach to Exel-Laca algebras and C ∗-algebras associated to graphs. J. Operator Theory, 50:345 -- 368, 2003. 19. M. Tomforde. Stability of C ∗-algebras affiliated to graphs. Proc. Amer. Math. Soc., 132:1787 -- 1795, 2004.
1612.04088	3	1612	2017-12-31T14:02:56	Rigidity theory for $C^*$-dynamical systems and the "Pedersen Rigidity Problem"	[ "math.OA" ]	Let $G$ be a locally compact abelian group. By modifying a theorem of Pedersen, it follows that actions of $G$ on $C^*$-algebras $A$ and $B$ are outer conjugate if and only if there is an isomorphism of the crossed products that is equivariant for the dual actions and preserves the images of $A$ and $B$ in the multiplier algebras of the crossed products. The rigidity problem discussed in this paper deals with the necessity of the last condition concerning the images of $A$ and $B$. There is an alternative formulation of the problem: an action of the dual group $\hat G$ together with a suitably equivariant unitary homomorphism of $G$ give rise to a generalized fixed-point algebra via Landstad's theorem, and a problem related to the above is to produce an action of $\hat G$ and two such equivariant unitary homomorphisms of $G$ that give distinct generalized fixed-point algebras. We present several situations where the condition on the images of $A$ and $B$ is redundant, and where having distinct generalized fixed-point algebras is impossible. For example, if $G$ is discrete, this will be the case for all actions of $G$.	math.OA	math	RIGIDITY THEORY FOR C∗-DYNAMICAL SYSTEMS AND THE "PEDERSEN RIGIDITY PROBLEM" S. KALISZEWSKI, TRON OMLAND, AND JOHN QUIGG Abstract. Let G be a locally compact abelian group. By mod- ifying a theorem of Pedersen, it follows that actions of G on C∗- algebras A and B are outer conjugate if and only if there is an isomorphism of the crossed products that is equivariant for the dual actions and preserves the images of A and B in the multiplier algebras of the crossed products. The rigidity problem discussed in this paper deals with the necessity of the last condition concerning the images of A and B. of the dual group (cid:98)G together with a suitably equivariant unitary produce an action of (cid:98)G and two such equivariant unitary homo- homomorphism of G give rise to a generalized fixed-point algebra via Landstad's theorem, and a problem related to the above is to There is an alternative formulation of the problem: an action morphisms of G that give distinct generalized fixed-point algebras. We present several situations where the condition on the images of A and B is redundant, and where having distinct generalized fixed-point algebras is impossible. For example, if G is discrete, this will be the case for all actions of G. 1. Introduction When presented with a C∗-dynamical system, which we call simply an action, one of our first impulses is to form the crossed product C∗- algebra. A fundamental question arises: how do we recover the action from the crossed product? The short answer is: we cannot the crossed- product C∗-algebra is not enough information. So, the next questions are: (1) what extra data do we need to recover the action, and (2) in what sense can the action be recovered? Crossed-product duality is largely devoted to these questions. Date: December 29, 2017. 2010 Mathematics Subject Classification. Primary 46L55. Key words and phrases. action, crossed-product, exterior equivalence, outer con- jugacy, generalized fixed-point algebra. The second author is funded by the Research Council of Norway through FRINATEK, project no. 240913. 1 2 KALISZEWSKI, OMLAND, AND QUIGG We consider actions of a fixed group G, and we want to recover (A, α). If we are only given A (cid:111)α G, we cannot even recover the C∗-algebra A up to Morita equivalence. In other words, there exist non-Morita- equivalent C∗-algebras A and B carrying actions of a group G such that A (cid:111) G (cid:39) B (cid:111) G. Instances of this are surprisingly hard to find in the literature, but for example [KOQ14, Remark 4.3] shows how to do it with A and B commutative and G discrete abelian. We will assume throughout that G is abelian. Most of the following is true for nonabelian G, but this involves coactions, which would tend to obscure the heart of the matter. If G is an abelian group, A is a C∗-algebra, and α : G → Aut A is an action, then there is an action (cid:98)α, called the dual action, of the dual group (cid:98)G on the crossed product A(cid:111)αG, and the (Takesaki-)Takai theorem says that the crossed product by the dual action is isomorphic to A⊗K(L2(G)). Moreover, the double dual action corresponds to the tensor product of α with conjugation by the right regular representation ρ of G. Thus, Takai duality recovers the original action (A, α) of G up to Morita equivalence from the dual action (A (cid:111)α G,(cid:98)α) of (cid:98)G. To recover more about α, we need more information about the crossed product. Raeburn's abstract definition of the crossed product, using universal properties, gives a covariant homomorphism (iA, iG) of (A, G) in the multiplier algebra M (A (cid:111)α G) such that every other covariant homomorphism (π, U ) in M (B) factors through a homomor- phism π × U : A (cid:111) G → M (B). Landstad duality says that the action (A, α) can be recovered up to conjugacy (equivariant isomorphism) from the data (A (cid:111)α G,(cid:98)α, iG). Abstractly, we are given an equivariant action (C, ζ, V ) of (cid:98)G, i.e., an action (C, ζ) of (cid:98)G and a strictly contin- and the action αG of (cid:98)G on C∗(G) determined by αG γ ∈ (cid:98)G, s ∈ G where we freely identify elements of G with unitaries uous unitary homomorphism V : G → M (C) that is equivariant for ζ γ (s) = γ(s)s for in M (C∗(G)). The rough idea is to construct iA(A) as a generalized fixed-point algebra C ζ,V of the equivariant action. In [KOQ16] we explored a duality intermediate between Takai and Landstad duality, that we called outer duality, but which perhaps de- serves to be christened "Pedersen duality". This intermediate dual- ity is based upon Pedersen's theorem: two actions α and β of G on A are exterior equivalent if and only if there is an isomorphism A. Escap- ing from the single C∗-algebra A, an alternative version of Pedersen's theorem (see Theorem 2.1 below) says that two actions (A, α) and (B, β) of G are outer conjugate if and only if there is an isomorphism (cid:39)−→ (A (cid:111)β G,(cid:98)β) such that Θ ◦ iα Θ : (A (cid:111)α G,(cid:98)α) A = iβ Θ : (A (cid:111)α G,(cid:98)α) RIGIDITY FOR C∗-DYNAMICAL SYSTEMS (cid:39)−→ (B (cid:111)β G,(cid:98)β) such that Θ(iA(A)) = iB(B). Thus, 3 Pedersen shows how to recover the action (A, α) up to outer conju- gacy if we know the dual action (C, ζ) and the generalized fixed-point algebra C ζ,V , but perhaps not the equivariant homomorphism V itself. After proving the outer duality theorem, we naturally wanted to give examples exploring the boundaries of Pedersen's theorem. More pre- cisely, we searched for examples of two actions (A, α) and (B, β) of G such that the dual actions(cid:98)α and (cid:98)β are conjugate, but α and β are not outer conjugate. Equivalently, we want there to be an isomorphism between the dual actions, but not one that preserves the generalized fixed-point algebras. Somehow surprisingly, we were not able to pro- duce any examples of this phenomenon, and moreover, we discovered a complete absence of such examples in the literature. This is strik- ing, since it is tempting to conjecture that one of the first questions researchers must have asked about crossed products is, how much in- formation do we get from only knowing the dual actions? It seems to us that this investigation is long overdue. Thus we are led to what we call the Pedersen rigidity problem: either find explicit examples of two non-outer-conjugate actions with conju- gate dual actions, or prove that this cannot happen. Such a theorem would (at least for abelian groups) result in a significant simplification of Pedersen's theorem, namely removing the clause about the general- ized fixed-point algebras; we suspect that counterexamples do in fact exist. Our ultimate goal is a definitive answer to this question, but so far we have made only partial progress, namely we have found a number of "no-go theorems": the phenomenon of non-outer-conjugate actions with conjugate dual actions cannot occur if (1) G is discrete, (2) A and B are stable, (3) A and B are commutative, (4) α or β is inner, (5) G is compact and α and β are faithful and ergodic, or (6) A is covered by invariant ideals on which the action is strongly Pedersen rigid in the sense that the dual action has only one generalized fixed- point algebra. In light of these no-go theorems, it is apparent that the phenomenon of multiple generalized fixed-point algebras is delicate. We begin in Section 2 by recording our conventions regarding ac- tions, crossed products, outer conjugacy, and Pedersen's theorem(s). In subsequent sections we prove the no-go theorems. The final such theorem includes the case of locally unitary actions on continuous-trace C∗-algebras, and we explain in a remark that this context was in fact the germ of the idea leading to the last no-go theorem. Some of this research was done during a visit of the third author to the University of Oslo, and he thanks Erik B´edos, Nadia Larsen, and 4 KALISZEWSKI, OMLAND, AND QUIGG Tron Omland for their hospitality. Other parts of the work were done during the second author's visit to Arizona State University, and he is grateful to the analysis group at ASU for their hospitality. We thank the referee for comments that improved the quality of this paper. 2. Preliminaries Throughout, G will be a locally compact abelian group, A a C∗- algebra, and α : G → Aut A an action of G on A. Since G will be fixed, we just say that (A, α) is an action. We adopt the conventions of [EKQR06, Appendix A] for actions and crossed products, and here we recall the basic notation and results we will need. If (A, α) is an action and B is a C∗-algebra, a covariant homomorphism of (A, α) in M (B) is a pair (π, U ), where π : A → M (B) is a nondegenerate homomorphism and U : G → M (B) is a strictly continuous unitary homomorphism such that π ◦ αs(a) = Ad Us ◦ π(a) = Usπ(a)U∗ s for all s ∈ G, a ∈ A. The crossed product of (A, α) is a triple (A (cid:111)α G, iA, iG), where A (cid:111)α G is a C∗-algebra and (iA, iG) is a covariant homomorphism of (A, α) in M (A (cid:111)α G) with the universal property that for any covariant homo- morphism (π, U ) of (A, α) in M (B) there is a unique nondegenerate homomorphism π × U : A (cid:111)α G → M (B), called the integrated form of (π, U ), such that (π × U ) ◦ iA = π and (π × U ) ◦ iG = U. A, iα Sometimes we write (iα G) if ambiguity is possible, and on the other hand we sometimes write A (cid:111) G if α is understood. By definition, the crossed product is unique up to isomorphism in the sense that if (B, π, U ) is another crossed product then there is a unique isomorphism θ : A (cid:111)α G (cid:39)−→ B such that θ ◦ iA = π and θ ◦ iG = U . Given an action (A, α), because G is abelian there is a unique action (cid:98)α of the dual group (cid:98)G on A (cid:111)α G such that (cid:98)αγ ◦ iA = iA and (cid:98)αγ ◦ iG(s) = γ(s)iG(s) for all s ∈ G, γ ∈ (cid:98)G. A conjugacy, or isomorphism, between two actions (A, α) and (B, β) (cid:39)−→ B that is α−β equivariant in the sense that is an isomorphism φ : A φ◦αs = βs◦φ for all s ∈ G, and α and β are conjugate, or isomorphic, if such a φ exists. Given a conjugacy φ, the pair (iB ◦ φ, iβ G) is a covariant homomorphism of (A, α), and the integrated form φ(cid:111) G is a conjugacy (A (cid:111)α G,(cid:98)α) (cid:39)−→ (B (cid:111)β G,(cid:98)β). RIGIDITY FOR C∗-DYNAMICAL SYSTEMS 5 (Takesaki-)Takai duality [Tak75, Theorem 3.4] says that the double dual action (cid:0)A (cid:111)α G (cid:111)(cid:98)α G,(cid:98)(cid:98)α(cid:1) is conjugate to (A ⊗ K(L2(G)), id ⊗ Ad ρ) (the part about the double dual action is not stated in [Tak75], but appears in [Ped79, Theo- rem 7.9.3], for example). Two actions (A, α) and (B, β) of G are Morita equivalent if there exist an A − B imprimitivity bimodule X and an α − β compatible action u of G on X, i.e., for all s ∈ G and x, y ∈ X we have A(cid:104)x, y(cid:105)(cid:1) = A(cid:104)us(x), us(y)(cid:105) (cid:0) (cid:0)(cid:104)x, y(cid:105)B (cid:1) = (cid:104)us(x), us(y)(cid:105)B. αs βs By [RW98, Remark 7.3], the above properties imply that u is also compatible with the bimodule structure, i.e., us(ax) = αs(a)ux(x) and us(xb) = us(x)βs(b) this Morita equivalence is proved in [Com84, Section 6, Theorem] and for all a ∈ A, x ∈ X, b ∈ B. If α and β are Morita equivalent, then so are the dual actions (cid:98)α and (cid:98)β; for the crossed-product C∗-algebras, [CMW84, Theorem 1], and the existence of an(cid:98)α−(cid:98)β compatible action duality that the double dual action (A(cid:111)αG(cid:111)(cid:98)αG,(cid:98)(cid:98)α) is Morita equivalent on the crossed-product imprimitivity bimodule follows from [EKQR06, Proposition 3.5] (see also [Kus08, Lemma 3.3]). It follows from Takai to (A, α). An inner action of G on A is one of the form Ad u, where u : G → In this case M (A) is a strictly continuous unitary homomorphism. there is a conjugacy (A (cid:111)α G,(cid:98)α) (cid:39) (A ⊗ C∗(G), idA ⊗ αG), where αG is the unique action of (cid:98)G on C∗(G) such that αG for all γ ∈ (cid:98)G, s ∈ G. Note that the above terminology is somewhat γ (s) = γ(s)s inconsistent within the literature. What we call an inner action is sometimes called a "unitary action". Moreover, we will still use the notion "locally unitary action" later on, e.g. in Corollary 4.12. Landstad duality [Lan79, Theorem 3] (stated for abelian groups in [Ped82, Theorem 28], and in somewhat more detail in [KOQ16, The- orem 2.2]) says that, for a C∗-algebra C, there exist an action (A, α) (cid:39)−→ C if and only if there ex- of G and an isomorphism θ : A (cid:111)α G ous unitary homomorphism V : G → M (C). Moreover, given such a ist an action ζ of (cid:98)G on C and a αG − ζ equivariant strictly continu- triple (C, ζ, V ), which we call an equivariant action of (cid:98)G, the action 6 KALISZEWSKI, OMLAND, AND QUIGG (A, α) and the isomorphism θ can be chosen such that θ is (cid:98)α − ζ equi- (cid:39)−→ C is a (cid:98)β − ζ equivariant isomorphism such that variant and θ ◦ iG = V ; with such a choice, if (B, β) is any action and σ : V (cid:111)β G (cid:39)−→ (B, β) such σ ◦ iG = V , then there exists a conjugacy ϕ : (A, α) that σ◦ (ϕ (cid:111) G) = θ. In fact, we can take A to be the C∗-subalgebra of M (C) consisting of all multipliers m satisfying Landstad's conditions (1) ζγ(m) = m for all γ ∈ (cid:98)G; (2) mV (f ), V (f )m ∈ C for all f ∈ Cc(G); (3) s (cid:55)→ Ad Vs(m) is norm continuous from G to C, ant actions (C, ζ, V ) and (D, ε, W ) of (cid:98)G, if Θ : (C, ζ) and we can let α be the restriction of (the extension to M (C) of) the inner action Ad V . Then, letting ι : A → M (C) be the inclusion map, the pair (ι, V ) is a universal covariant homomorphism of the action (A, Ad V ). We write C ζ,V = A, and call it the generalized fixed-point algebra of the equivariant action (C, ζ, V ). Note that, given equivari- (cid:39)−→ (D, ε) is a conjugacy such that Θ ◦ V = W , then Θ(C ζ,V ) = Dε,W . In [KOQ16, Corollary 2.6] we recorded a routine consequence of Landstad dual- ity that we sometimes find useful (which in the following rendition we translate into the context of abelian groups): if (C, ζ, V ) is an equivari- ant action of (cid:98)G and ϕ : A → C ζ,V is an isomorphism, then there exist an action α of G on A and a conjugacy Θ : (A (cid:111)α G,(cid:98)α) (cid:39)−→ (C, ζ) such that Θ ◦ iG = V and Θ ◦ iA = ϕ. A cocycle for an action (A, α) is a strictly continuous unitary map u : G → M (A) such that ust = usαs(ut) for all s, t ∈ G. In this case Ad u◦ α is also an action of G on A, which is called exterior equivalent to α. In particular, a strictly continuous unitary homomorphism is a cocycle for the trivial action ι, and an inner action is exterior equiv- alent to ι. Two actions (A, α) and (B, β) are outer conjugate if β is conjugate to an action on A that is exterior equivalent to α. If α and β are outer conjugate, then the dual actions (cid:98)α and (cid:98)β are conjugate. Also, outer conjugacy is stronger than Morita equivalence (strictly so, because (A, α) and (B, β) outer conjugate implies A (cid:39) B). Pedersen's theorem [Ped82, Theorem 35] (stated more precisely in [RR88, Theorem 0.10], and generalized to nonabelian groups in [KOQ16, Theorem 3.1]) says that two actions α and β of G on A are exte- (cid:39)−→ A. We take this opportunity to record rior equivalent if and only if there is a conjugacy Θ : (A (cid:111)α G,(cid:98)α) (A (cid:111)β G,(cid:98)β) such that Θ ◦ iα A = iβ RIGIDITY FOR C∗-DYNAMICAL SYSTEMS 7 an alternative version, which seems not to be explicitly recorded in the literature, but which is an immediate consequence of [KOQ16, proof of Theorem 5.9]: Theorem 2.1 (Pedersen's theorem). Two actions (A, α) and (B, β) of G are outer conjugate if and only if there is an isomorphism Θ : (A (cid:111)α G,(cid:98)α) (cid:39)−→ (B (cid:111)β G,(cid:98)β) such that (2.1) Θ(iA(A)) = iB(B). Proof. The forward direction is an obvious corollary of Pedersen's the- orem, so suppose that Θ is an isomorphism satisfying (2.1). Let V = Θ ◦ iα G. Then (B (cid:111)β G,(cid:98)β, V ) is an equivariant action of (cid:98)G, with (B (cid:111)β G)(cid:98)β,V = Θ(cid:0)(A (cid:111)α G)(cid:98)α,iα Ψ : (B (cid:111)ζ G,(cid:98)ζ) G(cid:1) = Θ(iA(A)) = iβ (cid:39)−→ (B (cid:111)β G,(cid:98)β) and iβ are an action ζ of G on B and a conjugacy B(B) is an isomorphism, so by Landstad duality there B : B → iβ B(B), such that Ψ ◦ iζ G = V and Ψ ◦ iζ B = iβ B. Thus by Pedersen's theorem the actions β and ζ are exterior equivalent. On the other hand, we have a conjugacy Ψ−1 ◦ Θ : (A (cid:111)α G,(cid:98)α) (cid:39)−→ (B (cid:111)ζ G,(cid:98)ζ), taking iα conjugate. Therefore α and β are outer conjugate. G, so again by Landstad duality the actions α and ζ are (cid:3) G to iζ Definition 2.2. We will refer to (2.1) as Pedersen's condition. Thus, Pedersen's theorem says that two actions of G are outer con- jugate if and only if there is a conjugacy between the dual actions satisfying Pedersen's condition. It is useful to compare Pedersen's theorem (Theorem 2.1) to Land- it follows from the latter that two actions (A, α) and stad duality: (B, β) of G are conjugate if and only if there is an isomorphism Θ : (A (cid:111)α G,(cid:98)α) (cid:39)−→ (B (cid:111)β G,(cid:98)β) such that Θ ◦ iα G = iβ G. 8 KALISZEWSKI, OMLAND, AND QUIGG 3. The Pedersen rigidity problem The alternative form of Pedersen's theorem (Theorem 2.1) leads to The Pedersen rigidity problem. Does there exist an example of two non-outer-conjugate actions α and β of G such that the dual actions (cid:98)α and (cid:98)β are conjugate? Equivalently, is Pedersen's condition Θ(iA(A)) = iB(B) in Theorem 2.1 redundant? We have only found a few references to this question in the literature; for example, Buss and Echterhoff say in [BE15, Remark 3.13 (e)] that it "is not clear to us" whether "there might exist two different structures" giving "different generalized fixed-point algebras". First, we remark the following: suppose that we have an isomorphism (cid:39)−→ B and two actions β and γ on B that are not exterior ϕ : A equivalent. Then define an action α on A by setting αg(a) = ϕ−1(βg(ϕ(a))). Then (A, α) and (B, γ) are not outer conjugate. In other words, there are two ways to produce non-outer-conjugate actions; either via non- isomorphic C∗-algebras, or via non-exterior-equivalent actions. In the following sections we present various no-go theorems, giving general conditions under which Pedersen's condition is redundant. Our no-go theorems seem to come in two flavors, which we characterize via the following definitions: Definition 3.1. An action (C, ζ) of (cid:98)G is strongly fixed-point rigid if it has a unique generalized fixed-point algebra, i.e., for any two (cid:98)G- equivariant strictly continuous unitary homomorphisms V, W : G → M (C) we have C ζ,V = C ζ,W . An action (A, α) of G is strongly Pedersen rigid if its dual action is strongly fixed-point rigid. Definition 3.2. An action (C, ζ) of (cid:98)G is fixed-point rigid if the au- fixed-point algebras, i.e., for any two (cid:98)G-equivariant strictly continuous tomorphism group of (C, ζ) acts transitively on the set of generalized unitary homomorphisms V, W : G → M (C) there is an automorphism Θ of (C, ζ) such that Θ(C ζ,V ) = C ζ,W . An action (A, α) of G is Pedersen rigid if its dual action is fixed-point rigid. the following properties: RIGIDITY FOR C∗-DYNAMICAL SYSTEMS Lemma 3.3. Let (C, ζ, V ) be an equivariant action of (cid:98)G, and consider (2) For every equivariant action (D, ε, W ) of (cid:98)G, if Θ : (C, ζ) (4) For every equivariant action (D, ε, W ) of (cid:98)G, if ζ and ε are (D, ε) is a conjugacy then Θ(C ζ,V ) = Dε,W . (1) ζ is strongly fixed-point rigid. (3) ζ is fixed-point rigid. (cid:39)−→ 9 conjugate then there is a conjugacy Θ : (C, ζ) that Θ(C ζ,V ) = Dε,W . Then (1) ⇔ (2) ⇒ (3) ⇔ (4). Proof. Assume (1), and let Θ : (C, ζ) Θ−1◦ W is a (cid:98)G-equivariant strictly continuous unitary homomorphism, (cid:39)−→ (D, ε) be a conjugacy. Then (cid:39)−→ (D, ε) such so = Θ−1(cid:0)Dε,W ), C ζ,V = C ζ,Θ−1◦W (since (C, ζ) is strongly fixed-point rigid) and we have shown (2). Conversely, (2) ⇒ (1) follows by taking (D, ε) = (C, ζ) and Θ = idC. (cid:3) (1) ⇒ (3) is trivial, and (3) ⇔ (4) is similar to (1) ⇔ (2). (1) α is strongly Pedersen rigid. (cid:39)−→ (B (cid:111)β G,(cid:98)β) Corollary 3.4. Let (A, α) be an action of G, and consider the following properties: is a conjugacy then Θ(iA(A)) = iB(B). (3) α is Pedersen rigid. (4) If (B, β) is any action of G, then α and β are outer conjugate (2) For every action (B, β) of G, if Θ : (A(cid:111)α G,(cid:98)α) if and only if (cid:98)α and (cid:98)β are conjugate. Then (1) ⇔ (2) ⇒ (3) ⇔ (4). Proof. Recall that iA(A) = (A (cid:111)α G)(cid:98)α,iα action (B, β) of G, if (cid:98)α and (cid:98)β are conjugate then there is a conjugacy (cid:39)−→ (B (cid:111)β G,(cid:98)β) such that Θ(iA(A)) = iB(B). Now Θ : (A (cid:111)α G,(cid:98)α) apply Lemma 3.3 with (C, ζ, V ) = (A(cid:111)αG,(cid:98)α, iα if (D, ε, W ) is any equivariant action of (cid:98)G, then by Landstad duality (cid:39)−→ (B(cid:111)βG,(cid:98)β) G, and similarly for iB(B). By Pedersen's theorem, condition (4) is equivalent to the following: for any there is an action (B, β) of G and a conjugacy Ψ : (D, ε) such that Ψ ◦ W = iβ G. G), keeping in mind that (cid:3) Thus, the Pedersen rigidity problem is equivalent to the following: 10 KALISZEWSKI, OMLAND, AND QUIGG The Pedersen rigidity problem (alternative formulation). Is every action of G Pedersen rigid? But now with the stronger type of rigidity, we can ask for more: The strong Pedersen rigidity problem. Is every action of G strongly Pedersen rigid? In fact, our no-go theorems seem to hint that this stronger rigidity might hold. In the next section, we will discuss the rigidity problem when restrict- ing to certain types of groups and actions, and therefore we introduce the following: Definition 3.5. We say that a class C of actions of G is Pedersen rigid if two actions in C are outer conjugate if and only if their dual actions are conjugate. Remark 3.6. Suppose we are given an action (C, ζ) of (cid:98)G. Let us write H(G, C, ζ) for the set of all (cid:98)G-equivariant strictly continuous unitary homomorphisms V : G → M (C). Note that the group Aut(C, ζ) acts on this set by composition: (θ, V ) (cid:55)→ θ ◦ V : Aut(C, ζ) × H(G, C, ζ) → H(G, C, ζ). Now let us write GF P A(C, ζ) for the set of all generalized fixed-point algebras of the action (C, ζ). Then we have a surjection V (cid:55)→ C ζ,V : H(G, C, ζ) → GF P A(C, ζ), and the action of Aut(C, ζ) descends to an action on GF P A(C, ζ). By definition, the action (C, ζ) is fixed-point rigid if and only if this action on GF P A(C, ζ) is transitive. This leads us to consider a stronger property: can the action of Aut(C, ζ) on H(G, C, ζ) be transitive? We will explain here that the answer is generally negative. For every V ∈ H(G, C, ζ), by Landstad duality there are an action (A, α) of G and a conjugacy Θ : (A (cid:111)α G,(cid:98)α) (A, α) of G for which (A (cid:111)α G,(cid:98)α, iα W ∈ H(G, C, ζ) and another action (B, β) of G with (B (cid:111)β G,(cid:98)β, iβ In other words, the elements V ∈ H(G, C, ζ) correspond to actions G) = (C, ζ, V ). If we take another G) = (C, ζ, W ), then it follows from Landstad duality that (A, α) (cid:39) (B, β) if and only if there exists θ ∈ Aut(C, ζ) such that θ ◦ V = W . We can (cid:39)−→ (C, ζ) such that Θ ◦ iα G = V. RIGIDITY FOR C∗-DYNAMICAL SYSTEMS 11 arrange for (B, β) to be outer conjugate, but not conjugate, to (A, α), and then there is no automorphism of (C, ζ) taking V to W . As we mentioned above, our no-go theorems hint at the possibility that every action of G is actually strongly Pedersen rigid. If so, then the set GF P A(C, ζ) would be a singleton, but again we could easily have the action of Aut(C, ζ) on H(G, C, ζ) be nontransitive. arbitrary locally compact group Γ instead of (cid:98)G, but then the homomor- Remark 3.7. The above definition of fixed-point rigidity (as well as strong fixed-point rigidity) can be phrased in terms of actions of an phism V : G → M (C) would have to be replaced by a nondegenerate homomorphism from C0(Γ) to M (C) that is equivariant for ζ and the action of Γ on C0(Γ) given by right translation. Moreover, Theorem 2.1 also holds for non-abelian groups, again by applying [KOQ16, proof of Theorem 5.9]. Hence, the Pedersen rigidity problem can be formulated for arbitrary locally compact groups as well. Remark 3.8. It is interesting to note that the rigidity theory for C∗- dynamical systems discussed in this paper bears resemblance to recent works involving diagonal-preserving isomorphisms between graph C∗- algebras, see [Mat16, Theorem 1.5] and [CR16, Theorem 4.1]. 4.1. Discrete groups. If G is discrete, then (cid:98)G is compact, so the 4. No-go theorems dual action has a genuine fixed-point algebra, and hence all generalized fixed-point algebras coincide. Consequently, we get the following result (which we have not found in the literature): Proposition 4.1 (No-Go Theorem 1). If G is discrete, then every action of G is strongly Pedersen rigid. Thus, by Pedersen's theorem, if G is discrete, then two actions of G are outer conjugate if and only if the dual actions are conjugate. 4.2. Stable C∗-algebras. If (A(cid:111)α G,(cid:98)α) (cid:39) (B (cid:111)β G,(cid:98)β), then by Takai duality the actions α and β must at least be Morita equivalent. Moreover, [Com84, Section 8 Proposition] says that if A and B are stable and have strictly positive elements (which is satisfied if they are separable, for example) then α and β are Morita equivalent if and only if they are outer conjugate. This leads to Proposition 4.2 (No-Go Theorem 2). The class of actions of G on stable C∗-algebras possessing strictly positive elements is Pedersen rigid. 12 KALISZEWSKI, OMLAND, AND QUIGG Thus (assuming, for example, that we restrict our attention to sepa- rable C∗-algebras) we will not find any examples of multiple generalized fixed-point algebras unless at least one of A and B is nonstable. This indicates that the phenomenon of multiple generalized fixed-point al- gebras is delicate in some sense, since it would not be possible, for example, if we replace the original actions by their double duals, since double crossed products are always stable. 4.3. Commutative C∗-algebras. For commutative algebras, we have a stronger version of Proposition 4.2. Proposition 4.3 (No-Go Theorem 3). If A and B are commutative, then actions (A, α) and (B, β) of G are conjugate if and only if the dual actions are conjugate. In particular, the class of actions of G on commutative C∗-algebras is Pedersen rigid. Proof. It suffices to show that if the actions α and β are Morita equiv- alent then they are conjugate. Suppose we have an α − β equivariant A− B imprimitivity bimodule. Then the associated Rieffel homeomor- an α − β equivariant isomorphism A (cid:39) B. (cid:3) Remark 4.4. If we have an action (A, α) and we know that A is com- phism (cid:98)B (cid:39) (cid:98)A [Rie74, Corollary 6.27] is G-equivariant, and this gives mutative, then the dual action (cid:98)α contains all the information about α (up to conjugacy), so weaker forms of equivalence are of interest. In a recent paper by Li [Li15], a notion of continuous orbit equivalence for topological dynamical systems is discussed: Let X and Y be locally compact Hausdorff spaces on which G acts. Then (X, G) and (Y, G) are said to be continuously orbit equivalent if there exists a homeomorphism ϕ : X → Y and continuous maps a : G × X → G and b : G × Y → G such that ϕ(g · x) = a(g, x) · ϕ(x) and ϕ−1(g · y) = b(g, y)· ϕ−1(y). This is clearly weaker than conjugacy (which is obtained by setting a(g, x) = b(g, y) = g). Moreover, if G is discrete and the actions are topologically free, then [Li15, Theorem 1.2] says that (X, G) and (Y, G) are continuously orbit equivalent if and only if there exists an isomorphism Φ : C0(X) (cid:111) G → C0(Y ) (cid:111) G satisfying Pedersen's condition. equivariant actions: The no-go theorems Propositions 4.2 and 4.3 have consequences for Corollary 4.5. Let (C, ζ) be an action of (cid:98)G, and let V, W : G → M (C) be (cid:98)G-equivariant strictly continuous unitary homomorphisms. If both C ζ,V and C ζ,W are stable and have strictly positive elements, or both RIGIDITY FOR C∗-DYNAMICAL SYSTEMS 13 are commutative, then there is a ζ-equivariant automorphism of C that takes C ζ,V to C ζ,W . Proof. By Landstad duality, there exist actions (A, α) and (B, β) of G and conjugacies (A (cid:111)α G,(cid:98)α) θ (cid:39) / / (C, ζ) ψ (cid:39)o (B (cid:111)β G,(cid:98)β) such that θ ◦ iα G = V, Then we have a conjugacy θ(iA(A)) = C ζ,V , ψ ◦ iβ G = W, ψ(iB(B)) = C ζ,W . and hence, by Propositions 4.2 and 4.3, (A, α) and (B, β) are outer conjugate. Thus there exists a conjugacy ψ−1 ◦ θ : (A (cid:111)α G,(cid:98)α) σ : (A (cid:111)α G,(cid:98)α) (cid:39)−→ (B (cid:111)β G,(cid:98)β), (cid:39)−→ (B (cid:111)β G,(cid:98)β) (possibly different from ψ−1 ◦ θ) such that σ(iA(A)) = iB(B). Then ψ ◦ σ ◦ θ−1 is a ζ-equivariant automorphism of C such that ψ ◦ σ ◦ θ−1(C ζ,V ) = ψ ◦ σ(iA(A)) = ψ(iB(B)) = C ζ,W . (cid:3) 4.4. Inner actions. Any inner action α = Ad u determined by a uni- tary homomorphism u : G → M (A) is exterior equivalent to the trivial action ι. Thus we can take the dual actions to be the same: (A (cid:111)α G,(cid:98)α) = (A (cid:111)ι G,(cid:98)ι). G : G → M (A (cid:111)ι G) maps into the center, and The homomorphism iι G : G → M (A (cid:111)ι G). Therefore, by [QR95, hence it commutes with iα Lemma 1.6] (see also [BE15, Proposition 3.12] for a slightly more gen- eral result) the generalized fixed-point algebras coincide: By transitivity, if we are given another action (B, β) such that G = (A (cid:111)ι G)(cid:98)ι,iι (A (cid:111)α G)(cid:98)α,iα (A (cid:111)α G,(cid:98)α) (cid:39) (B (cid:111)β G,(cid:98)β), G. then with a bit more work we will have: Proposition 4.6 (No-Go Theorem 4). Every inner action is strongly Pedersen rigid. o 14 KALISZEWSKI, OMLAND, AND QUIGG Proof. We will apply Lemma 3.4 (1) ⇔ (2). Let (A, α) and (B, β) be actions of G such that α is inner and (cid:98)α is conjugate to (cid:98)β. Up to isomorphism, we can take (A (cid:111)ι G,(cid:98)ι) = (A (cid:111)α G,(cid:98)α) = (B (cid:111)β G,(cid:98)β). Then we have three equivariant unitary homomorphisms iι G, iα G, iβ G : G → M (A (cid:111)ι G). The first one maps into the center, and hence commutes with the other two. Thus (A (cid:111)α G)(cid:98)α,iα G = (A (cid:111)ι G)(cid:98)ι,iι G = (B (cid:111)β G)(cid:98)β,iβ G. G,(cid:98)α) bras: Now we revert back to the original situation, without the "up to iso- morphism" reduction, and we find that any isomorphism Θ : (A (cid:111)α (cid:39)−→ (B (cid:111)β G,(cid:98)β) will preserve the generalized fixed-point alge- Θ(cid:0)(A (cid:111)α G)(cid:98)α,iα G(cid:1) = (B (cid:111)β G)(cid:98)β,iβ G. (cid:3) Remark 4.7. Proposition 4.6 could be generalized by using the full force of [BE15, Proposition 3.12]: given two actions (A, α) and (B, β), if there is an isomorphism Θ : (A (cid:111)α G,(cid:98)α) (cid:39)−→ (B (cid:111)β G,(cid:98)β) such that the two sets of products G(C∗(G)))iβ Θ(iα G(C∗(G)) and iβ G(C∗(G))Θ(iα G(C∗(G))) coincide, then α and β are outer conjugate. 4.5. Ergodic actions of compact groups. Suppose that G is com- pact, A is a unital C∗-algebra, and α is an action of G on A that is ergodic, that is, the fixed-point algebra is Aα = C1. We refer to [OPT80] for the following facts concerning ergodic actions of compact abelian groups. For each γ ∈ (cid:98)G let Aγ be the associated spectral subspace Aγ = {a ∈ A : αs(a) = γ(s)a for all s ∈ G}. We impose the further hypothesis that the action α is faithful (i.e., αs = αt implies s = t), and consequently Aγ (cid:54)= {0} for all γ ∈ (cid:98)G. Then for all γ ∈ (cid:98)G there is a unitary uγ ∈ A such that Moreover, there is a 2-cocycle ω : (cid:98)G × (cid:98)G → T such that Aγ = Cuγ. (4.1) uγuχ = ω(γ, χ)uγχ; RIGIDITY FOR C∗-DYNAMICAL SYSTEMS 15 unitary-valued maps u of G satisfying (4.1) are called ω-representations (see [Wil07, Appendix D.3], for example), and we say that ω is compat- ible with (A, α). Recall that a map ω : (cid:98)G×(cid:98)G → T is called a 2-cocycle if ω(γ, χ)ω(γχ, σ) = ω(γ, χσ)ω(χ, σ) for all γ, χ, σ ∈ (cid:98)G. for all γ ∈ (cid:98)G. We will assume that all our 2-cocycles are normalized, namely ω(1, γ) = ω(γ, 1) = 1 ω(γ, χ) = z(γ)z(χ)z(γχ) for all γ, χ ∈ (cid:98)G. The set Z 2((cid:98)G, T) of all 2-cocycles is an abelian group under pointwise multiplication. A cocycle ω ∈ Z 2((cid:98)G, T) is called a coboundary if there is a map z : (cid:98)G → T such that The set B2((cid:98)G, T) of all coboundaries is a subgroup of Z 2((cid:98)G, T), and the quotient group H 2((cid:98)G, T) := Z 2((cid:98)G, T)/B2((cid:98)G, T) is the second co- homology group of (cid:98)G with values in T. Two 2-cocycles ω and ζ of (cid:98)G are called cohomologous if ωB2((cid:98)G, T) = ζB2((cid:98)G, T), i.e., there is a map z : (cid:98)G → T such that Faithful ergodic actions of G are classified by H 2((cid:98)G, T) in the following for all γ, χ ∈ (cid:98)G. ω(γ, χ) = z(γ)z(χ)z(γχ)ζ(γ, χ) sense: two faithful ergodic actions (A, α) and (B, β) of G, with com- patible 2-cocycles ω and ζ, are conjugate if and only if ω and ζ are cohomologous. cocycle, and let λω be the associated left regular ω-representation of (cid:98)G on (cid:96)2((cid:98)G), given by Let α be a faithful ergodic action of G on A, let ω be an α-compatible for γ, χ ∈ (cid:98)G, ξ ∈ (cid:96)2((cid:98)G). (λω γ ξ)(χ) = ω(γ, γχ)ξ(γχ) Then there is an isomorphism (A (cid:111)α G,(cid:98)α) (cid:39) (K((cid:96)2((cid:98)G)), Ad λω) (see [OPT80, Remark 6.7]). Proposition 4.8 (No-Go Theorem 5). If G is compact, then faithful ergodic actions (A, α) and (B, β) of G are conjugate if and only if the dual actions are conjugate. In particular, the class of faithful ergodic actions of G is Pedersen rigid. Proof. It suffices to prove that if (A, α) and (B, β) are faithful ergodic actions of G such that (A (cid:111)α G,(cid:98)α) (cid:39) (B (cid:111)β G,(cid:98)β), 16 then (A, α) (cid:39) (B, β). Let ω and ζ be cocycles of (cid:98)G compatible with KALISZEWSKI, OMLAND, AND QUIGG α and β, respectively. By Lemma 4.9 below, the cocycles ω and ζ are cohomologous, and hence the actions (A, α) and (B, β) are conjugate. (cid:3) We thank Magnus Landstad for conversations that ultimately led to the above no-go theorem. In the above proof we referred to the following lemma, which we state in abstract form, for a possibly nonabelian discrete group. The result is contained in [Wil07, Proposition D.27], but we give a short proof for convenience. Lemma 4.9. Let ω1 and ω2 be 2-cocycles of a discrete group Γ, and let πi be an ωi-representation for i = 1, 2. Suppose that the actions (K((cid:96)2(Γ)), Ad π1) and (K((cid:96)2(Γ)), Ad π2) are conjugate. Then ω1 and ω2 are cohomologous. Proof. Since all automorphisms of K((cid:96)2(Γ)) are inner, we can choose a unitary u ∈ U((cid:96)2(Γ)) such that π2(g)uxu∗π2(g)∗ = uπ1(g)xπ1(g)∗u∗ for all g ∈ Γ and all x ∈ K((cid:96)2(Γ)). This means that π1(g)∗u∗π2(g)ux = xπ1(g)∗u∗π2(g)u, for all g ∈ Γ and all x ∈ K((cid:96)2(Γ)), hence there are scalars z(g) ∈ C such that for all g ∈ Γ, i.e., such that π1(g)∗u∗π2(g)u = z(g)1 π2(g) = z(g)uπ1(g)u∗ for all g ∈ Γ. Define the ω1-projective representation π(cid:48) uπ1(g)u∗ for all g ∈ Γ. Then π2(g) = z(g)π(cid:48) 1 by π(cid:48) 1(g) for all g ∈ Γ, and 1(g) = ω2(g, h)z(gh)π(cid:48) 1(gh) = ω2(g, h)π2(gh) = π2(g)π2(h) = z(g)z(h)π(cid:48) 1(h) = z(g)z(h)ω1(g, h)π(cid:48) 1(g)π(cid:48) 1(gh). for all g, h ∈ Γ. Hence, ω2(g, h) = z(g)z(h)z(gh)ω1(g, h) for all g, h ∈ Γ, so ω1 and ω2 are cohomologous. (cid:3) RIGIDITY FOR C∗-DYNAMICAL SYSTEMS 17 4.6. Local rigidity. Proposition 4.10 (No-Go Theorem 6). Let (A, α) be an action, and let I be a family of α-invariant ideals of A such that A = spanI. If for each I ∈ I the restricted action αI is strongly Pedersen rigid, then α is strongly Pedersen rigid. Proof. Let (B, β) be an action, and suppose that we have a conjugacy Θ : (A (cid:111)α G,(cid:98)α) (cid:39)−→ (B (cid:111)β G,(cid:98)β). By Lemma 3.4, it suffices to show that Θ(iA(A)) = iB(B). Let I ∈ I, and let αI denote the restricted action on I. Then I (cid:111)αI G is an (cid:98)α- invariant ideal of A (cid:111)α G, so K := Θ(I (cid:111) G) is a (cid:98)β-invariant ideal of B (cid:111)β G. Then by [GL89, Theorem 3.4] (since G is amenable) there is a β-invariant ideal JI of B such that (JI (cid:111) G,(cid:99)βJI ) = (K,(cid:98)βK). Thus, by Lemma 3.4 again, But iI is the restriction of iA to the ideal I, and similarly iJI = iBJI , so we have Θ(iI(I)) = iJI (JI). Θ(iA(I)) = iB(JI). Now, we have a family of β-invariant ideals {JI}I∈I of B, and B (cid:111)β G = Θ(A (cid:111)α G) (cid:19) (cid:18) = Θ = span I∈I = span I∈I I (cid:111) G span I∈I Θ(I (cid:111) G) (since Θ is an isomorphism between C∗-algebras) JI (cid:111) G, Θ(iA(A)) = Θ iA (cid:16) (cid:16) span{I : I ∈ I}(cid:17)(cid:17) = span{Θ(iA(I)) : I ∈ I} = span{iB(JI) : I ∈ I} = iB (span{JI : I ∈ I}) = iB(B). so by Lemma 4.11 below we have B = spanI∈I JI. Therefore, In the above proof we referred to the following general lemma, which is presumably folklore: (cid:3) 18 KALISZEWSKI, OMLAND, AND QUIGG Lemma 4.11. Let (B, β) be an action, and let {JI}I∈I be a family of β-invariant ideals of B such that B (cid:111)β G = span I∈I JI (cid:111) G. Then B = spanI∈I JI. Proof. Put Then J is a β-invariant ideal of B, and J = span I∈I JI. JI (cid:111) G (by [Gre78, Proposition 9]) J (cid:111) G = span I∈I = B (cid:111)β G, and so we must have J = B, by [Gre78, Proposition 11]. (cid:3) We will apply Proposition 4.10 to locally unitary actions. Recall from [RW98, Section 7.5] that an action (A, α) is called locally unitary if A is continuous-trace and for each π ∈ (cid:98)A there exist a compact (cid:98)A is locally compact Hausdorff it is equivalent to require that there neighborhood F of π and a strictly continuous unitary homomorphism u : G → M (AF ) (where AF is the corresponding quotient of A) such that αF = Ad u (where αF denotes the quotient action of G). Since exist an open set N containing π and a strictly continuous unitary homomorphism u : G → M (AN ) (where AN is the corresponding ideal of A) such that α = Ad u on AN . To see this, note that, given such a compact neighborhood F , we could take N to be the interior of F , and then, identifying A with the algebra of continuous sections of the associated C∗-bundle that vanish at infinity, the corresponding ideal AN is the set of sections vanishing outside N , while conversely given an open set N , since (cid:98)A is locally compact Hausdorff we can assume without loss of generality that N has compact closure F , and then the corresponding quotient AF is the set of continuous sections on the restricted C∗-bundle over F . Then by Proposition 4.6 the restricted actions on the AN 's are strongly Pedersen rigid, and A is the closed span of the AN 's. Thus by Proposition 4.10 we have Corollary 4.12. Any locally unitary action on a continuous-trace C∗- algebra is strongly Pedersen rigid. namely that if I is a family of α-invariant ideals such that(cid:84)I = {0} Question 4.13. Is there an analogue of Proposition 4.10 for quotients, and for each I ∈ I the quotient action on A/I is strongly Peder- sen rigid, then α is strongly Pedersen rigid? This would allow us to strengthen Corollary 4.12 to pointwise unitary actions. RIGIDITY FOR C∗-DYNAMICAL SYSTEMS 19 Remark 4.14. Our original motivation for studying the rigidity question for locally (or pointwise) unitary actions was [PR84, Proposition 2.5], which characterizes exterior equivalence for locally unitary actions in terms of isomorphism of the associated principal (cid:98)G-bundles (cid:92)A (cid:111) G → (cid:98)A (and [OR90, Corollary 1.11] is a version for pointwise unitary ac- tions). These characterizations of exterior equivalence made us wonder whether we could find examples of different generalized fixed-point al- gebras using these types of actions, but subsequent investigation led us instead to the above no-go theorem. Thus, we have additional neg- ative evidence for the existence of examples, alternatively, additional evidence in support of an affirmative answer to the Pedersen rigidity problem. References [BE15] [CR16] [GL89] [Com84] [Cun08] A. Buss and S. Echterhoff, Imprimitivity theorems for weakly proper actions of locally compact groups, Ergodic Theory Dynam. Systems 35 (2015), no. 8, 2412 -- 2457. T. Carlsen and J. Rout, Diagonal-preserving gauge-invariant isomor- phisms of graph C∗-algebras, J. Funct. Anal. 273 (2017), no. 9, 2981 -- 2993. F. Combes, Crossed products and Morita equivalence, Proc. London Math. Soc. (3) 49 (1984), no. 2, 289 -- 306. J. Cuntz, C∗-algebras associated with the ax + b-semigroup over N, K- theory and noncommutative geometry, EMS Ser. Congr. Rep., Eur. Math. Soc., Zurich, 2008, pp. 201 -- 215. [CMW84] R. Curto, P. Muhly, and D. Williams, Cross products of strongly Morita equivalent C∗-algebras, Proc. Amer. Math. Soc. 90 (1984), no. 4, 528 -- 530. [EKQR06] S. Echterhoff, S. Kaliszewski, J. Quigg, and I. Raeburn, A categori- cal approach to imprimitivity theorems for C∗-dynamical systems, vol. 180, Mem. Amer. Math. Soc., no. 850, American Mathematical Society, Providence, RI, 2006. E. Gootman and A. Lazar, Applications of non-commutative duality to crossed product C∗-algebras determined by an action or coaction, Proc. London Math. Soc. 59 (1989), 593 -- 624. P. Green, The local structure of twisted covariance algebras, Acta Math. 140 (1978), 191 -- 250. T. Kajiwara, Continuous crossed product of Hilbert C∗-bimodules, In- ternat. J. Math. 11 (2000), no. 7, 969 -- 981. S. Kaliszewski, T. Omland, and J. Quigg, Cuntz-Li algebras from a-adic numbers, Rev. Roumaine Math. Pures Appl. 59 (2014), no. 3, 331 -- 370. S. Kaliszewski, T. Omland, and J. Quigg, Three versions of categorical crossed-product duality, New York J. Math. 22 (2016), 293 -- 339. [KOQ16] [KOQ14] [Gre78] [Kaj00] 20 [Kus08] M. Kusuda, Duality for crossed products of Hilbert C∗-modules, J. Op- KALISZEWSKI, OMLAND, AND QUIGG erator Theory 60 (2008), no. 1, 85 -- 112. [Lan79] M. B. Landstad, Duality theory for covariant systems, Trans. Amer. [Li15] [Mat16] Math. Soc. 248 (1979), 223 -- 267. X. Li, Continuous orbit equivalence rigidity, Ergodic Theory Dynam. Systems, to appear, arXiv:1503.01704. K. Matsumoto, Uniformly continuous orbit equivalence of Markov shifts and gauge actions on Cuntz-Krieger algebras, Proc. Amer. Math. Soc. 145 (2017), no. 3, 1131 -- 1140. [OPT80] D. Olesen, G. K. Pedersen, and M. Takesaki, Ergodic actions of compact [OR90] [Ped82] [Ped79] [PR84] [QR95] [RR88] [RW98] [Rie74] [Tak75] [Wil07] abelian groups, J. Operator Theory 3 (1980), no. 2, 237 -- 269. D. Olesen and I. Raeburn, Pointwise unitary automorphism groups, J. Funct. Anal. 93 (1990), no. 2, 278 -- 309. G. K. Pedersen, Dynamical systems and crossed products, Operator al- gebras and applications, Part I (Kingston, Ont., 1980), Proc. Sympos. Pure Math. 38, Amer. Math. Soc., Providence, RI, 1982, pp. 271 -- 283. , C∗-algebras and their automorphism groups, London Mathe- matical Society Monographs, vol. 14, Academic Press, Inc. [Harcourt Brace Jovanovich, Publishers], London-New York, 1979. J. Phillips and I. Raeburn, Crossed products by locally unitary auto- morphism groups and principal bundles, J. Operator Theory 11 (1984), no. 2, 215 -- 241. J. C. Quigg and I. Raeburn, Induced C∗-algebras and Landstad duality for twisted coactions, Trans. Amer. Math. Soc. 347 (1995), 2885 -- 2915. I. Raeburn and J. Rosenberg, Crossed products of continuous-trace C∗- algebras by smooth actions, Trans. Amer. Math. Soc. 305 (1988), no. 1, 1 -- 45. I. Raeburn and D. P. Williams, Morita equivalence and continuous-trace C∗-algebras, Mathematical Surveys and Monographs, vol. 60, American Mathematical Society, Providence, RI, 1998. M. A. Rieffel, Induced representations of C∗-algebras, Adv. Math. 13 (1974), 176 -- 257. H. Takai, On a duality for crossed products of C∗-algebras, J. Funct. Anal. 19 (1975), 25 -- 39. D. P. Williams, Crossed products of C∗-algebras, Mathematical Surveys and Monographs, vol. 134, American Mathematical Society, Providence, RI, 2007. School of Mathematical and Statistical Sciences, Arizona State University, Tempe, Arizona 85287 E-mail address: [email protected] Department of Mathematics, University of Oslo, P.O. Box 1053 Blindern, NO-0316 Oslo, Norway E-mail address: [email protected] School of Mathematical and Statistical Sciences, Arizona State University, Tempe, Arizona 85287 E-mail address: [email protected]
1201.4653	1	1201	2012-01-23T09:21:05	La conjecture de Baum-Connes \`a coefficients pour les groupes hyperboliques	[ "math.OA" ]	This paper gives a proof of the Baum-Connes conjecture with coefficients for hyperbolic groups. More precisely the injectivity of the Baum-Connes map was established by Kasparov and Skandalis and we prove the surjectivity.	math.OA	math	La conjecture de Baum-Connes à coefficients pour les groupes hyperboliques Vincent Lafforgue 21 décembre 2013 Le but de cet article est de montrer la conjecture de Baum-Connes à co- efficients pour les groupes hyperboliques au sens de Gromov [Gro87, CDP90, GdlH90]. Soit G un groupe localement compact. La conjecture de Baum-Connes, formulée en 1982 [BC82] affirme que (G) → K∗ (C ∗ red (G)) est un isomorphisme de groupes abéliens. La conjecture de Baum-Connes à coefficients [BCH94] affirme que pour toute G-C ∗ -algèbre A, red : K top µG ∗ µG,A red : K top ∗ (G, A) → K∗ (C ∗ red(G, A)) est un isomorphisme de groupes abéliens. Pour les groupes hyperboliques, l’injectivité de µG,A red a été montrée par Kasparov et Skandalis [KS94, KS03]. La conjecture sans coefficients a été démontrée dans [Laf02] et [MY02]. La conjecture à coefficients commutatifs a été démontrée dans [Laf07] (mais la méthode de [Laf07] ne permet pas de montrer la conjecture de Baum-Connes à coefficients commutatifs pour un produit de deux groupes hyperboliques). Pour énoncer le théorème principal nous avons besoin de quelques défini- tions. Définition 0.1 Soit δ ≥ 0. Un espace métrique (X, d) est dit δ -hyperbolique si pour tout quadruplet (x, y , z , t) de points de X on a d(x, z) + d(y , t) ≤ max (cid:0)d(x, y ) + d(z , t), d(x, t) + d(y , z)(cid:1) + δ. (Hδ (x, y , z , t)) 1 Définition 0.2 Soit δ ≥ 0. Un espace métrique (X, d) est dit faiblement δ -géodésique si pour tous x, y ∈ X et pour tout s ∈ [0, d(x, y ) + δ ] il existe z ∈ X tel que d(x, z) ≤ s et d(z , y ) ≤ d(x, y ) − s + δ . Un espace métrique (X, d) est dit hyperbolique (resp. faiblement géodé- sique) s’il existe δ ≥ 0 tel que (X, d) soit δ -hyperbolique (resp. faiblement δ -géodésique). Lorsque x ∈ X et r ∈ R+ , on note B (x, r) = {y ∈ X, d(x, y ) ≤ r}. Définition 0.3 Un espace métrique (X, d) est dit uniformément localement fini si pour tout r ∈ R+ il existe K ∈ N tel que, pour tout x ∈ X , B (x, r) contienne au plus K points. Le théorème principal est le suivant. Théorème 0.4 Soit G un groupe localement compact agissant de façon iso- métrique, continue et propre sur un espace métrique hyperbolique, faiblement géodésique et uniformément localement fini. Alors G vérifie la conjecture de Baum-Connes à coefficients, c’est-à-dire que pour toute G-C ∗ -algèbre A, µG,A (G, A) → K∗(C ∗ red : K top red (G, A)) est une bijection. ∗ On rappelle que l’injectivité de µG,A red est démontrée dans [KS94, KS03]. Tout groupe hyperbolique Γ muni de la métrique d invariante à gauche associée à la longueur des mots déterminée par un système fini de générateurs est un espace métrique hyperbolique, faiblement géodésique et uniformément localement fini. Donc le théorème 0.4 implique la conjecture de Baum-Connes à coefficients pour les groupes hyperboliques. On aimerait remplacer dans le théorème 0.4 l’hypothèse “uniformement localement fini” par l’hypothèse “à géométrie grossière bornée” qui est stric- tement plus faible (on renvoie à [KS03] pour cette notion et on note que dans [KS03] l’injectivité de µG,A red est démontrée sous cette hypothèse plus faible). Cependant cela rendrait la démonstration encore plus technique et nous y avons renoncé. On aimerait aussi traiter le cas général des groupoïdes hyperboliques, au moins à base compacte, mais cela serait très difficile, car il faudrait adapter la construction de cet article (qui est assez combinatoire) aux techniques de Jean-Louis Tu dans [Tu99], consistant à pondérer par des coefficients tendant vers 0 les éléments du groupoïde qui sont près de disparaître. En revanche, la conjecture sans coefficients pour ces groupoïdes est beaucoup plus accessible par les méthodes de [Laf07], elle est d’ailleurs démontrée dans certains cas dans [Laf07]. 2 D’après [HLS02] la conjecture de Baum-Connes à coefficients est fausse pour certains groupes aléatoires construits par Gromov dans [Gro03]. Ces groupes sont des limites inductives de groupes hyperboliques, la limite étant indexée par N, et les morphismes de transition étant surjectifs. Comme le membre de gauche de la conjecture de Baum-Connes commute aux limites inductives, on voit bien que ces contre-exemples sont “dus” au fait que C ∗ red n’est pas fonctoriel en les morphismes de groupes (non nécessairement injec- tifs). Plus précisément si G → H est un morphisme de groupes, et A une H - C ∗ -algèbre, donc aussi une G-C ∗ -algèbre, le morphisme Cc(G, A) → Cc(H, A) ne se prolonge pas en général par continuité en un morphisme C ∗ red (G, A) → C ∗ red (H, A). Voici maintenant quelques indications sur la démonstration du théorème 0.4, qui occupe tout l’article. On commence par des rappels sur la méthode “Dirac- dual Dirac”, inventée par Kasparov puis développée par Kasparov et Skanda- lis, et Higson et Kasparov. Cette méthode s’applique à une très large classe C de groupes localement compacts, dont la définition est rappelée dans l’intro- duction de [Laf02], et qui contient en particulier les groupes hyperboliques, les sous-groupes fermés des groupes réductifs sur un corps local, et les groupes ayant la propriété de Haagerup, c’est-à-dire possédant une action affine conti- nue et propre sur un espace de Hilbert. Pour tout groupe G dans la classe C , on possède un idempotent γ ∈ KKG (C, C) tel que pour toute G-C ∗ -algèbre A, µG,A red soit injectif et que l’image de l’action de γ sur K∗(C ∗ red (G, A)) soit égale à l’image de µG,A red . Donc si G appartient à C , la conjecture de Baum- Connes à coefficients pour G équivaut au fait que γ agit par l’identité sur K∗ (C ∗ red (G, A)). Si G a la propriété de Haagerup, Higson et Kasparov ont montré dans [HK01] que γ = 1 dans KKG (C, C). Soit G un groupe hyperbolique. Comme certains groupes hyperboliques ont la propriété (T) de Kazhdan, on ne peut pas espérer montrer que γ = 1 dans KKG (C, C). D’un autre côté on sait d’après [Laf02] que γ est égal à 1 dans KK ban G,sℓ(C, C) pour tout s > 0 : cela permet de montrer la conjecture sans coefficients grâce à la propriété (RD) de Jolissaint [Laf02], et aussi à coefficients commutatifs grâce à un autre argument de stabilité par calcul fonctionnel holomorphe un peu plus subtil [Laf07], mais cela ne permet pas de montrer la conjecture à coefficients arbitraires. L’idée pour montrer la conjecture de Baum-Connes à coefficients pour les groupes hyperboliques est que ceux-ci, même s’ils ont la propriété (T), ne vé- rifient pas la propriété (T) renforcée au sens de la définition 0.1 de [Laf08]. De façon un peu imprécise un groupe localement compact G n’a pas la propriété (T) renforcée s’il existe une longueur ℓ sur G (comme dans la définition 1.1) telle que pour tout s > 0 il existe C ∈ R+ tel que la représentation triviale ne soit pas isolée parmi les représentations continues π de G dans des es- 3 paces de Hilbert vérifiant kπ(g )k ≤ eC+sℓ(g) pour tout g ∈ G. Le théorème 1.4 de [Laf08] affirme que si un groupe localement compact possède la pro- priété (T) renforcée toute action continue et isométrique de ce groupe sur un espace métrique hyperbolique, faiblement géodésique et uniformément locale- ment fini a des orbites bornées. La démonstration du théorème 1.4 de [Laf08] montre même que pour un groupe hyperbolique Γ muni de la longueur ℓ associée à un système fini de générateurs, il existe un polynôme P tel que la représentation triviale ne soit pas isolée parmi les représentations π de Γ dans des espaces de Hilbert vérifiant kπ(g )k ≤ P (ℓ(g )) pour tout g ∈ Γ. Notons que Ozawa [Oza08] a montré que les groupes hyperboliques sont faiblement moyennables, ce qui amène à se demander si dans la phrase précédente on ne pourrait pas prendre pour P un polynôme constant. Cependant cela n’appor- terait rien pour la conjecture de Baum-Connes car la seule chose qui compte est que P soit une fonction sous-exponentielle. Pour montrer le théorème 0.4 nous construisons une homotopie de 1 à γ en utilisant des représentations (continues) de G dans des espaces de Hilbert qui ne sont pas unitaires mais à croissance exponentielle arbitrairement petite. Plus précisément nous fixons une longueur ℓ sur G et nous montrons que pour tout s > 0 il existe C ∈ R+ tel que l’on puisse construire une homotopie de 1 à γ en utilisant des représentations π de G dans des espaces de Hilbert qui vérifient (1) kπ(g )k ≤ eC+sℓ(g) pour tout g ∈ G. Le théorème 1.3 affirme l’existence pour tout s > 0 d’une telle homotopie et le théorème 1.2 (qui repose sur des idées de Nigel Higson) montre que cela implique la conjecture de Baum-Connes à coefficients pour G. La preuve du théorème 1.3 est ramenée à celle du théorème 1.5 où l’on suppose que G agit proprement sur un espace hyperbolique X vérifiant certaines propriétés supplémentaires (essentiellement que la métrique est associée à une structure de graphe). La preuve du théorème 1.5 repose sur l’acyclicité du complexe d’homologie simpliciale 0 ← C(∆0 ) ∂← C(∆1 ) ∂← C(∆2 ) . . . ∂← C(∆pmax ) ← 0 du complexe de Rips, où l’ensemble ∆p des faces de dimension p − 1 est formé des parties de X de cardinal p et de diamètre ≤ N , pour N assez grand. On fixe un point base x ∈ X . La partie difficile est la construction de Jx : C(∆p ) → C(∆p+1 ) tel que ∂Jx + Jx∂ = 1 et de normes de Hilbert sur C(∆p ) vérifiant (1) et telles que ∂ et Jx soient continus. L’homotopie de 1 vers γ se fait alors en conjuguant ∂ + Jx par etρ♭ , où ρ♭ (a) est égal à d(x, a) à une constante près et est obtenu par un procédé de moyenne garantissant l’équivariance 4 à compacts près des opérateurs conjugués. Le lecteur qui voudrait se faire une idée rapide de la construction est invité à lire les paragraphes 1 et 2, l’introduction du paragraphe 3, et les sous-paragraphes 4.1 et 4.2. La méthode que nous utilisons est semblable à celle utilisée par Pierre Julg pour montrer la conjecture de Baum-Connes à coefficients pour S p(n, 1) (voir [Jul02]). En fait l’idée de Julg d’utiliser des représentations non unitaires dans des espaces de Hilbert est très ancienne : dans [Jul97] Julg proposait de construire une homotopie de 1 à γ en utilisant des représentations uni- formément bornées de S p(n, 1). L’idée d’utiliser des représentations non pas uniformément bornées mais à petite croissance exponentielle a été dégagée lors de discussions avec Julg et Higson en 1999. Pour conclure voici un petit aperçu du statut actuel de la conjecture de Baum-Connes à coefficients BCcoeff pour des groupes G de la classe C : – “non T” : si G a la propriété de Haagerup, G vérifie BCcoeff d’après [HK01] – “T possible mais non T renforcé” : BCcoeff est vrai si G = S p(n, 1) d’après [Jul02] ou si G est un groupe hyperbolique par le présent article – “T renforcé” : dans ce cas, qui comprend probablement tous les groupes simples sur des corps locaux de rang déployé ≥ 2 (et au moins ceux qui contiennent un SL3 d’après [Laf08]), BCcoeff est totalement ouvert et ne pourra être résolu qu’avec des idées nouvelles comme le principe d’Oka (on renvoie à [Laf10] pour plus de détails). Je remercie Georges Skandalis pour son aide et toutes les discussions que j’ai eues avec lui. Je remercie aussi Miguel Bermudez pour m’avoir indiqué le logiciel JPicEdit, avec lequel les dessins ont été réalisés, et Thomas Delzant pour m’avoir parlé du lemme d’approximation par les arbres, qui simplifie certaines démonstrations. Enfin je remercie vivement le rapporteur qui a tout lu en détail et indiqué de nombreuses corrections. 1 Structure de la démonstration Le but de ce paragraphe est de ramener la démonstration du théorème 0.4 à celle du théorème 1.5. Les paragraphes 3, 4 et 5 seront consacrés à la démonstration du théorème 1.5. Le théorème 1.2 ci-dessous affirme en gros que pour un groupe G agissant proprement sur un espace hyperbolique, l’existence d’homotopies de 1 à γ , utilisant des représentations dans des espaces de Hilbert dont la croissance est contrôlée par une exponentielle arbitrairement petite, implique la surjectivité de l’application de Baum-Connes à coefficients (l’injectivité est déjà connue grâce à [KS03]). 5 Définition 1.1 Soit G un groupe localement compact. On appel le longueur sur G une fonction continue ℓ : G → R+ vérifiant ℓ(g−1) = ℓ(g ) et ℓ(g1g2) ≤ ℓ(g1) + ℓ(g2) pour tous g , g1 , g2 ∈ G. Soit G un groupe localement compact et ℓ une longueur sur G. Pour toutes G-C ∗ -algèbres A et B on définit EG,ℓ(A, B ) comme l’ensemble des classes d’isomorphisme de (E , π , T ) où E est un (A, B )-bimodule hilbertien Z/2Z-gradué muni d’une action continue de G vérifiant kπ(g )k ≤ eℓ(g) pour tout g ∈ G, et d’un opérateur T borné impair tel que pour tout a ∈ A les opérateurs [a, T ] et a(T 2 − 1) soient compacts et que l’application g 7→ a(g (T ) − T ) soit une application normiquement continue de G dans KB (E ). On définit ensuite KKG,ℓ(A, B ) comme l’ensemble des classes d’homotopie dans EG,ℓ(A, B ) : deux éléments sont homotopes si ils sont les évaluations en 0 et 1 d’un élément de EG,ℓ (A, B [0, 1]). On rappelle que B [0, 1] = C ([0, 1], B ) muni de la norme du sup. On peut montrer que la somme directe munit KKG,ℓ (A, B ) d’une structure de groupe abélien. En particulier EG,ℓ (C, C) est l’ensemble des classes d’isomorphisme de (H, π , T ) où H est un espace de Hilbert Z/2Z-gradué muni d’une action continue de G vérifiant kπ(g )k ≤ eℓ(g) pour tout g ∈ G, et d’un opérateur T borné impair tel que (T 2 − 1) soit compact et que l’application g 7→ g (T ) − T soit une application normiquement continue de G dans K(H ). Théorème 1.2 Soit G un groupe localement compact agissant de façon iso- métrique, continue et propre sur un espace métrique (X, d) hyperbolique, fai- blement géodésique et uniformément localement fini. Soit γ ∈ KKG (C, C) l’élément défini sous ces hypothèses par Kasparov et Skandalis [KS03]. Soit x un point de X et ℓ la longueur sur G définie par ℓ(g ) = d(x, gx). Suppo- sons que pour tout s > 0 il existe C ∈ R+ tel que l’image de 1 − γ dans KKG,sℓ+C (C, C) soit nul le. Alors G vérifie la conjecture de Baum-Connes à coefficients, c’est-à-dire que pour toute G-C ∗ -algèbre A, µG,A red : K top (G, A) → ∗ K∗ (C ∗ red (G, A)) est une bijection. Ce théorème est le corollaire 2.12 de [Laf10], dont la preuve repose sur des idées de Higson. En fait nous avons remplacé l’hypothèse “à géométrie grossière bornée” du corollaire 2.12 de [Laf10] par l’hypothèse “uniformément localement fini” qui est strictement plus forte, car nous appliquerons ce théo- rème à des espaces uniformément localement finis et qu’il n’est donc pas nécessaire de rappeler la notion de géométrie grossière bornée. Grâce au théorème 1.2, le théorème 0.4 est une conséquence du théorème suivant. 6 Théorème 1.3 Soit G un groupe localement compact agissant de façon iso- métrique, continue et propre sur un espace métrique (X, d) hyperbolique, fai- blement géodésique et uniformément localement fini. Soit γ ∈ KKG (C, C) l’élément défini sous ces hypothèses par Kasparov et Skandalis [KS03]. Soit x ∈ X et ℓ la longueur sur G définie par ℓ(g ) = d(x, gx). Alors pour tout s > 0 il existe C ∈ R+ tel que l’image de 1 − γ dans KKG,sℓ+C (C, C) est nul le. Remarque. Le théorème 1.3 est encore vrai sans l’hypothèse de propreté de l’action de G sur X (nous avons inclus cette hypothèse pour que γ soit un “élément γ ” pour G). De toute façon le théorème avec l’hypothèse de propreté de l’action implique le théorème sans cette hypothèse car le groupe des automorphismes de (X, d) est localement compact et agit proprement sur X . Nous allons voir maintenant que le théorème 1.3 résulte du théorème 1.5 ci-dessous qui utilise des hypothèses plus fortes sur l’espace métrique (X, d). Définition 1.4 On dit qu’un espace métrique (X, d) est un bon espace hy- perbolique discret si – d prend ses valeurs dans N et est géodésique, c’est-à-dire vérifie ∀a, b ∈ X, ∀k ∈ {0, ..., d(a, b)}, ∃c ∈ X, d(a, c) = k , d(c, b) = d(a, b) − k , autrement dit d provient d’une structure de graphe connexe sur X , – (X, d) est uniformément localement fini (comme d est géodésique, cela équivaut à dire que le nombre de points à distance 1 d’un point, est borné indépendamment du point), – (X, d) est hyperbolique. Remarquons que si G est un groupe hyperbolique, et d est la distance in- variante à gauche associée à la longueur des mots, pour un système fini de générateurs, (G, d) est un bon espace hyperbolique discret, muni d’une ac- tion isométrique de G par translations à gauche (d provient de la structure de graphe de Cayley sur G associée à ce système de générateurs). Soit δ ∈ R∗ + , et (X, d) un espace métrique δ -hyperbolique, faiblement δ - géodésique, uniformément localement fini, et muni d’une action isométrique d’un groupe G. Munissons X de la distance suivante : d′ (a, b) = min{i ∈ N, ∃ a0 , ..., ai , tels que a0 = a, ai = b, et ∀j ∈ {0, ..., i − 1}, d(aj , aj+1) ≤ δ + 1}. Autrement dit d′ provient de la structure de graphe sur X pour laquelle deux points distincts x, y ∈ X sont voisins si d(x, y ) ≤ δ + 1. On a alors les propriétés suivantes : 7 – l’action de G sur (X, d′ ) est isométrique, – d′ est quasi-isométrique à d : on a d ≤ (δ + 1)d′ , et, en utilisant le fait que (X, d) est faiblement δ -géodésique, on montre facilement que d′ ≤ d + 1, – (X, d′ ) est un bon espace hyperbolique discret. L’hyperbolicité de (X, d′ ) résulte de la conservation de l’hyperbolicité par quasi-isométrie pour des espaces faiblement géodésiques. Pour les es- paces géodésiques la démonstration figure dans [GdlH90, CDP90] et pour les espaces faiblement géodésiques il n’y a pas grand chose à modifier. On peut aussi invoquer le théorème 3.18 de [Vai05] ainsi que la remarque 3.19 de [Vai05] appliquée à (X, d) et à l’espace total du graphe considéré précé- demment (je remercie Yves Stalder qui m’a indiqué cette référence). Donc le théorème 1.3 pour (X, d) résulte du théorème 1.5 ci-dessous ap- pliqué à (X, d′ ). Théorème 1.5 Soit G un groupe localement compact agissant de façon iso- métrique, continue et propre sur un espace métrique (X, d) qui est un bon espace hyperbolique discret. Soit γ ∈ KKG (C, C) l’élément défini sous ces hypothèses par Kasparov et Skandalis [KS03]. Soit x ∈ X et ℓ la longueur sur G définie par ℓ(g ) = d(x, gx). Alors pour tout s > 0 il existe C ∈ R+ tel que l’image de 1 − γ dans KKG,sℓ+C (C, C) est nul le. Pour montrer le théorème 0.4 nous sommes donc ramenés à montrer le théorème 1.5. Les paragraphes 3, 4 et 5 sont consacrés à la démonstration du théorème 1.5. 2 Le cas des arbres Le but de ce paragraphe est de démontrer le théorème 1.5 dans le cas où X est un arbre, afin d’introduire dans un cas simple les idées de la démonstration du théorème 1.5. Soit X un arbre et q un entier tel que chaque sommet de l’arbre ait au plus q + 1 voisins. Soit G un groupe localement compact agissant de façon isométrique, continue et propre sur X . Soit γ ∈ KKG (C, C) l’élément γ de Julg et Valette, dont la construction est rappelée ci-dessous. Soit x ∈ X et ℓ la longueur sur G définie par ℓ(g ) = d(x, gx). La proposition suivante est le théorème 1.5 dans le cas où X est un arbre. Proposition 2.1 Pour tout s > 0 il existe C ∈ R+ tel que l’image de 1 − γ dans KKG,sℓ+C (C, C) est nul le. 8 Bien sûr Julg et Valette ont montré dans [JV84] que γ = 1 dans KKG (C, C), sans aucune hypothèse sur l’arbre (c’est-à-dire un résultat plus fort en par- tant d’une hypothèse plus faible). Rappelons la construction de l’élément γ de Julg et Valette. Notons ∆1 = X l’ensemble des sommets de l’arbre et ∆2 l’ensemble des arêtes. Notons C(∆1 ) l’ensemble des combinaisons finies d’éléments de ∆1 . Pour simplifier les notations, nous choisissons pour chaque arête une orientation, ce qui permet d’identifier C(∆2 ) au sous-espace vectoriel de Λ2(C(∆1 ) ) engendré par les ea ∧eb pour {a, b} ∈ ∆2 . De même ℓ2(∆2 ) s’identifie au sous-espace vectoriel fermé de Λ2 (ℓ2(∆1 )) engendré par les ea ∧ eb pour {a, b} ∈ ∆2 . L’élément γ ∈ KKG (C, C) (associé au choix de x comme origine) est représenté par l’espace de Hilbert Z/2Z-gradué ℓ2 (∆1) ⊕ ℓ2 (∆2) et par l’opé- rateur T = (cid:18)0 u v 0(cid:19), où u : ℓ2 (∆2) → ℓ2 (∆1) et v : ℓ2(∆1 ) → ℓ2(∆2 ) sont définis de la façon suivante : si {a, b} est une arête telle que b se trouve sur la géodésique entre x et a, alors u(ea ∧ eb ) = −ea et v (ea ) = −ea ∧ eb et enfin v (ex ) = 0. On voit que 1 − v ◦ u = 0 et que 1 − u ◦ v est le pro jecteur orthogonal sur ex . De plus, si g ∈ G, g (u) − u et g (v ) − v sont de rang fini. Dans la suite nous écrirons toujours T = u + v au lieu de T = (cid:18)0 u v 0(cid:19). Démonstration de la proposition 2.1. Nous allons montrer que pour tout s > 0 il existe C ∈ R+ tel que l’image de 1 − γ dans KKG,2sℓ+C (C, C) soit nulle. Soit s ∈ R∗ + . En fait nous allons montrer qu’il existe un polynôme P et une homotopie de 1 à γ faisant intervenir des représentations π de G dans des espaces de Hilbert tels que kπ(g )k ≤ P (ℓ(g ))esℓ(g). Notons ∂ : C(∆2 ) → C(∆1 ) l’opérateur bord de l’homologie simpliciale. Pour toute arête {a, b} on a ∂ (ea ∧ eb ) = eb − ea . Alors ∂ est injectif et son image est de codimension 1. Pour le montrer définissons h : C(∆1 ) → C(∆2 ) de la façon suivante : pour tout point a ∈ X , notons n = d(a, x) et a0 , . . . , an la suite de points reliant a à x (c’est-à-dire que a0 = a, an = x et ai est voisin de ai−1 pour tout i ∈ {1, . . . , n}) et posons h(ea ) = −(ea0 ∧ ea1 + ... + ean−1 ∧ ean ). En particulier h(ex ) = 0. Alors h ◦ ∂ = IdC(∆2 ) et IdC(∆1 ) − ∂ ◦ h est de rang 1. Plus généralement pour tout t ∈ [0, 1] on définit des opérateurs ut : C(∆2 ) → C(∆1 ) et vt : C(∆1 ) → C(∆2 ) de sorte que u1 = ∂ , v1 = h, et u0 et v0 soient les restrictions de u et v à C(∆2 ) ⊂ ℓ2 (∆2) et C(∆1 ) ⊂ ℓ2(∆1 ). La formule est la suivante : pour toute arête {a, b} telle que b se trouve sur la géodésique entre x et a on pose ut(ea ∧ eb ) = −(ea − teb ). Pour tout point a ∈ X , on note n = d(a, x) et a0 , . . . , an la suite de points reliant a à x (c’est- à-dire que a0 = a, an = x et ai est voisin de ai−1 pour tout i ∈ {1, . . . , n}) et 9 on pose et f (a, b)esd(x,a) . vt (ea ) = −(ea0 ∧ ea1 + tea1 ∧ ea2 + t2 ea2 ∧ ea3 + ... + tn−1ean−1 ∧ ean ). Nous allons compléter C(∆2 ) et C(∆1 ) pour certaines normes de Hilbert telles que kπ(g )k ≤ P (ℓ(g ))esℓ(g) pour un certain polynôme P et telles que les opérateurs ut et vt soient continus, uniformément en t ∈ [0, 1]. Rappelons que dans [Laf02] nous introduisons les espaces ℓ1 x,s(∆1 ) et x,s(∆2 ) comme les complétés de C(∆1 ) et C(∆2 ) pour les normes ℓ1 pondérées ℓ1 suivantes : f (a)ea(cid:13)(cid:13)(cid:13)ℓ1 (cid:13)(cid:13)(cid:13) Xa∈∆1 = Xa∈∆1 f (a)esd(x,a) x,s (∆1 ) f (a, b)ea ∧ eb(cid:13)(cid:13)(cid:13)ℓ1 (cid:13)(cid:13)(cid:13) X{a,b}∈∆2 , b∈géod(x,a) X{a,b}∈∆2 , b∈géod(x,a) = x,s (∆2 ) x,s (∆1 )) = 1 + te−s et On voit que pour tout t ∈ [0, 1] on a kutkL(ℓ1 x,s (∆2 ),ℓ1 x,s (∆2 )) = (1 − te−s )−1 . kvtkL(ℓ1 x,s (∆1 ),ℓ1 Rappelons la construction de [Laf02] qui montre que l’image de 1 − γ dans G,sℓ (C, C) est nulle. On note Ex,s(∆1 ) la C-paire (c0,x,s(∆1), ℓ1 KK ban x,s(∆1 )), où c0,x,s(∆1 ) est le complété de C(∆1 ) pour la norme kf k = supa∈∆1 f (a)e−sd(x,a) x,s(∆1 ) est donné par hf , f ′ i = Pa∈∆1 f (a)f ′(a). et où le crochet entre c0,x,s(∆1 ) et ℓ1 On introduit de la même façon la C-paire Ex,s (∆2). Pour toute C-paire E = (E < , E > ) on note E [0, 1] la C-paire (E < [0, 1], E > [0, 1]) (où E < [0, 1] = C ([0, 1], E <) muni de la norme du sup, et de même pour E > [0, 1]). Alors (Ex,s (∆1)[0, 1]⊕Ex,s(∆2 )[0, 1], (ut+vt )t∈[0,1] ) définit un élément de E ban G,sℓ(C, C[0, 1]). D’autre part (Ex,s (∆1) ⊕ Ex,s(∆2 ), u1 + v1 ) est égal à 1 dans KK ban G,sℓ (C, C) grâce au lemme 1.4.2 de [Laf02] et (Ex,s(∆1 ) ⊕ Ex,s (∆2), u0 + v0 ) est égal à γ dans KK ban G,sℓ(C, C) (pour l’homotopie entre ces deux éléments, on garde les opérateurs, on complète C(∆1 ) pour la norme tk.kEx,s (∆1 ) + (1 − t)k.kℓ2 (∆1 ) , on considère la C-paire formée de cet espace de Banach et du complété de C(∆1 ) pour la norme duale et on fait de même pour ∆2 ). Par conséquent on voit que l’image de 1 − γ dans KK ban G,sℓ(C, C) est nulle. Pour montrer que l’image de 1 − γ dans KKG,2sℓ+C (C, C) est nulle (avec C une constante assez grande), la première idée qui vient est de remplacer les normes ℓ1 par des normes ℓ2 pour avoir des espaces de Hilbert. 10 et 2 Nous définissons donc ℓ2 x,s(∆1 ) et ℓ2 x,s(∆2 ) comme les complétés de C(∆1 ) et C(∆2 ) pour les normes ℓ2 pondérées suivantes : 2 f (a)ea(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13) Xa∈∆1 = Xa∈∆1 f (a)2e2sd(x,a) ℓ2 x,s (∆1 ) f (a, b)ea ∧ eb(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13) X{a,b}∈∆2 , b∈géod(x,a) X{a,b}∈∆2 , b∈géod(x,a) = ℓ2 x,s (∆2 ) Il est clair que ut : ℓ2 x,s(∆2 ) → ℓ2 x,s(∆1 ) est continu (et que sa norme est ma jorée de façon uniforme en t), mais ce n’est pas le cas de vt , si s est petit et t proche de 1, pour la raison suivante. Soit {z , z ′} une arête de X (telle que z ′ ∈ géod(x, z)) et X{a,b}∈∆2 , b∈géod(x,a) f (a, b)ea ∧ eb 7→ f (z , z ′ ) ξz ,z ′ : la forme linéaire sur ℓ2 x,s(∆2 ) qui donne le coefficient de cette arête. Cette forme linéaire est de norme e−sd(x,z ) . Or la forme linéaire tvt(ξz ,z ′ ) = ξz ,z ′ ◦ vt sur ℓ2 x,s(∆1 ) est f (a, b)2e2sd(x,a) . f (a)ea 7→ −(cid:16) Xa tel que z∈géod(x,a) td(z ,a) f (a)(cid:17). Xa Si chaque sommet de l’arbre a exactement q + 1 voisins, en notant k = x,s (∆1 )∗ = Pn≥k t2(n−k) e−2snqn−k . On voit donc que d(x, z) on a ktvt (ξz ,z ′ )k2 ℓ2 cette forme linéaire tvt (ξz ,z ′ ) n’est bornée que si e−st√q < 1. Nous allons compléter C(∆1 ) et C(∆2 ) pour d’autres normes de Hilbert de telle sorte que les opérateurs ut et vt soient continus pour t ∈ [0, 1]. Pour tout n ∈ N, on note S n x = {a ∈ X, d(x, a) = n} la sphère de rayon n et de centre x. Pour k , n ∈ N avec k ≤ n et pour tout z ∈ S k x on note I n,k ,x = {a ∈ S n x , z ∈ géod(x, a)} de sorte que S n I n,k ,x x = Sz∈S k est une z z x partition de la sphère S n x . Lorsque k = n c’est la partition par les singletons et lorsque k = 0 c’est la partition grossière. 11 S n x S k x z x I n,k ,x z . 2 Hx,s (∆1 ) e2sn On définit alors Hx,s(∆1) comme l’espace de Hilbert, complétion de C(∆1 ) pour la norme n 2 x (cid:12)(cid:12)(cid:12) Xa∈I n,k,x f (a)ea(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13) Xa f (a)(cid:12)(cid:12)(cid:12) Xk=0 Xz∈S k = Xn∈N z Alors pour tout t ∈ [0, 1], vt : Hx,s(∆1 ) → ℓ2 x,s(∆2 ) est continu. En effet vt = Pj∈N vt,j où vt,j est défini de la manière suivante : pour tout point a ∈ X , en notant n = d(a, x) et a0 , . . . , an la suite de points reliant a à x (c’est-à-dire que a0 = a, an = x et ai est voisin de ai−1 pour tout i ∈ {1, . . . , n}), on pose vt,j (ea ) = −tj eaj ∧ eaj+1 si j ≤ n − 1 et vt,j (ea ) = 0 si j ≥ n. On vérifie facilement que vt,j est continu de Hx,s(∆1 ) vers ℓ2 x,s(∆2) de norme inférieure ou égale à tj e−sj . La raison est que k.k2 x,s (∆2 ) est une somme pondérée des ℓ2 carrés des formes linéaires ξz ,z ′ , qu’en notant k = d(x, z), tvt,j (ξz ,z ′ ) est le f (a) et que k.k2 produit par −tj de la forme linéaire f 7→ Pa∈I j+k,k,x Hx,s (∆1 ) z est une somme pondérée des carrés de telles formes linéaires. Bien sûr ∂ et plus généralement ut : ℓ2 x,s(∆2 ) → Hx,s(∆1 ) ne sont plus continus et nous devons remplacer ℓ2 x,s(∆2 ) par un espace analogue à Hx,s(∆1 ). Pour k ∈ N, n ∈ N∗ avec k ≤ n et z ∈ S k x on pose x , b ∈ S n−1 = {(a, b), {a, b} ∈ ∆2 , a ∈ S n J n,k ,x , z ∈ géod(x, a)}. z x J n,k ,x On remarque que Sz∈S k est une partition de l’ensemble des arêtes à z x distance n − 1 de x. Lorsque k = n c’est la partition par les singletons et lorsque k = 0 c’est la partition grossière. On définit alors Hx,s(∆2 ) comme la complétion de C(∆2 ) pour la norme n 2 2 x (cid:12)(cid:12)(cid:12) X(a,b)∈J n,k,x f (a, b)(cid:12)(cid:12)(cid:12) f (a, b)ea∧eb(cid:13)(cid:13)(cid:13) Xk=0 Xz∈S k = Xn∈N∗ X{a,b}∈∆2 , b∈géod(x,a) Hx,s (∆2 ) z (cid:13)(cid:13)(cid:13) e2sn . 12 Il est alors très facile de voir que pour tout t ∈ [0, 1], les opérateurs ut : Hx,s(∆2 ) → Hx,s(∆1 ) et vt : Hx,s(∆1) → Hx,s(∆2 ) sont continus, de normes ma jorées uniformément en t. Pour vt la raison est la suivante : l’image par tvt,j de la forme linéaire f 7→ P(a,b)∈J n,k,x f (a, b) qui apparaît dans la z formule ci-dessus est égale au produit par −tj de la forme linéaire f 7→ f (a) qui apparaît dans la formule pour k.k2 Pa∈I n+j,k,x Hx,s (∆1 ) et on en déduit z facilement kvt,j kL(Hx,s (∆1 ),Hx,s (∆2 )) ≤ tj e−sj . Pour conclure il ne reste donc plus qu’à établir le résultat suivant : il existe un polynôme P tel que pour tout s > 0 les représentations de G sur Hx,s(∆1 ) et Hx,s(∆2 ) vérifient kπ(g )k ≤ P (ℓ(g ))esℓ(g) pour tout g ∈ G. Démontrons ce résultat pour Hx,s(∆1). Soit g ∈ G. On pose x′ = g (x), si bien que d(x, x′ ) = ℓ(g ). Il résulte de la définition de Hx,s(∆1 ) que l’on a une isométrie Θx,s : Hx,s(∆1 ) → ℓ2({(n, k , z), 0 ≤ k ≤ n, z ∈ S k x }) définie par f (a)ea ) = (cid:16)esn Xa∈I n,k,x f (a)(cid:17)(n,k ,z ) Θx,s(Xa . z Pour tout f ∈ C(∆1 ) on a kπ(g )f kHx′ ,s (∆1 ) = kf kHx,s (∆1 ) . D’autre part on a une isométrie Θx′ ,s de Hx′ ,s (∆1) dans ℓ2({(n′ , k ′ , z ′ ), 0 ≤ k ′ ≤ n′ , z ′ ∈ S k ′ x′ }). Le lemme suivant permet de comparer les normes de Hx,s et de Hx′ ,s . Lemme 2.2 Pour (n, k , z) ∈ {(n, k , z), 0 ≤ k ≤ n, z ∈ S k x } on peut écrire comme une réunion disjointe finie d’au plus (q − 1)(d(x, x′ ) + 1) + 2 I n,k ,x z parties I n′ ,k ′ ,x′ pour (n′ , k ′ , z ′ ) ∈ {(n′ , k ′ , z ′ ), 0 ≤ k ′ ≤ n′ , z ′ ∈ S k ′ x′ }, de sorte z ′ que pour chaque (n′ , k ′ , z ′ ) la partie I n′ ,k ′ ,x′ intervienne au plus d(x, x′ ) + 2 z ′ fois dans ces décompositions (c’est-à-dire que l’ensemble des (n, k , z) tels que la partie I n′ ,k ′ ,x′ intervienne dans la décomposition de I n,k ,x est fini et de z ′ z cardinal inférieur ou égal à d(x, x′ ) + 2) Ce lemme est le lemme 1.5 de [Laf08] mais nous en rappelons la démonstra- tion (avec une petite correction). = I n′ ,k ′ ,x′ Démonstration. Si z n’appartient pas à géod(x, x′ ), on a I n,k ,x z ′ z avec z ′ = z , k ′ = d(x′ , z), n′ = n + k ′ − k . Si z appartient à géod(x, x′ ), on peut écrire I n,k ,x comme la réunion disjointe z – des I n′ ,k ′ ,x′ , avec z ′ n’appartenant pas à géod(x, x′ ) mais à distance 1 z ′ d’un point de géod(z , x′ ), k ′ = d(x′ , z ′ ) et n′ = n − d(x, x′ ) + 2(k ′ − 1), sous réserve que n′ ≥ 0, – et du singleton I k ′ ,k ′ ,x′ = {z ′}, où z ′ ∈ géod(z , x′ ) vérifie d(x, z ′ ) = n, z ′ et avec k ′ = d(x′ , z ′ ), si n ≤ d(x, x′ ). 13 I n,k ,x z = I n′ ,k ′ ,x′ z ′ z = z ′ x x′ I n′ ,k ′ ,x′ z ′ z ′ • {z ′} = I k ′ ,k ′ ,x′ z ′ x′ x′ x x z • z • Le dessin illustre les deux cas envisagés dans la démonstration (le dessin de gauche correspond au cas où z 6∈ géod(x, x′ ) et les deux dessins de droite correspondent au cas où z ∈ géod(x, x′ )). (cid:3) Fin de la démonstration de la proposition 2.1 en admettant le lemme 2.3. Soit A la matrice de ℓ2 ({(n′ , k ′ , z ′ ), 0 ≤ k ′ ≤ n′ , z ′ ∈ S k ′ ℓ2({(n, k , z), 0 ≤ k ≤ n, z ∈ S k x′ }) dans x }) dont le coefficient vaut es(n−n′ ) si I n′ ,k ′ ,x′ intervient dans la décomposition z ′ de I n,k ,x dans le lemme précédent, et 0 sinon. On a alors A ◦ Θx′ ,s = Θx,s . z Dans chaque ligne de A il y a au plus (q − 1)(d(x, x′ ) + 1) + 2 coefficients non nuls, dans chaque colonne au plus d(x, x′ ) + 2 et ces coefficients ont une valeur absolue inférieure ou égale à esd(x,x′ ) (car pour n et n′ comme dans le lemme avec I n′ ,k ′ ,x′ intervenant dans la décomposition de I n,k ,x , on z ′ z a n − n′ ≤ d(x, x′ )). D’après le lemme 2.3 ci-dessous, une matrice telle que dans chaque ligne il y ait au plus C1 coefficients non nuls, dans chaque colonne au plus C2 et que ses coefficients aient une norme inférieure ou égale à C3 a une norme d’opérateur inférieure ou égale à √C1C2C3 . Donc kAk ≤ p(q − 1)(d(x, x′ ) + 1) + 2pd(x, x′ ) + 2esd(x,x′ ) . Pour tout f ∈ C(∆1 ) , on a kf kHx,s ≤ kAkkf kHx′ ,s et donc kπ(g )f kHx,s ≤ kAkkπ(g )f kHx′,s = kAkkf kHx,s . On voit que π(g ) est continu de Hx,s dans lui-même, de norme inférieure ou égale à p(q − 1)(ℓ(g ) + 1) + 2pℓ(g ) + 2esℓ(g) . 14 On termine maintenant la démonstration de la proposition 2.1. On décide que Hx,s(∆1) est pair et Hx,s(∆2 ) impair, et alors (cid:16)(cid:0)Hx,s(∆1 ) ⊕ Hx,s(∆2)(cid:1)[0, 1], (ut + vt )t∈[0,1](cid:17) appartient à KKG,2sℓ+C (C, C[0, 1]) pour une constante C assez grande, et réalise l’homotopie entre 1 et γ : en effet (Hx,s(∆1)⊕Hx,s (∆2 ), u1+v1 ) est égal à 1 dans KKG,2sℓ+C (C, C) par la proposition 1.4.2 de [Laf02] et (Hx,s(∆1 ) ⊕ Hx,s(∆2 ), u0 + v0 ) est homotope à (ℓ2 (∆1) ⊕ ℓ2 (∆2), u0 + v0) qui représente γ . Comme d’habitude cette homotopie est réalisée en introduisant les normes intermédiaires qtk.k2 Hx,s + (1 − t)k.k2 ℓ2 , et en conservant l’opérateur u0 + v0 pendant l’homotopie. (cid:3) Lemme 2.3 Soit B une C ∗ -algèbre, (Ei )i∈I des B -modules hilbertiens et E = L Ei . Pour i, j ∈ I on se donne aij ∈ LB (Ej , Ei ). Alors si i∈I (cid:0) Xj∈I j∈I (cid:0) Xi∈I kaij kLB (Ej ,Ei )(cid:1) kaij kLB (Ej ,Ei ) (cid:1) et sup sup sont finies, A = (aij )i,j∈I appartient à LB (E ) et on a kaij kLB (Ej ,Ei ) (cid:1)(cid:17). kaij kLB (Ej ,Ei ) (cid:1)(cid:17)(cid:16) sup LB (E ) ≤ (cid:16) sup j∈I (cid:0) Xi∈I i∈I (cid:0) Xj∈I kAk2 En particulier si A, considérée comme une matrice par blocs, possède au plus C1 blocs non nuls dans chaque ligne et C2 blocs non nuls dans chaque colonne et si ces blocs ont une norme inférieure ou égale à C3 , alors kAkLB (E ) ≤ √C1C2C3 . (2) Démonstration. Il suffit de montrer l’inégalité (2) pour I fini. Comme i∈I (cid:0) Xj∈I kaij kLB (Ej ,Ei )(cid:1) A 7→ sup est une norme d’algèbre sur LB (E ), on a alors pour tout n ∈ N∗ , en notant (A∗A)n = (bij )i,j∈I , LB (E ) = k(A∗A)nkLB (E ) ≤ I (cid:16) sup kbij kLB (Ej ,Ei ) (cid:1)(cid:17) i∈I (cid:0) Xj∈I kAk2n kaij kLB (Ej ,Ei )(cid:1)(cid:17)n(cid:16) sup kaij kLB (Ej ,Ei )(cid:1)(cid:17)n ≤ I (cid:16) sup i∈I (cid:0) Xj∈I j∈I (cid:0) Xi∈I 15 , d’où l’inégalité (2) en faisant tendre n vers l’infini. (cid:3) Remarque. Les espaces Hx,s(∆1 ) ressemblent aux espaces de Hilbert intro- duits par Julg et Valette dans [JV84] pour construire l’homotopie entre γ et 1. Plus précisément on rappelle que pour λ ∈]0, +∞[ Julg et Valette notent b i = e−λd(a,b) , où ξ λ Hλ le complété de C(∆1 ) pour le produit scalaire hξ λ a , ξ λ a est la fonction qui vaut 1 en a et 0 ailleurs. Nous écrirons ici H J V au lieu de Hλ λ pour éviter les confusions. Alors pour tout λ > 0 et tout n ∈ N, et f ∈ C(∆1 ) supporté sur S n x , on a = Xa,b∈S n e−λd(a,b) f (a)f (b) kf k2 H J V λ x 2 2 f (a)(cid:12)(cid:12)(cid:12) x (cid:12)(cid:12)(cid:12) Xa∈I n,k,x f (a)(cid:12)(cid:12)(cid:12) + e−2λn (cid:12)(cid:12)(cid:12) Xa∈S n = (1 − e−2λ )e−2λn X1≤k≤n e2λk Xz∈S k x z et comme S n x = I n,0,x on voit que cette formule ressemble à la formule pour x Hx,s (∆1 ) , bien que la présence des facteurs (1 − e−2λ ) et e2λk empêche de kf k2 les identifier, quelles que soient les valeurs de s et de λ. Plus précisément les et k.k2 restrictions de k.k2 Hx,s (∆1 ) à l’espace des fonctions supportées sur H J V λ S n x s’expriment comme des sommes des carrés des mêmes formes linéaires f 7→ Pa∈I n,k,x f (a), mais avec des pondérations un peu différentes. D’autre z part les opérateurs ut + vt que nous utilisons sont quasiment les mêmes que ceux introduits par Julg et Valette. On notera cependant que, dans [JV84], C(∆2 ) est complété pour la norme de ℓ2 (∆2) alors que nous le complétons pour la norme de Hx,s(∆2 ). 3 Construction des espaces et des opérateurs Ce paragraphe et les deux suivants sont consacrés à la démonstration du thérorème 1.5. Soit G un groupe localement compact agissant de façon isométrique, continue et propre sur un espace métrique (X, d) qui est un bon espace hyperbolique discret. Soit δ ∈ N∗ tel que (X, d) soit δ -hyperbolique. Dans toute la suite de cet article on utilisera les notations suivantes : si A et B sont des parties finies de X , on note d(A, B ) = mina∈A,b∈B d(a, b) et dmax (A, B ) = maxa∈A,b∈B d(a, b). Si A est un singleton {x} on note d(x, B ) et dmax (x, B ) au lieu de d({x}, B ) et dmax ({x}, B ). Soit x ∈ X et ℓ la longueur sur G définie par ℓ(g ) = d(x, gx). Le point x sert d’origine dans la construction. Cependant on utilisera en d’autres points que x les constructions faites pour x (notamment pour vérifier les propriétés d’équivariance). On utilisera aussi x comme variable dans certains lemmes. 16 Soit γ ∈ KKG (C, C) l’élément défini sous ces hypothèses par Kasparov et Skandalis [KS03]. Soit s ∈]0, 1]. On va construire, pour une certaine constante C assez grande, une homotopie de 1 à γ dans EG,2sℓ+C (C, C[0, 1]). Cette homotopie sera composée d’une partie difficile (partant de 1), et d’une partie facile (aboutissant à γ ) dont la construction est reléguée au paragraphe 5. Le but de ce paragraphe est de construire des espaces vectoriels et des opérateurs pour la partie difficile de l’homotopie. Dans le paragraphe 4 nous construirons les normes sur ces espaces. On commence par fixer un entier N tel que (HN ) : N est assez grand en fonction de δ. Plus précisément nous utiliserons un nombre fini de fois l’inégalité N ≥ C δ , avec C un entier. On note K un entier tel que pour tout point a de X le nombre de points de X à distance 1 de a soit inférieur ou égal à K . Dans toute la suite on notera C (δ ), C (δ, K ), C (δ, K, N ) ... des constantes qui ne dépendent que des variables indiquées, mais varient d’une formule à l’autre. On note ∆ l’ensemble des parties de X dont le diamètre est inférieur ou égal à N . On note pmax le cardinal maximal d’une partie de X de dia- mètre ≤ N . On a pmax ≤ C (δ, K, N ). Pour p ∈ {1, ..., pmax} on note ∆p l’ensemble des parties de X à p éléments de diamètre ≤ N . On note C(X ) le C-espace vectoriel formé des fonctions à support fini de X dans C, dont la base canonique est notée (ea )a∈X . On choisit une orientation de chaque simplexe associé à un élément de ∆ \ {∅}, ce qui permet de noter, pour tout p ∈ {1, . . . , pmax}, C(∆p ) le sous-espace vectoriel de Λp (C(X ) ) engendré par les ea1 ∧ · · · ∧ eap pour {a1 , ..., ap} ∈ ∆p . Pour S = {a1 , ..., ap} ∈ ∆p on notera aussi eS = ±ea1 ∧ · · · ∧ eap (suivant l’orientation choisie pour ce simplexe). Ces choix d’orientation ont simplement pour but d’alléger les notations. On a C(∆1 ) = C(X ) . On note encore ∆0 = {∅}, C(∆0 ) = C = Λ0(C(X ) ) et e∅ = 1. On note CX l’espace des fonctions de X dans C, qui est le dual algé- brique de C(X ) . Pour tout p ∈ {0, . . . , pmax − 1} on note ∂ : C(∆p+1 ) → C(∆p ) la contraction à gauche par (. . . , 1, 1, . . . ) ∈ CX , c’est-à-dire que pour {a0 , ..., ap} ∈ ∆p+1 , on a p Xi=0 (−1)i ea0 ∧ · · · ∧ eai−1 ∧ eai+1 ∧ · · · ∧ eap . ∂ (ea0 ∧ · · · ∧ eap ) = En d’autres termes ∂ : C(∆1 ) → C(∆0 ) = C est défini par ∂ (Pa∈X f (a)ea ) = ∂← C(∆2 ) . . . ∂← C(∆pmax )) est le complexe d’homologie Pa∈X f (a) et (C(∆1 ) simpliciale. 17 Le complexe suivant est exact : 0 ← C(∆0 ) ∂← C(∆1 ) ∂← C(∆2 ) . . . ∂← C(∆pmax ) ← 0. En fait nous allons construire, pour tout point x ∈ X , un premier para- metrix Hx pour ce complexe (le même que dans [Laf02]), puis un deuxième parametrix ux , puis un troisième parametrix Jx qui est un mélange de Hx et ux , et c’est celui-là qui nous servira. Ici comme dans la suite on emploie le terme “parametrix” plutôt que “homotopie” pour éviter que ce mot ait un double sens. Pour construire l’homotopie de 1 à γ on commencera par représenter l’image de 1 dans KKG,2sℓ+C (C, C) (avec C assez grand) par l’opérateur ∂ + Jx agissant sur le complété de Lpmax p=1 C(∆p ) pour certaines normes de Hilbert k.kHx,s très compliquées construites au paragraphe sui- vant, puis on déformera l’opérateur ∂ + Jx en le conjuguant par eτ θ♭ x où p=1 C(∆p ) → Lpmax x : Lpmax x (S )eS et où ρ♭ x (eS ) = ρ♭ p=1 C(∆p ) est défini par θ♭ θ♭ x est une variante moyennée de la distance à x : c’est la partie difficile de l’homotopie, qui est l’ob jet de ce paragraphe et du suivant. Pour τ assez x est continu sur Lpmax grand l’opérateur ∂ + Jx conjugué par eτ θ♭ p=1 ℓ2 (∆p ), ce qui permet de remplacer les normes compliquées par les normes ℓ2 et d’ar- river ensuite à γ : c’est la partie facile de l’homotopie, qui est traitée au paragraphe 5. L’opérateur Jx vérifiera les deux conditions suivantes, qui ne sont pas for- mulées de manière précise et servent seulement d’heuristique pour la construc- tion de Jx . Il existe une constante C telle que (C1) Jx rapproche de l’origine, plus précisément si S0 , S1 , ..., Sn est une suite sans répétition d’éléments de ∆ telle que eSi+1 apparaît avec un coefficient non nul dans ∂ (eSi ) ou dans Jx(eSi ), la suite S0 , S1 , ..., Sn se rapproche de x en restant à distance ≤ C de la réunion des géodésiques entre x et les points de S0 , (C2) Jx est une intégrale sur un paramètre α d’opérateurs Jx,α tels qu’il existe des parties Yx,α,S (pour S ∈ ∆) vérifiant les propriétés suivantes : – Yx,α,S est une partie finie de X , ne dépendant que de x, α, S , de cardinal ≤ C , et contenant x et S – tous les points de Yx,α,S sont à distance ≤ C de la réunion des géodé- siques entre x et les points de S , – les distances entre les points de Yx,α,S sont déterminées à C près par x, α, S (ce qui fait que le nombre de possibilités pour l’ensemble des distances entre les points de Yx,α,S est borné par une constante) – pour T ∈ ∆, le coefficient de eT dans Jx,α(eS ) est nul si T n’est pas inclus dans Yx,α,S et il ne dépend que de la connaissance des distances 18 entre les points de Yx,α,S , c’est-à-dire, plus précisément : si on se donne S = {a1 , ..., ap−1}, T = {b1 , ..., bp} ⊂ Yx,α,S , p−1} et T ′ = {b′ x′ , S ′ = {a′ 1 , ..., a′ 1 , ..., b′ p} ⊂ Yx′ ,α,S ′ tels qu’il existe une isométrie de Yx,α,S dans Yx′ ,α,S ′ qui envoie 1 , ..., b′ p−1 , b′ 1 , ..., a′ x, a1 , ..., ap−1 , b1 , ..., bp sur x′ , a′ p , (3) 1 ∧ ... ∧ ea′ 1 ∧ ... ∧ eb′ p dans Jx′ ,α(ea′ alors le coefficient de eb′ p−1 ) est égal au coefficient de eb1 ∧ ... ∧ ebp dans Jx,α(ea1 ∧ ... ∧ eap−1 ). Dans la condition (C2) la donnée des points de S sert à lever l’ambiguïté de signe sur eS et de même pour T , S ′ , T ′ . Ces propriétés (C1) et (C2) serviront pour montrer la continuité de Jx . D’après la formule (36) ci-dessous, pour f ∈ C(∆p ) , on aura Hx,s = XZ kf k2 κZ ξZ (f )2 , où la somme porte sur certaines classes d’équivalences Z d’uplets (a1 , ..., ap , Y ), avec {a1 , ..., ap} ∈ ∆ et Y une partie finie de X dont tous les points sont à dis- tance ≤ C de la réunion des géodésiques entre x et les points de {a1 , ..., ap} (la somme sur Z est infinie et le cardinal de Y ne dépend que de Z mais n’est pas borné indépendamment de Z ). La relation d’équivalence est telle que Z détermine les distances entre les points de {x} ∪ {a1 , ..., ap} ∪ Y . Enfin κZ ∈ R∗ + est une pondération et ξZ (f ) = X(a1 ,...,ap ,Y )∈Z où f (a1 , ..., ap ) désigne le coefficient de ea1 ∧ ... ∧ eap dans f . De plus si Z est comme ci-dessus et Jx,α est comme dans (C2), tJx,α(ξZ ) sera une com- binaison finie de formes linéaires ξZ ′ , avec Z ′ une classe d’équivalence de p−1 , Y ′ ) tels qu’il existe (a1 , ..., ap , Y ) ∈ Z vérifiant 1 , ..., a′ (a′ 1 ∧ ... ∧ ea′ – ea1 ∧ ... ∧ eap peut apparaître dans Jx,α(ea′ ), en particulier p−1 a1 , ..., ap sont à distance ≤ C de la réunion des géodésiques entre x et les points de {a′ 1 , ..., a′ p−1}, – Y ′ contient Y ∪ {a1 , ..., ap} ∪ Yx,α,{a′ p−1} . 1 ,...,a′ Donc k.k2 Hx,s sera choisie de telle sorte que si ξZ apparaît dans la formule pour k.k2 Hx,s , ces nouvelles formes linéaires ξZ ′ apparaissent aussi dans la formule Hx,s . La condition (C2) fournit une constante C ′ telle que tJx,α(ξZ ) pour k.k2 soit une combinaison d’au plus C ′ formes linéaires ξZ ′ , ce qui permettra f (a1 , ..., ap ) 19 d’appliquer Cauchy-Schwarz pour ma jorer ξZ (Jx,α(f ))2 par une combinai- son des ξZ ′ (f )2 . Comme Jx sera une intégrale de tels opérateurs Jx,α on montrera, en utilisant encore Cauchy-Schwarz, que kJxf kHx,s ≤ C kf kHx,s pour une certaine constante C . La formule (36) pour k.k2 Hx,s que nous don- nerons plus loin est plus ou moins déterminée par la condition que si une forme linéaire ξZ figure dans la formule pour k.k2 Hx,s , les formes linéaires ξZ ′ servant à décomposer tJx,α(ξZ ) doivent apparaître dans k.k2 Hx,s (ainsi que la x Jxe−τ θ♭ x et eτ θ♭ x ∂ e−τ θ♭ même condition pour les autres opérateurs comme eτ θ♭ x ). La condition (C1) garantit que ce procédé ne diverge pas, en particulier que pour f ∈ C(∆p ) , PZ κZ ξZ (f )2 < ∞. Nous allons voir que Hx satisfait (C1) mais pas (C2). Inversement ux satisfera (C2) mais pas (C1). Heureusement Jx vérifiera (C1) et (C2). 3.1 Construction du premier parametrix Hx Dans toute la suite nous aurons besoin de la notation suivante. Définition 3.1 Si (Y , d) est un espace métrique, ǫ ∈ R+ , et x, y ∈ Y , on note ǫ-géod(x, y ) l’ensemble des points z de Y tels que d(x, z)+d(z , y ) ≤ d(x, y )+ ǫ. On note géod(x, y ) au lieu de 0-géod(x, y ). Donc géod(x, y ) désigne simplement l’ensemble des points z de Y tels que d(x, z) + d(z , y ) = d(x, y ). Les deux lemmes suivants sont valables pour tout espace métrique car ils n’utilisent que l’inégalité triangulaire. Lemme 3.2 Soient (Y , d) un espace métrique, x, x′ , y , y ′, z , z ′ ∈ Y et α ∈ R+ tels que z ∈ α-géod(x, y ). Alors z ′ ∈ (cid:0)α + 2d(x, x′ ) + 2d(y , y ′) + 2d(z , z ′ )(cid:1)-géod(x′ , y ′). Démonstration. Cela résulte des inégalités évidentes d(z ′ , y ′ ) ≤ d(z , y ) + d(y , y ′) + d(z , z ′ ) d(x′ , z ′ ) ≤ d(x, z) + d(x, x′ ) + d(z , z ′ ), et d(x′ , y ′) ≥ d(x, y ) − d(x, x′ ) − d(y , y ′). (cid:3) Lemme 3.3 Soient (Y , d) un espace métrique, x, a, b, c ∈ Y et α, β ∈ R+ tels que b ∈ α-géod(x, a) et c ∈ β -géod(x, b). Alors a) c ∈ (α + β )-géod(x, a), b) b ∈ (α + β )-géod(a, c), c) si d(b, c) ≥ α + β , d(a, c) ≥ d(a, b). 20 c β b α a x Démonstration. On a d(a, b) + d(b, c) + d(c, x) ≤ d(a, b) + d(b, x) + β ≤ d(a, x) + α + β , d’où l’on déduit a) et b) à l’aide des inégalités triangulaires pour les triangles abc et acx respectivement. Enfin c) résulte immédiatement de b). (cid:3) Voici maintenant un lemme qui utilise le fait que (X, d) est δ -hyperbolique. Lemme 3.4 Soient x, a, b, c ∈ X . a) Soit β ∈ N tel que b ∈ β -géod(a, c). Alors d(x, b) ≤ max (cid:0)d(x, a) − d(a, b), d(x, c) − d(c, b)(cid:1) + β + δ. b) En particulier si b appartient à géod(a, c) on a d(x, b) ≤ max (cid:0)d(x, a) − d(a, b), d(x, c) − d(c, b)(cid:1) + δ. a (H β δ (x, a, b, c)) (H 0 δ (x, a, b, c)) x b β c Démonstration. Le a) est une conséquence directe de la propriété d’hyper- bolicité (Hδ (x, a, b, c)) de la définition 0.1 et b) est le cas particulier de a) où β = 0. (cid:3) Nous devons commencer par rappeler certains lemmes de [KS03] et [Laf02]. Comme nous avons affaire à des espaces hyperboliques et géodésiques, et non pas seulement boliques et faiblement géodésiques, les énoncés et les démons- trations des lemmes se simplifient, et nous repartirons donc de zéro. En plus nous éliminerons les paramètres k et t de [Laf02]. On rappelle que ∆ est l’ensemble des parties de X dont le diamètre est inférieur ou égal à N . Pour S ∈ ∆ \ {∅} on note US = ∩a∈S B (a, N ) = {z ∈ X, {z} ∪ S ∈ ∆}. 21 Lemme 3.5 (cf le LEMME 6.2 de [KS03], et le lemme 2.1.4 de [Laf02]) Soient x ∈ X et S ∈ ∆ \ {∅}. Pour z , z ′ ∈ US on a d(z , z ′ ) ≤ (d(x, z) − d(x, US )) + (d(x, z ′ ) − d(x, US )) + 4δ. En particulier le diamètre de {z ∈ US , d(x, z) ≤ d(x, US ) + δ} est inférieur ou égal à 6δ . Grâce à (HN ) on peut supposer N ≥ 6δ , ce que l’on fait. Le lemme 3.5 implique alors que le diamètre de {z ∈ US , d(x, z) ≤ d(x, US )+δ} est inférieur ou égal à N . Démonstration. On commence par un résultat trivial. Lemme 3.6 Soit x ∈ X , r ∈ N, z1 , z3 ∈ B (x, r) et z2 ∈ géod(z1 , z3) avec d(z1 , z2) ≥ δ et d(z2 , z3 ) ≥ δ . Alors z2 ∈ B (x, r). Démonstration. Par (H 0 δ (x, z1 , z2 , z3)) on a d(x, z2) ≤ max(d(x, z1) − d(z1 , z2 ), d(x, z3) − d(z2 , z3)) + δ. (cid:3) Fin de la démonstration du lemme 3.5. Soient z1 , z3 ∈ US avec d(z1 , z3) ≥ (d(x, z1) − d(x, US )) + (d(x, z3) − d(x, US )) + 4δ. Nous allons aboutir à une contradiction. Il existe z2 ∈ géod(z1 , z3 ) tel que d(z1 , z2) = (d(x, z1) − d(x, US )) + 2δ. Alors on a d(z3 , z2) ≥ (d(x, z3 ) − d(x, US )) + 2δ . D’après le lemme 3.6, on a z2 ∈ US . Mais par (H 0 δ (x, z1 , z2 , z3)) on a d(x, z2) ≤ d(x, US ) − δ , d’où une contradiction. (cid:3) Lemme 3.7 (cf le lemme 6.3 de [KS03] et le lemme 2.1.5 de [Laf02]) Soient x ∈ X et S, T ∈ ∆ \ {∅}. Supposons que tout point a dans la différence symétrique de S et T vérifie d(x, a) ≤ d(x, US ) + N − 5δ . Alors a) d(x, UT ) = d(x, US ), b) pour tout b dans la différence symétrique de US et UT , d(x, b) > d(x, US ) + δ. Démonstration. Nous suivons la preuve du lemme 6.3 de [KS03]. Par récur- rence sur le cardinal de la différence symétrique de S et T on peut supposer que la différence symétrique de S et T est un singleton {a}. 22 Supposons d’abord a ∈ T . Alors UT ⊂ US . Comme T = S ∪{a} appartient à ∆, a ∈ US . Par le lemme 3.5, d(a, {z ∈ US , d(x, z) ≤ d(x, US ) + δ}) ≤ N . Donc {z ∈ US , d(x, z) ≤ d(x, US ) + δ} ⊂ UT ⊂ US et les assertions a) et b) du lemme en résultent. Supposons maintenant a ∈ S . Alors US ⊂ UT et comme S = T ∪ {a} appartient à ∆, a ∈ UT . Il suffit de montrer a) car en échangeant les rôles de S et T , b) en résulte. Raisonnons par l’absurde et supposons d(x, UT ) < d(x, US ). Soit b ∈ UT tel que d(x, b) = d(x, UT ). On a alors b 6∈ US donc d(a, b) > N . On rappelle que N ≥ 6δ . Soit c ∈ géod(a, b), d(a, c) = N − 3δ . Par le lemme 3.6, comme a, b ∈ UT , d(a, c) ≥ δ, d(b, c) ≥ δ et c ∈ géod(a, b), on a c ∈ UT . Mais d(a, c) ≤ N donc c ∈ US . Or (H 0 δ (x, a, c, b)) donne d(x, c) ≤ max(d(x, a) − d(a, c), d(x, b) − d(b, c)) + δ < d(x, US ), ce qui est contradictoire. Pour x ∈ X , S ∈ ∆ \ {∅}, on définit AS,x = {z ∈ US , d(x, z) ≤ d(x, US ) + δ} (cid:3) et pour r ∈ N, YS,x,r = {z ∈ US , ∃y ∈ δ -géod(x, z), d(x, y ) ≤ r, d(y , z) = d(y , US )}. y ≤ r δ x z US Il est clair que r 7→ YS,x,r est une application croissante, c’est-à-dire que si r ≤ r ′ on a YS,x,r ⊂ YS,x,r ′ . De plus YS,x,0 est non vide. Si r ≤ d(x, US ) on a YS,x,r = {z ∈ US , ∃y ∈ δ -géod(x, z), d(x, y ) = r, d(y , z) = d(y , US )}. (4) En effet pour z ∈ YS,x,r et y ∈ δ -géod(x, z) vérifiant d(x, y ) ≤ r et d(y , z) = d(y , US ), {d(x, y ′), y ′ ∈ géod(y , z)} est un intervalle contenant d(x, y ) et d(x, z), et comme d(x, y ) ≤ r et d(x, z) ≥ d(x, US ) ≥ r , il existe y ′ ∈ 23 géod(y , z) tel que d(x, y ′) = r et alors d(y ′ , z) = d(y ′ , US ) et y ′ ∈ δ -géod(x, z) par le a) du lemme 3.3. Plus tard on aura besoin des conventions : A∅,x = A{x},x = B (x, δ ) et Y∅,x,r = Y{x},x,r , qui est égal à {x} pour r = 0. Cet ensemble YS,x,r est très proche de celui défini par Kasparov et Skan- dalis dans [KS03], il sert pour l’astuce des ensembles emboîtés. Comme le diamètre de US est inférieur ou égal à 2N , il est clair que US donc aussi AS,x et YS,x,r , ont des cardinaux bornés par C (δ, K, N ). En fait le lemme 3.5 montre que le diamètre de AS,x est inférieur ou égal à 6δ , donc son cardinal est borné par C (δ, K ). Lemme 3.8 Si r ≤ d(x, US ) − N ou si r = 0, on a YS,x,r ⊂ AS,x et YS,x,r = {z ∈ AS,x , ∃y ∈ δ -géod(x, z), d(x, y ) = r, d(y , z) = d(y , AS,x)}. Démonstration. Si r = 0, c’est évident. Supposons donc r ≤ d(x, US ) − N . Soit z ∈ YS,x,r . Grâce à (4) on a z ∈ US et il existe y ∈ δ -géod(x, z), tel que d(x, y ) = r et d(y , z) = d(y , US ). On a d(x, z) ≥ d(x, US ) ≥ r + N donc d(y , z) ≥ N . r x ≥ N z US y δ z ′ On va montrer que d(x, z) ≤ d(x, US ) + δ . Soit z ′ ∈ US . On a d(z , z ′ ) ≤ 2N et d(y , z ′ ) ≥ d(y , z) ≥ N . Par (Hδ (x, y , z , z ′ )) on a d(x, z) ≤ max(d(x, z ′ ) + d(y , z) − d(y , z ′), d(x, y ) + d(z , z ′ ) − d(y , z ′ )) + δ. Mais d(x, z ′ ) + d(y , z) − d(y , z ′ ) ≤ d(x, z ′ ) et d(x, y ) + d(z , z ′ ) − d(y , z ′ ) ≤ r + 2N − N ≤ d(x, US ). Donc d(x, z) ≤ d(x, z ′ ) + δ pour tout z ′ ∈ US . On a montré YS,x,r ⊂ AS,x . Soit maintenant z ∈ AS,x et y ∈ δ -géod(x, z) tels que d(x, y ) = r et d(y , z) = d(y , AS,x). On veut montrer d(y , AS,x) = d(y , US ). Si ce n’est pas 24 vrai, il existe z ′ ∈ US \ AS,x tels que d(y , z ′) = d(y , US ). Comme d(x, z ′ ) ≥ d(x, US ) + δ ≥ d(x, z) et d(y , z ′) ≤ d(y , z), on a y ∈ δ -géod(x, z ′ ). Alors z ′ ∈ YS,x,r ⊂ AS,x , ce qui amène une contradiction. (cid:3) L’astuce des ensembles emboîtés repose sur le lemme suivant. Lemme 3.9 Si r ≤ d(x, US ) − d(x, x′ ) − δ , on a YS,x,r ⊂ YS,x′ ,r+2d(x,x′ )+δ . Démonstration du lemme 3.9 en admettant le lemme 3.11. Soit z ∈ YS,x,r . Par hypothèse il existe y ∈ δ -géod(x, z), d(x, y ) ≤ r , d(y , z) = d(y , US ). On a d(y , z) ≥ d(x, z) − d(x, y ) ≥ d(x, US ) − r ≥ d(x, x′ ) + δ. Soit y ′ ∈ géod(y , z), d(y , y ′) = d(x, x′ ) + δ . y y ′ z US x x′ Alors d(y ′ , z) = d(y ′ , US ), et par le lemme 3.11 (avec ǫ = δ ) on a y ′ ∈ δ -géod(x′ , z). Enfin d(x′ , y ′) ≤ d(x′ , x) + d(x, y ) + d(y , y ′) ≤ r + 2d(x, x′ ) + δ. (cid:3) Lemme 3.10 Pour tout ǫ > 0 et x, z , y , y ′ ∈ X , si y ∈ ǫ-géod(x, z), y ′ ∈ géod(y , z), d(y , y ′) ≥ ǫ/2, alors y ′ ∈ δ -géod(x, z). x y ′ z y 25 δ (x, y , y ′ , z)) on a Démonstration. Par (H 0 d(x, y ′) ≤ max(d(x, y ) − d(y , y ′), d(x, z) − d(z , y ′)) + δ. D’où d(x, y ′ ) + d(y ′ , z) ≤ max(d(x, y ) − d(y , y ′) + d(y ′ , z), d(x, z)) + δ . Or d(x, y )−d(y , y ′)+d(y ′, z) = (d(x, y )+d(y , z))−2d(y , y ′) ≤ d(x, z)+ǫ−2d(y , y ′). Donc d(x, y ′) + d(y ′ , z) ≤ d(x, z) + δ . (cid:3) Lemme 3.11 Pour tous ǫ > 0, x, x′ , z , y , y ′ ∈ X , si y ∈ ǫ-géod(x, z), y ′ ∈ géod(y , z), d(y , y ′) ≥ ǫ/2 + d(x, x′ ), alors y ′ ∈ δ -géod(x′ , z). Démonstration. D’après le lemme 3.2, y ∈ (ǫ + 2d(x, x′ ))-géod(x′ , z). On applique alors le lemme 3.10 à (x′ , z , y , y ′) au lieu de (x, z , y , y ′) et ǫ+ 2d(x, x′ ) au lieu de ǫ. (cid:3) Dans [KS03] Kasparov et Skandalis définissent une mesure ψS,x de masse 1 en normalisant une moyenne paramétrée par r des fonctions caractéristiques de YS,x,r . Nous allons faire de même, à ceci près que nous prendrons plutôt une moyenne sur r de la fonction caractéristique de YS,x,r normalisée. Pour tout ensemble non vide A on note νA = 1 ♯A χA , où χA est la fonction caractéristique de A. On pose alors ψS,x = νYS,x,r . 1 max(1, d(x, US ) − N ) max(0,d(x,US )−N −1) Xr=0 D’après le lemme 3.8 le support de ψS,x est inclus dans AS,x . Plus loin nous aurons besoin de la notation suivante : on définit ψS,x,t = νYS,x,max(0,E (t(d(x,US )−N ))) pour t ∈ [0, 1] de sorte que ψS,x = R 1 0 ψS,x,tdt. Dans tout cet article on note E (.) la partie entière. Lemme 3.12 Pour tout t ∈ [0, 1], le support de ψS,x,t est inclus dans AS,x . Démonstration. Cela résulte immédiatement du lemme 3.8. (cid:3) Pour r ∈ {0, ..., max(0, d(x, US ) − N )}, on a, grâce au lemme 3.8, YS,x,r = {z ∈ AS,x , ∃y ∈ δ -géod(x, z), d(x, y ) = r, d(y , z) = d(y , AS,x)}. 26 Donc YS,x,r ne dépend que de la connaissance des points de (5) S ∪ AS,x ∪ {y , ∃z ∈ AS,x , y ∈ δ -géod(x, z), d(x, y ) = r} ∪ {x} et des distances mutuelles entre tous ces points. De façon plus précise, si x′ et S ′ sont tels qu’il existe une isométrie de (5) vers l’ensemble correspondant pour x′ et S ′ , qui envoie x et S sur x′ et S ′ , alors YS ′ ,x′ ,r ⊂ AS ′ ,x′ est l’image de YS,x,r ⊂ AS,x par cette isométrie. En anticipant un peu justifions l’intérêt de cette propriété. D’après la dé- monstration du lemme 3.20, l’ensemble (5) est inclus dans l’ensemble figurant dans l’énoncé du lemme 3.20 et on déduit du lemme 3.13 que le cardinal de (5) est borné par une constante C (δ, K, N ), et de plus la connaissance de r et de d(x, US ) détermine les distances entre les points de (5) à une constante C (δ, N ) près. Donc connaissant r et d(x, US ), il n’y a qu’un nombre fini, borné par C (δ, K, N ), de possibilités pour la donnée de toutes les distances entre les points de (5). Enfin ψS,x est une moyenne entre 0 et max(0, d(x, US ) − N − 1) de νYS,x,r . Au contraire avec la définition de Kasparov et Skandalis, où la normalisation est faite après la moyennne, le calcul de ψS,x nécessiterait la connaissance simultanée de tous les points y tels qu’il existe z ∈ AS,x vérifiant y ∈ δ -géod(x, z). Le lemme suivant, que nous venons d’utiliser, servira à de nombreuses reprises. Lemme 3.13 Pour tout r ∈ N il existe C dépendant seulement de δ, K et r tel que pour x, y ∈ X , on ait, – pour tout l ∈ {0, . . . , d(x, y ) + r}, ♯(cid:8)z ∈ r-géod(x, y ), d(x, z) = l(cid:9) ≤ C – et ♯(cid:0)r-géod(x, y )(cid:1) ≤ C (d(x, y ) + 1). Démonstration. Pour x, y ∈ X on pose d = d(x, y ) et on choisit x0 = x, x1 , . . . , xd = y des points de géod(x, y ) tels que d(x, xi ) = i. Soit z ∈ r-géod(x, y ) et i = min(d, d(x, z)). Alors par (H 0 δ (z , x, xi , y )) on a d(z , xi ) ≤ max(d(z , x) − d(x, xi ), d(z , y ) − d(y , xi)) + δ ≤ r + δ car d(x, xi ) = i, d(xi , y ) = d − i, d(x, z) ∈ [i, i + r ] et d(z , y ) ∈ [d − i, d − i + r ]. On pose C = 1 + K + K 2 + · · · + K r+δ . Alors maxx∈X ♯B (x, r + δ ) ≤ C . On voit que la première assertion est vraie et pour la deuxième assertion on remarque que r-géod(x, y ) ⊂ Sd i=0 B (xi , r + δ ). (cid:3) Par l’astuce des ensembles emboîtés nous allons montrer que pour x, x′ ∈ X , kψS,x − ψS,x′ k1 tend vers 0 en dehors des parties finies de ∆ (où k.k1 désigne la masse totale d’une mesure). Cela résulte du lemme suivant qui est plus fort. 27 Lemme 3.14 Il existe C = C (δ, K, N ) tel que la mesure de l’ensemble des t ∈ [0, 1] tels que ψS,x,t 6= ψS,x′ ,t soit ≤ C d(x,x′ ) 1+d(x,S ) . Démonstration. Le lemme est vrai si x = x′ , donc on suppose d(x, x′ ) ≥ 1. On prendra C ≥ 2N + 2 si bien que C d(x,x′ ) 1+d(x,S ) ≥ 1 si d(x, S ) ≤ d(x, x′ ) + 2N . Donc on suppose d(x, S ) ≥ d(x, x′ ) + 2N + 1, et il suffit de montrer le lemme sous cette hypothèse. On a alors min(d(x, US ) − N , d(x′ , US ) − N ) ≥ d(x, S ) − 2N − d(x, x′ ) ≥ 1. Comme le cardinal de YS,x,r est non nul et borné par une constante C0 = C (δ, K, N ), et que l’application r 7→ YS,x,r est croissante, l’intervalle [0, d(x, US ) − N − 1] se découpe en au plus C0 intervalles où l’application r 7→ YS,x,r est constante. Il résulte du lemme 3.9 que l’application r 7→ YS,x′ ,r coïncide avec la précé- dente sur des intervalles (éventuellement vides) obtenus à partir des précé- dents en raccourcissant chaque extrémité de 2d(x, x′ ) + δ . Il en résulte que l’application croissante t 7→ YS,x,E (t(d(x,US )−N )) prend au plus C0 valeurs et pour chaque valeur l’image inverse est un inter- valle, et l’image inverse de cette même valeur par l’application t 7→ YS,x′ ,E (t(d(x′ ,US )−N )) est un autre intervalle dont les extrémités diffèrent au plus de 3d(x,x′ )+δ+1 d(x,S )−2N des extrémités du premier. En effet pour t, t′ ∈ [0, 1], l’inégalité E (t(d(x, US ) − N )) − E (t′ (d(x′ , US ) − N )) ≤ 2d(x, x′ ) + δ implique t − t′ ≤ 3d(x,x′ )+δ+1 d(x,US )−N car d(x, US ) − d(x′ , US ) ≤ d(x, x′ ) et on utilise le fait que d(x, US ) − N ≥ d(x, S ) − 2N . On en déduit que la mesure de l’ensemble des t ∈ [0, 1] tels que µr,t(x, a) 6= µr,t(x′ , a) est inférieure ou égale à 2C0 (3d(x,x′ )+δ+1) . Enfin on a d(x,S )−2N 2C0(3d(x, x′ ) + δ + 1) d(x, S ) − 2N 2C0(4 + δ )(2N + 2)d(x, x′ ) 1 + d(x, S ) ≤ 28 car on a supposé d(x, x′ ) ≥ 1 et d(x, S ) ≥ 2N + 1. On prend alors C = 2C0(4 + δ )(2N + 2). (cid:3) On définit un opérateur hx de degré 1 sur ⊕pmax p=0 C(∆p ) par la formule suivante : hx (eS ) = ψS,x ∧ eS . Plus loin nous aurons besoin de la notation suivante : pour t ∈ [0, 1], hx,t est l’opérateur défini par hx,t(eS ) = ψS,x,t ∧ eS , de sorte que hx = R 1 0 hx,tdt. On note que pour tout t ∈ [0, 1], hx,t(e∅) = ex . Lemme 3.15 Il existe C = C (δ, K, N ) tel que la mesure de l’ensemble des t ∈ [0, 1] tels que (hx,t − hx′ ,t )(eS ) 6= 0 soit ≤ C d(x,x′ ) 1+d(x,S ) . Démonstration. Cela résulte immédiatement de (6) et du lemme 3.14. (cid:3) Le lemme suivant servira tout à fait à la fin de l’article. (6) Lemme 3.16 On a h2 x = 0. Démonstration. D’après les lemmes 3.7 et 3.12, si eT apparaît dans hx (eS ) on a AT ,x = AS,x . Le lemme 3.8 montre alors que ψT ,x = ψS,x . (cid:3) Nous voulons montrer que l’opérateur ∂hx+hx∂ est inversible de ⊕pmax p=0 C(∆p ) dans lui-même. En effet Hx = hx (∂hx + hx∂ )−1 sera alors un parametrix pour ∂ , c’est-à-dire que l’on aura ∂Hx + Hx∂ = 1. Pour cela nous introduisons, comme dans le paragraphe 2.5.1 de [Laf02], la “distance moyenne tronquée de S à x” : ζx(S ) = 1 pmax (cid:16) Xa∈S max(d(x, US ), d(x, a) − δ ) + (pmax − ♯S )d(x, US )(cid:17), si S 6= ∅ et ζx (∅) = 0. Cette distance est tronquée au sens où tous les points de S dont la distance à x est comprise entre d(x, US ) et d(x, US ) + δ contribuent de la même façon. On a clairement d(x, US ) ≤ ζx (S ) ≤ d(x, US ) + N − δ . Lemme 3.17 Pour tout S ∈ ∆ on a MT tel que ζx (T )≤ζx (S ) ∂ (eS ) ∈ MT tel que ζx (T )≤ζx (S ) M (1 − ∂hx − hx∂ )(eS ) ∈ T tel que ζx (T )<ζx (S )− N −6δ pmax hx (eS ) ∈ CeT , CeT , CeT . 29 Démonstration. La première assertion est évidente. La deuxième assertion est vraie car le support de ψS,x est inclus dans AS,x , en vertu du lemme 3.12 et pour T = S ∪ {a} avec a ∈ AS,x on a d(x, UT ) = d(x, US ) d’après le lemme 3.7. Pour montrer la dernière assertion on remarque que (1 − ∂hx − hx∂ )(eS ) = Xa∈S Si d(x, a) ≤ d(x, US ) + N − 5δ , on a d(x, US \{a} ) = d(x, US ) et AS \{a},x = AS,x par le lemme 3.7, d’où ψS \{a},x = ψS,x par le lemme 3.8. Si d(x, a) > d(x, US ) + N − 5δ , comme les supports de ψS,x et ψS \{a},x sont inclus dans AS,x et AS \{a},x , (ψS,x − ψS \{a},x ) ∧ eS \{a} est une combinaison de eS \{a}∪{b} avec ±(ψS,x − ψS \{a},x ) ∧ eS \{a} . (cid:3) d(x, b) ≤ max(d(x, US \{a} ) + δ, d(x, US ) + δ ) = d(x, US ) + δ, ce qui fait que ζx (S \ {a} ∪ {b}) < ζx(S ) − N −6δ . pmax Définition 3.18 Pour p ∈ {1, . . . , pmax} et f = PS∈∆p f (S )eS , on note [S tel que f (S )6=0 supp(f ) = S. On a hx = R 1 0 hx,tdt et les lemmes suivants permettent, pour t ∈ [0, 1] et S ∈ ∆, d’estimer le support de hx,t(eS ) et de savoir de quoi hx,t(eS ) dépend. Lemme 3.19 Soit x ∈ X et S ∈ ∆ \ {∅}. a) Pour tout t ∈ [0, 1] le support de hx,t(eS ) est inclus dans S ∪ AS,x . b) On a AS,x ⊂ (cid:16)B (x, 2δ ) ∩ US (cid:17) ∪ [a∈S,a6∈B (x,2δ) En particulier on a toujours AS,x ⊂ Sa∈S 4δ -géod(x, a). Démonstration. Comme hx,t (eS ) = ψS,x,t ∧ eS , le a) résulte du lemme 3.12. Pour montrer le b) on rappelle que AS,x = {y ∈ US , d(x, y ) ≤ d(x, US ) + δ}. Soit y ∈ AS,x n’appartenant pas à B (x, 2δ ). Soit z ∈ géod(x, y ) tel que d(y , z) = 2δ . Comme d(x, US ) ≥ d(x, y ) − δ et d(x, z) = d(x, y ) − 2δ , on a z 6∈ US . Donc il existe a ∈ S tel que d(a, z) > N . {y ∈ 3δ -géod(x, a), d(y , a) ∈]N − 2δ, N ]} 30 x z 2δ y a US Comme d(y , z) = 2δ , on a d(a, y ) > N − 2δ . Ensuite (H 0 δ (a, x, z , y )) donne d(a, z) ≤ max(d(a, y ) − 2δ, d(a, x) − d(x, y ) + 2δ ) + δ. Comme d(a, z) > N et d(a, y ) ≤ N on a nécessairement d(a, z) ≤ d(a, x) − d(x, y ) + 3δ, et comme d(a, y ) ≤ N < d(a, z) on en déduit y ∈ 3δ -géod(x, a). Par (7) on a aussi d(x, a) ≥ d(x, y ) + N − 3δ ≥ N − 3δ > 2δ et a 6∈ B (x, 2δ ). La seconde assertion de b) résulte immédiatement de la première : comme S est non vide, B (x, 2δ ) ⊂ Sa∈S 4δ -géod(x, a). (cid:3) Lemme 3.20 Pour t ∈ [0, 1] et S ∈ ∆, hx,t(eS ) ne dépend que de la connais- sance des points de B (x, 2δ ) ∪ S ∪ [a∈S ∪ [a∈S {y ∈ 5δ -géod(x, a), d(x, y ) ∈ [td(x, S ) − 2N − 1, td(x, S )]} et des distances entre tous ces points. Plus précisément, si on note S = {a1 , ..., ap−1}, si T = {b1 , ..., bp} est inclus dans (8) (sans quoi le coefficient de eT dans hx,t(eS ) est nul), et si x′ ∈ X , S ′ = {a′ 1 , ..., a′ p−1} et T ′ = {b′ 1 , ..., b′ p} sont tels qu’il existe une isométrie de (8) dans l’ensemble correspondant pour x′ et S ′ , qui envoie x, a1 , ..., ap−1 , b1 , ..., bp sur x′ , a′ 1 , ..., b′ p−1 , b′ 1 , ..., a′ p , alors le coefficient de eb1 ∧ ... ∧ ebp dans hx,t(ea1 ∧ ... ∧ eap−1 ) est égal au coefficient de 1 ∧ ... ∧ ea′ 1 ∧ ... ∧ eb′ ). p dans hx′ ,t(ea′ eb′ p−1 {y ∈ 3δ -géod(x, a), d(y , a) ∈]N − 2δ, N ]} (7) (8) 31 Démonstration. On rappelle que ψS,x,t = νYS,x,r avec r = max(0, E (t(d(x, US ) − N ))) et que, d’après le lemme 3.8, YS,x,r = {z ∈ AS,x , ∃y ∈ δ -géod(x, z), d(x, y ) = r, d(y , z) = d(y , AS,x)}. Donc YS,x,r ne dépend que de la connaissance des points de {y ∈ 3δ -géod(x, a), d(y , a) ∈]N − 2δ, N ]}. S ∪ AS,x ∪ {y , ∃z ∈ AS,x , y ∈ δ -géod(x, z), d(x, y ) = r} ∪ {x} et des distances entre tous ces points. D’après le lemme 3.19, AS,x ⊂ B (x, 2δ ) ∪ [a∈S Soit maintenant z ∈ AS,x et y ∈ δ -géod(x, z) tel que d(x, y ) = r . D’après le lemme 3.19, il existe a ∈ S tel que z ∈ 4δ -géod(x, a). Or y ∈ δ -géod(x, z) et z ∈ 4δ -géod(x, a) impliquent y ∈ 5δ -géod(x, a) par le a) du lemme 3.3. Comme d(x, US ) ∈ [d(x, S ) − N , d(x, S )] on a r ∈ [td(x, S ) − 2N − 1, td(x, S )]. (cid:3) Le lemme suivant est un lemme général sur les espaces hyperboliques, qui généralise le lemme 3.10. Lemme 3.21 Soient α, β ∈ N et x, a, b, c ∈ X tels que b ∈ α-géod(a, x), c ∈ β -géod(b, x). Alors c ∈ (cid:0) max(α + 2β − 2d(b, c), β ) + δ(cid:1)-géod(a, x), en particulier c ∈ (β + δ )-géod(a, x) si d(b, c) ≥ α+β 2 . c β b x α a Démonstration. En effet (H β δ (a, b, c, x)) s’écrit d(a, c) ≤ max(d(a, b) − d(b, c), d(a, x) − d(x, c)) + β + δ, d’où l’on déduit d(a, c) + d(c, x) − d(a, x) ≤ max(d(a, b) − d(b, c) + d(c, x) − d(a, x), 0) + β + δ 32 et d’autre part d(a, b) + d(b, c) + d(c, x) ≤ d(a, b) + d(b, x) + β ≤ d(a, x) + α + β . (cid:3) Nous allons montrer qu’en appliquant ∂ et hx de façon répétée à eS (pour S ∈ ∆ \ {∅}), on reste à distance bornée de la réunion des géodésiques reliant x aux points de S . Lemme 3.22 Soit n ∈ N et S0 , . . . , Sn une suite d’éléments de ∆ tel le que pour tout i, eSi+1 apparaît avec un coefficient non nul dans ∂ (eSi ) ou dans hx (eSi ). Alors pour tout point yn ∈ Sn n’appartenant ni à B (x, 2δ ) ni à S0 , il existe y0 ∈ S0 n’appartenant pas à B (x, 2δ ) tel que yn ∈ 4δ -géod(x, y0) et d(y0 , yn) > N − 2δ . Démonstration. D’après le lemme 3.19 il existe yn−1 ∈ Sn−1 ,..., y0 ∈ S0 n’appartenant pas à B (x, 2δ ) tels que yi = yi+1 ou bien yi+1 ∈ 3δ -géod(x, yi ) et d(yi , yi+1) > N − 2δ. Par récurrence sur i ∈ {1, ..., n} on montre yi ∈ 4δ -géod(x, y0) en appliquant le lemme 3.21 à (y0 , yi−1 , yi) au lieu de (a, b, c) et (4δ, 3δ ) au lieu de (α, β ) et en utilisant le fait que N − 2δ ≥ 7δ/2. On suppose N ≥ 9δ , ce qui est permis par (HN ). D’après le c) du lemme 3.3 appliqué à (y0 , yi , yi+1) au lieu de (a, b, c) et à (4δ, 3δ ) au lieu de (α, β ), on a d(y0 , yi) ≤ d(y0 , yi+1) pour tout i, et comme d(y0 , yi) > N − 2δ si i est le plus petit entier tel que yi 6= y0 on en déduit d(y0 , yn ) > N − 2δ . (cid:3) Il résulte du lemme 3.20 que pour x ∈ X et S ∈ ∆ \ {∅}, la connaissance de hx (eS ) dépend seulement de celle des points de Sa∈S 5δ -géod(x, a). Grâce au lemme 3.22, on en déduit le corollaire suivant. Corollaire 3.23 Si S0 , . . . , Sn est une suite d’éléments de ∆ tel le que S0 6= ∅ et que pour tout i, eSi+1 apparaît avec un coefficient ci non nul dans ∂ (eSi ) ou dans hx (eSi ), alors S0 ∪ · · · ∪ Sn ⊂ Sa∈S0 4δ -géod(x, a) et la connaissance de tous les ci ne dépend que de la connaissance des points de Sa∈S0 9δ -géod(x, a) (et de leurs distances mutuel les). Démonstration. Si x, a, b, c sont tels que b ∈ 4δ -géod(x, a) et c ∈ 5δ -géod(x, b) alors c ∈ 9δ -géod(x, a) par le a) du lemme 3.3. (cid:3) On rappelle que ∂hx + hx∂ est inversible de ⊕pmax p=0 C(∆p ) dans lui-même et que l’on a posé Hx = hx (∂hx + hx∂ )−1 . 33 Corollaire 3.24 Pour p ∈ {1, ..., pmax} et S ∈ ∆p , le support de Hx (eS ) est inclus dans Sa∈S 4δ -géod(x, a) et Hx (eS ) ne dépend que de la connaissance des points de Sa∈S 9δ -géod(x, a) (et de leurs distances mutuel les). Démonstration. C’est une conséquence immédiate du corollaire 3.23. (cid:3) On note par ailleurs que Hx (e∅ ) = ex . 3.2 Une conséquence de la construction de Hx Pour la construction du deuxième parametrix ux , nous aurons besoin du lemme suivant qui est une variante du lemme 2.5.6 de [Laf02]. On rappelle que pour f = PS∈∆p f (S )eS , on note supp(f ) = SS tel que f (S )6=0 S . Lemme 3.25 Pour tout p ∈ {1, . . . , pmax}, il existe une application G-équivariante mais non nécessairement linéaire Φp : {f ∈ C(∆p ) , ∂p−1(f ) = 0} → C(∆p+1 ) tel le que pour tout f dans l’ensemble de départ de Φp , – ∂p (Φp (f )) = f , – Φp (λf ) = λΦp (f ) pour λ ∈ C, – supp(Φp (f )) est inclus dans Sy ,z∈supp(f ) 4δ -géod(y , z), – Φp (f ) ne dépend que de f et de la connaissance des points de [y ,z∈supp(f ) 9δ -géod(y , z) et de leurs distances mutuel les, autrement dit si f ′ est une autre fonc- tion une isométrie de (9) vers Sy ,z∈supp(f ′ ) 9δ -géod(y , z) envoyant f sur f ′ envoie Φp (f ) sur Φp (f ′ ), et tel le que pour tout R ∈ R+ il existe C ∈ R+ ne dépendant que de δ, K, N , R, tel que pour tout f ∈ C(∆p ) avec ∂p−1(f ) = 0 et diam(supp(f )) ≤ R on ait kΦp (f )kℓ1 (∆p+1 ) ≤ C kf kℓ1 (∆p ) . Démonstration. On donne une formule explicite pour Φp , qui est la même que dans [Laf02] : (9) Φp (f ) = 1 ♯(supp(f )) Xz∈supp(f ) Pour montrer les propriétés de Φp , on applique le corollaire 3.24. Pour la suite de l’article, on fixe de telles applications Φp . Hz (f ). (cid:3) 34 3.3 Construction d’un deuxième paramétrix ux Le paramétrix Hx ne nous convient pas car nous ne savons pas construire de normes sur Lpmax p=0 C(∆p ) telles qu’il soit continu ainsi que ∂ et que kπ(g )k ≤ P (ℓ(g ))esℓ(g) pour un certain polynôme P . En fait nous avions déjà un pro- blème dans [Laf02], puisqu’il avait fallu prendre l’opérateur Hx associé à une valeur plus grande de N et le modifier grâce au lemme 2.5.6. Ici le problème est encore aggravé : pour a ∈ X le coefficient dans Hx (ea ) d’une arête conte- nant x dépend au moins de la connaissance de tous les points de 9δ -géod(x, a), et le nombre de possibilités pour les distances entre tous ces points est une exponentielle en d(x, a), comme le montre l’exemple suivant. Exemple. Soit Γ le produit libre de Z/3Z avec Z. On note e1 et e2 les générateurs de Z/3Z et Z. Soit ℓ la longueur des mots sur Γ et soit X = Γ muni de la distance d(a, b) = ℓ(a−1b). Alors pour tout a ∈ X , la réunion des cycles de longueur 3 passant par a contient ae1 et ae−1 1 mais ni ae2 et ae−1 2 . Donc la classe d’isométrie de 9δ -géod(x, a) détermine l’emplacement de e∓1 1 et e∓1 2 dans l’écriture de x−1a comme mot réduit, et il y a 2d(x,a) possibilités. Autrement dit Hx ne vérifie pas la condition (C2) (en revanche Hx vérifie la condition (C1) grâce au lemme 3.17 et au corollaire 3.24). Nous allons maintenant construire un nouveau parametrix ux vérifiant la condition (C2) (mais pas la condition (C1)). Nous commençons par définir, pour tous x, a ∈ X et r ∈ {1, ..., d(x, a) − 1} une mesure µr (x, a) de masse 1, supportée par {y ∈ δ -géod(x, a), d(a, y ) = r}. D’après le lemme 3.13, le cardinal de cet ensemble est borné par une constante C (δ, K ). Nous voulons que la propriété suivante soit satisfaite : ∀ρ > 0, ∀ǫ > 0, ∃R > 0, si d(x, x′ ) ≤ ρ et r ≤ d(a, x) − R, alors kµr (x′ , a) − µr (x, a)k1 ≤ ǫ. Dans la propriété ci-dessus, la condition que d(a, x)−r est suffisamment grand est évidemment la plus faible possible, car d(a, x) − r est essentiellement la distance entre x et les points du support de µr (x, a) et cette distance doit nécessairement être grande pour que µr (x, a) varie peu en fonction de x. On rappelle que pour toute partie A non vide de X , on note νA la mesure de masse 1 égale au produit par (♯A)−1 de la fonction caractéristique de A. Voici la formule : (10) µr (x, a) = d(a,x)−r−1 1 Xk=0 d(a, x) − r Ax,a,r,k = {y , d(a, y ) = r, ∃z ∈ δ -géod(x, a), d(a, z) = r + k , y ∈ géod(z , a)}. νAx,a,r,k où pour k ≤ d(a, x) − r on pose 35 x z k y r a µr,t(x, a)dt où µr,t(x, a) = νAx,a,r,E (t(d(a,x)−r)) . On a donc µr (x, a) = Z 1 0 D’après le a) du lemme 3.3, Ax,a,r,k ⊂ {y ∈ δ -géod(x, a), d(a, y ) = r} et d’après le lemme 3.13 le cardinal de ces ensembles est borné par une constante de la forme C (δ, K ). Pour tout k ≤ d(a, x) − r , Ax,a,r,k est non vide, puisque X est géodésique. Lemme 3.26 a) Pour tout k ∈ {1, ..., d(x, a) − r} on a Ax,a,r,k ⊂ Ax,a,r,k−1. b) Pour x′ ∈ X et k ∈ N vérifiant d(x, x′ ) + δ ≤ k ≤ d(x, a) − r on a Ax,a,r,k ⊂ Ax′ ,a,r,k−d(x,x′ )−δ . Démonstration. Montrons a). Soient y , z tels que d(a, y ) = r, z ∈ δ -géod(x, a), d(a, z) = r + k , y ∈ géod(z , a). Il existe un point z ′ ∈ géod(y , z) à distance 1 de z . Alors z ′ vérifie d(a, z ′ ) = r + k − 1, y ∈ géod(z ′ , a) et z ′ ∈ δ -géod(x, a) par le a) du lemme 3.3. Montrons b). Soient y , z vérifiant (11). Comme d(y , z) = k ≥ d(x, x′ ) + δ , il existe z ′ ∈ géod(y , z) à distance d(x, x′ ) + δ de z . z z ′ (11) y x x′ a Alors z ′ vérifie d(a, z ′ ) = r + k − d(x, x′ ) − δ, y ∈ géod(z ′ , a) de façon évidente et z ′ ∈ δ -géod(x′ , a) d’après le lemme 3.11 appliqué à (x, x′ , a, z , z ′ ) au lieu de (x, x′ , z , y , y ′) et δ au lieu de ǫ. (cid:3) La propriété (10) résulte immédiatement du lemme suivant. 36 Lemme 3.27 Il existe C = C (δ, K ) tel que pour a, x, x′ ∈ X et r ∈ N vérifiant r < min(d(a, x), d(a, x′ )) la mesure de l’ensemble des t ∈ [0, 1] tels que µr,t(x, a) 6= µr,t(x′ , a) est ≤ C d(x,x′ ) d(a,x)−r . Démonstration. On utilise encore l’astuce des ensembles emboîtés. Si C0 = C (δ, K ) ma jore le cardinal des ensembles Ax,a,r,k pour tous a, x, r, k , l’ap- plication décroissante k 7→ Ax,a,r,k de {0, ..., d(x, a) − r} dans l’ensemble des parties non vides de X prend au plus C0 valeurs, et pour chaque va- leur, l’image inverse est un intervalle, et l’image inverse de cette même va- leur par l’application k 7→ Ax′ ,a,r,k de {0, ..., d(x′ , a) − r} dans l’ensemble des parties de X est un autre intervalle dont les extrémités diffèrent au plus de d(x, x′ ) + δ des extrémités du premier (il peut être vide si la lon- gueur du premier intervalle est inférieure ou égale à 2(d(x, x′ ) + δ )). Comme d(a, x) − d(a, x′ ) ≤ d(x, x′ ), il en résulte que l’application décroissante t 7→ Ax,a,r,E (t(d(a,x)−r)) prend au plus C0 valeurs et pour chaque valeur l’image inverse est un intervalle, et l’image inverse de cette même valeur par l’ap- plication t 7→ Ax′ ,a,r,E (t(d(a,x′ )−r)) est un autre intervalle dont les extrémités diffèrent au plus de 2d(x,x′ )+δ+1 des extrémités du premier. En effet pour t, t′ ∈ d(a,x)−r [0, 1] vérifiant l’inégalité E (t(d(a, x)−r))−E (t′ (d(a, x′ )−r)) ≤ d(x, x′ )+δ on a t − t′ ≤ 2d(x,x′ )+δ+1 . On en déduit que la mesure de l’ensemble des t ∈ [0, 1] d(a,x)−r tels que µr,t(x, a) 6= µr,t(x′ , a) est inférieure ou égale à 2C0 (2d(x,x′ )+δ+1) . Elle d(a,x)−r est donc inférieure ou égale à 2C0 (δ+3)d(x,x′ ) car elle est nulle si x = x′ . On d(a,x)−r peut prendre C = 2C0(δ + 3). (cid:3) Cette formule pour µr (x, a) paraît artificiellement compliquée mais son gros avantage pour nous est que µr (x, a) est une certaine moyenne sur k de νAx,a,r,k et que Ax,a,r,k ne dépend que de la connaissance des points de {a, x} ∪ {y ∈ δ -géod(x, a), d(a, y ) = r} ∪ {z ∈ δ -géod(x, a), d(a, z) = r + k} et de leurs distances mutuelles. Or le nombre de ces points est borné par une constante C (δ, K ) et les distances entre ces points sont égales à 0 ou d(x, a) ou d(x, a) − r ou d(x, a) − r − k ou r ou k ou r + k à une constante C (δ ) près. Donc lorsque d(a, x), r et k sont fixés, le nombre de ces points et la donnée de leurs distances mutuelles n’admettent qu’un nombre fini de possibilités, borné par C (δ, K ). On pose enfin µ0,t(x, a) = ea et µd(x,a),t (x, a) = ex pour tout t ∈ [0, 1]. Nous mettons la remarque précédente sous forme d’un lemme, qui servira ensuite. Lemme 3.28 Pour x, a ∈ X , r ∈ {0, . . . , d(a, x)} et t ∈ [0, 1], µr,t(x, a) est supporté par {y ∈ δ -géod(x, a), d(a, y ) = r} 37 et dépend seulement de la connaissance des points de {a, x} ∪ {y ∈ δ -géod(x, a), d(a, y ) = r} ∪{z ∈ δ -géod(x, a), d(a, z) = r + E (t(d(a, x) − r))} et de leurs distances mutuel les. Démonstration. La preuve est incluse dans la remarque précédente. (cid:3) Nous allons maintenant construire un nouveau parametrix ux pour ∂ . Dans les trois pages qui suivent nous écrivons up x : C(∆p−1 ) → C(∆p ) au lieu de ux pour rendre plus claire la construction, qui se fait par récurrence sur p. Autrement dit on va construire des morphismes upmax x→ C(∆1 ) u2 C(∆0 ) u1 x→ C(∆pmax ) x→ C(∆2 ) . . . tels que up x∂ + ∂up+1 x = IdC(∆p ) pour tout p ∈ {0, ..., pmax}. En fait on aura x = Zt∈[0,1] up up x,tdt et up x,t sera construit en même temps qu’une famille indéxée par r ∈ N d’en- domorphismes (v p x,r,t)p≥1 du complexe 0 ← C(∆0 ) ∂← C(∆1 ) ∂← C(∆2 ) . . . ∂← C(∆pmax ) ← 0 qui consistent en gros à rapprocher de l’origine x d’une longueur r , à l’aide des mesures µr,t(x, a). On aura v p x,0,t = IdC(∆p ) et quand r tend vers l’infini v p x,r,t : C(∆p ) → C(∆p ) tendra vers une limite v p x,∞,t égale à IdC(∆0 ) si p = 0, au morphisme de rang 1 de C(∆1 ) donné par ea 7→ ex si p = 1 et à 0 si p > 1 (autrement dit v p x,∞,t est indépendant de t et associé à la contraction de X sur x). Pour r ∈ N∗ et p ≥ 2 on va construire up x,r,t : C(∆p−1 ) → C(∆p ) tel x,r−1,t − v p x,r,t = up x,r,t∂ + ∂up+1 x,r,t = 0 on ait v p qu’en posant u1 x,r,t pour tout p. On en déduira que ∞ ∞ Xr=1 Xr=1 up+1 up Id − v p x,r,t(cid:1) pour tout p. x,r,t(cid:1)∂ + ∂ (cid:0) x,∞,t = (cid:0) x,t = P∞ r=1 up On posera alors up x,r,t pour p ≥ 2 et u1 x,t(e∅) = ex , de sorte que l’on aura ∂up+1 x,t + up x,t∂ = IdC(∆p ) pour tout p (grâce à (12) et au fait que ∂u1 x,t∂ = v 1 x,∞,t et u1 x,t = v 0 x,∞,t). x , up x,t, up On passe maintenant à la construction des opérateurs up x,r,t et v p x,r,t . (12) 38 On définit u1 x : C(∆0 ) → C(∆1 ) par u1 x (e∅) = ex et on pose u1 x,t = u1 x . Voici la formule pour u2 x : C(∆1 ) → C(∆2 ) : d(x,a) x (ea) = Z 1 Xr=1 u2 0 u2 x,r,t(ea )dt où (13) u2 x,r,t(ea ) = Φ1 (µr−1,t(x, a) − µr,t(x, a)). x (ea) = ea − ex = (1 − u1 Il est évident que ∂u2 x∂ )(ea ). Voici maintenant la formule pour u3 x : C(∆2 ) → C(∆3 ) . Par commodité nous étendons la fonction r 7→ µr,t(x, a) par µr,t(x, a) = ex si r > d(x, a). La formule est, pour {a, b} ∈ ∆2 : max(d(x,a),d(x,b)) x(ea ∧ eb ) = Z 1 Xr=1 u3 u3 x,r,t(ea ∧ eb )dt où 0 x,1,t(ea ∧ eb ) = Φ2(cid:16)ea ∧ eb + Φ1 (cid:0)ea − µ1,t(x, a)(cid:1) − Φ1 (cid:0)eb − µ1,t(x, b)(cid:1) u3 −Φ1 (cid:0)µ1,t(x, b) − µ1,t(x, a)(cid:1)(cid:17), et pour r > 1, (14) x,r,t(ea ∧ eb ) = Φ2(cid:16)Φ1 (cid:0)µr−1,t(x, b) − µr−1,t(x, a)(cid:1) + Φ1(cid:0)µr−1,t(x, a) − µr,t(x, a)(cid:1) u3 −Φ1 (cid:0)µr−1,t(x, b) − µr,t(x, b)(cid:1) − Φ1 (cid:0)µr,t(x, b) − µr,t(x, a)(cid:1)(cid:17). (15) On vérifie que (∂u3 x + u2 x∂ )(ea ∧ eb ) = ea ∧ eb . Afin de motiver la construction pour les plus grandes valeurs de p on va x,r,t : C(∆1 ) → réécrire les formules (13), (14), (15) à l’aide des opérateurs v 1 x,r,t : C(∆2 ) → C(∆2 ) définis par C(∆1 ) et v 2 x,0,t(ea) = ea , v 1 v 1 x,r,t(ea ) = µr,t(x, a), v 2 x,0,t(ea ∧ eb ) = ea ∧ eb , v 2 x,r,t(ea ∧ eb ) = Φ1(v 1 x,r,t(∂ (ea ∧ eb )) = Φ1 (µr,t(x, b) − µr,t(x, a)). On a alors x,r−1,t(ea ) − v 1 x,r,t(ea) = Φ1(v 1 u2 x,r,t(ea )), x,r,t(ea ∧ eb ) − u2 x,r−1,t(ea ∧ eb ) − v 2 x,r,t(ea ∧ eb ) = Φ2 (v 2 u3 x,r,t(∂ (ea ∧ eb ))). 39 u2 x,1,t(ea) v 1 x,1,t(ea ) ✛✘ u2 x,2,t(ea) ✚✙ v 2 x,1,t(ea ∧ eb ) v 1 x,2,t(ea ) ✛✘ ✚✙ v 2 x,2,t(ea ∧ eb ) a ea ∧ eb •x b u2 x,1,t(eb ) ✛✘ ✚✙ u2 x,2,t(eb ) v 1 x,1,t(eb ) ✛✘ ✚✙ v 1 x,2,t(eb ) Le dessin illustre les calculs précédents. Les 4 petits cercles contiennent les supports de v 1 x,r,t(ea ) = µr,t(x, a) et v 1 x,r,t(eb ) = µr,t(x, b) pour r = 1, 2 (en réalité ces supports ne seront pas disjoints mais on l’a supposé pour la clarté du dessin). Les six éléments v 2 x,r,t(ea ∧ eb ), u2 x,r,t(ea ), u2 x,r,t(eb ) (pour r = 1, 2) sont des combinaisons d’arêtes comme il est indiqué sur le dessin. Enfin les deux éléments u3 x,r,t(ea ∧ eb ) (pour r = 1, 2), qui ne sont pas représentés sur le dessin, sont des combinaisons de triangles remplissant les deux grands carrés. En général on définit up x : C(∆p−1 ) → C(∆p ) pour p ∈ {1, . . . , pmax} par ré- currence ascendante sur p. La récurrence part de p = 1 et fournit les formules ci-dessus pour p = 2, 3. On construit des applications linéaires up x,r,t : C(∆p−1 ) → C(∆p ) pour r ∈ N∗ et t ∈ [0, 1] et v p x,r,t : C(∆p ) → C(∆p ) pour r ∈ N et t ∈ [0, 1] telles que up x,r,t(ea1 ∧ ... ∧ eap−1 ) soit nul si r > max(d(x, a1 ), ..., d(x, ap−1)) et p ≥ 2 et v p x,r,t(ea1 ∧ ... ∧ eap ) soit nul si r ≥ max(d(x, a1), ..., d(x, ap)) et p ≥ 2. Ces applications sont définies par v p x,0,t = IdC(∆p ) pour p ≥ 1, et pour r ≥ 1, par récurrence ascendante sur p à l’aide des formules suivantes : u1 x,r,t = 0, v 1 x,r,t(ea ) = µr,t(x, a) pour r ≥ 1 40 pour initialiser et x,r,t(ea1 ∧ ... ∧ eap ) = Φp−1 (v p−1 v p x,r,t(∂ (ea1 ∧ ... ∧ eap ))) pour r ≥ 1 et p ≥ 2 x,r−1,t(ea1 ∧ ... ∧ eap−1 ) − v p−1 x,r,t(ea1 ∧ ... ∧ eap−1 ) = Φp−1(cid:0)v p−1 up x,r,t(ea1 ∧ ... ∧ eap−1 ) −up−1 x,r,t(∂ (ea1 ∧ ... ∧ eap−1 ))(cid:1) pour r ≥ 1 et p ≥ 2. Notons que ces formules impliquent les deux égalités x,r,t = v p−1 ∂ ◦ v p x,r,t ◦ ∂ pour p ≥ 2 et r ≥ 0 et x,r,t ◦ ∂ = v p−1 x,r−1,t − v p−1 x,r,t + up−1 ∂ ◦ up x,r,t pour p ≥ 2 et r ≥ 1. Pour p ∈ {2, . . . , pmax}, on définit ux : C(∆p−1 ) → C(∆p ) par ∞ ux = Z 1 Xr=1 up x,r,tdt 0 (la somme sur r est toujours finie, plus précisément pour ux(ea1 ∧ ... ∧ eap−1 ) la somme s’arrête à max(d(x, a1), ..., d(x, ap−1))). Alors ux et ∂ sont des en- domorphismes de degrés 1 et −1 de Lpmax p=0 C(∆p ) et on a ∂ux + ux∂ = 1. Remarque. Comme les applications Φi ne sont pas linéaires, les formules précédentes pour ux,r,t en fonction de µr,t(x, a) ne donnent pas de formule pour ux en fonction de µr (x, a). Autrement dit µr (x, a) ne sert à rien. Plus loin pour montrer l’équivariance à compact près des opérateurs nous n’utiliserons pas la propriété (10) mais le lemme 3.27. Le gros avantage du paramétrix ux sur Hx apparaît dans la proposition suivante, où l’on voit qu’il vérifie la condition (C2). Une propriété semblable sera vraie aussi pour le paramétrix définitif Jx et jouera un rôle crucial dans la construction de normes telles que Jx soit continu. Proposition 3.29 Pour p ∈ {2, . . . , pmax}, t ∈ [0, 1], S = {a1 , . . . , ap−1} ∈ ∆p−1 , \ supp(ux,r,t(ea1 ∧ ... ∧ eap−1 )) ⊂ a∈S,y∈δ-géod(x,a),d(y ,a)=r et ux,r,t(ea1 ∧ ... ∧ eap−1 ) ne dépend que de la connaissance des points de l’ensemble \ a∈S,y∈δ-géod(x,a),d(y ,a)=r {y ∈ δ -géod(x, a), d(y , a) = r + E (t(d(x, a) − r))} S ∪ {x} ∪ ∪ [a∈S ∪ [a∈S {y ∈ δ -géod(x, a), d(y , a) = r − 1 + E (t(d(x, a) − (r − 1)))} et des distances entre ces points. (16) B (y , N + 5pδ ) B (y , N + 5δ + 5pδ ) 41 Le nombre de points dans l’ensemble ci-dessus est borné par une constante C (δ, K, N ) et les distances entre ces points sont elles-mêmes déterminées par la connaissance de r , t et d(x, S ) à une constante C (δ, K, N ) près. Donc le nombre total de possibilités pour le nombre de ces points et leurs distances mutuelles (c’est-à-dire le nombre de classes d’équivalence de (16) par les isométries qui préservent ses parties S et {x}) est borné par C (δ, K, N ). Pour montrer la proposition 3.29 nous aurons besoin de trois lemmes. Lemme 3.30 Soient α, β ∈ N et x, a, b, y , z ∈ X vérifiant y ∈ α-géod(x, a) et z ∈ β -géod(x, b). Alors d(y , z) ≤ d(a, b) + d(a, y ) − d(b, z) + α + β + 2δ. x α β y z a b Démonstration. Par (H α δ (b, a, y , x)), on a d(b, y ) ≤ max(d(b, a) − d(a, y ), d(b, x) − d(x, y )) + α + δ et par (H β δ (y , b, z , x)) on a d(y , z) ≤ max(d(y , b) − d(b, z), d(y , x) − d(x, z)) + β + δ ≤ max (cid:0)d(a, b)−d(a, y )−d(b, z)+α+β+2δ, d(b, x)−d(x, y )−d(b, z)+α+β+2δ, d(x, y ) − d(x, z) + β + δ(cid:1) ≤ d(a, b) + d(a, y ) − d(b, z) + α + β + 2δ car d(b, x) − d(x, y ) − d(b, z) ≤ d(a, b) + d(a, x) − d(x, y ) − d(b, z) ≤ d(a, b) + d(a, y ) − d(b, z) ≤ d(a, b) + d(a, y ) − d(b, z), 42 et d(x, y ) − d(x, z) ≤ d(x, a) − d(a, y ) + α − d(x, b) + d(b, z) ≤ d(a, b) + d(a, y ) − d(b, z) + α. (cid:3) Lemme 3.31 Soit R, α ∈ R+ et x, y , z , t ∈ X vérifiant y , z ∈ B (x, R) et t ∈ α-géod(y , z). Alors t ∈ B (x, R + α + δ ). Démonstration. Par (H α δ (x, y , t, z)), on a d(x, t) ≤ max(d(x, y ) − d(y , t), d(x, z) − d(z , t)) + α + δ ≤ R + α + δ. (cid:3) Lemme 3.32 Pour tout p ∈ {1, . . . , pmax}, y ∈ X , R ∈ R+ , et f ∈ C(∆p ) tel le que supp(f ) ⊂ B (y , R), on a supp(Φp (f )) ⊂ B (y , R + 5δ ) et Φp (f ) ne dépend que de f et de la connaissance des points de B (y , R + 10δ ) et de leurs distances mutuel les. Démonstration. Cela résulte des lemmes 3.25 et 3.31. (cid:3) Démonstration de la proposition 3.29. La réunion des supports des mesures µr,t(x, a1 ), ..., µr,t(x, ap−1), µr−1,t(x, a1 ), ..., µr−1,t(x, ap−1) est incluse dans {y ∈ δ -géod(x, a), d(y , a) = r} [a∈S ∪ [a∈S {y ∈ δ -géod(x, a), d(y , a) = r − 1}. Le lemme 3.30 (avec α = β = δ , a, b ∈ S , y ∈ δ -géod(x, a) à distance r ou r − 1 de a, et z ∈ δ -géod(x, b) à distance r ou r − 1 de b) implique que cet ensemble est de diamètre ≤ N + 4δ + 1 ≤ N + 5δ et qu’il est donc inclus dans \ B (y , N + 5δ ). a∈S,y∈δ-géod(x,a),d(y ,a)=r Grâce au lemme 3.32, on montre par récurrence sur p l’assertion de la propo- sition 3.29 en même temps que l’assertion analogue pour vx,r,t , à savoir que pour p ∈ {2, . . . , pmax}, t ∈ [0, 1], S = {a1 , . . . ap} ∈ ∆p , \ supp(vx,r,t(ea1 ∧ ... ∧ eap )) ⊂ a∈S,y∈δ-géod(x,a),d(y ,a)=r ou r+1 B (y , N + 5pδ ) 43 et vx,r,t(ea1 ∧ ... ∧ eap ) ne dépend que de la connaissance des points de l’en- semble \ a∈S,y∈δ-géod(x,a),d(y ,a)=r ou r+1 {y ∈ δ -géod(x, a), d(y , a) = r + E (t(d(x, a) − r))} S ∪ {x} ∪ ∪ [a∈S et des distances entre tous ces points. B (y , N + 5δ + 5pδ ) (cid:3) Lemme 3.33 Il existe C = C (δ, K, N ) tel le que pour tout p ∈ {1, ..., pmax} et pour tout S ∈ ∆p−1 , r ∈ N et t ∈ [0, 1] on ait kux,r,t(eS )kℓ1 (∆p ) ≤ C . Démonstration. Cela résulte de la dernière assertion du lemme 3.25 et du fait que dans les formules pour ux,r,t on applique les Φi à des fonctions dont le support est de diamètre ≤ 2(N + 5δpmax ) d’après la démonstration de la proposition 3.29. (cid:3) 3.4 Construction du paramétrix définitif Jx Nous allons introduire un opérateur Hx , sorte de troncature de Hx , et Jx sera donné par la formule Jx = Hx + ux (1 − ∂ Hx − Hx∂ ). En effet le gros inconvénient de ux est que l’on sait seulement, par la proposition 3.29 et le lemme 3.2, que, pour S ∈ ∆ \ {∅}, ux (eS ) est sup- porté par Ta∈S (cid:0)2(N + 5pmaxδ ) + δ(cid:1)-géod(x, a). Autrement dit ux ne vérifie pas la condition (C1) et la seule chose que l’on puisse affirmer est que ux “n’éloigne pas de x de plus que de N + 5pmaxδ + δ ”. Cela compliquerait beau- coup la construction de normes pour lesquelles ∂ et ux soient continus, voire la rendrait impossible. Au contraire nous verrons dans la proposition 3.37 que l’opérateur Jx vérifie les conditions (C1) et (C2). L’idée pour montrer que Jx vérifie (C1) est que (1 − ∂ Hx − Hx∂ ) rapproche davantage de x que ux n’en éloigne. L’idée pour montrer que Jx vérifie (C2) est que, contrairement à Hx , Hx vérifie (C2) grâce au fait que c’est une troncature. On peut écrire formellement Hx sous la forme Hx = +∞ Xq=1 Cette somme est infinie (mais pour chaque S ∈ ∆, Hx (eS ) est donné par une somme finie, grâce au lemme 3.17). On fixe un entier Q tel que hx (1 − (∂hx + hx∂ ))q−1 . (HQ) : Q soit assez grand en fonction de δ, K, N . 44 Dans la suite nous utiliserons un nombre fini de fois l’inégalité Q ≥ C avec C de la forme C (δ, K, N ). On pose Hx = hx (1 − (∂hx + hx∂ ))q−1 et Kx = 1 − (∂ Hx + Hx∂ ). Q Xq=1 Il est clair que Kx = (1 − (∂hx + hx∂ ))Q . Le lemme suivant résume les propriétés de Hx qui nous serviront ensuite. La propriété principale est que Kx rapproche strictement de l’origine. On pose Kx,0 = 1 et pour q ∈ {1, ..., Q} et (t1 , . . . , tq ) ∈ [0, 1]q , on note Kx,q ,(t1 ,...,tq ) = (1 − (∂hx,tq + hx,tq ∂ ))...(1 − (∂hx,t1 + hx,t1 ∂ )) Hx,q ,(t1 ,...,tq ) = hx,tq Kx,q−1,(t1 ,...,tq−1 ) et = hx,tq (1 − (∂hx,tq−1 + hx,tq−1 ∂ ))...(1 − (∂hx,t1 + hx,t1 ∂ )) de sorte que Kx,Q,(t1 ,...,tQ )dt1 . . . dtQ . B (a, r). (17) Hx,q ,(t1 ,...,tq )dt1 . . . dtq Q Xq=1 Z(t1 ,...,tq )∈[0,1]q Hx = et Kx = Z(t1 ,...,tQ )∈[0,1]Q Notation. Dans toute la suite de l’article, si A est une partie de X et r ∈ N, on notera B (A, r) = {y ∈ X, d(y , A) ≤ r} = [a∈A Lemme 3.34 1) Pour q ∈ {1, ..., Q}, (t1 , . . . , tq ) ∈ [0, 1]q et S ∈ ∆ \ {∅}, a) Hx,q ,(t1 ,...,tq )(eS ) est une combinaison de eT où T vérifie : T ⊂ S ∪ (cid:16)B (x, 2δ ) ∩ B (S, qN )(cid:17) ∪ [a∈S,a6∈B (x,2δ) (cid:8)y ∈ 4δ -géod(x, a), d(y , a) ∈]N − 2δ, qN ](cid:9), b) Hx,q ,(t1 ,...,tq ) (eS ) ne dépend que de la connaissance des points de B (x, 7δ ) ∪ [a∈S ∪ [a∈S,i∈{1,...,q} {y ∈ 9δ -géod(x, a), d(x, y ) − tid(x, a) ≤ (q + 2)N + 1} (18) {y ∈ 4δ -géod(x, a), d(y , a) ≤ qN } 45 {y ∈ 4δ -géod(x, a), d(y , a) ∈ [Q et des distances entre ces points. 2) Pour (t1 , . . . , tQ ) ∈ [0, 1]Q et S ∈ ∆ \ {∅}, a) Kx,Q,(t1 ,...,tQ )(eS ) est une combinaison de eT où T vérifie : N − 6δ T ⊂ [a∈S pmax − 2N , QN ]}. b) Kx,Q,(t1 ,...,tQ )(eS ) ne dépend que de la connaissance des points de B (x, 7δ ) ∪ [a∈S ∪ [a∈S,i∈{1,...,Q} {y ∈ 9δ -géod(x, a), d(x, y ) − tid(x, a) ≤ (Q + 2)N + 1} et des distances entre ces points. {y ∈ 4δ -géod(x, a), d(y , a) ≤ QN } Démonstration. Montrons 1)a). C’est essentiellement le lemme 3.22 mais on doit en répéter les arguments parce qu’on a ici hx,ti et non hx . Soit S0 = S et S1 , . . . , Sq tels que eSi apparaisse avec un coefficient non nul dans (1−(∂hx,ti + hx,ti ∂ ))(eSi−1 ) pour i = 1, . . . , q − 1 et que eSq apparaisse avec un coefficient non nul dans hx,tq (eSq−1 ). Soit yq ∈ Sq n’appartenant pas à B (x, 2δ ). Par le lemme 3.19, il existe yq−1 ∈ Sq−1 ,..., y0 ∈ S0 n’appartenant pas à B (x, 2δ ) tels que yi = yi+1 ou bien yi+1 ∈ 3δ -géod(x, yi ) et d(yi , yi+1) ∈]N − 2δ, N ] pour i ∈ {0, . . . , q − 1}. En répétant la preuve du lemme 3.22 on montre que yq ∈ 4δ -géod(x, y0) et d(y0 , yq ) > N − 2δ si y0 6= yq . Comme d(yi , yi+1) ≤ N pour i ∈ {0, . . . , q − 1} on a d(y0 , yq ) ≤ qN . Montrons b). Soient q ∈ {1, ..., Q}, S0 = S et S1 , . . . , Sq−1 tels que eSi apparaisse avec un coefficient non nul dans (1 − (∂hx,ti + hx,ti ∂ ))(eSi−1 ) pour i = 1, . . . , q − 1. Soit i ∈ {1, . . . , q}. Grâce au lemme 3.20, la connaissance de (1 − (∂hx,ti + hx,ti ∂ ))(eSi−1 ) si i < q ou de hx,ti (eSi−1 ) si i = q ne dépend que de la connaissance des points de B (x, 2δ ) et de la réunion pour b ∈ Si−1 des ensembles {b} ∪ {y ∈ 3δ -géod(x, b), d(y , b) ∈]N − 2δ, N ]} ∪{y ∈ 5δ -géod(x, b), d(x, y ) ∈ [tid(x, Si−1) − 2N − 1, tid(x, Si−1) + N ]} (19) (en effet si T ∈ ∆ \ {∅} est tel que eT apparaisse avec un coefficient non nul dans ∂ (eSi−1 ) on a T ⊂ Si−1 donc d(x, T ) ∈ [d(x, Si−1), d(x, Si−1) + N ]). De plus la preuve de a) montre que pour tout b ∈ Si−1 , on a b ∈ S0 ∪ B (x, 2δ ) ou il existe a ∈ S0 tel que a 6∈ B (x, 2δ ), b ∈ 4δ -géod(x, a) et d(a, b) ∈]N − 2δ, (q − 1)N ]. (20) 46 Si b ∈ S0 ∪ B (x, 2δ ) on vérifie facilement que l’ensemble (19) est inclus dans l’ensemble (18) car d(x, Si−1) − d(x, a) ≤ iN ≤ qN . Si a ∈ S0 vérifie (20), l’ensemble (19) est également inclus dans l’ensemble (18) car – si y ∈ 3δ -géod(x, b) et d(y , b) ∈]N − 2δ, N ], en appliquant le lemme 3.21 avec α = 4δ , β = 3δ on a y ∈ 4δ -géod(x, a), et de plus d(a, y ) ≤ qN , – si y ∈ 5δ -géod(x, b) et d(x, y ) ∈ [tid(x, Si−1) − 2N − 1, tid(x, Si−1) + N ] le a) du lemme 3.3 montre y ∈ 9δ -géod(x, a) et comme d(x, Si−1) − d(x, a) ≤ qN on a d(x, y ) − tid(x, a) ≤ (q + 2)N + 1. La preuve de 2) est tout à fait similaire à celle de 1), sauf que 2)a) nécessite un argument supplémentaire. Si eT apparaît avec un coefficient non nul dans Kx,Q,(t1 ,...,tQ ) (eS ), le lemme 3.17 implique ζx(T ) ≤ ζx(S ) − Q N −6δ . Le sous- pmax lemme suivant, appliqué à M = Q N −6δ , montre alors que pour tout y ∈ T pmax et a ∈ S on a d(a, y ) ≥ Q N −6δ pmax − 2N . Sous-lemme 3.35 Soient M ∈ R+ et S, T ∈ ∆. Si ζx(T ) ≤ ζx(S ) − M , pour y ∈ T et a ∈ S , on a d(x, y ) ≤ d(x, a) − M + 2N . Démonstration. On a d(x, y ) ≤ d(x, UT ) + N ≤ ζx(T ) + N et d(x, a) ≥ d(x, US ) ≥ ζx(S ) − N . (cid:3) Fin de la démonstration du lemme 3.34. On suppose Q N −6δ pmax − 2N > N , ce qui est permis par (HQ), d’où y 6∈ S . Enfin S est non vide, donc B (x, 2δ ) ⊂ Sa∈S 4δ -géod(x, a). (cid:3) Comme nous l’avons déjà dit nous posons Jx = Hx + uxKx . Montrons que ∂Jx + Jx∂ = 1. D’abord comme ∂ 2 = 0, ∂ commute à Kx = 1 − ∂ Hx − Hx∂ et donc ∂Jx + Jx∂ = (∂ Hx + Hx∂ ) + (∂ux + ux∂ )Kx = 1 puisque ∂ux + ux∂ = 1. On a Q Xq=1 Z(t1 ,...,tq )∈[0,1]q +∞ Xr=1 Zt,(t1 ,...,tQ )∈[0,1]Q+1 ux,r,tKx,Q,(t1 ,...,tQ )dtdt1 . . . dtQ . Il résulte de tout ce qui précède que Hx,q ,(t1 ,...,tq ) (eS ) et ux,r,tKx,Q,(t1 ,...,tQ )(eS ) ne dépendent que de la connaissance d’un nombre fini de points (borné par Hx,q ,(t1 ,...,tq )dt1 . . . dtq Jx = + 47 T ⊂ B (z , N + 5pmaxδ ), −2N −5δ+r,QN +r ] C (δ, K, N , Q)) et des distances entre ces points. De plus ces distances sont déterminées par x, S, q , r, t et les ti à une constante C (δ, K, N , Q) près. Plus précisément le 1) du lemme 3.34 apporte toutes les informations relatives à Hx,q ,(t1 ,...,tq ) (eS ), et le lemme suivant les donne pour ux,r,tKx,Q,(t1 ,...,tQ ) (eS ). Lemme 3.36 Pour t ∈ [0, 1], (t1 , . . . , tQ) ∈ [0, 1]Q , r ∈ N∗ et S ∈ ∆ \ {∅}, a) ux,r,tKx,Q,(t1 ,...,tQ )(eS ) est une combinaison de eT où T vérifie : [ a∈S,z∈5δ-géod(x,a),d(z ,a)∈[Q N −6δ pmax b) ux,r,tKx,Q,(t1 ,...,tQ )(eS ) ne dépend que de la connaissance des points de B (x, 7δ ) ∪ [a∈S ∪ [a∈S,i∈{1,...,Q} {y ∈ 9δ -géod(x, a), d(x, y ) − tid(x, a) ≤ (Q + 2)N + 1} [ ∪ a∈S,z∈5δ-géod(x,a),d(z ,a)∈[Q N −6δ −2N −5δ+r,QN +r ] pmax ∪ [a∈S (cid:8)z ∈ 5δ -géod(x, a), d(x, z) ∈ [(1 − t)(d(x, a) − QN − r), (1 − t)(d(x, a) − r) + 2 + δ ](cid:9) et des distances entre ces points. Démonstration. Montrons a). D’après le 2)a) du lemme 3.34 et la proposi- tion 3.29, {y ∈ 4δ -géod(x, a), d(y , a) ≤ QN } B (z , N + 5pmaxδ + 5δ ) ux,r,tKx,Q,(t1 ,...,tQ ) (eS ) est supporté par la réunion des B (z , N + 5pmaxδ ) pour les z tels que z ∈ δ -géod(x, y ) et d(y , z) = r , avec y ∈ 4δ -géod(x, a), a ∈ S et d(y , a) ∈ [Q N −6δ pmax − 2N , QN ]. Alors z ∈ 5δ -géod(x, a) par le a) du lemme 3.3. On a aussi y ∈ 5δ -géod(z , a) par le b) du lemme 3.3, d’où N − 6δ d(z , a) ∈ [Q pmax − 2N − 5δ + r, QN + r ]. Montrons b). D’après le 2)b) du lemme 3.34, Kx,Q,(t1 ,...,tQ ) (eS ) dépend de la connaissance des points de B (x, 7δ ) ∪ [a∈S ∪ [a∈S,i∈{1,...,Q} {y ∈ 9δ -géod(x, a), d(x, y ) − tid(x, a) ≤ (Q + 2)N + 1}. {y ∈ 4δ -géod(x, a), d(y , a) ≤ QN } 48 D’après le 2)a) du lemme 3.34 et la proposition 3.29, ux,r,tKx,Q,(t1 ,...,tQ ) (eS ) ∪ B (z , N + 5pmaxδ + 5δ ) N − 6δ pmax − 2N , QN ]} dépend de la connaissance de Kx,Q,(t1 ,...,tQ )(eS ) et des points de {x} ∪ [a∈S {y ∈ 4δ -géod(x, a), d(y , a) ∈ [Q [ a∈S,y∈4δ-géod(x,a),d(y ,a)∈[Q N −6δ −2N ,QN ],z∈δ-géod(x,y),d(y ,z )=r pmax [ −2N ,QN ] (cid:8)z ∈ δ -géod(x, y ), ∪ a∈S,y∈4δ-géod(x,a),d(y ,a)∈[Q N −6δ pmax d(y , z) = r + E (t(d(x, y ) − r)) ou d(y , z) = (r − 1) + E (t(d(x, y ) − (r − 1)))(cid:9) Les deux derniers ensembles de la réunion ci-dessus sont inclus dans les deux derniers ensembles de la réunion figurant dans b) du lemme 3.36, pour les raisons suivantes. La preuve de a) montre que pour a ∈ S les conditions N − 6δ y ∈ 4δ -géod(x, a), d(y , a) ∈ [Q pmax −2N , QN ], z ∈ δ -géod(x, y ), d(y , z) = r impliquent N − 6δ z ∈ 5δ -géod(x, a), d(z , a) ∈ [Q pmax − 2N − 5δ + r, QN + r ]. D’autre part pour a ∈ S les conditions N − 6δ y ∈ 4δ -géod(x, a), d(y , a) ∈ [Q pmax − 2N , QN ], z ∈ δ -géod(x, y ), d(y , z) = r + E (t(d(x, y ) − r)) ou d(y , z) = (r − 1) + E (t(d(x, y ) − (r − 1))) impliquent z ∈ 5δ -géod(x, a), r + t(d(x, y ) − r) − 2 ≤ d(y , z) ≤ r + t(d(x, y ) − r), d(x, z) ≥ d(x, y ) − d(y , z) ≥ (1 − t)(d(x, y ) − r), d(x, z) ≤ d(x, y ) − d(y , z) + δ ≤ (1 − t)(d(x, y ) − r) + 2 + δ, d(x, y ) ≥ d(x, a) − d(y , a) ≥ d(x, a) − QN et N − 6δ d(x, y ) ≤ d(x, a) − d(y , a) + 4δ ≤ d(x, a) − (Q pmax − 2N ) + 4δ ≤ d(x, a) car on suppose (Q N −6δ pmax − 2N ) ≥ 4δ (ce qui est permis par (HQ)). 49 (cid:3) Le 1) et le 2) de la proposition suivante récapitulent la partie des lemmes 3.34 et 3.36 qui nous sera utile ensuite sous une forme plus lisible. On suppose N − 6δ N ≥ 6δ + 1, Q ≥ 2 et Q pmax ≥ 2(3N + 5δpmax + 10δ ), ce qui est permis par (HN ) et (HQ ). On pose (21) (22) F = 15δ + 2N + 10δpmax . On a donc F ≤ C (δ, K, N ). Il est important de souligner que F est une simple notation permettant d’alléger les formules. Au contraire les constantes N , Q et d’autres qui seront introduites ensuite, sont des paramètres dans notre construction et chacun de ces paramètres doit être choisi suffisamment grand par rapport à ceux introduits auparavant. Proposition 3.37 1) Pour q ∈ {1, ..., Q}, (t1 , . . . , tq ) ∈ [0, 1]q et S ∈ ∆ \ {∅}, a) Hx,q ,(t1 ,...,tq )(eS ) est une combinaison de eT où T vérifie : T ⊂ S ∪ (cid:16)B (x, 2δ ) ∩ B (S, qN )(cid:17) ∪ [a∈S (cid:8)y ∈ 4δ -géod(x, a), d(y , a) ∈]N − 2δ, qN ](cid:9), b) Hx,q ,(t1 ,...,tq ) (eS ) ne dépend que de la connaissance des points de B (x, 7δ ) ∪ B (S, QN ) ∪ [a∈S,i∈{1,...,q} {y ∈ F -géod(x, a), d(x, y ) − tid(x, a) ≤ QF } et des distances entre ces points. 2) Pour t ∈ [0, 1], (t1 , . . . , tQ) ∈ [0, 1]Q , r ∈ N∗ et S ∈ ∆ \ {∅}, a) ux,r,tKx,Q,(t1 ,...,tQ )(eS ) est une combinaison de eT où T vérifie : T ⊂ [a∈S b) ux,r,tKx,Q,(t1 ,...,tQ )(eS ) ne dépend que de la connaissance des points de B (x, F ) ∪ B (S, QN ) ∪ [a∈S ∪ [a∈S,i∈{1,...,Q} {y ∈ F -géod(x, a), d(x, y ) − tid(x, a) ≤ QF } ∪ [a∈S {y ∈ F -géod(x, a), d(x, y ) − (1 − t)(d(x, a) − r) ≤ QF } {y ∈ F -géod(x, a), d(y , a) ∈ [r, r + QF ]} Q F {z ∈ F -géod(x, a), d(z , a) ∈ [ + r, QF + r ]}, 50 et des distances entre ces points. 3) Il existe C = C (δ, K, N , Q) tel que dans les notations de 1) et 2) on ait k Hx,q ,(t1 ,...,tq )(eS )kℓ1 ≤ C et kux,r,tKx,Q,(t1 ,...,tQ ) (eS )kℓ1 ≤ C. On remarque que le nombre de points des ensembles apparaissant dans l’énoncé est borné par C (δ, K, N , Q) et que les distances entre ces points sont déterminées à C (δ, K, N , Q) près par d(x, S ) et les divers paramètres (à savoir q , t1 , . . . , tq dans 1) et r, t, t1 , . . . , tQ dans 2)). Donc le nombre de com- binaisons possibles pour le nombre de ces points et leurs distances mutuelles est borné par C (δ, K, N , Q). Par des arguments similaires à ceux de la preuve du lemme 3.22 on peut montrer facilement, à l’aide de 1)a) et 2)a) du lemme précédent, que Jx vérifie la condition (C1) (en fait ces arguments seront cachés dans l’étude des normes menée au paragraphe 4). D’autre part 1)b) et 2)b) garantissent que Jx vérifie (C2). Démonstration. Le 1) résulte du 1) du lemme 3.34 et le 2) résulte du lemme 3.36. En particulier pour montrer 2) on utilise les inégalités suivantes, qui découlent de (21) et (22) : QN + N + 5δpmax + 5δ ≤ QF , (Q + 2)N + 1 ≤ QF N − 6δ N − 6δ 2pmax ≥ pmax − 2N − 5δ − N − 5δpmax − 5δ ≥ Q et Q On utilise aussi le fait que [ a∈S,z∈5δ-géod(x,a),d(z ,a)∈[Q N −6δ pmax ⊂ [a∈S {z ∈ (5δ + 2(N + 5pmaxδ + 5δ ))-géod(x, a), N − 6δ pmax − 2N − 5δ + r − (N + 5pmaxδ + 5δ ), d(z , a) ∈ [Q QN + r + (N + 5pmaxδ + 5δ )]} grâce au lemme 3.2, ainsi que d’autres inclusions analogues ou plus faciles. Enfin 3) résulte du lemme 3.33 et du fait que khx,t(eS )kℓ1 ≤ 1 pour tout S puisque hx,t(eS ) = ψS,x,t ∧ eS et que ψS,x,t est une mesure de probabilité. (cid:3) B (z , N + 5pmaxδ + 5δ ) −2N −5δ+r,QN +r ] Q F . 3.5 Construction d’une distance moyennée pour conju- guer les opérateurs On montrera dans le paragraphe 4 que l’image de 1 dans KKG,2sℓ+C (C, C) est représenté par l’opérateur impair ∂ + Jx agissant sur le complété 51 de Lpmax p=1 C(∆p ) pour une certaine norme de Hilbert k.kHx,s . Pour construire la première partie de l’homotopie de 1 à γ (qui fait l’ob jet de ce paragraphe et du p=1 C(∆p ) → Lpmax x : Lpmax suivant), on conjuguera ∂ +Jx par eτ θ♭ p=1 C(∆p ) est x où θ♭ défini par θ♭ x (eS ) = ρ♭ x (S )eS , où ρ♭ x est une variante moyennée de la distance à x et où τ varie de 0 à T (avec T assez grand). On doit moyenner la distance à x pour qu’elle se comporte mieux quand on change l’origine x en g (x) pour g ∈ G. Cela est nécessaire pour que l’opérateur ∂ + Jx conjugué par eτ θ♭ x soit G-équivariant à compact près. La suite de l’homotopie (qui est reléguée au paragraphe 5) sera facile : x (∂ + Jx )e−T θ♭ comme T est assez grand, l’opérateur eT θ♭ x est continu pour la norme ℓ2 , donc par une homotopie et tout en gardant cet opérateur on peut remplacer la norme k.kHx,s par la norme ℓ2 sur Lpmax p=1 C(∆p ) et il est alors très simple de terminer l’homotopie en aboutissant à γ dans KKG (C, C). Le but de ce sous-paragraphe est la construction de ρ♭ x . La proposition suivante renforce le théorème 17 de [MY02]. On rappelle que (X, d) est un bon espace discret δ -hyperbolique. Proposition 3.38 Il existe une distance G-invariante d′′ sur X tel le que d − d′′ est borné et que + , ∃R ∈ R+ , ∀x, x′ , y , y ′ ∈ X, d(x, x′ ) ≤ r, d(y , y ′) ≤ r, ∀r ∈ R+ , ∀ǫ ∈ R∗ d(x, y ) ≥ R, on ait d′′ (x, y ) − d′′ (x, y ′) − d′′ (x′ , y ) + d′′ (x′ , y ′) ≤ ǫ. (23) La proposition 3.38 est une conséquence du lemme suivant. Lemme 3.39 Il existe une famil le de mesures positives (µ(x, y ))(x,y)∈X×X de masse 1 sur X de sorte que – pour tout g ∈ G, µ(gx, gy ) = g∗µ(x, y ), – µ(x, y ) = µ(y , x), – le support de µ(x, y ) est inclus dans 7δ -géod(x, y ), – et ∀r ∈ R+ , ∀ǫ ∈ R∗ + , ∃R ∈ R+ , ∀x, x′ , y , y ′ ∈ X, avec d(x, x′ ) ≤ r, d(y , y ′) ≤ r, d(x, y ) ≥ R, on a kµ(x, y ) − µ(x′ , y ′)k1 ≤ ǫ. La première condition est simplement la condition naturelle de G-équivariance. Démonstration de la proposition 3.38 en admettant le lemme 3.39. On pose d′ (x, y ) = ZX (d(x, z) + d(z , y ))dµ(x, y )(z). (24) 52 On a d(x, y ) ≤ d′ (x, y ) ≤ d(x, y ) + 7δ pour x, y ∈ X et d′ vérifie la condition (23) de la proposition 3.38. En effet soit r ∈ R+ , ǫ ∈ R∗ + . Soit R, et x, y , x′ , y ′ comme dans la dernière assertion du lemme 3.39. Pour z ∈ 7δ -géod(x, y ), on a z ∈ (7δ +2r)-géod(x, y ′), z ∈ (7δ +2r)-géod(x′ , y ), et z ∈ (7δ +4r)-géod(x′ , y ′) par le lemme 3.2, donc d′ (x′ , y ) − ZX (d(x′ , z) + d(z , y ))dµ(x, y )(z) = (cid:12)(cid:12)(cid:12) ZX (d(x′ , z) + d(z , y ))(dµ(x′ , y ) − dµ(x, y ))(z)(cid:12)(cid:12)(cid:12) = (cid:12)(cid:12)(cid:12) ZX (d(x′ , z) + d(z , y ) − d(x′ , y ))(dµ(x′ , y ) − dµ(x, y ))(z)(cid:12)(cid:12)(cid:12) ≤ ǫ(7δ + 2r) d′(x, y ′ ) − ZX (d(x, z) + d(z , y ′))dµ(x, y )(z) ≤ ǫ(7δ + 2r) et de même d′ (x′ , y ′) − ZX (d(x′ , z) + d(z , y ′))dµ(x, y )(z) ≤ ǫ(7δ + 4r). et D’autre part (d(x, z) + d(z , y )) − (d(x′ , z) + d(z , y )) − (d(x, z) + d(z , y ′)) +(d(x′ , z) + d(z , y ′ )) = 0 pour tout z ∈ X . On en déduit d′ (x, y ) − d′ (x, y ′) − d′ (x′ , y ) + d′ (x′ , y ′) ≤ ǫ(21δ + 8r). Cependant d′ n’est pas nécessairement une distance. Comme d(x, y ) ≤ d′ (x, y ) ≤ d(x, y ) + 7δ pour x, y ∈ X on a d′ (u, w) ≤ d′ (u, v ) + d′ (v , w) + 7δ pour u, v , w ∈ X . On pose alors d′′ (x, y ) = 0 si x = y et d′′ (x, y ) = d′ (x, y ) + 7δ si x 6= y . Alors d′′ est une distance et on a montré la proposition 3.38 en admettant le lemme 3.39. (cid:3) Avant de montrer le lemme 3.39 on commence par un lemme général très utile. Lemme 3.40 Soient x, y ∈ X , α, γ ∈ N, a ∈ α-géod(x, y ), c ∈ γ -géod(x, y ), 2 et d(b, c) ≥ γ et b ∈ géod(a, c), avec d(a, b) ≥ α 2 , alors b ∈ 3δ -géod(x, y ). a c ≥ α/2 ≥ γ /2 b γ α x y 53 Démonstration. Par (H 0 δ (x, a, b, c)) et (H 0 δ (y , a, b, c)) on a d(x, b) ≤ max(d(x, a) − d(a, b), d(x, c) − d(b, c)) + δ d(y , b) ≤ max(d(y , a) − d(a, b), d(y , c) − d(b, c)) + δ. En additionnant ces deux inégalités on obtient et d(x, b) + d(b, y ) ≤ max(d(x, a) + d(a, y ) − 2d(a, b), d(x, c) + d(c, y ) − 2d(b, c), d(x, a) + d(y , c) − d(a, c), d(x, c) + d(y , a) − d(a, c)) + 2δ. Comme d(a, b) ≥ α 2 on a d(x, a) + d(a, y ) − 2d(a, b) ≤ d(x, y ) et comme d(b, c) ≥ γ 2 on a d(x, c) + d(c, y ) − 2d(b, c) ≤ d(x, y ), donc d(x, b) + d(b, y ) ≤ max (cid:0)d(x, y ) + 2δ, d(x, a) + d(y , c) − d(a, c) + 2δ, d(x, c) + d(y , a) − d(a, c) + 2δ(cid:1). Si d(x, a)+d(y , c) ≤ d(a, c)+d(x, y )+δ et d(x, c)+d(y , a) ≤ d(a, c)+d(x, y )+δ on a fini. Dans le cas contraire, supposons par exemple d(x, a) + d(y , c) > d(a, c) + d(x, y ) + δ. Grâce à (Hδ (x, y , a, c)) qui s’écrit d(x, a) + d(y , c) ≤ max(d(x, y ) + d(a, c), d(x, c) + d(y , a)) + δ on a alors d(x, c) + d(y , a) ≥ d(x, a) + d(y , c) − δ > d(a, c) + d(x, y ) d’où (cid:0)d(x, a) + d(y , c)(cid:1) + (cid:0)d(x, c) + d(y , a)(cid:1) > 2d(a, c) + 2d(x, y ) + δ. Or (cid:0)d(x, a) + d(y , c)(cid:1) + (cid:0)d(x, c) + d(y , a)(cid:1) ≤ 2d(x, y ) + α + γ ≤ 2d(a, c) + 2d(x, y ). Cette contradiction achève la démonstration du lemme 3.40. (cid:3) Démonstration du lemme 3.39. Comme le lemme 3.39 ne servira pas dans la suite nous abrégeons sa démonstration. On rappelle que pour toute partie A non vide de X , νA désigne la mesure de masse 1 égale au produit par (♯A)−1 de la fonction caractéristique de A. On pose alors φ(t) = E (t/8), et µ(x, y ) = 1 (φ(d(x, y )) + 1)4 54 φ(d(x,y)) Xk ,k,l,l=0 νBx,y ,k,k,l,l où pour k , k , l, l ∈ {0, . . . , φ(d(x, y ))} on note Bx,y ,k ,k,l,l = {z , ∃( x, y) ∈ 3δ -géod(x, y )2 , d(x, x) ≤ k , d( x, z) ≥ k , d(y , y) ≤ l, d(z , y) ≥ l, z ∈ géod( x, y)}. x 3δ ≥ k z ≥ l y 3δ ≤ l y ≤ k x Il est clair que µ(x, y ) = µ(y , x). Montrons que le support de µ(x, y ) est in- clus dans 7δ -géod(x, y ). Soient z , x et y comme dans la définition de Bx,y ,k ,k,l,l . Si d( x, z) ≥ 2δ et d( y , z) ≥ 2δ le lemme 3.40 (avec α = β = 3δ ) implique z ∈ 3δ -géod(x, y ). Si d( x, z) ≤ 2δ ou d( y , z) ≤ 2δ on a z ∈ 7δ -géod(x, y ) par le lemme 3.2. Pour montrer la dernière assertion du lemme 3.39 on uti- lise l’astuce des ensembles emboîtés, qui apparaît dans la démonstration de la proposition 6.9 de [KS03]. Le cardinal de Bx,y ,k ,k,l,l est toujours compris entre d(x,y) et C2d(x, y ) pour deux constantes C1 et C2 du type C (δ, K ). De plus C1 (k , k , l, l) 7→ Bx,y ,k ,k,l,l est une application croissante en k et l et décroissante en k et l. Enfin étant donnés x, y , x′ , y ′ ∈ X , en posant d = d(x, x′ ) + d(y , y ′) + 2δ on a Bx,y ,k ,k,l,l ⊂ Bx′ ,y ′ ,k+2d,k−d,l+2d,l−d pour k , l ≥ d. En effet soit z ∈ Bx,y ,k ,k,l,l et x, y ∈ 3δ -géod(x, y )2 tels que d(x, x) ≤ k , d( x, z) ≥ k , d(y , y) ≤ l, d(z , y) ≥ l, z ∈ géod( x, y). Par le lemme 3.2, x et y appartiennent à (cid:0)3δ + 2d(x, x′ ) + 2d(y , y ′)(cid:1)-géod(x′ , y ′). Soit x′ un point de géod( x, z) à distance d de x et y ′ un point de géod( y , z) à distance d de y . D’après le lemme 3.40, x′ et y ′ appartiennent à 3δ -géod(x′ , y ′) et donc z appartient à Bx′ ,y ′ ,k+2d,k−d,l+2d,l−d . Pour conclure la démonstration du lemme 3.39 on applique à ∆ = d(x, y ), ∆′ = d(x′ , y ′), nk ,k,l,l = ♯Bx,y ,k ,k,l,l et n′ k ,k,l,l = ♯Bx′ ,y ′ ,k ,k,l,l le sous-lemme suivant. Sous-lemme 3.41 Soient d ∈ N, C1 , C2 ∈ R∗ + . Soit ǫ > 0. Il existe ∆0 ∈ N tel que pour pour ∆, ∆′ ∈ N vérifiant ∆ ≥ ∆0 , ∆′ − ∆ ≤ d et pour (nk ,k,l,l )k ,k,l,l∈{0,...,φ(∆)} des éléments de N ∩ [ ∆ , C2∆] et (n′ k ,k,l,l )k ,k,l,l∈{0,...,φ(∆′ )} C1 des éléments de N ∩ [ ∆′ , C2∆′ ], tels que C1 55 – (k , k , l, l) 7→ nk ,k,l,l soit croissant en k et l et décroissant en k et l . – (k , k , l, l) 7→ n′ k ,k,l,l soit croissant en k et l et décroissant en k et l . – nk ,k,l,l ≤ n′ k+2d,k−d,l+2d,l−d et n′ k ,k,l,l ≤ nk+2d,k−d,l+2d,l−d quand ces nombres ont un sens, alors la proportion de (k , k , l, l) dans {0, . . . , φ(∆)}4 tels que l’expression nk,k,l,l ait un sens et appartienne à [1 − ǫ, 1] est supérieure ou égale n′ k+2d,k−d,l+2d,l−d à 1 − ǫ. Démonstration. La démonstration est facile, et laissée au lecteur car ce sous-lemme ne servira pas dans la suite. (cid:3) Fin de la démonstration du lemme 3.39. Si A et B sont deux parties finies de X avec A ⊂ B et ♯A/♯B ≥ 1− ǫ on a kνA − νB k1 ≤ 2ǫ. Le lemme 3.39 en résulte facilement. (cid:3) x (y ) = d′ (x, y ) mais cela ne nous convient pas car Nous pourrions poser ρ♭ d′ (x, y ) fait intervenir une moyenne sur Bx,y ,k ,k,l,l , donc nécessite de connaître le cardinal de cet ensemble. Nous allons construire d♭ jouissant de propriétés analogues à celles de d′ mais telle que (d♭ − d)(x, y ) soit une moyenne (sur un ensemble d’indices ne dépendant que de d(x, y )) d’une fonction à valeurs dans [0, 7δ ], qui ne dépend (c’est là le point important) que de la connaissance d’un nombre de points borné par C (δ, K ) et des distances entre eux. Plus précisément ces points seront x, y , les z ∈ 3δ -géod(x, y ) vérifiant d(x, z) = r et les t ∈ 3δ -géod(x, y ) vérifiant d(y , t) = s, pour 3(3δ + 1) valeurs différentes de r et de s. Nous définirons alors ρ♭ x en posant ρ♭ x (y ) = d♭ (x, y ). Le lecteur peut donc oublier la construction de d′ qui précède, c’est-à-dire la proposition 3.38 et les lemmes 3.39 et 3.41 (en revanche le lemme 3.40 sera réutilisé). L’idée de la construction de d♭ est de remplacer la moyenne par µ(x, y ) qui sert à construire d′ dans la formule (24) de la preuve de la proposition 3.38 par une moyenne sur un “point virtuel” défini comme la donnée de ses “distances” à certains autres points. Ces distances seront à valeurs dans Z et de plus on quotientera par une action naturelle de Z consistant à a jouter un entier relatif à certaines distances et à le retrancher aux autres. En particulier les distances d’un point virtuel à x et y ne seront pas bien définies, mais leur somme le sera. Les ensembles de points virtuels auront les deux propriétés suivantes : être de cardinal ≤ C (δ, K ) et même temps varier très peu souvent lorsqu’on bouge un peu x ou y . Ces deux propriétés seraient contradictoires pour des points réels (dans la preuve du lemme 3.39 c’est la deuxième propriété qui est en défaut car le cardinal du support de µ(x, y ) tend vers l’infini quand d(x, y ) tend vers l’infini). La fonction d♭ : X × X → R+ sera définie dans la formule (27) ci-dessous. Jusqu’à (27) on suppose d(x, y ) ≥ 6δ . 56 ) − 3δ}, on note Y r x,y l’ensemble des Pour tout entier r ∈ {0, ..., E ( d(x,y) 2 points z ∈ 3δ -géod(x, y ) tels que d(x, z) ∈ {r, ..., r + 3δ}. Lemme 3.42 a) Etant donnés trois entiers r, r ′ , r ′′ ∈ {0, ..., E ( d(x,y) ) − 3δ} 2 x,y et z ′′ ∈ Y r ′′ x,y , géod(z , z ′′ ) rencontre Y r ′ avec r ≤ r ′ ≤ r ′′ , et z ∈ Y r x,y . b) Etant donnés deux entiers r, r ′ ∈ {0, ..., E ( d(x,y) ) − 3δ} avec r ≤ r ′ , et 2 z ′ ∈ Y r ′ x,y , géod(x, z ′ ) rencontre Y r x,y . Démonstration. Pour a), on doit montrer l’existence de z ′ ∈ géod(z , z ′′ ) appartenant à Y r ′ x,y . Si z ou z ′′ appartient à Y r ′ x,y on prend z ′ = z ou z ′ = z ′′ . Supposons donc d(x, z) < r ′ et d(x, z ′′ ) > r ′ + 3δ . Soit z ′ ∈ géod(z , z ′′ ) tel que d(x, z ′ ) = r ′ + E (3δ/2) (un tel z ′ existe car lorsque z ′ parcourt géod(z , z ′′ ) les valeurs prises par d(x, z ′ ) forment un intervalle de N). On a d(z , z ′ ) ≥ 3δ/2 et d(z ′ , z ′′ ) ≥ 3δ/2 donc le lemme 3.40 implique z ′ ∈ 3δ -géod(x, y ). Pour b) on doit montrer l’existence de z ∈ géod(x, z ′ ) appartenant à Y r x,y . Si d(x, z ′ ) ≤ r + 3δ on prend z = z ′ . Sinon soit z ∈ géod(x, z ′ ) tel que d(x, z) = r . Comme d(z , z ′ ) ≥ 3δ , le lemme 3.10 montre que z appartient à δ -géod(x, y ) et donc à Y r x,y . (cid:3) Ensuite pour r, s ∈ {0, ..., E ( d(x,y) ) − 3δ} on note Λy ,s x,r l’ensemble des fa- 2 milles d’entiers relatifs indexées par Y r x,y ∪Y s y ,x , que nous notons (cz )z∈Y r y ,x , x,y ∪Y s telles que x,y , ∀t ∈ Y s ∀z ∈ Y r y ,x , cz + ct ≥ d(z , t), cz − cz ′ ≤ d(z , z ′ ), ∀z , z ′ ∈ Y r x,y , ct − ct′ ≤ d(t, t′ ). ∀t, t′ ∈ Y s y ,x , En quelque sorte pour z ∈ Y r x,y ∪ Y s y ,x , cz est la “distance” d’un point virtuel à z (on écrit “distance” car elle appartient à Z mais pas nécessairement à N). Maintenant nous allons imposer une condition qui impliquera en particu- lier que le point virtuel “appartient” à 3δ -géod(x, y ). L’idée naïve serait de demander qu’il existe z ∈ Y r x,y et t ∈ Y s y ,x tels que cz + ct = d(z , t), c’est-à- dire que le point virtuel “appartient” à géod(z , t). Cependant cette condition nous empêcherait d’appliquer l’astuce des ensembles emboîtés. En effet on x,r → Λy ,s′ peut définir une application αs′←s x,r ′ , pour r, r ′ , s, s′ vérifiant r ′←r : Λy ,s 0 ≤ r ′ ≤ r ≤ E ( d(x, y ) 2 ) − 3δ et 0 ≤ s′ ≤ s ≤ E ( d(x, y ) 2 ) − 3δ, (25) 57 r ′←r (c) ∈ Λy ,s′ x,r on associe c′ = αs′←s de la façon suivante : à c ∈ Λy ,s x,r ′ tel que z ′ ∈ Y r ′ d(z ′ , z) + cz x,y , c′ pour z ′ = min z∈Y r x,y t′ ∈ Y s′ y ,x , c′ t′ = min t∈Y s y ,x En d’autres termes, pour définir αs′←s r ′←r on fait comme si la géodésique entre le point virtuel et tout point de Y r ′ x,y (resp. Y s′ x,y (resp. Y s y ,x ) rencontrait Y r y ,x), comme dans le dessin ci-dessous. d(t′ , t) + ct . et pour x Y r ′ x,y Y r x,y "point virtuel" • Y s y ,x y Y s′ y ,x Le a) du lemme 3.42 montre que ces applications se composent bien : si r, r ′ , r ′′ , s, s′ , s′′ vérifient d(x, y ) 2 d(x, y ) ) − 3δ et 0 ≤ s′′ ≤ s′ ≤ s ≤ E ( 0 ≤ r ′′ ≤ r ′ ≤ r ≤ E ( ) − 3δ, 2 r ′′←r ′ ◦ αs′←s r ′←r = αs′′←s on a αs′′←s′ r ′′←r . Si on imposait en plus la condition évo- quée ci-dessus, on ne posséderait plus de telles applications. On va imposer cette condition en d’autres entiers r ′ , s′ , en utilisant précisément l’applica- tion αs′←s r ′←r , et dans l’astuce des ensembles emboîtés, r et r ′ , respectivement s et s′ , varieront en sens inverse l’un de l’autre (c’est pourquoi on ne peut pas les réunir en une seule variable). Pour r, r ′ , s, s′ vérifiant (25) on définit r ′←r (c) ∈ Λy ,s′ donc Λy ,s′ ,s x,r dont l’image c′ = αs′←s x,r ′ ,r comme l’ensemble des c ∈ Λy ,s x,r ′ vérifie la condition introduite ci-dessus : il existe z ′ ∈ Y r ′ x,y et t′ ∈ Y s′ y ,x tels t′ = d(z ′ , t′ ). z ′ + c′ que c′ Lemme 3.43 Pour r, s ∈ {0, . . . , E ( d(x,y) ) − 3δ} on a des inclusions 2 x,0,r ⊂ . . . ⊂ Λy ,0,s Λy ,0,s x,r−1,r ⊂ Λy ,0,s x,r,r ∩ ∩ ∩ . . . . . . . . . . . . . . . ∩ ∩ ∩ . . . ⊂ . . . ⊂ Λy ,s−1,s Λy ,s−1,s x,r−1,r ⊂ Λy ,s−1,s x,0,r x,r,r ∩ ∩ ∩ . . . Λy ,s,s x,0,r ⊂ . . . ⊂ Λy ,s,s x,r−1,r ⊂ Λy ,s,s x,r,r dans Λy ,s x,r et toutes ces parties sont non vides. 58 Démonstration. Soit r ′ ∈ {0, . . . , r − 1}, s′ ∈ {0, . . . , s} et montrons l’in- x,r ′+1,r . Soit c ∈ Λy ,s′ ,s x,r ′ ,r ⊂ Λy ,s′ ,s clusion horizontale Λy ,s′ ,s x,r ′ ,r . On note c′ = αs′←s r ′←r (c) ∈ Λy ,s′ r ′+1←r (c) ∈ Λy ,s′ x,r ′ et c′′ = αs′←s x,r ′+1 , si bien que c′ = αs′←s′ r ′←r ′+1 (c′′ ). Par hypothèse il existe u′ ∈ Y r ′ x,y et v ′ ∈ Y s′ v′ = d(u′ , v ′ ). Par définition de u′ + c′ y ,x tels que c′ r ′←r ′+1 il existe u′′ ∈ Y r ′+1 l’application αs′←s′ u′′ + d(u′ , u′′). Les u′ = c′′ tel que c′ x,y inégalités u′′ + c′′ v′ = d(u′ , u′′ ) + c′′ u′′ + d(u′ , u′′ ) + c′ v′ = c′′ u′ + c′ d(u′ , v ′ ) = c′ v′ ≥ d(u′ , u′′) + d(u′′ , v ′ ) ≥ d(u′ , v ′ ) sont toutes des égalités, donc c′′ u′′ + c′′ v′ = d(u′′ , v ′ ) et c appartient bien à Λy ,s′ ,s x,r ′+1,r . Les inclusions verticales se démontrent de la même manière. Il reste à montrer que Λy ,0,s x,0,r est non vide. Soit t ∈ géod(x, y ) tel que d(x, t) ∈ x,r défini par cz = d(z , t) pour z ∈ Y r [r, d(x, y )−s] et soit c ∈ Λy ,s x,y ∪Y s y ,x . Alors c appartient à Λy ,0,s x,0,r car c′ = α0←s 0←r (c) vérifie c′ x = d(x, t) et c′ y = d(t, y ) (puisque y ,x) et donc c′ x + c′ géod(x, t) rencontre Y r x,y et géod(t, y ) rencontre Y s y = d(x, y ). Par conséquent Λy ,0,s x,0,r est non vide. (cid:3) Enfin étant donné r1 , r2 , r3 , s1 , s2 , s3 vérifiant 0 ≤ r1 ≤ r2 ≤ r3 ≤ E ( d(x, y ) 2 ) − 3δ et 0 ≤ s1 ≤ s2 ≤ s3 ≤ E ( d(x, y ) 2 ) − 3δ, (26) x,r1 ,r2 ,r3 comme l’image de Λy ,s2 ,s3 on définit Λy ,s1 ,s2 ,s3 x,r2 ,r3 dans Λy ,s1 x,r1 par l’application r1←r3 . Pour tout entier r ∈ {0, ..., E ( d(x,y) αs1←s3 )} − 3δ on a une application β y ,s x,r : 2 Λy ,s x,r → N définie par β y ,s d(x, z) + cz + ct + d(t, y ) x,r (c) = min x,y ,t∈Y s z∈Y r y ,x y ,x (cid:0)ct + d(t, y )(cid:1). x,y (cid:0)d(x, z) + cz (cid:1) + min = min t∈Y s z∈Y r y ,x (cid:0)ct + x,y (cid:0)d(x, z) + cz (cid:1), respectivement mint∈Y s On doit comprendre minz∈Y r d(t, y )(cid:1), comme la “distance” de x, respectivement y , au point virtuel. Si r, r ′ , s, s′ sont des entiers vérifiant (25), on a β y ,s′ x,r ′ ◦ αs′←s r ′←r = β y ,s x,r par le b) x,r se factorise par Λy ,s x,r , où Λy ,s du lemme 3.42. Cette application β y ,s x,r est le quotient de Λy ,s x,r par la relation d’équivalence suivante : deux éléments c et c′ sont équivalents s’il existe k ∈ Z avec c′ (z) = c(z) + k pour z ∈ x,y et c′ (t) = c(t) − k pour t ∈ Y s Y r y ,x . Toutes les constructions ci-dessus passent au quotient de cette façon et on note en particulier Λy ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 l’image de Λy ,s2 ,s3 x,r2 ,r3 dans Λy ,s1 x,r2 ,r3 dans Λy ,s1 x,r1 (qui est aussi l’image de Λy ,s2 ,s3 x,r1 ). On re- marque que si r, r ′ , s, s′ vérifient (25), Λy ,s′ ,s x,r ′ ,r est un ensemble fini, dont le 59 cardinal est borné par une constante de la forme C (δ, K ). Il en va donc de même pour Λy ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 , lorsque r1 , r2 , r3 , s1 , s2 , s3 vérifient (26). Grâce à , αs3←s3 αs1←s1 r3−1←r3 , αs3−1←s3 r1−1←r1 , αs1−1←s1 et au lemme 3.43 on possède, pour r3←r3 r1←r1 (r1 , r2 , r3 , s1 , s2 , s3) vérifiant (26), x,r1 ,r2 ,r3 → Λy ,s1 ,s2 ,s3 – une application surjective Λy ,s1 ,s2 ,s3 x,r1−1,r2 ,r3 si r1 ≥ 1, x,r1 ,r2 ,r3 → Λy ,s1−1,s2 ,s3 – une application surjective Λy ,s1 ,s2 ,s3 si s1 ≥ 1, x,r1 ,r2 ,r3 x,r1 ,r2 ,r3 → Λy ,s1 ,s2 ,s3 x,r1 ,r2+1,r3 et Λy ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 → Λy ,s1 ,s2 ,s3 – des applications injectives Λy ,s1 ,s2 ,s3 x,r1 ,r2 ,r3−1 si r2 < r3 , – des applications injectives Λy ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 → Λy ,s1 ,s2 ,s3−1 et Λy ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 → Λy ,s1 ,s2+1,s3 x,r1 ,r2 ,r3 x,r1 ,r2 ,r3 si s2 < s3 . De plus ces applications sont compatibles entre elles. On est maintenant en mesure de construire d♭ . On pose d♭(x, y ) = d(x, y ) si d(x, y ) < 6δ et si d(x, y ) ≥ 6δ on pose ∆x,y = E (d(x, y )/6) − δ et on définit 1 X d♭(x, y ) = (∆x,y + 1)6 r1 ,s1∈{0,...,∆x,y},r2 ,s2∈{∆x,y ,...,2∆x,y},r3 ,s3∈{2∆x,y ,3∆x,y } 1 x,r1 ,r2 ,r3 ) Xc∈ Λ β y ,s1 x,r1 (c). ♯( Λy ,s1 ,s2 ,s3 y ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 Pour u1 , u2 , u3 , v1 , v2 , v3 ∈ [0, 1[ et x, y ∈ X on pose d♭ v1 ,v2 ,v3 u1 ,u2 ,u3 (x, y ) = d(x, y ) si d(x, y ) < 6δ et si d(x, y ) ≥ 6δ on pose x,u1 ,u2 ,u3 = Λy ,E ((∆x,y+1)v1 ),∆x,y+E ((∆x,y+1)v2 ),2∆x,y+E ((∆x,y +1)v3 ) Ay ,v1 ,v2 ,v3 x,E ((∆x,y+1)u1 ),∆x,y+E ((∆x,y+1)u2 ),2∆x,y+E ((∆x,y+1)u3 ) , 1 et d♭ v1 ,v2 ,v3 β y ,E ((∆x,y+1)v1 ) x,u1 ,u2 ,u3 ) Xc∈A x,E ((∆x,y+1)u1 )(c) u1 ,u2 ,u3 (x, y ) = ♯(Ay ,v1 ,v2 ,v3 y ,v1 ,v2 ,v3 x,u1 ,u2 ,u3 de sorte que d’après la formule (27) on a toujours d♭(x, y ) = Zu1 ,u2 ,u3 ,v1 ,v2 ,v3∈[0,1[ Dans la formule précédente on intègre sur [0, 1[6 au lieu de [0, 1]6 car l’ex- pression pourrait ne pas avoir de sens pour u1 = 1 et u2 = 0 par exemple. d♭ v1 ,v2 ,v3 u1 ,u2 ,u3 (x, y )du1du2du3dv1dv2dv3 . (27) (28) Lemme 3.44 a) Pour r1 , s1 ∈ {0, ..., ∆x,y }, r2 , s2 ∈ {∆x,y , ..., 2∆x,y }, r3 , s3 ∈ {2∆x,y , ..., 3∆x,y }, c ∈ Λy ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 on a β y ,s1 x,r1 (c) ∈ [d(x, y ), d(x, y ) + 7δ ]. b) Pour u1 , u2 , u3 , v1 , v2 , v3 ∈ [0, 1[ et x, y ∈ X on a d(x, y ) ≤ d♭ v1 ,v2 ,v3 u1 ,u2 ,u3 (x, y ) ≤ d(x, y ) + 7δ. 60 Démonstration. On montre seulement a) car b) en résulte immédiatement. Il existe z ∈ Y r2 x,y et t ∈ Y s2 y ,x tels que β y ,s1 x,r1 (c) ≤ d(x, z) + d(z , t) + d(t, y ). On rappelle que z , t appartiennent à 3δ -géod(x, y ) et que d(x, z) ≤ d(x,y) et 2 d(t, y ) ≤ d(x,y) . Par (Hδ (z , x, t, y )) on a 2 d(z , t) + d(x, y ) ≤ max(d(x, z) + d(t, y ), d(x, t) + d(z , y )) + δ = d(x, t) + d(z , y ) + δ d’où d(x, z) + d(z , t) + d(t, y ) ≤ d(x, z) + d(x, t) + d(z , y ) + d(t, y ) + δ − d(x, y ) ≤ d(x, y ) + 7δ. (cid:3) Lemme 3.45 Pour tout ρ ∈ N, il existe C = C (δ, K, ρ) tel que pour x, x′ , y , y ′ ∈ X verifiant d(x, x′ ) ≤ ρ et d(y , y ′) ≤ ρ la mesure de l’ensemble des (u1 , u2 , u3 , v1 , v2 , v3) ∈ [0, 1[6 tels que d♭ v1 ,v2 ,v3 u1 ,u2 ,u3 (x, y ) − d♭ v1 ,v2 ,v3 u1 ,u2 ,u3 (x′ , y ) − d♭ v1 ,v2 ,v3 u1 ,u2 ,u3 (x, y ′) + d♭ v1 ,v2 ,v3 u1 ,u2 ,u3 (x′ , y ′) 6= 0 est ≤ C 1+d(x,y) . Démonstration. La démonstration consiste encore en une astuce d’ensembles emboîtés. Sous-lemme 3.46 Soient x, y , x, y ∈ X et d ≥ 2(d(x, x) + d(y , y) + 3δ ). a) Etant donnés r, r ′ , r ′′ ∈ {0, ..., E ( d(x,y) ) − 3δ} tels que r + d ≤ r ′ ≤ r ′′ − d, 2 x,y , géod(z , z ′′ ) rencontre Y r ′ x,y et z ′′ ∈ Y r ′′ z ∈ Y r x, y . b) Etant donnés r, r ′ , r ′′ ∈ {0, ..., E ( d(x,y) ) − 3δ} tels que r ≤ r ′ ≤ r ′′ − d, 2 x, y et z ′′ ∈ Y r ′′ x,y , géod(z , z ′′ ) rencontre Y r ′ z ∈ Y r x, y . c) Etant donnés r, r ′ , r ′′ ∈ {0, ..., E ( d(x,y) ) − 3δ} tels que r + d ≤ r ′ ≤ r ′′ , 2 x, y et z ′′ ∈ Y r ′′ x,y , géod(z , z ′′ ) rencontre Y r ′ z ∈ Y r x,y . d) Etant donnés r, r ′ ∈ {0, ..., E ( d(x,y) ) − 3δ} tels que r + d ≤ r ′ , z ′ ∈ Y r ′ x,y , 2 géod( x, z ′ ) rencontre Y r x, y . Démonstration. Pour a) on prend z ′ ∈ géod(z , z ′′ ) tel que d( x, z ′ ) = r ′ . Cela est possible car d( x, z) ≤ d(x, x) + r + 3δ ≤ r ′ − d 2 et d( x, z ′′ ) ≥ r ′′ − d(x, x) ≥ r ′ + d 2 . On a z , z ′′ ∈ (d − 3δ )-géod( x, y) par le lemme 3.2 et d(z , z ′ ) ≥ d 2 , d(z ′ , z ′′ ) ≥ d 2 ce qui permet d’appliquer le lemme 3.40 avec α = β = d − 3δ . D’où z ′ ∈ Y r ′ x, y . Pour b) on prend z ′ = z si d( x, z) ≥ r ′ donc on suppose d( x, z) < r ′ . On a d d d( x, z ′′ ) ≥ d(x, z ′′ ) − d(x, x) ≥ r ′′ − + 3δ ≥ r ′ + 2 2 + 3δ 61 donc il existe z ′ ∈ géod(z , z ′′ ) tel que d( x, z ′ ) = r ′ + 3δ . On a alors z ∈ 3δ -géod( x, y), z ′′ ∈ (d − 3δ )-géod( x, y), d(z , z ′ ) ≥ 3δ et d(z ′ , z ′′ ) ≥ d 2 , ce qui permet d’appliquer le lemme 3.40 avec α = 3δ et β = d − 3δ . D’où z ′ ∈ Y r ′ x, y . Pour c) on prend z ′ = z ′′ si d(x, z ′′ ) ≤ r ′ + 3δ donc on suppose d(x, z ′′ ) > r ′ + 3δ . On a . d d 2 ≤ r ′ − d(x, z) ≤ d(x, x) + d( x, z) ≤ r + 3δ + d(x, x) ≤ r + 2 Donc il existe z ′ ∈ géod(z , z ′′ ) tel que d(x, z ′ ) = r ′ . On a z ∈ (d−3δ )-géod(x, y ), z ′′ ∈ 3δ -géod(x, y ), d(z , z ′ ) ≥ d 2 , d(z ′ , z ′′ ) ≥ 3δ ce qui permet d’appliquer le lemme 3.40 avec α = d − 3δ et β = 3δ . D’où z ′ ∈ Y r ′ x,y . Pour d) on prend z ∈ géod( x, z ′ ) vérifiant d( x, z) = r . Cela est possible car d( x, z ′ ) ≥ d(x, z ′ ) − d(x, x) ≥ r ′ − d 2 ≥ r + d 2 . Comme z ′ ∈ (d − 3δ )-géod( x, y), z ∈ géod( x, z ′ ) et d(z , z ′ ) ≥ d 2 , le lemme 3.10 appliqué à ǫ = d − 3δ montre z ∈ 3δ -géod( x, y) d’où z ∈ Y r x, y . (cid:3) Suite de la démonstration du lemme 3.45. Pour x, y , x, y ∈ X et d ≥ 2(d(x, x) + d(y , y) + 3δ ) comme dans le sous-lemme 3.46, et pour r, s ∈ x,r → Λ y ,s−d {d, . . . , E ( d(x,y) ) − 3δ}, on a une application γ : Λy ,s x,r−d qui envoie c 2 sur c′ défini par cz + d(z ′ , z) pour z ′ ∈ Y r−d c′ z ′ = min x, y z∈Y r x,y ct + d(t′ , t) pour t′ ∈ Y s−d et c′ y , x . t′ = min t∈Y s y ,x On note γ sans indices pour alléger les formules car les indices sont déterminés par l’ensemble de départ et l’ensemble d’arrivée. De la même façon, pour x,r → Λy ,s−d r, s ∈ {d, . . . , E ( d( x, y) ) − 3δ}, on a une application γ : Λ y,s x,r−d . Ces 2 applications sont compatibles avec les applications αs′←s r ′←r (associées à (x, y ) et ( x, y)). Par exemple, lorsque les applications ont un sens, γ γ → Λy ,s−2d → Λ y ,s−d x,r−2d coïncide avec αs−2d←s – la composée Λy ,s r−2d←r (grâce à a) x,r x,r−d du sous-lemme 3.46), – les deux composées αs′−d←s−d αs′←s γ γ r ′←r→ Λy ,s′ → Λ y ,s′−d r ′−d←r−d→ Λ y ,s′−d → Λ y ,s−d Λy ,s et Λy ,s x,r ′−d x,r ′ x,r ′−d x,r x,r x,r−d sont égales car, grâce à b) et c) du sous-lemme 3.46 l’image c′ ∈ Λ y ,s′−d x,r ′−d de c ∈ Λy ,s x,r par chacune de ces deux composées est donnée par cz + d(z ′ , z) pour z ′ ∈ Y r ′−d x, y ct + d(t′ , t) pour t′ ∈ Y s′−d et c′ y , x . t′ = min t∈Y s y ,x c′ z ′ = min z∈Y r x,y 62 Plus généralement la compatibilité signifie que deux composées construites à partir des applications α et γ sont égales lorsqu’elles possèdent le même ensemble de départ et le même ensemble d’arrivée. En effet on démontre que si l’ensemble de départ est Λ y ,s x,r et l’ensemble d’arrivée est Λ y ,s x,r (avec ( x, y) et ( x, y) égaux à (x, y ) ou ( x, y)), les deux composées sont égales à l’application x,r associe c ∈ Λ y ,s qui à c ∈ Λ y ,s x,r défini par les formules c z + d( z , z) pour z ∈ Y r c z = min x, y z∈Y r x, y ct + d(t, t) pour t ∈ Y s et ct = min y , x . t∈Y s y, x Sous-lemme 3.47 Pour r, r ′ , s, s′ comme dans (25) vérifiant r ′ + 2d ≤ r et γ x,r ′ ,r dans Λ y ,s′+d,s−d x,r−d envoie Λy ,s′ ,s → Λ y ,s−d s′ + 2d ≤ s, l’application Λy ,s x,r ′+d,r−d . x,r x,r → Λy ,s′ Démonstration. D’après ce qui précède, αs′←s r ′←r : Λy ,s x,r ′ est la compo- sée αs′ +d←s−d γ γ → Λy ,s′ r ′+d←r−d→ Λ y ,s′+d → Λ y ,s−d Λy ,s x,r ′ . x,r ′+d x,r x,r−d x,r ′+d d’image c′ = γ (c) ∈ Λy ,s′ De plus pour c ∈ Λ y ,s′+d x,r ′ , s’il existe z ∈ Y r ′ x,y et t = d(z , t), alors il existe z ∈ Y r ′+d et t ∈ Y s′+d t ∈ Y s′ z + c′ y ,x tels que c′ tels que x, y y , x c z + ct = d( z , t) par le même argument que dans la preuve du lemme 3.43. (cid:3) On déduit du sous-lemme 3.47 que pour (r1 , r2 , r3 , s1 , s2 , s3) vérifiant (26) et (r1 , r2 , r3 , s1 , s2 , s3) vérifiant d( x, y) d( x, y) 0 ≤ r1 ≤ r2 ≤ r3 ≤ E ( ) − 3δ 0 ≤ s1 ≤ s2 ≤ s3 ≤ E ( ) − 3δ et 2 2 – si r1 ≤ r1 − d, s1 ≤ s1 − d, r2 ≥ r2 + d, s2 ≥ s2 + d, r3 ≤ r3 − d, γ → Λ y ,s1 ,s2 ,s3 s3 ≤ s3 − d, on a une application Λy ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 , et en notant x,r1 ,r2 ,r3 c = γ (c) ∈ Λ y ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 l’image de c ∈ Λy ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 on a, grâce au d) du sous-lemme 3.46, β y ,s1 x,y (cid:0)d( x, z) + cz (cid:1) + min y ,x (cid:0)ct + d(t, y)(cid:1), x,r1 (c) = min r1 s1 z∈Y t∈Y – si r1 ≤ r1 − d, s1 ≤ s1 − d, r2 ≥ r2 + d, s2 ≥ s2 + d, r3 ≤ r3 − d, γ s3 ≤ s3 − d, on a une application Λ y ,s1 ,s2 ,s3 → Λy ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 verifiant une x,r1 ,r2 ,r3 propriété semblable à (29). On rappelle que ∆x,y = E (d(x, y )/6) − δ , ∆ x, y = E (d( x, y)/6) − δ , et que pour (u1 , u2 , u3 , v1 , v2 , v3 ) ∈ [0, 1[6 , x,u1 ,u2 ,u3 = Λy ,E ((∆x,y+1)v1 ),∆x,y+E ((∆x,y+1)v2 ),2∆x,y+E ((∆x,y +1)v3 ) Ay ,v1 ,v2 ,v3 x,E ((∆x,y+1)u1 ),∆x,y+E ((∆x,y+1)u2 ),2∆x,y+E ((∆x,y+1)u3 ) , x,u1 ,u2 ,u3 = Λ y ,E ((∆ x, y +1)v1 ),∆ x, y+E ((∆ x, y+1)v2 ),2∆ x, y+E ((∆ x, y +1)v3 ) A y ,v1 ,v2 ,v3 x,E ((∆ x, y+1)u1 ),∆ x, y+E ((∆ x, y+1)u2 ),2∆ x, y +E ((∆ x, y+1)u3 ) . (31) (29) (30) 63 Soit η ∈]0, 1 2 [ tel que d(x, y ) ≥ 9d η +6δ . On a alors pour tout (u1 , u2 , u3 , v1 , v2 , v3 ) ∈ [η , 1 − η [6 , les six inégalités E ((∆ x, y + 1)(u1 − η )) ≤ E ((∆x,y + 1)u1) − d, ∆ x, y + E ((∆ x, y + 1)(u2 + η )) ≥ ∆x,y + E ((∆x,y + 1)u2) + d, ..., 2∆ x, y + E ((∆ x, y + 1)(v3 − η )) ≤ 2∆x,y + E ((∆x,y + 1)v3) − d. Ces inégalités ont lieu car ∆x,y − ∆ x, y ≤ d(x, y ) − d( x, y) d(x, x) + d(y , y) + 1 ≤ + 1 ≤ 6 6 d(x, y ) − 6δ 3d et (∆x,y + 1)η ≥ 2 ≥ d + 3∆x,y − ∆ x, y . η ≥ 6 On montre aussi les 6 inégalités analogues obtenues en permutant les rôles de x, y et x, y . On possède donc, pour (u1 , u2 , u3 , v1 , v2 , v3) ∈ [η , 1 − η [6 , des applications d 6 Ay ,v1 ,v2 ,v3 x,u1 ,u2 ,u3 γ γ → A y ,v1−η,v2+η,v3−η → Ay ,v1−η,v2+η,v3−η x,u1−η,u2+η,u3−η et A y ,v1 ,v2 ,v3 x,u1−η,u2+η,u3−η . x,u1 ,u2 ,u3 Les phrases précédentes, le a) du lemme 3.42 et les parties a), b), c) du sous-lemme 3.46 garantissent que ces applications vérifient les conditions de compatibilité supposées dans le sous-lemme suivant (en prenant Av1 ,v2 ,v3 u1 ,u2 ,u3 = x,u1 ,u2 ,u3 et Av1 ,v2 ,v3 u1 ,u2 ,u3 = A y ,v1 ,v2 ,v3 Ay ,v1 ,v2 ,v3 x,u1 ,u2 ,u3 ). Sous-lemme 3.48 Soit C ∈ N∗ . Alors pour tout η ∈]0, 1 4 [ et pour toutes famil les d’ensembles finis non vides, de cardinaux inférieurs ou égaux à C , u1 ,u2 ,u3 )(u1 ,u2 ,u3 ,v1 ,v2 ,v3 )∈[0,1[6 et ( Av1 ,v2 ,v3 (Av1 ,v2 ,v3 u1 ,u2 ,u3 )(u1 ,u2 ,u3 ,v1 ,v2 ,v3 )∈[0,1[6 munis u1 ,u2 ,u3 → Av′ u1 ,u2 ,u3 → Av′ 1 ,v2 ,v3 1 ,v2 ,v3 et Av1 ,v2 ,v3 – d’applications surjectives Av1 ,v2 ,v3 u′ u′ 1 ,u2 ,u3 1 ,u2 ,u3 1 ≤ u1 et v ′ pour u′ 1 ≤ v1 , u1 ,u2 ,u3 → Av1 ,v′ 2 ,v′ u1 ,u2 ,u3 → Av1 ,v′ 2 ,v′ et Av1 ,v2 ,v3 – d’applications injectives Av1 ,v2 ,v3 3 3 2 ,u′ u1 ,u′ 2 ,u′ u1 ,u′ 3 3 pour u′ 2 ≥ u2 , v ′ 2 ≥ v2 , u′ 3 ≤ u3 et v ′ 3 ≤ v3 , – et pour (u1 , u2 , u3 , v1 , v2 , v3) ∈ [η , 1 − η [6 , d’applications u1 ,u2 ,u3 → Av1−η,v2+η,v3 −η Av1 ,v2 ,v3 u1 ,u2 ,u3 → Av1−η,v2 +η,v3−η Av1 ,v2 ,v3 u1−η,u2+η,u3−η et u1−η,u2+η,u3−η tel les que toutes ces applications soient compatibles entre el les (c’est-à-dire que deux composées d’applications comme ci-dessus sont égales lorsqu’el les ont le même ensemble de départ et le même ensemble d’arrivée), alors la me- sure de l’ensemble des (u1 , u2 , u3 , v1 , v2 , v3 ) ∈ [η , 1 − η [6 tels que l’application u1 ,u2 ,u3 → Av1−η,v2+η,v3−η Av1 ,v2 ,v3 u1−η,u2+η,u3−η soit bijective est ≥ 1 − 50C η . 64 Démonstration. Pour (u1 , u2 , u3 , v1 , v2 , v3 ) ∈ [2η , 1 − 2η [6 on considère les applications Av1+η,v2−η,v3+η u1+η,u2−η,u3+η g f h→ Av1−2η,v2 +2η,v3−2η → Av1−η,v2 +η,v3−η → Av1 ,v2 ,v3 u1−2η,u2+2η,u3−2η . u1−η,u2+η,u3−η u1 ,u2 ,u3 Si g ◦ f et h ◦ g sont bijectives, g est bijective. Or g ◦ f est égale à la composée Av1 +η,v2−η,v3+η u1+η,u2−η,u3+η → Av1+η,v2 −η,v3+η u1−η,u2−η,u3+η → Av1+η,v2−η,v3 +η u1−η,u2+η,u3+η → Av1+η,v2−η,v3 +η u1−η,u2+η,u3−η → Av1−η,v2−η,v3 +η u1−η,u2+η,u3−η → Av1−η,v2+η,v3+η u1−η,u2+η,u3−η → Av1−η,v2+η,v3−η (32) u1−η,u2+η,u3−η et h ◦ g est égale à la composée u1 ,u2 ,u3 → Av1 ,v2 ,v3 u1−2η,u2 ,u3 → Av1 ,v2 ,v3 u1−2η,u2+2η,u3 → Av1 ,v2 ,v3 Av1 ,v2 ,v3 u1−2η,u2+2η,u3−2η → Av1−2η,v2 ,v3 u1−2η,u2+2η,u3−2η → Av1−2η,v2+2η,v3 u1−2η,u2+2η,u3−2η → Av1 −2η,v2+2η,v3 −2η u1−2η,u2+2η,u3−2η . Pour u2 , u3 , v1 , v2 , v3 ∈ [2η , 1 − 2η [, la mesure de l’ensemble des u1 ∈ [2η , 1 − 2η [ tels que (33) u1+η,u2−η,u3+η → Av1+η,v2−η,v3+η Av1+η,v2−η,v3+η u1−η,u2−η,u3+η (qui est la première application dans (32)) ne soit pas une bijection est infé- rieure ou égale à 4η (C − 1). En effet l’application (34) est surjective donc est une bijection en cas d’égalité des cardinaux des deux parties et d’autre part l’application de [0, 1[ dans N qui à u associe le cardinal de Av1+η,v2−η,v3+η u,u2−η,u3+η est décroissante et prend ses valeurs dans {1, ..., C }. On a 11 autres énoncés correspondant aux autres applications de (32) et (33). Donc la mesure de l’ensemble des (34) (u1 , u2 , u3 , v1 , v2 , v3 ) ∈ [2η , 1 − 2η [6 tels que g soit une bijection est supérieure ou égale à (1 − 4η )6 − 48η (C − 1)(1 − 4η )5 ≥ 1 − 24η − 48η (C − 1) ≥ 1 − 48C η ≥ 1 − 50C η et le sous-lemme 3.48 est démontré. (cid:3) Fin de la démonstration du lemme 3.45. Soit C = C (δ, K ) tel que pour x, y ∈ X et r1 , r2 , r3 , s1 , s2 , s3 vérifiant (26) on ait ♯( Λy ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 ) ≤ C . Soit ρ ∈ N. On pose d = 4ρ + 6δ . Soient x, y , x′ , y ′ ∈ X vérifiant d(x, x′ ) ≤ ρ et d(y , y ′) ≤ ρ. Il suffit de montrer le lemme 3.45 en supposant d(x, y ) > 36d+6δ . On choisit alors η ∈]0, 1 4 [ vérifiant d(x, y ) = 9d 9d η + 6δ , c’est-à-dire η = d(x,y)−6δ . On applique le sous-lemme 3.48 avec ( x, y) égal à (x, y ), (x′ , y ), (x, y ′ ) ou u1 ,u2 ,u3 ) et ( Av1 ,v2 ,v3 (x′ , y ′) et avec les familles (Av1 ,v2 ,v3 u1 ,u2 ,u3 ) égales à (Ay ,v1 ,v2 ,v3 x,u1 ,u2 ,u3 ) et (A y ,v1 ,v2 ,v3 x,u1 ,u2 ,u3 ). 65 Il existe donc une partie J ⊂ [η , 1 − η [6 de mesure ≥ 1 − 200C η telle que pour (u1 , u2 , u3 , v1 , v2 , v3 ) ∈ J les applications x,u1 ,u2 ,u3 → Ay ,v1−η,v2+η,v3−η x,u1 ,u2 ,u3 → Ay ,v1−η,v2 +η,v3−η Ay ,v1 ,v2 ,v3 x,u1−η,u2+η,u3−η , Ay ,v1 ,v2 ,v3 x′ ,u1−η,u2+η,u3−η , x,u1 ,u2 ,u3 → Ay ′ ,v1−η,v2+η,v3 −η x,u1 ,u2 ,u3 → Ay ′ ,v1−η,v2+η,v3−η x,u1−η,u2+η,u3−η et Ay ,v1 ,v2 ,v3 Ay ,v1 ,v2 ,v3 x′ ,u1−η,u2+η,u3−η soient bijectives. Soient (u1 , u2 , u3 , v1 , v2 , v3) ∈ J , c ∈ Ay ,v1 ,v2 ,v3 x,u1 ,u2 ,u3 et cx,y ∈ Ay ,v1−η,v2+η,v3−η x,u1−η,u2+η,u3−η , cx′ ,y ∈ Ay ,v1−η,v2 +η,v3−η x′ ,u1−η,u2+η,u3−η , cx,y ′ ∈ Ay ′ ,v1−η,v2+η,v3−η x,u1−η,u2+η,u3−η et cx′ ,y ′ ∈ Ay ′ ,v1−η,v2+η,v3−η x′ ,u1−η,u2+η,u3−η les images de c par les quatre bijections ci-dessus. Alors x′ ,u1−η (cx′ ,y ) − β y ′ ,v1−η x,u1−η (cx,y ′ ) + β y ′ ,v1−η β y ,v1−η x,u1−η (cx,y ) − β y ,v1−η x′ ,u1−η (cx′ ,y ′ ) = 0 car par (29) on a β y ,v1−η x,u1−η (c x, y ) = (cid:0)d( x, z) + cz (cid:1) + (cid:0)ct + d(t, y)(cid:1) min min E ((∆x,y+1)u1 ) E ((∆x,y +1)v1 ) z∈Y t∈Y x,y y ,x pour ( x, y) égal à (x, y ), (x′ , y ), (x, y ′) ou (x′ , y ′). Pour tout (u1 , u2 , u3 , v1 , v2 , v3) ∈ J on a donc d♭ v1−η,v2+η,v3−η u1−η,u2+η,u3−η (x, y ) − d♭ v1−η,v2+η,v3−η u1−η,u2+η,u3−η (x′ , y ) −d♭ v1−η,v2+η,v3−η u1−η,u2+η,u3−η (x, y ′ ) + d♭ v1−η,v2+η,v3 −η u1−η,u2+η,u3−η (x′ , y ′) = 0. Comme la mesure de J est supérieure ou égale à 1800C d 1800C (4ρ + 6δ ) 1 − 200C η = 1 − = 1 − d(x, y ) − 6δ d(x, y ) − 6δ cela termine la démonstration du lemme 3.45. (cid:3) Proposition 3.49 Pour tout x, y ∈ X on a d(x, y ) ≤ d♭(x, y ) ≤ d(x, y ) + 7δ et d♭ vérifie la condition (23). , Il résulte de la première assertion que l’on pourrait facilement remplacer d♭ par une vraie distance (comme on avait obtenu d′′ à partir de d′ dans la preuve de la proposition 3.38), mais cela n’est pas nécessaire pour la suite. Démonstration. La première assertion résulte du lemme 3.44. La seconde assertion résulte des lemmes 3.44 et 3.45. (cid:3) On définit alors, pour x ∈ X , la fonction ρ♭ x : X → R+ par x (a) = d♭(x, a). ρ♭ On l’étend ensuite en une fonction ρ♭ x : ∆ → R+ par la formule x (S ) = Pa∈S ρ♭ (a) si S est non vide et ρ♭ ρ♭ x (∅) = 0. ♯S 66 4 Construction des normes Soit s ∈]0, 1]. Nous allons donner la formule pour la norme pré-hilbertienne k.kHx,s (∆p ) sur C(∆p ) qui servira pour la partie difficile de l’homotopie de 1 à γ . Nous montrerons d’abord que cette norme est bien définie. Nous montre- rons ensuite la continuité des opérateurs, puis les propriétés d’équivariance de cette norme, et enfin l’équivariance à compacts près des opérateurs. 4.1 Formule pour la norme On rappelle que la constante F définie dans (22) est ma jorée par une constante de la forme C (δ, K, N ). On fixe un entier P ∈ N∗ tel que (HP ) : P soit divisible par 3 et assez grand en fonction de δ, K, N , Q. Dans la suite nous utiliserons un nombre fini de fois l’inégalité P ≥ C avec C de la forme C (δ, K, N , Q). Soient p ∈ {1, ..., pmax} et k , m, l0 , ..., lm ∈ N. Définition 4.1 On note Y p,k ,m,(l0 ,...,lm) l’ensemble des (p + m + 1 + Pm i=0 li )- x uplets (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } ) {z ∈ F -géod( x, a), d(z , a) ≥ {y ∈ 4δ -géod( x, a), d(y , a) ∈]N − 2δ, QN ]} Q F }, tels que – i) a1 , . . . , ap ∈ X sont deux à deux distincts, S0 = {a1 , . . . , ap}, S0 appartient à ∆p , Si ∈ ∆ \ {∅} pour i ∈ {1, . . . , m}, et pour tout i ∈ {0, . . . , m − 1}, on a Si+1 ⊂ Si ∪ [x∈B (x,k),a∈Si ∪ [x∈B (x,k),a∈Si – ii) pour tout i ∈ {1, . . . , m}, d(x, Si ) > k + P , – iii) pour i ∈ {0, . . . , m − 1} et j ∈ {1, . . . , li}, Y j i est une partie non vide de X de diamètre inférieur ou égal à P et i ⊂ [y∈Si ,z∈Si+1 Y j – iv) pour tout j ∈ {1, ..., lm}, Y j m est une partie non vide de X de dia- mètre inférieur ou égal à P et [y∈Sm , x∈B (x,k) 2P -géod( x, y ) et d(x, Y j Y j m) ≥ k + 3P . m ⊂ P -géod(y , z). 67 Remarque. La condition iv) implique que lm = 0 si d(x, Sm) ≤ k (ce qui ne peut se produire que si m = 0 à cause de ii)). En effet pour y ∈ B (x, k + N ) et x ∈ B (x, k), on a 2P -géod( x, y ) ⊂ B (x, k + N + 2P + δ ) par le lemme 3.31 et on suppose P > N + δ (ce qui est permis par (HP )). Dans la définition précédente, la seule raison pour laquelle on veut connaître (a1 , . . . , ap) en plus de S0 est que cela détermine ea1 ∧ · · ·∧eap alors que S0 per- met seulement de le connaître au signe près (le choix de eS0 = ±ea1 ∧ · · · ∧ eap , qui avait pour but de simplifier certaines formules, ne doit pas être utilisé ici, bien sûr). En contrepartie, chaque partie S0 est comptée p! fois, mais ce n’est pas grave car p est borné par pmax , qui est de la forme C (δ, K, N ). Dans la définition précédente la donnée de S0 est redondante mais nous préférons la garder car elle fournit une notation commode pour {a1 , . . . , ap}. On fixe un entier M ∈ N∗ tel que (HM ) : M soit pair et soit assez grand en fonction de δ, K, N , Q, P . Dans la suite nous utiliserons un nombre fini de fois l’inégalité M ≥ C avec C de la forme C (δ, K, N , Q, P ). On introduit maintenant une partition de Y p,k ,m,(l0 ,...,lm) pour la relation x d’équivalence suivante : (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } ) et (a1 , . . . , ap , S0 , ..., Sm , ( Y j i )i∈{0,...,m},j∈{1,...,li } ) sont en relation s’il existe une isométrie de [i∈{0,...,m} [ i∈{0,...,m},j∈{1,...,li } B (Si , M ) ∪ B (Y j i , M ) ∪ B (x, k + 2M ) vers B ( Si , M ) ∪ B ( Y j i , M ) ∪ B (x, k + 2M ) [i∈{0,...,m} [ i∈{0,...,m},j∈{1,...,li } qui envoie ai sur ai pour i ∈ {1, . . . , p}, Si sur Si pour i ∈ {0, . . . , m}, Y j i sur Y j i pour i ∈ {0, . . . , m}, j ∈ {1, . . . , li} et est l’identité sur B (x, k + 2M ). On rappelle que la notation B (A, r) pour A une partie de X a été introduite dans (17). p,k ,m,(l0 ,...,lm) le quotient de Y p,k ,m,(l0 ,...,lm) On note Y x x d’équivalence, et π p,k ,m,(l0 ,...,lm) l’application quotient. x p,k ,m,(l0 ,...,lm) Notations. Pour Z ∈ Y on note r0(Z ), . . . , rm(Z ), s0(Z ), . . . , sm(Z ) x les entiers tels que pour cette relation 68 – ri (Z ) = d(x, Si ) pour i ∈ {0, . . . , m}, – si (Z ) = d(Si , Si+1) + 2M pour i ∈ {0, . . . , m − 1}, – sm (Z ) = d(x, Sm) − k i )i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) pour tout (a1 , . . . , ap , S0 , ..., Sm , (Y j x On fixe B ∈ R∗ + et α ∈]0, 1[ tels que (HB ) : B soit assez grand en fonction de δ, K, N , Q, P , M , s et )−1(Z ). (Hα) : α soit assez petit en fonction de δ, K, N , Q, P , M , s, B . p,k ,m,(l0 ,...,lm) x Pour Z ∈ Y ξZ (f ) = on note ξZ la forme linéaire sur C(∆p ) définie par X (a1 ,...,ap ,S0 ,...,Sm ,(Y j i )i∈{0,...,m},j∈{1,...,li } )∈(π f (a1 , ..., ap ). p,k,m,(l0 ,...,lm ) x )−1 (Z ) (35) Dans cette formule, comme dans la suite, pour f ∈ C(∆p ) on note f (a1 , ..., ap) le coefficient de ea1 ∧...∧eap lorsque qu’on écrit f dans la base (±eS )S∈∆p (dont les vecteurs sont définis au signe près). On écrira aussi f (S ) = ±f (a1 , ..., ap) si S = {a1 , ..., ap}. On munit alors C(∆p ) de la norme pré-hilbertienne, définie par la formule suivante : . (36) B−(m+Pm i=0 li ) XZ ∈Y Hx,s (∆p ) = Xk ,m,l0 ,...,lm∈N kf k2 p,k,m,(l0 ,...,lm ) x m si (Z )−li (cid:17)♯(cid:0)(π p,k ,m,(l0 ,...,lm) e2s(r0 (Z )−k)(cid:16) )−1(Z )(cid:1)−α (cid:12)(cid:12)ξZ (f )(cid:12)(cid:12) 2 Yi=0 x Remarque. On a vu que lm = 0 lorsque sm(Z ) ≤ 0 (ce qui ne peut se produire que si m = 0) et dans ce cas on convient que sm (Z )lm = 1. En revanche pour i ∈ {0, . . . , m − 1}, on a toujours si (Z ) ≥ 1. Remarque. On verra dans la démonstration des propositions 4.21 et 4.30 Hx,s (∆p ) ou de kJx (f )k2 que l’on ma jore la partie de k∂ (f )k2 Hx,s (∆p ) correspon- dant à une valeur donnée de m par la partie de kf k2 Hx,s (∆p ) correspondant à m + 1 (en fait c’est un peu plus compliqué mais on renvoie aux propo- sitions 4.21 et 4.30 pour les détails). D’autre part la partie de kf k2 Hx,s (∆p ) correspondant à m = 0 et l0 = 0 est essentiellement égale au carré de la norme introduite dans le premier paragraphe de [Laf08] pour montrer que les groupes hyperboliques n’ont pas la propriété (T) renforcée. Enfin les Y j i as- surent la connaissance des points intermédiaires sur lesquels on moyenne dans 69 la construction de hx , ux , ρ′ x (le but de ces moyennes est d’assurer l’équiva- riance à compacts près des opérateurs par l’astuce des ensembles emboîtés, comme dans [KS03]). On ne peut pas demander que les parties Y j i soient des singletons car ces parties seront construites de façon naturelle dans les preuves et la nécessité de choisir un point dans chacune empêcherait les preuves de fonctionner. Remarque. Le facteur e2s(r0 (Z )−k) est utile pour la continuité de Jx , comme on le verra dans la proposition 4.30. Le facteur Qm i=0 si (Z )−li est motivé par le fait qu’étant donnés a1 , . . . , ap , S0 , . . . , Sm le nombre de possibilités pour (Y j i )i∈{0,...,m},j∈{1,...,li } vérifiant (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x est borné par CPm i=0 li Qm i=0 si (Z )li , avec C = C (δ, K, N , Q, P ), comme on le verra dans le lemme 4.15. La constante 2M qui apparaît dans la définition de si (Z ) pour i ∈ {0, ..., m − 1} trouvera son utilité dans la démonstration du lemme 4.66. Le facteur ♯((π p,k ,m,(l0 ,...,lm) )−1 (Z ))−α servira pour montrer que x les opérateurs ∂ et Jx conjugués par eτ θ♭ x sont équivariants à compact près. Enfin le facteur B−(m+Pm i=0 li ) (avec B assez grand) est nécessaire pour que la somme converge, comme le montre la preuve de la proposition 4.2. 4.2 Enoncé des résultats La proposition suivante montre que la norme pré-hilbertienne définie dans le sous-paragraphe précédent a bien un sens. Proposition 4.2 Pour tout p ∈ {1, . . . , pmax} et pour tout f ∈ C(∆p ) , kf kHx,s (∆p ) est fini. Cette proposition sera démontrée dans le sous-paragraphe 4.3 et la dé- monstration utilisera l’hypothèse (qui est une partie de (HB )) selon laquelle B est assez grand en fonction de δ, K, N , Q, P , M (sans cette hypothèse la proposition serait fausse). On note Hx,s(∆p ) le complété de C(∆p ) pour la norme pré-hilbertienne k.kHx,s (∆p ) . On note Hx,s = Lpmax p=1 Hx,s(∆p ). La proposition suivante sera démontrée dans le sous-paragraphe 4.7. Proposition 4.3 L’action de G sur Lpmax p=1 C(∆p ) s’étend en une action conti- nue de G sur Hx,s et il existe une constante C tel le que pour tout g ∈ G on ait kπ(g )kL(Hx,s ) ≤ e2sℓ(g)+C . 70 La démonstration de cette proposition utilisera entièrement l’hypothèse (HB ) (y compris la condition que B est assez grand en fonction de s). La proposition suivante est le résultat principal de ce paragraphe. x : C(∆p ) → C(∆p ) par θ♭ Pour p ∈ {1, . . . , pmax}, on définit θ♭ x (eS ) = x (S )eS pour tout S ∈ ∆p . Pour tout τ ∈ R on note eτ θ♭ x : C(∆p ) → C(∆p ) ρ♭ x (eS ) = eτ ρ♭ l’opérateur défini par eτ θ♭ x (S )eS . Proposition 4.4 Pour tout T ∈ R+ , l’opérateur x (∂ + Jx∂Jx )e−τ θ♭ (eτ θ♭ x )τ ∈[0,T ] s’étend en un opérateur continu sur le C[0, T ]-module hilbertien Hx,s [0, T ], et x (∂ + Jx∂Jx )e−τ θ♭ (Hx,s [0, T ], (eτ θ♭ x )τ ∈[0,T ]) appartient à KKG,2sℓ+C (C, C[0, T ]). En particulier (Hx,s , ∂ + Jx∂Jx ) est un élément de KKG,2sℓ+C (C, C). Proposition 4.5 Cet élément est égal à l’image de 1 ∈ KKG (C, C). Démonstration. La preuve est analogue à celle de la proposition 1.4.2 de [Laf02]. On reprend les arguments car on travaille ici dans un cadre hil- bertien plutôt que banachique. Soit Hx,s = C ⊕ Hx,s (avec une graduation inversée pour Hx,s ). On veut montrer que ( Hx,s , ∂ + Jx∂Jx ) est nul dans KKG,2sℓ+C (C, C). Il résulte de l’égalité ∂Jx + Jx∂ = Id que ∂Jx et Jx∂ com- mutent et Jx∂∂Jx = 0 car ∂ 2 = 0. Donc ∂Jx et Jx∂ sont deux pro jecteurs qui commutent, de somme Id, de produit nul, pairs. On pose T = Jx∂Jx . Alors T 2 = 0. On a ∂T = ∂Jx et T ∂ = Jx∂ . D’autre part, ∂ et ∂T ont même image, de même que T et T ∂ , d’où l’identification Hx,s = ∂ ( Hx,s ) ⊕ T ( Hx,s) en tant qu’espaces de Hilbert à équivalence des normes près (car les deux idempotents ∂Jx et Jx∂ ne sont pas nécessairement auto-adjoints). Dans cette décompo- sition, l’action de g ∈ G s’écrit sous la forme (cid:18)c1,1 (g ) c1,2(g ) c2,2(g )(cid:19). Considérons 0 alors le C[0, 1]-module hilbertien Z/2Z-gradué Hx,s [0, 1], où l’action de g ∈ G c2,2(g ) (cid:19) (cid:17)t∈[0,1] est donnée par (cid:16) (cid:18)c1,1(g ) tc1,2 (g ) . Alors ( Hx,s [0, 1], ∂ + T ) fournit 0 une homotopie entre ( Hx,s , ∂ + T ), en t = 1 et un élément dégénéré, en t = 0. En effet, quand l’action de G est diagonale, les opérateurs ∂ et T commutent exactement à cette action. (cid:3) Donc (Hx,s [0, T ], (eτ θ♭ x (∂ +Jx∂Jx )e−τ θ♭ x )τ ∈[0,T ]) réalise une homotopie entre 1 et (Hx,s , eT θ♭ x (∂ + Jx∂Jx )e−T θ♭ x ) et montre donc l’égalité entre ces deux 71 éléments dans KKG,2sℓ+C (C, C). On fixera T assez grand et cela constituera la partie difficile de l’homotopie de 1 à γ . La partie facile (qui fera l’ob jet du paragraphe 5), sera une homotopie entre (Hx,s , eT θ♭ x (∂ + Jx∂Jx )e−T θ♭ x ) et γ (qui montrera donc l’égalité entre ces deux éléments dans KKG,2sℓ+C (C, C)). 4.3 Premières propriétés de la norme Le but de ce sous-paragraphe est de montrer la proposition 4.2. Les lemmes 4.6, 4.7, 4.8, 4.9, 4.10, 4.11, 4.12 et 4.15 sont des préliminaires à la preuve de la proposition 4.2. Nous commençons par rappeler le lemme d’approximation par les arbres. Lemme 4.6 Soit (Y , dY ) un espace métrique fini et δ -hyperbolique, et w ∈ Y un point base. Soit l ∈ N tel que ♯Y ≤ 2l + 2. Alors il existe un arbre métrique fini (T , dT ) et une application Ψ : Y → T tel le que – pour y ∈ Y , dT (Ψw , Ψy ) = dY (w , y ), – pour y , z ∈ Y , dY (y , z) − lδ ≤ dT (Ψy , Ψz) ≤ dY (y , z). Démonstration. C’est exactement le (i) du théorème 12 du chapitre 2 de [GdlH90] car un espace métrique est δ -hyperbolique au sens de la dé- finition 0.1 si et seulement s’il est δ 2 -hyperbolique au sens de la définition 3 (reformulée dans 4) du chapitre 2 de [GdlH90]. (cid:3) Lemme 4.7 Dans les notations du lemme précédent, soient y , z , t ∈ Y . a) Si t ∈ α-géod(y , z), alors Ψt ∈ (α + lδ )-géod(Ψy , Ψz) et si y = w , Ψt ∈ α-géod(Ψy , Ψz). b) Si Ψt ∈ α-géod(Ψy , Ψz), alors t ∈ (α + 2lδ )-géod(y , z) et si t = w , t ∈ α-géod(y , z). c) On a d(t, géod(y , z)) − d(Ψt, géod(Ψy , Ψz)) ≤ (l + 1)δ + 1. Démonstration. Seul c) demande une démonstration. On a d(Ψt, géod(Ψy , Ψz)) = d(Ψt, Ψy ) + d(Ψt, Ψz) − d(Ψy , Ψz) 2 . Il est évident que d(t,y)+d(t,z )−d(y ,z ) 2 ≤ d(t, géod(y , z)). Enfin d(t, y ) + d(t, z) − d(y , z) d(t, géod(y , z)) ≤ 2 car si v ∈ géod(y , z) est tel que d(y , v ) = E ( d(t,y)+d(y ,z )−d(t,z ) ) on a d(t, v ) ≤ 2 d(t,y)+d(t,z )−d(y ,z ) + δ + 1 par (H 0 δ (t, y , v , z)). Le c) en résulte facilement. (cid:3) 2 + δ + 1 72 Lemme 4.8 Soient α, β ∈ N, ρ ∈ Z et x, y , a, b, ∈ X tels que a ∈ α-géod(x, y ), b ∈ β -géod(x, y ) et d(x, b) ≤ d(x, a) + ρ. Alors b ∈ (max(α + 2ρ, β ) + δ )-géod(x, a). b x β α a y Démonstration. Par (Hδ (a, x, b, y )) on a d(a, b) ≤ max(d(a, y ) + d(x, b) − d(x, y ), d(a, x) + d(b, y ) − d(x, y )) + δ donc d(a, b) + d(b, x) − d(a, x) ≤ max (cid:0)d(a, y ) + 2d(x, b) − d(a, x) − d(x, y ), d(b, x) + d(b, y ) − d(x, y )(cid:1) + δ ≤ max (cid:0)α + 2ρ, β (cid:1) + δ. Lemme 4.9 Soient k ∈ N et x, y , z , t des points de X tels que z appartienne à B (x, k) et que t soit un point de B (x, k) à distance minimale de y . Alors t ∈ δ -géod(z , y ). Démonstration. L’énoncé est clair si y ∈ B (x, k) car alors t = y . (cid:3) z x t y Sinon on a t ∈ géod(x, y ), d(x, z) ≤ k = d(x, t) et d(y , z) ≥ d(y , t) d’où par (H 0 δ (z , x, t, y )), d(z , t) ≤ max(d(z , x) − d(x, t), d(z , y ) − d(t, y )) + δ = d(z , y ) − d(t, y ) + δ d’où t ∈ δ -géod(z , y ). (cid:3) Lemme 4.10 Soient k , µ, ν ∈ N, x, y , y ′ ∈ X et z ∈ B (x, k) vérifiant 73 – y ′ 6∈ B (x, k + µ+δ 2 ), – y ′ ∈ µ-géod(z , y ) et d(z , y ′) ≤ d(z , y ) − ν . Alors y ′ ∈ (µ + δ )-géod(x, y ). d(x, y ′ ) ≤ d(x, y ) − ν + δ et z x y ′ y Démonstration. On applique le lemme 4.6 à {x, y , y ′ , z} avec l = 1 et y ′ comme point base. Soit T et Ψ comme dans le lemme 4.6. Ψz Ψx Ψy ′ t Ψy Soit t le point de géod(Ψz , Ψy ) à distance minimale de Ψy ′ . On a donc t ∈ géod(Ψz , Ψy ′ ) t ∈ géod(Ψz , Ψy ) et On a Ψy ′ ∈ (µ+δ )-géod(Ψz , Ψy ) par le a) du lemme 4.7, donc d(t, Ψy ′) ≤ µ+δ 2 . Comme d(Ψx, Ψy ′) = d(x, y ′) > k + µ+δ 2 par hypothèse, on a d(Ψx, t) > k et comme d(Ψx, Ψz) ≤ k , t appartient à géod(Ψx, Ψy ′ ). à géod(Ψx, Ψy ) et (37) (38) Il résulte de (37) et (38) que d(Ψx, Ψy ′) − d(Ψx, Ψy ) = d(t, Ψy ′) − d(t, Ψy ) = d(Ψz , Ψy ′) − d(Ψz , Ψy ) et donc d(x, y ′) − d(x, y ) ≤ d(z , y ′) − d(z , y ) + δ ≤ −ν + δ où la dernière inégalité a lieu par hypothèse. Enfin par la première partie de (38), et comme d(t, Ψy ′) ≤ µ+δ 2 , on a Ψy ′ ∈ (µ + δ )-géod(Ψx, Ψy ), d’où y ′ ∈ (µ + δ )-géod(x, y ) par le b) du lemme 4.7. (cid:3) Le lemme suivant est une conséquence du précédent. 74 . et d(z , yi+1) ≤ d(z , yi ) − ν1 , Lemme 4.11 Soient µ1 , ν1 , µ2 , ν2 ∈ N vérifiant 3δ 3δ et ν1 > µ1 + ν2 > µ2 + 2 2 Soient k ∈ N et x ∈ X . Soit y0 , ..., yj une suite de points de X tel le que y1 , . . . , yj n’appartiennent pas à B (x, k + max(µ1 ,µ2 )+δ ) et que pour tout i ∈ 2 {0, ..., j − 1}, – ou bien yi+1 = yi , – ou bien il existe z ∈ B (x, k) tel que yi+1 ∈ µ1 -géod(z , yi ) et – ou bien il existe z ∈ B (x, k) tel que yi+1 ∈ µ2 -géod(z , yi ) et d(z , yi+1) ≤ d(z , yi ) − ν2 . a) On a d(x, yj ) ≤ · · · ≤ d(x, y1) ≤ d(x, y0) et pour i ∈ {0, . . . , j − 1} on a d(x, yi+1) < d(x, yi ) si yi 6= yi+1 . b) Soit h le nombre de valeurs prises par la suite y0 , . . . , yj . Alors yj ∈ (cid:0) max(µ1 , µ2) + 2δ(cid:1)-géod(x, y0) d(x, yj ) ≤ d(x, y0) − (h − 1)(min(ν1 , ν2) − δ ). Démonstration. Montrons a). Soit i ∈ {0, . . . , j − 1} tel que yi 6= yi+1 . On va montrer d(x, yi+1) < d(x, yi). Il existe c ∈ {1, 2} et z ∈ B (x, k) tels que yi+1 ∈ µc -géod(z , yi ) et d(z , yi+1) ≤ d(z , yi ) − νc . Comme yi+1 6∈ B (x, k + µc+δ 2 ), en appliquant le lemme 4.10 à (yi , yi+1 , z) au lieu de (y , y ′, z) et (µc , νc ) au lieu de (µ, ν ), on obtient d(x, yi+1) ≤ d(x, yi ) − νc + δ < d(x, yi) puisque νc > δ . Pour montrer b), on procède par récurrence ascendante. On pose λ = max(µ1 , µ2) + 2δ. Pour i ∈ {0, . . . , j } on note hi le nombre de valeurs prises par la suite y0 , . . . , yi , de sorte que 1 = h0 ≤ h1 ≤ · · · ≤ hj = h. Par l’hypothèse de récurrence on a yi ∈ λ-géod(x, y0) et d(x, yi ) ≤ d(x, y0) − (hi − 1)(min(ν1 , ν2) − δ ). Si yi+1 = yi , on a hi+1 = hi et yi+1 satisfait l’hypothèse de récurrence. Sinon, soit c ∈ {1, 2} et z ∈ B (x, k) tel que yi+1 ∈ µc -géod(z , yi ) et d(z , yi+1) ≤ d(z , yi ) − νc . 75 (39) En appliquant le lemme 4.10 à (yi , yi+1 , z) au lieu de (y , y ′, z) et (µc , νc ) au lieu de (µ, ν ) on obtient yi+1 ∈ (µc + δ )-géod(x, yi ) et d(x, yi+1) ≤ d(x, yi) − νc + δ. Comme d(yi , yi+1) ≥ d(z , yi ) − d(z , yi+1) ≥ νc et grâce aux premières parties de (39) et (40), le lemme 3.21 appliqué à (x, y0 , yi , yi+1) au lieu de (x, a, b, c) et (λ, µc + δ ) au lieu de (α, β ) montre que yi+1 ∈ λ-géod(x, y0) puisque max(λ + 2(µc + δ ) − 2νc , µc + δ ) + δ ≤ λ. D’autre part les deuxièmes parties de (39) et (40) impliquent immédiatement (40) d(x, yi+1) ≤ d(x, y0) − (hi+1 − 1)(min(ν1 , ν2) − δ ). (cid:3) On rappelle que pour x ∈ X et S ∈ ∆, on note dmax (x, S ) = maxy∈S d(x, y ). Lemme 4.12 Il existe une constante D = C (δ, K, N , Q, P ) tel le que le ré- sultat suivant soit vrai. Soit S0 ∈ ∆, x ∈ X , k ∈ N. Alors pour tout m ∈ N et pour toute suite S1 , . . . , Sm de ∆ vérifiant les conditions i) et ii) de la définition 4.1, c’est-à-dire – i) pour tout i ∈ {0, . . . , m − 1}, Si+1 ⊂ Si ∪ [x∈B (x,k),a∈Si {y ∈ 4δ -géod( x, a), d(y , a) ∈]N − 2δ, QN ]} Q ∪ [x∈B (x,k),a∈Si F }, – ii) pour tout i ∈ {1, . . . , m}, d(x, Si ) > k + P , on a S0 ∪ · · · ∪ Sm ⊂ [a∈S0 (cid:0)F + 2δ(cid:1)-géod(x, a), dmax (x, S0 ) ≥ dmax (x, S1) ≥ · · · ≥ dmax (x, Sm ), {z ∈ F -géod( x, a), d(z , a) ≥ (41) (42) et – si d(x, S0) ≤ k on a m = 0 – si d(x, S0) > k , le nombre de valeurs prises par la suite S0 , . . . , Sm est inférieur ou égal à D(d(x, S0) − k) et le nombre de possibilités pour (S1 , . . . , Sm ) est fini et majoré par eD(d(x,S0 )−k+m) . 76 Démonstration. On applique le lemme 4.11 à (Ci ) -ou bien µ1 = 4δ, ν1 = N − 6δ, µ2 = F et ν2 = Q/F − F . Pour tout i ∈ {0, . . . , m − 1} et tout yi+1 ∈ Si+1 il existe yi ∈ Si tel que  -ou bien yi+1 = yi , il existe x ∈ B (x, k) tel que -ou bien yi+1 ∈ µ1 -géod( x, yi ) et d( x, yi+1) ≤ d( x, yi ) − ν1 , il existe x ∈ B (x, k) tel que  yi+1 ∈ µ2 -géod( x, yi ) et d( x, yi+1) ≤ d( x, yi ) − ν2 . 2 et P ≥ max(µ1 ,µ2 )+δ On suppose ν1 > µ1 + 3δ 2 , ν2 > µ2 + 3δ , ce qui est permis 2 par (HN ), (HQ) et (HP ) respectivement. Soit i ∈ {1, ..., m} et yi ∈ Si . Il existe yi−1 ∈ Si−1 , ..., y0 ∈ S0 tels que les conditions (Ci−1), ..., (C0) soient satisfaites. Le a) du lemme 4.11 montre alors que d(x, yi ) ≤ d(x, yi−1) et comme yi ∈ Si est arbitraire il résulte que dmax (x, Si ) ≤ dmax (x, Si−1) et on a montré (42). Comme max(µ1 , µ2) + 2δ = F + 2δ , le b) du lemme 4.11 montre que yi ∈ [a∈S0 (cid:0)F + 2δ(cid:1)-géod(x, a) et on a montré (41). Pour montrer la suite de l’énoncé on suppose d’abord d(x, S0 ) ≤ k . Alors m = 0 par (42) et par la condition ii), car P ≥ N . On suppose d(x, S0) > k dans toute la suite de la démonstration. Soit ym ∈ Sm et soient ym−1 ∈ Sm−1 , ..., y0 ∈ S0 tels que les conditions (Cm−1 ), ..., (C0) soient satisfaites. On note h le nombre de valeurs différentes prises par la suite y0 , . . . , ym . On suppose min(ν1 , ν2) = N − 6δ , ce qui est permis par (HQ). Le b) du lemme 4.11 montre que ym ∈ (cid:0)F + 2δ(cid:1)-géod(x, y0) et d(x, ym) ≤ d(x, y0) − (h − 1)(N − 7δ ). On a d(x, ym) ≥ k + P puisque m ≥ 1. On en déduit (h − 1)(N − 7δ ) ≤ d(x, S0 ) + N − (k + P ). (43) On suppose N − 7δ ≥ 1 et P ≥ N + 1, ce qui est permis par (HN ) et (HP ). Alors (43) implique h ≤ d(x, S0) − k . Il existe une constante C1 = C (δ, K, N ) telle que tout point de X appartienne au plus à C1 éléments de ∆. En notant l le nombre de valeurs prises par la suite S0 , . . . , Sm , on a h ≥ l , d’où C1 l ≤ C1h ≤ C1(cid:0)d(x, S0) − k(cid:1). 77 Il existe une constante C2 = C (δ, K, N ) telle que, pour i ∈ {0, ..., m − 1}, connaissant Si et dmax (x, Si+1) le nombre de possibilités pour Si+1 vérifiant les conditions de l’énoncé est inférieur ou égal C2 . En effet pour yi+1 ∈ Si+1 il existe yi ∈ Si tel que yi+1 ∈ (F + 2δ )-géod(x, yi ) d’après (41) appliqué à (Si , ..., Sm ), et on déduit l’existence de C2 du lemme 3.13 appliqué à (x, yi ) au lieu de (x, y ). Etant donné m, le nombre de possibilités pour les entiers dmax (x, Si ) (pour i = 1, . . . , m) qui vérifient nécessairement dmax (x, S0) ≥ dmax (x, S1 ) ≥ · · · ≥ dmax (x, Sm) ≥ k + P est inférieur ou égal à (cid:19)(cid:17) ≤ 2d(x,S0 )−k+m max (cid:16)1, (cid:18)m + dmax (x, S0) − k − P m car P ≥ N . Le nombre de possibilités pour (S1 , . . . , Sm ) est donc inférieur ou égal à (C2 )m2d(x,S0 )−k+m . Ceci termine la démonstration du lemme 4.12. (cid:3) Les deux lemmes suivants sont des conséquences du lemme 4.12 et servi- ront ultérieurement. Lemme 4.13 Pour m, l0 , ..., lm ∈ N et (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x on a pour i ∈ {0, ..., m} et j ∈ {1, ..., li}, dmax (x, Y j i ) ≤ dmax (x, Si ) + 2P + δ. Démonstration. Le lemme 3.31 appliqué à α = 2P donne – pour i ∈ {0, ..., m − 1}, dmax (x, Y j i ) ≤ max(dmax (x, Si ), dmax (x, Si+1)) + 2P + δ, d’où le résultat puisque dmax (x, Si ) ≥ dmax (x, Si+1) d’après le lemme 4.12, – pour i = m, dmax (x, Y j m ) ≤ max(dmax (x, Sm ), k) + 2P + δ d’où le ré- sultat puisque lm = 0 si dmax (x, Sm ) ≤ k , par la remarque qui suit la définition 4.1. (cid:3) Lemme 4.14 Soient p ∈ {1, ..., pmax}, k , m, l0 , ..., lm ∈ N et (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x 78 . Soit b ∈ S0 et u un point de B (x, k) à distance minimale de b. Alors a) S0 ∪ · · · ∪ Sm ⊂ 2F -géod(b, u) ⊂ P -géod(b, u), b) Sj∈{1,...,lm} Y j m ∈ Sa∈Sm (2P + δ )-géod(x, a), c) pour tout i ∈ {0, . . . , m}, Sj∈{1,...,li } Y j i ⊂ 4P -géod(b, u). d) pour tout i ∈ {0, . . . , m−1}, et tout j ∈ {1, . . . , li}, d(x, Y j i ) ≥ dmax (x, Si+1 )− 4P Démonstration. On commence par traiter le cas où b ∈ B (x, k). Alors d(x, S0 ) ≤ k , d’où m = 0 par le lemme 4.12. De plus l0 = 0 par la remarque qui suit la définition 4.1, et les assertions a), b), c) et d) sont évidentes. On suppose maintenant que b 6∈ B (x, k). En particulier d(x, u) = k et d(u, b) = d(x, b) − k . Montrons a). D’après le lemme 4.12, on a S0 ∪ · · · ∪ Sm ⊂ [a∈S0 (cid:0)F + 2δ(cid:1)-géod(x, a). Comme d(a, b) ≤ N pour tout a ∈ S0 , le lemme 3.2 montre que S0 ∪ · · · ∪ Sm ⊂ (cid:0)F + 2N + 2δ(cid:1)-géod(x, b). Pour tout y ∈ S0 on a d(y , b) ≤ N , donc y ∈ 2F -géod(b, u) car F ≥ N par (22). Soit donc i ≥ 1 et y ∈ Si , et montrons y ∈ 2F -géod(b, u). On a y ∈ (cid:0)F + 2N + 2δ(cid:1)-géod(x, b) et d(x, y ) ≥ d(x, u) + P . y x u Par (H F +2N +2δ δ (u, x, y , b)) on a b d(u, y ) ≤ max(d(u, x) − d(x, y ), d(u, b) − d(b, y )) + F + 2N + 3δ. Or d(u, x) − d(x, y ) + F + 2N + 3δ ≤ −P + F + 2N + 3δ et on suppose −P + F + 2N + 3δ < 0, ce qui est permis par (HP ). Donc d(u, y ) ≤ d(u, b) − d(b, y ) + F + 2N + 3δ , c’est-à-dire y ∈ (F + 2N + 3δ )-géod(u, b). On en déduit y ∈ 2F -géod(u, b) puisque F ≥ 2N + 3δ par (22). Enfin on suppose P ≥ 2F , ce qui est permis par (HP ). On va montrer maintenant b), ainsi que c) dans le cas où i = m. Soit j ∈ {1, ..., lm} et y ∈ Y j m . On a d(x, y ) ≥ k + 3P et il existe a ∈ Sm et x ∈ B (x, k) tels que y ∈ 2P -géod( x, a). On a a ∈ P -géod(u, b) par le a). Ensuite (H 2P δ (x, x, y , a)) implique d(x, y ) ≤ max(k , d(x, a) − d(a, y )) + 2P + δ . Comme 79 d(x, y ) ≥ k+3P et P > δ on en déduit d(x, y ) ≤ d(x, a)−d(a, y )+2P +δ , c’est- à-dire y ∈ (2P + δ )-géod(x, a), ce qui montre déjà b). Par (H 2P +δ (u, x, y , a)) δ on a d(u, y ) ≤ max(d(u, x) − d(x, y ), d(u, a) − d(a, y )) + 2P + 2δ et comme d(u, x) − d(x, y ) ≤ −3P et P > 2δ on en déduit d(u, y ) ≤ d(u, a) − d(a, y ) + 2P + 2δ , c’est-à-dire y ∈ (2P + 2δ )-géod(a, u). y a x x u b Comme a ∈ P -géod(u, b), le a) du lemme 3.3 montre alors y ∈ (3P + 2δ )-géod(u, b) d’où y ∈ 4P -géod(u, b) car P ≥ 2δ . On montre maintenant c) dans le cas où i ∈ {0, ..., m − 1}, ainsi que d). Soit j ∈ {1, ..., li} et t ∈ Y j i . Il existe y ∈ Si et z ∈ Si+1 tels que t ∈ P -géod(y , z). Les hypothèses y , z ∈ P -géod(u, b) et t ∈ P -géod(y , z) suffisent à impliquer t ∈ (3P + δ )-géod(u, b). z t y x u b En effet quitte à permuter y et z on peut supposer d(u, z) + d(y , b) ≤ d(u, y ) + d(z , b). Alors (Hδ (z , u, y , b)) implique d(z , y ) ≤ d(z , b) + d(u, y ) − d(u, b) + δ donc d(u, z) + d(z , y ) + d(y , b) ≤ (d(u, z) + d(z , b)) +(d(u, y ) + d(y , b)) − d(u, b) + δ ≤ d(u, b) + 2P + δ et d(u, t) + d(t, b) ≤ d(u, z) + d(z , t) + d(t, y ) + d(y , b) ≤ d(u, z) + d(z , y ) + P + d(y , b) ≤ d(u, b) + 3P + δ. (44) 80 On a donc t ∈ (3P + δ )-géod(u, b) donc t ∈ 4P -géod(u, b) puisque P ≥ δ . Il reste à montrer d). On garde les notations du dessin ci-dessus. Par (44) on a d(x, u) + d(u, z) + d(z , t) + d(t, b) ≤ d(x, b) + 3P + δ , d’où d(x, z) + d(z , t) ≤ d(x, t) + 3P + δ . Comme d(x, z) ≥ dmax (x, Si+1 ) − N , on en déduit d(x, t) ≥ dmax (x, Si ) − N − 3P − δ ≥ dmax (x, Si+1 ) − 4P car on suppose P ≥ N + δ . (cid:3) Le lemme suivant sera utile pour la démonstration de la proposition 4.2. Lemme 4.15 Il existe une constante C = C (δ, K, N , Q, P ) tel le que pour tous x ∈ X , k , m, l0 , . . . , lm ∈ N, et a1 , . . . , ap , S0 , . . . , Sm ∈ ∆ vérifiant les conditions i) et ii) de la définition 4.1, – pour tout i ∈ {0, . . . , m} et j ∈ {1, . . . , li}, si on se donne d(x, Y j i ), le nombre de possibilités pour Y j i vérifiant la condition iii) ou iv) de la définition 4.1 (selon que i < m ou i = m) est inférieur ou égal à C , – pour tout i ∈ {0, . . . , m} et j ∈ {1, . . . , li} le nombre de possibilités pour Y j i vérifiant la condition iii) ou iv) de la définition 4.1 est inférieur ou égal à C si (Z ). – le cardinal de l’ensemble des (Y j i )i∈{0,...,m},j∈{1,...,li } ) tels que (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) est majoré par CPm appartienne à Y p,k ,m,(l0 ,...,lm) i=0 li Qm i=0 si (Z )li x Démonstration. La première assertion résulte immédiatement du c) du i ⊂ 4P -géod(x, b), du fait que diam(Y j lemme 4.14, qui implique que Y j i ) ≤ P , et du lemme 3.13 appliqué à (x, b) au lieu de (x, y ). On montre d’abord la deuxième assertion pour i ∈ {0, ..., m − 1}. Soit a ∈ Si et b ∈ Si+1 . Pour j ∈ {1, ..., li} on a i ⊂ [y∈Si ,z∈Si+1 Y j par le lemme 3.2. Comme diam(Y j i ) ≤ P , le lemme 3.13 appliqué à (a, b) au lieu de (x, y ) montre alors la deuxième assertion. On montre maintenant la deuxième assertion pour i = m. Soit a ∈ Sm . On a m ⊂ [y∈Sm Y j (2P + δ )-géod(x, y ) par le b) du lemme 4.14, d’où Y j m ⊂ (2P +δ+2N )-géod(x, a) par le lemme 3.2. Comme diam(Y j i ) ≤ P et d(x, Y j m) > k + 3P , le lemme 3.13 appliqué à (x, a) au lieu de (x, y ) montre la deuxième assertion. Enfin la troisième assertion résulte facilement de la deuxième. (cid:3) P -géod(y , z) ⊂ (P + 4N )-géod(a, b) 81 e2s(d(x,S0 )−k) f (S0)2 Démonstration de la proposition 4.2. Soit p ∈ {1, . . . , pmax} et f ∈ C(∆p ) . Soit R = max dmax (x, S ) où le maximum est pris sur les S tels que eS apparaisse dans f avec un coefficient non nul. Soit k ≥ R. Si m, l0 , ..., lm ∈ N et (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x sont tels que f (a1 , ..., ap) 6= 0, alors dmax (x, S0) ≤ k , donc m = 0 et l0 = 0 par le lemme 4.12 et la remarque qui suit la définition 4.1. De plus pour k ≥ R la relation d’équivalence sur la partie de Y p,k ,0,(0) telle que eS0 apparaisse dans f x avec un coefficient non nul, est triviale, donc la partie de (36) correspondant à k ≥ R se réécrit p! XS0 ,k≥R et elle est finie et ma jorée par p!(1 − e−2s )−1kf k2 ℓ2 (∆p ) . On fixe k ∈ {0, . . . , R − 1}. Il reste donc à montrer que la partie corres- pondante de (36) est une somme convergente. La somme B−(m+Pm i=0 li ) XZ ∈Y Xm∈N,(l0 ,...,lm )∈Nm+1 p,k,m,(l0 ,...,lm ) x m e2s(r0 (Z )−k)(cid:16) si (Z )−li (cid:17)♯((π p,k ,m,(l0 ,...,lm) 2 Yi=0 )−1(Z ))−α (cid:12)(cid:12)ξZ (f )(cid:12)(cid:12) x est ma jorée par m Yi=0 si (Z )−li (cid:17)(cid:12)(cid:12)ξZ (f )(cid:12)(cid:12) 2 (45) Soit D comme dans le lemme 4.12. Grâce au lemme 4.12, pour tout m ∈ N et pour tout S0 ∈ ∆ vérifiant S0 ⊂ B (x, R), le nombre de possibilités pour (S1 , . . . , Sm) est inférieur ou égal à max(1, eD(R−k+m) ) ≤ eD(R+m) . Soit C égal à la constante C du lemme 4.15 (qui est de la forme C (δ, K, N , Q, P )). p,k ,m,(l0 ,...,lm) Pour tous m, (l0 , . . . , lm ), et pour tout Z ∈ Y l’application de x (π p,k ,m,(l0 ,...,lm) )−1(Z ) dans ∆p qui à x (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } ) associe S0 a des fibres de cardinal ≤ p!eD(R+m)CPm i=0 li , car connaissant S0 on a p! possibilités pour (a1 , . . . , ap ), au plus eD(R+m) possibilités pour (S1 , ..., Sm), e2sR Xm∈N,(l0 ,...,lm )∈Nm+1 B−(m+Pm i=0 li ) XZ ∈Y p,k,m,(l0 ,...,lm ) x (cid:16) 82 e2sR plus CPm i=0 li grâce et au 4.15, lemme au pour possibilités (Y j i )i∈{0,...,m},j∈{1,...,li } , puisque Z détermine, pour tous i ∈ {0, . . . , m} et j ∈ {1, . . . , li}, l’entier d(x, Y j i ). Donc dans (45) on a toujours (cid:12)(cid:12)ξZ (f )(cid:12)(cid:12) ≤ p!eD(R+m)CPm i=0 li kf kℓ1 (∆p ) . Donc la somme (45) est ma jorée par B−(m+Pm i=0 li ) XZ ∈Y Xm∈N,(l0 ,...,lm )∈Nm+1 p,k,m,(l0 ,...,lm ) x m (cid:16) si (Z )−li (cid:17)(p!)2e2D(R+m)C 2 Pm Yi=0 i=0 li kf k2 ℓ1 (∆p ) Il existe D ′ = C (δ, K ) tel que le nombre de S0 ∈ ∆p inclus dans B (x, R) soit inférieur ou égal à eD ′R pour tout R ∈ N. Par Cauchy-Schwarz, on ℓ1 (∆p ) ≤ eD ′Rkf k2 en déduit kf k2 ℓ2 (∆p ) . Pour tous m, (l0 , . . . , lm) le cardinal de p,k ,m,(l0 ,...,lm) Y p,k ,m,(l0 ,...,lm) , sont ma jorés par , et donc a fortiori celui de Y x x m i=0 li (cid:16) si (Z )li (cid:17), p!eD(R+m)+D ′ RCPm Yi=0 car on a au plus eD ′R possibilités pour S0 , p! possibilités pour (a1 , . . . , ap ), au plus eD(R+m) possibilités pour (S1 , ..., Sm), et grâce au lemme 4.15, au plus CPm i=0 li Qm i=0 si (Z )li possibilités pour (Y j i )i∈{0,...,m},j∈{1,...,li } . Donc la somme (46) est ma jorée par Xm∈N,(l0 ,...,lm )∈Nm+1 (p!)3e2sR On suppose B > C 3 + e3D , ce qui est permis par (HB ). Donc cette somme converge et est ma jorée par B−m e3Dm (p!)3e(2s+3D+2D ′ )R Xm∈N (1 − B−1C 3 )m+1 kf k2 ℓ2 (∆p ) = 1 1 (p!)3e(2s+3D+2D ′ )R 1−B−1C 3 kf k2 ℓ2 (∆p ) 1 − B−1 e3D 1 − B−1C 3 (p!)3e(2s+3D+2D ′ )R 1 − B−1(C 3 + e3D ) kf k2 ℓ2 (∆p ) . Cela termine la démonstration de la proposition 4.2. i=0 li ) e3D(R+m)+2D ′ RC 3 Pm B−(m+Pm i=0 li kf k2 ℓ2 (∆p ) . = 83 (46) (cid:3) Au total, pour p ∈ {1, . . . , pmax} et f ∈ C(∆p ) et en notant R = max dmax (x, S ) où le maximum est pris sur les S tels que eS apparaisse dans f avec un co- efficient non nul, on a (47) R−1 (p!)3 e(2s+3D+2D ′ )R Xk=0 Hx,s (∆p ) ≤ p!(1 − e−2s )−1kf k2 kf k2 1 − B−1 (C 3 + e3D ) kf k2 ℓ2 (∆p ) + ℓ2(∆p ) (p!)3Re(2s+3D+2D ′ )R ≤ (cid:16)p!(1 − e−2s )−1 + 1 − B−1 (C 3 + e3D ) (cid:17)kf k2 ℓ2 (∆p ) . Le lemme suivant donne au contraire une minoration de la norme de Hx,s (∆p ). Lemme 4.16 Pour tout p ∈ {1, . . . , pmax} et pour tout f ∈ C(∆p ) on a Hx,s (∆p ) ≥ p!(1 − e−2s )−1kf k2 kf k2 ℓ2 (∆p ) . Démonstration. Pour tout S0 ∈ ∆p , pour tout k ≥ d(x, S0 ) et pour toute énumération (a1 , . . . , ap ) des points de S0 , Y p,k ,0,(0) contient (a1 , . . . , ap , S0) x et comme M ≥ N , le singleton {(a1 , . . . , ap , S0 )} est une classe d’équivalence p,k ,0,(0) dans Y . On a donc x Hx,s (∆p ) ≥ p! XS0∈∆p Xk≥d(x,S0 ) kf k2 e2s(d(x,S0 )−k) f (S0)2 = p!(1 − e−2s )−1kf k2 ℓ2(∆p ) . (cid:3) 4.4 Autres propriétés de la norme Les propriétés que nous allons établir dans ce sous-paragraphe sont des préliminaires indispensables pour les sous-paragraphes suivants. De façon un peu imprécise nous allons montrer qu’il existe une constante C de la forme C (δ, K, N , Q, P , M ) telle que pour p ∈ {1, ..., pmax}, k , m, l0 , ..., lm ∈ N et (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) , x pour connaître les distances entre les points de [ [i∈{0,...,m} B (Y j B (Si , M ) ∪ i , M ) ∪ B (x, k + 2M ) i∈{0,...,m},j∈{1,...,li } il suffit de connaître certaines de ces distances, de telle sorte que pour chaque point de cet ensemble n’appartenant pas à B (x, k + 2M ), le nombre de dis- tances à connaître depuis de ce point soit inférieur ou égal à C . La raison est que cet ensemble est une réunion de boules de grands rayons dont les centres sont à peu près alignés le long d’une géodésique (grâce au lemme 4.14) et qu’il suffit donc de connaître les distances entre les points de boules voisines. 84 + . (48) Lemme 4.17 Soient c, d ∈ X , I un ensemble fini, et pour i ∈ I , αi ∈ N, ρi ∈ N∗ et wi ∈ αi -géod(c, d). On suppose que pour tout i ∈ I , ρi ≥ 4δ + αi 2 . a) Pour connaître les distances entre les points de Si∈I B (wi , ρi ) il suffit de connaître les distances entre les points de B (wi , ρi ) et B (wj , ρj ) pour tous les couples (i, j ) ∈ Λ où Λ ⊂ I 2 est l’ensemble des couples (i, j ) tels qu’il n’existe pas de k ∈ I vérifiant αj αi αk d(c, wj ) − d(c, wk ) ≥ 2ρj + et d(c, wk ) − d(c, wi) ≥ 2ρi + 2 2 2 b) Soit R ∈ N∗ tel que αi ≤ R et ρi ≤ R pour tout i ∈ I . Soit i ∈ I . Alors l’ensemble des d(c, wj ) pour j ∈ I tel que (i, j ) ∈ Λ ou (j, i) ∈ Λ est inclus dans la réunion de [d(c, wi) − 3R, d(c, wi) + 3R] et de deux interval les de longueur ≤ 3R. c) Soit R comme dans b). Il existe une constante C = C (δ, K, R) (indépen- dante de ♯I en particulier) tel le que pour toute partie J ⊂ I , connaissant les distances entre les points de Si∈I \J B (wi , ρi ), les distances entre les points de Si∈J B (wi , ρi ) et ceux de Si∈I \J B (wi , ρi ) soient déterminées par la donnée des distances entre les points de Si∈J B (wi , ρi ) et les points d’une partie de Si∈I \J B (wi , ρi ) – dont le cardinal est borné par C (♯J ), – qui est déterminée par la connaissance des distances entre les points de [i∈I \J et par les entiers d(c, wi) pour i ∈ J . Remarque. Dans toutes les situations où on appliquera le c) de ce lemme, ♯J sera ma joré par une constante de la forme C (δ, K, N , Q, P , M ). Démonstration. On commence par montrer a). On suppose I = {1, ..., r} et w1 , ..., wr ordonnés de telle sorte que d(c, w1) ≤ · · · ≤ d(c, wr ). Soit i < k < j des entiers vérifiant (48). Soient y ∈ B (wi , ρi ) et z ∈ B (wj , ρj ). On a αk αi 2 − d(c, y ) ≤ d(c, wi) + ρi ≤ d(c, wk ) − ρi − 2 αj et d(c, z) ≥ d(c, wj ) − ρj ≥ d(c, wk ) + ρj + . 2 Donc il existe v ∈ géod(y , z) tel que d(c, v ) = d(c, wk ) − E ( αk 2 ). On choisit un tel v . B (wi , ρi ) ∪ {c} 85 z wj v wk y wi c d On a d(y , v ) ≥ d(c, v ) − d(c, y ) ≥ ρi + αi/2 et d(z , v ) ≥ d(c, z) − d(c, v ) ≥ ρj + αj /2. D’après le lemme 3.2, on a y ∈ (αi+2ρi )-géod(c, d) et z ∈ (αj+2ρj )-géod(c, d). Le lemme 3.40 implique alors v ∈ 3δ -géod(c, d). Comme wk ∈ αk -géod(c, d) et d(c, v ) = d(c, wk ) − E ( αk 2 ), (Hδ (v , c, wk , d)) implique d(v , wk ) ≤ max(d(c, v ) + d(wk , d) − d(c, d), d(c, wk ) + d(v , d) − d(c, d)) + δ ≤ αk αk max(d(c, wk ) − E ( ) + d(v , d) − d(c, d)) ) + d(wk , d) − d(c, d), d(c, v ) + E ( 2 2 αk αk αk +δ ≤ max(αk − E ( 2 ≤ ρk , ) + 3δ ) + δ ≤ 4δ + ), E ( 2 2 donc v ∈ B (wk , ρk ) ∩ géod(y , z). On a donc montré que pour i < k < j vérifiant (48) toute géodésique entre un point de B (wi , ρi ) et un point de B (wj , ρj ) intersecte B (wk , ρk ), ce qui implique que la connaissance des dis- tances entre les points de B (wk , ρk ) et ceux de B (wi , ρi ) ∪ B (wj , ρj ) permet de déterminer les distances entre les points de B (wi , ρi ) et ceux de B (wj , ρj ). Ceci termine la preuve du a). Le b) résulte facilement du a). Le cas particulier de c) où ♯J = 1 résulte du b) et du lemme 3.13. Pour montrer c) dans le cas général on se ramène au cas particulier déjà démontré de la façon suivante. Pour tout j ∈ J on note I ′ = I \ J ∪ {j } et J ′ = {j } et on applique le cas particulier de c) déjà démontré avec I ′ et J ′ au lieu I et J . (cid:3) Lemme 4.18 Soient x, b ∈ X, l ∈ N, α ∈ N et u un point de B (x, l) à distance minimale de b. Pour tout z ∈ α-géod(b, u), les distances entre z et les points de B (x, l) sont déterminées par les distances entre z et les points de B (u, α + 4δ ) ∩ B (x, l). Plus précisément pour tout x ∈ B (x, l) et pour tout z ∈ α-géod(b, u), géod( x, z) intersecte B (u, α + 4δ ) ∩ B (x, l). Démonstration. Le lemme est évident si b ∈ B (x, l). On suppose donc b 6∈ B (x, l). Soit x ∈ B (x, l) et z ∈ α-géod(b, u). On veut montrer qu’il 86 existe w ∈ géod( x, z) ∩ B (u, α + 4δ ) ∩ B (x, l). Si d(u, x) ≤ α + 4δ on prend w = x. On suppose donc d(u, x) > α + 4δ . D’après le lemme 4.9 appliqué à (x, b, u, x) au lieu de (x, y , t, z) et l au lieu de k , on a u ∈ δ -géod( x, b). Donc d(b, x) ≥ d(b, u) + α + 3δ et comme de plus d(b, z) ≤ d(b, u) + α, il existe w ∈ géod(z , x) tel que d(b, w) = d(b, u) + α + 2δ . On choisit un tel w . w x z x u b Comme u ∈ δ -géod( x, b) et z ∈ α-géod(b, u) le a) du lemme 3.3 montre que z ∈ (α + δ )-géod( x, b). On en déduit que w ∈ (α + δ )-géod( x, b). Alors (Hδ (w , x, u, b)) montre que d(w , u) ≤ max(d(w , x) − d(b, x) + d(b, u) + δ, d(w , b) − d(b, x) + d( x, u) + δ ). Or d(w , x) − d(b, x) + d(b, u) + δ ≤ −d(b, w) + α + δ + d(b, u) + δ = 0 et d(w , b) − d(b, x) + d( x, u) + δ ≤ d(w , b) − d(b, u) + 2δ = α + 4δ . On en déduit d(w , u) ≤ α + 4δ . De plus (Hδ (x, x, w , b)) montre que d(x, w) ≤ max(d(x, x) − d(b, x) + d(b, w) + δ, d(x, b) − d(b, x) + d( x, w) + δ ). Or d(x, x) − d(b, x) + d(b, w) + δ ≤ l car d(x, x) ≤ l et d(b, x) ≥ d(b, w) + δ et d(x, b) − d(b, x) + d( x, w) + δ ≤ d(x, b) − d(b, w) + α + 2δ = l car d(b, w) = d(b, u) + α + 2δ et d(x, b) = l + d(u, b). Donc on a bien w ∈ B (x, l). (cid:3) Le lemme suivant est une conséquence des deux précédents. Lemme 4.19 Il existe une constante C = C (δ, K, N , Q, P , M ) tel le que pour p ∈ {1, ..., pmax}, k , m, l0 , ..., lm ∈ N et (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x les distances entre les points de B (S0 , M ) ∪ [j∈{1,...,l0 } B (Y j 0 , M ) , (49) 87 B (Si , M ) ∪ [ i∈{1,...,m},j∈{1,...,li } B (Y j i , M ) ∪ B (x, k + 2M ) et ceux de [i∈{1,...,m} sont déterminées par – a) les distances entre les points de (50), – b) les entiers d(x, S0) et d(x, Y j 0 ), – c) les distances entre les points de (49) et C (1 + l0) points de (50) (qui sont eux-mêmes déterminés par a) et b)) et de plus les distances entre les points de (49) et ceux de (50) sont détermi- nées à C près par a) et b). (50) Remarque. Dans toutes les situations où on appliquera ce lemme, l0 sera ma joré par une constante de la forme C (δ, K, N , Q, P , M ). Démonstration. Soient p ∈ {1, ..., pmax}, k , m, l0 , ..., lm ∈ N et (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x On applique le lemme 4.17 (complété par la remarque que, dans les notations de ce lemme, 4P -géod(u, b) ⊂ 4P -géod(x, b)) avec – (x, b) au lieu de (c, d), – {wi , i ∈ I } égal à Si∈{0,...,m} Si ∪ Si∈{0,...,m},j∈{1,...,li } Y j i , – J la partie de I telle que {wj , j ∈ J } = S0 ∪ Sj∈{1,...,l0 } Y j 0 , – et (αi , ρi ) égal à (4P , M ) pour tout i (les hypothèses sont satisfaites grâce au lemme 4.14). Puis on applique le lemme 4.18 à z parcourant (49), α = 4P + 2M et l = k + 2M (grâce aux lemmes 4.14 et 3.2, (49) est inclus dans (4P +2M )-géod(b, u)). (cid:3) . 4.5 Continuité de ∂ et Jx On introduit d’abord une variante k.kH→ x,s de la norme k.kHx,s et on montre que ces deux normes sont équivalentes. Soient p ∈ {1, ..., pmax} et k , m, l0 , ..., lm ∈ N. On note Y →,p,k ,m,(l0 ,...,lm) x l’ensemble défini de la même façon que Y p,k ,m,(l0 ,...,lm) mais en a joutant la x condition – v) ou bien d(x, S0 ) > k + P , ou bien k = 0, m = 0, l0 = 0, d(x, S0) ≤ P . →,p,k ,m,(l0 ,...,lm) et l’application π→,p,k ,m,(l0,...,lm) On définit le quotient Y : x x →,p,k ,m,(l0 ,...,lm) p,k ,m,(l0 ,...,lm) Y →,p,k ,m,(l0,...,lm) → Y et de la même façon que Y x x x 88 π p,k ,m,(l0 ,...,lm) x Y Y →,p,k ,m,(l0 ,...,lm) x ↓ →,p,k ,m,(l0 ,...,lm) x . Cela fournit un diagramme commutatif ֒→ Y p,k ,m,(l0 ,...,lm) x ↓ p,k ,m,(l0 ,...,lm) ֒→ Y x où les flèches verticales sont des surjections. Ce diagramme est cartésien au sens où les flèches horizontales induisent des bijections sur les fibres des flèches verticales. x,s (∆p ) la norme sur C(∆p ) donnée par la formule (36) en On note k.kH→ p,k ,m,(l0 ,...,lm) par Y →,p,k ,m,(l0 ,...,lm) et π p,k ,m,(l0 ,...,lm) remplaçant Y p,k ,m,(l0 ,...,lm) , , Y x x x x →,p,k ,m,(l0,...,lm) et π→,p,k ,m,(l0 ,...,lm) . On remarque que les sous-espaces de C(∆p ) Y x x engendrés par les eS pour d(x, S ) ≤ P , resp. d(x, S ) > P sont orthogonaux pour la norme pré-hilbertienne k.kH→ x,s (∆p ) , et que sur le premier la norme est donnée par kf k2 x,s (∆p ) = p! PS e2sd(x,S ) f (S )2 , car si S ∈ ∆p vérifie H→ d(x, S ) ≤ P , on a S ⊂ B (x, M ). En effet on suppose M ≥ P + N , ce qui est permis par (HM ). . Lemme 4.20 Les normes k.kHx,s (∆p ) et k.kH→ x,s (∆p ) sont équivalentes. Démonstration. D’abord il est évident que k.kH→ x,s (∆p ) ≤ k.kHx,s (∆p ) . Soit (a1 , . . . , ap , S0 , ..., Sm , (Y j −Y →,p,k ,m,(l0,...,lm) i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x x On a alors d(x, S0) ≤ k + P et d(x, Si ) > k + P pour i ≥ 1. Le lemme 4.12 montre que dmax (x, S0 ) ≥ dmax (x, S1) ≥ · · · ≥ dmax (x, Sm ). Donc S0 ∪ · · · ∪ Sm ⊂ B (x, k + N + P ). Le lemme 4.13 implique alors [ i∈{0,...,m},j∈{1,...,li } Grâce à (HM ) on suppose M ≥ N + 3P + δ , d’où [ [i∈{0,...,m} B (Y j B (Si , M ) ∪ i , M ) ⊂ B (x, k + 2M ). i∈{0,...,m},j∈{1,...,li } Donc l’ensemble des points entre lesquels on veut connaître les distances, à savoir Y j i ⊂ B (x, k + N + 3P + δ ). [i∈{0,...,m} B (Si , M ) ∪ [ i∈{0,...,m},j∈{1,...,li } B (Y j i , M ) ∪ B (x, k + 2M ) 89 B−(m+Pm i=0 li ) est simplement égal à B (x, k + 2M ). Il en résulte que le singleton {(a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } )} p,k ,m,(l0 ,...,lm) est une classe d’équivalence dans Y . On a donc x x,s (∆p ) ≤ XS0∈∆p Xk≥d(x,S0 )−P e2s(d(x,S0 )−k) Xm,l0 ,...,lm∈N Hx,s (∆p ) − kf k2 kf k2 H→ si (Z )−li (cid:17)!f (S0)2 . m (cid:16) X Yi=0 (a1 ,...,ap ,S0 ,...,Sm ,(Y j i )i∈{0,...,m},j∈{1,...,li } ) p,k,m,(l0 ,...,lm ) tel que S0={a1 ,...,ap } ∈Y x Soient D et C comme dans les lemmes 4.12 et 4.15 (on rappelle que ce sont des constantes de la forme C (δ, K, N , Q, P )). Etant donnés S0 , k , m, l0 , . . . , lm tels que k ≥ d(x, S0) − P , le nombre de possibilités pour (S1 , . . . , Sm) est inférieur ou égal à eD(P +m) d’après le lemme 4.12 et le nombre de possibilités i )i∈{0,...,m},j∈{1,...,li } est inférieur ou égal à CPm i=0 li Qm pour (Y j i=0 si (Z )li d’après le lemme 4.15. Donc x,s (∆p ) ≤ p! XS0∈∆p Xk≥d(x,S0 )−P Hx,s (∆p ) − kf k2 kf k2 H→ i=0 li !f (S0)2 B−(m+Pm i=0 li ) eD(P +m)CPm e2s(d(x,S0 )−k) f (S0)2 e2s(d(x,S0 )−k) Xm,l0 ,...,lm∈N eDP 1 − B−1 (C + eD ) XS0∈∆p Xk≥d(x,S0 )−P = p! e(D+2s)P (1 − B−1(C + eD ))(1 − e−2s ) kf k2 ≤ p! ℓ2 (∆p ) . Pour (a1 , . . . , ap , S0) ∈ Y p,k ,0,(0) tel que S0 vérifie x – ou bien d(x, S0) = k + P + 1 – ou bien k = 0 et d(x, S0 ) ≤ P , →,p,k ,0,(0) le singleton {(a1 , . . . , ap , S0)} est une classe d’équivalence dans Y x En limitant la somme qui définit kf k2 x,s (∆p ) à ces éléments-là, on voit que H→ x,s (∆p ) ≥ p!kf k2 kf k2 ℓ2 (∆p ) . H→ On déduit des inégalités (51) et (52) qu’il existe une constante C = C (δ, K, N , Q, P , M , s, B ) telle que kf k2 Hx,s (∆p ) ≤ C kf k2 x,s (∆p ) pour tout f ∈ H→ C(∆p ) . (cid:3) (51) . (52) 90 B−(m+Pm i=0 li ) Soit P le pro jecteur orthogonal sur le sous-espace vectoriel de H→ x,s(∆p ) engendré par les eS pour S ∈ ∆p tel que d(x, S ) ≤ P , de sorte que (P f )(S ) = f (S ) si d(x, S ) ≤ P et (P f )(S ) = 0 sinon. Pour f ∈ C(∆p ) on a X x,s (∆p ) = Xk ,m,l0 ,...,lm∈N k(1 − P )f k2 H→ p,k,m,(l0 ,...,lm ) Z ∈Y ,r0 (Z )>k+P x m e2s(r0 (Z )−k)(cid:16) si (Z )−li (cid:17)♯(cid:0)(π p,k ,m,(l0 ,...,lm) )−1(Z )(cid:1)−α (cid:12)(cid:12)ξZ (f )(cid:12)(cid:12) 2 Yi=0 x Cette formule est la raison pour laquelle on a introduit la norme k.kH→ x,s (∆p ) . En effet pour montrer la continuité de ∂ ou de Jx on cherchera à ma jorer ξZ (∂f )2 ou ξZ (Jxf )2 par une combinaison de ξ Z (f )2 avec Z vérifiant notamment r1( Z ) = r0(Z ). Comme la condition r1( Z ) > k + P est imposée par la condition ii) de la définition 4.1, il est très utile d’avoir r0 (Z ) > k + P . (53) Proposition 4.21 Pour tout p ∈ {1, . . . , pmax}, ∂ se prolonge en un opéra- teur continu de Hx,s(∆p ) dans Hx,s(∆p−1). Démonstration. Supposons d’abord p = 1. On va montrer qu’il existe une constante C = C (δ, K, N , Q, P , M , s, B ) telle que Pa∈X f (a) ≤ C kf kHx,s (∆1 ) . D’après la formule (36), kf k2 Hx,s (∆1 ) est une somme sur k , m, l0 , . . . , lm et en limitant cette somme à k = 0, m = 0, l0 = 0 on voit que 2 f (a)(cid:12)(cid:12)(cid:12) )−1(Z )(cid:1)−α (cid:12)(cid:12)(cid:12) Hx,s (∆1 ) ≥ XZ ∈Y X kf k2 e2sr0 (Z )♯(cid:0)(π 1,0,0,(0) x 1,0,0,(0) (a,{a})∈(π1,0,0,(0) x x 1,0,0,(0) De plus Y s’identifie au quotient de X pour la relation d’équivalence x suivante : a et b sont équivalents s’il existe une isométrie de B (a, M ) ∪ B (x, 2M ) vers B (b, M ) ∪ B (x, 2M ) qui est l’identité sur B (x, 2M ) et ap- plique a sur b. Cette relation d’équivalence détermine d(x, a) et inversement il existe C = C (δ, K, N , Q, P , M ) telle que pour tout r ∈ N, l’ensemble {a ∈ X, d(x, a) = r} est réunion d’au plus C classes d’équivalences de 1,0,0,(0) . Il existe D ′ = C (δ, K ) telle que pour tout r ∈ N, le cardinal de Y x l’ensemble {a ∈ X, d(x, a) = r} soit inférieur ou égal à eD ′ r . On a donc, par Cauchy-Schwarz, 2 f (a)(cid:12)(cid:12)(cid:12) e(2s−αD ′ )r (cid:12)(cid:12)(cid:12) Xa∈X,d(x,a)=r 91 Hx,s (∆1 ) ≥ C −1 Xr∈N kf k2 )−1 (Z ) . . Grâce à (Hα) on suppose 2s − αD ′ ≥ s. Par Cauchy-Schwarz, 2(cid:17) e(2s−αD ′ )r (cid:12)(cid:12)(cid:12) Xa∈X,d(x,a)=r f (a)(cid:12)(cid:12)(cid:12) e−(2s−αD ′ )r (cid:17)(cid:16) Xr∈N f (a)2 ≤ (cid:16) Xr∈N Xa∈X C C 1 − e−(2s−αD ′ ) kf k2 1 − e−s kf k2 ≤ Hx,s (∆1 ) ≤ Hx,s (∆1 ) . Soit maintenant p ∈ {2, . . . , pmax}. Grâce au lemme 4.20, il suffit de mon- trer qu’il existe une constante C = C (δ, K, N , Q, P , M , s, B ) telle que k∂f kH→ x,s (∆p−1 ) ≤ C kf kHx,s (∆p ) pour tout f ∈ C(∆p ) . Grâce à (47) et au lemme 4.16, il est clair que kP (∂f )kH→ x,s (∆p−1 ) ≤ C kf kHx,s (∆p ) pour une constante C = C (δ, K, N , Q, P , M , s, B ). Il reste donc à montrer qu’il existe une constante C = C (δ, K, N , Q, P , M , s, B ) telle que (54) 2 B−(m+Pm i=0 li ) k(1 − P )(∂f )kH→ x,s (∆p−1 ) ≤ C kf kHx,s (∆p ) pour tout f ∈ C(∆p ) . Par (53) on a X x,s (∆p−1 ) = Xk ,m,l0 ,...,lm∈N k(1 − P )(∂f )k2 H→ p−1,k,m,(l0 ,...,lm ) Z ∈Y ,r0 (Z )>k+P x m si (Z )−li (cid:17)♯(cid:0)(π p−1,k ,m,(l0 ,...,lm) e2s(r0 (Z )−k)(cid:16) )−1(Z )(cid:1)−α (cid:12)(cid:12)ξZ (∂f )(cid:12)(cid:12) Yi=0 x p−1,k ,m,(l0 ,...,lm) Soient k , m, l0 , . . . , lm ∈ N et Z ∈ Y vérifiant r0(Z ) > k + P . x On pose l0 = 0, li = li−1 pour i ∈ {1, . . . , m + 1}. Alors on a X (∂f )(a1 , ..., ap−1) = XZ∈ΛZ ξZ (∂f ) = (a1 ,...,ap−1 ,S0 ,...,Sm,(Y j i )i∈{0,...,m},j∈{1,...,li } ) p−1,k,m,(l0 ,...,lm ) )−1 (Z ) ∈(π x X f (a1 , . . . , ap) = XZ ∈ΛZ (55) ξ Z (f ) (a1 ,...,ap , S0 ,..., Sm+1,( Y j i )i∈{0,...,m+1},j∈{1,...,li } ) p,k,m+1,(l0 ,...,lm+1 ) )−1 ( Z ) x ∈(π où ΛZ est la partie de Y p,k ,m+1,(l0 ,...,lm+1) formée des Z tels que pour tout x i )i∈{0,...,m+1},j∈{1,...,li }) ∈ (π p,k ,m+1,(l0 ,...,lm+1) (a1 , . . . , ap , S0 , ..., Sm+1 , ( Y j x 92 )−1( Z ) )−1 (Z ). (56) (57) ξ Z , ξ Z (f )2 . B ( Si , M ) ∪ B ( Y j i , M ) ∪ B (x, k + 2M ) on ait S1 = {a2 , . . . , ap} et (a2 , . . . , ap , S1 , ..., Sm+1 , ( Y j i+1)i∈{0,...,m},j∈{1,...,li }) ∈ (π p−1,k ,m,(l0 ,...,lm) x Il existe C1 = C (δ, K, N , Q, P , M ) telle que ♯ΛZ ≤ C1 . En effet, grâce au lemme 4.19, pour connaître les distances entre les points de B ( S0 , M ) et ceux de [i∈{1,...,m+1} [ i∈{1,...,m+1},j∈{1,...,li } il suffit de connaître les distances entre les points de B ( S0 , M ) et C points de (56), avec C = C (δ, K, N , Q, P , M ) et comme S0 = S1 ∪{a1} et d(a1 , a2 ) ≤ N , ces distances sont déterminées à N + M près par les distances de a2 à ces C points (qui font partie de la donnée de Z ). Comme ♯ΛZ ≤ C1 , grâce à (55) et par Cauchy-Schwarz, on obtient 2 ≤ C1 XZ∈ΛZ (cid:12)(cid:12)ξZ (∂f )(cid:12)(cid:12) De plus pour Z ∈ ΛZ on a m m+1 Yi=0 Yi=0 ♯(π p,k ,m+1,(l0 ,...,lm+1) )−1( Z ) ≤ C ♯(π p−1,k ,m,(l0 ,...,lm) )−1 (Z ) x x avec C = C (δ, K, N ) et r0 (Z ) − r0 ( Z ) ≤ N . Enfin Z détermine Z , donc quand on somme sur Z , chaque Z ne peut apparaître qu’une fois. L’inégalité (54) en résulte facilement et ceci termine la démonstration de la proposi- tion 4.21. (cid:3) Remarque. Le coeur de la démonstration ci-dessus est formé par – l’égalité (55), que l’on peut mettre sous la forme t∂ (ξZ ) = P Z∈ΛZ – l’application de Cauchy-Schwarz qui fournit l’inégalité (57), – le fait que les normes sont des sommes pondérées des ξZ 2 . Dans la suite les arguments seront plus compliqués mais ils reposeront tous sur ce principe. Avant de montrer la continuité de Jx on va introduire des nouvelles normes sur C(∆p ) (pour µ0 , µ1 ∈ N) et montrer qu’elles pré-hilbertiennes k.kH ♮,µ0 ,µ1 x,s sont équivalentes à k.kHx,s . Ces normes seront obtenues en a joutant aux par- ties Y j i qui intervenaient dans la définition 4.1 de nouvelles parties, notées Z j i dans la définition ci-dessous. On verra dans la démonstration de la conti- nuité de Jx (proposition 4.30) que la connaissance des points de B (Z j i , M ) 93 si ( Z )−li = si (Z )−li , détermine exactement les différentes moyennes intervenant dans la formule pour Jx , d’où l’intérêt de ces parties supplémentaires. On va voir dans la preuve de l’équivalence des normes k.kH et k.kHx,s (lemme 4.24) que ♮,µ0 ,µ1 x,s i peut être oubliée ou reconsidérée comme une partie Y j ′ chaque partie Z j i′ supplémentaire. Cependant i′ est déterminé par d(x, Z j i ) d’une façon assez compliquée. Ces nouvelles normes rendent donc la preuve de la continuité de Jx (proposition 4.30) beaucoup plus lisible et elles resserviront de plus pour montrer la continuité des autres opérateurs (proposition 4.46) et l’équiva- riance à compact près de tous les opérateurs (proposition 4.68). Inversement on n’a pas inclus ces parties Z j i dans la définition 4.1 car elles auraient rendu beaucoup plus difficile la preuve des propriétés d’équivariance de la norme k.kHx,s (proposition 4.3, démontrée dans le sous-paragraphe 4.7). Soient p ∈ {1, ..., pmax} et k , m, l0 , ..., lm , λ0 , λ1 ∈ N vérifiant λ1 = 0 si m = 0. Définition 4.22 On note Y ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x Pm i=0 li + λ0 + λ1 )-uplets (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } , (Z j i )i∈{0,1},j∈{1,...,λi }) tels que i )i∈{0,...,m},j∈{1,...,li }) appartient à Y p,k ,m,(l0 ,...,lm) – (a1 , . . . , ap , S0 , ..., Sm , (Y j x c’est-à-dire vérifie les conditions i), ii), iii), iv) de la définition 4.1, – pour i ∈ {0, 1} et j ∈ {1, . . . , λi}, Z j i est une partie non vide de X de diamètre inférieur ou égal à P /3 et Z j i ⊂ Sa∈Si géod(x, a). On introduit une partition de Y ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 pour la relation d’équi- x valence suivante : i )i∈{0,...,m},j∈{1,...,li } , (Z j (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,1},j∈{1,...,λi }) l’ensemble des (p + m + 1 + , (a1 , . . . , ap , S0 , ..., Sm , ( Y j i )i∈{0,...,m},j∈{1,...,li } , ( Z j i )i∈{0,1},j∈{1,...,λi }) sont en relation s’il existe une isométrie de [ [i∈{0,...,m} i∈{0,...,m},j∈{1,...,li } B (Z j i , M ) ∪ B (x, k + 2M ) ∪ B (Y j i , M ) B (Si , M ) ∪ [ i∈{0,1},j∈{1,...,λi } B ( Si , M ) ∪ [i∈{0,...,m} et vers [ i∈{0,...,m},j∈{1,...,li } B ( Y j i , M ) 94 ∪ B ( Z j i , M ) ∪ B (x, k + 2M ) [ i∈{0,1},j∈{1,...,λi } qui envoie ai sur ai pour i ∈ {1, . . . , p}, Si sur Si pour i ∈ {0, . . . , m}, Y j i sur Y j i sur Z j i pour i ∈ {0, . . . , m}, j ∈ {1, . . . , li}, Z j i pour i ∈ {0, 1}, j ∈ {1, . . . , λi} et est l’identité sur B (x, k + 2M ). ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 le quotient de Y ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 On note Y pour cette x x relation d’équivalence, et π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 l’application quotient. x ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 Notations. Pour Z ∈ Y on note r0(Z ), . . . , rm(Z ), s0(Z ), . . . , x i,j (Z ))i∈{0,1},j∈{i,...,m+1} et (tj sm (Z ), (rmax i (Z ))i∈{0,1},j∈{1,...,λi } les entiers tels que – ri (Z ) = d(x, Si ) pour i ∈ {0, . . . , m}, – si (Z ) = d(Si , Si+1) + 2M pour i ∈ {0, . . . , m − 1}, – sm (Z ) = d(x, Sm) − k , – pour i ∈ {0, 1}, si dmax (x, Si ) ≤ k + 3P , rmax i,j (Z ) = dmax (x, Si ) pour tout j ∈ {i, ..., m + 1}, – pour i ∈ {0, 1}, si dmax (x, Si ) ≥ k+3P , rmax i,j (Z ) = max(k+3P , dmax (x, Sj )) pour tout j ∈ {i, ..., m} et rmax i,m+1(Z ) = k + 3P , i (Z ) = d(x, Z j – tj i ) pour i ∈ {0, 1}, j ∈ {1, . . . , λi} pour tout (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } , (Z j i )i∈{0,1},j∈{1,...,λi } ) ∈ (π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 )−1(Z ). x Pour clarifier le sens des notations ci-dessus, on rappelle que d’après le lemme 4.12, on a toujours dmax (x, S0 ) ≥ dmax (x, S1) ≥ · · · ≥ dmax (x, Sm ). D’autre part, pour i ∈ {0, 1}, on a toujours rmax rmax et i,i (Z ) = dmax (x, Si ) i,m+1(Z ) = min(k + 3P , dmax(x, Si )). (58) (59) Le lemme suivant indique quelques propriétés de ces entiers, qui nous seront utiles ensuite. Lemme 4.23 Pour k , m, l0 , ..., lm , λ0 , λ1 ∈ N (vérifiant λ1 = 0 si m = 0) et ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 Z ∈ Y on a x 1,1 (Z ) ≥ rmax 0,1 (Z ) ≥ ... ≥ rmax a) rmax 0,0 (Z ) ≥ rmax 0,m+1(Z ) et rmax 1,2 (Z ) ≥ ... ≥ rmax 1,m+1 (Z ), b) pour tout i ∈ {0, 1}, ri (Z ) ≤ rmax i,i (Z ) ≤ ri (Z ) + N , c) pour i ∈ {0, 1} et j ∈ {1, ..., λi}, tj i (Z ) ≤ rmax i,i (Z ), d) pour i ∈ {0, 1}, rmax i,m+1 (Z ) = min(k + 3P , rmax i,i (Z )). Démonstration. L’assertion a) découle de (58), b) est évidente et pour montrer c) on remarque que dans les notations précédentes on a d(x, Z j i ) ≤ 95 dmax (x, Si ) par la dernière condition de la définition 4.22. Enfin d) résulte de (59). (cid:3) on note ξZ la forme linéaire sur C(∆p ) définie Pour Z ∈ Y ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x par ξZ (f ) = X (a1 ,...,ap ,S0 ,...,Sm ,(Y j i )i∈{0,...,m},j∈{1,...,li } ,(Z j i )i∈{0,1},j∈{1,...,λi } ) ♮,p,k,m,(l0 ,...,lm ),λ0 ,λ1 )−1 (Z ) ∈(π x f (a1 , ..., ap ). (60) (∆p ) B−(m+Pm i=0 li ) e2s(r0 (Z )−k) kf k2 ♮,µ0 ,µ1 H x,s Pour µ0 , µ1 ∈ N on munit alors C(∆p ) de la norme pré-hilbertienne, définie par la formule suivante : X = Xk ,m,l0 ,...,lm ,λ0 ,λ1 ♮,p,k,m,(l0 ,...,lm ),λ0 ,λ1 Z ∈Y x m si (Z )−li (cid:17)(r0 (Z ) + 1)−λ0 (r1 (Z ) + 1)−λ1 (cid:16) Yi=0 )−1 (Z )(cid:1)−α (cid:12)(cid:12)ξZ (f )(cid:12)(cid:12) 2 ♯(cid:0)(π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x où la première somme porte sur k , m, l0 , ..., lm , λ0 , λ1 ∈ N vérifiant λ1 = 0 si m = 0 et satisfaisant les conditions (61) λ1 ≤ µ1 . λ0 ≤ µ0 En utilisant l’hypothèse (HB ) nous allons montrer le lemme suivant. et Lemme 4.24 Il existe C = C (δ, K, N , Q, P , M , s, B ) tel que pour µ0 , µ1 ∈ N et f ∈ C(∆p ) , Hx,s (∆p ) ≤ kf k2 kf k2 (∆p ) ≤ C µ0+µ1 kf k2 Hx,s (∆p ) . H♮,µ0 ,µ1 x,s Démonstration. L’inégalité de gauche est évidente, car la somme (36) qui donne kf k2 Hx,s (∆p ) est une partie de la somme (61) qui donne kf k2 ♮,µ0 ,µ1 H (∆p ) x,s (c’est la partie qui correspond à λ0 = λ1 = 0). Pour montrer l’inégalité de droite on a besoin de deux lemmes préliminaires. L’idée est simplement de i comme une partie Y j ′ reconsidérer chaque partie Z j i′ supplémentaire, avec i′ déterminé par tj i (Z ) = d(x, Z j i ), si tj i (Z ) > k + 3P , et d’oublier Z j i si tj i (Z ) ≤ k + 3P . On prend k , m, l0 , ..., lm , λ0 , λ1 ∈ N (vérifiant λ1 = 0 si m = 0). Soit σ ∈ {0, 1} (avec σ = 0 si m = 0). On pose – (λ0 , λ1) = (λ0 + 1, λ1) si σ = 0 – et (λ0 , λ1) = (λ0 , λ1 + 1) si σ = 1. 96 ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 Sous-lemme 4.25 Soit Z ∈ Y x (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } , (Z j i )i∈{0,1},j∈{1,...,λi }) ∈ (π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 )−1(Z ). x et Alors – si 0 ≤ tλσ +1 (Z ) ≤ rmax σ,m+1 (Z ), on a σ (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } , (Z j i )i∈{0,1},j∈{1,...,λi } ) ∈ Y ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x σ,i (Z ), pour i ∈ {σ, ..., m}, en posant lj = lj – si rmax (Z ) ≤ rmax σ,i+1 (Z ) < tλσ +1 σ pour j ∈ {0, ..., m} \ {i}, li = li + 1 et Y li+1 = Z λσ +1 on a σ i (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } , (Z j i )i∈{0,1},j∈{1,...,λi } ) ∈ Y ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x Remarque. Par le c) et le d) du lemme 4.23, la condition tλσ +1 (Z ) ≤ σ rmax σ,m+1 (Z ) qui détermine le premier cas est équivalente à tλσ +1 (Z ) ≤ k + 3P . σ Démonstration. D’après le c) du lemme 4.23, on a tλσ +1 (Z ) ≤ rmax σ,σ (Z ) σ donc on se trouve toujours exactement dans l’un des cas ci-dessus. Il n’y a rien à montrer dans le premier cas. Supposons maintenant (Z ) ∈]rmax σ,m+1 (Z ), rmax σ,m (Z )]. Alors nécessairement rmax σ,m+1 (Z ) < rmax σ,m (Z ), ce qui implique tλσ +1 σ σ,m (Z ) = dmax (x, Sm) et rmax rmax σ,m+1 (Z ) = k + 3P . On a donc d(x, Z λσ +1 ) > k + 3P . Montrons Z λσ +1 ⊂ Sy∈Sm 2P -géod(x, y ). σ σ Soit z ∈ Z λσ +1 . Il existe a ∈ S0 tel que z ∈ 2F -géod(x, a). Cela est clair si σ σ = 0. Si σ = 1 il existe b ∈ S1 tel que z ∈ géod(x, b). Soit a ∈ S0 . Alors le a) du lemme 4.14 montre que b ∈ 2F -géod(x, a), d’où z ∈ 2F -géod(x, a). Soit y ∈ Sm tel que d(x, y ) = dmax (x, Sm). On a y ∈ 2F -géod(x, a) d’après le a) du lemme 4.14. D’autre part d(x, z) ≤ d(x, y ) + P /3 puisque tλσ +1 (Z ) ≤ σ dmax (x, Sm) et Z λσ +1 est de diamètre inférieur ou égal à P /3. En appliquant σ le lemme 4.8 à (x, a, y , z) au lieu de (x, y , a, b) et (2F , 2F , P /3) au lieu de (α, β , ρ) on trouve z ∈ (2F + 2P /3 + δ )-géod(x, y ) ⊂ 2P -géod(x, y ) 97 car on suppose 2F + 2P /3 + δ ≤ 2P , ce qui est permis par (HP ). On a donc montré Z λσ +1 ⊂ Sy∈Sm 2P -géod(x, y ) et Y lm+1 = Z λσ +1 vérifie la condition σ m σ iv) de la définition 4.1. Supposons maintenant (Z ) ∈]rmax σ,i+1 (Z ), rmax tλσ +1 i ∈ {σ, ..., m − 1}. σ,i (Z )] avec σ σ,i+1 (Z ) < rmax Alors nécessairement rmax σ,i (Z ), ce qui implique σ,i (Z ) = dmax (x, Si ) et rmax rmax σ,i+1 (Z ) = max(k + 3P , dmax(x, Si+1)). On a donc tλσ +1 (Z ) ∈]dmax (x, Si+1 ), dmax (x, Si )] Soit z ∈ Z λσ +1 . On a vu σ σ qu’il existe a ∈ S0 tel que z ∈ 2F -géod(x, a). Soit y ∈ Si tel que d(x, y ) = dmax (x, Si ) et soit y ′ ∈ Si+1 . D’après le a) du lemme 4.14, y et y ′ appartiennent à 2F -géod(x, a). Il est clair que d(x, y ′ ) ≤ d(x, z) ≤ d(x, y ) + P /3. y y ′ x a z Le lemme 4.8 appliqué à (x, a, z , y ′ ) au lieu de (x, y , a, b) et (2F , 2F , 0) au lieu de (α, β , ρ) montre y ′ ∈ (2F + δ )-géod(x, z). Le lemme 4.8 appliqué à (x, a, y , z) au lieu de (x, y , a, b) et (2F , 2F , P /3) au lieu de (α, β , ρ) montre (62) z ∈ (2P /3 + 2F + δ )-géod(x, y ). Par le b) du lemme 3.3 on déduit de (62) et (63) que (63) z ∈ (2P /3 + 4F + 2δ )-géod(y ′ , y ) ⊂ P -géod(y ′ , y ) car on suppose 2P /3 + 4F + 2δ ≤ P , ce qui est permis par (HP ). On a P -géod(y , y ′) et Y li+1 donc montré Z λσ +1 = Z λσ +1 ⊂ Sy∈Si ,y ′∈Si+1 vérifie la σ σ i condition iii) de la définition 4.1. (cid:3) Soit σ ∈ {0, 1} (avec σ = 0 si m = 0). On pose – (λ0 , λ1) = (λ0 + 1, λ1) si σ = 0 – et (λ0 , λ1) = (λ0 , λ1 + 1) si σ = 1. 98 → Y ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x Y ♮,p,k ,m,(l0 ,...,li−1 ,li+1,li+1 ,...,lm),λ0 ,λ1 x On note κσ : Y ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x ∪ [i∈{σ,...,m} l’application définie par le sous-lemme 4.25. On vérifie facilement que κσ passe au quotient et définit ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 , 0 ≤ tλσ +1 κσ,∞ : {Z ∈ Y (Z ) ≤ k + 3P } → Y σ x x ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 σ,i+1(Z ) < tλσ +1 , rmax (Z ) ≤ rmax et κσ,i : {Z ∈ Y σ,i (Z )} σ x ♮,p,k ,m,(l0 ,...,li−1 ,li+1,li+1 ,...,lm),λ0 ,λ1 → Y pour i ∈ {σ, ..., m}. x Sous-lemme 4.26 Il existe C = C (δ, K, N , Q, P ) tel que ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 a) pour tout i ∈ {σ, ..., m}, κσ,i est injective, et pour Z ∈ Y x (Z ) ≤ rmax tel que rmax σ,i+1 (Z ) < tλσ +1 σ,i (Z ), κσ induit une bijection de σ (π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x )−1 (Z ) dans )−1(κσ,i (Z )), (π ♮,p,k ,m,(l0 ,...,li−1 ,li+1,li+1 ,...,lm),λ0 ,λ1 x ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 , κ−1 b) pour tout élément Z∞ ∈ Y σ,∞ (Z∞) est de cardinal x inférieur ou égal à C (rmax σ,m+1 (Z∞ ) + 1) et pour tout entier t ≤ k + 3P le ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 tels que tλσ +1 nombre de Z ∈ Y (Z ) = t et κσ,∞ (Z ) = Z∞ σ x ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 est inférieur ou égal à C . De plus pour tout Z ∈ Y x (Z ) ≤ k + 3P , κσ induit une bijection de (π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 tλσ +1 x σ (π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 )−1(κσ,∞ (Z )). x tel que )−1 (Z ) dans Démonstration. La preuve de a) est immédiate. Montrons b). Soit Z ∈ ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 tel que tλσ +1 (Z ) ≤ k + 3P . Pour tout Y x σ (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } , (Z j i )i∈{0,1},j∈{1,...,λi }) dans (π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 )−1(Z ) on a Z λσ +1 ⊂ B (x, k + 3P + P /3) donc x σ Z λσ +1 ⊂ B (x, k + M ) car on suppose 3P + P /3 ≤ M , ce qui est per- σ mis par (HM ), et donc B (Z λσ +1 , M ) ⊂ B (x, k + 2M ). Par conséquent, si σ on note Z∞ = κσ,∞ (Z ), la donnée de Z est équivalente à celle de Z∞ et de Z λσ +1 ⊂ B (x, k + 2M ). La condition Z λσ +1 ⊂ Sa∈Sσ géod(x, a) s’ex- σ σ prime uniquement en termes des distances entre les points de Sσ et ceux 99 de B (x, k + 2M ) (qui font partie de la donnée de Z∞) et le nombre de possibilités pour Z λσ +1 ⊂ B (x, k + 2M ) de diamètre ≤ P /3 vérifiant cette σ ) ≤ k + 3P est borné par C (rmax condition et d(x, Z λσ +1 σ,m+1 (Z∞ ) + 1) avec σ C = C (δ, N , K, Q, P ) à cause du lemme 3.13. Pour tout t, l’ensemble des Z λσ +1 vérifiant les conditions précédentes et d(x, Z λσ+1 ) = t ne dépend que σ σ de Z∞ et de t et le lemme 3.13 montre que le cardinal de cet ensemble est inférieur ou égal à C = C (δ, N , K, Q, P ). (cid:3) Suite de la démonstration du lemme 4.24. Grâce au sous-lemme 4.26, il existe C = C (δ, K, N , Q, P , M ) tel que pour k , m, l0 , . . . , lm , λ0 , λ1 ∈ N, σ ∈ {0, 1} (vérifiant λ1 = 0 et σ = 0 si m = 0), et en posant – (λ0 , λ1) = (λ0 + 1, λ1) si σ = 0 – et (λ0 , λ1) = (λ0 , λ1 + 1) si σ = 1 on ait l’inégalité suivante : B−(m+Pm i=0 li ) m si (Z )−li (cid:17)(r0(Z ) + 1)−λ0 e2s(r0 (Z )−k)(cid:16) X Yi=0 ♮,p,k,m,(l0 ,...,lm ), λ0 , λ1 Z ∈Y x )−1(Z )(cid:1)−α (cid:12)(cid:12)ξZ (f )(cid:12)(cid:12) (r1(Z ) + 1)−λ1 ♯(cid:0)(π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 2 x ≤ C B−(m+Pm (cid:16) rmax σ,m+1 (Z ) + 1 σ,σ (Z ) + 1 (cid:17)e2s(r0 (Z )−k) X i=0 li ) rmax ♮,p,k,m,(l0 ,...,lm ),λ0 ,λ1 Z ∈Y x m (cid:16) si (Z )−li (cid:17)(r0 (Z ) + 1)−λ0 (r1 (Z ) + 1)−λ1 Yi=0 2! )−1(Z )(cid:1)−α (cid:12)(cid:12)ξZ (f )(cid:12)(cid:12) ♯(cid:0)(π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x Xi=σ B−(m+Pm m X j=0 lj +1) +CB ♮,p,k,m,(l0 ,...,li−1 ,li+1,li+1 ,...,lm ),λ0 ,λ1 Z ∈Y x σ,i (Z ) − rmax (cid:16) rmax σ,i+1 (Z ) (cid:17)e2s(r0 (Z )−k)(cid:16)si (Z )−(li+1) Yj∈{0,...,m}\{i} sj (Z )−lj (cid:17) rmax σ,σ (Z ) + 1 (r0(Z ) + 1)−λ0 (r1 (Z ) + 1)−λ1 2!. )−1(Z )(cid:1)−α (cid:12)(cid:12)ξZ (f )(cid:12)(cid:12) ♯(cid:0)(π ♮,p,k ,m,(l0 ,...,li−1 ,li+1,li+1 ,...,lm),λ0 ,λ1 x En effet le b) du lemme 4.23 assure que rmax σ,σ (Z ) + 1 ≤ (N + 1)(rσ (Z ) + 1) d’où (rσ (Z ) + 1)−1 ≤ (N + 1)(rmax σ,σ (Z ) + 1)−1 . De plus le ième terme dans la somme qui constitue la deuxième moitié du membre de droite n’apparaît pas (64) 100 si rmax σ,i (Z ) = rmax σ,i+1(Z ) et si rmax σ,i (Z ) > rmax σ,i+1(Z ) on a σ,i (Z ) − rmax si (Z ) ≤ C (rmax σ,i+1(Z )) avec C = C (δ, K, N , Q, P ). Sous-lemme 4.27 Soient C ∈ R∗ + , m ∈ N, ǫ0 , ..., ǫm , η1 , ..., ηm ∈ [0, 1] vé- rifiant Pm i=0 ǫi ≤ 1 et Pm i=1 ηi ≤ 1. Soient µ0 , µ1 ∈ N et (Al0 ,...,lm ,λ0 ,λ1 ) une famil le d’éléments de R+ indexée par les (l0 , ..., lm , λ0 , λ1) ∈ Nm+3 vérifiant λ0 ≤ µ0 et λ1 ≤ µ1 . On suppose – pour λ0 ∈ {0, ..., µ0 − 1} et λ1 ∈ {0, ..., µ1}, m Xi=0 ǫiAl0 ,...,li−1 ,li+1,li+1 ,...,lm ,λ0 ,λ1 (cid:1), Al0 ,...,lm,λ0+1,λ1 ≤ C (cid:0)Al0 ,...,lm,λ0 ,λ1 + – et pour λ0 ∈ {0, ..., µ0} et λ1 ∈ {0, ..., µ1 − 1}, m Xi=1 ηiAl0 ,...,li−1,li+1,li+1 ,...,lm,λ0 ,λ1 (cid:1). Al0 ,...,lm ,λ0 ,λ1+1 ≤ C (cid:0)Al0 ,...,lm ,λ0 ,λ1 + Alors X l0 ,...,lm∈N,λ0∈{0,...,µ0 },λ1∈{0,...,µ1} Al0 ,...,lm ,λ0 ,λ1 ≤ (2C + 1)µ0+µ1 Xl0 ,...,lm∈N Al0 ,...,lm ,0,0 (65) où l’on sous-entend que si le membre de droite converge, le membre de gauche converge aussi. Démonstration. Posons pour λ0 ∈ {0, ..., µ0} et λ1 ∈ {0, ..., µ1}, Aλ0 ,λ1 = Xl0 ,...,lm Al0 ,...,lm ,λ0 ,λ1 . Alors pour λ0 ∈ {0, ..., µ0 − 1} et λ1 ∈ {0, ..., µ1}, on a Aλ0+1,λ1 ≤ 2CAλ0 ,λ1 et de même pour λ0 ∈ {0, ..., µ0} et λ1 ∈ {0, ..., µ1 − 1}, Aλ0 ,λ1+1 ≤ 2CAλ0 ,λ1 . Le sous-lemme en résulte facilement car 1 + (2C ) + ...(2C )µc ≤ (2C + 1)µc pour c = 0, 1. (cid:3) 101 ǫj = pour j = 1, ..., m pour j = 0, ..., m, ηj = Fin de la démonstration du lemme 4.24. On applique le sous-lemme 4.27 de la façon suivante. On fixe m, k et des entiers R0,0 ≥ ... ≥ R0,m+1 ≥ 0 et R1,1 ≥ ... ≥ R1,m+1 ≥ 0. On applique le sous-lemme 4.27 en prenant R0,j − R0,j+1 R1,j − R1,j+1 R0,0 + 1 R1,1 + 1 et, pour l0 , ..., lm ∈ N, λ0 ∈ {0, ..., µ0} et λ1 ∈ {0, ..., µ1}, Al0 ,...,lm ,λ0 ,λ1 = 0 si m = 0 et λ1 > 0, et sinon Al0 ,...,lm ,λ0 ,λ1 = B−(m+Pm X i=0 li ) ♮,p,k,m,(l0 ,...,lm ),λ0 ,λ1 Z ∈Y tel que x rmax i,j (Z )=Ri,j pour i∈{0,1} et j∈{i,...,m+1} m si (Z )−li (cid:17)(r0(Z ) + 1)−λ0 (r1(Z ) + 1)−λ1 e2s(r0 (Z )−k)(cid:16) Yi=0 )−1(Z )(cid:1)−α (cid:12)(cid:12)ξZ (f )(cid:12)(cid:12) 2 ♯(cid:0)(π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 . x rmax σ,m+1(Z )+1 σ,σ (Z )+1 ≤ 1, les hypothèses du sous-lemme 4.27 Grâce à (64) et comme rmax sont satisfaites pour une constante C = C (δ, K, N , Q, P , M , s, B ). Puis on somme l’inégalité (65) sur m, k , (R0,0 , ..., R0,m+1), (R1,1 , ..., R1,m+1). Ceci termine la démonstration du lemme 4.24. Le lemme suivant est une variante du lemme 4.19. (cid:3) Lemme 4.28 Il existe une constante C = C (δ, K, N , Q, P , M ) tel le que pour p ∈ {1, ..., pmax}, k , m, l0 , ..., lm , λ0 , λ1 ∈ N (vérifiant λ1 = 0 si m = 0) et i )i∈{0,...,m},j∈{1,...,li } , (Z j (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,1},j∈{1,...,λi }) ∈ Y ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 , x les distances entre les points de B (S0 , M ) ∪ [j∈{1,...,l0 } B (Y j 0 , M ) ∪ et ceux de [i∈{1,...,m} sont déterminées par B (Y j i , M ) ∪ B (x, k + 2M ) [ i∈{1,...,m},j∈{1,...,li } [ i∈{0,1},j∈{1,...,λi } B (Si , M ) ∪ B (Z j i , M ) (66) (67) 102 – a) les distances entre les points de (67), – b) les entiers d(x, S0), d(x, Y j 0 ) et d(x, Z j 0 ), d(x, Z j 1 ), – c) les distances entre les points de (66) et C (1 + l0 + λ0 + λ1) points de (67) (qui sont eux-mêmes déterminés par a) et b)) et de plus les distances entre les points de (66) et ceux de (67) sont détermi- nées à C près par a) et b). Remarque. Dans toutes les situations où on appliquera ce lemme, l0+λ0+λ1 sera ma joré par une constante de la forme C (δ, K, N , Q, P , M ). Démonstration. Soient p ∈ {1, ..., pmax}, k , m, l0 , ..., lm , λ0 , λ1 ∈ N (vérifiant λ1 = 0 si m = 0) et (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } , (Z j i )i∈{0,1},j∈{1,...,λi }) ∈ Y ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 . x Soit b ∈ S0 et u un point de B (x, k) à distance minimale de b. On commence par montrer que pour σ ∈ {0, 1} et j ∈ {1, ..., λσ }, on a Z j σ ⊂ 2F -géod(x, b) (68) et (69) B (Z j σ , M ) ⊂ B (x, k + 2M ) ou Z j σ ⊂ (2F + δ )-géod(u, b). Soit j ∈ {1, ..., λ0}. On a 0 ⊂ [a∈S0 Z j géod(x, a) ⊂ 2N -géod(x, b) ⊂ 2F -géod(x, b) car F ≥ N . Si d(x, Z j 0 ) ≤ k + 3P , B (Z j 0 , M ) ⊂ B (x, k + 2M ) car on suppose 3P + P /3 ≤ M , ce qui est permis par (HM ). Si d(x, Z j 0 ) > k + 3P , pour z ∈ Z j 0 on a z ∈ 2N -géod(x, b) et d(x, z) > d(x, u) + 3P . Soit j ∈ {1, ..., λ1}. On a Z j 1 ⊂ Sa∈S1 géod(x, a). Par le lemme 4.14, S1 ⊂ 2F -géod(u, b), donc Z j 1 ⊂ 2F -géod(x, b). On a déjà prouvé (68). Si d(x, Z j 1 ) ≤ k + 3P , B (Z j 1 , M ) ⊂ B (x, k + 2M ) (70) 103 car 3P + P /3 ≤ M . Si d(x, Z j 1 ) > k + 3P , soit z ∈ Z j 1 . On a donc z ∈ 2F -géod(x, b) et d(x, z) > d(x, u) + 3P . (71) Soit maintenant z vérifiant (70) ou (71). Alors z vérifie (71) car F ≥ N par (22). z x u b Par (H 2F δ (u, x, z , b)) on a d(u, z) ≤ max(d(u, x) − d(x, z), d(u, b) − d(b, z)) + 2F + δ. Or d(u, x) − d(x, z) + 2F + δ ≤ −3P + 2F + δ et on suppose −3P + 2F + δ < 0, ce qui est permis par (HP ). Donc d(u, z) ≤ d(u, b) − d(b, z) + 2F + δ , c’est- à-dire z ∈ (2F + δ )-géod(u, b). Ceci termine la preuve de (69). On suppose 2F + δ ≤ 4P , ce qui est permis par (HP ). Pour montrer le lemme 4.28 on répète alors les arguments de la preuve du lemme 4.19. Plus précisément on applique le lemme 4.17 avec – (x, b) au lieu de (c, d), – {wi , i ∈ I } égal à [i∈{0,...,m} – J la partie de I telle que [i∈{0,1},j∈{1,...,λi } {wj , j ∈ J } = S0 ∪ [j∈{1,...,l0 } – et (αi , ρi ) égal à (4P , M ) pour tout i (les hypothèses sont satisfaites grâce au lemme 4.14 et à (68) et car on a supposé 2F ≤ 4P ). Puis on applique le lemme 4.18 à z parcourant [ B (S0 , M ) ∪ [j∈{1,...,l0 } B (Y j 0 , M ) ∪ i∈{0,1},j∈{1,...,λi } B (Z j i ,M )6⊂B (x,k+2M ) avec α = 4P + 2M et l = k + 2M (grâce au lemme 4.14 et à (69) et comme on a supposé 2F + δ ≤ 4P , cet ensemble est inclus dans (4P + 2M )-géod(b, u)). (cid:3) Le lemme suivant nous sera utile ensuite. [ i∈{0,...,m},j∈{1,...,li } B (Z j i , M ) Y j 0 ∪ Z j i , Si ∪ Y j i ∪ [ i∈{0,1},j∈{1,...,λi } Z j i , 104 Lemme 4.29 Il existe une constante C = C (δ, K, N , Q, P ) tel le que pour p,k ,m,(l0 ,...,lm) m, l0 , ..., lm , λ0 , λ1 ∈ N (avec λ1 = 0 si m = 0) et pour Z ∈ Y et x ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 Z ∈ Y tels que pour tout x (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } , (Z j i )i∈{0,1},j∈{1,...,λi }) )−1 ( Z ) ∈ (π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x on ait (a1 , . . . , ap , S0 , ..., Sm , (Y j )−1(Z ) i )i∈{0,...,m},j∈{1,...,li } ) ∈ (π p,k ,m,(l0 ,...,lm) x )−1( Z )(cid:1) ≤ C λ0+λ1 ♯(cid:0)(π p,k ,m,(l0 ,...,lm) alors ♯(cid:0)(π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 )−1(Z )(cid:1). x x Démonstration. Soit (a1 , . . . , ap , S0 , ..., Sm , (Y j )−1(Z ) i )i∈{0,...,m},j∈{1,...,li } ) ∈ (π p,k ,m,(l0 ,...,lm) x et soient b0 ∈ S0 , b1 ∈ S1 . Si (Z j i )i∈{0,1},j∈{1,...,λi } sont tels que (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } , (Z j i )i∈{0,1},j∈{1,...,λi }) )−1 ( Z ) ∈ (π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x on a Z j i ⊂ 2N -géod(x, bi ) pour i ∈ {0, 1}, j ∈ {1, . . . , λi}. De plus la donnée de Z détermine tj i ( Z ) = d(x, Z j i ) et le diamètre de Z j i doit être inférieur ou égal à P /3. On applique alors le lemme 3.13. (cid:3) Voici quelques rappels et notations pour la proposition suivante. On a ux,rKx Jx = Hx + uxKx = Hx + +∞ Xr=1 en notant ux,r = R 1 0 ux,r,tdt (de sorte que ux = P+∞ r=1 ux,r ). Pour q ∈ {1, ..., Q} on note Hx,q = hx (1 − ∂hx − hx∂ )q−1 , de sorte que Q Hx,q = Z(t1 ,...,tq )∈[0,1]q Hx,q ,(t1 ,...,tq )dt1 . . . dtq Xq=1 et On rappelle aussi que Kx = R(t1 ,...,tQ )∈[0,1]Q Kx,Q,(t1 ,...,tQ )dt1 . . . dtQ . Proposition 4.30 Pour tout p ∈ {1, . . . , pmax}, Jx se prolonge en un opé- rateur continu de Hx,s(∆p−1 ) dans Hx,s(∆p ). Plus précisément, pour p ∈ Hx = Hx,q . 105 {2, . . . , pmax} il existe C = C (δ, K, N , Q, P , M , s, B ) tel que pour tout q ∈ {1, ..., Q}, k Hx,q kL(Hx,s (∆p−1 ),Hx,s (∆p )) ≤ C et pour tout r ∈ N, kux,rKxkL(Hx,s (∆p−1 ),Hx,s (∆p )) ≤ C e− s 2 r . On remarque que Hx,1 = hx et grâce à la proposition 4.21 la continuité de hx implique celle de Hx,q pour tout q ∈ {1, ..., Q}. Cependant nous préférons montrer directement la continuité de Hx,q pour tout q ∈ {1, ..., Q} car cela prépare à la démonstration de la continuité de ux,rKx . Démonstration. Le cas où p = 1 est trivial. Soit p ∈ {2, . . . , pmax} et q ∈ {1, ..., Q}. Comme dans la démonstration de la proposition 4.21, on note P le pro jecteur orthogonal sur le sous-espace vectoriel de H→ x,s(∆p ) engendré par les eS pour S ∈ ∆p tel que d(x, S ) ≤ P , de sorte que (P f )(S ) = f (S ) si d(x, S ) ≤ P et (P f )(S ) = 0 sinon. La proposition 4.30 résulte donc des lemmes 4.31, 4.32, 4.36 et 4.40 que nous allons montrer successivement. (cid:3) Lemme 4.31 Il existe C = C (δ, K, N , Q, P , M , s, B ) tel que kP Hx,q kL(Hx,s (∆p−1 ),H→ x,s (∆p )) ≤ C. Démonstration. Il suffit de montrer que pour tout U ∈ ∆p vérifiant d(x, U ) ≤ P , il existe C = C (δ, K, N , Q, P , M , s, B ) tel que pour f ∈ C(∆p ) , ( Hx,q (f ))(U ) ≤ C kf kHx,s (∆p−1 ) . Soit U ∈ ∆p vérifiant d(x, U ) ≤ P . L’inégalité (72) est évidente, car, d’après le 1)a) de la proposition 3.37, pour tout S ∈ ∆p−1 , Hx,q (eS ) est supporté par les T ∈ ∆p tels que T ⊂ Sa∈S B (a, QN ), donc ( Hx,q (eS ))(U ) est nul sauf si d(x, S ) ≤ QN + P et le nombre de telles parties S est ma joré par une constante C = C (δ, K, N , Q, P ). De plus pour une telle partie S , ( Hx,q (eS ))(U ) est également ma joré par une telle constante par le 3) de la proposition 3.37. On conclut en utilisant le lemme 4.16. (cid:3) (72) Lemme 4.32 Il existe C = C (δ, K, N , Q, P , M , s, B ) tel que, pour tout r ∈ N, x,s (∆p )) ≤ C e− s 2 r . kP ux,rKxkL(Hx,s (∆p−1 ),H→ 106 Démonstration. Il suffit de montrer que pour tout U ∈ ∆p vérifiant d(x, U ) ≤ P , il existe C = C (δ, K, N , Q, P , M , s, B ) tel que, pour r ∈ N et f ∈ C(∆p ) , (ux,rKx (f ))(U ) ≤ C e− s 2 r kf kHx,s (∆p−1 ) . (73) Soit U ∈ ∆p vérifiant d(x, U ) ≤ P et r ∈ N. Nous allons montrer (73). Soit t, t1 , . . . , tQ ∈ [0, 1]. D’après le 2)a) de la proposition 3.37, pour S ∈ ∆p−1 , ux,r,tKx,Q,(t1 ,...,tQ ) (eS ) est une combinaison de eT pour T vérifiant Q d(x, T ) ∈ [d(x, S ) − r − QF , d(x, S ) − r + N + F − F On suppose Q F ≥ F + N , ce qui est permis par (HQ). Pour que (ux,r,tKx,Q,(t1 ,...,tQ )(eS ))(U ) ]. soit non nul il est donc nécessaire que (74) (75) )−1 (Z ), et d(x, S ) ∈ [r, r + QF + P ]. ♮,p−1,0,0,(0),Q,0 formée des Z tels que On note Λt1 ,...,tQ la partie de Y x – r0 (Z ) ∈ [r, r + QF + P ], 0 )j∈{1,...,Q}) ∈ (π ♮,p−1,0,0,(0),Q,0 – pour tout (a1 , . . . , ap−1 , S0 , (Z j x pour tout j ∈ {1, . . . , Q}, on a 0 = [b∈S0 Z j {z ∈ géod(x, b), d(x, z) = E (tj r0(Z ))} (on rappelle que r0 (Z ) = d(x, S0)). La condition (75) implique que pour Z ∈ Λt1 ,...,tQ et j ∈ {1, . . . , Q} on a tj 0 (Z ) = E (tj r0(Z )). Sous-lemme 4.33 Soit (a1 , . . . , ap−1 , S0) ∈ Y p−1,0,0,(0) tel que d(x, S0 ) ∈ x 0 par (75). Alors Z j [r, r + QF + P ]. Pour j ∈ {1, ..., Q} on définit Z j 0 est de diamètre inférieur ou égal à P /3 et il existe Z ∈ Λt1 ,...,tQ tel que (a1 , . . . , ap−1 , S0 , (Z j )−1(Z ). 0 )j∈{1,...,q}) ∈ (π ♮,p−1,0,0,(0),Q,0 x Démonstration. Soient j ∈ {1, ..., Q} et z , z ′ ∈ Z j 0 . Soit b ∈ S0 . On a z , z ′ ∈ 2N -géod(x, b) et d(x, z) = d(x, z ′ ) donc par (Hδ (z , x, z ′ , b)), d(z , z ′ ) ≤ 2N + δ et on suppose 2N + δ ≤ P /3, ce qui est permis par (HP ). Comme les parties Z j 0 sont non vides l’argument que nous venons de donner montre aussi que la condition (75) est vérifiée par les autres éléments de la classe d’équivalence Z de (a1 , . . . , ap−1 , S0 , (Z j 0 )j∈{1,...,q}) (car on suppose P /3 ≤ M , ce qui est permis par (HM )) et donc que Z ∈ Λt1 ,...,tQ . (cid:3) 107 Sous-lemme 4.34 Soit Z ∈ Λt1 ,...,tQ et (a1 , . . . , ap−1 , S0 , (Z j 0 )j∈{1,...,Q}) ∈ (π ♮,p−1,0,0,(0),Q,0 x Alors ux,r,tKx,Q,(t1 ,...,tQ ) (ea1 ∧ ... ∧ eap−1 ) ne dépend que de la connaissance des points de )−1 (Z ). B (S0 , M ) ∪ B (x, 2M ) ∪ [j∈{1,...,Q} et des distances entre ces points. B (Z j 0 , M ) (76) Démonstration. Le 2)b) de la proposition 3.37 montre que (78) (79) {y ∈ F -géod(x, a), d(y , a) ∈ [r, r + QF ]} (77) ux,r,tKx,Q,(t1 ,...,tQ ) (ea1 ∧ ... ∧ eap−1 ) dépend seulement de la connaissance des points de B (x, F ) ∪ B (S0 , QN ) ∪ [a∈S0 ∪ [a∈S0 {y ∈ F -géod(x, a), d(x, y ) − (1 − t)(d(x, a) − r) ≤ QF } [a∈S0 ,j∈{1,...,Q} ∪ {y ∈ F -géod(x, a), d(x, y ) − tj d(x, a) ≤ QF } et des distances entre ces points. Il suffit donc de montrer que cet ensemble est inclus dans (76). On a B (x, F ) ⊂ B (x, 2M ) et B (S0 , QN ) ⊂ B (S0 , M ) car on suppose F ≤ 2M et QN ≤ M , ce qui est permis par (HM ). Comme d(x, S0) ∈ [r, r + QF + P ] on a [a∈S0 {y ∈ F -géod(x, a), d(y , a) ∈ [r, r + QF ]} ⊂ B (x, d(x, S0 ) + N + F − r) ⊂ B (x, QF + P + N + F ) ⊂ B (x, 2M ) car on suppose QF + P + N + F ≤ 2M , ce qui est permis par (HM ). Donc (77) est inclus dans (76). Comme d(x, S0) ∈ [r, r + QF + P ] on a [a∈S0 {y ∈ F -géod(x, a), d(x, y ) − (1 − t)(d(x, a) − r) ≤ QF } ⊂ B (x, N + 2QF + P ) ⊂ B (x, 2M ) 108 car on suppose N + 2QF + P ≤ 2M , ce qui est permis par (HM ). Donc (78) est inclus dans (76). Enfin, soit j ∈ {1, ..., Q}, a ∈ S0 , et y ∈ F -géod(x, a) vérifiant d(x, y ) − tj d(x, a) ≤ QF . Soit z ∈ géod(x, a) vérifiant d(x, z) = E (tj d(x, S0 )), si bien que z appartient à Z j 0 . Comme y , z ∈ F -géod(x, a) et d(x, y ) − d(x, z) ≤ QF + N + 1, (Hδ (y , x, z , a)) montre que d(y , z) ≤ (QF + N + 1) + F + δ . On suppose (QF + N + 1) + F + δ ≤ M , ce qui est permis par (HM ). Donc (79) est inclus dans Sj∈{1,...,Q} B (Z j 0 , M ) et a fortiori dans (76). (cid:3) Sous-lemme 4.35 Le cardinal de Λt1 ,...,tQ est majoré par une constante de la forme C (δ, K, N , Q, P , M ). Démonstration. Cela résulte du lemme 4.28 (ou même d’un argument plus simple car le cardinal de l’ensemble (76) est borné par C = C (δ, K, N , Q, P , M ) et les distances entre les points de (76) sont déterminées à C ′ = C (δ, K, N , Q, P , M ) près par la donnée de r, t1 , ..., tQ). (cid:3) Fin de la démonstration du lemme 4.32. On écrit U = {b1 , ..., bp} pour lever l’ambiguïté de signe. On rappelle, pour Z ∈ Λt1 ,...,tQ , la notation X ξZ (f ) = f (a1 , ..., ap−1). 0 )j∈{1,...,Q} )∈(π♮,p−1,0,0,(0),Q,0 (a1 ,...,ap−1 ,S0 ,(Z j )−1 (Z ) x On a f = 1 (p−1)! P(a1 ,...,ap−1) f (a1 , ..., ap−1)ea1 ∧ ... ∧ eap−1 où la somme porte que les (a1 , ..., ap−1) tel que {a1 , ..., ap−1} ∈ ∆p−1 . Le sous-lemme 4.33 montre donc que (ux,rKx,Q,(t1 ,...,tQ ) (f ))(b1 , ..., bp ) = 1 (p − 1)! XZ ∈Λt1 ,...,tQ αZ,(t1 ,...,tQ ),(b1 ,...,bp) ξZ (f ) (80) où αZ,(t1 ,...,tQ),(b1 ,...,bp) ∈ C est défini de la façon suivante : pour tout (a1 , . . . , ap−1 , S0 , (Z j 0 )j∈{1,...,Q}) ∈ (π ♮,p−1,0,0,(0),Q,0 )−1 (Z ), x αZ,(t1 ,...,tQ ),(b1 ,...,bp) = (cid:0)ux,rKx,Q,(t1 ,...,tQ ) (ea1 ∧ ... ∧ eap−1 )(cid:1)(b1 , ..., bp ) (d’après le sous-lemme 4.34 ce nombre ne dépend que de Z ). D’après le 3) de la proposition 3.37, αZ,(t1 ,...,tQ ),(b1 ,...,bp ) est ma joré par une constante de la forme C (δ, K, N , Q). Par Cauchy-Schwarz et grâce au sous-lemme 4.35, on a donc (ux,rKx,Q,(t1 ,...,tQ ) (f ))(U )2 ≤ C XZ ∈Λt1 ,...,tQ 109 ξZ (f )2 formée pour une certaine constante C = C (δ, K, N , Q, P , M ). On en déduit, pour t1 , . . . , tQ ∈ [0, 1], (ux,rKx,Q,(t1 ,...,tQ ) (f ))(U )2 )−1(Z )(cid:1)−α ξZ (f )2 ≤ C e−sr XZ ∈Λt1 ,...,tQ e2sr0 (Z ) ♯(cid:0)(π ♮,p−1,0,0,(0),Q,0 x pour une certaine constante C = C (δ, K, N , Q, P , M ). En effet il existe une constante D = C (δ, K, N , Q, P , M ) telle que pour tout Z ∈ Λt1 ,...,tQ , )−1 (Z )(cid:1) ≤ eD(r0 (Z )+1) ♯(cid:0)(π ♮,p−1,0,0,(0),Q,0 x et on suppose αD ≤ s, ce qui est permis par (Hα). De plus pour tout Z ∈ Λt1 ,...,tQ on a r ≤ r0(Z ), d’où e−sr ≥ e−sr0 (Z ) . ♮,p−1,0,0,(0),Q,0 On a vu que Λt1 ,...,tQ est inclus dans la partie de Y x des Z tels que – r0 (Z ) ∈ [r, r + QF + P ], – pour tout j ∈ {1, . . . , Q} on a tj 0 (Z ) = E (tj r0 (Z )). Pour r0 ∈ N, quand (t1 , . . . , tQ ) parcourt [0, 1]Q muni de la mesure de Le- besgue, (E (tj r0 ))j=1,...,Q parcourt (cid:8)0, . . . , max(0, r0 − 1)(cid:9)Q avec la probabilité uniforme max(1, r0)−Q et comme 2 max(1, r0)−1 ≤ , r0 + 1 par Cauchy-Schwarz on obtient l’inégalité (ux,rKx (f ))(U )2 ≤ C e−sr (cid:16) XZ ∈Y e2sr0 (Z ) (cid:0)r0(Z ) + 1(cid:1)−Q ♮,p−1,0,0,(0),Q,0 x )−1(Z )(cid:1)−α ξZ (f )2(cid:17) ♯(cid:0)(π ♮,p−1,0,0,(0),Q,0 x pour une certaine constante C = C (δ, K, N , Q, P , M ). A fortiori on a (ux,rKx (f ))(U )2 ≤ C e−sr kf k2 H♮,Q,0 x,s puisque l’expression entre parenthèses dans (81) est la partie de la somme (61) ♮,p−1,0,0,(0),Q,0 donnant kf k2 qui correspond à Y . Grâce au lemme 4.24 x H♮,Q,0 (∆p−1 ) x,s ceci termine la démonstration du lemme 4.32. (∆p−1 ) (81) (cid:3) Lemme 4.36 Il existe C = C (δ, K, N , Q, P , M , s, B ) tel que k(1 − P ) Hx,q kL(Hx,s (∆p−1 ),H→ x,s (∆p )) ≤ C. 110 Démonstration. Grâce au lemme 4.24, il suffit de montrer l’inégalité sui- vante : il existe C = C (δ, K, N , Q, P , M , B ) tel que pour tout f ∈ C(∆p−1 ) , k(1 − P )( Hx,q f )k2 x,s (∆p ) ≤ C kf k2 (82) . H→ H♮,q,0 x,s (∆p−1 ) On rappelle que , . B−(m+Pm k(1 − P )( Hx,q f )k2 i=0 li ) XZ ∈Y x,s (∆p ) = Xk ,m,l0 ,...,lm∈N H→ p,k,m,(l0 ,...,lm ) x r0 (Z )>k+P m e2s(r0 (Z )−k)(cid:16) Yi=0 (cid:0)si (Z )(cid:1)−li (cid:17)♯(cid:0)(π p,k ,m,(l0 ,...,lm) )−1(Z )(cid:1)−α (cid:12)(cid:12)ξZ ( Hx,q f )(cid:12)(cid:12) 2 x On va voir que l’inégalité (82) résulte de l’inégalité (83) ci-dessous. p,k ,m,(l0 ,...,lm) Soient k , m, l0 , . . . , lm ∈ N et Z ∈ Y vérifiant r0(Z ) > k + P . x On pose l0 = 0 et li = li−1 pour i ∈ {1, . . . , m + 1}. On va montrer qu’il existe C = C (δ, K, N , Q, P , M ) tel que 2 ≤ C XZ ∈ΛZ (cid:0)r0 ( Z ) + 1(cid:1)−q (cid:12)(cid:12)(cid:12) ξZ ( Hx,q f )(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) 2 ξ Z (f )(cid:12)(cid:12) ♮,p−1,k ,m+1,(l0 ,...,lm+1),q ,0 formée des Z vérifiant où ΛZ est la partie de Y x r0( Z ) − r1( Z ) ≤ (q + 1)N (84) (83) et tels que pour tout i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 )j∈{1,...,q}) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),q ,0 )−1( Z ) x il existe une énumération (a1 , . . . , ap) de S1 vérifiant (a1 , . . . , ap , S1 , ..., Sm+1 , ( Y j i+1)i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x )−1 (Z ). (85) On rappelle que ξ Z (f ) = X i )i∈{0,...,m+1},j∈{1,...,li } ,( Z j (a1 ,...,ap−1 , S0 ,..., Sm+1 ,( Y j 0 )j∈{1,...,q} ) ♮,p−1,k,m+1,(l0 ,...,lm+1 ),q,0 )−1 ( Z ) x ∈(π f (a1 , . . . , ap−1). 111 On justifie maintenant le fait que (83) implique (82). D’abord Z déter- mine Z à permutation près de a1 , ..., ap donc connaissant Z il y a au plus p! possibilités pour Z . Dans les notations précédentes, soit b ∈ S0 . D’après (84), pour tout y ∈ S1 on a d(x, y ) − d(x, b) ≤ (Q + 2)N et on a y ∈ 2F -géod(x, b) par le a) du lemme 4.14, donc d(y , b) ≤ (Q + 2)N + 2F . Connaissant S1 on a donc au plus C = C (δ, K, N , Q) possibilités pour S0 . En utilisant de plus le lemme 4.29 on en déduit que pour Z ∈ ΛZ on a ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),q ,0 )−1( Z )(cid:1) ≤ C ♯(cid:0)(π p,k ,m,(l0 ,...,lm) )−1(Z )(cid:1) x x i=0 si ( Z )−li = avec C = C (δ, K, N , Q, P , M ). Il est clair que pour Z ∈ ΛZ on a Qm+1 Qm i=0 si (Z )−li . Donc (83) implique (82). L’inégalité (83) résulte de l’inégalité plus précise (86) ci-dessous. Soient t1 , . . . , tq ∈ [0, 1]. On va montrer qu’il existe C = C (δ, K, N , Q, P , M ) tel que 2 2 ξZ ( Hx,q ,(t1 ,...,tq )f )(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12) ξ Z (f )(cid:12)(cid:12)(cid:12) ≤ C XZ ∈ΛZ,(t1 ,...,tq ) (cid:12)(cid:12)(cid:12) où ΛZ,(t1 ,...,tq ) est l’ensemble des Z ∈ ΛZ tels que pour tout i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 )j∈{1,...,q}) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),q ,0 )−1( Z ) x et pour tout j ∈ {1, ..., q} on ait {z ∈ géod(x, b), d(x, z) = E (tj r0 ( Z ))}. Z j 0 = [b∈ S0 La condition (87) implique que pour Z ∈ ΛZ,(t1 ,...,tq ) et j ∈ {1, ..., q} on a 0 ( Z ) = E (tj r0( Z )). tj (87) (86) Sous-lemme 4.37 Soit i )i∈{0,...,m+1},j∈{1,...,li } ) ∈ Y p−1,k ,m+1,(l0 ,...,lm+1) (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j x tel que d(x, S0) − d(x, S1) ≤ (q + 1)N et qu’il existe (a1 , . . . , ap ) vérifiant (85). Pour j ∈ {1, ..., q} on définit Z j 0 par (87). Alors Z j 0 est de diamètre inférieur ou égal à P /3 et il existe Z ∈ ΛZ,(t1 ,...,tq ) tel que (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j 0 )j∈{1,...,q}) appartienne à (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),q ,0 )−1( Z ). x (88) 112 0 . Soit b ∈ S0 . On a Démonstration. Soient j ∈ {1, ..., q} et z , z ′ ∈ Z j z , z ′ ∈ 2N -géod(x, b) et d(x, z) = d(x, z ′ ) donc par (Hδ (z , x, z ′ , b)), d(z , z ′ ) ≤ 2N + δ ≤ P /3. Comme les parties Z j 0 sont non vides et P /3 ≤ M , l’argument que nous venons de donner montre aussi que la condition (87) est vérifiée par les autres éléments de la classe d’équivalence Z de l’élément (88) et donc Z ∈ ΛZ,(t1 ,...,tq ) . (cid:3) Sous-lemme 4.38 Soit Z ∈ ΛZ,(t1 ,...,tq ) , et i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 )j∈{1,...,q}) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),q ,0 )−1 ( Z ). x Alors Hx,q ,(t1 ,...,tq )(ea1 ∧ ... ∧ eap−1 ) ne dépend que de la connaissance des points de B ( S0 , M ) ∪ B (x, k + 2M ) ∪ [j∈{1,...,q} et des distances entre ces points. B ( Z j 0 , M ) (89) Démonstration. D’après le 1)b) de la proposition 3.37, Hx,q ,(t1 ,...,tq ) (ea1 ∧ ... ∧ eap−1 ) ne dépend que de la connaissance des points de B (x, 7δ ) ∪ B ( S0 , QN ) ∪ [a∈ S0 ,j∈{1,...,q} {y ∈ F -géod(x, a), d(x, y ) − tj d(x, a) ≤ QF } et des distances entre ces points. Il suffit donc de montrer que cet ensemble est inclus dans (89). D’abord on suppose 7δ ≤ 2M et QN ≤ M , ce qui est permis par (HM ), et (90) est inclus dans (89). Soit a ∈ S0 , j ∈ {1, ..., q}, et y ∈ F -géod(x, a) vérifiant d(x, y ) − tj d(x, a) ≤ QF . Soit z ∈ géod(x, a) vérifiant d(x, z) = E (tj r0( Z )), si bien que z appartient à Z j 0 . On a tj d(x, a) − E (tj r0 ( Z )) ≤ N + 1 puisque d(x, a) − r0 ( Z ) ≤ N , d’où d(x, y ) − d(x, z) ≤ QF + N + 1, et comme y et z appartiennent à F -géod(x, a), (Hδ (y , x, z , a)) montre que (91) (90) d(y , z) ≤ (QF + N + 1) + F + δ. On suppose (QF + N + 1) + F + δ ≤ M , ce qui est permis par (HM ). Donc (91) est inclus dans (89). (cid:3) 113 Sous-lemme 4.39 Le cardinal de ΛZ,(t1 ,...,tq ) est majoré par une constante de la forme C (δ, K, N , Q, P , M ). Démonstration. Les parties Z j 0 sont déterminées de manière unique par (87) et grâce au lemme 4.28, pour connaître les distances entre les points de B ( S0 , M ) ∪ [j∈{1,...,q} B ( Z j 0 , M ) (92) (93) B ( Si , M ) ∪ B( Y j i , M ) ∪ B (x, k + 2M ) et ceux de [ [i∈{1,...,m+1} i∈{1,...,m+1},j∈{1,...,li } il suffit de connaître les distances entre les points de (92) et C points de (93), avec C = C (δ, K, N , Q, P , M ) et grâce à (84) ces distances sont déterminées à C ′ = C (δ, K, N , Q, P , M ) près par les distances de S1 à ces C points (qui font 0 ( Z ))j∈{1,...,q} , qui grâce à (84) sont partie de la donnée de Z ) et les entiers (tj eux-mêmes déterminés à C ′′ = C (δ, K, N , Q, P , M ) près par r0(Z ), t1 , ..., tq . (cid:3) Fin de la démonstration du lemme 4.36. On termine la démonstration de (86). Pour Z ∈ ΛZ,(t1 ,...,tq ) et i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 )j∈{1,...,q}) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),q ,0 )−1( Z ) x on considère ( Hx,q ,(t1 ,...,tq ) (ea1 ∧ ... ∧ eap−1 ))(b1 , ..., bp ), X(b1 ,...,bp) où la somme porte sur les énumérations (b1 , ..., bp ) de S1 telles que (b1 , . . . , bp , S1 , ..., Sm+1 , ( Y j i+1)i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x Comme la somme (94) a au plus p! termes, le 3) de la proposition 3.37 montre qu’elle est ma jorée par une constante de la forme C (δ, K, N , Q, P , M ). D’après le sous-lemme 4.38 la somme (94) ne dépend que de Z et on peut donc la noter αZ, Z ,(t1 ,...,tq ) . D’après le sous-lemme 4.37 on a 1 (p − 1)! XZ∈ΛZ,(t1 ,...,tq ) 114 ξZ ( Hx,q ,(t1 ,...,tq )f ) = αZ, Z ,(t1 ,...,tq )ξ Z (f ). )−1(Z ). (94) (95) Par Cauchy-Schwarz et grâce au sous-lemme 4.39 on en déduit que ξZ ( Hx,q ,(t1 ,...,tq )f )2 ≤ C XZ ∈ΛZ,(t1 ,...,tq ) avec C = C (δ, K, N , Q, P , M ). On a montré (86), donc (83) et (82). Ceci termine la preuve du lemme 4.36. (cid:3) ξ Z (f )2 Lemme 4.40 Il existe C = C (δ, K, N , Q, P , M , s, B ) tel que, pour tout r ∈ N, x,s (∆p )) ≤ C e− sr k(1 − P )ux,rKxkL(Hx,s (∆p−1 ),H→ 2 . Démonstration. Grâce au lemme 4.24, il suffit de montrer l’inégalité (96) ci-dessous. Soit r ∈ N. On va montrer qu’il existe C = C (δ, K, N , Q, P , M , B ) tel que pour f ∈ C(∆p−1 ) , x,s (∆p ) ≤ C e−sr kf k2 k(1 − P )(ux,rKxf )k2 H→ H♮,Q,1 x,s (96) (∆p−1 ) . On a , . B−(m+Pm i=0 li ) XZ ∈Y x,s (∆p ) = Xk ,m,l0 ,...,lm∈N k(1 − P )(ux,rKxf )k2 H→ p,k,m,(l0 ,...,lm ) x r0 (Z )>k+P m e2s(r0 (Z )−k)(cid:16) si (Z )−li (cid:17)♯(cid:0)(π p,k ,m,(l0 ,...,lm) )−1(Z )(cid:1)−α (cid:12)(cid:12)ξZ (ux,rKxf )(cid:12)(cid:12) 2 Yi=0 x On va voir que l’inégalité (96) résulte de l’inégalité (97) ci-dessous. p,k ,m,(l0 ,...,lm) Soient k , m, l0 , . . . , lm ∈ N et Z ∈ Y vérifiant r0(Z ) > k + P . x On pose l0 = 0 et li = li−1 pour i ∈ {1, . . . , m + 1}. On va montrer qu’il existe C = C (δ, K, N , Q, P , M ) tel que )−1 (Z )(cid:1)−α (cid:12)(cid:12)ξZ (ux,rKxf )(cid:12)(cid:12) 2 e2s(r0 (Z )−k)♯(cid:0)(π p,k ,m,(l0 ,...,lm) x ≤ C e−sr XZ ∈ΛZ (cid:0)r0 ( Z ) + 1(cid:1)−Q(cid:0)r1 ( Z ) + 1(cid:1)−1e2s(r0 ( Z )−k) )−1( Z )(cid:1)−α (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) 2 ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 x ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 formée des Z vérifiant où ΛZ est la partie de Y x r0( Z ) − r1 ( Z ) − r ≤ QF , (97) (98) 115 et tels que pour tout i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j 0 )j∈{1,...,Q} , Z 1 (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 1 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z ) x il existe une énumération (a1 , . . . , ap) de S1 vérifiant (a1 , . . . , ap , S1 , ..., Sm+1 , ( Y j i+1)i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x )−1 (Z ). (99) i=0 si ( Z )−li . De plus Z détermine Z à Pour Z ∈ ΛZ on a Qm i=0 si (Z )−li = Qm+1 permutation près de a1 , ..., ap donc connaissant Z il y a au plus p! possibilités pour Z tels que Z ∈ ΛZ . Donc en sommant sur Z on voit que (97) implique (96). L’inégalité (97) résulte de l’inégalité plus précise (100) ci-dessous (en re- prenant les arguments de la fin de la démonstration du lemme 4.32). Soient t, t1 , . . . , tQ ∈ [0, 1]. On va montrer qu’il existe C = C (δ, K, N , Q, P , M ) tel que )−1(Z )(cid:1)−α (cid:12)(cid:12)ξZ (ux,r,tKx,Q,(t1 ,...,tQ )f )(cid:12)(cid:12) 2 ≤ C e−sr e2s(r0 (Z )−k) ♯(cid:0)(π p,k ,m,(l0 ,...,lm) x )−1( Z )(cid:1)−α (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) e2s(r0 ( Z )−k) ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 2 XZ∈ΛZ,t,(t1 ,...,tQ ) (100) x où ΛZ,t,(t1 ,...,tQ ) est l’ensemble des Z ∈ ΛZ tels que pour tout 0 )j∈{1,...,Q} , Z 1 i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 1 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z ) x on ait {z ∈ géod(x, b), d(x, z) = E (tj r0 ( Z ))} pour j ∈ {1, ..., Q} (101) Z j 0 = [b∈ S0 Z 1 {z ∈ géod(x, b), d(x, z) = E ((1 − t)r1 ( Z ))}. 1 = [b∈ S1 et Pour Z ∈ ΛZ,t,(t1 ,...,tQ) , les conditions (101) et (102) impliquent 0 ( Z ) = E (tj r0( Z )) pour j ∈ {1, ..., Q} et 1( Z ) = E ((1 − t)r1( Z )). tj t1 (102) 116 Sous-lemme 4.41 Soit i )i∈{0,...,m+1},j∈{1,...,li } ) ∈ Y p−1,k ,m+1,(l0 ,...,lm+1) (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j x tel que d(x, S0) − d(x, S1) − r ≤ QF et qu’il existe (a1 , . . . , ap ) vérifiant (99). On définit ( Z j 0 )j∈{1,...,Q} et Z 1 1 par (101) et (102). Alors les parties Z j 0 et Z 1 1 sont de diamètre inférieur ou égal à P /3 et il existe Z ∈ ΛZ,t,(t1 ,...,tQ) tel que 0 )j∈{1,...,Q} , Z 1 i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j (103) 1 ) appartienne à (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z ). x Démonstration. Pour σ ∈ {0, 1} et z , z ′ ∈ Z j σ , on choisit b ∈ Sσ , d’où z , z ′ ∈ 2N -géod(x, b) et comme d(x, z) = d(x, z ′ ), (Hδ (z , x, z ′ , b)) donne d(z , z ′ ) ≤ 0 et Z 1 2N + δ ≤ P /3. Comme les parties Z j 1 sont non vides et P /3 ≤ M , l’argument que nous venons de donner montre aussi que les conditions (101) et (102) sont vérifiées par les autres éléments de la classe d’équivalence Z de l’élément (103) et donc Z ∈ ΛZ,t,(t1 ,...,tQ ) . (cid:3) Le sous-lemme suivant explique d’où vient la condition (98). Sous-lemme 4.42 Pour S ∈ ∆p−1 et T ∈ ∆p tels que eT apparaisse avec un coefficient non nul dans ux,r,tKx,Q,(t1 ,...,tQ ) (eS ), on a Q F , r + QF ]} {y ∈ F -géod(x, a), d(y , a) ∈ [r + d(x, S ) − d(x, T ) − r ≤ QF Démonstration. D’après le 2)a) de la proposition 3.37, on a T ⊂ [a∈S d’où l’énoncé du sous-lemme car Q F ≥ N + F . Sous-lemme 4.43 Soit Z ∈ ΛZ,t,(t1 ,...,tQ ) , et (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j 0 )j∈{1,...,Q} , Z 1 1 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z ). x Alors ux,r,tKx,Q,(t1 ,...,tQ ) (e S0 ) ne dépend que de la connaissance des points de 0 , M ) ∪ B ( Z 1 B (x, k + 2M ) ∪ B ( S0 , M ) ∪ B ( S1 , M ) ∪ [j∈{1,...,Q} B ( Z j 1 , M ) (cid:3) et des distances entre ces points. 117 (104) ∪ (105) (106) (108) (107) Démonstration. D’après le 2)b) de la proposition 3.37, ux,r,tKx,Q,(t1 ,...,tQ )(e S0 ) ne dépend que de la connaissance des points de B (x, F ) ∪ B ( S0 , QN ) [a∈ S0 ,j∈{1,...,Q} {y ∈ F -géod(x, a), d(x, y ) − tj d(x, a) ≤ QF } ∪ [a∈ S0 {z ∈ F -géod(x, a), d(z , a) ∈ [r, r + QF ]} ∪ [a∈ S0 (cid:8)z ∈ F -géod(x, a), d(x, z) − (1 − t)(d(x, a) − r) ≤ QF (cid:9) et des distances entre ces points. Il suffit donc de montrer que cet ensemble est inclus dans (104). On suppose F ≤ 2M et QN ≤ M , ce qui est permis par (HM ). Donc (105) est inclus dans (104). Soient a ∈ S0 , j ∈ {1, ..., Q}, y ∈ F -géod(x, a) vérifiant d(x, y ) − tj d(x, a) ≤ QF . Soit z ∈ géod(x, a) vérifiant d(x, z) = E (tj r0 ( Z )), si bien que z appartient à Z j 0 . On a tj d(x, a) − E (tj r0 ( Z )) ≤ N + 1, d’où d(x, y ) − d(x, z) ≤ QF + N + 1 et grâce à (Hδ (y , x, z , a)), d(y , z) ≤ (QF + N + 1) + F + δ . On suppose (QF + N + 1) + F + δ ≤ M , ce qui est permis par (HM ). Donc (106) est inclus dans Sj∈{1,...,Q} B ( Z j 0 , M ) et a fortiori dans (104). Soit a ∈ S0 et z ∈ F -géod(x, a) vérifiant d(z , a) ∈ [r, r + QF ]. Soit y ∈ S1 . Par le a) du lemme 4.14, on a y ∈ 2F -géod(a, x). La condition (98) implique d(x, a) − d(x, y ) − r ≤ QF + N . Comme d(x, a) − d(x, z) ∈ [r − F , r + QF ], on en déduit d(x, y ) − d(x, z) ≤ 2QF + N . Comme z ∈ F -géod(x, a) et y ∈ 2F -géod(x, a), (Hδ (z , x, y , a)) montre d(y , z) ≤ (2QF + N ) + 2F + δ. On suppose (2QF + N ) + 2F + δ ≤ M , ce qui est permis par (HM ). Donc (107) est inclus dans B ( S1 , M ) et a fortiori dans (104). 118 Enfin soit a ∈ S0 et z ∈ F -géod(x, a) vérifiant d(x, z) − (1 − t)(d(x, a) − r) ≤ QF . Soit b ∈ S1 et y ∈ géod(x, b) vérifiant d(x, y ) = E ((1 − t)r1 ( Z )), si bien que y appartient à Z 1 1 . Comme y ∈ géod(x, b) et b ∈ 2F -géod(x, a), on a y ∈ 2F -géod(x, a). Comme d(x, a) ∈ [r0 ( Z ), r0( Z ) + N ] et grâce à (98), on a d(x, z) − (1 − t)r1 ( Z ) ≤ 2QF + N , d’où d(x, y ) − d(x, z) ≤ 2QF + 1 + N . Comme z ∈ F -géod(x, a) et y ∈ 2F -géod(x, a), (Hδ (y , x, z , a)) montre d(y , z) ≤ (2QF + 1 + N ) + 2F + δ. On suppose (2QF + 1 + N ) + 2F + δ ≤ M , ce qui est permis par (HM ). Donc (108) est inclus dans B ( Z 1 1 , M ) et a fortiori dans (104). (cid:3) Sous-lemme 4.44 Le cardinal de ΛZ,t,(t1 ,...,tQ ) est majoré par une constante de la forme C (δ, K, N , Q, P , M ). Démonstration. Grâce au lemme 4.28, pour connaître les distances entre les points de B ( S0 , M ) ∪ [j∈{1,...,Q} 0 , M ) ∪ B ( Z 1 B ( Z j 1 , M ) (109) B ( Si , M ) ∪ (110) B ( Y j i+1 , M ) ∪ B (x, k + 2M ) et ceux de [i∈{1,...,m+1} [ i∈{0,...,m},j∈{1,...,li } il suffit de connaître les distances entre les points de (109) et C points de (110), avec C = C (δ, K, N , Q, P , M ) et grâce à (98) ces distances sont dé- terminées à C ′ = C (δ, K, N , Q, P , M ) près par les distances de S1 à ces C 0( Z ))j∈{1,...,Q} et points (qui font partie de la donnée de Z ) et les entiers r, (tj 1 ( Z ), qui sont eux-mêmes déterminés à C ′′ = C (δ, K, N , Q, P , M ) près par t1 r0 (Z ), r, t, t1 , ..., tQ . (cid:3) Suite de la démonstration du lemme 4.40. On termine maintenant la preuve de l’inégalité (100). Pour Z ∈ ΛZ,t,(t1 ,...,tQ ) , et (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j 0 )j∈{1,...,Q} , Z 1 1 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z ) x 119 (ux,r,tKx,Q,(t1 ,...,tQ ) (ea1 ∧ ... ∧ eap−1 ))(b1 , ..., bp ), on considère X(b1 ,...,bp) où la somme porte sur les énumérations de S1 = {b1 , ..., bp} telles que (b1 , . . . , bp , S1 , ..., Sm+1 , ( Y j )−1(Z ). i+1)i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x Comme la somme (111) a au plus p! termes, le 3) de la proposition 3.37 montre qu’elle est ma jorée par une constante de la forme C (δ, K, N , Q, P , M ). D’après le sous-lemme 4.43 la somme (111) ne dépend que de Z et on peut donc la noter αZ, Z ,t,(t1 ,...,tQ ) . D’après les sous-lemmes 4.41 et 4.42 on a 1 (p − 1)! XZ∈ΛZ,t,(t1 ,...,tQ ) Par Cauchy-Schwarz et grâce au sous-lemme 4.44, on en déduit que ξZ (ux,r,tKx,Q,(t1 ,...,tQ )f )2 ≤ C XZ∈ΛZ,t,(t1 ,...,tQ ) avec C = C (δ, K, N , Q, P , M ). ξZ (ux,r,tKx,Q,(t1 ,...,tQ )f ) = αZ, Z ,t,(t1 ,...,tQ) ξ Z (f ). (112) (111) ξ Z (f )2 (113) Sous-lemme 4.45 Il existe D ′ = C (δ, K ) et C = C (δ, K, N , Q, P , M ) tels que pour tout (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x le nombre de possibilités pour (a1 , . . . , ap−1) tels que (a1 , . . . , ap−1 , {a1 , . . . , ap−1}, S0 , ..., Sm , (Y j i−1)i∈{0,...,m+1},j∈{1,...,li } ) appartienne à Y p−1,k ,m+1,(l0 ,...,lm+1) et vérifie x d(x, {a1 , . . . , ap−1}) − d(x, S0) − r ≤ QF (114) soit ≤ C eD ′ r . Démonstration. Il existe D ′ = C (δ, K ) tel que tout y ∈ X et pour tout R ∈ N, ♯B (y , R) ≤ eD ′R . Pour tout a ∈ {a1 , . . . , ap−1}, grâce à (114) et au fait que S0 ⊂ 2F -géod(a, x) on a d(a, S0) ≤ r + QF + N + 2F . (cid:3) 120 Fin de la démonstration du lemme 4.40. Grâce à (98), on peut appliquer le sous-lemme 4.45 et grâce au lemme 4.29, on en déduit qu’il existe D ′ = C (δ, K ) et C = C (δ, K, N , Q, P , M ) tels que pour Z ∈ ΛZ,t,(t1 ,...,tQ ) , ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1 (Z )(cid:1)C eD ′ r . )−1( Z )(cid:1) ≤ ♯(cid:0)(π p,k ,m,(l0 ,...,lm) x x Grâce à (98) on a e2s(r0 (Z )−k) ≤ e2s(r0 ( Z )−k) eQF e−2sr . On suppose αD ′ ≤ s, ce qui est permis par (Hα). Donc il existe C = C (δ, K, N , Q, P , M ) tel que (cid:16)e2s(r0 (Z )−k)♯(cid:0)(π p,k ,m,(l0 ,...,lm) )−1(Z )(cid:1)−α(cid:17) x )−1( Z )(cid:1)−α(cid:17). ≤ C e−sr(cid:16)e2s(r0 ( Z )−k)♯(cid:0)(π p−1,k ,m+1,(l0 ,...,lm+1) x L’inégalité (100) résulte de (113) et (115). On a montré (100) et donc (97) et (96). Ceci termine la preuve du lemme 4.40. (cid:3) On a donc montré la proposition 4.30. (115) x Jxe−τ θ♭ x et eτ θ♭ x ∂ e−τ θ♭ 4.6 Continuité de eτ θ♭ x Le but de ce sous-paragraphe est de montrer la proposition 4.46. Proposition 4.46 Pour tout T ∈ R+ et tout r ∈ N, x hxe−τ θ♭ (eτ θ♭ x ∂ e−τ θ♭ x )τ ∈[0,T ] , (eτ θ♭ x Jxe−τ θ♭ x )τ ∈[0,T ] , (eτ θ♭ x )τ ∈[0,T ] x ux,rKxe−τ θ♭ et (eτ θ♭ x )τ ∈[0,T ] s’étendent en des opérateurs continus sur le C[0, T ]-module hilbertien Hx,s [0, T ]. On a inclus les opérateurs hx et ux,rKx dans l’énoncé de cette proposition pour un usage ultérieur. On rappelle que pour tout t ∈ R+ et pour p ∈ {1, . . . , pmax}, on a défini x : C(∆p ) → C(∆p ) par θ♭ x (eS ) = ρ♭ θ♭ x (S )eS pour tout S ∈ ∆p et que la fonction ρ♭ x : X → R+ avait été définie par ρ♭ x (a) = d♭ (x, a) et étendue en une fonction ρ♭ x : ∆ → R+ par la formule x (S ) = Pa∈S ρ♭ x (a) ρ♭ ♯S si S est non vide et ρ♭ x (∅) = 0. 121 si S est non vide et ρx (∅) = 0. De façon analogue on définit ρx : X → R+ en posant ρx (a) = d(x, a) et on étend cette fonction en ρx : ∆ → R+ par la formule ρx (S ) = Pa∈S ρx (a) ♯S D’après la proposition 3.49, pour x, y ∈ X , on a d(x, y ) ≤ d♭(x, y ) ≤ d(x, y ) + 7δ . Il en résulte que pour tout S ∈ ∆ on a ρx (S ) ≤ ρ♭ x (S ) ≤ ρx (S ) + 7δ . On définit l’opérateur θx : C(∆p ) → C(∆p ) par θx (eS ) = ρx (S )eS . Démonstration de la proposition 4.46 en admettant les lemmes 4.47 et 4.48. Pour τ ∈ R+ et p ∈ {1, . . . , pmax}, on introduit les opérateurs eτ (θ♭ x−θx ) : C(∆p ) → C(∆p ) et eτ θx : C(∆p ) → C(∆p ) en posant x−θx ) (eS ) = eτ (θ♭ eτ (θ♭ x (S )−θx (S )) eS et eτ θx (eS ) = eτ θx (S ) eS . Ces opérateurs commutent entre eux et on a bien sûr eτ θ♭ x = eτ (θ♭ x−θx ) eτ θx . Pour p = 0 on définit θ♭ x et θx comme 0 sur C(∆0 ) = C. La proposition 4.46 résulte des lemmes 4.47 et 4.48. (cid:3) Lemme 4.47 Pour tout T ∈ R+ et tout r ∈ N, (eτ θx ∂ e−τ θx )τ ∈[0,T ] , (eτ θx Jxe−τ θx )τ ∈[0,T ] , (eτ θx hxe−τ θx )τ ∈[0,T ] et (eτ θx ux,rKxe−τ θx )τ ∈[0,T ] s’étendent en des opérateurs continus sur le C[0, T ]-module hilbertien Hx,s [0, T ]. Pour la démonstration de ce lemme on a besoin de la notation suivante. p,k ,m,(l0 ,...,lm) Notation. Pour Z ∈ Y et x (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x ρx (S0) et ρx (S1) ne dépendent que de Z et on les note ρ0 x (Z ) et ρ1 x (Z ). On ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 adopte une notation similaire pour Z ∈ Y . x Remarque. Il est évidemment faux que ρ♭ x (S0 ) et ρ♭ x (S1) ne dépendent que de Z et c’est pour cette raison que la preuve de la proposition 4.46 n’est pas aussi simple que celle du lemme 4.47. Démonstration du lemme 4.47. On reprend la démonstration des proposi- tions 4.21 et 4.30 du sous-paragraphe précédent, qui affirmaient la continuité de ∂ , Jx , hx et ux,rKx . Les seuls ingrédients supplémentaires sont les faits suivants : – dans les notations ci-dessus ρx (S0) et ρx (S1) ne dépendent que de Z )−1 (Z ), 122 – d’après le 1)a) et le 2)a) de la proposition 3.37, il existe C = C (δ, K, N , Q) tel que si eT apparaît dans ∂ (eS ) ou Jx (eS ) avec un coefficient non nul, on a d(x, T ) ≤ d(x, S ) + C , d’où ρx (T ) ≤ ρx (S ) + C + N . Voici de façon plus précise les modifications à apporter : x ( Z )) ξ Z (f ) x ( Z )−ρ0 – pour la proposition 4.21, dans (55) on remplace ξ Z (f ) par eτ (ρ1 x ( Z ) − ρ0 x ( Z ) ≤ N , et on remarque que ρ1 – pour le lemme 4.32, dans (80) on remplace ξZ (f ) par eτ (ρx (U )−ρ0 x (Z )) ξZ (f ) et on remarque que ρx (U ) ≤ P + N , x ( Z )) ξ Z (f ) x ( Z )−ρ0 – pour le lemme 4.36, dans (95) on remplace ξ Z (f ) par eτ (ρ1 x ( Z ) − ρ0 x ( Z ) ≤ (q + 2)N grâce à (84), et on remarque que ρ1 x ( Z )) ξ Z (f ) x ( Z )−ρ0 – pour le lemme 4.40, dans (112) on remplace ξ Z (f ) par eτ (ρ1 x ( Z ) − ρ0 x ( Z ) ≤ QF + N par (98) et on remarque que ρ1 et en plus on remplace les opérateurs par les opérateurs conjugués à de nom- breux endroits (notamment dans les égalités (55), (80), (95) et (112)). En utilisant le fait que les fonctions à support fini sont denses dans Hx,s on montre que les opérateurs du lemme 4.47 sont continus en τ pour la topologie forte et leurs adjoints aussi. Ceci justifie le fait qu’ils s’étendent en des morphismes de C[0, T ]-modules hilbertiens. (cid:3) Lemme 4.48 Pour tout T ∈ R+ , l’opérateur (eτ (θ♭ x−θx ) )τ ∈[0,T ] s’étend en un automorphisme du C[0, T ]-module hilbertien Hx,s [0, T ]. Démonstration du lemme 4.48 en admettant le lemme 4.49. Soit p ∈ {1, . . . , pmax}. Le lemme 4.48 résulte immédiatement du lemme suivant. (cid:3) Lemme 4.49 L’opérateur θ♭ x−θx s’étend en un opérateur continu sur Hx,s(∆p ), dont la norme est bornée par une constante du type C (δ, N , K, Q, P , M , s, B ). Bien entendu cet énoncé n’est pas vrai pour les opérateurs θ♭ x et θx séparé- ment. Pour u1 , u2 , u3 , v1 , v2 , v3 ∈ [0, 1[ on définit (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 : X → R+ en po- u1 ,u2 ,u3 (a) = d♭ v1 ,v2 ,v3 sant (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 (x, a) et on l’étend en une fonction (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 : ∆ → R+ par la formule u1 ,u2 ,u3 (S ) = Pa∈S (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 (a) (ρ♭ x )v1 ,v2 ,v3 ♯S Pour u1 , u2 , u3 , v1 , v2 , v3 ∈ [0, 1[ et S ∈ ∆ on a ρx (S ) ≤ (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 (S ) ≤ ρx (S ) + 7δ. De plus, pour S ∈ ∆, x )(S ) = Zu1 ,u2 ,u3 ,v1 ,v2 ,v3∈[0,1[ (ρ♭ si S est non vide et (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 (∅) = 0. x )v1 ,v2 ,v3 (ρ♭ u1 ,u2 ,u3 (S )du1du2du3dv1dv2dv3 . 123 Lemme 4.50 Pour S ∈ ∆p , (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 (S ) ne dépend que de la connaissance des points de {x} ∪ S ∪ [a∈S,j∈{1,...,6} {y ∈ 3δ -géod(x, a), d(x, y ) − wj ≤ N + 6δ + 4} (116) et des distances entre ces points, où l’on note w1 = E ( u1 6 w4 = E ((1 − v1 6 1 6 u2 6 5 6 − v2 6 d(x, S )), w2 = E (( + )d(x, S )), w3 = E (( + )d(x, S )), w5 = E (( )d(x, S )), w6 = E (( 2 6 u3 6 4 6 − )d(x, S )), v3 6 )d(x, S )). (117) (118) Démonstration. L’énoncé est évident si d(x, S ) ≤ 6δ et on suppose donc d(x, S ) > 6δ . Par la construction même de d♭ , (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 (S ) ne dépend que de la connaissance de x, de S et de la réunion pour a ∈ S de ∪ Y 2∆x,a+E ((∆x,a+1)u3 ) ∪ Y ∆x,a+E ((∆x,a+1)u2 ) Y E ((∆x,a+1)u1 ) x,a x,a x,a ∪ Y 2∆x,a+E ((∆x,a+1)v3 ) ∪ Y ∆x,a+E ((∆x,a+1)v2 ) ∪Y E ((∆x,a+1)v1 ) . a,x a,x a,x Soit a ∈ S . On rappelle que ∆x,a = E ( d(x,a) ) − δ et que pour u, v ∈ X et 6 r ∈ {0, ..., E ( d(u,v) )− 3δ}, on note Y r u,v l’ensemble des points z ∈ 3δ -géod(u, v ) 2 tels que d(u, z) ∈ {r, ..., r + 3δ}. Il suffit donc de montrer que l’ensemble (118) est inclus dans l’ensemble [j∈{1,...,6} {y ∈ 3δ -géod(x, a), d(x, y ) − wj ≤ N + 6δ + 4}. On a (cid:12)(cid:12)∆x,a − d(x,a) (cid:12)(cid:12) ≤ δ + 1 et pour i ∈ {1, ..., 3}, 6 uid(x, a) vid(x, a) (cid:12)(cid:12)E ((∆x,a + 1)ui) − (cid:12)(cid:12) ≤ δ + 1 et (cid:12)(cid:12)E ((∆x,a + 1)vi ) − (cid:12)(cid:12) ≤ δ + 1. 6 6 Il en résulte facilement que pour i ∈ {1, ..., 3} et y ∈ Y (i−1)∆x,a+E ((∆x,a+1)ui ) x,a on a (119) d(x, y ) ∈ [(i − 1)∆x,a + E ((∆x,a + 1)ui ), (i − 1)∆x,a + E ((∆x,a + 1)ui) + 3δ ] ((i − 1) + ui)d(x, a) ((i − 1) + ui)d(x, a) − (3δ + 3), ⊂ [ + (6δ + 3)]. 6 6 D’autre part ((i−1)+ui )d(x,a) − wi ≤ N + 1 car d(x, a) − d(x, S ) ≤ N . On en 6 déduit que y appartient à (119). 124 Soit maintenant i ∈ {1, ..., 3} et y ∈ Y (i−1)∆x,a+E ((∆x,a+1)vi ) a,x appartient à . Donc d(a, y ) [(i − 1)∆x,a + E ((∆x,a + 1)vi ), (i − 1)∆x,a + E ((∆x,a + 1)vi ) + 3δ ] ((i − 1) + vi )d(x, a) ((i − 1) + vi )d(x, a) ⊂ [ − (3δ + 3), + (6δ + 3)]. 6 6 Comme y ∈ 3δ -géod(x, a) on a d(x, y ) ∈ [d(x, a)−d(a, y ), d(x, a)−d(a, y )+3δ ], et comme d(a, y ) appartient à (120) on en déduit que d(x, y ) appartient à (120) [ − (6δ + 3), (7 − i − vi )d(x, a) 6 (7 − i − vi )d(x, a) 6 D’autre part (7−i−vi )d(x,a) − wi+3 ≤ N + 1 car d(x, a) − d(x, S ) ≤ N . On 6 en déduit que y appartient à (119). (cid:3) u1 ,u2 ,u3 : C(∆p ) → C(∆p ) l’opérateur défini par x )v1 ,v2 ,v3 Enfin on note (θ♭ (θ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 (eS ) = (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 (S )eS , + (6δ + 3)]. x )v1 ,v2 ,v3 (θ♭ u1 ,u2 ,u3 du1du2du3dv1dv2dv3 . de sorte que x = Zu1 ,u2 ,u3 ,v1 ,v2 ,v3∈[0,1[ θ♭ Démonstration du lemme 4.49. Soit f ∈ C(∆p ) . Par définition B−(m+Pm i=0 li ) XZ ∈Y Hx,s (∆p ) = Xk ,m,l0 ,...,lm∈N x − θx )(f )k2 k(θ♭ p,k,m,(l0 ,...,lm ) x m e2s(r0 (Z )−k)(cid:16) si (Z )−li (cid:17)♯(cid:0)(π p,k ,m,(l0 ,...,lm) )−1(Z )(cid:1)−α (cid:12)(cid:12)ξZ (cid:0)(θ♭ 2 Yi=0 x − θx )(f )(cid:1)(cid:12)(cid:12) x On va voir que le lemme 4.49 résulte de l’inégalité (121) ci-dessous. p,k ,m,(l0 ,...,lm) Soient k , m, l0 , . . . , lm ∈ N et Z ∈ Y . On va montrer qu’il x existe C = C (δ, K, N , Q, P , M ) tel que 2 ≤ C XZ ∈ΛZ (cid:0)r0 ( Z ) + 1(cid:1)−6 (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) 2 (cid:12)(cid:12)ξZ (cid:0)(θ♭ x − θx )(f )(cid:1)(cid:12)(cid:12) ♮,p,k ,m,(l0 ,...,lm),6,0 formée des Z tels que pour tout où ΛZ est la partie de Y x (a1 , . . . , ap , S0 , ..., Sm , ( Y j i )i∈{0,...,m},j∈{1,...,li } , ( Z j 0 )j∈{1,...,6}) )−1( Z ) ∈ (π ♮,p,k ,m,(l0 ,...,lm),6,0 x (121) . 125 )−1(Z ) (122) (123) )−1 (Z ). on ait (a1 , . . . , ap , S0 , ..., Sm , ( Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x On justifie maintenant le fait que l’inégalité (121) implique l’énoncé du lemme. Le lemme 4.29 montre que pour Z ∈ ΛZ on a )−1( Z ) ≤ C ♯(π p,k ,m,(l0 ,...,lm) ♯(π ♮,p,k ,m,(l0 ,...,lm),6,0 x x avec C = C (δ, K, N , Q, P , M ). En sommant sur Z et en appliquant le lemme 4.24 à µ0 = 6 et µ1 = 0, on voit que l’inégalité (121) implique l’énoncé du lemme. L’inégalité (121) découle de l’inégalité (122) plus précise ci-dessous. Soient u1 , u2 , u3 , v1 , v2 , v3 ∈ [0, 1[. On va montrer qu’il existe C = C (δ, K, N , Q, P , M ) tel que 2 2 ≤ C XZ ∈(ΛZ ) x )v1 ,v2 ,v3 (cid:12)(cid:12)ξZ (cid:0)((θ♭ u1 ,u2 ,u3 (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) u1 ,u2 ,u3 − θx )(f )(cid:1)(cid:12)(cid:12) v1 ,v2 ,v3 u1 ,u2 ,u3 est l’ensemble des Z ∈ ΛZ tels que, en notant où (ΛZ )v1 ,v2 ,v3 u1 u3 1 u2 2 )r0( Z )), w3 = E (( r0 ( Z )), w2 = E (( w1 = E ( + + 6 6 6 6 6 v1 v2 5 4 )r0 ( Z )), w6 = E (( )r0( Z )), w5 = E (( w4 = E ((1 − 6 − 6 − 6 6 on ait, pour tout i )i∈{0,...,m},j∈{1,...,li } , ( Z j (a1 , . . . , ap , S0 , ..., Sm , ( Y j 0 )j∈{1,...,6}) )−1( Z ) ∈ (π ♮,p,k ,m,(l0 ,...,lm),6,0 x et pour tout j ∈ {1, ..., 6}, Z j 0 = [b∈ S0 La condition (123) implique que pour Z ∈ (ΛZ )v1 ,v2 ,v3 u1 ,u2 ,u3 et pour j ∈ {1, ..., 6}, 0 ( Z ) = wj . on a tj Sous-lemme 4.51 Soit (a1 , . . . , ap , S0 , ..., Sm , ( Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x 0 )j∈{1,...,6} par (123). Alors les parties Z j On définit ( Z j 0 sont de diamètre in- férieur ou égal à P /3 et il existe Z ∈ (ΛZ )v1 ,v2 ,v3 u1 ,u2 ,u3 tel que (a1 , . . . , ap , S0 , ..., Sm , ( Y j i )i∈{0,...,m},j∈{1,...,li } , ( Z j 0 )j∈{1,...,6}) appartienne à (π ♮,p,k ,m,(l0 ,...,lm),6,0 )−1( Z ). x {z ∈ géod(x, b), d(x, z) = wj }. )r0 ( Z )), v3 6 )r0( Z )) )−1 (Z ). (124) 126 Démonstration. Soit j ∈ {1, ..., 6} et z , z ′ ∈ Z j 0 . Soit b ∈ S0 . On a z , z ′ ∈ 2N -géod(x, b) et d(x, z) = d(x, z ′ ) donc par (Hδ (z , x, z ′ , b)), d(z , z ′ ) ≤ 2N + δ ≤ P /3. Comme les parties Z j 0 sont non vides et P /3 ≤ M , l’argument que nous venons de donner montre aussi que la condition (123) est vérifiée par les autres éléments de la classe d’équivalence Z de l’élément (124) et donc Z ∈ (ΛZ )v1 ,v2 ,v3 u1 ,u2 ,u3 . (cid:3) Sous-lemme 4.52 Pour Z ∈ (ΛZ )v1 ,v2 ,v3 u1 ,u2 ,u3 et (a1 , . . . , ap , S0 , ..., Sm , ( Y j i )i∈{0,...,m},j∈{1,...,li } , ( Z j 0 )j∈{1,...,6}) )−1 ( Z ), ∈ (π ♮,p,k ,m,(l0 ,...,lm),6,0 x u1 ,u2 ,u3 ( S0 ) ne dépend que de la connaissance des points de x )v1 ,v2 ,v3 (ρ♭ B ( Z j B (x, k + 2M ) ∪ B ( S0 , M ) ∪ [j∈{1,...,6} 0 , M ) et des distances entre ces points. (125) Démonstration. Il suffit de montrer que l’ensemble (116) figurant dans le lemme 4.50 (avec S0 au lieu de S ) est inclus dans (125). Soit a ∈ S0 , j ∈ {1, ..., 6}, et y ∈ 3δ -géod(x, a) vérifiant d(x, y ) − wj ≤ N + 6δ + 4. Soit z ∈ géod(x, a) vérifiant d(x, z) = wj , si bien que z appartient à Z j 0 . Comme d(x, y )−d(x, z) ≤ N +6δ+4 et que y , z ∈ 3δ -géod(x, a), (Hδ (y , x, z , a)) implique d(y , z) ≤ (N + 6δ + 4)+ 3δ + δ . On suppose (N + 6δ + 4)+ 3δ + δ ≤ M , ce qui est permis par (HM ). On a donc d(y , Z j 0 ) ≤ M . (cid:3) Sous-lemme 4.53 Le cardinal de (ΛZ )v1 ,v2 ,v3 u1 ,u2 ,u3 est majoré par une constante de la forme C (δ, K, N , Q, P , M ). Démonstration. Grâce au lemme 4.28, pour connaître les distances entre les points de [j∈{1,...,6} B ( Z j 0 , M ) (126) et ceux de [i∈{0,...,m} B ( Si , M ) ∪ [ i∈{0,...,m},j∈{1,...,li } B ( Y j i , M ) ∪ B (x, k + 2M ) (127) 127 il suffit de connaître les distances entre les points de (126) et C points de (127), avec C = C (δ, K, N , Q, P , M ) et ces distances sont déterminées à C ′ = C (δ, K, N , Q, P , M ) par Z et u1 , u2 , u3 , v1 , v2 , v3 (en fait le lemme 4.28 fournit un énoncé légèrement différent où B ( S0 , M ) figure dans (126) et non dans (127), mais il est clair que cet énoncé implique le nôtre). (cid:3) Fin de la démonstration du lemme 4.49. D’après le sous-lemme 4.52, pour Z ∈ (ΛZ )v1 ,v2 ,v3 u1 ,u2 ,u3 on a x )v1 ,v2 ,v3 u1 ,u2 ,u3 − θx )(f )(cid:1) = (α Z )v1 ,v2 ,v3 ξ Z (cid:0)((θ♭ u1 ,u2 ,u3 ξ Z (f ) p,k ,m,(l0 ,...,lm) u1 ,u2 ,u3 (cid:12)(cid:12) ≤ 7δ . Le sous-lemme 4.51 montre que pour Z ∈ Y avec (cid:12)(cid:12)(α Z )v1 ,v2 ,v3 x u1 ,u2 ,u3 − θx )(f )(cid:1) = XZ∈(ΛZ ) ξ Z (cid:0)((θ♭ x )v1 ,v2 ,v3 ξZ (cid:0)((θ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 − θx )(f )(cid:1). v1 ,v2 ,v3 u1 ,u2 ,u3 Par Cauchy-Schwarz et grâce au sous-lemme 4.53 on en déduit (122). On a montré (122) donc (121) et ceci termine la démonstration du lemme 4.49. (cid:3) , 4.7 Propriétés d’équivariance de la norme Ce sous-paragraphe a pour but de montrer la proposition 4.3. On recom- mande au lecteur de commencer par lire le premier paragraphe de [Laf08], car la démonstration de la proposition 4.3 repose sur les mêmes idées que celle de la proposition 1.10 de [Laf08] mais sur des calculs beaucoup plus compliqués. On fixe p ∈ {1, . . . , pmax} et x, x′ ∈ X . On rappelle qu’après la défi- nition 4.1, pour k , m ∈ N et (l0 , ..., lm ) ∈ Nm+1 on a introduit une rela- p,k ,m,(l0 ,...,lm) tion d’équivalence sur Y p,k ,m,(l0 ,...,lm) , et noté Y le quotient, et x x π p,k ,m,(l0 ,...,lm) l’application quotient. On va introduire maintenant une rela- x tion d’équivalence plus fine Y p,k ,m,(l0 ,...,lm) x π p,k,m,(l0 ,...,lm ) x,x′ → Y p,k ,m,(l0 ,...,lm) x,x′ et une autre encore plus fine Y p,k ,m,(l0 ,...,lm) x p,k,m,(l0 ,...,lm ) x,x′ ,⋆→ p,k ,m,(l0 ,...,lm) telles que tout élément de Y x p,k ,m,(l0 ,...,lm) dans Y et tout élément de Y x,x′ p,k ,m,(l0 ,...,lm) x,x′ ,⋆ dans Y π Y p,k ,m,(l0 ,...,lm) x,x′ ,⋆ a au plus 2d(x, x′ ) + 1 antécédents p,k ,m,(l0 ,...,lm) a au plus C antécédents x,x′ , pour une certaine constante C = C (δ, K, N , Q, P , M ). 128 On définit la relation d’équivalence Y p,k ,m,(l0 ,...,lm) x π p,k,m,(l0 ,...,lm ) x,x′ → Y p,k ,m,(l0 ,...,lm) x,x′ de la façon suivante : (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } ) et (a1 , . . . , ap , S0 , ..., Sm , ( Y j i )i∈{0,...,m},j∈{1,...,li } ) p,k ,m,(l0 ,...,lm) s’ils ont même image dans Y ont même image dans Y x,x′ et si d(x′ , S0) = d(x′ , S0). L’application (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) 7→ d(x′ , S0 ) p,k ,m,(l0 ,...,lm) se factorise donc en une application r ′ → N. On définit une 0 : Y x,x′ application p,k ,m,(l0 ,...,lm) x k ′ : Y p,k ,m,(l0 ,...,lm) x,x′ → N p,k ,m,(l0 ,...,lm) x,x′ )(cid:1) ))(cid:1) (128) 0(Z ) + d(x, x′ ) − r0(Z ) r ′ 2 en posant, pour Z ∈ Y 0(Z ), max(r ′ k ′ (Z ) = min (cid:0)r ′ 0(Z ) − r0(Z ) + k , E ( M . + 2 Voici le dessin dans le cas où X est un arbre et p = 1 (si bien que S0 = {a1} et on note b = a1 pour que le dessin serve de nouveau dans la suite) et où l’on prend M = 0 dans la formule précédente. Dans ce dessin on a choisi (b, {b}) ∈ (π 1,k ,0,(0) x,x′ on a représenté B (x, k) par une moitié de boule et on a noté u le point de B (x, k) à distance minimale de b. On définit t comme le point de \x∈B (x,k) le plus proche de x′ (ce point est unique et appartient à géod(x′ , b)). On re- marque que t est aussi le point central du triangle ubx′ . La formule précédente pour k ′ (Z ) devient k ′ (Z ) = min (cid:0)d(x′ , b), max(d(x′ , b) − d(x, b) + k , 129 d(x′ , b) + d(x, x′ ) − d(x, b) 2 géod( x, b) )−1(Z ), = d(x′ , t). La boule B (x′ , k ′ (Z )) (dont le bord contient t) est la boule de centre x′ de plus petit rayon telle que toute géodésique entre b et un point de B (x, k) la rencontre. On remarque que t dépend de S0 = {b} mais que d(x′ , t) ne dépend que de Z . Dans le dessin ci-dessous, le premier cas est le cas où b ∈ B (x, k) et le deuxième cas et le troisième cas sont distingués par l’appartenance ou non de u à géod(x, x′ ) (dans le premier et le deuxième cas x′ n’appartient pas forcément à B (x, k)). x u = t = b • x′ x x′ u = t • b x u t • x′ b Premier cas Deuxième cas Troisième cas On revient maintenant au cas général et on va distinguer trois cas comme p,k ,m,(l0 ,...,lm) dans le dessin pour les arbres. Soit Z ∈ Y x,x′ Premier cas. On suppose r0(Z ) ≤ k . Alors, par (128), M 2 Deuxième cas. On suppose r0 (Z ) > k et r0 (Z ) − r ′ 0(Z ) + d(x, x′ ) ≤ 2k . Cette dernière condition implique k ′ (Z ) = r ′ 0(Z ) + . (129) r ′ 0 (Z ) − r0 (Z ) + k ≥ E ( d’où, par (128), 0(Z ) − r0(Z ) + d(x, x′ ) r ′ 2 ) M k ′ (Z ) = r ′ 0 (Z ) − r0 (Z ) + k + 2 Troisième cas. On suppose r0(Z ) > k et r0 (Z ) − r ′ 0 (Z ) + d(x, x′ ) > 2k . 130 . (130) Cette dernière condition implique N 2 . (131) et 0(Z ) − r0(Z ) + d(x, x′ ) r ′ 2 ) r ′ 0 (Z ) − r0 (Z ) + k ≤ E ( et comme d(x, x′ ) ≤ r0(Z ) + r ′ 0 (Z ) + N on en déduit, par (128), 0(Z ) − r0(Z ) + d(x, x′ ) r ′ M (cid:12)(cid:12)(cid:12) 2 (cid:17)(cid:12)(cid:12)(cid:12) ≤ k ′ (Z ) − (cid:16)E (cid:0) (cid:1) + 2 On introduit maintenant une partition de Y p,k ,m,(l0 ,...,lm) pour la relation x suivante : (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } ) (a1 , . . . , ap , S0 , ..., Sm , ( Y j i )i∈{0,...,m},j∈{1,...,li } ) p,k ,m,(l0 ,...,lm) par π p,k ,m,(l0 ,...,lm) sont en relation s’ils ont même image Z dans Y x,x′ x,x′ et s’il existe une isométrie de B (Si , M ) ∪ [i∈{0,...,m}, [i∈{0,...,m} B (Y j i , M ) ∪ B (x, k + 2M ) ∪ B (x′ , k ′ (Z ) + 2M ) j∈{1,...,li } vers B ( Si , M ) ∪ [i∈{0,...,m}, [i∈{0,...,m} j∈{1,...,li } qui envoie ai sur ai pour i ∈ {1, . . . , p}, Si sur Si pour i ∈ {0, . . . , m}, Y j i sur Y j i pour i ∈ {0, . . . , m}, j ∈ {1, . . . , li} et est l’identité sur B (x, k + 2M ) et sur B (x′ , k ′ (Z ) + 2M ). p,k ,m,(l0 ,...,lm) le quotient de Y p,k ,m,(l0 ,...,lm) On note Y x x,x′ ,⋆ d’équivalence, et π p,k ,m,(l0 ,...,lm) l’application quotient. x,x′ ,⋆ B ( Y j i , M ) ∪ B (x, k + 2M ) ∪ B (x′ , k ′ (Z ) + 2M ) pour cette relation Lemme 4.54 a) Les fibres de l’application surjective p,k ,m,(l0 ,...,lm) → Y Y x,x′ sont toutes de cardinal ≤ 2d(x, x′ ) + 1. b) Il existe une constante C = C (δ, K, N , Q, P , M ) tel le que les fibres de l’application surjective p,k ,m,(l0 ,...,lm) x p,k ,m,(l0 ,...,lm) x,x′ ,⋆ soient toutes de cardinal ≤ C . Y p,k ,m,(l0 ,...,lm) x,x′ → Y 131 p,k,m,(l0 ,...,lm ) x )−1 (Z ) f (a1 , ..., ap ). Avant de montrer ce lemme, expliquons la stratégie de la preuve de la propo- sition 4.3. On veut ma jorer kf k2 Hx,s (∆p ) en fonction de kf k2 Hx′ ,s (∆p ) . D’abord kf k2 Hx,s (∆p ) est une somme pondérée indexée par certaines classes d’équiva- p,k ,m,(l0 ,...,lm) lence Z ∈ Y des carrés des formes linéaires x X ξZ : f 7→ (a1 ,...,ap ,S0 ,...,Sm ,(Y j i )i∈{0,...,m},j∈{1,...,li } )∈(π Grâce au lemme précédent on peut couper chaque classe d’équivalence Z en p,k ,m,(l0 ,...,lm) morceaux Z (paramétrés par Y ) dont le nombre est inférieur ou x,x′ ,⋆ égal à C (2d(x, x′ )+1). Par Cauchy-Schwarz on a donc pour tout f , ξZ (f )2 ≤ p,k ,m,(l0 ,...,lm) C (2d(x, x′ )+1) P Z ξ Z (f )2 où la somme porte sur les Z ∈ Y dont x,x′ ,⋆ p,k ,m,(l0 ,...,lm) est Z . On montrera ensuite (dans le lemme 4.61) l’image dans Y x que chaque forme linéaire ξ Z est proportionnelle à une forme linéaire ξZ ′ Hx′ ,s (∆p ) . Par définition Z est la donnée apparaissant dans la formule pour kf k2 des distances entre les points de B (x, k + 2M ), ceux de B (x′ , k ′ + 2M ) et les points à distance ≤ M de la réunion de toutes les parties Si et Y j i . Le dessin ci-dessous correspond au troisième cas du dessin pour les arbres (qui est le cas le plus intéressant) et les deux boules y représentent B (x, k) et B (x′ , k ′ ). Les parties Si sont représentées par des cercles plus grands que les parties Y j i dans un souci de clarté, bien que leur diamètre maximal soit plus petit (N au lieu de P ). parties éliminées parties conservées x Y j Sm ❥❡❡ m . . . . . . Sm′+1 Sm′ ❡❡❡❡❡❥ ❥ Y j Y j m′ m′ conservées éliminées •x′ S1 ❥ Y j ❡❡ 0 S0 ❥ La classe d’équivalence Z ′ sera la donnée des distances entre les points de B (x′ , k ′ + 2M ) et les points à distance ≤ M de la réunion des parties Si et Y j i situées à droite de la boule B (x′ , k ′ ), qui sont indiquées sur le dessin comme “parties conservées”. On montrera qu’il existe m′ tel que les parties conservées soient les Si pour i ∈ {0, ..., m′}, tous les Y j i pour i ∈ {0, ..., m′ −1} et certains des Y j m′ . La raison pour laquelle ξ Z est proportionnelle à ξZ ′ est 132 que la connaissance de Z et des parties Si et Y j i éliminées (qui se trouvent dans B (x, k) ou B (x′ , k ′ ) ou entre ces deux boules) détermine Z , car si Γ est une partie conservée et ∆ une partie éliminée toute géodésique entre B (Γ, M ) et B (∆, M ) ∪ B (x, k + 2M ) traverse B (x′ , k ′ + 2M ), comme on le verra dans le lemme 4.57. Par conséquent, pour connaître les distances entre B (Γ, M ) et B (∆, M ) ∪ B (x, k + 2M ) , il suffit de connaître les distances entre B (Γ, M ) et B (x′ , k ′ + 2M ), qui font partie de la donnée de Z . La difficulté sera ensuite que chaque ξZ ′ est proportionnelle à ξ Z pour une infinité de p,k ,m,(l0 ,...,lm) Z ∈ Sk ,m,(l0 ,...,lm) Y et on devra vérifier que, compte tenu des x,x′ ,⋆ pondérations et grâce à (HB ), la somme sur Z n’introduit pas de divergence. Démonstration du lemme 4.54. Le a) vient simplement du fait que pour (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x on a d(x′ , S0 ) ∈ [d(x, S0) − d(x, x′ ), d(x, S0) + d(x, x′ )]. Montrons b). Soit p,k ,m,(l0 ,...,lm) Z ∈ Y . On doit montrer que le nombre d’antécédents de Z dans x,x′ p,k ,m,(l0 ,...,lm) est ma joré par une constante C = C (δ, K, N , Q, P , M ). On Y x,x′ ,⋆ note u un point de B (x, k) tel que pour tout )−1(Z ) i )i∈{0,...,m},j∈{1,...,li } ) ∈ (π p,k ,m,(l0 ,...,lm) (a1 , . . . , ap , S0 , ..., Sm , (Y j x,x′ u soit à distance minimale de S0 (cela ne dépend que de Z car les distances entre les points de S0 et ceux de B (x, k) font partie de la donnée de Z ). On a alors d(u, S0) = max(0, r0(Z ) − k). On rappelle que d’après le lemme 4.14, pour tout i )i∈{0,...,m},j∈{1,...,li } ) ∈ (π p,k ,m,(l0 ,...,lm) (a1 , . . . , ap , S0 , ..., Sm , (Y j x,x′ et pour b ∈ S0 à distance minimale de u (si bien que u est un point de B (x, k) à distance minimale de b) on a )−1(Z ) S0 ∪ · · · ∪ Sm ∪ [ i∈{0,...,m},j∈{1,...,li } Y j i ⊂ 4P -géod(b, u) B (Si , M ) ∪ d’où par le lemme 3.2, [ [i∈{0,...,m} i∈{0,...,m},j∈{1,...,li } On va distinguer trois cas qui correspondent à peu près aux trois cas envisagés dans le dessin pour les arbres (auquel le lecteur peut se reporter pour lire la démonstration). B (Y j i , M ) ⊂ (2M + 4P )-géod(b, u). 133 )−1(Z ) Premier cas. On suppose r0(Z ) ≤ k . Soit i )i∈{0,...,m},j∈{1,...,li } ) ∈ (π p,k ,m,(l0 ,...,lm) (a1 , . . . , ap , S0 , ..., Sm , (Y j x,x′ et b ∈ S0 à distance minimale de u. On a alors b = u d’où [ [i∈{0,...,m} B (Y j B (Si , M ) ∪ i , M ) i∈{0,...,m},j∈{1,...,li } ⊂ (2M + 4P )-géod(u, u) = B (u, M + 2P ) ⊂ B (x, k + 2M ) où la dernière inclusion vient de l’inégalité 2P ≤ M que l’on suppose grâce à (HM ). Donc (π p,k ,m,(l0 ,...,lm) )−1(Z ) est un singleton et a fortiori l’image inverse x,x′ p,k ,m,(l0 ,...,lm) p,k ,m,(l0 ,...,lm) → Y de Z par l’application Y est un singleton. x,x′ ,⋆ x,x′ Deuxième cas. On suppose r0 (Z ) > k et r0 (Z ) − r ′ 0(Z ) + d(x, x′ ) ≤ 2k . Par (130) on a k ′ (Z ) = r ′ 0 (Z ) − r0 (Z ) + k + M 2 . Soit (132) )−1(Z ) i )i∈{0,...,m},j∈{1,...,li } ) ∈ (π p,k ,m,(l0 ,...,lm) (a1 , . . . , ap , S0 , ..., Sm , (Y j x,x′ et b ∈ S0 à distance minimale de u. On a donc d(x, b) − d(x′ , b) + d(x, x′ ) ≤ 2k + N , d(x, b) > k , M (cid:12)(cid:12)k ′ (Z ) − (d(x′ , b) − d(x, b) + k + )(cid:12)(cid:12) ≤ N . 2 Soit t ∈ B (x′ , k ′ (Z ) + 2M ) à distance minimale de b. On a u ∈ géod(x, b), et δ (x′ , x, u, b)) et (132) on a d(x, u) = k , d(u, b) = d(x, b) − k . Par (H 0 d(x′ , u) ≤ max(d(x, x′ ) − k , d(x′ , b) − d(x, b) + k) + δ ≤ d(x′ , b) − d(x, b) + k + N + δ = d(x′ , b) − d(u, b) + N + δ donc u ∈ (N + δ )-géod(x′ , b). On a t ∈ géod(x′ , b), d(x′ , t) = min(d(x′ , b), k ′ (Z ) + 2M ) et k ′ (Z ) − (d(x′ , b) − d(u, b) + M 2 ) ≤ N donc 5M d(x′ , t) ≤ k ′ (Z ) + 2M ≤ d(x′ , u) + 2 5M d(x′ , t) ≥ min(d(x′ , b), d(x′ , u) + 2 − 2N − δ ), + N , 134 et comme d(x′ , u) ≤ d(x′ , b) + N + δ , on en déduit 5M d(x′ , t) − d(x′ , u) ≤ 2 car on suppose N + δ ≤ 5M 2 , ce qui est permis grâce à (HM ). Comme u ∈ (N + δ )-géod(x′ , b) et t ∈ géod(x′ , b), et grâce à (133), (Hδ (u, x′ , t, b)) implique alors (133) + N 5M 2 + N ) + (N + δ ) + δ = d(u, t) ≤ ( Cette inégalité reflète l’égalité u = t dans le dessin pour les arbres qui cor- respond au deuxième cas. On en déduit + 2N + 2δ. 5M 2 (2M + 4P )-géod(u, b) ⊂ α-géod(t, b) avec α = 2( 5M 2 + 2N + 2δ ) + (2M + 4P ). Le lemme 4.18 montre que les distances entre un point de α-géod(t, b) et les points de B (x′ , k ′ (Z ) + 2M ) sont déterminées par les distances entre ce point et les points de B (x′ , k ′ (Z ) + 2M ) ∩ B (t, α + 4δ ) ⊂ B (t, α + 4δ ) ⊂ B (u, β ) 5M 5M avec β = ( + 2N + 2δ ) + α + 4δ = 3( + 2N + 2δ ) + 2M + 4P + 4δ. 2 2 (134) B (Y j i , M ) B (Si , M ) ∪ Donc les distances entre les points de l’ensemble [i∈{0,...,m} [ i∈{0,...,m},j∈{1,...,li } (qui est inclus dans (4P + 2M )-géod(u, b) ⊂ α-géod(t, b)) et les points de B (x′ , k ′ (Z ) + 2M ) sont déterminées par les distances entre les points de (134) et les points de B (u, β ). D’autre part le c) du lemme 4.17, appliqué à – (u, b) au lieu de (c, d), – {wi , i ∈ I } égal à {u} ∪ Si∈{0,...,m} Si ∪ Si∈{0,...,m},j∈{1,...,li } Y j i , – J ⊂ I le singleton tel que {wj , j ∈ J } = {u}, – (αi , ρi ) égal à (0, β ) pour i ∈ J et à (4P , M ) pour i ∈ I \ J montre que les distances entre les points de l’ensemble (134) et ceux de B (u, β ) sont déterminées par les distances entre C = C (δ, K, N , Q, P , M ) points de (134) (déterminés par la donnée de Z ) et les points de B (u, β ). De plus le cardinal de B (u, β ) est ma joré par une constante de la forme C (δ, N , K, Q, P , M ) et les distances entre ces C points de (134) et les points 135 Soit . N 2 0(Z ) − r0(Z ) + d(x, x′ ) r ′ 2 de B (u, β ) sont déterminées à β près par la donnée de Z . Cela termine l’étude du deuxième cas. Troisième cas. On suppose r0(Z ) > k et r0 (Z ) − r ′ 0 (Z ) + d(x, x′ ) > 2k . Par (131) on a M (cid:12)(cid:12)(cid:12) 2 (cid:17)(cid:12)(cid:12)(cid:12) ≤ k ′ (Z ) − (cid:16)E (cid:0) (cid:1) + i )i∈{0,...,m},j∈{1,...,li } ) ∈ (π p,k ,m,(l0 ,...,lm) (a1 , . . . , ap , S0 , ..., Sm , (Y j x,x′ et b ∈ S0 à distance minimale de u. On a donc d(x, b) − d(x′ , b) + d(x, x′ ) ≥ 2k − N , d(x, b) > k , d(x′ , b) − d(x, b) + d(x, x′ ) M (cid:12)(cid:12)k ′ (Z ) − (E ( )(cid:12)(cid:12) ≤ N . ) + 2 2 Comme d(x, b) − d(x′ , b) + d(x, x′ ) ≥ 2k − N , (H 0 δ (x′ , x, u, b)) donne d(x′ , u) ≤ max(d(x′ , x) − k , d(x′ , b) − d(x, b) + k) + δ ≤ d(x, x′ ) − k + N + δ = d(x, x′ ) − d(x, u) + N + δ donc u ∈ (N + δ )-géod(x′ , x). )−1(Z ) Donc d(x′ , b) + d(x′ , u) − d(u, b) M (cid:12)(cid:12)k ′ (Z ) − (E ( )(cid:12)(cid:12) ≤ 2N + δ. 2 2 Soit u′ ∈ B (x′ , k ′ (Z ) + 2M ) à distance minimale de u et b′ ∈ B (x′ , k ′ (Z ) + 2M ) à distance minimale de b. Soit w ∈ 4P -géod(u, b). On applique le lemme 4.6 à u, b, x′ , u′ , b′ , w avec l = 2 et x′ comme point base. On note t le point central du triangle de sommets Ψu, Ψb, Ψx′ . Dans le dessin ci-dessous le point de géod(Ψu, Ψb) le plus proche de Ψw est arbitraire. (135) ) + Ψu Ψw t • Ψu′ • Ψb′ Ψx′ On a ∈ [ d(Ψx′ , Ψu) + d(Ψx′ , Ψb) − d(Ψu, Ψb) d(Ψx′ , t) = 2 d(x′ , u) + d(x′ , b) − d(u, b) d(x′ , u) + d(x′ , b) − d(u, b) 2 2 , + δ ]. 136 Ψb (136) Grâce au a) du lemme 4.7 on a de plus Ψu′ ∈ géod(Ψx′ , Ψu), d(Ψx′ , Ψu′ ) = min(k ′ (Z ) + 2M , d(Ψx′ , Ψu)) et Ψb′ ∈ géod(Ψx′ , Ψb), d(Ψx′ , Ψb′ ) = min(k ′ (Z ) + 2M , d(Ψx′ , Ψb)). Par (135) on a k ′ (Z ) + 2M ≥ d(x′ ,u)+d(x′ ,b)−d(u,b) + δ car on suppose 5M 2 ≥ 2 2N + 2δ + 1, ce qui est permis par (HM ). On déduit alors de (136) que + 2N + δ + 2N + δ. 5M Ψu′ ∈ géod(t, Ψu), d(t, Ψu′ ) ≤ 2 5M et Ψb′ ∈ géod(t, Ψb), d(t, Ψb′ ) ≤ 2 Grâce au a) du lemme 4.7 on a Ψw ∈ (4P + 2δ )-géod(Ψb, Ψu) donc Ψw est à distance ≤ 2P + δ de géod(Ψb, Ψu). Donc l’une au moins des assertions suivante est vraie (suivant que le point de géod(Ψb, Ψu) le plus proche de Ψw appartient à géod(Ψb, t) ou géod(t, Ψu)) : – Ψw est à distance ≤ 5M 2 + 2N + 2P + 2δ de géod(Ψb′ , Ψb) – ou Ψw est à distance ≤ 5M 2 + 2N + 2P + 2δ de géod(Ψu′ , Ψu) Il résulte facilement de ce qui précède et du b) du lemme 4.7 que l’une au moins des assertions suivante est vraie – w ∈ (5M + 4N + 4P + 8δ )-géod(b′ , b), – ou w ∈ (5M + 4N + 4P + 8δ )-géod(u′ , u). On suppose 4N + 4P + 8δ ≤ M , ce qui est possible par (HM ). Il résulte de ce qui précède que w ∈ 6M -géod(b′ , b) ∪ 6M -géod(u′ , u). On applique ceci à w ∈ Si∈{0,...,m} Si ∪ Si∈{0,...,m},j∈{1,...,li } Y j i ⊂ 4P -géod(u, b). Donc l’ensemble [i∈{0,...,m} [ B (Y j B (Si , M ) ∪ (137) i , M ) i∈{0,...,m},j∈{1,...,li } est inclus dans 8M -géod(b′ , b) ∪ 8M -géod(u′ , u). En appliquant le lemme 4.18 à α = 8M , (u, u′) (resp. (b, b′ )) au lieu de (b, u) et (x′ , k ′ (Z ) + 2M ) au lieu de (x, l) on voit que les distances entre n’importe quel point y ∈ 8M -géod(u′ , u) (resp. y ∈ 8M -géod(b′ , b)) et les points de B (x′ , k ′ (Z ) + 2M ) sont déterminées par les distances entre y et les points de B (x′ , k ′ (Z ) + 2M ) ∩ B (u′ , 8M + 4δ ) ⊂ B (u′ , 8M + 4δ ) et B (x′ , k ′ (Z ) + 2M ) ∩ B (b′ , 8M + 4δ ) ⊂ B (b′ , 8M + 4δ ) respectivement. De plus d(u′ , b′ ) ≤ d(Ψu′ , Ψb′ ) + 2δ ≤ d(Ψu′ , t) + d(t, Ψb′ ) + 2δ ≤ 5M + 4N + 4δ 137 (138) (139) donc B (b′ , 8M + 4δ ) ⊂ B (u′ , 13M + 4N + 8δ ). On a Ψu′ ∈ géod(Ψu, Ψb) donc u′ ∈ 4δ -géod(u, b) par le b) du lemme 4.7. En appliquant le c) du lemme 4.17 à – (u, b) au lieu de (c, d), – {wi , i ∈ I } égal à {u′} ∪ Si∈{0,...,m} Si ∪ Si∈{0,...,m},j∈{1,...,li } Y j i , – J ⊂ I le singleton tel que {wj , j ∈ J } = {u′}, – (αi , ρi ) égal à (4δ, 13M +4N +8δ ) pour i ∈ J et à (4P , M ) pour i ∈ I \J , on obtient que les distances entre les points de l’ensemble (137) et ceux de B (u′ , 13M + 4N + 8δ ) sont déterminées par les distances entre C = C (δ, K, N , Q, P , M ) points de (137) (déterminés par la donnée de Z ) et les points de B (u′ , 13M + 4N + 8δ ). De plus ces distances sont déterminées à p,k ,m,(l0 ,...,lm) C ′ = C (δ, N , K, Q, P , M ) près par la donnée de Z ∈ Y . x,x′ Cela termine l’étude du troisième cas et achève donc la démonstration du lemme 4.54. (cid:3) On définit la norme pré-hilbertienne k.kHx,x′ ,⋆,s (∆p ) sur C(∆p ) de la même p,k ,m,(l0 ,...,lm) p,k ,m,(l0 ,...,lm) façon que k.kHx,s (∆p ) , mais en remplaçant Y par Y . x,x′ ,⋆ x p,k ,m,(l0 ,...,lm) Plus précisément, pour Z ∈ Y on note ξZ la forme linéaire sur x,x′ ,⋆ C(∆p ) définie par ξZ (f ) = X (a1 ,...,ap ,S0 ,...,Sm ,(Y j i )i∈{0,...,m},j∈{1,...,li } )∈(π p,k,m,(l0 ,...,lm ) x,x′ ,⋆ )−1 (Z ) f (a1 , ..., ap ). Puis pour f ∈ C(∆p ) on pose B−(m+Pm Hx,x′ ,⋆,s (∆p ) = Xk ,m,l0 ,...,lm∈N i=0 li ) XZ ∈Y kf k2 p,k,m,(l0 ,...,lm ) x,x′ ,⋆ m si (Z )−li (cid:17)♯(cid:0)(π p,k ,m,(l0 ,...,lm) e2s(r0 (Z )−k)(cid:16) )−1 (Z )(cid:1)−α (cid:12)(cid:12)ξZ (f )(cid:12)(cid:12) 2 Yi=0 x,x′ ,⋆ Lemme 4.55 En notant C la constante qui apparaît dans le b) du lemme 4.54, on a pour tout f ∈ C(∆p ) , Hx,s (∆p ) ≤ C (2d(x, x′ ) + 1)kf k2 kf k2 Hx,x′,⋆,s (∆p ) . Démonstration. Cela résulte immédiatement du lemme 4.54, de l’inéga- p,k ,m,(l0 ,...,lm) et Z ∈ lité de Cauchy-Schwarz et du fait que pour Z ∈ Y x p,k ,m,(l0 ,...,lm) antécédent de Z , c’est-à-dire que Y x,x′ ,⋆ (140) . (π p,k ,m,(l0 ,...,lm) x,x′ ,⋆ )−1( Z ) ⊂ (π p,k ,m,(l0 ,...,lm) x 138 )−1(Z ) on a évidemment ♯(cid:0)(π p,k ,m,(l0 ,...,lm) )−1 ( Z )(cid:1) ≤ ♯(cid:0)(π p,k ,m,(l0 ,...,lm) )−1(Z )(cid:1). (cid:3) x x,x′ ,⋆ Démonstration de la proposition 4.3 en admettant le lemme 4.56. Grâce au lemme 4.55, pour montrer la proposition 4.3 on est ramené à mon- trer le lemme suivant. (cid:3) Lemme 4.56 Il existe une constante C = C (δ, K, N , Q, P , M , s, B ), tel que pour tout f ∈ C(∆p ) , Hx,x′ ,⋆,s (∆p ) ≤ C e3sd(x,x′ )kf k2 kf k2 Hx′ ,s (∆p ) . La preuve du lemme 4.56 occupe toute la suite de ce sous-paragraphe. Afin de mieux comprendre le lemme suivant, on peut se référer au dessin pour les arbres, en remarquant que max(k , d(x,x′ )+r0 (Z )−r ′ 0 (Z ) ) vaut k dans les 2 deux premiers dessins et d(x, t) dans le troisième. Lemme 4.57 Soit (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x a) Il existe un unique m′ ∈ {0, ..., m} tel que d(x, x′ ) + r0(Z ) − r ′ 0 (Z ) 2 d(x, x′ ) + r0 (Z ) − r ′ 0 (Z ) 2 dmax (x, Si ) ≤ max(k , dmax (x, Si ) > max(k , et . ) + M pour i > m′ ) + M pour 1 ≤ i ≤ m′ . ) + M + 2P + δ. b) Pour i > m′ et j ∈ {1, ..., li} on a d(x, x′ ) + r0 (Z ) − r ′ 0 (Z ) dmax (x, Y j i ) ≤ max(k , 2 c) Soit J ⊂ {1, ..., lm′ } l’ensemble des j tels que d(x, x′ ) + r0(Z ) − r ′ 0(Z ) dmax (x, Y j m′ ) > max(k , 2 m′ } avec l ′ m′ = ♯J ∈ {0, ..., lm′ } et j1 < ... < jl′ On écrit J = {j1 , ..., jl′ m′ et on p,k ,m,(l0 ,...,lm) note k ′ = k ′ (Z ) où Z ∈ Y est tel que x,x′ i )i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) (a1 , . . . , ap , S0 , ..., Sm , (Y j x,x′ Alors ) + M + 2P + δ. )−1 (Z ). (a1 , . . . , ap , S0 , ..., Sm′ , (Y j i )i∈{0,...,m′−1},j∈{1,...,li } , (Y jλ m′ )λ∈{1,...,l′ m′ }) 139 . p,k ′ ,m′ ,(l0 ,...,lm′−1 ,l′ m′ ) appartient à Y x′ d) Pour tout z dans B (x, k + 2M ) ∪ [i∈{m′+1,...,m} i , M ) ∪ [j 6∈J [ B (Y j i∈{m′+1,...,m},j∈{1,...,li } ∪ B (Si , M ) B (Y j m′ , M ) B (Y j m′ , M ), B (Si , M ) ∪ i , M ) ∪ [j∈J B (Y j et pour tout z ′ dans [i∈{0,...,m′ } [ i∈{0,...,m′−1},j∈{1,...,li} géod(z , z ′ ) rencontre B (x′ , k ′ + 2M − 2δ ). e) Pour tout (a1 , . . . , ap , S0 , ..., Sm′ , ( Y j i )i∈{0,...,m′−1},j∈{1,...,li } , ( Y jλ m′ )λ∈{1,...,l′ m′ }) appartenant à la même classe que (a1 , . . . , ap , S0 , ..., Sm′ , (Y j i )i∈{0,...,m′−1},j∈{1,...,li } , (Y jλ m′ )λ∈{1,...,l′ m′ }) p,k ′ ,m′ ,(l0 ,...,lm′−1 ,l′ m′ ) , pour tout z dans x′ B (x, k + 2M ) ∪ [i∈{m′+1,...,m} [ i , M ) ∪ [j 6∈J B (Y j i∈{m′+1,...,m},j∈{1,...,li } B (Y j m′ , M ) B (Si , M ) dans Y ∪ et pour tout z ′ dans B ( Si , M ) ∪ [i∈{0,...,m′ } [ i∈{0,...,m′−1},j∈{1,...,li} géod(z , z ′ ) rencontre B (x′ , k ′ + 2M ). B ( Y j i , M ) ∪ [λ∈{1,...,l′ m′ } B ( Y jλ m′ , M ), dans Y Remarque. L’intérêt de e) est que la classe de (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } ) p,k ,m,(l0 ,...,lm) est déterminée par celle de x,x′ ,⋆ (a1 , . . . , ap , S0 , ..., Sm′ , (Y j i )i∈{0,...,m′−1},j∈{1,...,li } , (Y jλ m′ )λ∈{1,...,l′ m′ }) 140 p,k ′ ,m′ ,(l0 ,...,lm′−1 ,l′ m′ ) x′ dans Y et par la connaissance des parties “éliminées” i )i∈{m′+1,...,m},j∈{1,...,li } , (Y j (Si )i∈{m′+1,...,m} , (Y j m′ )j 6∈J , qui se trouvent en gros entre les boules B (x, k + 2M ) et B (x′ , k ′ + 2M ). Démonstration. Le a) et le b) résultent des lemmes 4.12 et 4.13. Comme la preuve de c) et d) est longue on commence par montrer que d) implique e). Cela résulte du sous-lemme suivant, appliqué à r = k ′ + 2M , et x′ au lieu de x. Sous-lemme 4.58 Soit r ≥ 2δ et z , z ′ , z ′ ∈ X tels que – géod(z , z ′ ) rencontre B (x, r − 2δ ), – pour tout y ∈ B (x, r), d(y , z ′ ) = d(y , z ′). Alors géod(z , z ′ ) rencontre B (x, r). z B (u, 2δ ) u u v • x z ′ z ′ B (x, r) Démonstration. Soit u ∈ géod(z , z ′ ) ∩ B (x, r − 2δ ). On veut montrer u ∈ géod(z , z ′ ). Supposons par l’absurde que cela ne soit pas vrai. Soit u ∈ B (z , d(z , u)) à distance minimale de z ′ . On a alors d( u, z ′ ) < d(u, z ′ ) = d(u, z ′ ) ≤ d( u, z ′ ) donc u 6∈ B (x, r) et en particulier d(u, u) ≥ 2δ + 1. Soit v ∈ géod(u, u) vérifiant d(u, v ) = δ + 1. Par (H 0 δ (z , u, v , u)) on a d(z , v ) ≤ δ ( z ′ , u, v , u)) on a d( z ′ , v ) < d( z ′ , u). Comme v ∈ B (x, r) on a d(z , u). Par (H 0 d(z ′ , v ) = d( z ′ , v ). On en déduit que d(z , v ) + d(z ′ , v ) < d(z , u) + d(z ′ , u) = d(z , z ′ ) ce qui est impossible. (cid:3) Suite de la démonstration du lemme 4.57. Il reste à montrer c) et d). Pour le faire on distingue trois cas comme dans le dessin pour les arbres et dans la démonstration du lemme 4.54. Premier cas. On suppose r0(Z ) ≤ k . 0 (Z ) + M Alors k ′ = r ′ 2 par (129). D’après le lemme 4.12 et la remarque qui suit la définition 4.1, on a m = 0 et l0 = 0, donc c) est évident. Comme 141 0 (Z ) + N ) on a B (S0 , M ) ⊂ B (x′ , k ′ + 2M − 2δ ) car M ≥ N + 2δ S0 ⊂ B (x′ , r ′ et d) en résulte, puisque pour z ′ comme dans d) on a z ′ ∈ B (x′ , k ′ + 2M − 2δ ). Deuxième cas. On suppose r0 (Z ) > k et r0 (Z )−r ′ 0 (Z )+d(x,x′ ) ≤ k . 2 Par (130) on a alors . M k ′ = r ′ 0 (Z ) − r0 (Z ) + k + 2 Soit u ∈ B (x, k) à distance minimale de S0 et b ∈ S0 à distance minimale de u. Pour la suite on note le fait suivant. Comme d(x, b) − d(x′ , b) + d(x, x′ ) ≤ δ (x′ , x, u, b)) implique 2k + N et u ∈ géod(x, b), (H 0 d(x′ , u) ≤ max(d(x, x′ ) − k , d(x′ , b) − d(x, b) + k) + δ ≤ d(x′ , b) − d(b, u) + N + δ et u ∈ (N + δ )-géod(x′ , b). On a k ′ ≥ d(x′ , b) − d(x, b) + k + M 2 − N = d(x′ , b) − d(u, b) + M 2 − N d’où, grâce à (141), (141) M 2 − 2N − δ (142) à ) = k , pour passer de k ′ ≥ d(x′ , u) + Comme max(k , d(x,x′ )+r0 (Z )−r ′ 0 (Z ) 2 (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } ) i )i∈{0,...,m′−1},j∈{1,...,li } , (Y jλ (a1 , . . . , ap , S0 , ..., Sm′ , (Y j m′ )λ∈{1,...,l′ m′ }) on enlève les Si et les Y j i pour i ≥ 1 vérifiant dmax (x, Si ) ≤ k + M et les Y j m′ vérifiant dmax (x, Y j m′ ) ≤ k + M + 2P + δ . Pour vérifier c) on va montrer que les conditions i), ii), iii) et iv) de la définition 4.1 sont satisfaites par (a1 , . . . , ap , S0 , ..., Sm′ , (Y j i )i∈{0,...,m′−1},j∈{1,...,li } , (Y jλ m′ )λ∈{1,...,l′ m′ }). On vérifie d’abord i). Soit i ∈ {0, ..., m′ − 1}, y ∈ Si+1 , a ∈ Si et x ∈ B (x, k) tels que – y ∈ 4δ -géod( x, a) et d(y , a) ∈]N − 2δ, QN ], – ou y ∈ F -géod( x, a) et d(y , a) ≥ Q F . On note α = 4δ ou F suivant le cas, de sorte que y ∈ α-géod( x, a). Comme d(x, Si+1) ≥ dmax (x, Si+1 ) − N et i + 1 ∈ {1, ..., m′}, on a d(x, y ) > k + M − N . (143) 142 x x x′ • u y a b x′ Pour vérifier i) il suffit de montrer qu’il existe x′ ∈ B (x′ , k ′ ) tel que y ∈ α-géod( x′ , a). Cela résulte du sous-lemme suivant appliqué à α = 4δ ou α = F et à β = M − N . On a F ≥ 4δ et on suppose F + 3δ+F 2 < M − N , ce qui est permis par (HM ). Les hypothèses du sous-lemme suivant sont donc satisfaites et cela termine la preuve de i). Sous-lemme 4.59 Soit α, β ∈ N vérifiant F + 3δ+α 2 < β . Soit a ∈ 2F -géod(u, b), x ∈ B (x, k) et y ∈ α-géod( x, a) tel que d(x, y ) ≥ k + β . Alors il existe x′ ∈ B (x′ , k ′ ) tel que y ∈ α-géod( x′ , a). Démonstration du sous-lemme 4.59. Comme a ∈ 2F -géod(b, u) et u ∈ δ -géod( x, b) par le lemme 4.9, on a a ∈ (2F + δ )-géod( x, b) par le lemme 3.3, et comme y ∈ α-géod( x, a) on en déduit y ∈ (2F + δ + α)-géod( x, b). On applique le lemme 4.6 à {x, u, b, x, y} avec l = 2 et x comme point base. Soit t le point de géod(Ψ x, Ψb) le plus proche de Ψy . Ψx Ψ x •Ψu Ψy t Ψb Grâce au a) du lemme 4.7 on a Ψy ∈ (2F + 3δ + α)-géod(Ψ x, Ψb), d’où d(Ψy , t) ≤ F + 3δ+α . De plus d(Ψx, Ψy ) = d(x, y ) ≥ k +β . Comme F + 3δ+α 2 < 2 β on a d(Ψx, t) > k = d(Ψx, Ψu) et comme d(Ψx, Ψ x) ≤ k on en déduit Ψu ∈ géod(Ψ x, Ψy ) et u ∈ 4δ -géod( x, y ) par le b) du lemme 4.7 . Soit x′ ∈ géod( x, y ) vérifiant d(y , x′) = min(d(y , u), d(y , x)). On a alors, δ (u, y , x′ , x)), d(u, x′) ≤ 5δ . Par (142) on a k ′ ≥ d(x′ , u) + M par (H 0 2 − 2N − δ 143 et on suppose 2N + 6δ ≤ M 2 , ce qui est permis par (HM ). Alors d(x′ , x′ ) ≤ k ′ et ceci termine la démonstration du sous-lemme 4.59. (cid:3) Fin de l’étude du deuxième cas. On vérifie maintenant ii). Soit i ∈ {1, ..., m′}. On a Si ⊂ 2F -géod(b, u) par le a) du lemme 4.14. On a dmax (x, Si ) > k +M par hypothèse. Soit a ∈ Si . On a alors d(x, a) > k + M − N d’où d(u, a) > M − N . Par (141) et le b) du lemme 3.3, on a u ∈ (2F + N + δ )-géod(x′ , a). D’où d(x′ , a) ≥ d(x′ , u) + d(u, a) − (2F + N + δ ) > d(x′ , u) + M − (2F + 2N + δ ). Or M k ′ ≤ d(x′ , b) − d(x, b) + k + 2 M = d(x′ , b) − d(b, u) + + N ≤ d(x′ , u) + 2 + N M 2 + N . D’où d(x′ , a) > k ′ + M 2 − (2F + 3N + δ ) ≥ k ′ + P car on suppose (2F + 3N + δ ) + P ≤ M 2 , ce qui est permis par (HM ). Cela achève la preuve de ii). La propriété iii) est immédiate. On vérifie maintenant iv). Soit j ∈ J . On commence par montrer d(x′ , Y j m′ ) ≥ k ′ + 3P . On rappelle que Y j m′ ⊂ 4P -géod(x, b) et d(x, Y j m′ ) ≥ k + M + P + δ . Soit y ∈ Y j m′ . On a alors d(b, y ) ≤ d(x, b) − d(x, y ) + 4P ≤ d(b, x) − k − M − δ + 3P et d(x′ , y ) ≥ d(x′ , b) − d(b, y ) ≥ d(x′ , b) − d(x, b) + k + M + δ − 3P M ≥ k ′ + 2 − N + δ − 3P ≥ k ′ + 3P . car on suppose M 2 − N + δ − 3P ≥ 3P , ce qui est permis par (HM ). Il reste donc à montrer que pour tout y ∈ Y j il existe a ∈ Sm′ et m′ x′ ∈ B (x′ , k ′ ) tels que y ∈ 2P -géod( x′ , a). Soit y ∈ Y j m′ . On va distinguer deux cas. On suppose d’abord m′ < m. Alors il existe a ∈ Sm′ et w ∈ Sm′+1 tels que y ∈ P -géod(w , a). Par le a) du lemme 4.14, w et a appartiennent 144 à 2F -géod(u, b). Comme u ∈ géod(x, b) le b) du lemme 3.3 montre u ∈ 2F -géod(x, w). Donc d(u, w) ≤ d(x, w) − k + 2F . y w a u b x x′ D’autre part d(x, w) ≤ dmax (x, Sm′+1) ≤ dmax (x, Sm′ ) ≤ d(x, a) + N et d(x, a)− k ≤ d(u, a). On en déduit d(u, w) ≤ d(u, a)+ (2F +N ). Le lemme 4.8 appliqué à (u, b, a, w) au lieu de (x, y , a, b) et à (2F , 2F , 2F + N ) au lieu de (α, β , ρ) montre alors w ∈ (6F + 2N + δ )-géod(u, a). On en déduit y ∈ (6F + 2N + δ + P )-géod(u, a) par le lemme 3.3 et donc y ∈ 2P -géod(u, a) car on suppose P ≥ 6F + 2N + δ grâce à (HP ). On suppose M 2 ≥ 2N + δ grâce à (HM ). Par (142) on a alors u ∈ B (x′ , k ′ ). En prenant x′ = u on a fini. On suppose maintenant m′ = m. Il existe a ∈ Sm et x ∈ B (x, k) tels que y ∈ 2P -géod( x, a). On veut montrer qu’il existe x′ ∈ B (x′ , k ′ ) tel que y ∈ 2P -géod( x′ , a). On a d(x, y ) ≥ k + M + P + δ . On applique le sous- lemme 4.59 avec α = 2P et β = M + P + δ . On suppose F + 3δ+α 2 < β , ce qui est permis par (HM ). Il reste à montrer d). Le premier ensemble de d) est inclus dans B (x, k + 2M +2P +δ ) et le second ensemble de d) est inclus dans (2M +4P )-géod(u, b). On a B (u, 5M 2 − 2N − 3δ ) ⊂ B (x′ , k ′ + 2M − 2δ ) grâce à (142). Il suffit donc de montrer que pour z ∈ B (x, k + 2M + 2P + δ ) et z ′ ∈ (2M + 4P )-géod(u, b), géod(z , z ′ ) rencontre B (u, 5M 2 − 2N − 3δ ). Par (H 0 δ (z , x, u, b)), on a u ∈ δ -géod(x, z) ou u ∈ δ -géod(z , b). Si u ∈ δ -géod(x, z), d(u, z) ≤ 2M + 2P + 2δ , donc z ∈ géod(z , z ′ ) appartient à B (u, 5M 2 − 2N − 3δ ) car on suppose M 2 ≥ 2P + 2N + 5δ grâce à (HM ). Si u ∈ δ -géod(z , b), comme z ′ ∈ (2M + 4P )-géod(u, b), on a u ∈ (2M + 4P + δ )-géod(z , z ′ ) par le b) du lemme 3.3, donc t ∈ géod(z , z ′ ) vérifiant d(z , t) = min(d(z , u), d(z , z ′ )) δ (u, z , t, z ′ )), satisfait, grâce à (H 0 d(u, t) ≤ 2M + 4P + 2δ ≤ 5M 2 − 2N − 3δ 145 + M N 2 . (144) car on suppose 2M + 4P + 2δ ≤ 5M 2 − 2N − 3δ grâce à (HM ). Cela termine la preuve de d) et donc l’étude du deuxième cas. Troisième cas. On suppose r0(Z ) > k et r0 (Z )−r ′ 0 (Z )+d(x,x′ ) > k . 2 Par (131) on a alors 0(Z ) − r0(Z ) + d(x, x′ ) r ′ M (cid:12)(cid:12)k ′ − (cid:0)E ( 2 (cid:1)(cid:12)(cid:12) ≤ 2 ) = d(x,x′ )+r0 (Z )−r ′ Comme max(k , d(x,x′ )+r0 (Z )−r ′ 0 (Z ) 0 (Z ) , pour passer de 2 2 (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } ) (a1 , . . . , ap , S0 , ..., Sm′ , (Y j i )i∈{0,...,m′−1},j∈{1,...,li } , (Y jλ à m′ )λ∈{1,...,l′ m′ }) on enlève les Si et les Y j i pour i ≥ 1 vérifiant r0(Z ) − r ′ 0(Z ) + d(x, x′ ) dmax (x, Si ) ≤ 2 ainsi que les Y j m′ vérifiant 0 (Z ) + d(x, x′ ) r0 (Z ) − r ′ dmax (x, Y j m′ ) ≤ 2 Soit u ∈ B (x, k) à distance minimale de S0 et b ∈ S0 à distance minimale de u. Il résulte de (144) que d(x′ , b) − d(x, b) + d(x, x′ ) M k ′ ≥ 2 − N + 2 d(x′ , b) − d(x, b) + d(x, x′ ) M et k ′ ≤ + N + 2 2 Pour vérifier c) on va montrer que les conditions i), ii), iii) et iv) de la définition 4.1 sont satisfaites par (a1 , . . . , ap , S0 , ..., Sm′ , (Y j i )i∈{0,...,m′−1},j∈{1,...,li } , (Y jλ m′ )λ∈{1,...,l′ m′ }). On vérifie d’abord i). Soit i ∈ {0, ..., m′ − 1}, y ∈ Si+1 , a ∈ Si et x ∈ B (x, k) tels que – y ∈ 4δ -géod( x, a) et d(y , a) ∈]N − 2δ, QN ], – ou y ∈ F -géod( x, a) et d(y , a) ≥ Q F . On note α = 4δ ou F suivant le cas, de sorte que y ∈ α-géod( x, a). Comme d(x, Si+1) ≥ dmax (x, Si+1 ) − N et i + 1 ∈ {1, ..., m′}, on a 0 (Z ) + d(x, x′ ) r0 (Z ) − r ′ + M − N d(x, y ) ≥ 2 d(x, b) − d(x′ , b) + d(x, x′ ) 3N ≥ + M − . 2 2 + M + 2P + δ. (145) (146) (147) ) + 146 Pour vérifier i) il suffit de montrer qu’il existe x′ ∈ B (x′ , k ′ ) tel que y ∈ α-géod( x′ , a). x x x′ y a u b x′ Cela résulte du sous-lemme suivant appliqué à α = 4δ ou α = F et à β = M − 3N 2 . On a F ≥ 4δ et on suppose F + 3δ+F 2 + δ + N ≤ M − 3N 2 , ce qui est permis par (HM ). Les hypothèses du sous-lemme suivant sont donc satisfaites et cela termine la preuve de i). Sous-lemme 4.60 Soit α, β ∈ N vérifiant F + 3δ+α 2 + δ + N ≤ β . Soit a ∈ 2F -géod(u, b), x ∈ B (x, k) et y ∈ α-géod( x, a) tel que d(x, b) − d(x′ , b) + d(x, x′ ) d(x, y ) ≥ + β . 2 Alors il existe x′ ∈ B (x′ , k ′ ) tel que y ∈ α-géod( x′ , a). Démonstration du sous-lemme 4.60. Comme a ∈ 2F -géod(b, u) et u ∈ δ -géod( x, b) par le lemme 4.9, on a a ∈ (2F + δ )-géod( x, b) par le lemme 3.3, et comme y ∈ α-géod( x, a) on en déduit y ∈ (2F + δ + α)-géod( x, b). On applique le lemme 4.6 à {x, b, x′ , x, y} avec l = 2 et x comme point base. Soit t ∈ géod(Ψ x, Ψb) à distance minimale de Ψy . Comme y ∈ (2F + δ + α)-géod( x, b), on a Ψy ∈ (2F + 3δ + α)-géod(Ψ x, Ψb) par le a) du lemme 4.7, donc d(Ψy , t) ≤ F + 3δ+α . Or 2 d(Ψx, Ψy ) = d(x, y ) ≥ N ≥ k − + β > k + (F + 2 d(x, b) − d(x′ , b) + d(x, x′ ) 2 3δ + α 2 + β ) 147 par hypothèse. Donc d(Ψx, t) > k et comme d(Ψx, Ψ x) ≤ k , t ∈ géod(Ψx, Ψb). Ψx Ψ x t t′ Ψy t Ψb ). (148) Ψx′ De ce qui précède on retient aussi que d(x, b) − d(x′ , b) + d(x, x′ ) 3δ + α d(Ψx, t) ≥ + β − (F + 2 2 Soit t′ le point de géod(Ψx, Ψb) à distance minimale de Ψx′ . On a donc d(Ψx, Ψb) − d(Ψx′ , Ψb) + d(Ψx, Ψx′ ) d(Ψx, t′ ) = 2 d(x, b) − d(x′ , b) + d(x, x′ ) d(x, b) − d(x′ , b) + d(x, x′ ) ∈ [ 2 2 par le lemme 4.6. Comme F + 3δ+α 2 + δ + N ≤ β par hypothèse, il ré- sulte de (148) et (149) que d(Ψx, t′ ) ≤ d(Ψx, t). Par ailleurs d(Ψx, t′ ) ≥ d(x,b)−d(x′ ,b)+d(x,x′ ) > k − N . Si t désigne le point de géod(Ψx, Ψb) à dis- 2 tance minimale de Ψ x, comme d(Ψx, Ψ x) ≤ k , on a d(Ψx, t) ≤ k , donc d(Ψx, t′ ) ≥ d(Ψx, t) − N (dans le dessin ci-dessus t′ pourrait être entre Ψx et t, mais dans ce cas à distance ≤ N de t). En tous cas on a d(t′ , géod(t, t)) ≤ N , d’où d(t′ , géod(Ψ x, Ψy )) ≤ N . On en déduit d(Ψx′ , géod(Ψ x, Ψy )) ≤ d(Ψx′ , t′ ) + d(t′ , géod(Ψ x, Ψy )) ≤ d(Ψx′ , t′ ) + N d(Ψx, Ψx′ ) + d(Ψx′ , Ψb) − d(Ψx, Ψb) d(x, x′ ) + d(x′ , b) − d(x, b) + N ≤ 2 2 (149) + δ ] , = + N . Par le c) du lemme 4.7, d(x′ , géod( x, y )) ≤ d(Ψx′ , géod(Ψ x, Ψy )) + 3δ + 1 d(x, x′ ) + d(x′ , b) − d(x, b) + N + 3δ + 1 ≤ k ′ ≤ 2 grâce à (145) et car on suppose 2N + 3δ + 1 ≤ M 2 , ce qui est permis par (HM ). Il existe donc x′ ∈ géod( x, y ) ∩ B (x′ , k ′ ). On a alors y ∈ α-géod( x′ , a). (cid:3) Fin de l’étude du troisième cas. On vérifie maintenant ii). Soit i ∈ {1, ..., m′}. On a Si ⊂ 2F -géod(b, u) ⊂ 2F -géod(x, b) 148 par le a) du lemme 4.14. On a dmax (x, Si ) > r0 (Z )−r ′ 0 (Z )+d(x,x′ ) 2 thèse. Soit a ∈ Si . On a donc r0 (Z ) − r ′ 0 (Z ) + d(x, x′ ) + M − N d(x, a) > 2 d(x, b) − d(x′ , b) + d(x, x′ ) + M − 2N . ≥ 2 Donc d(b, a) ≤ d(x, b) − d(x, a) + 2F < d(x,b)+d(x′ ,b)−d(x,x′ ) 2 et + M par hypo- (150) − M + (2N + 2F ) d(x′ , a) ≥ d(x′ , b) − d(b, a) > Grâce à (146) on en déduit d(x′ , b) + d(x, x′ ) − d(x, b) 2 + M − (2N + 2F ). d(x′ , a) > k ′ + M 2 − (3N + 2F ) ≥ k ′ + P car on suppose M 2 ≥ (3N + 2F ) + P , ce qui est permis par (HM ). Ceci achève la preuve de ii). La propriété iii) est immédiate. On vérifie maintenant iv). On note b et u comme dans la preuve de i). Soit j ∈ J . On commence par montrer d(x′ , Y j m′ ) ≥ k ′ + 3P . On rappelle que Y j m′ ⊂ 4P -géod(x, b) et r0(Z ) − r ′ 0(Z ) + d(x, x′ ) d(x, Y j m′ ) ≥ 2 d(x, b) − d(x′ , b) + d(x, x′ ) ≥ 2 + M + P + δ − + M + P + δ N 2 . (151) Soit y ∈ Y j m′ . On a alors, grâce à (151), d(x, b) + d(x′ , b) − d(x, x′ ) 2 d(b, y ) ≤ d(x, b)− d(x, y )+ 4P ≤ et −M + 3P − δ + N 2 d(x′ , b) − d(x, b) + d(x, x′ ) 2 d(x′ , y ) ≥ d(x′ , b) − d(b, y ) ≥ 3N M 2 ≥ k ′ + 3P ≥ k ′ + + δ − 3P − 2 où l’avant-dernière inégalité a lieu par (146) et où, pour la dernière, on a 2 + δ − 3P − 3N supposé M 2 ≥ 3P , ce qui est permis par (HM ). 149 + M + δ − 3P − N 2 Il reste donc à montrer que pour tout y ∈ Y j il existe a ∈ Sm′ et m′ x′ ∈ B (x′ , k ′ ) tels que y ∈ 2P -géod( x′ , a). Soit y ∈ Y j m′ . On va distinguer deux cas. On suppose d’abord m′ < m. Alors il existe a ∈ Sm′ et w ∈ Sm′+1 tels que y ∈ P -géod(w , a). On a w , a ∈ 2F -géod(u, b). D’autre part d(x, w) ≤ dmax (x, Sm′+1) ≤ dmax (x, Sm′ ) ≤ d(x, a) + N . Le lemme 4.8 appliqué à (x, b, a, w) au lieu de (x, y , a, b) et (2F , 2F , N ) au lieu de (α, β , ρ) montre alors w ∈ (2F + 2N + δ )-géod(x, a). y x w u v x′ a b δ (x′ , x, v , b)) ). Par (H 0 et Soit v ∈ géod(x, b) vérifiant d(x, v ) = E ( d(x,b)+d(x,x′ )−d(x′ ,b) 2 on a v ∈ géod(x, b) ∩ (δ + 1)-géod(x, x′ ) ∩ (δ + 1)-géod(x′ , b) d(x′ , b) − d(x, b) + d(x, x′ ) d(x′ , v ) ≤ + δ + 1 ≤ k ′ 2 où la dernière égalité a lieu grâce à (145) car on suppose N + δ + 1 ≤ M 2 , ce qui est permis par (HM ). On a donc v ∈ B (x′ , k ′ ). Comme y ∈ P -géod(w , a) et w ∈ (2F + 2N + δ )-géod(x, a) on a y ∈ (2F + 2N + δ + P )-géod(x, a). Par (150) on a d(x, a) ≥ d(x, v ) + M − 2N , et comme a ∈ 2F -géod(x, b) on en déduit d(b, a)−d(b, v ) ≤ −M +2N +2F < −δ car on suppose −M + 2N + 2F < −δ grâce à (HM ). Alors (H 0 δ (a, x, v , b)) donne v ∈ δ -géod(x, a). Grâce à (151) on a d(x, b) − d(x′ , b) + d(x, x′ ) d(x, y ) ≥ 2 N ≥ d(x, v ) − 2 (v , x, y , a)) on a + M + P + δ − + M + P + δ. Par (H 2F +2N +δ+P δ d(v , y ) ≤ max(d(v , x) − d(x, y ), d(v , a) − d(a, y )) + 2F + 2N + 2δ + P . 150 N 2 (152) Or d(v , x) − d(x, y ) + 2F + 2N + 2δ + P < 0 par (152) et car on suppose 2F + 2N + 2δ + P < − N 2 + M + P + δ , ce qui est permis par (HM ). D’où y ∈ (2F +2N +2δ+P )-géod(v , a). Grâce à (HP ) on suppose P ≥ 2F +2N +2δ , d’où y ∈ 2P -géod(v , a). En prenant x′ = v on a fini. On suppose maintenant m′ = m. Il existe a ∈ Sm′ et x ∈ B (x, k) tels que y ∈ 2P -géod( x, a). On veut montrer qu’il existe x′ ∈ B (x′ , k ′ ) tel que y ∈ 2P -géod( x′ , a). Grâce à (151) on peut appliquer le sous-lemme 4.60 avec α = 2P et β = M + P + δ − N 2 . On suppose F + 3δ+α 2 + δ + N ≤ β , ce qui est permis par (HM ). Donc les hypothèses du sous-lemme 4.60 sont satisfaites et l’existence de x′ est démontrée. Il reste à montrer d). Soit v ∈ géod(x, b) vérifiant d(x, b) + d(x, x′ ) − d(x′ , b) d(x, v ) = E ( 2 δ (x′ , x, v , b)) on a v ∈ géod(x, b)∩(δ +1)-géod(x, x′ )∩(δ +1)-géod(x′ , b) Par (H 0 et d(x′ , b) − d(x, b) + d(x, x′ ) d(x′ , v ) ≤ 2 Il résulte alors de (145) que k ′ ≥ d(x′ , v ) + M 2 − (N + δ + 1), donc 5M 2 − (N + 3δ + 1)) ⊂ B (x′ , k ′ + 2M − 2δ ). B (v , Le premier ensemble de d) est inclus dans d(x, x′ ) + r0 (Z ) − r ′ 0(Z ) 2 +2M +2P +δ ) ⊂ B (x, d(x, v )+2M +2P +N +δ ). + δ + 1. ). B (x, Si + M (153) d(x, x′ ) + r0(Z ) − r ′ 0(Z ) r0 (Z ) ≤ 2 on a k ′ ≥ r ′ 0 (Z )−r0 (Z )+d(x,x′ ) 0 (Z ) − M + M 2 − N ≥ r ′ 2 − N , d’où B (S0 , M ) ⊂ 2 B (x′ , k ′ + 2M ). Grâce au d) du lemme 4.14 le deuxième ensemble de d) (privé dans le cas où (153) a lieu, de B (S0 , M ) que l’on a déjà traité dans ce cas), est inclus dans d(x, x′ ) + r0 (Z ) − r ′ 0 (Z ) {z ′ ∈ (2M + 4P )-géod(x, b), d(x, z ′ ) ≥ − N − 4P }. 2 Comme d(x,x′ )+r0 (Z )−r ′ 0 (Z ) − N − 4P ≥ d(x, v ) − 2N − 4P , il suffit de montrer 2 que pour z ∈ B (x, d(x, v ) + 2M + 2P + N + δ ) z ′ ∈ (2M + 4P )-géod(x, b) vérifiant d(x, z ′ ) ≥ d(x, v ) − 2N − 4P , et 151 géod(z , z ′ ) rencontre B (v , 5M 2 − (N + 3δ + 1)). Par (H 0 δ (z , x, v , b)), on a v ∈ δ -géod(x, z) ou v ∈ δ -géod(z , b). Si v ∈ δ -géod(x, z), d(v , z) ≤ 2M + 2P + N + 2δ , donc z ∈ géod(z , z ′ ) 2 − (N + 3δ + 1)) car on suppose M appartient à B (v , 5M 2 ≥ (2P + N + 2δ ) + (N + 3δ + 1) grâce à (HM ). Supposons donc v ∈ δ -géod(z , b). z z ′ x Comme v b z ′ ∈ (2M + 4P )-géod(x, b) et d(x, z ′ ) ≥ d(x, v ) − 4P − 2N , on a d(b, z ′ ) ≤ d(b, v ) + 2M + 8P + 2N . Le lemme 4.8 appliqué à (b, x, v , z ′ ) au lieu de (x, y , a, b) et (0, 2M + 4P , 2M + 8P + 2N ) au lieu de (α, β , ρ) donne z ′ ∈ (4M + 16P + 4N + δ )-géod(v , b). Comme v ∈ δ -géod(z , b) le b) du lemme 3.3 montre v ∈ (4M + 16P + 4N + 2δ )-géod(z , z ′ ). Soit t ∈ géod(z , z ′ ) vérifiant d(z , t) = E ( d(z ,z ′ )+d(z ,v)−d(z ′ ,v) δ (v , z , t, z ′ )) ). Alors (H 0 2 montre d(v , t) ≤ 2M + 8P + 2N + 3δ + 1 ≤ 5M 2 − (N + 3δ + 1) car on sup- pose 2M + 8P + 2N + 3δ + 1 ≤ 5M 2 − (N + 3δ + 1) grâce à (HM ). Cela termine la preuve de d) et donc l’étude du troisième cas. On a donc montré le lemme 4.57. (cid:3) Lemme 4.61 Soit (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x Dans les notations du lemme 4.57, les entiers k ′ , m′ , l ′ m′ , la partie J ⊂ {1, ..., lm′ } et l’image Z ′ de . dans Y Z de i )i∈{0,...,m′−1},j∈{1,...,li } , (Y jλ (a1 , . . . , ap , S0 , ..., Sm′ , (Y j m′ }) m′ )λ∈{1,...,l′ p,k ′ ,m′ ,(l0 ,...,lm′−1 ,l′ p,k ′,m′ ,(l0 ,...,lm′−1 ,l′ m′ ) m′ ) par π x′ x′ ne dépendent que de l’image (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } ) 152 p,k ,m,(l0 ,...,lm) par π p,k ,m,(l0 ,...,lm) . dans Y x,x′ ,⋆ x,x′ ,⋆ De plus il existe L ne dépendant que de Z tel que tout élément de Z ′ a L antécédents par l’application de (154) )−1(Z ′ ) p,k ′ ,m′ ,(l0 ,...,lm′−1 ,l′ m′ ) (π p,k ,m,(l0 ,...,lm) )−1(Z ) dans (π x,x′ ,⋆ x′ (a1 , . . . , ap , S0 , ..., Sm , (Y j qui à i )i∈{0,...,m},j∈{1,...,li } ) associe i )i∈{0,...,m′−1},j∈{1,...,li } , (Y jλ (a1 , . . . , ap , S0 , ..., Sm′ , (Y j m′ )λ∈{1,...,l′ m′ }) et il existe C = C (δ, N , K, Q, P , M ) tel que l’on ait toujours 1 ≤ L ≤ C (m+Pm i=m′ li )−(m′ +l′ m′ ) . Démonstration. La première partie du lemme est évidente. L’existence de L vient du e) du lemme 4.57, qui est une propriété très importante car elle “découple” les parties éliminées et les parties conservées en rendant superflue la connaissance des distances entre les points à distance ≤ M d’une partie éliminée et ceux à distance ≤ M d’une partie conservée. De plus L est égal au nombre de possibilités pour les parties éliminées (dont la liste est rappelée dans (155) ci-dessous) telles que les distances entre les points de B (x, k+2M ), B (x′ , k ′ + 2M ) et ceux à distance ≤ M des parties éliminées prennent les va- leurs prescrites par la donnée de Z . Il reste à montrer (154). Les cardinaux des parties Si et Y j i sont bornés par une constante de la forme C (δ, K, N , Q, P ). Comme le nombre des parties éliminées est (m + Pm i=m′ li ) − (m′ + l ′ m′ ), il suffit de montrer que tout point y d’une partie Si ou Y j i éliminée ne peut prendre que C positions avec C = C (δ, K, N , Q, P , M ). Soit donc i ∪ [j 6∈J [ y ∈ [i∈{m′+1,...,m} Y j i∈{m′+1,...,m},j∈{1,...,li } Soit b ∈ S0 et u ∈ B (x, k) à distance minimale de b. On a alors y ∈ 4P -géod(x, b) par le a) et le c) du lemme 4.14, et Si ∪ Y j m′ . (155) d(x, x′ ) + r0 (Z ) − r ′ 0 (Z ) d(x, y ) ≤ max(k , 2 Si d(x, y ) ≤ k + M + 2P + δ on a y ∈ B (x, k + 2M ) car on peut supposer M ≥ 2P + δ par (HM ) donc y était déjà déterminé par la donnée de Z . Supposons donc ) + M + 2P + δ. d(x, y ) ≤ d(x, x′ ) + r0(Z ) − r ′ 0(Z ) 2 + M + 2P + δ. 153 Cela implique immédiatement d(x, x′ ) + d(x, b) − d(x′ , b) 2 (156) + M + 2P + δ + N . d(x, y ) ≤ On a alors d(x′ , y ) ≤ max(d(x′ , x) − d(b, x) + d(y , b), d(x′ , b) − d(b, x) + d(x, y )) + δ ≤ max(d(x′ , x) − d(x, y ) + 4P , d(x′ , b) − d(b, x) + d(x, y )) + δ ≤ d(x′ , x) − d(x, y ) + 2M + 4P + 2N + 3δ où la première inégalité vient de (Hδ (x′ , x, y , b)), la deuxième utilise le fait que y ∈ 4P -géod(x, b) et la dernière résulte de (156). On en déduit y ∈ (2M + 4P + 2N + 3δ )-géod(x, x′ ). Comme d(x, y ) fait partie de la donnée de Z , le lemme 3.13 montre que le nombre de possibilités pour y est borné par une constante C = C (δ, N , K, Q, P , M ). (cid:3) Dans les notations du lemme 4.61 on désigne par θ : [k ,m,l0 ,...,lm → [k ′ ,m′ ,l′ 0 ,...,l′ m′ l’application qui à (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) associe (a1 , . . . , ap , S0 , ..., Sm′ , (Y j i )i∈{0,...,m′−1},j∈{1,...,li } , (Y jλ m′ )λ∈{1,...,l′ m′ }) et on note p,k ′ ,m′ ,(l′ 0 ,...,l′ m′ ) x′ Y p,k ,m,(l0 ,...,lm) x,x′ ,⋆ Y p,k ′ ,m′ ,(l′ 0 ,...,l′ m′ ) x′ Y Y p,k ,m,(l0 ,...,lm) x,x′ ,⋆ ξZ ′ (f ) = p,k ′ ,m′ ,(l′ 0 ,...,l′ m′ ) x′ on note ξZ ′ la forme → [k ′ ,m′ ,l′ θ : [k ,m,l0 ,...,lm 0 ,...,l′ m′ l’application induite par θ, qui à Z associe Z ′ et dont l’existence résulte du lemme 4.61. m′ ∈ N et Z ′ ∈ Y 0 , ..., l ′ Pour k ′ , m′ , l ′ linéaire sur C(∆p ) définie par X ′ j i } )∈(π i )i∈{0,...,m′ },j∈{1,...,l′ Grâce au lemme 4.61, on a pour Z ′ = θ(Z ) et f ∈ C(∆p ) , )−1 (Z )(cid:1)−α (cid:12)(cid:12)ξZ (f )(cid:12)(cid:12) 2 ♯(cid:0)(π p,k ,m,(l0 ,...,lm) x,x′ ,⋆ 0 ,...,l′ p,k ′ ,m′ ,(l′ m′ ) )−1(Z ′ )(cid:1)−α (cid:12)(cid:12)ξZ ′ (f )(cid:12)(cid:12) 2 = (cid:0)L2−α (cid:1) ♯(cid:0)(π x′ 154 0 ,...,l′ p,k′ ,m′ ,(l′ m′ ) x′ 0 ,...,S ′ p ,S ′ (a′ 1 ,...,a′ m′ ,(Y 1 , ..., a′ f (a′ p). )−1 (Z ′ ) (157) avec L ≤ C (m+Pm i=m′ li )−(m′ +l′ m′ ) , où C est la constante du lemme 4.61. Il en résulte que pour tout f ∈ C(∆p ) , on a kf k2 Hx,x′ ,⋆,s (∆p ) ≤ ,...,l′ m′ ) p,k′ ,m′ ,(l′ 0 x′ k ,m,l0 ,...,lm∈N,Z ∈Y B−(m+Pm i=0 li ) X k ′ ,m′ ,l′ 0 ,...,l′ m′ ∈N,Z ′∈Y X p,k,m,(l0 ,...,lm ) tels que θ(Z )=Z ′ x,x′ ,⋆ m si (Z )−li (cid:17)! m′ )e2s(r0 (Z )−k)(cid:16) (C 2−α )(m+Pm i=m′ li )−(m′ +l′ Yi=0 p,k ′ ,m′ ,(l′ 0 ,...,l′ m′ ) )−1(Z ′ )(cid:1)−α (cid:12)(cid:12)ξZ ′ (f )(cid:12)(cid:12) 2 . ♯(cid:0)(π x′ B−(m′+Pm′ i ) e2s(r0 (Z ′ )−k ′ ) i=0 l′ D’autre part p,k′ ,m′ ,(l′ 0 x′ kf k2 Hx′ ,s (∆p ) = X ,...,l′ m′ ) k ′ ,m′ ,l′ 0 ,...,l′ m′ ∈N,Z ′∈Y m′ 0 ,...,l′ p,k ′ ,m′ ,(l′ (cid:16) i (cid:17)♯(cid:0)(π m′ ) )−1(Z ′ )(cid:1)−α (cid:12)(cid:12)ξZ ′ (f )(cid:12)(cid:12) 2 si (Z ′ )−l′ Yi=0 x′ où les entiers r0 (Z ′ ), s0(Z ′ ), ..., sm′ (Z ′ ) sont tels que pour tout i }) ∈ (π p,k ′ ,m′ ,(l′ 0 ,...,l′ ′ j m) p , S ′ 1 , . . . , a′ (a′ 0 , ..., S ′ m′ , (Y i )i∈{0,...,m′ },j∈{1,...,l′ x′ on a r0(Z ′ ) = d(x′ , S ′ 0), si (Z ′ ) = d(S ′ i , S ′ i+1) + 2M pour i ∈ {0, ..., m′ − 1}, m′ ) − k ′ . On remarque que pour Z ′ = θ(Z ) on a r0 (Z ′ ) = et sm′ (Z ′ ) = d(x′ , S ′ 0 (Z ) et pour i ∈ {0, ..., m′ − 1}, l ′ i = li et si (Z ′ ) = si (Z ). r ′ Démonstration du lemme 4.56 en admettant le lemme 4.62. On pose C0 = C 2−α où C est comme dans le lemme 4.61. Donc C0 ≤ C (δ, N , K, Q, P , M ). Pour montrer le lemme 4.56 on est ramené à montrer le lemme suivant. (cid:3) )−1(Z ′ ) k ,m,l0 ,...,lm∈N,Z ∈Y Lemme 4.62 Il existe une constante C = C (δ, K, N , P , Q, M , s, B ) tel le que 0 ,...,l′ p,k ′ ,m′ ,(l′ m′ ) m′ et Z ′ ∈ Y pour tous k ′ , m′ , l ′ 0 , ..., l ′ on ait x′ X p,k,m,(l0 ,...,lm ) x,x′ ,⋆ si (Z )−li (cid:17) ≤ C e3sd(x,x′ ) (C0B−1 )m′+l′ m′ e2s(r0 (Z ′ )−k ′ )sm′ (Z ′ )−l′ m′ . 155 (C0B−1)m+Pm i=m′ li e2s(r0 (Z )−k) m Yi=m′ tels que θ(Z )=Z ′ (cid:16) Démonstration du lemme 4.62 en admettant lemme 4.63. On rappelle que pour Z ′ = θ(Z ), r0 (Z ′ ) = r ′ 0(Z ). Par (128) on a d(x, x′ ) − r0 (Z ) − r ′ 0 (Z ) 2 k ′ − r ′ 0(Z ) = min(0, max(k − r0 (Z ), E ( d’où M 2 ))) + )) + . M 2 d(x, x′ ) − r0(Z ) − r ′ 0(Z ) 2 k ′ − r ′ 0(Z ) ≤ max(k − r0(Z ), E ( Comme r0(Z ) − r ′ 0(Z ) ≤ d(x, x′ ) on a d(x, x′ ) − r0(Z ) − r ′ 0(Z ) ≤ d(x, x′ ) − r0 (Z ) ≤ d(x, x′ ) + (k − r0(Z )) 2 0(Z ) − k ′ ) + d(x, x′ ) + M d’où (r0(Z ) − k) ≤ (r ′ 2 . Les valeurs de k pour lesquelles p,k ,m,(l0 ,...,lm) tels que θ(Z ) = Z ′ sont donc il existe m, l0 , ..., lm ∈ N et Z ∈ Y x,x′ ,⋆ incluses dans un intervalle [k0 , +∞[ avec k0 ∈ N vérifiant M (r0(Z ) − k0) ≤ (r ′ 0(Z ) − k ′ ) + d(x, x′ ) + 2 Comme la série 1 + e−2s + e−4s + ... converge, pour montrer le lemme 4.62 il suffit d’établir le lemme suivant. (cid:3) . m′ et Z ′ ∈ Y 0 , ..., l ′ Lemme 4.63 Pour tous k , k ′ , m′ , l ′ p,k ′ ,m′ ,(l′ 0 ,...,l′ m′ ) x′ i=m′ li (cid:16) (C0B−1 )m+Pm X p,k,m,(l0 ,...,lm ) tels que θ(Z )=Z ′ x,x′ ,⋆ m′ sm′ (Z ′ )−l′ ≤ 4esd(x,x′ )(C0B−1 )m′+l′ m′ . Démonstration du lemme 4.63 en admettant le lemme 4.64. D’après le lemme 4.61, pour Z ′ = θ(Z ), θ induit une surjection si (Z )−li (cid:17) m Yi=m′ m,l0 ,...,lm∈N,Z ∈Y on a de (π p,k ,m,(l0 ,...,lm) x,x′ ,⋆ )−1(Z ) dans 0 ,...,l′ p,k ′ ,m′ ,(l′ m′ ) (π x′ )−1(Z ′ ). Pour montrer le lemme 4.63 il suffit donc d’établir le lemme suivant. (cid:3) 0 , ..., l ′ Lemme 4.64 Pour tous k , k ′ , m′ , l ′ m′ et (a′ 1 , . . . , a′ p , S ′ 0 , ..., S ′ m′ , (Y ′j i }) ∈ Y i )i∈{0,...,m′ },j∈{1,...,l′ 0 ,...,l′ p,k ′ ,m′ ,(l′ m′ ) x′ 156 on a (C0B−1)(m−m′ )+(Pm i=m′ li )−l′ m′ X m,l0 ,...,lm ,(a1 ,...,ap ,S0 ,...,Sm ,(Y j i )i∈{0,...,m},j∈{1,...,li } ) m−1 m′ ) − k ′(cid:1)l′ (cid:16) Yi=m′ (cid:0)d(Si , Si+1) + 2M (cid:1)−li (cid:17)(cid:0)d(x, Sm) − k(cid:1)−lm (cid:0)d(x′ , S ′ m′ ≤ 4esd(x,x′ ) où la somme porte sur les m, l0 , ..., lm ∈ N et les (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x tels que k ′ , m′ , l ′ 0 , ..., l ′ m′ et (158) 0 , ..., S ′ p , S ′ 1 , . . . , a′ (a′ m′ , (Y ′ j i )i∈{0,...,m′ },j∈{1,...,l′ i }) soient associés à (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } ) p = ap , S ′ 1 = a1 , ..., a′ comme dans le lemme 4.57 (donc en particulier a′ 0 = ′ j i = Y j i = li pour i ∈ {0, ..., m′ − 1}, Y m′ = Sm′ , k ′ = k ′ (Z ), l ′ S0 , ..., S ′ i pour m′ = Y jλ i ∈ {0, ..., m′ − 1} et j ∈ {1, ..., li} et Y ′λ m′ pour λ ∈ {1, ..., l ′ m′ }). Démonstration du lemme 4.64 en admettant les lemmes 4.65 et ′ j 4.66. D’abord on suppose l ′ 0 = ... = l ′ m′−1 = 0 car en supprimant les Y i pour i ∈ {0, ..., m′ − 1} et j ∈ {1, ..., l ′ i} le nouvel énoncé est strictement équivalent à l’ancien. D’après le lemme 4.15 il existe une constante C1 = C (δ, K, N , Q, P ) telle que, connaissant S0 , ..., Sm , pour i ∈ {m′ + 1, ..., m} et j ∈ {1, ..., li} le nombre de possibilités pour Y j i est borné par C1 (d(Si , Si+1) + 2M ) si i < m et par C1(d(x, Sm) − k) si i = m. Comme (C1C0B−1 )Pm i=m′ +1 li = (1 − C1C0B−1 )−(m−m′ ) , Xlm′ +1 ,...,lm∈N pour montrer le lemme 4.64 il suffit de montrer les deux lemmes suivants, qui distinguent les cas m′ = m et m′ < m. Dans les deux lemmes on note l et l ′ au lieu de lm′ et l ′ m′ , puisque que les li et l ′ i pour i 6= m′ ont disparu. Dans les deux lemmes suivants on a supposé J = {1, ..., l ′} et remplacé la somme sur les parties J ⊂ {1, ..., l} de cardinal l ′ par la multiplication par (cid:0) l l′ (cid:1). (cid:3) 157 Lemme 4.65 Pour tous k , k ′ , m′ , l ′ ∈ N et m′ )j∈{1,...,l′ }) ∈ Y p,k ′ ,m′ ,(0,...,0,l′) (a1 , . . . , ap , S0 , ..., Sm′ , (Y j x′ ≤ 2 (159) on a l ′(cid:19) XY l′+1 Xl≥l′ (cid:18) l (C0B−1 )l−l′ (cid:0)d(x, Sm′ ) − k(cid:1)−l (cid:0)d(x′ , Sm′ ) − k ′(cid:1)l′ ,...,Y l m′ m′ où la somme porte sur les Y l′+1 , ..., Y l m′ tels que m′ m′ )j∈{1,...,l}) ∈ Y p,k ,m′ ,(0,...,0,l) (a1 , . . . , ap , S0 , ..., Sm′ , (Y j x et, en notant T = d(x,x′ )+d(x,S0 )−d(x′ ,S0 ) on ait 2 dmax (x, Sm′ ) > max(k , T ) + M si m′ > 0, dmax (x, Y j j ∈ {1, ..., l ′} et pour m′ ) > max(k , T ) + M + 2P + δ dmax (x, Y j j ∈ {l ′ + 1, ..., l}. m′ ) ≤ max(k , T ) + M + 2P + δ On rappelle que les constantes C0 et C1 (qui apparaissent avant les énon- cés des lemmes 4.62 et 4.65) sont ma jorées par C (δ, K, N , Q, P , M ). pour Lemme 4.66 Pour tous k , k ′ , m′ , l ′ ∈ N et m′ )j∈{1,...,l′ }) ∈ Y p,k ′ ,m′ ,(0,...,0,l′) (a1 , . . . , ap , S0 , ..., Sm′ , (Y j x′ on a (C0B−1)l−l′ C0B−1 1 − C0C1B−1 (cid:17)m−m′ l ′(cid:19)(cid:16) Xm>m′ ,l≥l′ (cid:18) l X Sm′ +1 ,...,Sm,Y l′+1 ,...,Y l m′ m′ (cid:0)d(Sm′ , Sm′+1 ) + 2M (cid:1)−l (cid:0)d(x′ , Sm′ ) − k ′ (cid:1)l′ ≤ 4esd(x,x′ ) où la deuxième somme porte sur les Sm′+1 , ..., Sm , Y l′+1 , ..., Y l m′ tels que m′ (a1 , . . . , ap , S0 , ..., Sm , (Y j m′ )j∈{1,...,l}) ∈ Y p,k ,m,(0,...,0,l,0,...,0) x et, en notant T = d(x,x′ )+d(x,S0 )−d(x′ ,S0 ) on ait 2 dmax (x, Sm′ ) > max(k , T ) + M si m′ > 0, dmax (x, Sm′+1) ≤ max(k , T ) + M , dmax (x, Y j j ∈ {1, ..., l ′} et pour m′ ) > max(k , T ) + M + 2P + δ dmax (x, Y j j ∈ {l ′ + 1, ..., l}. m′ ) ≤ max(k , T ) + M + 2P + δ (160) pour 158 Dans la démonstration des lemmes 4.65 et 4.66 le petit calcul suivant servira plusieurs fois. ≤ 2. Lemme 4.67 Soit A, B ∈ R+ , 2A + B ≤ 1. Alors pour tout l ′ ∈ N, l ′(cid:19)Al−l′ Xl≥l′ (cid:18) l B l′ Démonstration. On a la formule générale, pour x ∈ C avec x < 1 et k ∈ N, k (cid:19)xn = (1 − x)−(k+1) . Xn∈N (cid:18)n + k Bl′ l′(cid:1)Al−l′ B l′ = Donc Pl≥l′ (cid:0) l (1−A)l′ +1 ≤ 1 1−A ≤ 2 puisque B ≤ 1 − A et 1 − A ≥ 1/2. (cid:3) Démonstration du lemme 4.65. On va distinguer trois cas comme dans la démonstration du lemme 4.57, mais une partie de la démonstration est com- mune aux trois cas. On note T = d(x,S0 )−d(x′ ,S0 )+d(x,x′ ) comme dans l’énoncé 2 du lemme. On fixe b ∈ S0 . Soit y ∈ Sm′ . Alors d(x, b) + d(x, x′ ) − d(x′ , b) − N . d(x, y ) ≥ 2 En effet cela est vrai si m′ = 0 car d(x, y ) ≥ d(x, S0) ≥ d(x, b) − N et d(x, x′ ) ≤ d(x, b) + d(x′ , b), et cela est vrai si m′ > 0 car la condition dmax (x, Sm′ ) > max(k , T ) + M implique (161) (162) d(x, b) + d(x, x′ ) − d(x′ , b) d(x, y ) ≥ dmax (x, Sm′ )−N ≥ T +M −N ≥ 2 et on suppose M ≥ N , ce qui est permis par (HM ). On en déduit d(x′ , y ) − d(x′ , b) + d(x, b) ≤ d(x, y ) + P + 2N + δ. +M −2N (163) En effet d(x′ , y ) ≤ max(d(x′ , b) − d(x, b) + d(x, y ), d(x′ , x) − d(x, b) + d(b, y )) + δ ≤ max(d(x′ , b) − d(x, b) + d(x, y ), d(x′ , x) − d(x, y ) + P ) + δ ≤ d(x′ , b) − d(x, b) + d(x, y ) + P + 2N + δ où la première inégalité a lieu par (Hδ (x′ , x, y , b)), la deuxième inégalité utilise y ∈ P -géod(x, b) et la dernière inégalité résulte de (162). 159 Premier cas. On suppose d(x, S0) ≤ k . Alors m′ = 0 et l = 0 d’après le lemme 4.12 et la remarque qui suit la définition 4.1. Donc la somme est vide ou réduite à un élément et l’inégalité est triviale. Deuxième cas. On suppose d(x, S0 ) > k et T ≤ k . Alors pour j ∈ {l ′ + 1, ..., l} on a k + 3P ≤ dmax (x, Y j m′ ) ≤ k + M + 2P + δ où l’inégalité de gauche vient de la condition iv) de la définition 4.1. Comme Y j m′ ⊂ 4P -géod(x, b) le lemme 3.13 montre que le nombre de possibilités pour Y j m′ est borné par C2 = C (δ, K, N , Q, P , M ). On a k ′ = d(x′ , S0) − d(x, S0) + k + M 2 par (130) donc k ′ ≥ d(x′ , b) − d(x, b) + k + M 2 − N . (164) M 2 +P +3N +δ (165) (166) d’où M 2 + P + 3N + δ Soit y ∈ Sm′ . On a M d(x′ , y )−k ′ ≤ d(x′ , y )−d(x′ , b)+d(x, b)+N −k− 2 ≤ d(x, y )−k− où la première inégalité vient de (164) et la deuxième de (163). Comme cela est vrai pour tout y ∈ Sm′ on en déduit (cid:0)d(x′ , Sm′ ) − k ′ (cid:1) ≤ (cid:0)d(x, Sm′ ) − k(cid:1) − (cid:0)d(x′ , Sm′ ) − k ′(cid:1) ≤ (cid:0)d(x, Sm′ ) − k(cid:1) − 1 car on suppose M 2 ≥ P + 3N + δ + 1, ce qui est permis par (HM ). Le membre de gauche de (159) est donc ma joré par d(x, Sm′ ) − k (cid:17)(l−l′ )(cid:16) d(x′ , Sm′ ) − k ′ l ′(cid:19)(cid:16) C0C2B−1 d(x, Sm′ ) − k (cid:17)l′ Xl≥l′ (cid:18) l par le lemme 4.67. Les hypothèses du lemme 4.67 sont satisfaites car on suppose 2C0C2B−1 ≤ 1 grâce à (HB ) ce qui implique 2C2C0B−1 + (d(x′ , Sm′ ) − k ′ ) ≤ (cid:0)d(x′ , Sm′ ) − k(cid:1) + 1 ≤ (d(x, Sm′ ) − k). 160 ≤ 2 Troisième cas. On suppose d(x, S0) > k et T > k . Pour j ∈ {l ′ + 1, ..., l} on a k + 3P ≤ dmax (x, Y j m′ ) ≤ T + M + 2P + δ. Comme Y j m′ ⊂ 4P -géod(x, b) le lemme 3.13 montre que le nombre de possi- bilités pour Y j m′ est borné par C2 (T − k) avec C2 = C (δ, K, N , Q, P , M ). Par (131) on a d(x′ , S0) + d(x, x′ ) − d(x, S0) k ′ − ( 2 Soit y ∈ Sm′ . Il résulte de (163) que d(x′ , y ) − d(x′ , S0) + d(x, S0) ≤ d(x, y ) + P + 3N + δ. ) ≤ M 2 N 2 + 1. + (167) (168) Donc + + 1 N 2 M 2 − + (P + 7N/2 + δ + 1) d(x′ , S0) + d(x, x′ ) − d(x, S0) 2 M 2 d(x′ , y ) − k ′ ≤ d(x′ , y ) − ≤ d(x, y ) − T − où la première inégalité a lieu par (167) et la deuxième par (168). Comme cela est vrai pour tout y ∈ Sm′ on en déduit (d(x′ , Sm′ ) − k ′ ) M ≤ (d(x, Sm′ ) − k) − (T − k) − + (P + 7N/2 + δ + 1). 2 Le membre de gauche de (159) est alors ma joré par l ′(cid:19)(cid:16) C0C2B−1 (T − k) d(x, Sm′ ) − k (cid:17)(l−l′ )(cid:16) d(x′ , Sm′ ) − k ′ d(x, Sm′ ) − k (cid:17)l′ Xl≥l′ (cid:18) l par le lemme 4.67. Les hypothèses du lemme 4.67 sont satisfaites car on 2 ≥ P + 7N suppose 2C0C2B−1 ≤ 1 grâce à (HB ) et M 2 + δ + 1 grâce à (HM ) d’où (T − k) + (d(x′ , Sm′ ) − k ′ ) ≤ (d(x, Sm′ ) − k) grâce à (169). (cid:3) Démonstration du lemme 4.66. On distingue trois cas comme dans les démonstrations des lemmes 4.57 et 4.65. On note T = d(x,S0 )−d(x′ ,S0 )+d(x,x′ ) 2 comme dans l’énoncé du lemme 4.66. On fixe b ∈ S0 . En reprenant mot pour mot la preuve de (163), c’est-à-dire le début de la démonstration du lemme 4.65 jusqu’à la distinction des trois cas, on obtient (163), c’est-à-dire que pour tout y ∈ Sm′ , d(x′ , y ) − d(x′ , b) + d(x, b) ≤ d(x, y ) + P + 2N + δ. ≤ 2 (169) (170) 161 Premier cas. On suppose d(x, S0) ≤ k . Alors m = 0 d’après le lemme 4.12, et cela est impossible, puisque m > m′ . Deuxième cas. On suppose d(x, S0 ) > k et T ≤ k . Pour j ∈ {l ′ + 1, ..., l} on a k + 3P ≤ dmax (x, Y j m′ ) ≤ k + M + 2P + δ. Comme Y j m′ ⊂ 4P -géod(x, b) le lemme 3.13 montre que le nombre de possibili- tés pour Y j m′ est borné par C2 = C (δ, K, N , Q, P , M ). Pour i ∈ {m′ + 1, ..., m} on a k + P ≤ d(x, Si ) ≤ k + M . En effet l’inégalité de gauche a lieu par la condition ii) de la définition 4.1. Comme Si ⊂ P -géod(x, b) le lemme 3.13 montre que le nombre de possibilités pour Si est borné par C3 = C (δ, K, N , Q, P , M ). On a k ′ = d(x′ , S0) − d(x, S0) + k + M 2 d’après (130) donc M 2 − N . (171) + P + 3N + δ k ′ ≥ d(x′ , b) − d(x, b) + k + Soit y ∈ Sm′ . On a M M d(x′ , y )−k ′ ≤ d(x′ , y )−d(x′ , b)+d(x, b)+N −k− 2 ≤ d(x, y )−k− 2 où la première inégalité vient de (171) et la deuxième de (170). Comme cela est vrai pour tout y ∈ Sm′ on en déduit M (cid:0)d(x′ , Sm′ ) − k ′ (cid:1) ≤ (cid:0)d(x, Sm′ ) − k(cid:1) − 2 On a d(x, Sm′+1) ≤ k + M donc d(Sm′ , Sm′+1 ) ≥ d(x, Sm′ ) − k − M − N et d(Sm′ , Sm′+1) + 2M ≥ d(x, Sm′ ) − k + M − N ≥ (d(x′ , Sm′ ) − k ′ ) + 3M/2 − (P + 4N + δ ) ≥ (d(x′ , Sm′ ) − k ′ ) + 1 où l’avant-dernière inégalité a lieu par (172) et où la dernière inégalité a lieu car on suppose 3M/2 ≥ P + 4N + δ + 1, ce qui est permis par (HM ). Le membre de gauche de (160) est ma joré par C0C2B−1 Xm>m′ ,l≥l′ (cid:16) C0C3B−1 d(Sm′ , Sm′+1) + 2M (cid:17)l−l′ 1 − C0C1B−1 (cid:17)m−m′ (cid:18) l l ′(cid:19)(cid:16) d(x′ , Sm′ ) − k ′ ≤ 2 Xm>m′ (cid:16) C0C3B−1 d(Sm′ , Sm′+1) + 2M (cid:17)l′ 1 − C0C1B−1 (cid:17)m−m′ (cid:16) 162 +P +3N +δ (172) (173) = grâce au lemme 4.67. En effet les hypothèses du lemme 4.67 sont satisfaites car 2C0C2B−1 ≤ 1 par (HB ) donc 2C0C2B−1 + (d(x′ , Sm′ ) − k ′ ) ≤ d(Sm′ , Sm′+1) + 2M grâce à (173). Ensuite on calcule Xm>m′ (cid:16) C0C3B−1 C0C3B−1 1 − C0C1B−1 (cid:17)m−m′ 1 − C0(C1 + C3 )B−1 ≤ 1 où la dernière inégalité a lieu car on suppose B ≥ 2C0(C1 + C3) grâce à (HB ). Troisième cas. On suppose d(x, S0) > k et T > k . Soit j ∈ {l ′ + 1, ..., l} et w ∈ Y j m′ . Il existe y ∈ Sm′ et z ∈ Sm′+1 tels que w ∈ P -géod(y , z). On a d(x, w) ≥ d(x, y ) − d(y , w) et d(y , w) ≤ d(y , z) + P donc d(x, w) ≥ d(x, y ) − d(y , z) − P . On en déduit dmax (x, Y j m′ ) ≥ d(x, Sm′ ) − d(Sm′ , Sm′+1 ) − 2N − P . D’autre part comme j ∈ {l ′ + 1, ..., l} on a dmax (x, Y j m′ ) ≤ T + M + 2P + δ . Donc dmax (x, Y j m′ ) appartient à l’intervalle [d(x, Sm′ ) − d(Sm′ , Sm′+1 ) − 2N − P , T + M + 2P + δ ] qui est de longueur (M + 3P + 2N + δ ) + d(Sm′ , Sm′+1 ) + T − d(x, Sm′ ). On suppose 3P + 2N + δ + 1 ≤ M , ce qui est permis par (HM ). Comme Y j m′ ⊂ 4P -géod(x, b) le lemme 3.13 montre que le nombre de possibilités pour Y j m′ est borné par C2 (cid:0)2M + d(Sm′ , Sm′+1) + T − d(x, Sm′ )(cid:1) avec C2 = C (δ, K, N , Q, P ). Pour éviter des absurdités vérifions que 2M + d(Sm′ , Sm′+1) + T − d(x, Sm′ ) ≥ 1. Cela résulte des inégalités d(Sm′ , Sm′+1 ) ≥ d(x, Sm′ ) − d(x, Sm′+1) − N et d(x, Sm′+1) ≤ dmax (x, Sm′+1) ≤ T + M car on suppose M ≥ N + 1 grâce à (HM ). Par (131) on a d(x′ , S0) + d(x, x′ ) − d(x, S0) 2 k ′ − ( (174) + 1. + M 2 ) ≤ N 2 163 Soit y ∈ Sm′ . Il résulte de (170) que d(x′ , y ) − d(x′ , S0) + d(x, S0) ≤ d(x, y ) + P + 3N + δ. (175) Donc M 2 N 2 + + 1 − + (P + 7N/2 + δ + 1) d(x′ , S0) + d(x, x′ ) − d(x, S0) 2 M 2 d(x′ , y ) − k ′ ≤ d(x′ , y ) − ≤ d(x, y ) − T − où la première inégalité a lieu par (174) et la deuxième par (175). Comme cela est vrai pour tout y ∈ Sm′ on en déduit M d(x′ , Sm′ ) − k ′ ≤ d(x, Sm′ ) − T − 2 D’autre part on a + (P + 7N/2 + δ + 1). (176) k + P ≤ dmax (x, Sm ) ≤ ... ≤ dmax (x, Sm′+1) ≤ T + M donc le nombre de possibilités pour ces entiers est borné par (cid:0)(T −k)+M +(m−m′ ) (cid:1). m−m′ Comme Sm′+1 , ..., Sm sont inclus dans P -géod(x, b), le lemme 3.13 montre que, connaissant les entiers dmax (x, Sm), ..., dmax (x, Sm′+1), le nombre de pos- sibilités pour Sm′+1 , ..., Sm est borné par C m−m′ avec C3 = C (δ, K, N , Q, P ). 3 Le membre de gauche de (160) est alors ma joré par Xm>m′ (cid:16) C0C3B−1 1 − C0C1B−1 (cid:17)m−m′ (cid:18)(T − k) + M + (m − m′ ) (cid:19) m − m′ l ′(cid:19)(cid:16) C0C2B−1(cid:0)2M + d(Sm′ , Sm′+1) + T − d(x, Sm′ )(cid:1) (cid:17)l−l′ Xl≥l′ (cid:18) l d(Sm′ , Sm′+1) + 2M d(x′ , Sm′ ) − k ′ d(Sm′ , Sm′+1) + 2M (cid:17)l′ (cid:16) 1 − C0C1B−1 (cid:17)m−m′ (cid:18)(T − k) + M + (m − m′ ) ≤ 2 Xm>m′ (cid:16) C0C3B−1 (cid:19) m − m′ où la dernière inégalité résulte du lemme 4.67. Les hypothèses du lemme 4.67 sont satisfaites car on suppose 2C0C2B−1 ≤ 1 grâce à (HB ) et car (cid:0)2M + d(Sm′ , Sm′+1) + T − d(x, Sm′ )(cid:1) + (d(x′ , Sm′ ) − k ′ ) 3M ≤ d(Sm′ , Sm′+1 ) + + (2P + 7N/2 + δ + 1) ≤ d(Sm′ , Sm′+1 ) + 2M 2 164 où la première inégalité a lieu par (176) et la deuxième a lieu car on suppose M 2 ≥ 2P + 7N/2 + δ + 1, ce qui est permis par (HM ). Or grâce à (161), Xm≥m′ (cid:16) C0C3B−1 1 − C0C1B−1 (cid:17)m−m′ (cid:18)(T − k) + M + (m − m′ ) (cid:19) m − m′ = (cid:16)1 − (cid:16) C0C3B−1 1 − C0C1B−1 (cid:17)(cid:17)−((T −k)+M +1) ≤ 2esd(x,x′ ) car (T − k) ≤ T ≤ d(x, x′ ) et car on suppose 1 − (cid:16) C0C3B−1 1 − C0C1B−1 (cid:17) ≥ max(e−s , 2−(M +1)−1 grâce à (HB ). Ceci termine l’étude du troisième cas. (cid:3) On a démontré les lemmes 4.65 et 4.66 et donc aussi les lemmes 4.64, 4.63, 4.62, 4.56, et la proposition 4.3. ) 4.8 Equivariance des opérateurs à compacts près. Pour terminer la démonstration de la proposition 4.4, il suffit de montrer la proposition suivante. Proposition 4.68 Soit T ∈ R+ et x, x′ ∈ X vérifiant d(x, x′ ) = 1. Alors x − eτ θ♭ x′ ∂ e−τ θ♭ x ∂ e−τ θ♭ a) (eτ θ♭ x′ )τ ∈[0,T ] appartient à K(Hx,s [0, T ]), x′ hx′ e−τ θ♭ x − eτ θ♭ x hxe−τ θ♭ b) (eτ θ♭ x′ )τ ∈[0,T ] appartient à K(Hx,s [0, T ]), x′ ux′ ,rKx′ e−τ θ♭ x − eτ θ♭ x ux,rKxe−τ θ♭ c) pour tout r ∈ N, (eτ θ♭ x′ )τ ∈[0,T ] appartient à K(Hx,s [0, T ]). La proposition 4.3, qui a été établie au sous-paragraphe précédent, assure l’équivalence des normes de Hx,s et Hx′ ,s . On en déduit, grâce à la proposi- tion 4.46, que tous les opérateurs apparaissant dans la proposition 4.68 sont continus. La proposition 4.68 est donc seulement un énoncé de compacité. Démonstration de la proposition 4.4 en admettant la proposition 4.68. Comme X est géodésique, l’énoncé de la proposition 4.68 implique évidem- ment le même énoncé pour x, x′ quelconques (c’est-à-dire sans l’hypothèse q=1 hx (1 − ∂hx − hx∂ )q−1 + P∞ d(x, x′ ) = 1). Comme Jx = PQ r=1 ux,rKx , on déduit des propositions 4.46 et 4.68 que pour T ∈ R+ et x, x′ ∈ X , x′ (∂ + Jx′ ∂Jx′ )e−τ θ♭ (eτ θ♭ x (∂ + Jx∂Jx )e−τ θ♭ x − eτ θ♭ x′ )τ ∈[0,T ] est un opérateur compact sur le C[0, T ]-module hilbertien Hx,s [0, T ]. (cid:3) 165 On va voir que la proposition 4.68 résulte du lemme suivant. On com- mence par remarquer que θx′ − θx est un opérateur borné (et même de norme p,k ,m,(l0 ,...,lm) ≤ 1) sur Hx,s : en effet pour tous p, k , m, l0 , ..., lm , pour Z ∈ Y et x pour )−1(Z ) (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li } ) ∈ (π p,k ,m,(l0 ,...,lm) x d(x, S0 )− d(x′ , S0) appartient à {−1, 0, 1} et est déterminé par Z (on rappelle que S0 = {a1 , ..., ap}). D’autre part θ♭ x − θx est un opérateur borné sur Hx,s d’après le lemme 4.49 et de même θ♭ x′ − θx′ est un opérateur borné sur Hx′ ,s . Grâce à l’équivalence des normes de Hx,s et Hx′ ,s on en déduit que l’opérateur x − θ♭ θ♭ x′ est borné sur Hx,s . On note [u, v ] = uv − vu. Lemme 4.69 Soit T ∈ R+ et x, x′ ∈ X vérifiant d(x, x′ ) = 1. Alors a) (eτ θx (hx − hx′ )e−τ θx )τ ∈[0,T ] appartient à K(Hx,s [0, T ]), b) pour tout r ∈ N, (eτ θx (ux,rKx−ux′ ,rKx′ )e−τ θx )τ ∈[0,T ] appartient à K(Hx,s [0, T ]), x′ ), eτ θx ∂ e−τ θx ])τ ∈[0,T ] appartient à K(Hx,s [0, T ]), c) ([(θ♭ x − θ♭ x′ ), eτ θx hxe−τ θx ])τ ∈[0,T ] appartient à K(Hx,s [0, T ]), x − θ♭ d) ([(θ♭ x′ ), eτ θx ux,rKxe−τ θx ])τ ∈[0,T ] appartient à K(Hx,s [0, T ]). x−θ♭ e) pour tout r ∈ N, ([(θ♭ Démonstration de la proposition 4.68 en admettant le lemme 4.69. Montrons d’abord que le c) du lemme 4.69 implique le a) de la propo- x′ )i , eτ θx ∂ e−τ θx ])τ ∈[0,T ] appartient à sition 4.68. Pour tout i ∈ N, ([(θ♭ x − θ♭ K(Hx,s [0, T ]) et donc x−θ♭ ([eτ (θ♭ x′ ) , eτ θx ∂ e−τ θx ])τ ∈[0,T ] ∈ K(Hx,s [0, T ]). Or on a eτ θ♭ x ∂ e−τ θ♭ x − eτ θ♭ x′ ∂ e−τ θ♭ x′ = eτ (θ♭ x−θ♭ x′ −θx ) [eτ (θ♭ x′ ) , eτ θx ∂ e−τ θx ]e−τ (θ♭ x−θx ) ce qui montre le a) de la proposition 4.68. Ensuite le a) et le d) du lemme 4.69 impliquent le b) de la proposition 4.68. En effet, on a x′ hxe−τ θ♭ x′ )+(eτ θ♭ x′ (hx−hx′ )e−τ θ♭ x −eτ θ♭ x hxe−τ θ♭ x′ = (eτ θ♭ x′ hx′ e−τ θ♭ eτ θ♭ x −eτ θ♭ x hxe−τ θ♭ x′ ) x′ hx′ e−τ θ♭ x − eτ θ♭ donc la compacité de (eτ θ♭ x hxe−τ θ♭ x′ )τ ∈[0,T ] que l’on cherche à établir résulte de la compacité des deux termes du membre de droite. D’abord en conjugant par eτ (θ♭ x′ −θx ) on voit que le a) du lemme 4.69 implique que x′ (hx − hx′ )e−τ θ♭ l’opérateur (eτ θ♭ x′ )τ ∈[0,T ] est compact. Ensuite on a l’égalité x−θ♭ x′ −θx ) [eτ (θ♭ x′ = eτ (θ♭ x − eτ θ♭ x′ hxe−τ θ♭ eτ θ♭ x hxe−τ θ♭ x′ ) , eτ θx hxe−τ θx ]e−τ (θ♭ x−θx ) 166 donc le d) du lemme 4.69 montre la compacité de x′ hxe−τ θ♭ x − eτ θ♭ x hxe−τ θ♭ (eτ θ♭ x′ )τ ∈[0,T ] . Ceci termine la preuve du b) de la proposition 4.68. Enfin par un argument similaire (en remplaçant hx par ux,rKx et hx′ par ux′ ,rKx′ ) le b) et le e) du lemme 4.69 impliquent le c) de la proposition 4.68. (cid:3) Démonstration du lemme 4.69 en admettant le lemme 4.70. Pour tout p ∈ {1, ..., pmax} et n ∈ N on note Pn le pro jecteur orthogonal sur le sous-espace vectoriel de H→ x,s(∆p ) engendré par les eS pour S ∈ ∆p tel que d(x, S ) ≤ n, de sorte que (Pnf )(S ) = f (S ) si d(x, S ) ≤ n et (Pnf )(S ) = 0 si d(x, S ) > n. Dans les notations adoptées jusqu’ici on a donc P = PP . On supposera toujours n ≥ P , de sorte que (1 − Pn )(1 − P ) = 1 − Pn . Il est évident que le lemme 4.69 résulte du lemme suivant. (cid:3) Lemme 4.70 Soit T ∈ R+ et x, x′ ∈ X vérifiant d(x, x′ ) = 1 et p ∈ {2, ..., pmax}. Alors a) supτ ∈[0,T ] k(1−Pn)eτ θx (hx−hx′ )e−τ θx kL(Hx,s (∆p−1 ),H→ x,s (∆p )) tend vers 0 quand n → ∞, b) pour tout r ∈ N, τ ∈[0,T ] k(1 − Pn )eτ θx (ux,rKx − ux′ ,rKx′ )e−τ θx kL(Hx,s (∆p−1 ),H→ sup x,s (∆p )) tend vers 0 quand n → ∞, x′ ), eτ θx ∂ e−τ θx ]kL(Hx,s (∆p ),H→ x − θ♭ c) supτ ∈[0,T ] k(1 − Pn )[(θ♭ x,s (∆p−1 )) tend vers 0 quand n → ∞, x′ ), eτ θx hxe−τ θx ]kL(Hx,s (∆p−1 ),H→ d) supτ ∈[0,T ] k(1 − Pn )[(θ♭ x − θ♭ x,s (∆p )) tend vers 0 quand n → ∞, e) pour tout r ∈ N, x′ ), eτ θx ux,rKxe−τ θx ]kL(Hx,s (∆p−1 ),H→ x − θ♭ τ ∈[0,T ] k(1 − Pn )[(θ♭ sup x,s (∆p )) tend vers 0 quand n → ∞. Démonstration du a) du lemme 4.70. La preuve qui suit est assez voisine de la démonstration du lemme 4.36. On déduira a) de (182) qui est une variante de l’inégalité (86). Sous-lemme 4.71 Il existe C1 = C (δ, K, N ) tel que pour tout S ∈ ∆p−1 , la mesure de l’ensemble des t ∈ [0, 1] tels que (hx,t − hx′ ,t)(eS ) 6= 0 est ≤ C1 1+d(x,S ) . 167 Démonstration. C’est une conséquence immédiate du lemme 3.15. Le sous-lemme suivant est une conséquence évidente du lemme 3.20. (cid:3) Sous-lemme 4.72 Pour tout S ∈ ∆p−1 , hx,t(eS ) et hx′ ,t (eS ) ne dépendent que de la connaissance des points de B (x, 2δ + 1) ∪ B (S, N ) (177) {y ∈ (5δ + 2)-géod(x, a), d(x, y ) ∈ [td(x, S ) − 2N − 3, td(x, S ) + 2]} ∪ [a∈S et des distances entre ces points. Démonstration. On applique le lemme 3.20 à x et x′ et on remarque que 5δ -géod(x′ , a) ⊂ (5δ + 2)-géod(x, a) et d(x, S ) − d(x′ , S ) ≤ 1. (cid:3) Suite de la démonstration du a). Soient k , m, l0 , . . . , lm ∈ N et p,k ,m,(l0 ,...,lm) Z ∈ Y vérifiant r0 (Z ) > k + P . x On pose l0 = 0 et li = li−1 pour i ∈ {1, . . . , m + 1} et on note ΛZ la partie ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 formée des Z vérifiant de Y x r0 (Z ) ≤ r0 ( Z ) ≤ r0 (Z ) + N (178) et tels que pour tout (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , Z 1 0 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 )−1( Z ) x il existe une énumération (a1 , . . . , ap) de S1 vérifiant (a1 , . . . , ap , S1 , ..., Sm+1 , ( Y j i+1)i∈{0,...,m},j∈{1,...,li } ) ∈ (π p,k ,m,(l0 ,...,lm) x Soit t ∈ [0, 1]. On note ΛZ,t l’ensemble des Z ∈ ΛZ tels que pour tout (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , Z 1 0 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 )−1( Z ) x )−1(Z ). (179) on ait {z ∈ géod(x, b), d(x, z) = E (tr0 ( Z ))}. Z 1 0 = [b∈ S0 La condition (180) implique que pour Z ∈ ΛZ,t on a t1 0 ( Z ) = E (tr0( Z )). 168 (180) Sous-lemme 4.73 Soit i )i∈{0,...,m+1},j∈{1,...,li } ) ∈ Y p−1,k ,m+1,(l0 ,...,lm+1) (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j x tel que d(x, S1) ≤ d(x, S0) ≤ d(x, S1) + N et qu’il existe (a1 , . . . , ap) vérifiant (179). On définit Z 1 0 par (180). Alors Z 1 0 est de diamètre inférieur ou égal à P /3 et il existe Z ∈ ΛZ,t tel que (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , Z 1 0 ) appartienne à (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 )−1 ( Z ). x Démonstration. Soient z , z ′ ∈ Z 1 0 . Soit b ∈ S . On a z , z ′ ∈ 2N -géod(x, b) et d(x, z) = d(x, z ′ ) donc par (Hδ (z , x, z ′ , b)), d(z , z ′ ) ≤ 2N + δ ≤ P /3. Comme Z 1 0 est non vide et P /3 ≤ M , l’argument que nous venons de donner montre aussi que la condition (180) est vérifiée par les autres éléments de la classe d’équivalence Z de l’élément (181) et donc Z ∈ ΛZ,t . (cid:3) Suite de la démonstration du a). Notre but est maintenant de montrer l’inégalité suivante, qui est une variante de (86) : il existe une constante C2 = C (δ, K, N , Q, P , M , T ) telle que τ ∈[0,T ] ξZ (eτ θx (hx,t − hx′ ,t)e−τ θx f )2 ≤ C2 XZ ∈Λ 6= sup Z,t Z,t est l’ensemble des Z ∈ ΛZ,t tels que pour tout où Λ 6= i )i∈{0,...,m+1},j∈{1,...,li } , Z 1 (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 )−1( Z ) x ξ Z (f )2 (181) (182) on ait (183) (hx,t − hx′ ,t )(e S0 ) 6= 0. Nous allons montrer (182) et en même temps justifier que la condition (183) ne dépend que de Z (c’est-à-dire que pour Z ∈ ΛZ,t elle est vérifiée ou non simultanément pour tous les éléments de (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 )−1( Z )). x Sous-lemme 4.74 Soit Z ∈ ΛZ,t et (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , Z 1 0 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 )−1( Z ). x 169 Alors hx,t(e S0 ) et hx′ ,t (e S0 ) ne dépendent que de la connaissance des points de B ( S0 , M ) ∪ B (x, k + 2M ) ∪ B ( Z 1 0 , M ). et des distances entre ces points. (184) Remarque. On devrait plutôt noter hx,t (ea1 ∧ ... ∧ eap−1 ) au lieu de hx,t(e S0 ) mais à partir de maintenant nous commettrons cet abus. Démonstration. Grâce au sous-lemme 4.72, hx,t(e S0 ) et hx′ ,t(e S0 ) ne dé- pendent que de la connaissance des points de B (x, 2δ + 1) ∪ B ( S0 , N )∪ {y ∈ (5δ + 2)-géod(x, a), d(x, y ) ∈ [tr0( Z ) − 2N − 3, tr0( Z ) + 2]} (185) [a∈ S0 et des distances entre ces points. Il suffit donc de montrer que (185) est inclus dans (184). Il est évident que B (x, 2δ + 1) ∪ B ( S0 , N ) est inclus dans (184). Soient a ∈ S0 , y ∈ (5δ + 2)-géod(x, a) vérifiant d(x, y ) ∈ [tr0 ( Z ) − 2N − 3, tr0( Z ) + 2]. Soit z ∈ géod(x, a) vérifiant d(x, z) = E (tr0 ( Z )), si bien que z appartient à Z 1 0 . Alors d(x, y ) − d(x, z) ≤ 2N + 3 donc (Hδ (y , x, z , a)) montre d(y , z) ≤ (2N + 3) + (5δ + 2) + δ et on suppose (2N + 3) + (5δ + 2) + δ ≤ M grâce à (HM ). Donc l’ensemble (185) est inclus dans (184). (cid:3) Le sous-lemme 4.74 implique immédiatement que la condition (183) ne dépend que de Z . Sous-lemme 4.75 Le cardinal de ΛZ,t est majoré par une constante de la forme C (δ, K, N , Q, P , M ). Démonstration. Grâce au lemme 4.28, pour connaître les distances entre les points de B ( S0 , M ) ∪ B ( Z 1 0 , M ) (186) B ( Si , M ) ∪ (187) et ceux de [i∈{1,...,m+1} [ i∈{1,...,m+1},j∈{1,...,li } il suffit de connaître les distances entre les points de (186) et C ′ points de (187), avec C ′ = C (δ, K, N , Q, P , M ) et ces distances sont déterminées à B ( Y j i , M ) ∪ B (x, k + 2M ) 170 C ′′ = C (δ, K, N , Q, P , M ) près par les distances de S1 à ces C ′ points (qui 0 ( Z ), qui est lui-même déterminé font partie de la donnée de Z ) et l’entier t1 à C ′′′ = C (δ, K, N , Q, P , M ) près par r0(Z ) et t. (cid:3) Suite de la démonstration du a). On termine maintenant la preuve de (182). Pour i )i∈{0,...,m+1},j∈{1,...,li } , Z 1 (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 )−1( Z ). x on considère (188) )−1(Z ). X(b1 ,...,bp) (cid:0)(hx,t − hx′ ,t)(ea1 ∧ ... ∧ eap−1 )(cid:1)(b1 , ..., bp ), où la somme porte sur les énumérations (b1 , ..., bp ) de S1 telles que (b1 , . . . , bp , S1 , ..., Sm+1 , ( Y j i+1)i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x Comme la somme (188) a au plus p! termes, le 3) de la proposition 3.37 montre qu’elle est ma jorée par une constante de la forme C (δ, K, N , Q, P , M ). D’après le sous-lemme 4.74 la somme (188) ne dépend que de Z et on peut donc la noter αZ, Z ,t . D’après le sous-lemme 4.73 on a 1 (p − 1)! XZ ∈Λ 6= Z,t x ( Z ) − ρ0 x ( Z ) ≤ 2N . Par Cauchy-Schwarz et grâce au Grâce à (178) on a ρ1 sous-lemme 4.75 on en déduit (182). Montrons maintenant a) à l’aide de (182). Soit C2 comme dans (182). Soient k , m, l0 , . . . , lm ∈ N. On pose l0 = 0 et li = li−1 pour i ∈ {1, . . . , m + 1} comme précédemment. ξZ (eτ θx (hx,t − hx′ ,t)e−τ θx f ) = x ( Z )−ρ0 x ( Z ))αZ, Z ,tξ Z (f ). eτ (ρ1 Sous-lemme 4.76 Il existe C3 = C (δ, K, N , Q, P , M ) tel que pour tout (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x le nombre de possibilités pour (a1 , . . . , ap−1) tels que (a1 , . . . , ap−1 , {a1 , . . . , ap−1}, S0 , ..., Sm , (Y j i−1)i∈{0,...,m+1},j∈{1,...,li } ) appartienne à Y p−1,k ,m+1,(l0 ,...,lm+1) et vérifie d(x, S0) ≤ d(x, {a1 , . . . , ap−1}) ≤ x d(x, S0 ) + N soit ≤ C3 . 171 Démonstration. Pour a ∈ S0 et i ∈ {1, ..., p − 1} on a a ∈ 2F -géod(x, ai ) et d(x, a) − d(x, ai ) ≤ 2N , d’où d(a, ai ) ≤ 2F + 2N . (cid:3) Fin de la démonstration du a). Soit p,k ,m,(l0 ,...,lm) Z ∈ Y vérifiant r0 (Z ) > k + P . x Grâce au sous-lemmes 4.71 et 4.76, et comme Z 1 0 est déterminé par (180) et r0 (Z ) ≤ r0 ( Z ), on a alors Z 1 0 (cid:16) XZ∈Λ 6= )−1( Z )(cid:1)(cid:17)dt ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 x Z,t C1C3 )−1(Z )(cid:1). ♯(cid:0)(π p,k ,m,(l0 ,...,lm) ≤ x r0 (Z ) + 1 Notons IZ l’ensemble des t ∈ [0, 1] tels qu’il existe Z ∈ Λ 6= Z,t vérifiant ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 )−1( Z )(cid:1) ≥ (r0(Z ) + 1)− 1 )−1(Z )(cid:1). 2 ♯(cid:0)(π p,k ,m,(l0 ,...,lm) x x La mesure de IZ est donc ≤ C1C3(r0 (Z ) + 1)− 1 2 . Grâce à Cauchy-Schwarz et à (182) on obtient que ξZ (cid:16)eτ θx (cid:16) Zt∈IZ τ ∈[0,T ] (cid:12)(cid:12)(cid:12) (hx,t − hx′ ,t )dt(cid:17)e−τ θx f (cid:17)(cid:12)(cid:12)(cid:12) sup τ ∈[0,T ] Zt∈[0,1] (cid:12)(cid:12)ξZ (eτ θx (hx,t − hx′ ,t)e−τ θx f )(cid:12)(cid:12) 2dt ≤ C1C3 (r0(Z ) + 1)− 1 2 sup ≤ C1C2C3(r0(Z ) + 1)− 1 (r0( Z ) + 1)−1 ξ Z (f )2 . 2 XZ∈ΛZ D’après le sous-lemme 4.76 et le lemme 4.29, il existe une constante C4 = C (δ, K, N , Q, P , M ) telle que pour Z ∈ ΛZ on ait ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 )−1( Z )(cid:1) ≤ C4♯(cid:0)(π p,k ,m,(l0 ,...,lm) )−1(Z )(cid:1). x x D’autre part pour t 6∈ IZ et Z ∈ Λ 6= Z,t on a ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 )−1( Z )(cid:1) ≤ (r0(Z ) + 1)− 1 )−1(Z )(cid:1). 2 ♯(cid:0)(π p,k ,m,(l0 ,...,lm) x x Par Cauchy-Schwarz on déduit alors de (182) que ξZ (cid:16)eτ θx (cid:16) Zt6∈IZ 2 )−1(Z )(cid:1)−α (cid:12)(cid:12)(cid:12) (hx,t − hx′ ,t)dt(cid:17)e−τ θx f (cid:17)(cid:12)(cid:12)(cid:12) τ ∈[0,T ] (cid:0)♯(π p,k ,m,(l0 ,...,lm) sup x ≤ C2(r0(Z ) + 1)− α 2 )−1( Z )(cid:1)−α ξ Z (f )2 . (r0( Z ) + 1)−1(cid:0)♯(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 XZ ∈ΛZ (191) x 172 (189) 2 (190) Comme (hx,t − hx′ ,t)dt, hx − hx′ = Zt∈IZ (hx,t − hx′ ,t )dt + Zt6∈IZ en combinant les inégalités (189), (190) et (191) et par Cauchy-Schwarz on obtient que )−1(Z )(cid:1)−α ξZ (eτ θx (hx − hx′ )e−τ θx f )2 τ ∈[0,T ] (cid:0)♯(π p,k ,m,(l0 ,...,lm) sup x ≤ 2(cid:16)C1C2C3C α 2 (cid:17) XZ∈ΛZ 4 (r0(Z ) + 1)− 1 2 + C2(r0(Z ) + 1)− α (r0( Z ) + 1)−1 )−1( Z )(cid:1)−α ξ Z (f )2 . (cid:0)♯(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),1,0 x i=0 si ( Z )−li . Pour calculer De plus pour Z ∈ ΛZ on a Qm i=0 si (Z )−li = Qm+1 la norme de (1 − Pn )eτ θx (hx − hx′ )e−τ θx on peut se limiter aux Z tels que r0 (Z ) ≥ n et on déduit donc de (192) que τ ∈[0,T ] k(1 − Pn )eτ θx (hx − hx′ )e−τ θx k2 sup L(H♮,1,0 x,s (∆p−1 ),H→ x,s (∆p )) 2 (cid:17) ≤ 2p!B(cid:16)C1C2C3C α 2 + C2(n + 1)− α 4 (n + 1)− 1 où le facteur p! est dû au fait que Z détermine Z à permutation près de a1 , ..., ap . Ceci termine la preuve de a). Démonstration du b) du lemme 4.70. La preuve de b) n’introduit aucune idée nouvelle par rapport à celle de a), donc on sera bref. Soit r ∈ N. On déduira b) de (197) qui est une variante de l’inégalité (100), de la même façon que l’on avait montré le lemme 4.40 à l’aide de (100). (192) Sous-lemme 4.77 Il existe C = C (δ, K, N , r) tel que pour tout a ∈ X , la mesure de l’ensemble des t ∈ [0, 1] tels que µr,t(x, a) 6= µr,t(x′ , a) est ≤ C 1+d(x,a) . Démonstration. Cela résulte du lemme 3.27. (cid:3) Sous-lemme 4.78 Il existe C1 = C (δ, K, N , Q, r) tel que pour tout S ∈ ∆p−1 , la mesure de l’ensemble des (t, t1 , ..., tQ) ∈ [0, 1]Q+1 tels que ux,r,tKx,Q,(t1 ,...,tQ )(eS ) 6= ux′ ,r,tKx′ ,Q,(t1 ,...,tQ ) (eS ) est ≤ C1 1+d(x,S ) . 173 Démonstration. Il existe C = C (δ, K, N , Q, r) tel que pour S ∈ ∆p−1 la connaissance de ux,r,tKx,Q,(t1 ,...,tQ ) (eS ) et celle de ux′ ,r,tKx′ ,Q,(t1 ,...,tQ ) (eS ) ne dépendent que de la connaissance des hx,ti (eS ′ ) et hx′ ,ti (eS ′ ) pour i = 1, ..., Q et S ′ ∈ ∆ vérifiant d(S, S ′) ≤ C et des µr,t(x, a) et µr,t(x′ , a) pour a ∈ X vérifiant d(a, S ) ≤ C . On applique alors le sous-lemme 4.71 à ces parties S ′ et le sous-lemme 4.77 à ces points a. (cid:3) Suite de la démonstration du b). Soient k , m, l0 , . . . , lm ∈ N et Z ∈ p,k ,m,(l0 ,...,lm) vérifiant r0 (Z ) > k + P . On pose l0 = 0 et li = li−1 pour Y x ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 i ∈ {1, . . . , m + 1} et on note ΛZ la partie de Y formée x des Z vérifiant r0( Z ) − r1( Z ) − r ≤ QF (193) et tels que pour tout (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j 0 )j∈{1,...,Q} , Z 1 1 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z ) x il existe une énumération (a1 , . . . , ap) de S1 vérifiant (a1 , . . . , ap , S1 , ..., Sm+1 , ( Y j i+1)i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x )−1 (Z ). (194) Soient t, t1 , . . . , tQ ∈ [0, 1]. On note ΛZ,t,(t1 ,...,tQ ) l’ensemble des Z ∈ ΛZ tels que pour tout 0 )j∈{1,...,Q} , Z 1 i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 1 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z ) x on ait {z ∈ géod(x, b), d(x, z) = E (tj r0 ( Z ))} pour j ∈ {1, ..., Q} (195) Z j 0 = [b∈ S0 {z ∈ géod(x, b), d(x, z) = E ((1 − t)r1 ( Z ))}. Z 1 1 = [b∈ S1 et Pour Z ∈ ΛZ,t,(t1 ,...,tQ) , les conditions (195) et (196) impliquent 0 ( Z ) = E (tj r0( Z )) pour j ∈ {1, ..., Q}, 1( Z ) = E ((1 − t)r1( Z )). tj t1 et Remarque. Les notations ΛZ et ΛZ,t,(t1 ,...,tQ ) que nous venons d’introduire sont les mêmes que dans la preuve du lemme 4.40 et les conditions (193), (194), (195) et (196) coïncident avec les conditions (98), (99), (101) et (102). (196) 174 Sous-lemme 4.79 Soit i )i∈{0,...,m+1},j∈{1,...,li } ) ∈ Y p−1,k ,m+1,(l0 ,...,lm+1) (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j x tel que d(x, S0) − d(x, S1) − r ≤ QF et qu’il existe (a1 , . . . , ap ) vérifiant 0 )j∈{1,...,Q} et Z 1 1 par (195) et (196). Alors les parties Z j (194). On définit ( Z j 0 et Z 1 1 sont de diamètre inférieur ou égal à P /3 et il existe Z ∈ ΛZ,t,(t1 ,...,tQ ) tel que 0 )j∈{1,...,Q} , Z 1 i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 1 ) appartienne à (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z ). x Démonstration. C’est exactement le sous-lemme 4.41. (cid:3) Suite de la démonstration du b). Notre but est maintenant de montrer l’inégalité suivante, qui est une variante de (100) : il existe une constante C2 = C (δ, K, N , Q, P , M , r, T ) telle que τ ∈[0,T ] (cid:12)(cid:12)ξZ (cid:0)eτ θx (ux,r,tKx,Q,(t1 ,...,tQ ) − ux′ ,r,tKx′ ,Q,(t1 ,...,tQ ) )e−τ θx f (cid:1)(cid:12)(cid:12) sup 2 ≤ C2 XZ∈Λ 6= Z,t,(t1 ,...,tQ ) (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) . Z,t,(t1 ,...,tQ ) est l’ensemble des Z ∈ ΛZ,t,(t1 ,...,tQ ) tels que pour tout où Λ 6= (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j 0 )j∈{1,...,Q} , Z 1 1 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z ) x (197) 2 on ait (cid:0)ux,r,tKx,Q,(t1 ,...,tQ ) − ux′ ,r,tKx′ ,Q,(t1 ,...,tQ ) (cid:1)(e S0 ) 6= 0. Le sous-lemme suivant indique d’où vient la condition (193). (198) Sous-lemme 4.80 Pour S ∈ ∆p−1 et T ∈ ∆p tels que eT apparaisse avec un coefficient non nul dans ux,r,tKx,Q,(t1 ,...,tQ )(eS ) ou dans ux′ ,r,tKx′ ,Q,(t1 ,...,tQ )(eS ) on a d(x, S ) − d(x, T ) − r ≤ QF . 175 Q F (cid:3) , r + QF ]} {y ∈ (F + 2)-géod(x, a), d(y , a) ∈ [r + Démonstration. D’après le 2)a) de la proposition 3.37 on a T ⊂ [a∈S et on suppose Q F + QF ≥ N + (F + 2), ce qui est permis par (HQ ). Sous-lemme 4.81 Soit Z ∈ Λ 6= Z,t,(t1 ,...,tQ ) et 0 )j∈{1,...,Q} , Z 1 i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 1 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z ). x Alors ux,r,tKx,Q,(t1 ,...,tQ )(e S0 ) et ux′ ,r,tKx′ ,Q,(t1 ,...,tQ )(e S0 ) ne dépendent que de la connaissance des points de B (x, k + 2M ) ∪ B ( S0 , M ) ∪ B ( S1 , M ) 0 , M ) ∪ B ( Z 1 B ( Z j ∪ [j∈{1,...,Q} 1 , M ) et des distances entre ces points. (199) ∪ Démonstration. La réunion des ensembles figurant dans le 2)b) de la pro- position 3.37 avec x et x′ au lieu de x et S0 au lieu de S (et qui a la propriété que ux,r,tKx,Q,(t1 ,...,tQ )(e S0 ) et ux′ ,r,tKx′ ,Q,(t1 ,...,tQ ) (e S0 ) ne dépendent que de la connaissance des distances entre les points de cette réunion) est inclus dans B (x, F + 1) ∪ B ( S0 , QN ) (200) [a∈ S0 ,j∈{1,...,Q} {y ∈ (F + 2)-géod(x, a), d(x, y ) − tj d(x, a) ≤ QF + 2} (201) ∪ [a∈ S0 {z ∈ (F + 2)-géod(x, a), d(z , a) ∈ [r, r + QF ]} ∪ [a∈ S0 (cid:8)z ∈ (F + 2)-géod(x, a), d(x, z) − (1 − t)(d(x, a) − r) ≤ QF + 2(cid:9). Il suffit donc de montrer que cet ensemble est inclus dans (199). On suppose F + 1 ≤ 2M et QN ≤ M , ce qui est permis par (HM ). Donc (200) est inclus dans (199). (202) (203) 176 Soit a ∈ S0 , j ∈ {1, ..., Q} et y ∈ (F + 2)-géod(x, a) vérifiant d(x, y ) − tj d(x, a) ≤ QF + 2. Soit z ∈ géod(x, a) vérifiant d(x, z) = E (tj r0 ( Z )), si bien que z appartient à Z j 0 . On a tj d(x, a) − E (tj r0 ( Z )) ≤ N + 1, d’où d(x, y ) − d(x, z) ≤ QF + N + 3 et grâce à (Hδ (y , x, z , a)), d(y , z) ≤ (QF + N + 3) + (F + 2) + δ. On suppose (QF + N + 3) + (F + 2) + δ ≤ M , ce qui est permis par (HM ). Par conséquent (201) est inclus dans Sj∈{1,...,Q} B ( Z j 0 , M ) et donc il est inclus dans (199). Par le a) du lemme 4.14, on a pour tout a ∈ S0 , S1 ⊂ 2F -géod(a, x). Soit a ∈ S0 et y ∈ S1 . L’inégalité (193) implique d(x, a) − d(x, y ) − r ≤ QF + N . Soit z ∈ (F + 2)-géod(x, a) vérifiant d(z , a) ∈ [r, r + QF ]. Cela implique d(x, a) − d(x, z) ∈ [r − F − 2, r + QF ]. On en déduit d(x, y ) − d(x, z) ≤ 2QF + N . Comme z ∈ (F +2)-géod(x, a), y ∈ 2F -géod(x, a) et 2F ≥ F +2, (Hδ (z , x, y , a)) montre d(y , z) ≤ (2QF + N ) + 2F + δ d’où [a∈S {z ∈ (F + 2)-géod(x, a), d(z , a) ∈ [r, r + QF ]} ⊂ B ( S1 , 2QF + N + 2F + δ ) ⊂ B ( S1 , M ) car on suppose 2QF + N + 2F + δ ≤ M , ce qui est permis par (HM ). Donc (202) est inclus dans (199). Enfin soit a ∈ S0 et z ∈ (F + 2)-géod(x, a) vérifiant d(x, z) − (1 − t)(d(x, a) − r) ≤ QF + 2. Soit b ∈ S1 et y ∈ géod(x, b) vérifiant d(x, y ) = E ((1 − t)r1 ( Z )), si bien que y appartient à Z 1 1 . Comme y ∈ géod(x, b) et b ∈ 2F -géod(x, a), on a y ∈ 2F -géod(x, a). Comme d(x, a) ∈ [r0 ( Z ), r0( Z ) + N ] et grâce à (98), on a d(x, z) − (1 − t)r1( Z ) ≤ 2QF + 2 + N , 177 d’où d(x, y ) − d(x, z) ≤ 2QF + 3 + N . Comme z ∈ (F +2)-géod(x, a), y ∈ 2F -géod(x, a) et F +2 ≤ 2F , (Hδ (y , x, z , a)) montre d(y , z) ≤ (2QF + 3 + N ) + 2F + δ. On suppose (2QF + 3 + N ) + 2F + δ ≤ M , ce qui est permis par (HM ). Par conséquent (203) est inclus dans B ( Z 1 1 , M ) et donc dans (199). Ceci termine la preuve du sous-lemme 4.81. (cid:3) Le sous-lemme 4.81 justifie le fait que la condition (198) ne dépend que de Z (c’est-à-dire que pour Z ∈ ΛZ,t,(t1 ,...,tQ ) elle est vérifiée ou non simulta- nément pour tous les éléments de (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z )). x Sous-lemme 4.82 Le cardinal de ΛZ,t,(t1 ,...,tQ ) est majoré par une constante de la forme C (δ, K, N , Q, P , M ). Démonstration. Grâce au lemme 4.28, pour connaître les distances entre les points de B ( S0 , M ) ∪ [j∈{1,...,Q} 0 , M ) ∪ B ( Z 1 B ( Z j 1 , M ) (204) B ( Si , M ) ∪ (205) B ( Y j i+1 , M ) ∪ B (x, k + 2M ) et ceux de [i∈{1,...,m+1} [ i∈{0,...,m},j∈{1,...,li } il suffit de connaître les distances entre les points de (204) et C points de (205), avec C = C (δ, K, N , Q, P , M ) et grâce à (193) ces distances sont dé- terminées à C ′ = C (δ, K, N , Q, P , M ) près par les distances de S1 à ces C 0 ( Z ))j∈{1,...,Q} et points (qui font partie de la donnée de Z ) et les entiers (tj 1 ( Z ). t1 (cid:3) Suite de la démonstration du b). On termine maintenant la preuve de l’inégalité (197). Pour i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 )j∈{1,...,Q} , Z 1 1 ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z ) x on considère X(b1 ,...,bp ) (cid:0)(ux,r,tKx,Q,(t1 ,...,tQ ) − ux′ ,r,tKx′ ,Q,(t1 ,...,tQ ) )(ea1 ∧ ... ∧ eap−1 )(cid:1)(b1 , ..., bp ), (206) 178 où la somme porte sur les énumérations S1 = {b1 , ..., bp} telles que (b1 , . . . , bp , S1 , ..., Sm+1 , ( Y j )−1(Z ). i+1)i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x Comme la somme (206) a au plus p! termes, le 3) de la proposition 3.37 montre qu’elle est ma jorée par une constante de la forme C (δ, K, N , Q, P , M ). D’après le sous-lemme 4.81 la somme (206) ne dépend que de Z et on peut donc la noter αZ, Z ,t,(t1 ,...,tQ ) . D’après les sous-lemmes 4.79 et 4.80 on a ξZ (eτ θx (ux,r,tKx,Q,(t1 ,...,tQ ) − ux′ ,r,tKx′ ,Q,(t1 ,...,tQ ) )e−τ θx f ) 1 x ( Z )) ξ Z (f ). x ( Z )−ρ0 αZ, Z ,t,(t1 ,...,tQ ) eτ (ρ1 (p − 1)! XZ∈Λ 6= = Z,t,(t1 ,...,tQ ) x ( Z ) − ρ0 x ( Z ) ≤ QF + N . Par Cauchy-Schwarz et grâce Grâce à (193) on a ρ1 au sous-lemme 4.82 on en déduit (197). Montrons maintenant b) à l’aide de (197). Soit C2 comme dans (197). Soient k , m, l0 , . . . , lm ∈ N. On pose l0 = 0 et li = li−1 pour i ∈ {1, . . . , m + 1} comme précédemment. Sous-lemme 4.83 Il existe C3 = C (δ, K, N , Q, P , M , r) tel que pour tout (a1 , . . . , ap , S0 , ..., Sm , (Y j i )i∈{0,...,m},j∈{1,...,li }) ∈ Y p,k ,m,(l0 ,...,lm) x le nombre de possibilités pour (a1 , . . . , ap−1) tels que (a1 , . . . , ap−1 , {a1 , . . . , ap−1}, S0 , ..., Sm , (Y j i−1)i∈{0,...,m+1},j∈{1,...,li } ) appartienne à Y p−1,k ,m+1,(l0 ,...,lm+1) et vérifie x d(x, {a1 , . . . , ap−1}) − d(x, S0) − r ≤ QF soit ≤ C3 . Démonstration. C’est une conséquence immédiate du sous-lemme 4.45. (cid:3) p,k ,m,(l0 ,...,lm) Fin de la démonstration du b). Soit Z ∈ Y vérifiant r0 (Z ) > x k + P . Grâce aux sous-lemmes 4.78 et 4.83, on a Z(t,t1 ,...,tQ )∈[0,1]Q+1 (cid:16) XZ ∈Λ 6= ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 x Z,t,(t1 ,...,tQ ) C1C3 )−1(Z )(cid:1). ♯(cid:0)(π p,k ,m,(l0 ,...,lm) ≤ x r0 (Z ) + 1 179 )−1 ( Z )(cid:1)(cid:17)dtdt1 ...dtQ Notons IZ l’ensemble des (t, t1 , ..., tQ ) ∈ [0, 1]Q+1 tels qu’il existe Z ∈ Λ 6= Z,t,(t1 ,...,tQ ) vérifiant ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z )(cid:1) ≥ (r0(Z ) + 1)− 1 )−1(Z )(cid:1). 2 ♯(cid:0)(π p,k ,m,(l0 ,...,lm) x x La mesure de IZ est donc ≤ C1C3(r0 (Z ) + 1)− 1 2 . Grâce à Cauchy-Schwarz et à (197) on obtient que ξZ (cid:16)eτ θx (cid:16) Z(t,t1 ,...,tQ )∈IZ τ ∈[0,T ] (cid:12)(cid:12)(cid:12) (ux,r,tKx,Q,(t1 ,...,tQ ) sup 2 −ux′ ,r,tKx′ ,Q,(t1 ,...,tQ ) )dtdt1 ...dtQ(cid:17)e−τ θx f (cid:17)(cid:12)(cid:12)(cid:12) τ ∈[0,T ] Z(t,t1 ,...,tQ )∈[0,1]Q+1 ≤ C1C3(r0(Z ) + 1)− 1 2 sup 2 (cid:12)(cid:12)ξZ (eτ θx (ux,r,tKx,Q,(t1 ,...,tQ ) − ux′ ,r,tKx′ ,Q,(t1 ,...,tQ ) )e−τ θx f )(cid:12)(cid:12) dtdt1 ...dtQ 2 ≤ 2Q+1C1C2C3(r0 (Z ) + 1)− 1 (r0( Z ) + 1)−Q (r1( Z ) + 1)−1 (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) 2 XZ ∈ΛZ (207) D’après le sous-lemme 4.83 et le lemme 4.29, il existe C4 = C (δ, K, N , Q, P , M , r) tel que pour Z ∈ ΛZ on ait ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z )(cid:1) ≤ C4♯(cid:0)(π p,k ,m,(l0 ,...,lm) )−1(Z )(cid:1). x x D’autre part pour (t, t1 , ..., tQ) 6∈ IZ et Z ∈ Λ 6= Z,t,(t1 ,...,tQ ) on a ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 )−1( Z )(cid:1) ≤ (r0(Z ) + 1)− 1 2 ♯(cid:0)(π p,k ,m,(l0 ,...,lm) x x Par Cauchy-Schwarz on déduit alors de (197) que )−1(Z )(cid:1)−α τ ∈[0,T ] (cid:0)♯(π p,k ,m,(l0 ,...,lm) sup x ξZ (cid:16)eτ θx (cid:16) Z(t,t1 ,...,tQ )6∈IZ (cid:12)(cid:12)(cid:12) (ux,r,tKx,Q,(t1 ,...,tQ ) 2 −ux′ ,r,tKx′ ,Q,(t1 ,...,tQ ) )dtdt1 ...dtQ(cid:17)e−τ θx f (cid:17)(cid:12)(cid:12)(cid:12) ≤ 2Q+1C2 (r0(Z ) + 1)− α (r0( Z ) + 1)−Q (r1( Z ) + 1)−1 2 XZ ∈ΛZ )−1( Z )(cid:1)−α (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) (cid:0)♯(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 2 . x 180 )−1(Z )(cid:1). (209) . (208) . (210) (ux,r,tKx,Q,(t1 ,...,tQ ) − ux′ ,r,tKx′ ,Q,(t1 ,...,tQ ) ) Comme ux,rKx − ux′ ,rKx′ = Z(t,t1 ,...,tQ )∈IZ dtdt1 ...dtQ + Z(t,t1 ,...,tQ )6∈IZ (ux,r,tKx,Q,(t1 ,...,tQ ) − ux′ ,r,tKx′ ,Q,(t1 ,...,tQ ) )dtdt1 ...dtQ , en combinant les inégalités (207), (208) et (209) et par Cauchy-Schwarz on obtient que )−1 (Z )(cid:1)−α ξZ (eτ θx (ux,rKx − ux′ ,rKx′ )e−τ θx f )2 τ ∈[0,T ] (cid:0)♯(π p,k ,m,(l0 ,...,lm) sup x ≤ 2Q+2(cid:16)C1C2C3C α 2 (cid:17) XZ∈ΛZ 4 (r0(Z ) + 1)− 1 2 + C2 (r0(Z ) + 1)− α (r0( Z ) + 1)−Q )−1( Z )(cid:1)−α (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) 2 (r1 ( Z ) + 1)−1(cid:0)♯(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,1 x i=0 si ( Z )−li . Pour calculer la De plus pour Z ∈ ΛZ on a Qm i=0 si (Z )−li = Qm+1 norme de (1 − Pn )eτ θx (ux,rKx − ux′ ,rKx′ )e−τ θx on peut se limiter aux Z tels que r0(Z ) ≥ n et on déduit donc de (210) que τ ∈[0,T ] k(1 − Pn )eτ θx (ux,rKx − ux′ ,rKx′ )e−τ θx k2 sup L(H♮,Q,1 (∆p−1 ),H→ x,s (∆p )) x,s 2 (cid:17) ≤ 2Q+2p!B e2(QF −r)s(cid:16)C1C2C3C α 2 + C2(n + 1)− α 4 (n + 1)− 1 où le facteur p! est dû au fait que Z détermine Z à permutation près de a1 , ..., ap . Ceci termine la preuve de b). Démonstration du c) du lemme 4.70. La preuve de c) est quasiment identique à celle de d) et légèrement plus simple (du fait que ∂ ne fait pas in- tervenir une moyenne, contrairement à hx ). Nous choisissons donc de montrer d) seulement. Démonstration du d) du lemme 4.70. Soient k , m, l0 , . . . , lm ∈ N et p,k ,m,(l0 ,...,lm) Z ∈ Y vérifiant r0 (Z ) > k + P . x On pose l0 = 0 et li = li−1 pour i ∈ {1, . . . , m + 1}. On note ΛZ la partie de ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 formée des Z vérifiant Y x r0 (Z ) ≤ r0 ( Z ) ≤ r0 (Z ) + N (211) et tels que pour tout (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j 0 )j∈{1,...,7} ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 )−1( Z ) x 181 il existe une énumération (a1 , . . . , ap) de S1 vérifiant (a1 , . . . , ap , S1 , ..., Sm+1 , ( Y j i+1)i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x )−1 (Z ). (212) Soient t ∈ [0, 1] et u1 , u2 , u3 , v1 , v2 , v3 ∈ [0, 1[. On note (ΛZ,t)v1 ,v2 ,v3 u1 ,u2 ,u3 l’en- semble des Z ∈ ΛZ tels que, en notant u1 u2 1 r0( Z )), w2 = E (( w1 = E ( + 6 6 6 5 v1 )r0 ( Z )), w5 = E (( 6 − w4 = E ((1 − 6 on ait, pour tout i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 )j∈{1,...,7} ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 )−1( Z ), x )r0( Z )), w3 = E (( v2 6 2 6 )r0 ( Z )), w6 = E (( )r0( Z )), v3 6 u3 6 4 6 − )r0( Z )) + les égalités Z j 0 = [b∈ S0 et (214) (213) {z ∈ géod(x, b), d(x, z) = wj } pour j ∈ {1, ..., 6} Z 7 {z ∈ géod(x, b), d(x, z) = E (tr0( Z ))}. 0 = [b∈ S0 Pour Z ∈ (ΛZ,t)v1 ,v2 ,v3 u1 ,u2 ,u3 , les conditions (213) et (214) impliquent 0 ( Z ) = wj pour j ∈ {1, ..., 6}, et tj Sous-lemme 4.84 Soit i )i∈{0,...,m+1},j∈{1,...,li } ) ∈ Y p−1,k ,m+1,(l0 ,...,lm+1) (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j x tel que d(x, S1) ≤ d(x, S0) ≤ d(x, S1) + N et qu’il existe (a1 , . . . , ap) vérifiant (212). On définit ( Z j 0 )j∈{1,...,7} par (213) et (214). Alors les parties Z j 0 sont de diamètre inférieur ou égal à P /3 et il existe Z ∈ (ΛZ,t)v1 ,v2 ,v3 u1 ,u2 ,u3 tel que i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 )j∈{1,...,7}) appartienne à (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 )−1 ( Z ). x 0 ( Z ) = E (tr0( Z )). t7 (215) 182 Démonstration. En effet soit b ∈ S0 . Pour z , z ′ ∈ Z j 0 , on a z , z ′ ∈ 2N -géod(x, b) et comme d(x, z) = d(x, z ′ ), (Hδ (z , x, z ′ , b)) donne d(z , z ′ ) ≤ 2N + δ ≤ P /3. Comme les parties Z j 0 sont non vides et P /3 ≤ M , l’argument que nous ve- nons de donner montre aussi que les conditions (213) et (214) sont vérifiées par les autres éléments de la classe d’équivalence Z de l’élément (215) et donc Z ∈ (ΛZ,t)v1 ,v2 ,v3 u1 ,u2 ,u3 . (cid:3) Suite de la démonstration du d). On déduira d) de l’inégalité suivante : il existe C2 = C (δ, K, N , Q, P , M , T ) tel que 2 u1 ,u2 ,u3 , eτ θx hx,t e−τ θx ]f )(cid:12)(cid:12) x′ )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ x )v1 ,v2 ,v3 τ ∈[0,T ] (cid:12)(cid:12)ξZ ([(θ♭ sup 2 ≤ C2 XZ∈(Λ 6= u1 ,u2 ,u3 (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) v1 ,v2 ,v3 Z,t ) u1 ,u2 ,u3 est l’ensemble des Z ∈ (ΛZ,t)v1 ,v2 ,v3 où (Λ 6= Z,t)v1 ,v2 ,v3 u1 ,u2 ,u3 tels que pour tout i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 )j∈{1,...,7} ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 )−1( Z ) x (216) on ait (217) u1 ,u2 ,u3 ( S0) u1 ,u2 ,u3 ( S0) − (ρ♭ x′ )v1 ,v2 ,v3 (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 ( S1) + (ρ♭ u1 ,u2 ,u3 ( S1) 6= 0. −(ρ♭ x )v1 ,v2 ,v3 x′ )v1 ,v2 ,v3 Nous allons maintenant montrer (216) et justifier que la condition (217) ne dépend que de Z (c’est-à-dire que pour Z ∈ (ΛZ,t)v1 ,v2 ,v3 u1 ,u2 ,u3 elle est vérifiée ou non simultanément pour tous les éléments de (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 )−1( Z )). x Sous-lemme 4.85 Pour Z ∈ (ΛZ,t)v1 ,v2 ,v3 u1 ,u2 ,u3 et i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 )j∈{1,...,7} ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 )−1( Z ) x u1 ,u2 ,u3 , eτ θx hx,te−τ θx ](e S0 ) et le x′ )v1 ,v2 ,v3 le coefficient de e S1 dans [(θ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ membre de gauche de (217) ne dépendent que de la connaissance des points de B ( S0 , M ) ∪ B ( S1 , M ) ∪ B (x, k + 2M ) ∪ [j∈{1,...,7} et des distances entre ces points. B ( Z j 0 , M ) (218) 183 Démonstration. Le sous-lemme 4.74 (que l’on applique avec Z 7 0 au lieu de Z 1 0 et en oubliant Z j 0 pour j ∈ {1, ..., 6}) montre que hx,t(e S0 ) ne dépend que de la connaissance des points de B ( S0 , M ) ∪ B (x, k + 2M ) ∪ B ( Z 7 0 , M ). Pour montrer le sous-lemme, il suffit donc de montrer que pour x ∈ B (x, 1), u1 ,u2 ,u3 ( S0 ) et (ρ♭ u1 ,u2 ,u3 ( S1) ne dépendent que de la connaissance des (ρ♭ x )v1 ,v2 ,v3 x )v1 ,v2 ,v3 points de (218) et des distances entre ces points (en effet on applique ceci à x = x et x = x′ ). En vertu du lemme 4.50, il suffit de montrer que pour x ∈ B (x, 1), σ ∈ {0, 1}, a ∈ Sσ et j ∈ {1, ..., 6}, l’ensemble {y ∈ 3δ -géod( x, a), d( x, y ) − wj (σ, x) ≤ N + 6δ + 4} est inclus dans (218), où l’on note (219) u1 6 + w1(σ, x) = E ( w3(σ, x) = E (( u2 1 d( x, Sσ )), w2(σ, x) = E (( 6 6 u3 2 )d( x, Sσ )), w4(σ, x) = E ((1 − 6 6 v3 4 v2 5 )d( x, Sσ )), w6(σ, x) = E (( 6 − 6 − 6 6 Soit x ∈ B (x, 1), σ ∈ {0, 1}, a ∈ Sσ et j ∈ {1, ..., 6}. Soit b ∈ S0 . On a w5 (σ, x) = E (( + )d( x, Sσ )), v1 6 )d( x, Sσ )). )d( x, Sσ )), d(a, b) ≤ 2N + 2F . En effet c’est évident si σ = 0 et si σ = 1 cela résulte du fait que d(x, S1) = r0 (Z ) ∈ [d(x, S0 ) − N , d(x, S0)] ⊂ [d(x, b) − 2N , d(x, b)] (220) (221) (222) et S1 ⊂ 2F -géod(x, b). Soit y dans l’ensemble (219). Comme y ∈ 3δ -géod( x, a) et x ∈ B (x, 1) et grâce à (221), le lemme 3.2 montre que y ∈ (3δ + 4N + 4F + 2)-géod(x, b). D’autre part il résulte de (222) que wj (σ, x) − wj ≤ N + 1. On a donc d(x, y ) − wj ≤ 2N + 6δ + 6. (224) (223) 184 Soit z ∈ géod(x, b) vérifiant d(x, z) = wj , si bien que z appartient à Z j 0 . Comme y et z appartiennent à (3δ + 4N + 4F + 2)-géod(x, b) et que d(x, y ) − d(x, z) ≤ 2N + 6δ + 6, (Hδ (y , x, z , b)) implique d(y , z) ≤ (2N + 6δ + 6) + (3δ + 4N + 4F + 2) + δ. On suppose (2N + 6δ + 6) + (3δ + 4N + 4F + 2) + δ ≤ M , ce qui est permis par (HM ). Donc d(y , z) ≤ M et y appartient à (218). Ceci termine la preuve du sous-lemme 4.85. (cid:3) Le sous-lemme 4.85 implique immédiatement que la condition (217) ne dépend que de Z . Sous-lemme 4.86 Le cardinal de (Λ 6= Z,t)v1 ,v2 ,v3 u1 ,u2 ,u3 est majoré par une constante de la forme C (δ, K, N , Q, P , M ). Démonstration. Grâce au lemme 4.28, pour connaître les distances entre les points de B ( S0 , M ) ∪ [j∈{1,...,7} B ( Z j 0 , M ) (225) B ( Si , M ) ∪ (226) B ( Y j i+1 , M ) ∪ B (x, k + 2M ) et ceux de [i∈{1,...,m+1} [ i∈{0,...,m},j∈{1,...,li } il suffit de connaître les distances entre les points de (225) et C points de (226), avec C = C (δ, K, N , Q, P , M ) et grâce à (211) ces distances sont dé- terminées à C ′ = C (δ, K, N , Q, P , M ) près par les distances de S1 à ces C 0 ( Z ))j∈{1,...,7} . (cid:3) points (qui font partie de la donnée de Z ) et les entiers (tj Fin de la démonstration du d). On termine maintenant la preuve de (216). Pour (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j 0 )j∈{1,...,7} ) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 )−1( Z ) x on considère X(b1 ,...,bp) (cid:0)[(θ♭ u1 ,u2 ,u3 , eτ θx hx,te−τ θx ](ea1 ∧ ... ∧ eap−1 )(cid:1)(b1 , ..., bp ), x′ )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ x )v1 ,v2 ,v3 (227) 185 où la somme porte sur les énumérations S1 = {b1 , ..., bp} telles que (b1 , . . . , bp , S1 , ..., Sm+1 , ( Y j )−1(Z ). i+1)i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x Comme la somme (227) a au plus p! termes, le 3) de la proposition 3.37 montre qu’elle est ma jorée par une constante de la forme C (δ, K, N , Q, P , M , T ). D’après le sous-lemme 4.85 la somme (227) ne dépend que de Z et on peut donc la noter (αZ, Z ,t,τ )v1 ,v2 ,v3 u1 ,u2 ,u3 . D’après le sous-lemme 4.84 on a = u1 ,u2 ,u3 , eτ θx hx,t e−τ θx ]f ) x′ )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ x )v1 ,v2 ,v3 ξZ ([(θ♭ 1 (p − 1)! XZ∈(Λ 6= (αZ, Z ,t,τ )v1 ,v2 ,v3 u1 ,u2 ,u3 ξ Z (f ). v1 ,v2 ,v3 Z,t ) u1 ,u2 ,u3 Par Cauchy-Schwarz et grâce au sous-lemme 4.86 on en déduit (216). Montrons maintenant d) à l’aide de (216). Il résulte de (221) que pour S0 , S1 comme ci-dessus on a d( S0 , S1 ) ≤ 2N + 2F . Donc le lemme 3.45 im- plique facilement qu’il existe C1 = C (δ, K, N ) tel que pour S0 , S1 comme ci-dessus la mesure de l’ensemble des (u1 , u2 , u3 , v1 , v2 , v3) ∈ [0, 1[6 vérifiant (217) est ≤ C1 1+r0 (Z ) . Soit C2 comme dans (182). Soient k , m, l0 , . . . , lm ∈ N. On pose l0 = 0 et li = li−1 pour i ∈ {1, . . . , m + 1} comme précédemment. Soit C3 = C (δ, K, N , Q, P , M ) comme dans le sous-lemme 4.76. Soit Z ∈ Y p,k ,m,(l0 ,...,lm) x vérifiant r0 (Z ) > k + P . On a alors Z(u1 ,u2 ,u3 ,v1 ,v2 ,v3 )∈[0,1[6 (cid:16) XZ∈(Λ 6= ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 x v1 ,v2 ,v3 Z,t ) u1 ,u2 ,u3 C1C3 )−1(Z )(cid:1). ♯(cid:0)(π p,k ,m,(l0 ,...,lm) ≤ x r0 (Z ) + 1 Notons IZ l’ensemble des )−1( Z )(cid:1)(cid:17)du1 ...dv3 (t, u1 , u2 , u3 , v1 , v2 , v3) ∈ [0, 1] × [0, 1[6 tels qu’il existe Z ∈ (Λ 6= Z,t)v1 ,v2 ,v3 u1 ,u2 ,u3 vérifiant ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 )−1( Z )(cid:1) ≥ (r0(Z ) + 1)− 1 2 ♯(cid:0)(π p,k ,m,(l0 ,...,lm) x x )−1(Z )(cid:1). 186 (229) La mesure de IZ est donc ≤ C1C3(r0(Z ) + 1)− 1 2 . Grâce à Cauchy-Schwarz on déduit de (216) que ξZ (cid:16)(cid:16) Z(t,u1 ,u2 ,u3 ,v1 ,v2 ,v3 )∈IZ τ ∈[0,T ] (cid:12)(cid:12)(cid:12) u1 ,u2 ,u3 , eτ θx hx,te−τ θx ] x′ )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ x )v1 ,v2 ,v3 [(θ♭ sup 2 dt...dv3(cid:17)f (cid:17)(cid:12)(cid:12)(cid:12) 2 . ≤ 2766C1C2C3 (r0(Z ) + 1)− 1 (r0( Z ) + 1)−7 (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) 2 XZ∈ΛZ (228) Grâce au sous-lemme 4.76 et au lemme 4.29, il existe C4 = C (δ, K, N , Q, P , M ) tel que pour Z ∈ ΛZ on ait ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 )−1( Z )(cid:1) ≤ C4♯(cid:0)(π p,k ,m,(l0 ,...,lm) )−1(Z )(cid:1). x x D’autre part pour (t, u1 , u2 , u3 , v1 , v2 , v3) 6∈ IZ et Z ∈ (Λ 6= Z,t)v1 ,v2 ,v3 u1 ,u2 ,u3 on a ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 )−1( Z )(cid:1) ≤ (r0(Z ) + 1)− 1 )−1(Z )(cid:1). 2 ♯(cid:0)(π p,k ,m,(l0 ,...,lm) x x Par Cauchy-Schwarz on déduit alors de (216) que )−1(Z )(cid:1)−α (cid:0)♯(π p,k ,m,(l0 ,...,lm) sup x τ ∈[0,T ] ξZ (cid:16)(cid:16) Z(t,u1 ,u2 ,u3 ,v1 ,v2 ,v3 )6∈IZ 2 u1 ,u2 ,u3 , eτ θx hx,te−τ θx ]dt...dv3(cid:17)f (cid:17)(cid:12)(cid:12)(cid:12) x′ )v1 ,v2 ,v3 [(θ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ ≤ 2766C2(r0(Z ) + 1)− α (r0 ( Z ) + 1)−7 2 XZ ∈ΛZ )−1( Z )(cid:1)−α (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) (cid:0)♯(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 2 . (230) x En combinant les inégalités (228), (229) et (230) et par Cauchy-Schwarz on obtient que )−1(Z )(cid:1)−α x′ ), eτ θx hxe−τ θx ])f )2 τ ∈[0,T ] ξZ (([(θ♭ (cid:0)♯(π p,k ,m,(l0 ,...,lm) x − θ♭ sup x ≤ 2866(cid:16)C1C2C3C α 2 (cid:17) XZ∈ΛZ 4 (r0(Z ) + 1)− 1 2 + C2 (r0(Z ) + 1)− α (r0( Z ) + 1)−7 )−1( Z )(cid:1)−α ξ Z (f )2 . (cid:0)♯(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),7,0 x i=0 si ( Z )−li . Pour calculer la De plus pour Z ∈ ΛZ on a Qm i=0 si (Z )−li = Qm+1 x′ ), eτ θx hxe−τ θx ]) on peut se limiter aux Z tels que norme de (1 − Pn )([(θ♭ x − θ♭ (231) (cid:12)(cid:12)(cid:12) 187 r0 (Z ) ≥ n et on déduit donc de (231) que x′ ), eτ θx hxe−τ θx ])k2 τ ∈[0,T ] k(1 − Pn )([(θ♭ x − θ♭ sup L(H♮,7,0 x,s (∆p−1 ),H→ x,s (∆p )) 2 (cid:17) ≤ 2866p!(cid:16)C1C2C3C α 2 + C2(n + 1)− α 4 (n + 1)− 1 où le facteur p! est dû au fait que Z détermine Z à permutation près de a1 , ..., ap . Ceci termine la preuve de d). Démonstration du e) du lemme 4.70. La preuve de e) est plus subtile que celle de d) pour la raison suivante. L’entier r ∈ N est fixé mais peut être beaucoup plus grand que M . On ne peut donc pas espérer que, dans les no- tations de la preuve de d), (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 (e S0 ) et (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 (e S1 ) soit déterminés par la connaissance des points de B ( S0 , M ) ∪ B (x, k + 2M ) ∪ [j∈{1,...,6} et des distances entre ces points, pour certaines parties Z j 0 . Au contraire si des parties ( Z j 1 )j∈{1,...,6} sont choisies de telle sorte qu’elles déterminent (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 (e S1 ), on peut affirmer (en utilisant de nouveau (HM )) qu’elles x )v1 ,v2 ,v3 déterminent (ρ♭ u1 , u2 , u3 (e S0 ) pour certains u1 , u2 , u3 , v1 , v2 , v3 que nous allons calculer. D’abord S1 est situé en gros (c’est-à-dire modulo des constantes de la forme C (δ, K, N , Q, P )) sur une géodésique entre x et S0 , à distance r de S0 . Pour que les parties ( Z j 1 )j∈{1,...,6} déterminent (ρ♭ x )v1 ,v2 ,v3 u1 ,u2 ,u3 (e S1 ), elles doivent être situées en gros sur une géodésique entre x et S1 , à des distances de x égales à B ( Z j 0 , M ) u1 d(x, S1), 6 6 − v1 6 1 + u2 2 + u3 d(x, S1), d(x, S1), 6 6 5 − v2 4 − v3 6 6 u1 , u2 , u3 (e S0 ) les parties ( Z j x )v1 ,v2 ,v3 Mais pour déterminer (ρ♭ 1 )j∈{1,...,6} doivent égale- ment être situées en gros sur une géodésique entre x et S0 , à des distances de x égales à d(x, S1), d(x, S1). d(x, S1), u1 d(x, S0), 6 6 − v1 6 d(x, S0), 1 + u2 d(x, S0), 6 5 − v2 6 d(x, S0), 2 + u3 d(x, S0), 6 4 − v3 6 d(x, S0). 188 de sorte que d(x, S0) est en gros égal à (1 + κ)d(x, S1). r On pose κ = 1+d(x, S1 ) On obtient donc les relations u1 1 + κ v1 + 6κ 1 + κ u3 − 2κ 1 + κ v3 + 4κ 1 + κ u2 − κ 1 + κ v2 + 5κ 1 + κ , u2 = , u3 = , . u1 = v1 = , v2 = , v3 = (232) Pour adapter la preuve de d) on utilisera la variante suivante du lemme 3.45 (que l’on appliquera avec ρ en gros égal à r , κ comme ci-dessus, y ∈ S1 et y ′ ∈ S0). Sous-lemme 4.87 Pour tout ρ ∈ N, il existe C = C (δ, K, ρ) tel que pour x, x′ , y , y ′ ∈ X verifiant d(x, x′ ) ≤ ρ et d(y , y ′) ≤ ρ et pour κ ∈ [0, 1 10 ] vérifiant κd(x, y ) ≤ ρ, la mesure de l’ensemble des (u1 , u2 , u3 , v1 , v2 , v3 ) ∈ [5κ, 1 − 5κ[6 tels que, en définissant u1 , u2 , u3 , v1 , v2 , v3 comme dans (232), d♭ v1 ,v2 ,v3 u1 ,u2 ,u3 (x, y ) − d♭ v1 ,v2 ,v3 u1 ,u2 ,u3 (x′ , y ) − d♭ v1 ,v2 ,v3 u1 , u2 , u3 (x, y ′) + d♭ v1 ,v2 ,v3 u1 , u2 , u3 (x′ , y ′) 6= 0 est ≤ C 1+d(x,y) . Démonstration. La démonstration est une adaptation de celle du lemme 3.45. En particulier on applique le sous-lemme 3.48 aux familles Ab1 ,b2 ,b3 a1 ,a2 ,a3 = A y ,v1 ,v2 ,v3 x,u1 ,u2 ,u3 si y = y et Ab1 ,b2 ,b3 a1 ,a2 ,a3 = A y ,v1 ,v2 ,v3 x, u1 , u2 , u3 si y = y ′ , pour a1 , a2 , a3 , b1 , b2 , b3 ∈ [0, 1[, avec ui = 5κ + (1 − 10κ)ai et vi = 5κ + (1 − 10κ)bi pour i = 1, 2, 3 et u1 , u2 , u3 , v1 , v2 , v3 comme dans (232). Suite de la démonstration du e). Soient k , m, l0 , . . . , lm ∈ N et p,k ,m,(l0 ,...,lm) Z ∈ Y vérifiant r0 (Z ) > k + P . x On pose l0 = 0 et li = li−1 pour i ∈ {1, . . . , m + 1}. On note ΛZ la partie de ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 formée des Z vérifiant Y x r0( Z ) − r1 ( Z ) − r ≤ QF , (233) (cid:3) et tels que pour tout (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j 0 )j∈{1,...,Q} , ( Z j 1 )j∈{1,...,7}) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 )−1( Z ) x 189 il existe une énumération (a1 , . . . , ap) de S1 vérifiant (a1 , . . . , ap , S1 , ..., Sm+1 , ( Y j )−1(Z ). (234) i+1)i∈{0,...,m},j∈{1,...,li } ) ∈ (π p,k ,m,(l0 ,...,lm) x r 1+r0 (Z ) . Soient t, t1 , . . . , tQ ∈ [0, 1] et u1 , u2 , u3 , v1 , v2 , v3 ∈ On pose κ = u1 ,u2 ,u3 l’ensemble des Z ∈ ΛZ tels que, en [5κ, 1 − 5κ[. On note (ΛZ,t,(t1 ,...,tQ ))v1 ,v2 ,v3 notant u1 u3 2 u2 1 r1( Z )), w2 = E (( 6 6 6 6 6 5 4 v1 )r1 ( Z )), w6 = E (( )r1 ( Z )), w5 = E (( 6 − 6 − w4 = E ((1 − 6 on ait, pour tout (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j 0 )j∈{1,...,Q} , ( Z j 1 )j∈{1,...,7}) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 )−1( Z ) x )r1( Z )), w3 = E (( v2 6 )r1( Z )), v3 6 )r1( Z )) w1 = E ( + + les égalités {z ∈ géod(x, b), d(x, z) = E (tj r0( Z ))} pour j ∈ {1, ..., Q}, Z j 0 = [b∈ S0 Z j 1 = [b∈ S1 {z ∈ géod(x, b), d(x, z) = wj } pour j ∈ {1, ..., 6}, Z 7 {z ∈ géod(x, b), d(x, z) = E ((1 − t)r1 ( Z ))}. 1 = [b∈ S1 et Pour Z ∈ (ΛZ,t,(t1 ,...,tQ ))v1 ,v2 ,v3 u1 ,u2 ,u3 , les conditions (235), (236) et (237) im- pliquent (235) (236) (237) 0( Z ) = E (tj r0( Z )) pour j ∈ {1, ..., Q}, tj 1 ( Z ) = E ((1 − t)r1 ( Z )). 1 ( Z ) = wj pour j ∈ {1, ..., 6} tj t7 et Sous-lemme 4.88 Soit i )i∈{0,...,m+1},j∈{1,...,li } ) ∈ Y p−1,k ,m+1,(l0 ,...,lm+1) (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j x tel que d(x, S0) − d(x, S1) − r ≤ QF et qu’il existe (a1 , . . . , ap ) vérifiant (234). On définit ( Z j 0 )j∈{1,...,Q} et ( Z j 1 )j∈{1,...,7} par (235), (236) et (237). 0 et Z j Alors les parties Z j 1 sont de diamètre inférieur ou égal à P /3 et il existe Z ∈ (ΛZ,t,(t1 ,...,tQ ))v1 ,v2 ,v3 u1 ,u2 ,u3 tel que i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 )j∈{1,...,Q}, ( Z j 1 )j∈{1,...,7}) (238) appartienne à (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 x )−1( Z ). 190 Démonstration. Pour σ ∈ {0, 1} et z , z ′ ∈ Z j σ , on choisit b ∈ Sσ , d’où z , z ′ ∈ 2N -géod(x, b) et comme d(x, z) = d(x, z ′ ), (Hδ (z , x, z ′ , b)) donne 0 et Z j d(z , z ′ ) ≤ 2N + δ ≤ P /3. Comme les parties Z j 1 sont non vides et P /3 ≤ M , l’argument que nous venons de donner montre aussi que les condi- tions (235), (236) et (237) sont vérifiées par les autres éléments de la classe d’équivalence Z de l’élément (238) et donc Z ∈ (ΛZ,t,(t1 ,...,tQ ))v1 ,v2 ,v3 u1 ,u2 ,u3 . (cid:3) Suite de la démonstration du e). On déduira e) de l’inégalité suivante : il existe C2 = C (δ, K, N , Q, P , M , r, T ) tel que en notant u1 , u2 , u3 , v1 , v2 , v3 comme dans (232), on ait τ ∈[0,T ] (cid:12)(cid:12)(cid:12) ξZ (cid:16)(cid:16)(cid:0)(θ♭ u1 ,u2 ,u3 (cid:1)eτ θx ux,r,tKx,Q,(t1 ,...,tQ )e−τ θx − x′ )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ x )v1 ,v2 ,v3 sup 2 u1 , u2 , u3 (cid:1)(cid:17)f (cid:17)(cid:12)(cid:12)(cid:12) x )v1 ,v2 ,v3 x′ )v1 ,v2 ,v3 eτ θx ux,r,tKx,Q,(t1 ,...,tQ )e−τ θx (cid:0)(θ♭ u1 , u2 , u3 − (θ♭ 2 X u1 ,u2 ,u3 (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) ≤ C2 v1 ,v2 ,v3 Z∈(Λ 6= Z,t,(t1 ,...,tQ ) ) u1 ,u2 ,u3 est l’ensemble des Z ∈ (ΛZ,t,(t1 ,...,tQ ))v1 ,v2 ,v3 où (Λ 6= Z,t,(t1 ,...,tQ ))v1 ,v2 ,v3 u1 ,u2 ,u3 tels que pour tout (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 0 )j∈{1,...,Q} , ( Z j i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j 1 )j∈{1,...,7}) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 )−1( Z ) x (239) on ait (240) u1 , u2 , u3 ( S0) − (ρ♭ u1 , u2 , u3 ( S0) x )v1 ,v2 ,v3 x′ )v1 ,v2 ,v3 (ρ♭ u1 ,u2 ,u3 ( S1) 6= 0. u1 ,u2 ,u3 ( S1) + (ρ♭ x′ )v1 ,v2 ,v3 −(ρ♭ x )v1 ,v2 ,v3 Nous allons maintenant montrer (239). Nous allons justifier aussi que la condition (240) ne dépend que de Z (c’est-à-dire que pour Z ∈ (ΛZ,t,(t1 ,...,tQ ) )v1 ,v2 ,v3 u1 ,u2 ,u3 elle est vérifiée ou non simultanément pour tous les éléments de (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 )−1 ( Z )). x Sous-lemme 4.89 Pour Z ∈ (ΛZ,t,(t1 ,...,tQ ) )v1 ,v2 ,v3 u1 ,u2 ,u3 et 0 )j∈{1,...,Q} , ( Z j i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j 1 )j∈{1,...,7}) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 )−1( Z ) x le coefficient de e S1 dans (cid:16)(cid:0)(θ♭ u1 ,u2 ,u3 (cid:1)eτ θx ux,r,tKx,Q,(t1 ,...,tQ ) e−τ θx − x′ )v1 ,v2 ,v3 x )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ u1 , u2 , u3 (cid:1)(cid:17)(e S0 ) x )v1 ,v2 ,v3 x′ )v1 ,v2 ,v3 eτ θx ux,r,tKx,Q,(t1 ,...,tQ )e−τ θx (cid:0)(θ♭ u1 , u2 , u3 − (θ♭ 191 et le membre de gauche de (240) ne dépendent que de la connaissance des points de B (x, k + 2M ) ∪ B ( S0 , M ) ∪ B ( S1 , M ) B ( Z j B ( Z j ∪ [j∈{1,...,Q} 0 , M ) ∪ [j∈{1,...,7} 1 , M ) et des distances entre ces points. (241) Démonstration. Le coefficient de e S1 dans eτ θx ux,r,tKx,Q,(t1 ,...,tQ )e−τ θx (e S0 ) ne dépend que de la connaissance des points de (241) et de leurs distances mutuelles. Cela résulte du sous-lemme 4.43 ou du sous-lemme 4.81 (que l’on applique avec Z 7 0 au lieu de Z 1 0 et en oubliant Z j 0 pour j ∈ {1, ..., 6}). Pour montrer le sous-lemme il suffit donc de montrer que pour x ∈ u1 , u2 , u3 ( S0) ne dépendent que de la connais- u1 ,u2 ,u3 ( S1) et (ρ♭ x )v1 ,v2 ,v3 x )v1 ,v2 ,v3 B (x, 1), (ρ♭ sance des points de (241) et des distances entre ces points (en effet on ap- plique ceci à x = x et x = x′ ). En vertu du lemme 4.50 il suffit de montrer que pour x ∈ B (x, 1), σ ∈ {0, 1}, a ∈ Sσ et j ∈ {1, ..., 6}, l’ensemble {y ∈ 3δ -géod( x, a), d( x, y ) − wj (σ, x) ≤ N + 6δ + 4} est inclus dans (241), où l’on note (242) λ1 = λ1 = u1 6 u1 6 , λ2 = , λ3 = 5 − v2 6 − v1 2 + u3 1 + u2 6 6 6 6 5 − v2 6 − v1 2 + u3 1 + u2 , λ2 = , λ6 = , λ5 = , λ4 = , λ3 = 6 6 6 6 wj (0, x) = E (λj d( x, S0)) et wj (1, x) = E (λj d( x, S1 )). , λ4 = , λ5 = , λ6 = 4 − v3 6 4 − v3 6 , , On rappelle que wj = E (λj d(x, S1)). On commence par le cas où σ = 1. Soit x ∈ B (x, 1), a ∈ S1 , j ∈ {1, ..., 6} et y dans l’ensemble (242). Soit z ∈ géod(x, a) vérifiant d(x, z) = wj , si bien que z appartient à Z j 1 . Comme d(x, S1)−d( x, S1) ≤ 1 on a wj (1, x)−wj ≤ 1 et donc d(x, y ) − d(x, z) ≤ N + 6δ + 6. Comme y et z appartiennent à (3δ+2)-géod(x, a), (Hδ (y , x, z , a)) implique d(y , z) ≤ (N +6δ+6)+(3δ+2)+δ . On suppose (N + 6δ + 6) + (3δ + 2) + δ ≤ M , ce qui est permis par (HM ). On a donc d(y , z) ≤ M et y appartient à (241). On considère maintenant le cas où σ = 0. Soit x ∈ B (x, 1), a ∈ S0 , j ∈ {1, ..., 6} et y dans l’ensemble (242). On commence par montrer wj (0, x) − wj ≤ QF + 2. (243) 192 1+κ et r1 ( Z )+1+r On a λj = λj 1+κ = r1( Z ) + 1, d’où λj (r1( Z ) + 1 + r) = λj (r1 ( Z ) + 1). D’après (233) on a (r0( Z ) + 1) − (r1( Z ) + 1 + r) ≤ QF . Donc λj (r0( Z ) + 1) − λj (r1( Z ) + 1) ≤ QF , et on en déduit immédiatement λj r0( Z ) − λj r1 ( Z ) ≤ QF + 1. Comme r0( Z ) = d(x, S0 ) et r1( Z ) = d(x, S1), (243) en résulte aisément. Soit b ∈ S1 et z ∈ géod(x, b) vérifiant d(x, z) = wj si bien que z appartient à Z j 1 . Il résulte facilement de (243) que (244) d(x, y ) − d(x, z) ≤ QF + N + 6δ + 7. Comme b ∈ 2F -géod(x, a) on a z ∈ 2F -géod(x, a). On a y ∈ (3δ+2)-géod(x, a) et comme 2F ≥ 3δ + 2, y et z appartiennent à 2F -géod(x, a). Grâce à (244), (Hδ (y , x, z , a)) implique d(y , z) ≤ (QF + N + 6δ + 7) + 2F + δ . On sup- pose (QF + N + 6δ + 7) + 2F + δ ≤ M , ce qui est permis par (HM ). On a donc d(y , z) ≤ M et y appartient à (241). Ceci termine la démonstration du sous-lemme 4.89. (cid:3) Une conséquence immédiate du sous-lemme 4.89 est que la condition (240) ne dépend que de Z . Sous-lemme 4.90 Le cardinal de (ΛZ,t,(t1 ,...,tQ) )v1 ,v2 ,v3 u1 ,u2 ,u3 est majoré par une constante de la forme C (δ, K, N , Q, P , M ). Démonstration. En effet, grâce au lemme 4.28, pour connaître les distances entre les points de B ( S0 , M ) ∪ [j∈{1,...,Q} B ( Z j 0 , M ) ∪ [j∈{1,...,7} B ( Z j 1 , M ) (245) B ( Si , M ) ∪ et ceux de [i∈{1,...,m+1} [ i∈{0,...,m},j∈{1,...,li } il suffit de connaître les distances entre les points de (245) et C points de (246), avec C = C (δ, K, N , Q, P , M ) et grâce à (233) ces distances sont dé- terminées à C ′ = C (δ, K, N , Q, P , M ) près par les distances de S1 à ces C B ( Y j i+1 , M ) ∪ B (x, k + 2M ) (246) 193 0 ( Z ))j∈{1,...,Q} et points (qui font partie de la donnée de Z ) et par les entiers (tj 1 ( Z ))j∈{1,...,7} , qui sont eux-mêmes déterminés à C ′′ = C (δ, K, N , Q, P , M ) (tj près par r0 (Z ), r, t, t1 , ..., tQ , u1 , u2 , u3 , v1 , v2 , v3 . (cid:3) Suite de la démonstration du e). On termine maintenant la preuve de (239). Pour 0 )j∈{1,...,Q} , ( Z j (a1 , . . . , ap−1 , S0 , ..., Sm+1 , ( Y j i )i∈{0,...,m+1},j∈{1,...,li } , ( Z j 1 )j∈{1,...,7}) ∈ (π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 )−1( Z ) x (247) )−1(Z ). on considère X(b1 ,...,bp) (cid:16)(cid:16)(cid:0)(θ♭ u1 ,u2 ,u3 (cid:1)eτ θx ux,r,tKx,Q,(t1 ,...,tQ ) e−τ θx − u1 ,u2 ,u3 − (θ♭ x′ )v1 ,v2 ,v3 x )v1 ,v2 ,v3 u1 , u2 , u3 (cid:1)(cid:17) x )v1 ,v2 ,v3 x′ )v1 ,v2 ,v3 eτ θx ux,r,tKx,Q,(t1 ,...,tQ ) e−τ θx (cid:0)(θ♭ u1 , u2 , u3 − (θ♭ (ea1 ∧ ... ∧ eap−1 )(cid:17)(b1 , ..., bp ), où la somme porte sur les énumérations (b1 , ..., bp ) de S1 telles que (b1 , . . . , bp , S1 , ..., Sm+1 , ( Y j i+1)i∈{0,...,m},j∈{1,...,li }) ∈ (π p,k ,m,(l0 ,...,lm) x Comme la somme (247) a au plus p! termes, le 3) de la proposition 3.37 montre qu’elle est ma jorée par une constante de la forme C (δ, K, N , Q, P , M , T ). D’après le sous-lemme 4.89 la somme (247) ne dépend que de Z et on peut donc la noter (αZ, Z ,t,τ ,(t1 ,...,tQ ))v1 ,v2 ,v3 u1 ,u2 ,u3 . D’après le sous-lemme 4.88 et le sous- lemme 4.42 (ou 4.80), on a ξZ (cid:16)(cid:16)(cid:0)(θ♭ u1 ,u2 ,u3 (cid:1)eτ θx ux,r,tKx,Q,(t1 ,...,tQ ) e−τ θx − x′ )v1 ,v2 ,v3 x )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ u1 , u2 , u3 (cid:1)(cid:17)f (cid:17) x )v1 ,v2 ,v3 x′ )v1 ,v2 ,v3 eτ θx ux,r,tKx,Q,(t1 ,...,tQ ) e−τ θx (cid:0)(θ♭ u1 , u2 , u3 − (θ♭ 1 X (αZ, Z ,t,τ ,(t1 ,...,tQ ))v1 ,v2 ,v3 = u1 ,u2 ,u3 ξ Z (f ). (p − 1)! v1 ,v2 ,v3 Z∈(Λ 6= Z,t,(t1 ,...,tQ ) ) u1 ,u2 ,u3 Par Cauchy-Schwarz et grâce au sous-lemme 4.90 on en déduit (239). Montrons maintenant e) à l’aide de (239). Le sous-lemme 4.87 appli- qué à ρ = r + C avec C = C (δ, K, N , Q) implique facilement qu’il existe C1 = C (δ, K, N , Q, r) tel que pour S0 , S1 comme ci-dessus la mesure de l’en- semble des (u1 , u2 , u3 , v1 , v2 , v3 ) ∈ [5κ, 1 − 5κ[6 vérifiant (240) est ≤ C1 1+r0 (Z ) . Soit C2 comme dans (239). Soient k , m, l0 , . . . , lm ∈ N et posons l0 = 0 194 et li = li−1 pour i ∈ {1, . . . , m + 1} comme précédemment. Soit C3 = C (δ, K, N , Q, P , M , r) comme dans le sous-lemme 4.83. Soit p,k ,m,(l0 ,...,lm) x Z ∈ Y vérifiant r0 (Z ) > k + P . On a alors Z(u1 ,u2 ,u3 ,v1 ,v2 ,v3 )∈[5κ,1−5κ[6 (cid:16) du1 ...dv3 ≤ ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 X x v1 ,v2 ,v3 Z ∈(Λ 6= Z,t,(t1 ,...,tQ ) ) u1 ,u2 ,u3 C1C3 )−1 (Z )(cid:1), ♯(cid:0)(π p,k ,m,(l0 ,...,lm) x r0 (Z ) + 1 )−1( Z )(cid:1)(cid:17) Notons IZ l’ensemble des (t, t1 , ..., tQ , u1 , u2 , u3 , v1 , v2 , v3) ∈ [0, 1]Q+1 × [5κ, 1 − 5κ[6 tels qu’il existe Z ∈ (Λ 6= Z,t,(t1 ,...,tQ ) )v1 ,v2 ,v3 u1 ,u2 ,u3 vérifiant ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 )−1( Z )(cid:1) ≥ (r0(Z ) + 1)− 1 )−1(Z )(cid:1). 2 ♯(cid:0)(π p,k ,m,(l0 ,...,lm) x x La mesure de IZ est donc ≤ C1C3(r0(Z ) + 1)− 1 2 . Grâce à Cauchy-Schwarz on déduit de (239) que ξZ (cid:16)(cid:16) Z(t,t1 ,...,tQ,u1 ,u2 ,u3 ,v1 ,v2 ,v3 )∈IZ τ ∈[0,T ] (cid:12)(cid:12)(cid:12) sup (cid:16)(cid:0)(θ♭ u1 ,u2 ,u3 (cid:1)eτ θx ux,r,tKx,Q,(t1 ,...,tQ )e−τ θx x )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ x′ )v1 ,v2 ,v3 2 u1 , u2 , u3 (cid:1)(cid:17)dt...dv3(cid:17)f (cid:17)(cid:12)(cid:12)(cid:12) x )v1 ,v2 ,v3 x′ )v1 ,v2 ,v3 −eτ θx ux,r,tKx,Q,(t1 ,...,tQ ) e−τ θx (cid:0)(θ♭ u1 , u2 , u3 − (θ♭ 2 . ≤ 2Q+766C1C2C3(r0 (Z ) + 1)− 1 (r0 ( Z ) + 1)−Q (r1( Z ) + 1)−7 (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) 2 XZ ∈ΛZ (248) D’après le sous-lemme 4.83 et le lemme 4.29, il existe C4 = C (δ, K, N , Q, P , M , r) tel que pour Z ∈ ΛZ on ait ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 )−1( Z )(cid:1) ≤ C4♯(cid:0)(π p,k ,m,(l0 ,...,lm) )−1(Z )(cid:1). (249) x x D’autre part pour (t, t1 , ..., tQ , u1 , u2 , u3 , v1 , v2 , v3 ) ∈ (cid:0)[0, 1]Q+1×[5κ, 1−5κ[6\IZ (cid:1) et Z ∈ (Λ 6= Z,t,(t1 ,...,tQ ) )v1 ,v2 ,v3 u1 ,u2 ,u3 on a ♯(cid:0)(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 )−1( Z )(cid:1) ≤ (r0(Z ) + 1)− 1 )−1(Z )(cid:1). 2 ♯(cid:0)(π p,k ,m,(l0 ,...,lm) x x 195 Par Cauchy-Schwarz on déduit alors de (239) que )−1 (Z )(cid:1)−α (cid:0)♯(π p,k ,m,(l0 ,...,lm) x ξZ (cid:16)(cid:16) Z(t,t1 ,...,tQ,u1 ,u2 ,u3 ,v1 ,v2 ,v3 )∈(cid:0)[0,1]Q+1×[5κ,1−5κ[6 \IZ (cid:1) τ ∈[0,T ] (cid:12)(cid:12)(cid:12) sup (cid:16)(cid:0)(θ♭ u1 ,u2 ,u3 (cid:1)eτ θx ux,r,tKx,Q,(t1 ,...,tQ )e−τ θx x′ )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ x )v1 ,v2 ,v3 2 u1 , u2 , u3 (cid:1)(cid:17)dt...dv3(cid:17)f (cid:17)(cid:12)(cid:12)(cid:12) x′ )v1 ,v2 ,v3 x )v1 ,v2 ,v3 −eτ θx ux,r,tKx,Q,(t1 ,...,tQ ) e−τ θx (cid:0)(θ♭ u1 , u2 , u3 − (θ♭ ≤ 2Q+766C2(r0(Z ) + 1)− α (r0 ( Z ) + 1)−Q (r1( Z ) + 1)−7 2 XZ ∈ΛZ )−1( Z )(cid:1)−α (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) 2 (cid:0)♯(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 (250) . x En combinant les inégalités (248), (249) et (250) et par Cauchy-Schwarz on obtient que )−1(Z )(cid:1)−α (cid:0)♯(π p,k ,m,(l0 ,...,lm) sup x τ ∈[0,T ] ξZ (cid:16)(cid:16) Z(u1 ,u2 ,u3 ,v1 ,v2 ,v3 )∈[5κ,1−5κ[6 (cid:16)(cid:0)(θ♭ (cid:12)(cid:12)(cid:12) u1 ,u2 ,u3 (cid:1)eτ θx ux,rKxe−τ θx x′ )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ x )v1 ,v2 ,v3 2 u1 , u2 , u3 (cid:1)(cid:17)du1 ...dv3(cid:17)f (cid:17)(cid:12)(cid:12)(cid:12) x )v1 ,v2 ,v3 x′ )v1 ,v2 ,v3 −eτ θx ux,rKxe−τ θx (cid:0)(θ♭ u1 , u2 , u3 − (θ♭ ≤ 2Q+866(cid:16)C1C2C3C α 2 (cid:17) XZ∈ΛZ 4 (r0(Z ) + 1)− 1 2 + C2(r0(Z ) + 1)− α (r0( Z ) + 1)−Q )−1( Z )(cid:1)−α (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) 2 (r1 ( Z ) + 1)−7(cid:0)♯(π ♮,p−1,k ,m+1,(l0 ,...,lm+1),Q,7 . x i=0 si ( Z )−li . En sommant sur De plus pour Z ∈ ΛZ on a Qm i=0 si (Z )−li = Qm+1 les Z tels que r0 (Z ) = n, et en posant κ = r 1+n , on déduit donc de (251) que pour tout n > P , (Pn − Pn−1 )(cid:16)(cid:16) Z[5κ,1−5κ[6 (cid:0)(θ♭ τ ∈[0,T ] (cid:13)(cid:13)(cid:13) u1 ,u2 ,u3 (cid:1)du1 ...dv3(cid:17) x′ )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ x )v1 ,v2 ,v3 sup eτ θx ux,rKxe−τ θx − eτ θx ux,rKxe−τ θx (cid:16) Z[5κ,1−5κ[6 (cid:0)(θ♭ 2 u1 , u2 , u3 (cid:1)du1 ...dv3(cid:17)(cid:17)(cid:13)(cid:13)(cid:13) x )v1 ,v2 ,v3 x′ )v1 ,v2 ,v3 u1 , u2 , u3 − (θ♭ L(H♮,Q,7 (∆p−1 ),H→ x,s (∆p )) x,s 2 (cid:17) ≤ 2Q+866p!B e2(QF −r)s(cid:16)C1C2C3C α 2 + C2(n + 1)− α 4 (n + 1)− 1 où ( u1 , u2 , u3 , v1 , v2 , v3) est défini par (232). Le facteur p! est dû au fait que Z détermine Z à permutation près de a1 , ..., ap . On rappelle que Pn − Pn−1 196 (251) est le pro jecteur orthogonal défini par (Pn − Pn−1 )(eS ) = eS si d(x, S ) = n et (Pn − Pn−1 )(eS ) = 0 sinon. x,s (∆p−1) et H→ D’après les lemmes 4.24 et 4.20, les normes de H♮,Q,7 x,s(∆p ) sont équivalentes à celles de Hx,s(∆p−1) et Hx,s(∆p ). Donc il existe C = C (δ, K, N , Q, P , M , s, B , r, T ) tel que pour tout n > P , (Pn − Pn−1 )(cid:16)(cid:16) Z[5κ,1−5κ[6 (cid:0)(θ♭ τ ∈[0,T ] (cid:13)(cid:13)(cid:13) u1 ,u2 ,u3 (cid:1)du1 ...dv3(cid:17) x′ )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ x )v1 ,v2 ,v3 sup eτ θx ux,rKxe−τ θx − eτ θx ux,rKxe−τ θx (cid:16) Z[5κ,1−5κ[6 (cid:0)(θ♭ 2 u1 , u2 , u3 (cid:1)du1 ...dv3(cid:17)(cid:17)(cid:13)(cid:13)(cid:13) x )v1 ,v2 ,v3 x′ )v1 ,v2 ,v3 u1 , u2 , u3 − (θ♭ L(Hx,s (∆p−1 ),Hx,s (∆p )) ≤ C (n + 1)− α 2 où κ = r 1+n et où ( u1 , u2 , u3 , v1 , v2 , v3) est défini par (232). (252) Sous-lemme 4.91 Il existe C = C (δ, K, N , Q, P , M , s, B ) tel que pour p ∈ {1, ..., pmax} et κ ∈]0, 1 10 [, (cid:13)(cid:13)(cid:13) Z[5κ,1−5κ[6 (cid:0)(θ♭ x )v1 ,v2 ,v3 x′ )v1 ,v2 ,v3 u1 , u2 , u3 − (θ♭ u1 , u2 , u3 (cid:1)du1 ...dv3− Z[0,1[6 (cid:0)(θ♭ 2 u1 ,u2 ,u3 (cid:1)du1 ...dv3(cid:13)(cid:13)(cid:13) x′ )v1 ,v2 ,v3 x )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ L(Hx,s (∆p ),Hx,s (∆p )) ≤ C κ et (cid:13)(cid:13)(cid:13) Z[5κ,1−5κ[6 (cid:0)(θ♭ x′ )v1 ,v2 ,v3 x )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ u1 ,u2 ,u3 (cid:1)du1 ...dv3− Z[0,1[6 (cid:0)(θ♭ 2 u1 ,u2 ,u3 (cid:1)du1 ...dv3(cid:13)(cid:13)(cid:13) x′ )v1 ,v2 ,v3 u1 ,u2 ,u3 − (θ♭ x )v1 ,v2 ,v3 L(Hx,s (∆p ),Hx,s (∆p )) ≤ C κ. Démonstration. Comme kθx−θx′ kL(Hx,s (∆p ),Hx,s (∆p )) ≤ 1, il suffit de montrer qu’il existe C = C (δ, K, N , Q, P , M , s, B ) tel que pour x ∈ {x, x′} on ait (cid:13)(cid:13)(cid:13) Z[5κ,1−5κ[6 (cid:0)(θ♭ x )v1 ,v2 ,v3 u1 , u2 , u3 − θ x(cid:1)du1 ...dv3− Z[0,1[6 (cid:0)(θ♭ 2 u1 ,u2 ,u3 − θ x(cid:1)du1 ...dv3(cid:13)(cid:13)(cid:13) x )v1 ,v2 ,v3 L(Hx,s (∆p ),Hx,s (∆p )) ≤ C κ et (cid:13)(cid:13)(cid:13) Z[0,1[6 \[5κ,1−5κ[6 (cid:0)(θ♭ 2 u1 ,u2 ,u3 − θ x(cid:1)du1 ...dv3(cid:13)(cid:13)(cid:13) x )v1 ,v2 ,v3 L(Hx,s (∆p ),Hx,s (∆p )) ≤ C κ. (254) 197 (253) (255) f (u1 , u2 , u3 , v1 , v2 , v3 )du1 ...dv3 Grâce à l’équivalence des normes de Hx,s(∆p ) et Hx′ ,s(∆p ), il suffit de montrer (253) et (254) pour x = x. Soient hκ,1 et hκ,2 les fonctions mesurables sur [0, 1[6 définies de la manière suivante : hκ,1 est telle que pour toute fonction continue f sur [0, 1]6 on a Z[5κ,1−5κ[6 f ( u1 , u2 , u3 , v1 , v2 , v3 )du1 ...dv3 − Z[0,1[6 = Z[0,1[6 f hκ,1du1 ...dv3 et hκ,2 est la fonction caractéristique de [0, 1[6\[5κ, 1 − 5κ[6 . Il suffit donc de montrer qu’il existe C = C (δ, K, N , Q, P , M , s, B ) tel que pour i ∈ {1, 2} on ait (cid:13)(cid:13)(cid:13) Z[0,1[6 (cid:0)(θ♭ 2 u1 ,u2 ,u3 − θx(cid:1)hκ,i (u1 , u2 , u3 , v1 , v2 , v3 )du1 ...dv3(cid:13)(cid:13)(cid:13) x )v1 ,v2 ,v3 L(Hx,s (∆p ),Hx,s (∆p )) ≤ C κ. A partir de maintenant on reprend les notations ΛZ et (ΛZ )v1 ,v2 ,v3 u1 ,u2 ,u3 de la démonstration du lemme 4.49. Ces notations sont définies entre les formules (121) et (123). Pour montrer (255) il suffit de montrer qu’il existe C = C (δ, K, N , Q, P , M , s, B ) p,k ,m,(l0 ,...,lm) et f ∈ C(∆p ) tel que pour i ∈ {1, 2}, k , m, l0 , . . . , lm ∈ N, Z ∈ Y x on ait ξZ (cid:16)(cid:16) Z[0,1[6 (cid:0)(θ♭ (cid:12)(cid:12)(cid:12) u1 ,u2 ,u3 − θx(cid:1)hκ,i (u1 , u2 , u3 , v1 , v2 , v3 )du1 ...dv3(cid:17)(f )(cid:17)(cid:12)(cid:12)(cid:12) x )v1 ,v2 ,v3 ≤ C κ XZ ∈ΛZ (cid:0)r0 ( Z ) + 1(cid:1)−6 (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) 2 . (256) Or (122) montre qu’il existe C = C (δ, K, N , Q, P , M , s, B ) tel que Z[0,1[6 (cid:12)(cid:12)(cid:12) 2 u1 ,u2 ,u3 − θx(cid:1)(f )(cid:17)(cid:12)(cid:12)(cid:12) ξZ (cid:16)(cid:0)(θ♭ du1 ...dv3 ≤ C XZ ∈ΛZ (cid:0)r0 ( Z ) + 1(cid:1)−6 (cid:12)(cid:12)ξ Z (f )(cid:12)(cid:12) 2 x )v1 ,v2 ,v3 (257) D’autre part on vérifie facilement qu’il existe une constante C telle que pour i ∈ {1, 2}, on ait Z[0,1[6 hκ,i (u1 , u2 , u3 , v1 , v2 , v3)2du1 ...dv3 ≤ C κ. 198 2 . (258) Par Cauchy-Schwarz et grâce à (257) et (258), on obtient (256). Ceci termine la démonstration du sous-lemme 4.91. (cid:3) Fin de la démonstration du e). D’après le lemme 4.47, il existe C = C (δ, K, N , Q, P , M , s, B , T ) tel que τ ∈[0,T ] keτ θx ux,rKxe−τ θx kL(Hx,s (∆p−1 ),Hx,s (∆p )) ≤ C. sup (259) (260) 2 L(Hx,s (∆p−1 ),Hx,s (∆p )) En combinant (252), le sous-lemme 4.91 et (259) on voit qu’il existe C = C (δ, K, N , Q, P , M , s, B , r, T ) tel que pour tout n > P , x′ ), eτ θx ux,rKxe−τ θx ](cid:13)(cid:13) x − θ♭ τ ∈[0,T ] (cid:13)(cid:13)(Pn − Pn−1)[(θ♭ sup ≤ C (n + 1)− α 2 . Pour tout i ∈ Z soit Ti ∈ L(Hx,s(∆p−1), Hx,s(∆p )) défini par (Pn − Pn−1)Ti (Pn′ − Pn′−1 ) x′ ), eτ θx ux,rKxe−τ θx ](Pn′ − Pn′−1) si n′ − n = i x − θ♭ = (Pn − Pn−1 )[(θ♭ (Pn − Pn−1)Ti (Pn′ − Pn′−1 ) = 0 sinon. et D’après le sous-lemme 4.42 (ou 4.80) on a Ti = 0 sauf si i − r ≤ QF . Il est x′ ), eτ θx ux,rKxe−τ θx ] = Pi Ti . x − θ♭ clair que [(θ♭ D’après (260) il existe donc C = C (δ, K, N , Q, P , M , s, B , r, T ) tel que pour tout i ∈ Z, 2 L(Hx,s (∆p−1 ),Hx,s (∆p )) ≤ C (n + 1)− α (cid:13)(cid:13)(1 − Pn )Ti(cid:13)(cid:13) 2 . En sommant sur i ∈ {r − QF , ..., r + QF } on en déduit qu’il existe C = C (δ, K, N , Q, P , M , s, B , r, T ) tel que pour tout n > P , 2 L(Hx,s (∆p−1 ),Hx,s (∆p )) ≤ C (n + 1)− α x′ ), eτ θx ux,rKxe−τ θx ](cid:13)(cid:13) (cid:13)(cid:13)(1 − Pn )[(θ♭ x − θ♭ 2 . Ceci termine la preuve de e) et donc celle du lemme 4.70. (cid:3) On a donc montré les lemmes 4.69 et 4.68, et la proposition 4.4 qui était l’énoncé principal de ce paragraphe. 5 Fin de l’homotopie Ce paragraphe n’offre guère d’intérêt car il recopie quasiment la fin du paragraphe 2.3.4 et le paragraphe 2.3.5 de [Laf02]. 199 Lemme 5.1 Pour T assez grand – a) on a keT θ♭ xKxe−T θ♭ ℓ2 (∆p )) ≤ 1 x kL(Lpmax 2 , p=1 x Hxe−T θ♭ – b) il existe C tel que keT θ♭ ℓ2 (∆p )) ≤ C et x kL(Lpmax p=1 x ux,re−T θ♭ keT θ♭ ℓ2 (∆p )) ≤ C 2−r . x kL(Lpmax p=1 Démonstration. Grâce à 2) a) du lemme 3.34, Kx (eS ) est une combinaison de eT où T vérifie d(x, T ) ≤ d(x, S ) − (Q N −6δ pmax − 2N − 4δ ), d’où ρ♭ x (T ) ≤ x (S ) − (Q N −6δ pmax − 3N − 11δ ). On suppose (Q N −6δ ρ♭ pmax − 3N − 11δ ) ≥ 1, ce qui est permis par (HQ ). Ceci permet de montrer que la condition a) est réalisée pour T assez grand et pour b) on utilise 1)a) et 3) de la proposition 3.37, la proposition 3.29 et le lemme 3.33. (cid:3) On prend T assez grand pour que les conditions du lemme 5.1 soient x est continu sur Lpmax satisfaites. En particulier eT θ♭ x (∂ + Jx∂Jx )e−T θ♭ p=1 ℓ2(∆p ). On peut relier les espaces de Hilbert Hx,s et Lpmax p=1 ℓ2 (∆p ) par un champ continu d’espaces de Hilbert (Hx,s,α)α∈[0,1] , défini par k.k2 Hx,s,α = αk.kℓ2 + (1 − α)k.kHx,s sur Lpmax p=1 C(∆p ) . Lemme 5.2 ((Hx,s,α)α∈[0,1] , eT θ♭ x (∂ + Jx∂Jx )e−T θ♭ x ) appartient à KKG,2sℓ+C (C, C[0, 1]) et réalise donc une homotopie entre x (∂ + Jx∂Jx )e−T θ♭ (Hx,s , eT θ♭ x ) pmax x (∂ + Jx∂Jx )e−T θ♭ ℓ2 (∆p), eT θ♭ Mp=1 x ) ( x (∂ + Jx∂Jx )e−T θ♭ Démonstration. La continuité de eT θ♭ x résulte de la pro- position 4.46, du lemme 5.1 et du fait que pour tout opérateur U on a (261) et (262) (263) kU kL(Hx,s,α ) ≤ max(kU kL(Hx,s ) , kU kL(ℓ2 ) ). Pour montrer que eT θ♭ x (∂ + Jx∂Jx )e−T θ♭ x est équivariant à compact près, on voit, en reprenant l’argument de la démonstration de la proposition 4.68 et du lemme 4.69, qu’il suffit de montrer l’énoncé analogue à celui du lemme 4.70 obtenu en prenant τ = T mais en remplaçant les normes k.kHx,s par les normes k.kHx,s,α et en prenant le supremum sur α. Par exemple l’énoncé analogue à a) du lemme 4.70 est que α∈[0,1] k(1 − Pn )eT θx (hx − hx′ )e−T θx kL(Hx,s,α (∆p−1 ),Hx,s,α (∆p )) sup 200 tend vers 0 quand n → ∞. Par (263) il suffit de montrer que k(1 − Pn )eT θx (hx − hx′ )e−T θx kL(ℓ2 (∆p−1 ),ℓ2 (∆p )) → 0 quand n → ∞. (264) Cela résulte du lemme 3.14 (et même simplement du fait, mentionné avant le lemme 3.14, que kψS,x − ψS,x′ k1 tend vers 0 en dehors des parties finies de ∆). Il reste à montrer les énoncés analogues à (264) correspondant aux opérateurs de b), c), d) et e) du lemme 4.70. Pour cela on utilise de plus la propriété (10) et le lemme 3.45 (ou même l’énoncé plus faible qui est la propriété (23) avec d♭ au lieu de d′′ ). (cid:3) En fait (262) appartient à KKG,0(C, C) (c’est-à-dire qu’il vérifie les mêmes conditions qu’un élément de KKG (C, C) sauf celle qui assure que l’opé- rateur est auto-adjoint à compact près). De plus l’image de (262) dans KKG,2sℓ+C (C, C) est égale à 1 d’après le lemme 5.2 et les propositions 4.4 et 4.5. Pour terminer la preuve du théorème 1.5, il suffit donc de montrer le lemme suivant, dont la preuve suit de très près le paragraphe 2.3.5 de [Laf02]. Lemme 5.3 L’élément (262) est égal à l’image de γ ∈ KKG (C, C) dans KKG,0 (C, C). Démonstration. D’après le a) du lemme 5.1 et comme Kx = (1 − ∂hx − hx∂ )Q , on a ρ(eT θ♭ x (1 − ∂hx − hx∂ )e−T θ♭ x ) ≤ 2−Q−1 où ρ désigne le rayon spectral dans L(ℓ2(∆p )). On rappelle que Hx = hx (∂hx + hx∂ )−1 . x agit continûment sur Lpmax x Hxe−T θ♭ Donc eT θ♭ p=1 ℓ2(∆p ). Sous-lemme 5.4 L’élément pmax x (∂ + Hx )e−T θ♭ ℓ2(∆p ), eT θ♭ Mp=1 x ) ( appartient à KKG,0 (C, C) et (262) est homotope à (266). (265) (266) Démonstration. La continuité de eT θ♭ x (∂ + Hx )e−T θ♭ x a été justifiée avant le lemme et l’équivariance à compact près résulte de (264). On a H 2 x = 0 car on a vu dans le lemme 3.16 que h2 x = 0. L’homotopie entre (262) et (266) résulte du lemme 1.4.1 de [Laf02]. (cid:3) On pose Dx = ∂ (∂hx + hx∂ )−1 . 201 Sous-lemme 5.5 L’élément pmax Mp=1 appartient à KKG,0 (C, C) et (266) est homotope à (267). x (hx + Dx)e−T θ♭ ℓ2 (∆p ), eT θ♭ x ) ( (267) Démonstration. La continuité de eT θ♭ x (∂ + Dx )e−T θ♭ x vient de nouveau de (265) et son équivariance à compact près résulte de (264). Pour réaliser l’homotopie entre (266) et (267) on procède comme dans le lemme 2.3.11 de [Laf02]. Pour α ∈ [0, 1] on définit grâce à (265) et en utilisant la détermi- nation principale du logarithme, x = hx (∂hx + hx∂ )−α et Dα x = ∂ (∂hx + hx∂ )−α . H α Alors pmax ℓ2(∆p )(cid:1)[0, 1], (eT θ♭ x )e−T θ♭ Mp=1 x (H 1−α x + Dα (cid:0)(cid:0) x )α∈[0,1](cid:1) appartient à KKG,0(C, C[0, 1]) et réalise une homotopie entre (266) et (267). (cid:3) (268) On rappelle maintenant quelques notations de [Laf02]. On note φS,x = pψS,x . On note f , h′ les opérateurs de ℓ2(∆p ) dans ℓ2(∆p+1) donnés par 1 et h′ (eS ) = f (eS ) = φS,x ∧ eS ψS,x ∧ eS . kφ3 S,xk1 On note g , g ′ les opérateurs de ℓ2 (∆p) dans ℓ2(∆p−1) donnés par g (eS ) = φS,xyeS 1 kφ3 S,xk1 On note aussi h = hx pour être cohérent avec [Laf02]. Alors (267) est homotope à φS,xyeS . et g ′(eS ) = pmax Mp=1 grâce au lemme 1.4.1 de [Laf02]. Puis (269) est homotope à x (h + g ′ )e−T θ♭ ℓ2(∆p ), eT θ♭ x ) ( ℓ2 (∆p ), h + g ′) ( pmax Mp=1 202 (269) (270) par l’homotopie évidente pmax ℓ2(∆p )(cid:1)[0, T ], (cid:0)eτ θ♭ x (h + g ′)e−τ θ♭ Mp=1 (cid:0)(cid:0) x (cid:1)τ ∈[0,T ](cid:1). On note que l’homotopie entre (267) et (270) correspond au lemme 2.3.12 de [Laf02]. Enfin (270) est homotope à (Lpmax p=1 ℓ2(∆p ), h′ + g ) puis à pmax Mp=1 ℓ2(∆p ), f + g ) ( (voir les lemmes 2.3.13 et 2.3.14 de [Laf02]). Or (271) est égal à l’image de γ ∈ KKG (C, C) dans KKG,0(C, C). On renvoie à [KS03] pour la construction complète de γ , qui est rappelée au début de la section 2 de [Laf02]. Il y a une toute petite subtilité due au fait que dans [KS03] la construction utilise des mesures ψK S S,x qui sont légèrement différentes des mesures ψS,x que nous avons définies. Cependant on peut les relier par une homotopie α 7→ (1−α)ψS,x +αψK S S,x . En effet grâce au lemme 3.7 et au lemme 6.3 de [KS03], pour tout T tel que eT apparaisse dans ψS,x ∧ eS , ψS,xyeS , ψK S S,x ∧ eS ou ψK S yeS , on a ψS,x = ψT ,x et ψK S S,x = ψK S T ,x . Donc les S,x opérateurs analogues à f , g , g ′ , h, h′ construits à l’aide de (1 − α)ψS,x + αψK S S,x sont de carré nul. Ceci termine la preuve du lemme 5.3. (cid:3) On a donc terminé la preuve de théorème 1.5. (271) Références [BC82] P. Baum and A. Connes. Geometric K -theory for Lie groups and foliations (preprint de 1982). Enseign. Math. (2), 46 :3–42, 2000. [BCH94] P. Baum, A. Connes, and N. Higson. Classifying space for proper actions and K -theory of group C ∗ -algebras. In C ∗ -algebras : 1943– 1993 (San Antonio, TX, 1993), volume 167 of Contemp. Math., pages 240–291. Amer. Math. Soc., Providence, RI, 1994. [CDP90] M. Coornaert, T. Delzant, and A. Papadopoulos. Géométrie et théorie des groupes, volume 1441 of Lecture Notes in Mathema- tics. Springer-Verlag, 1990. Les groupes hyperboliques de Gromov. [Gromov hyperbolic groups], With an English summary. [GdlH90] E. Ghys and P. de la Harpe, editors. Sur les groupes hyperboliques d’après Mikhael Gromov, volume 83 of Progress in Mathematics. Birkhäuser Boston Inc., Boston, MA, 1990. Papers from the Swiss Seminar on Hyperbolic Groups held in Bern, 1988. 203 [Jul97] [JV84] [Jul02] [Gro87] M. Gromov. Hyperbolic groups. Essays in group theory, 75–263, Math. Sci. Res. Inst. Publ., 8, Springer, New York, 1987. [Gro03] M. Gromov. Random walk in random groups. Geom. Funct. Anal., 13(1) :73–146, 2003. [HK01] N. Higson and G. Kasparov. E -theory and KK -theory for groups Invent. which act properly and isometrically on Hilbert space. Math., 144(1) :23–74, 2001. [HLS02] N. Higson, V. Lafforgue, and G. Skandalis. Counterexamples to the Baum-Connes conjecture. Geom. Funct. Anal., 12(2) :330–354, 2002. P. Julg. Remarks on the Baum-Connes conjecture and Kazhdan’s property T . In Operator algebras and their applications (Waterloo, ON, 1994/1995), volume 13 of Fields Inst. Commun., pages 145– 153. Amer. Math. Soc., 1997. P. Julg. La conjecture de Baum-Connes à coefficients pour le groupe Sp(n, 1). C. R. Math. Acad. Sci. Paris, 334(7) :533–538, 2002. P. Julg and A. Valette. K-theoretic amenability for SL2(Qp ), and the action on the associated tree. J. Funct. Anal., 58(2) :194–215, 1984. [Kas88] G.G. Kasparov. Equivariant KK-theory and the Novikov conjec- ture. Invent. Math. 91, pages 147–201, 1988. [KS94] G. Kasparov and G. Skandalis. Groupes boliques et conjecture de Novikov. C.R.A.S,Paris, Série I(319) :815–820, 1994. [KS03] G. Kasparov and G. Skandalis. Groups acting properly on “bolic” spaces and the Novikov conjecture. Ann. of Math. (2), 158(1) :165– 206, 2003. [Laf02] V. Lafforgue. K -théorie bivariante pour les algèbres de Banach et conjecture de Baum-Connes. Invent. Math., 149(1) :1–95, 2002. [Laf07] V. Lafforgue (avec un appendice de Hervé Oyono-Oyono). K - théorie bivariante pour les algèbres de Banach, groupoïdes et conjecture de Baum-Connes. J. Inst. Math. Jussieu 6(3) : 415–451, 2007. [Laf08] V. Lafforgue. Un renforcement de la propriété (T). Duke Math. J., 143(3) :559–602, 2008. [Laf10] V. Lafforgue. Propriété (T ) renforcée et conjecture de Baum- Connes. Quanta of maths, Clay Math. Proc. 11, Amer. Math. Soc., Providence, RI, 2010, 323–345. 204 [MY02] I. Mineyev and G. Yu. The Baum-Connes conjecture for hyperbolic groups. Invent. Math., 149(1) :97–122, 2002. [Oza08] N. Ozawa. Weak amenability of hyperbolic groups. Groups Geom. Dyn., 2(2) : 271–280, 2008. [Tu99] [Vai05] J.-L. Tu. La conjecture de Novikov pour les feuilletages hyperbo- liques. K -Theory, 16(2) :129–184, 1999. J. Väisälä. Gromov hyperbolic spaces. Expo. Math. 23(3) :187–231, 2005. Index des notations. Page 1 : Hδ (x, y , z , t) Page 2 : B (x, r) Page 16 : d(A, B ), dmax (A, B ), Page 17 : N , (HN ), ∆, ∆p , C(∆p ) , pmax Page 17 : ∂ Page 21 : (H β δ (x, a, b, c)), (H 0 δ (x, a, b, c)), US Page 23 : AS,x , YS,x,r , A∅,x , Y∅,x,r Page 26 : χA , νA , ψS,x , ψS,x,t, E (.) Page 29 : hx , hx,t Page 29 : Hx , ζx (S ) Page 34 : Φp Page 36 : Ax,a,r,k , µr (x, a), µr,t(x, a) Page 38 : ux , ux,r,t, up x,r,t, v p x,r,t Page 45 : (HQ), Hx , Kx Page 45 : Kx,q ,(t1 ,...,tq ) , Hx,q ,(t1 ,...,tq ) , B (A, r) Page 47 : Jx Page 50 : F x,r , αs′←s x,y , Λy ,s Page 57 : Y r r ′←r Page 58 : Λy ,s′ ,s x,r ′ ,r x,r , Λy ,s1 ,s2 ,s3 Page 59 : Λy ,s1 ,s2 ,s3 x,r1 ,r2 ,r3 , β y ,s x,r1 ,r2 ,r3 x,u1 ,u2 ,u3 , d♭(x, y ), d♭ v1 ,v2 ,v3 Page 60 : Ay ,v1 ,v2 ,v3 u1 ,u2 ,u3 (x, y ) Page 66 : ρ♭ x (a), ρ♭ x (S ) Page 67 : (HP ), Y p,k ,m,(l0 ,...,lm) , Y j x i p,k ,m,(l0 ,...,lm) , π p,k ,m,(l0 ,...,lm) Page 68 : (HM ), Y x x Page 69 : (HB ), (Hα), ξZ , Hx,s(∆p ) Page 71 : θ♭ x , r0(Z ), . . . , rm(Z ), s0(Z ), . . . , sm(Z ) 205 →,p,k ,m,(l0 ,...,lm) x , π→,p,k ,m,(l0 ,...,lm) x , Z j i , π ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x i,j (Z ), tj , rmax i (Z ) , Y Page 72 : Ψ Page 88 : Y →,p,k ,m,(l0 ,...,lm) x Page 89 : k.kH→ x,s (∆p ) Page 91 : P Page 94 : Y ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 x ♮,p,k ,m,(l0 ,...,lm),λ0 ,λ1 Page 95 : Y x Page 96 : H♮,µ0 ,µ1 (∆p ) x,s Page 99 : κσ , κσ,∞ , κσ,i Page 122 : ρx , θx , ρ0 x (Z ), ρ1 x (Z ) x )v1 ,v2 ,v3 Page 123 : (ρ♭ u1 ,u2 ,u3 p,k ,m,(l0 ,...,lm) Page 129 : π p,k ,m,(l0 ,...,lm) , Y x,x′ x,x′ p,k ,m,(l0 ,...,lm) , π p,k ,m,(l0 ,...,lm) Page 131 : Y x,x′ ,⋆ x,x′ ,⋆ Page 167 : Pn , r ′ 0(Z ), k ′ (Z ) 206
1703.07997	2	1703	2017-10-10T09:30:16	Polynomials in operator space theory: matrix ordering and algebraic aspects	[ "math.OA" ]	We extend the $\lambda$-theory of operator spaces given by Defant and Wiesner (2014), that generalizes the notion of the projective, Haagerup and Schur tensor norm for operator spaces to matrix ordered spaces and Banach $$-algebras. Given matrix regular operator spaces and operator systems, we introduce cones related to $\lambda$ for the algebraic tensor product that respect the matricial structure of matrix regular operator spaces and operator systems, respectively. The ideal structure of $\lambda$-tensor product of $C^$-algebras has also been discussed.	math.OA	math	POLYNOMIALS IN OPERATOR SPACE THEORY: MATRIX ORDERING AND ALGEBRAIC ASPECTS PREETI LUTHRA, AJAY KUMAR∗, AND VANDANA RAJPAL Abstract. We extend the λ-theory of operator spaces given in [4], that gen- eralizes the notion of the projective, Haagerup and Schur tensor norm for operator spaces to matrix ordered spaces and Banach ∗-algebras. Given ma- trix regular operator spaces and operator systems, we introduce cones related to λ for the algebraic tensor product that respect the matricial structure of matrix regular operator spaces and operator systems, respectively. The ideal structure of λ-tensor product of C ∗-algebras has also been discussed. . A O h t a m [ 2 v 7 9 9 7 0 . 3 0 7 1 : v i X r a 1. Introduction C∗-algebras are rich objects as they come along with matrix norms that are not only uniquely related to algebraic structure but are also known to have matricial cone structures being closely related to those norm. Although, operator spaces and their tensor products are primarily defined in terms of appropriate matrix norms, over the years it has been observed that some operator space tensor products of C∗-algebras still possess few algebraic properties that can be characterized in terms of the individual algebras ([1, 14]). Regarding ordering, although operator spaces may possess some order structure unrelated to the matrix norms, it was Schreiner [18] who defined matrix regular operator spaces to be the spaces where there is a relationship between norm and order. In matrix regular operator spaces, there are enough positive elements so that each element can be written as a linear combination of positive elements. Recently introduced tensor product theory for (unital) operator systems category ([13]) shows that this matrix order-matrix norm relation is successfully carried over. Defant and Wiesner in [4] (see also [19]) have given a λ-theory which generalizes the definitions of the projective, Haagerup and Schur tensor norm for operator It is thus natural to ask for appropriate matrix ordering and algebraic spaces. structure that is compatible with this generalized λ-theory. In [9] and [17], the projective and Schur operator space tensor product of matrix ordered operator spaces are shown to be matrix ordered respectively. Further, Han in [8] successfully introduced cones at each matrix level of the tensor product of operator spaces that are closely related to projective and injective operator space tensor norms thereby, constructing two extremal tensor products of matrix regular operator space. Section 2 discusses the prerequisites. Next, we introduce conditions (O1)-(O3) in Section 3 that enables generalization of Han's ([8]) operator space tensor product Key words and phrases. Matrix regular operator spaces, operator systems, tensor products, ideals. Mathematics Subject Classification (2010): Primary 46L06, 46L07; Secondary 46L05, 47L25 . ∗Corresponding author 1 2 P. LUTHRA, A. KUMAR, AND V. RAJPAL matrix regularity results to λ-theory of operator spaces. In Section 4, we show that the cones defined in Section 3 also preserve the operator system structure. Finally in Section 5, we show that the techniques to study ideal structure of operator space tensor product of C∗-algebras can be extended to λ-theory. 2. Preliminaries 2.1. The λ-theory [4][19]. Let V1, V2, . . . , Vm; W be operator spaces and let φ be an m-linear mapping on V1 × V2 × · · · × Vm into W . Given a sequence of matrix products λ = (λk), for each k, λk is an m-linear mapping: λk : Mk × · · · × Mk → Mτ (k), where τ (k) ∈ N is a natural number only depending on k, tensorizing λk with φ leads to the m-linear mapping φλk := λk ⊗ φ : Mk(V1) ⊗ · · · ⊗ Mk(Vm) → Mτ (k)(W ), (α1 ⊗ v1, · · · , αm ⊗ vm) 7→ λk(α1, · · · , αm) ⊗ φ(v1, v2, · · · , vm). Further and kφkcb,λ := sup k∈N {kφλk (x1, . . . , xm)kMτ (k)(W ) : kxikMk(Vi) ≤ 1} CBλ(V1, . . . , Vm; W ) := {φ ∈ L(V1, . . . , Vm; W ) : kφkcb,λ < ∞}. Since m-fold tensor product on V1 ×V2 ×· · ·×Vm is an m-linear map onto ⊗m i=1Vi, the natural map obtained as above by tensorizing with λk is represented by ⊗λk : i=1Vi), (α1 ⊗ v1, · · · , αm ⊗ vm) 7→ λk(α1, · · · , αm) ⊗ v1 ⊗ v2 ⊗ · · · ⊗ vm. ⊗λk : Mk(V1) ⊗ · · · ⊗ Mk(Vm) → Mτ (k)(⊗m In [4, 19], a tensor norm λ was defined as: kukλ,k = inf{kαkkv1kkv2k · · · kvmkkβk} (1) for any element u ∈ Mk(⊗m positions u = α ⊗λj (v1, v2, · · · , vm)β, α ∈ Mk,τ (j), β ∈ Mτ (j),k, vt ∈ Mj(Vt). i=1Vi), where the infimum is taken over arbitrary decom- Keeping the notations from [19, 4] unchanged, e.g. εi,j := ε[k,l] i,j ∈ Mk,l de- notes the matrix which is 1 in the (i, j)-th coordinate and zero elsewhere, ε[k] i,j := ε[k,k] i,j = 0 if (i, j) /∈ {1, . . . , k} × {1, . . . , l}, i,j we state the three technical conditions (E1)-(E3) that were isolated on the family λ = (λn)n∈N to assure that the k · kλ,k, generates an operator space structure on V1 ⊗ V2 ⊗ · · · ⊗ Vm [4, Proposition 4.1]: := εi,i, ε[k] i , εi := ε[k,k] i,i and ε[k,l] (E1) For all k ∈ N there exist p ∈ N and matrices S ∈ Mk,τ (p) , T ∈ Mτ (p),k, a1, · · · , ak ∈ Mp such that for all j1, · · · , jm ∈ {1, · · · , k}: Sλp(aj1 , · · · , ajm )T =(cid:26) ε[k] j 0 if j1 = j2 = · · · = jm = j, otherwise. (E2) For all r, s ∈ N there exist matrices P ∈ Mτ (r)+τ (s),τ (r+s), with kP k ≤ 1 such that for all (ik, jk) ∈ {1, · · · , r}2 ∪ {r + 1, · · · , r + s}2 with 1 ≤ k ≤ m: MATRIX ORDERING AND ALGEBRAIC ASPECTS OF λ-THEORY 3 P λr+s(ε[r+s] i1,j1 , . . . , ε[r+s] im,jm = diag(cid:16)λr(ε[r] i1,j1 , · · · , ε[r] (E3) λ1(1, 1, · · · , 1) = 1 and sup k∈N )P ∗ im,jm ), λs(ε[s] i1−r,j1−r, · · · , ε[s] kλkk < ∞. im−r,jm−r)(cid:17). If in addition λ satisfies: (N1) then ⊗λC = C completely isometric [19, Proposition 4.13]. τ (1) = 1 and (N2) kλjk = 1 for all j ∈ N, For j ∈ {1, . . . , m}, if λ further satisfy conditions: (W1) For all γ ∈ Mp there exists matrices P ∈ Mp,τ (p), Q ∈ Mτ (p),p with kP k, kQk ≤ 1 such that γ = P λp(Ip, · · · , Ip, , Ip, · · · , Ip)Q j-th position γ {z} (W2) For all α1, · · · , αm ∈ Mp, β1, · · · , βm ∈ Mq there exist matrices S ∈ Mτ (p)τ (q),τ (pq), T ∈ Mτ (pq),τ (p)τ (q) with kSk, kT k ≤ 1 such that λp(α1, · · · , αm) ⊗ λq(β1, · · · , βm) = Sλpq(α1 ⊗ β1, · · · , αm ⊗ βm)T then the mapping Φ(j) : (V1 ⊗ · · · ⊗ Mp(Vj ) ⊗ · · · ⊗ Vm, k · kλ) → Mp(⊗λVi) v1 ⊗ · · · ⊗ (α ⊗ vj) ⊗ · · · vm 7→ α ⊗ (v1 ⊗ · · · ⊗ vm) is completely contractive [19, Proposition 12.2]. If λ satisfies (N1)-(N2), (E1)-(E3) and (W1)-(W2) then ⊗λVi, the completion of ⊗λVi with respect to k · kλ norm, is an operator space tensor product denoted by λ-operator space tensor product in the sense of [3]. The Kronecker product, matrix product and mixed product fulfill all the above conditions. We assume throughout that λ satisfies all the prescribed conditions. 2.2. Matrix regular operator space and operator systems. An operator space V is called a matrix ordered operator space if: (1) (V, {Mn(V )+}∞n=1) is a matrix ordered vector space i.e. for each n ∈ N, Mn(V ) is a ∗-ordered vector space with cone Mn(V )+ and A ∈ Mn,m implies A∗Mn(V )+A ⊆ Mm(V )+. (2) the ∗-operation is an isometry on Mn(V ). (3) the cones Mn(V )+ are closed. A matrix ordered operator space V is called matrix regular [18, Definition 3.1.9] if for each n ∈ N and for all v ∈ Mn(V )sa, the following conditions hold : (1) u ∈ Mn(V )+ and −u ≤ v ≤ u implies that kvkn ≤ kukn. 4 P. LUTHRA, A. KUMAR, AND V. RAJPAL (2) kvkn ≤ 1 implies that there exists u ∈ Mn(V )+ such that kukn ≤ 1 and −u ≤ v ≤ u. Next result from [18] giving a necessary and sufficient for a matrix ordered op- erator space V to be matrix regular is quite useful: Theorem 2.1. [18, Theorem 3.4] A matrix ordered operator space V is matrix regular if and only if the following condition holds: for all x ∈ Mn(V ), kxkn < 1 if and only if there exist a, d ∈ Mn(V )+, kakn < 1 and kdkn < 1, such that (cid:18) a x∗ x d(cid:19) ∈ M2n(V )+. The positive cone of a matrix regular operator space is always proper. Adopting the methodology of [8], the norms on matrix regular operator spaces are not assumed to be complete. For a matrix ordered operator space V and its dual space V ∗, the positive cone on Mn(V ∗) for each n ∈ N is defined by Mn(V ∗)+ = CB(V, Mn) ∩ CP (V, Mn). The operator space dual V ∗ with this positive cone is a matrix ordered operator space [18, Corollary 3.2]. An (abstract) operator system ([13, Definition 2.2]) is a triple (V, {Cn}∞n=1, e), where V is a complex ∗-vector space, {Cn}∞n=1 is a matrix ordering on V, and e ∈ Vsa is an Archimedean matrix order unit, i.e. for all v ∈ Mn(V )sa, (1) there exists a real number r > 0 such that ren > v and (2) for each n ∈ N and en = v ∈ Cn. e . . . , sen + v ∈ Cn for all s > 0 implies e 3. λ-theory and Matrix regularity In this section, we provide three additional conditions on λ = (λn)n∈N to in- troduce an order structure to λ-theory that preserves matrix regularity. Further, using our conditions (O1)-(O3) defined below, we prove that the results of [8] hold true in a more general setting introduced by [4, 19]. For a sequence λ = (λn)n∈N of m-linear mappings λk ∈ L(mMk; Mτ (k)) consider the following three properties: (O1) For each r ∈ N, λr(ε[r] j2,i2 , · · · , ε[r] for all (ik, jk) ∈ {1, . . . , r} × {1, . . . , r}, and k = 1, 2, . . . m. i2,j2 , · · · , ε[r] ) = λr(ε[r] j1,i1 , ε[r] i1,j1, ε[r] im,jm jm,im )∗ ∈ Mτ (r), (O2) For r ∈ N, the permutation matrix P ∈ M2τ (r),τ (2r) with kP k ≤ 1 obtained in (E2) and (ik, jk) ∈ R ∪ S, where R := {1, · · · , r} × {r + 1, r + 2, · · · , 2r} and S := {r + 1, r + 2, · · · , 2r} × {1, · · · , r} P λ2r(ε[2r] i1,j1 , · · · , ε[2r] im,jm )P ∗ = adiag(cid:0)λr(ε[r] i1,j1−r, · · · , ε[r] i1,j1−r), λr(ε[r] i1−r,j1, · · · , ε[r] im−r,jm )(cid:1); MATRIX ORDERING AND ALGEBRAIC ASPECTS OF λ-THEORY 5 adiag being an anti-diagonal matrix, where all the entries are zero except those on the diagonal going from the upper right corner to the lower left corner. (O3) For each r ∈ N, the map p1...pm ⊗λr = p1...pm ⊗ ⊗ λr : Mr(Mp1) ⊗ Mr(Mp2 ) ⊗ . . . ⊗ Mr(Mpm ) → Mτ (r)(Mp1p2...pm) obtained by tensorizing λr with the Kronecker product on matrix algebras Mp1 , . . . , Mpm p1...pm ⊗ : Mp1 × Mp2 × . . . × Mpm → Mp1p2...pm (α1, α2, . . . , αm) 7→ α1 ⊗ α2 ⊗ . . . ⊗ αm, is positive for all pi ∈ N (i = 1, 2, . . . , m). Thus, p1...pm ⊗λr (α1 ⊗ β1, . . . , αm ⊗ βm) = λr(β1 ⊗ . . . ⊗ βm) ⊗ α1 ⊗ . . . ⊗ αm ∈ M + τ (r)p1p2...pm , whenever αi ⊗ βi ∈ (Mpi ⊗ Mr)+, pi ∈ N (i = 1, 2, . . . m). Recall from [18, Proposition 4.1] (see also [19, Proposition 4.2]), given a se- quence λ = (λn) of m-linear maps and operator spaces V1, V2, . . . , Vm any ele- ment u ∈ Mk(V1 ⊗ . . . Vm) has a representation u = α ⊗λr (v1, v2, . . . , vm)β where α ∈ Mn,τ (r), β ∈ Mτ (r),n, vi ∈ Mr(Vi), r ∈ N. Next, we analyze the above conditions in view of their applications to matrix ordered spaces: Lemma 3.1. Let λ = (λn) be sequence of m-linear maps and V1, . . . Vm be matrix ordered operator spaces. For any α ⊗λr (v(1), v(2), . . . , v(m))β ∈ Mn(⊗λVi); α ∈ Mn,τ (r), β ∈ Mτ (r),n, v(i) ∈ Mr(Vi), i = 1, . . . , m, r ∈ N, we have: (i) If λ satisfies (O1), then ∗-map defined as is a well defined involution. (ii) If λ satisfies (O2), then for u(i), u(i) ∈ Mr(Vi), i = 1, . . . , m (cid:0)α ⊗λr (v(1), v(2), . . . , v(m))β(cid:1)∗ = β∗ ⊗λr(cid:0)(v(1))∗, (v(2))∗, . . . , (v(m))∗(cid:1)α∗, (cid:18) ⊗λr(cid:0)(v(1))∗, (v(2))∗, . . . , (v(m))∗(cid:1) ⊗λr (u(1), u(2), . . . , u(m))(cid:19) = P ⊗λ2r (cid:18) u(1) u(m)(cid:19)!P ∗. u(1)(cid:19) , . . . ,(cid:18) u(m) ⊗λr (u(1), u(2), . . . , u(m)) ⊗λr (v(1), v(2), . . . , v(m)) (v(1))∗ (v(m))∗ v(1) v(m) (iii) For λ and µ, if (µp)λr is a positive map (p, r ∈ N), then for v(i) ∈ Mr(Vi)+ and completely positive maps φ(i) : Vi → Mpi , i = 1, . . . , m, we have In particular, if λ satisfies (O3), (cid:0) ⊗µp (φ(1), . . . , φ(m))(cid:1)n(α ⊗λr (v(1), v(2), . . . , v(m))α∗) ∈ M + (cid:0)(φ(1) ⊗ . . . ⊗ φ(m))(cid:1)n(α ⊗λr (v(1), v(2), . . . , v(m))α∗) ∈ M + Proof. To obtain (i) one can easily verify that the ∗-operation is conjugate linear and involutive. np1...pm. np1...pm. 6 P. LUTHRA, A. KUMAR, AND V. RAJPAL (ii) We have ⊗λr (v(1), v(2), . . . , v(m)) ⊗λr (u(1), u(2), . . . , u(m)) (cid:18) ⊗λr ((v(1))∗, (v(2))∗, . . . , (v(m))∗) ⊗λr (u(1), u(2), . . . , u(m))(cid:19) = diag(cid:0) ⊗λr (u(1), u(2), . . . , u(m)), ⊗λr (u(1), u(2), . . . , u(m)(cid:1)+ adiag(cid:0) ⊗λr (v(1), v(2), . . . , v(m)), ⊗λr ((v(1))∗, (v(2))∗, . . . , (v(m))∗)(cid:1). let u(t) := X(kt,lt)∈R1 (v(t))∗ = X(kt,lt)∈R3 Setting R1 := {1, 2, . . . , r}2, R2 := {1, 2, . . . , r} × {r + 1, r + 2, . . . , 2r}, R3 := {r + 1, r + 2, . . . , 2r} × {1, 2, . . . , r} and R4 := {r + 1, r + 2, . . . , 2r}2, ε[r] kt,lt−r ⊗ v(t) Define x(t) := diag(u(t), u(t)) and y(t) := adiag(v(t), (v(t))∗), so that ε[r] kt−r,lt ε[r] kt−r,lt ⊗ (v(t) kt,lt ⊗ u(t) kt,lt ⊗ u(t) kt,lt ε[r] kt,lt kt,lt , . , v(t) := X(kt,lt)∈R2 )∗ and u(t) := X(kt,lt)∈R4 k,l =( v(t) k,l (v(t))∗k,l if (k, l) ∈ R1, if (k, l) ∈ R4 ) and y(t) if (k, l) ∈ R2, if (k, l) ∈ R3 ) . k,l =( u(t) k,l u(t) k,l x(t) Then, λr(ε[r] adiag(cid:0) ⊗λr (v(1), v(2), . . . , v(m)), 0) · · · X(k1,l1)∈R2 = adiag(cid:0) X(km,lm)∈R2 = X(km,lm)∈R2 = X(km,lm)∈R2 · · · X(k1,l1)∈R2 · · · X(k1,l1)∈R2 (O2) adiag(cid:0)λr(ε[r] Also, k1,l1−r, · · · , ε[r] k1,l1−r, · · · , ε[r] k1,l1 km,lm−r) ⊗ v(1) km,lm−r), 0(cid:1) ⊗ y(1) ⊗ · · · ⊗ v(m) km,lm , 0(cid:1) ⊗ · · · ⊗ y(m) km,lm k1,l1 P λ2r(ε[2r] k1,l1 , · · · , ε[2r] km,lm )P ∗ ⊗ y(1) k1,l1 ⊗ · · · ⊗ y(m) km,lm λr(ε[r] adiag(cid:0)0, ⊗λr((v(1))∗, (v(2))∗, . . . , (v(m))∗)) · · · X(k1,l1)∈R3 = adiag(cid:0)0, X(km,lm)∈R3 = X(km,lm)∈R3 · · · X(k1,l1)∈R3 adiag(cid:0)0, λr(ε[r] · · · X(k1,l1)∈R3 = X(km,lm)∈R3 P λ2r(ε[2r] k1,l1 (O2) k1−r,l1 k1−r,l1 , · · · , ε[r] km−r,lm ) ⊗ (v(1) k1,l1 )∗ ⊗ · · · ⊗ (v(m) km,lm , · · · , ε[r] km−r,lm(cid:1) ⊗ y(1) k1,l1 ⊗ · · · ⊗ y(m) km,lm )∗(cid:1) , · · · , ε[2r] km,lm )P ∗ ⊗ y(1) k1,l1 ⊗ · · · ⊗ y(m) km,lm Thus, adiag(cid:0) ⊗λr (v(1), v(2), . . . , v(m)), ⊗λr ((v(1))∗, (v(2))∗, . . . , (v(m))∗)(cid:1) = X(km,lm)∈R2∪R3 · · · X(k1,l1)∈R2∪R3 P λ2r(ε[2r] k1,l1 , · · · , ε[2r] )P ∗ ⊗ y(1) k1,l1 km,lm Similarly, using (E2), ⊗ · · · ⊗ y(m) km,lm MATRIX ORDERING AND ALGEBRAIC ASPECTS OF λ-THEORY 7 Therefore, , · · · , ε[2r] P λ2r(ε[2r] k1,l1 ⊗λr (v(1), v(2), . . . , v(m)) ⊗λr (u(1), u(2), . . . , u(m)) · · · X(k1,l1)∈R1∪R4 diag(cid:0) ⊗λr (u(1), u(2), . . . , u(m)), ⊗λr (u(1), u(2), . . . , u(m)(cid:1) = X(km,lm)∈R1∪R4 (cid:18) ⊗λr ((v(1))∗, (v(2))∗, . . . , (v(m))∗) ⊗λr (u(1), u(2), . . . , u(m)(cid:19) = P(cid:16) X(km,lm)∈∪4 = P ⊗λ2r (cid:18) u(1) (iii) Note that, if v(t) := X(kt,lt)∈R · · · X(k1,l1)∈∪4 u(1)(cid:19) , . . . ,(cid:18) u(m) u(m)(cid:19)!P ∗. ⊗λ2r(cid:0)ε[2r] ⊗ x(1) k1,l1 · · · , ε[2r] ⊗ v(t) kt,lt ⊗ x(m) (v(m))∗ (v(1))∗ ε[r] kt,lt v(m) i=1Ri km,lm km,lm i=1Ri v(1) k1,l1 then, (⊗µp (φ(1), . . . , φ(m))τµ(p)(⊗λr (v(1), v(2), · · · , v(m))) )P ∗ ⊗ x(1) k1,l1 ⊗ · · · ⊗ x(m) km,lm km,lm + ε[2r] k1,l1 ⊗ y(1) k1,l1 , · · · + ε[2r] km,lm ⊗ y(m) km,lm(cid:1)(cid:17)P ∗ ; R := {1, . . . , r}2, t = 1, 2, . . . , m . . .Xk1,l1 = Xkm,lm . . .Xk1,l1 = Xkm,lm = Xkm,lm . . .Xk1,l1 = Xkm,lm . . .Xk1,l1 . . .Xk1,l1 = Xkm,lm = (µp)λr(cid:0)φ(1) ∈ M + τ (r)p1p2...pm ⊗ . . . ⊗ v(m) k1,l1 ⊗ . . . ⊗ v(m) km,lm km,lm(cid:1) )(cid:1) (⊗µp (φ(1), . . . , φ(m))τµ(p)(cid:0)λr(εk1,l1 , . . . , εkm,lm) ⊗ v(1) λr(εk1,l1, . . . , εkm,lm) ⊗(cid:0) ⊗µp (φ(1), . . . , φ(m))(v(1) λr(εk1,l1, . . . , εkm,lm) ⊗ µp(cid:0)φ(1)(v(1) (µp)λr(cid:0)εk1,l1 ⊗ φ(1)(v(1) (µp)λr(cid:0)φ(1) ), . . . , εkm,lm ⊗ φ(m)(v(m) r (εk1,l1 ⊗ v(1) k1,l1 (εkm,lm ⊗ v(m) ), . . . , φ(m) km,lm km,lm k1,l1 k1,l1 k1,l1 r )(cid:1) ))(cid:1) ) ⊗ . . . ⊗ φ(m)(v(m) km,lm )(cid:1) r (v(1)), . . . , φ(m) r (v(m))(cid:1) so that, (⊗µp (φ(1), . . . , φ(m))n(α ⊗λr (v(1), . . . , v(m))α∗)) = α(cid:0)(⊗µ(p)(φ(1), . . . , φ(m)))τµ(p)(⊗λj (v1, v2, · · · , vm))(cid:1)α∗ ∈ M + p1...pm If λ satisfies (O3), µp = ⊗ gives the desired result. nk1k2...km . (cid:3) 8 P. LUTHRA, A. KUMAR, AND V. RAJPAL Verification of Properties (O1)-(O3): • Kronecker product: Property (O1) reduces to j1,i1 ⊗ ε[r] = (ε[r] im,jm j2,i2 ⊗ · · · ⊗ ε[r] i2,j2 ⊗ · · · ⊗ ε[r] i1,j1 ⊗ ε[r] ε[r] which is true. To check for the condition (O3), recall that Kronecker product of two posi- tive matrices is positive, but Kronecker product does not commute, in fact for any square matrices A and B, there exists a permutation matrix S such that B ⊗A = S(A⊗B)S∗. Therefore, for some suitable permutation matrix S we have: jm,im )∗, p1...pm ⊗⊗r (α1 ⊗ β1, . . . , αm ⊗ βm) = (β1 ⊗ . . . ⊗ βm) ⊗ α1 ⊗ . . . ⊗ αm = S(cid:0)(α1 ⊗ β1) ⊗ · · · ⊗ (αm ⊗ βm)(cid:1)S∗ rmp1p2...pm, ∈ M + whenever αi ⊗ βi ∈ Mr(Mpi)+, pi ∈ N (i = 1, 2, . . . m). In order to verify (O2), we use the same notations as in proof of [4, Propo- sition 4.2], let ∆ : M1 → M1,m, x 7→ (x, x, . . . , x) and set ε[∆r,∆2r] p,p P1 := Xr∈{1,2,...,r}m and let P =(cid:18)P1 P2(cid:19) and P2 := Xr∈{r+1,r+2,...,2r}m ε[∆r,∆2r] p−∆r,p , ε[∆r,∆r] (i1,··· ,im),(j1,j2,··· ,jm)−∆r (i1,··· ,im)−∆r,(j1,j2,··· ,jm)−∆r ! ε[∆r,∆r] if (ik, jk) ∈ {1, · · · , r} × {r + 1, · · · , 2r}, if (ik, jk) ∈ {r + 1, · · · , 2r} × {1, · · · , r} else i1,j1, · · · , ε[2r] im,jm )P ∗ (i1,...,im),(j1,...,jm)P ∗ (i1,··· ,im),(j1,j2,··· ,jm) ε[∆r,∆r] (i1,··· ,im)−∆r,(j1,j2,··· ,jm) adiag(cid:0)ε[∆r,∆r] adiag(cid:0)0, ε[∆r,∆r] (i1,··· ,im),(j1,j2,··· ,jm)−∆r, 0(cid:1) (i1,··· ,im)−∆r,(j1,j2,··· ,jm)(cid:1) P ⊗2r (ε[2r] = P ε[∆2r,∆2r] = ε[∆r,∆r] = 0 • Schur product: Here property (O1) takes the form im,jm = (ε[r] j1,i1 ⊙ ε[r] j2,i2 ⊙ · · · ⊙ ε[r] i2,j2 ⊙ · · · ⊙ ε[r] i1,j1 ⊙ ε[r] ε[r] which is true. To check for the condition (O3), recall that Schur product of two positive matrices is positive, for any square matrices A, B, C and D of order n, (A⊙B)⊗(C ⊙D) = (A⊗C)⊙(B ⊗D) (see [19, Proposition 10.5]) and there exist a matrix E ∈ Mn,k, A′, B′ ∈ Mk such that (A ⊗ B) = E(A′ ⊙ B′)E∗. Therefore we have, jm,im )∗, p1...pm ⊗⊙r (α1 ⊗ β1, . . . , αm ⊗ βm) = (β1 ⊙ . . . ⊙ βm) ⊗ α1 ⊗ . . . ⊗ αm ∈ M + rp1p2...pm , MATRIX ORDERING AND ALGEBRAIC ASPECTS OF λ-THEORY 9 whenever αi ⊗ βi ∈ Mr(Mpi)+, pi ∈ N (i = 1, 2, . . . m). Again as for (E2), for r ∈ N and (iq, jq) ∈ [{1, · · · , r} × {r + 1, · · · , 2r}] ∪ [{r + 1, · · · , 2r} × {1, · · · , r}], let P := I2r, we have P ⊙2r (ε[2r] im,jm i1,j1, · · · , ε[2r] 0 ε[r] i1−r,j1 ε[r] i1,j1−r 0 )P ∗ ! ⊙ · · · ⊙ 0 ε[r] i1−r,j1 ⊙ · · · ⊙ ε[r] im−r,jm = = 0 ε[r] im−r,jm ε[r] im,jm−r 0 i1,j1−r ⊙ · · · ⊙ ε[r] ε[r] im,jm−r 0 ! ! implying that (O2) holds. • Matrix Product: One can easily see that this product may not satisfy (O1), (O2) and (O3). • Mixed Product: One can mix above listed products to construct a new one, for example [19, Chapter 9], λ = (λk)k with λk : Mk × Mk × Mk × Mk → Mk2 , (α1, α2, α3, α4) 7→ (α1 • α2) ⊗ (α3 • α4) Clearly, it does not satisfy any of the (O1)-(O3) as • does not. One can similarly talk of λ = (λk)k with λk : Mk × Mk × Mk × Mk → Mk2 , (α1, α2, α3, α4) 7→ (α1 ⊙ α2) ⊗ (α3 ⊙ α4), which clearly satisfies all the conditions (O1)-(O3). (cid:3) The self-adjoint elements in Mn(⊗λVi) have a special representation: Proposition 3.2. Let Vi, 1 ≤ i ≤ m, be matrix ordered operator spaces and let λ = (λn)n∈N be a sequence of m-multilinear mappings satisfying (O1) and (O2). If u ∈ Mn(⊗λVi)sa, then u has a representation: u = α ⊗λj (x1, x2, · · · , xm)α∗ where α ∈ Mn,τ (j), xt ∈ Mj(Vt)sa, t = 1, 2, · · · , m. Moreover, kukλ,n = inf(cid:8)kαk2kx1kkx2k · · · kxmk : u = α ⊗λj (x1, x2, · · · , xm)α∗, α ∈ Mn,τ (j), xt ∈ Mj(Vt)sa, j ∈ N(cid:9). Proof. Suppose u ∈ Mn(⊗λVi)sa. Given ǫ > 0, there exist α ∈ Mn,τ (j), β ∈ Mτ (j),n, xt ∈ Mj(Vt) such that u = α ⊗λj (x1, x2, · · · , xm)β with kukλ,n ≤ kαkkx1kkx2k · · · kxmkkβk ≤ kukλ,n + ǫ. 10 P. LUTHRA, A. KUMAR, AND V. RAJPAL As u is self adjoint by Lemma 3.1(ii), for any µ > 0, we have u = = 1 2 1 2 (u + u∗) (µα ⊗λj (x1, x2, · · · , xm)µ−1β + µ−1β∗ ⊗λj (x∗1, x∗2, · · · , x∗m)µα∗) 0 √2 ⊗λj (x1, x2, · · · , xm) =(cid:16) µ−1β ∗ = αP(cid:0) ⊗λ2j (v1, v2, · · · , vm)(cid:1)P ∗α∗ where vt = (cid:18) 0 µα√2 (cid:17)(cid:18) 0 (cid:19), and α = (cid:16) µ−1β ∗ √2 x∗t xt ⊗λj (x∗1, x∗2, · · · , x∗m) 0 (cid:19) µ−1β√2 √2 ! µα∗ µα√2 (cid:17), R = {1, 2, · · · , r} × {r + 1, · · · , 2r} and S = {r + 1, r + 2, · · · , 2r} × {1, · · · , r}. Thus, kukλ,n ≤ 1 self adjoint element. 2 (µ2kβk2 + µ−2kαk2)]kv1kkv2k · · · kvmk, where for each t, vt is a (µ2kβk2 + µ−2kαk2) = kβkkαk, given δ > 0 we can choose µ0 > 0 Since, min µ>0 1 2 such that kβkkαk+δ > 1 kβkkαkkx1kkx2k · · · kxmk. Thus, we get the desired norm condition. 0 kαk2). Therefore, kukλ,n ≤ k αk2kv1kkv2k · · · kvmk ≤ 0kβk2+µ−2 2 (µ2 (cid:3) We are now in a position to define an appropriate cone structure : Definition 3.3. Let V1, · · · , Vm be matrix ordered operator spaces and λ = (λn)n∈N be a sequence satisfying (O1)-(O2). We define (a) Cn={α⊗λj (v1, v2, · · · , vm)α∗ : vt ∈ Mj(Vt)+, α ∈ Mn,τ (j), j ∈ N, t = 1, 2, · · · , m} ⊂ Mn(⊗λVi)sa and (b) Mn(⊗λVi)+ := C−k·kλ,n n . Proposition 3.4. For matrix ordered operator spaces Vi, 1 ≤ i ≤ m,(cid:0) ⊗m kλ,n}∞n=1, Mn(⊗λVi)+(cid:1) is a matrix ordered operator space. Proof. From Proposition 3.2, the involution is an isometry on (Mn(⊗λVi))sa hence Mn(⊗λVi)+ is a cone provided Cn is. i=1 Vi, {k · Let u1 = α1 ⊗λk (v1, v2, · · · , vm)α∗1 ∈ Cn1, u2 = α2 ⊗λl (w1, w2, · · · , wm)α∗2 ∈ Cn, then using Lemma 3.1(ii), u1 + u2 = ( α1 α2 )(cid:18) ⊗λk (v1, v2, · · · , vm) ⊗λl(w1, w2, · · · , wm) (cid:19)(cid:18) α∗1 α∗2 (cid:19) where xt = diag(vt, wt) ∈ Mk+l(V )+ and α =(cid:0)α1 α2(cid:1), hence the family {Cn} is = αP (⊗λk+l(x1, x2, · · · , xm))P ∗α∗ ∈ Cn, closed under addition. Now, for t ≥ 0 0 0 tu1 = t1/2α1 ⊗λk (v1, v2, · · · , vm)t1/2α∗1 ∈ Cn. Also, for γ ∈ Mm,n and α ⊗λk (v1, v2, · · · , vm)α∗ ∈ Cm, γ∗α ⊗λk (v1, v2, · · · , vm)α∗γ = γ∗α ⊗λk (v1, v2, · · · , vm)(γ∗α)∗ ∈ Cn. (cid:3) Note that if u ∈ Cn, then kzkΛ,n := inf{max{kukλ,n, ku′kλ,n} :(cid:18) u 1(cid:19) u(cid:0)1 1(cid:1) ∈ C2n. (cid:18)u u u u(cid:19) =(cid:18)1 z z∗ u′(cid:19) ∈ C2n}. MATRIX ORDERING AND ALGEBRAIC ASPECTS OF λ-THEORY 11 z z and so using Proposition 3.2 z∗ u′(cid:19) ∈ C2n, implies that 0(cid:19) ∈ Cn Remark 3.5. Note that(cid:18) u 0(cid:1)(cid:18) u z∗ u′(cid:19)(cid:18)1n z∗ u′(cid:19)(cid:18) 0 u =(cid:0)1n kukλ,n = inf(cid:8)kαk2kx1kkx2k · · · kxmk : u = α ⊗λj (x1, x2, · · · , xm)α∗, ku′kλ,n = inf(cid:8)kα′k2kx′1kkx′2k · · · kx′mk : u′ = α′ ⊗λr (x′1, x′2, · · · , x′m)α′∗, u′ =(cid:0)0 1n(cid:1)(cid:18) u α ∈ Mn,τ (j), xt ∈ Mj(Vt)+, j ∈ N(cid:9) α′ ∈ Mn,τ (r), x′t ∈ Mr(Vt)+, j ∈ N(cid:9). Motivated by this and Han's [8, Definition 3.2], we relate a suitable norm to the 1n(cid:19) ∈ Cn, and z cone Cn defined above that behaves well with matrix regular operator spaces. Definition 3.6. Let Vi, 1 ≤ i ≤ m be matrix ordered operator spaces and λ = (λn)n∈N be a sequence satisfying (O1)-(O3). Then for z in Mn(⊗λVi), we define: Therefore k · kΛ,n ≤ k · kλ,n on Cn. The set(cid:26) max{kukλ,n, ku′kλ,n} :(cid:18) u C2n(cid:27) is non empty from Proposition 3.7(i) and (ii) proved below. z z∗ u′(cid:19) ∈ Proposition 3.7. If Vi, 1 ≤ i ≤ m are matrix regular operator spaces and λ = (λn)n∈N a sequence satisfying (O1)-(O3), then (i) Cn is a proper cone in Mn(⊗m (ii) For z ∈ Mn(⊗m i=1Vi), there exist elements u1, u2 in Cn such that(cid:18)u1 z∗ u2(cid:19) ∈ i=1Vi) for all n ∈ N. z C2n. (iii) k · kΛ,n is a norm on Mn(⊗m i=1Vi). Proof. (i) Let z ∈ Cn ∩ −Cn, then z = α ⊗λj (v1, v2, · · · , vm)α∗ with vt ∈ Mj(Vt)+, α ∈ Mn,τ (j). By Lemma 3.1(iii), for continuous c.p. maps φt : Vt → Mkt , t = 1, . . . , m, we have (⊗m t=1φt)n(z) ∈ M + nk1k2...km ∩ −M + nk1k2...km = {0}. Now each Vt being matrix regular its dual V ∗t is also matrix regular ([18, Corollary 6.7]), hence each completely bounded linear map from Vt into a matrix algebra is actually a linear combination of some c.p. maps. Thus, even for c.b. maps f (t) : Vt → Mkt, t = 1, . . . , m; (⊗m t=1f (t))n(z) ∈ M + nk1k2...km ∩ −M + nk1k2...km = {0}, which further implies the operator space injective tensor norm is given by, kzkMn( ⊗Vi) = 0, giving z = 0. 12 P. LUTHRA, A. KUMAR, AND V. RAJPAL (ii) Using [4, Proposition 4.1], for any z ∈ Mn(⊗m z = α ⊗λj (v1, v2, · · · , vm)β∗ for vt ∈ Mj(Vt), α, β ∈ Mn,τ (j), t = 1, 2, · · · , m. i=1Vi) can be written as: As each Vi is a matrix regular operator space, there exist v1 such that t , v2 t ∈ Mj(Vt)+ xt =(cid:18)v1 t v∗t vt v2 t(cid:19) ∈ M2j(Vt)+, t = 1, 2, · · · , m. Using Lemma 3.1(ii), we have β ⊗λj (v∗1 , v∗2, · · · , v∗m)α∗ β ⊗λj (v2 1, v2 2, · · · , v2 m)α∗ α ⊗λj (v1, v2, · · · , vm)β∗ 1, v1 2, · · · , v1 (cid:18)α ⊗λj (v1 =(cid:18)α 0 0 β(cid:19)(cid:18)⊗λj (v1 =(cid:18)α 0 0 β(cid:19) ⊗λ2j(cid:18)(cid:18)v1 1 v∗1 1, v1 2, · · · , v1 m) ⊗λj (v1, v2, · · · , vm) ⊗λj (v∗1 , v∗2, · · · , v∗m) ⊗λj (v2 1, v2 2, · · · , v2 v1 v2 1(cid:19) ,(cid:18)v1 2 v∗2 v2 v2 2(cid:19) , · · · ,(cid:18)v1 m)β∗(cid:19) m)(cid:19)(cid:18)α∗ β∗(cid:19) m(cid:19)(cid:19)(cid:18)α∗ β∗(cid:19) ∈ C2n. m vm v∗m v2 0 0 0 0 (iii) k · kΛ,n clearly satisfies homogeneity. i=1Vi) be any elements, then by (ii) above there exist Let z1, z2 ∈ Mn(⊗m ui, u′i ∈ Cn such that (cid:18)ui z∗i From zi u′i(cid:19) ∈ C2n and kuikλ,n, ku′ikλ,n < kzkΛ,n + ǫ. (cid:18)u1 + u2 z∗1 + z∗2 z1 + z2 u′1 + u′2(cid:19) ∈ C2n it follows that kz1 + z2kΛ,n ≤ max{ku1 + u2kλ,n, ku′1 + u′2kλ,n} ≤ max{ku1kλ,n + ku2kλ,n, ku′1kλ,n + ku′2kλ,n} < kz1kΛ,n + kz2kΛ,n + 2ǫ Let kzkΛ,n = 0. Given ǫ > 0, there exist u, u′ ∈ Cn such that and kukλ,n, ku′kλ,n < ǫ. Again by Lemma 3.1(iii), for c.c.p. maps f (t) : Vt → Mkt , t = 1, . . . , m; z (cid:18) u z∗ u′(cid:19) ∈ C2n (cid:18) (⊗m (⊗m t=1f (t))n(u) t=1f (t))n(z∗) t=1f (t))2n (cid:18) u (⊗m (⊗m t=1f (t))n(z) t=1f (t))n(u′)(cid:19) z∗ u′(cid:19)! ∈ M + z = (⊗m 2nk1k2...km . It follows that k(⊗m t=1f (t))n(z)k ≤ max{k ⊗m t=1 f (t))n(u)k, k ⊗m t=1 f (t))n(u′)k} ≤ max{kukλ,n, ku′kλ,n} < ǫ. Thus (⊗m 0, which implies z = 0. t=1f (t))n(z) = 0 and using matrix regularity as in case (i), kzkMn( ⊗Vi) = (cid:3) MATRIX ORDERING AND ALGEBRAIC ASPECTS OF λ-THEORY 13 The positive cones of matrix ordered operator spaces are closed. Therefore, we consider Mn(⊗ΛVi)+ := C−k·kΛ,n (cid:18) u z∗ u′(cid:19) ∈ C2n if and only if (cid:18)u′ z z n . From the definition of k · kΛ,n, since 1 1 z z∗ u(cid:19) = (cid:18)0 1 0(cid:19)(cid:18) u z∗ u′(cid:19)(cid:18)0 1 have kz∗kΛ,n = kzkΛ,n. Thus, the involution is an isometry on (Mn(⊗m kΛ,n). Hence Mn(⊗ΛVi)+ is a proper cone. 0(cid:19) ∈ C2n, we i=1Vi), k · Recall from [19, Definition 4.7] that an operator space matrix norm k · kα is said to be λ-subcross if k ⊗p (v1, v2, . . . , vm)k(α,τ (p)) ≤ kv1k . . . kvmk for all p ∈ N and vi ∈ Mp(Vi). In case of equality the norm is called λ-cross. Theorem 3.8. For matrix regular operator spaces Vi and λ = (λn)n∈N satisfying (O1)-(O3), (⊗m i=1Vi, {k · kΛ,n}∞n=1, Mn(⊗ΛVi)+) is a matrix regular operator space with λ-subcross matrix norm. Proof. We first prove that Mn(⊗m kΛ,n}∞n=1 of matrix norms. Given z1 ∈ Mn1(⊗m i=1Vi) is an operator space with the family {k · i=1Vi) and ǫ > 0, choose ui, u′i ∈ Cn such i=1Vi), z2 ∈ Mn2(⊗m that(cid:18)ui z∗i zi u′i(cid:19) ∈ C2ni and kuikλ,ni, ku′ikλ,ni < kzikΛ,ni + ǫ, i = 1, 2. By definition, there exist representations ui = αi ⊗λji (vi 1, vi 2, · · · , vi m)α∗i αi ∈ Mni,τ (ji), vi t ∈ Mji(Vt)+. Since k · kλ,n is an operator space norm, using Ruan's first condition (M1) [5] for operator space, we have ku1 ⊕ u2kλ,2n ≤ max{ku1kλ,n, ku2kλ,n} As, we have, 0 u2 0 z∗2 1 0 0 0 0 1 0 0 u1 0 z∗1 0  = (cid:13)(cid:13)(cid:13)(cid:13)(cid:18)z1 0 z1 0 u 1 0 ′ 0 z2 0 u ′ 2 0 1 0 0 0 0 0 1  z2(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)Λ,2n 0 u1 z1 z∗1 u′1 0 0 0 0   ≤ max(cid:26)(cid:13)(cid:13)(cid:13)(cid:13)(cid:18)u1 0 0 0 0 0 z2 u2 z∗2 u′2 0 0 1 0 1 0 0 0   u2(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)λ,2n 0 0 0 1 0 0 0 0 1 ,(cid:13)(cid:13)(cid:13)(cid:13)(cid:18)u′1 0  ∈ C2(n1+n2) u′2(cid:19)(cid:13)(cid:13)(cid:13)(cid:13)λ,2n(cid:27) 0 ≤ max{ku1kλ,n, ku2kλ,n, ku3kλ,n, ku4kλ,n} < max{kz1kΛ,n, kz2kΛ,n} + ǫ. Let z ∈ Mn(⊗m i=1Vi) and α, β ∈ Mm,n, then there exist u, u′ ∈ Cn such that (cid:18) u z∗ u′(cid:19) ∈ C2n and kukλ,n, ku′kλ,n < kzkΛ,n + ǫ. Assuming, kαk = kβk by homo- z geneity, since (cid:18) αuα∗ β∗z∗α∗ β∗u′β(cid:19) =(cid:18)α 0 β∗(cid:19)(cid:18) u z∗ u′(cid:19)(cid:18)α∗ αzβ 0 0 z 0 β(cid:19) ∈ C2n 14 P. LUTHRA, A. KUMAR, AND V. RAJPAL we have, kαzβkΛ,n ≤ max{kαuα∗kλ,n, kβ∗u′βkΛ,n} ≤ max{kαk2kukλ,n, kβk2ku′kλ,n} < kαkkβk(kzkΛ,n + ǫ). Hence,(cid:0) ⊗m i=1 Vi, {k · kΛ,n}∞n=1(cid:1) is an operator space. Let vi ∈ Mj(Vi) with kvik < 1 for i = 1, . . . , m. Then there exist ui, u′i ∈ Mj(Vi)+ k·k≤1 such that i = 1, 2, . . . , m. vi v∗i u′i(cid:19) ∈ M2j(Vi)+ (cid:18)ui (cid:18) ⊗λj (u1, . . . , um) ⊗λj (v1, . . . , vm) ⊗λj (v1, . . . , vm)∗ ⊗λj (u′1, . . . , u′m)(cid:19) = P ⊗λ2j(cid:0)(cid:18)u1 v∗1 u′1(cid:19) , . . . ,(cid:18)um vm v1 v∗m u′m(cid:19)(cid:1)P ∗ ∈ C2τ (j), Now, it follows that k ⊗λj (v1, . . . , vm)kΛ,2j ≤ max{k ⊗λj (u1, . . . , um)kλ,j, k ⊗λj (u′1, . . . , u′m)kλ,j} ≤ max{ku1k . . . kumk, ku′1k . . . ku′mk} < 1 Therefore, the family of matrix norms {k · kΛ,n}∞n=1 is λ-subcross. As kukΛ,n ≤ kukλ,n, if kzkΛ,n < 1 then there exist u, u′ ∈ Cn such that z (cid:18) u z∗ u′(cid:19) ∈ C2n and kukλ,n, ku′kλ,n < 1. Thus, and matrix regularity follows. (cid:3) kukΛ,n < 1 and ku′kΛ,n < 1, 4. λ-operator system tensor product We now prove that the cones Cn associated with λ under the conditions (O1)- (O3) also preserve the operator system structure defined in [13]. The techniques are again same as that for the max operator system tensor product defined in [13]. Theorem 4.1. Let (S, {Mn(S)+}∞n=1, 1S) and (T , {Mn(T )+}∞n=1, 1T ) be operator systems. The family {Cn}∞n=1 associated with a sequence λ = (λn)n∈N satisfying (O1)-(O3), is a matrix ordering on S ⊗ T with order unit 1S ⊗ 1T . Proof. From Proposition 3.7, we know that {Cn}∞n=1 is a family of proper compatible cones on Mn(⊗λSi). We only need to check that 1 ⊗ 1 is a matrix order unit. Let α⊗λj (s, t)α∗ ∈ (S1⊗S2)sa with s ∈(cid:0)Mj(S)(cid:1)sa , t ∈(cid:0)Mj(T )(cid:1)sa and α ∈ M1,τ (j), 1S and 1T being Archimedean order unit for S and T respectively, then we can find K large enough such that K(1S)j + s ∈ Mj(S)+ K(1T )j + t ∈ Mj(T )+ and and K(1S)j − s ∈ Mj(S)+, K(1T )j − t ∈ Mj(T )+. MATRIX ORDERING AND ALGEBRAIC ASPECTS OF λ-THEORY 15 So that, Further, Cn ∋α ⊗λj(cid:0)K(1S)j + s, K(1T )j + t(cid:1)α∗ + α ⊗λj(cid:0)K(1S)j − s, K(1T )j − t(cid:1)α∗ ⊗λj(cid:0)K(1S)j + s, K(1T )j + t(cid:1), ⊗λj(cid:0)K(1S)j − s, K(1T )j − t(cid:1) ∈ Cn. = α(cid:16) ⊗λj(cid:0)K(1S)j, K(1T )j(cid:1) + ⊗λj(cid:0)s, t(cid:1)(cid:17)α∗ = α(cid:16)(cid:16)K 2λj(1, 1) ⊗ 1S ⊗ 1T(cid:1) + ⊗λj(cid:0)s, t(cid:1)(cid:17)α∗ = α(cid:16)(cid:0)K 2Iτ (j) ⊗ 1S ⊗ 1T(cid:17) + ⊗λj(cid:0)s, t(cid:1)(cid:17)α∗ = α(cid:16)K 2(1S ⊗ 1T )τ (j) + ⊗λj(cid:0)s, t(cid:1)(cid:17)α∗ = K 2α(1S ⊗ 1T )τ (j)α∗ + α(cid:16) ⊗λj(cid:0)s, t(cid:1)(cid:17)α∗ = (K 2αα∗)1S ⊗ 1T + α(cid:16) ⊗λj(cid:0)s, t(cid:1)(cid:17)α∗ which proves that 1S ⊗ 1T is an order unit. Similarly, one can prove that 1S ⊗ 1T is in fact a matrix order unit. (cid:3) Definition 4.2. Let λ = (λn)n∈N fulfills conditions (O1)-(O3), and Cλ n := {P ∈ Mn(S ⊗ T ) : r(1S ⊗ 1T )n + P ∈ Cn, ∀r > 0} be the Archimedeanization([16]) of the matrix ordering Cn for all n ≥ 1. We call the n}∞n=1, 1S ⊗ 1T ) the λ- operator system tensor product operator system (S ⊗ T , {Cλ of S and T and denote it by S ⊗λ T . Theorem 4.3. The mapping λ : O × O → O sending (S, T ) to S ⊗λ T is an operator system tensor product in the sense of [13]. Proof. Observe that, (T1) By definition (S ⊗ T , {Cλ (T2) For P ∈ Mk(S)+ and Q ∈ Ml(T )+, since n(S ⊗ T )}∞n=1, 1S ⊗ 1T ) is an operator system. P ⊗ Q = α ⊗λk+l(cid:0)Ik+l ⊗ P, Ik+l ⊗ Q(cid:1)α∗ ∈ Ckl, where α = (Ik+l, 0, · · · , 0) ∈ Mkl,τ (k+l), we have property (T2). (T3) For unital completely positive maps φ ∈ S → Mn and ψ ∈ T → Mm, using Lemma 3.1(iii) we have (φ ⊗ ψ)n(Cn) ⊆ M + n , thus (T3) follows. (T4) Let φ ∈ U CP (S1, S2) and ψ ∈ U CP (T1, T2). Then for any element A ⊗λj (P, Q)A∗ ∈ Cn, where A ∈ Mn,τ (j), P ∈ Mj(S1)+ and Q ∈ Mj(T1)+: (φ ⊗ ψ)n(A ⊗λj (P, Q)A∗) = A(cid:0)(φ ⊗ ψ)τ (j) ⊗λj (P, Q)(cid:1)A∗ = A ⊗λj(cid:0)φj(P ), ψj(Q)(cid:1)A∗ ∈ Mn(S2 ⊗λ T2)+. Thus, (φ⊗ ψ)n(Cn(S1, T1)) ⊆ Cn(S2, T2), and using [13, Lemma 2.5] φ⊗ ψ ∈ U CP (S2, T2). Remark 4.4. If λ is either Kronecker or Schur Product, the cone Cλ with Cmax n ([13, 15]). n = Cs n coincides (cid:3) 16 P. LUTHRA, A. KUMAR, AND V. RAJPAL 5. λ-tensor product of C∗-algebras We now move on to the algebraic structures for the λ-theory. For this we make use of the condition (W2) (Section 2.1). An associative algebra A over C is said to be a completely contractive Banach algebra if it is a complete operator space for which the multiplication map mA : A× A → A (a, b) → ab is jointly completely contractive, i.e. k[aijbkl]k ≤ k[aij]kk[bkl]k for all [aij] ∈ Mn(A) and [bkl] ∈ Mn(A). Theorem 5.1. For completely contractive Banach algebras A1, A2, · · · , Am, ⊗λAi is a Banach algebra if λ satisfies (W2). Further if each Ai is a Banach ∗-algebra and λ also satisfies (O1), then ⊗λAi is a Banach ∗-algebra. Moreover if each Ai is approximately unital then ⊗λAi is also approximately unital. Proof. Let x, y ∈ ⊗λAi with x = α ⊗λr (u(1), u(2), . . . , u(m))β and y = γ ⊗λs (v(1), v(2), . . . , v(m))δ, where for each t = 1, 2, · · · , m α ∈ M1,τ (r), β ∈ Mτ (r),1, γ ∈ M1,τ (s), δ ∈ Mτ (s),1 ⊗ v(t) kt,lt ⊗ u(t) it,jt ε[s] kt,lt ε[r] it,jt . , v(t) := X(kt,lt) u(t) := X(it,jt) Then, using Property (W2), there exists S ∈ Mτ (r)τ (s),τ (rs), T ∈ Mτ (rs),τ (r)τ (s) with kSk, kT k ≤ 1 such that αλr(ε[r] i1,j1, . . . , εim,jm )β ⊗ u(1) i1,j1 ⊗ . . . ⊗ u(m) γλs(ε[s] k1,l1 , . . . , εkm,lm)δ ⊗ v(1) k1,l1 ⊗ . . . ⊗ v(m) im,jm(cid:1) km,lm(cid:1) . . .Xi1,j1 xy =(cid:0) Xim,jm . . .Xk1,l1 (cid:0) Xkm,lm . . .Xi1,j1 Xkm,lm = Xim,jm = (α ⊗ γ)S(Xim,jm = (α ⊗ γ)S(Xim,jm (W2) (β ⊗ δ) ⊗ u(1) i1,j1 . . .Xk1,l1(cid:0)(α ⊗ γ)(λr(ε[r] . . .Xk1,l1 . . .Xi1,j1 Xkm,lm . . .Xk1,l1 . . .Xi1,j1 Xkm,lm ⊗ u(1) i1,j1 v(1) k1,l1 (λrs(ε[rs] v(1) k1,l1 (λrs(ε[r] i1,j1 , . . . , εim,jm) ⊗ λs(ε[s] k1,l1 , . . . , εkm,lm)) ⊗ . . . ⊗ u(m) i1,j1 ⊗ ε[s] k1,l1 v(m) km,lm(cid:1) im,jm , . . . , εim,jm ⊗ ε[s] ) k1,l1 ⊗ . . . ⊗ u(m) im,jm v(m) km,lm )T (β ⊗ δ) (i1−1)s+k1,(j1−1)s)+l1 , . . . , ε[rs] (im−1)s+km,(jm−1)s)+lm = (α ⊗ γ)S(⊗λrs(z(1), . . . , z(m)))T (β ⊗ δ); ) ⊗ u(1) i1,j1v(1) k1,l1 ⊗ . . . ⊗ u(m) im,jm v(m) km,lm )T (β ⊗ δ) where z(t) = Pit,jtPkt,lt 1, . . . , m. Thus, ε[rs] (it−1)s+kt,(jt−1)s)+lt ⊗ u(t) i1,j1 v(t) kt,lt (2) = u(t) ⊗ v(t); t = kxykλ,1 ≤ kα ⊗ γkkSkkz(1)kkz(2)k · · · kz(m)kkT kkβ ⊗ δk ≤ kαkkγkkv(1)kkw(1)k . . . kv(m)kkw(m)kkβkkδk MATRIX ORDERING AND ALGEBRAIC ASPECTS OF λ-THEORY 17 making ⊗m λ Ai, and hence ⊗λAi a Banach algebra. If λ satisfies (O1), ∗-part follows from Lemma 3.1(i) and definition of k · kλ,1. One can easily verify that k · kλ,1 ≤ k · kγ, implying that k · kλ,1 is an admissi- ble cross norm on ⊗mAi . Therefore, ⊗λAi has a bounded approximate identity whenever each Ai is approximately unital. (cid:3) In particular, we have the following well known result (see[14, 17]): Corollary 5.2. ⊗⊗Ai, the projective tensor product and ⊗⊙Ai, the Schur tensor product are Banach ∗-algebras with a bounded approximate identity, however, ⊗•Ai, the Haagerup tensor product is a Banach algebra. In general, λ-tensor product of operator spaces is not injective. Since, (⊗λAi)∗ = CBλ(A1 × A2 · · · Am, C)[19, Proposition 4.11] completely isometrically, so the proof of [14, Theorem 5] can be adopted in this case to show the injectivity of λ-tensor product for the closed ideals, i.e. Proposition 5.3. Let Ii be closed two-sided ideals in C∗-algebras Ai for i = 1, 2, · · · , m, then ⊗λIi is a closed two-sided ∗-ideal of ⊗λAi. Lemma 5.4. Let Wi, 1 ≤ i ≤ m be completely complemented subspaces of the operator spaces Vi, 1 ≤ i ≤ m complemented by cb projection having cb norm equal to 1, respectively, then ⊗λWi is a closed subspace of ⊗λVi. Proof. Using the assumption, there are cb projections P1, P2, · · · , Pm from V1 onto W1, V2 onto W2 · · · , Vm onto Wm with kP1kcb = kP2kcb = · · · = kPmkcb = 1. There- fore, by the functoriality of the λ- tensor product([19, Proposition 6.1], ⊗m i=1Pi : ⊗λVi → ⊗λWi is a completely bounded map and k ⊗m i=1 Pikcb ≤ 1. Since, for u ∈ ⊗m i=1Pi(u) = u, so kuk⊗λWi ≤ kuk⊗λVi , hence ⊗λWi is a closed sub- space of ⊗λVi. i=1Vi, ⊗m (cid:3) Since, there is a conditional expectation from a C∗-algebra A onto a finite di- mensional C∗-subalgebra of A, so by the above Lemma for finite dimensional C∗- algebras, λ-tensor product of operator spaces is injective. In general k · kλ need not be injective. Proposition 5.5. Let A0 and B0 be closed ∗-subalgebras of A and B, respectively, ≤ 2kukA b⊗B for u ∈ A0 ⊗ B0. with kf k = 1. Let φ0 be the jcb bilinear form on A0 × B0 corresponding to f . By ([7, Corollary 3.10]), φ0 : A0 × B0 → C extends to a jcb bilinear form φ : A × B → C such that However, for b⊗, we have something partial: then A0b⊗B0 is (isomorphic to) closed ∗-subalgebra of Ab⊗B. Proof. Let I denote the closure of A0⊗B0 in Ab⊗B, so that I is a closed ∗-subalgebra of Ab⊗B. We first claim that kukA b⊗B ≤ kukA0 Choose f ∈ (A0b⊗B0)∗ such that f (u) = kukA0 kφkjcb ≤ 2kφ0kjcb. Therefore k f k ≤ 2, where f is the linear functional on Ab⊗B claim, so it can be extended toei : A0b⊗B0 → Ab⊗B. We now show that A0b⊗B0 is isomorphic to I. For the injectivity ofei, by [10, Theorem 2], it is enough to Again, by the last inequality,ei−1 is continuous. For onto-ness, let u ∈ I. There is corresponding to φ, and thus the claim. Now consider the identity map i : (A0 ⊗ ) → (A ⊗ B, k · kA b⊗B) which is linear and continuous by the last B0, k · kA0 show that it is injective on A0 ⊗ B0 but this follows directly by the last inequality. b⊗B0 b⊗B0 b⊗B0 18 P. LUTHRA, A. KUMAR, AND V. RAJPAL a sequence un ∈ A0 ⊗ B0 converging to u in k · kA b⊗B-norm. The sequence {un} becomes Cauchy with respect to k · kA0 -norm by the last claim, so it converges, say, to u′. Clearly,ei(u′) = u. Thus A0b⊗B0 can be regarded as a closed ∗-subalgebra of Ab⊗B. Proposition 5.6. For C∗-algebras A and B, any λ-cb bilinear form φ on A × B can be extended uniquely to φ on A∗∗ × B∗∗ such that kφkλ = k φkλ. b⊗B0 (cid:3) Proof. Since φ : A×B → C is λ-cb bilinear form. It is in particular bounded bilinear form and thus determines a unique separately normal bilinear form φ : A∗∗ ×B∗∗ → C by [6, Corollary 2.4]. For k ∈ N, consider the map φk : Mk(A∗∗) × Mk(B∗∗) → Mτ (k) taking φk(a1⊗m, a2⊗m′) = λk(a1, a2)⊗ φ(m, m′). Let a∗∗ = [a∗∗ij ] ∈ Mk(A∗∗) and b∗∗ = [b∗∗ij ] ∈ Mk(B∗∗) with ka∗∗k ≤ 1 and kb∗∗k ≤ 1. Since the unit ball of Mk(A) is w∗-dense in the unit ball of Mk(A∗∗), so we obtain a net (aλ) = (aλ ij ) (resp., (bν)) in Mk(A) (resp., Mk(B)) which is w∗-convergent to a∗∗ (resp., b∗∗) ij is w∗-convergent to a∗∗ij for each ǫpl ⊗ with kaλk ≤ 1 (resp., kbνk ≤ 1). Therefore, caλ i, j. Now by the separate normality of φ, we have kλk ⊗ φ(Xi,j b∗∗pl )k = k Xi,j,p,l k Xi,j,p,l ǫij ⊗ a∗∗ij ,Xp,l λk(ǫij , ǫpl) ⊗ φ(a∗∗ij , b∗∗pl )k = k lim λk(ǫij, ǫpl) ⊗ φ(aλ lim λ kφkk ≤ kφkλ for every k ∈ N. Clearly, kφkλ ≤ k φkλ as φ being the restriction of φ. Hence kφkλ = k φkλ. (cid:3) kl)k ≤ kλk ⊗ φk for each k ∈ N. Thus k φkk ≤ ν Xi,j,p,l λk(ǫij, ǫpl) ⊗ φ(aλ pl)k = ij , bν ij, bν lim λ lim ν For a tensor norm α and a closed ideal J of A ⊗α B, we try to find out whether a ⊗ b ∈ Jmin implies that a ⊗ b ∈ J. This question stems from the study of the elusive nature of the Haagerup tensor product of C∗-algebras, it was resolved for the Haagerup tensor product in ([1, Theorem 4.4]) and for the operator space projective tensor product in [14, Theorem 6]. We present here a unified approach. For C∗-algebras A and B, assume that k · kλ ≥ k · kmin on A ⊗ B, so there will be a identity map i from A ⊗λ B into A ⊗min B. Lemma 5.7. Let M and N be von Neumann algebras and let L be a closed ideal in M ⊗λ N , where the sequence λ = (λn)n∈N satisfies condition (W2). If 1 ⊗ 1 ∈ Lmin, then 1 ⊗ 1 ∈ L and L equals M ⊗λ N . Proof. Since, 1 ⊗ 1 ∈ Lmin, for a given ǫ = representation in M ⊗λ N where for t = 1, 2, . . . , rt ∈ N, 1 2 . Let w = ki(w) − 1 ⊗ 1kmin < ∞Xt=1 αt ∈ M1,τ (rt), βt ∈ Mτ (rt),1, u(t) := X(it,jt) αt ⊗λrt , there exists w ∈ L such that 1 2 (u(t), v(t))βt; be a norm convergent ε[rt] it,jt ⊗ u(t) it,jt ε[rt] kt,lt ⊗ v(t) kt,lt . , v(t) := X(kt,lt) By [11, Theorem 8.3.5], there exist sequences {x(t) it,jt {φn} ∈ P (M ) and {ψn} ∈ P (N ) such that } ∈ Z(M ), {y(t) it,jt } ∈ Z(N ), kφn(u(t) it,jt ) − x(t) it,jt lim n→∞ k = lim n→∞ kψn(v(t) kt,lt ) − y(t) it,jt k = 0 (t = 1, 2, . . .) (3) MATRIX ORDERING AND ALGEBRAIC ASPECTS OF λ-THEORY 19 where P (M ) denotes the set of all mappings φ : M → M such that, for m ∈ M , φ(m) is in the convex hull of the set {umu∗ : u ∈ U (M )}. For each n ∈ N, using the contractive maps φn ⊗ ψn on M ⊗λ N ([19, Proposition 6.1]) and invariance of ideal L under φn ⊗ ψn, we have for all positive integers k ≤ l , ε[rt] kt,lt ) ⊗ x(t) it,jt ⊗ y(t) λrt (ε[rt] it,jt , ε[rt] kt,lt ) ⊗ x(t) it,jt kt,lt(cid:1)βt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)λ,1 kt,lt(cid:1)βt ⊗ y(t) λrt (ε[rt] it,jt , ε[rt] kt,lt ) ⊗ φn(u(t) it,jt ) ⊗ ψn(v(t) kt,lt λrt (ε[rt] it,jt , ε[rt] kt,lt ) ⊗ φn(u(t) it,jt ) ⊗ ψn(v(t)) λrt (ε[rt] it,jt (cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) lXt=k αt(cid:0) X(kt,lt) X(it,jt) ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) lXt=k αt(cid:0) X(kt,lt) X(it,jt) lXt=k αt(cid:0) X(kt,lt) X(it,jt) +(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) lXt=k αt(cid:0) X(kt,lt) X(it,jt) lXt=k αt(cid:0) X(kt,lt) X(it,jt) − ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) z = ∞Xt=1 αt(cid:0) X(kt,lt) X(it,jt) λrt (ε[rt] it,jt , ε[rt] kt,lt ) ⊗ u(t) it,jt ⊗ v(t) −→ 0 as n → ∞, being the partial sum of w. Therefore, one can define an element λrt (ε[rt] it,jt , ε[rt] kt,lt ) ⊗ x(t) it,jt ⊗ y(t) )(cid:1)βt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)λ,1 kt,lt(cid:1)βt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)λ,1 (Using 3) kt,lt(cid:1)βt(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)λ,1 kt,lt(cid:1)βt ∈ Z(M ) ⊗λ Z(N ). For sufficiently large choice of n and ǫ > 0, we deduce easily that Since, L is left invariant by φn ⊗ ψn for each n, so kφn ⊗ ψn(w) − zkλ < ǫ (4) z = lim n→∞ (φn ⊗ ψn(w) ∈ L ∩ (Z(M ) ⊗λ Z(N )) It is easy to show that i ◦ (φn ⊗λ ψn) = (φn ⊗min ψn) ◦ i. k(φn ⊗min ψn)(i(w)) − 1 ⊗ 1kmin = k(φn ⊗min ψn)(i(w)) − (φn ⊗min ψn)(i(1 ⊗ 1))kmin ≤ ki(w) − 1 ⊗ 1kmin < 1 2 . By (4), we have ki ◦ (φn ⊗λ ψn)(w) − i(z)kmin < ǫ, for sufficiently large n. Then the inequality ki(z) − 1 ⊗ 1kmin ≤ 1 2 < 1. (5) (6) (7) is a consequence of (5), (6) and triangle inequality, and so i(z) is invertible in Lmin ∩ (Z(M ) ⊗min Z(N )). Using similar arguments as in ([12, Theorem 2.11.6, Lemma 2.11.1] and the fact that Z(M ) is a nuclear C∗-algebra [2, Proposition 20 P. LUTHRA, A. KUMAR, AND V. RAJPAL 1] we get Z(M ) ⊗λ Z(N ) is semisimple. The regularity of Z(M ) ⊗λ Z(N ) follows from [12, Lemma 4.2.19]. Since i(z) is invertible in Lmin ∩ (Z(M ) ⊗min Z(N )), i.e. invertible in both Lmin and Z(M ) ⊗min Z(N ), so 0 /∈ σZ(M)⊗minZ(N )(i(z)). So by [12, Exercise 4.8.12] 0 /∈ σZ(M)⊗λZ(N )(z). Hence, z is invertible in Z(M ) ⊗λ Z(N ), so there exists w ∈ Z(M ) ⊗λ Z(N ) such that zw = wz = 1 ⊗ 1. Since, z ∈ L and L being an ideal, so 1 ⊗ 1 ∈ L. (cid:3) Theorem 5.8. Let λ = (λn)n∈N fulfills (W2), A and B be C∗-algebras and let J be a closed ideal in A ⊗λ B. If a ⊗ b ∈ Jmin then a ⊗ b ∈ J. In particular A ⊗λ B is a -semi-simple Banach algebra provided λ = (λn)n∈N further fulfills (O1). Proof. Suppose that a, b ≥ 0 and a ⊗ b ∈ Jmin but not in J. So by Hahn Banach theorem there exists φ ∈ (A ⊗λ B)∗ such that φ(J) = 0 and φ(a ⊗ b) 6= 0. Since, (A ⊗λ B)∗ = CBλ(A × B, C) so φ(x ⊗ y) = φ(x, y) for some φ ∈ CBλ(A × B, C) and for all x ∈ A, y ∈ B. By Proposition 5.6, we have φ∗∗ : A∗∗ × B∗∗ → C a λ- completely bounded operator satisfying k φ∗∗kλ = k φkλ . Let L be the closed ideal in A∗∗ ⊗λ B∗∗ generated by J. Let u = αj ⊗λkj (xj 1, xj 2)βj be a norm convergent sum in A⊗λB representing a fixed but an arbitrary element of J. Since φ annihilates J, so αjφλkj (⊗λkj (xj 1, xj 2))βj = αjλkj (ai1 , ai2 ) ⊗ φ(xj i1 ⊗ xj i2)βj = 0. Let ∞Xj=1 ∞Xj=1 ∞Xj=1Xi1,i2 ∞Xj=1Xi1,i2 u, v ∈ A and s, t ∈ B then we have αjλkj (ai1 , ai2 ) ⊗ φ(uxj i1 v ⊗ sxj i2 t)βj = 0. Let M = A∗∗∗ and N = B∗∗ be the von Neumann algebras generated by A and B. For each n ∈ N and u ∈ M , define wn(u) = αjλkj (ai1 , ai2 ) ⊗ φ(uxj i1 v ⊗ sxj i2 t)βj. We will claim that {wn} is a Cauchy sequence. To see this, let m < n, wn(u) − wm(u) = αjλkj (ai1 , ai2 ) ⊗ φ(uxj i1 v ⊗ sxj i2 t)βj ≤ nXj=m+1Xi1,i2 nXj=1Xi1,i2 kukkskktkkvkk nXj=m+1 αj ⊗λkj limit w ∈ M∗ given by w(u) = in [1], we obtain φ annihilates L. ∞Xj=1Xi1,i2 (xj 1, xj 2)βjkλ, and so {wn} is a Cauchy sequence with αjλkj (ai1 , ai2 ) ⊗ φ(uxj i1 v ⊗ sxj i2 t). Again as Now for ǫ > 0, let pǫ ∈ M and qǫ ∈ N be the spectral projections of a and b respectively for the closed interval [ǫ, ∞). Since there is a conditional expectation from M onto pǫM pǫ, so pǫM pǫ ⊗λ qǫN qǫ is a closed subalgebra of M ⊗λ N by Lemma 5.4. Let L0 = L ∩ (pǫM pǫ ⊗λ qǫN qǫ), a closed ideal in pǫM pǫ ⊗λ qǫN qǫ, and so (L0)min is a closed ideal in pǫM pǫ ⊗min qǫN qǫ. Now as in [1, Theorem 4.4] and [14, Theorem 6], we get (L0)min contains pǫ ⊗ qǫ and so by the above Lemma 5.7, pǫ ⊗ qǫ ∈ L0. Hence L0 = pǫM pǫ ⊗λ qǫN qǫ, which further implies that pǫa⊗ qǫb ∈ L, and so φ(pǫa ⊗ qǫb) = 0. Letting ǫ → 0, we have φ(a ⊗ b) = 0, contrary to the choice of φ. In the case when both a and b are arbitrary elements, then one may apply the (cid:3) similar technique as given in [1] to obtain the result. MATRIX ORDERING AND ALGEBRAIC ASPECTS OF λ-THEORY 21 Acknowledgements. The authors would like to thank Andreas Defant for providing a copy of [19]. The authors would also like to thank the referee for useful comments and suggestions. References [1] S. D. Allen, A. M. Sinclair, and R. R. Smith. The ideal structure of the Haagerup tensor product of C ∗-algebras. J. Reine Angew. Math, 442:111 -- 148, 1993. [2] D. P. Blecher. Geometry of the tensor product of C-algebras. In Mathematical Proceedings of the Cambridge Philosophical Society, volume 104, pages 119 -- 127. Cambridge Univ Press, 1988. [3] D. P. Blecher and V. I. Paulsen. Tensor products of operator spaces. J. Funct. Anal., 99(2):262 -- 292, 1991. [4] A. Defant and D. Wiesner. Polynomials in operator space theory. J. Funct. Anal., 266(9):5493 -- 5525, 2014. [5] E. G. Effros and Z. J. Ruan. Operator spaces. Clarendon Press, 2000. [6] U. Haagerup. The Grothendieck inequality for bilinear forms on C-algebras. Advances in Mathematics, 56(2):93 -- 116, 1985. [7] U. Haagerup and M. Musat. The Effros -- Ruan conjecture for bilinear forms on C-algebras. Inventiones mathematicae, 174(1):139 -- 163, 2008. [8] K. H. Han. The predual of the space of decomposable maps from a C ∗-algebra into a von Neumann algebra. J. Math. Anal. Appl., 402(2):463 -- 476, 2013. [9] T. Itoh. Completely positive decompositions from duals of C ∗-algebras to von Neumann algebras. Math. Japon., 51:89 -- 98, 2000. [10] R. Jain and A. Kumar. Operator space tensor products of C*-algebras. Mathematische Zeitschrift, 260(4):805 -- 811, 2008. [11] R. V. Kadison and J. R. Ringrose. Fundamentals of the Theory of Operator Algebras II, volume 2. American Mathematical Soc., 1996. [12] E. Kaniuth. A course in commutative Banach algebras, volume 246. Springer Science & Business Media, 2008. [13] A. S. Kavruk, V. I. Paulsen, I. G. Todorov, and M. Tomforde. Tensor products of operator systems. J. Funct. Anal., 261(2):267 -- 299, 2011. [14] A. Kumar. Operator space projective tensor product of C ∗-algebras. Math. Z., 237(2):211 -- 217, 2001. [15] W. H. Ng. Two characterizations of the maximal tensor product of operator systems. arXiv preprint arXiv:1503.07097, 2015. [16] V. I. Paulsen and M. Tomforde. Vector spaces with an order unit. Indiana Univ. Math. J., (58):1319 -- 1359, 2009. [17] V. Rajpal, A. Kumar, and T. Itoh. Schur tensor product of operator spaces. In Forum Math., volume 27, pages 3635 -- 3655, 2015. [18] W. J. Schreiner. Matrix regular operator spaces. J. Funct. Anal., 152(1):136 -- 175, 1998. [19] D. Wiesner. Polynomials in operator space theory. PhD thesis, Der Andere Verlag, Tonning, Lubeck, Marburg, 2009. Department of Mathematics, University of Delhi, Delhi-110007, INDIA E-mail address: [email protected] Department of Mathematics, University of Delhi, Delhi-110007, INDIA E-mail address: [email protected] Department of Mathematics, Shivaji College, University of Delhi, Delhi, India. E-mail address: [email protected]
1706.01877	1	1706	2017-06-06T08:47:20	On the continuity and differentiability of the (dual) core inverse in C*-algebras	[ "math.OA", "math.FA" ]	The continuity of the core inverse and the dual core inverse is studied in the setting of C-algebras. Later, this study is specialized to the case of bounded Hilbert space operators and to complex matrices. In addition, the differentiability of these generalized inverses is studied in the context of C-algebras.	math.OA	math	On the continuity and differentiability of the (dual) core inverse in C ∗-algebras Julio Ben´ıtez, Enrico Boasso, Sanzhang Xu Abstract The continuity of the core inverse and the dual core inverse is studied in the setting of C ∗- algebras. Later, this study is specialized to the case of bounded Hilbert space operators and to complex matrices. In addition, the differentiability of these generalized inverses is studied in the context of C ∗-algebras. Keywords: Core inverse, Dual core inverse, C ∗-algebra, Hilbert space, Matrices. AMS Classification: Primary 46L05, 47A05, Secondary 46K05, 15A09. 1 Introduction The core inverse and the dual core inverse of a matrix were introduced in [1]. These generalized inverses have been studied by several authors, in particular they have been extended to rings with involution ([20]) and to Hilbert space operators ([21]). It is worth noticing that the inverses under consideration are closely related to the group inverse and the Moore-Penrose inverse; to learn more results concerning these notions, see for example [1, 20, 21, 26]. So far the properties of the (dual) core inverse that have been researched are mainly of algebraic nature and the setting has been essentially the one of rings with involution. The objective of the present article is to study the continuity and the differentiability of these inverses in the context of C ∗-algebras. In fact, in section 3, after having recalled several preliminary results in section 2, the continuity of the core inverse and of the dual core inverse will be studied. Two main char- acterizations will be presented. The first one relates the continuity of the aforementioned notions to the continuity of the group inverse and of the Moore-Penrose inverse. The second characterization uses the notion of the gap between subspaces; a similar approach has been used to study the continuity of the Drazin inverse and of the Moore-Penrose inverse, see for example [15, 22, 23] and [7, Chapter 4]. In section 4 results regarding the continuity of the (dual) core inverse of Hilbert space operators and matrices will be presented. Finally, in sec- tion 5 the differentiability of the generalized inverses under consideration will be researched. Furthermore, some results concerning the continuity and the differentiability of the group inverse and the Moore-Penrose inverse will be also proved. It is noteworthy to mention that the core inverse and the dual core inverse are two partic- ular cases of the (b, c)-inverse ([9]), see [20, theorem 4.4]. Therefore the representations and other results presented in [6, Section 7] can be applied to these generalized inverses. 1 2 Preliminary Definitions Since properties of C ∗-algebra elements will be studied in what follows, although the main notions considered in this article can be given in the context of rings with involution, all the definition will be presented in the frame of C ∗-algebras. From now on A will denote a unital C ∗-algebra with unity 1. In addition, A−1 will stand for the set of all invertible elements in A. Given a ∈ A, the image ideals and the null ideals defined by a ∈ A are the following sets: aA = {ax : x ∈ A}, a◦ = {x ∈ A : ax = 0}, Aa = {xa : x ∈ A}, ◦a = {x ∈ A : xa = 0}. Recall that a ∈ A is said to be regular, if there exists b ∈ A such that a = aba. In addition, b ∈ A is said to be an outer inverse of a ∈ A, if b = bab. The notion of invertible element has been generalized or extended in several ways. One of the most important notion of generalized inverse is the Moore-Penrose inverse. An element a ∈ A is said to be Moore-Penrose invertible, if there is x ∈ A such that the following equations hold: axa = a, xax = x, (ax)∗ = ax, (xa)∗ = xa. It is well known that if such an x exists, then it is unique, and in this case x, the Moore- Penrose inverse of a, will be denoted by a†. Moreover, the subset of A composed of all the Moore-Penrose invertible elements of A will be denoted by A†. It is worth noticing that according to [11, Theorem 6], a necessary and sufficient condition for a ∈ A† is that a ∈ A is regular, which in turn is equivalent to aA is closed ([11, Theorem 8]). Moreover, if a ∈ A†, then it is not difficult to prove that a†A = a∗A and Aa† = Aa∗. To learn more properties of the Moore-Penrose inverse in the frame of C ∗-algebras, see [7, 11, 12, 17, 18, 24]. Another generalized inverse which will be central for the purpose of this article is the group inverse. An element a ∈ A is said to be group invertible, if there is x ∈ A such that axa = a, xax = x, ax = xa. It can be easily proved that if such x exists, then it is unique. The group inverse is customarily denoted by a#. The subset of A composed by all the group invertible elements in A will be denoted by A#. Next follows one of the main notions of this article (see [20, Definition 2.3], see also [1] for the original definition in the context of matrices). Definition 2.1. Given a unital C ∗-algebra A, an element a ∈ A will be said to be core invertible, if there exists x ∈ A such that the following equalities hold: axa = a, xA = aA, Ax = Aa∗. 2 According to [20, Theorem 2.14], if such an element x exists, then it is unique. This element will be said to be the core inverse of a ∈ A and it will be denoted by a #(cid:13). In addition, the set of all core invertible elements of A will be denoted by A #(cid:13). Recall that according to [20, Theorem 2.14], when a #(cid:13) exists (a ∈ A), it is an outer inverse of a, i.e., a #(cid:13)aa #(cid:13) = a #(cid:13). Moreover, in [20, Theorem 2.14], the authors characterized the core invertibility in terms of equalities. This characterization was improved in [26, Theorem 3.1]. Specifically, a ∈ A is core invertible if and only if there exists x ∈ A such that ax2 = x, xa2 = a, (ax)∗ = ax. Furthermore, if such x exists, then x = a #(cid:13). Another generalized inverse, which is related with the core inverse, was defined in [20]. Definition 2.2. Given A a unital C ∗-algebra, an element a ∈ A is said to be dual core invertible, if there is x ∈ A such that axa = a, xA = a∗A, and Ax = Aa. As for the core inverse, it can be proved that this x is unique, when it exists; thus it will be denoted by a #(cid:13) and A #(cid:13) will stand for the set of all dual core invertible elements of A. Note that a ∈ A is core invertible if and only if a∗ is dual core invertible and in this case, (a∗) #(cid:13) = (a #(cid:13))∗. In addition, according to [20, Theorem 2.15], when a ∈ A #(cid:13), a #(cid:13) is an outer inverse of a, i.e., a #(cid:13) = a #(cid:13)aa #(cid:13). Observe that according to Definition 2.1 (respectively Definition 2.2), if a ∈ A is core invertible (respectively dual core invertible), then it is regular, and hence a is Moore-Penrose invertible ([11, Theorem 6]). Moreover, if a ∈ A #(cid:13) ∪A #(cid:13), then a is group invertible ([20, Remark 2.16]). Furthermore, according to [20, Theorem 2.19], if a ∈ A #(cid:13), then the following equalities hold: a# = (a #(cid:13))2a, a# = a #(cid:13)aa #(cid:13), a† = a #(cid:13)aa #(cid:13), a #(cid:13) = a#aa†, a #(cid:13) = a†aa#. To learn more on the properties of the core and dual core inverse, see [1, 20, 26]. On the other hand, X will stand for a Banach space and L(X) for the algebra of all operators defined on and with values in X. When A ∈ L(X), the range and the null space of A will be denoted by R(A) and N(A), respectively. When dim X < ∞ and A ∈ L(X), the dimension of R(A) will be denoted with rk(A). Evidently, if A ∈ Cn, the set of complex n × n n), the rank of the complex matrix A coincides with matrices, by considering that A ∈ L(C the previously defined rk(A); consequently, the same notation will be used for both notions. One of most studied generalized inverses is the outer inverse with prescribed range and null space. This generalized inverse will be introduced in the Banach frame. Let X be a Banach space and consider A ∈ L(X) and T, S two closed subspaces in X. If there exists an operator B ∈ L(X) such that BAB = B, N(B) = S, and R(B) = T, then such B is unique ([7, Theorem 1.1.10]). In this case, B will be said to be the A(2) T,S outer inverse of A. To prove several results of this article, the definition of the gap between two subspaces need to be recalled. Let X be a Banach space and consider M and N two closed subspaces in 3 X. If M = 0, then set δ(M, N) = 0, otherwise set δ(M, N) = sup{dist(x, N) : x ∈ M, kxk = 1}, where dist(x, N) = inf{kx − yk : y ∈ N}. The gap between the closed subspaces M and N is bδ (M, N) = max{δ(M, N), δ(N, M)}. See [7, 10, 13] for a deeper insight of this concept. Another notion needed to study the continuity of the (dual) core inverse is the following. Let p and q be self-adjoint idempotents in a C ∗-algebra A. The maximal angle between p and q is the number ψ(p, q) ∈ [0, π/2] such that kp − qk = sin ψ(p, q); see [4, Definition 2.3]. In what follows, given x ∈ A†, ψx will stand for the maximal angle between xx† and x†x, i.e., ψx = ψ(xx†, x†x). 3 Continuity of the (dual) core inverse In first place a preliminary result need to be presented. Theorem 3.1. Let A be a unital C ∗-algebra and consider a ∈ A. The following statements are equivalent (i) a is core invertible. (ii) a is dual core invertible. (iii) a∗ is core invertible. (iv) a∗ is dual core invertible. (v) a is group invertible and Moore-Penrose invertible. In particular, A #(cid:13) = A #(cid:13) = A# ∩ A† = A#. Proof. The equivalence between statements (i) and (iv) and between statements (ii) and (iii) can be derived from Definition 2.1 and Definition 2.2. Note that to conclude the proof, it is enough to prove the last statement of the Theorem. In fact, this statement implies that statement (i) and (ii) are equivalent. According to [20, Remark 2.16], A #(cid:13) ∪ A #(cid:13) ⊆ A#. Moreover, according to [11, Theorem 6], A #(cid:13) ∪ A #(cid:13) ⊆ A†. Therefore, according to [20, Remark 2.16], A #(cid:13) ⊆ A# ∩ A† = A #(cid:13) ∩ A #(cid:13) ⊆ A #(cid:13), A #(cid:13) ⊆ A# ∩ A† = A #(cid:13) ∩ A #(cid:13) ⊆ A #(cid:13) Finally, according to [11, Theorem 6], A# ∩ A† = A#. 4 Note that under the same conditions in Theorem 3.1, as for the group inverse and the Moore-Penrose inverse, (A #(cid:13))∗ = A #(cid:13) and (A #(cid:13))∗ = A #(cid:13), where if X ⊆ A is a set, X ∗ stands for the following set: X ∗ = {x∗ : x ∈ X}. However, in contrast to the case of the group inverse and the Moore-Penrose inverse (when a# (respectively a†) exits, (a∗)# = (a#)∗ (respectively (a∗)† = (a†)∗), a ∈ A), recall that to obtain the core inverse (respectively the dual core inverse) of a∗ it is necessary to consider the dual core (respectively the core) inverse of a: (a∗)#(cid:13) = (a #(cid:13))∗, (a∗) #(cid:13) = (a #(cid:13))∗. To prove the first characterization of this section some preparation is needed. Lemma 3.2. Let A be a unital C ∗-algebra and consider a ∈ A. (i) If a ∈ A #(cid:13), then aa†a #(cid:13) = a #(cid:13). (ii) If a ∈ A #(cid:13), then (aa† + a†a − 1)a #(cid:13) = a†. (iii) Suppose that a ∈ A is regular. The element aa† + a†a − 1 is invertible if and only if a is core invertible. Moreover, in this case, (aa† + a†a − 1)−1 = a #(cid:13)a + (a #(cid:13)a)∗ − 1. (iv) If a ∈ A #(cid:13), then a# = a(a #(cid:13))2. Proof. Recall that according to [11, Theorem 6], if a ∈ A #(cid:13), then a† exists. The proof of statement (i) can be derived from the fact that a #(cid:13) ∈ aA. To prove statement (ii), recall that according to [20, Theorem 2.19 (v)], a #(cid:13) = a#aa†. Therefore, (aa† + a†a − 1)a #(cid:13) = aa†a #(cid:13) + a†aa #(cid:13) − a #(cid:13) = a†aa #(cid:13) = a†aa#aa† = a†. Now statement (iii) will be proved. Note that according to [11, Theorem 6], a† exists. Recall that according to [3, Theorem 2.3], aa† + a†a − 1 ∈ A−1 is equivalent to a ∈ A#. Thus, according to Theorem 3.1, necessary and sufficient for aa† + a†a − 1 ∈ A−1 is that a ∈ A #(cid:13). Next the formula of the inverse of aa† + a†a − 1 will be proved. Recall that according to [26, Theorem 3.1], a #(cid:13)aaa† = aa†. [a #(cid:13)a + (a #(cid:13)a)∗ − 1][aa† + a†a − 1] =(cid:2)a #(cid:13)a + (a #(cid:13)a)∗ − 1(cid:3) aa† +(cid:2)a #(cid:13)a + (a #(cid:13)a)∗ − 1(cid:3) a†a −(cid:2)a #(cid:13)a + (a #(cid:13)a)∗ − 1(cid:3) = aa† + (a #(cid:13)a)∗(aa†)∗ − aa† + a #(cid:13)a + (a #(cid:13)a)∗(a†a)∗ − a†a − a #(cid:13)a − (a #(cid:13)a)∗ + 1 = (aa†a #(cid:13)a)∗ + (a†aa #(cid:13)a)∗ − a†a − (a #(cid:13)a)∗ + 1 = (a #(cid:13)a)∗ + (a†a)∗ − a†a − (a #(cid:13)a)∗ + 1 = 1. Since aa† + a†a − 1 is invertible, (aa† + a†a − 1)−1 = a #(cid:13)a + (a #(cid:13)a)∗ − 1. 5 To prove statement (iv), recall that according Theorem 3.1, a∗ ∈ A #(cid:13). In addition, accord- ing to the paragraph between Theorem 3.1 and the present Lemma, a #(cid:13) = ((a∗) #(cid:13))∗. However, according to [20, Theorem 2.19], (a∗)# = ((a∗) #(cid:13))2a∗. Thus, a# = a(((a∗)#(cid:13))∗)2 = a(a #(cid:13))2. Note that given a ring with involution R, Lemma 3.2 holds in such a context provided that a ∈ R is Moore-Penrose invertible, In the next theorem the continuity of the (dual) core inverse will be characterized. It is worth noticing that a ∈ A will be not assumed to be core invertible, dual core invertible, group invertible or Moore-Penrose invertible. Note also that the following well known result will be used in the proof of the theorem: given A a unital Banach algebra, b ∈ A and (bn)n∈ ⊂ A−1 a sequence such that (bn)n∈ converges to b, if (b−1 n )n∈ is a bounded sequence, then b is invertible and the sequence (b−1 n )n∈ converges to b−1. Theorem 3.3. Let A be a unital C ∗-algebra and consider a ∈ A. Let (an)n∈ ⊂ A #(cid:13) = A #(cid:13) be such that (an)n∈ converges to a. The following statements are equivalent. (i) The element a ∈ A #(cid:13) and (a #(cid:13) n )n∈ converges to a #(cid:13). (ii) The element a ∈ A #(cid:13) and (an #(cid:13))n∈ converges to a #(cid:13). (iii) The element a ∈ A# and (a# n )n∈ converges to a#. (iv) The element a ∈ A #(cid:13) and (an #(cid:13))n∈ is a bounded sequence. (v) The element a ∈ A #(cid:13) and (an #(cid:13))n∈ is a bounded sequence. (vi) The element a ∈ A†, (a† n)n∈ converges to a†, and (an #(cid:13)an)n∈ is a bounded sequence. (vii) The element a ∈ A†, (a† n)n∈ converges to a†, and (anan #(cid:13))n∈ is a bounded sequence. (viii) The element a ∈ A†, (a† n)n∈ converges to a†, and there exists ψ ∈ [0, π 2 ) such that ψn = ψan ≤ ψ for all n ∈ . Proof. Note that according to Theorem 3.1, (an)n∈ ⊂ A #(cid:13) ∩ A #(cid:13) ∩ A# ∩ A†. First the equivalence between statements (i) and (iii) will be proved. Suppose that state- n an)n∈ converges to n an. n an)n∈ converges to a#a, which according to [15, Theorem 2.4], implies that ment (i) holds. Then according to Theorem 3.1, a ∈ A#. In addition, (a #(cid:13) a #(cid:13)a. However, according to [20, Remark 2.17], a#a = a #(cid:13)a, and for each n ∈ , a# Consequently, (a# (a# n )n∈ converges to a#. n an = a #(cid:13) Suppose that statement (iii) holds. Note that according to [11, Theorem 6], a ∈ A†. In particular, according to Theorem 3.1, a ∈ A #(cid:13). Moreover, according to [3, Corollary 2.1 (ii)] and [4, Equation (2.1)], ka† nk = k(ana† n + a† nan − 1)a# n (ana† n + a† nan − 1)k ≤ kana† n + a† nan − 1k2ka# n k ≤ ka# n k. Consequently, (a† n)n∈ is a bounded sequence. 6 Now two cases need to be considered. to [20, Theorem 2.19], an (a† #(cid:13) = a# n)n∈ are bounded sequences, (an n ana† If a = 0, then a #(cid:13) = 0. However, according n )n∈ and n. Since (an)n∈ converges to 0 and (a# #(cid:13))n∈ converges to 0. Now suppose that a 6= 0. Since (an)n∈ converges to a, there exists and n0 ∈ such that an 6= 0, n ≥ n0. Without loss of generality, it is possible to assume that (an)n∈ ⊂ A #(cid:13) \ {0}. Thus, according to [14, Theorem 1.6], (a† n)n∈ converges to a†. However, according again to [20, Theorem 2.19], a #(cid:13) = a#aa† and for each n ∈ , a #(cid:13) n )n∈ converges to a #(cid:13). n. Therefore, (a #(cid:13) n = a# n ana† To prove the equivalence between statements (ii) and (iii), apply a similar argument to the one used to prove the equivalence between statements (i) and (iii). In particular, use the following identities, which holds for b ∈ A #(cid:13): (α) b#b = bb #(cid:13) ([20, Remark 2.17]); (β) b #(cid:13) = b†bb# ([20, Theorem 2.19]). It is evident that statement (i) implies statement (iv). Now suppose that statement (iv) holds. It will be proved that statement (iii) holds. According to Theorem 3.1, a ∈ A#. In addition, according to [20, Theorem 2.19], for each n ∈ , a# #(cid:13))2an. In particular, (a# n )n∈ converges to a#, equivalently, statement (iii) holds. n )n∈ is a bounded sequence. Consequently, according to [15, Theorem 2.4], (a# n = (an The equivalence between statements (ii) and (v) can be proved applying a similar argument to the one used to prove the equivalence between statements (i) and (iv), using in particular Lemma 3.2 (iv). Next it will be proved that statement (iv) implies statement (vi). Suppose then that #(cid:13)an)n∈ is a bounded sequence. In addition, according to statement (iv) holds. Then, (an Theorem 3.1, a ∈ A†. Now two cases need to be considered. Suppose first that a = 0. Since statement (iv) and (v) are equivalent, (an #(cid:13))n∈ is a bounded sequence. According to [20, Theorem 2.19], for each n ∈ , a† #(cid:13))n∈ and (an #(cid:13))n∈ are bounded sequences, (a† n)n∈ converges to 0 = a†. n = an #(cid:13)anan #(cid:13). Since (an If a 6= 0, as when it was proved that statement (iii) implies statement (i), it is possible to assume that (an)n∈ ⊂ A \ {0}. According to Lemma 3.2 (ii), k a† n k≤k ana† n + a† nan − 1 kk an #(cid:13) k≤ 3 k an #(cid:13) k . In particular, (a† (a† n)n∈ converges to a†. n)n∈ is a bounded sequence. However, according to [14, Theorem 1.6], Suppose that statement (vi) holds. It will be proved that statement (vi) implies statement n)n∈ and n. Since (a† (iv). According to [20, Theorem 2.19], for each n ∈ , an (an #(cid:13)an)n∈ are bounded sequences, (an #(cid:13))n∈ is a bounded sequence. Note that according to Lemma 3.2 (iii), for each n ∈ , bn = ana† nan − 1 is invertible and b−1 In fact, according to [16, Lemma 2.3], k b−1 #(cid:13)an k. Now, since (bn)n∈ converges to n b = aa† + a†a − 1, the element b is invertible, which in view of Lemma 3.2 (iii), is equivalent to a ∈ A #(cid:13). In addition, the sequence (b−1 n + a† n )n∈ is bounded. #(cid:13)an)∗ − 1. #(cid:13)an + (an n = an #(cid:13)ana† k=k an #(cid:13) = an According to [20, Theorem 2.19], for each n ∈ , an #(cid:13)an = anan #(cid:13). Thus, statement (vii) is an equivalent formulation of statement (vi). Finally, statements (vi) and (viii) will be proved to be equivalent. In fact, note that if nan − 1 is invertible, an = 0, then ψn = 0. In addition, according to Lemma 3.2 (iii), ana† n + a† 7 and when an 6= 0, according to Lemma 3.2 (iii), [4, Theorem 2.4 (iii)] and [16, Lemma 2.3], 1 cos ψn =k (ana† n + a† nan − 1)−1 k=k an #(cid:13)an + (an #(cid:13)an)∗ − 1 k=k an #(cid:13)an k . In particular, (an for all n ∈ . #(cid:13)an)n∈ is bounded if and only if there exists ψ ∈ [0, π 2 ) such that ψn ≤ ψ Theorem 3.3 shows that the continuity of the group inverse and of the Moore-Penrose inverse are central for the continuity of the core inverse and the dual core inverse. To learn more on the continuity of the group inverse and the Moore-Penrose inverse, see for example [2, 4, 15, 23, 24] and [12, 14, 17, 22, 24], respectively, see also [7, Chapter 4]. Observe that the conditions in statement (vi) of Theorem 3.3, (α) a ∈ A†, a† n → a†, and #(cid:13)an} is a bounded sequence, are independent from each other, as the following two (β) {an examples show. Example 3.4. Consider C as a C ∗-algebra. Let an = 1/n and a = 0. It is evident that an → a, a† n)n∈ does not converge to a† = 0. However, it should be clear that a #(cid:13) n = n. Therefore, a #(cid:13) #(cid:13)an)n∈ is a bounded sequence. n an = 1, and thus, (an n = n, and (a† Example 3.5. Consider the set of 2×2 complex matrices as a C ∗-algebra. Take the conjugate transpose of the matrix as the involution on this matrix. Let (ψn)n∈ be a sequence in (0, π/2) such that ψn → π/2 and let An =(cid:20) cos ψn 0 sin ψn 0 (cid:21) , 0 0 (cid:21) . A =(cid:20) 0 1 It is simple prove that 0 (cid:21) , n =(cid:20) cos ψn 0 sin ψn A† 1 0 (cid:21) , A† =(cid:20) 0 0 Therefore, (A† n)n∈ converges to A† and 0 (cid:21) . n =(cid:20) 1/ cos ψn 0 0 A #(cid:13) n An =(cid:20) 1 tan ψn 0 0 A #(cid:13) (cid:21) , which shows that (A #(cid:13) sequence. n An)n∈ is not bounded. Note also that (An #(cid:13))n∈ is not a convergent Observe also that if A is a unital C ∗-algebra and (an)n∈ ⊂ A #(cid:13) is such that (an)n∈ #(cid:13)an)n∈ is a convergent converges to a ∈ A, Example 3.4 also shows that the condition (an sequence does not implies that (an #(cid:13))n∈ is convergent. It is worth noticing that Example 3.5 also proves that A #(cid:13) = A #(cid:13) is not in general a closed set. In fact, using the same notation as in Example 3.5, (An)n∈ ⊂ A #(cid:13), (An)n∈ converges to A but A /∈ A #(cid:13) (A2 = 0, rk(A2) = 0 6= 1 = rk(A), i.e., A is not group invertible). Next an extension of [4, Theorem 2.7] will be derived from Theorem 3.3. 8 Corollary 3.6. Let A be a unital C ∗-algebra and consider a ∈ A. Suppose that the sequence (an)n∈ ⊂ A# is such that (an)n∈ converges to a. Then, the following statements are equivalent. (i) The element a ∈ A# and (a# n )n∈ converges to a#. (ii) The sequence (a# n )n∈ is bounded. (iii) The element a ∈ A†, (a† n)n∈ converges to a†, and the sequence (a# n an)n∈ is bounded. (iv) The element a ∈ A†, (a† n)n∈ converges to a†, and there exists ψ ∈ [0, π 2 ) such that ψn = ψan ≤ ψ for all n ∈ . Proof. Statement (ii) is a consequence of statement (i). Suppose that statement (ii) holds. Then, (a# that a ∈ A† and (a† of [4, Theorem 2.7] (see statement (ii) implies statement (iii) in [4, Theorem 2.7]). n an)n∈ is a bounded sequence. To prove n)n∈ converges to a†, proceed as in the corresponding part of the proof Suppose that statement (iii) holds. First note that if an = 0, then ψn = 0. In addition, according to [4, Theorem 2.5], if an 6= 0, then, kana# n k = 1 cos ψan . Therefore, the sequence (a# ψn = ψan ≤ ψ for all n ∈ . n an)n∈ is bounded if and only if there exists ψ ∈ [0, π 2 ) such that To prove that statement (iv) implies statement (i), apply Theorem 3.3 (equivalence be- tween statements (iii) and (viii)). In Theorem 3.3 and Corollary 3.6 the general case has been presented for sake of complete- ness. However, the case a = 0 is particular and it deserves to be studied. Recall that given a unital C ∗-algebra A, if (an)n∈ ⊂ A−1 is such that (an)n∈ converges to 0, then the sequence n )n∈ is unbounded. Next the case of a sequence (an)n∈ ⊂ A# = A #(cid:13) = A #(cid:13) ⊆ A† (a−1 such that it converges to 0 will be studied. In first place the Moore-Penrose inverse will be considered. Remark 3.7. Let A be a unital C ∗-algebra and consider a ∈ A† and (an)n∈ ⊂ A† such that (an)n∈ converges to a. Recall that according to [14, Theorem 1.6], the following statements are equivalent. (i) The sequence (a† n)n∈ converges to a†. (ii) The sequence (ana† n)n∈ converges to aa†. (iii) The sequence (a† nan)n∈ converges to a†a. (iv) The sequence (a† n)n∈ is bounded. Now when a = 0, according to [14, Theorem 1.3], the following equivalence holds: 9 (v) A necessary and sufficient condition for (a† n)n∈ to converge to 0 is that the sequence (a† n)n∈ is bounded. However, concerning the convergence of (ana† n k< 1, then ana† a self-adjoint idempotent, if k ana† similar result can be derived for the convergence of (a† statements are equivalent. n)n∈, note that given n ∈ , since ana† n is n = 0, which implies that an = 0; a nan)n∈. Consequently, the following (vi) The sequence (a† n)n∈ converges to 0. (vii) There exists n0 ∈ such that for n ≥ n0, an = 0. Therefore, according to statements (v)-(vii), given (an)n∈ ⊂ A† such that (an)n∈ converges to 0, there are only two possibilities. (viii) There exists n0 ∈ such that for n ≥ n0, an = 0; or (ix) the sequence (a† n)n∈ is unbounded. In the following proposition, sequences of group invertible or (dual) core invertible elements that converge to 0 will be studied. Proposition 3.8. Let A be a unital C ∗-algebra and consider a sequence (an)n∈ ⊂ A# = A #(cid:13) = A #(cid:13). such that (an)n∈ converges to 0. The following statements are equivalent. (i) The sequence (a# n )n∈ converges to 0. (ii) The sequence (an #(cid:13))n∈ converges to 0. (iii) The sequence (an #(cid:13))n∈ converges to 0. (iv) The sequence (a# n )n∈ is bounded. (v) There exists n0 ∈ such that for n ≥ n0, an = 0. In addition, there exist only two possibilities for the sequence (an)n∈. (vi) There exists n0 ∈ such that for n ≥ n0, an = 0; or (vii) the sequence (a# n )n∈ is unbounded. Moreover, statement (vii) is equivalent to the following two statements. (viii) the sequence (an #(cid:13))n∈ is unbounded. (ix) the sequence (an #(cid:13))n∈ is unbounded. Proof. According to Theorem 3.3, statements (i)-(iii) are equivalent. It is evident that statement (i) implies statement (iv). Suppose that statement (iv) holds. According to [11, Theorem 6], (an)n∈ ⊂ A# ⊂ A†. In addition according to [3, Corollary 2.1 (ii)], k a† n k≤k a# n kk ana† n + a† nan − 1 k2≤ 9 k a# n k . 10 In particular, the sequence (a† converges to 0. However, according to Remark 3.7 (vi)-(vii), statement (v) holds. n)n∈ is bounded. Thus, according to Remark 3.7 (v), (a† n)n∈ It is evident that statement (v) implies statement (i). Statements (vi) and (vii) can be derived from what has been proved. According to Theorem 3.3, statements (vii)-(ix) are equivalent. To prove the second characterization of this section some preparation is needed. Remark 3.9. Let A be a unital C ∗-algebra and consider a ∈ A #(cid:13) = A #(cid:13). If La : A → A and Ra : A → A are the left and the right multiplication operators defined by a, i.e., for x ∈ A, La(x) = ax, Ra(x) = xa, respectively, then according to [20, Theorem 2.14], La #(cid:13) LaLa #(cid:13) = La #(cid:13) , Ra #(cid:13) RaRa #(cid:13) = Ra #(cid:13) . Note also that according to Definition 2.1, R(La #(cid:13) ) = aA, R(Ra #(cid:13) ) = Aa∗, N(La #(cid:13) ) = (a∗)◦. N(Ra #(cid:13) ) = ◦a. Therefore, La #(cid:13) = (La)(2) In addition, since Laa #(cid:13) = LaLa #(cid:13) , Raa #(cid:13) = Ra #(cid:13) Ra ∈ L(A) are idempotents, observe that according to Definition 2.1 and [20, Theorem 2.14], aA,(a∗)◦ and Ra #(cid:13) = (Ra)(2) Aa∗,◦a. R(Laa #(cid:13) ) = aA, N(Laa #(cid:13) ) = (a∗)◦, R(Raa #(cid:13) ) = Aa∗, N(Raa #(cid:13) ) = ◦a. Similar arguments prove the following facts: La #(cid:13) = (La)(2) a∗ A,a◦, Ra #(cid:13) = (Ra)(2) Aa,◦(a∗) and R(La #(cid:13)a) = a∗A, N(La #(cid:13)a) = a◦, R(Ra #(cid:13)a) = Aa, N(Ra #(cid:13)a) = ◦(a∗). Next follows the second characterization of the continuity of the (dual) core inverse. In this case, the notion of the gap between subspaces will be used. Theorem 3.10. Let A be a unital C ∗-algebra and consider a ∈ A #(cid:13) = A #(cid:13), a 6= 0. Consider a sequence (an)n∈ ⊂ A #(cid:13) = A #(cid:13) such that (an)n∈ converges to a. The following statements are equivalent. (i) (a #(cid:13) n )n∈ converges to a #(cid:13). (ii) (ana #(cid:13) n )n∈ converges to aa #(cid:13). (iii) (bδ (anA, aA))n∈ and (bδ ((a∗ n)◦, (a∗)◦))n∈ converge to 0. 11 (iv) (bδ (Aa∗ n, Aa∗))n∈ and (bδ (◦an, ◦a))n∈ converge to 0. (v) (an #(cid:13))n∈ converges to a #(cid:13). (vi) (an #(cid:13)an)n∈ converges to a #(cid:13)a. n (vii) (bδ (a∗ A, a∗A))n∈ and (bδ (a◦ (viii) (bδ (Aan, Aa))n∈ and (bδ (◦(a∗ n, a◦))n∈ converge to 0. n), ◦(a∗)))n∈ converge to 0. Proof. It is evident that statement (i) implies statement (ii). Suppose that statement (ii) holds. According to Remark 3.9, aA = R(Laa #(cid:13) ), (a∗)◦ = N(Laa #(cid:13) ), anA = R(Lana ) and (a∗ ) (n ∈ ). However, according to [15, Lemma 3.3], statement (iii) holds. n)◦ = N(Lana #(cid:13) Suppose that statement (iii) holds. Recall that according to Remark 3.9, #(cid:13) n n La #(cid:13) = (La)(2) aA,(a∗)◦ , La #(cid:13) n = (Lan)(2) an A,(a∗ n)◦ , for each n ∈ . Let κ = kLakkLa #(cid:13) k = kakka #(cid:13)k and consider n0 ∈ such that for all n ≥ n0, rn =bδ(cid:16)N(cid:16)(Lan)(2) sn =bδ(cid:16)R(cid:16)(Lan)(2) an A,(a∗ anA,(a∗ n)◦(cid:17) , N(cid:16)(La)(2) n)◦(cid:17) , R(cid:16)(La)(2) aA,(a∗)◦(cid:17)(cid:17) =bδ ((a∗ aA,(a∗)◦(cid:17)(cid:17) =bδ (anA, aA) < n)◦, (a∗)◦) < 1 , 3 + κ 1 (1 + κ)2 , and tn = kLa #(cid:13) kkLa − Lank = ka #(cid:13)kka − ank < 2κ (1 + κ)(4 + κ) . Thus, according to [10, Theorem 3.5], n − a #(cid:13)(cid:13)(cid:13) =(cid:13)(cid:13)(cid:13)La (cid:13)(cid:13)a #(cid:13) #(cid:13) n − La #(cid:13)(cid:13)(cid:13)(cid:13) ≤ which implies statement (i). (1 + κ)(sn + rn) + (1 + rn)tn 1 − (1 + κ)sn − κrn − (1 + rn)tn ka #(cid:13)k, Statements (i), (ii) and (iv) are equivalent. To prove this fact, apply a similar argument to the one used to prove the equivalence among statements (i), (ii) and (iii), using in particular Ra #(cid:13) = (Ra)(2) instead of the respectively left multiplication operators (Remark 3.9, n ∈ ). n,◦an, Raa #(cid:13) and Rana = (Ran)(2) Aa∗ Aa∗,◦a, Ra #(cid:13) n #(cid:13) n Statements (i) and (v) are equivalent (Theorem 3.3). To prove the equivalence among statements (v) and (viii), apply a similar argument to the one used to prove that statements (i)-(iv) are equivalent, using in particular Remark 3.9 and [10, Theorem 3.5]. Next, some bounds for ka #(cid:13) n − a #(cid:13)k will be proved, when (an)n∈ ⊂ A converges to a ∈ A in a C ∗-algebra A. Before, a technical lemma is presented. 12 Lemma 3.11. Let A be a unital C ∗-algebra and let a, b ∈ A #(cid:13) = A #(cid:13). Then (i) (ii) b #(cid:13) − a #(cid:13) = b #(cid:13)b(b† − a†)(1 − aa #(cid:13)) + b #(cid:13)(a − b)a #(cid:13) + (1 − b #(cid:13)b)(b − a)a†a #(cid:13). b #(cid:13) − a #(cid:13) = (1 − a #(cid:13)a)(b† − a†)bb #(cid:13) + a #(cid:13)(a − b)b #(cid:13) + a #(cid:13)a†(b − a)(1 − bb #(cid:13)). Proof. To prove statement (i), recall that since a and b are core invertible, a and b are Moore- In addition, according to [26, Theorem 3.1], b = b #(cid:13)b2. Penrose invertible (Theorem 3.1). Thus, according to Lemma 3.2 (i), (1 − b #(cid:13)b)(b − a)a†a #(cid:13) = −(1 − b #(cid:13)b)aa†a #(cid:13) = −(1 − b #(cid:13)b)a #(cid:13) = b #(cid:13)ba #(cid:13) − a #(cid:13). Now, according to [20, Theorem 2.19], b #(cid:13) = b #(cid:13)bb†. In addition, a∗aa #(cid:13) = a∗(aa #(cid:13))∗ = (aa #(cid:13)a)∗ = a∗, i.e., a∗(1 − aa #(cid:13)) = 0. Moreover, since a† = a†aa† = a†(aa†)∗ = a†(a†)∗a∗, a†(1 − aa #(cid:13)) = 0. Therefore, b #(cid:13)b(b† − a†)(1 − aa #(cid:13)) = b #(cid:13)bb†(1 − aa #(cid:13)) = b #(cid:13)(1 − aa #(cid:13)) = b #(cid:13) − b #(cid:13)aa #(cid:13). As a result, b #(cid:13) − a #(cid:13) = b #(cid:13) − b #(cid:13)aa #(cid:13) + b #(cid:13)aa #(cid:13) − b #(cid:13)ba #(cid:13) + b #(cid:13)ba #(cid:13) − a #(cid:13) = b #(cid:13)b(b† − a†)(1 − aa #(cid:13)) + b #(cid:13)(a − b)a #(cid:13) + (1 − b #(cid:13)b)(b − a)a†a #(cid:13). To prove statement (ii), use that x #(cid:13) = ((x∗)#(cid:13))∗ (x ∈ A), and apply statement (i). Next the aforementioned bounds will be given. Theorem 3.12. Let A be a unital C ∗-algebra and consider a ∈ A #(cid:13) = A #(cid:13). The following statements holds. (i) If b ∈ A #(cid:13) = A #(cid:13), b 6= 0, then kb #(cid:13) − a #(cid:13)k ≤ kb† − a†k cos ψb +(cid:20)kb #(cid:13)k + ka†k cos ψb(cid:21) ka #(cid:13)kka − bk. (ii) In addition, kb #(cid:13) − a #(cid:13)k ≤ kb† − a†k cos ψb +(cid:20)kb #(cid:13)k + ka†k cos ψb(cid:21) ka #(cid:13)kka − bk. (iii) If also a 6= 0, then kb #(cid:13) − a #(cid:13)k = kb #(cid:13) − a #(cid:13)k ≤ kb† − a†k cos ψb ka†k(cid:0)kb†k + ka†k(cid:1) cos ψa cos ψb + ka − bk. 13 (iv) In particular, if a ∈ A #(cid:13) = A #(cid:13), a 6= 0, and (an)n∈ ⊂ A #(cid:13) = A #(cid:13), an 6= 0 for all n ∈ , then kan #(cid:13) − a #(cid:13)k = kan #(cid:13) − a #(cid:13)k ≤ ka† n − a†k cos ψn + where ψn = ψan. ka†k(cid:16)ka† nk + ka†k(cid:17) cos ψa cos ψn ka − ank, Proof. To prove statement (i), note that according to [16, Lemma 2.3], Lemma 3.2 (iii) and [4, Theorem 2.4 (iii)] k1 − b #(cid:13)bk = kb #(cid:13)bk = k(b #(cid:13)b) + (b #(cid:13)b)∗ − 1k = k(bb† + b†b − 1)−1k = 1 cos ψb . Observe that 1 − aa #(cid:13) is a self-adjoint idempotent, Hence k1 − aa #(cid:13)k = 1, and according to Lemma 3.11, kb #(cid:13) − a #(cid:13)k ≤ kb #(cid:13)bkkb† − a†kk1 − aa #(cid:13)k +hkb #(cid:13)kka #(cid:13)k + k1 − b #(cid:13)bkka†a #(cid:13)ki ka − bk = ≤ kb† − a†k cos ψb kb† − a†k cos ψb ka†a #(cid:13)k +(cid:20)kb #(cid:13)kka #(cid:13)k + +(cid:20)kb #(cid:13)k + cos ψb (cid:21) ka − bk. cos ψb(cid:21) ka #(cid:13)kka − bk. ka†k Statement (ii) can be derived from statement (i). In fact, recall that given x ∈ A #(cid:13) = A #(cid:13), x #(cid:13) = ((x∗) #(cid:13))∗. Moreover, if x ∈ A† \ {0}, then note that ψx∗ = ψx† = ψx. Now apply statement (i) to a∗ and b∗. Now observe that if a = 0 in statement (i), then kb #(cid:13)k ≤ kb†k cos ψb . Thus, if a 6= 0, then ka #(cid:13)k ≤ ka†k cos ψa . To prove statement (iii) for the core inverse, apply these inequalities to statement (i). To prove statement (iii) for the dual core inverse, proceed as in the proof of statement (ii). Statement (iv) can be derived from statement (iii). Remark 3.13. As it was used in the proof of Theorem 3.12, given a ∈ A #(cid:13) = A #(cid:13), a 6= 0, Theorem 3.12 (i) (respectively Theorem 3.12 (ii)) gives a relationship between the norm of a #(cid:13) (respectively of a #(cid:13)) and the norm of the a†: ka #(cid:13)k ≤ ka†k cos ψa ). Moreover, under the same hypothesis of Theorem 3.3, when a 6= 0, Theorem 3.12 (iv) gives an estimate of the convergence of (an cos ψa (respectively ka #(cid:13)k ≤ ka†k #(cid:13))n∈ and (an #(cid:13))n∈ to a #(cid:13) and a #(cid:13), respectively. 4 Continuity of (dual) core invertible Hilbert space operators Let H be a Hilbert space and consider A ∈ L(H). The definition of core invertible Hilbert space operators was given in [21, Definition 3.2]. In fact, A ∈ L(H) is said to be core invertible, if there exists X ∈ L(H) such that A = AXA, R(X) = R(A), N(X) = N(A∗). 14 Thus, when A ∈ L(H), two definitions of the core inverse of A has been given: as an element of the C ∗-algebra L(H) and as Hilbert space operator. However, as the following proposition shows, both definitions coincide in the Hilbert space context. Proposition 4.1. Let H be a Hilbert space and consider A ∈ L(H). The following statements are equivalent. (i) The core inverse of A exists. (ii) There exists an operator X ∈ L(H) such that AXA = A, R(X) = R(A) and N(X) = N(A∗). Moreover, in this case X = A #(cid:13) = A(2) R(A),N(A∗). Proof. Suppose that A #(cid:13) exists. Then, A = AA #(cid:13)A and there are operator S, T , U , V ∈ L(H) such that A #(cid:13) = AS, A = A #(cid:13)T, A #(cid:13) = U A∗, A∗ = V A #(cid:13). In particular, R(A #(cid:13)) = R(A) and N(A #(cid:13)) = N(A∗). Now suppose that statement (ii) holds. Then, there exists X ∈ L(H) such that R(X) = R(A). According to [8, Theorem 1], there are L, K ∈ L(H) such that A = XL and X = AK. In particular, X L(H) = AL(H). In addition, since R(A) is closed, R(X) is closed, which is equivalent to the fact that X is regular. Now since A∗ is regular, according to [5, Remark 6], there exist operators M , N ∈ L(H) such that X = M A∗ and A∗ = N X. In particular, L(H)X = L(H)A∗. Since A = AXA and the core inverse is unique, when it exists ([20, Theorem 2.14]), X = A #(cid:13). Finally, since according again to [20, Theorem 2.14], A #(cid:13) is an outer inverse, according to what has been proved, A #(cid:13) = A(2) R(A),N(A∗). As for the core inverse case, a definition of dual core invertible Hilbert space operators was given in [21, Definition 3.3]. In the following proposition the equivalence between Definition 2.2 and [21, Definition 3.3] will be considered. Proposition 4.2. Let H be a Hilbert space and consider A ∈ L(H). The following statements are equivalent. (i) The dual core inverse of A exists. (ii) There exists an operator X ∈ L(H) such that AXA = A, R(X) = R(A∗) and N(X) = N(A). Moreover, in this case X = A #(cid:13) = A(2) R(A∗),N(A). Proof. Apply a similar argument to the one used in Proposition 4.1. Note that the relationship between the (dual) core inverse and the outer inverse with prescribed range and null space for the case of square complex matrices was studied in [19, Theorem 1.5] (apply [20, Theorem 4.4]). Next the continuity of the (dual) core inverse will be characterized using the gap between subspaces. The next theorem is the Hilbert space version of Theorem 3.10 15 Theorem 4.3. Let H be a Hilbert space and consider A ∈ L(H), A 6= 0, such that A is (dual) core invertible. Suppose that there exists a sequence of operators (An)n∈ ⊂ L(H) such that for each n ∈ , An is (dual) core invertible and (An)n∈ converges to A. Then, the following statements are equivalent. (i) The sequence (An #(cid:13))n∈ converges to A #(cid:13). (ii) The sequence (An #(cid:13))n∈ converges to A #(cid:13). (iii) The sequence (An #(cid:13)An)n∈ converges to A #(cid:13)A. (iv) The sequence (AnAn #(cid:13))n∈ converges to A #(cid:13)A. #(cid:13)), R(A #(cid:13))))n∈ converges to 0. #(cid:13)), N(A #(cid:13))))n∈ converges to 0. (v) The sequence (bδ (R(An (vi) The sequence (bδ (R(An), R(A)))n∈ converges to 0. (vii) The sequence (bδ (N(An (viii) The sequence (bδ (N(A∗ (ix) The sequence (bδ (R(An #(cid:13)), R(A #(cid:13))))n∈ converges to 0. (x) The sequence (bδ (R(A∗ (xi) The sequence (bδ (N(An #(cid:13)), N(A #(cid:13))))n∈ converges to 0. (xii) The sequence (bδ (N(An), N(A)))n∈ converges to 0. n), N(A∗)))n∈ converges to 0. n), R(A∗)))n∈ converges to 0. Proof. First of all recall that L(H) #(cid:13) = L(H) #(cid:13) (Theorem 3.1). Statements (i)-(iv) are equivalent (Theorem 3.10). According to [15, Lemma 3.3], state- ment (iii) implies statement (v) and according to Proposition 4.1 and [13, Chapter 4, Section 2, Subsection 3, Theorem 2.9], Statements (v)-(viii) are equivalent. Now suppose that statement (vi) holds. Thus, according to what has been proved, n), N(A∗)))n∈ converge to 0 (recall that the sequences (bδ (R(An), R(A)))n∈ and (bδ (N(A∗ according to [13, Chapter 4, Section 2, Subsection 3, Theorem 2.9], bδ (R(An), R(A)) = bδ ((N(A∗ n), N(A∗)), n ∈ ). In addition, according to Proposition 4.1, for each n ∈ , A #(cid:13) n = (An)(2) R(An),N(A∗ n), A #(cid:13) = A(2) R(A),N(A∗). Let κ =k A kk A #(cid:13) k and consider n0 ∈ such that for all n ≥ n0, wn =bδ(cid:16)N((An)(2) R(An),N(A∗ n)), N(A(2) =bδ (R(An), R(A)) =bδ(cid:16)R((An)(2) 2κ zn =k A #(cid:13) kk A − An k< R(A),N(A∗ ))(cid:17) =bδ ((N(A∗ n)), R(A(2) R(An),N(A∗ n), N(A∗)) R(A),N(A∗))(cid:17) < 1 (3 + κ)2 , . (1 + κ)(4 + κ) 16 Since 1 (3+κ)2 ≤ min{ 1 3+κ , 1 (1+κ)2 }, according to [10, Theorem 3.5], k A #(cid:13) n − A #(cid:13) k≤ 2(1 + κ)wn + (1 + wn)zn 1 − (1 + 2κ)wn − (1 + wn)zn k A #(cid:13) k, which implies statement (i). Now, according to [15, Lemma 3.3], statement (iv) implies statement (xi) and according to Proposition 4.2 and [13, Chapter 4, Section 2, Subsection 3, Theorem 2.9], Statements (ix)-(xii) are equivalent. Suppose that statement (x) holds. Since then statement (xii) also holds, to prove that statement (ii) holds, it is enough to apply an argument similar to the one used to prove that statement (vi) implies statement (i), interchanging in particular A with A∗, An with A∗ n, A #(cid:13) R(A∗),N(A), with A #(cid:13), An and κ with κ′ =k A kk A #(cid:13) k. R(A),N(A∗) with A(2) #(cid:13) with An #(cid:13), (An)(2) n),N(An), A(2) R(An),N(A∗ n) with (An)(2) R(A∗ Next the continuity of the (dual) core inverse will be studied in a particular case. To this end, two results from [24] need to be extended first. Proposition 4.4. Let X be a Banach space and consider A ∈ L(X) such that A is group invertible and the codimension of R(A) is finite. Suppose that there exists a sequence of operators (An)n∈ ⊂ L(X) such that for each n ∈ , An is group invertible and (An)n∈ converges to A. Then the following statements are equivalent. (i) The sequence (A# n )n∈ converges to A#. (ii) For all sufficiently large n ∈ , codim R(An) = codim R(A). Proof. Recall that A ∈ L(X) is group invertible if and only if A∗ ∈ L(X∗) is group invertible. In addition, dim N(A∗) is finite A similar statement holds for each An ∈ L(X) (n ∈ ). n)n∈ ⊂ L(X∗) converges to A∗. Thus, according to [24, Theorem 3], statement (i) is and (A∗ n) = dim N(A∗), which in equivalent to the fact that for all sufficiently large n ∈ , dim N(A∗ turn is equivalent to statement (ii). Proposition 4.5. Let H be a Hilbert space and consider A ∈ L(H) such that A is Moore- Penrose invertible and the codimension of R(A) is finite. Suppose that there exists a sequence of operators (An)n∈ ⊂ L(H) such that for each n ∈ N, An is Moore-Penrose invertible and (An)n∈ converges to A. Then the following statements are equivalent. (i) The sequence (A† n)n∈ converges to A†. (ii) For all sufficiently large n ∈ , codim R(An) = codim R(A). Proof. Apply a similar argument to the one in the proof of Proposition 4.4, using in particular [24, Corollary 10] instead of [24, Theorem 3]. Corollary 4.6. Let H be a Hilbert space and consider A ∈ L(H) such that A is group invertible and either the codimension of R(A) is finite or dim N(A) is finite. Suppose that there exists a sequence of operators (An)n∈ ⊂ L(H) such that for each n ∈ N, An is group invertible and (An)n∈ converges to A. Then, the following statements are equivalent. 17 (i) The sequence (A# n )n∈ converges to A#. (ii) The sequence (A† n)n∈ converges to A†. Proof. Recall that given and operator S ∈ L(H) such that S is group invertible, then S is Moore-Penrose invertible ([11, Theorem 6]). To conclude the proof apply, when dim N(A) is finite, [24, Theorem 3] and [24, Corollary 10], and when codimension of R(A) is finite, Proposition 4.4 and Proposition 4.5. Now a characterization of the continuity of the (dual) core inverse for a particular case of Hilbert spaces operators will be presented. Theorem 4.7. Let H be a Hilbert space and consider A ∈ L(H) such that A is (dual) core invertible and either the codimension of R(A) is finite or dim N(A) is finite. Suppose that there exists a sequence of operators (An)n∈ ⊂ L(H) such that for each n ∈ , An is (dual) core invertible and (An)n∈ converges to A. The following statements are equivalent. (i) The sequence (A #(cid:13) n )n∈ converges to A #(cid:13). (ii) The sequence (An #(cid:13))n∈ converges to A #(cid:13). (iii) The sequence (A† n)n∈ converges to A†. When dim N(A) is finite, statements (i)-(iii) are equivalent to the following statement. (iv) For all sufficiently large n ∈ , dim N(An) = dim N(A). When codim (A) is finite, statements (i)-(iii) are equivalent to the following statement. (v) For all sufficiently large n ∈ , codim R(An) = codim R(A). Proof. Apply Theorem 3.3, Corollary 4.6, [24, Theorem 3] and Proposition 4.4. For the case A = 0, apply Remark 3.7 and Proposition 3.8. Now the finite dimensional case will be derived from Theorem 4.7. It is worth noticing that the following corollary also provides a different proof of a well known result concerning the continuity of the Moore-Penrose inverse in the matricial setting, see [25, Theorem 5.2]. Corollary 4.8. Let A ∈ Cm be a (dual) core invertible matrix. Suppose that exists a sequence (An)n∈ ⊂ Cm of (dual) core invertible matrices such that (An)n∈ converges to A. The following statements are equivalent. (i) The sequence (A #(cid:13) n )n∈ converges to A #(cid:13). (ii) The sequence (An #(cid:13))n∈ converges to A #(cid:13). (iii) The sequence (A† n)n∈ converges to A†. (iv) There exists n0 ∈ such that rk(An) = rk(A), for n ≥ n0. Proof. Apply Theorem 4.7. 18 5 Differentiability of the (dual) core inverse To prove the main results of this section, some preparation is needed. Let U ⊆ R be an open set and consider a : U → A a function such that a(U ) ⊆ A #(cid:13). Since according to Theorem 3.1, A #(cid:13) = A #(cid:13) = A# ⊂ A†, it is possible to consider the functions which are defined as follows. Given u ∈ U , a #(cid:13), a #(cid:13), a#, a† : U → A, a #(cid:13)(u) = (a(u)) #(cid:13), a#(u) = (a(u))#, a #(cid:13)(u) = (a(u)) #(cid:13), a†(u) = (a(u))†. Since in this section functions instead of sequence will be considered and the notion of continuity will be central in the results concerning differentiability, Theorem 3.3 will be reformulated for functions. Theorem 5.1. Let A be a unital C ∗-algebra and consider U ⊆ R an open set and a function a : U → A such that a(U ) ⊆ A #(cid:13) and a is continuous at t0 ∈ U . The following statements are equivalent. (i) The element a(t0) ∈ A #(cid:13) and the function a #(cid:13) is continuous at t0. (ii) The element a(t0) ∈ A #(cid:13) and the function a #(cid:13) is continuous at t0. (iii) The element a(t0) ∈ A# and the function a# is continuous at t0. (iv) The element a(t0) ∈ A #(cid:13) and there exists an open set V ⊆ U such that t0 ∈ V and the function a #(cid:13) is bounded on V . (v) The element a(t0) ∈ A #(cid:13) and there exists an open set W ⊆ U such that t0 ∈ W and the function a #(cid:13) is bounded on W . (vi) The element a(t0) ∈ A†, the function a† is continuous at t0, and there exists and open set I ⊆ U such that t0 ∈ I and the function a #(cid:13)a is bounded on I. (vii) The element a(t0) ∈ A†, the function a† is continuous at t0 , and there exists and open set J ⊆ U such that t0 ∈ J and the function aa #(cid:13) is bounded on J. (viii) The element a(t0) ∈ A†, the function a† is continuous at t0, and there exist an open set Z such that t0 ∈ Zand ψ ∈ [0, π 2 ) such that when a(t) 6= 0 (t ∈ Z), ψt = ψa(t) ≤ ψ. Proof. Apply Theorem 3.3. Remark 5.2. Note that under the same hypotheses of Theorem 5.1, when a(t0) = 0, the continuity of the function a #(cid:13) (respectively a #(cid:13), a#, a†) at t0 is equivalent to the following condition: there exists an open set K ⊆ U , t0 ∈ K and a(t) = 0, for all t ∈ K (Remark 3.7, Proposition 3.8). 19 To study the differentiability of the (dual) core inverse, the differentiability of the Moore- Penrose inverse need to be considered first. Remark 5.3. Let A be a unital C ∗-algebra and consider an open set U and a : U → A a function such that a(U ) ⊂ A† and there is t0 such that a is differentiable at t0. Thus, a necessary and sufficient condition for a† to be differentiable at t0 is that a† is continuous at t0. In fact, if a(t0) 6= 0, there is an open set V ⊆ U such that t0 ∈ V and a(t) 6= 0 for t ∈ V , and then according to [14, Theorem 2.1], this equivalence holds. On the other hand, if a(t0) = 0, according to Remark 3.7 (vi)-(vii), the function a† is continuous at t0 if and only if there exists an open set W such that t0 ∈ W and a(t) = 0 for t ∈ W , which implies that a† is differentiable at t0. As a result, in [14, Theorem 2.1] it is not necessary to assume that a(t) 6= 0 for t in a neighbourhood of t0. In the following theorem the differentiability of the (dual) core inverse will be studied. Note that the following notation will be used. Given a unital C ∗-algebra A, if U ⊆ R is an open set and b : U → A is a function, then b∗ : U → A will denote the function b∗(t) = (b(t))∗ (t ∈ U ). In addition, if b : U → A is differentiable at t0 ∈ U , then b′(t0) will stand for the derivative of b at t0. Theorem 5.4. Let A be a unital C ∗-algebra and consider U ⊆ R an open set and a : U → A a function such that is differentiable at t0 ∈ U and a(U ) ⊂ A #(cid:13) = A #(cid:13) = A#. The following statements are equivalent. (i) The function a #(cid:13) is continuous at t0. (ii) The function a #(cid:13) is differentiable at t0. (iii) The function a #(cid:13) is differentiable at t0. (iv) The function a# is differentiable at t0. Furthermore, the following formulas hold. (v) (vi) (vii) (a #(cid:13))′(t0) = a #(cid:13)(t0)a(t0)(a†)′(t0)(1 − a(t0)a #(cid:13)(t0)) − a #(cid:13)(t0)a′(t0)a #(cid:13)(t0) + (1 − a #(cid:13)(t0)a(t0))a′(t0)a†(t0)a #(cid:13)(t0). (a #(cid:13))′(t0) = (1 − a #(cid:13)(t0)a(t0))(a†)′(t0)a(t0)a #(cid:13)(t0) − a #(cid:13)(t0)a′(t0)a #(cid:13)(t0) + a #(cid:13)(t0)a†(t0)a′(t0)(1 − a(t0)a #(cid:13)(t0)). (a#)′(t0) = 2a #(cid:13)(t0)(a #(cid:13))′(t0)a(t0) + (a #(cid:13)(t0))2a′(t0) = a′(t0)(a #(cid:13)(t0))2 + 2a(t0)a #(cid:13)(t0)(a #(cid:13))′(t0) = (a #(cid:13))′(t0)a(t0)a #(cid:13)(t0) + a #(cid:13)(t0)a′(t0)a #(cid:13)(t0) + a #(cid:13)(t0)a(t0)(a #(cid:13))′(t0). 20 Proof. According to Lemma 3.11, a #(cid:13)(t) − a #(cid:13)(t0) = a #(cid:13)(t)a(t)(a†(t) − a†(t0))(1 − a(t0)a #(cid:13)(t0)) + a #(cid:13)(t)(a(t0) − a(t))a #(cid:13)(t0) + (1 − a #(cid:13)(t)a(t))(a(t) − a(t0))a†(t0)a #(cid:13)(t0). Now suppose that statement (i) holds. According to Theorem 5.1, the function a† is continuous at t0, and according to [14, Theorem 2.1] and Remark 5.3, the function a† is differentiable at t0. Thus, a #(cid:13)(t)a(t)(a†(t) − a†(t0))(1 − a(t0)a #(cid:13)(t0)) t − t0 converges to a #(cid:13)(t0)a(t0)(a†)′(t0)(1 − a(t0)a #(cid:13)(t0)). In addition, a #(cid:13)(t)(a(t0) − a(t))a #(cid:13)(t0) t − t0 converges to −a #(cid:13)(t0)a′(t0)a #(cid:13)(t0), and (1 − a #(cid:13)(t)a(t))(a(t) − a(t0))a†(t0)a #(cid:13)(t0) t − t0 converges to (1 − a #(cid:13)(t0)a(t0))a′(t0)a†(t0)a #(cid:13)(t0). Consequently statements (ii) and (v) hold. It is evident that statement (ii) implies statement (i). Now observe that the function a∗ : U → A is differentiable at t0 and a∗(U ) ⊂ A #(cid:13) (Theorem 3.1). Suppose that statement (i) holds. According to the identity (a∗) #(cid:13)(t) = (a #(cid:13))∗(t) and Theorem 5.1, the function (a∗) #(cid:13) : U → A is continuous at t0. Thus, according to what has been proved, the function (a∗) #(cid:13) : U → A is differentiable at t0. Therefore, the function a #(cid:13) : U → A #(cid:13) is differentiable at t0. Consequently, statement (iii) holds. Furthermore, since (a #(cid:13))′(t0) = (((a)∗ #(cid:13))′)∗(t0), to prove statement (vi), apply statement (v). On the other hand, if statement (iii) holds, then the function a #(cid:13) is continuous at t0. According to Theorem 5.1, statement (i) holds. Suppose that statement (i) holds. According to [20, Theorem 2.19] and Lemma 3.2 (iv), the following identities hold. a# = (a #(cid:13))2a = a(a #(cid:13))2 = a #(cid:13)aa #(cid:13). Therefore, according to what has been proved, the function a# is differentiable at t0. Fur- thermore, from these idenetities statement (vii) can be derived. On the other hand, according to Theorem 5.1, statement (iv) implies statement (i). Remark 5.5. Under the same hypothesis of Theorem 5.4, the following facts should be noted. (i). When a(t0) = 0, according to Remark 5.2, (a #(cid:13))′(t0) = (a #(cid:13))′(t0) = (a#)′(t0) = (a†)′(t0) = 0. 21 (ii). Recall that in [14, Theorem 2.1], a formula concerning the derivative of the function a† at t0 was given. (iii). Note that according to Theorem 5.1, a necessary and sufficient condition for the function a #(cid:13) (respectively a#) to be differentiable at t0 is that a #(cid:13) (respectively a#) is continous at t0. In fact, the continuity of one of the functions a #(cid:13), a #(cid:13) and a# at a point t0 is equivalent to the continuity and the differentiability of the three functions under consideration at t0 (Theorem 5.1 and Theorem 5.4). (iv). According to [20, Theorem 2.19], a† = a #(cid:13)aa #(cid:13), a #(cid:13) = a#aa†, a #(cid:13) = a†aa#. Thus, the derivative of a†, a #(cid:13) and a #(cid:13) at t0 can also be computed as follows: (a†)′(t0) = (a #(cid:13))′(t0)a(t0)a #(cid:13)(t0) + a #(cid:13)(t0)a′(t0)a #(cid:13)(t0) + a #(cid:13)(t0)a(t0)(a #(cid:13))′(t0). (a #(cid:13))′(t0) = (a#)′(t0)a(t0)a†(t0) + a#(t0)a′(t0)a†(t0) + a#(t0)a(t0)(a†)′(t0). (a #(cid:13))′(t0) = (a†)′(t0)a(t0)a#(t0) + a†(t0)a′(t0)a#(t0) + a†(t0)a(t0)(a#)′(t0). References [1] O. M. Baksalary, G. Trenkler, Core inverse of matrices, Linear Multilinear Algebra 58 (2010) 681 -- 697. [2] J. Ben´ıtez, X. Liu, On the continuity of the group inverse, Oper. Matrices 6 (2012) 859-868. [3] J. Ben´ıtez, D. Cvetkovi´c-Ili´c, On the elements aa† and a†a in a ring, Appl. Math. Comput. 222 (2013) 478 -- 489. [4] J. Ben´ıtez, D. Cvetkovi´c-Ili´c, X. Liu, On the continuity of the group inverse in C ∗-algebras Banach J. Math. Anal. 8 (2014) 204 -- 213. [5] E. Boasso, Drazin spectra of Banach space operators and Banach algebra elements, J. Math. Anal. Appl. 359 (2009) 48 -- 55. [6] E. Boasso, G. Kant´un-Montiel, The (b, c)-inverse in rings and in the Banach context, Mediterr. J. Math. 14 (2017), doi:10.1007/s00009-017-0910-1. [7] D. Djordjevi´c, V. Rakocevi´c, Lectures on Generalized Inverses, Faculty of Sciences and Mathematics, University of Nis, Nis, Serbia, 2008. [8] R. G. Douglas, On majorization, factorization and range inclusion of operators on Hilbert spaces, Proc. Amer. Math. Soc. 17 (1966) 413-415. [9] M. P. Drazin, A class of outer generalized inverses, Linear Algebra Appl. 436 (2012) 1909-1923. 22 [10] F. Du, Y. Xue, Perturbation analysis of A(2) T,S on Banach spaces, Electron. J. Linear Algebra 23 (2012) 586 -- 598. [11] R.E. Harte, M. Mbekhta, On generalized inverses in C ∗-algebras, Studia Math. 103 (1992) 71-77. [12] R.E. Harte, M. Mbekhta, Generalized inverses in C ∗-algebras II, Studia Math. 106 (1993) 129 -- 138. [13] T. Kato, Perturbation Theory for Linear Operators, Springer Verlarg, Berlin, Heidelberg, New York, 1980. [14] J. J.Koliha, Continuity and diffrentiability of the Moore-Penrose inverse in C ∗-algebras, Math. Scand. 88 (2001) 154-160. [15] J.J. Koliha, V. Rakocevi´c, Continuity of the Drazin inverse II, Studia Math. 131 (1998) 167-177. [16] J.J. Koliha, V. Rakocevi´c, On the norm of idempotents in C ∗-algebras, Rocky Mountain J. Math. 34 (2004) 685 -- 697. [17] M. Mbekhta, Conorme et inverse g´en´eralis´e dans les C ∗-alg`ebres, Canad. Math. Bull. 35 (1992) 515 -- 522. [18] R. Penrose, A generalized inverse for matrices, Math. Proc. Cambridge Philos. Soc. 3 (1955) 406-413. [19] D. S. Raki´c, A note on Rao and Mitra's constrained inverse and Drazin's (b, c) inverse, Linear Algebra Appl. 523 (2017) 102-108. [20] D. S. Raki´c, N. C. Dinci´c, D. S. Djordjevi´c, Group, Moore-Penrose, core and dual core inverse in rings with involution, Linear Algebra Appl. 463 (2014) 115 -- 133. [21] D. S. Raki´c, N. C. Dinci´c, D. S. Djordjevi´c, Core inverse and core partial order of Hilbert space operators, Appl. Math. Comp. 244 (2014) 283-302. [22] V. Rakocevi´c, On the continuity of the Moore-Penrose inverse in Banach algebras, Facta Univ. Ser. Math. Inform. 6 (1991) 133-138. [23] V. Rakocevi´c, Continuity of the Drazin inverse, J. Operator Theory 41 (1999) 55-68. [24] S. Roch, B. Silbermann, Continuity of generalized inverses in Banach algebras, Studia Math. 136 (1999) 197 -- 266. [25] G. W. Stewart, On the continuity of the generalized inverse, SIAM J. Appl. Math. 17 (1969) 33-45. [26] S. Z. Xu, J.L. Chen, X. X. Zhang, New characterizations for core and dual core inverses in rings with involution, Front. Math. China 12 (2017) 231 -- 246. 23 Julio Ben´ıtez E-mail address: [email protected] Enrico Boasso E-mail address: enrico [email protected] Sanzhang Xu [email protected] 24
1803.11075	4	1803	2019-01-02T09:04:28	Fixed-points in the cone of traces on a C*-algebra	[ "math.OA" ]	Nicolas Monod introduced the class of groups with the fixed-point property for cones, characterized by always admitting a non-zero fixed point whenever acting (suitably) on proper weakly complete cones. He proved that his class of groups contains the class of groups with subexponential growth and is contained in the class of supramenable groups. In this paper we investigate what Monod's results say about the existence of invariant traces on (typically non-unital) C*-algebras equipped with an action of a group with the fixed-point property for cones. As an application of these results we provide results on the existence (and non-existence) of traces on the (non-uniform) Roe algebra.	math.OA	math	Fixed-points in the cone of traces on a C ∗-algebra Mikael Rørdam∗ Dedicated to the memory of John Roe Abstract Nicolas Monod introduced in [16] the class of groups with the fixed-point property for cones, characterized by always admitting a non-zero fixed point whenever acting (suitably) on proper weakly complete cones. He proved that his class of groups contains the class of groups with subexponential growth and is contained in the class of supramenable groups. In this paper we investigate what Monod's results say about the existence of invariant traces on (typically non-unital) C ∗-algebras equipped with an action of a group with the fixed-point property for cones. As an application of these results we provide results on the existence (and non-existence) of traces on the (non-uniform) Roe algebra. 1 Introduction Whenever a discrete amenable group acts on a unital C ∗-algebra with at least one tracial state, then the C ∗-algebra admits an invariant tracial state, and the crossed product C ∗-al- gebra admits a tracial state. This, moreover, characterizes amenable groups. The purpose of this paper is to find statements, similar to this well-known fact, about the existence of invariant traces on non-unital C ∗-algebras using the results of the recent paper by Monod, [16], in which the class of groups with the fixed-point property for cones is introduced and developed. In [16], a group is said to have the fixed-point property for cones if whenever it acts continuously on a proper weakly complete cone (embedded into a locally convex topological vector space), such that the action is of cobounded type and locally bounded, then there is a non-zero fixed point in the cone. Being of cobounded type is an analog of an action on a locally compact Hausdorff space being co-compact, see [16, Definition 2]. The action is locally bounded if there is a non-zero bounded orbit, see [16, Definition 1]. Even the group of integers, Z, can fail to leave invariant any non-zero trace when acting on a non-unital C ∗-algebra. For example, the stabilization of the Cuntz algebra O2 is the crossed product of the stabilized CAR-algebra A with an action of Z that scales the traces ∗Supported by the Danish Council for Independent Research, Natural Sciences, and the Danish National Research Foundation (DNRF) through the Centre for Symmetry and Deformation at the University of Copenhagen. 1 on A by a factor of 2. In particular, there are no non-zero invariant traces on A. The cone of (densely defined lower semi-continuous) traces on A is isomorphic to the cone [0, ∞), and the induced action of Z on this cone is multiplication by 2, which of course fails to be locally bounded. The action of any group G on a locally compact Hausdorff space X induces a locally bounded action on the cone of Radon measures on X (which again is the same as the cone of densely defined lower semi-continuous traces on C0(X)). However, any infinite group can act on the locally compact non-compact Cantor set K∗ in a non co-compact way so that there are no non-zero invariant Radon measures, cf. [15, Section 4]. Such an action of G on the Radon measures on K∗ fails to be of cobounded type. The two examples above explain why one must impose conditions on the action, such as being of cobounded type and being locally bounded, to get meaningful results on when invariant traces exist. Monod proves in [16] that the class of groups with the fixed-point property for cones contains the class of groups of subexponetial growth and is contained in the class of supra- menable groups introduced by Rosenblatt in [23]. It is not known if there are supramenable groups of exponential growth, so the three classes of groups could coincide, although the common belief seems to be that they all are different. Monod proved a number of per- manence properties for his class of groups, prominently including that it is closed under central extensions, see [16, Theorem 8]. He also shows that the property of having the fixed-point property for cones can be recast in several ways, including the property that for each non-zero positive function f ∈ ℓ∞(G) there is a non-zero invariant positive linear functional (called an invariant integral in [16]) on the subspace ℓ∞(G, f ) of all bounded functions G-dominated by f . We begin our paper in Section 2 by recalling properties of possibly unbounded positive traces on (typically non-unital) C ∗-algebras, including when they are lower semi-continuous and when they are singular. By default, all traces in this paper are assumed to be positive. Interestingly, many of the traces predicted by Monod turn out to be singular. Traces on C ∗-algebras were treated systematically already by Dixmier in [7]. Elliott, Robert and Santiago consider in [8] the cone of lower semi-continuous traces as an invariant of the C ∗-algebra, they derive useful properties of this cone, and they make the point that such traces most conveniently are viewed as maps defined on the positive cone of the C ∗-alge- bra taking values in [0, ∞]. While this indeed is a convenient way to describe unbounded traces, and one we in part shall use here, the cone structure from the point of view of this paper is sometimes better portrayed when traces are viewed as linear functionals on a suitable domain: a hereditary symmetric algebraic ideal in the C ∗-algebra. By a theorem of G.K. Pedersen one can identify the cone of lower semi-continuous densely defined traces on a C ∗-algebra with the cone of traces defined on its Pedersen ideal. In the case where the primitive ideal space of the C ∗-algebra is compact, we show that there are non-zero lower semi-continuous densely defined traces if and only if the stabilization of the C ∗-algebra contains no full properly infinite projections, thus extending well-known results from both the unital and the simple case. A similar compactness condition appears in our reformulation of coboundedness of the action of the group on the cone of traces. 2 In Section 3 we explain when a C ∗-algebra equipped with an action of a group G with the fixed-point property for cones admits an invariant densely defined lower semi- continuous trace. Given that the cone of densely defined lower semi-continuous traces is always proper and weakly complete, all we have to do is to explain when the action of the group on this cone is of cobounded type, respectively, when it is locally bounded. The former can be translated into a compactness statement, as mentioned above, while the latter just means that there exists a non-zero trace which is bounded on all G-orbits. In Section 4 we examine the situation where the group G acts on ℓ∞(G) with the aim of describing for which positive f ∈ ℓ∞(G) the invariant integrals on ℓ∞(G) normalizing f are lower semi-continuous, respectively, singular. As it turns out, frequently they must be singular. We use this to give an example of a G-invariant densely defined trace on a C ∗-algebra that does not extend to a trace on the crossed product. Finally, in Section 5, we consider the particular example of invariant traces on ℓ∞(G, K) and traces on the Roe algebra ℓ∞(G, K)⋊G. We show that ℓ∞(G, K) only has the "obvious" densely defined lower semi-continuous traces, and hence that it never has non-zero G- invariant ones, when G is infinite. In many cases, however, there are invariant lower semi-continuous traces with smaller domains, such as domains defined by projections in ℓ∞(G, K). Specifically we show that any projection in ℓ∞(G, K) whose dimension (as a function on G) grows subexponentially is normalized by an invaraint lower semi-continuous trace if G has the fixed-point property for cones. In general, for any non-locally finite group G, there are (necessarily exponentially growing) projections in ℓ∞(G, K) not normalized by any invariant trace, and which are properly infinite in the Roe algebra, while the Roe algebra of a locally finite group is always stably finite. I thank Nicolas Monod, Nigel Higson, Guoliang Yu, and Eduardo Scarparo for useful discussions related to this paper. I also thank the referee for suggesting improvements of the exposition and for pointing out a couple mistakes in an earlier version of this paper. 2 Hereditary ideals and cones of traces By a hereditary ideal I in a C ∗-algebra A we shall mean an algebraic two-sided self-adjoint if 0 ≤ a ≤ b, b ∈ I and a ∈ A, then a ∈ I. If ideal satisfying the hereditary property: x∗x ∈ I whenever xx∗ ∈ I, for all x ∈ A, then we say that I is symmetric. All closed two-sided ideals are hereditary and symmetric. If a group G acts on the C ∗-algebra A, then we refer to an ideal I as being G-invariant (or just invariant) if it is invariant under the group action. If M is a subset of A, then let IA(M), respectively, IA(M), denote the smallest hereditary ideal in A, respectively, the smallest closed two-ideal in A, which contains M; and let similary I G A(M) denote the smallest G-invariant hereditary ideal, respectively, the smallest G-invariant closed two-ideal in A which contains M. A (M) and I G Example 2.1. The Pedersen ideal, Ped(A), of a C ∗-algebra A is the (unique) smallest dense ideal in A, see [18, Section 5.6]. It is a hereditary symmetric ideal, and, even better: 3 for each x ∈ Ped(A), the hereditary sub-C ∗-algebra, xAx∗, of A is contained in Ped(A). In particular, the Pedersen ideal of A is closed under continuous function calculus on its normal elements, as long as the continuous function vanishes at 0. If x ∈ A is such that x∗x ∈ Ped(A), then x ∈ Ped(A), which shows that the Pedersen ideal is also symmetric. Example 2.2. Not all self-adjoint two-sided ideals in a C ∗-algebra are hereditary. Consider for example the commutative C ∗-algebra A = C([−1, 1]), the element f ∈ C([−1, 1]) given by f (t) = t, for t ∈ [−1, 1], and the (two-sided) self-adjoint ideal I = Af in A. The function g(t) = max{− 1 2 t, t}, t ∈ [−1, 1], then satisfies 0 ≤ g ≤ f , but g /∈ I. Hereditary ideals need not be spanned (or even generated) by their positive elements. Indeed, take again A = C([−1, 1]) and let I = A ι, where ι(t) = t, for t ∈ [−1, 1]. If f ∈ I ∩ A+, then f = g · ι, for some g ∈ P, where P is the set of functions g ∈ A such that g(t) ≤ 0, for t ∈ [−1, 0], and g(t) ≥ 0, for t ∈ [0, 1]. Since g(0) = 0, for all g ∈ P, it is not possible to write ι ∈ I as a linear combination of functions in I ∩ A+. To see that I is hereditary, suppose that 0 ≤ h ≤ f , where f ∈ I ∩ A+ and h ∈ A. Then f = g · ι, for some g ∈ P. Hence h(t)/t ≤ g(t), for all t 6= 0, from which we see that h ∈ I. Let us also note that (algebraic) two-sided ideals need not be self-adjoint. Consider the commutative C ∗-algebra A = C(D), where D is the closed unit disk in the complex plane, the function f (z) = z, z ∈ D, and the (two-sided) ideal I = Af in A. Then f belongs to I, but f ∗ = ¯f does not. when it already is unital. Lemma 2.3. Let A be a C ∗-algebra, and let C be a subcone of A+ satisfying: For a C ∗-algebra A let eA denote the unitization of A, when it is non-unital, or A itself (i) If a ∈ C and x ∈ eA, then x∗ax ∈ C, (ii) C is hereditary: if 0 ≤ a ≤ b, a ∈ A and b ∈ C, then a ∈ C. Let I be the linear span of C. Then I is a hereditary ideal in A and I ∩ A+ = C. If C is symmetric (in the sense that x∗x ∈ C implies xx∗ ∈ C, for all x ∈ A), then so is I. Conversely, if I is a hereditary ideal in A and if C = I ∩ A+, then C is a subcone of A+ satisfying (i) and (ii) above. The span, I0, of C is a hereditary subideal of I; and I0 = I if and only if I is generated as a hereditary ideal by its positive elements. Proof. It follows from (i) that if x ∈ A and a ∈ C, then xa + ax∗ = (x + 1)a(x + 1)∗ − xax∗ − a, i(xa − ax∗) = (x − i)a(x − i)∗ − xax∗ − a, belong to I, whence xa and ax∗ belong to I. This shows that I is an ideal in A. Clearly, I is self-adjoint. A subcone C of A+ satisfies span(C)∩A+ = C if and only if whenever a, b ∈ C are such that a ≤ b, then b − a ∈ C. Hereditary cones clearly have this property, so I ∩ A+ = C. This also shows that I is hereditary (because C is hereditary). 4 It is clear that C = I ∩ A+ has the stated properties if I is a hereditary ideal of A; and I0 is a hereditary ideal of A by the first part of the lemma. It is contained in I and contains by definition all positive elements of I. That proves the last claim of the lemma. Corollary 2.4. Let A be a C ∗-algebra and let M be a non-empty subset of A+. Let C be the set of all elements a ∈ A+ for which there exist n ≥ 1, e1, e2, . . . , en ∈ M, and x1, x2, . . . , xn ∈ eA such that a ≤Pn j=1 x∗ j ejxj. Then: (i) C is a subcone of A+ satisfying (i) and (ii) of Lemma 2.3; (ii) IA(M) = span(C); (iii) IA(M) ∩ A+ = C. Proof. It is clear that (i) holds, so Lemma 2.3 implies that I := span(C) is a hereditary ideal in A satisfying I ∩ A+ = C. As M ⊆ C ⊆ I we conclude that IA(M) ⊆ I. Conversely, C ⊆ IA(M), so I ⊆ IA(M). Corollary 2.5. Let A be a C ∗-algebra, let α be an action of a group G on A, and let M be a non-empty subset of A+. Then (i) I G A (M) = IA(G.M), where G.M = {αt(e) : t ∈ G, e ∈ M}. (ii) An element a ∈ A+ belongs to I G A (M) if and only if there exist n ≥ 1, t1, t2, . . . , tn ∈ G, y1, y2, . . . , yn ∈ eA, and e1, e2, . . . en ∈ M such that a ≤Pn A (M) we see that IA(G.M) ⊆ I G Proof. (i). As G.M ⊆ I G G-invariant, it contains I G A (M). Part (ii) follows from (i) and from Corollary 2.4. j=1 y∗ j αtj (ej)yj. A(M). Conversely, as IA(G.M) is For each non-empty subset M of A+ denote by JA(M) the smallest symmetric hereditary ideal in A containing M. If a group G acts on A, then denote by J G A (M) the smallest symmetric hereditary G-invariant ideal in A containing M. Since closed two-sided ideals in a C ∗-algebra always are symmetric, we have IA(M) ⊆ JA(M) ⊆ IA(M), I G A (M) ⊆ J G A (M) ⊆ I G A(M). (2.1) Lemma 2.6. Let A be a C ∗-algebra and let M be a non-empty subset of projections in A, then IA(M) = JA(M) = Ped(A0), where A0 = IA(M). If A is equipped with an action of a group G, then I G A (M) = Ped(A1), where A1 = I G A (M) = J G A(M). Proof. We have M ⊆ Ped(A0) ⊆ IA(M) ⊆ JA(M) ⊆ A0 (the first inclusion holds because each projection in a C ∗-algebra belong to its Pedersen ideal, and second inclusion holds because IA(M) is a dense ideal in A0). As Ped(A0) is a hereditary symmetric ideal, which contains M, cf. Example 2.1, JA(M) ⊆ Ped(A0), so Ped(A0) = IA(M) = JA(M). The second part of the statement follows from the first part applied to G.M (instead of M). 5 Lemma 2.7. Let A be a C ∗-algebra and let M be a non-empty set of positive elements in A. Then JA(M) is the linear span of its positive elements. If A is equipped with an action of a group G, then the same holds for J G A (M). Proof. Set C = J G ideal in A by Lemma 2.3. Since J G C, and hence for J0. Moreover, M ⊆ C ⊆ J0 ⊆ J G with these properties, J0 = JA(M). The first claim is proved in a similar manner. A (M) ∩ A+ and let J0 be the linear span of C. Then J0 is a hereditary A (M) is symmetric and G-invariant, the same holds for A (M) is the smallest ideal A (M). As J G Definition 2.8. Let A be a C ∗-algebra. Denote by T +(A) the cone of traces on the positive cone of A, i.e., the set of additive homogeneous maps τ : A+ → [0, ∞] satisfying τ (x∗x) = τ (xx∗), for all x ∈ A. For each hereditary symmetric ideal I in A, let T (I, A) denote the cone of linear traces on I, i.e., the set of positive linear maps τ : I → C satisfying τ (x∗x) = τ (xx∗), whenever x ∈ A is such that x∗x (and hence xx∗) belong to I. We refer to I as the domain of τ . If the domain of τ is a dense ideal of A (in which case it will contain the Pedersen ideal of A), then τ is said to be a densely defined trace on A. The cone of traces, here denoted by T +(A), is in [8] denoted by T (A). Note that all traces by default are assumed to be positive. The following easy fact will be used repeatedly: Lemma 2.9. Let τ be a trace on a C ∗-algebra A. Then τ (x∗ax) ≤ kxk2τ (a), for all a ∈ A+ (in the domain of τ ) and all x in the unitization of A. The "ε-cut-down" (a − ε)+ of a ∈ A+ appearing in part (i) of the lemma below is defined by applying the continuous positive function t 7→ max{t − ε, 0} to a. A trace in T +(A), or a linear trace on A defined on a given domain, is said to be lower semi-continuous if one of the equivalent conditions in the following lemma holds. Lemma 2.10. Let A be a C ∗-algebra. The following conditions are equivalent for each trace τ in T +(A) (or for each linear trace on A): (i) τ (a) = supε>0 τ ((a − ε)+) for each a ∈ A+ (in the domain of τ ), (ii) whenever {an}∞ n=1 is an increasing sequence in A+ (in the domain of τ ) converging in norm to a ∈ A+ (in the domain of τ ), then τ (a) = limn→∞ τ (an), (iii) whenever {an}∞ n=1 is a sequence in A+ (in the domain of τ ) converging in norm to a ∈ A+ (in the domain of τ ), then τ (a) ≤ lim inf n→∞ τ (an). Proof. (ii) ⇒ (i). To verify (i) one needs only show that τ (a) = limn→∞ τ ((a − εn)+) for all sequences {εn} decreasing to 0; but this is just a special case of (ii). (iii) ⇒ (ii). If {an}∞ a, then τ (an) ≤ τ (a), for all n, by positivity of τ . τ (a) = limn→∞ τ (an). n=1 is an increasing sequence of positive elements converging to If (iii) holds, then this entails that (i) ⇒ (iii). Let ε > 0 be given. Choose n0 ≥ 1 such that kan − ak < ε, for all n ≥ n0. nandn, for some contractions dn It then follows from [14, Lemma 2.2] that (a − ε)+ = d∗ 6 in A, for all n ≥ n0. Hence τ ((a − ε)+) ≤ τ (an), by Lemma 2.9. This shows that lim inf n→∞ τ (an) ≥ τ ((a − ε)+). It follows that lim inf n→∞ τ (an) ≥ supε>0 τ ((a − ε)+), which proves that (i) implies (iii). Theorem 2.11 (G.K. Pedersen, [19, Corollary 3.2]). The restriction of any densely defined trace on a C ∗-algebra A to the Pedersen ideal of A is automatically lower semi-continuous. Definition 2.12. Denote by Tlsc(A) the cone of linear traces on A whose domain is the Pedersen ideal of A. In other words, Tlsc(A) = T (Ped(A), A). We can identify Tlsc(A) with the set of densely defined lower semi-continuous traces on A as follows: Each trace in Tlsc(A) is clearly densely defined, and it is lower semi-continuous by Pedersen's theorem. Conversely, if τ is a lower semi-continuous densely defined trace, then its restriction τ0 to the Pedersen ideal of A belongs to Tlsc(A), and τ is uniquely determined on its domain by τ0 by Lemma 2.10 (i), because (a − ε)+ ∈ Ped(A) for all positive a ∈ A and all ε > 0. A trace in Tlsc(A) can usually be extended to a lower semi-continuous trace on a larger domain than the Pedersen ideal; and such an extension is unique, see Proposition 2.13 below and the subsequent discussion. It follows from Theorem 2.11 and Lemma 2.6 that each linear trace on a C ∗-algebra A A (M) (when A has a G-action) is lower semi-continuous whenever with domain JA(M) or J G M is a subset of projections in A. Observe that Tlsc(B(H)) = {0}, where B(H) is the bounded operators on a separable in- finite dimensional Hilbert space H, while Tlsc(K(H)) and the cone of lower semi-continuous traces in T +(B(H)) both are equal to the one-dimensional cone spanned by the canonical trace on B(H), in the latter case viewed as a function B(H)+ → [0, ∞]. The Dixmier trace is an example of a singular trace on K(H). It belongs to T +(B(H)) and to T (I, K(H)), where I ⊂ K(H) is its domain, and it is zero on the finite rank operators. Consider a general (not necessarily densely defined) linear trace τ on A with domain I. The closure, I, of I is a closed two-sided ideal in A, and hence, in particular, a C ∗-algebra; and τ is of course densely defined relatively to this C ∗-algebra. We have the following inclusions: Ped(I) ⊆ I ⊆ I. The restriction of τ to Ped(I) is lower semi-continuous by Theorem 2.11. If this restriction is zero, then τ is said to be singular. Each trace τ on A with domain I can in a unique way be written as the sum τ = τ1 + τ2 of a lower semi-continuous trace τ1 and a singular trace τ2, both with domain I. The lower semi-continuous part is obtained by restricting τ to the Pedersen ideal (which is lower semi-continuous) and then extending to a lower semi-continuous trace τ1 defined on I as described in (2.3) and the subsequent comments. One can smoothly and uniquely pass from a trace in T +(A) to a linear trace defined on its natural (maximal) domain: Proposition 2.13. Let A be a C ∗-algebra, and let τ ′ ∈ T +(A). Let C be the set of positive elements a ∈ A with τ ′(a) < ∞, and let I be the linear span of C. Then I is a hereditary 7 symmetric ideal in A, I ∩ A+ = C, and there is a unique linear trace τ with domain I that agrees with τ ′ on C. We can recover τ ′ from τ via the formula τ ′(a) =(τ (a), a ∈ C, ∞, a ∈ A+ \ C. (2.2) If τ ′ is lower semi-continuous, then so is τ . Proof. Observe that the set C is a symmetric subcone of A+ satisfying conditions (i) and (ii) of Lemma 2.3 (use Lemma 2.9 to see that Lemma 2.3 (i) holds). It therefore follows from Lemma 2.3 that I is a symmetric hereditary ideal in A and that I + = I ∩ A+ = C. By additivity and homogeneity of τ ′, its restriction to C extends (uniquely) to a linear map τ : I → C. If a ∈ I is positive, then a ∈ C, so τ (a) = τ ′(a) ≥ 0, which shows that τ is positive. Let x ∈ A be such that x∗x ∈ I. Then x∗x and xx∗ are positive elements in I, so both belong to C, whence τ (x∗x) = τ ′(x∗x) = τ ′(xx∗) = τ (xx∗), so τ is a trace on I. It is clear that (2.2) holds. If τ ′ is lower semi-continuous, then so is its restriction to C, which shows that τ is lower semi-continuous. Whenever we talk about a trace on a C ∗-algebra A, we shall mean a trace defined on the cone of positive elements of that C ∗-algebra taking values in [0, ∞], i.e., a trace in T +(A), and, at the same time, a linear trace on the domain defined in the proposition above, or some other domain to be specified in the context. As a converse to the proposition above, consider a linear trace τ defined on a hereditary symmetric ideal I in A. Then τ ′ given by (2.2) above, with C = I ∩A+, belongs to T +(A), and it agrees with τ on C. If we apply Proposition 2.13 to τ ′, then we get back a new linear trace τ0 defined on some symmetric hereditary ideal J0 of A, which contains the sub-ideal I0 of I defined in Lemma 2.3 (but perhaps not I itself); and τ and τ0 agree on I0. However, this extension of τ to a trace τ ′ defined on the positive cone of A is not unique, and τ ′ need not be lower semi-continuous, even when τ is lower semi-continuous. If τ is lower semi-continuous, then the map τ ′ : A+ → [0, ∞] defined by τ ′(a) = sup{τ (a0) : 0 ≤ a0 ≤ a, a0 ∈ I}, a ∈ A+, (2.3) is a lower semi-continuous trace in T +(A), and it is the unique such that extends τ . In the sequel, when considering a lower semi-continuous trace, we may at wish view it either as a linear trace defined on its domain, or as a trace defined on the positive cone, via (2.3). There is a canonical way of extending a lower semi-continuous trace τ defined on some hereditary symmetric ideal I of A to its maximal domain: first extend τ to a lower semi- continuous trace τ ′ : A+ → [0, ∞] as in (2.3) above; and then take the linearization τ of τ ′ defined in Proposition 2.13. We quote the following well-known result for later reference, see, eg., [20, Lemma 5.3] for a proof. 8 Lemma 2.14. Let A be a C ∗-algebra equipped with an action of a group G, and let τ be a G- invariant lower semi-continuous trace on A. It follows that τ ◦E is a lower semi-continuous trace on the (reduced) crossed product A ⋊ G that extends τ , where E : A ⋊ G → A is the standard conditional expectation. If τ is densely defined, then so is τ ◦ E. One should here view τ and τ ◦ E as traces defined on the positive cone of A, respectively, A ⋊ G. For the claim that τ ◦ E is densely defined when τ is, use that τ ◦ E is finite on the positive cone of Ped(A), and the hereditary ideal in A ⋊ G generated by Ped(A) is dense in A ⋊ G. It is a curious fact that an invariant densely defined trace τ on A need not in general extend to a trace on the crossed product A ⋊ G; in particular, τ ◦ E need not be a trace if τ is not lower semi-continuous. See Example 4.4. We end this section by considering when a C ∗-algebra admits a non-zero densely defined trace. Blackadar and Cuntz proved in [2] that a stable simple C ∗-algebra either contains a properly infinite projection or admits a non-zero dimension function (defined on its In the latter case, assuming moreover that the C ∗-algebra is exact, it Pedersen ideal). admits a non-zero densely defined trace. (This step follows from the work of Blackadar- Handelman [3], Haagerup, [10], and Kirchberg, [12], as explained in the last part of the proof of the theorem below.) Also, it is well-known that a unital exact C ∗-algebra admits a tracial state if and only if no matrix algebra over it is properly infinite. A common feature of simple and of unital C ∗-algebras is that their primitive ideal spaces are compact. Recall that the primitive ideal space, Prim(A), of a C ∗-algebra A is compact if and only if for all upward directed families {Iα} of closed two-sided ideals in A whose union is dense in A there is α such that A = Iα. Theorem 2.15. Let A be an exact C ∗-algebra whose primitive ideal space is compact. Then A admits a non-zero densely defined lower semi-continuous trace, i.e., Tlsc(A) 6= {0}, if and only if the stabilization of A does not contain a full properly infinite projection. Proof. The proof is most naturally phrased via dimension functions (as defined by Cuntz in [6]) and the Cuntz semigroup, see, eg., [5]. Observe first that the cone of densely defined lower semi-continuous traces and the primitive ideal space are not changed by stabilizing the C ∗-algebra, so may assume that A is stable. The class of (closed two-sided) ideals of the form I A((e − ε)+), where e ∈ A+ and ε > 0, is upwards directed and its union is dense in A. Hence A = I A((e − ε0)+), for some e ∈ A+ and some ε0 > 0 by compactness of the primitive ideal space of A. Since (e − ε)+ belongs to the Pedersen ideal and since IA((e − ε)+) is a dense ideal in A, for all 0 < ε ≤ ε0, it follows that IA((e − ε)+) = Ped(A), for this e ∈ A+ and for all 0 < ε ≤ ε0. Let uε = h(e − ε)+i be the corresponding element the Cuntz semigroup Cu(A) of A. j ejxj(cid:11) ≤ Pn j=1heji in Cu(A), for all positive ej and all xj in eA, that for each positive a in Ped(A) It follows from Corollary 2.4, and the fact that (cid:10)Pn there exists k ≥ 1 such that hai ≤ kuε. In other words, uε is an order unit for the sub- semigroup Cu0(A), consisting of all classes hai, where a is a positive element in Ped(A). In particular, uε0 ≤ uε ≤ kuε0, for some integer k ≥ 1 (that depends on ε). Fix 0 < ε ≤ ε0. j=1 x∗ 9 Consider first the case that nuε0 is properly infinite, for some integer n ≥ 1. Upon replacing e by an n-fold direct sum of e with itself (which is possible since A is assumed to be stable), we may assume that uε0 itself is properly infinite, i.e., that kuε0 ≤ uε0, for all integers k ≥ 1. By the discussion in the previous paragraph, we can then conclude that uε is properly infinite and that x ≤ uε, for all x ∈ Cu0(A) and for all 0 < ε ≤ ε0. We can now follow the argument of [17, Proposition 2.7], which uses the notion of scaling elements introduced by Blackadar and Cuntz, [2], to construct a full properly infinite projection p ∈ A: Fix 0 ≤ ε < ε0. Then (e − ε)+ is properly infinite, so by [13, Proposition 3.3] there exist positive elements b1, b2 in the hereditary sub-C ∗-algebra of A generated by (e − ε0)+ such that b1 ⊥ b1 and (e − ε0)+ - bj, for j = 1, 2. In particular, b1, b2 ∈ Ped(A). As explained in [17, Remark 2.5] there exists x ∈ A such that x∗x(e − ε0)+ = (e − ε0)+ and xx∗ belongs to the hereditary sub-C ∗-algebra of A generated by b1. This shows that x is a scaling element (cf. [17, Remark 2.4]) satisfying x∗xb2 = b2 and xx∗b2 = 0. By [2], see also [17, Remark 2.4], we get a projection p ∈ A satisfying b2p = b2. As uε0 ≤ hb2i ≤ hpi ≤ uε0, we conclude that hpi is a properly infinite order unit of Cu0(A), whence p is a full properly infinite projection in A. Suppose now that there is no integer n ≥ 1 such that nuε0 is properly infinite. We proceed to show that Tlsc(A) 6= {0} in this case. As shown above, nuε is not properly infinite, for any n ≥ 1 and for any 0 < ε ≤ ε0. Fix 0 < ε1 < ε0, and observe that nuε1 ≤ muε1 implies n ≤ m, for all integers n, m ≥ 0 (since no multiple of uε1 is properly infinite). The map f0 : N0uε1 → R+, given by f0(nuε1) = n, for n ≥ 0, is therefore a positive additive map on the sub-semigroup N0uε1 of Cu0(A) (where N0uε1 is equipped with the relative order arising from Cu0(A)). By [4, Corollary 2.7] we can extend f0 to a positive additive map (state) f : Cu0(A) → R+ (since uε1 is an order unit for Cu0(A)). Let d : Ped(A)+ → R+ be the associated dimension function given by d(a) = f (hai), and let ¯d : Cu0(A) → R+ be the corresponding lower semi-continuous dimension function given by ¯d(a) = limε>0 d((a − ε)+), for a ∈ Ped(A)+, cf. [22, Proposition 4.1]. Then d((e − ε0)+) ≤ ¯d((e − ε1)+) ≤ d((e − ε1)+), and d(e−ε0)+) > 0 because d is non-zero and uε0 = h(e−ε0)+i is an order unit for Cu0(A). This shows that ¯d is non-zero. It follows from Blackadar -- Handelman, [3, Theorem II,2,2], that the lower semi-contin- uous dimension function ¯d (called rank function in [3]) lifts to a lower semi-continuous i.e., ¯d = dτ , where dτ (a) = 2-quasitrace τ defined on the "pre-C ∗-algebra" Ped(A), limn→∞ τ (a1/n), for all positive elements a ∈ Ped(A). Finally, by Kirchberg's extension, [12], to the non-unital case of Haagerup's theorem, [10], that any 2-quasitrace on an exact C ∗-algebra is a trace, τ is a non-zero lower semi-continuous densely defined trace on A. It remains unresolved when a C ∗-algebra with non-compact primitive ideal space admits a non-zero densely defined lower semi-continuous trace. Clearly, Tlsc(A) is non-zero for all commutative C ∗-algebras A, while the primitive ideal space of a commutative C ∗-alge- bra is compact only when it is unital. On the other hand, absence of full properly infinite projections is not sufficient to guarantee existence of non-zero lower semi-continuous traces. 10 Take, for example, the suspension (or the cone over) any purely infinite C ∗-algebra, cf. [13, Proposition 5.1]. In [15, Section 4] it was shown that any infinite group G admits a (free) action on the locally compact non-compact Cantor set K∗ with no non-zero invariant Radon measures. Accordingly, C0(K∗) ⋊ G has no non-zero densely defined lower semi-continuous trace, although C0(K∗) ⋊ G admits an approximate unit consisting of projections, and, if G is supramenable, eg., if G = Z, then no projection in the (stabilization of) C0(K∗) ⋊ G is properly infinite. The latter example is covered by the proposition below. When p and q are projections in a C ∗-algebra A and n ≥ 1 is an integer, then denote by p ⊗ 1n the n-fold direct sum of p with itself, and write p ≺≺ q if p ⊗ 1n - q, for all n ≥ 1. Proposition 2.16. Let A be a C ∗-algebra admitting an approximate unit consisting of projections. Suppose that for each projection p ∈ A there exists a projection q in A with p ≺≺ q. Then Tlsc(A) = {0}. Proof. Suppose that τ ∈ Tlsc(A), let p be a projection in A and let q ∈ A be another projection such that p ≺≺ q. Since p and q belong to the Pedersen ideal of A, and hence to the domain of τ , we find that τ (q) < ∞, which entails that τ (p) = 0. As A has an approximate unit consisting of projection, this implies that τ = 0. Here is an elementary example of an exact stably finite1 C ∗-algebra satisfying the condi- tions of Proposition 2.16, and which accordingly admits no non-zero lower semi-continuous densely defined trace: Let A be the inductive limit of the sequence A1 → A2 → A3 → · · · , where A1 = K, the C ∗-algebra of compact operators on a separable Hilbert space, where An+1 = eAn ⊗ K, for n ≥ 1, and where the inclusion An → An+1 is given by a 7→ a ⊗ e ∈ An ⊗ K ⊂ An+1, for some fixed 1-dimensional projection e ∈ K. 3 Invariant unbounded traces on C ∗-algebras We shall here use Monod's characterization of groups with the fixed-point property for cones to say something about when a (typically non-unital) C ∗-algebra A with an action of a group G admits an invariant trace. We are mostly interested in the existence of a (non-zero) invariant densely defined lower semi-continuous trace, i.e., an invariant non- zero trace in the cone Tlsc(A) defined in Section 2. But we shall also address the existence of more general traces (including singular traces and not densely defined traces). Recall from Definition 2.8 that T (I, A) is the cone of positive traces on A with domain I, whenever I is a hereditary symmetric ideal in A. The cone T (I, A) is embedded in the complex vector space L(I) of all linear functionals on I equipped with the locally convex weak topology induced by I. The dual space L(I)∗ of L(I) is naturally isomorphic to I, cf. [24, 3.14] (and as remarked in [16]), i.e., L(I)∗ = {ϕa : a ∈ I}, where ϕa denotes the functional ϕa(ρ) = ρ(a), for ρ ∈ L(I) and a ∈ I. The dual space L(I)∗ is equipped with 1A C ∗-algebra A is said to be stably finite if its stabilization A ⊗ K contains no infinite projections. This definition is meaningful when the C ∗-algebra has an approximate unit consisting of projections. 11 the preordering given by T +(I, A), whereby an element ϕ ∈ L(I)∗ is positive if ϕ(τ ) ≥ 0, for all τ ∈ T (I, A). Observe that ϕ ≥ 0 and −ϕ ≥ 0 if and only if ϕ(τ ) = 0, for all τ ∈ T (I, A). The map a 7→ ϕa is a positive isomorphism, but not necessarily an order embedding, since ϕa ≥ 0 does not necessarily imply that a ≥ 0. Monod considers real vector spaces in his paper [16], while our vector spaces are complex by the nature of C ∗-algebras. To translate some properties from Monod's paper to our language we shall occasionally consider the real vector space of all self-adjoint functionals ϕ in L(I)∗, and we note that ϕa is self-adjoint if and only if a ∈ I is self-adjoint. The cone T (I, A) is said to be proper if T (I, A) ∩ −T (I, A) = {0}, or, equivalently, if 0 is the only trace in T (I, A) that vanishes on I ∩ A+. This will hold if I is the span of its positive elements. Most ideals considered in this paper have this property, including the Pedersen ideal Ped(A), or any of the ideals IA(M), I G A (M), when M is any non-empty subset of A+, cf. Example 2.1, Corollary 2.4, Corollary 2.5 and Lemma 2.7. Recall also that Tlsc(A) = T (Ped(A), A). We allow for the possibility that the cones A (M), JA(M) or J G T (I, A) and Tlsc(A) are trivial, that is, equal to {0}, unless otherwise stated. Proposition 3.1. For each C ∗-algebra A and for each hereditary symmetric ideal I in A, the cone T (I, A) is weakly complete. In particular, Tlsc(A) is weakly complete. Proof. We must show that each weak Cauchy net in T (I, A) is weakly convergent, i.e., if (τi)i is a net in T (I, A) such that (ϕ(τi))i is Cauchy in C, for all ϕ ∈ L(I)∗, then the net converges weakly in T (I, A). Since ϕa(τi) = τi(a), for all a ∈ I, being weakly Cauchy implies that (τi(a))i is Cauchy and hence convergent in C, for all a ∈ I. Set τ (a) = limi τi(a), for all a ∈ I. It is easy to check that τ : I → C is in fact a trace, so it belongs to T (I, A), and since ϕa(τi) → ϕa(τ ), for all a ∈ I, τ is the weak limit of the net (τi)i, as desired. If I is a G-invariant hereditary Consider an action α of a (discrete) group G on A. symmetric ideal in A, then G induces an action of the cone T (I, A) by t.τ = τ ◦ α−1 , for t ∈ G and τ ∈ T (I, A). It is clear that this action of G on T (I, A) is continuous. Each automorphism of A leaves the Pedersen ideal invariant, so each group action on A induces an action on the cone Tlsc(A). t The action of G on T (I, A) is in [16] said to be of cobounded type if there exists a positive functional ϕ in L(I)∗ which G-dominates2 any other self-adjoint functional in L(I)∗. This condition is automatically satisfied when I = J G A (e), for some positive element e ∈ A+, cf. Corollary 3.4 below, but not always when I is the Pedersen ideal of A. However, in the latter case we can reformulate the coboundedness condition into more familiar statements for C ∗-algebras. A positive element e ∈ I is said to G-dominate a self-adjoint element a ∈ I if there are j=1 αtj (e); and e is said to tracially G-dominate a j=1 τ (αtj (e)), for all τ ∈ T (I, A). group elements t1, . . . , tn such that a ≤Pn if there are group elements t1, . . . , tn such that τ (a) ≤Pn 2If ϕ and ψ are self-adjoint functionals in L(I)∗, then ψ is G-dominated by ϕ if ψ ≤ Pn some n ≥ 1 and some t1, t2, . . . , tn ∈ G. j=1 tj.ϕ, for 12 The latter holds if and only if ϕa ≤ Pn j We can summarize these remarks as follows: j=1 t−1 .ϕe; in other words, if ϕe G-dominates ϕa. Lemma 3.2. Let A be a C ∗-algebra equipped with an action of a group G, and let I be an invariant hereditary symmetric ideal in A. The induced action of G on the cone T (I, A) is of cobounded type if and only if there is a positive element e ∈ I, which tracially G-dominates each self-adjoint element a ∈ I. Lemma 3.3. Let I be a G-invariant hereditary symmetric ideal in a C ∗-algebra A, and let e be a positive element in I. Then the functional ϕa is G-dominated by ϕe, for each self-adjoint element a ∈ J G A (e). Proof. Let C be the set of positive elements a ∈ I such that ϕa is G-dominated by e. We claim that C is a G-invariant symmetric cone in A+, which satisfies (i) and (ii) of Lemma 2.3. Since e clearly belongs to C, it will then follow from Lemma 2.3 that J G A (e) ∩ A+ ⊆ C, and this will prove the lemma. The set of positive ϕ ∈ L(I)∗ that are G-dominated by ϕe is a G-invariant hereditary cone in the positive cone of L(I)∗. As the map a 7→ ϕa is linear, order preserving and satisfies ϕαt(a) = t.ϕa, for a ∈ I + and t ∈ G, we conclude that C is a hereditary G- invariant cone in A+. For each x ∈ A, for which x∗x (and hence xx∗) belong to I, we have ϕx∗x ≤ ϕxx∗ ≤ ϕx∗x, which implies that C is symmetric. It remains to show that x∗ax belongs to C when a belongs to C and x belongs to eA. To see this, recall from Lemma 2.9 that τ (x∗ax) ≤ kxk2τ (a), so ϕx∗ax ≤ kxk2ϕa, and the latter is G-dominated by ϕe since a ∈ C (and since C is a cone). The corollary below follows immediately from Lemma 3.3. Corollary 3.4. The action of a group G on the cone T (J G whenever A is a C ∗-algebra with an action of G and e is a positive element in A. A (e), A) is of cobounded type The action of a group on the Pedersen ideal of a C ∗-algebra is not always of cobounded type, as illustrated in the proposition below, that covers the case of commutative C ∗-al- gebras, and which paraphrases and expands a remark on page 71 in [16]. We remind the reader that the action of a group G on a locally compact Hausdorff space is co-compact if X = G.K, for some compact subset K of X. Proposition 3.5. Let X be a locally compact Hausdorff space equipped with a continuous action of a group G. Then the following conditions are equivalent: (i) The action of G on X is co-compact. (ii) X is compact in the (non-Hausdorff ) topology on X consisting of the G-invariant open subsets of X. (iii) The action of G on the cone of Radon measures on X equipped with the vague topology is of cobounded type. 13 (iv) The action of G on the cone Tlsc(C0(X)) is of cobounded type. Proof. (i) ⇒ (ii). Let K be a compact subset of X witnessing co-compactness of the action. Let {Ui}i∈I be a collection of invariant open sets that covers X. Select a finite subset F ⊆ I (ii) ⇒ (i). Let {Ui}i∈I be the collection of all open pre-compact subsets of X. Then such that {Ui}i∈F covers K. ThenSi∈F Ui = X, being a G-invariant set that contains K. X = Si∈I Ui, because X is locally compact. For each i ∈ I, set Vi = St∈G t.Ui. The X = Vi, for some i ∈ I. Hence X =St∈G t.K, when K is the (compact) closure of Ui. families {Ui}i∈I and {Vi}i∈I are both upwards directed (both are closed under forming finite unions). It follows by compactness of X in the topology of invariant open sets that (i) ⇒ (iv). The cone Tlsc(C0(X)) is embedded into the vector space L(Cc(X)) equipped with the weak topology from Cc(X); and the dual space, L(Cc(X))∗, is equal to Cc(X). By co-compactness of the action we can find sets U ⊆ K ⊆ X, such that K is compact, U is open, and G.U = X. Let f ∈ Cc(X) be such that 1K ≤ f ≤ 1. Then any real valued function g ∈ Cc(X) is G-dominated by f . Indeed, if F is a finite subset of G such that the (iv) ⇒ (i). Following the set-up of the proof above we can find a positive function f ∈ Cc(X) which G-dominates any other real valued function in Cc(X). Let K be the support of f . Let x ∈ X and choose a positive function g ∈ Cc(X) such that g(x) = 1. support of g is contained inSt∈F t.U, then g ≤ kgk∞Pt∈F t.f . Then g ≤Pt∈F t.f for some finite subset F of G. Thus t.x ∈ K, for some t ∈ K. This (iii) ⇔ (iv). By Riesz' theorem there is a one-to-one correspondance between Radon shows that X = G.K. measures on X and positive linear functionals (hence traces) on Cc(X) = Ped(C0(X)). We have previously, in Theorem 2.15, considered C ∗-algebras A whose primitive ideal space, Prim(A), is compact, which happens if whenever {Iα} is an upwards directed net α. Let us say that a C ∗-algebra A equipped with an action of a group G is G-compact if whenever {Iα} is an upwards directed net of G-invariant closed two-sided ideals in A of closed two-sided ideals in A such that Sα Iα is dense in A, then A = Iα, for some such thatSα Iα is dense in A, then A = Iα, for some α. In the commutative case this is equivalent to condition (ii) of Proposition 3.5. In the non-commutative case we have the following: Proposition 3.6. Let A be a C ∗-algebra with an action of a group G, and suppose that A is G-compact. Then the induced action of G on the cone Tlsc(A) is of cobounded type. A (e)} of G-invariant ideals in A, where e ∈ Ped(A)+, is upward Proof. The family {J G directed and its union is dense in A. Hence by the G-compactness assumption there is a positive element e in Ped(A) such that the hereditary ideal J G A (e) is dense in A. As e ∈ Ped(A), this entails that J G A (e) = Ped(A). We can therefore conclude from Lemma 3.3 that each self-adjoint ϕ ∈ L(Ped(A))∗ is G-dominated by ϕe. Examples of G-compact C ∗-algebras include any simple C ∗-algebra and, more generally, any C ∗-algebra with no non-trivial G-invariant closed two-sided ideals. Also any C ∗-al- gebra A, which contains a G-full projection, i.e., a projection p such that I G A (p) = A, is G-compact. 14 It is not hard to show that the conclusion of Proposition 3.6 still holds under the weaker assumption that A/I0 is G-compact, where I0 is the closed two-sided G-invariant ideal I0 = \τ ∈Tlsc(A) {x ∈ A : τ (x∗x) = 0}. Following Monod, [16], a subset M of T (I, A) is bounded if for each open neighborhood 0 ∈ U ⊆ L(I) there is r > 0 such that M ⊆ r U. Lemma 3.7. Let A be a C ∗-algebra and let I be a hereditary symmetric ideal in A. A subset M of T (I, A) is bounded if and only if {τ (a) : τ ∈ M} is bounded for all a ∈ I. In particular, M ⊆ Tlsc(A) is bounded if and only if {τ (a) : τ ∈ M} is bounded for all a ∈ Ped(A). Proof. For each self-adjoint element a ∈ I, the set Ua := {ϕ ∈ L(I) : ϕ(a) < 1} is an open neighborhood of 0. Hence, if M is bounded, then M ⊆ r Ua for some r > 0, which entails that τ (a) < r, for all τ ∈ M. Conversely, suppose that {τ (a) : τ ∈ M} is bounded, for all a ∈ I, and let U be an open neighborhood of 0. Then there are a1, . . . , an in I such that Ua1,...,an ⊆ U, where Ua1,...,an = {ϕ ∈ L(I) : ϕ(ai) < 1, for i = 1, 2, . . . , n}. Set ri = sup{τ (ai) : τ ∈ M} and set r = 1 + maxi ri. Then τ (ai) < r, for all i and for all τ ∈ M, whence M ⊆ r Ua1,a2,...,an ⊆ r U, as desired. A trace τ ∈ T (I, A) will be said to be locally bounded if it is bounded on the G-orbit of each a ∈ I. Clearly, any bounded trace is G-bounded (regardless of the properties of the action). By Monod, [16], the action of G on T (I, A) is locally bounded if T (I, A) contains a non-zero bounded orbit. Lemma 3.8. Let G be a group acting on a C ∗-algebra A, let I be a hereditary symmetric invariant ideal in A, and let e be a positive element in A. (i) The induced action of G on T (I, A) is locally bounded if and only if there is a non- zero locally bounded trace in T (I, A). (ii) A trace in T (J G A (e), A) is locally bounded if it is bounded on the orbit of e. Proof. (i) is an immediate reformulation of Lemma 3.7. To prove (ii), let τ ∈ T (J G A (e), A) be bounded on the orbit of e, and let C be the set of positive elements a ∈ J G A (e) such that τ is bounded on the orbit of a. Then C satisfies conditions (i) and (ii) of Lemma 2.3, cf. Lemma 2.9, e ∈ C, and C is G-invariant, so C = J G A (e) ∩ A+ by Lemma 2.3. Groups with the fixed-point property for cones allow invariant lower semi-continuous densely defined traces on a C ∗-algebra as follows: 15 Theorem 3.9. Let A be a C ∗-algebra with Tlsc(A) 6= {0}, and let G be a group with the fixed-point property for cones, which acts on A making A G-compact. Then there is a non-zero invariant, necessarily lower semi-continuous, trace in Tlsc(A) if and only if there is a non-zero locally bounded trace in Tlsc(A). If this is the case, then Tlsc(A ⋊ G) is non-zero, i.e., A ⋊ G admits a non-zero lower semi-continuous densely defined trace. Proof. The cone Tlsc(A) is weakly complete by Proposition 3.1, the action of G on Tlsc(A) is affine and continuous. By the assumption on G there is a non-zero invariant trace in Tlsc(A) if the action of G on Tlsc(A) is of cobounded type and locally bounded. The former holds by Proposition 3.6, since A is assumed to be G-compact, and the latter holds by Lemma 3.8 if there is a non-zero locally bounded trace. Conversely, any invariant trace in Tlsc(A) is trivially G-bounded. The last claim follows from Lemma 2.14. Consider the class of groups G for which Theorem 3.9 holds. This class contains the class of groups with the fixed-point propety for cones, and it is contained in the class of supramenable groups, cf. the proposition below, that paraphrases [11, Theorem 1.1]. We do not know if Theorem 3.9 characterizes groups with the fixed-point property for cones, or if it characterizes the class of supramenable groups, or some intermediate class of groups. Proposition 3.10. The following conditions are equivalent for each group G: (i) G is supramenable. (ii) Whenever G acts on a commutative C ∗-algebra A, such that A is G-compact, then there is a non-zero invariant trace in Tlsc(A). (iii) Whenever G acts on a commutative C ∗-algebra A, then for each projection p ∈ A there is an invariant lower semi-continuous trace τ ∈ T +(A) with τ (p) = 1. (iv) Whenever G acts on a commutative C ∗-algebra A, then for each projection p ∈ A there is a lower semi-continuous trace τ ∈ T +(A ⋊ G) with τ (p) = 1. Proof. (i) ⇒ (ii). Let X be the spectrum of A, so that A = C0(X). If A is G-compact, then the action of G on X is co-compact, cf. Proposition 3.5, so by [11, Theorem 1.1] there is a non-zero invariant Radon measure on X, since G is supramenable. Integrating with respect to this measure gives a non-zero invariant trace in Tlsc(A). (ii) ⇒ (iii). The C ∗-algebra B = I G A (p) is G-compact being generated by a projection, so (ii) implies that there is a non-zero invariant τ ∈ Tlsc(B). We must show that 0 < τ (p) < ∞. The latter inequality holds because p ∈ Ped(B). If τ (p) = 0, then τ (x) = 0 for all x ∈ I G B (p) = Ped(B), cf. Lemma 2.6, contradicting that τ 6= 0. Finally, we can extend τ to a lower semi-continuous trace in T +(A) by (2.3). B (p), and I G (iii) ⇒ (iv). This follows from Lemma 2.14. (iv) ⇒ (i). If G is non-supramenable, then, by [11, Theorem 1.1], it admits a minimal, free, purely infinite action on the locally compact non-compact Cantor set K∗, making the crossed product C0(K∗)⋊G purely infinite (and simple). In particular, there is no non-zero lower semi-continuous trace on C0(K∗) ⋊ G. 16 Remark 3.11. If τ is a trace on a C ∗-algebra B such that τ (p) = 1, for some projection p ∈ B, then the restriction of τ to the unital corner C ∗-algebra pBp is a tracial state. Conversely, each tracial state τ on pBp extends to a lower semi-continuous trace in T +(B) normalizing p. To see this, consider the family {Hα}α of all σ-unital hereditary sub-C ∗-algebras of IB(p) containing pBp, and observe that pBp is a full hereditary sub-C ∗-algebra of Hα, for each α. By Brown's theorem, the inclusion pBp → pBp ⊗ K, where K is the C ∗-algebra of compact operators on a separable Hilbert space, extends to an inclusion Hα → pBp ⊗ K, for each α. The tracical state τ on pBp extends (uniquely) to a lower semi-continuous trace on the positive cone of pBp ⊗ K, and thus restricts to a lower semi-continuous trace τα on the positive cone of Hα. The extension of τ to τα on Hα is unique. The familiy {Hα}α is upwards directed with union Sα Hα = IB(p). To see the first claim, recall that a C ∗-algebra is σ-unital precisely if it contains a strictly positive ele- ment. Let bα ∈ Hα be strictly positive. Then bα + bβ is a strictly positive element in (bα + bβ)B(bα + bβ), and the latter is therefore a σ-unital sub-C ∗-algebras of IB(p) con- taining Hα and Hβ. For the second claim, if a ∈ IB(p) is positive, then (a + p)B(a + p) is a σ-unital hereditary sub-C ∗-algebra of IB(p) containing pBp and a. There exists therefore a (unique) lower semi-continuous trace τ in T +(IB(p)) that extends each τα, and hence τ . Extend τ further to a lower semi-continuous trace in T +(B) using (2.3). Finally, if we wish, we can linearize τ using Proposition 2.13. If we depart from lower semi-continuous densely defined traces and consider possibly singu- lar traces, then we do obtain a C ∗-algebraic characterization of groups with the fixed-point property for cones. The theorem below is a non-commutative analog of the equivalence between (1) and (4) of Theorem 7 in [16]. Theorem 3.12. The following conditions are equivalent for any group G: (i) G has the fixed-point property for cones. (ii) Whenever G acts on a C ∗-algebra A and whenever e is a positive element in A for A (e), A), which is bounded on the G-orbit which there exists a non-zero trace in T (J G of e, then there exists an invariant trace in T (J G A (e), A) normalizing e. (iii) Whenever G acts on a C ∗-algebra A and whenever e is a positive element in A for which there exists a trace in T +(A) which is non-zero and bounded on the G-orbit of e, then there is an invariant trace in T +(A) normalizing e. Proof. (i) ⇒ (ii). Set I = J G A (e). Then T (I, A) is weakly complete by Proposition 3.1, the action of G on T (I, A) is locally bounded, by Lemma 3.8, and of cobounded type by Corollary 3.4. The cone of traces T (I, A) therefore has a non-zero fixed point τ0 since G has the fixed-point property for cones. As τ0 is non-zero and invariant we can use Lemma 2.9, Lemma 2.5 and the fact that e ∈ I to conclude that 0 < τ0(e) < ∞. (ii) ⇒ (i). Property (ii), applied to A = ℓ∞(G) (see also Section 4 below), says that condition (4) in [16, Theorem 7] holds, which again, by that theorem, is equivalent to (i). 17 (ii) ⇔ (iii). This follows from Proposition 2.13 and the remarks below that proposition. If A is commutative, or, more generally, if A admits a separating family of bounded traces, then the condition in Theorem 3.12 (ii) and (iii), that there exists a non-zero locally bounded traces, is always satisfied independent on the action of G on A. Below we specialize Theorem 3.12 to the case where the positive element is a projection. Corollary 3.13. Let G be a group with the fixed-point property for cones acting on a C ∗- algebra A, and let p ∈ A be a non-zero projection. The following conditions are equivalent: (i) There is a non-zero lower semi-continuous trace with domain J G A (p), which is bounded on the G-orbit of p. (ii) There is an invariant lower semi-continuous trace in T +(A) normalized on p. (iii) There is a lower semi-continuous trace in T +(A ⋊ G) normalized on p. If, in addition, A ⋊ G is exact, then the conditions above are equivalent to: (iv) The C ∗-algebra p(A ⋊ G)p ⊗ Mn is not properly infinite, for all n ≥ 1. Proof. (i) ⇒ (ii). domain J G automatically lower semi-continuous. A (p) normalizing p. As remarked below Theorem 2.11, each trace on J G It follows from Theorem 3.12 that there is an invariant trace τ with A (p) is (ii) ⇒ (iii). Any invariant lower semi-continuous trace τ on A with τ (p) = 1 extends to a lower semi-continuous trace on A ⋊ G, by Lemma 2.14. (iii) ⇒ (i). Take the restriction to A of the trace whose existence is claimed in (iii) and linearize as in Proposition 2.13. (iii) ⇔ (iv). It is well-known that this equivalence holds when A ⋊ G is exact, see Remark 3.11 and the comments above Theorem 2.15. We can rephrase the corollary above as follows. A projection p in a C ∗-algebra A with a G-action can fail to be normalized by an invariant trace on A for two reasons. Either A does not have a trace that is non-zero and bounded on the G-orbit of p, or the group G possesses some amount of "paradoxicality" materialized in failing to have the fixed-point property for cones. In Lemma 5.7 we give non-obvious examples of projections for which there exist a trace that is non-zero and bounded on the G-orbit of the projection. While we do not know that every group without the fixed-point property for cones has the ability of acting on a C ∗-algebra A in such a way that it obstructs the existence of an invariant trace normalizing a given projection p in A, for which there is a trace on A that is non-zero and bounded on the orbit of p, it does follow from Proposition 3.10 that all non-supramenable groups have this quality. In conclusion, we do not know if Corollary 3.13 characterizes the class of groups with the fixed-point property for cones, the class of supramenable groups, or some intermediate class of groups. 18 4 Invariant integrals on ℓ∞(G) g on G whose absolute value is G-bounded by f , i.e., for which g ≤Pn The bounded (complex valued) functions, ℓ∞(G), on a (discrete) group G is a unital C ∗-al- gebra equipped with an action of the group G by left-translation. Following the notation of Monod, [16], for each positive f ∈ ℓ∞(G), let ℓ∞(G, f ) denote the set of bounded functions j=1 tj.f , for some n ≥ 1 and some t1, . . . , tn ∈ G, where t.f is the left-translate of f ∈ ℓ∞(G) by t ∈ G. In the language of C ∗-algebras, ℓ∞(G, f ) is the smallest (automatically symmetric) hereditary G- invariant ideal in ℓ∞(G) containing f , denoted by I G ℓ∞(G)(f ) in the previous sections. The Pedersen ideal of the uniform closure of ℓ∞(G, f ) is denoted by ℓ∞ c (G, f ), and we have the following inclusions: ℓ∞ c (G, f ) ⊆ ℓ∞(G, f ) ⊆ ℓ∞(G, f ). An invariant integral on G normalized for f is a G-invariant positive linear functional µ on ℓ∞(G, f ) satisfying µ(f ) = 1. In the language of C ∗-algebras, an (invariant) integral on ℓ∞(G, f ) is an (invariant) trace on ℓ∞(G) with domain ℓ∞(G, f ). Monod observed in [16, Theorem 7] that G has the fixed-point property for cones if and only if for each positive function f in ℓ∞(G) there is an invariant integral on G normalized for f . (This result is extended to general C ∗-algebras in our Theorem 3.12.) Let µ be an invariant integral on ℓ∞(G) normalized on some positive function f ∈ ℓ∞(G). We say that µ is lower semi-continuous if µ(h) = limn→∞ µ(hn), whenever {hn}∞ n=1 is an increasing seqence of positive functions in ℓ∞(G, f ) converging uniformly to h ∈ ℓ∞(G, f ), cf. Lemma 2.10. The restriction of µ to ℓ∞ c (G, f ) is automatically lower semi- continuous by Theorem 2.11 (Pedersen). If the restriction of µ to ℓ∞ c (G, f ) is zero, then µ is said to be singular. We shall in this section find conditions that will ensure that an (invariant) integral is lower semi-continuous and also exhibit situations where such integrals necessarily are singular. We start by rephrasing what Proposition 3.10 says about the existence of lower semi-continuous integrals: Proposition 4.1. Let G be a supramenable group, let f be a positive function in ℓ∞(G), and suppose that ℓ∞(G, f ) is G-compact (see above Proposition 3.6). Then there is a non-zero invariant lower semi-continuous integral on ℓ∞ c (G, f ). Example 4.2. Let G be a group and let f be a positive function in ℓ∞(G). Then ℓ∞(G, f ) is G-compact if and only if there exists δ > 0 such that A(f, ε) ∝G A(f, δ), for all 0 < ε < δ; where A(f, η) = {t ∈ G : f (t) > η}, and where A ∝G B, for subsets A, B of G, means that A ⊆St∈F tB, for some finite subset F of G. We will not go into the details of the proof of this, but just mention that to prove the "if" part, one observes that A(f, ε) ∝G A(f, δ) implies that (f − ε)+ ∈ I G ℓ∞(G)((f − δ′)+), for all 0 < δ′ < δ. The condition above ensuring G-compactness of ℓ∞(G, f ) can further be rewritten as n−1}, for all n ≥ 1. Then ℓ∞(G, f ) is G-compact n=1 An, for all N ≥ N0. This n=1 An = A(f, 1/N0), for all k > N0. In other words, ℓ∞(G, f ) if and only if there exists N0 ≥ 1 such that SN condition holds if Ak ∝GSN0 n=1 An ∝G SN0 follows: set An = {t ∈ G : 1 n < f (t) ≤ 1 19 is G-compact if the set of t ∈ G where f (t) is "very small" can be controlled by the set where f (t) is "small enough". Proposition 4.1 and the example above does not give information about the existence of invariant lower semi-continuous integrals normalizing the given positive function f ∈ ℓ∞(G). See Example 4.4 below for more about this problem. Using an example from [15], we proceed to show that the conclusion of Proposition 4.1 fails without the assumption on G-compactness. Proposition 4.3. For each countably infinite group G there is a positive function f ∈ ℓ∞(G) such that ℓ∞ c (G, f ) admits no non-zero invariant integral. In particular, if an in- variant integral on G normalized on f exists, as is the case whenever G has the fixed-point property for cones, then it is singular. Proof. Let G be a countably infinite group. By [15, Proposition 4.3] and its proof there is an increasing sequence {An}n≥1 of (infinite) subsets of G such that if Kn is the (compact- open) closure of (the open set) An in βG, and if Xn =St∈G t.Kn, then X :=Sn≥1 Xn is an open G-invariant subset of βG that admits no non-zero invariant Radon measure. Let ϕ : ℓ∞(G) → C(βG) be the canonical ∗-isomorphism of C ∗-algebras, and let I be the closed G-invariant ideal of ℓ∞(G) such that ϕ(I) = C0(X). Observe that C0(X) is the closed G-invariant ideal in C(βG) generated by the indicator functions 1Kn, n ≥ 1. Since ϕ(1An) = 1Kn, we conclude that I is the closed G-invariant ideal in ℓ∞(G) generated by the projections 1An, n ≥ 1, or by the positive function f =Pn≥1 n−21An ∈ ℓ∞(G). Hence I = ℓ∞(G, f ). Since X admits no non-zero invariant Radon measure, ℓ∞ c (G, f ) ∼= Cc(X) admits no non-zero invariant integrals. Example 4.4 (On the ideal c0(G)). For a countably infinite group G, the subspace c0(G) is a closed G-simple invariant ideal of ℓ∞(G), and the Pedersen ideal of c0(G) is cc(G). It follows that if f is a positive non-zero function in c0(G), then ℓ∞ c (G, f ) = cc(G), ℓ∞(G, f ) = c0(G), while ℓ∞(G, f ) is some invariant hereditary ideal between these two ideals. Being G-simple, the ideal c0(G) is G-compact. The only lower semi-continuous invariant integrals defined on cc(G) are multiples of the counting measure on G. Hence, if f is a positive function in c0(G), then there is an invariant lower semi-continuous integral on G normalized on f if and only if f belongs to ℓ1(G). If a positive function f in c0(G) \ ℓ1(G) is normalized by an invariant integral on G, which is the case if G has the fixed-point property for cones, then this integral is necessarily singular. Consider now a countably infinite group G with the fixed-point property for cones, and take a non-zero positive function f ∈ c0(G). The crossed product C ∗-algebra c0(G) ⋊ G is isomorphic to the C ∗-algebra K of compact operators (on a separable infinite dimensional Hilbert space). There is an invariant integral on G normalized for f , and we can view this 20 integral as a invariant densely defined trace τ on c0(G) with τ (f ) = 1. The image of f in c0(G) ⋊ G is a positive compact operator with eigenvalues {f (t)}t∈G. It was shown in [1] that for a positive compact operator T with eigenvalues {λn}∞ n=1, listed in decreasing order (and with multiplicity), there exists a densely defined trace on K normalized for T if and only if lim inf n→∞ σ2n/σn = 1, (4.1) where σn =Pn j=1 λj. Let s > 0 and choose f ∈ c0(G) such that the eigenvalues of the positive compact operator f ∈ c0(G) ⋊ G is the sequence {n−s}∞ n=1. Then f is trace class if and only if s > 1; and the limit in (4.1) is 1 if and only if s ≥ 1. If s = 1, then f is normalized by the Dixmier trace on K. If 0 < s < 1, then there is no trace on c0(G)⋊G (lower semi-continuous or not) that normalizes f . For such a choice of f , the invariant densely defined trace τ on c0(G) does not extend to a trace on c0(G) ⋊ G, thus showing that the conclusion of Lemma 2.14 fails without the assumption that the trace is lower semi-continuous. By the remark in the example above, that for no positive functions f in c0(G) \ ℓ1(G) does there exist an invariant lower semi-continuous integral µ on ℓ∞(G, f ) with µ(f ) = 1, we get the corollary below, which implies that at least some integrals witnessing the fixed-point property for cones for infinite groups necessarily must be singular. Corollary 4.5. Let G be a countable group with the property that for each positive func- tion f in ℓ∞(G) there exists a G-invariant lower semi-continuous integral µ on ℓ∞(G, f ) normalized for f . Then G must be a finite group. 5 The Roe algebra As an application of the results developed in the previous sections we shall in this last section prove some results about existence (and non-existence) of traces on the Roe algebra ℓ∞(G, K) ⋊ G associated with a (countably infinite) group G, where K denotes the C ∗- algebra of compact operators on a separable infinite dimensional Hilbert space. These C ∗-algebras originates from the thesis of John Roe, published in [21], where the index of elliptic operators are computed using traces on a (variant of) what is now called the Roe algebra. The C ∗-algebra ℓ∞(G, K) is equipped with the natural action of G given by left- translation. We shall consider existence (and non-existence) of invariant traces on this C ∗- algebra. First we give a complete description of the densely defined lower semi-continuous traces on these C ∗-algebras (there are not so many). The group G plays no role in Lemma 5.1 and Proposition 5.2 other than as a set, and it could be replaced with the set of natural numbers N. Lemma 5.1. The Pedersen ideal of ℓ∞(G, K) is equal to ℓ∞(G, F ), where F = Ped(K) is the set of finite rank operators. 21 Proof. The "point evaluation" at t ∈ G gives a surjective ∗-homomorphism ℓ∞(G, K) → K that maps Ped(ℓ∞(G, K)) onto Ped(K) = F . This shows that Ped(ℓ∞(G, K)) ⊆ ℓ∞(G, F ). Conversely, if x ∈ ℓ∞(G, F ), then x(t) has finite rank, for each t ∈ G, and so there exists a finite dimensional projection p(t) ∈ K with p(t)x(t) = x(t). The function t 7→ p(t) defines a projection p in ℓ∞(G, K) satisfying px = x. As p belongs to the Pedersen ideal, being a projection, it follows that also x belongs to Ped(ℓ∞(G, K)). It is worth mentioning that the Pedersen ideal of a C ∗-algebra of the form C(T, K), where T is a compact Hausdorff space, may be properly contained in C(T, F ), cf. [9]. For each s ∈ G, let τs be the (lower semi-continuous densely defined) trace on ℓ∞(G, K) given by τs(f ) = Tr(f (s)), for f either in ℓ∞(G, K)+ or in ℓ∞(G, F ), where Tr is the standard trace on the compact operators K. In the proof of the proposition below we shall view the C ∗-algebra ℓ∞(G, K) as an ℓ∞(G)-algebra via the natural (unital) embedding of ℓ∞(G) into the center of the multiplier algebra of ℓ∞(G, K). Proposition 5.2. For each countable group G, the cone, Tlsc(ℓ∞(G, K)), of densely defined lower semi-continuous traces on ℓ∞(G, K) is equal to the cone of finite positive linear where cs ≥ 0 and cs 6= 0 for only finitely many s ∈ G. combinations of the traces τs, s ∈ G, defined above, i.e., traces of the form Ps∈G csτs, Proof. Clearly, any trace of the form Ps∈G csτs, with cs ≥ 0 and cs 6= 0 only for finitely many s ∈ G, is lower semi-continuous and densely defined. For the converse direction, take τ in Tlsc(ℓ∞(G, K)). Fix a one-dimensional projection e ∈ K. For each s ∈ G, let es ∈ ℓ∞(G, F ) be given by es(s) = e and es(t) = 0, when t 6= s. Set cs = τ (es) ≥ 0. We show first that the set F = {s ∈ G : cs 6= 0} is finite. Suppose it were infinite, and let {s1, s2, s3, . . . } be an enumeration of the elements in F . Let p ∈ ℓ∞(G, K) be a projection such that Tr(p(sn)) ≥ nc−1 sn , for all n ≥ 1. As p ≥ p · 1{s}, for all s ∈ G, we get τ (p) ≥ τ (p · 1{sn}) ≥ csnTr(p(sn)) ≥ n. As this cannot be true for all n ≥ 1, we conclude that F is finite. Set τ0 =Ps∈F csτs, and observe that τ0(f ) = τ (f · 1F ), for all f ∈ ℓ∞(G, K). It follows that τ ′ := τ − τ0 is a positive trace on ℓ∞(G, K), satisfying τ ′(f ) = τ (f · 1F c), for all f ∈ ℓ∞(G, K). We show that τ ′ = 0. Assume, to reach a contradiction, that τ ′ 6= 0. Notice that τ ′ vanishes on cc(G, K) by construction of τ0. Since the Pedersen ideal of ℓ∞(G, K) is generated (as a hereditary ideal) by its projections, cf. the proof of Lemma 5.1, there n=1 Fn, where {Fn}n≥1 is a strictly increasing sequence of finite subsets of G. Find a sequence {pk}k≥1 of pairwise orthogonal projections in ℓ∞(G, K) such that each pk is equivalent to p. Let q ∈ ℓ∞(G, K) be the projection given by is a projection p ∈ ℓ∞(G, K) such that τ ′(p) > 0. Write G =S∞ q(s) = p1(s) + p2(s) + · · · + pn(s), s ∈ Fn+1 \ Fn, for n ≥ 0 (with the convention F0 = ∅). Then q · 1F c n ≥ (p1 + p2 + · · · + pn) · 1F c n, 22 for all n ≥ 1; and as τ ′(g · 1Ec) = τ ′(g), for all g ∈ ℓ∞(G, K) and all finite subsets E ⊆ G, we conclude that τ ′(q) = τ ′(q · 1F c n) ≥ τ ′((p1 + p2 + · · · + pn) · 1F c n) = τ ′(p1 + p2 + · · · + pn) = nτ ′(p), for all n ≥ 1, which is impossible. A non-zero trace of the form as in Proposition 5.2 can clearly not be G-invariant when G is infinite, so we obtain the following: Corollary 5.3. The C ∗-algebra ℓ∞(G, K) admits no non-zero lower semi-continuous in- variant densely defined traces whenever G is a countably infinite group, and, consequently, the Roe algebra ℓ∞(G, K) ⋊ G admits no non-zero densely defined lower semi-continuous trace. When combining the conclusion of the corollary above with Theorem 3.9, we see that the action of G on Tlsc(ℓ∞(G, K)) either must fail to be of cobounded type, or there is no non-zero locally bounded trace in Tlsc(ℓ∞(G, K)). If fact, both fail! That the latter fails follows easily from the description of Tlsc(ℓ∞(G, K)) in Proposition 5.2. Lemma 5.4. If G is a countably infinite group, then the action of G on Tlsc(ℓ∞(G, K)) is not of cobounded type. Proof. Let e be a positive contraction in Ped(ℓ∞(G, K)). We must find another positive contraction a in Ped(ℓ∞(G, K)) that is not tracially G-dominated by e, cf. Lemma 3.2. In j=1 τ (tj.e) will fail other words, for all finite sets t1, t2, . . . , tn ∈ G, the inequality τ (a) ≤Pn for at least one τ ∈ Tlsc(ℓ∞(G, K)), or, taking Proposition 5.2 into account, Tr(a(s)) ≤ Tr(e(t−1 j s)), nXj=1 will fail for at least one s ∈ G. It follows from Lemma 5.1 and its proof that e is dominated by a projection in ℓ∞(G, K), so we may without loss of generality assume that e itself is a projection. Set f (s) = Tr(e(s)), for s ∈ G. Let G = {s1, s2, s3, . . . } be an enumeration of the elements in G, and let {uj}∞ j=1 be a j=1 uj.f , for each N ≥ 1, and let g : G → N0 be given by g(sN ) = fN (sN ) + 1, for N ≥ 1. For each j=1 tj.f ≤ fN , which entails that sequence in which each element of G is repeated infinitely often. Set fN =PN finite set t1, t2, . . . , tn ∈ G there exists N ≥ 1 such thatPn g(sN ) (cid:2)Pn Let now a ∈ ℓ∞(G, K) be a projection such that Tr(a(t)) = g(t), for all t ∈ G. Then a j=1 tj.f (sN ), for each N ≥ 1. is not tracially G-dominated by e. Although there are no densely defined lower semi-continuous invariant traces on ℓ∞(G, K), when G is infinite, there are still interesting lower semi-continuous traces on the Roe In the results to follow we describe the class of projections p ∈ ℓ∞(G, K) for algebra. 23 which there exists a lower semi-continuous trace on ℓ∞(G, K) which is bounded and non- zero on the orbit of p, respectively, which is invariant and normalizes p. These two classes of projections agree when G has the fixed-point property for cones, cf. Corollary 3.13. First we note the following general result that holds for locally finite groups: Proposition 5.5. The Roe algebra ℓ∞(G, K) ⋊ G of any locally finite group G is stably finite. Moreover, for each non-zero projection p ∈ ℓ∞(G, K) there exists a lower semi- continuous trace τ ∈ T +(ℓ∞(G, K) ⋊ G) with τ (p) = 1, and hence an invariant lower semi-continuous trace on ℓ∞(G, K) normalizing p. Proof. Write G = S∞ n=1 Gn as an increasing union of finite groups. The C ∗-algebra ℓ∞(G, K) is stably finite, as can be witnessed by the separating family of densely defined traces from Proposition 5.2. The crossed product ℓ∞(G, K) ⋊ Gn is stably finite because it embeds into the stably finite C ∗-algebra ℓ∞(G, K) ⊗ B(ℓ2(Gn)). It follows that ℓ∞(G, K) ⋊ G = lim−→ ℓ∞(G, K) ⋊ Gn is stably finite, being an inductive limit of stably finite C ∗-algebras. Let next p ∈ ℓ∞(G, K) be a non-zero projection. Then p(ℓ∞(G, K) ⋊ G)p = lim−→ p(ℓ∞(G, K) ⋊ Gn)p. (5.1) The unital C ∗-algebra p(ℓ∞(G, K)⋊Gn)p admits a tracial state, for each n ≥ 1. To see this, choose m ≥ n such that the restriction p′ of p to ℓ∞(Gm, K) is non-zero. The restriction mapping ℓ∞(G, K) → ℓ∞(Gm, K) is Gn-equivariant and therefore extends to a ∗-homo- morphism ℓ∞(G, K) ⋊ Gn → ℓ∞(Gm, K) ⋊ Gn, and in turns to a unital ∗-homomorphism p(ℓ∞(G, K) ⋊ Gn)p → p′(ℓ∞(Gm, K) ⋊ Gn)p′. Composing this ∗-homomorphism with any tracial state on the (finite dimensional) C ∗-algebra p′(ℓ∞(Gm, K) ⋊ Gn)p′ gives a tracial state on p(ℓ∞(G, K)⋊Gn)p. As the inductive limit of a sequence of unital C ∗-algebras with unital connecting mappings (e.g., as in (5.1)) admits a tracial state if (and only if) each C ∗-algebra in the sequence does, we conclude that p(ℓ∞(G, K) ⋊ G)p admits a tracial state. By Remark 3.11 there is a lower semi-continuous trace on ℓ∞(G, K) ⋊ G normalizing the projection p. The restriction of this trace to ℓ∞(G, K) is an invariant lower semi-continuous trace still normalizing p. The Roe algebra of any infinite, locally finite group provides yet another example of a stably finite C ∗-algebra with an approximate unit consisting of projections which has no densely defined lower semi-continuous trace, cf. Theorem 2.15 and Proposition 2.16. For the more general class of groups G with the fixed-point property for cones, it follows from Corollary 3.13 that if p is a projection in ℓ∞(G, K), then there is a lower semi- continuous invariant trace on ℓ∞(G, K) that normalizes p (and hence there is a trace on the Roe algebra ℓ∞(G, K) ⋊ G normalizing p), if and only if there is a trace on ℓ∞(G, K), which is non-zero and bounded on the orbit {t.p}t∈G. If p has uniformly bounded dimension, i.e., if supt∈G Tr(p(t)) < ∞, then such a trace clearly exists, take for example τs (defined above Proposition 5.2), for any fixed s ∈ G. One can do a bit better: In Proposition 5.10 24 below it is shown that for each projection in ℓ∞(G, K) with uniformly bounded dimension there is a trace on the Roe algebra normalizing this projection provided that the group G is supramenable (a formally weaker condition than having the fixed-point property for cones). Let G be a countably infinite group, and let ℓ : G → N0 be a proper length function on G, i.e., ℓ(t) = 0 if and only if t = e, ℓ(s + t) ≤ ℓ(s) + ℓ(t), for all s, t ∈ G, and Wn := {t ∈ G : ℓ(t) ≤ n} is finite, for all n ≥ 0. Such proper length functions always exist, and if G is finitely generated, then we can take ℓ to be the word length function with respect to some finite generating set for G. Set αn = max t∈Wn Tr(p(t)), Zn = {t ∈ Wn : Tr(p(t)) = αn}. (5.2) Let τn be the (lower semi-continuous densely defined) trace on ℓ∞(G, K) given by τn(f ) = 1 ZnXt∈Zn α−1 n Tr(f (t)), for f in ℓ∞(G, K)+ or in ℓ∞(G, F ). Observe that τn(p) = 1, for all n ≥ 1. Let ω be a free ultrafilter on N0, and define a trace τω,p on the positive cone ℓ∞(G, K)+ by τω,p(f ) = lim ω τn(f ), f ∈ ℓ∞(G, K)+, (where the limit along the ultrafilter is taken in the compact set [0, ∞]). Let Cω,p be the cone of positive functions f in ℓ∞(G, K), for which τω,p(f ) < ∞, and let Iω,p be the linear span of Cω,p. Then Iω,p is a hereditary symmetric ideal in ℓ∞(G, K), and τω,p defines a linear trace on Iω,p, cf. Proposition 2.13. We say that the projection p ∈ ℓ∞(G, K) has subexponentially growing dimension if lim inf n→∞ αn+m αn = 1, (5.3) for all m ≥ 0. (This definition may depend on the choice of proper length function ℓ, and should be understood to be with respect to some proper length function.) Lemma 5.6. Let {αn}∞ isfying (5.3). Then there is a free ultrafilter ω on N0 such that n=0 be an increasing sequence of strictly positive real numbers sat- lim ω αn+m αn = 1, for all m ≥ 1. Proof. For each m ≥ 0 and ε > 0, set Am,ε = {n ≥ 0 : αn+m/αn ≤ 1 + ε}. By the assumption that (5.3) holds, each of the sets Am,ε is infinite. The collection of sets Am,ε is downwards directed, since Am1,ε1 ⊆ Am2,ε2, when m1 ≥ m2 and ε1 ≤ ε2. It follows that the intersection of any finite collection of these sets is infinite. We can therefore find a free ultrafilter ω which contains all the sets Am,ε; and any such ultrafilter will satisfy the conclusion of the lemma. 25 Lemma 5.7. Let G be a countably infinite group, let p ∈ ℓ∞(G, K) be a projection of subexponentially growing dimension, and let ω be a free ultrafilter as in Lemma 5.6 for the sequence {αn}∞ n=0 associated with the projection p as in (5.2). Then τω,p(p) = 1 and τω,p(t.p) ≤ 1, for all t ∈ G. Proof. We have already noted that τn(p) = 1, for all n ≥ 1, which implies that τω,p(p) = 1. Let m ≥ 0 and let s ∈ Wm. For n ≥ 0 and t ∈ Zn ⊆ Wn, we have st ∈ Wn+m, so that Tr(p(st)) ≤ αn+m, which shows that τn(s−1.p) = 1 ZnXt∈Zn α−1 n Tr(p(st)) ≤ αn+m αn . By the assumption that p has subexponentially growing dimension, by Lemma 5.6, and by the choice of ω, we conclude that τω,p(s−1.p) ≤ 1. Theorem 5.8. Let G be a countably infinite group with the fixed-point property for cones and let p ∈ ℓ∞(G, K) be a projection of subexponentially growing dimension. Then there is an invariant lower semi-continuous trace on ℓ∞(G, K) normalized on p, and there is a lower semi-continuous trace on the Roe algebra ℓ∞(G, K) ⋊ G also normalized on p. Proof. This follows from Corollary 3.13, where condition (i) is satisfied with τ = τp,ω, cf. Lemma 5.7. We proceed to examine the case of projections in ℓ∞(G, K) of uniformly bounded dimension. In the lemma below we embed ℓ∞(G) into ℓ∞(G) ⋊ G, which again embeds into (the upper left corner of) (ℓ∞(G) ⋊ G) ⊗ Mn, for each n ≥ 1. Note that projections in ℓ∞(G) are indicator functions 1E, for some subset E of G. Lemma 5.9. If G is an exact group and n ≥ 1 is an integer. Then for each projection p in (ℓ∞(G) ⋊ G) ⊗ Mn there is a projection r in ℓ∞(G) such that p and r generate the same closed two-sided ideal in (ℓ∞(G) ⋊ G) ⊗ Mn. Proof. Let n ≥ 1, let p ∈ (ℓ∞(G) ⋊ G) ⊗ Mn be a projection, and let I be the closed two- sided ideal in (ℓ∞(G) ⋊ G) ⊗ Mn generated by p. It follows from [26, Theorem 1.16] that I is the closed two-sided ideal in (ℓ∞(G) ⋊ G) ⊗ Mn generated by J := I ∩ (ℓ∞(G) ⊗ Mn). Arguing as in the proof of [11, Proposition 5.3] we find a projection q ∈ J which generates the ideal I in (ℓ∞(G)⋊G)⊗Mn. By [28], since ℓ∞(G) is of real rank zero, q is equivalent to a diagonal projection diag(q1, q2, . . . , qn) in ℓ∞(G) ⊗ Mn. Let r ∈ ℓ∞(G) be the supremum of the projections q1, q2, . . . , qn. Then q and r generate the same ideal of ℓ∞(G) ⊗ Mn, and so r and p generate the same ideal in (ℓ∞(G) ⋊ G) ⊗ Mn. The result below extends the characterization in [11] of supramenable groups in terms of non-existence of properly infinite projections in the uniform Roe algebra. Proposition 5.10. Let G be a group. The following conditions are equivalent: (i) G is supramenable. 26 (ii) Each non-zero projection in the stabilized uniform Roe algebra (ℓ∞(G) ⋊ G) ⊗ K is normalized by a lower semi-continuous trace in T +((ℓ∞(G) ⋊ G) ⊗ K). (iii) Each projection in ℓ∞(G, K) of uniformly bounded dimension is normalized by an in- variant lower semi-continuous trace in T +(ℓ∞(G, K)) and by a lower semi-continuous trace in T +(ℓ∞(G, K) ⋊ G) (iv) The stabilized uniform Roe algebra (ℓ∞(G) ⋊ G) ⊗ K contains no properly infinite projections. Proof. (i) ⇒ (ii). Since (ℓ∞(G) ⋊ G) ⊗ K is the inductive limit of C ∗-algebras of the form (ℓ∞(G)⋊G)⊗Mn, for n ≥ 1, it suffices to show that each projection p in (ℓ∞(G)⋊G)⊗Mn is normalized by a lower semi-continuous trace on this C ∗-algebra. By Lemma 5.9 there is a projection q = 1E ∈ ℓ∞(G) that generates the same closed two-sided ideal in (ℓ∞(G) ⋊ G) ⊗ Mn as p. Since G is supramenable, the set E must be non-paradoxical, so by Tarski's theorem there is an invariant trace τ0 on ℓ∞(G, q) which normalizes q (see [11, Proposition 5.3]). Extend τ0 to a trace τ on J := J(ℓ∞(G)⋊G)⊗Mn(q) satisfying τ (q) = 1. Now, p ∈ J and 0 < τ (p) < ∞, by the assumption on p and q, so upon rescaling we obtain a trace τ on J satisfying τ (p) = 1; and τ is lower semi- continuous by the remarks below Theorem 2.11. Finally, we can extend τ to a trace in T +((ℓ∞(G) ⋊ G) ⊗ Mn) using (2.3). (ii) ⇒ (iii). Let p be a projection in ℓ∞(G, K) such that n := supt∈G Tr(p(t)) < ∞. Let e ∈ K be a projection of dimension n, and let e ∈ ℓ∞(G, K) be the projection given by e(t) = e, for all t ∈ G. Then e is fixed by the action of G, and we have isomorphisms e(ℓ∞(G, K))e ∼= ℓ∞(G) ⊗ Mn, e(ℓ∞(G, K) ⋊ G)e ∼= (ℓ∞(G) ⋊ G) ⊗ Mn. Moreover, p is equivalent to a projection p0 in e(ℓ∞(G, K))e, which, under the isomorphism above, corresponds to a projection p1 ∈ ℓ∞(G) ⊗ Mn. By (ii) there is a lower semi- continuous trace on (ℓ∞(G) ⋊ G) ⊗ Mn normalizing p1. Hence there is a lower semi- continuous trace on e(ℓ∞(G, K) ⋊ G)e normalizing p0. Arguing as in Remark 3.11 we can extend this trace to a lower semi-continuous trace in T +(ℓ∞(G, K) ⋊ G). The restriction of τ to ℓ∞(G, K) becomes an invariant lower semi-continuous trace. (iii) ⇒ (iv) is clear (no properly infinite projection can be normalized by a trace). If G is non-supramenable, then ℓ∞(G) ⋊ G contains a propery infinite projection, cf. [20, Proposition 5.5], in which case (iv) cannot be true, which proves (iv) ⇒ (i). It was shown in [20] that the uniform Roe algebra ℓ∞(G) ⋊ G is properly infinite, i.e., its unit is a properly infinite projection, if and only if G is non-amenable; and in [11] it was shown that the uniform Roe algebra contains a properly infinite projection if and only if G is non-supramenable. The proposition above allows us to conclude that no matrix algebras over the uniform Roe algebra contains a properly infinite projections when G is supramenable. It was shown by Wei in [27] and Scarparo in [25], answering a question in [11], that ℓ∞(G) ⋊ G is finite if and only if G is a locally finite group. Using a similar idea as in [25] 27 we show below that the Roe algebra ℓ∞(G, K) ⋊ G contains a properly infinite projection whenever G is not locally finite. We may therefore strengthen Proposition 5.5 as follows: The Roe algebra ℓ∞(G, K) ⋊ G is stably finite if and only if G is locally finite; and if ℓ∞(G, K) ⋊ G is not stably finite, then it contains a properly infinite projection. Lemma 5.11. Let G be a non-locally finite group. Then there is a finite subset S of G such that for each integer N ≥ 1 there exists a non-zero function : G → N0 satisfying N ≤Xs∈S s.. (5.4) Proof. Let G0 be an infinite, finitely generated subgroup of G, and let S be a finite sym- metric generating set for G0. By [29, Lemma 1] there is a one-sided geodesic {tn}∞ n=0 in G0, where tnt−1 n+1 ∈ S, for all n ≥ 0; and n 7→ tn is injective. Fix N ≥ 1 and define : G → N0 by (tn) = N n, for all n ≥ 0; and set (t) = 0 if t /∈ {tn : n ≥ 0}. Fix n ≥ 0 and set s = tnt−1 n+1 ∈ S. Then s.(tn) = (tn+1) = N(tn). As (t) = 0, when t /∈ {tn : n ≥ 0}, we see that (5.4) holds. Proposition 5.12. For each countable non-locally finite group G there is a projection p ∈ ℓ∞(G, K) satisfying: (i) p is properly infinite in the Roe algebra ℓ∞(G, K) ⋊ G. (ii) There is no trace on the Roe algebra ℓ∞(G, K) ⋊ G normalizing the projection p. (iii) Each trace on ℓ∞(G, K) is either unbounded or zero on the orbit {t.p}t∈G. Proof. Note first that if e and f are projections in ℓ∞(G, K), then e ∼ f , respectively, e - f , in ℓ∞(G, K) if and only if Tr(e(t)) = Tr(f (t)), respectively, Tr(e(t)) ≤ Tr(f (t)), for all t ∈ G. Let : G → N0 be as in Lemma 5.11 with respect to a finite subset S of G and with N ≥ 2S. Find a projection q ∈ ℓ∞(G, K) with Tr(q(t)) = (t), for all t ∈ G. Let p and p′ be the the S-fold, respectively, the N-fold, direct sum of q with itself in ℓ∞(G, K), and set e =Ls∈S s.q. Then Tr(p′(t)) = N(t) and Tr(e(t)) =Ps∈S s.(t), for all t ∈ G, so p′ - e and p ⊕ p - p′ in ℓ∞(G, K). Moreover, p ∼ e in ℓ∞(G, K) ⋊ G (since q ∼ t.q in ℓ∞(G, K) ⋊ G, for all t ∈ G). Hence, in ℓ∞(G, K) ⋊ G, which shows that (i) and (ii) hold. p ⊕ p - p′ - e ∼ p, (iii) follows from (ii) when G has the fixed-point property for cones, cf. Corollary 3.13, on functions σ : G → N0, to the left and right hand side of (5.4) k − 1 times, we get that but requires a separate argument for general groups. Applying the operator σ 7→Ps∈S s.σ, N k ≤Ps∈Sk s., for all k ≥ 1, where Sk is the set of k-tuples of elements from S, and s ∈ G is the product of the k elements in the k-tuple s ∈ Sk. For k ≥ 1, set ek =Ls∈Sk s.q, Tr(ek(t)) =Ps∈Sk s.(t), for all t ∈ G, so p′ k - ek. and let p′ k be the N k-fold direct sum of q with itself. Then Tr(p′ k(t)) = N k(t) and 28 Let τ be any trace in T +(ℓ∞(G, K)). Then N kτ (q) = τ (p′ k) ≤ τ (ek) = Xs∈Sk τ (s.q) ≤ Sk sup t∈G τ (t.q). As this holds for all k ≥ 1, either τ (q) = 0 or supt∈G τ (t.q) = ∞. Suppose that {τ (t.p)}t∈G is bounded. Then {τ (t.q)}t∈G is also bounded. Fix t ∈ G, and let τ ′ be the trace on ℓ∞(G, K) given by τ ′(f ) = τ (t.f ), for f ∈ ℓ∞(G, K)+. As τ ′ also is bounded on the orbit {s.q}s∈G, the argument above implies that 0 = τ ′(q) = τ (t.q), so τ (t.p) = Sτ (t.q) = 0. This proves that τ is zero on the orbit {t.p}t∈G. Any projection satisfying the conclusions of Proposition 5.12 (iii) above must have expo- nentially growing dimension with respect to any proper length function on the group, cf. Lemma 5.7. References [1] S. Albeverio, D. Guido, A. Ponosov, and S. Scarlatti, Singular traces and compact operators, J. Funct. Anal. 137 (1996), no. 2, 281 -- 302. [2] B. Blackadar and J. Cuntz, The stucture of stable algebraically simple C ∗-algebras, American J. Math. 104 (1982), 813 -- 822. [3] B. Blackadar and D. Handelman, Dimension functions and traces on C ∗-algebras, J. Funct. Anal. 45 (1982), 297 -- 340. [4] B. Blackadar and M. Rørdam, Extending states on preordered semigroups and the existence of quasitraces on C ∗-algebras, J. Algebra 152 (1992), 240 -- 247. [5] K. T. Coward, G. A. Elliott, and C. Ivanescu, The Cuntz semigroup as an invariant for C ∗-algebras, J. Reine Angew. Math. 623 (2008), 161 -- 193. [6] J. Cuntz, Dimension functions on simple C ∗-algebras, Math. Ann. 233 (1978), 145 -- 153. [7] J. Dixmier, Traces sur les C ∗-alg`ebres, Ann. Inst. Fourier (Grenoble) 13 (1963), no. fasc. 1, 219 -- 262. [8] G. A. Elliott, L. Robert, and L. Santiago, The cone of lower semicontinuous traces on a C ∗-algebra, Amer. J. Math. 133 (2011), no. 4, 969 -- 1005. [9] R. M. Gillette and D. C. Taylor, A characterization of the Pedersen ideal of C0(T, B0(H)) and a counterexample, Proc. Amer. Math. Soc. 68 (1978), no. 1, 59 -- 63. [10] U. Haagerup, Quasitraces on exact C ∗-algebras are traces, C. R. Math. Acad. Sci. Soc. R. Can. 36 (2014), no. 2-3, 67 -- 92. [11] J. Kellerhals, N. Monod, and M. Rørdam, Non-supramenable groups acting on locally com- pact spaces, Doc. Math. 18 (2013), 1597 -- 1626. 29 [12] E. Kirchberg, On the existence of traces on exact stably projectionless simple C ∗-algebras, Operator Algebras and their Applications (P. A. Fillmore and J. A. Mingo, eds.), Fields Institute Communications, vol. 13, Amer. Math. Soc., 1995, pp. 171 -- 172. [13] E. Kirchberg and M. Rørdam, Non-simple purely infinite C ∗-algebras, American J. Math. 122 (2000), 637 -- 666. [14] , Infinite non-simple C ∗-algebras: absorbing the Cuntz algebra O∞, Advances in Math. 167 (2002), no. 2, 195 -- 264. [15] H. Matui and M. Rørdam, Universal properties of group actions on locally compact spaces, J. Funct. Anal. 268 (2015), no. 12, 3601 -- 3648. [16] N. Monod, Fixed points in convex cones, Trans. Amer. Math. Soc. Ser. B 4 (2017), 68 -- 93. [17] C. Pasnicu and M. Rørdam, Purely infinite C ∗-algebras of real rank zero, J. Reine Angew. Math. 613 (2007), 51 -- 73. [18] G. K. Pedersen, C ∗-algebras and their automorphism groups, Academic Press, London, 1979. [19] G.K. Pedersen, Measure theory for C ∗ algebras, Math. Scand. 19 (1966), 131 -- 145. [20] M. Rørdam and A. Sierakowski, Purely infinite C ∗-algebras arising from crossed products, Ergodic Theory Dynam. Systems 32 (2012), no. 1, 273 -- 293. [21] J. Roe, An index theorem on open manifolds. I, II, J. Differential Geom. 27 (1988), no. 1, 87 -- 113, 115 -- 136. [22] M. Rørdam, On the Structure of Simple C ∗-algebras Tensored with a UHF-Algebra, II, J. Funct. Anal. 107 (1992), 255 -- 269. [23] J. M. Rosenblatt, Invariant measures and growth conditions, Trans. Amer. Math. Soc. 193 (1974), 33 -- 53. [24] W. Rudin, Functional analysis, second ed., International Series in Pure and Applied Math- ematics, McGraw-Hill, Inc., New York, 1991. [25] E.P. Scarparo, Characterizations of locally finite actions of groups on sets, Glasgow Math. J., to appear. [26] A. Sierakowski, The ideal structure of reduced crossed products, Munster J. Math. 3 (2010), 237 -- 261. [27] S. Wei, On the quasidiagonality of Roe algebras, Sci. China Math. 54 (2011), no. 5, 1011 -- 1018. [28] S. Zhang, Diagonalizing projections in multiplier algebras and in matrices over a C ∗-algebra, Pacific J. Math. 145 (1990), no. 1, 181 -- 200. [29] A. Zuk, On an isoperimetric inequality for infinite finitely generated groups, Topology 39 (2000), no. 5, 947 -- 956. 30 Mikael Rørdam Department of Mathematical Sciences University of Copenhagen Universitetsparken 5, DK-2100, Copenhagen Ø Denmark [email protected] 31
1512.04288	2	1512	2015-12-30T10:55:07	A Cuntz algebra approach to the classification of near-group categories	[ "math.OA", "math.CT", "math.QA", "math.RT" ]	We classify C$^$ near-group categories by using Vaughan Jones theory of subfactors and the Cuntz algebra endomorphisms. Our results show that there is a sharp contrast between two essentially different cases, integral and irrational cases. When the dimension of the unique non-invertible object is an integer, we obtain a complete classification list, and it turns out that such categories are always group theoretical. When it is irrational, we obtain explicit polynomial equations whose solutions completely classify the C$^$ near-group categories in this class.	math.OA	math	A CUNTZ ALGEBRA APPROACH TO THE CLASSIFICATION OF NEAR-GROUP CATEGORIES MASAKI IZUMI Dedicated to Professor Vaughan Jones for his sixtieth birthday. Abstract. We classify C∗ near-group categories by using Vaughan Jones theory of subfactors and the Cuntz algebra endomorphisms. Our results show that there is a sharp contrast between two essentially different cases, integral and irrational cases. When the dimension of the unique non-invertible object is an integer, we obtain a complete classification list, and it turns out that such categories are always group theoretical. When it is irrational, we obtain explicit polynomial equations whose solutions completely classify the C∗ near-group categories in this class. Contents Irrational case Introduction 1. 2. Preliminaries 3. Basic ingredients 4. Reconstruction 5. The case of m = G − 1 6. The noncommutative case 7. 8. Polynomial equations for the irrational case 9. The case of m = G 10. The case of m = 2G 10.1. Possible cases 10.2. Case I 10.3. Case II 10.4. Case III 10.5. The case of G = Z2 10.6. The case of G = Z3 10.7. The case of G = Z4 10.8. The case of G = Z2 × Z2 11. 2G 12. Orbifold construction I (de-equivariantization) 12.1. Untwisted case 12.2. Twisted case l 1 subfactors 2 5 12 29 36 39 50 56 70 74 74 75 81 85 89 90 94 96 100 106 106 115 2010 Mathematics Subject Classification. Primary 22E66; Secondary 46L99. Key words and phrases. Supported in part by the Grant-in-Aid for Scientific Research (B) 22340032, and 15H03623, JSPS. subfactors, fusion categories, Cuntz algebras. 1 2 MASAKI IZUMI 13. Orbifold construction II (equivariantization) 13.1. Automorphism groups 13.2. The case of m > G 13.3. The case of m = G 14. Appendix References 117 117 124 128 130 133 1. Introduction In his celebrated paper [35], Vaughan Jones introduced the notion of index for subfactors, and opened up a totally new subject related to various fields in mathe- matics and mathematical physics, such as low dimensional topology, conformal field theory, quantum groups etc. The theory of subfactors is often compared with the classical Galois theory of field extensions, and in fact a subfactor gives rise to tensor categories, which play an analogous role of the Galois group. More precisely, to a subfactor N ⊂ M of finite index and finite depth, we can associate two kinds of fusion categories, consisting of M − M bimodules and N − N bimodules respectively, and two kinds of module categories of them consisting of N − M bimodules and M − N bimodules. Fusion categories may be considered as a generalization of the representation cat- egories of finite groups, and commonly appear in the above mentioned fields related to subfactors. Their axioms were formulated by Etingof, Nikshych, and Ostrik [12] as follows: a fusion category over an algebraically closed field k is a rigid semisim- ple k-linear tensor category with finitely many simple objects and finite dimensional morphism spaces such that the unit object 1 is simple. Other than the representation category of a finite group G, a typical example of a fusion category, which may not be commutative in general, is given by G itself. Namely, if every simple object of a fusion category C is invertible, the set of equivalence classes of simple objects in C is identified with a finite group G obeying the fusion rules (1.1) g ⊗ h ∼= gh, g, h ∈ G. It is known that such fusion categories, called pointed categories, are parametrized by the third cohomology group H 3(G, k×) arising from the associativity constraint. A near-group category, formally introduced by Siehler [50], is one step beyond the pointed categories, and has only one non-invertible simple object. Let C be a near- group category with the group G of the invertible simple objects, and let ρ be the unique non-invertible object in C. Then the only possible fusion rules, apart from Eq.(1.1), are g ⊗ ρ ∼= ρ ⊗ g ∼= ρ, ρ ⊗ ρ ∼=Mg∈G z g ⊕ } { g ∈ G, m times ρ ⊕ ρ ⊕ · · · ⊕ ρ . Thus in the level of fusion rules, the multiplicity m of ρ in ρ ⊗ ρ is the only free parameter, which we call the multiplicity parameter of C. The based ring ZG + Zρ THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 3 obeying the above relations is denoted by K(G, m) in [47]. The classification of the near-group categories having a prescribed based ring K(G, m) is a fundamental question in the subject. Many near-group categories appear in nature. For example, the Ising model in conformal field theory is a near-group category with (G, m) = (Z2, 0), and the repre- sentation category of the symmetric group S3 is a near-group category with (G, m) = (Z2, 1). The E6 subfactor provides a more exotic example with (G, m) = (Z2, 2). The recent developments of the classification of subfactors [36], [1] show that quadratic categories, slight generalization of near-group categories, are rather typical (see [20]), and it is hard to construct subfactors neither related to (quantum) groups nor qua- dratic categories. So far the only known exception is the extended Haagerup subfactor constructed in [2]. The first systematic classification result of a class of near-group categories was obtained by Tambara-Yamagami [52], and they completely classified the near-group categories with m = 0, now called Tambara-Yamagami categories. They showed that such categories exist if and only if G is abelian, and obtained complete classification invariants in terms of G. Another extreme case was treated by Ostrik [46], where he showed that a near-group category with trivial G exists if and only if m = 1. The proof requires a much more sophisticated argument than one would think at first sight, and it uses braiding in a crucial way. Since then, there have been several attempts to classify various classes of near-group categories (see [50], [10], [23], [47], [45], [16], [39]). Fusion categories arising from subfactors form a special class, called C∗ fusion cat- egories, whose morphism spaces are complex Banach spaces with ∗-structure obeying the C∗ condition. Throughout this note, we focus on near-group categories satisfying this condition, which we call C∗ near-group categories. It is known that every C∗ fusion category is uniquely realized in the category of bimodules of the hyperfinite II1 factor, and similarly in the category of endomorphisms End(M) of the hyperfi- nite type III1 factor M (see [25], [48]). In the latter category, invertible objects are nothing but automorphisms of M, and one can utilize well-known results on group actions on operator algebras to analyse the structure of C∗ near-group categories. In fact, the third cohomology class of a pointed category for a finite group G mentioned above can be identified with the Connes obstruction of a G-kernel (see [5], [34], [51]). Moreover, the author pointed out in [28] and [32] that C∗ near group categories re- alized in End(M) can be reconstructed from Cuntz algebra endomorphisms, which provides us a handy way to construct new examples. The main purpose in [32] was to compute the Drinfeld centers for concrete examples of tensor categories, and for that we deduced polynomial equations whose solutions give C∗ near group categories via Cuntz algebra endomorphisms. One of the purposes of this paper is to further pursue this approach, and to classify the C∗ near-group categories by using operator algebra techniques. We summarize the main results of this paper now, which we prove in Section 3-6 (see Theorem 3.9 and Theorem 6.1). 4 MASAKI IZUMI Theorem 1.1. Let C be a C∗ near-group category with G, ρ, and m 6= 0 as above, and let d = be the dimension of the object ρ. Then the following hold. m+√m2+4G If d is rational, either of the following two cases occurs: 2 (1) G is abelian and m = G − 1. (2) G is an extra-special 2-group of order 22a+1 and m = 2a with a natural num- ber a. For each extra-special 2-group, there exist exactly 3 C∗ near group- categories. If d is irrational, G is abelian and m is a multiple of G. Siehler [50, Theorem 1.2] showed, under the implicit assumption of G being abelian, that if m < G, the group G is cyclic, m = G− 1, and G + 1 is a prime power. On the other hand, Etingof-Gelaki-Ostrik [10, Corollary 7.4] completely classified such fusion categories, and showed that there exist exactly three categories for G = Z2, two categories for each of Z3 and Z7, and there exists a unique category for every other cyclic group Zq−1 with a prime power q. These results hold without the C∗ condition. The case (2) was overlooked in [50] for the reason stated above. The proof of the statement for the irrational case can be found in [45] (and in [16, Theorem 2(a)] with an extra assumption). In the case of m = G, Evans-Gannon [16, Theorem 4] showed that the solutions of the polynomial equations obtained in [32] completely classify the C∗ near-group categories in this class. Moreover, they obtained a number of new solutions and computed the Drinfeld centers of the corresponding C∗ near-group categories following the approach established in [31], [32]. We will deduce the polynomial equations for the general irrational case in Section 7-8, and show that their solutions completely classify such categories in Theorem 8.6. We will treat the special case of m = G in Section 9, and the case of m = 2G in Section 10. Recently an example with (G, m) = (Z3, 6) was discovered, and the latter case also draws attention of specialists (see [19], [39], [40]). Theorem 10.19 shows that there exist exactly two C∗ near group categories with (G, m) = (Z3, 6), and they are complex conjugate to each other. This paper is an extended version of author's personal note written around 2008, which roughly corresponds to the contents in Section 3-9 in the present version. Since then there have been several new developments about near-group categories, and some results in the original note are no longer new as already mentioned above. Nevertheless, the author believes that it is still worth publishing the old part because it lays the foundation for the other part of this note, and it still contains new results, for example, the case (2) in Theorem 1.1. Section 2 and Section 10-13 are newly written in 2015. As already mentioned above, it is a folklore result that any C∗ fusion category uniquely embeds into End(M) for the hyperfinite type III1 factor, and the uniqueness part is based on Popa's classification result for subfactors [48]. In Section 2, we clarify in what sense the uniqueness statement holds. The author would like to thank Roberto Longo for pointing out the author's vague understanding of the statement before, and Luca Giorgetti for pointing out a flaw in Theorem 2.2 in the first version. THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 5 In [28], [32], we showed that a C∗ near-group category with m = G gives rise to so called a 2G1 subfactor. Generalizing this observation, we show in Section 11 that there exists a one-to-one correspondence between the C∗ near-group categories in the irrational case and a certain class of subfactors with specific principal graphs. We show that these subfactors are self-dual under a mild assumption. Section 12 is devoted to de-equivariantization of C∗ near group categories, which is a systematic account inspired by Ostrik's wonderful observation that the Haagerup category is related to a near-group category with (G, m) = (Z3 × Z3, 9) via de- equivariantization (see Example 12.13). As a byproduct, we find new C∗ near-group categories with (G, m) = (Z2 × Z2 × Z3, 12) missing in [16, Table 2] (see Example 12.18). In subsection 13, we discuss equivariantization. For that purpose, we determined the structure of the automorphism groups of C∗ near-group categories in the irrational case in subsection 13.1. The author would like thank David Evans, Terry Gannon, Pinhas Grossman, Vaughan Jones, Zhengwei Liu, Scott Morrison, Sebastien Palcoux, David Penneys, Victor Ostrik, and Noah Snyder for stimulating discussions and encouragement. 2. Preliminaries Our basic references are [11] for tensor categories, [18] for operator algebras and subfactors, and [3] for the category of endomorphisms of von Neumann algebras. Every von Neumann algebra in this note is assumed to have separable predual. For a Hilbert space H, we denote by B(H) the set of bounded operators on H, and by U(H) the set of unitaries on H. The identity operator of H is denoted by IH or simply by I. For a unital C∗-algebra A, we denote by U(A) the set of unitaries in A. The unit of A is denoted by IA or simply by I. Let M be a properly infinite factor. Then the set of unital endomorphisms End(M) forms a tensor category with the monoidal product ρ⊗σ of two objects ρ, σ ∈ End(M) given by the composition ρ ◦ σ, and the morphism space from ρ to σ given by HomEnd(M )(ρ, σ) = {T ∈ M; T ρ(x) = σ(x)T, ∀x ∈ M}. For simplicity, we denote (ρ, σ) = HomEnd(M )(ρ, σ). monoidal product T1 ⊗ T2 of two morphisms Ti ∈ (ρi, σi), i = 1, 2, are given by In this tensor category, the T1ρ1(T2) = σ1(T2)T1. This is graphically expressed as ρ1 ρ2 ρ1 ρ2 T2 T1 T1 = T2 . σ1 σ2 σ1 σ2 By definition, two objects ρ, σ are equivalent if and only if there exists a unitary U ∈ U(M) satisfying ρ = Ad U ◦ σ, where Ad U is the inner automorphism of M 6 MASAKI IZUMI given by Ad U(x) = UxU−1. We denote by [ρ] the equivalence class of ρ. The self- morphism space (ρ, ρ) is nothing but the relative commutant M ∩ ρ(M)′, and when this space consists of only scalars, we say that ρ is irreducible (or simple). The morphism space (ρ, σ) inherits the Banach space structure from M, and the ∗-operation of M sends (ρ, σ) to (σ, ρ), which makes End(M) a C∗ tensor category (see [3, Section 1]). Moreover, if ρ is irreducible, the space (ρ, σ) is a Hilbert space with an inner product given by T ∗1 T2 = hT1, T2iIM for T1, T2 ∈ (ρ, σ). Throughout the paper, we assume that any functor between C∗ fusion categories preserves the ∗-structure. , where [M : ρ(M)]0 is the minimal index of ρ(M) in M. We denote by End0(M) the set of ρ ∈ End(M) with finite d(ρ). The dimension function End0(M) ∋ ρ 7→ d(ρ) is additive with respect to the direct sum operation and multiplicative with respect to the monoidal product operation. The tensor category End0(M) is rigid in the following sense: for any ρ ∈ End0(M), there exist ρ ∈ End0(M), called the conjugate endomorphism of ρ, and two isometries Rρ ∈ (id, ρ ◦ ρ), Rρ ∈ (id, ρ ◦ ρ) satisfying For ρ ∈ End(M), its dimension d(ρ) is defined by [M : ρ(M)]1/2 0 R∗ ρρ(Rρ) = R∗ρρ(Rρ) = 1 . d(ρ) If ρ is self-conjugate, either Rρ = Rρ or Rρ = −Rρ occurs. We say that ρ is real in the former case, and ρ is pseudo-real in the latter case. If we replace End(M) with the set of unital homomorphisms between two type III factors, the dimension function and conjugate morphisms still make sense, and we use the same notation as above (see [30], [3]). Every C∗ fusion category is realized as a category of bimodules of the hyperfinite II1 factor (see [25]), which implies the following statement by a tensor product trick. Theorem 2.1. Every C∗ fusion category is realized as a subcategory of End0(M) for any hyperfinite type III factor M. For uniqueness, we have the following statement, which is a consequence of Popa's classification theorem for amenable subfactors. Recall that a monoidal functor from a strict fusion category C to another strict fusion category D is a pair (F, L) consisting of a functor F : C → D and natural isomorphisms Lρ,σ ∈ HomD(F (ρ) ⊗ F (σ), F (ρ ⊗ σ)) satisfying Lρ⊗σ,τ ◦ (Lρ,σ ⊗ IF (τ )) = Lρ,σ⊗τ ◦ (IF (ρ) ⊗ Lσ,τ ) for any ρ, σ, τ ∈ C (see [11, Definition 2.4.1]). We may and do assume F (1C) = 1D and L1C,ρ = Lρ,1C = IF (ρ). When C and D are C∗ categories, we further assume that Lρ,σ is a unitary. Theorem 2.2. Let M and P be hyperfinite type III1 factors, and let C and D be C∗ fusion categories embedded in End(M) and End(P ) respectively. Let (F, L) be a monoidal functor from C to D that is an equivalence of the two C∗ fusion categories C THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 7 and D. Then there exists a surjective isomorphism Φ : M → P and unitaries Uρ ∈ P for each object ρ ∈ C satisfying F (ρ) = Ad Uρ ◦ Φ ◦ ρ ◦ Φ−1, F (X) = UσΦ(X)U∗ρ , X ∈ (ρ, σ), Lρ,σ = Uρ◦σΦ ◦ ρ ◦ Φ−1(U∗σ)U∗ρ = Uρ◦σU∗ρ F (ρ)(U∗σ). Before proving the statement, let us recall Popa's classification theorem in the special case of finite depth subfactors of the hyperfinite type III1 factor (see [48], [43]). Note that hyperfinite type III1 factors are mutually isomorphic due to Haagerup [24]. Let N ⊂ M be an inclusion of hyperfinite type III1 factors of finite index and finite depth, and let l be an integer larger than the depth of the inclusion. Let M ⊃ N = N0 ⊃ N1 ⊃ · · · ⊃ Nl be the downward basic construction. In each step, the subfactor Nk+1 is uniquely determined up to inner conjugacy in Nk. The standard invariant of N ⊂ M is deter- mined by the following nested system of finite dimensional von Neumann algebras: M ∩ N′ ⊂ M ∩ N′1 ⊂ · · · ⊂ M ∩ N′l ⊂ N ∩ N′1 ⊂ · · · ⊂ N ∩ N′l ∪ C ∪ ∪ . Popa showed that there is a continuation of the downward basic construction Nl ⊃ Nl+1 ⊃ Nl+2 ⊃ · · · so that M = M st ⊗ R and N = N st ⊗ R hold, where M st = (M ∩ N′k), N st = (N ∩ N′k), ∞_k=0 ∞_k=0 and R is the relative commutant of M st in M, which is hyperfinite of type III1. Since the core inclusion N st ⊂ M st is completely determined by the standard invariant, it classifies N ⊂ M. Now we recall the precise statement we need in the proof of Theorem 2.2. Assume that Q ⊂ P is another inclusion of hyperfinite type III1 factors with the same standard invariant as that of N ⊂ M. Let P ⊃ Q ⊃ Q1 ⊃ · · · ⊃ Ql be an arbitrary downward basic construction up to l step. Popa's theorem implies that if Φ is an isomorphism from M ∩ N′l onto P ∩ Q′l with Φ(M ∩ N′k) = P ∩ Q′k and Φ(N ∩ N′k) = Q ∩ Q′k for any 1 ≤ k ≤ l, it extends to an isomorphism from M onto P mapping N onto Q. Proof of Theorem 2.2. In the following arguments, whenever we apply the functor F to morphisms, we need a special care because the same operator may belong to different morphism spaces. For example, an operator X ∈ (ρ, σ) at the same time belongs to (ρ ◦ µ, σ ◦ µ) as it is identified with X ⊗ Iµ. On the other hand, we have F (X ⊗ Iµ) = Lσ,µ ◦ (F (X) ⊗ IF (µ)) ◦ L−1 ρ,µ. 8 MASAKI IZUMI Let Irr(C) be a complete system of representatives of the set of equivalence classes of irreducible endomorphisms in C. We may assume id ∈ Irr(C), and F (id) = id. We choose an object γ = Mξ∈Irr(C) ξ ∈ C. Then γ(M) ⊂ M is a finite depth subfactor. Since γ is self-conjugate, M ⊃ γ(M) ⊃ γ2(M) ⊃ γ3(M) is a downward basic construction, and the standard invariant of the inclusion γ(M) ⊂ M is determined by (γ, γ) ⊂ (γ2, γ2) ⊂ (γ3, γ3) ∪ ⊂ γ((γ, γ)) ⊂ γ((γ2, γ2)) C ∪ ∪ In the tensor category language, this is expressed as . EndC(γ) ⊗ Iγ ⊗ Iγ ⊂ EndC(γ ⊗ γ) ⊗ Iγ ⊂ EndC(γ ⊗ γ ⊗ γ) ⊂ Iγ ⊗ EndC(γ) ⊗ Iγ ⊂ Iγ ⊗ EndC(γ ⊗ γ) CIγ ⊗ Iγ ⊗ Iγ ∪ ∪ ∪ . Let or in the tensor category language, Lγ,γ,γ := Lγ2,γLγ,γ = Lγ,γ2F (γ)(Lγ,γ), Lγ,γ,γ = Lγ⊗γ,γ ◦ (Lγ,γ ⊗ IF (γ)) = Lγ,γ⊗γ ◦ (IF (γ) ⊗ Lγ,γ). Then we have Ad Lγ,γ,γ ◦ F (γ)3 = F (γ3), and Ad L∗γ,γ,γ induces an isomorphism from (F (γ3), F (γ3)) onto (F (γ)3, F (γ)3). We denote by Φ0 the composition of F restricted to (γ3, γ3) and Ad L∗γ,γ,γ, which is an isomorphism from (γ3, γ3) onto (F (γ)3, F (γ)3). We claim that Φ0 induces an isomorphism of the standard invariants of γ(M) ⊂ M and F (γ)(P ) ⊂ P . Indeed, for Y ∈ (γ2, γ2), we have Φ0(Y ⊗ Iγ) = (L∗γ,γ ⊗ IF (γ))L∗γ⊗γ,γF (Y ⊗ Iγ)Lγ⊗γ,γ(Lγ,γ ⊗ IF (γ)) Φ0(Iγ ⊗ Y ) = (IF (γ) ⊗ L∗γ,γ)L∗γ,γ⊗γF (Iγ ⊗ Y )Lγ,γ⊗γ(IF (γ) ⊗ Lγ,γ) = L∗γ,γF (Y )Lγ,γ ⊗ IF (γ), = IF (γ) ⊗ L∗γ,γF (Y )Lγ,γ. In the same way, for X ∈ EndC(γ), we have Φ0(X ⊗ Iγ ⊗ Iγ) = F (X) ⊗ IF (γ) ⊗ IF (γ), Φ0(Iγ ⊗ X ⊗ Iγ) = IF (γ) ⊗ F (X) ⊗ IF (γ), Φ0(Iγ ⊗ Iγ ⊗ X) = IF (γ) ⊗ IF (γ) ⊗ F (X), which shows the claim. Thus Φ0 extends to an isomorphism Φ from M onto P satisfying Φ(γ(M)) = F (γ)(P ). By construction and the above computation, we have Φ(Y ) = L∗γ,γF (Y )Lγ,γ and Φ(γ(Y )) = F (γ)(Φ(Y )) for Y ∈ (γ2, γ2), and we have Φ(X) = F (X) and Φ(γ(X)) = F (γ)(Φ(X)) for X ∈ (γ, γ). THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 9 We choose an isometry Vξ ∈ (ξ, γ) for each ξ ∈ Irr(C). Then we have γ(x) = Xξ∈Irr(C) Vξξ(x)V ∗ξ , Since VξV ∗ξ ∈ (γ, γ), we get Φ(VξV ∗ξ ) = F (VξV ∗ξ ) = F (Vξ)F (Vξ)∗, and Φ(Vξ) and F (Vξ) are isometries with the same range projection. Thus there exists a unitary Wξ ∈ U(P ) for each ξ ∈ Irr(C) satisfying Φ(Vξ) = F (Vξ)Wξ. ing Φ ◦ γ = F (γ) ◦ ϕ. On one hand, Since Φ(γ(M)) = F (γ)(P ), there exists an isomorphism ϕ from M onto P satisfy- F (Vξ)F (ξ)(ϕ(x))F (Vξ)∗, F (γ) ◦ ϕ(x) = Xξ∈Irr(C) Φ(Vξ)Φ ◦ ξ(x)Φ(Vξ)∗ = Xξ∈Irr(C) and on the other hand, Φ ◦ γ(x) = Xξ∈Irr(C) F (Vξ)WξΦ ◦ ξ(x)W ∗ξ F (Vξ)∗. This implies F (ξ) ◦ ϕ = Ad Wξ ◦ Φ ◦ ξ, and in particular ϕ = Ad Wid ◦ Φ in the case with ξ = id. Thus we obtain F (ξ) = Ad Uξ ◦ Φ ◦ ξ ◦ Φ−1 with Uξ = WξΦ ◦ ξ ◦ Φ−1(W ∗id) = F (ξ)(W ∗id)Wξ. Let Z ∈ (γ, γ2). Then Y1 = ZV ∗id belongs to (γ2, γ2), and F (Y1) = Lγ,γΦ(Y1)L∗γ,γ. Since Y1 should be interpreted as Z ◦ (V ∗id⊗ Iγ), the left-hand side is F (Z)F (Vid)∗L∗γ,γ. Thus we get F (Z) = Lγ,γΦ(Z)Φ(Vid)∗F (Vid) = Lγ,γΦ(Z)W ∗id. On the other hand, the operator Y2 = Zγ(V ∗id) belongs to (γ2, γ2), and we have F (Y2) = Lγ,γΦ(Y2)L∗γ,γ. Since Y2 should be interpreted as Z ◦ (Iγ ⊗ V ∗id), the left hand side is F (Z)F (γ)(Vid)∗L∗γ,γ, and we get F (Z) = Lγ,γΦ(Z)Φ(γ(Vid)∗)F (γ(Vid)) = Lγ,γΦ(Z)F (γ)(WidΦ(Vid)∗W ∗id)F (γ)(F (Vid)) = Lγ,γΦ(Z)F (γ)(F (Vid)∗W ∗idF (Vid)), and Setting Z = Vid and multiplying the both sides by Φ(Vid)∗ from left, we get Φ(Z)F (γ)(F (Vid)∗W ∗idF (Vid)) = Φ(Z)W ∗id. F (γ)(F (Vid)∗WidF (Vid)) = Wid, which implies (2.1) F (Vη)∗WidF (Vη) = F (η)(F (Vid)∗WidF (Vid)). Let ξ, η, ζ ∈ Irr(C), and let X ∈ (ζ, ξ◦ η). Since Φ(X) ∈ (Φ◦ ζ ◦ Φ−1, Φ◦ ξ◦ η◦ Φ−1), Φ ◦ ζ ◦ Φ−1 = Ad U∗ζ ◦ F (ζ) and Φ ◦ ξ ◦ η ◦ Φ−1 = Ad(U(ξ)∗F (ξ)(U(η))∗) ◦ F (ξ) ◦ F (η), 10 we have MASAKI IZUMI F (ξ)(Uη)UξΦ(X)U∗ζ = UξΦ ◦ ξ ◦ Φ−1(Uη)Φ(X)U∗ζ ∈ (F (ζ), F (ξ) ◦ F (η)). We claim that this coincides with L∗ξ,ηF (X). Indeed, since Z = γ(Vη)VξXV ∗ζ ∈ (γ, γ2), we have F (Z) = Lγ,γΦ(Z)F (γ)(F (Vid)∗W ∗idF (Vid)). Since Z should be under- stood as (Iγ ⊗ Vη) ◦ (Vξ ⊗ Iη) ◦ X ◦ V ∗ζ , the left-hand side is (Lγ,γF (γ)(F (Vη))L∗γ,η)(Lγ,ηF (Vξ)L∗ξ,η)F (X)F (Vζ)∗ = Lγ,γF (γ)(F (Vη))F (Vξ)L∗ξ,ηF (X)F (Vζ)∗. The right-hand side is Lγ,γΦ(γ(Vη))Φ(Vξ)Φ(X)Φ(Vζ)∗F (γ)(F (Vid)∗W ∗idF (Vid)) = Lγ,γF (γ)(WidΦ(Vη)W ∗id)F (Vξ)WξΦ(X)W ∗ζ F (Vζ)∗F (γ)(F (Vid)∗W ∗idF (Vid)) = Lγ,γF (γ)(WidF (Vη)WηW ∗id)F (Vξ)WξΦ(X)W ∗ζ F (Vζ)∗F (γ)(F (Vid)∗W ∗idF (Vid)) Thus L∗ξ,ηF (X) = F (ξ)(F (Vη)∗WidF (Vη)WηW ∗id)WξΦ(X)W ∗ζ F (ζ)(F (Vid)∗W ∗idF (Vid)) = F (ξ)(F (η)(F (Vid)∗WidF (Vid))WηW ∗id)WξΦ(X)W ∗ζ F (ζ)(F (Vid)∗W ∗idF (Vid)) = F (ξ) ◦ F (η)(F (Vid)∗WidF (Vid)Wid)F (ξ)(Uη)UξΦ(X)U∗ζ × F (ζ)(W ∗idF (Vid)∗W ∗idF (Vid)) = F (ξ)(Uη)UξΦ(X)U∗ζ , where we used Eq.(2.1), and F (ξ)(Uη)UξΦ(X)U∗ζ ∈ (F (ζ), F (ξ) ◦ F (η)). This shows that the claim holds. of (ξ, ρ) For any object ρ ∈ C, we choose an orthonormal basis {V (ξ, ρ)i}dim(ξ,ρ) i=1 and define a unitary Uρ ∈ U(P ) by Uρ =Xξ,i F (V (ξ, ρ)i)UξΦ(V (ξ, ρ)i)∗. Then it does not depend on the choice of the orthonormal basis, and coincides with the previous definition if ρ ∈ Irr(C). It is straightforward to show F (ρ) = Ad Uρ ◦ Φ ◦ ρ ◦ Φ−1. Let ρ, σ ∈ C, and let X ∈ (ρ, σ). For ξ, η ∈ Irr(C), we have V (ξ, σ)∗i XV (η, ρ)j ∈ (η, ξ), which vanishes if ξ 6= η. Since ξ is irreducible, the restriction of Φ on (ξ, ξ) = CIρ coincides with that of F on (ξ, ξ), and F (V (ξ, σ)i)∗F (X)F (V (η, ρ)j) = δξ,ηΦ(V (ξ, σ)∗i )Φ(X)Φ(V (η, ρ)j). THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 11 Thus F (X) = Xξ,η∈Irr(C)Xi,j = Xξ∈Irr(C)Xi,j = Xξ∈Irr(C)Xi,j = UσΦ(X)U∗ρ . F (V (ξ, σ)i)F (V (ξ, σ)i)∗F (X)F (V (η, ρ)j)F (V (η, ρ)j)∗ F (V (ξ, σ)i)Φ(V (ξ, σ)∗i )Φ(X)Φ(V (ξ, ρ)j)F (V (ξ, ρ)j) F (V (ξ, σ)i)UξΦ(V (ξ, σ)∗i )Φ(X)Φ(V (ξ, ρ)j)U∗ξ F (V (ξ, ρ)j) Let ζ ∈ Irr(C). Then {V (ξ, ρ)iξ(V (η, σ)j)V (ζ, ξ◦η)k}ξ,η,i,j,k is an orthonormal basis of (ζ, ρ ◦ σ), and we have Uρ◦σ = Xξ,η,ζ∈Irr(C)Xi,j,k F (V (ξ, ρ)iξ(V (η, σ)j)V (ζ, ξ ◦ η)k) × UζΦ(V (ζ, ξ ◦ η)∗kξ(V (η, σ)j)∗V (ξ, ρ)∗i ). Here V (ξ, ρ)iξ(V (η, σ)j)V (ζ, ξ ◦ η)k should be interpreted as (V (ξ, ρ)i ⊗ Iσ) ◦ (Iξ ⊗ V (η, σ)j) ◦ V (ζ, ξ ◦ η)k, and we have F (V (ξ, ρ)iξ(V (η, σ)j)V (ζ, ξ ◦ η)k) = (Lρ,σF (V (ξ, ρ)i)L∗ξ,σ)(Lξ,σF (ξ)(F (V (η, σ)j))L∗ξ,η)F (V (ζ, ξ ◦ η)k) = Lρ,σF (V (ξ, ρ)i)F (ξ)(F (V (η, σ)j))L∗ξ,ηF (V (ζ, ξ ◦ η)k). Note that we have already seen L∗ξ,ηF (V (ζ, ξ ◦ η)k) = F (ξ)(Uη)UξΦ(V (ζ, ξ ◦ η)k)U∗ζ . Thus F (V (ξ, ρ)i)F (ξ)(F (V (η, σ)j))F (ξ)(Uη)UξΦ(V (ζ, ξ ◦ η)k)U∗ζ F (V (ξ, ρ)i)F (ξ)(F (V (η, σ)j)UηΦ(V (η, σ)j)∗)UξΦ(V (ξ, ρ)∗i ) Uρ◦σ = Lρ,σ Xξ,η,ζ∈Irr(C)Xi,j,k × UζΦ(V (ζ, ξ ◦ η)k)∗U∗ξ F (ξ)(Φ(V (η, σ)j))∗UξΦ(V (ξ, ρ)∗i ) = Lρ,σ Xξ,η∈Irr(C)Xi,j = Lρ,σ Xξ∈Irr(C)Xi F (ρ)(Uσ)F (V (ξ, ρ)i)UξΦ(V (ξ, ρ)∗i ) = Lρ,σF (ρ)(Uσ)Uρ. This finishes the proof. (cid:3) Applying the above theorem to the case with M = P and C = D, we obtain the following statement. 12 MASAKI IZUMI Corollary 2.3. Let M be a hyperfinite type III1 factor, let C ⊂ End0(M) be a C∗ fusion category. Then up to a natural isomorphism, every automorphism of C is it is given by ρ 7→ induced by an automorphism Φ of M in the following sense: Φ ◦ ρ ◦ Φ−1 for an object ρ and by X 7→ Φ(X) for a morphism X. One of the main tools in this note is the Cuntz algebra On, and we summarize the main feature of it here. Let n be an integer larger than 1. The Cuntz algebra On is the universal C∗-algebra with generators {Si}n S∗i Sj = δi,jI, i=1 and relations SiS∗i = I. nXi=1 The most peculiar property of the Cuntz algebra is that it is at the same time universal and simple (see [6]). Therefore if {Ti}n i=1 are noncommutative polynomials of the generators obeying the same relation as the defining relation, then there exists a unique endomorphism σ ∈ End(On) satisfying σ(Si) = Ti. 3. Basic ingredients Let G be a finite group of order n. We would like to classify a C∗ near group cat- egory C with group G and the multiplicity parameter m. Throughout this note, we assume that G is not trivial and m 6= 0 because the two cases are completely under- stood as we mentioned in Introduction. In this section, we deduce basic ingredients to determine the structure of such C. Thanks to Theorem 2.1 and Theorem 2.2, we may assume that C is a subcategory of End(M) where M is the hyperfinite type III1 factor, and C is generated by a single irreducible endomorphism ρ ∈ End0(M) satisfying the following fusion rules: [ρ]2 =Mg∈G [αg] ⊕ m[ρ], [αg][αh] = [αgh], [αg][ρ] = [ρ][αg] = [ρ], where the map α : G → Aut(M) induces an injective homomorphism from G into Out(M). Since [αg][ρ] = [ρ], we can arrange α so that αg ◦ ρ = ρ holds for all g ∈ G. Then we have αg ◦ αh = αgh, that is, the map α is an action of G on M, and in particular, the pointed subcategory generated by αG has trivial third cohomology. Indeed, from the fusion rules, there exists a unitary U(g, h) ∈ M satisfying αg ◦ αh = Ad U(g, h) ◦ αgh. On the other hand, we have ρ = αg ◦ αh ◦ ρ = Ad U(g, h) ◦ αgh ◦ ρ = Ad U(g, h) ◦ ρ. Since ρ is irreducible, the unitary U(g, h) is a scalar, and we get the claim. We set d = d(ρ) = m + √m2 + 4n , 2 which is the dimension of ρ satisfying d2 = n + md. Throughout this note, we keep using the symbols G, m, n, d in this sense. THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 13 We fix an isometry Se ∈ (id, ρ2). Since ρ is self-conjugate, we have (3.1) S∗e ρ(Se) = ǫ d , ǫ ∈ {1,−1}. When ǫ = 1 (resp. Graphically, we have ǫ = −1), we say that ρ is a real (resp. pseudo-real) sector. √dSe = ρ ǫ√dS∗e = , ρ . We set Sg = αg(Se). Then (αg, ρ2) = CSg. Let K = (ρ, ρ2), and let {Ti}m i=1 be an i=1 satisfies the Cuntz algebra On+m orthonormal basis of K. Then {Sg}g∈G ∪ {Ti}m relation, and in particular, (3.2) Xg∈G We set P =Pg∈G SgS∗g and Q =Pm mXi=1 SgS∗g + TiT ∗i = I. i=1 TiT ∗i , which are projections. Let KsK∗t be the linear span of {Ti1Ti2 · · · TisT ∗jtT ∗jt−1 · · · T ∗j1}. We identify Ks with K⊗s via the identification of Ti1Ti2 · · · Tis with Ti1 ⊗ Ti2 ⊗ · · · ⊗ Tis. We identify KsK∗t with B(Kt,Ks) by left multiplication, and K2K2∗ with B(K) ⊗ B(K). For example, we denote Ti1T ∗j1 ⊗Ti2T ∗j2 = Ti1Ti2T ∗j2T ∗j1. We abuse this notation and denote TiT ∗j ⊗ x = TixT ∗j for any x ∈ On+m. With this notation, we have (3.3) (ρ2, ρ2) =Mg∈G CSgS∗g ⊕ B(K). Since (ρ, ρ ◦ αg) ⊂ (ρ2, ρ2), we can choose a (a priori projective) unitary represen- tation {U(g)}g∈G in (ρ2, ρ2) such that (ρ, ρ ◦ αg) = CU(g). Since U(g)Se ∈ (id, ρ2), we can normalize U(g) so that U(g)Se = Se holds. Then {U(g)} is a genuine repre- sentation of the form U(g) =Xh∈G χh(g)ShS∗h + UK(g), (3.4) (3.5) (3.6) where χh ∈ Hom(G, T) and {UK(g)}g∈G is a unitary representation of G in B(K). Since αg ◦ ρ = ρ, we have αg(K) = K, and there exists a unitary representation {V (g)}g∈G in B(K) such that αg(T ) = V (g)T for all T ∈ K and g ∈ G. For T ∈ K, we set j1(T ) = √dT ∗ρ(Se) ∈ K, j2(T ) = √dρ(T )∗Se ∈ K. Then the Frobenius reciprocity ([30]) implies O O 14 MASAKI IZUMI Lemma 3.1. The maps j1 and j2 are anti-linear isometries of K satisfying (3.7) 1 = j2 j2 2 = ǫ, (3.8) (3.9) V (g)j1 = j1V (g), UK(g)j2 = j2V (g), g ∈ G, g ∈ G. In particular, the two unitary representations UK and V are unitarily equivalent with an intertwining unitary j2 ◦ j−1 1 . Remark 3.2. For those readers who would like to reproduce our arguments in this paper without assuming the C∗ condition on near-group categories, we briefly give graphical expressions of the intertwiners appearing so far. Let C be a pivotal near- group category whose simple objects are G∪{ρ}. We assume that C is strict. We first choose and fix a non-zero homomorphism Se ∈ Hom(id, ρ⊗ρ) and isomorphisms fg ∈ Hom(g⊗ρ, ρ) for g ∈ G. Then there exist unique isomorphisms mg,h ∈ Hom(g⊗h, gh) for g, h ∈ G to make the following diagrams commutative: g ⊗ h ⊗ ρ Ig⊗fh−−−→ g ⊗ ρ mg,h⊗Iρy gh ⊗ ρ −−−→fgh . yfg ρ With this family {mg,h}g,h∈G, we can show that the following diagrams are commu- tative, Ig⊗mh,k −−−−−→ g ⊗ gh g ⊗ h ⊗ k mg,h⊗Iky gh ⊗ k −−−→mgh,k , ymg,hk ghk and in consequence, we can see that the group part has trivial third cohomology. The homomorphism Sg ∈ Hom(g, ρ ⊗ ρ) is given by g ρ Sg = (fg ⊗ Iρ) ◦ (Ig ⊗ Se) = fg ρ Se zttttt . ρ We choose S∗e ∈ Hom(ρ ⊗ ρ, id) satisfying S∗e ◦ Se = 1, and set S∗g = (Ig ⊗ S∗e ) ◦ (f−1 g ⊗ Iρ) ∈ Hom(ρ ⊗ ρ, g). Then we have S∗g ◦ Sg = Ig. z THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 15 Setting V (g) : Hom(ρ, ρ⊗ρ) ∋ T 7→ (fg⊗Iρ)◦(Ig⊗T )◦f−1 g = g ρ f−1 g ρ T ✑✑✑✑✑ ρ ✬✬✬✬✬✬✬✬✬✬✬✬ ρ fg ρ ∈ Hom(ρ, ρ⊗ρ), we get a representation {V (g)}g∈G of G on Hom(ρ, ρ ⊗ ρ). condition The homomorphism U(g) ∈ Hom(ρ, ρ ⊗ g) is determined by the normalization Se ✞✞✞ Se = (Iρ ⊗ fg) ◦ (U(g) ⊗ Iρ) ◦ Se = U(g) $❏❏❏❏❏ fg ✯✯✯✯✯✯✯✯✯ ρ ρ g . Then they satisfy the following relation ρ ρ (Iρ ⊗ mg,h) ◦ (U(g) ⊗ Ih) ◦ U(h) = U(gh), ρ = U(gh) . ρ U(h) ✞✞✞ ρ U(g) $❏❏❏❏❏ ✯✯✯✯✯✯✯✯✯ g h mg,h ✶✶✶✶✶✶✶✶✶✶ gh Using this, we can see that ρ gh ρ ρ G ∋ g 7→ (Iρ ⊗ fg) ◦ (U(g) ⊗ Iρ) = U(g) g ❄❄❄❄❄ fg ρ ∈ Hom(ρ ⊗ ρ, ρ ⊗ ρ) gives a representation of G. It is easy to show ρ ρ (1 ⊗ S∗e ) ◦ (Iρ ⊗ fg ⊗ Iρ) ◦ (U(g) ⊗ Iρ ⊗ Iρ) ◦ (Iρ ⊗ Se) = δg,eIρ, $ $ 16 MASAKI IZUMI ρ Se U(g) g ❄❄❄❄❄ ρ ✒✒✒✒✒✒✒✒✒✒ #●●●●●● S∗e ρ fg ρ = δg,e , ρ ρ which is essentially the right categorical trace of (1 ⊗ fg) ◦ (U(g) ⊗ Iρ). Choosing a basis {Ti}m i=1 of Hom(ρ ⊗ ρ, ρ) satisfying T ∗i ◦ Tj = δi,jIρ, we have i=1 of Hom(ρ, ρ ⊗ ρ) and a basis {T ∗i }m Iρ⊗ρ =Xg∈G Sg ◦ S∗g + Ti ◦ T ∗i . mXi=1 Although neither j1 nor j2 can be defined without C∗ condition, they can be replaced by linear maps from Hom(ρ⊗ρ, ρ) to Hom(ρ, ρ⊗ρ) given by the Frobenius reciprocity. We get back to our original situation with a C∗ near-group category C realized inside End(M). Lemma 3.3. For any g ∈ G, we have (3.10) ρ(Se) = ǫ dXh∈G Sh + 1 √d mXi=1 Tij1(Ti), (3.11) Proof. The second statement follows from Ad U(g) ◦ ρ = ρ ◦ αg and Sg = αg(Se), From Eq.(3.2), we obtain ρ(Sg) = U(g)ρ(Se)U(g)∗. ShS∗hρ(Se) + αh(SeS∗e ρ(Se)) + ρ(Se) =Xh∈G dXh∈G = ǫ αh(Se) + mXi=1 TiT ∗i ρ(Se) =Xh∈G mXi=1 dXh∈G Tij1(Ti) = ǫ 1 √d 1 √d mXi=1 Tij1(Ti) Sh + 1 √d mXi=1 Tij1(Ti). (cid:3) Lemma 3.4. There exists a unique linear map l : K → K2K∗ such that for any T ∈ K (3.12) 1 (T ))ShS∗h + l(T ). αh(j2 ◦ j−1 The maps j1, j2 and l satisfy the following for all T, T ′ ∈ K: (3.13) Shαh(j2(T )∗) +Xh∈G √dXh∈G αg(l(T )) = l(T ), ρ(T ) = 1 g ∈ G, # THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 17 1 dXh∈G (3.14) (3.15) (3.16) (3.17) (3.18) ǫ j2(T )∗V (g)∗ + dXg∈G V (h)j2(T ′)j2(T )∗V (h)∗ + l(T ′)∗l(T ) = hT, T ′i j1(Ti)∗T ∗i l(T ) = 0, mXi=1 TiT ∗i , mXi=1 l(j1(T )) = l(T )∗TiTjj1(Tj)T ∗i , mXi,j=1 mXi=1 (j2 ◦ j1)3 = I. l(j2(T )) = Til(T )∗j1(Ti), Proof. To show that ρ is of the above form, we determine P ρ(T ) and Qρ(T ) sepa- rately. For P ρ(T ), we have P ρ(T ) =Xh∈G ShS∗hρ(T ) =Xh∈G √dXh∈G = 1 Shj2(T )∗V (h)∗. Shαh(S∗e ρ(T )) = 1 √dXh∈G Shαh(j2(T )∗) Since K∗ρ(K) ⊂ (ρ2, ρ2), there exist linear maps lh : K → K for h ∈ G and l : K → K2K∗ such that lh(T )ShS∗h + l(T ). Qρ(T ) =Xh∈G We compute ρ(j1(T )) by using the definition of j1: Since αg ◦ ρ = ρ, we obtain lh(T ) = αh(le(T )) and αg(l(T )) = l(T ). ShShS∗h + √dρ(T )∗ Tiρ(Se)T ∗i mXi=1 hTi, V (h)le(T )iShS∗hρ(Se)T ∗i αh(ρ(T )∗Se)ShS∗h + √d ρ(j1(T )) = √dρ(T ∗ρ(Se)) = √dρ(T )∗Xh∈G mXi=1Xh∈G = √dXh∈G mXi=1 + √d =Xh∈G mXi,j=1 √dXh∈G l(T )∗TiTjj1(Tj)T ∗i . αh(j2(T ))ShS∗h + l(T )∗Tiρ(Se)T ∗i + ǫ Shle(T )∗V (h)∗ This shows le = j2 ◦ j−1 1 and Eq.(3.16). ǫ ǫ = = Shαh(T ∗) + Shαh(T ∗) + ρ(j2(T )) = √dρ(ρ(T )∗Se) = mXi=1 √dXh∈G mXi=1Xh∈G √dXh∈G mXi=1 √dXh∈G mXi=1 Shαh(T ∗) +Xh∈G Til(T )∗j1(Ti). Til(T )∗j1(Ti) + = + ǫ hj1(Ti), V (h)j2 ◦ j−1 1 (T )iTiShS∗h V (h)j−1 1 ◦ j2 ◦ j−1 1 (T )ShS∗h 18 MASAKI IZUMI We compute ρ(j2(T )) by using the definition of j2: ǫ √dXh∈G ρ2(T ∗)Sh + mXi=1 ρ2(T ∗)Tij1(Ti) Tiρ(T )∗j1(Ti) 1 ◦ j2 = j−1 1 ◦ j2 ◦ j−1 Remark 3.5. Let l(1) This implies Eq.(3.17), and j2 ◦ j−1 1 , which shows Eq.(3.18). Eq.(3.14) and (3.15) follow from ρ(Se)∗ρ(T ) = 0 and ρ(T ′)∗ρ(T ) = hT, T ′iI. i,j=1 l(1) ij ∈ B(K) be linear maps defined by l(T ) =Pm ij (T )TiT ∗j ij (T )T ∗j respectively. Then Eq.(3.16) and Eq.(3.17) are equiv- 1 (l(1) ij , l(2) i,j=1 Til(2) ij (j2(T )) = j−1 and l(T ) =Pm ij (j1(T )) = j1(l(2) ji (T )) respectively. ji (T )) and l(2) alent to l(1) (cid:3) Remark 3.6. The above lemma shows that every morphism between objects gener- ated by ρ and αg are noncommutative polynomials of the Cuntz algebra generators {Sg}g∈G ∪ {Ti}m i=1 and their adjoints, and the structure of C, or more precisely the 6j symbols of C, are completely determined by the tuple (K, j1, j2, V, UK, χ, l). To obtain this tuple from C, we made the following choices: (1) the representative ρ from the class [ρ], (2) the parametrization of the group {[αg]}g∈G, (3) the choice of Se from TSe. A different choice in (1) only ends up with a unitarily equivalent tuple in an appro- priate sense, and that in (2) allows us to insert an group automorphism of G in the variables of V, UK, χ. Replacing Se with ωSe results in replacing j1 and j2 with ωj1 and ωj2 respectively, which is a special case of unitary equivalence. In terms of U(g), V (g), and l, the relation ρ(αg(T )) = U(g)ρ(T )U(g)∗ is expressed as follows: Lemma 3.7. Let the notation be as above. Then (3.19) (3.20) UK(g)V (h) = χh(g)V (h)UK(g), l(V (g)T ) = U(g)l(T )U(g)∗, g ∈ G, T ∈ K. THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 19 Proof. We compute ρ(αg(T )) first: ρ(αg(T )) = + l(V (g)T ) 1 √dXh∈G √dXh∈G 1 Shj2(V (g)T )∗V (h)∗ +Xh∈G Shj2(T )∗UK(g)∗V (h)∗ +Xh∈G = + l(V (g)T ). On the other hand, V (h)j2 ◦ j−1 1 (V (g)T )ShS∗h V (h)UK(g)j2 ◦ j−1 1 (T )ShS∗h U(g)ρ(T )U(g)∗ = U(g)Shj2(T )∗V (h)∗UK(g)∗ 1 √dXh∈G +Xh∈G UK(g)V (h)j2 ◦ j−1 √dXh∈G +Xh∈G χh(g)UK(g)V (h)j2 ◦ j−1 = 1 χh(g)Shj2(T )∗V (h)∗UK(g)∗ which shows the statement. Corollary 3.8. For g, h ∈ G, the following hold: χh(g) = χg(h), (3.21) 1 (T )ShS∗hU(g)∗ + UK(g)l(T )UK(g)∗ 1 (T )ShS∗h + UK(g)l(T )UK(g)∗, (cid:3) (3.22) Proof. From Lemma 3.1, we have j2V (g)j∗2 = j∗2V (g)j2 = UK(g). Thus Eq.(3.19) implies αh(U(g)) = χh(g)U(g). and so j2UK(g)j∗2j2V (h)j∗2 = χh(g)j2V (h)j∗2 j2UK(g)j∗2, which shows χh(g) = χg(h). Using this, we obtain the second statement from V (g)UK(h) = χh(g)UK(h)V (g), αh(U(g)) =Xk∈G =Xk∈G =Xk∈G χk(g)αh(SkS∗k) + αh(UK(g)) χk(g)ShkS∗hk + V (h)UK(g)V (h)∗ χh−1k(g)SkS∗k + χh(g)UK(g). (cid:3) We denote by λ the left regular representation of G. 20 MASAKI IZUMI Theorem 3.9. When d = m+√m2+4n m is a multiple of n, and 2 is irrational, then the group G is always abelian, (3.23) (3.24) χh ∼= λ, Mh∈G V ∼= UK ∼= m n λ. When d is rational, there exist two natural numbers s and t such that n = st2, m = (s − 1)t, and d = st. Moreover, (i) When t = 1, the group G is abelian. In this case, the character χh is trivial for all h ∈ G and (3.25) 1 ⊕ V ∼= 1 ⊕ UK ∼= λ. (ii) When t > 1, the group G is non-abelian. In this case, the order of Hom(G, T) is t2 and (3.26) Mh∈G χh ≡ s Mχ∈Hom(G,T) χ. Let G† be the set of equivalence classes of irreducible unitary representations of G whose dimensions are greater than 1. Then t divides dim π for all π ∈ G†, and (3.27) V ∼= UK ∼= Mπ∈ G† dim π t π. Proof. Since we have δg,e = ρ(Se)∗ρ(Sg) = ρ(Se)∗U(g)ρ(Se)U(g)∗ for g ∈ G, we have δg,e = ρ(Se)∗U(g)ρ(Se) = = = 1 d2Xh∈G d2Xh∈G 1 χh(g) + χh(g) + 1 d 1 d χh(g) + 1 d mXi,j=1 j1(Tj)∗T ∗j UK(g)Tij1(Ti) hUK(g)Ti, Tjihj1(Ti), j1(Tj)i 1 d2Xh∈G mXi,j=1 Tr UK(g). This implies (3.28) Tr λ(g) = n d2Xh∈G χh(g) + n d Tr UK(g). For χ ∈ Hom(G, T), we denote by aχ and bχ the multiplicities of χ inLh∈G χh and UK respectively. THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 21 First, we assume that d is irrational. Then Eq.(3.28) implies d2 = naχ + nbχd for χ ∈ Hom(G, T). Since d2 = n + md, we obtain aχ = 1 and bχn = m. This implies Mh∈G χh ∼= Mχ∈Hom(G,T) χ, and so G is abelian. Assume now that d is rational. Then d is an integer. The inclusion M G ⊃ ρ(M) is of finite depth, and its index is d2/n = 1 + md/n. Since the index of a finite depth inclusion is an algebraic integer, the number 1 + md/n is indeed an integer, which we denote by s. Then d2 = n + md implies n(s − 1)2 = sm2, and so m is a multiple of s − 1. Letting t = m/(s − 1), we get n = st2 and d = st. Now Eq.(3.28) is of the form (3.29) Tr λ(g) = χh(g) + t Tr UK(g). 1 sXh∈G Assume that t = 1. Then n = d = s and m = s − 1. Let χ ∈ Hom(G, T). Since 0 ≤ aχ ≤ s and Eq.(3.28) implies 1 = aχ/n + bχ, either aχ = 0 and bχ = 1, or aχ = n and bχ = 0 hold. Since χe = 1, we get a1 = s, which implies that χh = 1 for all h ∈ G. Thus we get Tr λ(g) = 1 + Tr UK(g), and 1 ⊕ V ∼= 1 ⊕ UK ∼= λ. Let v0(g) = V (g), v1(g) = UK(g), and w = j1j2. Since χh = 1 for any h ∈ G, Eq(3.19) implies that v0 and v1 commute with each other. Moreover, since w3 = 1 and v1(g) = w∗v0(g)w, if we define v2(g) by w∗v1(g)w, the three representations v0, v1, and v2 commute with each other. Since 1 ⊕ vi ∼= λ for i = 0, 1, 2, the group G is abelian. Indeed, since the dimension of the commutant v0(G)′ of v0(G) is n − 1, we can see that v1(G)′′ is the commutant of v0(G)′′. Thus v0(G)′ ∩ v1(G)′ coincides with the center of v0(G)′′. Since v2 is a faithful representation of G in v0(G)′ ∩ v1(G)′, we conclude that G is abelian. Assume t > 1 now. Let π be an irreducible representation of G contained in UK. Then Eq.(3.29) implies dim π ≥ t, and so G is non-abelian. The rest of the statements in (ii) follow from a similar reasoning as above. Remark 3.10. Under naive identification of ρ and ρ, the map j2 ◦ j1 would be graph- ically expressed as (cid:3) ρ ρ j2 ◦ j1 : T ✹✹✹✹✹✹✹✹✹ ρ ✡✡✡✡✡✡✡✡✡ ρ 7→ ǫT . ρ ρ This means that 360◦ rotation in this picture ends up with multiplying by ǫ. To avoid this awkward convention in the case of ǫ = −1, we need to properly take the pivotal structure into account to define rotation by 120◦, which should be j2 ◦ j1 instead O O 22 MASAKI IZUMI of j2 ◦ j−1 1 = ǫj2 ◦ j1. In fact, we can deduce Eq.(3.18) from [44, Theorem 5.1] by identifying (ρ, ρ2) with (id, ρ3). Since we don't not rely on graphical calculus at all in this paper, instead of seriously pursuing it, we give a short and general argument, based on Longo's observation in [42], giving another proof of (j2 ◦ j1)3 = I here. Let φρ be the left inverse of ρ defined by φρ(x) = S∗e ρ(x)Se. Following the notation in [44], we denote by E(n) : (id, ρn) → (id, ρn) the restriction of dφρ. We first claim that E(n) is a unitary. Indeed, for W1, W2 ∈ (id, ρn), we have ρ ρ hE(n) ρ W1, E(n) ρ W2i = d2S∗e ρ(W2)∗SeS∗e ρ(W1)Se. Since ρ(W2)∗SeS∗e ρ(W1) ∈ (ρ, ρ) is already a scalar, we have hE(n) ρ W1, E(n) ρ W2i = d2φρ(ρ(W2)∗SeS∗e ρ(W1)) = d2W ∗2 φρ(SeS∗e )W1 = hW1, W2i. As was already observed in [39], the n-th power of E(n) ρ is positive for h(E(n) ρ )nW, Wi = dnW ∗φn ρ (W ) = dnφρ(ρn(W ∗)W ) = dnφn ρ (W W ∗) ≥ 0. Thus (E(n) Then Φ−1 = dS∗e ρ(·), and ρ )n = id. Let Φ : (ρ, ρ2) → (id, ρ3) be a unitary defined by Φ(T ) = ρ(T )Se. ρ ◦ Φ(T ) = dS∗e ρ(dφρ(ρ(T )S)) Φ−1 ◦ E(3) = d2S∗e ρ(T φρ(Se)) = dS∗e ρ(T S) = √dj2(T )∗ρ(S) = j1 ◦ j2(T ), which proves (j2 ◦ j1)3 = (j1 ◦ j2)−3 = I again. Remark 3.11. Our explicit formula for E(n) as above allows us to compute easily the higher Frobenius-Schur indicators νn,r(ρ) = tr((E(n) ρ )r). Indeed, we have ν2,1(ρ) = ǫ, which should be treated as a given datum. Since E(3) is unitarily equivalent to j1◦ j2, ρ we have ν3,1(ρ) = tr(j1 ◦ j2). Since {SgSe}g∈G ∪ {TiTjSe}1≤i,j≤m is an orthonormal ρ THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 23 basis of (id, ρ4), we have S∗e T ∗j T ∗i j2(Ti)∗ρ(TjSe)Se S∗e T ∗j T ∗i φρ(TiTjSe) S∗e T ∗j T ∗i S∗e ρ(TiTjSe)Se S∗e S∗g φρ(SgSe) + dXi,j ν4,1(ρ) = dXg∈G S∗e S∗g S∗e ρ(SgSe)Se + dXi,j = dXg∈G S∗e S∗g S∗e U(g)ρ(Se)U(g)∗ρ(Se)Se + √dXi,j = dXg∈G S∗e S∗g U(g)∗ρ(Se)Se +Xi,j =Xg∈G dXg∈G χg(g) +Xi,j dXg∈G χg(g) +Xi,j dXg∈G χg(g) +Xi,j hj1(Tj), (T ∗i j2(Ti)∗ρ(Tj))∗i hj1((T ∗i j2(Ti)∗ρ(Tj))∗), Tji T ∗i j2(Ti)∗ρ(Tj)j1(Tj). = 1 1 1 = = S∗e T ∗j j1((T ∗i j2(Ti)∗ρ(Tj))∗)Se It is easy to evaluate these for concrete examples. We get back to the original setting where our near-group categories live in End(M). About ρ(U(g)), we have Lemma 3.12. (3.30) ρ(U(g)) =Xh∈G ShS∗hg + j2 ◦ j−1 1 UK(g)(j2 ◦ j−1 1 )∗ ⊗ U(g). Proof. Since ρ(U(g))Se ∈ (αg−1, ρ2) there exists a complex number c of modulus 1 such that ρ(U(g))Se = cSg−1. To show that c = 1, we compute both sides of S∗e ρ2(U(g))ρ(Se) = cS∗e ρ(Sg−1). The left-hand side is S∗e ρ2(U(g))ρ(Se) = U(g)S∗e ρ(Se) = ǫ d U(g). On the other hand, the right-hand side is cS∗e ρ(Sg−1) = cS∗e U(g)∗ρ(Se)U(g) = cS∗e ρ(Se)U(g) = c ǫ n U(g), showing c = 1. Thus we have Shg−1 = αh(ρ(U(g))Se) = ρ(U(g))Sh and This implies that P and Q commute with ρ(U(g)) and ρ(U(g))P =Xh∈G ρ(U(g)) =Xh∈G ρ(U(g))ShS∗h =Xh∈G Shg−1S∗h. Shg−1S∗h + Qρ(U(g))Q. 24 MASAKI IZUMI Note that T ∗i ρ(U(g))Tj ∈ (ρ, ρ◦ αg) = CU(g). This means that there exists a unitary representation U′ of G in B(K) such that Qρ(U(g))Q = U′(g) ⊗ U(g), and it is determined by S∗e T ∗i ρ(U(g))TjSe = T ∗i U′(g)Tj. Indeed, expanding U(g) as U(g) =Xh∈G χh(g)ShS∗h + mXp,q=1 UK(g)pqTpT ∗q , S∗e T ∗i ρ(U(g))TjSe χh(g)S∗e T ∗i ρ(ShS∗h)TjSe + mXp,q=1 UK(g)pqS∗e T ∗i ρ(TpT ∗q )TjSe UK(g)pqhj2 ◦ j−1 1 (Tp), TiihTj, j2 ◦ j−1 1 (Tq)i, =Xh∈G mXp,q=1 = we get and U′(g) = S∗e T ∗i ρ(U(g))TjSeTiT ∗j mXi,j=1 mXp,q=1 = = mXi,j=1 mXp,q=1 = j2 ◦ j−1 UK(g)pqhj2 ◦ j−1 1 (Tp), TiihTj, j2 ◦ j−1 1 (Tq)iTiT ∗j UK(g)pqj2 ◦ j−1 1 UK(g)(j2 ◦ j−1 1 (Tp)j2 ◦ j−1 1 )∗. 1 (Tq)∗ Remark 3.13. We can graphically express the three unitary representations of G on K as follows: ρ (cid:3) V (g)T := αg(T ) = αg T ⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧ ρ ρ , ❄❄❄❄❄❄❄❄❄❄ ρ UK(g)T = , T ⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧ αg ρ ❄❄❄❄❄❄❄❄❄❄ ρ ? ? THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 25 (j2 ◦ j−1 1 )UK(g)(j2 ◦ j−1 1 )−1T = ρ(U(g))T U(g)∗ = T ρ . αg ❄❄❄❄❄❄❄❄❄❄ ρ ⑧⑧⑧⑧⑧⑧⑧⑧⑧⑧ This explains why they are related by the rotation map j2 ◦ j1. ρ The relation implies I =Xg∈G ρ(Sg)ρ(Sg)∗ + ρ(Ti)ρ(Ti)∗ mXi=1 Lemma 3.14. Let GV be the set of irreducible representations of G contained in V , and let (UK,K) = Mπ∈GV mπMa=1 (π,Ka π) be the irreducible decomposition. We choose an orthonormal basis {T a so that π,i}dim π i=1 of Ka π UK(g) = Xπ∈ GV mπXa=1 π(g)ijT a π,iT a π,j∗. Then (3.31) (3.32) n d Xπ,a,b,i,j 1 dim π T a π,ij1(T a π,j)j1(T b π,j)∗T b π,i ∗ + mXi=1 l(Ti)l(Ti)∗ = Q ⊗ Q, l(Ti)V (h)j2(Ti) + mXi=1 ǫn d Xπ,a δχh,πT a π j1(T a π ) = 0. e e 26 MASAKI IZUMI Proof. By using the Peter-Weyl theorem, we obtain U(g)ρ(Se)ρ(S∗e )U(g)∗ χh(g)χk(g)ShS∗hρ(Se)ρ(Se)∗SkS∗k ρ(Sg)ρ(Sg)∗ =Xg∈G Xg∈G = Xg,h,k∈G + Xg,k∈G Xπ,a,i,j +Xg∈G Xπ,σ,i,j,p,q d2 Xh,k∈G d Xπ,a,b,i,j dim π + = n n 1 δχh,χkShS∗k + π(g)ijχk(g)T a π,iT a π(g)ijσ(g)pqT a π,iT a π,j∗ρ(Se)ρ(Se)∗SkS∗k + (· · · )∗ π,j∗ρ(Se)ρ(Se)∗T b σ,bT b σ,p ∗ ǫn d√dXk,π,a δχk,πT a π j1(T a π )S∗k + (· · · )∗ T a π,ij1(T a π,j)j1(T b π,j)∗T b π,i ∗. On the other hand, ρ(Ti)ρ(Ti)∗ = hV (h−1k)j2(Ti), j2(Ti)iShS∗k 1 d mXi=1 Xh,k∈G V (h)j2 ◦ j−1 1 (Ti)ShS∗hj2 ◦ j−1 1 (Ti)∗V (h)∗ + l(Ti)l(Ti)∗ mXi=1 + + mXi=1 mXi=1Xh∈G mXi=1Xh∈G 1 √d d Xh,k∈G 1 mXi=1 = + l(Ti)V (h)j2(Ti)S∗h + (· · · )∗ Tr V (h−1k)ShS∗k + TiShS∗hT ∗i mXi=1Xh∈G 1 √d mXi=1Xh∈G l(Ti)l(Ti)∗ + l(Ti)V (h)j2(Ti)S∗h + (· · · )∗. Since V is unitarily equivalent to UK, Eq.(3.29) and Eq.(3.21) imply Tr V (h−1k) = Tr UK(h−1k) = Tr(λ(h−1k)) − d n 1 dXg∈G χk(g)χh(g) = dδh,k − n d δχh,χk, THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 27 and Since we get ρ(Ti)ρ(Ti)∗ = P − δχh,χkShS∗k TiShS∗hT ∗i + l(Ti)l(Ti)∗ n d2 Xh,k∈G mXi=1 + mXi=1 mXi=1Xh∈G mXi=1Xh∈G 1 √d + l(Ti)V (h)j2(Ti)S∗h + (· · · )∗. I =Xg∈G ρ(Sg)ρ(Sg)∗ + ρ(Ti)ρ(Ti)∗, mXi=1 Q = TiShS∗hT ∗i n mXi=1Xh∈G d Xπ,a,b,i,j mXi=1Xh∈G 1 √d d√dXh,π,a ǫn + + + 1 dim π T a π,ij1(T a π,j)j1(T b π,j)∗T b π,i ∗ + mXi=1 l(Ti)l(Ti)∗ l(Ti)V (h)j2(Ti)S∗h + (· · · )∗ δχh,πT a π )S∗h + (· · · )∗, π j1(T a which shows the statement. (cid:3) Lemma 3.15. In terms of l(T ), Eq.(3.30) is expressed as (3.33) 1 )∗ ⊗ UK(g) 1 1 UK(g)(j2 ◦ j−1 j2 ◦ j−1 d Xπ,σ,a,b,i,j,p,q = hχgπij, σpqiT a where π,ij1(T a π,j)j1(T b σ,q)∗T b σ,p ∗ + Xπ,a,i,j π(g)ijl(T a π,i)l(T a π,j)∗, Proof. Since ρ(U(g)) is a unitary, Eq.(3.30) holds if and only if the following two hold: hχgπij, σpqi =Xh∈G χg(h)πij(h)σpq(h). P ρ(U(g))P =Xh ShS∗hg, 28 MASAKI IZUMI The first one does not give any condition on l as Qρ(U(g))Q = j2 ◦ j−1 1 UK(g)(j2 ◦ j−1 1 )∗ ⊗ U(g). χh(g)P ρ(ShS∗h)P + mXi,j=1 UK(g)ijP ρ(Ti)ρ(Tj)∗P χh(g)P U(h)ρ(SeS∗e )U(h)∗P UK(g)ijhV (t)j2(Tj), V (s)j2(Ti)iSsS∗t χh(t−1sg)SsS∗t + UK(g)ijhj∗2V (t−1s)j2Ti, TjiSsS∗t χh(t−1sg)SsS∗t + Tr UK(t−1sg)SsS∗t 1 d 1 d mXi,j=1Xs,t∈G mXi,j=1 1 = + 1 d P ρ(U(g))P =Xh∈G =Xh,s,t mXi,j=1Xs,t∈G d2Xh,s,t d2Xh,s,t =Xs∈G Qρ(U(g))Q =Xh∈G SsS∗sg. = 1 For the second, we have χh(g)QU(h)ρ(SeS∗e )U(h)∗Q + Xπ,a,i,j π(g)ijQρ(T a π,iT a π,j∗)Q. The first term above is χh(g)QU(h)ρ(SeS∗e )U(h)∗Q Xh∈G = Xπ,σ,a,b,i,j,p,qXh∈G d Xπ,σ,a,b,i,j,p,q = 1 The second term is π(g)ijQρ(T a π,iT a π,j∗)Q χg(h)π(h)ijσ(h)pqT a π,iT a π,j∗ρ(SeS∗e )T b σ,qT b σ,p ∗ hχgπij, σpqiT a π,ij1(T a π,j)j1(T b σ,q)∗T b σ,p ∗ Xπ,a,i,j = Xπ,a,i,jXh∈G + Xπ,a,i,j =Xh∈G =Xh∈G π(g)ijV (h)j2 ◦ j−1 1 (T a π,i)ShS∗hj2 ◦ j−1 1 (T a π,j)∗V (h)∗ π(g)ijl(T a π,i)l(T a π,j)∗ V (h)j2 ◦ j−1 1 UK(g)(j2 ◦ j−1 j2 ◦ j−1 1 UK(g)(j2 ◦ j−1 1 )∗V (h)∗ ⊗ ShS∗h + Xπ,a,i,j 1 )∗ ⊗ χh(g)ShS∗h + Xπ,a,i,j π(g)ijl(T a π,i)l(T a π,j)∗ π(g)ijl(T a π,i)l(T a π,j)∗, THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 29 where we use the fact that j2 ◦ j1 has period three. Thus the statement follows. (cid:3) 4. Reconstruction In the previous section, we showed that every information of a C∗ near-group category C with a finite group G is encoded in two anti-linear isometries j1, j2 of a finite dimensional Hilbert space K, two unitary representations V, UK of G on K, characters {χh}h∈G of G, and a linear map l : K → B(K,K ⊗ K). In this section, we deduce a necessary and sufficient condition for the tuple (K, j1, j2, V, UK, χ, l) to recover the original C∗ near-group category C. Let G be a finite group of order n, and let m be a natural number satisfying the condition in the conclusion of Theorem 3.9. Let On+m be the Cuntz algebra with the i=1 TiT ∗i , and canonical generators {Sg}g∈G∪{Ti}m K = span{Ti}m i=1. Let ǫ ∈ {1,−1}. We choose anti-linear isometries j1 and j2 of K and unitaries representations V and UK in B(K) satisfying Eq.(3.7),(3.8),(3.9),(3.18). We assume that χh ∈ Hom(G, T), h ∈ G, satisfy the condition in the conclusion of Theorem 3.9 and Eq.(3.19),(3.21). Under the above assumption, we can introduce a unitary representation U of G in On+m by i=1. We set P =Pg∈G SgS∗g , Q =Pm U(g) =Xh∈G χh(g)ShS∗h + UK(g), and an action α of G on On+m by αg(Sh) = Sgh and αg(T ) = V (g)T for g ∈ G and T ∈ K. Choosing a linear map l : K → B(K,K ⊗ K) = K2K∗ satisfying Eq.(3.13),(3.14),(3.15), (3.16),(3.17),(3.20), and (3.33), we can introduce a unital endomorphism ρ of On+m by ρ(Se) = ǫ dXh∈G Sh + 1 √d mXi=1 Tij1(Ti), ρ(Sg) = U(g)ρ(Se)U(g)∗, ρ(T ) = 1 √dXh∈G Shαh(j2(T )∗) +Xh∈G αh(j2 ◦ j−1 1 (T ))ShS∗h + l(T ). Indeed, we first define ρ on the canonical generators {Sg, Ti} of the Cuntz algebra On+m. Then {ρ(Sg), ρ(Ti)} are isometries with mutual orthogonal ranges, and so E =Xg∈G ρ(Sg)ρ(Sg)∗ + ρ(Ti)ρ(Ti)∗ mXi=1 is a projection. The proof of Lemma 3.15 in the case of g = 0 implies P EP = P and QEQ = Q, which shows that E = I. Thus ρ extends to a unital endomorphism of On+m, and the proof of Lemma 3.14 implies that Eq.(3.32) holds (note that Eq.(3.31) is a special case of Eq.(3.33)). Now it is easy to show Eq.(3.5), Eq.(3.6), αg ◦ ρ = ρ, and Ad U(g) ◦ ρ = ρ ◦ αg. The proof of Lemma 3.15 shows that Eq.(3.30) holds. 30 MASAKI IZUMI Lemma 4.1. Let X be an isometry of On+m. If S∗e ρ2(X)Se = X and T ∗i ρ2(X)Ti = ρ(X) for all i, then ρ2(X) =Xh∈G Shαh(X)S∗h + Tiρ(X)T ∗i . mXi=1 Proof. Applying αg to S∗e ρ2(X)Se = X, we get S∗g ρ2(X)Sg = αg(X). Since I = ρ2(X)∗ρ2(X), we have I = S∗g ρ2(X)∗ρ2(X)Sg too, and S∗g ρ2(X)∗TiT ∗i ρ2(X)Sg S∗g ρ2(X)∗ShS∗hρ2(X)Sg I =Xh∈G S∗g ρ2(X)∗ShS∗hρ2(X)Sg +Xi=1 = S∗g ρ2(X)∗SgS∗g ρ2(X)Sg + Xh∈G\{g} +Xi=1 = I + Xh∈G\{g} S∗g ρ2(X)∗ShS∗hρ2(X)Sg +Xi=1 S∗g ρ2(X)∗TiT ∗i ρ2(X)Sg S∗g ρ2(X)∗TiT ∗i ρ2(X)Sg. Thus S∗hρ2(X)Sg = 0 for g 6= h and T ∗i ρ2(X)Sg = 0. In a similar way, we have Sgρ2(X)Ti = 0 and T ∗j ρ2(X)Ti = 0 for j 6= i. These relations imply P ρ2(X)Q = Qρ2(X)P = 0 and ρ2(X) = (P + Q)ρ2(X)(P + Q) = P ρ2(X)P + Qρ2(X)Q =Xh∈G Shαh(X)S∗h + Tiρ(X)T ∗i . mXi=1 Lemma 4.2. For g, h ∈ G, the equation S∗hρ2(Sg)Sh = Shg holds. Proof. It suffices to show the statements for h = e because the others follow from them by applying αh to them. (cid:3) S∗e ρ2(Se)Se = = = Now Eq.(3.30) implies ǫ ǫ dXk∈G dXk∈G d2Xk∈G 1 S∗e ρ(Sk)Se + S∗e ρ(Ti)ρ(j1(Ti))Se 1 √d mXi=1 j2(Ti)∗ρ(j1(Ti))Se 1 d mXi=1 S∗e U(k)ρ(Se)U(k)∗Se + Se + 1 d mXi=1 Se = Se. S∗e ρ(Sg)Se = S∗e ρ(U(g))ρ2(Se)ρ(U(g))∗Se = S∗g ρ2(Se)Sg = Sg. (cid:3) THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 31 Proof. We first show the statement for g = e. Thanks to the above two lemmas, it i=1 Tiρ(Se)T ∗i . Instead of the orthogonal basis {Ti}m i=1, π,p} sometimes. By the definition of ρ, we have mXi=1 mXi=1 Lemma 4.3. ǫ Shαh(Sg)S∗h + Qρ2(Se)Q = we use {T a ρ2(Sg) =Xh∈G suffices to show Qρ2(Se)Q =Pm dXg∈G 1 √d dXg∈G Xπ,σ,a,b,i,j,p,q d Xπ,a,b,i,j d√d Xπ,a,b,i,j UK(g)ρ(Se)UK(g)∗ = dXg∈G The first term is Qρ(Sg)Q + dim π dim π nǫ nǫ = = 1 1 ǫ ǫ Tiρ(Sg)T ∗i . Qρ(Ti)ρ(j1(Ti))Q. π(g)ijσ(g)pqT a π,iT a π,j∗ρ(Se)T b σ,qT b σ,p ∗ π,iT a T a π,j∗ρ(Se)T b π,jT b π,i ∗ T a π,ij1(T a π,j)T b π,jT b π,i ∗. The second term is Second term = = αh(j2 ◦ j−1 √dXh∈G 1 √d ǫ d mXi=1(cid:8) 1 mXj=1Xh∈G TjShT ∗j + 1 √d 1 (Ti))Shαh(j2 ◦ j1(Ti)∗) + l(Ti)l(j1(Ti))(cid:9) mXi=1 l(Ti)l(j1(Ti)). Using Eq.(3.16),(3.31), we have 1 √dXi,p,q l(Ti)l(Ti)∗TpTqj1(Tq)T ∗p l(Ti)l(j1(Ti)) = TpTqj1(Tq)T ∗p 1 n = − mXi=1 1 √d √dXp,q d√dXp,q Xπ,a,b,i,j √dXp,q √dXp,q = = 1 1 TpTqj1(Tq)T ∗p − TpTqj1(Tq)T ∗p − n d√d Xπ,a,b,i,j d√d Xπ,a,b,i,j nǫ 1 dim π T a π,ij1(T a π,j)j1(T b π,j)∗T b π,i ∗TpTqj1(Tq)T ∗p 1 dim π 1 dim π T a π,ij1(T a π,j)j2 1(T b π,j)T b π,i ∗ T a π,ij1(T a π,j)T b π,jT b π,i ∗. Thus we obtain the statement for g = e. The general statement follows from this and Eq.(3.30). (cid:3) 32 MASAKI IZUMI Lemma 4.4. The following conditions are equivalent: (1) Sg ∈ (αg, ρ2) for all g ∈ G. (2) S∗e ρ(l(T ))Se = (1 − 2n/d2)T for all T ∈ K. (3) Let l(2) : K → K be as in Remark 3.5. Then ij 1 d ij (T ))j2(Tj) = (1 − j2(Ti)∗l(l(2) (4.1) mXi,j=1 2n d2 )T, T ∈ K. Proof. It is easy to show that (2) and (3) are equivalent. Thanks to the above lemmas, it suffices to show that (2) is equivalent to S∗e ρ2(T )Se = T . Indeed, S∗e ρ2(T )Se = 1 √dXh∈G S∗e ρ(αh(Sej2(T )∗))Se +Xh∈G S∗e ρ(αh(j2 ◦ j−1 1 (T )SeS∗e ))Se + S∗e ρ(l(T ))Se. The first term is 1 √dXh∈G n √d = S∗e U(h)ρ(Sej2(T )∗)U(h)∗Se S∗e ρ(Sej2(T )∗)Se = nǫ d2 j2 2 (T ) = n d2 T. The second term is S∗e U(h)ρ(j2 ◦ j−1 Xh∈G = nS∗e ρ(j2 ◦ j−1 = n d2 T. 1 (T )SeS∗e )U(h)∗Se 1 (T )SeS∗e )Se = nǫ d√d j2 2 ◦ j−1 1 (T )∗ρ(Se) = n d√d j−1 1 (T )∗ρ(Se) Thus the statement is proved. (cid:3) Lemma 4.5. Assume that Eq.(4.1) holds. Then the following conditions are equiv- alent: (1) K = (ρ, ρ2). (2) For any T, T ′, T ′′ ∈ K, the equality ρ(T ′′)∗ρ2(T )T ′ = ρ(T ′′)∗T ′ρ(T ) holds. Proof. We only show that (2) implies (1). Assume that (2) holds. Thanks to the above lemmas, it suffices to show ρ2(T )T ′ = T ′ρ(T ). For this, it suffices to show ρ(Sg)∗ρ2(T )T ′ = ρ(S∗g )T ′ρ(T ) for we have ρ(P ) + ρ(Q) = I. Since we have ρ(Sg)∗ρ2(T )T ′ = U(g)ρ(Se)∗ρ2(T )U(g)∗T ′, ρ(S∗g )T ′ρ(T ) = U(g)ρ(S∗e )U(g)∗T ′ρ(T ), we only take care of the case with g = e. THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 33 For ρ(Se)∗ρ2(T )T ′, we have ρ(Se)∗ρ2(T )T ′ = ρ(S∗e ρ(T ))T ′ = 1 √d ρ(j2(T ))∗T ′ = 1 √dXh∈G For ρ(Se)∗T ′ρ(T ), we have hT ′, V (h)j2 ◦ j−1 1 ◦ j2(T )iShS∗h + 1 √d l(j2(T ))∗T ′. j1(T ′)∗ρ(T ) ρ(Se)∗T ′ρ(T ) = = = 1 √d √dXh∈G 1 √dXh∈G 1 hV (h)j2 ◦ j−1 1 (T ), j1(T ′)iShS∗h + j1(T ′)∗l(T ) hT ′, V (h)j1 ◦ j2 ◦ j1(T )iShS∗h + j1(T ′)∗l(T ). 1 √d 1 √d Now the statement follows from Eq.(3.17),(3.18). (cid:3) Theorem 4.6. The endomorphism ρ satisfies ρ2(x) =Xg∈G Sgαg(x)S∗g + Tiρ(x)T ∗i mXi=1 if and only if Eq.(4.1) and the following three equations hold for all h, k ∈ G and T, T ′, T ′′ ∈ K: (4.2) (4.3) T ′∗l(Ti)j2(l(T )∗j2(T ′′)Ti) mXi=1 1 dXh∈G 1 (T ′)j2 ◦ j−1 = ǫhT ′′, T ′iT − hUK(h)T ′′, j1(T )ij1(UK(h)T ′), l(T ′′)∗j2 ◦ j−1 1 (T ) = j2 ◦ j−1 1 (T ′′∗l(T )T ′), (4.4) l(T ′′)∗T ′l(T ) = l(T ′′∗l(T )Ti)l(Ti)∗T ′ mXi=1 + 1 dXh∈G mXi=1 hV (h)j2 ◦ j−1 1 (T ), T ′′iUK(h)Tij1(Ti)j1(UK(h)∗T ′)∗. Proof. It suffices to show that ρ(T ′′∗ρ(T ))T ′ = ρ(T ′′)∗T ′ρ(T ) is equivalent to the above three. We compute ρ(T ′′)∗ρ2(T )T ′ first: ρ(T ′′∗ρ(T ))T ′ =Xh∈G hV (h)j2 ◦ j−1 1 (T ), T ′′iρ(ShS∗h)T ′ + ρ(T ′′∗l(T ))T ′. 34 MASAKI IZUMI The first term is 1 (T ), T ′′iU(h)ρ(SeS∗e )U(h)∗T ′ hV (h)j2 ◦ j−1 1 (T ), T ′′iχk(h)Skj1(UK(h)∗T ′)∗ hV (h)j2 ◦ j−1 1 (T ), T ′′iU(h)Tij1(Ti)j1(UK(h)∗T ′)∗. 1 = Xh∈G hV (h)j2 ◦ j−1 d√d Xh,k∈G ǫ mXi=1 dXh∈G mXi=1 + The second term is ρ(T ′′∗l(T ))T ′ = ρ(T ′′∗l(T )Ti)ρ(Ti)∗T ′ hT ′, V (h)j2 ◦ j−1 1 (Ti)iρ(T ′′∗l(T )Ti)ShS∗h + ρ(T ′′∗l(T )Ti)l(Ti)∗T ′ mXi=1 hj1 ◦ j−1 2 V (h)∗T ′, TiiV (h)j2 ◦ j−1 1 (T ′′∗l(T )Ti)ShS∗h Skj2(T ′′∗l(T )Ti)∗V (k)∗l(Ti)∗T ′ + l(T ′′∗l(T )Ti)l(Ti)∗T ′ mXi=1 V (h)j2 ◦ j−1 1 (T ′′∗l(T )j1 ◦ j−1 2 (V (h)∗T ′))ShS∗h Skj2(T ′′∗l(T )Ti)∗l(Ti)∗V (k)∗T ′V (k)∗ + = = mXi=1Xh∈G mXi=1Xh∈G mXi=1Xk∈G + 1 √d =Xh∈G 1 √d + mXi=1Xk∈G On the other hand, l(T ′′∗l(T )Ti)l(Ti)∗T ′. mXi=1 hT ′, V (k)j2 ◦ j−1 1 (T ′′)iSkS∗kρ(T ) + l(T ′′)∗T ′ρ(T ) 1 (T ′′)iSkj2(T )∗V (k)∗ 1 (T )ShS∗h + l(T ′′)∗T ′l(T ). 1 = ρ(T ′′)∗T ′ρ(T ) =Xk∈G √dXk∈G hT ′, V (k)j2 ◦ j−1 +Xh∈G l(T ′′)∗T ′V (h)j2 ◦ j−1 hV (k)j2 ◦ j−1 = 1 (T ′′), T ′ij2(T ) T ′∗V (k)l(Ti)j2(T ′′∗l(T )Ti) mXi=1 dXh∈G ǫ Therefore we get (4.5) (4.6) + χk(h)hT ′′, V (h)j2 ◦ j−1 1 (T )iV (k)∗j1(UK(h)∗T ′), l(T ′′)∗T ′V (h)j2 ◦ j−1 1 (T ) = V (h)j2 ◦ j−1 1 (T ′′∗l(T )j1 ◦ j−1 2 (V (h)∗T ′)), THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 35 and Eq.(4.4). Thanks to Eq.(3.13), we have V (h)∗l(T ′′)∗ = l(T ′′)(V (h)∗ ⊗ V (h)∗), and Eq.(4.6) is equivalent to Eq.(4.3) if T ′ is replaced with V (h)j2 ◦ j−1 1 (T ′). Since V (k)∗j1(UK(h)∗T ′) = j1(V (k)∗UK(h)∗T ′) = j1(χk(h)UK(h)∗V (k)∗T ′), if T ′ is replaced with V (k)T ′, Eq.(4.5) is equivalent to hj2 ◦ j−1 = 1 (T ′′), T ′ij2(T ) T ′∗l(Ti)j2(T ′′∗l(T )Ti) + mXi=1 ǫ dXh∈G hT ′′, V (h)j2 ◦ j−1 1 (T )ij1(UK(h)∗T ′), If we replace T ′′ with j1 ◦ j−1 2 (T ′′) and T with j−1 2 (T ), this is equivalent to 2 (T ′′)∗l(j2(T ))Ti) T ′∗l(Ti)j2(j1 ◦ j−1 hj1 ◦ j−1 2 (T ′′), V (h)j2 ◦ j−1 1 ◦ j−1 2 (T )ij1(UK(h)∗T ′) = ǫ T ′∗l(Ti)j2(l(T )∗j2(T ′′)Ti) hT ′′, T ′iT = ǫ + + ǫ mXi=1 dXh∈G mXi=1 dXh∈G mXi=1 1 hj1 ◦ j−1 2 (T ′′), V (h)j1 ◦ j2 ◦ j1(T )ij1(UK(h)∗T ′) = ǫ T ′∗l(Ti)j2(l(T )∗j2(T ′′)Ti) + ǫ dXh∈G hUK(h)∗T ′′, j1(T )ij1(UK(h)∗T ′), where we use Eq.(3.17),(3.18). Thus we are done. (cid:3) Remark 4.7. Replacing T ′′ with j2(T ′′) and T ′ with j1(T ′), we see that in Eq.(4.4) is equivalent to (4.7) T ′∗l(T ′′)l(T ) = l(j2(T ′′)∗l(T )j2(Ti))T ′∗l(Ti) mXi=1 hT ′′, UK(h)j1(T )iUK(h)Tij1(Ti)j1(UK(h)∗j1(T ′))∗. + 1 dXh∈G mXi=1 36 MASAKI IZUMI Likewise, replacing T with j1(T ) in Eq.(4.4), it is equivalent to (4.8) l(T ′′)∗T ′l(T )∗TiTjj1(Tj)T ∗i = mXi,j=1 dXh∈G + 1 mXi=1 l(T ′′∗l(T )∗TiTjj1(Tj))l(Ti)∗T ′ mXi,j=1 hV (h)j2(T ), T ′′iUK(h)Tij1(Ti)j1(UK(h)∗T ′)∗. Definition 4.8. We say that a tuple (K, j1, j2, V, UK, χ, l) is admissible if it satis- fies Eq.(3.7),(3.8),(3.9),(3.18), the condition in the conclusion of Theorem 3.9, and Eq.(3.19),(3.21),(3.13),(3.14),(3.15), (3.16),(3.17),(3.20),(3.33),(4.1),(4.2), (4.3),(4.4). We say that two tuples (K, j1, j2, V, UK, l) and (K′, j′1, j′2, V ′, U′ , l′) are equivalent if K there exist a unitary W : K → K′ and a group automorphism ϕ ∈ Aut(G) satis- fying j′1W = W j1, j′2W = W j2, UK′(g)W = W UK(ϕ(g)), V ′(g)W = W V (ϕ(g)), χ′h(g) = χϕ(h)(ϕ(g)), and l′(W T )W = (W ⊗ W )l(T ), T ∈ K. We have seen that starting from an admissible tuple (K, j1, j2, V, UK, χ, l), we can construct the Cuntz algebra endomorphism ρ ∈ End(Om+n) and the G-action α satisfying relevant properties. As in [28], we can choose an appropriate representation of Om+n so that ρ and α extend to the weak closure of Om+n, which is a hyperfinite type IIIλ factor, without changing morphism spaces (see Appendix and [28]). This finishes the reconstruction process from (K, j1, j2, V, UK, χ, l) to C. Let C and D be C∗ near-group categories with finite group G and a multiplic- ity parameter m realized in End0(M), which give rise to two admissible tuples (K, j1, j2, V, UK, l) and (K′, j′1, j′2, V ′, U′ , l′) respectively. In view of Theorem 2.2 and K Remark 3.6, we see that the two C∗ near-group categories C and D are equivalent if and only if the two corresponding tuples are equivalent. In conclusion, we get the following result. Theorem 4.9. The association C 7→ (K, j1, j2, V, UK, l) gives a one-to-one correspon- dence between the set of equivalence classes of C∗ near-group categories with finite group G and a multiplicity parameter m, and the set of equivalence classes of admis- sible tuples. 5. The case of m = G − 1 In this section, we briefly give an account of the classification of near-group cate- gories with a finite group G and the multiplicity parameter m = G − 1. We have seen in Theorem 3.9 that G is abelian under the C∗ condition. In fact, we have the following classification result without this additional assumption. A fusion category is said to be group theoretical if it is categorically Morita equiv- alent to a pointed category (see [11, Definition 9.7.1]). Theorem 5.1. Let C be a near group category with a finite group G and the multi- plicity parameter m = G − 1. Then the group G is cyclic and G + 1 is a prime power q. If G = Z2, there are three such categories, if G = Z3 or G = Z7, there are THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 37 two such categories, and if G = Zq−1 in the other cases, there is one such category. All these fusion categories are group theoretical. Example 5.2. For G = Z2, we have dimK = 1, and Eq.(3.31) implies l = 0. We may choose a basis {T} of K so that j1(T ) = T , and so ǫ = 1. In this case, the only choices of j2 are j2(T ) = ζT with ζ 3 = 1. All the conditions in Definition 4.8 are easy to verify in this case, and there are certainly three near-group categories. This example was already discussed in [28]. For the proof of the above theorem, Siehler [50, Theorem 1.2] was the first to show that G is cyclic and G + 1 is a prime power under an additional assumption of G being abelian (though this assumption is not explicitly written in [50, Theorem 1.2]). Etingof-Gelaki-Ostrik [10, Corollary 7.4] showed the statement under the assumption that G is cyclic. Recently Nikshych-Ostrik [45] showed that G is cyclic and the classification was completed. Let Fq be the finite field of order q, and let F×q be its multiplicative group. We regard Fq as an additive group on which F×q acts by multiplication. Etingof-Gelaki- Ostrik reduced the classification to a counting argument of the group H 3(Fq, T)F× of F×q -invariant cohomology classes in H 3(Fq, T). We reproduce their reduction ar- gument from the view point of operator algebras now. The following argument was developed in [29], [33, Theorem 9.8,(i)], which could serve as an introduction to more complicated arguments in the next section. q Let N = ρ(M), which is an irreducible subfactor of index d2 = G2. Since αg ◦ ρ = ρ, the fixed point algebra is an intermediate subfactor between M and N with M G = {x ∈ M; αg(x) = x, ∀g ∈ G}, [M : M G] = [M G : N] = G. Let ι : M G ֒→ M and κ : N ֒→ M G be inclusion maps. Then we have the decompo- sition ρ = ι ◦ κ ◦ ρ0, where ρ0 is ρ regarded as an isomorphism from M onto N. Let α′g = ρ0 ◦ αg ◦ ρ−1 0 , which is an outer action of G on N, and let N G be the fixed point algebra N G = {x ∈ N; α′g(x) = x, ∀g ∈ G}. Since Ad U(g) ◦ ρ = ρ ◦ αg, we have α′g(y) = Ad U(g)(y) for any y ∈ N, and the von Neumann algebra generated by N and {U(g)}g∈G is identified with the crossed product N ⋊α′ G. Eq.(3.22) and Theorem 3.9 show αg(U(h)) = U(h) for any g, h ∈ G, and we get N ⋊α′ G ⊂ M G. Since [M G : N] = G = [N ⋊α′ G : N], we get M G = N ⋊α′ G. Since the image of ρ0 ◦ ι is N G, the duality between the fixed point inclusion N G ⊂ N and the crossed product inclusion N ⊂ N ⋊α′ G implies that the endomorphism κ ◦ ρ0 ◦ ι ∈ End0(M G) contains an automorphism, which we denote θ ∈ Aut(M G). Then the Frobenius reciprocity implies [κρ0] = [θι], and we get [ρ] = [ικρ0] = [ιθι]. 38 MASAKI IZUMI Since G is abelian, there exists an outer action β : G → Aut(M G) of the dual group G of G such that M = M G ⋊β G and α is the dual action of β. Thus [ιι] =Mχ∈ G [βχ]. We denote by L the group generated by [θ] and [β G] in Out(M G). Lemma 5.3. Let the notation be as above. (1) For χ1, χ2, τ1, τ2 ∈ G, we have [βχ1θβχ2] = [βτ1θβτ2] if and only if χi = τi for (2) L = [β G] ⊔ [β G][θ][β G]. i = 1, 2. Proof. Since ρ is irreducible, we have 1 = dim(ρ, ρ) = dim(ιθι, ιθι) = dim(ιιθ, θιι) = Xχ,τ∈ G (βχθ, θ, βτ ), which shows (1). Since ρ is self-conjugate, we [ιθι] = [ιθ−1ι]. This implies that [θ−1] is contained in [ιιθιι], and it is an element in [β G][θ][β G]. Let L1 be the right-hand side of (2). The fusion rules of the category implies Xg∈G [αg] + (G − 1)[ρ] = [ρ2] = [ιθιιθι] =Xχ∈ G [ιθβχθι]. This means that [θβχθ] belongs to L1, which shows that L1 is already a group. Therefore L = L1. (cid:3) In what follows, we identify G with [β G] for simplicity. The above double coset decomposition implies that L acting on L/ G is a sharply 2-transitive permutation group with the abelian point stabilizer G, which allows us to identify the pair (L, G) with (Fq ⋊ F×q , F×) (see [26, Chapter XII, §9]). The embedding L ⊂ Out(L) carries a cohomology class in ω ∈ H 3(L, T) whose restriction to G is trivial as it comes from the genuine group action β. Such a class ω is identified with a class in H 3(Fq, T)F× by restriction thanks to the Lyndon-Hochschild-Serre spectral sequence. q Now we reverse the above process. Assume that we are given a cohomology class ω ∈ H 3(Fq ⋊ F×q , T) whose restriction to F×q is trivial. Let P be a hyperfinite type III1 factor. Then there exists an embedding of Fq ⋊ F×q into Out(P ) carrying the class ω, which is unique up to conjugacy in Out(P ) thanks to Theorem 2.2 (or see [37]). We choose a lifting γ : Fq ⋊ F⋊ q → Aut(P ) of it. Since the restriction of ω to F×q is an action, which we denote by β. Since F×q is cyclic, the second cohomology H 2(F×q , T) is trivial, and such β is unique on F×q up to 1-cocycle perturbation. We denote by G the dual group of F×q . Let M = P ⋊β F×q , and let ι : P ֒→ M be the inclusion map. We choose an arbitrary h ∈ Fq ⋊ F×q \ F×q , and set ρ = ιγhι, whose equivalence class does not depend on the is trivial, we may assume that γ restricted to F×q THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 39 choice of h. The same computation as above shows that ρ is irreducible and [ρ2] =Xg∈G [ βg] + (q − 2)[ρ], where β is the dual action of β. Therefore we get a C∗ near-group category with the group G ∼= Zq−1 and the multiplicity parameter m = q − 2. 6. The noncommutative case In this section, we classify C∗ near group categories with noncommutative G. In this case, Theorem 3.9 implies that d is an integer and there exist natural numbers s, t with t > 1 satisfying n = G = st2, m = (s − 1)t, and d = st, where we use the same notation as in Section 3. Let G be the set of equivalence classes of unitary representations of G, and let G† = G \ Hom(G, T). Then Theorem 3.9 implies that t divides dim π for every π ∈ G†, and # Hom(G, T) = t2. We denote by [G, G] the commutator subgroup of G. Since Hom(G, T) is identified with the dual group of G/[G, G], we have #[G, G] = s. Let p be a prime number. A p-group G is said to be extra-special if Z(G) = [G, G] ∼= Zp, where Z(G) is the center of G. Our goal in this section is to prove the following theorem. Theorem 6.1. Let G be a noncommutative finite group. Then a C∗ near-group category with G exists if and only if s = 2 and G is an extra-special 2-group. In particular t is a power of 2, n = G = 2t2, m = t, and d = 2t. For each extra-special 2-group, there exist exactly three different C∗ near-group categories. Remark 6.2. As we will see in the proof below, the three fusion categories for a given extra-special 2-group G are distinguished by the third Frobenius-Schur indicator ν3,1(ρ). Example 6.3. The representation category of the bicrossed product Hopf algebra arising from the symmetric group S4 = S3 · Z/4Z is an example of such a fusion category with the dihedral group G = D8 of order 8 (see [33, Theorem 14.40,II]). We will prove Theorem 6.1 in several steps. Assume that C is a C∗ near group category with a noncommutative G and we use the same notation as in Section 3 and Theorem 3.9. Let N = ρ(M), which is an irreducible subfactor of index d2 = s2t2. Let α′ be the outer action of G on N defined by α′g = ρ◦αg◦ρ−1. Since ρ◦αg = Ad U(g)◦ρ, we can identify the von Neumann algebra generated by N and U(G) with the crossed product N ⋊α′ G. We denote by M G the fixed point subalgebra of M under the G-action α, that is, M G = {x ∈ M; αg(x) = x, ∀g ∈ G}. Since αg ◦ ρ = ρ, the fixed point algebra M G is an intermediate subfactor between M and N with index [M : M G] = G = st2. Lemma 6.4. Let the notation be as above. Then N ⋊α′ G = M [G,G] and N ⋊α′ [G, G] = M G. 40 MASAKI IZUMI Proof. Let ι1 : N ⋊α′ G ֒→ M, ι2 : N ⋊α′ [G, G] ֒→ N ⋊α′ G, and ι3 : N ֒→ N ⋊α′ [G, G] be the inclusion maps. Then we have ρ = ι1ι2ι3ρ0, where ρ0 is ρ regarded as an isomorphism from M onto N. We have d(ι1) = √s, d(ι2) = t, d(ι3) = √s. Since N ⋊α′ [G, G] ⊂ N ⋊α′ G is a crossed product inclusion by a G/[G : G]-action, and G/[G, G] is an abelian group of order t2, the endomorphism ι2ι2 is decomposed into t2 automorphisms. Since [ρ2] = [ρρ] contains [ι1(ι2ι2)ι1], if γ is an automorphism contained in ι2ι2, then ι1γι1 is contained in ρ2. Since d(ι1γι1) = s < d(ρ) and [ρ2] =Xg∈G [αg] + (s − 1)t[ρ], the endomorphism ι1γι1 is decomposed into automorphisms, and it is the case for ι1ι2ι2 ι1 as well. Since d(ι1ι2) = √st = √#G, we get [αg]. [ι1ι2ι2 ι1] =Xg∈G This means that N ⋊ [G, G] coincides with the fixed point algebra M G. Since ρ is self-conjugate, we can show M [G,G] = N ⋊α′ G too switching the roles of (cid:3) the crossed products and fixed point algebras. Let N [G,G] be the fixed point algebra N [G,G] = {x ∈ N; α′g(x) = x, ∀g ∈ [G, G]}. Since ρ0(M [G,G]) = N [G,G], the image of ρ0ι1 is noting but N [G,G]. By the duality between the crossed product subfactor N ⊂ N ⋊α′ [G, G] = M G and the fixed point algebra subfactor N [G,G] ⊂ N, the homomorphism ι3ρ0ι1 from M [G,G] to M G contains an isomorphism from M [G:G] onto M G, say ϕ, and we have ι3ρ0 = ϕι1 by the Frobenius reciprocity. Thus (6.1) [ρ] = [ι1ι2ϕι1] = [ι1ϕ−1ι2 ι1]. Let \[G, G] be the set of equivalence classes of unitary representations of [G, G]. Then we have the following irreducible decomposition (6.2) (6.3) [ι2 ι1ι1ι2] = Mτ∈Hom(G,T) [ι1ι1] = Mσ∈\[G,G] [βτ ] ⊕Mπ∈ G† dim σ[γσ]. dim π[βπ], Lemma 6.5. With the above notation, the following hold. (1) the homomorphism βπϕγσ is irreducible for all π ∈ G and σ ∈ \[G, G]. (2) for π, π′ ∈ G and σ, σ′ ∈ \[G, G], the two homomorphisms βπϕγσ and βπ′ϕγσ′ (3) for π, π′ ∈ G, we have dim(ι1ϕ−1βπ, ι1ϕ−1βπ′) = δπ,π′. (4) for σ, σ′ ∈ \[G, G], we have dim(ι1ι2ϕγσ, ι1ι2ϕγσ′) = δσ,σ′. are equivalent if and only if π = π′ and σ = σ′. THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 41 Proof. Since ρ is irreducible, 1 = dim(ι1ι2ϕι1, ι1ι2ϕι1) = dim(ι2 ι1ι1ι2ϕ, ϕι1ι1) dim π dim σ dim(βπϕ, ϕγσ), =Xπ∈ G Xσ∈\[G,G] which shows that [βπϕ] = [ϕγσ] if and only if π = 1 and σ = 1. Since the right-hand side of dim(βπϕγσ, βπ′ϕγσ′) = dim(βπ′βπϕ, ϕγσ′γσ), is 1 if π = π′ and σ = σ′, and it is 0 otherwise, we obtain (1) and (2). (3) and (4) follow from (1),(2) and the Frobenius reciprocity. (cid:3) Let \[G, G]† = \[G, G] \ {1}. Lemma 6.6. There exists a unique bijection Φ : \[G, G]† → G† such that [ι1ι2ϕγσ] = [ι1ϕ−1βΦ(σ)ϕ] and dim Φ(σ) = t dim σ. Moreover, there exists Ψ(σ) ∈ Hom([G, G], T) for σ ∈ \[G, G]† satisfying [ι2ϕγσ] = [γΨ(σ)ϕ−1βΦ(σ)ϕ]. Proof. On one hand, we have [ρ2] =Xg∈G [αg] + (s − 1)t[ρ], and on the other hand, Since dim σ[ι1ι2ϕγσϕ−1ι2 ι1]. [ρ2] = [ι1ι2ϕι1ι1ϕ−1ι2 ι1] = Xσ∈\[G,G] [ι1ι2ϕϕ−1ι2 ι1] =Xg∈G [αg], for σ 6= 1, we have [ι1ι2ϕγσϕ−1ι2 ι1] = t dim σ[ρ]. Thus for σ 6= 1, t dim σ = dim(ι1ι2ϕγσϕ−1ι2 ι1, ρ) = dim(ι1ι2ϕγσ, ι1ϕ−1ι2 ι1ι1ι2ϕ) dim π dim(ι1ι2ϕγσ, ι1ϕ−1βπϕ). =Xπ∈ G Thanks to Lemma 6.5(3),(4), there exists unique π, which we call Φ(σ) such that [ι1ι2ϕγσ] = [ι1ϕ−1βπϕ] and dim π = t dim σ. Moreover Φ is an injection. Since Xσ∈\[G,G] (dim σ)2 = [G, G] − 1 = s − 1, † we have Xσ∈\[G,G] † (dim Φ(σ))2 = t2(s − 1) = #G − # Hom(G, T) = Xπ∈ G† (dim π)2, 42 MASAKI IZUMI and we see that Φ is a surjection. Since 1 = dim(ι1ι2ϕγσ, ι1ϕ−1βΦ(σ)ϕ) = dim(ι2ϕγσ, ι1ι1ϕ−1βΦ(σ)ϕ) dim σ′ dim(ι2ϕγσ, γσ′ϕ−1βΦ(σ)ϕ), = Xσ′∈\[G,G] (cid:3) there exists a unique σ′ ∈ Hom([G, G], T) such that [ι2ϕγσ] = [γσ′ϕ−1βΦ(σ)ϕ]. Lemma 6.7. The commutator subgroup [G, G] is abelian, and dim π = t for all π ∈ G†. Proof. Suppose that [G, G] is non-abelian. Then there exists an irreducible repre- sentation σ ∈ \[G, G] such that σ ⊗ σ contains a non-trivial irreducible represen- tation µ. Since [ι2ϕγσ] = [γΨ(σ)ϕ−1βΦ(σ)ϕ], the endomorphism γµ is contained in ϕ−1βΦ(σ)βΦ(σ)ϕ, and there exists µ ∈ G such that [γµ] = [ϕ−1βµϕ]. However, this implies [ϕγµ] = [βµϕ], which contradicts Lemma 6.5 because µ 6= 1. We recall a well-known fact that the dimension of any irreducible projective repre- sentation of a finite group does not exceed the square root of the order of the group (see [27, Problem 11.7]). (cid:3) Lemma 6.8. The commutator subgroup [G, G] coincides with the center Z(G) of G. Proof. Let σ ∈ \[G, G]†. Since [ι2ϕγσ] = [γΨ(σ)ϕ−1βΦ(σ)ϕ] and ι2ι2 is decomposed into t2 automorphisms, the endomorphism βΦ(σ)βΦ(σ) contains t2 automorphisms. Since dim Φ(σ) = t, this means that π ⊗ π = Mτ∈Hom(G,T) τ, for all π ∈ G†. Let g ∈ [G, G]. Then π(g) is a scalar for all π ∈ G† because τ (g) = 1 for all τ ∈ Hom(G, T). This implies g ∈ Z(G), and [G, G] ⊂ Z(G). Since π ∈ G† can be regarded as an irreducible projective representation of G/Z(G), the dimension of π does not exceed the square root of the order of G/Z(G). Since dim π = t, we have t2 ≤ #G/Z(G) ≤ #G/[G, G] = t2, and Z(G) = [G, G]. (cid:3) Recall that [Z(G)† = [Z(G) \ {1}. Lemma 6.9. For each character σ ∈ [Z(G)†, the induced representation IndG Z(G)σ is decomposed as IndG Z(G)σ ∼= πσ ⊕ · · · ⊕ πσ, t } { where πσ is an irreducible representation of dimension t. For σ1, σ2 ∈ [Z(G)† with σ1 6= σ2, the corresponding two induced representations are mutually disjoint. More- over, we have G† = {πσ}σ∈\Z(G) πσ1 ⊗ πσ2 ∼= πσ1σ2 ⊕ · · · ⊕ πσ1σ2, for σ1 6= σ2, †, and t z { z } THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 43 τ ∈ Hom(G, T), πσ ⊗ τ ∼= πσ, πσ ⊗ πσ ∼= Mτ∈Hom(G,T) τ, πσ ∼= πσ. a representation of G/Z(G) = G/[G, G]. Since dim π > 1 and G/[G, G] is abelian, Proof. For π ∈ G† and g ∈ Z(G) = [G, G], we have seen in the proof of the previous lemma that π(g) is a scalar. Thus the restriction of π to Z(G) is decomposed as t copies of some σ ∈ [Z(G). If σ were trivial, the representation π would reduce to this is impossible, and σ ∈ [Z(G)†. By the Frobenius reciprocity, the induced repre- sentation IndG Z(G)σ contains t copies of π. Since #G/Z(G) = t2 and dim π = t, we obtain IndG Z(G)σ ∼= π ⊕ · · · ⊕ π . z t } { This means that π ∈ G† is completely determined by its restriction to Z(G) and the other statements follow. (cid:3) Lemma 6.10. The group Z(G) is cyclic, and s + 1 is a prime power. Proof. Since t = dim πσ, Theorem 3.9 implies that the representation (V,K) has the following irreducible decomposition: (V,K) = Mσ∈\Z(G) (πσ,Kσ). † Eq.(3.8) implies j1Kσ = Kσ. Thanks to Eq.(3.19), the operator UK(h) for h ∈ G is an intertwiner between V and χh ⊗ V . Since the above lemma shows that χh ⊗ πσ is equivalent to πσ, we have UK(h)Kσ = Kσ. Thus Eq.(3.9) implies that there exists a permutation θ of order two of [Z(G)† such that j2Kσ = Kθ(σ). Eq.(3.18) implies σ ∈ [Z(G)†. Eq.(3.15) shows that if T ∈ Kσ is an isometry, we have θ(θ(θ(σ))) = σ, (6.4) Thus if s = 2, we have l(T ) = 0. In this case Z(G) ∼= Z2 and the statement holds. Assume s 6= 2 now. Then l(T ) 6= 0 for any T ∈ Kσ \ {0}. Eq(3.13) and Eq.(3.20) implies IKθ(σ) + l(T )∗l(T ) = Q. l(Kσ) ⊂ Mµ,ν∈\Z(G) † KµKνµK∗ν. , σ=θ(ν)θ(µ) Eq.(3.17) implies θ(θ(ν)θ(µ)) = θ(νµ)θ(µ). Thus [50, Theorem 6.1] implies that Z(G) is cyclic, and s + 1 is a prime power. (cid:3) Lemma 6.11. The number s is a prime and G/Z(G) is an elementary abelian s- group. In consequence, the group G is an extra-special s-group. 44 MASAKI IZUMI Proof. Since dim π = t and #G/Z(G) = t2, if we regard πσ for σ ∈ [Z(G)† as a projective representation of G/Z(G), it is faithful. Indeed, if it were not the case, there would be a normal subgroup H of G strictly containing Z(G) such that we can regard πσ as an irreducible projective representation of G/H. However, this is impossible because t = dim πσ cannot exceed the square root of the order of G/H. For each g ∈ G/Z(G), we choose a lift g, and set ω(g, h) = ghg−1h−1 ∈ Z(G) for g, h ∈ G/Z(G). Then ω : (G/Z(G))2 → Z(G) is an anti-symmetric bihomomorphism, which is independent of the choice of the lifts. Let p be a prime dividing t. We choose g ∈ G/Z(G) whose order is p. Assume that s is not equal to p. Then there exists a character σ ∈ [Z(G) such that σp 6= 1. Since σp(ω(g, h)) = 1 for all h ∈ G/Z(G), the matrix πσp(g) is a scalar. However, this would imply g = e, which is contradiction, and we obtain s = p. Since this is the case for all prime p dividing t, we see that G/Z(G) is a p-group. If there existed g ∈ G/Z(G) satisfying gp 6= e, for any σ ∈ [Z(G)† and h ∈ G/Z(G), we would have σ(ω(gp, h)) = 1, and ω(gp, h) = e. This means gp ∈ Z(G), and is again contradiction. Thus we conclude that G/Z(G) is an elementary abelian p-group. (cid:3) Lemma 6.12. The number s is 2, and j2j1 is a scalar. In particular, the group G is an extra-special 2-group. Proof. We choose an orthonormal basis {T (σ)i}t πσ as i=1 of Kσ for σ ∈ [Z(G)†, and express Since j2V (g)j∗2 = UK(g), we have the irreducible decomposition of UK as πσ(g)ijj2(T (σ)i)j2(T (σ)j)∗. πσ(g)ijT (σ)iT (σ)∗j . πσ(g) = tXi,j=1 UK(g) = Xσ∈\Z(G) † χg(h)πσ(h)ijπσ(h)pqj2T (σ)iT (σ)∗pj∗2 ⊗ j1j2T (σ)jT (σ)∗qj∗2j∗1 Since χh ⊗ πσ ∼= πσ, Eq.(3.33) takes the following form now: which is equivalent to 1 πσ(g)ijl(j2T (σ)i)l(j2T (σ)j)∗, j2j1UK(g)j∗1j∗2 ⊗ UK(g) st Xh,σ,i,j,p,q = +Xσ,ij j1j2V (g)j∗2j∗1 ⊗ V (g) st Xh,σ,i,j,p,q = +Xσ,ij πσ(g)ij(j2 ⊗ j2)l(j2T (σ)i)l(j2T (σ)j)∗(j2 ⊗ j2)∗. 1 χg(h)πσ(h)ijπσ(h)pqT (σ)iT (σ)∗p ⊗ j2j1j2T (σ)jT (σ)∗qj∗2 j∗1j2 THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 45 χg(h)πσ(h)ijπσ(h)pqT (σ)iT (σ)∗p ⊗ j2j1j2T (σ)jT (σ)∗qj∗2j∗1 j2, This and Lemma 6.9 imply and = Xσ j1j2πσ(g)j∗2j∗1 ⊗ πσ(g) st Xh,σ,i,j,p,q 1 Xσ6=τ j1j2πσ(g)j∗2j∗1 ⊗ πτ (g) =Xσ,ij πσ(g)ij(j2 ⊗ j2)l(j2T (σ)i)l(j2T (σ)j)∗(j2 ⊗ j2)∗. Recall that for σ ∈ [Z(G)†, the map G × G ∋ (g, h) 7→ σ(ghg−1h−1) ∈ T induces a non-degenerate anti-symmetric bicharacter of G/[G, G]. Thus for each g ∈ G, there exists a map ϕσ : G → G satisfying χg(h) = σ(ϕσ(g)hϕσ(g)−1h−1). Note that the element ϕσ(g) is unique up to a multiple of a central element of G. Using this, we have Xh∈G =Xh∈G =Xh∈G =Xh∈G = tXa,b=1 n dim πσ χg(h)πσ(h)ijπσ(h)pq σ(ϕσ(g)hϕσ(g)−1h−1)πσ(h)ijπσ(h)pq πσ(ϕσ(g)hϕσ(g)−1h−1h)ijπσ(h)pq πσ(ϕσ(g))iaπσ(h)abπσ(ϕσ(g))jbπσ(h)pq πσ(ϕσ(g))ipπσ(ϕσ(g))jq. Thus we obtain Xσ∈\Z(G) † j1j2πσ(g)j∗2j∗1 ⊗ πσ(g) = Xσ∈\Z(G) πσ(ϕσ(g)) ⊗ j2j1j2πσ(ϕσ(g))j∗2j∗1j∗2. † This implies that there exists a scalar cσ(g) ∈ T such that πσ(ϕσ(g)) = cσ(g)j1j2πθ(σ)(g)j∗2j∗1, j2j1j2πσ(ϕσ(g))j∗2j∗1 j∗2 = cσ(g)πθ(σ)(g). Since (j2j1)3 = I and j1πσ(g)j∗1 = πσ(g), the above two equations are equivalent. In the above argument, we have seen that χg(h)πσ(h) = πσ(ϕσ(g)hϕσ(g)−1) holds. On the other hand, Eq.(3.19) implies that χg(h)πσ(h) = j2πθ(σ)(g−1)j∗2πσ(h)j2πθ(σ)(g)j∗2 46 MASAKI IZUMI holds, and so j2πθ(σ)(g−1)j∗2 ∼ j1j2πθ(σ)(g)j∗2j∗1 , where ∼ means that two matrices are proportional. Let R = j2j1, which is a unitary of period three. Then if σ is replaced with θ(σ) and g is replaced with g−1 in the above, we obtain πσ(g) ∼ Rj2πθ(θ(σ))(g−1)j∗2R∗ = Rj2πθ(σ)(g−1)j∗2R∗ = R∗πθ(σ)(g−1)R. Since R3 = I and θ(θ(θ(σ))) = σ, this implies πσ(g) ∼ πσ(g−1), and g2 ∈ Z(G) for all g ∈ G. Therefore s = 2. Let σ be the unique element in [Z(G)†, and let π = πσ. The above argument shows that there exists τ ∈ Hom(G, T) such that R∗π(g)R = τ (g)π(g). Since R has period three, we have τ 3 = 1, and τ = 1 as Hom(G, T) = \G/Z(G) is an elementary 2-group. This implies that R is a scalar. (cid:3) We have seen that G is an extra-special 2-group of order st2 = 2t2. Let σ be a unique non-trivial character of Z(G) ∼= Z2. Then πσ is a unique irreducible represen- tation of dimension t, and we denote (πσ,Kσ) = (π,Kπ) for simplicity. The set G† is a singleton {π}, and we can identify (V,K) with (π,Kπ). The period three unitary j2j1 is a scalar, which we denote by ζ ∈ T, and we have j2 = ǫζj1. This implies that UK = V , and χh(g) is determined by π(g)π(h) = χh(g)π(h)π(g). Eq.(6.4) implies l = 0. In summary, every information of C is encoded in the tuple (K = Kπ, V = π, UK = π, j1, j2 = ǫζj1, χ, l = 0). Now we proceed to the reconstruction part. Recall that there are exactly two isomorphism classes of extra-special 2-groups for a given order 22k+1, where k is a natural number (see [49, Exercise 5.3.7]). In the both cases, there exists only one irreducible representation whose dimension is larger than one, and its dimension is necessarily 2k. Lemma 6.13. Let G be an extra-special 2-group of order 22k+1, and let (π,Kπ) be the unique irreducible representation G of dimension 2k. Let j be an anti-unitary on Kπ satisfying j2 = ǫ ∈ {1,−1} and jπ(g) = π(g)j for all g ∈ G, and let ζ be a third root of unity. Let χ be a bicharacter of G determined by π(g)π(h) = χh(g)π(h)π(g). Then (K = Kπ, V = π, UK = π, j1 = j, j2 = ǫζj, χ, l = 0), is an admissible tuple. Proof. We will show only Eq.(3.21),(3.33),(4.1),(4.2),(4.3),(4.4) because the other conditions are easy to verify. Recall that we have n = G = 22k+1, m = 2k, d = 2k+1. Since ghg−1h−1 ∈ Z(G) ∼= Z2, Eq.(3.21) holds. Since d2 = 2n, Eq.(4.1) is auto- matically satisfied. Since l = 0, Eq.(4.3) is automatically satisfied, and the left-hand sides of Eq.(4.2) and Eq.(4.4) are 0. Since π⊗π⊗π contains no trivial representation, THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 47 the right-hand side of Eq.(4.4) is 0. For an orthonormal basis {Ti}2k i=1 of K = Kπ, let π(g) = π(g)ijTiT ∗j . 2kXi,j=1 Then hUK(h)T ′′, j1(T )ij1(UK(h)T ′) π(h)ijπ(h)abhTi, j1(T )ihT ′′, TjihTb, T ′ij1(Ta) δi,aδj,bhTi, j1(T )ihT ′′, TjihTb, T ′ij1(Ta) = hT ′′, T ′ij2 1(T ) = ǫhT ′′, T ′iT, and the right-hand side of Eq.(4.2) is 0. The left-hand side of Eq.(3.33) is π(g) ⊗ π(g). On the other hand, the right-hand side of Eq.(3.33) is 1 dXh∈G = 1 d = 2kXi,j,a,b=1Xh∈G 2kXi,j,a,b=1 n d dim π 1 d 1 d 2kXi,j,x,y=1Xh∈G 2kXi,j,x,y=1Xh∈G 2kXi,j,x,y,a,b=1Xh∈G 2kXi,j,x,y,a,b=1 d dim π 1 d n = = = = π(g) ⊗ j1π(g)j∗1 = π(g) ⊗ π(g), χh(g)π(h)ijπ(h)xyTiT ∗x ⊗ j1TjT ∗y j∗1 π(ghg−1)ijπ(h)xyTiT ∗x ⊗ j1TjT ∗y j∗1 π(g)iaπ(h)abπ(g)jbπ(h)xyTiT ∗x ⊗ j1TjT ∗y j∗1 δa,xδb,yπ(g)iaπ(g)jbTiT ∗x ⊗ j1TjT ∗y j∗1 (cid:3) which finishes the proof. Proof of Theorem 6.1. We use the same notation as in the above lemma. Since π is irreducible, if j′ is another anti-unitary on Kπ satisfying j′π(g) = π(g)j′ for any g ∈ G, there exists c ∈ T with j′ = cj. Choosing a square root c1/2 of c, we get j′ = c1/2jc−1/2. Therefore the equivalence class of the tuple in the above lemma is determined by the third root of unity ζ. Theorem 4.9 shows that there are exactly three C∗ near-group categories for a given extra-special 2-group G as far as j exists. Therefore to finish the proof, it suffices to show the existence of j satisfying the condition of the above lemma. 48 MASAKI IZUMI When G = 8, we have dim π = 2, and the anti-unitary j with j2 = 1 is unitarily equivalent to the complex conjugation of C2. Thus the condition jπ(g) = π(g)j is equivalent to the existence of a real representation equivalent to π. The dihedral group D8 of order 8 satisfies this condition. When dim π = 2 and j2 = −1, the condition jπ(g) = π(g)j is equivalent to π(G) ⊂ SU(2). The quaternion group Q8 satisfies this condition. Now we consider the general case. Note that every extra-special 2-group of order 22k+1 is a central product of either k copies of D8, or k − 1 copies of D8 and 1 copy of Q8 (see [49, Exercise 5.3.7]). Moreover, the irreducible representation π can be constructed from 2-dimensional irreducible representations of D8 and Q8 by tensor product. Thus there exists j as in the previous lemma with ǫ = 1 in the former case, and ǫ = −1 in the latter case. This finishes the proof. (cid:3) We denote by CG,ζ the C∗ near-group category arising from the tuple (K = Kπ, V = π, UK = π, j1 = j, j2 = ǫζj, χ, l = 0). Corollary 6.14. The fusion category CG,ζ is group theoretical. Proof. We may assume CG,ω ⊂ End0(M) and we use the same notation as before. We claim that there exists an abelian normal subgroup H ⊳ G of order 2k+1 with Z(G) ⊂ H and χh1(h2) = 1 for any h1, h2 ∈ H. Indeed, since G is a central product of copies of D8 and Q8, it suffices to verify the claim for G = D8 and G = Q8, which is straightforward. Let M H and N H be the fixed point algebras We claim M H = N ⋊α′ H. Indeed, we have M H ⊂ M Z(G) = N ⋊α′ G, and every element in N ⋊α′ G is uniquely expands as {x ∈ M; αh(x) = x, ∀h ∈ H}, {x ∈ N; α′h(x) = x, ∀h ∈ H}. U(g)ρ(xg). Xg∈G Thus Eq.(3.22) implies N ⋊α′ H ⊂ M H . Since, [M H : N] = [M : N]/H = 2k+1 = H = [N ⋊α′ H : N], we get the claim. Let ν : M H ֒→ M and µ : N ֒→ M H be the inclusion maps. Then we have ρ = νµρ0, and the image of ρ0ν is N H . By the duality between the crossed product inclusion N ⊂ N ⋊α′ H = M H and the fixed point algebra inclusion N H ⊂ N, the endomorphism µρ0ν ∈ End0(M H ) contains an automorphism, say θ ∈ Aut(M H ). Thus the Frobenius reciprocity implies that [µρ0] = [θν] and we get [ρ] = [νθν]. Since H is abelian, the endomorphism νν is decomposed into automorphisms, and so is νρν. Since αg normalizes αH, it globally preserves M H , and there exists βg ∈ Aut(M H ) satisfying αgν = νβg. This implies that ναgν = ννβg is decomposed into automorphisms too. Thus the fusion category generated by νρν and ναgν is a pointed category, which is categorically Morita equivalent to CG,ζ. (cid:3) THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 49 Remark 6.15. Let the notation be as in the above proof. (1) The subgroup H is not unique. For D8, there are three possibilities; two of them are isomorphic to Z2 × Z2 and the other is isomorphic to Z4. For Q8 there are three possibilities, and they are all isomorphic to Z4. (2) Let L be the group generated by the automorphisms contained in νρν and ναgν in Out(M H ). Since the global dimension of CG,ζ is G + d2 = 22k+1 × 3, the order of L is 22k+1×3. The group Z2k 2 ⋊ S3 is a canonical candidate of L as we see below. When k = 1, this group is isomorphic to S4. 2 ⋊SL(2, 2) ∼= Z2k Example 6.16. Let F2 be the finite field of order 2, and we consider the following subgroups of SL(2 + k, 2): L = { z u (cid:19) ∈ SL(2, 2)},  ∈ SL(2 + k, 2); a, b ∈ Fk 2, (cid:18) x y  ∈ SL(2 + k, 2); y ∈ F2, a ∈ Fk a 0k Ik 2}. . The normalizer of A in L is y 1 0T k 2 ⋊ SL(2, 2) ∼= Z2k 2 ⋊ S3 and A ∼= Zk+1 1 0 0T k y 1 0T k a b Ik 2 Then L ∼= Z2k and NL(A) = A ⋊ K with 2, y ∈ F2},  ∈ L; a, b ∈ Fk  ∈ L; b ∈ Fk  ∈ SL(2 + k, 2); a ∈ Fk 2} ∼= Zk 2. 0k b Ik 2}. 1 0 0T k 0 1 0T k 0 1 0T k a 0k Ik x z 0T k y u 0T k a b Ik 1 0 0T k A = { NL(A) = { K = { B = { 1 0 0T k Let Let P be a type III factor and let β : L → Aut(P ) be a map that induces an injective homomorphism from L into Out(P ). We assume that the restriction of β to A is an action and βk normalizes βA for any k ∈ K. Let M = P ⋊β A, and let ν : P ֒→ M be the inclusion map, let and let θ = βf . We claim that ρ = νθν generates a C∗ near-group category with a noncommutative G with G = 22k+1. Indeed, since f−1 = f , the endomorphism ρ is self-conjugate. Since 0k 0k Ik 0 1 0T k 1 0 0T k f =  ∈ L, dim(ρ, ρ) = dim(ννβf , βf νν) = Xg,h∈A (βf gf , βh) = 1, 50 MASAKI IZUMI ρ is irreducible. Since βk for k ∈ K normalizes βA, it extends to an automorphism βk ∈ Aut(M). We denote by β the dual action of β restricted to A. Then since f we get 1 0 0T k 1 1 0T k a 0k Ik 0 1 0T k a 0k Ik 1 0 0T k 0 1 0T k 1 0 0T k 0 1 0T k 1 1 0T k f  f = [ρ2] = [νθννθν] =Xg∈A =Xk∈K [ βkνν] +Xh∈B  f =  = [νβf gf ν] =Xk∈K [νθν] = Xτ∈ A, k∈K 1 0 0T k 0k a Ik 0k a Ik  ∈ K,  f [νβkν] +Xh∈B a 0k Ik 1 1 0T k [ βk βτ ] + 2k[ρ]. 1 0 0T k 1 1 0T k a 0k Ik  , [νβhf hν] This shows the claim with G equal to {[ βk tension βτ ]}τ∈ A, k∈K ⊂ Out(M), which is an ex- 0 → A → G → K → 0. It is an interesting problem to compute a cohomological invariant for β, and identify it with the third root of unity ζ in CG,ζ. The extra-special 2-group G obtained in this way is always a central product of k copies of D8. To obtain the other type of an extra-special 2-group of the same order, we should replace A with a subgroup A1 isomorphic to Z4 × Zk−1 . For example, we could choose A1 to be the group generated by B and 2  1 0 0T k 1 1 0T k 0k e1 Ik  , e1 = (1, 0,· · · , 0) ∈ Fk 2. Note that the group G is not uniquely determined by the pair (L, A1), but it also de- pends on the choice of β. The pointed category generated by βL carries a cohomology class in H 3(L, T). On the other hand, there is a restriction map from H 3(L, T) to H 3(SL(2, 2), T) ∼= Z6. We conjecture that H 3(SL(2, 2), T) parameterizes 6 different C∗ near-group categories CG,ζ with G = 22k+1. 7. Irrational case In the rest of this note, we investigate the structure of C∗ near-group categories in the irrational case. This section is devoted to laying the basis for the classification of such categories. Assume that we have a C∗ near group category as in Section 3 with irrational d. Then Theorem 3.9 implies that G is abelian, the multiplicity parameter m is a multiple of n = G, and the symmetric bicharacter h·,·i : G × G → T defined by hg, hi = χg(h) is non-degenerate. In what follows, we use additive notation for G. We have two unitary representations V (g), UK(g) in K, equivalent to the regular representation of G, and two anti-unitaries j1, j2 satisfying the conditions stated in THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 51 Section 3. Let v0(g) = V (g), v1(g) = UK(g), v2(g) = (j1j2)∗UK(g)j1j2, and w = j1j2. Then they satisfy the following relations: w3 = I, j2 1 = v−i(g), j1wj−1 1 = ǫ ∈ {1,−1}, j1vi(g)j−1 1 = w∗, w∗vi(g)w = vi+1(g), vi+1(g)vi(h) = hh, givi(h)vi+1(g), where i is understood as an element of Z/3Z. Lemma 7.1. Let G be a finite abelian group of order n, and let h·,·i : G × G → T be a non-degenerate symmetric bicharacter. Let H(G) be the universal C∗-algebra generated by three unitary representations v0, v1, v2 of G and a unitary w satisfying the following commutation relations: (7.1) (7.2) (7.3) vi+1(g)vi(h) = hh, givi(h)vi+1(g), w∗vi(g)w = vi+1(g), w3 = 1, where i is understood as an element of Z/3Z. Then there exist exactly 3n irreducible representations of H(G), and they are of the following form: The representation space is ℓ2(G) and (7.4) (7.5) (7.6) (7.7) πa,c(v0(g))f (h) = hg, hif (h), πa,c(v1(g))f (h) = f (h + g), πa,c(v2(g))f (h) = a(h)a(h − g)f (h − g), a(h)hh, kif (k), πa,c(w)f (h) = c √nXk where a : G → T and c ∈ T satisfy (7.8) a(g + h)hg, hi = a(g)a(h), (7.9) a(g) = √n. c3Xg∈G Proof. Note that the dimension of the ∗-algebra generated by v0,v1, v2, and w is 3n3. Thus it suffices to show that there exist 3n irreducible representations of H(G) of dimension n as above. Since h·,·i is a symmetric 2-cocycle, it is a coboundary and a function a satisfying Eq.(7.8) certainly exists. We choose one of them and denote it by a. Let a′ : G → T. Then a′ satisfies Eq.(7.8) if and only if a′/a is a character, and so there are exactly n such functions. For a function f on G, we define its Fourier transform F f (h) = f (h) by f (g) = 1 √nXh∈G hg, hif (h). 52 MASAKI IZUMI Then Eq.(7.8) implies a(g) = a(0)a(g). (7.10) Since the Fourier transform preserves the ℓ2-norm, we have a(0) = 1. Thus for a given a, there exists exactly three c ∈ T satisfying Eq.(7.1). We choose one of them. Let µ(g)f (h) = hg, hif (h) and (g)f (h) = f (h + g). Then the commutation relations (g)µ(h) = hg, hiµ(h)(g) and µ(g)F = F (g) hold. Let Af (h) = a(h)f (h). Then we have µ(g)AF = AF (g). To show that πa,c is a well-defined irreducible representation, it suffices to show c3(AF )3 = 1 because if it is the case, we can put πa,c(w) = cAF , πa,c(v0(g)) = µ(g), πa,c(v1(g)) = (g), and πa,c(v2(g)) = πa,c(w)∗πa,c(v1(g))πa,c(w). Indeed, F AF f (g) = a(g + k)f (k) a(g)hg, kia(k)f (k) hg, hia(h)hh, kif (k) = a(0) a(g + k)f (k) = 1 n Xh,k∈G √n Xk∈G = a(0)A∗F∗A∗f (g). a(0) = 1 √nXk∈G √n Xk∈G This shows (AF )3 = a(0). We now show that if πa,c and πa′,c′ are equivalent then (a, c) = (a′, c′). Note that they coincide on C∗{v0, v1}, which are already irreducible. Thus if they are equivalent, they must coincide on H(G), which implies (a, c) = (a′, c′). Remark 7.2. We always have the following relation: (cid:3) πa,c(v2(g)) = a(g)πa,c(v1(−g))πa,c(v0(−g)). We introduce an anti-linear involutive ∗-isomorphism κ of H(G) by setting κ(vi(g)) = v−i(g), κ(w) = w−1. We say that (π, j,K) is a covariant representation of (H(G), κ) if (π,K) is a repre- sentation of H(G) and j is an anti-unitary of K such that π(κ(x)) = jπ(x)j−1 holds for all x ∈ H(G) and j2 = ǫ is a scalar. We say that a covariant representation is even (resp. odd) if ǫ = 1 (resp. ǫ = −1). We would like to classify all even and odd covariant representations. We fix a : G → T satisfying Eq.(7.8) and a(g) = a(−g) for all g ∈ G, and set aχ(g) = a1(g)χ(g) for all χ ∈ G. Let cχ,i, i = 0, 1, 2 be the three solutions of Eq.(7.1) for aχ. Note that baχ(0) = daχ−1(0) holds. Indeed, choosing h ∈ G satisfying χ(g) = hg, hi for all g ∈ G, we obtain Thus we may and do assume cχ,i = cχ−1,i. baχ(0) = a(−h) = a(0)a(−h) = a(0)a(h) = a(h) = daχ−1(0). THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 53 Let (π,K) be a representation of H(G). Then Lemma 7.1 shows that up to unitary equivalence, we may assume the following: K = ℓ2(G) ⊗ K0, π(v0(g)) = µ(g) ⊗ I, π(v1(g)) = (g) ⊗ I, where eg is the projection from ℓ2(G) onto Cδg. K0 is decomposed as π(w) =Xg∈G (eg ⊗ CA(g))(F ⊗ I), K0 = Mχ∈ G, i=1,2,3 Kχ,i Let j be an anti-linear unitary of K satisfying j2 = ǫ. Then (π, K, j) is a covariant and A(g)Kχ,i = aχ(g)I, CKχ,i = cχ,iI. representation of (H(G), j) if and only if (7.11) jπ(v0(g)) = π(v0(g))j, (7.12) jπ(w) = π(w)−1j. Indeed, it is clear that the two conditions Eq.(7.11),(7.12) are necessary. Assume conversely that they are satisfied. Then jπ(v1(g))j−1 = jπ(w)−1π(v0(g))π(w)j−1 = π(w)jπ(v0(g))j−1π(w)−1 = π(w)π(v0(g))π(w)−1 = π(v2(g)), which shows that (π, H, j) is a covariant representation. The equation Eq.(7.11) is equivalent to the condition that there exists a family of anti-unitaries j(g) on H such that jδh ⊗ ξ = δ−h ⊗ j(h)ξ. The condition j2 = ǫ is equivalent to j(−g)j(g) = ǫ. Under these conditions, we have (7.13) jπ(w)δh ⊗ ξ = hk, hiδk ⊗ CA(k)ξ = 1 √n jXk∈G 1 hk, hiδ−k ⊗ j(k)CA(k)ξ, √nXk∈G √nXk∈G hk, hiδ−k ⊗ A(−h)∗C∗j(h)ξ. j(k)CA(k) = A(−h)∗C∗j(h). 1 (7.14) π(w)−1jδh ⊗ ξ = π(w)−1δ−h ⊗ j(−h)ξ = Thus Eq.(7.12) is equivalent to We claim that Eq.(7.15) is equivalent to the following two: (7.15) (7.16) (7.17) j(0)C = C∗j(0), j(h) = A(−h)j(0) = j(0)A(h)∗. Assume Eq.(7.15) first. Then for h = k = 0, we get Eq.(7.16). This together with Eq.(7.15) for the case k = 0 and the case h = 0 implies Eq.(7.17). It is straightforward to show that Eq.(7.16) and (7.17) imply Eq.(7.15), and the claim is shown. Assuming these equivalent conditions, we known that j(−h)j(h) = ǫ is equivalent to j(0)2 = ǫ. Summing up the above argument, now we have the following lemma: 54 MASAKI IZUMI Lemma 7.3. Every even covariant representation of (H(G), κ) is a direct sum of the following covariant representations (π, H, j): (1) H = ℓ2(G), π = πχ,c with χ2 = 1, and jδh = aχ(h)δ−h. (2) H = ℓ2(G) ⊕ ℓ2(G), π = πχ,c ⊕ πχ−1,c with χ2 6= 1 and j(xδh ⊕ yδk) = yaχ−1(k)δ−k ⊕ xaχ(h)δ−h. Every odd covariant representation of (H(G), κ) is a direct sum of the following covariant representations (π, H, j): H = ℓ2(G) ⊕ ℓ2(G), π = πχ,c ⊕ πχ−1,c and j(xδh ⊕ yδk) = −yaχ−1(k)δ−k ⊕ xaχ(h)δ−h. We get back to the C∗ near-group category C with irrational d. We may identify K with ℓ2(G) ⊗ K0 as above. We denote by Th(ξ) ∈ K the element corresponding to δh ⊗ ξ. Lemma 7.4. For a given C∗ near group category with a finite abelian group G and irrational d as in Section 3, there exist a non-degenerate symmetric bicharacter h·,·i : G × G → T, an anti-unitary J, and mutually commuting unitaries A(g), C acting on K0 satisfying J 2 = ǫ, JC = C∗J, JA(g) = A(−g)∗J, A(g)A(h) = hg, hiA(g + h), 1 √nXg∈G A(g) = C−3, V (g)Th(ξ) = hg, hiTh(ξ), UK(g)Th(ξ) = Th−g(ξ), j1(Th(ξ)) = T−h(A(−h)Jξ), ǫ 1 j2(Th(ξ)) = j1j2(Th(ξ)) = √nXk∈G √nXk∈G √nXk∈G 1 )UK(g)(j2 ◦ j−1 j2j1(Th(ξ)) = 1 hh, kiTk(C∗Jξ), hh, kiTk(CA(k)ξ), hh, kiTk(C∗A(h)∗ξ), (j2 ◦ j−1 = Th+g(A(g + h)A(h)∗ξ) = UK(−g)V (−g)Th(A(g)ξ). 1 )∗Th(ξ) Λ ∋ t 7→ t ∈ Λ satisfying A(g)et = a(g)χt(g)et, and Cet = ctet, Jet = ǫtet with Moreover we can choose an orthonormal basis {et}t∈Λ of K0 with an involution (1) a(−g) = a(g), a(g)a(h) = hg, hia(g + h), (2) χt ∈ G, χt = χ−1 (3) ct = ct,Pg∈G a(g)χt(g) = √nc−3 (4) ǫt ∈ {1,−1}, ǫtǫt = ǫ. , , t t When ǫ = 1, we can arrange the basis so that ǫt = 1 for all t ∈ Λ. THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 55 Now we determine the form of ρ(Tg(ξ)). Since ρ(Tg(ξ)) = ρ(U(−g))ρ(T0(ξ)), we have l(Tg(ξ)) = ((j2 ◦ j−1 1 )UK(−g)(j2 ◦ j−1 1 )∗ ⊗ UK(−g))l(T0(ξ)) = (AUK(g)A∗ ⊗ UK(−g))l(T0(ξ)), where ATh(ξ) = Th(A(h)ξ). Thus l(Tg(ξ)) is determined by l(T0(ξ)). Moreover, l(T0(ξ)) satisfies αg(l(T0(ξ))) = l(T0(ξ)) and UK(g)l(T0(ξ))UK(g)∗ = l(T0(ξ)). Let Kh be the linear span of {Th(ξ); ξ ∈ K0}, and let Qh be the projection from K onto Kh. This notation for h = 0 is consistent with the previous one as we identify ξ ∈ K0 in the previous sense with T0(ξ) in what follows. Then l(T0(ξ)) =Xk∈G =Xk∈G l(T0(ξ))Qk =Xk∈G UK(−k)l(T0(ξ))Q0UK(k). l(T0(ξ))UK(−k)Q0UK(k) Therefore ρ(Tg(ξ)) is determined by l(T0(ξ))Q0 ∈ K2K∗0. Since αg(l(T0(ξ))Q0) = l(T0(ξ))Q0, we have l(T0(ξ))Q0 ∈Mh∈G KhK−hK∗0, and there exists a family of linear maps Bh : K0 → K0K0K∗0 satisfying (UK(−h) ⊗ UK(h))(A(h) ⊗ I)Bh(ξ). l(T0(ξ))Q0 =Xh∈G Now we can write down l(Tg(ξ)) in terms of Bh(ξ). Lemma 7.5. There exists a family of linear maps Bh : K0 → K2 (7.18) 0K∗0 satisfying hg, ki(UK(−h − k) ⊗ UK(h))(A(h) ⊗ I)Bg+h(ξ)UK(k). l(Tg(ξ)) =Xh,k Proof. This follows from × (UK(−h) ⊗ UK(h))(A(h) ⊗ I)Bh(ξ)UK(k) l(Tg(ξ)) =Xh,k (cid:0)AUK(g)A∗UK(−k) ⊗ UK(−g)(cid:1) =Xh,k =Xh,k =Xh,k hg, ki(UK(−h − k) ⊗ UK(h))(A(h) ⊗ I)Bg+h(ξ)UK(k). (UK(−h − k) ⊗ UK(h))(A(h + k)A(g + h + k)∗A(h + g) ⊗ I)Bg+h(ξ)UK(k) hg + h, ki(UK(−h − k) ⊗ UK(h))(A(h + k)A(k)∗ ⊗ I)Bg+h(ξ)UK(k) (cid:3) 56 MASAKI IZUMI 8. Polynomial equations for the irrational case Let G be a finite abelian group of order n, let m be a multiple of n, and let ǫ ∈ {1,−1}. Let h·,·i, K, K0, A(g), C, and J satisfy the conditions in the statement of Lemma 7.4, and let Bg : K0 → K0K0K∗0 be a linear map. We defined l → KKK∗ by Eq.(7.18). We will deuce the polynomial equations classifying the corresponding C∗ near-group categories in terms of the above data. We set Bg(ξ) = 1 √nXh∈G hg, hiBh(ξ). Lemma 8.1. Eq.(3.14), (3.15), (3.16), and (3.17) are equivalent to T0(Jes)∗T0(es)∗Bh(ξ) = − Xs,h Bh+q(η)∗Bh+p(ξ) = δp,qhξ, ηiQ0 − Xh √n d T0(C∗Jξ)∗. 1 d T0(C∗Jη)T0(C∗Jξ)∗. (8.1) (8.2) (8.3) (8.4) Proof. Eq.(3.14) is T0(es)Bg(C∗Jξ)∗T0(Jes) j1(Tl(es))∗Tl(es)∗l(Tg(ξ)) (X(h) ⊗ X(h))B−g(X(h)∗A(−g)Jξ)X(h)∗ =Xs,t Bg(ξ)∗T0(es)T0(et)T0(Jet)T0(es)∗. Bg(ξ) =Xs j2(Tg(ξ))∗Q0 +Xl,s T0(C∗Jξ)∗ +Xl,s T0(C∗Jξ)∗ +Xl,s T0(C∗Jξ)∗ +Xl,s T0(C∗Jξ)∗ +Xl,s T−l(A(−l)Jes)∗Tl(es)∗l(Tg(ξ)) 0 = = = = = ǫn d √n d √n d √n d √n d T0(A(−l)Jes)∗T0(es)∗(A(l) ⊗ I)Bg+l(ξ) T0(JA(l)∗es)∗T0(A(l∗)es)∗Bg+l(ξ) T0(Jes)∗T0(es)∗Bg+l(ξ). THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 57 Eq.(3.15) is Qkj2(Tq(η))j2(Tp(ξ))∗Qk n dXk δp,qhξ, ηiQ = l(Tq(η))∗l(Tp(ξ)) + =Xh,k hp − q, kiUK(k)∗Bh+q(η)∗Bh+p(ξ)UK(k) dXk hp − q, kiTk(C∗Jη)Tk(C∗Jξ)∗, + 1 which is equivalent to Bh+q(η)∗Bh+p(ξ) = δp,qhξ, ηiQ0 − 1 d T0(C∗Jη)T0(C∗Jξ)∗. Xh Eq.(3.16) is l(Tg(ξ))∗Th+k(es)T−h(et)Th(A(h)Jet)Th+k(es)∗ hg, kiUK(k)∗Bg+h(ξ)∗T0(A(h)∗es)T0(et)Th(A(h)Jet)T0(es)∗UK(h + k) l(T−g(A(−g)Jξ)) = Xh,k,s,t = Xh,k,s,t = Xh,k,s,t × T0(A(−h)∗es)T0(et)T−h(A(−h)Jet)T0(es)∗UK(k) = Xh,k,s,t × T0(es)T0(et)T0(Jet)T0(es)∗A(−h)∗UK(k), hg, h + kiUK(h + k)∗Bg−h(ξ)∗ hg, h + ki(UK(−h − k) ⊗ UK(h))(I ⊗ A(−h))Bg−h(ξ)∗ which is equivalent to (A(h) ⊗ I)Bh−g(A(−g)Jξ) = hg, hi(I ⊗ A(−h))Bg−h(ξ)∗Xs,t Replacing g with g + h, we get T0(es)T0(et)T0(Jet)T (es)∗A(−h)∗. hh, hi(A(h) ⊗ A(−h)∗)B−g(A(−h)A(−g)Jξ)A(−h) = Bg(ξ)∗Xs,t T0(es)T0(et)T0(Jet)T0(es)∗. 58 MASAKI IZUMI Eq.(3.17) is hg + h, kiT−h−k(es)UK(−k)Bg+h(ξ)∗(A(k)A(h + k)∗)T0(A(h + k)Jes)UK(−h) Tl(es)l(Tg(ξ)∗)T−l(A(−l)Jes) hg + h, kiT−h−k(es)UK(−k)Bg+h(ξ)∗T0(A(k)Jes)UK(−h) hg − k, hi(UK(−h − k) ⊗ UK(h))T0(es)Bg−k(ξ)∗T0(A(−h)Jes)UK(k) l(j2(Tg(ξ))) =Xl,s =Xh,k,s =Xh,k,s =Xh,k,s = ǫXh,k,s hg − k, hi(UK(−h − k) ⊗ UK(h))T0(A(h)Jes)Bg−k(ξ)∗T0(es)UK(k). √nXl √nXl √nXh,k,l = ǫXh,k hg, lil(Tl(C∗Jξ)) hg, liXh,k hg − k, l − hi(Uk(−h − k) ⊗ UK(h))(A(h) ⊗ I)Bl(C∗Jξ)UK(k) hg − k, hi(Uk(−h − k) ⊗ UK(h))(A(h) ⊗ I) Bg−k(C∗Jξ)UK(k), hl, ki(Uk(−h − k) ⊗ UK(h))(A(h) ⊗ I)Bh+l(C∗Jξ)UK(k) ǫ = = ǫ ǫ The left-hand side is which shows that Eq.(3.17) is equivalent to Bg(C∗Jξ) =Xs T0(Jes)Bg(ξ)∗T0(es). (cid:3) Remark 8.2. By Fourier transform, Eq.(8.1) and (8.2) are equivalent to the following two: T0(Jes)∗T0(es)∗ B0(ξ) = − 1 d T0(C∗Jξ)∗, Xs (8.6) Q0 − Eq.(8.3) is equivalent to the following two: Bh(η)∗ Bh(ξ) = hξ, ηi n δh,0 d T0(C∗Jη)T0(C∗Jξ)∗. (8.5) (8.7) (8.8) Bg(A(h)ξ) = (A(h) ⊗ A(h))Bg(ξ)A(h)∗, T0(es)T0(Jet)∗T0(et)∗T0(es)∗B−g(A(−g)Jξ). Bg(ξ)∗ = ǫXs,t THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 59 Eq.(8.4) is the equivalent to the following under the presence of Eq.(8.8): (8.9) Bg(ξ) = ǫXs,t,r(cid:0)T0(et)∗T0(es)∗B−g(A(−g)Cξ)T0(er)(cid:1)T0(Jer)T0(es)T0(Jet)∗. Note that T0(et)∗T0(es)∗B−g(A(−g)Cξ)T0(er) is already a scaler. Lemma 8.3. Eq.(3.33) is equivalent to (8.10) Xr Bg(er) Bg(er)∗ = Proof. Note that we have 1 n δg,0 Q0 ⊗ Q0 − d Xs,t hg, hiShS∗h +Xh,t U(g) =Xh Th−g(et)Th(et)∗. T0(es)T0(Jes)T0(Jet)∗T0(et)∗. In view of the proof of Lemma 3.15, it suffices to show that the above equality is equivalent to j2 ◦ j−1 1 UK(g)(j2 ◦ j−1 1 )∗ ⊗ UK(g) =Xl hg, liQρ(SlS∗l )Q +Xl,s l(Tl−g(es))l(Tl(es))∗. The left-hand side is Xh,k,s,t Tg+h(A(g + h)A(h)∗es)Th(es)∗ ⊗ Tk−g(et)Tk(et)∗ = Xh,k,s,t hg, hiTg+h(A(g)es)Th(es)∗ ⊗ Tk−g(et)Tk(et)∗. The first term of the right-hand side is = = Xl hg, liUK(l)ρ(S0S∗0 )UK(l)∗ d Xl,p,q,s,t 1 d Xl,p,q,s,t d Xh,k,p,s,t d Xh,k,p,s,t = 1 1 1 hg, liUK(l)Tp(es)T−p(A(−p)Jes)T−q(A(−q)Jet)∗Tq(et)∗UK(l)∗ hg, liTp−l(es)Tq−l(et)∗ ⊗ T−p(A(−p)Jes)T−q(A(−q)Jet)∗ = hg, h + kiTh+k+p(es)Th(et)∗ ⊗ T−p(A(−p)Jes)Tk(A(k)Jet)∗ hg, h + kiTh+p(es)Th(et)∗ ⊗ Tk−p(A(k − p)Jes)Tk(A(k)Jet)∗ 60 MASAKI IZUMI The second term of the right-hand side is (UK(−h − k′) ⊗ UK(k′))(hl − g, hiA(k′) ⊗ I)Bk′+l−g(es) hg, hi(UK(−h − k′) ⊗ UK(k′))(A(k′) ⊗ I) Xh,k,k′,l,s × Bl+k(es)∗(hl,−hiA(k)∗ ⊗ I)(UK(h + k) ⊗ UK(−k)) = Xh,k,k′,l,s × Bl+k′−k−g(es)Bl(es)∗(A(k)∗ ⊗ I)(UK(h + k) ⊗ UK(−k)) = Xh,k,p,l,s × Bl+p−g(es)Bl(es)∗(A(−k)∗ ⊗ I)(UK(h) ⊗ UK(k)). hg, h + ki(UK(−h − p) ⊗ UK(−k + p))(A(p − k) ⊗ I) Thus Eq.(3.33) is equivalent to Bl+p−g(es)Bl(es)∗ Xl,s = δg,phg, kiA(g − k)∗A(g)A(−k) ⊗ Q0 dXs,t − 1 = δg,pQ0 ⊗ Q0 − 1 dXs,t T0(A(p − k)∗es)T0(A(−k)∗et)∗ ⊗ T0(A(k − p)Jes)T0(A(k)Jet)∗ T0(es)T0(et)∗ ⊗ T0(Jes)T0(Jet)∗. (cid:3) Lemma 8.4. Eq.(4.1) is equivalent to (8.11) ( 1 )T0(ξ) d T0(A(−h)∗C∗Jes)∗B−h(T0(es)∗Bh(ξ)T0(et))T0(C∗Jet). m n − =Xh,s,t THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 61 Proof. The left-hand side of Eq.(4.1) for Ti = Th+k(es) and Tj = Tk(et) is hh, q − giTq−g(C∗Jes)∗l(T−h(T0(A(h)∗es)∗Bg+h(ξ)T0(et)))Tq(C∗Jet) 1 = 1 = hh, giUK(−g) hh + k, pihg, kihk, qi j2(Th+k(es))∗l(Th+k(es)∗l(Tg(ξ))Tk(et))j2(Tk(et)) d Xh,k,s,t nd Xh,k,p,q,s,t × Tp(C∗Jes)∗l(UK(h)∗T0(A(h)∗es)∗Bg+h(ξ)T0(et))Tq(C∗Jet) d Xh,q,s,t = d Xh,q,s,t × T0(A(−g)∗C∗Jes)∗B−h−g(T0(A(h)∗es)∗Bg+h(ξ)T0(et))T0(C∗Jet) dXh,s,t = × T0(A(−g)∗C∗Jes)∗B−h(T0(A(h − g)∗es)∗Bh(ξ)T0(et))T0(C∗Jet) dXh,s,t = × T0(A(−g)∗A(g − h)∗C∗Jes)∗B−h(T0(es)∗Bh(ξ)T0(et))T0(C∗Jet) dXh,s,t = hg − h, giUK(−g) hg − h, giUK(−g) 1 1 n n n UK(−g)T0(A(−h)∗C∗Jes)∗B−h(T0(es)∗Bh(ξ)T0(et))T0(C∗Jet). Since d n (1 − 2n d2 ) = d2 − 2n nd = md − n nd = m n − 1 d , we get the statement. Lemma 8.5. Eq.(4.2) is equivalent to (cid:3) (8.12) δg,0hη, ζiT0(ξ) − 1 dhξ, ζiT0(η) =Xh,s T0(Jη)∗Bg−h(es)T0(C∗JBh(ξ)∗T0(A(h)∗C∗ζ)T0(es)). Eq.(4.3) is equivalent to (8.13) 1 ǫC∗A(k)∗T0(A(−k)∗ζ)∗Bg−k(ξ)T0(η) √nXh = hg, hiBh−k(ζ)∗T0(C∗A(h)∗η)T0(C∗A(g)∗ξ). 62 MASAKI IZUMI Under the presence of Eq.(8.7), Eq.(4.4) is equivalent to (8.14) Xg,t = hh, kiBk(A(−h)∗ζ)∗T0(η)B−h(ξ) − Bg(T0(A(−g − h)∗ζ)∗B−g−h(ξ)T0(et))Bg+k(et)∗T0(η) d√nhC∗ξ, ζiK0Xt ǫ T0(et)T0(Jet)T0(Jη)∗. Proof. We set T = Tg(ξ), T ′ = Tg′(η), and Tg′′(ζ) in Eq.(4.2), (4.3, (4.4)). Then the left-hand side of Eq.(4.2)is hg′′, h + kihk, giTg′(η)∗l(T−h(es))j2(Tk(Bg+h(ξ)∗T0(A(h)∗C∗Jζ)T0(es))) ǫ = 1 = hg + g′′ − f, kihh, g′′i Tg′(η)∗l(Ti)j2(l(Tg(ξ))∗j2(Tg′′(ζ))Ti) Tg′(η)∗l(T−h(es))j2(l(Tg(ξ))∗ ǫ √nhg′′, h + kiTh+k(C∗Jζ)T−h(es)) Xi =Xh,k,s √nXh,k,s n Xh,k,f,s × Tg′(η)∗l(T−h(es))Tf (C∗JBg+h(ξ)∗T0(A(h)∗C∗Jζ)T0(es)) =Xh,k,s hh, g′′iTg′(η)∗l(T−h(es))Tg+g′′(C∗JBg+h(ξ)∗T0(A(h)∗C∗Jζ)T0(es)) =Xh,k,s × T0(C∗JBg+h(ξ)∗T0(A(h)∗C∗Jζ)T0(es)) =Xh,k,s × T0(C∗JBg+h(ξ)∗T0(A(h + g)∗A(g)C∗Jζ)T0(es)) =Xh,k,s × T0(C∗JBh(ξ)∗T0(A(h)∗A(g)C∗Jζ)T0(es)). hh, giUK(g′ − g′′ − g)T0(A(g′ − g′′ − g)∗η)∗Bg′−g′′−g−h(es) UK(g′ − g′′ − g)T0(A(g′ − g′′ − g)∗η)∗Bg′−g′′−g−h(es) UK(g′ − g′′ − g)T0(A(g′ − g′′ − g)∗η)∗Bg′−g′′−h(es) The right-hand side is ǫδg′,g′′hζ, ηiTg(ξ) − 1 1 dXh hTg′′−h(ζ), T−g(A(−g)Jξ)iTh−g′(A(h − g′)Jη) dXh hζ, A(−g)Jξiδh,g+g′′Th−g′(A(h − g′)Jη) ǫ dhξ, A(g)JζiTg+g′′−g′(A(g + g′′ − g′)Jη). = ǫδg′,g′′hζ, ηiTg(ξ) − = ǫδg′,g′′hζ, ηiTg+g′′−g′(ξ) − THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 63 Thus we get δg′,0hη, ζiT0(ξ) − 1 dhξ, A(g)ζiT0(A(g − g′)η) T0(A(g′ − g)∗Jη)∗Bg′−h(es)T0(C∗JBh(ξ)∗T0(A(h)∗A(g)C∗ζ)T0(es)). =Xh,s Replacing η with A(g − g′)∗η and ζ with A(g)∗ζ, we get Eq.(8.12). Note that we have j2 ◦ j−1 1 (Tg(ξ)) = ǫj2(T−g(A(−g)Jξ)) = hg, hiTh(C∗JA(−g)Jξ) 1 √nXh hg, hiTh(C∗A(g)∗ξ). = ǫ √nXh The left-hand side of Eq.(4.3) is hg′ − g, hihg′ − g′′, kiUK(−k)Bg′′+h(ζ)∗T0(C∗A(h)∗A(g′)∗η)T0(C∗A(g)∗ξ) 1 1 = nXh,k l(Tg′′(ζ))∗hg′, h + kihg,−hiTh+k(C∗A(g′)∗η)T−h(C∗A(g)∗ξ) nXh,k nXh,k nXh,k = = 1 1 hg, hihg′ − g′′, kiUK(−k)Bg′′+h(ζ)∗T0(C∗A(h + g′)∗η)T0(C∗A(g)∗ξ) hg, h − g′ihg′ − g′′, kiUK(−k)Bh+g′′−g′(ζ)∗T0(C∗A(h)∗η)T0(C∗A(g)∗ξ) The right-hand side of Eq.(4.3) is 1 (Tg′′(ζ)∗l(Tg(ξ))Tg′(η)) j2 ◦ j−1 = hg′, gij2 ◦ j−1 ǫhg′, gi√n Xk = 1 (Tg′−g′′(T0(A(g′′ − g′)∗ζ)∗Bg+g′′−g′(ξ)T0(η))) hk, g′ − g′′iTk(C∗A(g′ − g′′)∗T0(A(g′′ − g′)∗ζ)∗Bg+g′′−g′(ξ)T0(η)). Let f = g′ − g′′. Then Eq.(4.3) is equivalent to 1 ǫC∗A(f )∗T0(A(−f )∗ζ)∗Bg−f (ξ)T0(η) √nXh = hg, hiBh−f (ζ)∗T0(C∗A(h)∗η)T0(C∗A(g)∗ξ). 64 MASAKI IZUMI The left-hand side of Eq.(4.4) is l(Tg′′(ζ))∗Tg′(η)l(Tg(ξ)) hg′′, kiUK(−k)Bg′′+h(ζ)∗(A(h)∗UK(h + k) ⊗ UK(−h))Tg′(η)l(Tg(ξ)) hg′′, kiUK(−k)Bg′′+h(ζ)∗Tg′−h−k(A(h)∗η)UK(−h)l(Tg(ξ)) =Xk,h =Xk,h =Xk hg′′, kiUK(−k)Bg′+g′′−k(ζ)∗T0(A(g′ − k)∗η)UK(k − g′)l(Tg(ξ)) = Xk,k′,h′hg, k′ihg′′, kiUK(−k)Bg′+g′′−k(ζ)∗T0(A(g′ − k)∗η)UK(k − g′) × (UK(−h′ − k′)A(h′) ⊗ UK(h′))Bg+h′(ξ)UK(k′) =Xk,k′hg, k′ihg′′, kiUK(−k)Bg′+g′′−k(ζ)∗T0(A(g′ − k)∗η) × (A(k − k′ − g′) ⊗ UK(k − k′ − g′))Bg−g′+k−k′(ξ)UK(k′) THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 65 The first term of the right-hand side is l(Tg′′(ζ)∗l(Tg(ξ))Tf (et))l(Tf (et))∗Tg′(η) hg, fihq, g′′i(UK(−p − q)A(p) ⊗ UK(p)) hg, fihq, f − g′′i(UK(−p − q)A(p) ⊗ UK(p)) hg, fil(UK(g′′ − f )T0(A(g′′ − f )∗ζ)∗Bg+g′′−f (ξ)T0(et))l(Tf (et))∗Tg′(η) hg, fil(Tg′′(ζ)∗(UK(−h − f )A(h) ⊗ UK(h))Bg+h(ξ)T0(et))l(Tf (et))∗Tg′(η) Xf,t =Xf,t,h =Xf,t = Xf,t,p,q × Bf−g′′+p(T0(A(g′′ − f )∗ζ)∗Bg+g′′−f (ξ)T0(et))UK(q)l(Tf (et))∗Tg′(η) = Xf,t,p,q,r × Bf−g′′+p(T0(A(g′′ − f )∗ζ)∗Bg+g′′−f (ξ)T0(et))UK(q) × UK(−q)Bf +r(et)∗(A(r)∗UK(r + q) ⊗ UK(−r))Tg′(η) = Xf,t,p,r hg, fihr − g′, g′′i(UK(−p + r − g′)A(p) ⊗ UK(p)) × Bf−g′′+p(T0(A(g′′ − f )∗ζ)∗Bg+g′′−f (ξ)T0(et))Bf +r(et)∗T0(A(r)∗η)UK(−r) = Xf,t,k,r hg, fihr − g′, g′′i(UK(−k)A(k + r − g′) ⊗ UK(k + r − g′)) × Bf +k+r−g′−g′′(T0(A(g′′ − f )∗ζ)∗Bg+g′′−f (ξ)T0(et))Bf +r(et)∗T0(A(r)∗η)UK(−r) = Xf,t,k,k′hg, fihk′ + g′, g′′i(UK(−k)A(k − k′ − g′) ⊗ UK(k − k′ − g′)) × Bf +k−k′−g′−g′′(T0(A(g′′ − f )∗ζ)∗Bg+g′′−f (ξ)T0(et))Bf−k′(et)∗T0(A(−k′)∗η)UK(k′) The second term of the right-hand side is 1 1 (Tg(ξ)), Tg′′(ζ)iK hh, g′′ihg, g′′ihC∗A(g)∗ξ, ζiK0 dXh,k,t hV (h)j2 ◦ j−1 × UK(h)Th+k(et)j1(Th+k(et))j1(UK(h)∗Tg′(η))∗ d√nXh,k,t = × Tk(et)T−h−k(A(−h − k)Jet)T−h−g′(A(−h − g′)Jη)∗ d√nXk,k′,t = × Tk(et)Tg′−k+k′(A(g′ − k + k′)Jet)Tk′(A(k′)Jη)∗. hk′ + g′, g′′ihg, g′′ihC∗A(g)∗ξ, ζiK0 ǫ ǫ 66 MASAKI IZUMI Thus we get T0(et)T0(Jet)T0(A(k′)Jη)∗ Xf,t hg, fiBf +k−k′−g′−g′′(T0(A(g′′ − f )∗ζ)∗Bg+g′′−f (ξ)T0(et)) × Bf−k′(et)∗T0(A(−k′)∗η) d√nhg, g′′ihC∗A(g)∗ξ, ζiK0Xt ǫ = hg′, g′′ihg + g′′, k′ihg′′, kiA(k − k′ − g′)∗Bg′+g′′−k(ζ)∗T0(A(g′ − k)∗η) × A(k − k′ − g′)Bg−g′+k−k′(ξ) = hg′, g′′ihg + g′′, k′ihg′′, kiBg′+g′′−k(A(k − k′ − g′)∗ζ)∗ × T0(A(k − k′ − g′)∗A(g′ − k)∗η)Bg−g′+k−k′(ξ) = hg′′, g′ − k + k′ihg, k′ihk − k′ − g′,−g′ + ki × Bg′+g′′−k(A(k − k′ − g′)∗ζ)∗T0(A(−k′)∗η)Bg−g′+k−k′(ξ), where we used Eq.(8.7). Replacing k, η, and ζ with k′′ + k′ + g′, A(−k′)η, and A(g)∗ζ respectively, and multiplying hg, g′′i make the first term of the left-hand side of this as Xf,t Bf +k′′−g′′(T0(A(g + g′′ − f )∗ζ)∗Bg+g′′−f (ξ)T0(et))Bf−k′(et)∗T0(η) =Xx,t Bx(T0(A(g + k′′ − x)∗ζ)∗Bg+k′′−x(ξ)T0(et))Bx−k′′−k′+g′′(et)∗T0(η). The second term becomes ǫ The right-hand side becomes d√nhC∗ξ, ζiK0Xt T0(et)T0(Jet)T0(Jη)∗. hg′′, k′′ihg, k′ − g′′ihk′′, k′ + k′′iBg′′−k′′−k′(A(k′′)∗A(g)∗ζ)∗T0(η)Bg+k′′(ξ) = hg, k′′ihg′′, k′′ihg, k′ − g′′ihk′′, k′ + k′′iBg′′−k′′−k′(A(g + k′′)∗ζ)∗T0(η)Bg+k′′(ξ) = hg′′ − k′ − k′′,−g − k′iBg′′−k′′−k′′(A(g + k′′)∗ζ)∗T0(η)Bg+k′′(ξ). This is equivalent to Eq.(8.14). (cid:3) Theorem 4.9 and our computation so far show the following classification theorem. Theorem 8.6. C∗ near-group categories with a finite abelian group G and irrational d is completely classified by (ǫ,h·,·i,K0, A(g), Bg, C, J) satisfying the conditions in the statement of Lemma 7.4 and Eq.(8.5)-(8.14) up to equivalence in the following sense: we say that two tuples (ǫ,h·,·i,K0, A(g), Bg, C, J), (ǫ′,h·,·i′,K′0, A′(g), B′g, C′, J′) THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 67 are equivalent if there exist a unitary W : K0 → K′0 and a group automorphism ϕ ∈ Aut(G) satisfying hg, hi′ = hϕ(g), ϕ(h)i, A′(g) = W A(ϕ(g))W ∗, B′ϕ(g)(W ξ)W = (W ⊗ W )Bg(ξ), C′ = W CW ∗, and J′ = W JW ∗. Now we assume that an orthonormal basis {et}t∈Λ of K0 satisfies the conditions in Lemma 7.4, and define bs,t r,p(g) ∈ C by br,s t,u(g) = T (es)∗T (er)∗Bg(eu)T (et). Then we have br,s t,u(g)T0(er)T0(es)T0(et)∗. Theorem 8.7. In terms of br.s t,u(g), Eq.(8.5)-(8.14) altogether are equivalent to t,u(h) = ǫǫrǫtcua(g)χu(g)bs,t r,u(g), (8.15) (8.16) (8.17) (8.18) (8.19) (8.20) (8.21) (8.22) (8.23) (8.24) Bg(eu) =Xr,s,t √nXh 1 hg, hibr,s Xr Xs br,s r,u(0) = − δs,u d , br,s t,s(0) = − δr,t d , t,u(g) = br,s′ t,u′(g)br,s Xr,t Xs,u χrχs 6= χtχu =⇒ br,s t,u(g)br′,s br,s t′,u(g) = δs,s′δu,u′ n δg,0δs,uδs′,u′ d , − δr,r′δt,t′ n − t,u(g) = 0, δg,0δr,tδr′,t′ d , ∀g ∈ G, br,s t,u(g) = ǫsǫua(g)χu(g)bt,s r,u(−g), t,u(g) = ǫtǫrcrcta(g)χr(g)br,u br,s t,s (−g), br,s t,u(g) = ǫrǫsǫtǫuctcrχr(g)χs(g)bt,u r,s(g), t,s (g + h)bp,x q,t (g + k) ǫtctbv,w q,s (g)br,u crǫr Xg,q,s,t = ǫuǫwχr(h)χu(h)hh, kiXy v,r(k)bx,w bp,y y,u (h) − cuǫpǫv d√n δr,uδw,vδx,p. 68 MASAKI IZUMI Proof. Eq.(8.9) is equivalent to (8.15). Under the presence of this condition, we have equivalence of Eq.(8.5),(8.6),(8.7),(8.8),(8.10) and Eq.(8.16),(8.18)-(8.21). Eq.(8.11) is equivalent to ( m n − 1 d )δp,u = Xg,r,s,t ǫrǫtcrcta(g)χr(g)br,p t,s (−g)br,s t,u(g), which follows from Eq.(8.18),(8.21),(8.23). We show, on the other hand, that this together with Eq.(8.18),(8.21) implies Eq.(8.23). Eq.(8.18) implies Xs Xh Xr,r t,u(g)2 =Xs Xh br,s ( 1 n − δg,0δs,u d ) =Xs (1 − δs,u d ) = n m − 1 d . In the same way, we have Xs Xh Xr,r ǫrǫtcrcta(g)χr(g)br,p t,s (−g)2 = n m − 1 d . Thus the Cauchy-Schwartz inequality implies t,s (−g) = ǫrǫtctcra(g)χr(g)br,s br,p t,u(g), which together with Eq.(8.21) implies Eq.(8.23). Although Eq.(8.17),(8.22) are redundant, we put them in the statement to empha- size that the equations are symmetric in left and right variables. In fact, Eq.(8.17) follows from Eq.(8.16),(8.19), and Eq.(8.22) follows from Eq.(8.21),(8.23). In the rest of the proof, we assume the conditions we have obtained so far, and show that Eq.(8.12),(8.13) follow from them, and Eq.(8.14) is equivalent to Eq.(8.24). Eq.(8.12) is equivalent to δg,0δr,vδp,u − δu,rδp,v d =Xh,s,t ǫtǫvcrcta(h)χr(h)br,s t,u(h)bv,p t,s (g − h), and by Eq.(8.22) and the Fourier transform, the right-hand side is equal to Xh,s,t ǫrǫvbr,u t,s (−h)br,p t,s (g − h) =Xk,s,t ǫrǫvcbr,u t,s (k)cbr,p t,s (k). Thanks to Eq.(8.15),(8.19), this is equal to the left-hand side. Eq.(8.13) is equivalent to ǫcsχr(k)χs(k)br,s t,u(g) = ctcua(g + k)a(k)χu(g + k)χt(k)hg + k, ki 1 √nXh hg, hia(h)χt(h)bt,u s,r(h). THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 69 Thanks to a(g + k)a(k)hg + k, ki = a(g)a(k)2hk, ki = a(g) and Eq.(8.22), the right- hand side is ctcua(g)χu(g + k)χt(k) s,r(h) 1 √nXh hg, hia(h)χt(h)bt,u √nXh 1 hg, hibt,r s,u(−h). = ǫsǫtcscua(g)χu(g + k)χt(k) This coincides with the left-hand side thanks to Eq.(8.15),(8.20). We set ξ = eu, η = ep, ζ = er in Eq.(8.14). Then left-hand side is Xg,s,t a(−g − h)χr(−g − h)br,s t,u(−g − h)Bg(T0(es))Bg+k(et)∗T0(ep). Thanks to Eq.(8.22), this is equal to ǫtǫrctcrbr,u t,s (g + h)Bg(T0(es))Bg+k(et)∗T0(ep) ǫtǫrctcrbv,w q,s (g)br,u t,s (g + h)bp,x q,t (g + k)T0(ev)T0(ew)T (ex)∗. Xg,s,t = Xg,s,t,v,w,x The first term of the right-hand side is hh, kia(h)χr(h)Bk(er)∗T (ep)B−h(eu) = hh, kia(h)χr(h) Xv,w,x,y bp,y v,r(k)by,w = ǫuhh, kiχu(h)χr(h) Xv,w,x,y ǫwbp,y x,u (−h)T0(ev)T0(ew)T0(ex)∗ v,r(k)bx,w y,u (h)T0(ev)T0(ew)T0(ex)∗, where we used Eq.(8.21). The second term is − cuδr,uǫp d√n Xv,w,x ǫvδw,vδx,pT0(ev)T0(ew)T0(ex)∗. Thus Eq.(8.14) is equivalent to Eq.(8.24). (cid:3) Since h·,·i is non-degenerate, there exists unique gr ∈ G for each r ∈ Λ satisfying χr(g) = hg, gri for any g ∈ G. Lemma 8.8. Eq.(8.15) and (8.23) imply (8.25) br,s t,u(g) = crcucsctbs,r u,t(g + gs − gu). 70 MASAKI IZUMI Proof. Assuming Eq.(8.15),(8.20), and (8.23), we get r,u t,u)(h) = ǫǫrǫtcua(h)χu(h)bs,t F−1(br,s = ǫǫrǫtcua(h)χu(h)ǫsǫtǫrǫucrcsχs(h)χt(h)br,u = ǫǫsǫucucrcsa(h)χu(h)χt(h)χs(h)br,u = cucrcsctχu(h)χs(h)F−1(bs,r = cucrcscthh, gu − gsiF−1(bs,r u,t)(h) u,t)(h), s,t (h) s,t (h) which show the statement. (cid:3) We denote by G(A, C, J) the set of all unitaries acting on K0 commuting with A(g), C, J, and call it the gauge group. An element v in the gauge group acts on the solution of the above polynomial equations as B′g = (v ⊗ v)Bg(v∗·)v∗, or equivalently, b′r′,s′ t′,u′ (g) = Xr,s,t,u vr′rvs′svt′tvu′ubr,s t,u(g). It is convenient to introduce a matrix B(g) for each g ∈ G with an index set Λ× Λ whose(cid:0)(r, t), (s, u)(cid:1)-entry is br,s t,u(g). Then B(g)∗B(g) = B(g)B(g)∗ = 1 n I − δg,0 d δ ⊗ δ, where δ ∈ CΛ×Λ is a vector whose (r, t)-component is δr,t. In fact δ is a common eigenvector of B(0) and B(0)∗ with an eigenvalue −1/d. Thus √nB(g) is a unitary for g 6= 0, and √nB(0) is a unitary on {δ}⊥. An element v of the gauge group acts on B(g) as B′(g) = (v ⊗ v)B(g)(vT ⊗ v∗). 9. The case of m = G In this section, we give a brief account of the most tractable case m = n among the irrational case, which was essentially done in [32] and [16]. Since dimK0 = 1, we can choose e ∈ K0 with kek = 1 and Je = e, and so ǫ = 1. Such e is unique up to sign. We denote Tg = Tg(e). Then V (g)Th = hg, hiTh and UK(g)Th = Th−g. We can arrange a(g) so that A(g) = a(g)I, and C is a scaler, which we denote by c. Thus j1(Th) = a(h)T−h, j2(Th) = c √nXk∈G hh, kiTk. There exists b : G → C satisfying Bg(e) = b(g)T0T0T ∗0 , and l(Tg) =Xh,k a(h)b(h + g)hg, kiTh+kT−hT ∗k . THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 71 Theorem 9.1. The C∗ near-group categories with a finite abelian group G with m = G are completely classified by (h·,·i, a, b, c) satisfying the conditions in the statement of Lemma 7.4 and the following equations up to the group automorphisms of G. (9.1) (9.2) (9.3) (9.4) (9.5) Xg∈G b(g) = ca(g)b(−g), b(0) = − 1 d , δg,0 d , 1 n − b(g)2 = b(g) = a(g)b(−g), b(g + h)b(g + k)b(g) = hh, kib(h)b(k) − c d√n . Proof. Note that replacing e with −e does not change the above equations. Thus the statement follows from Theorem 8.6 and Theorem 8.7. (cid:3) Remark 9.2. We rephrase the equations in the above theorem so that we can eas- ily guess Galois conjugate solutions in the non-unitary case. We use a parameter c′ = c/√n instead of c. For a given pair (h·,·i, a(g)), the polynomial equations are equivalent to (9.6) (9.7) (9.8) (9.9) (9.10) (9.11) (9.12) c′3 = 1 n2Xg∈G a(g), d2 = dn + n, Rc′b = b, 1 b(0) = − d a(g)b(g)b(−g) = 1 n − δg,0 d , a(g)b(−g)b(g + h)b(g + k) = hh, ki−1b(h)b(k) − c′−1 dn , Xg∈G Jb = b, (9.13) where Rc′ is a period 3 unitary on ℓ2(G) given by d > 0, Rc′f (g) = c′a(g)−1Xh∈G hg, hif (h), 72 MASAKI IZUMI and J is a period 2 anti-unitary on ℓ2(G) given by J f (g) = a(g)f (−g). They satisfy J Rc′ = R2 c′J , and the eigenspaces of Rc′ are preserved by J . We suspect that the last two equations come from the unitarity of the category, and the other equations are already good enough to give near-group categories in the general case. We show that known solutions of the whole equations have Galois conjugate solutions that satisfy all the equations except for the last two. Example 9.3. For G = Z2, there is a unique non-degenerate symmetric bicharacter hg, hi = (−1)gh, and there is unique a(g) up to complex conjugate given by a(1) = i. Then there is a unique solution of Eq.(9.6)-(9.13): c′ = 12 e− 7πi √2 = 1 − √3 − (1 + √3)i d = 1 + √3, , 1 − i 2 For only Eq.(9.6)-(9.11), there is another solution b(1) = = . c′ = πi 12 e √2 = 1 + √3 + (√3 − 1)i , d = 1 − √3, b(1) = = 1 − i 2 . 4 e− πi √2 4 4 e− πi √2 4 For only Eq.(9.6)-(9.11), there is another solution Example 9.4. For G = Z3, there is a unique non-degenerate symmetric bicharacter up to complex conjugate, and we consider hg, hi = ζ gh n . For this there is a unique a(g) given by a(1) = a(2) = ζ3. For this pair, there exists a unique solution of Eq.(9.6)-(9.13) up to group automorphism given by 3 + √21 3 , where ζn = e πi 6 2πi √3i 6 ∈ Q(ζ3), d = , 2 c′ = e √3 = + 1 2 +s 3 + √21 2 c′ = πi 6 e √3 = +s−3 + √21 2 b(1) = ζ3(cid:0)−3 + √21 12 b(1) = ζ3(cid:0)−3 − √21 12 i 12 2√3(cid:1), b(2) = ζ3(cid:0)−3 + √21 3 − √21 2√3(cid:1), b(3) = ζ3(cid:0)−3 − √21 √3i 6 d = 12 + 2 1 , , 1 2 −s3 + √21 2 i 2√3(cid:1). −s−3 + √21 2 1 2√3(cid:1). Example 9.5. For Z4, there is a unique non-degenerate symmetric bicharacter up to complex conjugate, and we consider hg, hi = igh. Then there exists a unique solution of Eq.(9.6)-(9.13) up to group automorphism given by a(1) = a(3) = e− πi 4 , a(2) = −1, THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 73 c′ = e− 3πi 4 2 , b(1) = ζ16(p4 − 2√2 4 + d = 2 + 2√2, b(2) = ζ16(p4 − 2√2 4 i 25/4 ), b(2) = −i 2 . For only Eq.(9.6)-(9.11), there is another solution c′ = b(1) = ζ 5 16(cid:0)p4 + 2√2 4 + a(1) = a(3) = −e− πi 4 , a(2) = −1, πi 4 e 2 , d = 2 − 2√2, b(3) = ζ 5 1 25/4(cid:1), b(2) = −i 2 . 16(cid:0)p4 + 2√2 4 i 25/4 ), − − 1 25/4(cid:1), Example 9.6. For G = Z2 × Z2 = {0, g1, g2, g3}, there is a unique C∗ near-group category. There exist exactly two non-degenerate symmetric bicharacters on Z2 × Z2 up to group automorphisms, and we denote by h·,·i1 the one given by the following table. Up to flipping g1 and g2, there are exactly two a(g) for h·,·i1. For a(g0) = g0 g1 g2 -1 -1 1 -1 -1 1 -1 -1 1 g0 g1 g2 −a(g1) = i, a(g2) = 1, there exists a unique solution for Eq.(9.6)-(9.13): c′ = 1 2 , d = 2 + 2√2, b(g1) = 3πi 4 e 2 , b(g2) = −3πi 4 e 2 , b(g3) = 1 2 . For only Eq.(9.6)-(9.11) there is another solution c′ = 1 2 , d = 2 − 2√2, b(g1) = −πi 4 e 2 , b(g2) = πi 4 e 2 , b(g3) = 1 2 . For a(g0) = a(g1) = i, a(g2) = −1, there is no solution. table though there is no solution for it. The other non-degenerate symmetric bicharacter h·,·i2 is given by the following For the complete list of the solutions of Eq.(9.6)-(9.13) for G with G ≤ 13, see [16, Table 2] (see Example 12.18 below for two additional solutions for G = Z2× Z2× Z3). 74 MASAKI IZUMI g0 g1 g2 -1 -1 1 -1 1 -1 1 -1 -1 g0 g1 g2 10. The case of m = 2G In this section we assume m = 2n and write down the polynomial equations Eq(8.15)-(8.23) in a more accessible form. The author's experience tells that there are only finitely many solutions for Eq(8.15)-(8.23) up to equivalence, and they are likely to satisfy Eq.(8.24) automatically, which in practice could be verified by computer. We show that there are exactly two solutions, up to equivalence, of the polynomial equations for G = Z3. 10.1. Possible cases. We choose the index set Λ of the orthonormal basis of K0 as Λ = {1, 2}. When χ2 1 = 1, we may replace a(g) with a(g)χ1(g), and we may and do assume χ1 = 1. Lemma 7.3 shows that the only possible cases are the following: • Case I. χ1 = χ2 = 1, ǫ = 1, Je1 = e1, Je2 = e2. • Case II. χ1 = χ2 = 1, ǫ = −1, Je1 = e2, Je2 = −e2, c1 = c2. • Case III. χ1 = 1, χ2 6= 1, χ2 2 = 1, ǫ = 1, Je1 = e1, Je2 = e2. • Case IV. χ2 = χ−1 1 6= 1. Je1 = e2, Je2 = ǫe1, c1 = c2. 1 , χ2 We use the lexicographic order of the set Λ2 = {1, 2}2 to express B(g) as a matrix. t,u(g) is a solution for Eq.(8.15)-Eq.(8.24) in Case IV. We Lemma 10.1. Case IV never occurs. Proof. We assume that br,s introduce a unitary operator R1 of period 3 on ℓ2(G) by hg, hif (h). √n Xh∈G 2,1(g) = ǫb1,1 1b1,2 2,1(g) and R2 1,1k = kb1,2 kb2,1 2,1k = kb1,1 2,1k, where kfk denotes the ℓ2-norm of f ∈ ℓ2(G). Then Eq.(8.15) implies R1b1,2 Eq.(8.20) implies that the matrix B(g) is of the form R1f (g) = 2,1(g) = b2,1 c1a(g) 1,1(g), which shows and Eq.(8.18) implies B(g) = ∗ 0 0 ∗ 0 ∗ ∗ 0 0 ∗ ∗ 0 ∗ 0 0 ∗  , 2,1(g)2 + b2,2 b1,2 b1,1 1,1(g)2 + b2,1 1 1,1(g)2 = n 1 n − 2,1(g)2 = , δg,0 d . THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 75 Thus 2kb1,2 2,1k2 < 1/2. 2,1k2 = kb1,1 If χ4 1,1k2 + kb2,1 2,1k2 =Xg∈G 1 n − 1,1(g) = 0 and b2,1 1 6= 1, Eq.(8.20) implies b2,2 1 = 1, Eq.(8.23) implies b2,2 1,1(g) = χ1(g)2b2,2 ) = 1 − and kb1,2 1,2(g)2 = 1/n, 1,1(g), and b2,2 which is contradiction. If χ4 1,1 is supported by H = {g ∈ G; χ1(g)2 = 1}. Note that H is an index 2 subgroup of G. For g ∈ G \ H, we have b1,2 δg,0 d 1 d ( , 2,1(g)2 = 1/n, and 2,1k2 ≥ Xg∈G\H b1,2 kb1,2 2,1(g)2 = 1 2 , (cid:3) which is contradiction too. 10.2. Case I. We first assume only χ1 = χ2 = 1. We choose c ∈ T with c3a(0) = 1, and set ωr = ct/c. Then ω3 a unitary R ∈ B(ℓ2(G)) of period 3 by t = 1. We introduce and anti-unitary J of period 2 by Then they satisfy RJ = J R2. Eq.(8.15) and Eq.(8.16) now become Rf (g) = hg, hif (h), ca(g) √n Xh J f (g) = a(g)f (−g). r,u(g), t,u(g) = ωubs,t t,u(g) = ǫsǫubt,s Rbr,s J br,s r,u(g). In particular, the function br,r Now we assume ǫ = 1, Je1 = e1, Je2 = e2. When c1 = c2, the gauge group r,u is an eigenvector of R for the eigenvalue ωu. G(A, C, J) is O(2). When c1 6= c2, it is Z2 × Z2. It is straightforward to show the following lemma. 1,2(g), µ(g) = b1,2 2,2(g), η1(g) = b2,2 2,1(g), η2(g) = b1,1 Lemma 10.2. Assume χ1 = χ2 = 1, ǫ = 1, Je1 = e1, Je2 = e2. Let ξ1(g) = b1,1 ξ2(g) = b2,2 are equivalent to the following: (1) The matrix B(g) = (br,s ξ1(g) 2η2(g) ω2R2µ(g) ω2 ω1ω2 ω2 1ω2η2(g) ω2 µ(g) η2(g) 1Rµ(g) ω1ω2 η1(g) 2Rµ(g) ω1R2µ(g) ω2 η1(g) µ(g) 2η1(g) 1ω2η1(g) ξ2(g) 1,1(g), 1,2(g). Then Eq.(8.15)-(8.23) t,u(g))(r,t),(s,u) is expressed as η2(g) B(g) =  . (2) For any g ∈ G, (10.1) (10.2) Rξ1(g) = ω1ξ1(g), J ξ1(g) = ξ1(g), Rξ2(g) = ω2ξ2(g), J ξ2(g) = ξ2(g), 76 (10.3) (10.4) (10.5) MASAKI IZUMI Rη1(g) = ω1η1(g), J η1(g) = η1(g), Rη2(g) = ω2η2(g), J η2(g) = η2(g), J µ(g) = µ(g), J Rµ(g) = R2µ(g), (where the second one follows from the first one in the last equation though). (3) ξ1(0) + µ(0) = ξ2(0) + µ(0) = −1/d, η1(0) + η2(0) = 0. (4) For any g ∈ G. ξ1(g)2 + 2η2(g)2 + µ(g)2 = ξ2(g)2 + 2η1(g)2 + µ(g)2 = 1 n − 1 n − δg,0 d δg,0 d , , η1(g)2 + η2(g)2 + Rµ(g)2 + R2µ(g)2 = δg,0 d 2η2(g)η1(g) + µ(g)ξ1(g) + µ(g)ξ2(g) = − , 1 n , η2(g)ξ1(g) + η1(g)µ(g) + (ω1ω2Rµ(g) + ω1ω2R2µ(g))η2(g) = 0, η1(g)ξ2(g) + η2(g)µ(g) + (ω1ω2Rµ(g) + ω1ω2R2µ(g))η1(g) = 0, η1(g)2 + η2(g)2 + ω1ω2Rµ(g)R2µ(g) + ω1ω2Rµ(g)R2µ(g) = 0. Remark 10.3. The condition (4) above imply (10.6) (10.7) (10.8) (10.9) (10.10) (10.11) (10.12) (10.13) (10.14) (10.15) (10.16) (10.17) (10.18) (10.19) kξ1k2 + 2kη2k2 + kµk2 = 1 − kξ2k2 + 2kη1k2 + kµk2 = 1 − kη1k2 + kη2k2 + 2kµk2 = 1, 1 d 1 d , , 1 d , 2hη2, η1iℓ2(G) + hµ, ξ1iℓ2(G) + hξ2, µiℓ2(G) = − hη2, ξ1iℓ2(G) + hη1, µiℓ2(G) + (ω1ω2 + ω1ω2)hµ, η2iℓ2(G) = 0, hη1, ξ2iℓ2(G) + hη2, µiℓ2(G) + (ω1ω2 + ω1ω2)hµ, η1iℓ2(G) = 0, kη1k2 + kη2k2 + ω1ω2hµ,Rµiℓ2(G) + ω1ω2hRµ, µiℓ2(G) = 0. Eq.(10.15) and Eq.(10.19) imply (10.20) 2kµk2 − ω1ω2hRµ, µiℓ2(G) − ω1ω2hR2µ, µiℓ2(G) = 1. Lemma 10.4. The solutions of the equations in Lemma 10.2 satisfy either of the following two: THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 77 (1) η1(0) = −η2(0) ∈ R with η1(0) ≤ 1/(2√n), and there exist κ1, κ2 ∈ {1,−1} satisfying ξ1(0) = ξ2(0) = − µ(0) = − 1 2d + 1 2d − , 2√n κ1p1 − 4nη1(0)2 κ1p1 − 4nη1(0)2 κ1p1 − 4nη1(0)2 + κ2i 2√n 2√n , . Rµ(0) = R2µ(0) = ω1ω2 (2) There exist κ1, κ2 ∈ {1,−1} satisfying ξ1(0) = ξ2(0) = − κ1 2√n , 1 2d − η1(0) = η2(0) = 0, κ1 2√n µ(0) = − 1 2d + , Rµ(0) = R2µ(0) = ω1ω2−κ1 + κ2i 2√n . Proof. Lemma 10.2 (2) shows that ξr(0), ηr(0), and µ(0) are all real and R2µ(0) = Rµ(0). Lemma 10.2 (3) shows that η1(0) + η2(0) = 0 and there exist real numbers x and y satisfying ξ1(0) = ξ2(0) = − 1 2d − x, µ(0) = − 1 2d + x. Now Lemma 10.2 (4) with g = 0 is equivalent to x2 + η1(0)2 = 1 4n , η1(0)2 + Rµ(0)2 = 1 2n , η1(0)(ω1ω2Rµ(0) + ω1ω2Rµ(0) − 2x) = 0, 2η1(0)2 + (ω1ω2Rµ(0))2 + (ω1ω2Rµ(0))2 = 0. Solving these, we get the statement. (cid:3) Let ζ3 = e2πi/3. We can expand µ satisfying J µ = µ as µ = µ0 + µ1 + µ2 with 3µi and J µi = µi because we have J R = R2J . The index i in µi will be Rµi = ζ i understood as an element of Z3. Lemma 10.5. If ω1 = ω2 = 1, there exists a solution of the equations in Lemma 10.2 only if dim ker(R − 1) ≥ 2. The solutions satisfy kµ1k2 + kµ2k2 = 1/3, and either of the following two: 78 MASAKI IZUMI (1) η1(0) = −η2(0) ∈ R with η1(0) ≤ 1/(2√n), and there exist κ1, κ2 ∈ {1,−1} satisfying ξ1(0) = ξ2(0) = − 1 6d µ0(0) = − 1 2d − 2√n κ1p1 − 4nη1(0)2 κ1p1 − 4nη1(0)2 , + , + µ1(0) = − 1 6d 1 6d − (2) There exist κ1, κ2 ∈ {1,−1} satisfying ξ1(0) = ξ2(0) = − µ2(0) = − 2√n κ2 2√3n κ2 2√3n , . κ1 2√n , 1 2d − η1(0) = η2(0) = 0, κ1 6√n µ0(0) = − , 1 6d − κ1 3√n κ1 3√n − + + + κ2 2√3n κ2 2√3n , . µ1(0) = − µ2(0) = − 1 6d 1 6d Proof. The equality kµ1k2 + kµ2k2 = 1/3 follows from Eq.(10.20). Assume that (1) in Lemma 10.4 occurs. Then we have µ0(0) + µ1(0) + µ2(0) = − µ0(0) + ζ3µ1(0) + ζ3µ2(0) = µ0(0) + ζ3µ1(0) + ζ3µ2(0) = which imply (1). 1 2d , + 2√n κ1p1 − 4nη1(0)2 κ1p1 − 4nη1(0)2 + κ2i κ1p1 − 4nη1(0)2 − κ2i 2√n 2√n , , We further assume dim(R − 1) = 1 and get contradiction. Since (ξ1(0), µ0(0)) 6= (0, 0), there exists a unique f ∈ ker(R − 1) satisfying J f = f and f (0) = 1. Then we have ξ1 = ξ2 = ξ1(0)f , η1 = −η2 = η1(0)f , and µ0 = µ0(0)f . Eq.(10.18) implies η1(0)(ξ1(0) + µ0(0)) = 0, and we get η1(0) = 0. Thus η1 = η2 = 0. Eq.(10.15) implies kµk2 = 1/2. Since kµ1k2 + kµ2k2 = 1/3, we get kµ0k2 = 1/6 and µ0(0)2kfk2 = 1/6. Eq.(10.13) and Eq.(10.16) imply 1 1 ξ1(0)2kfk2 = 2 − d 1 µ0(0)ξ1(0)kfk2 = − 2d , . THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 79 and ( ξ1(0) µ0(0) )2 = 3(1 − 3 d = − , ξ1(0) µ0(0) 2 d ), which is contradiction. Now assume that (2) in Lemma 10.4 occurs. Then we have µ0(0) + µ1(0) + µ2(0) = − 1 2d + κ1 2√n , µ0(0) + ζ3µ1(0) + ζ3µ2(0) = −κ1 + κ2i 2√n µ0(0) + ζ3µ1(0) + ζ3µ2(0) = −κ1 − κ2i 2√n , , which implies (2). We further assume dim(R − 1) = 1 and get contradiction. In this case, we would have ξ1 = ξ2 = 3µ0, η1 = η2 = 0, which contradicts Eq.(10.16). (cid:3) Remark 10.6. If ω1 = ω2 = ζ i R, and the same statement replacing (µ0, µ1, µ2) with (µi, µ1+i, µ2+i) holds. 3, we can apply the same argument to ζ−i 3 R instead of In the same way, we can show Lemma 10.7. If ω1 = ζ3 and ω2 = ζ−1 10.2 satisfy 3 , the solutions of the equations in Lemma (10.21) kµ1k2 + kµ2k2 = 1 3 , and either of the following two: (1) η1(0) = −η2(0) ∈ R with η1(0) ≤ 1/(2√n), and there exist κ1, κ2 ∈ {1,−1} satisfying ξ1(0) = ξ2(0) = − µ0(0) = − 1 6d + , 1 2d − 2√n κ1p1 − 4nη1(0)2 κ1p1 − 4nη1(0)2 , 2√n κ2 2√3n κ2 2√3n , . µ1(0) = − µ2(0) = − 1 6d + 1 6d − 80 MASAKI IZUMI (2) There exist κ1, κ2 ∈ {1,−1} satisfying ξ1(0) = ξ2(0) = − κ1 2√n , 1 2d − η1(0) = η2(0) = 0, 1 κ1 6√n µ0(0) = − 6d − κ1 3√n + + κ1 3√n − 1 6d 1 6d + µ1(0) = − µ2(0) = − , κ2 2√3n κ2 2√3n , . We can show the following lemma in the same way except for Eq.(10.23), which follows from Eq.(10.14), (10.15), and (10.22). Lemma 10.8. If ω1 = 1 and ω2 = ζ±1 satisfy 3 , the solutions of the equations in Lemma 10.2 (10.22) (10.23) kµ0k2 + kµ±1k2 = 1 3 , kξ2k2 − 2kη2k2 − 3kµ∓1k2 = − 1 d , and either of the following two: (1) η1(0) = −η2(0) ∈ R with η1(0) ≤ 1/(2√n), and there exist κ1, κ2 ∈ {1,−1} satisfying (2) There exist κ1, κ2 ∈ {1,−1} satisfying ξ1(0) = ξ2(0) = − 1 2d − η1(0) = η2(0) = 0, κ1 2√n , + 1 µ0(0) = − 6d 1 µ±1(0) = − 6d µ∓1(0) = − , + κ1 3√n κ1 3√n − κ1 1 6√n 6d − κ2 2√3n κ2 2√3n . + , ξ1(0) = ξ2(0) = − κ1p1 − 4nη1(0)2 2√n , 1 2d − 1 µ0(0) = − 6d 1 µ±1(0) = − 6d − + , κ2 2√3n κ2 2√3n , µ∓1(0) = − 1 6d + κ1p1 − 4nη1(0)2 2√n . THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 81 10.3. Case II. We assume χ1 = χ2 = 1, Je1 = e2, Je2 = −e1. Since c1 = c2, we denote ω = cr/c. In this case the gauge group G(A, C, J) is SU(2). Lemma 10.9. Let the notation be as above. Let ξ(g) = b1,1 µ(g) = b1,2 2,1(g). Then Eq.(8.15)-(8.23) are equivalent to the following: 2,2(g), η(g) = b1,1 2,1(g), (1) The matrix B(g) = (br,s t,u(g))(r,t),(s,u) is expressed as B(g) = (2) For any g ∈ G, −ωR2µ(g) η(g) −J η(g) ωRµ(g) −η(g) J η(g) −ωR2µ(g) η(g) −J η(g) µ(g) ξ(g) µ(g) −J ξ(g) ωRµ(g) −η(g) J η(g)  . Rξ = ωξ, Rη = ωη, (10.24) (10.25) (10.26) (10.27) (10.28) (10.29) (10.30) (10.31) (10.32) (where the second one follows from the first one in the last equation though). J µ = −µ, J Rµ = −R2µ, (3) ωRµ(0) + ωRµ(0) = −1/d. (4) For any g ∈ G, Rµ(g)2 + Rµ(−g)2 + η(g)2 + η(−g)2 = ξ(g)2 + µ(g)2 + 2η(g)2 = 1 n , 1 n − δg,0 d , ωRµ(g)R2µ(g) + ωR2µ(g)Rµ(g) + η(g)2 + η(−g)2 = η(g)ξ(g) − J η(g)µ(g) − (ωRµ(g) + ωR2µ(g))η(g) = 0, η(g)µ(g) + J η(g)J ξ(g) + (ωRµ(g) + ωR2µ(g))J η(g) = 0, δg,0 d , 2η(g)J η(g) + µ(g)J ξ(g) − ξ(g)µ(g) = 0. Remark 10.10. The condition (4) above imply (10.33) (10.34) (10.35) (10.36) (10.37) 1 2kµk2 + 2kηk2 = 1 − d kξk2 + kµk2 + 2kηk2 = 1, , ωhµ,Rµiℓ2(G) + ωhRµ, µiℓ2(G) + 2kηk2 = 1 d , hη, ξiℓ2(G) = hµ, ηiℓ2(G), hη,J ηiℓ2(G) = hξ, µiℓ2(G). 82 MASAKI IZUMI Eq(10.33) and Eq(10.35) imply (10.38) 2kµk2 − ωhRµ, µiℓ2(G) − ωhRµ, µiℓ2(G) = 1 − 2 d . Lemma 10.11. The solutions of the equations in Lemma 10.9 satisfy either of the following two: (1) η(0) ≤ 1/(2√n), and there exist κ1, κ2 ∈ {1,−1} satisfying ξ(0) = η(0) η(0) µ(0) = Rµ(0) = ω(cid:0) − i, 2√n κ1 + κ2p1 − 4nη(0)2 κ1 − κ2p1 − 4nη(0)2 κ2p1 − 4nη(0)2 1 2d − 2√n 2√n i, i(cid:1), µ(0) = , κi √n 1 2d − Rµ(0) = ω(cid:0) − κi 2√n(cid:1). where η(0)/η(0) is interpreted as an arbitrary phase if η(0) = 1. (2) η(0) = ξ(0) = 0, and there exist κ ∈ {1,−1} satisfying Proof. Lemma 10.9,(2) shows µ(0) ∈ iR and R2µ(0) = −Rµ(0), and Lemma 10.9,(4) with g = 0 shows Rµ(0)2 + η(0)2 = 1 2n − 1 2d , −(ωRµ(0))2 − (ωRµ(0))2 + 2η(0)2 = 1 d , η(0)ξ(0) + η(0)µ(0) − (ωRµ(0) − ωRµ(0))η(0) = 0, −η(0)µ(0) + η(0)ξ(0) + (ωRµ(0) − ωRµ(0))η(0) = 0, η(0)2 + µ(0)ξ(0) = 0. ξ(0)2 + µ(0)2 + 2η(0)2 = 2d ±p1 − 4nη(0)2 Rµ(0) = ω(cid:0) − 2√n 1 n 1 . i(cid:1). The first two with Lemma 10.9,(3) is equivalent to If η(0) = 0, we get (2) or the η(0) = 0 case of (1). Assume η(0) 6= 0 now. Then the third and fourth equalities implies ωRµ(0) − ωRµ(0) = η(0) η(0) ξ(0) + µ(0) = − η(0) η(0) ξ(0) + µ(0). THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 83 Thus we can introduce a real parameter l satisfying η(0) i, ξ(0) = l η(0) µ(0) = η(0)2 i. l Iterating these into the last equality, we get and l2 + η(0)4 l2 + 2η(0)2 = κ1√n l + η(0)2 = l , 1 n , with κ2 1 = 1. Solving this, we get the statement. (cid:3) We can expand µ ∈ ℓ2(G) satisfying J µ = −µ as µ = µ0 + µ1 + µ2 with Rµi = ζ i 3µi and J µi = −µi. Lemma 10.12. There exists a solution of the equations in Lemma 10.9 only if The solutions with ω = 1 satisfy dim ker(R − ω) ≥ 2. kµ1k2 + kµ2k2 = 1 3 − 2 3d , and either of the following two: (1) η(0) ≤ 1/(2√n), and there exist κ1, κ2 ∈ {1,−1} satisfying ξ(0) = η(0) η(0) µ0(0) = ( 2√n κ1 + κ2p1 − 4nη(0)2 κ2p1 − 4nη(0)2 κ1 6√n − i, )i, µ1(0) = ( µ2(0) = ( + κ1 6√n κ1 6√n − 2√n 1 2√3d 1 2√3d )i, )i. where η(0)/η(0) is interpreted as an arbitrary phase if η(0) = 1. (2) η(0) = ξ(0) = 0, and there exist κ ∈ {1,−1} satisfying µ0(0) = 0, µ1(0) = ( µ2(0) = ( + κ 2√n κ 2√n − 1 2√3d 1 2√3d )i, )i. The second case could occur only if dim{f ∈ ker(R − 1); f (0) = 0} ≥ 2. 84 MASAKI IZUMI Proof. We assume ω = 1. The proof for the general case can be obtained by applying the same argument to ωR, and replacing (µ0, µ1, µ2) with (µ1, µ2, µ0) or (µ2, µ0, µ1). The first equation follows from Eq.(10.38). Assume that (1) in Lemma 10.11 occurs. Then we have µ0(0) + µ1(0) + µ2(0) = 2√n i, κ1 − κ2p1 − 4nη(0)2 κ2p1 − 4nη(0)2 κ2p1 − 4nη(0)2 1 2d − 1 2d − 2√n 2√n i, i, µ0(0) + ζ3µ1(0) + ζ3µ2(0) = − µ0(0) + ζ3µ1(0) + ζ3µ2(0) = which implies (1). We further assume dim ker(R− 1) = 1 and get contradiction. Since (ξ(0), µ0(0)) 6= (0, 0), we can find f ∈ ker(R − 1) with J f = f and f (0) = 1. Then we have ξ = ξ(0)f , η = η(0)f , µ0 = µ0(0)f , and Eq.(10.36),(10.37) imply η(0)ξ(0) = µ0(0)η(0), η(0)2 = ξ(0)µ0(0). Since µ0(0) ∈ Ri, this implies either ξ = η = 0 or µ0 = η = 0. The first case contradicts Eq.(10.33),(10.34). The second case with Eq.(10.33) gives kµk2 = (1 − 1/d)/2 on one hand, and on the other hand we have kµ2k = kµ1k2 + kµ2k2 = 1 3 − 2 3d , which is contradiction too. Now we assume (2) in Lemma 10.11 occurs. Then we have µ0(0) + µ1(0) + µ2(0) = κ √n i, 1 µ0(0) + ζ3µ1(0) + ζ3µ2(0) = − 2d − 1 2d − µ0(0) + ζ3µ1(0) + ζ3µ2(0) = κ 2√n κ 2√n i, i, which implies (2). In a similar way as above, we can show that the condition dim{f ∈ ker(R − 1); f (0) = 0} ≤ 1 does not allow any solution. (cid:3) Remark 10.13. If ω = ζ i and the same statement replacing (µ0, µ1, µ2) with (µi, µ1+i, µ2+i) holds. 3, we can apply the same argument to ζ−i 3 R instead of R, THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 85 10.4. Case III. We assume χ1 = 1, and χ2 6= 1, χ2 2 = 1, Je1 = e1, Je2 = e2. For simplicity we denote χ = χ2. There exists unique gχ ∈ G satisfying χ(g) = hg, gχi for all g ∈ G. Note that we have the following for aχ(g) = χ(g)a(g): hgχ, hia(h) = a(−gχ) = a(0)a(gχ). 1 Thus baχ(0) = √nXh∈G 2 = baχ(0) = a(0)a(gχ) = c3 c3 1a(gχ). Since gχ has order 2, we have a(gχ)2 = hgχ, gχi = χ(gχ), and a(gχ)4 = 1, which shows 2 = (c1a(gχ))3. c3 Lemma 10.14. For r = 1, 2, we introduce unitaries Rr, r = 1, 2, of period 3 on ℓ2(G) by Rrf (g) = crχr(g)a(g) √n Xh∈G hg, hif (h), and anti-unitaries Jr of period 2 on ℓ2(G) by Then Eq.(8.15), (8.20), (8.23), (8.25) are equivalent to the following conditions: Jrf (g) = χr(g)a(g)f (−g). (1) Let ξ1(g) = b1,1 1,1(g), ξ2(g) = b2,2 2,2(g), µ(g) = b1,2 (br,s t,u(g))(r,t),(s,u) is expressed as 1,2(g). The matrix B(g) = B(g) = 0 ξ1(g) 0 R2 0 R2µ(g) R2 µ(g) 0 2µ(g) R1µ(g) 1µ(g) 0 0 µ(g) 0 0 ξ2(g)  , and the following hold: R1ξ1(g) = ξ1(g), J1ξ1(g) = ξ1(g), R2ξ2(g) = ξ2(g), J2ξ2(g) = ξ2(g). (2) χ(g)µ(g) = µ(g). Proof. Eq.(8.25) implies b12 12(g) = b21 Eq.(8.23), (8.25) are equivalent to 21(g). Then Eq.(8.15), (8.20) are equivalent to (1). µ(g) = χ(g)µ(g), R1µ(g) = c1c2χ(g)R2µ(g) = c2 1c2 2R2µ(g + gχ), 1c2 R2µ(g) = c1c2χ(g)R1µ(g) = c2 2R1µ(g + gχ), 2µ(g) = R2 1µ(g + gχ), 1µ(g) = R2 1µ(g) = c1c2R2 R2 R2 2µ(g) = c1c2R2 2µ(g + gχ). 86 MASAKI IZUMI We show that all the conditions follow from the first one. Note that we have the relation R2f (g) = c1c2χ(g)R1f (g). One the other hand, the unitary R2 is given by r = R−1 r R2 rf (g) = cr√nXh∈G hg, hia(h)χr(h)f (h), 2f (g) = c2R2 and in particular it implies that we have c1R2 satisfying χ(g)f (g) = f (g). Note that the first condition implies 1f (g) for any function f Rrµ(g) = ar(g + gχ)ar(g)Rrµ(g + gχ) = ar(gχ)χ(g)Rrµ(g + gr), R2 rµ(g) = R2 rµ(g + gχ). 2 = c3 Thus the remaining conditions are equivalent to c3 holds in general. 1a(gχ) = c3 1a(gχ)hgχ, gχi, which (cid:3) We arbitrarily fix c ∈ T with c3a(0) = 1, and define R and J as in the previous 2 = 1. Note that using notation in the proof of the above subsections. We set cχ = ca(gχ), which satisfies c3 1 = ω3 ω2 = c2/cχ. Then ω3 lemma, we have R1 = ω1R. Lemma 10.15. Let the notation be as above. Let ξ1(g) = b1,1 µ(g) = b1,2 χbaχ(0) = 1. We set ω1 = c1/c and 1,2(g). Then Eq.(8.15)-(8.23) are equivalent to the following: 1,1(g), ξ2(g) = b2,2 2,2(g), (1) The matrix B(g) = (br,s t,u(g))(r,t),(s,u) is expressed as B(g) = (2) For any g ∈ G, ξ1(g) 0 0 µ(g) 0 0 µ(g) ω2a(gχ)R2µ(g) ω1Rµ(g) ω2a(gχ)χ(g)Rµ(g) ω1R2µ(g) 0 0 0 0 ξ2(g) Rξ1(g) = ω1ξ1(g), J ξ1(g) = ξ1(g),  . Rξ2(g) = ω2a(gχ)χ(g)ξ2(g), J ξ2(g) = χ(g)ξ2(g), χ(g)µ(g) = µ(g), J µ(g) = µ(g). (3) ξ1(0) + µ(0) = ξ2(0) + µ(0) = −1/d. (4) For any g ∈ G. (10.39) (10.40) (10.41) (10.42) (10.43) (10.44) (10.45) ξ1(g)2 + µ(g)2 = ξ2(g)2 + µ(g)2 = 1 n − 1 n − δg,0 d δg,0 d , , Rµ(g)2 + R2µ(g)2 = 1 n , THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 87 (10.46) (10.47) µ(g)ξ1(g) + µ(g)ξ2(g) = − δg,0 d , a(gχ)ω1ω2Rµ(g)R2µ(g) + a(gχ)ω1ω2χ(g)R2µ(g)Rµ(g) = 0. Proof. The proof of the previous lemma shows that any solution satisfies R2µ(g) = ω2a(gχ)χ(g)Rµ(g), 2µ(g) = ω2a(gχ)R2µ(g). R2 Now the statement follows from the previous lemma and a straightforward argument. (cid:3) (10.48) Remark 10.16. The condition (4) above imply kξ1k2 = kξ2k2 = 1 2 kµk2 = (10.49) 1 2 − 1 d , , Lemma 10.17. Any solutions of the equations in Lemma 10.15 satisfy the following two conditions: (1) There exists κ ∈ {1,−1} satisfying ξ1(0) = ξ2(0) = − κ 2√n , 1 2d − κ 2√n . µ(0) = − 1 2d + (2) For every order 2 element g ∈ G, there exists κg ∈ {1,−1} satisfying and in particular ω1ω2a(gχ)a(g)κgi(1+χ(g))/2 (Rµ(g))2 = (√2nRµ(g))12 = hgχ, gχihg, gi(−1) 2n , 1+χ(g) 2 . Consequently, when a(gχ) = ±1, we have (√2nRµ(0))12 = −1, and when a(gχ) = ±i we have (√2nRµ(0))12 = 1. Proof. Since JRµ(g) = R2µ(g), we have R2µ(g) = a(g)Rµ(−g). (1) follows from the g = 0 case of Lemma 10.15. If g ∈ G has order 2, we have R2µ(g) = a(g)Rµ(g). Thus Lemma 10.15,(4) implies Rµ(g)2 = 1/(2n) and ω1ω2a(gχ)a(g)(Rµ(g))2 + χ(g)ω1ω2a(gχ)a(g)Rµ(g) 2 = 0, (cid:3) which implies the statement. Note that Case III could occur only if G is an even group because there is no character of order 2 for odd groups. For small even groups we have Lemma 10.18. Case III could occur only if #G ≥ 8. 88 MASAKI IZUMI Proof. Assume that we have a solution (ξ1, ξ2, µ) for the equations in Lemma 10.15. Since χ(g)µ(g) = µ(g), the function µ is supported by {χ}⊥ = {g ∈ G; χ(g) = 1}. For G = Z2, we have = − On the other hand since {χ}⊥ = {0}, we have µ(0) = − + 1 2d κ 2√n 1 2(2 + √6) + κ1 2√2 . 1 2 = kµk2 = µ(0)2, which is impossible. For G with #G = 4, we have #{χ}⊥ = 2, and we set {χ}⊥ = {0, g⊥}. Since g⊥ has order 2 and J µ(g⊥) = µ(g⊥), we have µ(g⊥) = a(g⊥)µ(g⊥). Thus 1 8 = Rµ(0)2 = µ(0) + µ(g⊥) 2 2 = µ(0)2 + µ(g⊥)2 + (1 + a(g⊥))µ(g⊥) . 4 Since µ(0)2 + µ(g⊥)2 = kµk2 = 1 2 , and µ(g⊥) 6= 0, we get a(g⊥) = −1. This implies µ(0) ±q 1 2 − µ(0)2i 2 , Rµ(0) = c and on the other hand (2√2Rµ(0))24 = 1. This is contradiction as we have c24 = a(0)8 = 1, and µ(0) = 2 − √5 + κ1 4 . For G = Z6, we have χ(g) = (−1)g and {χ}⊥ = {0, 2, 4}. Up to complex conjugate, 6 . There are only two possibilities of a(g), 6 as the other one is a(g)χ(g). Since J µ(g) = µ(g), we may assume hg, hi = ζ gh and we can choose a(g) = e− πig2 we have µ(4) = a(2)µ(2) = ζ3µ(2). Now we have 6 , where ζ6 = e 2πi Rµ(0) = c 6 µ(2) + ζ6µ(2). Then Let x = ζ−1 µ(0) + ζ6(ζ−1 6 µ(2) + ζ6µ(2)) √6 . 1 12 and µ(0)2 + x2 + µ(0)x 6 = Rµ(0)2 = x = −µ(0) ±p2 − 3µ(0)2 2 . , THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 89 This implies Rµ(0) = c 3µ(0) ±p2 − 3µ(0)2 + (−µ(0) ±p2 − 3µ(0)3)√3i 4√6 . On the other hand we have (√12Rµ(0))24 = 1, which is a contradiction as we have c24 = 1 and µ(0) = √6 − √7 + κ1 2√6 . (cid:3) To see if Case III really occurs, the first test case is order 8 abelian groups Z8, Z4 × Z2, and Z2 × Z2 × Z2. 10.5. The case of G = Z2. Ostrik [47] showed that there is no near-group category for G = Z2 with m ≥ 3. We give a proof of this fact in the case of m = 4. There is a unique non-degenerate symmetric bicharacter hg, hi = (−1)gh for Z2, and there are exactly two a(g). We may assume a(0) = 1, a(1) = i as the other one is 12 . We identify ℓ2(Z2) its complex conjugate. Then a(0) = e with C2, the set of column vectors with two components, and express R as a matrix 4 , and we choose c = e− πi πi πi 12(cid:18) 1 0 −i (cid:19) 1 0 R = e J is given by Let √3 + 1 + (√3 − 1)i (cid:18) 1 −i 1 i (cid:19) . 1 4 √2(cid:18) 1 1 −1 (cid:19) = J(cid:18) f (0) f (1) (cid:19) =(cid:18) f (0) −if (1) (cid:19) . f1 = 4 ! , 1 e− πi √3−1√2 √3+1√2 1 − f0 = 4 ! . e− πi Then we have Rf0 = f0, Rf1 = ζ3f1, Jf0 = f0, Jf1 = f1, kf0k2 = 3 − √3, kf1k2 = 3 + √3. Lemma 10.12 shows that Case II never occurs, and Lemma 10.5 and Lemma 10.7 show that the only possible case is Case I with (ω1, ω2) = (1, ζ3) as ζ 2 3 is not an eigenvalue of R. Since µ2 = 0, the case (1) of Lemma 10.7 is the only possibility. Note that we have µ0 = µ0(0)f0 and µ1 = µ1(0)f1. Eq.(10.22) implies 1 = µ0(0)2kf0k2 + µ1(0)2kf1k2 3 = (3 + √3)( )2 + (3 − 1 6d − κ2 2√6 √3)( 1 6d + κ2 2√6 )2, which is contradiction as d = 2 + √6. 90 MASAKI IZUMI 10.6. The case of G = Z3. Larson [39] showed that there exists no near-group category for G = Z3 with m ≥ 7. On the other hand, it is known that there exists at least one C∗ near-group category for G = Z3 with m = 6, which was first observed by Zhengwei Liu and Noah Snyder ([40, page 59], see also [19, page 14]). Theorem 10.19. There exist exactly two C∗ near-group categories for Z3 with m = 6, and they are complex conjugate to each other. There are exactly two symmetric bicharacters of Z3 and they are complex conjugate 3 . For this, there is a unique a(g), given by to each other. We choose hg, hi = ζ gh a(1) = a(2) = ζ3. Since a(0) = i, we choose c = e− πi 6 . Then R = e Let 1 ζ 2 3 ζ 2 3 1 1 ζ3 1 ζ3 1  , πi 6  1 0 0 ζ 2 3 0 0 f0 = 1 − ζ3 − ζ3 2 2 0 0 ζ 2 3 f (0) f (1) 1 1 1 ζ3 1 ζ 2 3  1 √3 f (2)  = J  , f′0 = 1 ζ 2 3 ζ3  = f (0) ζ−1 3 f (2) ζ−1 3 f (1) 0 ζ3i −ζ3i  , e πi 6 √3  . f1 = 1 ζ3 ζ3  . Then Rf0 = f0, Rf′0 = f′0, Rf1 = ζ3f1, and J f0 = f0, J f′0 = f′0, J f1 = f1. They form a (non-normalized) orthogonal basis of ℓ2(Z3) as well as the real vector space We prove the above theorem by showing the following lemma. {f ∈ ℓ2(Z3) J f = f}. Lemma 10.20. For R and J as above, there is no solution of the equations in Lemma 10.9, and there is a unique solutions of the equations in Lemma 10.2 up to equivalence given as follows: ω1 = ω2 = 1, and √3 − 1 √3 − 1 2 2 ξ1(0) = − ξ2(0) = − , , ξ1(1) = ζ3( ξ2(1) = ζ3( √3 − 1 √3 − 1 4 4 η1(0) = 0, η1(0) = 0, η1(1) = yζ3i, η2(1) = −yζ3i, µ(0) = , µ(1) = ζ3( √3 − 1 2√3 + xi), ξ1(2) = ζ3( + xi), ξ2(2) = ζ3( η1(2) = −yζ3i, η2(2) = yζ3i, √3 − 1 √3 − 1 4 4 − xi), − xi), √3 + 1 4√3 − xi), µ(2) = ζ3( 1 − √3 + 2i 1 − √3 − 2i − xi), Rµ(2) = ζ3( 4√3 √3 + 1 4√3 1 − √3 + 2i 4√3 1 − √3 − 2i − xi), R2µ(2) = ζ3( + xi), 4√3 4√3 + xi), + xi), Rµ(0) = −1 + i 2√3 1 + i R2µ(0) = − 2√3 , Rµ(1) = ζ3( , R2µ(1) = ζ3( 2 − √3 , κ1p1 − 12η1(0)2 = 1 + 2√3 η1(0)2 = 3 , ξ1(0) = ξ2(0) = − . 54 2(2 − √3) 3√3 2(2 − √3) 3√3 THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 91 where x, y ∈ R with x2 + y2 = √3 24 . Proof. Since ζ 2 3 is not an eigenvalue of R, we have µ2(0) = 0, and Lemma 10.12 shows that Case II never occur. Lemma 10.5 shows that the only possibilities in Case I are (ω1, ω2) = (1, ζ3) and (ω1, ω2) = (1, 1). First we assume that (ω1, ω2) = (1, ζ3) holds. Since µ2(0) = 0, Lemma 10.7 shows that only the case (1) of Lemma 10.7 is possible, and we have Since ξ2 = ξ1(0)f1, η2 = −η1(0)f1, µ2 = 0, Eq.(10.23) implies )2 − 6 = ξ1(0)2kf1k2 − 2η1(0)2kf1k2 = 3( − 1 d 1 + 2√3 54 = 3 − 2√3, which is contradiction as we have d = 3 + 2√3. Now we solve the equations for (ω1, ω2) = (1, 1). Since µ2(0) = 0, only the case (2) in Lemma 10.5 with κ1 = κ2 = 1 is possible, and ξ1(0) = ξ2(0) = 3µ0(0) = − √3 − 1 2 , η1(0) = η2(0) = 0, µ1(0) = 1/3. Since ξ1, ξ2, η1, η2, µ0 belong to the real linear space {f ∈ ker(R − 1); J f = f}, there exist real numbers x1, x2, y1, y2, p satisfying ξr = ξr(0)f0 + xrf′0, ηr = yrf′0 for r = 1, 2, and µ0 = µ0(0)f0 + pf′0. Since dim ker(R − ζ3) = 1, we have µ1 = 1 3f1. Eq.(10.13), (10.14), (10.15) imply Thus kξ1k2 + kξ2k2 − 2kµk2 = − 2 d . 2ξ1(0)2kf0k2 + (x2 1 + x2 2)kf′0k2 − 2(µ0(0)2kf0k2 + p2kf′0k2 + 1 3 ) = − 2 d , and we get x2 1 + x2 2 = 2p2. Eq.(10.17), (10.18) imply py1 + (2p + x1)p = 0, (2p + x2)y1 + py2 = 0, and so either y1 = y2 = 0 or p2 = (2p + x1)(2p + x2). We claim that x1 = x2 = −p, y1 = −y2, and x2 1 + y2 1 = 1 8√3 . 92 MASAKI IZUMI Indeed, assume first that (y1, y2) 6= (0, 0) holds. Then 0 = 3p2 + 2(x1 + x2)p + x1x2 1 + x2 2) = 3p2 + 2(x1 + x2)p + (x1 + x2)2 − (x2 6p2 + 4(x1 + x2)p + (x1 + x2)2 − 2p2 (2p + x1 + x2)2 2 2 = = , 2 and we get x1 + x2 = 2p. Therefore we have x1 = x2 = −p and p(y1 + y2) = 0. Eq.(10.13), (10.14) imply which shows y2 1 = y2 2. Eq.(10.13),(10.16) imply kξ1k2 − kξ2k2 + 2(kη2k2 − kη1k2) = 0, 4y2 2 + 4p2 = −4y1y2 + 4p2 = 1 2√3 , which shows the claim. Assume now that y1 = y2 = 0 holds. Then η1 = η2 = 0, and the equations in Remark 10.3 are equivalent to 1 kµk2 = 2 kξ1k2 = kξ2k2 = , 3 2 − 2 √3 , hµ, ξ1iℓ2(Z3) + hξ2, µiℓ2(Z3) = 1 − 2 √3 . In terms of p, x1, x2, these are equivalent to p2 = x2 1 = x2 2 = p(x1 + x2) = − , 1 8√3 1 4√3 , which shows the claim again. Let x = x1 and y = y1. Then ξr, ηr, µ are as in the statement. It is straightforward to show that they satisfy the conditions in Lemma 10.2. Since (ω1, ω2) = (1, 1), the gauge group is O(2), and we show that they act on the solutions transitively. Indeed, it is easy to show that acts on the solutions as (x, y) 7→ (x,−y). Let 0 (cid:18) 1 0 −1 (cid:19) R(θ) =(cid:18) cos θ − sin θ cos θ (cid:19) . sin θ THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 93 Note that we have B(g) √3 − 1 6 1 3 = − 3 0 0 1 0 1 1 0 0 1 1 0 1 0 0 3  + f1(g) 0 0 0 ζ 2 3 0 ζ3 0 1 + f′0(g)[x(X ⊗ X − Y ⊗ Y ) − y(X ⊗ Y + Y ⊗ X)], Y =(cid:18) 0 1 1 0 (cid:19) . f0(g) X =(cid:18) 1 0 0 −1 (cid:19) , 0 1 ζ3 0 ζ 2 3 0 0 0  where The first two terms commute with R(θ) ⊗ R(θ). Since R(θ)XR(θ)−1 = cos 2θX + sin 2θY, we have R(θ)Y R(θ)−1 = − sin 2θX + cos 2θY, (R(θ) ⊗ R(θ))(X ⊗ X − Y ⊗ Y )(R(θ)−1 ⊗ R(θ)−1) = cos 4θ(X ⊗ X − Y ⊗ Y ) + sin 4θ(X ⊗ Y + Y ⊗ X), (R(θ) ⊗ R(θ))(X ⊗ Y + Y ⊗ X)(R(θ)−1 ⊗ R(θ)−1) = − sin 4θ(X ⊗ X − Y ⊗ Y ) + cos 4θ(X ⊗ Y + Y ⊗ X). Therefore R(θ) acts on the solutions as (x, y) 7→ (x cos 4θ + y sin 4θ,−x sin 4θ + y cos 4θ). (cid:3) Remark 10.21. We already know that there is a C∗ near-group category for Z3 with m = 6, and we do not need to verify Eq.(8.24). Among the equivalence class of the solutions above, we can choose a representative with y = 0. Such a solution is invariant under the subgroup of O(2) generated by 0 (cid:18) 1 0 −1 (cid:19) , (cid:18) 0 −1 0 (cid:19) , 1 which is the dihedral group D8 of order 8. Thus there is a D8-action γ on O9 commut- ing with αg and ρ corresponding such a solution. We can perform equivariantization with this D8-action, and the resulting category includes a near-group category for G = Z2 × Z2 × Z3 with m = 12 as a subcategory (see Example 13.15). e− 3πi 4 2 R = Let f0 = 1 − ζ16 2 cos 3π 8 tan 3π 8 i − ζ16 2 cos 3π 8 ,  1 1 1 i −1 −i 1 −1 i 1 1 1 −1 1 −i −1  , πi 4 1 0 0 e 0 0 0 0 0 −1 0 0  J f′0 = f (0) f (1) f (2) f (3) 0 ζ16i 0 −ζ16i πi f (0) 4 f (3) e −f (2) e 4 f (1) πi πi 4 0 0 0 e    = f1 =  , .  1 ζ16 2 cos 17π 24 tan 17π 24 i 2 cos 17π 24 ζ16 ,  f2 = 1 ζ16 2 cos π 24 tan π 24i ζ16 2 cos π 24 .  94 MASAKI IZUMI 10.7. The case of G = Z4. Theorem 10.22. There is no C∗ near-group category for Z4 with m = 8. There are exactly two non-degenerate symmetric bicharacters for Z4, which are complex conjugate to each other, and we may assume hg, hi = igh to show the state- ment. For this, there are exactly two a(g), and one of them is a(g) = e− g2πi 4 . In this case we can choose c = e 4 , and 3πi Then Rf0 = f0, Rf′0 = f′0, Rf1 = ζ3f1, Rf2 = ζ 2 3 f2, and they are invariant under J . They form a (non-normalized) orthogonal basis of ℓ2(Z4) as well as the real subspace of J -invariant vectors. We have 3 6√2 3 kf1k2 = kf2k2 = 2 cos2 17π 24 3 1 + cos 7π 12 3 = = = = , 1 + 2√2 − √3 1 + 2√2 + √3 6√2 . 1 + cos π 12 The other choice of a(g) is a(g) = (−1)ge− g2πi 4 . In this case we can choose c = e− πi 4 , 2 cos2 π 24 and πi 4 e 2 R = πi 4 1 0 0 −e 0 0 0 0  J f (0) f (1) f (2) f (3) 0 0 0 0 −1 0 0 −e πi 4  =   1 1 1 i −1 −i 1 −1 i 1 1 1 −1 1 −i −1  , πi f (0) −e 4 f (3) −f (2) −e 4 f (1) πi .  THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 95 Let f0 = 1 ζ 5 16 2 cos 7π 8 tan 7π 8 i ζ 5 16 2 cos 7π 8  , f′0 = 0 ζ 5 16i 0 −ζ 5 16i  , f1 = 1 ζ 5 16 2 cos 5π 24 tan 5π 24 i ζ 5 16 2 cos 5π 24  , f2 = 1 ζ 5 16 2 cos 11π 24 − tan 11π 24 i ζ 5 16 2 cos 11π 24 .  Then Rf0 = f0, Rf′0 = f′0, Rf1 = ζ3f1, Rf2 = ζ 2 3 f2, and they are invariant under J . They form a (non-normalized) orthogonal basis of ℓ2(Z4) as well as the real subspace of J -invariant vectors. kf1k2 = 6√2 = = 3 3 2 cos2 5π 24 1 + cos 5π 12 kf2k2 = 3 2 cos2 11π 24 = 3 1 + cos 11π 12 = , −1 + 2√2 + √3 −1 + 2√2 − √3 6√2 . We give a unified argument for Z4 and Z2 × Z2 to show that certain cases never occur. Note that we have d = 4 + 2√5 in the both cases. Lemma 10.23. Assume that #G = 4, dim ker(R − ζ r exist fr ∈ ker(R − ζ r √5 /∈ Q(√3,kf1k2,kfk2 2). 3) = 1 for r = 1, 2, and there 3) satisfying Jfr = fr and fr(0) = 1 for r = 1, 2. We assume (1) Neither the case (1) of Lemma 10.5 nor the case (1) of Lemma 10.7 occurs. (2) The case (2) of Lemma 10.5 or the case (2) of Lemma 10.7 could occur only if κ1 = −1 and kf1k2 = kf2k2 = 3. (3) The case (1) of Lemma 10.12 never occurs. (4) The case (2) of Lemma 10.12 could occur only if kf1k2 = kf2k2 = 2. Proof. Note that we have µ1 = µ1(0)f1 and µ2 = µ2(0)f2. The statements follow from for the cases of Lemma 10.5 and Lemma 10.7, and µ1(0)2kf1k2 + µ2(0)2kf2k2 = 1 3 , µ1(0)2kf1k2 + µ2(0)2kf2k2 = 1 3 (1 − 2 d ), for the case of Lemma 10.12. (cid:3) Proof of Theorem 10.22. Thanks to the previous lemma, the only possibilities are Case I with (ω1, ω2) = (1, ζ±1 3 ). We assume (ω1, ω2) = (1, ζ3) as the other case can be treated in the same way. Lemma 10.2,(4) with g = 2 implies ξ1(2)2 + 2η2(2)2 = ξ2(2)2 + 2η1(2)2. Since dim ker(R − ζ3) = 1, we have ξ2 = ξ2(0)f1 = ξ1(0)f1 and η2 = η2(0)f1 = −η1(0)f1. Since f′0(2) = 0, we have ξ1(2) = ξ1(0)f0(2) and η1(2) = η1(0)f0(2). Thus (ξ0(0)2 − 2η1(0)2)(f0(2) − f1(2)) = 0, 96 MASAKI IZUMI and ξ0(0) = ±√2η1(0). This shows that the case (2) of Lemma 10.8 never occurs. Since ξ2 = ξ1(0)f1 and η2 = −η1(0)f1, we have ξ2 = ∓√2η2, and Eq.(10.23) implies √5 − 2 . 6 1 3d = µ2(0)2kf1k2 = kµ2k2 = Now the equations in (1) of Lemma 10.8 imply κ1p1 − 16η1(0)2 √5 − 2 − κ1p1 − 16η1(0)2 √5 − 2 (− − 12 + 4 4 4 = ± √2η1(0), √5 − 2 . 6 )2kf1k2 = Direct computation shows that for kf1k2 as above, these is no η1(0) satisfying these two equations. 10.8. The case of G = Z2 × Z2. Theorem 10.24. There is no C∗ near-group category for Z2 × Z2 with m = 8. (cid:3) We use the notation in Example 9.6. Up to complex conjugate, there are exactly two a(g) for each of h·,·i1 and h·,·i2, and we show the statement for these 4 cases separately. Lemma 10.25. There is no C∗ near-group category for h·,·i1 and a(g0) = a(g1) = i, a(g2) = −1. Proof. In this case we can choose c = i, and R = −i 2 0 1 0 0 −i 0 0 −i 0 0 0  Let f0 = 1 − ζ −1 8√2 − ζ −1 8√2 i ,  1 1 1 −1 1 −1 −1 1 1 1 1 −1 1 1 −1 −1  f (0) f (g0) f (g1) f (g2) 0 0 0 0 −1  J f′0 = 0 ζ−1 8 i −ζ−1 8 i 0  =  , f1 = f (0) −if (g0) −if (g1) −f (g2)  1 2  =  .  1 ζ −1 8 ζ −1 8 2 cos 7π 12 2 cos 7π 12 tan 7π 12 i  −i −i −i −i −1 1 −1 1 −1 −1 1 1 i −i −i i  , , f2 =  1 ζ −1 8 2 cos π 12 ζ −1 8 2 cos π 12 − tan π 12i .  Then Rf0 = f0, Rf′0 = f′0, Rf1 = ζ3f1, Rf2 = ζ 2 3 f2, and they are invariant under J . They form a (non-normalized) orthogonal basis of ℓ2(Z2 × Z2) as well as the real subspace of J -invariant vectors. We have 3 kf1k2 = 2 cos2 7π 12 = 6(2 + √3), THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 97 kf2k2 = 3 2 cos2 π 12 = 6(2 − √3). Since f′0(g2) = 0, we can show the statement in exactly the same way as in the (cid:3) case of Z4. Lemma 10.26. There is no C∗ near-group category for h·,·i1 and a(g0) = −a(g1) = i, a(g2) = 1. Proof. In this case we can choose c = 1, and R = 1 2 0 1 0 0 0 −i 0 0 i 0 0 0 0 1 0 0  1 1 1 1 −i i −i i i −i −i i 1 −1 −1 1  ,  = f (0) −if (g0) if (g1) f (g2) 1 1 1 −1 1 −1 −1 1   J  0 ζ −1 8√2 ζ8√2 −1 1 1 1 −1 1 1 −1 −1 f (0) f (g0) f (g1) f (g2)  = , f1 = 1 2   .  , f2 = Let f0 =  , f′0 = 1 ζ −1 8 2√2 ζ8 2√2 1 2 1 12 ζ−1 −2 cos π 8 2 sin π 12 ζ8 −1 Then Rf0 = f0, Rf′0 = f′0, Rf1 = ζ3f1, Rf2 = ζ 2 3 f2, and they are invariant under J . They form a (non-normalized) orthogonal basis of ℓ2(Z2 × Z2) as well as the real subspace of J -invariant vectors. We have kf1k2 = kf2k2 = 6. Thanks to Lemma 10.23, the only remaining case is Case I with (ω1, ω2) = (1, ζ±1 3 ). We assume (ω1, ω2) = (1, ζ3) because the other case can be treated in the same way. Eq.(10.23) implies 1 12ζ−1 2 sin π 8 −2 cos π 12 ζ8 −1  . and so (ξ1(0)2 − 2η1(0)2)kf1k2 − 3µ2(0)2kf2k2 = − √5 − 2 3µ2(0)2 − ξ1(0)2 + 2η1(0)2 = 12 . 1 d , This shows that case (2) in Lemma 10.8 never occurs because the left-hand side would be negative in that case. We assume that the case (1) in Lemma 10.8 holds. Then we would get √5 − 2 12 = 3µ2(0)2 − ξ1(0)2 + 2η1(0)2 κ1p1 − 16η1(0)2 √5 − 2 )2 − ( = 3(− (√5 − 2)κ1p1 − 16η1(0)2 = −3 + 2√5 − 12 12 + 4 4 , √5 − 2 4 κ1p1 − 16η1(0)2 4 + )2 + 2η1(0)2 1 1 1 1 1 −1 −1 1 1 −1 1 −1 −1 1 −1 1  , R = 1 2  1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 −1  = f (0) f (g0) f (g1) −f (g2)   J  , f′0 = 1 1 1 1 −1 −1 1 −1 1 1 1 1 −1 1 −1 −1 f (0) f (g0) f (g1) f (g2)  =  , f1 = 1 2   .  , f2 = 98 and This is contradiction. MASAKI IZUMI κ1p1 − 16η1(0)2 = 3 + √5 3 > 1. (cid:3) Up to group automorphism there are exactly two a(g) for h·,·i2, the one given by a(g0) = a(g1) = 1, a(g2) = −1, and the one given by a(g0) = a(g1) = a(g2) = −1. Lemma 10.27. There is no C∗ near-group category for h·,·i2 and a(g0) = a(g1) = 1, a(a2) = −1. Proof. In this case, we can choose c = 1, and Let  . f0 = 1 1 2 1 2 0 0 1 −1 0 1 −1 −1 −√3i 1 −1 −1√3i Then Rf0 = f0, Rf′0 = f′0, Rf1 = ζ3f1, Rf2 = ζ 2 3 f2, and they are invariant under J . They form a (non-normalized) orthogonal basis of ℓ2(Z2 × Z2) as well as the real subspace of J -invariant vectors. We have kf1k2 = kf2k2 = 6. Since f′0(g3) = 0, we can show the statement in exactly the same way as in the case of Z4. (cid:3) Lemma 10.28. There is no C∗ near-group category for h·,·i2 and a(g0) = a(g1) = a(g2) = −1. Proof. In this case, we can choose c = −1, and 1 R = −1 1 2 2 −1 −1 −1 −1 1 −1 −1 1 1 −1 1 −1 1 −1 −1 1  ,  1 0 0 0 −1 0 0 −1 0 0 0 0 0 0 0 −1    J 1 1 1 −1 −1 1 −1 1 1 1 1 −1 1 −1 −1 f (0) f (g0) f (g1) f (g2)  = f (0) −f (g0) −f (g1) −f (g2)  =  . THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 99 Let 0 1 ζ 2 3 ζ3 0 1 ζ3 ζ 2 3 f0 =  , f1 =  , f′0 = 1 − i√3 − i√3 − i√3 Then Rf0 = f0, Rf′0 = f′0, Rf1 = ζ3f1, Rf2 = ζ 2 3 f2, and J f0 = −f′0, J f′0 = −f0, J f1 = f1, J f2 = f2. They form a (non-normalized) orthogonal basis of ℓ2(Z2 × Z2) as well as the real subspace of J -invariant vectors. We have kf1k2 = kf2k2 = 2. Lemma 10.23 and f0(0) = f′0(0) = 0 imply that the only possible case is the case (2) of Lemma 10.12. Thus there exist p ∈ C and κ ∈ {1,−1} such that µ0 = pf0+pf′0, and  , f2 =  . 1 i√3 i√3 i√3 √5 − 2 4√3 κ 4 + )if1 + ( √5 − 2 4√3 )if2. Let pi = pζ i 3 + pζ−i κ 4 − √5 − 2 6 , Now Lemma 10.2,(4) with g = gi is equivalent to µ = pf0 + pf′0 + ( 3 ∈ R. Then µ(0) = κ 2 √5 − 2 √5 − 2 4 4 Rµ(0) = − R2µ(0) = i, µ(gi) = pi + κ i, Rµ(gi) = pi − − 4 κ i, R2µ(gi) = pi − 4 √5 36 − √5 − 2 6 , p2 i − + 12 √5 − 2 √5 − 2 12 − κ 4 κ 4 i, i. ξ(gi)2 + 2η(gi)2 + p2 3 i + pi + η(gi)2 = √5 − 2 √5 − 2 √5 − 2 3 6 ξ(gi)η(gi) = (pi − )η(gi), η(gi)2 + ξ(gi)(pi + ) = 0. √5 9 , pi = (10.50) (10.51) (10.52) (10.53) From Eq.(10.52) and (10.53), we have 0 = η(gi)2η(gi) + ξ(gi)η(gi)(pi + √5 − 2 ) 6 √5 − 2 √5 − 2 ) = η(gi)2η(gi) + η(gi)(pi − = η(gi)(η(gi)2 + p2 i − 3 √5 − 2 6 )(pi + 6 9 − 4√5 ). 18 pi − and η(gi) = 0. Eq.(10.53) implies either ξ(gi) = 0 or pi = −(√5 − 2)/6, which in either case is not compatible with Eq.(10.50), (10.51). (cid:3) 100 MASAKI IZUMI 11. 2G l 1 subfactors Let n and l be natural numbers. A bipartite graph is said to be 2n l 1 if the following conditions hold: (1) the set of even vertices is {vi}n−1 (2) the set of odd vertices is {w}n−1 (3) the only non-zero entries of the incidence matrix Γ are i=0 ∪ {vρ}, i=0 ∪ {wπ}, Γvi,wi = Γwi,vi = 1, Γwi,vρ = Γvρ,wi = l, Γvρ,wπ = Γwπ,v = 1. Let N ⊂ M be a subfactor whose principal graph is 2n l 1. More precisely, by the principal graph, we mean the induction reduction graph for M − M and M − N sectors. Then the vertices {vi}n−1 i=0 correspond to automorphisms, forming a group G in Out(M), and we denote them by {αg}g∈G. We call such a subfactor 2G l 1. When l = 1, we just call it a 2G1 subfactor. For example, the E6 subfactor is 2Z21. Let ρ be the endomorphism of M corresponding to vρ. Then {αg}g∈G ∪ {ρ} generate a C∗ near-group category for G with m = ln. In particular G is always abelian. In this section, we determine the structure of 2G l 1 subfactors in terms of the corresponding C∗ near group categories. id ι α2ι α2 π ρ αι α Figure 1. 2Z3 2 1 subfactor Let N ⊂ M be a 2G l 1 subfactor and let {αg}g∈G and ρ be as above. We may and do assume αg ◦ ρ = ρ, and we use the same notation as in the previous sections for the restriction of αg and ρ to the intertwiner spaces (αg, ρ2), and (ρ, ρ2). Let ι : N ֒→ M be the inclusion map. Then odd vertices {wi}n−1 i=0 correspond to {αg ◦ ι}g∈G. Let π : N → M be the homomorphism corresponding to wπ. Direct computation shows d(ι) = d/√n, d(π) = √n, and [ρ] = [πι]. Since ρ is self-conjugate, we have [ρ] = [ιπ] too. Since [αg ◦ π] = [π] and d(π) = √n, we have [ππ] =Mg∈G From [ρ] = [ιπ], we may assume ρ = ιπ by replacing ρ with Ad v ◦ ρ and αg by Ad v ◦ αg ◦ Ad v∗ for a unitary v ∈ M if necessary. This means that we have M ⊃ N ⊃ ρ(M). Lemma 11.1. Let the notation be as above. [αg]. THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 101 (1) N = ρ(M) ∨ {U(g)}′′g∈G. (2) The dual inclusion of M ⊃ N is isomorphic to M α ⊃ ρ(M), where M α is the fixed point subalgebra M α = {x ∈ M; αg(x) = x, ∀g ∈ G}. Proof. (1) Since [π◦αg] = [π], there exist unitaries ug ∈ N satisfying π◦αg = Ad ug◦π. On the other hand, we have Ad U(g) ◦ ρ = ρ ◦ αg = ι ◦ π ◦ αg = ι ◦ Ad ug ◦ π, and ug is a multiple of U(g). This means that U(g) ∈ N, and we have the inclusion relation N ⊃ ρ(M) ∨ {U(g)}g∈G. Let ρ0 be ρ regarded as an isomorphism from M onto ρ(M), and let βg = ρ0 ◦ αg ◦ ρ−1 0 , which is an outer action of G on ρ(M). Since ρ(M) ∨ {U(g)}g∈G is identified with the crossed product ρ(M) ⋊β G, its index in M is d2/n, which coincides with [M : N]. Thus we have the equality N = ρ(M) ∨ {U(g)}g∈G. (2) Since [ρ] = [πι], we may assume ρ = π◦ι by replacing π with an equivalent sector if necessary. Since [αg ◦ π] = [π], there exist unitaries vg ∈ M such αg ◦ π = Ad vg ◦ π. On the other hand, we have αg ◦ ρ = ρ, and so vg is a scalar and αg ◦ π = π. Since d(π) = √n, this implies that the image of π coincides with M α. Thus the dual inclusion N ⊃ ι(N) is isomorphic to M α ⊃ ρ(M). Theorem 11.2. For an abelian group G, there is a one-to-one correspondence be- tween the isomorphism classes of 2G l 1 subfactors and the set of equivalence classes of a C∗ near-group category for G with m = ln. (cid:3) Proof. It is obvious that isomorphic 2G l 1 subfactors give rise to equivalent C∗ near- group categories. Moreover the previous lemma shows that if the two resulting C∗ near-group categories are equivalent, the two 2G l 1 subfactors are isomorphic. It remains to show that every C∗ near-group category for G with m = ln gives rise to a 2G l 1 subfactor. We may assume that such a category is realized by {αg}g∈G and ρ acting on a type III factor M, and we use the same notation as before. We set N = ρ(M) ∨ {U(g)}′′g∈G. Then it is identified with the crossed product ρ(M) ⋊ G, and its index in M is d2/n = 1 + ld. Since ρ(M) is irreducible in M, so is N too. We denote by ι the inclusion map ι : N ֒→ M. All we have to show is dim(ρ, ιι) = l, or equivalently dim(ι, ρι) = l, which will show Since (ι, ρι) ⊂ (ρ, ρ2), we have and in view of Eq.(3.30), we have [ιι] = [id] ⊕ l[ρ]. (ι, ρι) = {T ∈ K; ∀g ∈ G, T U(g) = ρ(U(g))T}, (ι, ρι) = {T ∈ K; ∀g ∈ G, (j1 ◦ j2)U(g)(j1 ◦ j2)∗T = T}, which shows dim(ι, ρι) = l. (cid:3) Grossman-Jordan-Snyder [22, Section 4] discussed Z2-graded extensions of near- group categories C in the case of prime m = n. They constructed a non-trivial 102 MASAKI IZUMI Z2-graded extension of C under the assumption that the outer automorphism group Out(C) is trivial and the corresponding 2G1 subfactor is self-dual. We will determine the structure of Out(C) in Section 13. The second condition turns out to always hold. Theorem 11.3. Let N ⊂ M be a 2G l 1 subfactor whose associated C∗ near-group category satisfies the following condition: the operator A(g) in Lemma 7.4 is a scalar for any g ∈ G. Then N ⊂ M is self-dual. In particular, every 2G1 subfactor is self-dual. l 1 subfactor N ⊂ M as before. Let ι′ : Proof. We use the same notation for a 2G ρ(M) ֒→ M α and κ : M α ֒→ M be the inclusion maps. Since ρ(M) ⊂ M α, the restriction of ρ to M α makes sense as an endomorphism of M α, which will be denoted by ρ′. Since αg(U(h)) = hg, hiU(h), the restriction of Ad U(−g) to M α makes sense too as a G-action on M α, which will be denoted by α′. By definition, we have ρ◦ κ = κ ◦ ρ′ and κ◦ α′g = Ad U(−g) ◦ κ. We also have α′g ◦ ρ′ = ρ′ as Ad U(−g) ◦ ρ = ρ ◦ αg. We first claim that that ρ′ is irreducible. Indeed, we have (ρ′, ρ′) = (κ ◦ ρ′, κ ◦ ρ′) ∩ M α = (ρ ◦ κ, ρ ◦ κ) ∩ M α, and we first determine (ρ ◦ κ, ρ ◦ κ). Since Ad U(g) ◦ ρ = ρ ◦ αg, we have U(g) ∈ (ρ ◦ κ, ρ ◦ κ). On the other hand, dim(ρ ◦ κ, ρ ◦ κ) = dim(κκ, ρ2) = dim(Mg∈G αg, ρ2) = n, and so (ρ ◦ κ, ρ ◦ κ) = span{U(g)}g∈G. This shows that ρ′ is irreducible. Next we observe [ι′ι′] = [id] ⊕ l[ρ′], or equivalently dim(ι′, ρ′ ◦ ι′) = l. This follows from (ι′, ρ′ ◦ ι′) = (ρ, ρ2) ∩ M α = {T ∈ K; ∀g ∈ G, V (g)T = T}. Now it suffices to show that the two categories generated by {αg}g∈G ∪{ρ} and by {α′g}g∈G ∪ {ρ′} are mutually equivalent, and for this we show that the their Cuntz algebra models are the same. Let S′0 = Then for any x ∈ M α, we have 1 ρ′2(x)S′0 = 1 √nXg∈G Sg ∈ M α. Sgαg(x) = 1 √nXg∈G Sgx = S′0x, and S′0 ∈ (id, ρ′2). We set S′g = α′g(S′0) = √nXg∈G √nXh∈G 1 U(−g)ShU(−g)∗ = 1 √nXh∈G hg, hiShU(g), THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 103 which is in (α′g, ρ′2). Note that we have S′gα′g(x)S′∗g = Xg∈G On the other hand, 1 n Xg,h,k∈G hg, k − hiShxS∗k =Xh∈G ShxS∗h. Tg(er)ρ(x)Tg(er)∗, Tg(er)U(g)ρ(x)U(g)∗Tg(er)∗ Sgαg(x)S∗g +Xg,r ρ′2(x) =Xg =Xg =Xg SgxS∗g +Xg,r S′gα′g(x)S′g∗ +Xg,r Tg(er)U(g)ρ′(x)U(g)∗Tg(er)∗. Since Tg(ξ)U(g) ∈ M α, we see that {Tg(er)U(g)}g,r form an orthonormal basis of (ρ′, ρ′2). Since ρ(U(−g)) ∈ M α, and Ad ρ(U(−g)) ◦ ρ′ = ρ′ ◦ α′g, we set U′(−g) = ρ(U(g)). Then thanks to Eq.(3.30), we have U′(g)S′h = 1 √nXk hh, kiρ(U(−g))SkU(h) = 1 √nXk hh, kiSk+gU(h) = hh, giS′h. By assumption, the operator A(g) in Lemma 7.4 is a scalar a(g). We choose a square root of a(0) ∈ T and fix it, and set Th(ξ)U(h) ∈ (ρ′, ρ′2) ∩ (M α)α′ , T ′0(ξ) = a(0)1/2 √n Xh∈G T ′g(ξ) = U′(−g)T ′0(ξ) = a(0)1/2 a(0)1/2 a(0)1/2 √n Xh∈G √n Xh∈G √n Xh∈G √n Xh∈G a(0)1/2a(g) = = = ρ(U(g))Th(ξ)U(h) (j1 ◦ j2)∗U(g)(j1 ◦ j2)Th(ξ)U(h + g) Th+g(A(h + g)A(h)∗ξ)U(h + g) hg, hiTh(ξ)U(h). Then {T ′g(er)}g,r is an orthonormal basis of (ρ′, ρ′2), and we have α′g(T ′h(er)) = hg, hiT ′g(er) and U′(g)T ′h(er) = T ′h−g(er). 104 MASAKI IZUMI For T ′ ∈ (ρ′, ρ′2), we defined j′1(T ′) and j′2(T ′) as in the definition of j1 and j2 by n = a(0)−1/2a(g)√d a(0)−1/2a(g)√d hg, hiU(h)∗Th(ξ)∗ρ(Sk) replacing ρ and S0 with ρ′ and S′0. Then j′1(T ′g(ξ)) = √dT ′g(ξ)∗ρ′(S0) Xh,k∈G Xh,k∈G n Xh,k∈G hg, hiU(h)∗j1(Th+k(ξ))U(k)∗ n Xh,k∈G a(k + h)hg, hiT−k(Jξ)U(k)∗ a(0)−1/2a(g) a(0)−1/2a(g) = = = n hg, hiU(h)∗Th(ξ)∗U(k)ρ(S0)U(k)∗ Since we get 1 √nXh hg, hia(k + h) = hg, kia(−g) = hg, kia(0)a(−g), j′1(T ′g(ξ)) = hg, kiT−k(Jξ)U(k)∗ = T ′ −g(A(−g)Jξ), a(0)1/2 √n Xk∈G which shows that j′1 has the same form as j1. For j′2, we have n = = = a(0)−1/2a(g) ǫa(0)−1/2a(g) a(0)−1/2a(g)√d hg, hiρ(U(h)∗Th(ξ)∗)Sk j′2(T ′g(ξ)) = √dρ(T ′g(ξ)S′0) Xh,k n Xh,k hg, hiρ(U(h)∗)αk(j2(Th(ξ))) n√n Xh,k,p √n Xh √n Xh √n Xh hg, hia(h)T−h(C∗Jξ)U(h)∗ a(g − h)T−h(C∗Jξ)U(h)∗. hg, hiρ(U(h)∗)T0(C∗Jξ) ǫa(0)−1/2a(g) ǫa(0)−1/2a(g) ǫa(0)−1/2 = = = hg − p, hihk, piρ(U(h)∗)Tp(C∗Jξ) THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 105 On the other hand, ǫ √nXk ǫa(0)1/2 hg, kiT ′k(C∗Jξ) n Xh,k √n Xh √n Xh ǫa(0)−1/2 ǫa(0)1/2 = = = hg + h, kia(k)Th(C∗Jξ)U(h) a(−g − h)Th(C∗Jξ)U(h) a(g − h)Th(C∗Jξ)U(h). Thus j′2 has the same form as j2. Now it suffices to show that Bh(ξ) for ρ′ takes the same form as the original one. Note that we have T ∗g (er)ρ(T0(eu))T0(et) = a(g)Xs br,s t,u(g)T−g(es). On the other hand, T ′g(er)∗ρ(T ′0(eu))T ′0(et) = = = = = a(0)1/2 a(0)1/2 √n Xh √n Xh n√n Xh,p,q n√n Xh,p,q a(0)1/2a(g) a(0)1/2 a(0)1/2 √n Xh T ′g(er)∗ρ(Th(eu)U(h))T ′0(et) T ′g(er)∗U′(h)ρ(T0(eu))U′(−h)T ′0(et) T ′g+h(er)∗ρ(T0(eu))T ′h(et) a(g + h)a(h)hq, g + hihp, hiU(q)∗Tq(er)∗ρ(T0(eu))Tp(et)U(p) hq − p − g, hihq, giU(q)∗Tq−p(er)∗ρ(T0(eu))T0(et)U(p), where we used Tq(ζ)∗ρ(T0(ξ))Tp(η) = Tq(ζ)∗ρ(T0(ξ))U(−p)T0(η) = Tq(ζ)∗U(−p)ρ(αp(T0(ξ)))T0(η) = Tq−p(ζ)∗ρ(T0(ξ))T0(η). 106 MASAKI IZUMI Thus we get T ′g(er)∗ρ(T ′0(eu))T ′0(et) hp + g, giU(p + g)∗Tg(er)∗ρ(T0(eu))T0(et)U(p) br,s t,u(g)hp + g, giU(p + g)∗T−g(es)U(p) = = = a(0)1/2a(g) a(0)1/2a(g)2 √n Xp √n Xp,s √n Xp,s = a(g)Xs a(0)1/2 br,s t,u(g)T ′ br,s t,u(g)hp, giTp(es)U(p) −g(es), which finishes the proof. (cid:3) 12. Orbifold construction I (de-equivariantization) Orbifold construction for subfactors was first introduced in [17] to construct new subfactors from given ones with group actions (see [38] too). Purely categorical versions (see [4], [11], [53] for example) of it are now called equivariantization and de-equivariantization, which are dual operations to each other via Takesaki dual- ity. In the final two sections, we systematically investigate these operations for 2G l 1 subfactors and near-group categories in the irrational case. Victor Ostrik is the first to observe that a near-group category for Z3 × Z3 with m = 9 produces the Haagerup category via de-equivariantization (see Example 12.13 below). In this section, we systematically pursue this kind of phenomena. We con- centrate on the case with m = G, though our argument makes sense for the case of m > G under a mild assumption. A C∗ near group category in this class is com- pletely described by the data (h·,·i, a, b, c) as in Section 9. The pair (h·,·i, a) is often called a quadratic form on G in the literature, where a is the complex conjugate of a. Let N ⊂ M be a 2G1 subfactor whose even part is a C∗ near-group category given by α and ρ with αg ◦ ρ = ρ. We assume that it has an invariant (h·,·i, a, b, c), and we use the same notation as before for the Cuntz algebra model. Now Eq.(3.30) takes the form ρ(U(g)) =Xh∈G Sh−gS∗h +Xh∈G a(h)a(h − g)ThU(g)T ∗h−g. 12.1. Untwisted case. We fix a subgroup H ⊂ G, and consider the crossed product M ⋊α H, which is a factor generated by M and a unitary representation {λh} of H with the relation λhx = αh(x)λh for x ∈ M. Then we can extend ρ and α to M ⋊ H by setting ρ(λh) = a(h)U(h)λh, and αg(λh) = hg, hiλh. Note that we have THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 107 Ad(a(h)U(h)λh)(ρ(x)) = ρ(αh(x)) for x ∈ M, and (a(h)U(h)λh)(a(k)U(k)λk) = a(h)a(k)U(h)αh(U(k))λh+k = a(h)a(k)hh, kiU(h + k)λh+k = a(h + k)U(h + k)λh+k. Lemma 12.1. With the notation as above, we have the following for all x ∈ M ⋊α H. ρ2(x) =Xg∈G Sg αg(x)S∗g +Xg∈G Tg ρ(x)T ∗g , Proof. It suffices to show the statement for x = λh. For x = λh the left-hand side is a(h)2ρ(U(h))U(h)λh, and ρ ◦ αg = Ad U(g) ◦ ρ. S∗g ρ2(λh)Sg = a(h)2S∗g ρ(U(h))U(h)Sg+hλh = a(h)2hh, g + hiS∗g ρ(U(h))Sg+hλh = hh, giλg = αg(λh), T ∗g ρ2(λh)Tg = a(h)2T ∗g ρ(U(h))U(h)λhTg = a(h)2hh, giT ∗g ρ(U(h))U(h)Tgλh = a(h)2hh, giT ∗g ρ(U(h))Tg−hλh = a(h)2a(g)a(g − h)hh, giU(h)λh = a(h)U(h)λh = ρ(λh), which show the first statement. For the second statement, we have and on the other hand, ραg(λh) = hg, hia(h)U(h)λh, Ad U(g)ρ(λh) = a(h)U(g)U(h)λhU(g)∗ = a(h)U(g + h)αh(U(g)∗)λh = a(h)hh, giU(h)λh. (cid:3) For a subgroup H ⊂ G, we set H⊥ = {g ∈ G; hg, hi = 1}. Since Ad λ∗h ◦ αh is trivial on M, it comes from the dual action of α. If moreover H ⊂ H⊥, it is trivial, that is αh = Ad λh. Lemma 12.2. If H ⊂ H⊥, then (ρ, ρ) = {λh}′′h∈H. Proof. Note that we have (ρ, ρ) ⊂ M ⋊α H ∩ ρ(M)′. Since αh ◦ ρ = ρ, and ρ is irreducible, we get M ⋊α H ∩ ρ(M)′ = {λh}′′h∈H. Now we have (ρ, ρ) = {λh}′′h∈H ∩ {ρ(λk)}′k∈H = {λh}′′h∈H ∩ {a(k)U(k)λk}′k∈H, which is again {λh}′′h∈H because λh commutes with U(k) thanks to αh(U(k)) = hh, kiU(k) = U(k). (cid:3) 108 MASAKI IZUMI For the character group H of H, we use the additive notation, that is, for χ1, χ2 ∈ H, we denote (χ1 + χ2)(h) = χ1(h)χ2(h). For χ ∈ H, we denote by eχ the corre- sponding minimal projection in {λh}′′h∈H, that is Assume H ⊂ H⊥. Then since eχ = 1 pHXh∈H (ρ, ρ) =Mχ∈ H χ(h)λh. Ceχ, the endomorphism ρ is decomposed into irreducible components σχ parametrized by χ ∈ H. More precisely, we choose an isometry Vχ ∈ M ⋊α H for each χ ∈ H satisfying VχV ∗χ = eχ, and set σχ(x) = V ∗χ ρ(x)Vχ. For simplicity, we denote σ = σ0. Then we have [ρ] =Mχ∈ H [σχ]. Note that {σχ}χ∈ H are all inequivalent. Recall that we may assume N = ρ(M) ∨ {U(g)}′′g∈G, and N is regarded as the crossed product of ρ(M) by the G-action ρ0 ◦ αg ◦ ρ−1 0 , where ρ0 is ρ regarded as an isomorphism from M onto ρ(M). Since αh(U(h)) = hg, hiU(g), the restriction of α to N is identified with the dual action for this crossed product, and in consequence α on N is outer too. Thus N ⋊α H is a subfactor of M ⋊α H. Theorem 12.3. Let κ : N ⋊α H ֒→ M ⋊α H be the inclusion map. Then [κκ] = [id] ⊕Mχ∈ H [σχ]. Proof. Since [M ⋊α H : N ⋊α H] = [M : N], it suffices to show that σχ is contained in κκ, or equivalently (κ, σχκ) 6= {0}. Let ι : N ֒→ M be the inclusion map. In the proof of Theorem 11.2, we showed (ι, ρι) = {T ∈ K; ∀g ∈ G, (j1 ◦ j2)U(g)(j1 ◦ j2)∗T = T} = C(j1 ◦ j2)∗T0. Note that we have (j1 ◦ j2)U(g)(j1 ◦ j2)∗ = a(g)U(−g)V (−g). We claim that (j1 ◦ j2)∗T0λh = ρ(λh)(j1 ◦ j2)∗T0 holds. Indeed, the right-hand side is a(h)U(h)λh(j1 ◦ j2)∗T0 = a(h)U(h)αh((j1 ◦ j2)∗T0)λh = a(h)U(h)V (h)(j1 ◦ j2)∗T0λh, THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 109 and the claim follows. Now we have V ∗χ (j1 ◦ j2)∗T0 ∈ (κ, σχκ). It remains to show V ∗χ (j1 ◦ j2)∗T0 6= 0, which follows from (V ∗χ (j1 ◦ j2)∗T0)∗V ∗χ (j1 ◦ j2)∗T0 = χ(h)T ∗0 j1 ◦ j2λh(j1 ◦ j2)∗T0 1 1 HXh∈H HXh∈H HXh∈H 1 1 H = = = χ(h)T ∗0 j1 ◦ j2αh((j1 ◦ j2)∗T0)λh χ(h)T ∗0 j1 ◦ j2V (h)(j1 ◦ j2)∗T0λh . (cid:3) Remark 12.4. From Theorem 11.3 and its proof, we can see that the subfactor N ⋊α H ⊂ M ⋊α H is self-dual too. the fusion rules for the categories generated by {σχ}χ∈ H. To determine the principal graph of N ⋊α H ⊂ M ⋊α H, it suffices to determine For g ∈ G, we denote by χg ∈ H the character of H determined by χg(h) = hh, gi. Since every character of H extends to a character of G and the bicharacter h·,·i is non-degenerate, the map G ∋ g 7→ χg ∈ H is a surjection, giving an isomorphism from G/H⊥ onto H. Direct computation shows that αg(eχ) = eχ−χg, and eχU(g) = U(g)eχ+χg. This implies [ αgσχ] = [σχ−χg] and [σχ αg] = [σχ+χg ]. In particular, when k ∈ H⊥, we have [ αkσχ] = [σχ αk] = [σχ]. Theorem 12.5. Assume that H ⊂ H⊥. Then there exists ga ∈ G with a(h) = hh, gai for all h ∈ H, and [σ][σ] = Mk∈H ⊥/H [ αk−ga] ⊕ H⊥/H Mg∈G/H ⊥ [ αgσ], [ αg][σ] = [σ][ α−g]. Note that since [ αg+h] = [ αg] for any h ∈ H and [ αkσ] = [σ] for any k ∈ H⊥, the above expression makes sense. Proof. For h1, h2 ∈ H, we have a(h1)a(h2) = hh1, h2ia(h1 + h2) = a(h1 + h2), and the restriction of a to H is a character. Thus there exists ga ∈ G satisfying a(h) = hh, gai for any h ∈ H. We choose a transversal {ki}i∈H ⊥/H ⊂ H⊥ for H. Note that we have σ2(x) = V ∗0 ρ(V ∗0 )ρ2(x)ρ(V0)V0 =Xg∈G V ∗0 ρ(V ∗0 )Sg αg(x)S∗g ρ(V0)V0 +Xg∈G V ∗0 ρ(V ∗0 )Tg ρ(x)T ∗g ρ(V0)V0, 1 H 1 HXh∈H e0Xh∈H Xg∈G = Xk∈H ⊥ = Xi∈H ⊥/HXh∈H = Xi∈H ⊥/HXh∈H = H Xi∈H ⊥/H a(h)U(h)Sg = hh, ga + giSg = 1H ⊥(ga + g)e0Sg, where 1H ⊥ is the indicator function of H⊥. Now V ∗0 ρ(V ∗0 )Sg αg(x)S∗g ρ(V0)V0 V ∗0 ρ(V ∗0 )e0Sk−ga αk−ga(x)S∗k−gae0 ρ(V0)V0 V ∗0 ρ(V ∗0 )e0Sh+ki−ga αh+ki−ga(x)S∗h+ki−gae0 ρ(V0)V0 V ∗0 ρ(V ∗0 )e0λhSki−ga αki−ga(x)S∗ki−gaλ∗he0 ρ(V0)V0 V ∗0 ρ(V ∗0 )Ski−ga αki−ga(x)S∗ki−ga ρ(V0)V0. 110 MASAKI IZUMI and the range projection of ρ(V0)V0 is ρ(V0)e0 ρ(V0)∗ = ρ(V0V ∗0 )e0 = ρ(e0)e0 = = We have a(h)U(h)e0 = a(h)e−χhU(h) = 1 HXh∈H a(h)U(h)λhe0 1 H e0Xh∈H a(h)U(h). 1 HXh∈H e0Xh∈H 1 H mutually orthogonal ranges. For the second term we have V ∗0 ρ(V ∗0 )Tg ρ(x)T ∗g ρ(V0)V0 a(h1)a(h2)V ∗0 ρ(V ∗0 )e0U(h1)Tg ρ(x)T ∗g U(h2)∗e0 ρ(V0)V0 It is straightforward to see that {pHV ∗0 ρ(V ∗0 )Ski−ga}i∈H ⊥/H are isometries with Xg∈G H2Xg∈G Xh1,h2∈H H4Xg∈G Xh1,h2,h3,h4∈H H4Xg∈G Xh1,h2,h3,h4∈H H2Xg∈G Xh1,h2∈H H2Xg∈G Xh1,h2∈H a(h1)a(h2)hh3 + h4, giV ∗0 ρ(V ∗0 )Tg−h1λh3+h4 ρ(x)T ∗g−h2 ρ(V0)V0 a(h1)a(h2)V ∗0 ρ(V ∗0 )λh3Tg−h1 ρ(x)T ∗g−h2λh4 ρ(V0)V0 a(h1)a(h2)V ∗0 ρ(V ∗0 )Tg−h1e−χg ρ(x)T ∗g−h2 ρ(V0)V0 a(h1)a(h2)V ∗0 ρ(V ∗0 )Tg−h1V−χgσ−χg(x)V ∗ −χg T ∗g−h2 ρ(V0)V0. = 1 1 1 1 1 = = = = THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 111 We choose a transversal {gj}j∈G/H ⊥ ⊂ G for H⊥. Then we get 1 = V ∗0 ρ(V ∗0 )Tg ρ(x)T ∗g ρ(V0)V0 Xg∈G H2 Xj∈G/H ⊥ Xi∈H ⊥/H Xh0,h1,h2∈H × V ∗0 ρ(V ∗0 )Tgj+ki+h0−h1V−χgj H Xj∈G/H ⊥ Xi∈H ⊥/H Xh1,h2∈H = × V ∗0 ρ(V ∗0 )Tgj+ki−h1V−χgj σ−χgj 1 a(h1)a(h2) (x)V ∗ σ−χgj a(h1)a(h2) T ∗gj+ki+h0−h2 ρ(V0)V0. −χgj (x)V ∗ T ∗gj+ki−h2 ρ(V0)V0. −χgj a(h)V ∗0 ρ(V ∗0 )Tgj +ki+hV−χgj}j∈G/H ⊥, i∈H ⊥/H , Direct computation shows that { 1 pHXh∈H is a family of isometries with mutually orthogonal ranges. Therefore the statement is proved. (cid:3) Remark 12.6. (1) The above theorem in particular shows that σ is self-conjugate if and only if a(h) = 1 for any h ∈ H. Since a(h)a(h) = hh, hi = 1 for h ∈ H, we have 2ga = 0, and non-trivial ga could occur only if H is an even group. (2) The original near-group category is Morita equivalent to the fusion category generated by σ and the dual action α of α. A subgroup H ⊂ G is called Lagrangian if H = H⊥ and the restriction of a to H is 1 (see [13]). Corollary 12.7. If H is a Lagrangian, we have [σ][σ] = [id] ⊕ Mg∈G/H [ αg][σ] = [σ][ α−g]. [ αgσ], In [32], we gave a Cuntz algebra construction of the fusion categories with the above type of fusion rules, other than near-group categories, which was further pursued in [15]. In a forthcoming paper, we systematically investigate such fusion categories for arbitrary finitely abelian groups. In this note, we just give a converse construction of the above corollary in the case of odd groups. Let R be a type III factor and K be an odd abelian group, which will play the role of H as well as G/H. Assume that we are given a map β : K → Aut(R) inducing an injective homomorphism from K into Out(R), and irreducible σ ∈ End(R) satisfying [βk][σ] = [σ][β−k], [σ][σ] = [id] ⊕ lMk∈K [βk][σ]. Note that the Frobenius reciprocity implies that {βkσ}k∈K are all inequivalent. 112 MASAKI IZUMI Since K ∋ k 7→ [βk] ∈ Out(R) is a homomorphism, there exist u(k1, k2) ∈ U(R) satisfying βk1 ◦ βk2 = Ad u(k1, k2) ◦ βk1+k2. The associativity βk1 ◦ (βk2 ◦ βk3) = (βk1 ◦ βk2) ◦ βk3 implies that there exists a 3-cocycle ω ∈ Z 3(K, T) satisfying βk1(u(k2, k3))u(k1, k2 + k3) = ω(k1, k2, k3)u(k1, k2)u(k1 + k2, k3). Recall that the opposite algebra Ropp of R is a linear space R equipped with a new product x · y = yx. To realize Ropp as a von Neumann algebra, we identify it with the commutant R′ of R under identification of x ∈ R and JRx∗JR ∈ R′ where JR is the modular conjugation of R. We denote by j the anti-linear multiplicative map : R ∋ x 7→ JRxJR ∈ Ropp, and define βopp (x) = j ◦ βk ◦ j−1 ∈ Aut(Ropp). Let R = R ⊗ Ropp, let γk = βk ⊗ βopp k ∈ R, and let v(k1, k2) = u(k1, k2) ⊗ j(u(k1, k2)). Then γ and v satisfy the 2-cocycle relation k γk1(v(k2, k3))v(k1, k2 + k3) = v(k1, k2)v(k1 + k2, k3), and we can construct the twisted crossed product M = R ⋊γ,v K, that is a factor generated by R and unitaries {λv k for x ∈ R and k1λv λv k1+k2. Let κ : R ֒→ M be the inclusion map. Then the fusion categories generated by {κ(βk ⊗ id)κ}k∈K is nothing but the quantum double Dω(K) (see [31]). k}k∈K satisfying λv k2 = v(k1, k2)λv kx = γk(x)λv Let θk1(k2, k3) = ω(k2, k1, k3) ω(k1, k2, k3)ω(k2, k3, k1) . Then it is known that θk is a 2-cocycle for any fixed k ∈ K (see [7], [8]). We assume that θk is a coboundary for any k ∈ K, which automatically holds for cyclic groups. This is equivalent to the condition that Dω(D) is pointed, that is, every simple object of Dω(K) is invertible, which in our situation means that βk ⊗ id extends to M as we see now. Thanks to this assumption, there exists νk1(k2) satisfying θk(k′, k′′) = νk(k′ + k′′)νk(k′)νk(k′′). Lemma 12.8. Let the notation be as above. The automorphism βk⊗id on R extends to M by βk(λv k′) = νk(k′)(u(k, k′)u(k′, k)∗ ⊗ 1)λv k′. Proof. It suffices to verify the following two relations: k′(βk ⊗ id)(x) νk(k′)(u(k, k′)u(k′, k)∗ ⊗ 1)λv = (βk ⊗ id) ◦ γk′(x)νk(k′)(u(k, k′)u(k′, k)∗ ⊗ 1)λv k′, ∀x ∈ R, νk(k′)(u(k, k′)u(k′, k)∗ ⊗ 1)λv = νk(k′ + k′′)(βk ⊗ 1)(w(k′, k′′))(u(k, k′ + k′′)u(k′ + k′′, k)∗ ⊗ 1)λv k′νk(k′′)(u(k, k′′)u(k′′, k)∗ ⊗ 1)λv k′′ k′+k′′. The first one is easy to verify and the second one is equivalent to νk(k′)νk(k′′)u(k, k′)u(k′, k)∗βk′(u(k, k′′)u(k′′, k)∗)u(k′, k′′) = νk(k′ + k′′)βk(u(k′ + k′′))u(k, k′ + k′′)u(k′ + k′′, k)∗. THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 113 Using the defining relation of ω, we get βk(u(k′ + k′′))u(k, k′ + k′′)u(k′ + k′′, k)∗ = ω(k, k′, k′′)u(k, k′)u(k + k′, k′′)u(k′ + k′′, k)∗ = ω(k, k′, k′′)u(k, k′)u(k′, k)∗(u(k′, k)u(k + k′, k′′))u(k′ + k′′, k)∗ = ω(k, k′, k′′)ω(k′, k, k′′)u(k, k′)u(k′, k)∗βk′(u(k, k′′))u(k′, k + k′′)u(k′ + k′′, k)∗ = ω(k, k′, k′′)ω(k′, k, k′′)ω(k′, k′′, k)u(k, k′)u(k′, k)∗βk′(u(k, k′′)u(k′′, k)∗)u(k′k′′) = θk(k′, k′′)u(k, k′)u(k′, k)∗βk′(u(k, k′′)u(k′′, k)∗)u(k′k′′), which shows the statement. (cid:3) Note that the group of the equivalence classes of the simple objects of Dω(K) is {[ βk ◦ γχ]}k∈K, χ∈ K, where γ is the dual action of γ. We denote this group by G. Thus G is an extension: 0 → K → G → K → 0. Theorem 12.9. Let the notation be as above. The fusion category generated by κ(σ ⊗ id)κ is a C∗ near-group category for G with m = lK2. Proof. Let ρ = κ(σ ⊗ id)κ. Then ρ is irreducible because dim(βkσ ⊗ βopp k , σβk′ ⊗ βopp k′ ) = 1. For ρ2, we have dim(ρ, ρ) = dim(κκ(σ ⊗ id), (σ ⊗ id)κκ) = Xk,k′∈K )κ] =Mk∈K [ρ2] = [κ(σ ⊗ id)κκ(σ ⊗ id)κ] =Mk∈K [κ(σβkσ ⊗ βopp [κ(β−2k ⊗ id)κ] ⊕ l Mk,k′∈K [κγk(β−2kσ2 ⊗ id)κ] =Mk∈K =Mk∈K =Mk∈K [ β−2kκκ] ⊕ lKMk′∈K [κ(βk′σ ⊗ id)κ]. k where we used κ ◦ βk = βk ◦ κ. Since K is an odd group, the first term is [κ(β−kσ2 ⊗ βopp k )κ] [κ(β−2kβk′σ ⊗ id)κ] Mk∈K, χ∈ K [ βkγχ]. dim(κ(βkσ ⊗ id)κ, κ(σ ⊗ id)κ) = dim(κκ(βkσ ⊗ id), (σ ⊗ id)κκ) dim(βk+k′σ ⊗ βopp k′ ), ((σβk′′ ⊗ βopp dim(βk+k′σ, β−k′σ) k′′ )) = Xk′∈K For the second term, = Xk′,k′′∈K = Xk′∈K dim(βk+2k′σ, σ) = 1. 114 MASAKI IZUMI This shows [ρ2] = Mk∈K, χ∈ K [ βk γχ] + lK2[ρ]. (cid:3) Remark 12.10. Let Cσ be the fusion category generated by σ, and let Cρ be the near-group category generated by ρ in Theorem 12.9. Then the above construction shows that Cρ is categorically Morita equivalent to Cσ ⊠ Vecω K is the category of K-graded vector spaces with twist ω. In consequence, we have equivalence Z(Cρ) ∼= Z(Cσ)⊠Z(Vecω K), where Z(C) denotes the Drinfeld center of C. This explains why the modular data of Z(Cρ) factors as the product of those of Z(Cσ) and Dω(K) in Example 12.13 and Example 12.14 below (see [15, Section2 and Section 3] and [16, p.636]). Example 12.11. For G = Z2 × Z2, there is a unique C∗ near-group category with m = 4, and a non-trivial subgroup H with H ⊂ H⊥ is unique, which is H = {0, g3} in Example 9.6. The subgroup H is Lagrangian, and by de-equivariantization, we get the even part of the A7-subfactor. K, where Vecω Example 12.12. There are two C∗ near-group categories for G = Z4, which are mutually complex conjugate, and H = Z2 satisfies H = H⊥, though it is not La- grangian. In this case, de-equivariantization by H gives two 1-supertransitive subfac- tors constructed Liu-Morrison-Penneys [41]. In fact they already observed that the two subfactors they constructed give rise to the 2Z41 subfactors by equivariantization (see [41, Section 4.3]). Example 12.13. Let G = Z3 × Z3. Evans-Gannon [16, Table 2] showed that there is a unique C∗ near-group category for G = Z3 × Z3, and the quadratic form in the solution is given as h(x1, x2), (y1, y2)i = ζ x1x2−y1y2 . There are exactly two Lagrangians: H1 = {(0, 0), (1, 1), (2, 2)} and H2 = {(0, 0), (1, 2), (2, 1)}, and they produce two different fusion categories by de-equivariantization. A priori, two different subgroups could produce equivalent categories, but this is not the case now because of the following reason. In the case of H = H1, we have , a(x1, x2) = ζ x2 3 1−y2 1 3 [σ2] = [id] ⊕Mh∈H2 [ αhσ], and α restricted to H2 is an action, and therefore there is no third cohomology obstruction for the Z3 part. When H = H2, the situation is the same. In fact, Grossman-Snyder [21] showed that there are exactly two C∗-fusion categories with the same fusion rules as the Haagerup category (the even part of the Haagerup subfactor) and with trivial third cohomology obstruction for the Z3 part. Thanks to Theorem 12.9, which is the converse construction of Corollary 12.7, the two groups H1 and H2 indeed produce two different fusion categories. Example 12.14. Evans-Gannon [15] showed that there are exactly two C∗ near- group categories for Z9, and they are complex conjugate to each other. In this case, the subgroup H =< 3 >∼= Z3 is a Lagrangian, which produces two more fusion THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 115 categories with the same fusion rules as the Haagerup category. The Z3 part of the resulting fusion categories is generated by α1, and its third cohomology obstruction, known as Connes obstruction [5], can be easily computed as follows. Indeed, we have α3 1 = λ3 and α1(λ3) = h1, 1i3λ3, which shows that the Connes obstruction is either ζ3 or ζ−1 3 . The existence of these two categories has been expected since Evans- Gannon [15, Question 5] obtained the corresponding hypothetical modular data of the Drinfeld center. Example 12.15. It is known that there exist fusion categories as in Theorem 12.9 for G = Zn, n = 5, 7, 9, with l = 1 and trivial third cohomology obstruction on the group part (see [32], [15]). It follows that there exist C∗ near-group categories for G = Zn × Zn with m = n2 for n = 5, 7, 9. 12.2. Twisted case. In the tensor category setting (see [11]), de-equivariantization with respect to a subgroup H ⊂ G can be defined for an algebra structure of the object αh, Mh∈H because it is known that αh has a unique half-braiding (see [32, Lemma 6.1]), and hence αh uniquely lifts to the Drinfeld center. In operator algebra setting, an algebra structure of the above object corresponds to a lifting of {[αh]}h∈H ⊂ Out(M) to In the case of near-group categories, we have a privileged lifting a group action. determined by αh ◦ ρ = ρ, and we essentially discussed de-equivariantization with this algebra structure in the previous subsection. The other algebra structures are in one-to-one correspondence with H 2(H, T) \ {0}, and the corresponding operation in terms of operator algebras is cocycle twisted crossed product. Let G, M, ρ, and α be as in the previous subsection. Now we do not assume that H satisfies the condition H ⊂ H⊥. Instead, we assume that H ∼= Z2s for a natural number s, and that the restriction of h·,·i to H is non-degenerate. We further assume that there exists a 2-cocycle ω ∈ Z 2(H, T) satisfying ω(h, k)ω(k, h) = hh, ki ∈ R. Let M ⋊α,ω H be the twisted crossed product, that is, the von Neumann algebra generated by M and unitaries {λω h for any x ∈ M and λω h1+h2. As before we can extend ρ and α to M ⋊α,ω H by setting ρ(λω h. As in Lemma 12.1, we have the following for all x ∈ M ⋊α H. h}h∈H satisfying λω h ) = hg, hiλω h2 = ω(h1, h2)λω h) = a(h)U(h)λω h x = αh(x)λω h , and αg(λω h1λω 2 ρ2(x) =Xg∈G Sg αg(x)S∗g +Xg∈G ρ ◦ αg = Ad U(g) ◦ ρ. Tg ρ(x)T ∗g , We claim that λω h commutes with ρ(λω k ) = λω λω h ρ(λω h a(k)U(k)λω = a(k)hh, kiω(h, k)U(k)λω = ρ(λω h = ρ(λω k )λω h. k )λω k ) for any h, k ∈ H. Indeed, we have k = a(k)αh(U(k))λω h+k = a(k)hh, kiω(h, k)ω(k, h)U(k)λω h λω k k λω h Lemma 12.16. Let the notation be as above. 116 MASAKI IZUMI (1) αh = Ad λω (2) (ρ, ρ) = {λω h for any h ∈ H. h}′′h∈H. Proof. (1) Note that the restriction of h·,·i to H is real-valued. Since Ad λω trivial on M, it suffices to show h∗ ◦ αh(λω h∗ h+kλω k ) = hh, kiλω Ad λω k λω = hh, kiω(h, k)ω(k, h)λω hλω h λω h∗ = hh, kiω(h, k)λω k λω h∗ = hh, ki2λω k = λω k . (2) As in the proof of Lemma 12.2, we have (M ⋊α,ω H) ∩ ρ(M)′ = {λω (ρ, ρ) = {λω h}′′h∈H ∩ {ρ(λω k )}′k∈H = {λω h}′′h∈H. h∗ ◦ αh is h}′′h∈H, and (cid:3) Since h·,·i restricted to H is non-degenerate, the twisted group algebra {λω h}′′h∈H is isomorphic to the full matrix algebra M2s(C). Thus (2) above shows that there exists an irreducible σ ∈ End(M ⋊α,ω H) satisfying [ρ] = 2s[σ]. On the other hand we have [ρ2] =Mg [ αg] ⊕ n[ρ] = H Mg∈G/H [ αg] ⊕ n2s[σ]. This implies Theorem 12.17. Let the notation be as above. Then [σ2] = Mg∈G/H [ αg] ⊕ 2sG/H[σ]. In particular σ generates a C∗ near-group category for G/H with m = 2sG/H. We have already seen in Subsection 10.6 that there are exactly two C∗ near-group categories for Z3 with m = 6 (resp. 2Z3 2 1 subfactors), and there is a D8-actions on each of them, where D8 is the dihedral group of order 8. We will see in Example 13.15 that equivariantization by this action produces a fusion category containing a C∗ near- group category for Z2×Z2×Z3 with m = 12 (resp. a 2Z2×Z2×Z31 subfactor). Theorem 12.17 should be considered as a converse construction of this equivariantization in spite of a superficial discrepancy: twisted crossed product is used instead of ordinary crossed product, and Z2 × Z2 is not the dual group of D8. The first point can be easily fixed because it is well-known that every outer cocycle action of a finite group is known to be equivalent to an ordinary action. It means that there is a family of unitaries {wh}h∈H in M satisfying whαh(wk) = ω(h, k)wh+k. Now the twisted crossed product in Theorem 12.17 is identified with the ordinary crossed product by the new action Ad wh ◦ αh. For the second point, let N ⊂ M be one of the 2Z3 2 1 subfactors, and let γ be the D8-action, which we will see in Section 13. Let Z(D8) ∼= Z2 be the center of D8. Then it is easy to show that the inclusion N ⋊γ Z(D8) ⊂ M ⋊γ Z(D8) is still 2Z3 2 1, and it is the fixed point algebra pair of N ⋊γ D8 ⊂ M ⋊γ D8 under the dual action γ restricted to the group part Hom(D8, T) ∼= Z2 × Z2 of the unitary dual D8. This reasoning indicates that there must be two missing solutions of the polynomial equations for G = Z2 × Z2 × Z3 with m = 12 in Evans-Gannon's list [16, Table 2], THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 117 for which the restriction of h·,·i to Z2 × Z2 should be h·,·i2 in Example 9.6. Since they are complex conjugate to each other, we give one of them now. Example 12.18. Let G = H × K with H = Z2 × Z2 and K = Z3, and let h(h, k), (h′, k′)i = hh, h′i2ζ kk′ 3 , a(h, k) = aH (h)ζ k2 3 , with aH(0) = aH (g1) = aH(g2) = 1, aH (g3) = −1, where we use the parametrization H = {0, g0, g1, g2} as in Example 9.6. Then there is a unique solution for the equations in Theorem 9.1, which is given by c = e−πi/6, d = 6 + 4√3 and  + 1 2p2√3  0 1 −1 0  ⊗ 0 ζ3i −ζ3i  1 − √3 3 b = + 1 6  1 −1 −1 −√3i  1 1 2 1 2 0  ⊗  ⊗ 1 ζ3 ζ3 1 − ζ3 − ζ3 2 2  , where we identify ℓ2(H × K) with ℓ2(H) ⊗ ℓ2(K), and ℓ2(H) and ℓ2(K) with C4 and C3 via orthonormal bases < δ0, δg0, δg1, δg2 > and < δ0, δ1, δ2 > respectively. We can apply Theorem 12.17 to this solution with ω given by following table, and obtain the 0 g0 g1 g2 1 1 0 1 -i g0 1 1 g1 1 -i i 1 i g2 1 1 i 1 -i C∗ near-group category for Z3 with m = 6 discussed in Subsection 10.6. Note that the above solution is invariant under θ ∈ Aut(G) given by θ(0, x) = (0, x), θ(g0, x) = (g1,−x), (g1, x) = (g0,−x), (g2, x) = (g2, x). 13. Orbifold construction II (equivariantization) 13.1. Automorphism groups. In this subsection, we investigate the structure of the outer automorphism group of a C∗ near-group category C in the irrational case. To realize group actions on C as those on a concrete operator algebra, we use the Cuntz algebra model discussed in Section 4. We fix a solution (ǫ,h·,·i, A, C, J, Bg) of the equations in Theorem 8.7, and define a G-action α on On+m and an endomorphism ρ as before. We introduce a subgroup Autρ(On+m) of Aut(On+m) by Autρ(On+m) = {γ ∈ Aut(On+m); γ ◦ ρ = ρ ◦ γ}. 118 MASAKI IZUMI Lemma 13.1. For any γ ∈ Autρ(On+m), there exist ω ∈ T, θ ∈ Aut(G), and u ∈ U(K0) satisfying hθ(g), θ(h)i = hg, hi, A(θ(g)) = uA(g)u∗, uC = Cu, uJ = Ju, and Bθ(g)(ξ) = (u ⊗ u)Bg(u∗ξ)u∗, such that γ(Sg) = ω2Sθ(g), γ(Tg(ξ)) = ωTθ(g)(uξ). On the other hand, for any triple (ω, θ, u) ∈ T × Aut(G) × U(K0) satisfying the above condition, there exists γ(ω,θ,u) ∈ Autρ(On+m) satisfying γ(ω,θ,u)(Sg) = ω2Sθ(g), γ(ω,θ,u)(Tg(ξ)) = ωTθ(g)(uξ). Moreover two different (ω, θ, u) and (ω′, θ′, u′) give the same automorphism if and only if ω′ = −ω, θ′ = θ, and u′ = −u. Proof. It is straightforward to show the second part, and we show only the first part here. Let γ ∈ Autρ(On+m). Since (id, ρ2) = CS0 and γ commutes with ρ, we have γ(S0) ∈ (id, ρ2), and γ(S0) is a multiple of S0. Replacing γ with γ(eit,id,I) ◦ γ for appropriate t, we may assume γ(S0) = S0. We claim that there exists θ ∈ Aut(G) satisfying γ ◦ αg ◦ γ−1 = αθ(g) and γ(Sg) = Sθ(g). Since γ(ρ2, ρ2) = (ρ2, ρ2), and (ρ2, ρ2) =Mg∈G CSgS∗g ⊕ KK∗, γ ◦ αg ◦ γ−1 ◦ ρ = ρ = αg ◦ ρ, there exists a permutation θ of G fixing 0 such that γ(SgS∗g ) = Sθ(g)S∗θ(g). Thus S∗θ(g)γ(Sg) is a unitary belonging to (γ ◦ αg ◦ γ−1, αθ(g)). Since we have (γ ◦ αg ◦ γ−1, αθ(g)) ⊂ (ρ, ρ) = C, which implies that γ ◦ αg ◦ γ−1 = αθ(g) and θ is a group automorphism of G. Moreover, γ(Sg) = γ(αg(S0)) = αθ(g)(γ(S0)) = αθ(g)(S0) = Sθ(g). Since Ad γ(U(g)) ◦ ρ = γ ◦ Ad U(g) ◦ ρ ◦ γ−1 = γ ◦ ρ ◦ αg ◦ γ−1 = ρ ◦ αθ(g), γ(U(g))S0 = γ(U(g)S0) = γ(S0) = S0, We have γ(U(g)) = U(θ(g)), which together with αg(U(h)) = hg, hiU(h) implies hθ(g), θ(h)i = hg, hi. Since γ commutes with ρ, it induces a unitary transformation of (ρ, ρ2) = K. Since αg(γ(T0(ξ))) = γ(αθ−1(g)(T0(ξ))) = γ(T0(ξ)), we have γ(T0(ξ)) ∈ K0, and there exists a unitary u ∈ U(K0) satisfying γ(T0(ξ)) = T0(uξ). For Tg(ξ), we have γ(Tg(ξ)) = γ(U(−g)T0(ξ)) = U(−θ(g))T0(uξ) = Tθ(g)(uξ). Since γ(S0) = S0, we get γ ◦ j1 = j1 ◦ γ and γ ◦ j2 ◦ γ, and in consequence Lemma 7.4 implies that A(θ(g)) = uA(g)u∗, uC = Cu, and uJ = Ju. It remains to show that the condition γ(l(T )) = l(γ(T )) is equivalent to Bθ(g)(ξ) = (u ⊗ u)Bg(u∗ξ)u∗ under the other conditions we have verified so far. Thanks to THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 119 Lemma 7.18, the former is equivalent to γ(Bg(ξ)) = Bθ(g)(uξ). Since Bg(ξ) ∈ K2 we get γ(Bg(ξ)) = (u ⊗ u)Bg(ξ)u∗, and we are done. 0K∗0, (cid:3) Recall that the periodic 1-parameter automorphism group {γ(eit,id,I)}t∈R, which is a central subgroup of Autρ(On+m), has a unique KMS state, whose GNS construction produces a type III factor M extending On+m (see Appendix and [28]). We use the same symbols α, ρ, γ(ω,θ,u) for their extensions to M as before. We denote by C ⊂ End(M) the fusion category generated by ρ. To avoid possible confusion, we make the only exception in this note here, and use the tensor symbol ⊗ for the monoidal product of C instead of composition, and do not omit the symbol HomC for the spaces of intertwiners in the following arguments. To emphasize that X ∈ M is regarded as an element in HomC(µ, ν) = {X ∈ M; Xµ(x) = ν(x)X, ∀x ∈ M}, we use the notation (νXµ) following [9, Section IV]. (Caution: the order of µ and ν are switched in (νXµ) and HomC(µ, ν).) Note that we have Iσ ⊗ (νXµ) = (σ ⊗ νσ(X)σ ⊗ µ) ∈ HomC(σ ⊗ µ, σ ⊗ ν), (νXµ) ⊗ Iσ = (ν ⊗ σXµ ⊗ σ) ∈ HomC(µ ⊗ σ, ν ⊗ σ). An automorphism of C is a tensor auto-equivalence of C as a C∗-category, which is a pair (F, L) consisting of an auto-equivalence F and natural isomorphisms Lσ1,σ2 : F (σ1) ⊗ (σ2) → F (σ1 ⊗ σ2) satisfying Lσ1⊗σ2,σ3 ◦ (Lσ1,σ2 ⊗ IF (σ3)) = Lσ1,σ2⊗σ3 ◦ (Iσ1 ⊗ Lσ2,σ3). g The automorphism group Aut(C) of C is the collection of automorphisms of C modulo tensor natural isomorphisms. Since we are working on C∗ categories, an isomorphism between two objects are always assumed to be unitary. An automorphism is inner if it is equivalent to the one given by β ⊗ · ⊗ β−1 for an invertible object β in C, which is always equivalent to αg ⊗ · ⊗ α−1 for some g ∈ G in our case. The outer automorphism group Out(C) of C is Aut(C) modulo the normal subgroup of inner automorphisms. Every γ ∈ Autρ(On+m) extends to M and induces an automorphism of C, which is denoted by Fγ (we omit the natural isomorphisms L : Fγ(σ1)⊗ Fγ(σ2) → Fγ(σ1 ⊗ σ2) as it is trivial now). More precisely, we define Fγ(σ) = γ ◦ σ◦ γ−1 for an object σ ∈ C, and define Fγ on the Hom-spaces by the restriction of γ. For γ = γ(ω,θ,u), we have Fγ(αg) = αθ(g), Fγ(ρ) = ρ, Lemma 13.2. An automorphism γ ∈ Autρ(On+m) induces a trivial automorphism of C if and only if γ = γ(ω,id,IK0 ) for ω ∈ T. Proof. Since Fγ is a trivial automorphism only if it fixes the equivalence class of each simple object, we assume γ = γ(ω,id,u). Assume that η is a natural tensor isomorphism between Fγ and id. We identify ηαg ∈ HomC(αg, αg) and ηρ ∈ HomC(ρ, ρ) with complex numbers of modulus 1. Note that we have HomC(αg+h, αg ⊗ αh) = CIM, HomC(ρ, αg ⊗ ρ) = CIM, HomC(ρ, ρ ⊗ αg) = CU(g), HomC(αg, ρ ⊗ ρ) = CSg, and 120 MASAKI IZUMI HomC(ρ, ρ ⊗ ρ) = K. Since γ acts on the first three trivially, we get ηαg = Iαg . For Sg ∈ HomC(αg, ρ ⊗ ρ) and Tg(ξ) ∈ HomC(ρ, ρ ⊗ ρ), we have (ηρ ⊗ ηρ) ◦ Fγ((ρ ⊗ ρSgαg)) = (ρ ⊗ ρSgαg) ◦ ηαg , (ηρ ⊗ ηρ) ◦ Fγ((ρ ⊗ ρTg(ξ)ρ)) = (ρ ⊗ ρTg(ξ)ρ) ◦ ηρ, which are equivalent to η2 may assume u = IK0, and γ = γ(η−1 ρ ,id,IK0 ). ρω2 = 1 and ηρωu = IK0. Since γ(ω,id,u) = γ(−ω,id,−u), we Tracing back the above argument, we can see that γ(ω,id,IK0 ) induces a trivial au- (cid:3) tomorphism of C. We set Then we have Autρ,0(On+m) = {γ ∈ Autρ(On+m); γ(S0) = S0} = {γ(1,θ,u) ∈ Autρ(On+m)}. Autρ(On+m)/T ∼= Autρ,0(On+m)/Z2, where T is identified with {γ(eit,id,I)}t∈R/2πZ and Z2 is identified with {γ(1,id,±IK0)}. The above lemma shows that there is an injective homomorphism from Autρ,0(On+m)/Z2 into Aut(C). Recall that the gauge group G(A, C, J) is the set of all unitaries in U(K0) commut- ing with A(g), C, and J. Theorem 13.3. With the above notation, we have Aut(C) ∼= Autρ(On+m)/T ∼= Autρ,0(On+m)/Z2 ∼= {(θ, u) ∈ Aut(G) × U(K0); hθ(g), θ(h)i = hg, hi, uC = Cu, uJ = Ju, A(θ(g)) = uA(g)u∗, Bθ(g)(ξ) = (u ⊗ u)Bg(u∗ξ)u∗}/{(id,±IK0)}. In particular, the kernel of the canonical homomorphism from Aut(C) to Aut(G) is isomorphic to {u ∈ G(A, C, J); Bg(ξ) = (u ⊗ u)Bg(u∗ξ)u∗}/{±IK0}. Proof. It suffices to show that the association Autρ,0(On+m) ∋ γ 7→ Fγ induces a map onto Aut(C). Let (F, L) be an automorphism of C. Up to a tensor natural iso- morphism, we may assume F (αg) = αθ(g) and F (ρ) = ρ with a group automorphism θ ∈ Aut(G). Then we have F (αg) ⊗ F (αh) = αθ(g) ⊗ αθ(h) = αθ(g+h) = F (αg+h) = F (αg ⊗ αh), We claim that we may further assume F (αg) ⊗ F (ρ) = αθ(g) ⊗ ρ = ρ = F (ρ) = F (αg ⊗ ρ). F (ρ ⊗ αg) = F (ρ) ⊗ F (αg), F (ρ ⊗ ρ) = F (ρ) ⊗ F (ρ), Indeed, there exist x(g, h) ∈ T and and Lαg ,αh, Lαg ,ρ, Lρ,αg , and Lρ,ρ are trivial. y(g) ∈ T satisfying Lαg,αh = (αθ(g+h)x(g, h)αθ(g+h)) and Lαg ,ρ = (ρy(g)ρ), and the relation Lαg⊗αh,ρ ◦ (Lαg,αh ⊗ IF (ρ)) = Lαg ,αh⊗ρ ◦ (IF (αg) ⊗ Lαh,ρ) THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 121 implies x(g, h)y(g + h) = y(g)y(h). Now passing from (F, L) to (F ′, L′) by a tensor natural isomorphism η : F → F ′ with ηαg = (αθ(g)y(g)αθ(g)), ηρ = (ρIMρ), ηρ⊗αg = y(g)L−1 ρ,ρ, we get the claim. ρ,αg, and ηρ⊗ρ = L−1 Note that we have F ((αg ⊗ αhIMαg+h)) = F (αg+hIMαg+h) = (αθ(g+h)IMαθ(g+h)) = (αθ(g) ⊗ αθ(h)IMαθ(g+h)), F ((αg ⊗ ρIMρ)) = F (ρIMρ) = (ρIMρ) = (αθ(g) ⊗ ρIMρ). We will freely use these relations and their adjoints in what follows. Since F ((ρ⊗ ρS0id)) is an isometry belonging to HomC(id, ρ⊗ ρ), it is of the form ω(ρ ⊗ ρS0id) with ω ∈ T. Thus replacing F with Fγ(ω−1,id,IK) ◦ F , we may further assume F ((ρ ⊗ ρS0id)) = (ρ ⊗ ρS0id). Since Sg = αg(S0), we have (ρ ⊗ ρSgαg) = ((ρ1αg ⊗ ρ) ⊗ Iρ) ◦ (Iαg ⊗ (ρ ⊗ ρS0id)). Applying F to this, we get F ((ρ ⊗ ρSgαg)) = (ρ ⊗ ρSθ(g)αθ(g)) and its adjoint. we introduce an automorphism γ ∈ Aut(On+m) by Since F restricted to HomC(ρ, ρ⊗ ρ) is a unitary transformation of HomC(ρ, ρ⊗ ρ), F (((ρ ⊗ ρSgαg))) = (ρ ⊗ ργ(Sg)αθ(g)), F ((ρ ⊗ ρTρ)) = (ρ ⊗ ργ(T )ρ). The above computation shows γ(Sg) = Sθ(g). Recall that we have HomC(ρ, ρ ⊗ αg) = CU(g). Thus F ((ρ ⊗ αgU(g)ρ)) is a multiple of (ρ ⊗ αθ(g)U(θ(g))ρ). Since U(g)S0 = S0, we have (ρ ⊗ ρS0id) = (Iρ ⊗ (ρIMαg ⊗ ρ)) ◦ ((ρ ⊗ αgU(g)ρ) ⊗ Iρ) ◦ (ρ ⊗ ρS0id). Applying F to this, we get F ((ρ⊗ αgU(g)ρ)) = (ρ⊗ αθ(g)U(θ(g))ρ) and its adjoint. We claim F ((ρ ⊗ αgU(g)ρ)) = (ρ ⊗ αθ(g)γ(U(g))ρ), or equivalently γ(U(g)) = U(θ(g)). Indeed, applying F to the two equalities (ρ ⊗ ρU(g)Shαh) = (Iρ ⊗ (ρIMαg ⊗ ρ)) ◦ ((ρ ⊗ αgU(g)ρ) ⊗ Iρ) ◦ (ρ ⊗ ρShαh), (ρ ⊗ ρU(g)Tρ) = (Iρ ⊗ (ρIMαg ⊗ ρ)) ◦ ((ρ ⊗ αgU(g)ρ) ⊗ Iρ) ◦ (ρ ⊗ ρTρ), we get γ(U(g)Sh) = U(θ(g))γ(Sh) and γ(U(g)T ) = U(θ(g))γ(T ). Choosing an or- thonormal basis {Ti}m i=1 of HomC(ρ, ρ ⊗ ρ), we get γ(U(g)Sh)γ(Sh)∗ + γ(U(g)Ti)γ(Ti)∗ γ(U(g)) =Xh∈G =Xh∈G mXi=1 mXi=1 U(θ(g))γ(ShS∗h) + U(θ(g))γ(TiT ∗i ) = U(θ(g)), which shows the claim. 122 MASAKI IZUMI We have already seen that F coincides with γ on the following morphism spaces: HomC(αg+h, αg ⊗ αh), HomC(ρ, αg ⊗ ρ), HomC(ρ, ρ ⊗ αg), HomC(αg, ρ ⊗ ρ), HomC(ρ, ρ ⊗ ρ), and their adjoints. To finish the proof, it suffices to show γ ∈ Autρ,0(On+m). Before showing it, we claim γ ◦ αg = αθ(g) ◦ γ. This is easily shown for Sh. Applying F to (ρ ⊗ ραg(T )ρ) = ((ρIMαg ⊗ ρ) ⊗ Iρ) ◦ ((Iαg ⊗ (ρ ⊗ ρTρ)) ◦ (αg ⊗ ρIMρ), we get γ(αg(T )) = αθ(g)(γ(T )). Applying F to the two equations (ρ ⊗ ρj1(T )ρ) = √d((ρ ⊗ ρTρ)∗ ⊗ Iρ) ◦ (Iρ ⊗ (ρ ⊗ ρS0id)), (ρ ⊗ ρj2(T )ρ) = √d(Iρ ⊗ (ρ ⊗ ρTρ)∗) ◦ ((ρ ⊗ ρS0id) ⊗ Iρ), we see that γ commutes with j1 and j2. Thus Eq.(3.3) implies and γ(ρ(S0)) = ρ(S0) = ρ(γ(S0)), γ(ρ(Sg)) = γ(U(g)ρ(S0)U(g)∗) = U(θ(g))ρ(S0)U(θ(g))∗ = ρ(Sθ(g)) = ρ(γ(Sg)). It remains to show γ(ρ(T )) = ρ(γ(T )) for T ∈ HomC(ρ, ρ ⊗ ρ). We recall the two i=1 TiT ∗i , which satisfies γ(P ) = P , γ(Q) = Q, and P + Q = IM. Then Eq.(3.12) implies γ(ρ(T ))P = ρ(γ(T ))P and P γ(ρ(T )) = P ρ(γ(T )). Note that Qρ(T )Q = l(T ) ∈ K2K∗. Let T ′, T ′′, T ′′′ ∈ HomC(ρ, ρ ⊗ ρ). Then T ′′∗T ′∗ρ(T )T ′′′ is a scalar. Denoting it by ξ, we get projections P =Pg∈G SgS∗g and Q =Pm ξIρ = (ρT ′′ρ ⊗ ρ)∗ ◦ ((ρT ′ρ ⊗ ρ)∗ ⊗ Iρ) ◦ (Iρ ⊗ (ρTρ ⊗ ρ)) ◦ (ρT ′′′ρ ⊗ ρ). Applying F to it, we get γ(T ′′)∗γ(T ′)∗ρ(γ(T ))γ(T ′′′) = ξ, and so T ′′∗T ′∗γ−1(ρ(γ(T )))T ′′′ = ξ = T ′′∗T ′∗ρ(T )T ′′′. This shows Qγ−1(ρ(γ(T )))Q = Qρ(T )Q and equivalently Qρ(γ(T )))Q = Qγ(ρ(T ))Q, which finishes the proof. (cid:3) Theorem 13.4. When A(g) is a scalar a(g)I for any g ∈ G, the inner automorphism group of C is trivial. In consequence, Out(C) = Aut(C) ∼= Autρ(On+m)/T ∼= Autρ,0(On+m)/Z2 ∼= {(θ, u) ∈ Aut(G) × G(A, C, J); a(θ(g)) = a(g), Bθ(g)(ξ) = (u ⊗ u)Bg(u∗ξ)u∗}/{(id,±I)}. Proof. Let F = αh ⊗ · ⊗ α−1 that η determined by h . Then F (αg) = αg and F (ρ) = Ad U(h)−1 ◦ ρ. We show ηαg = hg, hiIαg ∈ HomC(αh ⊗ αg ⊗ α−1 h , αg), ηρ = (ρa(h)U(h) Ad U(h)−1 ◦ ρ) ∈ HomC(αh ⊗ ρ ⊗ α−1 h , ρ), is a tensor natural isomorphism from F to id. Indeed, it suffices to verify (ηαg1 ⊗ ηαg2 ) ◦ F ((αg1 ⊗ αg2IMαg1+g2)) = (αg1 ⊗ αg2IMαg1+g2) ◦ ηαg1+g2 , THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 123 (ηαg ⊗ ηρ) ◦ F ((αg ⊗ ρIMρ)) = (αg ⊗ ρIMρ) ◦ ηρ, (ηρ ⊗ ηαg) ◦ F ((ρ ⊗ αgU(g)ρ)) = (ρ ⊗ αgU(g)ρ) ◦ ηρ, (ηρ ⊗ ηρ) ◦ F ((ρ ⊗ ρSgαg)) = (ρ ⊗ ρSgαg) ◦ ηαg , which are equivalent to (ηρ ⊗ ηρ) ◦ F ((ρ ⊗ ρTρ)) = (ρ ⊗ ρTρ) ◦ ηρ, hg1, hihg2, hi = hg1 + g2, hi, αg(a(h)U(h))hg, hi = a(h)U(h), hh, gia(h)U(h)αh(U(g)) = U(g)a(h)U(h), ρ(a(h)U(h))a(h)U(h)αh(Sg) = hg, hiSg, ρ(a(h)U(h))a(h)U(h)αh(T ) = T a(h)U(h). It is straightforward to show these equalities. (cid:3) Corollary 13.5. When m = n, Out(C) = Aut(C) ∼= {θ ∈ Aut(G); hθ(g), θ(h)i = hg, hi, a(θ(g)) = a(g), b(θ(g)) = b(g)}. Let ΓC = {(θ, u) ∈ Aut(G) × G(A, C, J); Bθ(g)(ξ) = (u ⊗ u)Bg(u∗ξ)u∗}, which is isomorphic to Autρ,0(On+m). For x ∈ ΓC, we denote γx instead of γ(1,x) for simplicity. Remark 13.6. Assume that A(g) is a scalar for any g ∈ G. Then there exists a Out(C)-graded extension C′ of C such that Out(C) on C ⊂ C′ is implemented by inner automorphisms of C′. Indeed, let µ = γ(id,−I), and consider the crossed product M ⋊µ Z2, which is the von Neumann algebra generated by M and a period two unitary λ satisfying λx = µ(x)λ for any x ∈ M. We can extend ρ, α, and γ to M ⋊µ Z2 by ρ(λ) = λ, αg(λ) = λ, and γ(λ) = λ. We have γ(id,−I) = Ad λ. Let µ be the dual action of µ, which is a period two automorphism of M ⋊µ Z2 given by −λ, µ(x) =(cid:26) x, Sg αg(x)S∗g +Xg,r x ∈ M x = λ . Tg(er)µ ◦ ρ(x)Tg(er)∗. Then it is straightforward to show ρ2(x) =Xg∈G Now we can see that the category generated by µ ◦ ρ is equivalent to C, and we can set C′ to be the category generated by µ ◦ ρ and γΓC . Example 13.7. Any solutions of Eq.(9.1)-(9.5) given in Section 9 have trivial sym- metry, and the corresponding C∗ near-group has trivial automorphism group. 124 MASAKI IZUMI 5 5 , there is a unique a(g), which is given by . There is a unique solution of Eq.(9.1)-(9.5) as follows (see [32, Example Example 13.8. For G = Z5 and hg, hi = ζ gh a(g) = ζ 2g2 A 4] and [16, Table 2]): c = −1, d = (3 + 3√5)/2, b(0) = −1/d, ζ5√5 b(2) = b(3) = b(1) = b(4) = . ζ−1 5√5 , In particular, the solution has a symmetry a(g) = a(−g) and b(g)) = b(−g), and therefore we have Out(C) ∼= Z2 for the corresponding near-group category C for G = Z5 and m = 5. Example 13.9. The solution for G = Z2 × Z2 × Z3 with m = 12 given in Example 12.18 has a Z2-symmetry, and Out(C) ∼= Z2 for the corresponding near-group category C. Example 13.10. As we have seen in Section 10, there are exactly two C∗ near-group categories for G = Z3 with m = 6. For each of the two categories, the solutions of Eq.(8.15)-(8.25) are parametrized by {(x, y) ∈ R2; x2 + y2 = √3/24}. The group automorphism Z3 ∋ g 7→ −g ∈ Z3 takes (x, y) to (−x,−y). The gauge group G(A, C, J) is O(2), and acts on the solutions as (x, y) 7→ (x,−y), and the rotation 0 L =(cid:18) 1 0 −1 (cid:19) R(θ) =(cid:18) cos θ − sin θ cos θ (cid:19) sin θ acts on the solutions as (x, y) 7→ (cos 4θx + sin 4θy,− sin 4θx + cos 4θy). We fix a solution Bg corresponding to (q√3/24, 0). Identifying Autρ,0(O9) with {(θ, u) ∈ Aut(Z3) × O(2); Bθ(g)(ξ) = (u ⊗ u)Bg(u∗ξ)u∗}, we see that it is generated by the two elements (−1, R( π 4 )), (id, L). Thus Autρ,0(O9) is isomorphic to the dihedral group D16 of order 16, and the au- tomorphism group Aut(C), as well as the outer automorphism group Out(C), of the corresponding near-group category C are isomorphic to D8. 13.2. The case of m > G. In this section, we assume that Γ is a finite subgroup of the group {u ∈ G(A, C, J); Bg(ξ) = (u ⊗ u)Bg(u∗ξ)u∗}. We regard γ as an outer action of Γ on M via γu = γ(1,id,u). Recall that N = ρ(M) ∨ {U(g)}g∈G ⊂ M is a 2G l 1 subfactor with l = m/n. The purpose of this subsection is to determine the structure of the subfactor N ⋊γ Γ ⊂ M ⋊γ Γ. Let {λu}u∈Γ be the implementing unitary representation in M ⋊γ Γ. Then we can extend α and ρ to M ⋊γ Γ by THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 125 αg(λu) = λu and ρ(λu) = λu. We would like to describe the structure of the fusion category generated by ρ. Lemma 13.11. Let the notation be as above. (1) The two commuting actions α and γ give an outer action of G × Γ on M. (2) ρ is irreducible. (3) α is an outer action of G. Proof. (1) The statement follows from Theorem 14.1 below. (2) Assume X ∈ (ρ, ρ). Then X is uniquely expanded as X = Pu∈Γ Xuλu with Xu ∈ M. For any x ∈ M, we have Xu∈Γ ρ(x)Xuλu =Xu∈Γ Xuλuρ(x) =Xu∈Γ Xuγu ◦ ρ(x)λu, and so ρ(x)Xu = Xuγu◦ρ(x). The Frobenius reciprocity and (1) show that (γuρ, ρ) ∼= (γu, ρ2) is trivial except for u = I, and X ∈ (ρ, ρ) = C, which means that ρ is irreducible. (3) Assume that Y ∈ M ⋊γ Γ satisfies Y x = αg(x)Y for any x ∈ M ⋊γ Γ. Yu ∈ (γu, αg) = Cδu,Iδg,0 as before. This implies that Y = 0 unless g = 0, and so α is an outer action. Then Y is expanded as Y = Pu∈Γ Yuλu with Yu ∈ M, and we can show that (cid:3) Since we are working on a non-commutative group Γ, we recall the dual action of γ as a Roberts action of the dual Γ of Γ. Recall that a Hilbert space H in M is a closed subspace of M in the weak operator topology such that W ∗V is a scalar for any V, W ∈ H. The space H is equipped with the inner product hV, Wi for V, W ∈ H given by W ∗V . Let {Vi}i be an orthonormal basis of a Hilbert space H in M. The support of H is defined by ViV ∗i , Xi ρH(x) =Xi VixV ∗i , which does not depend on the choice of the orthonormal basis. When the support is I, we say that H has full support. For a full support Hilbert space H in M, we can define a unital endomorphism ρH of M by whose definition does not depend on the choice of the orthonormal basis either. When H is globally invariant under the Γ-action γ, the endomorphism ρH preserves the fixed point algebra MΓ = {x ∈ M; γu(x) = x, ∀u ∈ Γ}, and we denote by γH the restriction of ρH to MΓ. The equivalence class of γH depends only on the unitary equivalence class of H as a representation space of Γ. Since γ is outer, if the representation H is irreducible, so is γH. For any finite dimensional (not necessary irreducible) unitary representation π of Γ, there exists a full support Hilbert space Hπ in M so that the Γ-action on H is equivalent to π. We use the notation γπ instead of γHπ for simplicity. Let ν : MΓ ֒→ M be the inclusion map. Then we have 126 MASAKI IZUMI [νγπ] = dim π[ν] by definition, and so [νν] =Mπ∈Γ dim π[γπ], Let {V (π)i}dim π by the Frobenius reciprocity. We call {γπ}π∈Γ the pre-dual action of γ. i=1 be an orthonormal basis of Hπ. Applying the above argument to the second dual action γu ⊗ Ad u on M ⊗ B(ℓ2(Γ)), where {u}u∈Γ is the right regular representation of Γ, we can define the dual action γπ ∈ End(M ⋊γ Γ), which is explicitly given by the following formula: γπ(x) =(cid:26) Pdim π λu, i=1 V (π)ixV (π)∗i , x ∈ M x = λu . which is a unitary belonging to (ργπ, γπ ρ). (2) can be shown in a similar way. (3) By the Frobenius reciprocity, we have dim(γπ1 αg1, γπ2 αg2) = dim(γπ2⊗π1, αg2−g1). Thus it suffices to show dim(γτ , αg) = δτ,1δg,0 for τ ∈ Hom(Γ, T) and g ∈ G. Thanks to Remark 13.12, we can further replace γτ with Ad V ∗τ ◦γτ . Assume that X ∈ M⋊γ Γ satisfies X Ad V ∗τ ◦ γτ (x) = αg(x)X for any x ∈ M ⋊γ Γ. Then X is expanded as This can been seen by identifying λu ∈ M ⋊γ Γ with I ⊗ λu ∈ M ⊗ B(ℓ2(Γ)), where {λu} in the latter is the left regular representation of Γ, and M with πγ(M) ⊂ M ⊗ B(ℓ2(Γ)), where πγ(x)ξ(u) = γu−1(x)ξ(u), and by using the Hilbert space {V ⊗ I; V ∈ Hπ} in M ⊗ B(ℓ2). In fact V (π)i ⊗ I commutes with I ⊗ λu and we have dim πXi=1 (V (π)i ⊗ I)πγ(x)(V (π)∗i ⊗ I) = πγ( V (π)ixV (π)∗i ). dim πXi=1 Remark 13.12. When τ ∈ Hom(Γ, T) is a 1-dimensional representation, the Hilbert space Hτ is spanned by a single unitary Vτ with γu(Vτ ) = τ (u)Vu. Thus we have Ad V ∗τ ◦ γτ (x) =(cid:26) x, τ (u)λu, x ∈ M x = λu , which recovers the usual dual action in the commutative case. Lemma 13.13. Let the notation be as above. (1) [γπ ρ] = [ργπ]. (2) [γπ αg] = [ αg γπ]. (3) For π1, π2 ∈ Γ and g1, g2 ∈ G, we have dim(γπ1 αg1, γπ2 αg2) = δπ1,π2δg1,g2. Proof. (1) Choosing a basis {V (π)i}dim π i=1 of Hπ, we set V (π)iρ(V (π)∗i ), dim πXi=1 THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 127 X = Pu∈Γ Xuλu with Xu ∈ M, and Xu ∈ (γu, αg) = Cδu,Iδg,0. Therefore X is a scalar multiple of δg,e. Setting x = λu, we get τ = 1 if X 6= 0. (cid:3) We denote by π0 the defining representation of Γ on K0, which is not necessarily irreducible. Theorem 13.14. Let the notation be as above. (1) ρ2 is decomposed as (2) Let κ : N ⋊γ Γ ֒→ M ⋊γ Γ be the inclusion map. Then [ρ2] =Mg∈G [ αg] ⊕ n[γπ0 ρ]. [κκ] = [id] ⊕ [γπ0 ρ]. (3) For any π1, π2 ∈ Γ, we have dim(γπ1 ρ, γπ2 ρ) = δπ1,π2. Proof. (1) Since γu acts on {Sg}g∈G trivially, we can easily see Sg ∈ ( αg, ρ2). As in Section 8, we choose an orthonormal basis {Tg(ei)}l Tg(ei)Th(ei)∗. i=1 of K0, and set eh,g = Then {eg,h}g,h∈G is a system of matrix units in Mγ. We choose an isometry R ∈ Mγ whose range projection is e0,0, and set V (π0)i = R∗T0(ei). Let Hπ0 be the linear span of {V (π0)i}l i=1. Then Hπ0 is a Hilbert space in M with full support such that the restriction of γ to Hπ0 is equivalent to π0. Therefore we may assume that γπ0, which is defined up to equivalence in any case, is given by i=1 V (π0)ixV (π0)∗i , x ∈ M x = λu . γπ0(x) =(cid:26) Pdim π λu, lXi=1 We set Rg = eg,0R, which is an isometry in Mγ. We claim Rg ∈ (γπ0 ρ, ρ2). Indeed, for x ∈ M, we have ρ2(x)Rg = ρ2(x)eg,0R0 = Tg(ei)ρ(x)T0(ei)∗R0 lXi=1 eg,0RR∗T0(ei)ρ(x)T0(ei)∗R0 = Rg γπ0 ◦ ρ(x), = lXi=1 and which shows the claim. Since ρ2(λu)Rg = λuRg = Rgλu = Rg γπ0 ◦ ρ(λu), RgR∗g =Xg∈G eg,0e0,0e0,g =Xg∈G eg,g =Xg,i Xg∈G Tg(ei)Tg(ei)∗, we get the statement. 128 MASAKI IZUMI (2) Let ι : N ֒→ M be the inclusion map. Then the proof of Theorem 11.2 shows (3) The statement follows from (1) because (ι, ρι) = (j1 ◦ j2)∗K0. This and the above argument shows the statement. dim(γπ1 ρ, γπ2 ρ) = dim(γπ2⊗π1, ρ2) =Xg∈G dim(γπ2⊗π1, αg) + n dim(γπ0⊗π2⊗π1, ρ) =Xg∈G dim(γπ1, γπ2 αg) = δπ1,π2. (cid:3) Example 13.15. We apply the above theorem to Example 13.10. As Γ, we take the group generated by 0 (cid:18) 1 0 −1 (cid:19) , (cid:18) 0 −1 0 (cid:19) , 1 which is the dihedral group D8. The defining representation π0 is real and irreducible, and the tensor product π0 ⊗ π0 is decomposed into 1-dimensional representations, π0 ⊗ π0 ∼= Mτ∈Hom(D8,T) Since τ ⊗ π0 ∼= π0 for any τ ∈ Hom(D8, T), we get [(γπ0 ρ)2] = [γπ0⊗π0 ρ2] = Xg∈Z3, τ∈Hom(D8,T) τ. [γτ αg] + 3 Xτ∈Hom(D8,T) [γτ⊗π0 ρ] = Xg∈Z3, τ∈Hom(D8,T) [γτ αg] + 12[γπ0 ρ]. This means that the fusion category generated by γπ0 ρ is a near-group category with the group Z2 × Z2 × Z3 and the multiplicity parameter m = 12. 13.3. The case of m = G. In this subsection, we assume that C is a C∗ near-group category with a finite abelian group G and m = n, and we use the same notation as in Section 9. In this case, the group Aut(C) = Out(C) is isomorphic to the set of group automorphisms of G leaving the triple (h·,·i, a, b) invariant. For simplicity, we assume that we have such an automorphism θ of order two, which covers all the known examples (see Example 13.8, Example 13.9). Then γ = γ(θ,I) is an outer automorphism of the factor M of order two commuting with ρ, and it satisfies the relation γ ◦ αg = αθ(g) ◦ γ. Thus Theorem 14.1 implies that α and γ give an outer action of G ⋊ Z2 on M. We denote by λ the implementing unitary in the crossed product M ⋊γ Z2. As before ρ extends to ρ ∈ End(M ⋊γ Z2) by ρ(λ) = λ. Let Gθ be the set of automor- phisms of G fixed by γ. Then αg with g ∈ Gθ also extends to αg ∈ Aut(M ⋊γ Z2) by αg(λ) = λ. We denote by γ the dual action of γ, which is identified with the generator of the dual action too. As in the previous subsection, we can show that ρ is irreducible with [γ ρ] = [ργ] 6= [ρ], and that α and γ give an outer action of Gθ × Z2. Lemma 13.16. (ρ2, ρ2) = (ρ2, ρ2)γ. THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 129 Proof. Assume X + Y λ ∈ (ρ2, ρ2) with X, Y ∈ M. Then X ∈ (ρ2, ρ2)γ and (cid:3) Recall that we have Y ∈ (ρ2γ, ρ2) ∼= (γ, ρ4) = {0}. (ρ2, ρ2) =Mg∈G CSgS∗g ⊕Mg∈Γ CSgS∗g ⊕ B(K). We choose a subset Λ ⊂ G\ Gθ satisfying Λ∩ θ(Λ) = ∅ and Λ∪ θ(Λ) = G\ Gθ. Then the above lemma shows (ρ2, ρ2) = Mg∈Gθ C(SgS∗g + Sθ(g)S∗θ(g)) ⊕ B(K)γ. Thus for g ∈ Λ, the projection SgS∗g + Sθ(g)S∗θ(g) is minimal in (ρ2, ρ2), and it cor- responds to an irreducible component of ρ2. We choose an isometry Rg ∈ Mγ whose range projection is SgS∗g + Sθ(g)S∗θ(g), and defined πg ∈ End(M ⋊γ Z2) by πg(x) = R∗g ρ2(x)Rg. Then πg(x) =(cid:26) R∗g(Sgαg(x)S∗g + Sθ(g)αθ(g)(x)S∗θ(g))Rg, λ, For g ∈ Gθ, we have Tg ∈ (ρ, ρ2). For g ∈ G \ Gθ, we have x ∈ M x = λ . 1 √2 1 √2 (Tg + Tθ(g)) ∈ (ρ, ρ2), (Tg − Tθ(g)) ∈ (γ ρ, ρ2). Thus it is straightforward to show the following. Theorem 13.17. Let the notation be as above. (1) ρ2 is decomposed as [ρ2] = Xg∈Gθ [ αg] +Xg∈Λ [πg] + G + Gθ 2 [ρ] + G − Gθ 2 [γ ρ]. (2) Let κ : N ⋊γ Z2 ֒→ M ⋊γ Z2 be the inclusion map. Then [κκ] = [id] + [ρ]. About the fusion rules, we have the following. Theorem 13.18. Let the notation be as above. the representation category of G ⋊θ Z2. (1) The tensor category generated by {γ} ∪ { αg}g∈Gθ ∪ {πg}g∈Λ is equivalent to (2) For g ∈ Gθ, we have [ αg][ρ] = [ρ][ αg] = [ρ]. (3) For g ∈ Λ, we have [πg][ρ] = [ρ][πg] = [ρ] + [γ ρ]. 130 MASAKI IZUMI Proof. (1) Let MG be the fixed point algebra MG = {x ∈ M; αg(x) = x, ∀g ∈ G}, and let ι1 : MG ֒→ M and ι2 : M ֒→ M ⋊γ Z2 be the inclusion maps. Then it is easy to show that MG ⊂ M ⋊γ Z2 is a crossed product inclusion by a G ⋊θ Z2-action, and the tensor category generated by ι2ι1(ι2ι1) is equivalent to the representation category of G ⋊θ Z2. By construction, we have [ αg][ι2] = [ι2][αg] for g ∈ Gθ, and [πg][ι2] = [ι2][αg] + [ι2][αθ(g)], for g ∈ Λ. Thus we get the following for g ∈ Λ by the Frobenius reciprocity: [πg] = [ι2αgι2] = [ι2αθ(g)ι2]. Now we get [ι2ι1(ι2ι1)] =Xg∈G = Xg∈Gθ [ αg] + Xg∈Gθ [ι2αgι2] = Xg∈Gθ [ αg γ] +Xg∈Λ 2[πg], [ αg][ι2ι2] +Xg∈Λ 2[πg] which shows the statement. (2) We have αg ◦ ρ = ρ, and taking its adjoint, we get the statement. (3) By construction, we have 1 √2 1 √2 R∗g(Sg − Sθ(g)) ∈ (γ ρ, πg ρ), R∗g(Sg + Sθ(g)) ∈ (ρ, πg ρ), which shows [πg ρ] = [ρ] + [γ ρ]. Taking its adjoint, we get the statement. (cid:3) Remark 13.19. The endomorphism ι2αgι2 corresponds to the induced representation Ind G⋊θZ2 G g, where g ∈ G = G is regarded as a one-dimensional representation of G. From the above two theorems, it is easy to determine the principal graphs of the subfactor N ⋊γ Z2 ⊂ M ⋊γ Z2, in particular, for Example 13.8 and Example 13.9. We leave it for the reader. 14. Appendix Let On+m be the Cuntz algebra with canonical generators {Si}i∈J, where J = {1, 2, . . . , n + m}. Let J1 = {1, 2, . . . , m} and J2 = {m + 1, m + 2, . . . , n + m}. We consider a weighted gauge action γ : T → Aut(On+m) defined by γt(Si) =(cid:26) eitSi, e2itSi, i ∈ J1 i ∈ J2 . THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 131 2 Then γ has a unique KMS-state ϕ with the inverse temperature log d, where d = m+√m2+4n (see [14]). The weak closure M of On+m in the GNS representation of ϕ is the Powers factor of type III1/d. With identification {Sg} = {Si}i∈I2 and K = span{Si}i∈I1, the Cuntz algebra model (α, ρ) of a C∗ near-group category C can extend to M, which gives a realization of C in End(M) (see [28]). The purpose of this appendix is to show that the symmetry group Autρ(On+m) is also realized in Aut(M) ⊂ End(M), and is even faithfully realized in Out(M). Let β ∈ Aut(On+m) be an automorphism globally preserving span{Si}i∈I1 and span{Si}i∈I2. Since β commutes with γt, it preserves ϕ and extends to M, which is still denoted by β. Theorem 14.1. Let β ∈ Aut(M) be as above. Then β is inner if and only if β = id. Proof. We use the same symbol ϕ and γ for their normal extension to M. Then we have σϕ t = γ−t log d, and the centralized Mϕ is the hyperfinite II1 factor with a unique trace τ = ϕMϕ. We set For a k-tuple ξ = (ξ1, ξ2, . . . , ξk) ∈ J k, we set Sξ = Sξ1Sξ2 · · · Sξk , r(ξ) = ξk, and 2, w(i) =(cid:26) 1, kXi=1 w(ξ) = i ∈ J1 i ∈ J2 . w(ξi). Then we have γt(Sξ) = eiw(ξ)tSξ. We introduce finite dimensional ∗-subalgebras of Mϕ by Ak,1 = span{SξS∗η ; w(ξ) = w(η) = k}, Ak,2 = span{SξS∗η; w(ξ) = w(η) = k + 1, r(ξ), r(η) ∈ I2}, Ak = Ak,1 ⊕ Ak,2. Then {Ak}∞k=1 is an increasing sequence of finite dimensional ∗-subalgebras of Mϕ whose union is dense in Mϕ in the strong ∗-topology (see [28]). We denote by Ek the τ -preserving conditional expectation from Mϕ onto Ak. Assume that β is inner, that is, there exists a unitary u ∈ M satisfying ux = β(x)u for any x ∈ M. Since β commutes with γt, its Fourier coefficient uk = e−iktγt(u)dt 1 2πZ 2π 0 satisfies ukx = β(x)uk too. This implies that un is a multiple of a unitary. On the other hand, the KMS condition implies ϕ(uku∗k) = 1 dk ϕ(uku∗k), which implies that uk = 0 for any k 6= 0. Thus u belongs to Mϕ. Let vk = Ek(u). Then we have vkx = β(x)vk for any x ∈ Ak, and the sequence {vk}∞k=1 of contractions converges to u in the L2-topology with respect to τ . We may 132 MASAKI IZUMI assume β(Si) = ωiSi for any i ∈ J by choosing appropriate orthonormal bases of span{Si}i∈J1 and span{Si}i∈J2. For ξ = (ξ1, ξ2, . . . , ξk) ∈ J k, we set Then there exists two numbers ck,1, ck,2 ∈ C satisfying ωξ = ωξ1ωξ2 · · · ωξk. On the other hand, we have the martingale condition Ek(vk+1) = vk. Simple compu- tation shows ωξωiSξSiS∗i S∗ξ . 1 dSξSxi∗, SξSiS∗i S∗ξ , 1 d2 SξS∗ξ , ωξSξS∗ξ + ck,2 Xw(ξ)=k−1, i∈J2 vk = ck,1 Xw(ξ)=k Ek(SξSiS∗i S∗ξ ) = ωξEk(SξS∗ξ ) + ck+1,2 Xw(ξ)=k, i∈J2 ωξωiEk(SξSiS∗i S∗ξ ) + ck+1,1 Xw(ξ)=k−1, i∈J2 w(ξ) = k, i ∈ I1 w(ξ) = k − 1, i ∈ I2 w(ξ) = k, i ∈ I2 ωξωiEk(SξSiS∗i S∗ξ ), ωξωiSξS∗ξ ωξωiSξS∗ξ + ck+1,1 Xw(ξ)=k−1, i∈J2 ωξωiSξS∗ξ , ωξωiSξSiS∗i S∗ξ vk = Ek(vk+1) + ck+1,2 = ck+1,1 Xw(ξ)=k+1 = ck+1,1 Xw(ξ)=k, i∈J1 d2 Xw(ξ)=k, i∈J2 d Xw(ξ)=k, i∈J1 d2 Xw(ξ)=k, i∈J2 ck+1,2 ck+1,1 = + ωξωiEk(SξSiS∗i S∗ξ ) Thus and so where (cid:18) ck,1 ck,2 (cid:19) =(cid:18) mz d 1 1 nw d2 0 (cid:19)(cid:18) ck+1,1 ck+1,2 (cid:19) , mXi∈J1 nXi∈J2 ωi. ωi, 1 z = w = Let Γ be the above matrix. Then the two eigenvalues of Γ are Unless z = 1 and w = z2, we have λ± < 1, and λ± = . 2d mz ± √m2z2 + 4nw ck+l,2 (cid:19) →(cid:18) 0 0 (cid:19) (l → ∞), ck,2 (cid:19) = Γl(cid:18) ck+l,1 (cid:18) ck,1 THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 133 for we have ck+l,1 ≤ 1 and ck+l,2 ≤ 1, which is contradiction. Assume z = 1 and w = z2 now. Then we have ωi = z for any i ∈ J1 and ωi = z2 for any i ∈ J2, and the restriction of β to Mϕ is trivial. This implies that u ∈ Mϕ ∩ M′ϕ = C, and β = id. (cid:3) References [1] Afzaly N.; Morrison S.; Penneys D. The classification of subfactors with index at most 5 1 2 . preprint, arXiv:1509.00038. [2] Bigelow, S.; Peters, E.; Morrison, S.; Snyder, N. Constructing the extended Haagerup planar algebra. Acta Math. 209 (2012), no. 1, 29 -- 82. [3] Bischoff, M.; Kawahigashi, Y.; Longo, R.; Rehren, K.-H. Tensor categories and endomorphisms of von Neumann algebras (with applications to Quantum Field Theory. Springer Briefs in Math- ematical Physics Vol. 3, 2015. [4] Brugui`eres, A. Cat´egories pr´emodulaires, modularisations et invariants des vari´et´es de dimen- sion 3. Math. Ann. 316 (2000), 215 -- 236. [5] Connes, A. Periodic automorphisms of the hyperfinite factor of type II1. Acta Sci. Math. (Szeged) 39 (1977), 39 -- 66. [6] Cuntz, J. Simple C∗-algebras generated by isometries. Comm. Math. Phys. 57 (1977), no. 2, 173 -- 185. [7] Dijkgraaf, R.; Pasquier, V.; Roche, P. Quasi Hopf algebras, group cohomology and orbifold models. Recent advances in field theory (Annecy-le-Vieux, 1990). Nuclear Phys. B Proc. Suppl. 18B (1990), 60 -- 72 (1991). [8] Dijkgraaf, R.; Witten, E. Topological gauge theories and group cohomology. Comm. Math. Phys. 129 (1990), 393 -- 429. [9] Doplicher, S.; Haag, R.; Roberts, J. E. Local observables and particle statistics. I. Comm. Math. Phys. 23 (1971), 199 -- 230. [10] Etingof, P.; Gelaki, S.; Ostrik, V. Classification of fusion categories of dimension pq. Int. Math. Res. Not. 2004, no. 57, 3041 -- 3056. [11] Etingof, P.; Gelaki, S.; Nikshych, D.; Ostrik, V. Tensor categories. Mathematical Surveys and Monographs, 205. American Mathematical Society, Providence, RI, 2015. [12] Etingof, P.; Nikshych, D.; Ostrik, V. On fusion categories. Ann. of Math. 162 (2005), 581 -- 642. [13] Etingof, P.; Nikshych, D.; Ostrik, V. Fusion categories and homotopy theory. With an appendix by Ehud Meir. Quantum Topol. 1 (2010), 209 -- 273. [14] Evans, D. E. On On. Publ. Res. Inst. Math. Sci. 16 (1980), 915 -- 927. [15] Evans, D. E.; Gannon, T. The exoticness and realisability of twisted Haagerup-Izumi modular data. Comm. Math. Phys. 307 (2011), 463 -- 512. [16] Evans, D. E.; Gannon, T. Near-group fusion categories and their doubles. Adv. Math. 255 (2014), 586 -- 640. [17] Evans, D. E.; Kawahigashi, Y. Orbifold subfactors from Hecke algebras. Comm. Math. Phys. 165 (1994), 445 -- 484. [18] Evans, D. E.; Kawahigashi, Y. Quantum symmetries on operator algebras. Oxford Mathematical Monographs. Oxford Science Publications. T he Clarendon Press, Oxford University Press, New York, 1998. [19] Evans D. E.; Pugh, M. Braided subfactors, spectral measures, planar algebras, and Calabi-Yau algebras associated to SU(3) modular invariants. Progress in operator algebras, noncommuta- tive geometry, and their applications, 17 -- 60, Theta Ser. Adv. Math., 15, Theta, Bucharest, 2012. [20] Grossman P.; Izumi, M.; Snyder, N. The Asaeda-Haagerup fusion categories. to appear in J. Reine Angew. Math. arXiv:1501.07324. [21] Grossman, P.; Snyder, N. Quantum subgroups of the Haagerup fusion categories. Comm. Math. Phys. 311 (2012), 617 -- 643. 134 MASAKI IZUMI [22] Grossman, P.; Jordan, D.; Snyder, N. Cyclic extensions of fusion categories via the Brauer- Picard groupoid. Quantum Topol. 6 (2015), no. 2, 313 -- 331. [23] Hagge, T.; Hong, Seung-Moon. Some non-braided fusion categories of rank three. Commun. Contemp. Math. 11 (2009), no. 4, 615 -- 637. [24] Haagerup, U. Connes' bicentralizer problem and uniqueness of the injective factor of type III1. Acta Math. 158 (1987), no. 1-2, 95 -- 148. [25] Hayashi, T.; Yamagami, S. Amenable tensor categories and their realizations as AFD bimodules. J. Funct. Anal. 172 (2000), 19 -- 75. [26] Huppert, B.; Blackburn, N. Finite groups. III. Grundlehren der Mathematischen Wis- senschaften, 243. Springer-Verlag, Berlin-New York, 1982. [27] Isaacs, I. M. Character theory of finite groups. Dover Publications, Inc., New York, 1994. [28] Izumi, M. Subalgebras of infinite C∗-algebras with finite Watatani indices. I. Cuntz algebras. Comm. Math. Phys. 155 (1993), 157 -- 182. [29] Izumi, M. Goldman's type theorems in index theory. Operator algebras and quantum field theory (Rome, 1996), 249 -- 269, Int. Press, Cambridge, MA, 1997. [30] Izumi, M. Subalgebras of infinite C∗-algebras with finite Watatani indices. II. Cuntz-Krieger algebras. Duke Math. J. 91 (1998), no. 3, 409 -- 461. [31] Izumi, M. The structure of sectors associated with Longo-Rehren inclusions. I. General theory. Comm. Math. Phys. 213 (2000), 127 -- 179. [32] Izumi, M. The structure of sectors associated with Longo-Rehren inclusions. II. Examples. Rev. Math. Phys. 13 (2001), 603 -- 674. [33] Izumi, M.; Kosaki, H. Kac algebras arising from composition of subfactors: general theory and classification. Mem. Amer. Math. Soc. 158 (2002), no.750. [34] Jones, V. F. R. Actions of finite groups on the hyperfinite type II1 factor. Mem. Amer. Math. Soc. 28 (1980), no. 237. [35] Jones, V. F. R. Index for subfactors. Invent. Math. 72 (1983), no. 1, 1 -- 25. [36] Jones, V. F. R.; Morrison, S.; Snyder, N. The classification of subfactors of index at most 5. Bull. Amer. Math. Soc. (N.S.) 51 (2014), no. 2, 277 -- 327. [37] Katayama, Y.; Takesaki, M. Outer actions of a discrete amenable group on approximately finite [38] Kawahigashi, Y. On flatness of Ocneanu's connections on the Dynkin diagrams and classifica- dimensional factors. II. The IIIλ-case, λ 6= 0. Math. Scand. 100 (2007), no. 1, 75 -- 129. tion of subfactors. J. Funct. Anal. 127 (1995), 63 -- 107. [39] Larson, H. Pseudo-unitary non-self-dual fusion categories of rank 4. J. Algebra 415 (2014), 184 -- 213. [40] Liu, Zhengwei. Singly generated planar algebras of small dimension, Part IV. preprint, arXiv:1507.06030. [41] Liu, Zhengwei; Morrison, S.; Penneys, D. 1-supertransitive subfactors with index at most 6 1 5 . Commun. Math. Phys. 334, (2015), 889 -- 922. [42] Longo, R. Index of subfactors and statistics of quantum fields. II. Correspondences, braid group statistics and Jones polynomial. Comm. Math. Phys. 130 (1990), 285 -- 309. [43] Masuda, Toshihiko, An analogue of Connes-Haagerup approach for classification of subfactors of type III1. J. Math. Soc. Japan 57 (2005), 959 -- 1001. [44] Ng, Siu-Hung; Schauenburg, P. Higher Frobenius-Schur indicators for pivotal categories. Hopf algebras and generalizations, 63 -- 90, Contemp. Math., 441, Amer. Math. Soc., Providence, RI, 2007. [45] Nikshych, D.; Ostrik, V. On the structure of near-group categories. in preparation. [46] Ostrik, V. Fusion categories of rank 2. Math. Res. Lett. 10 (2003), 177 -- 183. [47] Ostrik, V. Pivotal fusion categories of rank 3. to appear in Moscow Math J. arXiv:1309.4822. [48] Popa, S. Classification of subfactors and their endomorphisms. CBMS Regional Conference Series in Mathematics, 86. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 1995. THE CLASSIFICATION OF NEAR-GROUP CATEGORIES 135 [49] Robinson, D. J. S. A course in the theory of groups. Graduate Texts in Mathematics, 80. Springer-Verlag, New York, 1993. [50] Siehler, J. Near-group categories. Algebr. Geom. Topol. 3 (2003), 719 -- 775. [51] Sutherland, C. E. Cohomology and extensions of von Neumann algebras. I, II. Publ. Res. Inst. Math. Sci. 16 (1980), no. 1, 105 -- 133, 135 -- 174. [52] Tambara, D.; and Yamagami, S. Tensor categories with fusion rules of self-duality for finite abelian groups. J. Algebra 209 (1998), 692 -- 707. [53] Yamagami, S. Group symmetry in tensor categories and duality for orbifolds. J. Pure Appl. Algebra 167 (2002), 83 -- 128. Department of Mathematics, Graduate School of Science, Kyoto University, Sakyo-ku, Kyoto 606-8502, Japan E-mail address: [email protected]
1811.06929	2	1811	2019-09-08T12:34:00	Classification of regular subalgebras of the hyperfinite II_1 factor	[ "math.OA", "math.DS" ]	We prove that the regular von Neumann subalgebras $B$ of the hyperfinite II_1 factor $R$ satisfying the condition $B'\cap R=Z(B)$ are completely classified (up to conjugacy by an automorphism of $R$) by the associated discrete measured groupoid $G$. We obtain a similar classification result for triple inclusions $A\subset B \subset R$, where $A$ is a Cartan subalgebra in $R$ and the intermediate von Neumann algebra $B$ is regular in $R$. A key step in proving these results is to show the vanishing cohomology for the associated cocycle actions of $G$ on $B$. We in fact prove two very general vanishing cohomology results for free cocycle actions of amenable discrete measured groupoids on arbitrary tracial von Neumann algebras $B$, resp. Cartan inclusions $A\subset B$. Our work provides a unified approach and generalizations to many known vanishing cohomology and classification results [CFW81], [O85], [ST84], [BG84], [FSZ88], [P18], etc.	math.OA	math	Classification of regular subalgebras of the hyperfinite II1 factor by Sorin Popa1, Dimitri Shlyakhtenko2 and Stefaan Vaes3 Dedicated to Alain Connes. Abstract We prove that the regular von Neumann subalgebras B of the hyperfinite II1 factor R satisfying the condition B ′ ∩ R = Z(B) are completely classified (up to conjugacy by an automorphism of R) by the associated discrete measured groupoid G = GB⊂R. We obtain a similar classification result for triple inclusions A ⊂ B ⊂ R, where A is a Cartan subal- gebra in R and the intermediate von Neumann algebra B is regular in R. A key step in proving these results is to show the vanishing cohomology for the associated cocycle actions (αB⊂R, uB⊂R) of G on B. We in fact prove two very general vanishing cohomology results for free cocycle actions (α, u) of amenable discrete measured groupoids G on arbitrary tracial von Neumann algebras B, resp. Cartan inclusions A ⊂ B. Our work provides a unified ap- proach and generalizations to many known vanishing cohomology and classification results [CFW81], [O85], [ST84], [BG84], [FSZ88], [P18], etc. 1 Introduction Connes's fundamental theorem ([C75]), establishing the equivalence between amenability and finite dimensional approximation (AFD) for finite von Neumann algebras, has two remark- able consequences. On the one hand, it gives a complete classification of the von Neu- mann subalgebras of the hyperfinite II1 factor R, showing that they are all AFD of the form B = ⊕n≥1(An ⊗ Mn×n(C)) ⊕ (A0 ⊗ R), where An, 0 ≤ n < ∞, are separable and abelian, pos- sibly equal to 0. On the other hand, it shows that any crossed product von Neumann algebra B ⋊α,u G, corresponding to a free cocycle action (α, u) of an amenable discrete groupoid G on a diffuse tracial AFD algebra B, which is ergodic on the center Z(B) of B, is isomorphic to R. During the late 1970s and early 1980s, much effort has been put into clarifying whether such a crossed product decomposition of R is in fact uniquely determined by the nature of B and G alone (thus not depending on (α, u)). Alternatively, the question is whether any two regular copies B ⊂ R of a specific AFD algebra B, with B′ ∩ R = Z(B) and with the same associated amenable discrete measured groupoid GB⊂R = G, are conjugate by an automorphism of R. The assumptions imply that B is necessarily homogeneous of the form B = L∞(X, µ) ⊗ N , where either N ≃ Mn×n(C), for some n ≥ 1, or N ≃ R. The normalizer NR(B) := {u ∈ U (R) uBu∗ = B} defines an amenable discrete measured groupoid G = GB⊂R together with a free cocycle action (α, u) of G on B (see Section 2 for a rigorous definition and detailed discussion of 1Mathematics Department, UCLA, Los Angeles, CA 90095-1555 (United States), [email protected] Supported in part by NSF Grant DMS-1700344 2Mathematics Department, UCLA, Los Angeles, CA 90095-1555 (United States), [email protected] Supported in part by NSF Grant DMS-1500035 3KU Leuven, Department of Mathematics, Leuven (Belgium), [email protected] Supported in part by European Research Council Consolidator Grant 614195, and by long term structural funding -- Methusalem grant of the Flemish Government. Part of this research was completed while the authors were visiting the Institute for Pure and Applied Mathematics (IPAM), which is supported by the National Science Foundation. 1 discrete measured groupoids and their free cocycle actions on tracial von Neumann algebras). The groupoid G accounts for an amenable ergodic countable equivalence relation "along" the space G(0) = X of units of G, with amenable countable isotropy groups Γx at each x ∈ X acting outerly on Bx ≃ N . When B is abelian, then B ≃ L∞(X, µ) and one calls it a Cartan subalgebra of R. In this case, G is just a countable equivalence relation R on X, with α intrinsic to R. If B is a factor, then B ≃ R and GB⊂R coincides with the group Γ = NR(B)/U (B), which is automatically countable amenable, with (α, u) the free cocycle action of Γ on B implemented by NR(B). The uniqueness problem has been solved in two important cases: when B is abelian, i.e., for Cartan subalgebras of R ([CFW81]); and when B is a factor, i.e., when B ≃ R ([O85]). Thus, it was shown in [CFW81] that any two Cartan subalgebras of R are conjugate by an automorphism of R. Equivalently, there is a unique amenable ergodic type II1 equivalence relation, which has vanishing 2-cohomology. This also implies that for any n ≥ 2, any two regular subalgebras of type In of R are conjugate by an automorphism of R. In turn, the result in [O85] shows hat any two free cocycle actions of the same countable amenable group Γ on B ≃ R are cocycle conjugate. Equivalently, given any countable amenable group Γ, there exists a unique (up to conjugacy by an automorphism of R) irreducible regular subfactor B ⊂ R with NR(B)/U (B) ≃ Γ. An important step towards solving the remaining case when B is an arbitrary II1 AFD algebra, B ≃ L∞(X, µ) ⊗ R, has been achieved in [ST84], where it was shown that any two free genuine actions α of the same amenable discrete measured groupoid G on B that are ergodic on Z(B) = L∞(X) = L∞(G(0)), are cocycle conjugate. However, this result does not settle the uniqueness of all cocycle actions (α, u) of G on arbitrary such B. We solve this last step here, by proving that in fact any such cocycle u untwists. When combined with [ST84], this shows the uniqueness of all cocycle actions of G on B. Equivalently, all regular embeddings B ⊂ R, with B′ ∩ R = Z(B), of a given II1 AFD algebra B that have the same groupoid GB⊂R ≃ G are conjugate by an automorphism of R. We in fact prove a very general vanishing cohomology result, showing that any free cocycle action (α, u) of any amenable discrete measured groupoid G on any (not necessarily AFD) tracial von Neumann algebra (B, τ ), untwists. We also prove a relative version of this result, by showing that if in addition (α, u) is assumed to normalize a Cartan subalgebra A of B, on which it acts freely, then the vanishing cohomology can be realized within the normalizer of A in B, i.e., as a coboundary of a map from G into the normalizer of A in B. More precisely, in Theorems 3.1 and 5.2, we show: Theorem A. Let G be a discrete measured groupoid with X = G(0) and (Bx)x∈X a measurable field of II1 factors with separable predual. Assume that G is amenable and that (α, u) is a free cocycle action of G on (Bx)x∈X. 1. The cocycle u is a co-boundary: there exists a measurable field of unitaries G ∋ g 7→ wg ∈ Bt(g) such that u(g, h) = αg(w∗ h) w∗ g wgh for all (g, h) ∈ G(2). 2. If (α, u) globally preserves a field of Cartan subalgebras (Ax ⊂ Bx)x∈X and induces an outer action on the associated equivalence relations, then wg can be chosen in NBt(g)(At(g)). We also show in Proposition 5.1 that if A ⊂ B ⊂ M is a triple inclusion of von Neumann algebras, with M a II1 factor and A Cartan in M , then B is regular in M iff A ⊂ B is regular in M (i.e. the normalizer of A ⊂ B in M , NM (A ⊂ B) := {u ∈ U (M ) uAu∗ = A, uBu∗ = B}, generates M ) and iff the corresponding inclusion of equivalence relations RA⊂B ⊂ RA⊂M is 2 strongly normal in the sense of [FSZ88]. When combined with the second part of Theorem A, this shows that if in addition M is assumed amenable relative to B, then such a triple A ⊂ B ⊂ M can be identified with A ⊂ B ⊂ B ⋊α G, where G = GB⊂M and α is a genuine action of G on B that leaves A invariant. Also the equivalence relation RA⊂M can then be written as a genuine semidirect product of RA⊂B and an action of G. Initiated in [FSZ88], the study of normal subequivalence relations saw a revival of interest in recent years, and this structural result about the co-amenable case may be relevant in this direction. As we mentioned before, while [CFW81] shows the uniqueness up to conjugacy by automor- phisms of R of the regular von Neumann subalgebras B ⊂ R of type In satisfying B′∩R = Z(B), the first part of Theorem A combined with [ST84] shows the uniqueness up to conjugacy by automorphisms of R of regular von Neumann subalgebras B ⊂ R of type II1 that satisfy B′ ∩ R = Z(B) and have the same groupoid. Since any groupoid G has a "model action" on B that normalizes a Cartan subalgebra A of B on which it acts freely, this uniqueness result also implies that any regular subalgebra B ⊂ R contains a Cartan subalgebra of R. We use the second part of Theorem A to also prove that any two Cartan subalgebras A1, A2 of R that are contained in B are conjugate by an automorphism of R that leaves B globally invariant. Altogether, in Theorem 3.2, Corollary 3.4 and Theorem 5.3, we obtain: Theorem B. Let R be the hyperfinite II1 factor. 1. Two regular von Neumann subalgebras B ⊂ R satisfying B′ ∩ R = Z(B) are conjugate by an automorphism of R if and only if they are of the same type and have isomorphic associated discrete measured groupoids GB⊂R. 2. Any such B contains a Cartan subalgebra of R and if A1, A2 ⊂ B are Cartan in R, there exists an automorphism θ of R satisfying θ(B) = B and θ(A1) = A2. Our proofs of Theorems A and B make use of the known 2-cocycle vanishing theorems for amenable groups on factors, resp. equivalence relations (see [O85, P18, FSZ88]), as well as the uniqueness, up to cocycle conjugacy, of free actions of amenable groups on the hyperfinite II1 factor, resp. equivalence relation (see [O85, BG84]). We apply these results to the isotropy groups Γx = {g ∈ G s(g) = t(g) = x} of an amenable groupoid G. In order to extend to the entire groupoid G, we have to make equivariant choices of 2-cocycle vanishing, resp. cocycle conjugacy, for the Γx, where the equivariance is w.r.t. to the isomorphisms Γx → Γy given by conjugation with an element g ∈ G with s(g) = x and t(g) = y. The proof of this later part depends on two key points. The first one is the technical Theorem 3.5, saying that such an equivariant choice exists, provided that the 2-cocycle vanishing, resp. cocycle conjugacy for Γx, can be done in an "approximately unique way". More precisely, assume that (α, u) is an outer cocycle action of the countable amenable group Γ on a II1 factor B. By [P18], we know that u is a co-boundary: we can choose unitary elements (wg)g∈Γ in B such that u(g, h) = αg(w∗ g wgh. Then, βg = Ad wg ◦ αg defines an outer action of Γ on B. Any other choice of w such that u = ∂w is of the form vgwg, where (vg)g∈Γ is a 1-cocycle for the action β, meaning that vgh = vg βg(vh). By the above "approximate uniqueness" of w, we mean that every such 1-cocycle v is approximately a co-boundary. h) w∗ The second key point is then to prove such approximate vanishing of 1-cocycles for arbitrary amenable groups (see Theorem 4.1). We state this result below because of its independent interest. Theorem C. Let Γ be a countable group. The following properties are equivalent. 3 (i) Γ is amenable. (ii) No free trace preserving action Γ yσ (B, τ ) on a tracial diffuse von Neumann algebra is strongly ergodic. (iii) For every free action Γ yσ N on a II1 factor N , the fixed point algebra of the ultrapower action σω on N ω is an irreducible subfactor of N ω. (iv) Given any free action Γ yσ N on a II1 factor N , any 1-cocycle w = (wg)g∈Γ for σ for every finite subset F ⊂ Γ and every ε > 0, there is approximately a co-boundary: exists u ∈ U (N ) such that kuσg(u∗)− wgk2 ≤ ε for all g ∈ F . Equivalently, there exists a g (U ∗) = wg, ∀g ∈ Γ, i.e., w is a co-boundary as a 1-cocycle unitary U ∈ N ω such that U σω for σω. The proof of the above result is quite subtle, with the statement itself being somewhat sur- prising, given its generality. Note in this respect that 1-cocycles for a unitary representation of an amenable group need not be approximately co-boundary (see [S03]). We similarly prove an approximate vanishing of 1-cocycles with values in the normalizer of a Cartan subalgebra, which we need in the proof of the second part of Theorem A (see Theorem 4.7). In [P18], the class of countable groups satisfying 2-cohomology vanishing for cocycle actions on II1 factors is denoted as VC. In [P18, Remarks 4.5], it is speculated that VC might coincide with the class of treeable groups. We elaborate on this in Section 6 and this leads us to a new notion of treeability for countable pmp equivalence relations, which we call weak treeability. It is not known whether every treeable group is strongly treeable. We prove that this question is closely related to the question whether every weakly treeable equivalence relation is treeable. Similarly, we introduce a fixed price problem for equivalence relations and relate it to the fixed price problem for groups. In the final Section 7, we construct numerous families of amenable discrete measured groupoids. We show that even though there is a unique ergodic amenable equivalence relation of type II1, the classification problem of amenable discrete measured groupoids is strictly harder than the classification of amenable groups. Even restricting to groupoids G for which a.e. isotropy group is Z2, one may "encode" ergodic transformations of the interval [0, 1] up to conjugacy into such amenable groupoids up to isomorphism (see Corollary 7.2). 2 Discrete measured groupoids, cocycle actions and crossed products In this section, we summarize some of the basic terminology and well known results on discrete measured groupoids and their actions. Recall that a discrete measured groupoid G is a groupoid with the following extra structure. This concept goes back to [M63]. • G is a standard Borel space and the units G(0) ⊂ G form a Borel subset. • The source and target maps s, t : G → G(0) are Borel and countable-to-one. • Defining G(2) = {(g, h) ∈ G×G s(g) = t(h)}, the multiplication map G(2) → G : (g, h) 7→ gh is Borel. The inverse map G → G : g 7→ g−1 is Borel. 4 • G(0) is equipped with a probability measure µ such that, denoting by µs and µt the σ-finite measures on G given by integrating the counting measure over s, t : G → G(0), we have that µs ∼ µt. We say that G is probability measure preserving (pmp) if µs = µt. To every discrete measured groupoid G is associated the countable nonsingular equivalence relation R on (G(0), µ) given by R = {(t(g), s(g)) g ∈ G} . Note that G is pmp if and only if R is a pmp equivalence relation. We say that G is ergodic if the associated equivalence relation R is ergodic. Remark 2.1. In this paper, we only work in the measurable context, discarding sets of measure zero whenever useful. By von Neumann's measurable selection theorem (see e.g. [K95, Theorem 18.1]), we therefore never have problems choosing measurable sections. Also, all isomorphisms between measure spaces, fields of von Neumann algebras, measured groupoids, are defined up to sets of measure zero. Whenever we write for all, this should be interpreted as for almost every, whenever appropriate. Using [S85, Sections 1 and 2], we may consider measurable fields of any kind of 'separable struc- tures': von Neumann algebras with separable predual, standard probability spaces, countable groups, Polish spaces, Polish groups, etc. In particular, by [S85, Theorem 2.5], all natural definitions of such measurable fields are equivalent. Fix a discrete measured groupoid G and write X = G(0). Define G(2) as above and similarly define G(3). Definition 2.2. A cocycle action (α, u) of G on a measurable field (Bx)x∈X of von Neumann algebras with separable predual is given by • a measurable field of ∗-isomorphisms G ∋ g 7→ αg : Bs(g) → Bt(g), • a measurable field of unitaries G(2) ∋ (g, h) 7→ u(g, h) ∈ U (Bt(g)), satisfying αg ◦ αh = Ad(u(g, h)) ◦ αgh αg(u(h, k)) u(g, hk)) = u(g, h) u(gh, k) for all (g, h) ∈ G(2), for all (g, h, k) ∈ G(3), αg = id when g ∈ G(0), u(g, h) = 1 when g ∈ G(0) or h ∈ G(0). An automorphism α of a von Neumann algebra B is said to be free (or properly outer) if the only element v ∈ B satisfying vx = α(x)v for all x ∈ B is the zero element v = 0. When B is a factor, an automorphism α is free if and only if it is outer, meaning that there is no unitary element v ∈ U (B) such that α = Ad v. Denote by Γx = {g ∈ G s(g) = t(g) = x} the isotropy groups of G. The cocycle action (α, u) is said to be free if for almost every x ∈ X and all g ∈ Γx with g 6= e, the ∗-automorphism αg : Bx → Bx is free. The cocycle actions (α, u) of G on (Bx)x∈X and (β, Ψ) of G on (Dx)x∈X are said to be cocycle conjugate if there exists a measurable field of ∗-isomorphisms X ∋ x 7→ θx : Bx → Dx and a 5 measurable field of unitaries G ∋ g 7→ wg ∈ U (Dt(g)) satisfying θt(g) ◦ αg ◦ θ−1 s(g) = Ad wg ◦ βg θt(g)(u(g, h)) = wg βg(wh) Ψ(g, h) w∗ gh for all g ∈ G, for all (g, h) ∈ G(1). The cocycle crossed product construction associates to every cocycle action (α, u) a regular inclusion of von Neumann algebras B ⊂ M together with a faithful normal conditional expec- tation E : M → B. The cocycle crossed product M can be easily defined by first considering the full pseudogroup [[G]] of G consisting of all Borel sets U ⊂ G with the property that the restrictions to U of the source map s and the target map t are injective, and identifying U ,U ′ if they differ by a set of measure zero. The composition of U ,V ∈ [[G]] is defined as U · V = {gh g ∈ U , h ∈ V, s(g) = t(h)} . For every U ∈ [[G]], define the Borel map ϕU : s(U ) → t(U ) : ϕU (s(g)) = t(g) for all g ∈ U . Writing B = Z ⊕ X Bx dµ(x) , the cocycle action (α, u) of G on (Bx)x∈X can be reinterpreted as follows in terms of [[G]]. To every U ∈ [[G]], there corresponds a ∗-isomorphism αU : B1s(U ) → B1t(U ) given by (αU (b))(t(g)) = αg(b(s(g))) for all b ∈ B1s(U ), g ∈ U . To every U ,V ∈ [[G]], there corresponds a unitary element u(U ,V) ∈ B1t(U ·V) given by u(U ,V)t(gh) = u(g, h) for all g ∈ U and h ∈ V with s(g) = t(h). Then, the cocycle crossed product M is generated by B and partial isometries u(U ) for all U ∈ [[G]] satisfying the following properties. u(U )∗u(U ) = 1s(U ) u(U ) u(V) = u(U ,V) u(U · V) , u(U )bu(U )∗ = αU (b) E(u(U )) = 1U ∩G(0) . and u(U )u(U )∗ = 1t(U ) , for all b ∈ B1s(U ), Note that B′ ∩ M = L∞(X) if and only if almost every Bx is a factor and the cocycle action is free. Conversely, whenever M is a von Neumann algebra with separable predual and B ⊂ M is a regular von Neumann subalgebra with a faithful normal conditional expectation E : M → B and B′ ∩ M = Z(B), there is a canonical discrete measured groupoid G = GB⊂M and a cocycle action of G on the field of factors given by the central decomposition of B, such that B ⊂ M is isomorphic with the cocycle crossed product inclusion. To see this, define P as the set of partial isometries v ∈ M such that the projections p = v∗v and q = vv∗ belong to Z(B) and vBv∗ = Bq. Define the equivalence relation ∼ on P by v ∼ w if and only if v ∈ Bw. Using the usual correspondence between groupoids and inverse semigroups, one identifies P/∼ with the full pseudogroup [[G]] of an essentially unique discrete measured groupoid G with space of units G(0) satisfying L∞(G(0)) = Z(B). Writing X = G(0) and writing B as the direct integral of factors (Bx)x∈X, the above relation between P and the full pseudogroup [[G]] provides a cocycle action of G on (Bx)x∈X. By construction, B ⊂ M is isomorphic with the cocycle crossed product inclusion. 6 Two such cocycle crossed product inclusions are isomorphic if and only if the corresponding measured groupoids are isomorphic and their cocycle actions are cocycle conjugate through this isomorphism of groupoids. This bijective correspondence between regular inclusions B ⊂ M and cocycle crossed products can be viewed in different ways. In [DFP18], this is interpreted in the language of inverse semigroups, where cocycle actions of groupoids become extensions of inverse semigroups. Note that the cocycle crossed product of a free cocycle action of G on a field (Bx)x∈X of factors admits a faithful normal tracial state if and only if each Bx admits such a faithful normal tracial state and X = G(0) admits an invariant probability measure that is equivalent with µ. So, when (M, τ ) is a tracial von Neumann algebra and B ⊂ M is a regular von Neumann subalgebra satisfying B′ ∩ M = Z(B), we can decompose M as the cocycle crossed product of a cocycle action of a pmp discrete measured groupoid G on a field of tracial factors. We get that M is a factor if and only if G is ergodic. Recall from [AR00, Section 3.2] the notion of amenability for discrete measured groupoids. Also recall from [AR00, Section 5.3] that G is amenable if and only if the countable equivalence relation R is amenable and almost all isotropy groups Γx are amenable. 3 Vanishing of 2-cohomology for amenable groupoids In this section, we prove the first parts of Theorem A and B. So we first prove the following result. Theorem 3.1. Let G be a discrete measured groupoid with X = G(0) and (Bx)x∈X a measurable field of II1 factors with separable predual. Assume that G is amenable. When (α, u) is a free cocycle action of G on (Bx)x∈X, the cocycle u is a co-boundary: there exists a measurable field of unitaries G ∋ g 7→ wg ∈ Bt(g) such that u(g, h) = αg(w∗ h) w∗ g wgh for all (g, h) ∈ G(2). By Theorem 3.1, if (M, τ ) is a tracial von Neumann algebra and B ⊂ M is a regular von Neumann subalgebra satisfying B′∩ M = Z(B) such that the associated groupoid is amenable, then the inclusion B ⊂ M is isomorphic to a true crossed product inclusion, without 2-cocycle. So in combination with [ST84, Theorem 1.2], we obtain the following classification of regular subalgebras of amenable tracial von Neumann algebras. Theorem 3.2. Let R be the hyperfinite II1 factor. The regular von Neumann subalgebras B ⊂ R satisfying B′ ∩ M = Z(B) are completely classified by the associated discrete measured groupoid and the type of B. More precisely, given two such inclusions Bi ⊂ R with associated groupoids Gi, i = 1, 2, there exists an automorphism θ ∈ Aut(R) satisfying θ(B1) = B2 if and only if B1 and B2 have the same type and there exists a measure class preserving isomorphism G1 ∼= G2. Remark 3.3. One may formulate Theorem 3.2 without the factoriality assumption on R, i.e. for arbitrary amenable tracial von Neumann algebras (M, τ ) and their regular von Neumann subalgebras B ⊂ M satisfying B′ ∩ M = Z(B). In that case, denote for every n ≥ 1, by zn ∈ Z(B) the largest central projection such that Bzn is of type In. Let z0 = 1 − P∞ n=1 zn, so that Bz0 is of type II1. Since B ⊂ M is regular, we have that zn ∈ Z(M ) for every n ≥ 0. 7 The discrete measured groupoid G associated with B ⊂ M can then be viewed as the disjoint union of the groupoids G(n) = GBzn⊂M zn. Note that since nontrivial groups do not have free actions on type I factors, each of the groupoids G(n), n ≥ 1, has trivial isotropy groups, meaning that each G(n) is a countable equivalence relation. i . 2 1 ∼= G(n) The general statement of Theorem 3.2 then goes as follows. Given, for i = 1, 2, amenable tracial von Neumann algebras (Mi, τi) with regular von Neumann subalgebras Bi ⊂ Mi satisfying i ∩ Mi = Z(Bi), denote by G(n) B′ the associated discrete measured groupoids. Then there exists a ∗-isomorphism θ : M1 → M2 satisfying θ(B1) = B2 if and only if for every n ≥ 0, there is a measure class preserving isomorphism G(n) Proof of Theorem 3.1. Denote by Γx = {g ∈ G t(g) = s(g) = x} the field of isotropy groups. Denote by B ⊂ M the cocycle crossed product and define D = Z(B)′∩ M . Note that B ⊂ D ⊂ M and Z(B) = Z(D). By construction, we have a measurable field of cocycle actions Γx y Bx and D is the direct integral of the cocycle crossed products Bx ⋊ Γx. By [P18, Theorem 0.1], every cocycle action of an amenable group on a II1 factor can be perturbed to a genuine action. So, after a perturbation of (α, u), we may assume that Dx is the crossed product of Bx and a genuine action of Γx, which means that u(g, h) = 1 whenever s(g) = t(g). We denote by u(x, g), x ∈ X, g ∈ Γx, the canonical unitary elements realizing this crossed product decomposition Dx = Bx ⋊ Γx. Denote by R the countable, nonsingular equivalence relation on (X, µ) given by R = {(t(g), s(g)) g ∈ G} . Since G is amenable, R is an amenable equivalence relation. So, up to measure zero, X can be partitioned into R-invariant Borel subsets X = X0 ⊔ X1 such that the restriction of R to X0 has finite orbits and thus admits a fundamental domain, while the restriction of R to X1 is implemented by a free nonsingular action of the group Z. It follows that the groupoid morphism G → R : g 7→ (t(g), s(g)) admits a measurable lift q : R → G that is itself a morphism, meaning that q(x, y)q(y, z) = q(x, z) for all (x, y), (y, z) ∈ R. We may then view G as the semidirect product of the field (Γx)x∈X of groups and the measurable family of group isomorphisms δ(x,y) : Γy → Γx given by δ(x,y)(g) = q(x, y)gq(x, y)−1 for all (x, y) ∈ R and all g ∈ Γy. Note that δ(x,y) ◦ δ(y,z) = δ(x,z). For the same reason as above, we can perturb (α, u) so that the composition with q is a genuine action of R on the field (Bx)x∈X . This means that we have written M as the crossed product of D = Z ⊕ X (Bx ⋊ Γx) dµ(x) by an action of R given by a field of ∗-isomorphisms θ(x,y) : (By ⊂ By ⋊ Γy) → (Bx ⊂ Bx ⋊ Γx) of the form θ(x,y)(b) = αq(x,y)(b) θ(x,y)(u(y, g)) ∈ U (Bx) u(x, δ(x,y)(g)) for all (x, y) ∈ R and b ∈ By, for all (x, y) ∈ R and g ∈ Γy, and satisfying θ(x,y) ◦ θ(y,z) = θ(x,z). In order to prove the theorem, we have to show that there exist measurable families of unitaries v(x, y) ∈ U (Bx) (for (x, y) ∈ R) and w(y, g) ∈ U (By) (for y ∈ X, g ∈ Γy) such that 8 • v is a 1-cocycle for the action θ of R on (Bx)x∈X , in the sense that v(x, y) α(x,y)(v(y, z)) = v(x, z) for all (x, y), (y, z) ∈ R, • for all y ∈ X, we have that (w(y, g))g∈Γy is a 1-cocycle for the action Γy y By, • we have (Ad(v(x, y))◦θ(x,y))(cid:0) w(y, g)u(y, g)(cid:1) = w(x, δ(x,y)(g))u(x, δ(x,y)(g)) for all (x, y) ∈ R and g ∈ Γy. Indeed, once this is proved, we can perturb (α, u) in such a way that Dx is still the crossed product of Bx and a genuine action of Γx (with implementing unitary elements w(x, g)u(x, g), g ∈ Γx) and such that M is the crossed product by an action of R on D given by genuine conjugacies between these actions. This means that we have decomposed M as the crossed product of B and a genuine action of the groupoid G, which is exactly what we have to prove. So it remains to prove the statement above. Denote by Px the normalizer of Bx inside Dx. Note that Px ⊂ U (Dx) is a closed subgroup and that Px = {bu(x, g) b ∈ U (Bx), g ∈ Γx}. We obtain the natural homomorphism Px → Γx : bu(x, g) 7→ g. We denote by Px the Polish space of homomorphic lifts Γx → Px, which we can identify with the space of 1-cocycles for the action Γx y Bx. We have a natural action U (Bx) y Px given by conjugation. Given a lifting homomorphism γ : Γy → Py and given (x, y) ∈ R, we define the lifting homo- morphism β(x,y)(γ) : Γx → Px given by β(x,y)(γ) = θ(x,y) ◦ γ ◦ δ−1 (x,y) . We have found a measurable field of homeomorphisms β(x,y) : Py → Px satisfying β(x,y)◦β(y,z) = β(x,z). We also have the measurable field of Polish group isomorphisms θ(x,y) : U (By) → U (Bx) satisfying θ(x,y) ◦ θ(y,z) = θ(x,z). By construction, β(x,y)(b · γ) = θ(x,y)(b) · β(x,y)(γ) meaning that θ(x,y), β(x,y) form a measurable field of conjugacies between the continuous Polish group actions U (Bx) y Px. By Theorem 4.1 below, the actions U (Bx) y Px have dense orbits. By Theorem 3.5 below, we can thus find a measurable section X ∋ x 7→ π(x) ∈ Px and a 1-cocycle R ∋ (x, y) 7→ v(x, y) ∈ U (Bx) satisfying v(x, y) · β(x,y)(π(y)) = π(x) for all (x, y) ∈ R. Writing π(x)(g) = w(x, g)u(x, g), the theorem is proved. Corollary 3.4. Let (M, τ ) be an amenable tracial von Neumann algebra with separable predual and B ⊂ M a regular von Neumann subalgebra. Then B′ ∩ M = Z(B) if and only if there exists a Cartan subalgebra A ⊂ M with A ⊂ B. Proof. If A ⊂ M is a Cartan subalgebra with A ⊂ B, we have B′ ∩ M ⊂ A′ ∩ M = A ⊂ B. Therefore, B′ ∩ M = Z(B). Conversely, assume that B ⊂ M is a regular von Neumann subalgebra satisfying B′ ∩ M = Z(B). As in Remark 3.3, denote for every n ≥ 1, by zn ∈ Z(B) the largest central projection such that Bzn is of type In. Since B ⊂ M is regular, we have that zn ∈ Z(M ). Since it is straightforward to construct Cartan subalgebras An ⊂ M zn with An ⊂ Bzn (and such that Bzn ∼= Mn(C) ⊗ An), we may assume that B is of type II1. By Theorem 3.1, we can write B ⊂ M as the crossed product inclusion B ⊂ B ⋊ G, where G is a pmp amenable discrete measured groupoid acting freely on B. Write X = G(0) and choose 9 any free action of G on a field of Cartan subalgebras (Cx ⊂ Dx)x∈X, with each Dx being the hyperfinite II1 factor. Let C ⊂ D ⊂ N be the corresponding crossed product inclusion. By Theorem 3.2, there exists a ∗-isomorphism θ : N → M satisfying θ(D) = B. Writing A = θ(C), we have found the required Cartan subalgebra of M . Although the formulation of the following result looks quite different from the 'cohomology lemmas' in [JT82, Appendix] and [S85, Theorem 5.5], the proof is very similar to the proof of [S85, Theorem 5.5]. Theorem 3.5. Let R be a countable nonsingular equivalence relation on the standard probabil- ity space (X, µ). Let (Gx y Px)x∈X be a measurable field of continuous actions of Polish groups on Polish spaces, on which R is acting by conjugacies: we have measurable fields of group iso- morphisms R ∋ (x, y) 7→ δ(x,y) : Gy → Gx and homeomorphisms R ∋ (x, y) 7→ β(x,y) : Py → Px satisfying δ(x,y) ◦ δ(y,z) = δ(x,z) , β(x,y) ◦ β(y,z) = β(x,z) and β(x,y)(g · π) = δ(x,y)(g) · β(x,y)(π) for all (x, y), (y, z) ∈ R and g ∈ Gy, π ∈ Py. Let X ∋ x 7→ π(x) ∈ Px be a measurable section. Assume that R is amenable and assume that for all (x, y) ∈ R, the element π(x) belongs to the closure of Gx · β(x,y)(π(y)). Then, there exists a measurable family R ∋ (x, y) 7→ v(x, y) ∈ Gx and a section X ∋ x 7→ π′(x) ∈ Px satisfying the following properties. • v is a 1-cocycle: v(x, y) δ(x,y)(v(y, z)) = v(x, z) for all (x, y), (y, z) ∈ R. • v(x, y) · β(x,y)(π′(y)) = π′(x) for all (x, y) ∈ R. Proof. By [CFW81, Theorem 10], the equivalence relation R is hyperfinite. Fix subequivalence relations R0 ⊂ R1 ⊂ ··· having finite orbits such that R = Sn Rn, up to measure zero. We assume that R0 = {(x, x) x ∈ X} is the trivial subequivalence relation. To prove the theorem, we inductively define 1-cocycles Rn ∋ (x, y) 7→ vn(x, y) ∈ Gx and sections X ∋ x 7→ πn(x) ∈ Px with the following properties. • For every n ≥ 0, the restriction of vn+1 to Rn equals vn. • Defining βn,(x,y)(π) = vn(x, y) · β(x,y)(π) for all π ∈ Py, (x, y) ∈ Rn, n ≥ 0, we have βn,(x,y)(πn(y)) = πn(x) for all (x, y) ∈ Rn. • For all n ≥ 0 and all x ∈ X, we have that πn(x) ∈ Gx · π(x). • For all x ∈ X, the sequence πn(x) ∈ Px is convergent. Once these statements are proven, we can define π′(x) = limn πn(x) and v(x, y) = vn(x, y) whenever (x, y) ∈ Rn. Choose a measurable family of metrics dx on Px that induce the topology on Px. Define v0(x, x) = 1 and π0(x) = π(x). Then assume that n ≥ 0 and that vn and πn have been defined. Write βn,(x,y)(π) = vn(x, y) · β(x,y)(π). Choose a fundamental domain X0 ⊂ X for Rn. Every y ∈ X0 has a finite Rn-orbit. Therefore, since βn,(x,y) : Py → Px is a homeomorphism, we can choose a measurable function y 7→ εy > 0 such that dx(βn,(x,y)(π), βn,(x,y)(πn(y))) < 2−n whenever y ∈ X0, (x, y) ∈ Rn, π ∈ Py and dy(π, πn(y)) < εy. 10 Denote by S the restriction of Rn+1 to X0. Choose a fundamental domain X1 ⊂ X0 for S. Note that X1 also is a fundamental domain for Rn+1. Whenever y ∈ X1 and (x, y) ∈ S, by assumption, π(x) belongs to the closure of Gx · β(x,y)(π(y)) = Gx · β(x,y)(πn(y)) . Since πn(x) ∈ Gx · π(x), also πn(x) belongs to the closure of Gx · β(x,y)(πn(y)). Defining S1 = {(x, y) ∈ S x ∈ X0 \ X1, y ∈ X1}, we can thus choose a measurable map S1 ∋ (x, y) 7→ w(x, y) ∈ Gx such that dx(πn(x), w(x, y) · β(x,y)(πn(y))) < min{εx, 2−n} for all (x, y) ∈ S1. Since Rn+1 is the free product of Rn and S, we can define vn+1 as the unique 1-cocycle Rn+1 ∋ (x, y) 7→ vn+1(x, y) ∈ Gx satisfying vn+1(x, y) = vn(x, y) for all (x, y) ∈ Rn and vn+1(x, y) = w(x, y) for all (x, y) ∈ S. Define βn+1,(x,y)(π) = vn+1(x, y) · β(x,y)(π) for all (x, y) ∈ Rn+1 and π ∈ Py. Uniquely define the section X ∋ x 7→ πn+1(x) ∈ Px given by πn+1(x) = πn(x) πn+1(x) = βn+1,(x,y)(πn(y)) if x ∈ X1, and if x ∈ X \ X1, y ∈ X1, (x, y) ∈ Rn+1. By construction, βn+1,(x,y)(πn+1(y)) = πn+1(x) for all (x, y) ∈ Rn+1. When x ∈ X0 \ X1, take the unique y ∈ X1 with (x, y) ∈ Rn+1. Then (x, y) ∈ S1, so that πn+1(x) = w(x, y) · β(x,y)(πn(y)) and thus dx(πn+1(x), πn(x)) < min{εx, 2−n} . When x ∈ X \ X0, take the unique y ∈ X0 with (x, y) ∈ Rn. Since dy(πn+1(y), πn(y)) < εy and βn,(x,y)(πn(y)) = πn(x), we find that dx(βn,(x,y)(πn+1(y)), πn(x)) < 2−n . Because (x, y) ∈ Rn, we get that βn,(x,y)(πn+1(y)) = βn+1,(x,y)(πn+1(y)) = πn+1(x). So, we have proved that dx(πn+1(x), πn(x)) < 2−n for all x ∈ X. Continuing inductively, it follows that for all x ∈ X, the sequence πn(x) ∈ Px is convergent. This concludes the proof of the theorem. 4 Approximate vanishing of 1-cocycles Recall that a trace preserving action Γ yσ (B, τ ) on a tracial von Neumann algebra is called strongly ergodic if any sequence bn in the unit ball of B satisfying the approximate invariance property limn kσg(bn) − bnk2 = 0 for all g ∈ Γ must be approximately scalar, in the sense that limn kbn − τ (bn)1k2 = 0. Theorem 4.1. Let Γ be a countable group. The following properties are equivalent. (i) Γ is amenable. 11 (ii) No free trace preserving action Γ yσ (B, τ ) on a tracial diffuse von Neumann algebra is strongly ergodic. (ii') For every free trace preserving action Γ yσ (B, τ ) on a tracial diffuse von Neumann algebra and for every non-principal ultrafilter ω on N, the ultrapower action σω on B = Bω has a diffuse fixed point algebra Bσω . (ii") The Bernoulli Γ-action Γ yσ R = (M2×2(C))⊗Γ with base M2×2(C) is not strongly ergodic. (iii) For every free action Γ yσ N on a II1 factor N , every 1-cocycle (wg)g∈Γ is approximately for every finite subset F ⊂ Γ and every ε > 0, there exists u ∈ U (N ) a co-boundary: such that kuσg(u∗) − wgk2 ≤ ε for all g ∈ F . (iii') The previous statement with N being the hyperfinite II1 factor. (iv) Let Γ yσn Nn be a sequence of free Γ-actions on II1 factors. Let ω be a non-principal ultrafilter on N and denote by N = Qω Nn the corresponding ultraproduct II1 factor with σ = Qω σn the ultraproduct Γ-action, where σg((xn)n) = (σn,g(xn))n. Then, the fixed point algebra N σ is an irreducible subfactor of N . In the same setting as (iv), every 1-cocycle for σ = Qω σn is a co-boundary. (v) Proof. (i) ⇒ (iv). It is sufficient to prove that given any x ∈ (N )1, there exists a von Neumann subalgebra B ⊂ N σ such that EB ′∩N (x) = τ (x)1. Let x = (xn)n with xn ∈ (Nn)1 and let Fn ⊂ Γ be an increasing sequence of finite subsets exhausting Γ. By Lemma 4.4, for each n, there exists a finite dimensional abelian von Neumann subalgebra An ⊂ N such that kσg(y) − yk2 ≤ 2−n kEA′ n∩N (xn) − τ (xn)1k2 ≤ 2−n . for all y ∈ (An)1, g ∈ Fn , (4.1) (4.2) By (4.1), it follows that B = Qω An is contained in the fixed point algebra N σ, while (4.2) implies that EB ′∩N (x) = τ (x)1. (iv) ⇒ (iii). We use Connes' 2-by-2 matrix trick, as in the proof of [P18, Proposition 1.3.1◦]. Denote be σ the Γ-action on N = M2×2(N ) = N ⊗ M2×2(C) given by σg = σg ⊗ id. Denoting by {eij 1 ≤ i, j ≤ 2} the matrix units for M2×2(C) ⊂ N , we get that wg = e11 + wge22 is a 1-cocycle for σ, and thus also for the ultrapower action σω on N ω. Note that if we denote σ′ g = (Ad wg)◦ σg, then its ultrapower (σ′)ω coincides with (Ad wg)◦ σω . Thus, since we assume that (iv) holds, the fixed point algebra Q of the Γ-action (σ′)ω on N ω is an irreducible subfactor of N ω. Since e11, e22 are projections in Q, it follows that they are equivalent in Q via a partial isometry u∗e12 ∈ Q, for some u ∈ U (N ω). But the fact that u∗e12 g (u∗) for all g ∈ Γ, i.e., lies in the fixed point algebra Q is equivalent to the fact that wg = uσω w is a co-boundary for σω. Then, (iii) follows. (iii) ⇒ (v). By (iii), every 1-cocycle for σ is approximately a co-boundary. By the usual re-indexation argument, the 1-cocycle is a true co-boundary. (v) ⇒ (iii) follows by taking Nn = N for all n. (ii') ⇒ (ii) ⇒ (ii"), as well as (iii) ⇒ (iii'), are trivial. (iv) ⇒ (ii'). Let z ∈ Z(B) be the largest projection such that Z(B)z is diffuse. Then, z is Γ-invariant and it follows from [S81] that Z(B)σω z is diffuse. For the atomic part, it follows from (iv) that Bσω (1 − z) is diffuse. 12 The implication (ii") ⇒ (i) is contained in [J81]. (iii') ⇒ (i). For this, we follow a similar argument to [P18, Proposition 1.3.2◦] and [P01, Theorem 3.2]. Let (N0, ϕ0) be a copy of the 2 by 2 matrix algebra with the state given by the weights 1/3 and 2/3. Let Γ yα (N , ϕ) = ⊗g(N0, ϕ0)g be the Bernoulli Γ-action with base (N0, ϕ0) and denote by N = Nϕ the centralizer of the state ϕ. Note that N is isomorphic with the hyperfinite II1 factor. Denote by σ the corresponding Connes-Størmer Bernoulli action, given by restricting α to the II1 factor N . Let B ⊂ N be a copy of the 2 by 2 matrix algebra with matrix units (eij)1≤i,j≤2. By our choice of ϕ0 and because N is a factor, we can choose an isometry v ∈ N such that σϕ t (v) = 2−itv and vv∗ = e11. As in [P18, Proposition 1.3.2◦] and [P01, Theorem 3.2], define the 1-cocycle (wg)g∈Γ for σ given by wg = vαg(v∗) + e21vαg(v∗e12) . Since (iii') holds, w is approximately a co-boundary. Take a sequence of unitaries an ∈ U (N ) such that limn kwgσg(an) − ank2 = 0 for all g ∈ Γ. Working in L2(N , ϕ), it follows that limn kv∗wgσg(an) − v∗ank2 = 0 for all g ∈ Γ. But, v∗wg = αg(v∗) and we conclude that limn kαg(v∗an) − v∗ank2 = 0 for all g ∈ Γ. Since σϕ t (v∗an) = 2itv∗an, the sequence v∗an lies in L2(N , ϕ) ⊖ C1. Since kv∗ank2 = kvv∗ank2 = ke11ank2 = 1 √2 , we conclude that the unitary representation π of Γ on L2(N , ϕ) ⊖ C1 given by the action α weakly contains the trivial representation of Γ. As in [J81], π is weakly contained in the regular representation of Γ. So it follows that Γ is amenable. Lemma 4.2. Let Γ yσ N be a free action of a countable group Γ on a separable II1 factor N . Let ω be a non-principal ultrafilter on N and σω the ultrapower action of Γ on N ω defined by σg((xn)n) = (σg(xn))n for all (xn)n ∈ N ω and g ∈ Γ. g (uN u∗), g ∈ Γ, are freely independent. There exists a unitary element u ∈ N ω such that N, σω g (uN u∗) then σω leaves N invariant and the restriction of σω to Thus, if one denotes N = Wg σω N ∨ N is isomorphic to the diagonal free product Γ y N ∗ N ∗Γ between σ and the free Bernoulli Γ-action with base N . Proof. Since N has trivial relative commutant in M = N ⋊ Γ, by the main theorem of [P92] and [P13, Theorem 0.1], there exists a Haar unitary u ∈ N ω such that u is freely independent to the separable II1 factor M . But then, if we denote by (ug)g ∈ M the canonical unitaries implementing the action σ, it is easy to see that N, uguN u∗u∗ g, g ∈ Γ, are all mutually free von Neumann subalgebras. Since σ leaves Wg uguN u∗u∗ g invariant, and implements on it the free Lemma 4.3. Let Γ be a countable amenable group, N a II1 factor and Γ yρ N ∗Γ the free Bernoulli Γ-action with base N . Then ρ is not strongly ergodic. Bernoulli Γ-action, the lemma follows. Proof. Let a = a∗ ∈ N be a semi-circular element and denote by ag its identical copies in position g ∈ Γ inside the free product N ∗Γ. Thus, ρ acts on the set {ag}g by left translation: ρh(ag) = ahg. Let Kn ⊂ Γ be a sequence of Følner sets and denote bn = Kn−1/2Pg∈Kn ag. Then bn is also a semicircular element and one has kρh(bn) − bnk2 2 = hKn∆Kn/Kn → 0 for all h ∈ Γ . 13 Thus, the element b = (bn)n ∈ (N ∗Γ)ω is semicircular with ρh(b) = b for all h ∈ Γ, showing that ρ is not strongly ergodic. Note that the construction of bn in the proof of Lemma 4.3 can also be interpreted as follows. We apply the free Gaussian functor to the regular representation of Γ. Since Γ is amenable, the regular representation contains a sequence of almost invariant unit vectors. This provides approximately Γ-invariant semicircular elements, i.e. the elements bn. Lemma 4.4. Let Γ yσ N be a free action of a countable amenable group Γ on a II1 factor N . Given any finite subsets F0 ⊂ Γ, X0 ⊂ N and any δ > 0, there exists a d-dimensional abelian von Neumann subalgebra A0 ⊂ N with all minimal projections of trace 1/d, such that and kEA′ 0∩N (x) − τ (x)1k2 ≤ δ kσg(a) − ak2 ≤ δ for all a ∈ (A0)1, g ∈ F0 and x ∈ X0. Proof. By [P13], there exists a σ-invariant separable II1 subfactor N0 ⊂ N that contains X0 0 ⊂ N ω contains a σω-invariant II1 and on which Γ acts freely. By Lemmas 4.2 and 4.3, N ω subfactor P ⊂ N ω that is freely independent to N0 and on which ρ = σωP is free and non strongly ergodic. 0∩N ω (x) − Since P is freely independent to N0 and X0 ⊂ N0, we can fix d ≥ 1 such that kEB ′ τ (x)1k2 ≤ δ/2 for every x ∈ X0 and for every d-dimensional abelian von Neumann subalgebra B0 ⊂ P with all minimal projections having trace 1/d. Since the action ρ of Γ on P is not strongly ergodic, we can make such a choice of B0 ⊂ P such that kσg(b) − bk2 ≤ δ/2 for all b ∈ (B0)1 and g ∈ F0. Writing B0 as an ultraproduct of d-dimensional abelian von Neumann subalgebras B0,n ⊂ N with all minimal projections having trace 1/d, the subalgebra A0 = B0,n satisfies all required conditions when n is large enough. In [P18], VC is defined as the class of countable groups Γ with the property that for every free It is proved cocycle action (α, u) of Γ on a II1 factor B, the 2-cocycle u is a co-boundary. in [P18] that the class VC is closed under amalgamated free products over finite subgroups, but is not closed under amalgamated free products over amenable subgroups. One similarly defines the class VCω of countable groups for which every such 2-cocycle u is approximately a co-boundary. As a corollary to Theorem 4.1, we prove that VCω is closed under amalgamated free products over amenable subgroups. Corollary 4.5. The class VCω is closed under amalgamated free products over amenable sub- groups. In particular, if Γ1, Γ2 are free products of amenable groups and H ⊂ Γ1, Γ2 is an amenable subgroup, then Γ1 ∗H Γ2 ∈ VCω. Proof. The same proof as the one of [P18, Proposition 1.5] shows that if Γ1, Γ2 ∈ VCω and H ⊂ Γ1, Γ2 is a subgroup with the property that any 1-cocycle for σω is co-boundary, where σ is an arbitrary free H-action on a II1 factor, then Γ1∗H Γ2 lies in VCω as well. But by Theorem 4.1, the latter property is satisfied by any amenable group H. Remark 4.6. Note that Lemmas 4.2 and 4.3 already show that if Γ is a countable amenable group, then no free action of Γ on a II1 factor is strongly ergodic. This was shown for N having property Gamma in [B90], where one actually proves that in this case, one has (N ⋊ Γ)′∩ N ω 6= C1, by using results in [O85]. One can also give an alternative, more direct ("by hand") proof of the lack of strong ergodicity by constructing projections p ∈ N of trace 1/2 that are almost Γ-invariant by using the Følner condition and "local Rokhlin towers" for Γ y N and a maximality argument. 14 Recall that an action α of a countable group G on a countable pmp equivalence relation S on (X, µ) is called outer if (x, αg(x)) 6∈ S for all g ∈ Γ \ {e} and a.e. x ∈ X. Theorem 4.7. Let B be a II1 factor with separable predual and Cartan subalgebra A ⊂ B. Let G be a countable amenable group and α an action of G by automorphisms of A ⊂ B so that the action on the associated equivalence relation S is outer. Every 1-cocycle c : G → NB(A) for α is approximately a co-boundary: there exists a sequence of an ∈ NB(A) such that n kcg − anαg(a∗ lim n)k2 = 0 for all g ∈ G . 2 ≤ 8ε for every g ∈ F . Proof. Fix a finite subset F ⊂ G and ε > 0. We will construct an element a ∈ NB(A) such that kcg − aαg(a∗)k2 Denote βg = (Ad cg) ◦ αg. Since the action α of G on S is outer, both αg and βg act freely on A. By the version of the Ornstein-Weiss Rokhlin theorem proved in [KL16, Theorem 4.46], we can take n ∈ N and, for all i ∈ {1, . . . , n}, a finite subset Ti ⊂ G and projections pi, qi ∈ A such that the following conditions hold. • The projections {αh(pi) i = 1, . . . , n, h ∈ Ti} are orthogonal and their sum p has trace at least 1 − ε. • The projections {βh(qi) i = 1, . . . , n, h ∈ Ti} are orthogonal and their sum q has trace at least 1 − ε. • For every i ∈ {1, . . . , n}, we have that τ (pi) = τ (qi). • For every g ∈ F and i ∈ {1, . . . , n}, we have that gTi ∩ Ti ≥ (1 − ε)Ti. Since the equivalence relation S is ergodic, we can choose vi ∈ B such that v∗ and viAv∗ ww∗ = 1 − q and wAw∗ = A(1 − q). Define i = qi i = Aqi, for all i ∈ {1, . . . , n}. We can also choose w ∈ B such that w∗w = 1 − p, i vi = pi, viv∗ a = w + n Xi=1 Xh∈Ti ch αh(vi) . Note that a is a sum of partial isometries normalizing A and having orthogonal initial, resp. final projections. It then follows that a ∈ NB(A). Fix g ∈ F and define the projection f ∈ A given by Xi=1 Xh∈gTi∩Ti αh(pi) . f = n One checks that cgαg(a)f = af . It then follows that kcgαg(a) − ak2 2 = k(cgαg(a) − a)(1 − f )k2 2 ≤ 4τ (1 − f ) . We also have that τ (f ) = n Xi=1 gTi ∩ Ti τ (pi) ≥ (1 − ε) n Xi=1 Ti τ (pi) = (1 − ε) τ (p) ≥ 1 − 2ε . So, we have proven that kcg − aαg(a∗)k2 2 ≤ 8ε for all g ∈ F . 15 5 Regular inclusions containing a Cartan subalgebra Let (M, τ ) be a tracial von Neumann algebra with separable predual and Cartan subalgebra A ⊂ M . By [FM75], we can write A = L∞(X) and M = L(R, u) where R is a countable pmp equivalence relation on the standard probability space (X, µ) and u is a 2-cocycle on R with values in the circle T. There is a bijective correspondence between intermediate von Neumann algebras A ⊂ B ⊂ M and subequivalence relations S ⊂ R given by B = L(S, u). Note that automatically, B′ ∩ M ⊂ A′ ∩ M = A ⊂ B, so that B′ ∩ M = Z(B). Recall that an action α of a countable group Γ on a countable pmp equivalence relation S on (X, µ) is called outer if for all g ∈ Γ \ {e}, we have that (x, αg(x)) 6∈ S for a.e. x ∈ X. We start by sketching the proof of the following folklore result. Proposition 5.1. In the setting introduced above, we have that the inclusion B ⊂ M is regular if and only if the subequivalence relation S ⊂ R is strongly normal in the sense of [FSZ88, Definition 2.14]. In that case, writing A ⊂ B as the direct integral of the Cartan inclusions (Ay ⊂ By)y∈Y with the By being factors, we can write M as the cocycle crossed product of a pmp discrete measured groupoid G and a free cocycle action (α, u) on this field (Ay ⊂ By)y∈Y . This means that G(0) = Y and that we are given 1. a measurable field of ∗-isomorphisms G ∋ g 7→ αg : Bs(g) → Bt(g) satisfying αg(As(g)) = At(g), 2. a measurable field of unitaries G(2) ∋ (g, h) 7→ u(g, h) ∈ NBt(g)(At(g)), satisfying the conditions in Definition 2.2 and such that for every y ∈ Y , the action of the isotropy group Γy = {g ∈ G s(g) = y = t(g)} on the equivalence relation Sy associated with Ay ⊂ By is outer. Proof. By definition, strong normality of S ⊂ R means that R is generated by the graphs of the elements of F = {ϕ ∈ [R] ϕ ∈ Aut(S)}. Whenever ϕ ∈ F, the corresponding unitary element uϕ ∈ M = L(R, u) normalizes B = L(S, u). Therefore, strong normality of S ⊂ R implies the regularity of B ⊂ M . Conversely, assume that B ⊂ M is regular. Since B′ ∩ M ⊂ A′ ∩ M = A ⊂ B, we get that B′ ∩ M = Z(B). So, as in Section 2, we can write M as a cocycle crossed product of a pmp discrete measured groupoid G by a free cocycle action on B. To conclude the proof of the proposition, we have to prove that this cocycle action can be chosen in such a way that it globally preserves A ⊂ B. For this, it suffices to prove that for every u ∈ NM (B), there exists a unitary b ∈ U (B) such that bu ∈ NM (A). Fix u ∈ NM (B) and denote by β ∈ Aut(B), the automorphism given by β(x) = uxu∗ for all x ∈ B. Fix any v ∈ NM (A) and denote by α ∈ Aut(A), the automorphism given by α(a) = vav∗ for all a ∈ A. Define b = EB(vu∗). It follows that bβ(a) = α(a)b for all a ∈ A. Denoting by zv ∈ Z(B) the central support of the element b∗b ∈ B, we conclude that the Cartan subalgebras β(A)zv and Azv of Bzv are unitarily conjugate. Since A ⊂ M is regular, varying v ∈ NM (A), the join of all these central projections zv ∈ Z(B) is 1. Therefore β(A) and A are unitarily conjugate. So we can choose b ∈ U (B) such that bβ(A)b∗ = A and thus bu ∈ NM (A). Ignoring the scalar cocycles in Proposition 5.1, we can also write the equivalence relation R as the semidirect product of the subequivalence relation S and an outer cocycle action of the groupoid G by automorphisms of S. 16 The main goal of this section is to prove the following analogues of Theorems 3.1 and 3.2. These results were so far only known for amenable groups (see [FSZ88, Theorem 3.4], [BG84, Theorem 3.4] and [GS86]). Theorem 5.2. Let G be a discrete measured groupoid with Y = G(0) and (Ay ⊂ By)y∈Y a measurable field of Cartan subalgebras in II1 factors with separable predual. Assume that G is amenable. When (α, u) is a free cocycle action of G on (Ay ⊂ By)y∈Y , the cocycle u is a co-boundary: there exists a measurable field of unitaries G ∋ g 7→ wg ∈ NBt(g)(At(g)) such that u(g, h) = αg(w∗ h) w∗ g wgh for all (g, h) ∈ G(2). Applying Theorem 5.2 in a setup without scalar cocycles, we obtain the following result: if R is a countable pmp equivalence relation with strongly normal subequivalence relation S ⊂ R and if the quotient groupoid G is amenable, we can write R as the semidirect product of S by a free action of G on S. Proof of Theorem 5.2. The proof follows exactly the same lines as the proof of Theorem 3.1. We need the following two ingredients. • Vanishing of 2-cocycles for actions of amenable groups: if B is a separable II1 factor with Cartan subalgebra A ⊂ B and associated ergodic type II1 equivalence relation S and if (α, u) is a cocycle action of an amenable group Γ on A ⊂ B such that the corresponding action on S is outer, then the 2-cocycle u is a co-boundary. Theorem 3.4 in [FSZ88] provides 2-cocycle vanishing for the corresponding outer cocycle action on the equivalence relation S. So, we may perturb (α, u) such that u(g, h) ∈ U (A) for all g, h ∈ Γ. Writing A = L∞(X, µ), we now have that Γ y (X, µ) is a free action and we can view u as a scalar 2-cocycle on the associated orbit equivalence relation. Since this equivalence relation is hyperfinite, the scalar 2-cocycle u is a co-boundary. • Approximate vanishing of 1-cocycles for actions of amenable groups: this is Theorem 4.7 above. We then follow the same method as in the proof of Theorem 3.1. Applying Theorem 3.5, the result follows. Below, we then prove the following classification result. Theorem 5.3. Let R be the hyperfinite II1 factor with Cartan subalgebra A ⊂ R. The inter- mediate subalgebras A ⊂ B ⊂ R such that B ⊂ R is regular are completely classified by the type of B and the associated discrete measured groupoid. More precisely, given for i = 1, 2, von Neumann subalgebras Ai ⊂ Bi ⊂ R with Ai ⊂ R Cartan and Bi ⊂ R regular, there exists an automorphism θ ∈ Aut(R) satisfying θ(B1) = B2 and θ(A1) = A2 if and only if B1 and B2 have the same type and G1 ∼= G2. As in Remark 3.3, it is possible to formulate a version of Theorem 5.3 without assuming factoriality, i.e. for von Neumann subalgebras of an arbitrary amenable tracial von Neumann algebra. In order to deduce Theorem 5.3 from the vanishing of 2-cohomology in Theorem 5.2, we need to show that an amenable discrete measured groupoid G with G(0) = Y has only one outer 17 measure preserving action (up to cocycle conjugacy) on the field (Sy)y∈Y of ergodic hyperfinite type II1 equivalence relations. For outer actions of amenable groups on the ergodic hyperfinite type II1 equivalence relation, this was proved in [BG84, Theorem 3.4] and this should be considered as an equivalence relation version of Ocneanu's theorem [O85]. For completeness, we include a short proof of this result (see Theorem 5.4 and Lemma 5.5 below). We then establish the groupoid case in Theorem 5.7 (by invoking once more Theorem 3.5). Once Theorem 5.7 proven, Theorem 5.3 follows from Theorem 5.2. Theorem 5.4 (See [BG84, Theorem 3.4]). Let B be the hyperfinite II1 factor with Cartan subalgebra A ⊂ B and associated equivalence relation S. Let G be a countable amenable group and α, β actions of G on A ⊂ B such that the induced action on S is outer. Then there exists a ∗-automorphism θ ∈ Aut(B) and a 1-cocycle g 7→ cg ∈ NB(A) for the action α such that θ(A) = A and θ ◦ βg ◦ θ−1 = (Ad cg) ◦ αg for all g ∈ G . We prove Theorem 5.4 below, as an immediate consequence of the following lemma, which is a special case of [GS86, Theorem 1.5]. Lemma 5.5 (See [GS86, Theorem 1.5]). Let R be a countable pmp ergodic hyperfinite equiv- alence relation on the standard probability space (X, µ). Let α, β : R → G be 1-cocycles with values in a countable group G. Assume that α, β are surjective and that their kernels are ergodic subequivalence relations of R. Then, there exists θ ∈ Aut(R) such that for a.e. (x, y) ∈ R. α(x, y) = β(θ(x), θ(y)) Proof. In the course of the proof, we again tacitly assume that all statements are valid up to sets of measure zero. We call a σ-algebra B of Borel subsets of X atomic if B is generated by countably many minimal elements, called atoms. Note that B is atomic if and only if the algebra of bounded B-measurable functions is an atomic von Neumann subalgebra of L∞(X). We call decomposition of R any n-tuple P = (ϕ1, . . . , ϕn) of elements of [[R]] such that the range sets Xi = R(ϕi) form a partition of X with D(ϕi) = X1 for all i and ϕ1(x) = x for all x ∈ X1. Note that the graphs of the elements ϕi generate a finite subequivalence relation of R. We say that such a decomposition is compatible with an atomic σ-algebra B0 and a cocycle α : R → G if the following conditions hold. • For every i = 1, . . . , n, the set Xi belongs to B0 and a subset U of X1 belongs to B0 if and only if ϕi(U ) belongs to B0. • For every g ∈ G and i = 1, . . . , n, the set {x ∈ X1 α(x, ϕi(x)) = g} belongs to B0. We denote by G(P,B0) the pseudogroup of partial transformations ϕ with D(ϕ), R(ϕ) ∈ B0 and with the graph of ϕ contained in the finite subequivalence relation generated by P . Assume now that we have two compatible decompositions: P = (ϕ1, . . . , ϕn), B0, α, and Q = (ψ1, . . . , ψn), D0, β. Write X1 = D(ϕi) and Y1 = D(ψi). We call isomorphism between these two compatible decompositions any measure preserving isomorphism θ0 : B0 → D0 satisfying the following properties. • θ0(X1) = Y1 and θ0(R(ϕi)) = R(ψi) for all i. 18 • θ0(ϕi(U )) = ψi(θ(U )) whenever U ∈ B0 and U ⊂ X1. • For every g ∈ G and every i, θ0 maps the set {x ∈ X1 α(x, ϕi(x)) = g} to the set {y ∈ Y1 β(y, ψi(y)) = g}. 1, . . . , ϕ′ Further assume that (ϕ′ position P ′ of R consisting of the maps ϕiϕ′ We make the following two claims. 1. There exists an atomic σ-algebra B1 containing C such that the decomposition P ′ is com- m) is decomposition of RX1 . We then define the refined decom- j. Let C be any atomic σ-algebra containing B0. patible with B1 and α. 1, . . . , ψ′ m) such that the refined decomposition Q′ consisting of the maps ψiψ′ 2. There exists an atomic σ-algebra D1 containing D0 and RY1 admits a decomposition (ψ′ j is compat- ible with D1 and β, and such that θ0 can be extended to an isomorphism between P ′,B1, α and Q′,D1, β. Once these claims are proven, we inductively construct as follows finer and finer compatible decompositions Pn,Bn, α and Qn,Dn, β and isomorphisms θn between them. At step 0, we take the trivial decomposition consisting of the identity map on X and the trivial σ-algebras consisting of ∅ and X. At the n-th step for n odd, using the first part of the claim, we choose the refinements Pn and Bn in such a way that the pseudogroup G(Pn,Bn) approximately contains a given finite subset of [[R]]. This is possible because R is hyperfinite. We then use the second part of the claim to pick Qn,Dn and to extend the isomorphism θn−1 to θn. At the next even step, we first choose the refinements Qn+1,Dn+1 in such a way that G(Qn+1,Dn+1) approximately contains a given finite subset of [[R]]. We then pick Pn+1,Bn+1 and extend the isomorphism θn to θn+1. Taking the inductive limit, we find a pmp isomorphism θ : X → X with the following properties. θ ◦ ϕi = ψi ◦ θ and θ(cid:0){x ∈ D(ϕi) α(x, ϕi(x)) = g}(cid:1) = {y ∈ D(ψi) β(y, ψi(y)) = g} whenever ϕi, ψi belong to one of the decompositions Pn, Qn. The graphs of all ϕi, resp. all ψi, generate the entire equivalence relation R. The first property thus implies that θ ∈ Aut(R). The second property says that α(x, ϕi(x)) = β(θ(x), ψi(θ(x))). In combination with the first property, this means that α(x, y) = β(θ(x), θ(y)) for all (x, y) ∈ R and the theorem is proven. It remains to prove the two claims. To prove the first claim, it suffices to show the following: if P = (η1, . . . , ηr) is a decomposition of R and C is an atomic σ-algebra, there exists an atomic σ-algebra B containing C such that P , B and α are compatible. Write Zi = R(ηi). First refine C so that Zi ∈ C for all i. For every i, define as follows the atomic σ-algebras Ci and Di on Z1. Define Ci = η−1 (CZi) and define Di generated by the partition of Z1 into the subsets i Z1 = Gg∈G {x ∈ Z1 α(x, ηi(x)) = g} . The finitely many atomic σ-algebras Ci and Di generate an atomic σ-algebra on Z1 that we denote as E. Then define B as the σ-algebra generated by the σ-algebras ηi(E) on Zi. This concludes the proof of the first claim. Since θ0(X1) = Y1 and thus θ0(B0X1 ) = D0Y1, we can choose an atomic σ-algebra E on Y1 such that the restriction of θ0 to B0X1 extends to a measure preserving σ-algebra isomorphism θ : B1X1 → E. Write X′ 1 = Fn Un where the Un are atoms of the σ-algebra B1. Since the Un are atoms, we find elements gj,n ∈ G such that α(x, ϕ′ j (x)) = gj,n j) and write X′ j = R(ϕ′ 19 for all j and all x ∈ Un. Since the cocycle β is surjective, we find non-negligible ψj,n ∈ [[R]] such that β(y, ψj,n(y)) = gj,n. Since the kernel of β is ergodic, we can make this choice of ψj,n such that D(ψj,n) = θ(Un) and R(ψj,n) = θ(ϕ′ j(Un)). So, for a fixed j ∈ {1, . . . , m}, the domains D(ψj,n), n ∈ N, form a partition of θ(X′ 1) and the ranges R(ψj,n), n ∈ N, form a partition of θ(X′ j by gluing together the ψj,n, n ∈ N. This concludes the proof of the second claim. j). We define ψ′ Proof of Theorem 5.4. Write M = B ⋊α G and N = B ⋊β G. Denote by R the hyperfinite ergodic type II1 equivalence relation on the standard probability space (X, µ). Choose ∗- isomorphisms π1 : M → L(R) and π2 : N → L(R) satisfying π1(A) = L∞(X) = π2(A). Then, π1(B) = L(S1) and π2(B) = L(S2) where the Si ⊂ R are subequivalence relations that arise as the kernel of the natural cocycles R → G. By Lemma 5.5, there exists an automorphism γ ∈ Aut(R) satisfying γ(S2) = S1. We still denote by γ the induced automorphism of L(R). Then, θ = π−1 1 ◦ γ ◦ π2 is a ∗-isomorphism θ : N → M satisfying θ(A) = A and θ(B) = B. Defining cg ∈ NB(A) such that θ(ug) = cgug and restricting θ to B, we have found the required cocycle conjugacy between α and β. In order to prove a groupoid version of Theorem 5.4, we first need the following lemma on approximate innerness of a cocycle self-conjugacy. Lemma 5.6. Let B be the hyperfinite II1 factor with Cartan subalgebra A ⊂ B and associated equivalence relation S. Let G be a countable amenable group and α an action of G on A ⊂ B such that the induced action on S is outer. Whenever θ ∈ Aut(A ⊂ B) and cg ∈ NB(A) is a 1-cocycle for the action α satisfying θαgθ−1 = (Ad cg)αg for all g ∈ G, there exists a sequence wn ∈ NB(A) such that Ad wn → θ in Aut(B) and lim n kcg − wnαg(w∗ n)k2 = 0 for all g ∈ G. Proof. Write M = B ⋊αG and denote by (ug)g∈G the canonical unitary operators in this crossed product. We call compatible von Neumann subalgebra of M any von Neumann subalgebra M0 ⊂ M that is generated by an atomic von Neumann subalgebra A0 ⊂ A and finitely many partial isometries v1, . . . , vn ∈ M satisfying the following properties. • The projections pi = viv∗ • We have viA0v∗ • There exist projections pi,g ∈ A0 such that for every i, we have that and vipi,g ∈ Bug for all g ∈ G. i belong to A0 and sum up to 1. We have v∗ i vi = p1 for every i. i = A0pi and viAv∗ i = Api. pi,g = p1 Xg∈G We claim that M can be generated by an increasing sequence of compatible von Neumann subalgebras. To prove this claim, let R be the hyperfinite ergodic type II1 equivalence relation on the standard probability space (X, µ). Choose a ∗-isomorphism π : M → L(R) satisfying π(A) = L∞(X). Then, π(B) = L(S) for an ergodic subequivalence relation S ⊂ R that arises as the kernel of a surjective cocycle R → G. Under the isomorphism π, the compatible von Neumann subalgebras of M correspond to the compatible decompositions of R introduced in the proof of Lemma 5.5. Our claim then follows from claim 1 in that proof. Extend θ to an automorphism of M by defining θ(ug) = cgug. To prove the lemma, we need to prove that there exist elements wn ∈ NB(A) such that Ad wn → θ in Aut(M ). Let M0 ⊂ M be 20 a compatible von Neumann subalgebra generated by A0 ⊂ A and v1, . . . , vn as above. Given the claim above, it suffices to show that there exists an element w ∈ NB(A) such that θ(a) = waw∗ for all a ∈ A0 and θ(vi) = wviw∗ for all i ∈ {1, . . . , n}. First note that θ(A0) is a discrete von Neumann subalgebra of A whose minimal projections have the same trace values as those of A0. Since B is a factor, we can pick w1 ∈ NB(A) such that w∗ 1θ(a)w1 = a for all a ∈ A0. Write θ1 = (Ad w∗ 1) ◦ θ. We claim that w2 = n Xj=1 θ1(vj)v∗ j belongs to NB(A) and satisfies w2a = aw2 for all a ∈ A0 and w2vi = θ1(vi)w2 for all i ∈ {1, . . . , n}. Once this claim is proven, the element w = w1w2 in NB(A) satisfies the required properties. Define the projections qi,g = vipi,gv∗ qi,gvj = 0 if i 6= j. Also, i . By our assumptions, qi,g ∈ A0 and qi,gvi = vipi,g, while n Xi=1 Xg∈G qi,g = 1 . Since θ1 is the identity on A0, it follows that w2qi,g = θ1(vipi,g)(vipi,g)∗. Since vipi,g ∈ Bug and thus also θ1(vipi,g) ∈ Bug, it follows that w2 ∈ B. Since θ1(A) = A and since the vj normalize A, we have for every j ∈ {1, . . . , n} that w2 Apj w∗ j ) = θ1(Apj) = Apj . 2 = θ1(vj)v∗ j Avjθ1(v∗ j ) = θ1(vj) Ap1 θ1(v∗ j ) = θ1(vjAv∗ So, w2 ∈ NB(A). The formula w2a = aw2 for all a ∈ A0 follows because θ1 is the identity on A0 and the vj normalize A0. The formula w2vi = θ1(vi)w2 holds by definition. Theorem 5.7. Let G be an amenable discrete measured groupoid with G(0) = Y . Let (Ay ⊂ By)y∈Y be a measurable field of Cartan subalgebras with each By being isomorphic to the hy- perfinite II1 factor. Let α, β be free actions of G on (Ay ⊂ By)y∈Y . Then, α and β are cocycle conjugate: there exists a measurable field of ∗-automorphisms θy ∈ Aut(By) satisfying θy(Ay) = Ay and a measurable field G ∋ g 7→ cg ∈ NBt(g)(At(g)) such that chk = ch αh(ck) and θt(g) ◦ βg ◦ θ−1 s(g) = (Ad cg) ◦ αg for all g ∈ G and (h, k) ∈ G(2). Proof. As in the proof of Theorem 3.2, we denote by Γy = {g ∈ G s(g) = y = t(g)} the isotropy groups of G and we view G as the semidirect product of the field (Γy)y∈Y of groups and the measurable family of group isomorphisms δ(y,z) : Γz → Γy where (y, z) ∈ R and R is an amenable nonsingular countable equivalence relation on Y . Denote by Py the Polish space of cocycle conjugacies between the actions (αg)g∈Γy and (βg)g∈Γy on Ay ⊂ By. More precisely, Py := (cid:8)(cid:0)θ, (cg)g∈Γy(cid:1) (cid:12)(cid:12) θ ∈ Aut(By) , θ(Ay) = Ay , cg ∈ NBy (Ay), cgh = cgαg(ch) , θβgθ−1 = (Ad cg)αg for all g, h ∈ Γy(cid:9) . The topology on Py is given by the usual topology on Aut(By) and the topology of pointwise k · k2-convergence on the space of 1-cocycles. By Theorem 5.4, Py is nonempty. Writing 21 Gy = NBy (Ay), we have a natural action Gy y Py. By Lemma 5.6, this action has dense orbits. Define the measurable field of continuous group isomorphisms R ∋ (y, z) 7→ α(y,z) : Gz → Gy. Also define the homeomorphisms R ∋ (y, z) 7→ γ(y,z) : Pz → Py given by mapping (θ, (cg)g∈Γz ) ∈ Pz to the automorphism α(z,y)θβ(y,z) of Bz and the 1-cocycle g 7→ α(z,y)(cδ(y,z)(g)). By Theorem 3.5, we find a measurable field of cocycle conjugacies πy = (θy, (cg)g∈Γy ) and a 1-cocycle R ∋ (y, z) 7→ c(y,z) ∈ Gy such that πy = c(y,z) · πz. This precisely means that the (cg)g∈Γy and (c(y,z))(y,z)∈R combine into a 1-cocycle for α, which together with (θy)y∈Y forms the required cocycle conjugacy. 6 Remarks on treeability and cost of equivalence relations The class VC, introduced in [P18], is defined as the class of countable groups Γ satisfying 2- cohomology vanishing for cocycle actions on II1 factors. We still denote by VC the class of discrete measured groupoids G satisfying the conclusion of Theorem 3.1. Similarly, we denote by VCCartan the class of discrete measured groupoids satisfying the conclusion of Theorem 5.2. As in [P18, Remarks 4.5], one may speculate that for a pmp discrete measured groupoid G, we have that G ∈ VC iff G ∈ VCCartan iff G is treeable, in the sense discussed in this section. We elaborate on these speculations below and prove that a groupoid in VCCartan must be treeable (see Proposition 6.3). This also leads us to a new notion of treeability for countable In Proposition 6.4, we relate the pmp equivalence relations, which we call weak treeability. well known open problem whether every treeable countable group Γ is strongly treeable to the question whether every weakly treeable equivalence relation is treeable. Analogously, we introduce a fixed price question for equivalence relations and connect it to the same question for countable groups in Remark 6.6. We first formulate a definition and then explain the required terminology. Definition 6.1. A pmp discrete measured groupoid G is said to be treeable if G admits a free pmp action on a field of standard probability spaces such that the orbit equivalence relation is treeable. More concretely, G is treeable if and only if there exists a countable pmp equivalence relation S on (Y, η) that is treeable, a measure preserving factor map π0 : Y → G(0) and a nonsingular factor map π : S → G satisfying the following properties t(π(x, y)) = π0(x) , π(x, x) = π0(x) for a.e. (x, y) ∈ S, s(π(x, y)) = π0(y) , π(x, y) π(y, z) = π(x, z) π maps S · y × {y} bijectively onto s−1(π0(y)) for a.e. y ∈ Y . for a.e. (x, y), (y, z) ∈ S, Recall that a countable equivalence relation S on (Y, η) is called treeable if there exists a Borel graph T on Y with T ⊂ S and such that for a.e. y ∈ Y , the connected component of y in T is a tree with vertex set S · y. Remark 6.2. Definition 6.1 is entirely analogous to the definition of a treeable group. Nev- ertheless, this definition is potentially confusing: a countable pmp equivalence relation R can also be viewed as a groupoid. The treeability of R as a groupoid, in the sense of Definition 6.1, is in principle weaker than the treeability of R as an equivalence relation. 22 We therefore say that a pmp equivalence relation R is weakly treeable if it is treeable as a groupoid. So the countable pmp equivalence relation R on (X, µ) is weakly treeable if and only if there exists a countable pmp equivalence relation S on (Y, η) that is treeable and a measure preserving factor map π : Y → X mapping S · y bijectively onto R · π(y) for a.e. y ∈ Y . In the same way as the Connes-Jones cocycles of [CJ84] were used in [P18, Section 4] to obtain restrictions on the countable groups in VC, we can prove the following result. Proposition 6.3. If a pmp discrete measured groupoid G belongs to VCCartan, then G is treeable in the sense of Definition 6.1. Proof. Denote by (X, µ) the probability space of units of G. Choose, up to measure zero, countably many subsets Un ⊂ G such that sUn and tUn are bijections from Un onto X and such that G = Sn Un. Denote Γ = F∞ with free generators an ∈ Γ, n ∈ N. Choose a free pmp mixing action Γ y (Y, η), e.g. a Bernoulli action. For every n, define the measure preserving automorphism ϕn ∈ Aut(X, µ) given by ϕn = t ◦ (sUn)−1. Define the action Γ y (X, µ) given by a2n · x = ϕn(x) and a2n+1 · x = x for all n ∈ N, x ∈ X . Then consider the diagonal action Γ y X × Y and denote by S the orbit equivalence relation. Denote by π : S → G the unique groupoid morphism satisfying π(cid:0)(x, y), (x, y)(cid:1) = x , π(cid:0)a2n · (x, y), (x, y)(cid:1) = (sUn)−1(x) and π(cid:0)a2n+1 · (x, y), (x, y)(cid:1) = x . The kernel of π is given by a field of orbit equivalence relations (cid:0)R(Λx y Y )(cid:1)x∈X, where a2n+1 ∈ Λx for all x ∈ X, n ∈ N. Then π induces a cocycle action of G on this field of ergodic type II1 equivalence relations. Since G ∈ VCCartan, this cocycle action can be chosen to be a genuine action. This means that we find a Borel family of measure preserving automorphisms G ∋ g 7→ θg ∈ Aut(Y, η) satisfying the following properties. θg ◦ θh = θgh , (t(g), θg(y)) ∼S (s(g), y) , π(cid:0)(t(g), θg(y)), (s(g), y)(cid:1) = g for all (g, h) ∈ G(2) and y ∈ Y . It follows that S0 := (cid:8)(cid:0)(t(g), θg(y)), (s(g), y)(cid:1) g ∈ G(cid:9) is a subequivalence relation of S. Since S is treeable, it follows from [G99, Th´eor`eme 5] that also S0 is treeable. The restriction of π maps S0 onto G and maps S · (x, y) × {(x, y)} bijectively onto s−1(x). By definition, G is treeable. Recall that a countable group Γ is called strongly treeable if for every free pmp action Γ y (X, µ), the orbit equivalence relation is treeable. It is a wide open problem whether every treeable group is strongly treeable. Proposition 6.4. The following two statements are equivalent (see Remark 6.2 for a disam- biguation of terminology). 1. Every treeable countable group is strongly treeable. 2. If a countable pmp equivalence relation R is weakly treeable and is the orbit relation of a free action of a countable group, then R is treeable. 23 If one believes that there exist treeable groups that are not strongly treeable, it could therefore be easier to first find weakly treeable, but non treeable equivalence relations. Proof. 1 ⇒ 2. Assume that R = R(Γ y X) for some free pmp action Γ y (X, µ) and assume that R is weakly treeable. Choose a pmp action of R such that its orbit equivalence relation S on (Y, η) is treeable. Denote by π : Y → X the corresponding measure preserving factor map that maps S-orbits bijectively onto R-orbits. Since π is bijective on orbits, there is a unique action Γ y Y such that g· y ∈ S · y for all g ∈ Γ, y ∈ Y and π(g· y) = g· π(y). By construction, the action Γ y (Y, η) is free and pmp, and S is its orbit equivalence relation. Since S is a treeable equivalence relation, it follows that Γ is a treeable group. Since 1 holds, Γ is strongly treeable and it follows that also R is a treeable equivalence relation. 2 ⇒ 1. Let Γ be a countable group and let Γ y (X, µ), Γ y (Y, η) be free pmp actions. Assume that the orbit equivalence relation R(Γ y X) is treeable. We have to prove that also R(Γ y Y ) is a treeable equivalence relation. Using the factor map X × Y → X, the treeing of R(Γ y X) can be lifted to a treeing of the orbit equivalence relation R(Γ y X × Y ) for the diagonal action. Using the factor map X × Y → Y , we conclude that R(Γ y Y ) is weakly treeable. Since 2 holds, R(Γ y Y ) is treeable. In the proofs of Theorems 3.1, 5.2 and 5.7, we used the following result: if G is a discrete mea- sured groupoid such that the associated countable equivalence relation R on G(0) is amenable, then G can be written as the semidirect product of R acting on the field of isotropy groups of G. Also this is a 2-cohomology vanishing result, for cocycle actions on fields of countable groups. We finally prove that the same 2-cohomology holds for arbitrary treeable equivalence relations and is a characterization of treeability. Proposition 6.5. Let R be a countable nonsingular equivalence relation on the standard prob- ability space (X, µ). The following two statements are equivalent. 1. R is treeable. 2. Every discrete measured groupoid G with G(0) = X and with R = {(t(g), s(g)) g ∈ G} can be written as a semidirect product of R acting on the field of isotropy groups of G. Proof. 1 ⇒ 2. Since R is treeable, we can decompose R as a free product of amenable sub- equivalence relations. On each of these, the quotient map G → R can be lifted as a groupoid homomorphism and these lifts can be combined into a well defined groupoid homomorphism R → G by freeness. 2 ⇒ 1. Denote Γ = F∞ and choose a nonsingular action Γ y (X, µ) such that R = {(g · x, x) g ∈ Γ, x ∈ X}. Define the action groupoid G = Γ× X with G(0) = X, s(g, x) = x, t(g, x) = g · x and (g, h · x)· (h, x) = (gh, x). Since 2 holds, there is a groupoid homomorphism R → G lifting the quotient map G → R. So there exists a cocycle ω : R → Γ such that ω(x, y) · y = x for a.e. (x, y) ∈ R. Removing from X a Borel set of measure zero, we may assume that ω is a Borel cocycle and that ω(x, y) · y = x for all (x, y) ∈ R. Define on X × Γ the countable Borel equivalence relation S given by (x, g) ∼S (y, h) if and only if x ∼R y. Define the subequivalence relation T ⊂ S given by (x, g) ∼T (y, h) if and only if x ∼R y and g = ω(x, y)h. Since the restriction of T to each of the subsets X × {g} is the trivial equivalence relation, it follows that T admits a Borel fundamental domain Y ⊂ X × Γ. Define the Borel map π : X × Γ → Y such that π(x, g) ∼T (x, g). Then define the action of Γ on Y given by h · (x, g) = π(x, gh−1). By construction, this action is free and its orbit 24 equivalence relation coincides with SY . So, SY is a treeable Borel equivalence relation. By [JKL01, Proposition 3.3] (see also [G99, Th´eor`eme 5 and Proposition 2.6] in the pmp setting), it first follows that S is treeable and then that the restriction of S to X × {e} is treeable. The latter being equal to R, the proposition is proven. Remark 6.6. Recall that a countable group Γ is said to have fixed price if all orbit equivalence relations for free pmp actions of Γ have the same cost, in the sense of [G99]. The well known fixed price problem asks whether all countable groups have fixed price. Let R be a countable pmp equivalence relation on (X, µ). Whenever S on (Y, η) is the orbit relation for a free pmp action of R (equivalently, there is a measure preserving factor map Y → X that is bijective on a.e. orbit), any graphing for R lifts to a graphing of S, so that cost(S) ≤ cost(R). We can therefore ask the following fixed price question for the equivalence relation R: do we have cost(S) = cost(R) for each such orbit relation S ? With precisely the same argument as in the proof of Proposition 6.4, we get that the following two statements are equivalent. 1. Every countable group has fixed price. 2. Every countable pmp equivalence relation that is the orbit relation of a free action of a countable group, has fixed price. So exactly as above, if one believes that there exist countable groups that do not have fixed price, it might be easier to first find countable equivalence relations that do not have fixed price. Note here as well that in the above context, the L2-Betti numbers of S and R coincide. Indeed, viewing S as the orbit relation of a free pmp action of R, one may repeat, mutatis mutandis, the proofs of [G01, Corollaire 3.16] or [S03, Theorem 5.5] (see also [PSV15, Remark 9.24] for an alternative proof in terms of Cartan inclusions). So a counterexample for the fixed price question for equivalence relations would also provide a counterexample for the equally 1 (R) + 1 for every ergodic countable pmp wide open problem asking whether cost(R) = β(2) equivalence relation with infinite orbits. 7 (Non-)classification of discrete amenable groupoids Theorems 3.2 and 5.3 provide a complete classification of the regular von Neumann subalgebras B of the hyperfinite II1 factor R satisfying B′ ∩ R = Z(B) in terms of the associated discrete measured groupoid G. Every amenable discrete measured groupoid is the semi-direct product of a measurable field of amenable groups and an action of a countable amenable equivalence relation. Although there is a unique amenable ergodic equivalence relation of type II1, the classification of amenable discrete groupoids is strictly more complex than the classification of amenable groups, as we show in the following examples. All this is meant in the descriptive set theoretic sense of the word (see e.g. [H00]). 7.1 Measurable fields of amenable groups Let X0 be the uncountable standard Borel space and assume that we are given a Borel family (Γx)x∈X0 of amenable groups with the property that Γx 6∼= Γy if x 6= y. More concretely, one 25 could take X0 = {0, 1}P , where P denotes the set of prime numbers, and define for x ∈ X0, Γx = Mp∈P,xp=1 Z/pZ . For every probability measure µ on X0, we define G(µ) as the direct integral w.r.t. µ of the amenable groups (Γx)x∈X0. By construction, G(µ1) ∼= G(µ2) if and only if µ1 and µ2 are absolutely continuous. By [KS99, Theorem 2.1], the nonatomic probability measures on X0 (up to absolute continuity) are not classifiable by countable structures. Countable amenable groups form, by definition, a countable structure. In this sense, already these trivial examples show that the classification of amenable discrete measured groupoids is strictly harder than the classification of amenable groups. 7.2 Ergodic amenable discrete groupoids In order to arise from regular inclusions B ⊂ R with R being the hyperfinite II1 factor, we need to consider ergodic groupoids G. Unclassifiably many ergodic amenable discrete groupoids can be constructed as follows. Again take X0 = {0, 1}P , where P denotes the set of prime numbers. We now define, for x ∈ X0, the subgroup Λx ⊂ Q given by Λx = hp−k p ∈ P, xp = 1, k ∈ Ni . One checks that for x 6= y, every group homomorphism Λx → Λy is trivial. Then define X = X Z 0 and for every x ∈ X, put Γx = Mn∈Z Λxn . Consider the Bernoulli action Z y X given by (k · x)n = xn−k and denote by δk,x : Γx → Γk·x, the natural group isomorphism. We then define the Borel groupoid G0 as the union of all (Γx)x∈X with unit space X, and we define G as the semidirect product of G0 and Z, acting by the automorphisms that we just defined. For every probability measure µ on X0, the product measure µZ on X is Z-invariant, so that we obtain the ergodic, pmp, amenable, discrete measured groupoid G(µ). We claim that again G(µ1) ∼= G(µ2) if and only if the measures µ1 and µ2 are absolutely continuous. So also the discrete measured groupoids that are ergodic, pmp and amenable are unclassifiable by countable structures. To prove the claim, it suffices to show the following statement: if U ⊂ X0 is a Borel set such that µ1(U ) > 0 and µ2(U ) = 0, then for µZ 1 -a.e. x ∈ X and µZ 1 -a.e. x ∈ X, we have that there exists an n ∈ Z with xn ∈ U , so that one of the groups Λa, a ∈ U , is a quotient of Γx. For µZ 2 -a.e. y ∈ X, we have that yn 6∈ U for all n ∈ Z, so that none of the groups Λa, a ∈ U , is a quotient of Γy. Then the claim is proven. 2 -a.e. y ∈ X, we have that Γx 6∼= Γy. Fix such a Borel set U ⊂ X0. For µZ 7.3 Amenable discrete groupoids with constant isotropy groups In all the examples so far, we obtain "many" groupoids by exploiting that there are "many" amenable groups to choose from as isotropy groups Γx. But even the ergodic pmp amenable discrete groupoids G with constant isotropy groups Γx ∼= H for a fixed amenable group H can be unclassifiable. This however depends on the choice of H, as we prove in Corollary 7.2: when H = Z, there are only countably many such groupoids G up to isomorphism, while for H = Z2, 26 we encode ergodic transformations of the interval [0, 1] up to conjugacy into such amenable groupoids up to isomorphism. Fix a countable nonsingular equivalence relation R on the standard probability space (X, µ) and fix a countable group H. For every 1-cocycle δ : R → Aut(H) : (x, y) 7→ δ(x,y) , define the groupoid Gδ = H × R with G(0) δ = X and s(g, x, y) = y , t(g, x, y) = x and (g, x, y) · (h, y, z) = (gδ(x,y)(h), x, z) . Note that Gδ is amenable if and only if both H and R are amenable. Conversely, every discrete measured groupoid G with the properties that the isotropy groups Γx are isomorphic with a fixed group H for a.e. x ∈ X = G(0) and that the associated equivalence relation on (X, µ) is amenable, is isomorphic with Gδ for some 1-cocycle δ : R → Aut(H). As in [BG90, Definition 1.2], given a Polish group G, the 1-cocycles δ, δ′ : R → G are called weakly equivalent if there exists a ∆ ∈ Aut(R) such that δ is cohomologous with δ′ ◦ ∆. More concretely, ∆ and ∆′ are weakly equivalent if and only if there exists a nonsingular automorphism ∆ ∈ Aut(X, µ) and a Borel map ϕ : X → G such that ∆(R · x) = R · ∆(x) for a.e. x ∈ X, and δ′(∆(x), ∆(y)) = ϕ(x) δ(x, y) ϕ(y)−1 for a.e. (x, y) ∈ R. Proposition 7.1. Let R be an amenable countable nonsingular equivalence relation and let H be a countable group. Define the Polish group G = Aut(H)/Inn(H) and denote by p : Aut(H) → G the quotient map. 1. Let δ, δ′ : R → Aut(H) be 1-cocycles. Then Gδ is isomorphic with Gδ′ if and only if the G-valued 1-cocycles p ◦ δ and p ◦ δ′ are weakly equivalent. 2. For every 1-cocycle δ0 : R → G, there exists a 1-cocycle δ : R → Aut(H) satisfying δ0 = p◦δ. Proof. 1. A direct computation gives that Gδ ∼= Gδ′ if and only if there exists a ∆ ∈ Aut(R) and Borel maps ω : R → H : (x, y) 7→ ω(x, y) and θ : X → Aut(H) : x 7→ θx such that ω(x, z) = ω(x, y) δ(x,y)(ω(y, z)) Ad ω(x, y) ◦ δ(x,y) = θx ◦ δ′ (∆(x),∆(y)) ◦ θ−1 y for a.e. (x, y), (y, z) ∈ R, for a.e. (x, y) ∈ R, (7.1) (7.2) and with the isomorphism θ : Gδ′ → Gδ given by θ(g, ∆(x), ∆(y)) = (θx(g) ω(x, y), x, y). So if Gδ ∼= Gδ′, we immediately get that p ◦ δ and p ◦ δ′ are weakly equivalent. Conversely, assume that p ◦ δ and p ◦ δ′ are weakly equivalent. Since R is amenable, we can write, up to measure zero, X = X1⊔ X2 so that the restriction of R to X1 is of type I, while the restriction of R to X2 is the orbit relation of a free action of Z implemented by T ∈ Aut(X2, µ). Every 1-cocycle for RX1 is cohomologous to the trivial 1-cocycle, so that the restrictions of Gδ and Gδ′ to X1 are both isomorphic with the direct product of H and RX1. We may thus assume that X = X2. 27 Since p ◦ δ and p ◦ δ′ are weakly equivalent, using a Borel lift G → Aut(H) for the quotient homomorphism p, we find an automorphism ∆ ∈ Aut(R) and a Borel map θ : X → Aut(H) : x 7→ θx such that the 1-cocycle δ′′ defined by δ′′(x, y) = θx ◦ δ′ (∆(x),∆(y)) ◦ θ−1 y satisfies p ◦ δ = p ◦ δ′′. By [JT82, Appendix], because R is amenable, we can modify θ and choose a Borel map η : X → H such that Ad η(x) ◦ δ(T x,x) = θT x ◦ δ′ (∆(T x),∆(x)) ◦ θ−1 x (7.3) for a.e. x ∈ X. Uniquely define ω : R → H such that (7.1) holds and ω(T x, x) = η(x) for all x ∈ X. Then (7.2) follows from (7.3) and we obtain that Gδ ∼= Gδ′. 2. Let δ0 : R → G be a 1-cocycle. Also here, we may assume that R is the orbit relation of a free action of Z implemented by T ∈ Aut(X, µ). Using a Borel lift G → Aut(H), choose a Borel function η : X → Aut(H) such that p(η(x)) = δ0(T x, x) for a.e. x ∈ X. Define δ : R → Aut(H) as the unique 1-cocycle satisfying δ(T x, x) = η(x) for all x ∈ X. It follows that p ◦ δ = δ0. To every 1-cocycle δ : R → G with values in a lcsc group G is associated the Mackey action, defined as follows. We equip G with its left or right Haar measure and consider on X × G the equivalence relation Rδ given by (x, g) ∼Rδ (y, h) if and only if x ∼R y and g = δ(x, y)h. The group G acts by automorphisms of Rδ given by g·(x, h) = (x, hg−1). We define the nonsingular action G y (Y, η) in such a way that L∞(Y ) can be identified with the Rδ-invariant functions in L∞(X × G). If R is ergodic, the action G y (Y, η) is ergodic. It is called the Mackey action of the 1-cocycle δ. Clearly, weakly equivalent 1-cocycles give rise to conjugate Mackey actions. When R is amenable and pmp, by [GS91, Theorem 3.1], the converse holds for recurrent 1- cocycles. Here, δ is called recurrent if Rδ is not of type I. In this way, we get the following corollary to Proposition 7.1. Corollary 7.2. Fix a countable group H. Consider all discrete measured groupoids G such that almost every isotropy group is isomorphic with H and such that the associated equiva- lence relation R is the unique ergodic hyperfinite type II1 equivalence relation. Write G = Aut(H)/Inn(H). 1. If G is a compact group, the groupoids G are concretely classifiable in terms of the space of closed subgroups of G up to conjugacy. 2. If G is discrete and contains a copy of Z, the groupoids G are not classifiable by countable structures: to every weakly mixing measure preserving transformation T ∈ Aut([0, 1]), we can associate such a groupoid GT such that GT ∼= GS if and only if the transformations T, S are (flip) conjugate. Note that we are in the first situation when H = Z or when H is a finite group, while we are in the second situation when H = Zn with n ≥ 2. Proof. 1. By Proposition 7.1, the classification of the groupoids G is equivalent with the clas- sification of the 1-cocycles δ : R → G up to weak equivalence. The Mackey action of such a 1-cocycle is an ergodic action of the compact group G and thus of the form G y G/K for some closed subgroup K < G. Two such actions are conjugate if and only if the corresponding closed subgroups are conjugate. Using the canonical K-valued 1-cocycle for the action Z y K Z/K, where K is sitting diagonally in K Z and Z is acting as a shift, every action G y G/K arises 28 as the Mackey action of a 1-cocycle. Finally, for every 1-cocycle δ : R → G, the equivalence relation Rδ preserves a probability measure, so that every 1-cocycle δ : R → G is recurrent. By the discussion above, using [GS91, Theorem 3.1], the result follows. 2. Denote by p : Aut(H) → G the quotient homomorphism. Choose α ∈ Aut(H) such that a := p(α) is an element of infinite order in G. Fix a standard nonatomic probability space (X, µ). For every weakly mixing pmp transformation T ∈ Aut(X, µ), consider the orbit equivalence relation RT ∼= R of the associated Z-action and consider the 1-cocycle δT : RT → Aut(H) given by δT (T nx, x) = αn for all x ∈ X, n ∈ Z. Denote by GT the associated groupoid. We distinguish two cases. a. There exists an element g ∈ G such that gag−1 = a−1. b. There is no element g ∈ G such that gag−1 = a−1. We prove that in case a, the groupoids GT and GS, for weakly mixing pmp transformations T, S ∈ Aut(X, µ), are isomorphic if and only if there exists a ∆ ∈ Aut(X, µ) such that ∆T ∆−1 = S±1. In case b, we prove that GT ∼= GS if and only if there exists a ∆ ∈ Aut(X, µ) such that ∆T ∆−1 = S. When T and S are conjugate, it is immediate that GT ∼= GS. When ∆T ∆−1 = S−1 and gag−1 = a−1, it follows that ∆ is an isomorphism of RT onto RS and that g (p ◦ δS ◦ ∆)(x, y) g−1 = (p ◦ δT )(x, y) for all (x, y) ∈ RT . So, the 1-cocycles p ◦ δS and p ◦ δT are weakly equivalent. It follows from Proposition 7.1 that GT ∼= GS. Conversely, assume that GT ∼= GS. By Proposition 7.1, the 1-cocycles p ◦ δS and p ◦ δT are weakly equivalent. So, their Mackey actions are conjugate. These Mackey actions are precisely the induced actions G y G×S X, resp. G y G×T X, of Z y (X, µ) given by S, resp. T , where Z is identified with the subgroup of G generated by a. Denote by ∆ : G ×S X → G ×T X the conjugacy of the Mackey actions. Since S and T are weakly mixing transformations, we must have ∆({e} × X) = {g} × X with g−1aZg = aZ. If g−1ag = a, it follows that S and T are conjugate. If g−1ag = a−1, it follows that S and T −1 are conjugate. By [FW03] (see also [K10, Theorem 5.7.b]), the classification of weakly mixing pmp transfor- mations up to (flip) conjugacy is not classifiable by countable structures. So the groupoids G are not classifiable by countable structures. 7.4 Exotic examples Let R be the unique hyperfinite equivalence relation of type II1. In this section, we construct for every ergodic subequivalence relation R0 ⊂ R an ergodic pmp amenable discrete groupoid G(R0) and prove that G(R0) ∼= G(R1) if and only if there exists ϕ ∈ [R] such that (ϕ×ϕ)(R0) = R1. So, the classification of this family G(R0) of amenable discrete measured groupoids is equivalent with the classification, up to conjugacy by an element of the full group, of all ergodic subequivalence relations of the hyperfinite type II1 ergodic equivalence relation R. Put X = {0, 1}P and denote by µ the product of the uniform probability measure on {0, 1}. View X as the space of subsets of the set P of prime numbers. For every Q ∈ X, define the subgroup ΓQ ⊂ Q given by ΓQ = hp−1 p ∈ Qi . 29 By construction, ΓQ ∼= ΓQ′ if and only if Q △ Q′ < ∞. Realize R as the countable pmp equivalence relation on (X, µ) consisting of all (Q,Q′) with Q △ Q′ < ∞. We construct for every (Q,Q′) ∈ R a canonical isomorphism δ(Q,Q′) : ΓQ′ → ΓQ. Define the element q(Q,Q′) ∈ Q∗ given by q(Q,Q′) = (cid:16) Yp∈Q\Q′ p−1(cid:17) · (cid:16) Yp∈Q′\Q p(cid:17) . Then define δ(Q,Q′) given by multiplication with q(Q,Q′). One checks that for all (Q,Q′), (Q′,Q′′) ∈ R. δ(Q,Q′′) = δ(Q,Q′) ◦ δ(Q′,Q′′) So we can define the discrete measured groupoid G as the semidirect product of the field (ΓQ)Q∈X and the equivalence relation R, with the invariant probability measure µ. For every ergodic subequivalence relation R0 ⊂ R, we denote by G(R0) the subgroupoid of G given by only taking the semidirect product with R0. Since ΓQ ∼= ΓQ′ if and only if (Q,Q′) ∈ R, the family of groupoids G(R0) has the properties mentioned above. References [AR00] [B90] [BG84] [BG90] C. Anantharaman-Delaroche and J. Renault, Amenable groupoids. Monographies de L'En- seignement Math´ematique 36, L'Enseignement Math´ematique, Geneva, 2000. E. Bedos, Actions of amenable groups on II1 factors. J. Funct. Anal. 91 (1990), 404-414. S.I. Bezuglyı and V.Ya. Golodets, Outer conjugacy of actions of countable amenable groups on a space with measure. Izv. Akad. Nauk SSSR Ser. Mat. 50 (1986), 643-660. S.I. Bezuglyı and V.Ya. Golodets, Weak equivalence and the structures of cocycles of an ergodic automorphism. Publ. Res. Inst. Math. Sci. 27 (1991), 577-625. A. Connes, Classification of injective factors. Ann. of Math. 104 (1976), 73-115. [C75] [CFW81] A. Connes, J. Feldman and B. Weiss, An amenable equivalence relation is generated by a [CJ84] single transformation. Ergodic Theory Dynamical Systems 1 (1981), 431-450. A. Connes and V. Jones, Property T for von Neumann algebras. Bull. London Math. Soc. 17 (1985), 57-62. [DFP18] A.P. Donsig, A.H. Fuller and D.R. Pitts, Cartan triples. Preprint. arXiv:1810.05267 [FM75] J. Feldman and C.C. Moore, Ergodic equivalence relations, cohomology, and von Neumann algebras, I, II. Trans. Amer. Math. Soc. 234 (1977), 289-324, 325-359. J. Feldman, C.E. Sutherland and R.J. Zimmer, Subrelations of ergodic equivalence relations. Ergodic Theory Dynam. Systems 9 (1989), 239-269. [FSZ88] [FW03] M. Foreman and B. Weiss, An anti-classification theorem for ergodic measure preserving [G99] [G01] [GS86] [GS91] [H00] [JKL01] [J81] transformations. J. Eur. Math. Soc. (JEMS) 6 (2004), 277-292. D. Gaboriau, Cout des relations d'´equivalence et des groupes. Invent. Math. 139 (2000), 41-98. D. Gaboriau, Invariants ℓ2 de relations d'´equivalence et de groupes. Publ. Math. Inst. Hautes ´Etudes Sci. 95 (2002), 93-150. V.Ya. Golodets and S.D. Sinelshchikov, Outer conjugacy for actions of continuous amenable groups. Publ. Res. Inst. Math. Sci. 23 (1987), 737-769. V.Ya. Golodets and S.D. Sinelshchikov, Classification and structure of cocycles of amenable ergodic equivalence relations. J. Funct. Anal. 121 (1994), 455-485. G. Hjorth, Classification and orbit equivalence relations. Mathematical Surveys and Mono- graphs 75, American Mathematical Society, Providence, 2000. S. Jackson, A.S. Kechris and A. Louveau, Countable Borel equivalence relations. J. Math. Log. 2 (2002), 1-80. V.F.R. Jones, A converse to Ocneanu's theorem. J. Operator Theory 10 (1983), 61-64. 30 [JT82] [K95] [K10] [KL16] [KS99] [M63] [O85] [P92] [P01] [P13] [P18] V.F.R. Jones and M. Takesaki, Actions of compact abelian groups on semifinite injective factors. Acta Math. 153 (1984), 213-258. A.S. Kechris, Classical descriptive set theory. Graduate Texts in Mathematics 156, Springer- Verlag, 1995. A.S. Kechris, Global aspects of ergodic group actions. Mathematical Surveys and Monographs 160, American Mathematical Society, Providence, 2010. D. Kerr and H. Li, Ergodic theory. Independence and dichotomies. Springer Monographs in Mathematics, Springer, Cham, 2016. A.S. Kechris and N.E. Sofronidis, A strong generic ergodicity property of unitary and self- adjoint operators. Ergodic Theory Dynam. Systems 21 (2001), 1459-1479. G.W. Mackey, Ergodic theory, group theory, and differential geometry. Proc. Nat. Acad. Sci. U.S.A. 50 (1963), 1184-1191. A. Ocneanu, Actions of discrete amenable groups on factors. Springer Lecture Notes 1138, Berlin-Heidelberg-New York, 1985. S. Popa, Free independent sequences in type II1 factors and related problems. Ast´erisque 232 (1995), 187-202. S. Popa, Some rigidity results for non-commutative Bernoulli shifts. J. Funct. Anal. 230 (2006), 273-328. S. Popa, Independence properties in subalgebras of ultraproduct II1 factors. J. Funct. Anal. 266 (2014), 5818-5846. S. Popa, On the vanishing cohomology problem for cocycle actions of groups on II1 factors. To appear in Ann. Sci. ´Ec. Norm. Sup´er. arXiv:1802.09964 [PSV15] S. Popa, D. Shlyakhtenko and S. Vaes, Cohomology and L2-Betti numbers for subfactors and [S03] [S81] [S03] [S85] [ST84] quasi-regular inclusions. Int. Math. Res. Not. IMRN 8 (2018), 2241-2331. R. Sauer, L2-Betti numbers of discrete measured groupoids. Internat. J. Algebra Comput. 15 (2005), 1169-1188. K. Schmidt, Amenability, Kahzdan's property T, strong ergodicity and invariant means for ergodic group actions. Ergodic Theory Dynam. Systems 1 (1981), 223-236. Y. Shalom, Harmonic analysis, cohomology, and the large-scale geometry of amenable groups. Acta Math. 192 (2004), 119-185. C.E. Sutherland, A Borel parametrization of Polish groups. Publ. Res. Inst. Math. Sci. 21 (1985), 1067-1086. C.E. Sutherland and M. Takesaki, Actions of discrete amenable groups and groupoids on von Neumann algebras. Publ. Res. Inst. Math. Sci. 21 (1985), 1087-1120. 31
1302.5593	1	1302	2013-02-22T13:56:39	Affine buildings, tiling systems and higher rank Cuntz-Krieger algebras	[ "math.OA" ]	To an $r$-dimensional subshift of finite type satisfying certain special properties we associate a $C^*$-algebra $\cA$. This algebra is a higher rank version of a Cuntz-Krieger algebra. In particular, it is simple, purely infinite and nuclear. We study an example: if $\G$ is a group acting freely on the vertices of an $\wt A_2$ building, with finitely many orbits, and if $\Omega$ is the boundary of that building, then $C(\Om)\rtimes \G$ is the algebra associated to a certain two dimensional subshift.	math.OA	math	AFFINE BUILDINGS, TILING SYSTEMS AND HIGHER RANK CUNTZ-KRIEGER ALGEBRAS GUYAN ROBERTSON AND TIM STEGER Abstract. To an r-dimensional subshift of finite type satisfying certain special properties we associate a C ∗-algebra A. This algebra is a higher rank version of a Cuntz-Krieger algebra. In particular, it is simple, purely infinite and nuclear. We study an example: if Γ is a group acting freely on the vertices of an eA2 building, with finitely many orbits, and if Ω is the boundary of that building, then C(Ω) ⋊ Γ is the algebra associated to a certain two dimensional subshift. Introduction This paper falls into two parts. The self-contained first part develops the theory of a class of C ∗-algebras which are higher rank generalizations of the Cuntz-Krieger algebras [CK, C1, C2]. We start with a set of r-dimensional words, based on an alphabet A, we define transition matrices Mj in each of r directions, satisfying certain conditions (H0)-(H3). The C ∗-algebra A is then the unique C ∗-algebra generated by a family of partial isometries su,v indexed by compatible r- dimensional words u, v and satisfying relations (0.1) below. If r = 1 then A is a Cuntz -- Krieger algebra. We prove that the algebra A is simple, purely infinite and stably isomorphic to the crossed product of an AF-algebra by a Zr-action. The last part of the paper (Section 7) studies in detail one particularly interesting example. This example was the authors' motivation for introducing these algebras. Let B be an affine building of type eA2. Let Γ be a group of type rotating automorphisms of B which acts freely on the vertex set with finitely many orbits. There is a natural action of Γ on the boundary Ω of B, and we can form the universal crossed product algebra C(Ω) ⋊ Γ. This algebra is isomorphic to an algebra of the form A obtained by the preceding construction. In the case where Γ also acts transitively on the vertices of B the algebra C(Ω) ⋊ Γ was previously studied in [RS1], where simplicity was proved. In a sequel to this paper [RS2] we explicitly compute the K-theory of some of these algebras. In [Sp] J. Spielberg treated an analogous example in rank 1: Γ is a free group, the building is a tree, and C(Ω) ⋊ Γ is isomorphic to an ordinary Cuntz-Krieger algebra. In fact [Sp] deals more generally with the case where Γ is a free product of cyclic groups. The generalizations in this paper are motivated by Spielberg's work. We now introduce some basic notation and terminology. Let Z+ denote the set of nonnegative integers. Let [m, n] denote {m, m + 1, . . . , n}, where m ≤ n are integers. If m, n ∈ Zr, say that m ≤ n if mj ≤ nj for 1 ≤ j ≤ r, and when m ≤ n, let [m, n] = [m1, n1] × · · · × [mr, nr]. In Zr, let 0 denote the zero vector and let ej denote the jth standard unit basis vector. We fix a finite set A (an "alphabet"). A {0, 1}-matrix is a matrix with entries in {0, 1}. Choose nonzero {0, 1}-matrices M1, M2, . . . , Mr and denote their elements by Mj(b, a) ∈ {0, 1} for a, b ∈ A. If m, n ∈ Zr with m ≤ n, let W[m,n] = {w : [m, n] → A; Mj(w(l + ej), w(l)) = 1 whenever l, l + ej ∈ [m, n]}. Date: February 2, 1999. 1991 Mathematics Subject Classification. Primary 46L35; secondary 46L55, 22D25, 51E24. Key words and phrases. C ∗-algebra, subshift of finite type, affine building. This research was supported by the Australian Research Council. 1 Typeset by AMS-LATEX. 2 GUYAN ROBERTSON AND TIM STEGER Put Wm = W[0,m] if m ≥ 0. Say that an element w ∈ Wm has shape m, and write σ(w) = m. Thus Wm is the set of words of shape m, and we identify A with W0 in the natural way. Define the initial and final maps o : Wm → A and t : Wm → A by o(w) = w(0) and t(w) = w(m). Fix a nonempty finite or countable set D (whose elements are "decorations"), and a map δ : D → A. Let W m = {(d, w) ∈ D ×Wm; o(w) = δ(d)}, the set of "decorated words" of shape m, and identify D with W 0 via the map d 7→ (d, δ(d)). Let W =Sm Wm and W =Sm W m, the sets of all words and all decorated words respectively. Define o : W m → D and t : W m → A by o(d, w) = d and t(d, w) = t(w). Likewise extend the definition of shape to W by setting σ((d, w)) = σ(w). Given j ≤ k ≤ l ≤ m and a function w : [j, m] → A, define w[k,l] ∈ Wl−k by w[k,l] = w′ where w′(i) = w(i + k) for 0 ≤ i ≤ l − k. If w = (d, w) ∈ W m, define w[k,l] = w[k,l] ∈ Wl−k if k 6= 0, and w[0,l] = (d, w[0,l]) ∈ W l. If w ∈ Wl and k ∈ Zr, define τkw : [k, k + l] → A by (τkw)(k + j) = w(j). If w ∈ Wl where l ≥ 0 and if p 6= 0, say that w is p-periodic if its p-translate, τpw, satisfies τpw[0,l]∩[p,p+l] = w[0,l]∩[p,p+l]. Assume that the matrices Mi have been chosen so that the following conditions hold. (H0): Each Mi is a nonzero {0, 1}-matrix. (H1): Let u ∈ Wm and v ∈ Wn. If t(u) = o(v) then there exists a unique w ∈ Wm+n such that w[0,m] = u and w[m,m+n] = v. (H2): Consider the directed graph which has a vertex for each a ∈ A and a directed edge from a to b for each i such that Mi(b, a) = 1. This graph is irreducible. (H3): Let p ∈ Zr, p 6= 0. There exists some w ∈ W which is not p-periodic. Definition 0.1. In the situation of (H1) we write w = uv and say that the product uv exists. This product is clearly associative. The C ∗-algebra A is defined as the universal C ∗-algebra generated by a family of partial isome- tries {su,v; u, v ∈ W and t(u) = t(v)} satisfying the relations (0.1a) (0.1b) (0.1c) (0.1d) su,v ∗ = sv,u su,vsv,w = su,w su,v = Xw∈W ;σ(w)=ej, o(w)=t(u)=t(v) suw,vw, for 1 ≤ j ≤ r su,usv,v = 0, for u, v ∈ W 0, u 6= v. 1. Products of higher rank words Condition (H1) is fundamental to all that follows. How then, does one verify (H1)? Given {0, 1}-matrices Mi, 1 ≤ i ≤ r, the following three simple conditions will be seen to be sufficient. (H1a): MiMj = MjMi. (H1b): For i < j, MiMj is a {0, 1}-matrix. (H1c): For i < j < k, MiMjMk is a {0, 1}-matrix. Indeed, the first two conditions are also necessary. Lemma 1.1. Fix {0, 1}-matrices Mi, 1 ≤ i ≤ r. Then (H1) implies (H1a) and (H1b). Proof. Suppose (MiMj)(b, a) > 0. Then there exists c ∈ A so that Mj(c, a) = 1 = Mi(b, c). Let u ∈ Wej and v ∈ Wei be given by u(0) = a u(ej) = c v(0) = c v(ei) = b. 3 According to (H1) there is a unique w ∈ Wei+ej with w(0) = a, w(ej) = c, w(ei + ej) = b. There must then be a unique d ∈ A which can be used for the missing value of w, w(ei). That is, there must be a unique d ∈ A satisfying Mi(d, a) = 1 = Mj(b, d). Hence (MjMi)(b, a) = 1. We have seen that if (MiMj)(b, a) > 0, then (MjMi)(b, a) = 1. Likewise, if (MjMi)(b, a) > 0, then (MiMj)(b, a) = 1. It follows that MiMj and MjMi are equal and have entries in {0, 1}. (cid:3) Lemma 1.2. Fix {0, 1}-matrices Mi satisfying (H1a),(H1b), and (H1c). Let 1 ≤ j ≤ r. Let w ∈ Wm and choose a ∈ A so that Mj(a, t(w)) = 1. Then there exists a unique word v ∈ Wm+ej such that v[0,m] = w and t(v) = a. Proof. In the case r = 2 this follows from conditions (H1a) and (H1b) alone. The situation is illustrated in Figure 1, for j = 2. The assertion is that there is a unique word v defined on the outer rectangle [0, m + e2] with final letter v(m + e2) = a. The hypothesis is that there is a transition from w(m) to a, in the sense that M2(a, w(m)) = 1. Define v(m + e2) = a. For notational convenience, let n = m − e1. We have M1(w(m), w(n)) = 1, and the product matrix M2M1 defines a transition w(n) → w(m) → a. The conditions (H1a) and (H1b) assert that the product M1M2 defines a unique transition w(n) → b → a, for some b ∈ A. Define v(n + e2) = b. Continue the process inductively until v is defined uniquely on the whole of [0, m + e2]. This completes the proof if r = 2. n + e2 • • n • • m + e2 ↑ m [0, m] 0 Figure 1. The case r = 2. Now consider the case r = 3. Proceeding by induction as in the case r = 2, the extension problem reduces to that for a single cube. Consider therefore without loss of generality the unit cube based at 0 with m = e1 + e2 and j = 3, as illustrated in Figure 2. Then w is defined on the base of the cube [0, m] and it is required to extend w to a function v on the whole cube taking the value a at m + e3, under the assumption that there is a valid transition from w(m) to a. Now use the case r = 2 on successive faces of the cube. Working on the right hand face there is a unique possible value b for v(e1 + e3). Then, using this value for v(e1 + e3) on the near face we obtain the value c for v(e3). Similarly, working respectively on the back and left faces we obtain a value v(e2 + e3) = d and a second value, c′, for v(e3). Now suppose that c 6= c′. Working on the top face and using the values a, d, and c′, we obtain another value, b′ for v(e1 + e3). There are two possible transitions along the directed path 0 → e3 → e1 + e3 → m + e3, namely w(0) → c′ → b′ → a and w(0) → c → b → a. This contradicts the assumption that (M2M1M3)(a, w(0)) ∈ {0, 1}. The proof for general r now follows by induction. The uniqueness of the extension follows from the two dimensional considerations embodied in Figure 1. All the compatibility conditions required for existence follow from the three dimensional considerations of Figure 2. (cid:3) The next result follows by induction from Lemma 1.2. Lemma 1.3. Fix {0, 1}-matrices Mi, 1 ≤ i ≤ r, satisfying (H1a), (H1b), and (H1c). Let 1 ≤ j1, . . . , jp ≤ r, let a0, . . . , ap ∈ A and suppose that Mji(ai, ai−1) = 1 for 1 ≤ i ≤ p. Then there exists a unique word w ∈ W with σ(w) = ej1 + · · · + ejp, such that w(0) = a0 and w(ej1 + · · · + eji) = ai for 1 ≤ i ≤ p. 4 GUYAN ROBERTSON AND TIM STEGER • ................................................................................ •e3 • ................................................................................ • .. . .. . .. •e2 .. .. .. . .. .. .. .. . .. .. .. .. . .. .. .. .. . .. .. .. .. . .. .. .. .. . .. .. .. .. .. . .. .. .. .. . .. .. .. .. . .. .. .. .. . .. .. .. •0 .. . .. .. . •m . .. . .. .. .. .. .. . .. .. .. .. . .. .. .. .. . .. .. .. .. . .. .. .. .. . .. .. .. .. . .. .. .. .. .. . .. .. .. .. . .. .. .. .. . .. .. .. .. . .. .. .. .. . .. • e1 .. . Figure 2. The case r = 3. As a consequence we have Lemma 1.4. Fix {0, 1}-matrices Mi, 1 ≤ i ≤ r. If (H1a), (H1b), and (H1c) hold, then (H1) holds. Proof. Let u ∈ Wm and v ∈ Wn. Suppose t(u) = o(v). Choose j1, . . . , jp as in Lemma 1.3 so that ej1 + · · · + ejq = m ejq+1 + · · · + ejp = n for some q, 0 ≤ q ≤ p. Choose ai, 0 ≤ i ≤ q so as to force the w of Lemma 1.3 to satisfy w[0,m] = u. Thus aq = t(u) = o(v). Choose ai, q ≤ i ≤ p so as to force w to satisfy w[m,m+n] = v. The existence and uniqueness in (H1) follow from the existence and uniqueness in Lemma 1.3. (cid:3) Corollary 1.5. If u = (d, u) ∈ W m and v ∈ Wn with t(u) = o(v), then there exists a unique w ∈ W m+n such that In these circumstances we write w = uv, and say that the product uv exists. w[0,m] = u and w[m,m+n] = v. Proof. This is immediate, with uv = (d, uv). (cid:3) For the next two lemmas, and for the rest of the paper, suppose that matrices Mi have been chosen so that (H0) -- (H2) hold. Lemma 1.6. Let a, b ∈ A and n ∈ Zr and t(w) = b. +. There exists w ∈ W with σ(w) ≥ n such that o(w) = a Proof. By condition (H0), the matrix Mj is nonzero, so there exists at least one word of shape ej. Using this, choose words w1, . . . , wq so that σ(w1) + · · · + σ(wq) ≥ n. Using conditions (H1) and (H2), one can always find a word with a given origin and terminus. So choose s0 ∈ W with o(s0) = a, t(s0) = o(w1), choose sk ∈ W with o(sk) = t(wk), t(sk) = o(wk+1) for 1 ≤ k ≤ q − 1, and choose sq ∈ W with o(sq) = t(wq), t(sq) = b. Let w = s0w1s1w2 . . . wq−1sq−1wqsq. (cid:3) Lemma 1.7. Given u ∈ W and b ∈ A, there exists v ∈ W such that uv exists, σ(v) 6= 0 and t(v) = t(uv) = b. Proof. This follows immediately Lemma 1.6. (cid:3) 2. nonperiodicity Assume that Mi, 1 ≤ i ≤ r, have been chosen and that (H0) -- (H2) hold. In the large class of examples associated to affine buildings it is fairly easy to verify the nonperiodicity condition, (H3). However, in general it is hard to see how one can start with the matrices Mi and check (H3). In this section we present a condition which implies (H3), and show how it can in principle be checked. This material is not used in the remainder of the paper. (H3): Fix j, 1 ≤ j ≤ r. Let m ∈ Zr + with mj = 0. Let w ∈ Wm. Then there exist u, u′ ∈ Wm+ej such that u[0,m] = u′[0,m] = w but u(ej) 6= u′(ej). For l, m ∈ Zr, define l ∧ m = (l1 ∧ m1, . . . , lr ∧ mr), l ∨ m = (l1 ∨ m1, . . . , lr ∨ mr), l = l ∨ (−l). If w ∈ Wl where l ≥ 0 and if p 6= 0, recall that w is p-periodic if its p-translate, τpw, satisfies τpw[0,l]∩[p,p+l] = w[0,l]∩[p,p+l]. 5 l p + l . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 0 p Figure 3. The region [0, l] ∩ [p, p + l]. Lemma 2.1. Conditions (H0) -- (H2) and (H3) imply condition (H3). Proof. Observe that p-periodicity is invariant under replacement of p by −p. We may therefore assume that p has at least one positive component which we may take to be p1. Let p+ = p∨0 and p− = (−p) ∨ 0. We will construct w ∈ Wp. In this case w is defined on [0, p], τpw is defined on [p, p+p], and τpw and w are both defined on [p∨0, p∧(p+p)] = [p+, p+], a single point. The word w is p-periodic if and only if w(p+) = (τpw)(p+) = w(p+ − p) = w(p−). The situation is illustrated in Figure 4. Choose any v in Wp−e1 and let u = v[p+−e1,p−e1] ∈ Wp−. By condition (H3), there exist two different words x, x′ ∈ Wp−+e1 such that x[0,p−] = x′[0,p−] but x(e1) 6= x′(e1). At least one of x(e1) and x′(e1) differs from v(p−); we may assume that x(e1) 6= v(p−). Let w be defined by w[0,p−e1] = u, w[p+−e1,p] = x. Then w(p+) = x(e1) 6= v(p−) = w(p−), so w is not p-periodic. (cid:3) p− 0 p − e1 v x p+ − e1 p • p+ p p + p Figure 4. Proof of Lemma 2.1 in the case r = 2. Now we discuss the checkability of (H3). Fix j, 1 ≤ j ≤ r. For w ∈ Wm with σ(w)j = 0 let A(j, w) = {u(ej); u ∈ Wm+ej and u[0,m] = w}. Given w, one can calculate A(j, w) by considering, one at a time, the possible values of u(m + ej), and working back to find the possible values of u(ej) as in the proof of Lemma 1.2. The assertion of (H3) is that #A(j, w) ≥ 2 for any w with σ(w)j = 0. Let v, w ∈ W with σ(v) = ek, k 6= j, σ(w)j = 0, and suppose that vw is defined. Then A(j, vw) = {a ∈ A; Mj(a, o(v)) = 1 and Mk(b, a) = 1 for some b ∈ A(j, w)}. Thus one can calculate A(j, vw) from the knowledge of v and A(j, w). To check (H3) for the fixed value of j, one proceeds to construct, for each c ∈ A a complete list of possibilities for A(j, w) with o(w) = c. The first step in the algorithm is to insert in the lists all A(j, w) for w ∈ W0. Then proceeding cyclically through all words v with σ(v) = ek, for all k 6= j, the algorithm adds to the lists all possible values of A(j, vw) corresponding to values 6 GUYAN ROBERTSON AND TIM STEGER A(j, w) already on the lists. The algorithm terminates when a complete cycle through the words v generates no new possible values for A(j, w). The algorithm works because any w ∈ W can be written w = v1v2 . . . vq, with σ(vi) = eki. Is this algorithm practical for hand computation? for electronic computation? The authors have done no experiments, but they suspect that the lists of subsets of A will get out of hand rapidly as the cardinality of A increases. The situation is not entirely satisfactory. 3. The C ∗-algebra Assume conditions (H0)-(H3) hold. Define an abstract untopologized algebra A0 over C which u,v; u, v ∈ W and t(u) = t(v)}. The j=1, D, and δ. The generators of A0 are {s0 depends on A, (Mj)r relations defining A0 are v,w = s0 s0 u,vs0 s0 u,w u,v = Xw∈W ;σ(w)=ej, o(w)=t(u)=t(v) s0 uw,vw, for 1 ≤ j ≤ r It is trivial to verify that A0 has an antilinear antiautomorphism defined on the generators by s0 u,us0 v,v = 0 for u, v ∈ W 0, u 6= v. s0 u,v ∗ = s0 v,u. This makes A0 a ∗-algebra. Let A be the corresponding enveloping C ∗-algebra (c.f. [CK, p.256]) and let su,v be the image of s0 u,v in A. The generators of A are therefore and the defining relations are {su,v; u, v ∈ W and t(u) = t(v)} (3.1a) (3.1b) (3.1c) (3.1d) su,v ∗ = sv,u su,vsv,w = su,w su,v = Xw∈W ;σ(w)=ej, o(w)=t(u)=t(v) suw,vw, for 1 ≤ j ≤ r su,usv,v = 0 for u, v ∈ W 0, u 6= v. Remark 3.1. Suppose that u, v ∈ W and t(u) = t(v). Then su,v is a partial isometry with initial projection su,v ∗su,v = sv,v and final projection su,vsu,v ∗ = su,u. Lemma 3.2. Fix m ∈ Zr + and let u, v ∈ W with t(u) = t(v). Then su,v = Xw∈W ;σ(w)=m o(w)=t(u)=t(v) suw,vw. Proof. Using (H1), this follows by induction from (3.1c). (cid:3) Lemma 3.3. su,usv,v = 0 if σ(u) = σ(v) and u 6= v. Proof. The case σ(u) = σ(v) = 0 is exactly the relation (3.1d). Assume that the assertion is true whenever σ(u) = σ(v) = m. Let σ(u′) = σ(v′) = m + ej and let u = u′[0,m], v = v′[0,m]. o(w) = t(u). Since su′,u′ is one of the terms of the preceding sum we have su,u ≥ su′,u′. Similarly sv,v ≥ sv′,v′. If u 6= v, this proves that su′,u′sv′,v′ = 0, since by induction su,usv,v = 0. On the By relation (3.1c), we have that su,u = Pw suw,uw where the sum is over w ∈ Wej such that other hand, if u = v then su′,u′, sv′,v′ are distinct terms in the sum Pa suw,uw and are therefore orthogonal. (cid:3) Remark 3.4. If W 0 = D is finite then it follows from (3.1d) thatPu∈W 0 From Lemma 3.2 it follows that for any m, Pu∈W m su,u = Pu∈W 0 Lemma 3.3 it follows thatPu∈W 0 su,u is an identity for A. The next lemma is an immediate consequence of the definition of an enveloping C ∗-algebra. su,u is an idempotent. su,u. Hence from (3.1b) and Lemma 3.5. Let H be a Hilbert space and for each u, v ∈ W with t(u) = t(v) let Su,v ∈ B(H). If the Su,v satisfy the relations (3.1), then there is a unique -homomorphism φ : A → B(H) such that φ(su,v) = Su,v. (cid:3) 7 Lemma 3.6. Any product su1,v1su2,v2 can be written as a finite sum of the generators su,v. Proof. Choose m ∈ Zr with m ≥ σ(v1) and m ≥ σ(u2). Use Lemma 3.2 to write su1,v1 as a sum of terms su3,v3 with σ(v3) = m. Likewise, write su2,v2 as a sum of terms su4,v4 with σ(u4) = m. Now in the product su1,v1su2,v2 each term has the form su3,v3su4,v4 =(su3,v4 su3,v3sv3,v3su4,u4su4,v4 = 0 if v3 = u4, if v3 6= u4 by Lemma 3.3. The result follows immediately. Corollary 3.7. The C ∗-algebra A is the closed linear span of the set {su,v; u, v ∈ W and t(u) = t(v)}. (cid:3) (cid:3) Lemma 3.8. The algebra A is nonzero. Proof. We must construct a nonzero -homomorphism from A into B(H) for some Hilbert space H. Consider the set of infinite words W∞ = {w : Zr + → A; Mj(w(l + ej), w(l)) = 1 whenever l ≥ 0}, and define the product uv for u ∈ W and v ∈ W∞ exactly as in Definition 0.1. Let H = l2(W∞) and define (φ(su,v))(δw) =(δuw1 0 if w = vw1 for some w1 ∈ W∞, otherwise. It is easy to check that the operators {φ(su,v); u, v ∈ W and t(u) = t(v)} satisfy the relations 3.1 and it follows from Lemma 3.5 that φ extends to a -homomorphism of A. (cid:3) Remark 3.9. The Hilbert space H = l2(W∞) is not separable. However, since the algebra A is countably generated, there exist nonzero separable, A-stable subspaces of H, and in particular, there exist nontrivial representations of A on separable Hilbert space. Lemma 3.10. If φ is a nontrivial representation of A, and if u ∈ W then φ(su,u) 6= 0. particular su,u 6= 0. In Proof. Suppose φ(su,u) = 0. Choose any v ∈ W . Use Lemma 1.6 to find w ∈ W such that o(w) = t(u) and t(w) = t(v). By Lemma 3.2, we have 0 = φ(su,u) = P φ(suw′,uw′), the sum being taken over all w′ ∈ W such that σ(w′) = σ(w) and o(w′) = t(u). Thus φ(suw,uw) = 0. By Remark 3.1 it follows that that φ(suw,v) = 0, that φ(sv,v) = 0, and finally that φ(sv,v′) = 0 whenever t(v) = t(v′). Hence φ is trivial. (cid:3) Remark 3.11. When r = 1, the algebra A is a simple Cuntz-Krieger algebra. More precisely, if we write M = M t 1, then the Cuntz-Krieger algebra OM is generated by a set of partial isometries {Sa; a ∈ A} satisfying the relations S∗ b . If u ∈ W , let Su = Su(0)Su(1) . . . St(u) and if v ∈ W with t(u) = t(v), define Su,v = SuS∗ v (c.f. [CK, Lemma 2.2]). The map su,v 7→ Su,v establishes an isomorphism of A with OM . Tensor products of ordinary Cuntz-Krieger algebras aSa =Pb M(a, b)SbS∗ 8 GUYAN ROBERTSON AND TIM STEGER can be identified as higher rank Cuntz-Krieger algebras A. If A1, A2 are simple rank one Cuntz- Krieger algebras, with corresponding matrices M1, M2 and alphabets A1, A2 then A1 ⊗ A2 is the algebra A arising from the pair of matrices M1 ⊗ I, I ⊗ M2 and the alphabet A1 × A2. More interesting examples arise from group actions on affine buildings. The details for some r = 2 algebras arising in this way are given in Section 7. If m ∈ Zr +, let Fm denote the subalgebra of A generated by the elements su,v for u, v ∈ W m. 4. The AF subalgebra Lemma 4.1. There exists an isomorphism Fm Proof. Let u, v ∈ W m with t(u) = t(v) = a. Consider the map Ea δu,δv denotes a standard matrix unit in K(l2({w ∈ W ; σ(w) = m, t(w) = a})). This extends to a map which is an isomorphism according to equations (3.1a), (3.1b) and Lemma 3.3. (cid:3) δu,δv 7→ su,v, where Ea ∼=La∈A K(l2({w ∈ W ; σ(w) = m, t(w) = a})). The relations (3.1c) show that there is a natural embedding of Fm into Fm+ej . The C ∗-algebras {Fm : m ∈ Zr +} form a directed system of C ∗-algebras in the sense of [KR, p. 864]. By [KR, Proposition 11.4.1] there is an essentially unique C ∗-algebra F in which the union of the algebras Fm is dense, namely the direct limit of these algebras. We have the following commuting diagram of inclusions. Fm+ek −−−→ Fm+ej+ek x x We may equally well regard F as the closure ofS∞ F is an AF -algebra. Fm −−−→ Fm+ej j=1 Fjp, where p = (1, 1, . . . , 1). In particular Proposition 4.2. Let φ be a nonzero homomorphism from A into some C ∗-algebra. Then the restriction of φ to F is an isomorphism. Proof. Suppose that φ is not an isomorphism on F . Then φ is not an isomorphism on some Fm. SinceLa∈A K(l2({w ∈ W ; σ(w) = m, t(w) = a})) ∼= Fm (Lemma 4.1), it follows from simplicity of the algebra of compact operators that φ(su,u) = 0 for some u, contradicting Lemma 3.10. (cid:3) Define an action α of the r-torus Tr on A as follows. 1 tm2 If σ(u) − σ(v) = m ∈ Zr and t = (t1, . . . , tr) ∈ Tr, let αt(su,v) = tmsu,v, where tm = tm1 . . . tmr . The elements αt(su,v) satisfy the relations (3.1) and generate the C ∗-algebra A. By the universal property of A it follows that αt extends to an automorphism of A. It is easy to see that that t 7→ αt is an action. It is also clear that αt fixes all elements su,v with σ(u) = σ(v) and so fixes F pointwise. We now show that F = Aα, the fixed point subalgebra of A. Consider the linear map on A defined by 2 r (4.1) αt(x)dt, for x ∈ A where dt denotes normalized Haar measure on Tr. π(x) =ZTr Lemma 4.3. Let π, F be as above. (1) The map π is a faithful conditional expectation from A onto F . (2) F = Aα, the fixed point subalgebra of A. Proof. Since the action α is continuous, it is easy to see that π is a conditional expectation from A onto Aα, and that it is faithful. Since αt fixes F pointwise, F ⊂ Aα. To complete the proof we show that the range of π is contained in (and hence equal to) F . By continuity of π and Corollary 3.7, it is enough to show that π(su,v) ∈ F for all u, v ∈ W . This is so because 9 (4.2) π(su,v) = su,vZTr if σ(u) 6= σ(v), if σ(u) = σ(v). (cid:3) tσ(u)−σ(v)dt =(0 su,v 5. Simplicity We show that the C ∗-algebra A is simple. Consequently any nontrivial C ∗-algebra with gener- ators Su,v satisfying relations (3.1) is isomorphic to A. Several preliminary lemmas are necessary. Consequences of (H3), their theme is the existence of words lacking certain periodicities. Recall that for m ∈ Zr, m = (m1, . . . , mr). Lemma 5.1. Let m ∈ Zr with m ≥ 0 and let a ∈ A. There exists some l ≥ 0 and some w ∈ Wl satisfying (1) If p ≤ m and p 6= 0 then τpw[0,l]∩[p,p+l] 6= w[0,l]∩[p,p+l], (2) o(w) = a. Proof. Note that if w = uwpv, and if wp is not p-periodic, then neither is w. Apply (H3) to obtain for each nonzero p, p ≤ m, a word wp which is not p-periodic. The final word w is obtained by concatenating these words in some order using "spacers" whose existence is guaranteed by Lemma 1.6. The construction is illustrated in Figure 5, where the spacer sp,q is chosen so that o(sp,q) = t(wp) and t(sp,q) = o(wq). (cid:3) wq sp,q wp Figure 5. Part of the word w; a word sp,q is used to concatenate wp and wq. Lemma 5.2. One can find u, u′ ∈ W with σ(u) = σ(u′), o(u) = o(u′), but u 6= u′. Proof. Assume the contrary. Considering words of shape e1, we see that for fixed a, no more than one b ∈ A satisfies M1(b, a) = 1. By Lemma 1.6, at least one b ∈ A satisfies M1(b, a) = 1 and at least one c ∈ A satisfies M1(c, a) = 1. Consequently the directed graph associated to M1 must be a union of closed cycles. If g is the g.c.d of the cycle lengths, then every w ∈ W is ge1-periodic, contradicting (H3). (cid:3) Lemma 5.3. Let p ∈ Zr. Let w1, w2 ∈ Wl. There exist l′ ≥ l and w′ 1, w′ 2 ∈ Wl′ such that and w′ 1[0,l] = w1, w′ 2[0,l] = w2, τpw′ 1[0,l′]∩[p,p+l′] 6= w′ 2[0,l′]∩[p,p+l′]. 10 GUYAN ROBERTSON AND TIM STEGER Proof. Find two different words u, v with σ(u) = σ(v) and o(u) = o(v). Choose s so that p + σ(w1) + σ(s) ≥ 0, o(s) = t(w1) and t(s) = o(u) = o(v). Choose w′′ 2[0,l] = w2 and σ(w′′ If this is equal to u, let w′′ 1 = w1su. (This is illustrated in Figure 6.) Finally, let l′ = σ(w′′ (cid:3) 2) ≥ p + σ(w1) + σ(s) + σ(u). Consider w′′ 1 = w1sv, otherwise, let w′′ 2[p+σ(w1)+σ(s),p+σ(w1)+σ(s)+σ(u)]. 2 so that w′′ 2) and extend w′′ 2 to words w′ 1) ∨ σ(w′′ 1 and w′′ 1 and w′ 2 in Wl′. u or v s w1 Figure 6. The word w′′ 1. Lemma 5.4. Fix m ∈ Zr with m ≥ 0. For some l ≥ 0 there exists a subset S = {wa; a ∈ A} of Wl satisfying the two properties below. (1) For each a ∈ A, o(wa) = a. (2) Let a, b ∈ A. Let p 6= 0 be in Zr with p ≤ m. Then wa[0,l]∩[p,p+l] 6= τpwb[0,l]∩[p,p+l]. Proof. The elements of S are chosen as follows. For each a ∈ A, let wa ∈ W satisfy the conclusions of Lemma 5.1. By extending the words wa as necessary we may suppose that wa ∈ Wl for some l ≥ 0. If a, b ∈ A, a 6= b and p ∈ Zr with p ≤ m, we can apply Lemma 5.3 and extend wa and wb to w′ b do not agree on their common domain. Then one can extend all the other wc to w′ b. Apply this procedure once for each of the finitely many triples (a, b, p) ∈ A × A × Zr with p ≤ m and a 6= b, and the proof is done. (cid:3) c of the same shape as w′ b where τpw′ a and w′ a and w′ a and w′ Fix m ∈ Zr +. Choose l and S satisfying the conditions of Lemma 5.4. Define (5.1) Lemma 5.5. If l′ ≥ l, then (5.2) Proof. By (H1) and Lemma 3.2, Xw∈W ;σ(w)=m+l′ w[m,m+l]∈S Q = Xw∈W ;σ(w)=m+l w[m,m+l]∈S sw,w. Q = Xw∈W ;σ(w)=m+l′ w[m,m+l]∈S sw,w. sw,w = Xw1∈W ,w2∈W σ(w1)=m+l,σ(w2)=l′−l w1[m,m+l]∈S,t(w1)=o(w2) sw1w2,w1w2 = Q. Lemma 5.6. Suppose σ(u), σ(v) ≤ m and t(u) = t(v) = a. If σ(u) 6= σ(v), then Qsu,vQ = 0. (cid:3) Proof. Using Lemma 3.2, write su,v = Xv1;σ(v1)=m+l o(v1)=a suv1,vv1. K(l2({w ∈ W ; σ(w) = m, t(w) = a})) → A Ma∈A 11 Apply Lemma 5.5 with l′ = l + σ(v) so that for each word w in the sum (5.2) σ(w) = m + l + σ(v) = σ(vv1). By Lemma 3.3, suv1,vv1sw,w = 0 unless vv1 = w. Consequently suv1,vv1Q = 0 unless vv1[m,m+l] ∈ S, which is to say v1[m−σ(v),m−σ(v)+l] ∈ S. Similarly, Qsuv1,vv1 = 0 unless v1[m−σ(u),m−σ(u)+l] ∈ S. But if w1 = v1[m−σ(v),m−σ(v)+l] ∈ S and w2 = v1[m−σ(u),m−σ(u)+l] ∈ S, then w1 and w2 would fail condition (2) of Lemma 5.4, with p = σ(u) − σ(v). (cid:3) Remark 5.7. If x = Pi cisui,vi σ(ui), σ(vi) ≤ m then QxQ = Xσ(ui)=σ(vi) is a finite linear combination of the generators of A with ciQsui,viQ = Qπ(x)Q, by Lemma 5.6 and equation (4.2). Lemma 5.8. The map x 7→ QxQ is an isometric -algebra map from Fm into A. Proof. If σ(u) = σ(v) = m and t(u) = t(v) = a, then, with the notation of Lemma 5.4, we have Qsu,vQ = suwa,vwa. With the notation of Lemma 4.1, consider the map given by Ea δu,δv 7→ suwa,vwa, when t(u) = t(v) = a. This is easily checked to be an injective -algebra map, hence an isom- etry. The map in the statement of the Lemma is the composition of this isometry with that of Lemma 4.1. (cid:3) Theorem 5.9. The C ∗-algebra A is simple. Proof. Let φ be a nonzero -homomorphism from A to some C ∗-algebra. It is enough to show that φ is an isometry. Let x = Pi cisui,vi be a finite linear combination of the generators of A. Choose m ∈ Zr + so that σ(ui), σ(vi) ≤ m for all i. Choose l and S as in Lemma 5.4 and let Q be as in equation (5.1). By Remark 5.7, QxQ = Qπ(x)Q. Observe that π(x) ∈ F and so by Lemma 5.8, kQπ(x)Qk = kπ(x)k. Moreover Qπ(x)Q ∈ Fm+l, so by Proposition 4.2, kφ(Qπ(x)Q)k = kQπ(x)Qk. Thus kφ(x)k ≥ kφ(Q)φ(x)φ(Q)k = kφ(QxQ)k = kφ(Qπ(x)Q)k = kQπ(x)Qk = kπ(x)k. The inequality extends by continuity to all x ∈ A. It follows that φ is faithful since if φ(y) = 0 then 0 = kφ(y∗y)k ≥ kπ(y∗y)k. Therefore y = 0, by Lemma 4.3. (cid:3) Corollary 5.10. Let H be a Hilbert space and for each u, v ∈ W with t(u) = t(v) let Su,v ∈ B(H) be a nonzero partial isometry. If the Su,v satisfy the relations (3.1) and A denotes the C ∗-algebra which they generate, then there is a unique -isomorphism φ from A onto A such that φ(su,v) = Su,v. Proof. This follows immediately from Lemma 3.5 and Theorem 5.9. (cid:3) Proposition 5.11. The C ∗-algebra A is purely infinite. Proof. Note first that by Lemma 3.2, suw,uw is a subprojection of su,u. By Lemma 1.7, this implies that su,u is an infinite projection for any u. Any rank one projection in any Fl is equivalent to su,u for some u ∈ Wl, and hence is infinite. We must show that for every nonzero h ∈ A+, the C ∗-algebra hAh contains an infinite pro- jection. Since π is faithful, we may assume kπ(h2)k = 1. Let 0 < ǫ < 1. Approximate h by self-adjoint finite linear combinations of generators. The square of this approximation gives an element y =Pi cisui,vi with y ≥ 0 and ky − h2k ≤ ǫ. Fix m so that σ(ui), σ(vi) ≤ m for all i and then construct Q by Lemma 5.4 and equation (5.1). We have QyQ = Qπ(y)Q ∈ Fl for some l, QyQ ≥ 0 and kQyQk = kQπ(y)Qk = kπ(y)k ≥ kπ(h2)k − ǫ = 1 − ǫ. Since Fl is a direct sum of (finite or infinite dimensional) algebras of compact operators, there exists a rank one positive 12 GUYAN ROBERTSON AND TIM STEGER operator R1 ∈ Fl with kR1k ≤ (1 − ǫ)−1/2 so that R1QyQR1 = P is a rank one projection in Fl. Hence P is an infinite projection. It follows that kR1Qh2QR1 − P k ≤ kR2 1kkQk2ky − h2k ≤ ǫ/(1 − ǫ). By functional calculus, one obtains R2 ∈ A+ so that R2R1Qh2QR1R2 is a projection and kR2R1Qh2QR1R2 −P k ≤ 2ǫ/(1 −ǫ). For small ǫ one can then find an element R3 in A so that R3R2R1Qh2QR1R2R∗ 3 = P . Let R = R3R2R1Q, so that Rh2R∗ = P . Consequently, Rh is a partial isometry, whose initial projection hR∗Rh is a projection in hAh and whose final projection is P . Moreover, if V is a partial isometry in A such that V ∗V = P and V V ∗ < P , then (hR∗)V (Rh) is a partial isometry in hAh with initial projection hR∗Rh and final projection strictly less than hR∗Rh. (cid:3) We can now explain one of the reasons for introducing the set D of decorations. Recall that D is a countable or finite set. Denote by AD the algebra A corresponding to a given D. One special case of interest is when D = A and δ is the identity map. Then the algebra AA is a direct generalization of a Cuntz-Krieger algebra [CK]. There is an obvious notion of equivalence for decorations: two decorations δ1 : D1 → A and δ2 : D2 → A are equivalent if there is a bijection η : D1 → D2 such that δ1 = δ2η. Equivalent decorations give rise to isomorphic algebras. Given any set D of decorations we can obtain another set of decorations D × N, with the decorating map δ′ : D × N → A defined by δ′((d, i)) = δ(d). Lemma 5.12. There exists an isomorphism of C ∗-algebras AD×N ∼= AD ⊗ K. Proof. If u, v ∈ W , the isomorphism is given by s((d,i),u),((d′,j),v) 7→ s(d,u),(d′,v) ⊗ Ei,j, where the Ei,j are matrix units for B(l2(N)). The fact that this is an isomorphism follows from Corollary 5.10. (cid:3) This procedure is useful, because it provides a routine method of passing from A to A ⊗ K, a technique that is necessary to obtain the results of [CK]. + be any map. Define D′ = {(d, w) ∈ W ; σ(w) = l(d)} and define Lemma 5.13. Let l : D → Zr δ′ : D′ → A by δ′(w) = t(w). Then AD′ ∼= AD. Proof. Define φ : AD′ → AD by φ(s(w1,u1),(w2,u2)) = sw1u1,w2u2, for w1, w2 ∈ D′, u1, u2 ∈ W , o(ui) = δ′(wi) = t(wi), and t(u1) = t(u2). Relations (3.1) (for AD′) are satisfied for φ(s(w1,u1),(w2,u2)). By Corollary 5.10 the homomorphism φ exists and is injective. Relation (3.1c) for AD shows that each generator of AD is in the image of φ. Hence φ is an isomorphism. (cid:3) Corollary 5.14. For any (D, δ), AD is isomorphic to AD′ for some (D′, δ′) with δ′ : D′ → A surjective. Proof. By general hypothesis, D is nonempty and A is finite. Use Lemma 5.13 once to replace D with D′′ so that #(D′′) ≥ #(A), and use it again, in conjunction with Lemma 1.6, to construct the pair (D′, δ′). (cid:3) Corollary 5.15. For a fixed alphabet A and fixed transition matrices Mj, the isomorphism class of AD ⊗ K is independent of D. ∼= AD′ for some (D′, δ′) with δ′ : D′ → A surjective. By Lemma 5.12, Proof. By Corollary 5.14, AD AD′×N ∼= AD′ ⊗ K. Since δ′ is surjective, the decorating set D′ × N is equivalent to the decorating set A × N: the inverse image of each a ∈ A is countable. Thus AD ⊗ K ∼= AA×N. (cid:3) Decorating sets other than A and A × N arise naturally in the examples associated to affine buildings. 6. Construction of the algebra A ⊗ K as a crossed product A vital tool in [CK] was the expression of OA ⊗ K as the crossed product of an AF algebra by a Z-action [CK, Theorem 3.8]. The present section is devoted to an analogous result. In view of Lemma 5.12, this is done by establishing an isomorphism from AA×N onto the crossed product of an AF-algebra by a Zr-action. The AF-algebra will be isomorphic to the algebra F of Section 4, relative to the decorating set A × N, and the action of an element k ∈ Zr will map the subalgebra Fm onto Fm+k for each m ≥ 0. Let a C ∗-algebra A′ be defined just as A is, with D = A and with generators {s′ u,v; u, v ∈ W and o(u) = o(v)}. The relations are the same as those in (3.1) except that in the sum (3.1c), words are extended from the beginning. The full relations are 13 (6.1a) (6.1b) (6.1c) (6.1d) ∗ = s′ s′ u,v s′ u,vs′ v,w = s′ s′ v,u u,w u,v = Xw∈W ;σ(w)=ej, t(w)=o(u)=o(v) s′ wu,wv u,us′ s′ v,v = 0 for u, v ∈ W0, u 6= v. This may be thought of as using words with extension in the negative direction. Alternatively, j for each j in the definition of A results in an algebra j follow from the corresponding conditions for replacing Mj by the transpose matrix M t isomorphic to A′. Conditions (H0) -- (H3) for the M t the Mj. Consequently all the preceding results are valid for the algebra A′. By Theorem 5.9, A′ is a simple separable C ∗-algebra. Let ψ : A′ → B(H) be a nondegenerate representation of A′ on a separable Hilbert space H. For a ∈ A, let Ha be the range of the projection ψ(s′ a,a). Then H =La∈A Ha. By Lemmas 3.2 and 3.10, each Ha is infinite dimensional. Let C =La∈A K(Ha) ⊂ K(H). For each l ∈ Zr + define a map αl : C → C by (6.2) αl(x) = Xw∈Wl ψ(s′ w,o(w))xψ(s′ o(w),w). Note that ψ(s′ w,o(w)) is a partial isometry with initial space Ho(w) and final space lying inside Ht(w). Lemma 6.1. Let αl be as above. (1) αl has image in C. (2) αl is a C ∗-algebra inclusion. (3) For k, l ∈ Zr +, αkαl = αk+l. Proof. 1. Clearly αl(x) ∈ K(H). Moreover for fixed w ∈ W with t(w) = a, ψ(s′ K(Ha). w,o(w))xψ(s′ o(w),w) ∈ 2. Fix w ∈ Wl with o(w) = a. Observe that s′ sa,a. Moreover for two different words w1, w2 ∈ Wl, the range projections of s′ are orthogonal. The result is now clear. 3. If w1 ∈ Wk and w2 ∈ Wl then according to Lemmas 3.2 and 3.3, w,a is a partial isometry with initial projection w2,o(w2) w1,o(w1) and s′ (6.3) s′ w1,o(w1)s′ Therefore t(w3)=o(w1) w2,o(w2) = Xw3∈Wl αkαl(x) = Xw1∈Wk = Xw1∈Wk w2∈Wl w2∈Wl t(w2)=o(w1) s′ w3w1,w3s′ w2,o(w2) =(0 s′ w2w1,o(w2) if t(w2) 6= o(w1) if t(w2) = o(w1). ψ(s′ w1,o(w1))ψ(s′ w2,o(w2))xψ(s′ o(w2),w2)ψ(s′ o(w1),w1) ψ(s′ w2w1,o(w2))xψ(s′ o(w2),w2w1) = αk+l(x). (cid:3) 14 GUYAN ROBERTSON AND TIM STEGER For each m ∈ Zr let C(m) be an isomorphic copy of C, and for each l ∈ Zr +, let α(m) l : C(m) → C(m+l) be a copy of αl. Let E = lim−→ C(m) be the direct limit of the category of C ∗-algebras with objects C(m) and morphisms αl ([KR, Proposition 11.4.1]). Then E is an AF algebra. (See the discussion preceding Proposition 4.2.) If x ∈ C, let x(m) be the corresponding element of C(m). Then x(m) is identified with (αlx)(m+l) for all l ∈ Zr +. Define an action ρ of Zr on E by ρ(l)(x(m)) = x(m+l). Since Zr is amenable, the full crossed product of E by this action coincides with the reduced crossed product [Ped, Theorem 7.7.7], and we denote it simply by E ⋊ Zr. The defining property of the crossed product says that there is a unitary representation m 7→ U m of Zr into the multiplier algebra of E ⋊ Zr such that ρ(l)(x(m)) = U lx(m)U −l, that is (6.4) U lx(m) = x(m+l)U l. Theorem 6.2. There exists an isomorphism φ : AA×N → E ⋊ Zr where, moreover φ(F ) = E. Proof. Let D = A × N and δ(a, n) = a. Fix a map β : D → H so that {β(d); δ(d) = a} is an orthonormal basis of Ha. For w = (d, w) ∈ W m define β(w) = ψ(s′ w,o(w))β(d), in this way extending β to a map β : W → H. Observe that for a fixed w ∈ W with o(w) = a, {β(d, w); d ∈ D, δ(d) = a} is an orthonormal basis for the range of s′ w,w. Since, moreover the ranges of {s′ w,w; w ∈ Wm} are pairwise orthogonal and sum to all of H, we see that {β(w); w ∈ W m} is an orthonormal basis for H. For w ∈ W and u = (d, u) ∈ W , we have ψ(s′ w,o(w))β(u) = ψ(s′ w,o(w))ψ(s′ u,o(u))β(d) =(0 ψ(s′ uw,o(u))β(d) if o(w) 6= t(u) if o(w) = t(u), by equation (6.3). That is w,o(w))β(u) =(0 ψ(s′ β(uw) if o(w) 6= t(u) if o(w) = t(u). We will now define the map φ : AD → E ⋊ Zr. For u, v ∈ W with t(u) = t(v) and σ(u) = l and σ(v) = m define (6.5) φ(sv,u) = U m−l(cid:16)β(v) ⊗ β(u)(cid:17)(l) . We use the notation ξ ⊗ η to denote the rank one operator on a Hilbert space defined by ζ 7→ hζ, ηiξ, so that when ξ,η have norm one, ξ ⊗ η is a partial isometry with initial projection η ⊗ η and final projection ξ ⊗ ξ. If the vectors ξ, η vary through an orthonormal basis for a Hilbert space then the operators ξ ⊗ η form a system of matrix units for the compact operators on that Hilbert space. By equation (6.4) we have (6.6) φ(sv,u) =(cid:16)β(v) ⊗ β(u)(cid:17)(m) U m−l. We show that the partial isometries φ(sv,u) satisfy the relations (3.1). Relation (3.1d) is imme- diate from the definition of β(w), since if w ∈ W 0 then φ(sw,w) =(cid:16)β(w) ⊗ β(w)(cid:17)(0) Relation (3.1a) is satisfied since, by (6.6), . φ(sv,u)∗ =(cid:18)(cid:16)β(v) ⊗ β(u)(cid:17)(m) U m−l(cid:19)∗ = U l−m(cid:16)β(u) ⊗ β(v)(cid:17)(m) = φ(su,v). 15 If t(w) = t(v) and σ(w) = n, then (6.7) φ(sw,v)φ(sv,u) = U n−m(cid:16)β(w) ⊗ β(v)(cid:17)(m) U m−l(cid:16)β(v) ⊗ β(u)(cid:17)(l) = U n−mU m−l(cid:16)β(w) ⊗ β(v)(cid:17)(l)(cid:16)β(v) ⊗ β(u)(cid:17)(l) = U n−l(cid:16)β(w) ⊗ β(u)(cid:17)(l) = φ(sw,u). Thus (3.1b) is satisfied. Finally (3.1c) is a consequence of the following calculation. φ(sv,u) = U m−l(cid:16)β(v) ⊗ β(u)(cid:17)(l) w,o(w))β(v) ⊗ ψ(s′ = U m−l(cid:16)αk(β(v) ⊗ β(u))(cid:17)(l+k) = U m−l Xw∈Wk(cid:16)ψ(s′ = U m−l Xw∈Wk = Xw∈Wk o(w)=t(u)(cid:16)β(vw) ⊗ β(uw)(cid:17)(l+k) φ(svw,uw). o(w)=t(u) w,o(w))β(u)(cid:17)(l+k) All the relations (3.1) are satisfied by the partial isometries φ(sv,u). Therefore by Lemma 3.5, φ defines a -homomorphism. Clearly φ is not the zero map; hence it is an isometry, by Theorem 5.9. It only remains to show that φ is onto. Fix m, n ∈ Zr so that m, m + n ≥ 0. For u ∈ W m and v ∈ W m+n with t(u) = t(v) = a, we have (6.8) φ(sv,u) = U n(cid:16)β(v) ⊗ β(u)(cid:17)(m) . As the sets {β(v); v ∈ W m+n, t(v) = a} and {β(u); u ∈ W m, t(u) = a} are bases for Ha, the image of φ contains a dense subset of U n (K(Ha))(m). Therefore the image of φ contains U n (K(Ha))(m), for each a ∈ A. It therefore contains U nC(m). Also, for any k ≥ 0, φ(AA×N) ⊇ U nC(m) ⊇ U nα(m−k) k C(m−k) = U nC(m−k). It follows that φ(AA×N) = E ⋊ Zr. It is clear from the definitions that φ(Fm) = C(m) and that φ(F ) = E. (cid:3) Corollary 6.3. A ⊗ K ∼= E ⋊ Zr. Proof. This follows immediately from Theorem 6.2, Lemma 5.12 and Corollary 5.15. Corollary 6.4. A is nuclear. Proof. This follows because the class of nuclear C ∗-algebras is closed under stable isomorphism and crossed products by amenable groups, and contains the AF algebras. (cid:3) Remark 6.5. Suppose that D is finite, so that A is unital by Remark 3.4. Then it has been established that the separable unital C ∗-algebra A is simple (Theorem 5.9), nuclear (Corollary 6.4) and purely infinite (Proposition 5.11). Corollary 6.3 also shows that A belongs to the bootstrap class N , which contains the AF-algebras and is closed under stable isomorphism and crossed products by Z. Thus A satisfies the Universal Coefficient Theorem [Bl, Theorem 23.1.1]. The work of E. Kirchberg and C. Phillips [K1, K2],[Ph] therefore shows that A is classified by its K-groups. (cid:3) 16 GUYAN ROBERTSON AND TIM STEGER 7. Boundary actions on affine buildings of type eA2 Let B be a locally finite thick affine building of type eA2. This means that B is a chamber system consisting of vertices, edges and triangles (chambers). An apartment is a subcomplex of B isomorphic to the Euclidean plane tesselated by equilateral triangles. A sector (or Weyl chamber ) is a π 3 -angled sector made up of chambers in some apartment. Two sectors are equivalent (or parallel) if their intersection contains a sector. We refer to [Br1, G, Ron] for the theory of buildings. Shorter introductions to the theory are provided by [Br2, Ca, St]. The boundary Ω is defined to be the set of equivalence classes of sectors in B. In B we fix some vertex O, which we assume to have type 0. For any ω ∈ Ω there is a unique sector [O, ω) in the class ω having base vertex O [Ron, Theorem 9.6]. The boundary Ω is a totally disconnected compact Hausdorff space with a base for the topology given by sets of the form Ω(v) = {ω ∈ Ω : [O, ω) contains v} where v is a vertex of B [CMS, Section 2]. We note that if [O, ω) contains v then [O, ω) contains the parallelogram conv(O, v). Let Γ be a group of type rotating automorphisms of B that acts freely on the vertex set with finitely many orbits. See [CMSZ] for a discussion and examples in the case where Γ acts transitively on the vertex set. There is a natural induced action of Γ on the boundary Ω and we can form the universal crossed product algebra C(Ω) ⋊ Γ [Ped]. The purpose of this section is to identify C(Ω) ⋊ Γ with an algebra of the form A. Let a be a Coxeter complex of type eA2, which we shall use as a model for the apartments of B. Each vertex of a has type 0,1, or 2. Fix as the origin in a a vertex of type 0. Coordinatize the vertices by Z2 by choosing a fixed sector in a based at the origin and defining the positive coordinate axes to be the corresponding sector panels (cloisons de quartier ). The coordinate axes are therefore given by two of the three walls of a passing through the origin. Let t be a model tile in a and let pm be a model parallelogram in a of shape m = (m1, m2), as illustrated in Figure 7. As per Figure 7, assume that t and pm are both based at (0, 0). Thus t is the model parallelogram of shape (0, 0). (0,1) (1,1) .. . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . .. .. . . . . . . . . . . .. (0,0) (1,0) The model tile t. (0,m2+1) (m1+1,m2+1) .. .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . .. .. . . . . . . . . . . .. . .. . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . (0,0) (m1+1,0) A model parallelogram pm. Figure 7. Let T denote the set of type rotating isometries i : t → B, and let A = Γ\T. We will use the set A as an alphabet to define an algebra A. Let Pm denote the set of type rotating isometries p : pm → B, and let Wm = Γ\Pm. Let P =Sm Pm and W =Sm Wm. If p ∈ Pm, then define t(p) : t → B by t(p)(l) = p(m + l). Then t(p) is a type rotating isometry such that t(p)(t) lies in p(pm) with t(p)(1, 1) = p(m1 + 1, m2 + 1), as illustrated in Figure 8. Thus t(p) ∈ T. Similarly o(p) : t → B is defined by o(p) = pt. 17 t(p)(t) .. . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . .. . . . .. . . . . . . . . . . . . . . . . . . . . . . . .. . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . .. . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . ... .. . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . .. . . . . . . . . . . . . .. .. . . . . .. .. . . . . . . . . . . . . . . . . . . . . . .. . . .. . . . . . . . . . . . . . . . . . . . . . .. .. . . . . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. o(p)(t) p(pm) p(0, 0) Figure 8. The initial and terminal tiles. The matrices M1, M2 with entries in {0, 1} are defined as follows. If a, b ∈ A, say that M1(b, a) = 1 if and only if there exists p ∈ P(1,0) such that a = Γo(p) and b = Γt(p). Similarly, if c ∈ A then M2(c, a) = 1 if and only if there exists p ∈ P(0,1) such that a = Γo(p) and c = Γt(p). The definitions are illustrated in Figure 9, for suitable representative isometries ia,ib,ic in T. ic(t) .. .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . .. .. . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . .. ia(t) ib(t) .. .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . .. .. . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . .. . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . .. ia(t) M2(c, a) = 1 M1(b, a) = 1 Figure 9. Definition of the transition matrices. In order to apply the general results we need to verify that conditions (H0)-(H3) are satisfied. We address this question in the subsection 7.1. Until further notice we simply impose the following ASSUMPTION: Conditions (H0)-(H3) are satisfied. We can now define the set of words Wm of shape m ∈ Z+ based on the alphabet A and the transition matrices M1,M2, as in Section 1. There is also a natural map α : P → W , defined as follows. Given p ∈ Pm, construct α(p) = w according to the following procedure. For each n ∈ Z2 + with 0 ≤ n ≤ m, let w(n) = Γpn where pn ∈ T is defined by pn(l) = p(n + l), for (0, 0) ≤ l ≤ (1, 1). Since the translation l 7→ n + l is a type-rotating isometry of a, it follows that pn is type rotating, hence an element of T. Passing to the quotient by Γ gives a well defined map α : Wm → Wm and hence a map α : W → W . The use of α for the different maps should not cause confusion. If a = Γi ∈ A = W0 then α(a) = a, so it is clear that o(α(p)) = Γo(p) and t(α(p)) = Γt(p). Lemma 7.1. The map α is a bijection from Wm to Wm for each m ∈ Z2 +. Proof. Suppose that α(p) = α(p′). Then Γpn = Γp′ n for 0 ≤ n ≤ m. For each n there exists γn so that γnpn = p′ n. Since Γ acts freely on the vertices, each γn is uniquely determined. Moreover, since pn and pn+ej share a pair of vertices in their image, it must be true that γn = γn+ej . By induction, the γn have a common value, γ, and γp = p′. Thus Γp = Γp′. It remains to show that α is surjective. Let w ∈ Wm. Choose a path (n0, . . . , nk) of points in Z2 so that n0 = 0, nk = m, and each difference nj+1 − nj is either e1 or e2. Choose representative type rotating isometries i0, i1, . . . , ik from t into B with Γir = w(nr) so that ir(t) and ir+1(t) 18 GUYAN ROBERTSON AND TIM STEGER are adjacent tiles in B, according to the two possibilities in Figure 9. This defines a gallery G = {i0(t), i1(t), . . . , ik(t)}. Let G0 = {C0, C1, . . . , Ck} be the corresponding gallery in pm with C0 = t. It is clear from Figure 10 that G0 is a minimal gallery in a. (The elements of G and G0 are tiles rather than chambers, but this is immaterial.) The obvious map p : G0 → G is a strong isometry (preserves generalized distance). Therefore G is contained in an apartment and p extends to a strong isometry p from the convex hull conv(G0) = pm into that apartment. See [Br1, p. 90, Theorem] and [Br2, Appendix B]. Thus p ∈ Pm and since α(p) agrees with w at n0, n1, . . . , nk, (H1) implies that α(p) = w. (cid:3) . . . . .. . . . . . . . . . . . . .. . . . . . . . . . . .. . . . . . . . . .. . . . . . . .. . . . .. . . . . . . .. . . . . . .. . . . . . . .. . . .. . . . . . . .. . . . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . .. . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . .. . . . .. . . .. . . . . . . . . . . . . . . . . . . . . .. .. . . . . ... . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . .. . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . .. .. . . . . . . . . . .. .. . . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . .. . . .. .. . . . . . . . . . . . .. . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . .. . . . . . . . . . . . .. . . .. .. . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . .. . . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . .. . .. . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . .. . . . . . . . . .. .. . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . .. . . . . . . . . . . .. . . . . . . .. . . .. . . . . . . .. . . .. . . . . . . .. . . Figure 10. A minimal gallery in in a. Let Wm denote the set of type rotating isometries p : pm → B such that p(0, 0) = O and let W =Sm Wm. Let D denote the set of type-rotating isometries d : t → B such that d(0, 0) = O. Let δ : D → A be given by δ(d) = Γd. The map δ is injective since Γ acts freely on the vertices of B. Moreover δ is surjective if and only if Γ acts transitively on the vertices. Define α : W → W by α(p) = (o(p), α(Γp)). (Recall that o(p) = pt.) Lemma 7.2. The map α is a bijection from Wm onto W m for each m ∈ Z2 +. Proof. If α(p1) = α(p2) then o(p1) = o(p2); moreover Γp1 = Γp2, by Lemma 7.1. Since Γ acts freely on the vertices, it follows that p1 = p2. Therefore α is injective. To see that it is surjective, let w = (d, w) ∈ W m, where w ∈ Wm and d ∈ D. By Lemma 7.1, there exists p ∈ Pm such that α(Γp) = w. Then Γd = δ(d) = o(w) = o(α(Γp)) = Γo(p). Replacing p by γp for suitable γ ∈ Γ ensures that o(p) = d and hence p ∈ Wm and α(p) = w. (cid:3) If p ∈ W then conv(O, t(p)(t)) = conv(O, p(m1 + 1, m2 + 1)) and we introduce the notation Ω(p) = Ω(p(m1 + 1, m2 + 1)) = {ω ∈ Ω; t(p) ⊂ [O, ω)} = {ω ∈ Ω; p(pm) ⊂ [O, ω)} . Let i ∈ T, that is, suppose that i : t → B is a type rotating isometry. Let Ω(i) = {ω ∈ Ω; i(t) ⊂ [i(0, 0), ω)} , those boundary points represented by sectors which originate at i(0, 0) and contain i(t). Clearly Ω(γi) = γΩ(i). For p ∈ Wm we have Ω(p) = Ω(t(p)). Indeed, any sector originating at t(p)(0, 0) and containing t(p)(t) extends to a sector originating at O and containing p(pm). Fix w1, w2 ∈ W with t(w1) = t(w2) = a ∈ A. Let p1 = α−1(w1) and p2 = α−1(w2). Let γ ∈ Γ be the unique element such that γt(p1) = t(p2). Define a homomorphism φ : A → C(Ω) ⋊ Γ by (7.1) φ(sw2,w1) = γ1Ω(p1) = 1Ω(p2)γ. Note that by Lemma 3.5 this does indeed define a *-homomorphism of A because the operators of the form φ(sw2,w1) are easily seen to satisfy the relations (3.1). We now prove that φ is an isomorphism from A onto C(Ω)⋊Γ (Theorem 7.7 below). For this some preliminaries are necessary. 19 Lemma 7.3. For any m ∈ Z2 +, 1 = Xp∈Wm 1Ω(p). Proof. This follows from the discussion in [CMS, Section 2]. (cid:3) Lemma 7.4. The linear span of {1Ω(p); p ∈ W} is dense in C(Ω). Proof. This follows because the sets Ω(p) for p ∈ W form a basis for the topology of Ω [CMS, Section 2]. (cid:3) Lemma 7.5. Let p ∈ Pm where m = (m1, m2) ∈ Z2 +. Let x = p(0, 0) and y = p(m1 + 1, m2 + 1). Let y′ be another vertex of B whose graph distance to y in the 1-skeleton of B equals n. Suppose that n ≤ m1, m2. Then conv(x, y′) contains p(p(m1−n,m2−n)). Proof. Induction reduces us to the case n = 1. There is some apartment containing x and the edge from y to y′. In Figure 11 we show one of the six possible positions for y′. y • y′ • . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. . .. . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. . .. . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. . .. . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . .. . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ... . .. .. . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . .. . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . • x . Figure 11. Relative positions of x,y and y′ in an apartment. Now conv(x, y′) is the image of some p′ ∈ P(m1+1,m2−1). It is evident in this case (and in the (cid:3) other five cases) that p′(p(m1−1,m2−1)) = p(p(m1−1,m2−1)). Corollary 7.6. Let p ∈ Pm for m = (m1, m2). Let x = p(m1 + 1, m2 + 1) and y = p(0, 0). Let y′ be a third vertex of B at distance n to y. Suppose n ≤ m1, m2. Then conv(x, y′) is the image of some p′ ∈ P, where p′(0, 0) = y′ and t(p′) = t(p). Proof. See Figure 12. According to Lemma 7.5, the convex hull conv(x, y′) contains t(p)(t). Be- cause conv(x, y′) contains a chamber, it must be the image of a unique p′ ∈ P with p′(0, 0) = y′. Since t(p)(t) lies in the image of p′, namely conv(x, y′), and t(p)(1, 1) = x, it follows that t(p′) = t(p). (cid:3) Theorem 7.7. The map φ is an isomorphism from A onto C(Ω) ⋊ Γ. 20 GUYAN ROBERTSON AND TIM STEGER x t(p)(t) .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . .. . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . .. .. . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . . . .. . . .. . . . . . . . .. .. . . . . . . . . . . .. . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . .. ... .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . .. . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . .. . . . . . . .. . . . . . . . . . . . . . . . . . .. .. . . . . . .. . . . . . . .. . . . . . . . . . . . . . .. . . . .. . . . . . . .. . . . . . . . . . . .. . . . . . . . . . .. .. . . . . . . . . . .. . . . . . . . . . . .. . . . . . .. .. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . .. . .. .. . . . . . . . . . .. . . . . . . . . . . . . . . . . .. .. . . .. . . . . . . . . . . .. . . . . . . . . . . . . .. . . . .. . . . .. . . . . . . . . . . .. . . . . . . . . . . . . .. .. . . . . . . . . . . .. . . . . . . .. . . . . . . . . .. .. . . . . . . . . . . . . . . ... . . . . . . .. . . . . . . . . .. .. . . . . . . . . . . .. . . . . . . . . . . .. . . . . .. .. . . . . . . . . . . . . . . .. . . . . . .. . . . . . . . . . .. .. . . . . . . . . . . .. . . . . . . . . . . .. . . . . .. .. . . . . . . .. . . . . . . . . . . .. . . . . .. ... . . . . . . . . . . . . . . . . . . .. . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . .. . . . . . . . . . . . .. .. . .. . . . . . . . . . . . . .. . . . . . . . . . . . . . . . .. .. . .. . . . . . . . . .. . . . . . . . . . . . . . . . .. .. . .. . . . . . . . . . . . . .. . . . . . . . . . . . .. . . . .. . .. . . . . . . . . . .. . . . . . . . . . . . . . . . .. .. . . . .. . . . . . . . . . . .. . . . . . . . . . . . .. .. . . . . . . . .. . . . . . . .. . . . . . . . . . . . . . . . .. .. . . . . . . . .. . . .. . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . .. . . .. . . . . . . . . . . . .. . . . .. . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . .. . . .. . . . . . . . . . . . .. . . . .. . . . . . . . . . . . .. . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . y′ y Figure 12. Proof. Since A is simple, φ is injective. If p ∈ W then 1Ω(p) = φ(sw,w), where w = α(p) and so Lemma 7.4 shows that the range of φ contains C(Ω). It remains to show that it contains Γ. Fix γ1Ω(p). For γ ∈ Γ. Choose m = (m1, m2) so that d(O, γ−1O) ≤ m1, m2. Now γ = γ1 = Xp∈Wm p ∈ Wm we claim that γt(p) = t(p′) for some p′ ∈ W. Hence γ1Ω(p) = φ(cid:0)sα(p′),α(p)(cid:1). This shows the range of φ contains Γ and hence is surjective. To prove the claim above, apply Corollary 7.6 to find p′′ ∈ P with p′′(0, 0) = γ−1(O) and (cid:3) t(p′′) = t(p). Let p′ = γp′′. Then p′(0, 0) = O, so p′ ∈ W and t(p′) = γt(p′′) = γt(p). Remark 7.8. It follows from Theorem 7.7 and Remark 6.5 that C(Ω) ⋊ Γ is simple, nuclear and purely infinite. Simplicity and nuclearity had previously been proved in [RS1] under the additional assumption that Γ acts transitively and in a type rotating manner on the vertices of B. From simplicity it follows that C(Ω) ⋊ Γ is isomorphic to the reduced crossed product C(Ω) ⋊r Γ. See also [An, QS] for general conditions under which the full and reduced crossed products coincide. Their results apply, for example when Γ is a lattice in a linear group. 7.1. Conditions (H0)-(H3) for affine buildings of type eA2. Continue the notation and termi- nology used above, assuming throughout that Γ acts on B via type rotating automorphisms. We prove that conditions (H0), (H1), and (H3) are satisfied so long as Γ acts freely and with finitely many orbits on the vertices of B. Moreover, if B is the building of G = PGL3(K), where K is a nonarchimedean local field of characteristic zero and Γ is a lattice in PGL3(K) we prove that (H2) holds as well. There are several concrete examples in [CMSZ] where all these hypotheses are satisfied. Proposition 7.9. Suppose that Γ acts freely and with finitely many orbits on the vertices of B. Then the matrices M1, M2 of the previous section satisfy conditions (H0),(H1), and (H3). Proof. By definition M1 and M2 are {0, 1}-matrices. To say that they are nonzero is to say that P(1,0) and P(0,1) are nonempty, which the are. This proves (H0). Fix any nonzero j ∈ Z2. We will construct w ∈ W which is not j-periodic. Choose m ∈ Z2 large enough so that inside pm one can find a minimal gallery of chambers, (C0, C1, . . . , Cl) so that Cl is the j-translate of C0. Write τ : C0 → Cl for the identification by translation of the two chambers. Construct an isometry p from this minimal gallery to B by defining successively pC0, pC1, etc. Since the building is thick, one has at least two choices at each step. Once pCl−1 is fixed, no two of the choices for pCl can be in the same Γ-orbit, since Γ acts freely on the vertices of B. Therefore, one may choose p so that pC0 and p ◦ τ are in different Γ-orbits. Now extend p to an isometry p : pm → B. The element of W associated to p, that is α(Γp), is not j-periodic. This proves (H3). Condition (H1c) is vacuous for r = 2. Consider the configuration of Figure 13. Given the tiles a, b, and c, there is exactly one tile d which completes the picture. Since a, b, and c make up a minimal gallery, this follows by the same argument used in proving Lemma 7.1. Translating this fact to matrix terms, we have that if (M2M1)(c, a) > 0 then (M1M2)(c, a) = 1. Likewise, if (M1M2)(c, a) > 0, then (M2M1)(c, a) = 1. Conditions (H1a) and (H1b) follow, and by Lemma 1.4, so does condition (H1). (cid:3) 21 . . . .. . . . . . . . . . . . .. . . . . . . .. . . .. . . . . . . .. . . .. . . . . . . .. . . c . . . . .. . . . . . . . .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. . .. . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . .. . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . .. . .. .. . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. ... .... . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . .. .. .. .. . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. .. . . . . . . . . . . . . .. a d b Figure 13. Uniqueness of extension. It is not much harder to prove (H1) directly, bypassing conditions (H1a)-(H1c). It remains to prove condition (H2). The next result and its corollary prove a strong version of condition (H2). Theorem 7.10. Let B be the building of G = SL3(K), where K is a local field of characteristic zero. Let Γ be a lattice in SL3(K) which acts freely on the vertices of B with finitely many orbits. Let the alphabet A = Γ\T and the transition matrices M1,M2 be defined as at the beginning of Section 7. Then for each i = 1, 2, the directed graph with vertices a ∈ A and directed edges (a, b) whenever Mi(b, a) = 1 is irreducible. Proof. We use an idea due to S. Mozes [M2, Proposition 3]. Fix a model half-infinite strip s of tiles in a based at (0, 0) and let sk = p(k,0) be the initial segment consisting of k + 1 tiles. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . .. . . . . . . . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . .. . . . . . . . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . .. . . . . . . . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . .. . . . . . . . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . .. . . . . . . . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . . . .. . . . . . . . .. . . . . .. . . . . . . Figure 14. The strip s. Let S0 [respectively Sk, k ≥ 1] be the set of type-preserving isometric embeddings s of s [respectively sk] into the building B. Thus Sk is the subset of type-preserving maps in P(k,0) and if s ∈ Sk, then o(s) and t(s) are defined as before. Also, if s ∈ S define o(s) = st and let sk = t(ssk ) for k ≥ 0. It is desired to prove that given a, b ∈ A, there exists k ∈ Z+ and s ∈ Sk such that a = Γo(s) and b = Γt(s). The group G acts transitively on the set of apartments of B [St, Section 5]. Moreover the stabilizer of an apartment acts transitively on the set of sectors of the apartment. It follows that G acts transitively on S0. Therefore S0 = G/H0 for some subgroup H0. In fact Say that two elements of S0 are equivalent if, beyond a certain point dependent on the two elements, they agree on all tiles of s. Let S be the space of equivalence classes. Since G acts transitively on S0, a fortiori it acts transitively on S. Thus S = G/H for some H. In fact a1 0 0 H0 = { H = {  ∈ G; aj = 1, bj, c ≤ 1, and d < 1}.  ∈ G; aj = 1, c ≤ 1, and d < 1}. b2 a2 d b3 c a3 a1 0 0 b3 b2 a2 c d a3 22 GUYAN ROBERTSON AND TIM STEGER The only relevant facts about H are that it is closed and noncompact. On S0 and S we put the topologies and measures obtained from the isomorphisms with G/H0 and G/H. The Howe -- Moore Theorem [Z, Theorems 10.1.4 and 2.2.6] shows that Γ acts ergodically on S. Suppose that there exist a, b ∈ A which cannot occur as a = Γo(s) and b = Γt(s) with s ∈ Sk, 0 to S. The sets S′ for any k. Let S′ 0 and S′ are clearly Γ-invariant. The set S′ is open, since S′ 0 is open. Therefore S′ is not of null measure. By the ergodicity of the Γ-action S′ has full measure. Also of full measure will be the inverse image of S′ in S0, which consists of all those s′ ∈ S0 which are equivalent to some s ∈ S0 with Γo(s) = a. A fortiori 0 = {s ∈ S0; Γo(s) = a} and let S′ be the projection of S′ {s ∈ S0; there exists K ≥ 0 such that b 6= Γsk for all k ≥ K} is of full measure in S0. Now, in Γ\S0 the set {Γs ∈ Γ\S0; there exists K ≥ 0 such that b 6= Γsk for all k ≥ K} will also be of full measure. On Γ\S0 use the measure obtained in the usual way from the unique (up to positive constant) positive G-invariant measure on S0. The following condition defines the new measure, d s, in terms of the old measure, ds: ZΓ\S0Xγ∈Γ F (γ(s)) d s =ZS0 F (s) ds for any F ∈ Cc(S0). Since Γ\G is of finite total measure, it follows that Γ\S0 is too. Assume that this total measure is one. One can easily verify that relative to d s the distribution of Γsk is independent of k. In fact, {Γs ∈ Γ\S0; Γsk = b} = 1 #A for all k ≥ 0. Let 0 < ǫ < 1 #A. The monotone convergence theorem implies that there exists K such that {Γs ∈ Γ\S0; b 6= Γsk for all k ≥ K} > 1 − ǫ. But this means that {Γs ∈ Γ\S0; ΓsK = b} ≤ ǫ. This contradicts {Γs ∈ Γ\S0; ΓsK = b} = 1 #A , and so proves the result. (cid:3) Corollary 7.11. Let B be the building of G = PGL3(K), where K is a local field of characteristic zero. Let Γ be a lattice in PGL3(K) which acts freely on the vertices of B. Then the conclusions of Theorem 7.10 hold. Proof. The image of SL3(K) in P GL3(K) has finite index. Let Γ′ be the pullback to SL3(K) of Γ. Then Γ′\SL3(K) also has finite volume and the proof of Theorem 7.10 applies. Moreover, the Γ-orbits of tiles of B are made up of unions of Γ′-orbits. So if we wish to construct s ∈ Sk having first and last tiles in certain Γ-orbits, we just pick Γ′-orbits contained in the two Γ-orbits and thereafter work with Γ′. (cid:3) Remark 7.12. We needed to use an indirect argument in the previous Corollary because the Howe -- Moore theorem does not apply in its simplest form to P GL3(K). Remark 7.13. In work which will appear elsewhere, it will be shown how to extend the methods of the proof of the Howe -- Moore theorem so as to prove the necessary ergodicity in greater generality. It is enough to suppose that Γ acts freely and with finitely many orbits on the vertices of a thick Theorem 7.7 is likewise true in this generality. Not only does this allow one to work with the buildings associated to PGL3(K) when K has building of type eA2. Since ergodicity implies (H2), and since (H0), (H1), and (H3) always hold, positive characteristic, but it also makes available those buildings of type eA2 which are associated the eA2 groups listed in [CMSZ]. to no linear group. Note finally that direct combinatorial proofs of (H2) can be constructed for 23 7.2. Examples of type eA1 × eA1. Analogous results hold for groups acting on buildings of type eA1×eA1. Consider by way of illustration a specific example studied in [M1] and generalized in [BM]. In [M1, Section 3], there is constructed a certain lattice subgroup Γ of G = P GL2(Qp)×P GL2(Qq), where p, q ≡ 1 (mod 4) are two distinct primes. The building ∆ of G is a product of two homogeneous trees T1, T2, so that the chambers of ∆ are squares and the apartments are copies of the euclidean plane tesselated by squares. If a ∈ Γ then there are automorphisms a1, a2 of T1, T2, respectively such that a(u, v) = (a1u, a2v) for each vertex (u, v) of ∆. However, even though each a ∈ Γ is a direct product of automorphisms, the group Γ is not a direct product of groups Γ1 and Γ2. The group Γ acts freely and transitively on the vertices of ∆. The preceding results all extend to this situation. The tiles are now squares instead of parallelograms. The boundary of ∆ is defined as before, using π 2 -angled sectors. The condition (H1) is a consequence of [M1, Theorem 3.2] and the irreducibility condition (H2) follows from [M2, Proposition 3]. References [An] C. Anantharaman-Delaroche, Syst`emes dynamiques non commutatifs et moyennabilit´e, Math. Ann. 279 [Bl] (1987), 297-315. B. Blackadar, K-theory for Operator Algebras, Second Edition, MSRI Publications 5, Cambridge University Press, Cambridge, 1998. [BM] M. Burger and S. Mozes, Finitely presented groups and products of trees, C. R. Acad. Sci. Paris, S´er. 1 324 (1997), 747 -- 752. [Br1] K. Brown, Buildings, Springer-Verlag, New York, 1989. [Br2] K. Brown, Five lectures on buildings, Group Theory from a Geometrical Viewpoint (Trieste 1990), 254 -- 295, World Sci. Publishing, River Edge, N.J., 1991. [Ca] D. I. Cartwright, A brief introduction to buildings, Harmonic Functions on Trees and Buildings (New York [C1] [C2] [CK] 1995), 45 -- 77, Contemp. Math. 206, Amer. Math. Soc., 1997. J. Cuntz, A class of C ∗-algebras and topological Markov chains: Reducible chains and the Ext-functor for C ∗-algebras, Invent. Math. 63 (1981), 23-50. J. Cuntz, K-theory for certain C ∗-algebras, Ann. of Math. 113 (1981), 181-197. J. Cuntz and W. Krieger, A class of C ∗-algebras and topological Markov chains, Invent. Math. 56 (1980), 251-268. [CMS] D. I. Cartwright, W. M lotkowski and T. Steger, Property (T) and eA2 groups, Ann. Inst. Fourier 44 (1993), [CMSZ] D. I. Cartwright, A. M. Mantero, T. Steger and A. Zappa, Groups acting simply transitively on the vertices 213 -- 248. [G] [K1] [K2] of a building of type eA2, I and II, Geom. Ded. 47 (1993), 143 -- 166 and 167 -- 223. P. Garrett, Buildings and Classical Groups, Chapman & Hall, London, 1997. E. Kirchberg, Exact C ∗-algebras, tensor products, and the classification of purely infinite algebras, Proceed- ings of the International Congress of Mathematicians (Zurich, 1994), Vol. 2, 943 -- 954, Birkhauser, Basel, 1995. E. Kirchberg, The classification of purely infinite C ∗-algebras using Kasparov's theory, in Lectures in Op- erator Algebras, Fields Institute Monographs, Amer. Math. Soc., 1998. [KR] R. V. Kadison and J. R. Ringrose, Fundamentals of the Theory of Operator Algebras, Volume II, Academic Press, New York, 1986. S. Mozes, Actions of Cartan subgroups, Israel J. Math. 90 (1995), 253 -- 294. S. Mozes, A zero entropy, mixing of all orders tiling system, Contemp. Math. 135 (1992), 319 -- 325. [M1] [M2] [Ph] N. C. Phillips, A classification theorem for purely infinite simple C ∗-algebras, preprint, Oregon 1995. [Ped] G. K. Pedersen, C ∗-algebras and their Automorphism Groups, Academic Press, New York, 1979. [QS] J. C. Quigg and J. Spielberg, Regularity and hyporegularity in C ∗-dynamical systems, Houston J. Math. 18 (1992), 139-152. [RS1] G. Robertson and T. Steger, C ∗-algebras arising from group actions on the boundary of a triangle building, Proc. London Math. Soc. 72 (1996), 613 -- 637. [RS2] G. Robertson and T. Steger, K-theory for rank two Cuntz-Krieger algebras, preprint. [Ron] M. Ronan, Lectures on Buildings, Perspectives in Mathematics, Vol. 7, Academic Press, New York, 1989. [Sp] J. Spielberg, Free product groups, Cuntz-Krieger algebras, and covariant maps, International J. Math. 2 (1991), 457-476. T. Steger, Local fields and buildings, Harmonic Functions on Trees and Buildings (New York 1995), 79 -- 107, Contemp. Math. 206, Amer. Math. Soc., 1997. [St] 24 [Z] GUYAN ROBERTSON AND TIM STEGER R. J. Zimmer, Ergodic Theory and Semisimple Groups, Birkhauser, Boston, 1984. Mathematics Department, University of Newcastle, Callaghan, NSW 2308, Australia E-mail address: [email protected] Istituto Di Matematica e Fisica, Universit`a degli Studi di Sassari, Via Vienna 2, 07100 Sassari, Italia E-mail address: [email protected]
1810.05267	1	1810	2018-10-11T21:51:30	Cartan Triples	[ "math.OA" ]	We introduce the class of Cartan triples as a generalization of the notion of a Cartan MASA in a von Neumann algebra. We obtain a one-to-one correspondence between Cartan triples and certain Clifford extensions of inverse semigroups. Moreover, there is a spectral theorem describing bimodules in terms of their support sets in the fundamental inverse semigroup and, as a corollary, an extension of Aoi's theorem to this setting. This context contains that of Fulman's generalization of Cartan MASAs and we discuss his generalization in an appendix.	math.OA	math	CARTAN TRIPLES ALLAN P. DONSIG, ADAM H. FULLER, AND DAVID R. PITTS Abstract. We introduce the class of Cartan triples as a generalization of the notion of a Car- tan MASA in a von Neumann algebra. We obtain a one-to-one correspondence between Cartan triples and certain Clifford extensions of inverse semigroups. Moreover, there is a spectral theorem describing bimodules in terms of their support sets in the fundamental inverse semigroup and, as a corollary, an extension of Aoi's theorem to this setting. This context contains that of Fulman's generalization of Cartan MASAs and we discuss his generalization in an appendix. 1. Introduction As observed in the seminal work of Feldman-Moore [14, 15], when a von Neumann algebra contains a Cartan MASA, strong structural results about the algebra may be obtained. However, many von Neumann algebras do not contain a Cartan MASA; the first examples were found in [32]. Determining which von Neumann algebras have a Cartan MASA and when it is unique is an important question and has attracted significant attention; for two examples, see [27, 28]. Part of the interest is that Cartan MASAs are closely connected to crossed product decompositions, as indeed is clear from the work of Feldman-Moore. Recall a Cartan MASA D in a von Neumann algebra M is a maximal abelian subalgebra with two additional properties: it is regular, that is, the span of its normalizers is weak-∗ dense in M; and there is a faithful, normal conditional expectation from M to D. In this paper, we study a much larger family of regular abelian von Neumann subalgebras of von Neumann algebras. Specifically, if M is a von Neumann algebra, we consider an abelian and regular subalgebra D ⊆ M such that there is a faithful normal conditional expectation onto the relative commutant Dc of D in M. Because Dc plays an important role in the structure of the algebras, we name it N and call (M, N, D) a Cartan triple. In our previous work [12], we showed that Cartan MASAs can be described in terms of certain ex- tensions of inverse semigroups. In the setting of Cartan triples, our main result is a correspondence between Cartan triples and a larger class of extensions of inverse semigroups P ֒→ G q ։ S. Further, we obtain a Spectral Theorem for N-Bimodules and a version of Aoi's theorem in this context of Cartan triples. Although some of the methods from [12] extend naturally, significant modifications are needed. For example, for Cartan triples, P is not usually an abelian inverse semigroup, but rather is Clifford, that is, the idempotents of P commute with all elements of P. Various generalizations of MASAs have been considered in the literature. For example in [13], Ruy Exel connects the existence of a suitable non-abelian generalization of a MASA in a separable C ∗-algebra to a reduced crossed product decomposition of the containing algebra. Instead of the inverse semigroup approach considered here, Exel considers a Fell bundle over an inverse semigroup as the classifying structure. The appropriate variant of Exel's notion of a generalized Cartan subalgebra in the von Neumann algebra setting is a full Cartan triple, meaning D is the center of 2010 Mathematics Subject Classification. Primary 46L10, Secondary 06E75, 20M18, 20M30, 46L51. Key words and phrases. Von Neumann algebra, Bimodule, Cartan MASA. This work was partially supported by a grant from the Simons Foundation (#316952 to David Pitts). 1 N (see Definition 2.3). Our approach using extensions of inverse semigroups for classification, while related to Fell bundles, is rather different. Another generalization, also related to crossed product decompositions, has been considered by Igor Fulman in [17]. Fulman's generalization of a Cartan MASA is, in our terms, a Cartan triple with an additional condition, the existence of a subgroup of the unitaries in M that normalizes D, contains the unitaries of N, and has a suitable fixed point property. We show in Appendix A that this additional condition can be characterized as the existence of a lift of an inverse semigroup homomorphism from S into the partial automorphisms of the Cartan triple. Crossed products by inverse semigroups were first introduced by N´andor Sieben in [29] using such a semigroup homomorphism into the partial automorphisms of a C∗-algebra. Thus, Fulman's condition can be interpreted naturally as saying that the containing algebra is a crossed product by a suitable inverse semigroup. Fulman's starting point was to generalize the Feldman-Moore characterization of Cartan subalgebras [14, 15] using measured equivalence relations. While some of our results resemble Fulman's, ours are more general, perhaps because of the comparative simplicity of the inverse semigroup approach used here. We now discuss our results and their motivation in more detail. We associate to each Cartan triple an extension of inverse semigroups, P ֒→ G → S where S is a fundamental inverse semigroup, that is, the only elements commuting with the idempotents of S, denoted E(S), is E(S) itself, and P is Clifford, meaning all elements of P commute with E(S). To be an extension, the restriction to idempotents of the maps above must be isomorphisms. It is well known that every inverse semigroup G may be represented as such an (idempotent-separating) extension; see [21, p. 141]. To construct the extension from a Cartan triple, take P to be the partial isometries in N that normalize D and G to the partial isometries in M that normalize D, with P ֒→ G the inclusion map. To construct S, we identify elements with the same action on the idempotents, i.e., we quotient by the Munn congruence. In [12], the inverse semigroup P was abelian and we required that the character space of E(P) was hyperstonean. In that case, it was easy to recover D from P, as the continuous functions on the character space of E(P). Here, we need a condition that allows us to again recover Dc = N from P: P arises as the partial isometries in a von Neumann algebra N which normalize a fixed von Neumann subalgebra of the center of N. In this case, we say that P is an N-Clifford inverse monoid (Definition 2.7). To see the need for this condition, consider the (degenerate) Cartan triple, (M, M, CI). In this id → U(M) → CI. However, there are von Neumann case, the associated extension has the form U(M) algebras not isomorphic to their opposite algebras [10]. The unitary groups of such an algebra and its opposite are isomorphic, so if the extension was defined purely in terms of inverse semigroups (and without our stronger condition) it would be possible for non-isomorphic triples to produce the same extension. With these definitions in hand and the construction of an extension from a triple (as outlined above), we show that Cartan triples are isomorphic if and only if their extensions are (in a suitable sense) isomorphic, Theorem 2.22. To obtain the converse, we construct a Cartan triple from an extension in Section 3. This is more subtle, and we build on the strategy of our previous paper. In particular, we use the order structure of S to construct a reproducing kernel Hilbert N-bimodule, A. We then define a representation, λ, of G, Theorem 3.2, by partial isometries on A. After tensoring this representation with a faithful normal representation of N to obtain a suitable representation of G (Theorem 3.2), we define the Cartan triple of an extension in terms of the double commutants of G, P, and their (common) idempotents (Definition 4.1). In Theorem 4.11 we complete the circle of ideas by showing that the extension associated to the Cartan triple constructed is (isomorphic to) the original extension. 2 In Section 5 we begin a study of the N-bimodules in a Cartan triple (M, N, D). For our strongest results we require that D be as large as possible, that is, D is the center Z(N) of N. We call such a Cartan triple full. When (M, D) is a Cartan pair, (M, D, D) is a full Cartan triple, and so the class of full Cartan triples properly includes Cartan pairs. Different examples arise when M is type I, Section 6.1, and when M = N ⋊α G is a crossed product of N by a discrete group G which acts by properly outer automorphisms on N and Z(N), Theorem 6.3. Let (M, N, D) be a full Cartan triple, with associated extension P ֒→ G q ։ S. We show in Theorem 5.2 that if B ⊆ M is a non-zero weak-∗ closed N-bimodule, then G ∩ B 6= {0}. Thus, every weak-∗ closed N-bimodule gives rise to a non-trivial subset q(B ∩ G) of S. We call such sets spectral sets, (Definition 5.1). If A ⊆ S is a spectral set, then span{q−1(A)} is an N-bimodule in M. Of course, it is conceivable that distinct weak-∗ closed N-bimodules have the same spectral sets. To study this, we use the Bures-topology on M, induced by the conditional expectation E : M → N. Whilst the weak-∗ and Bures topologies on M are not, in general, comparable, the Bures-closed N- bimodules are weak-∗ closed [6, Lemma 3.1]. The advantage of the Bures topology over the weak-∗ topology is that certain Fourier-type series often converge in the Bures topology, while they need not converge in the weak-∗ topology (or in any other "natural" topology). Indeed, Mercer showed in [23] that the Fourier series of elements in crossed-product von Neumann algebra converge in the Bures-topology, but need not converge in the weak-∗ topology. Analogously, when (M, N, D) is a Cartan triple, we show in Theorem 5.7 that if x ∈ M, then x is the Bures-limit of the net of finite sums Xu∈F ⊆GN(M,D) uE(u∗x). Similar results for x in a Cartan pair are given in [5, Proposition 2.4.4] and [24, Theorem 4.4]. In Proposition 5.8 we show that if B is a weak-∗ closed N-bimodule, B0 = spanwk∗ {B ∩ G}, and B1 = spanBures{B ∩ G}, then B0 ⊆ B ⊆ B1 and each of B0, B and B1 give the same spectral sets. We do not address when the bimodules B0, B and B1 are necessarily equal. In [5], if all weak-∗ bimodules are necessarily Bures closed, the Cartan pair is said to satisfy spectral synthesis. Even in the case of Cartan pairs, whether B0 = B1 remains an open problem. There are some special cases when the result is known. If (M, N, D) is a Cartan triple of the type studied by Fulman [17] discussed above, with the added condition that M is constructed from a hyperfinite equivalence relation, then it can be deduced from Theorem 5.10 and [17, Theorem 15.18] that all weak-∗ closed N-bimodules are necessarily Bures closed. Cameron and Smith [6, 8] studied a related problem in crossed-products. They showed that if G is a discrete group satisfying the AP condition, acting on a von Neumann algebra N by properly outer automorphisms, then the weak-∗ closed N-bimodules in N ⋊α G are necessarily Bures closed. We give a Spectral Theorem for Bimodules in Theorem 5.10, which gives a one-to-one corre- spondence between the Bures-closed N-bimodules and the spectral sets in S. We find it striking that Theorem 5.10 depends only on S and not on the extension G. Fuller and Pitts [16] had previ- ously observed a similar phenomenon: non-isomorphic Cartan pairs that have isomorphic lattices of Bures-closed bimodules. Theorem 5.10 generalizes the Spectral Theorem for Cartan pairs found in [12]; see also [5]. It should be noted that the study of bimodules in Cartan pairs was initiated in the seminal work of Muhly, Saito and Solel [25]. They present a spectral theorem for weak-∗ closed bimodules. Their work, however, has a gap. Though not explicitly stated as such, the gap in [25] amounts to assuming that the weak-∗ closed bimodules are necessarily Bures closed, see [5]. A class of N-bimodules of particular interest are the von Neumann algebras L such that N ⊆ L ⊆ M. In Theorem 5.12 we show that if (M, N, D) is a full Cartan triple and N ⊆ L ⊆ M, then (L, N, D) is again a Cartan triple. This extends Aoi's result for intermediate von Neumann 3 algebras in Cartan pairs [1]. A key step in the proof is showing that an intermediate subalgebra L is necessarily closed in the Bures topology. Thus, Theorem 5.12 together with Theorem 5.10 immediately give a one-to-one correspondence between the intermediate von Neumann subalgebras containing N, and the sub-inverse Cartan monoids of S, Corollary 5.14. We view this as a Galois correspondence-type result; although we do not have a group to hand, there is the Cartan inverse semigroup. Corollary 5.14 should be compared with the following well-known result: If N is a factor and G is a discrete group acting on N by (properly) outer automorphisms, then Izumi, Longo and Popa [18] show that if L is a von Neumann algebra satisfying N ⊆ L ⊆ N ⋊α G then there is a subgroup H of G such that L = N⋊αH; see also [6, 9]. That is, there is a one-to-one correspondence between subgroups of G and the von Neumann algebras M with N ⊆ M ⊆ N ⋊α G. A similar Galois correspondence-type theorem without an explicit group structure has been obtained in [2] for von Neumann algebras generated by a measured equivalence relation and an appropriate cocycle. Cameron and Smith have considered similar questions in [6, 7, 8]. They study crossed products by discrete groups and the bimodule and intermediate algebra structure therein, amongst other things. There is overlap with our work and [8], with neither work subsuming the other. There they let N be any von Neumann algebra and let G be a discrete group acting on N by properly outer automorphisms. If N is abelian, then N is a Cartan MASA in N ⋊α G and so both our settings cover this case. If N is not abelian, but G also acts on Z(N) by properly outer automorphisms then it is shown in Theorem 6.3 (N ⋊α G, N, Z(N)) is a Cartan triple. 2. Cartan triples and their extensions Our main goals in this section are the construction of the extension associated to a Cartan triple, Proposition 2.13, and the result that two such extensions are isomorphic if and only if they arise from isomorphic Cartan triples, Theorem 2.22. We begin by fixing some notation. For M a von Neumann algebra, Z(M) denotes its center and U(M) the unitary elements. For X ⊆ M, Xc denotes the relative commutant, that is, Xc := {m ∈ M : xm = mx for all x ∈ X}. Definition 2.1. Suppose M and L are von Neumann algebras with L ⊆ M. The groupoid normal- izer of L in M is the set, GN(M, L) := {v ∈ M : v is a partial isometry and v∗Lv ∪ vLv∗ ⊆ L}. If the linear span of GN(M, L) is weak-∗ dense in M, we say L is a regular subalgebra of M or that L is regular in M. Remark 2.2. When L ⊆ M is an abelian von Neumann subalgebra of M, it is more common to say L is regular in M if span{U ∈ U(M) : U DU ∗ = D} is weak-∗ dense. However, if L is abelian, then (2.1) thus the two definitions coincide in this case. For a proof of this statement see [5, p. 479, Inclu- sion 2.8]. span{U ∈ U(M) : U DU ∗ = D} = span GN(M, D), We now introduce our main topic of study. Definition 2.3. A Cartan triple is a triple (M, N, D) consisting of three von Neumann algebras satisfying: (a) D is an abelian and regular von Neumann subalgebra of M; (b) N is the relative commutant of D in M; and (c) there exists a faithful normal conditional expectation E : M → N. A Cartan triple (M, N, D) is full when D = Z(N). 4 Remarks 2.4. (a) For any Cartan triple (M, N, D), M ⊇ N ⊇ D because D is abelian. (b) If (M, N, D) is a Cartan triple, then (M, N, Z(N)) is a full Cartan triple. Indeed, N = Dc implies N = (Z(N))c, and since GN(M, D) ⊆ GN(M, Z(N )), Z(N) is regular in M. Section 6 is devoted to examples of Cartan triples. Here we content ourselves with making two simple observations regarding what occurs when two of the von Neumann algebras in a Cartan triple coincide. Examples 2.5. (a) Suppose (M, D, D) is a Cartan triple. Then (M, D, D) is full and D is a Cartan MASA in M. In this sense, the class of Cartan triples includes the class of Cartan pairs. (b) Now suppose (N, N, D) is a Cartan triple. Then Dc = N. When N has separable predual, Nx dµ(x) as a direct integral. When this is Nx dµ(x) such that we may write D = L∞(X, µ) and write N =R ⊕ done, GN(N, D) may be identified with the set of all functions f ∈R ⊕ for almost every x ∈ X, f (x) ∈ U(Nx) ∪ {0}. X X Example 2.5(b) shows how the inverse semigroup GN(N, D) can be used to describe a direct integral. Further, this inverse semigroup approach allows one to work with von Neumann algebras which do not have separable predual. This example discussed further in Example 2.15. We fix some notation for inverse semigroups next. For the most part, our notation follows Section 2 of [12], which also gives much of the inverse semigroup theory we will use. For an in-depth text on inverse semigroups see [21]. Throughout the paper: • E(S) will denote the idempotents of the inverse semigroup S; • we use s† to denote the inverse of the element s in an abstract inverse semigroup; however, for an inverse semigroup of partial isometries on a Hilbert space, the adjoint v∗ is the inverse of the element v and we typically use v∗ instead of v† in this setting. For our extensions, we need two special classes of inverse monoids: Cartan inverse monoids, defined in [12], and N-Clifford inverse monoids, which are new. Definition 2.6 ([12, Definition 2.11]). We call an inverse semigroup S a Cartan inverse monoid if (a) S is fundamental; (b) S is a complete Boolean inverse monoid; and space. (c) the character space dE(S) of the complete Boolean lattice E(S) is a hyperstonean topological Definition 2.7. Let N be a von Neumann algebra. An N-Clifford inverse monoid is an inverse monoid P such that P = GN(N, D), where D is a von Neumann subalgebra of Z(N). If in addition D = Z(N), we say P is a full N-Clifford inverse monoid. Remark 2.8. Suppose P = GN(N, D) is an N-Clifford inverse monoid. It is not difficult to show that P = {v ∈ N : v is a partial isometry with vv∗ = v∗v ∈ D} and E(P) = proj(D). In particular, P is a Clifford inverse semigroup of partial isometries. Since U(N) ⊆ P, every element of N is a linear combination of at most four elements of P. We need an appropriate notion of isomorphism of such inverse monoids. Definition 2.9. If for i = 1, 2, Pi are Ni-Clifford inverse monoids, a map α : P1 → P2 is an extendible isomorphism if there exists a normal ∗-isomorphism θ : N1 → N2 such that α = θP1; equivalently, there exists a normal isomorphism θ : N1 → N2 such that θ(D1) = D2. 5 Obviously, any extendible isomorphism is an isomorphism of inverse semigroups. Definition 2.10. For i = 1, 2, let Si be Cartan inverse monoids, let Pi be Ni-Clifford inverse q ։ Si are extensions of Si by Pi. These extensions are equivalent monoids, and suppose Pi if there are semigroup isomorphisms α : G1 → G2, α : S1 → S2 and an extendible isomorphism α : P1 → P2 such that ιi֒−→ Gi α ◦ ι1 = ι2 ◦ α and q2 ◦ α = α ◦ q1. Remark 2.11. When Pi is the set of partial isometries in C([E(Si)) and Ni := C([E(Si)), then this definition reduces to the notion of equivalence for extensions found in [12]. Definition 2.12. Let G be an inverse semigroup. The Munn congruence (also called the Munn relation) on G is the set RM := {(v1, v2) ∈ G × G : v1ev† 1 = v2ev† 2 for all e ∈ E(G)}. Then RM is the maximal idempotent separating congruence on G and the set of RM -equivalence classes equipped with the product [v][w] = [vw] and inverse [v]† = [v†] form a fundamental inverse semigroup [21, Proposition 5.2.5]. We now show how a Cartan triple gives rise to an extension of inverse semigroups. Proposition 2.13. Let (M, N, D) be a Cartan triple and set G := GN(M, D) and P := GN(N, D). Then G and P are inverse semigroups with P ⊆ G and E(P) = E(G) = proj(D). Moreover, the following statements hold. (a) P is a N-Clifford inverse monoid. (b) P is the set of elements of G Munn-related to an idempotent. (c) If S is the quotient of G by the Munn congruence, then S is a Cartan inverse monoid. Proof. Obviously, P ⊆ G and by definition, P is an N-Clifford inverse monoid. If v ∈ G, then v∗v ∈ D, so every idempotent of G is a projection in D. Also, every projection in D is an idempotent in G. It follows that G and P are von Neumann regular monoids for which the idempotents commute, so both are inverse monoids and E(P) = E(G) = proj(D). If v ∈ P, then (v, vv∗) ∈ RM , so every element of P is Munn-related to an idempotent. On the other hand, suppose v ∈ GN(M, D) is Munn-related to the idempotent e. Then vv∗ = e. Let p be a projection in D. Then vp = vpv∗v = epv = pev = pv. Hence v ∈ N, and so v ∈ P. Thus, P is the set of elements of G Munn related to an idempotent. We have already observed that S is a fundamental inverse monoid, and it clearly contains a zero element 0. As the Munn congruence is idempotent separating, E(S) is isomorphic to E(P). The proof that S is a Cartan inverse monoid now follows exactly as in the proof of [12, Proposition 3.5]. (cid:3) As noted in the introduction, any inverse semigroup G may be represented as an extension of a fundamental inverse semigroup S by a Clifford inverse semigroup C. Indeed, C may be taken to be the set of elements of G which are Munn-related to an idempotent, and S is the quotient of G by the Munn relation. We apply this construction to the inverse semigroup G of Proposition 2.13 to obtain the class of extensions studied in this paper. 6 Definition 2.14. Let (M, N, D) be a Cartan triple. Put G := GN(M, D), P := GN(N, D), S := G/RM , and let q : G → S the quotient map. The extension P ֒→ G q ։ S. (2.2) is called the extension associated to (M, N, D). Example 2.15. We return to the context of Example 2.5(b), that is, of a Cartan triple having the form (N, N, D). In this setting, P = G consists of the partial isometries in N whose initial and final spaces coincide and belong to D; S is the projection lattice of D; and q is the map v 7→ v∗v. Note that N is the linear span of P. When N∗ is separable and N is identified as the direct integral q ։ S as giving a description of the direct integral Nx dµ(x), we may view the extension P ֒→ P R ⊕ X in terms of the linear span of (cid:26)f ∈Z ⊕ X Mx dµ(x) : f (x) ∈ U(Mx) ∪ {0} for almost every x(cid:27) . The extension approach encodes the measure theory into D, and is a more operator theoretic view of N as opposed to the point based view of N as a direct integral. In the study of extensions, it is often useful to choose a section j for the quotient map q, that is, j is a map such that q ◦ j = idS. In our context, we will frequently need a section which is order preserving in the sense that j(1) = 1 and whenever s, t ∈ S and s ≤ t, we have j(s) ≤ j(t) (see [12, Definition 4.1]). Most of the following result was proved in [12] when P is the set of partial isometries in C ∗(E(S)), but the same proof holds for extensions of S by N-Clifford inverse monoids considered here. Recall that qE(G) is a complete Boolean algebra isomorphism of E(G) onto E(S). Also, as observed in [12, Remark 4.8], for any s, t ∈ S, (s†t ∧ 1) is the source idempotent for s ∧ t, that is, (s ∧ t)†(s ∧ t) = s†t ∧ 1. (2.3) q Proposition 2.16 (c.f. [12, Proposition 4.6]). Let P ֒→ G ։ S be an extension of the Cartan inverse monoid S by the N-Clifford inverse monoid P. The map (qE(S))−1 extends to an order preserving section j : S → G for q. This section has the property that for s1, s2 ∈ S, j(s1)†j(s2)j(s† 1s2 ∧ 1) = j(s† 1s2 ∧ 1). (2.4) Proof. Equation (2.4) was not proved in [12], so we provide a proof here. Using (2.3), observe j(s1)†j(s2)j(s† 1s2 ∧ 1) = j(s1)†j(s1 ∧ s2) =(cid:16)j(s1 ∧ s2)j(s1 ∧ s2)†j(s1)(cid:17)† = j(s1 ∧ s2)†j(s1 ∧ s2) = j(s† j(s1 ∧ s2) 1s2 ∧ 1). (cid:3) Definition 2.17. We say that the Cartan triples (M1, N1, D1) and (M2, N2, D2) are isomorphic if there exists a normal ∗-isomorphism θ : M1 → M2 such that θ(D1) = D2. Our goal is to show that Cartan triples, up to isomorphism, are uniquely determined by their associated extensions, up to equivalence. We do this in Theorem 2.22. We first need some technical lemmas. Throughout the remainder of the section, fix an order-preserving section j for the extension P ֒→ G q ։ S associated to the Cartan triple (M, N, D). 7 Lemma 2.18. Let (M, N, D) be a Cartan triple with conditional expectation E : M → N. Let P ֒→ G q ։ S be the associated extension. Then E(G) = P. Further, for v ∈ GN(M, D), E(v) = ve, where e = j(q(v) ∧ 1) ∈ E(G). Also, E preserves the natural inverse semigroup partial order on G and P in the sense that if v, w ∈ G with v ≤ w, then E(v) ≤ E(w). Proof. Each v ∈ G induces a normal ∗-isomorphism θv from v∗vD to vv∗D, given by θv(d) = vdv∗. Applying Frol´ık's Theorem [26, Proposition 2.11A] to θv we may find elements e0, e1, e2, e3 ∈ E(G) such that (a) for i 6= j, ei ∧ ej = 0; (b) e0 ∨ e1 ∨ e2 ∨ e3 = v∗v; (c) for i = 1, 2, 3, (vei)2 = 0; and (d) q(ve0) ∈ E(S). If w ∈ G and w2 = 0, then E(w) = ww∗E(w) = E(w)ww∗ = E(www∗) = 0. It follows that E(v) = E(ve0) = ve0. Let e := e0. Then q(e) = q(v) ∧ 1 because e0 corresponds to the ideal of D consisting of all elements fixed by θv. Finally, if v, w ∈ G and v ≤ w, we may find f ∈ E(G) so that v = wf . Then E(v) = E(w)f , so (cid:3) E(v) ≤ E(w). Lemma 2.19. Let (M, N, D) be a Cartan triple, and suppose y belongs to the linear span of GN(M, D). Then there exists a finite set {wk}m j wk) = 0 for j 6= k and k=1 ⊆ GN(M, D) such that E(w∗ y = wkE(w∗ ky). (2.5) mXk=1 Proof. Choose {vj}N j=1 ⊆ GN(M, D) and scalars {cj}N j=1 so that cjvj 6= 0 for each j and y = k=1 ⊆ S satisfying PN j=1 cjvj. Let si := q(vi) and apply [12, Lemma 4.15] to obtain a finite set {tk}m (a) for 1 ≤ j ≤ m, tj 6= 0; (b) for j 6= k, tj ∧ tk = 0; (c) for 1 ≤ j ≤ m and 1 ≤ n ≤ N , tj ∧ sn ∈ {0, tj }; (d) for 1 ≤ j ≤ m there exists 1 ≤ n ≤ N such that tj ∧ sn = tj; and (e) for each 1 ≤ n ≤ N , sn =W{tj : tj ≤ sn}. Let wk := j(tk). Lemma 2.18 implies that E(w∗ For 1 ≤ n ≤ N , let In := {i : ti ≤ sn}. Given n and i ∈ In, another application of Lemma 2.18 j wk) = 0 when j 6= k. gives Since sn =Wi∈In wiE(w∗ i vnj(t† i vn) = wiw∗ (ti ∧ sn), we obtain, Equation (2.5) follows. i sn ∧ 1) = vn(v∗ nwiw∗ i vn)j(t† i sn ∧ 1) = vnj(t† jsn ∧ 1). vn = mXi=1 wiE(w∗ i vn). 8 (cid:3) We now recall notation regarding weights on a von Neumann algebra used in [30]. Suppose M is a von Neumann algebra and φ is a weight on M. Recall that pφ := {x ∈ M+ : φ(x) < ∞}, nφ := {x ∈ M : φ(x∗x) < ∞} and mφ :=( NXk=1 kxk : n ∈ N, xk, yk ∈ nφ) . y∗ By [30, Lemma VII.1.2], pφ is a hereditary convex cone in M+, nφ is a left ideal of M, mφ is a hereditary ∗-algebra of M, and every element of mφ is a linear combination of four elements of pφ. The semi-cyclic representation πφ of M on the Hilbert space Hφ associated to φ will be denoted (πφ, Hφ, ηφ). When φ is faithful, normal and semi-finite, πφ is a faithful, normal representation of M. Lemma 2.20. Suppose M is a von Neumann algebra, L ⊆ M is a von Neumann subalgebra, and there exists a faithful normal conditional expectation E : M → L. Let ψ be a faithful, normal, semi-finite weight on L and let φ := ψ ◦ E. Then φ is a faithful normal semi-finite weight on M. Proof. Since mψ is a ∗-subalgebra of L, a corollary of the Kaplansky density theorem shows there exists a net (xλ) in pψ with 0 ≤ xλ ≤ I which converges σ-strongly to I. For any z ∈ M+, λ ∈ pψ. Therefore, xλzxλ ∈ pφ. Since limσ-strong xλzxλ = z, we have that pφ xλzxλ ≤ kzk p2 generates M. That is, φ is semi-finite on M. (cid:3) The following result is the key technical tool used in the proof of Theorem 2.22. Lemma 2.21. Let (M, N, D) be a Cartan triple, suppose ψ is a faithful normal semi-finite weight on N and let φ = ψ ◦ E. Then the linear span of {ηφ(vn) : v ∈ GN(M, D) and n ∈ nψ} is dense in Hφ. Proof. For m ∈ nφ and v ∈ GN(M, D), ψ(E(m∗v)v∗vE(v∗m)) = ψ(E(m∗v)E(v∗m)) ≤ ψ(E(m∗vv∗m)) ≤ ψ(E(m∗m)) = φ(m∗m) < ∞. This yields the following. (a) E(nφ) = nψ (take v = I). (b) For v ∈ GN(M, D), the map M ∋ m 7→ vE(v∗m) is idempotent and leaves nφ invariant; hence there is a projection Pv ∈ B(Hφ) whose action on ηφ(nφ) is given by Pvηφ(m) = ηφ(vE(v∗m)). In addition, notice that range Pv = {ηφ(vn) : n ∈ nψ}. To prove the lemma, it therefore suffices to show that if ξ ∈ Hφ and Pvξ = 0 for every v ∈ GN(M, D), then ξ = 0. We begin with a preliminary fact about approximating norms of vectors in Hφ. Let Φ := {τ ◦ E : τ ∈ N+ ∗ and τ (n) ≤ ψ(n) for all 0 ≤ n ∈ N}. Clearly Φ ⊆ M+ ∗ . For ω ∈ Φ, let (πω, Hω, ηω) be the semi-cyclic representation of M arising from ω. This representation is actually cyclic and nω = M because ω is a bounded positive linear functional on M. Define Tω : ηφ(nφ) → Hω by Tωηφ(m) = ηω(m). Write ω = ρ ◦ E for some ρ ∈ N+ ∗ . Then for m ∈ nφ, kηω(m)k2 = ρ(E(m∗m)) ≤ ψ(E(m∗m)) = kηφ(m)k2 . Thus Tω extends to a contraction belonging to B(Hφ, Hω), which we again denote by Tω. We claim that for any ξ ∈ Hφ, kξk = sup ω∈Φ kTωξk . 9 (2.6) To see this, fix ξ ∈ Hφ and choose a real number r such that r < kξk. Let ε > 0 satisfy 3ε < kξk−r. Choose m ∈ nφ such that kξ − ηφ(m)k < ε. By Haagerup's Theorem (see [30, Theorem VII.1.11]), ψ(E(m∗m)) = sup{τ (E(m∗m)) : τ ∈ N+ ∗ and τ (n) ≤ ψ(n) for all 0 ≤ n ∈ N}. Hence there exists ω ∈ Φ such that kTωηφ(m)k > kηφ(m)k − ε. Then kξk ≤ kξ − ηφ(m)k + kηφ(m)k < 2ε + kηφ(m)k − ε < 2ε + kTωηφ(m)k ≤ 3ε + kTωξk < kξk − r + kTωξk , whence r < kTωξk. Thus (2.6) holds. For each ω ∈ Φ and v ∈ GN(M, D), let P ω v be the projection on Hω determined by ηω(m) 7→ ηω(vE(v∗m)), m ∈ M. A routine calculation shows that for every ω ∈ Φ, m ∈ nφ and v ∈ GN(M, D), TωPvηφ(m) = P ω v Tωηφ(m), so TωPv = P ω v Tω. Fix ω ∈ Φ. We claim that if ζ ∈ Hω and P ω (2.7) v ζ = 0 for every v ∈ GN(M, D), then ζ = 0. Suppose ζ ∈ Hω is such a vector. Given ε > 0, there exists x ∈ M such that kζ − ηω(x)k < ε. Since D is regular in M, and ω ∈ M+ ∗ , there exists y ∈ span GN(M, D) such that kηω(x − y)k < ε. By Lemma 2.19, there exists a finite E-orthogonal set {vk}m k=1 ⊆ GN(M, D) such that y = vk is a k=1 is a pairwise orthogonal set of projections, Q := Pm projection. Therefore, ky). As {P ω Pm k=1 vkE(v∗ k=1 P ω vk }n So, ηω(y) = mXk=1 P ω vk ηω(y) = Qηω(y). kζk ≤ kζ − ηω(x)k + kηω(x − y)k + kηω(y)k < 2ε + kηω(y) − Qζk = 2ε + kQ(ηω(y) − ζ)k ≤ 2ε + kηω(y) − ζk < 4ε. Thus the claim holds. Now suppose ξ ∈ Hφ satisfies Pvξ = 0 for every v ∈ GN(M, D). Then for every ω ∈ Φ and v ∈ GN(M, D), 0 = TωPvξ = P ω v Tωξ. Hence Tωξ = 0 for every ω ∈ Φ, so ξ = 0 by (2.6). The proof is now complete. (cid:3) We come now to the main theorem of this section. Theorem 2.22. The Cartan triples (M1, N1, D1) and (M2, N2, D2) are isomorphic if and only if their associated extensions, P1 ֒→ G2 q2 ։ S2, are equivalent. q1 ։ S2 and P1 ֒→ G2 Proof. It is easy to see that if the triples are isomorphic, then their associated extensions are equivalent. Suppose now that the associated extensions are equivalent via the triple of maps (α, α, α). Then αP1 = α, q2 ◦ α = α ◦ q1 and α is an extendible isomorphism, say α = θP, where θ : N1 → N2 is a normal isomorphism with θ(D1) = D2. Let Ei : Mi → Ni be the conditional expectations. By Lemma 2.18, E2 ◦ α = (α ◦ E1)G1 , equivalently E2 ◦ α = (θ ◦ E1)G1. 10 Let ψ1 be a faithful normal weight on N1 and let ψ2 = ψ1 ◦ θ−1. Now let φi := ψi ◦ Ei. Then φi are faithful semi-finite normal weights on Mi. Let (πi, Hi, ηi) be the associated semi-cyclic representations and let ni := {x ∈ Mi : φi(x∗x) < ∞}. By Lemma 2.21, span{ηφi(vn) : v ∈ GN(Mi, Di) and n ∈ nψi} is dense in Hi. Let n ∈ N and suppose v1, . . . , vn ∈ G1 and c1, . . . , cn ∈ nψ1. Then α(vj) ∈ G2, and, since (α ◦ E1)G1 = E2 ◦ α, it follows from the definition of φ2 that φ2 nXi=1 α(vi)θ(ci)!∗ nXi=1 θ(ci)∗α(v∗ θ(ci)∗α(v∗ nXi,j=1 i vj)θ(cj) i vj)θ(cj) i vj))θ(cj) i vj))θ(cj) i vjcj) vici!! . vici!∗ nXi=1 E1(c∗ i v∗ θ(ci)∗E2(α(v∗ θ(ci)∗θ(E1(v∗ α(vi)θ(ci)!! = φ2 nXi,j=1 = ψ2E2 = ψ2 nXi,j=1 = ψ2 nXi,j=1 = ψ1 nXi,j=1 = φ1 nXi=1 α(vi)θ(ci)! vici! 7→ η2 nXi=1 Hence the map η1 nXi=1 extends to a unitary operator U : H1 → H2. π2(α(v))U . Therefore the map θ : M1 → M2 given by θ(x) = π−1 (M1, N1, D1) onto (M2, N2, D2). It is routine to verify that for v ∈ G1, U π1(v) = 2 (U π1(x)U ∗) is an isomorphism of (cid:3) 3. Representing an extension In this section we will show how to represent an extension as partial isometries on a right Hilbert- module. In Section 4 we will show how this gives rise to a Cartan triple. Throughout this section: • P ֒→ G q ։ S will be an idempotent separating extension of the Cartan inverse monoid S by the N-Clifford inverse monoid P; • j : S → G will be a fixed order-preserving section (see Proposition 2.16); and • D is the von Neumann subalgebra of Z(N) generated by E(P). We will sometimes use the fact that viewed as a C ∗-algebra, D is isomorphic to the universal C ∗-algebra C ∗(dE(S)) generated by the meet semilattice E(S), see [12, Proposition 2.2]. We now construct a right reproducing kernel Hilbert N-module. We begin by using the construc- tion of the right reproducing Hilbert D-module as done in [12, Section 4.2]. Recall that jE(S) is a complete lattice isomorphism of E(S) onto E(P) = E(G) ⊆ D. Define K : S × S → D by K(t, s) = j(s†t ∧ 1). 11 and for s ∈ S, define ks : S → D by ks(t) = K(t, s). For d ∈ D and s ∈ S, we use ksd to denote the map from S into D given by S ∋ t 7→ ks(t)d. Put A0 := span{ksd : s ∈ S and d ∈ D}. Let u, v ∈ A0. Lemma 4.11 and Proposition 4.12 of [12] show that: (a) if u =Pn i=1 ksidi and v =Pn j=1 ktj ej, then the formula hu, vi := d∗ i K(si, tj)ej (3.1) nXi,j=1 is independent of the choice of the representations for u and v and determines a well-defined D-valued inner product on A0 which is conjugate linear in the first variable; (b) for every s ∈ S and u ∈ A0, hks, ui = u(s); (c) the completion AD of A0 with respect to this inner product is a right Hilbert D-module of functions from S to D; and (d) span{ks : s ∈ S} is dense in AD. Next, we "fatten" AD to incorporate the fact that P is an N-Clifford semigroup, not a D- Clifford semigroup as in [12]. View N as a right Hilbert N-module, with hx, yiN := x∗y. Define a ∗-monomorphism ι : D → L(N) by ι(d)(x) = dx (where d ∈ D and x ∈ N). Put A := AD ⊗ι N, (3.2) see [19, pages 38 -- 44]. Then A is a right Hilbert N-module. This is the space on which we shall define a representation of G. Note that the inner product on the algebraic tensor product AD ⊙ι N is * NXi=1 ksi ⊗ xi, kti ⊗ yi+A NXi=1 = NXi,j=1 We denote the bounded, adjointable operators on A by L(A). x∗ i K(si, tj)yj. (3.3) We will presently describe the representation of G on A. First, we need a little more machinery derived from our extension. It is usual to describe idempotent-separating extensions in terms of a cocycle function γ : S × S → P. In the case when P is a abelian this is done explicitly by Lausch [20], leading to a one-to-one correspondence between extensions and the cohomology group H 2(S, P). In the case when P is not abelian D'Alarcao [11] has studied extensions, modelled on the Schreier extensions of groups. Though no cocycle is explicitly given, the construction again relies on functions from S × S to P. In our setting, where we are assuming we have an extension P ֒→ G ։ S, we instead work with a cocycle-like function from G×S into P. This leads to significant computational simplifications when we define our representation of G. To our knowledge, there is not a cohomological description of extensions when P is not abelian. As our cocycle-like function includes G a priori, our approach is unlikely to shed further light on that question. Definition 3.1. Define a cocycle-like function σ : G × S → P by σ(v, s) = j(q(v)s)†vj(s) = j(s†q(v†))vj(s). Since σ(v, s) ∈ P. Thus σ indeed maps G × S into P. Observe also that q(σ(v, s)) = s†q(v†v)s ∈ E(S), σ(v, s)∗σ(v, s) = j(s†q(v†v)s) = j(s)†v†vj(s). (3.4) 12 The following result gives the definition of the representation of G in L(A) and is the analog of [12, Theorem 4.16] suitable for our context. While the outline of the proof is the same as the proof of [12, Theorem 4.16], there are differences. Due to the importance of the result for our work, we provide most of the details of the proof. Theorem 3.2. For v ∈ G, s ∈ S and x ∈ N, the formula, λ(v)(ks ⊗ x) := kq(v)s ⊗ σ(v, s)x determines a partial isometry λ(v) ∈ L(A). Moreover, λ : G → L(A) is a one-to-one representation of G as partial isometries in L(A). Proof. Fix v ∈ G, and set r := q(v). Given s1, . . . , sN ∈ S, apply [12, Lemma 4.15] to obtain A ⊆ S satisfying: (a) 0 /∈ A; (b) if a, b ∈ A then a ∧ b = 0; (c) if a ∈ A then a∧ sn ∈ {0, a} for 1 ≤ n ≤ N ; and there exists 1 ≤ n ≤ N such that a∧ sn = a; (d) for each 1 ≤ n ≤ N , sn =W{a ∈ A : a ≤ sn}. Choose c1, . . . , cN ∈ N. For a ∈ A and 1 ≤ m ≤ N , put Am := {b ∈ A : b ≤ sm} and ca :=X{cn : a ≤ sn}. Since Am ⊆ A, the elements of Am are pairwise meet orthogonal. Further,W Am = sm. As in the proof of [12, Theorem 4.16], NXn=1 ksn ⊗ cn =Xa∈A ka ⊗ ca. (3.5) Secondly, with routine modifications to the proof of [12, Equation (4.4)], we obtain NXn=1 krsn ⊗ σ(v, sn)cn =Xa∈A kra ⊗ σ(v, a)ca. (3.6) Notice that if a, b ∈ A are distinct, then ra and rb are orthogonal, so for x, y ∈ N, hkra ⊗ σ(v, a)x, krb ⊗ σ(v, b)yi = x∗σ(v, a)∗K(ra, rb)σ(v, b)y = 0 = x∗K(a, b)y = hka ⊗ x, kb ⊗ yi . 13 Thus, as D ⊆ Z(N) and using (3.6), then (3.5), * NXn=1 krsn ⊗ σ(v, sn)cn, NXn=1 kra ⊗ σ(v, a)ca,Xa∈A ca2σ(v, a)∗j(a†r†ra)σ(v, a) kra ⊗ σ(v, a)ca+ krsn ⊗ σ(v, sn)cn+ =Xa∈A =Xa∈A =Xa∈A ≤Xa∈A =Xa∈A =* NXn=1 ≤(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) nXn=1 λ(v)(ksn ⊗ cn)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) NXn=1 (cid:13)(cid:13)(cid:13)(cid:13)(cid:13) ca2j(a†r†ra) ca2j(a†a) ksn ⊗ cn, ka ⊗ ca+ ka ⊗ ca,Xa∈A ksn ⊗ cn+ . NXn=1 ksn ⊗ cn(cid:13)(cid:13)(cid:13)(cid:13)(cid:13) . Therefore, It follows that we may extend λ(v) linearly to a contractive operator from the algebraic tensor product A0 ⊙ι N into A. Finally extend λ(v) by continuity to a contraction in B(A), the bounded operators on A. We next show that λ(v) is adjointable. As in the proof of the corresponding equality found in the proof of [12, Theorem 4.16], for s, t ∈ S, Therefore for any s, t ∈ S and x, y ∈ N, σ(v, s)†K(rs, t) = σ(v†, t)K(s, r†t). hλ(v)(ks ⊗ x), kt ⊗ yi = hkrs ⊗ σ(v, s)x, kt ⊗ yi = x∗σ(v, s)∗K(rs, t)y = x∗σ(v†, t)K(s, r†t)y = x∗K(s, r†t)σ(v†, t)y = hks ⊗ x, λ(v†)(kt ⊗ y)i. This equality implies that λ(v) is adjointable and λ(v)∗ = λ(v†). We now show that λ is a homomorphism. Suppose that v1, v2 ∈ G, x ∈ N and s ∈ S. Then λ(v1)(λ(v2)(ks ⊗ x)) = λ(v1)(kq(v2)s ⊗ σ(v2, s)x) = kq(v1v2)s ⊗ σ(v1, q(v2)s) σ(v2, s)x. But σ(v1, q(v2)s) σ(v2, s) = j(q(v1)q(v2)s))†v1j(q(v2)s) j(q(v2)s)†v2j(s) = j(q(v1v2)s))†v1j(q(v2)s) j(s†q(v2)†)v2j(s) = j(q(v1v2)s))†v1(v2j(ss†)v† = j(q(v1v2)s))†v1v2v† = j(q(v1v2)s))†v1v2j(s) = σ(v1v2, s). 2v2j(ss†)j(s) 2)v2j(s) 14 Hence λ(v1)λ(v2)(ks ⊗ x) = λ(v1v2)(ks ⊗ x). As span{ks ⊗ x : s ∈ S and x ∈ N} is dense in A, we conclude that λ(v1v2) = λ(v1)λ(v2). It follows that for every e ∈ E(G), λ(e) is a projection. Furthermore, for v ∈ G, λ(v) is a partial isometry because λ(v)∗ = λ(v†). It remains to show that λ is one-to-one. We first show that λE(G) is one-to-one. Suppose e, f ∈ E(S) and λ(j(e)) = λ(j(f )). Then for every s ∈ S, σ(j(e), s) = j(s†es) ∈ D and σ(j(f ), s) = j(s†f s) ∈ D. As the tensor product is balanced, kesj(s†es) ⊗ I = kes ⊗ σ(j(e), s) = λ(j(e))(ks ⊗ I) = λ(j(f ))(ks ⊗ I) = kf s ⊗ σ(j(f ), s) = kf sj(s†f s) ⊗ I, whence kesj(s†es) = kf sj(s†f s). Taking s = 1 gives kej(e) = kf j(f ). Evaluating these elements of AD at t = 1 gives j(e) = j(f ), so λE(G) is one-to-one. Now suppose v1, v2 ∈ G and λ(v1) = λ(v2). Then λ(v† 1v1) = λ(v† 1v2) = λ(v† 1v2)∗ = λ(v† 2v1) = λ(v† 2v2). Likewise, Hence v† λ(v1v† 1 = v2v† 2) = λ(v2v† 1) = λ(v1v† 2. For any e ∈ E(S), we have 1) = λ(v2v† 2). 1) = λ(v1v† 1v1j(e)v† 1v1v† 1) = λ(v1v† 2v2j(e)v† 2v2v† 1) = λ(v2v† 2v2j(e)v† 2v2v† 2) 2v2 and v1v† 1v1 = v† λ(v1j(e)v† = λ(v2j(e)v† 2). Hence v1j(e)v† that q(v1) = q(v2). 1 = v2j(e)v† 2. Since this holds for every e ∈ E(S) and S is fundamental, we conclude Let e := q(v† 1v1) and s := q(v1). Since the functions λ(v1)ke and λ(v2)ke agree, we obtain, ksj(s)†v1 = ksj(s)†v2. Evaluating these functions at t = s gives, j(s)†v1 = j(s)†v2. Multiplying each side of this equality on the left by j(s), we obtain v1 = v2. (cid:3) Let π : N → B(H) be a normal representation. Recall there is a ∗-representation π∗ : L(A) → B(A ⊗π H) given by (3.7) This representation is strictly continuous on the unit ball of L(A) and is faithful whenever π is faithful [19, p. 42]. As in [12, Corollary 4.17] we have the following corollary. π∗(T )(u ⊗ ξ) = (T u) ⊗ ξ. Corollary 3.3. Let π : N → B(H) be a ∗-representation of N on the Hilbert space H. Then λπ := π∗ ◦ λ is a representation of G by partial isometries on A ⊗π H. If π is faithful, then λπ is one-to-one. Remark 3.4. Our construction of A ⊗π H depends upon the order structure of S. We show presently that the range of λπ will generate a Cartan triple. When one starts with a Cartan triple and applies this construction to the extension associated to the pair, the Hilbert space A ⊗π H can be recognized as arising from the representation associated to a faithful normal weight φ on M such that φ ◦ E = φ. We will give the formal statement in Proposition 4.13 below. We close this section with some results which will be needed when constructing a Cartan triple from an extension in Section 4. They will also be used in Section 5 when we study the N-bimodule structure for a Cartan triple. Observe that the construction of the right Hilbert D-module AD above depends only upon S because jE(S) is the inverse of qE(P). The τ1-topology described in the following definition has been considered by several authors, see Section 3.5 of the survey article [22]. 15 Definition 3.5. Let M be a right Hilbert module over the von Neumann algebra N. (a) The τ1-topology on M is the topology generated by the seminorms, ξ 7→ φ(hξ, ξi)1/2, where φ is a normal state on N. (b) The τ1-strict topology on L(M) is the topology generated by the seminorms, T 7→ φ(hT ξ, T ξi)1/2 where ξ ∈ M and φ is a normal state on N. Notice that a net Tα → T in the τ1-strict topology if and only if for every ξ ∈ M, in the σ-strong topology of N. h(Tα − T )ξ, (Tα − T )ξi → 0 The following result can be proved directly in the same way as Proposition 5.2 of [12], but it is simpler to apply [12, Theorem 4.16 and Proposition 5.2]. Proposition 3.6. For s ∈ S, the map S ∋ t 7→ ks∧t extends to a projection Qs ∈ L(AD) whose range is span{kt : t ≤ s}. Furthermore, the following statements hold. (a) Let s, t ∈ S. If s ∧ t = 0, then QsQt = QtQs = 0; if s†t = st† = 0, then Qt + Qs = Qs∨t. (b) If B ⊆ S is a maximal meet disjoint subset and Λ is the set of all finite subsets of B directed by inclusion, then the net (Ps∈F Qs)F ∈Λ converges τ1-strictly to the identity operator in L(AD). Proof. By replacing σ(v, t) with the identity operator throughout the proof of [12, Thoerem 4.16] (or Theorem 3.2 above) one finds that for every s ∈ S, there exists a partial isometry λ0(s) ∈ L(AD) such that Calculations show that for every t ∈ S, λ0(s)kt = kst. λ0(s)PDλ0(s)∗kt = ks∧t, where PDkt := kt∧1 is the projection from [12, Proposition 5.2]. This establishes the existence of the projection Qs, with the desired range. The proof of (a) is routine and left to the reader. Let B be a maximal meet disjoint subset of S. Since the net (Ps∈F Qs)F ∈Λ is an increasing net of projections, it suffices to show that for each t ∈ S, the net QF kr = ktF . Denote by ¬ the NOT operation in the Boolean algebra E(S). We have that For F ∈ Λ, let QF :=Ps∈F Qs. For r ∈ S and s1, s2 ∈ F , r ∧ s1 and r ∧ s2 are disjoint elements of Ar := {t ∈ S : t ≤ r}, so (r ∧ s1) ∨ (r ∧ s2) is defined. Let tF :=Ws∈F (r ∧ s). Then tF ≤ r and Let b :=WF ∈Λ(t† hQF kr − kr, QF kr − kri = hktF − kr, ktF − kri = j(r†r ∧ ¬(t† F tF ). Clearly b ≤ r†r. Set a := r†r ∧ (¬b). Then for s ∈ B, a ∧ s†s = r†r ∧ s†s ∧ (¬b) ≤ r†r ∧ s†s ∧ ¬((r ∧ s)†(r ∧ s)) = 0. F tF )). Now ra ∧ s = ra(r†r ∧ s†s) = 0, so that ra is meet disjoint from every element of B. By maximality of B, we obtain ra = 0, whence a = 0. Thus b = r†r, from which it follows that j(t† F tF ) converges σ-strongly in D to j(r†r). Therefore, QF converges τ1-strictly to IL(AD). (cid:3) We now have the following corollary to Proposition 3.6. Corollary 3.7. The net QF ⊗ IN converges τ1-strictly to IL(A). 16 Xs∈F Qskt!F ∈Λ τ1-converges to kt. Proof. For n ∈ N and s ∈ S, we have h(QF ⊗ IN)(ks ⊗ n) − (ks ⊗ n), (QF ⊗ IN)(ks ⊗ n) − (ks ⊗ n)i = h(QF ks − ks) ⊗ n, (QF ks − ks) ⊗ ni = n∗ hQF ks − ks, QF ks − ksiAD n = n∗n hQF ks − ks, QF ks − ksiAD . As the last expression tends to zero in the σ-strong operator topology on N, the result follows. (cid:3) Now let π : N → B(H) be a faithful normal representation of N and for s ∈ S define projections on A ⊗π H by Proposition 3.8. Let B ⊆ S be a maximal meet disjoint subset. ThenPs∈B Ps,π converges strongly to I ∈ B(A ⊗π H). Ps,π := (Qs ⊗ IN) ⊗ IH. (3.8) Proof. For any h ∈ H and ξ ∈ A, the map L(A) ∋ T 7→ h((T ⊗ IH)(ξ ⊗ h), (T ⊗ IH)(ξ ⊗ h)i = hh, π(hT ξ, T ξiA)hi1/2 H is a τ1-strict seminorm on L(A). Therefore, if (Tα) is a bounded net in L(A) which converges τ1- strictly to T ∈ L(A), (Tα ⊗IH) converges strongly in B(H) to T ⊗I. An application of Corollary 3.7 completes the proof. (cid:3) Remark 3.9. Proposition 3.8 is similar to Lemma 2.21. The initial data for Proposition 3.8 is the q ։ S and the representation π : N → B(H); its conclusion may be interpreted extension P ֒→ G triple (M, N, D) and a semi-cyclic representation induced by a suitable weight; its conclusion may as the statement thatWs∈S Ps,π = IA⊗πH. On the other hand, Lemma 2.21 deals with the Cartan be interpreted as the statement that Wv∈GN(M,D) Pv = IHφ, where Pv is defined in the proof of q Lemma 2.21. There is a further relation between these results: when the extension P ֒→ G ։ S is the extension associated to the Cartan triple (M, N, D), Proposition 4.13 below implies that for any v ∈ GN(M, D), the projections Pv and Pq(v),πφ are unitarily equivalent. 4. The Cartan Triple Associated to an Extension Throughout this section we will consider the extension, P ֒→ G q ։ S, where S is a Cartan inverse monoid, and P is an N-Clifford inverse monoid. Assume throughout that a fixed order-preserving section j : S → G is given. Our goal, achieved in Theorem 4.10, is to show how the representation of G constructed in Corollary 3.3 gives rise to a Cartan triple with q associated extension P ֒→ G ։ S. In Theorem 4.11 we further show that the extension associated to the Cartan triple returns the original extension. We denote by A the right Hilbert N-module as defined in Equation (3.2). Let π be a faithful, normal representation of N, and let be the representation of G by partial isometries, as constructed in Theorem 3.2 and Corollary 3.3. λπ : G → B(A ⊗π H) Definition 4.1. Let Mq := (λπ(G))′′, Nq := (λπ(P))′′, and Dq := (λπ(E(G))′′. 17 We will show that (Mq, Nq, Dq) is a Cartan triple. The definitions of Mq, Nq and Dq depend upon the choice of π and, because λ : G → L(A) depends on the choice of j, Mq, Nq and Dq also depend on j. However, we shall see in Theorem 4.11 that the isomorphism class of (Mq, Nq, Dq) depends only on the extension P ֒→ G q ։ S and not upon π or j. The first step is to show that there is a faithful normal conditional expectation from Mq onto q in Proposition 4.9. As in [12], the expectation on Mq Nq. This will be used to show that Nq = Dc will be induced by the map s 7→ s ∧ 1 on S. Define ∆ : G → P by for all v ∈ G. First note that ∆(v) := vj(q(v) ∧ 1), Thus ∆(v) ∈ P for all v ∈ G. Further, if v ∈ P then q(v) ∈ E(S), thus q(∆(v)) = q(v)(q(v) ∧ 1) = q(v) ∧ 1 ∈ E(S). ∆(v) = vj(q(v) ∧ 1) = vj(q(v)) = v. We will show that, given v ∈ G, the formula, Eq(λπ(v)) := λπ(∆(v)) extends to a faithful conditional expectation Eq : Mq → Nq. Notation 4.2. Here is some notation. (a) Let PD ∈ L(AD) be the projection defined in Proposition 3.6, so that PDks = ks∧1. That is, PD = Q1 as defined in Proposition 3.6. Since A = AD ⊗ι N, the tensor product of PD with the identity of N gives a projection P ∈ L(A) so that, for s ∈ S and x ∈ N, (b) For x ∈ N, and y =Pn P (ks ⊗ x) = ks∧1 ⊗ x. (4.1) i=1 ksi ⊗ xnikA ≤ kxk kykA. It follows that the map ks ⊗ n 7→ ks ⊗ xn extends to a bounded linear map πℓ(x) on A. A computation shows that πℓ(x) is adjointable, so there exists a faithful ∗-representation πℓ : N → L(A). Tensoring with the identity map, we obtain a faithful normal representation πℓ∗ = πℓ ⊗ I of N on B(A ⊗π H). To be explicit, for x ∈ N, πℓ∗(x) is defined on elementary tensors (ks ⊗ n ⊗ ξ) ∈ A ⊗π H by i=1 ksi ⊗ ni ∈ AD ⊙ N, kPn πℓ∗(x)(ks ⊗ n ⊗ ξ) = ks ⊗ xn ⊗ ξ = ks ⊗ IN ⊗ π(xn)ξ. (4.2) Lemma 4.3. For s ∈ S, let Qs be the projection on L(AD) defined in Proposition 3.6 and let Ps,π := Qs ⊗ IN ⊗ IH ∈ B(A ⊗π H) be the projection defined in Equation (3.8). The following statements hold. (a) For v ∈ P, s ∈ S, n ∈ N and ξ ∈ H, (b) Ps,π ∈ N′ λπ(v)(ks ⊗ n ⊗ ξ) = ks ⊗ j(s)∗vj(s)n ⊗ ξ. q and for every v ∈ P, λπ(v)Ps,π = πℓ∗(j(s)∗vj(s))Ps,π. Proof. Since v ∈ P, q(v) = q(vv∗), so σ(v, s) = j(s†q(v)†)vj(s) = j(s)∗vj(s). Therefore, λπ(v)(ks ⊗ n ⊗ ξ) = kq(v)s ⊗ j(s)∗vj(s)n ⊗ ξ = kss†q(v)s ⊗ j(s)∗vj(s)n ⊗ ξ = ksj(s†q(v)s) ⊗ j(s)∗vj(s)n ⊗ ξ = ks ⊗ j(s†q(v)s)j(s)∗vj(s)n ⊗ ξ = ks ⊗ j(s)∗vj(s)n ⊗ ξ, where the third equality follows from [12, Corollary 4.9]. This gives part (a) and shows range(Ps,π) is invariant for every element of λπ(P). Thus, range(Ps,π) is invariant for Nq. As Nq is a ∗-algebra, Ps,π reduces Nq, whence Ps,π ∈ N′ q. 18 Now suppose t ∈ S, n ∈ N and ξ ∈ H. Note that j(s)j(s†t ∧ 1) = j(s(s†t ∧ 1)) = j(s ∧ t). Then, again using [12, Corollary 4.9], Ps,π(kt ⊗ n ⊗ ξ) = ks∧t ⊗ n ⊗ ξ = ks ⊗ j(s†t ∧ 1)n ⊗ ξ. A computation using part (a) and (4.3) now gives the formula in part (b). (4.3) (cid:3) With the obvious modifications to the proof of [12, Proposition 5.2], we obtain the following. Lemma 4.4. With P defined as in Equation (4.1), the following properties hold: (a) range P = span{ke ⊗ x : e ∈ E(S) and x ∈ N}; and (b) for v ∈ G, P λ(v)P = λ(∆(v))P. Modifications to the proof of [12, Proposition 5.3] yield the following result. Lemma 4.5. The map V : H → A ⊗π H given by V ξ = (k1 ⊗ IN) ⊗ ξ is an isometry. Moreover, the following properties hold: (a) for s ∈ S, x ∈ N and ξ ∈ H, V ∗(ks ⊗ x ⊗ ξ) = π(j(s ∧ 1)x)ξ; (b) V V ∗ = π∗(P ), where π∗ : L(A) → B(A ⊗π H) is defined by π∗(T )(u ⊗ ξ) = (T u) ⊗ ξ; (c) for v ∈ G, V ∗λπ(v)V = π(∆(v)). Lemma 4.6. We have V ∗MqV = π(N) = V ∗NqV. Proof. Lemma 4.5(c) shows that for x ∈ Mq, V ∗xV ∈ π(N), so V ∗MqV ⊆ π(N). On the other hand, for v ∈ P we have (4.4) so V ∗NqV ⊆ π(N). Since every element of N is a linear combination of at most four elements of P, we obtain the result. (cid:3) V ∗λπ(v)V = π(∆(v)) = π(v), Thus, the map Mq ∋ x 7→ π−1(V ∗xV ) is a normal, completely positive contraction of Mq onto N. We now show this map gives an isomorphism of Nq onto N. Lemma 4.7. The map α : Nq → N defined by α(x) = π−1(V ∗xV ) is a normal isomorphism of Nq onto N. Proof. The definition of α shows it is normal. Next we show that α is a homomorphism. For v1, v2 ∈ P, Lemma 4.5(c) gives V ∗(λπ(v1))V V ∗λπ(v2)V = π(v1)π(v2) = π(v1v2) = V ∗λπ(v1v2)V. Thus α is multiplicative on λπ(P). It follows that α is multiplicative on span(λπ(P)). As mul- tiplication is σ-strongly continuous on bounded sets, the Kaplansky density theorem ensures α multiplicative. Lemma 4.6 now shows α is a ∗-epimorphism. It remains to show α is one-to-one. To do this, we show α is isometric on span(λπ(P)). Suppose j=1 cjvj by Equation 4.4. n ∈ N, cj ∈ C, vj ∈ P and x =Pn By Lemma 4.3, for each s ∈ S, j=1 cjλπ(vj). Put y = α(x) so that y =Pn xPs,π = πℓ∗(j(s)∗yj(s))Ps,π, and hence kxPs,πk ≤ kyk. Now suppose B is a maximal meet-disjoint subset of S. Then for distinct s, t ∈ B, Ps,π and Pt,π are orthogonal projections. By Proposition 3.8, kxPs,πk ≤ kyk = kα(x)k ≤ kxk . kxk = sup s∈B So α is isometric on span(λπ(P)). At last, we can define the conditional expectation Eq. 19 (cid:3) Proposition 4.8. The formula, Eq(x) := α−1(π−1(V ∗xV )) gives a faithful normal conditional expectation of Mq onto Nq. Furthermore, for v ∈ G, Eq(λπ(v)) = vλπ(∆(v)). (4.5) (4.6) Proof. Lemmas 4.6 and 4.7 imply Eq is a normal conditional expectation of Mq onto Nq. It remains only to establish that Eq is faithful. The proof that Eq is faithful is modeled on the proof of [12, Proposition 5.9]. Let C denote the center of Mq. We claim that EqC is faithful. Let x ∈ C and suppose Eq(x∗x) = 0. The definition of Eq from Equation 4.5 shows that V ∗x∗xV = 0 and hence xV = 0. Notice that σ(j(s), 1) = j(s†s) so that for n ∈ N, λ(j(s))(k1 ⊗n) = ksj(s†s)⊗n = ks ⊗n (see [12, Corollary 4.9]). Hence for s ∈ S, n ∈ N and ξ ∈ H, x(ks ⊗ n ⊗ ξ) = xλπ(j(s))(k1 ⊗ n ⊗ ξ) = λπ(j(s))x(k1 ⊗ n ⊗ ξ) = λ(j(s))xV π(n)ξ = 0. Since the span of such vectors is a dense subspace of A ⊗π H, we conclude that x = 0. Let J := {x ∈ Mq : Eq(x∗x) = 0}. Then J is a left ideal of Mq. Compute as in the second part of [12, Lemma 5.8] to find that for x ∈ J and v ∈ G, Eq(λπ(v)∗x∗xλπ(v)) = λπ(v)∗Eq(x∗x)λπ(v) = 0. Thus, xλπ(v) ∈ J. It now follows that J is a two-sided ideal of Mq as well. Since J is weak-∗-closed, by [31, Proposition II.3.12], there is a projection Q ∈ C such that J = QMq. As Q ∈ J and EqC is faithful, we obtain Q = 0. Thus J = (0), that is, Eq is faithful. The equality (4.6) follows from Lemma 4.5. (cid:3) Proposition 4.9. The algebra Nq is the relative commutant of Dq in Mq. That is, Nq = Dc q. Proof. Notice that v ∈ G commutes with every element of E(G) if and only if v ∈ P. Since λπ is a isomorphism of G onto λπ(G), we obtain λπ(G) ∩ Dc q = λπ(P). Therefore, Nq ⊆ Dc q. Take x ∈ Dc q. Suppose w ∈ λπ(G) satisfies w2 = 0. Then Eq(w∗x) = w∗wEq(w∗x) = Eq(w∗x)w∗w = Eq(w∗xw∗w) = Eq((w∗)2wx) = 0. (4.7) Now choose an arbitrary v ∈ λπ(G). Our goal is to show that (again with x ∈ Dc q) (4.8) As in Lemma 2.18, v defines a map θv on Dq (d 7→ vdv∗). By Frol`ık's Theorem (see [26, Proposi- tion 2.11a]) there are orthogonal projections e0, e1, e2, e3 ∈ Dq such that Eq(v∗x) = Eq(v∗)Eq(x). θvDe0 = idDe0, and θv(ek)ek = 0 for k = 1, 2, 3. v = vek, 3Xk=0 As e0 is the largest projection in Dq on which θvDe0 = idDe0, it follows that ve0 = Eq(v). Also, as θv(ek)ek = 0, it follows that (vek)2 = 0, for k = 1, 2, 3. By (4.7), for k = 1, 2, 3, Eq(ekv∗x) = 0. Thus Eq(v∗x) = so (4.8) holds. 3Xk=0 Eq(ekv∗x) = Eq(e0v∗x) = Eq(v∗)Eq(x), Since λπ(G) spans a weak∗-dense subset of Mq, it follows that for x ∈ Dc q we have Eq(x∗x) = Eq(x∗)Eq(x). 20 Hence, if x ∈ Dc q we have Eq((x − Eq(x))∗(x − Eq(x)) = Eq(x − Eq(x))∗Eq(x − Eq(x)) = 0. Since Eq is faithful, it follows that x = Eq(x) ∈ Nq. (cid:3) Proposition 4.8 and Proposition 4.9 now immediately give the first main theorem of this section. Theorem 4.10. (Mq, Nq, Dq) is a Cartan triple. The second main theorem of this section is that the extension associated to (Mq, Nq, Dq) gives back the original extension P ֒→ G q ։ S. q Theorem 4.11. Let P be a N-Clifford inverse monoid and suppose P ֒→ G ։ S is an exten- sion of the Cartan inverse monoid S by P. Let (Mq, Nq, Dq) be the Cartan triple constructed in Theorem 4.10. The extension associated to (Mq, Nq, Dq) is equivalent to the extension P ֒→ G q ։ S from which (Mq, Nq, Dq) was constructed. Moreover, the isomorphism class of (Mq, Nq, Dq) depends only upon the equivalence class of the extension (and not on the choice of representation π or section j). Remark 4.12. In the proof of Theorem 4.11 and also in the proof of Theorem 5.2 below, we shall utilize a result of Arveson, [3, Theorem 6.2.2]. In [3], Arveson makes the blanket assumption that all Hilbert spaces are separable (see [3, Section 1.2]). However the proof of [3, Theorem 6.2.2] does not require separability. Proof. The argument below is a modification of the proof of [12, Theorem 5.12]. Let RM and RM,π be the Munn congruences for G and λπ(G) respectively. Since λπ is an isomorphism of G onto λπ(G), (v, w) belongs to RM if and only if (λπ(v), λπ(w)) belongs to RM,π. Let qπ : λπ(G) → λπ(G)/RM,π be the quotient map. The fact that S is fundamental implies that λπ := qπ ◦ λπ ◦ j is a multiplicative map of S onto λπ(G)/RM,π. In fact, λπ is an isomorphism satisfying λπ ◦ q = qπ ◦ λπ, and furthermore, λπP is an isomorphism of P onto λπ(P). Let v ∈ P. By Equation (4.4) (see Lemma 4.6), V ∗λπ(v)V = π(v). Thus in the notation of Lemma 4.7, α−1(v) = λπ(v). Therefore λπP = α−1P. It is now clear that the extensions P ֒→ G q ։ S and λπ(P) ֒→ λπ(G) qπ ։eλπ(S) are equivalent. For later use, note that in particular, λπE(G) is an isomorphism of E(G) onto E(λπ(G)). Our next task is to show that It will then follow immediately that λπ(P) ֒→ λπ(G) (Mq, Nq, Dq). Claim 1: If u ∈ GN(Mq, Dq), then uEq(u∗) is a projection in Dq, and (4.9) ։ eλπ(S) is the extension associated to λπ(G) = GN(Mq, Dq). qπ uEq(u∗) = Eq(uEq(u∗)) = Eq(u)Eq(u∗). (4.10) Let Λ be an invariant mean on the abelian group U(Dq). By [3, Theorem 6.2.2], uEq(u∗) = Λ (ugu∗)g∗ ∈ Dq. g∈U(Dq) 21 Next, uEq(u∗)uEq(u∗) = uEq(u∗uEq(u∗)) = uu∗uEq((Eq(u∗)) = uEq(u∗), so uEq(u∗) is a projection in Dq. The equality (4.10) is now obvious, so Claim 1 holds. By construction, λπ(G) ⊆ GN(Mq, Dq). To establish the reverse inclusion, fix v ∈ GN(Mq, Dq); without loss of generality, assume v 6= 0. Claim 2: There exists p ∈ λπ(E(G)) such that: a) vp ∈ λπ(G), b) p ≤ v∗v, and c) vp 6= 0. Since λπ(G)′′ = Mq, it follows (as in the proof of [5, Proposition 1.3.4]) that there exists w ∈ λπ(G) such that wEq(w∗v) 6= 0. Let p = v∗wEq(w∗v). By Claim 1, p ∈ Dq is a projection, so in particular, p ∈ λπ(E(G)). It is evident that p ≤ v∗v. By (4.10), Eq(v∗w)Eq(w∗v) = p, so x := Eq(w∗v) is a partial isometry in Nq with source projection p ∈ Dq. On the other hand, let p′ := w∗vEq(v∗w). Claim 1 gives p′ ∈ Dq and p′ = Eq(w∗v)Eq(v∗w). We have thus shown that both the source and range projections for x belong to Dq ⊆ Z(Nq). Therefore, p = x∗x = x∗(xx∗)x = (x∗x)(xx∗) = x(x∗x)x∗ = xx∗ = p′. Hence Eq(w∗v) is a partial isometry in Nq whose source and range projections both equal p ∈ Dq. Thus, Eq(w∗v) ∈ GN(Nq, Dq) = λπ(P). This gives wEq(w∗v) ∈ λπ(G). Since Eq(w∗v) = w∗v(v∗wEq(w∗v)), we obtain, 0 6= wEq(w∗v) = w(w∗v(v∗wEq(w∗v))) = vv∗wEq(w∗v) = vp. Thus Claim 2 holds. Now argue exactly as in the proof of [12, Theorem 5.12] to conclude that v ∈ λπ(G). Therefore, we have shown that λπ(G) = GN(Mq, Dq). Hence λπ(P) ֒→ λπ(G) qπ ։ qπeλπ(S) is the extension for (Mq, Nq, Dq). Suppose that π is a faithful normal representation of N and j : S → G is an order preserving section for q. Let ( Mq, Nq, Dq) be the Cartan triple constructed using π and j as in Theorem 4.10. Then the previous paragraphs show that the extensions associated to (Mq, Nq, Dq) and ( Mq, Nq, Dq) are equivalent extensions. By Theorem 2.22, (Mq, Nq, Dq) and ( Mq, Nq, Dq) are isomorphic Cartan triples. The proof is now complete. (cid:3) Let (M, N, D) be a Cartan triple and let φ be a faithful normal semi-finite weight on M satisfying φ ◦ E = φ. We end this section by relating the semi-cyclic representation (πφ, Hφ, ηφ) and the q reproducing kernel Hilbert N-module A ⊗ N constructed from the extension P ֒→ G ։ S associated to (M, N, D). Proposition 4.13. Let (M, N, D) be a Cartan triple, suppose ψ is a faithful, normal semi-finite weight on N, and put φ := ψ◦E. Let (πψ, Hψ, ηψ) and (πφ, Hφ, ηφ) be the semi-cyclic representations of N and M associated with ψ and φ respectively. Let P ֒→ G q ։ S be the extension associated to (M, N, D) and let j : S → G be an order-preserving section for q. For s ∈ S, n ∈ N and x ∈ nψ, j(s)nx ∈ nφ, and the map (AD ⊗ι N) ⊗πψ Hψ ∋ ks ⊗ n ⊗ ηψ(x) 7→ ηφ(j(s)nx) ∈ Hφ extends to a unitary operator W : A ⊗πψ Hψ → Hφ such that for every v ∈ G, W λπψ (v)W ∗ = πφ(v). 22 Proof. For i = 1, 2, let si ∈ S, ni ∈ N and xi ∈ nψ. Then φ((j(si)nixi)∗(j(si)nixi)) = ψ(x∗ i n∗ i E(j(si)∗j(si))nixi) ≤ knik2 ψ(x∗ i xi), so j(si)nixi ∈ nφ. Recall that for any v ∈ GN(M, D), E(v) = vj(q(v) ∧ 1). In particular, using Proposition 2.16, we have E(j(s1)∗j(s2)) = j(s1)∗j(s2)j(s† 1s2 ∧ 1) = j(s† 1s2 ∧ 1). So h(ks1 ⊗ n1 ⊗ ηψ(x1)), (ks2 ⊗ n2 ⊗ ηψ(x2))i = ψ(x∗ = ψ(x∗ = hηφ(j(s1)n1x1), ηφ(j(s2)n2x2)i . 1j(s† 1E(j(s1)∗j(s2))n2x1) 1s2 ∧ 1)n2x1) 1n∗ 1n∗ As every element in span{ks ⊗ n ⊗ x : s ∈ S, n ∈ N, x ∈ nψ} can be written asPa∈A ka ⊗ na ⊗ xa where A ⊆ S is a finite pairwise meet disjoint set, {na : a ∈ A} ⊆ N and {xa : a ∈ A} ⊆ nψ, it follows that ks ⊗ n ⊗ x 7→ ηφ(j(s)nx) extends to an isometry W : A ⊗πψ Hψ → Hφ. By Proposition 2.21, span{ηφ(j(s)nx) : s ∈ S, n ∈ N, x ∈ nψ} is dense in Hφ, so W is a unitary operator. If v ∈ GN(M, D), s ∈ S, n ∈ N and x ∈ nψ, W λπψ (v)(ks ⊗ n ⊗ x) = W (kq(v)s ⊗ σ(v, s)n ⊗ x) = ηφ(j(q(v)s)σ(v, s)nx) = ηφ(vj(s)nx) = πφ(v)W (ks ⊗ n ⊗ x). Thus, W λπψ (v)W ∗ = πφ(v). (cid:3) 5. The spectral theorem for bimodules and Aoi's theorem Throughout this section, (M, N, D) will be a Cartan triple with associated extension P ֒→ G q ։ S, and j : S → G will be a fixed choice of an order-preserving section for q. The goal in this section is study the N-bimodules in M. Recall the following definition from [12]. Definition 5.1. A subset A of a Cartan inverse monoid S is a spectral set if (a) s ∈ A and t ≤ s implies that t ∈ A; and (b) if {si}i∈I is a pairwise orthogonal family in A, thenWi∈I si ∈ A. In Theorem 5.10 we prove a Spectral Theorem for Bimodules. Will show a one-to-one corre- spondence between the spectral sets in S and a large class of weak-∗ closed N-bimodules: the Bures-closed N-bimodules (see Definition 5.5). We go on to study the intermediate von Neumann algebras N ⊆ L ⊆ M, giving a generalization of Aoi's Theorem [1] in Theorem 5.12. Several of these theorems require that the Cartan triple is a full Cartan triple. 5.1. N-bimodules. We begin by showing that weak-∗ closed N-bimodules give rise to non-empty spectral sets. In particular, Theorem 5.2 shows that when (M, N, D) is a full Cartan triple, any weak-∗ closed N-bimodule contains an abundance of elements of GN(M, D). Example 5.4 below gives a simple example showing fullness is necessary. Theorem 5.2. Let (M, N, D) be a full Cartan triple. Suppose (0) 6= B ⊆ M is a weak-∗-closed N-bimodule. Then {0} 6= GN(M, D) ∩ B. In fact, for every x ∈ B and v ∈ GN(M, D), vE(v∗x) is a linear combination of at most four elements of GN(M, D) ∩ B. 23 Proof. Let x ∈ B be non-zero. Since M is the weak-∗ closed span of GN(M, D), there exists v ∈ GN(M, D) such that E(v∗x) 6= 0, and hence vE(v∗x) 6= 0. By [3, Theorem 6.2.2] (see Remark 4.12 above), for any y ∈ M, Therefore, E(y) = Λ U ∗yU. U ∈U(D) vE(v∗x) = Λ (vU ∗v∗)xU ∈ B. U ∈U(D) Let J be the weak-∗ closed, two-sided ideal in N generated by E(v∗x). For any n1, n2 ∈ N we have v(n1E(v∗x)n2) = (vn1v∗)(vE(v∗x))n2 ∈ B. Since B is weak-∗ closed, it follows that vJ ⊆ B. Let p ∈ Z(N) = D be such that J = pN. Then, vp ∈ B ∩ GN(M, D). Since vE(v∗x) = vpE(v∗x), 0 6= vp. Since pN is a von Neumann algebra (with unit p), E(v∗x) is a linear combination of four unitary elements of pN. As p ∈ Z(N), U(pN) = {pw : w ∈ U(N)}. Also, for any unitary w ∈ U(N) we have vpw ∈ GN(M, D). Thus, vE(v∗x) is a linear combination of at most four elements of GN(M, D). (cid:3) The following corollary of the proof of Theorem 5.2 will be needed in the sequel. Corollary 5.3. Let (M, N, D) be a (not necessarily full) Cartan triple. For any x ∈ M and v ∈ GN(M, D), vE(v∗x) belongs to the weak-∗ closed D-bimodule generated by x. Example 5.4. Let M be any von Neumann algebra with non-trivial center. Then (M, M, CI) is a Cartan triple which is not full. Let p be a central projection in M with 0 < p < I. Then Mp is a weak-∗ closed M-bimodule. However Mp ∩ GN(M, CI) = {0}. Thus, the condition of fullness in Theorem 5.2 is necessary. A natural problem is to characterize the weak-∗ closed N-bimodules in M. Given a weak-∗ closed N-bimodule B ⊆ M, one might hope to use Theorem 5.2 to reconstruct a given element x ∈ B from the elements of B ∩ GN(M, D). However, for doing this, the weak-∗ topology is not generally the appropriate topology. Instead, as Mercer shows in [23], the Bures-topology turns out to be the "right" topology to handle such reconstruction problems. This phenomenon was also observed in studying bimodules in the Cartan pair case by Cameron, Pitts and Zarikian [5] (see also [12]), and in the crossed-product von Neumann algebras by Cameron and Smith [6, 8]. Our next goal is Theorem 5.7, which gives a method for reconstructing x ∈ B using GN(M, D) ∩ B when B ⊆ M is a Bures-closed N-bimodule. We begin with recalling the definition of the Bures-topology. Definition 5.5. Let L ⊆ M be an inclusion of von Neumann algebras and assume there is a faithful normal conditional expectation EL : M → L. The EL-Bures topology (or simply Bures topology when the context is clear) is the locally convex topology determined by the family of seminorms, M ∋ x 7→ ρ(E(x∗x))1/2, ρ ∈ L+ ∗ . The Bures topology was introduced in [4] in the case when M is a factor and L is abelian. By [6, Lemma 3.1], for any convex set C ⊆ M, the Bures closure of C contains the weak-∗ closure of C, that is, clweak-∗(C) ⊆ clBures(C). Take x ∈ M. We showed in Theorem 5.2 and Corollary 5.3 that for each v ∈ GN(M, D), vE(v∗x) is in the N-bimodule generated by x. We now show that, in the Bures topology, we can recover x from the elements of the form vE(v∗x). 24 ∗ } {M ∋ m 7→pτ (E(m∗m)) : τ ∈ N+ πφ(xF )ηφ(n) =Xu∈F =Xu∈F =Xu∈F πφ(xF )ξ =Xu∈F Hence for every ξ ∈ ηφ(nφ ∩ N), ηφ(uE(u∗xn)) πφ(u)P πφ(u)∗ηφ(xn) Pq(u)ηφ(xn) =Xu∈F Pq(u)πφ(x)ξ. Pq(u)πφ(x)ηφ(n). Definition 5.6. For a Cartan triple (M, N, D), a subset Y ⊆ GN(M, D) is E-orthogonal if whenever v, w ∈ Y with v 6= w, E(v∗w) = 0. topology to x. Theorem 5.7. Let (M, N, D) be a (not necessarily full) Cartan triple and let Y ⊆ GN(M, D) be a maximal E-orthogonal subset. Let Λ be the set of all finite subsets of Y directed by inclusion. For x ∈ M and F ∈ Λ, let xF :=Pu∈F uE(u∗x). Then the net (xF )F ∈Λ converges in the Bures q ։ S be the extension associated to (M, N, D), let ψ be a faithful, normal Proof. Let P ֒→ G semi-finite weight on N, let φ = ψ ◦ E, and let (πφ, Hφ, ηφ) be the semi-cyclic representation of M associated to φ. For any v ∈ G, the map M ∋ x 7→ vE(v∗x) leaves nφ invariant and depends only on s = q(v). Further, when x ∈ nφ, ηφ(x) 7→ ηφ(vE(v∗x)) is contractive, and extends to a projection Ps ∈ B(Hφ). (In the notation of Lemma 4.3 and Proposition 4.13, Ps = W Ps,πφW ∗). When s = 1, write P instead of P1. Arguing as in [5, Lemma 2.2], we find that the two families of semi-norms on M, coincide. These families of semi-norms define the Bures topology on M (see [5, Definition 2.2.3]). We now argue exactly as in the proof of [5, Proposition 2.4.4]. Let n ∈ nφ ∩ N. Then and {M ∋ m 7→ kπφ(m)ξk : ξ ∈ range(P )} By Proposition 3.8 and Proposition 4.13, I = Pu∈Y Pq(u) (where the sum converges strongly in B(Hφ)). Thus for every ξ ∈ ηφ(nφ ∩ N), πφ(xF )ξ → πφ(x)ξ. Therefore, xF Bures → x. (cid:3) We now show that the Bures closure of a weak-∗ closed N-bimodule B contains exactly the same groupoid normalizers as B itself. The reader should note that this result gives the versions of [5, Proposition 2.5.3 and Theorem 2.5.1] appropriate to our context. Proposition 5.8. Let B ⊆ M be a weak-∗ closed N-bimodule, and set B0 = spanw∗(GN(M, D) ∩ B) and B1 := spanBures(GN(M, D) ∩ B), Then B0 and B1 are weak-∗ closed N-bimodules satisfying B0 ⊆ B1 and GN(M, D) ∩ B0 = GN(M, D) ∩ B = GN(M, D) ∩ B1. Furthermore, when (M, N, D) is a full Cartan triple, B0 ⊆ B ⊆ B1 = B Bures . 25 Proof. Notice that if (xλ) is a net in M which Bures-converges to x ∈ M, then for any a ∈ M and b ∈ N, limBures axλb = axb. It follows that B1 is a weak-∗ closed N bimodule. That B0 is an N-bimodule follows from the fact that U(N)GN(M, D)U(N) = GN(M, D) and that span U(N) = N. Clearly B0 ⊆ B1. Suppose v ∈ GN(M, D)∩B1. If (xλ) is a net in B with limBures xλ = v, we find limBures v∗xλ = v∗v. As E is Bures continuous, we have that limBures E(v∗xλ) = E(v∗v) = v∗v. Since the relative Bures topology on N is the σ-strong topology on N, E(v∗xλ) converges weak-∗ to v∗v. By Corollary 5.3, vE(v∗xλ) is a net in B converging weak-∗ to v, showing that v ∈ GN(M, D) ∩ B. Thus, GN(M, D) ∩ B0 = GN(M, D) ∩ B = GN(M, D) ∩ B1. Now suppose (M, N, D) is a full Cartan triple. Clearly B0 ⊆ B ⊆ B1. Let L be the linear span . By Theorem 5.2, for each u ∈ GN(M, D), uE(u∗x) ∈ L, (cid:3) of GN(M, D) ∩ B and choose x ∈ B so Theorem 5.7 shows x ∈ B1, whence B1 = B Bures Bures . Notation 5.9. For a Bures-closed N-bimodule B ⊆ M, let GN(B, D) := GN(M, D) ∩ B. Define Θ(B) ⊆ S by Θ(B) = q(GN(B, D)). Further, define a map Ψ from the collection of spectral sets (see Definition 5.1) in S to Bures-closed N-bimodules in M by Ψ(A) = spanBuresq−1(A) = spanBures{j(a)n : a ∈ A, n ∈ N}, which is necessarily a Bures-closed N-bimodule. When (M, N, D) is full, Theorem 5.2 shows that GN(B, D) is non-zero whenever B 6= (0). We now extend the spectral theorem for bimodules in Cartan pairs (see [5, Theorem 2.5.8] and [12, Theorem 6.3]) to the context of Bures closed bimodules in a Cartan triple. Theorem 5.10 below should also be compared with [16, Theorem 4.3]. Suppose for i = 1, 2 that Pi are full Ni-Clifford inverse monoids, S is a Cartan inverse monoid, qi ։ S are extensions of S by Pi, and let (Mi, Ni, Di) be the corresponding Cartan triples. Pi ֒→ Gi Theorem 5.10 implies the striking fact that the lattice structure of the Bures-closed Ni-bimodules in Mi is isomorphic to the lattice of spectral sets in S. Thus, S completely determines the lattice structure of the Bures-closed Ni-bimodules regardless of the choice of extension of S. Theorem 5.10 (Spectral Theorem for Bimodules). Let (M, N, D) be a full Cartan triple. The map Θ is a lattice isomorphism between the family of Bures-closed N-bimodules in M and the family of spectral sets in S. Moreover, Θ−1 = Ψ. Proof. Let B be a Bures-closed N-bimodule in M and let A := Θ(B). We will first show that A is a spectral set in S. Suppose s ∈ A and t ≤ s. Then there exists an e ∈ E(S) such that t = se. Write s = q(v) for some v ∈ GN(B, D), and e = q(p) for some projection p ∈ D, we find t = q(vp), so the orthogonality of si and sk implies that j(si) and j(sk) are partial isometries with orthogonal t ∈ A. Next, suppose that {si}i∈I is a pairwise orthogonal family in A and let s =W si. For i 6= k, initial spaces and orthogonal range spaces. Therefore, the sumPi∈I j(si) converges strong-∗ to an element v ∈ GN(M, D). As the Bures topology is weaker than the strong-∗ topology, v ∈ GN(B, D). For every i ∈ I, q(vj(s† i si)) = si, and it follows that q(v) = s. Thus j(s) ∈ B, and hence s ∈ A. Therefore A = Θ(B) is a spectral set. Proposition 5.8 shows that B is generated as a N-bimodule by B ∩ GN(M, D). It follows that Ψ(Θ(B)) = B. We now prove that A = Θ(Ψ(A)). Clearly, A ⊆ Θ(Ψ(A)). Choose s ∈ Θ(Ψ(A)) and let B := Ψ(A). By definition, there exists v ∈ GN(B, D) such that q(v) = s. Let r = sup{p ∈ proj(D) : q(vp) ∈ A}. 26 Then q(r) is the maximal idempotent in E(S) such that s q(r) ∈ A. Thus if a ∈ A, s q(r⊥) ∧ a = 0. Therefore, for any n ∈ N, E((vr⊥)∗j(a)) = 0 = E((vr⊥)∗j(a)n). Hence for any x ∈ span(q−1(A)), E((vr⊥)∗x) = 0. As E is Bures continuous, we find that E((vr⊥)∗x) = 0 for every x ∈ B. As vr⊥ ∈ B and E is faithful, we obtain vr⊥ = 0. Hence v = vr. Applying q we obtain, s = s q(r) ∈ A, as desired. Finally, the order preserving properties follow by the definitions of Θ and Ψ. (cid:3) 5.2. Intermediate von Neumann algebras. Our next goal is to give a version of Aoi's Theorem appropriate to our context. We first note the following technical result. Proposition 5.11. Let M ⊇ N be an inclusion of von Neumann algebras and let D ⊆ Z(N) be a von Neumann subalgebra. Assume further that there exists a faithful, normal conditional expectation E : M → N. Let ψ be a faithful normal semi-finite weight on N and let φ = ψ ◦ E. Let σφ t be the modular automorphism group for φ. The following statements hold. t (x) = x ∀ t ∈ R}, for σφ (a) The centralizer, Mφ := {x ∈ M : σφ (b) If v ∈ GN(M, D), then for every t ∈ R, σφ t contains D. t (v) ∈ GN(M, D). Further σφ t (v) is Munn related to v; (c) If A is a von Neumann algebra such that N ⊆ A ⊆ M and D is regular in A, then there is a unique faithful normal conditional expectation EA : M → A such that φ = φ ◦ EA. In addition, EA has the following properties: (i) EAE = EEA = E; and (ii) EA is continuous when regarded as a map of (M, E-Bures) into (M, E-Bures). Proof. For x ∈ M and d ∈ D, E(x∗d∗dx) ≤ kdk2 E(x∗x) and E(d∗x∗xd) = E(x∗x)1/2d∗dE(x∗x)1/2 ≤ kdk2 E(x∗x). Thus nφ is a D-bimodule. Recalling that mφ := nXj=1 y∗ j xj : n ∈ N, xj, yj ∈ nφ , we see that mφ is also a D-bimodule. Furthermore, for any d ∈ D and x ∈ mφ, we have φ(xd) = ψ(E(xd)) = ψ(E(x)d) = ψ(dE(x)) = ψ(E(dx)) = φ(dx). An application of [30, Theorem VIII.2.6] now gives part (a). Now let v ∈ GN(M, D) and let w = σφ t (v). Using (a) we have w∗dw = σφ t (v∗dv) = v∗dv. Thus w ∈ GN(M, D) and w is Munn related to v, proving part (b). The regularity of D in A and part (b) show that σφ t (A) ⊆ A for every t ∈ R. Lemma 2.20 gives φA is a faithful, semi-finite normal weight on A. By [30, Theorem IX.4.2], there exists a unique normal conditional expectation EA : M → A such that φ ◦ EA = φ. Since N ⊆ A, EA ◦ E = E. Let Φ := E ◦ EA. Then Φ is a conditional expectation of M onto N which satisfies φ ◦ Φ = φ. The uniqueness assertion of [30, Theorem IX.4.2] gives Φ = E. We thus have the formula in part (c(i)). As E is faithful, so is EA. Finally, suppose (xλ) is a net in M converging to x in the E-Bures topology. Applying E to both sides of the inequality, (EA(xλ) − EA(x))∗(EA(xλ) − EA(x)) = EA(xλ − x)∗EA(xλ − x) ≤ EA((xλ − x)∗(xλ − x)) 27 and using the fact that EEA = E shows that EA(xλ) → EA(x) in the E-Bures topology. Thus EA is E-Bures continuous. (cid:3) Theorem 5.12 (Aoi's Theorem for Cartan Triples). Let (M, N, D) be a Cartan triple and suppose A is a von Neumann algebra such that N ⊆ A ⊆ M. Then A is Bures closed. Furthermore, if (M, N, D) is full, then (A, N, D) is a Cartan triple. Proof. Let A0 be the weak-∗ closure of span GN(A, Z(N)). Then A0 is a von Neumann algebra and, as U(N) ⊆ GN(A, Z(N)), A0 ⊇ N. Thus, A0 is a weak-∗ closed N-bimodule. By Proposition 5.11, there exists a faithful, normal conditional expectation EA0 : M → A0. Since EA0 is E-Bures continuous, it follows that A0 is E-Bures closed. By Proposition 5.8, A0 ⊆ A ⊆ A Bures = spanBuresGN(A, Z(N)) = A0 Bures = A0, so A is Bures closed. When (M, N, D) is full, that is, D = Z(N), the previous paragraph shows that D is regular in A, (cid:3) so (A, N, D) is a Cartan triple. Remark 5.13. With the notation of Theorem 5.12 and its proof, let A00 be the weak-∗ closure of span GN(A, D). Then N ⊆ A00 ⊆ A0, and A00 is Bures closed. However, we have been unable to show A00 = A0 in general, which is why we required the fullness hypothesis to conclude (A, N, D) is a Cartan triple. However, this hypothesis is rather mild, and is satisfied when D is a Cartan MASA in M. Thus Theorem 5.12 is indeed a generalization of Aoi's theorem for Cartan pairs. As an immediate corollary, we use the Spectral Theorem for Bimodules to parametrize the intermediate von Neumann algebras for a full Cartan triple. As with Bures-closed bimodules, this parametrization depends only on the Cartan inverse monoid and not the extension. Corollary 5.14. Let (M, N, D) be a full Cartan triple. Set vN(M, N, D) := {A : A is a von Neumann algebras such that N ⊆ A ⊆ M} and sub(S) := {T ⊆ S : T is a Cartan inverse submonoid of S with E(T) = E(S)}. Then the restriction of Θ to vN(M, N, D) gives a bijection between vN(M, N, D) and sub(S). In this section, we give several examples of Cartan triples. 6. Examples 6.1. Type I examples. Suppose a Hilbert space H is decomposed as as a direct sum, H = Li∈I Hi, where for all i, j ∈ I, dim Hi = dim Hj. Let M = B(H) and D be the von Neumann algebra generated by {Pi : Pi is the projection onto Hi, i ∈ I}. Then N = D′ = ⊕i∈I B(Hi) and (M, N, D) is a full Cartan triple. Indeed, M ∼= B(ℓ2(I))⊗B(H1), D ∼= D(ℓ2(I))⊗CIH1, and N ∼= D(ℓ2(I))⊗B(H1), (6.1) where D(ℓ2(I)) are the diagonal operators in B(ℓ2(I)). We now show every Cartan triple (M, N, D) with M = B(H) has the form outlined above, and hence is necessarily full. Showing that D is atomic is the key step. To start, let P be the projection onto the closure of the span of the ranges of the minimal projections in D. We argue by contradiction to show P = I. If P 6= I, fix a unit vector η ∈ P ⊥H and a positive integer n. Choose a maximal chain P in proj(P ⊥D) (with respect to the ordering ≤ in proj(P ⊥D)). The map from P into [0, 1] given by P ∋ R 7→ kRηk is onto [0, 1] since P has no atoms. So for 0 ≤ j ≤ n, let Rj ∈ P be such that kRjηk = j/n and for 1 ≤ j ≤ n put Qj := Rj − Rj−1. Thus kQjηk = 1/n for every j. 28 If X is the rank-one projection onto the span of η, then E(X) = E nXi=1 QiX! = QiE(X) = E nXi=1 nXi=1 QiXQi! . As kQiXQik = 1/n for each i, kE(X)k ≤ 1/n for all choices of n and so E(X) = 0, contradicting faithfulness of E. Thus P ⊥ = 0 and D is atomic. Finally, since GN(M, D) spans M, any two atoms of D, say A and B, must have the same dimension, since otherwise AMB ∩ GN(M, D) = {0}, contradicting the regularity of D in M. By the previous paragraph, for every non-zero minimal projection P ∈ Z(M), (MP, NP, DP ) is a full Cartan triple. As a consequence, we have the following observation. Proposition 6.1. If (M, N, D) is a Cartan triple with dim(M) < ∞, then (M, N, D) is full. 6.2. Tensoring Cartan pairs. Equations (6.1) decomposed a Cartan triple into tensor products, where M ∼= B(ℓ2(I))⊗B(H1) and N = D(ℓ2(I))⊗B(H1). Note that D(ℓ2(I)) is a Cartan subalgebra of B(ℓ2(I)). In fact, starting with any Cartan pair we can create a Cartan triple by tensoring with a von Neumann algebra. Suppose M is a von Neumann algebra, D ⊆ M is a Cartan MASA and let N be any von Neumann algebra. Consider the von Neumann algebras D ⊗ IN ⊆ D⊗N ⊆ M⊗N. Since D is regular in M it follows that D ⊗ IN is regular in M⊗N. Further, the conditional expectation E : M → D induces a faithful conditional expectation E⊗idN : M⊗N → D⊗N. By [31, Theorem IV.5.9 and Corollary IV.5.10], D⊗N = (D⊗IN)c. Thus (M⊗N, D⊗N, D ⊗ IN) is a Cartan triple. Further, if N is a factor (M⊗N, D⊗N, D ⊗ IN) is a full Cartan triple. 6.3. Crossed products by discrete groups. Cartan triples arise naturally as crossed product von Neumann algebras. In Section 6.3.1 we will show that if G is a discrete group acting on an abelian von Neumann algebra D then (D ⋊α G, Dc, D) will always give a Cartan triple. If a discrete group G acts on a (not necessarily abelian) von Neumann algebra N, and D = Z(N), we give necessary and sufficient conditions for (N ⋊α G, N, D) to be a Cartan triple in Section 6.3.2. Let G be a discrete group acting on a von Neumann algebra N by automorphisms α. Let M = N⋊αG. The von Neumann algebra M is generated by a copy of N and a unitary representation of G, {ug}g∈G such that αg(d) = ugdu∗ g. There is a faithful, normal conditional expectation EN from M onto N. Each element x ∈ M is uniquely determined by a Fourier series x =Xg∈G xgug, where xg := EN(xu∗ g) ∈ N. This series converges in the Bures-topology on M induced by EN [23]. Cameron and Smith [6, 8] have studied Bures-closed bimodules and intermediate von Neumann algebras in a large class of crossed products. We will see in Theorem 6.3 that there is overlap in our work and theirs. However, neither work subsumes the other. 6.3.1. Crossed products of abelian algebras. Theorem 6.2. Let D be an abelian von Neumann algebra and let G be a discrete group acting on D by automorphisms α. Let M := D ⋊α G and N = Dc. Then (M, N, D) is a Cartan triple. Proof. Since D is clearly regular in D ⋊α G, we only need to note that there is a faithful normal conditional expectation from M onto N. Since there is a faithful, normal conditional expectation ED from M onto D, D is regular in N, and D ⊆ N ⊆ M, this follows from Proposition 5.11(c). Alternatively, the existence of the conditional expectation onto N also follows from the proof of 29 Theorem 3.2 of [8]. In [8] it is assumed that the action of G is by properly outer automorphisms, though this is not needed in the proof. (cid:3) We give further details on the structure of this Cartan triple. For each g ∈ G, let pg be the largest projection in D such that αgDpg is the identity. We note pg is the Frol´ık projection e0 for adug on D described in the proofs of Lemma 2.18 and Proposition 4.9. By [26, Lemma 2.15], ugpg = pgug ∈ N and EN(ug) = ugpg. Since EN is Bures continuous, we can explicitly describe EN by ENXg∈G xgug =Xg∈G xgpgug. 6.3.2. Crossed products of non-abelian algebras. An automorphism α on a von Neumann algebra N is properly outer if there are no nonzero central projections z ∈ Z(N) such that αNz is inner. Equivalently α is properly outer if and only if yx = xα(y) for all y ∈ N implies that x = 0. In [8] crossed products by properly outer automorphisms are studied and the Bures-closed bimodules and intermediate von Neumann algebras are characterized. We show now that the crossed products studied in [8] give rise to full Cartan triples under the assumption that the restriction of the action to the center Z(N) is also properly outer. Theorem 6.3. Let N be a von Neumann algebra and let G be a discrete group acting on N by properly outer automorphisms α. Let M = N ⋊α G. Then (M, N, Z(N)) is a Cartan triple if and only if the action of G restricted to the center Z(N) is properly outer. Proof. Suppose x ∈ Z(N)′ ∩ M. Let x =Pg∈G xgug be the (Bures convergent) Fourier series for x. Since x ∈ Z(N)′ it follows that if xg 6= 0 then for d ∈ Z(N), g(ugdu∗ g)) = xgαg(d). dxg = EN(dxu∗ g) = EN(xu∗ (6.2) Let Jg be the two-sided ideal in N generated by xg. It follows from (6.2) that xd = xα(d) for all x ∈ Jg and all d ∈ Z(N). Since Jg is a two-sided ideal, there is a central projection zg ∈ Z(N) such that Jg = Nzg. Thus zgd = zgαg(d) for all d ∈ Z(N). That is, αgZ(N)zg = idZ(N)zg . It follows that N = Z(N)c if and only if for all g 6= e, αgZ(N) is properly outer. (cid:3) 6.4. Crossed products by equivalence relations. Igor Fulman in [17] studied a class of Cartan triples which he called crossed products by an equivalence relation. A crossed product by an equivalence relation is a Cartan triple satisfying the condition in Definition 6.4 below, which we also call Fulman's condition. In Appendix A we provide a conceptual framework in terms of inverse semigroups for Fulman's condition and show that Fulman's condition amounts to a lifting problem. Here we give a class of Cartan triples which satisfy Fulman's condition. Suppose (M, N, D) is a Cartan triple with associated extension, P ֒→ G q ։ S and fixed order preserving section j. Definition 6.4. A regularizer is a subgroup R ⊆ U(M) satisfying: (a) U(N) ⊆ R ⊆ GN(M, D); (b) span R is weak-∗ dense in M; (c) there is a homomorphism α : R → Aut(N) such that (a) if p is a projection in D such that αuDp = idDp then αuNp = idNp. (b) αu(d) = udu∗ for each u ∈ R and d ∈ D. 30 We will call a map α satisfying conditions (i) and (ii) of part (c) a regularizing map for R. When the Cartan triple (M, N, D) has a regularizer, we say (M, N, D) satisfies Fulman's condi- tion. Note that if R is a regularizer with regularizing map α, then ker α = U(N) ([17, Remark, pg. 41]). Example 6.5. Let N be a von Neumann algebra and let D = Z(N). Let G be a discrete group acting on N by properly outer automorphisms. Further assume that the restriction of the action of G to D is properly outer. Let M = N ⋊α G. Then by Theorem 6.3 (M, N, D) is a Cartan triple. Let R be the group generated by {ug : g ∈ G} ∪ {u ∈ N : u unitary}. Let R := {ug : g ∈ G} and let α : R → Aut(N) be ug 7→ adug . Since G acts by properly outer automorphisms on Z(N), α is a regularizing map for R so that R is a regularizer. Thus (M, N, D) satisfies Fulman's condition. Appendix A. An Inverse Semigroup Description of Fulman's Condition Fulman's condition as stated in Definition 6.4, is mysterious. Our goal in this appendix is to establish Theorem A.8, which shows Fulman's condition is equivalent to the statement that a rather natural lifting problem for inverse semigroups (Diagram A.1) has a positive solution. We begin with a definition. Definition A.1. Let (N, D) be a pair of von Neumann algebras with D a von Neumann subalgebra of Z(N). A partial automorphism of (N, D) is a triple (e, α, f ) consisting of projections e, f ∈ D and a normal ∗-isomorphism α : f N → eN satisfying α(eD) = f D. We will use pAut(N, D) for the set of all partial automorphisms of (N, D). If f = 0 (or e = 0) we say (e, α, f ) is the zero element of pAut(N, D). Further, define an involution and a product in pAut(N, D) via, (e, α, f )† := (f, α−1, e) and (e1, α1, f1)(e2, α2, f2) := (α1(f1e2), (α1 ◦ α2)α−1 2 (e2f1), α−1 2 (f1e2)). Then pAut(N, D) is an inverse monoid with 0. Also E(pAut(N, D)) = {(e, ideN, e) : e ∈ proj(D)} and hence may be identified with proj(D). For γ = (e, α, f ) ∈ pAut(N, D) and x ∈ f N, we write γ(x) for the value of α at x. We require the following notions for an inverse semigroup R. • Two elements s, t ∈ R are compatible if st† and s†t are idempotents; a subset A ⊆ R is compatible ([21, page 26] if every pair of elements of A is compatible. • R is infinitely distributive ([21, page 28]) if whenever I is an index set and {ri}i∈I ⊆ R is such thatWi∈I ri exists then for any s ∈ R, ris exist and s _i∈I • R is complete ([21, page 27]) if whenever A ⊆ R is a compatible subset,W A exists. ri! s =_i∈I sri, _i∈I Lemma A.2. The inverse semigroup pAut(N, D) is infinitely distributive and complete. _i∈I sri and _i∈I ri! =_i∈I ris. Proof. As proj(D) is a complete Boolean algebra, [21, Proposition 1.4.20] shows pAut(N, D) is an infinitely distributive inverse semigroup. We turn now to showing pAut(N, D) is complete. Given a = (ea, αa, fa) ∈ pAut(N, D), identify a†a with fa and aa† with ea, so that the source and range of a belong to proj(D). 31 First suppose that A ⊆ pAut(N, D) is a finite and orthogonal set. Let e = Wa∈A aa† and f =Wa∈A a†a. For n ∈ Nf , n =Pa∈A na†a and define α(n) :=Xa∈A αa(na). Then (e, α, f ) ∈ pAut(N, D). For a ∈ A, (e, α, f )(fa, idNfa, fa) = a so a ≤ (e, α, f ). Further, if for every a ∈ A, a ≤ (e′, α′, f ′), then (e, α, f ) = (e′, α′, f ′)(f, idNf , f ) = (e, α, f ). Thus, (e, α, f ) = Next, suppose A is a finite compatible set. Let B be the (finite) Boolean algebra generated by Let C := {ap : a ∈ A, p ∈ atom(B)}. Then C is a finite orthogonal set of elements in pAut(N, D). ap and W A. So joins exist for finite orthogonal sets. {a†a : a ∈ A}. The identity of B is f :=Wa∈A a†a. Let atom(B) be the (finite) set of atoms of B. Let (e, α, f ) := W C. Let a ∈ A and Pa := {p ∈ atom(B) : ap 6= 0}. Then a = Pp∈Pa a†a = fa =Pp∈Pa a ∈ A, then (e′, α′, f ′)(f, idNf , f ) = (e, α, f ) so (e, α, f ) =W A, showing joins exist for any finite subsets of A ordered by inclusion and for F ∈ F, let aF = W F . Notice that if F1 ⊆ F2, then aF1 ≤ aF2. Write aF = (eF , αF , fF ). Let e =W{aa† : a ∈ A} =WF ∈F eF and f =W{a†a : a ∈ A} = WF ∈F fF . For n ∈ Nf , the net αF (nfF ) converges strongly, and we define α(n) = lim αF (nfF ). Then (e, α, f ) =W A. This shows that for every a ∈ A, a ≤ (e, α, f ). On the other hand, if a ≤ (e′, α′, f ′) for every Finally, let A ⊆ pAut(N, D) be an arbitrary compatible subset. Let F be the set of all finite The inverse semigroup pAut(N, D) may be written as an extension, p ≤ f . So for every p ∈ Pa, (e, α, f )(p, idNp, p) = ap. Thus (e, α, f )(p, idNp, p) = ap, whence (e, α, f )(fa, idNfa, fa) = a. compatible set. (cid:3) Cliff(pAut(N, D)) ֒→ pAut(N, D) π ։ Fund(pAut(N, D)) where Cliff(pAut(N, D)) is the Clifford inverse subsemigroup of all elements of pAut(N, D) which are Munn related to an idempotent, and Fund(pAut(N, D)) is the quotient of pAut(N, D) by the Munn relation. Henceforth, fix a Cartan triple (M, N, D) with associated extension P ֒→ G q ։ S and order- preserving section j. We shall be interested in the semigroup pAut(N, D) arising from this Cartan triple. The idempotents of pAut(N, D) (and hence those of Fund(pAut(N, D))) may be identified with E(S). We shall show that for any Cartan triple, there is a one-to-one inverse semigroup homomor- phism θ : S → Fund(pAut(N, D)) which fixes idempotents. Our goal in this section is to show that Fulman's condition is satisfied if and only if there is a lifting of θ to an inverse semigroup homomorphism α so that the following diagram commutes: pAut(N, D) 7♣ ♣ ♣ α ♣ ♣ π ♣ S ♣ θ Fund(pAut(N, D)). (A.1) For (e, α, f ) ∈ pAut(N, D), let [e, α, f ] ∈ Fund(pAut(N, D)) denote the Munn equivalence class of (e, α, f ). It will be helpful to have an explicit description of the Munn relation on pAut(N, D). Lemma A.3. For i = 1, 2, let (ei, αi, fi) ∈ pAut(N, D). The following are equivalent. (a) (e1, α1, f1) is Munn related to (e2, α2, f2); (b) for every d ∈ proj(D), α1(df1) = α2(df2); 32 / / 7 (c) e1 = e2, f1 = f2 and α1f1D = α2f2D. Proof. Suppose (a) holds. Then for any d ∈ proj(D), (d, iddN, d) ∈ E(pAut(N, D)), so (ei, αi, fi)(d, iddN, d)(fi, α−1 i , ei) = (αi(dfi), idαi(dfi)N, αi(dfi)), (A.2) which yields (b). Now suppose (b) holds. Taking d ∈ {f1, f2, f1f2} gives α1(f1) = α2(f1f2) = α1(f1f2) = α2(f2), so that f1 = f2 and e1 = e2. Since fiD is generated by proj(fiD), we obtain (c). Finally, assume (c) holds. Let d ∈ proj(D). Examining (A.2) we obtain (e1, α1, f1)(d, iddN, d)(f1, α−1 1 , e1) = (α1(df1), idα1(df1)N, α1(df1)) = (α2(df2), idα2(df2)N, α2(df2)) = (e2, α2, f2)(d, iddN, d)(f2, α−1 2 , e2). Thus (a) holds and the proof is complete. (cid:3) We now observe that there is always a one-to-one inverse semigroup homomorphism of S into Fund(pAut(N, D)). Note that if v ∈ G, then v defines a partial automorphism in pAut(N, D). Indeed if we define adv by adv : v∗vN → vv∗N v∗vx 7→ vxv∗, then (vv∗, adv, v∗v) ∈ pAut(N, D). We define a map θ : S → Fund(pAut(N, D))) by θ(s) = [j(ss†), adj(s), j(s†s)]. By Lemma A.3, if v, w ∈ G and v and w are Munn equivalent, then [vv∗, adv, v∗v] = [ww∗, adw, w∗w]. Hence the map θ is independent of the choice of j. Indeed, for any w ∈ q−1{s}, θ(s) = [ww∗, adw, w∗w]. Thus we may use any of [j(ss†), αj(s), j(s†s)], [ss†, αs, s†s], or [j(ss†), αs, j(s†s)] to denote θ(s). Proposition A.4. The map θ : S → Fund(pAut(N, D)) given by θ(s) := [j(ss†), ads, j(s†s)] is a one-to-one homomorphism of inverse semigroups such that θE(S) is an isomorphism of E(S) onto E(Fund(pAut(N, D))). Proof. For e ∈ E(S), θ(e) = [j(e), idj(e)N, j(e)], so θE(S) is an isomorphism of E(S) onto E(Fund(pAut(N, D))). Take s1, s2 ∈ S. Then On the other hand, ad−1 [j(s1s† 1), ads1, j(s† θ(s1s2) = [j(s1s2s† 1s1s2s† 2s† 2), ads2, j(s† 2)) = j(s† s2 (j(s† 1s1)][j(s2s† 1), ads1s2, j(s† 2s† 1s1s2) and ads1(j(s† 2s† 1s1s2)]. 1s1s2s† 2s2)] = [j(s1s2s† 2s† 1), ads1 ◦(ads2 j(s† 2)) = j(s1s2s† 2s† 1s1s2)N), j(s† 2s† 1), so 2s† 1s1s2)]. Thus to show that θ is a homomorphism it suffices to show that ads1s2 = ads1 ◦(ads2 j(s† 1s1s2)N). 2s† Note that for each s ∈ S and e ∈ E(S), θ(s)(j(e)) = j(s)j(e)j(s)∗ = j(ses†). 33 Hence for e ∈ E(S), ads1s2(j(e)j(s† 1s1s2))). An application of Lemma A.3 now shows that θ is multiplicative on S. Hence θ is an inverse semigroup homo- morphism. 1s1s2)) = ads1(ads2(j(e)j(s† 2s† 2s† If θ(s1) = θ(s2), then for every e ∈ E(S), ads1 j(e)D = ads2 j(e)D, so that in particular, s1es† s2es† 2 for every e ∈ E(S). As S is fundamental, s1 = s2, whence θ is one-to-one. 1 = (cid:3) We now show that a regularizer may be viewed as a homomorphism of q(R) into Aut(N) satisfying Fulman's conditions. Lemma A.5. Suppose a regularizer R exists for (M, N, D). Let α : R → Aut(N) be a regularizing map. Then α induces a one-to-one group homomorphism α : q(R) → Aut(N, D) such that for every e ∈ E(P) and U ∈ R, αq(U )(e) = adU (e) = j(q(U ))ej(q(U ))∗. Proof. By condition (c)(ii) of Definition 6.4, (I, αU , I) ∈ Aut(N, D) for every U ∈ R. Applying condition (c)(i) of Definition 6.4, it follows that there exists a one-to-one group homomorphism α : q(R) → Aut(N, D). If e ∈ E(P), and U ∈ R, then αq(U )(e) = αU (e) = U eU −1. (cid:3) Lemma A.6. Let R be a regularizer for (M, N, D) and let R := {q(s)e : s ∈ R, e ∈ E(S)}. Then R is an inverse semigroup and S is isomorphic to the join completion of R. Proof. A calculation shows R is an inverse semigroup, and by definition, S is complete. Notice that every compatible order ideal of R is also a compatible order ideal of S. Thus by the proof of [21, Theorem 1.4.23], the join completion of R is contained in S. Take s ∈ S and let t =W{q(r)∧s : r ∈ R}. Suppose t 6= s. As {a ∈ S : a ≤ s} is a Boolean algebra, there is a u ∈ S such that u ∨ t = s and u ∧ t = 0. There is a w ∈ GN(M, D) such that q(w) = u. As R densely spans M, there is a U ∈ R such that E(U ∗w) 6= 0. Hence v = U E(U ∗w) 6= 0. Note that v ∈ GN(M, D) and q(v) = q(U )q(E(u∗w)) = q(U )(q(U ∗)q(w) ∧ 1) = q(w) ∧ q(U ) ≤ s ∧ q(U ). Hence q(v) ≤ t. However, q(v) ≤ u. Hence u = 0, and t = s. For every r ∈ R, q(r) ∧ s ∈ R. Hence the completion of R is S. (cid:3) Next we show that Fulman's condition implies that there is a homomorphism of S into pAut(N, D) which lifts the map θ described in Proposition A.4. Lemma A.7. Suppose Γ ⊆ S is a group (under the multiplication inherited from S) whose unit is 1 ∈ S. Assume that α : Γ → Aut(N, D) is a one-to-one homomorphism such that for every s ∈ Γ, Fulman's condition (c) is satisfied for αs, that is, (i) if p ∈ proj(D) satisfies αspD = idpD, then αspN = idpN; and (ii) for d ∈ D, αs(d) = j(s)dj(s)∗. Let SΓ ⊆ S be the smallest Cartan inverse submonoid of S containing Γ and E(S). Then α extends uniquely to a one-to-one homomorphism α′ : SΓ → pAut(N, D). In addition π ◦ α′ = θSΓ. Proof. Let R := {se : s ∈ Γ, e ∈ E(S)}. Since Γ is a group, R is an inverse semigroup. As in the proof of Lemma A.5(b), SΓ is the join completion of R. We shall show that there is a multiplicative map of α′ : R → pAut(N, D). Suppose s, t ∈ Γ. Fulman's condition (c) applied to s†t shows that if p ∈ proj(D) and αspD = αtpD, then αspN = αtpN. For s ∈ Γ and e ∈ E(S), define α′(se) := (j(ses†), αsj(e)N, j(e)). 34 Note that this is well-defined, for if se = tf for some idempotents e, f and t ∈ Γ, then αs†tf eD = idf eD, so αsef N = αtef N. Thus, α′ : R → pAut(N, D) is well-defined. For s, t ∈ Γ and e, f ∈ E(S) a calculation shows that α′((se)(tf )) = α′(st)α′(tf ), so α′ is a homomorphism. Also, for any s ∈ Γ and e ∈ E(S), π(α′(se)) = θ(se). By [21, Theorem 1.4.24], α′ extends uniquely to a join-preserving homomorphism of SΓ into pAut(N, D). Take s ∈ Γ. Recall θ(s) = [j(ss†), ads, j(s†s)] = [j(1), ads, j(1)]. Since αs(d) = j(s)dj(s)∗ for all d ∈ D, by Lemma A.3, π ◦ α(s) = θ(s). That π ◦ α′ = θΓS now follows from the definition of α′. Since θ is a one-to-one map it follows that α′ is one-to-one. (cid:3) We now are prepared to recast Fulman's condition as a lifting problem. q Theorem A.8. Let (M, N, D) be a Cartan triple with associated extension P ֒→ G ։ S. Then (M, N, D) satisfies Fulman's condition if and only if there exists a homomorphism of inverse semi- groups α : S → pAut(N, D) such that π ◦ α = θ. Proof. Suppose (M, N, D) satisfies Fulman's condition. Combining Lemmas A.5 and A.7 we obtain a homomorphism α : S → pAut(N, D) such that π ◦ α = θ. Conversely, suppose α : S → pAut(N, D) is a homomorphism satisfying π ◦ α = θ. Let R := U(M) ∩ G. Clearly U(N) ⊆ R ⊆ GN(M, D) and span R is weak-∗ dense in M. Let τ := α ◦ qR. Then τ : R → Aut(N) is a homomorphism. For u ∈ R write τu instead of τ (u). We claim that for u ∈ R and d ∈ D, τu(d) = udu∗. Since π◦τ = θ, we obtain π(τu) = θ(q(u)), that is, [1, τu, 1] = [1, adq(u), 1]. By Lemma A.3, we obtain τuD = adq(u) D. But, using Proposition A.4, for every d ∈ D, adq(u)(d) = udu∗. The claim follows. Suppose e ∈ proj(D) and τueD = ideD. Let s = q(ue) and note that s†s = q(e). For f ∈ E(S) we have sf s† = q(uej(f )pu∗) = q(τu(ej(f ))) = q(ej(f ))) = q(e)f q(e)†. Since S is fundamental, we obtain s = q(e). So s is an idempotent. Therefore α(s) ∈ pAut(N, D) is idempotent, which is to say that α(s) = ideN. Since α(s) = τueN, we find that τueN = ideN. This completes the proof. (cid:3) Remark A.9. While we do not presently have an example, it seems unlikely that for a general Cartan triple, this lifting problem will have a solution. Thus, we expect that there should be an example of a Cartan triple which is not a crossed product by an equivalence relation. A sufficient condition for a solution to the lifting problem is if the map j : S → G can be chosen so that j(st)∗j(s)j(t) ∈ D. In this case the map α : s 7→ (j(ss†), ads, j(s†s)) can be shown to be homomorphism. Clearly θ = π ◦ α. References 1. Hisashi Aoi, A construction of equivalence subrelations for intermediate subalgebras, J. Math. Soc. Japan 55 (2003), no. 3, 713 -- 725. MR 1978219 (2004c:46120) 2. Hisashi Aoi and Takehiko Yamanouchi, On the normalizing groupoids and the commensurability groupoids for inclusions of factors associated to ergodic equivalence relations -- subrelations, J. Funct. Anal. 240 (2006), no. 2, 297 -- 333. MR 2261685 3. William B. Arveson, Analyticity in operator algebras, Amer. J. Math. 89 (1967), 578 -- 642. MR 0223899 (36 #6946) 4. Donald Bures, Abelian subalgebras of von Neumann algebras, American Mathematical Society, Providence, R.I., 1971, Memoirs of the American Mathematical Society, No. 110. MR 0296706 (45 #5765) 5. Jan Cameron, David R. Pitts, and Vrej Zarikian, Bimodules over Cartan MASAs in von Neumann algebras, norming algebras, and Mercer's theorem, New York Journal of Mathematics 19 (2013), 455 -- 486. 6. Jan Cameron and Roger R. Smith, Bimodules in crossed products of von Neumann algebras, Adv. Math. 274 (2015), 539 -- 561. MR 3318160 35 7. , A Galois correspondence for reduced crossed products of unital simple C∗-algebras by discrete groups, Canad. J. Math. (2018), to appear. 8. Jan M. Cameron and Roger R. Smith, Intermediate subalgebras and bimodules for general crossed products of von Neumann algebras, Internat. J. Math. 27 (2016), no. 11, 1650091, 28. MR 3570376 9. Hisashi Choda, A Galois correspondence in a von Neumann algebra, Tohoku Math. J. (2) 30 (1978), no. 4, 491 -- 504. MR 516882 10. A. Connes, A factor not anti-isomorphic to itself, Ann. Math. (2) 101 (1975), 536 -- 554. MR 0370209 (51 #6438) 11. H. D'Alarcao, Idempotent-separating extensions of inverse semigroups, J. Austral. Math. Soc. 9 (1969), 211 -- 217. MR 0238970 (39 #330) 12. Allan P. Donsig, Adam H. Fuller, and David R. Pitts, Von Neumann algebras and extensions of inverse semi- groups, Proc. Edinb. Math. Soc. (2) 60 (2016), no. 1, 57 -- 97. MR 3589841 13. Ruy Exel, Noncommutative Cartan subalgebras of C ∗-algebras, New York J. Math. 17 (2011), 331 -- 382. MR 2811068 (2012f:46131) 14. Jacob Feldman and Calvin C. Moore, Ergodic equivalence relations, cohomology, and von Neumann algebras. I, Trans. Amer. Math. Soc. 234 (1977), no. 2, 289 -- 324. MR 58 #28261a 15. , Ergodic equivalence relations, cohomology, and von Neumann algebras. II, Trans. Amer. Math. Soc. 234 (1977), no. 2, 325 -- 359. MR 58 #28261b 16. Adam H. Fuller and David R. Pitts, Isomorphisms of lattices of Bures-closed bimodules over Cartan MASAs, New York Journal of Mathematics 19 (2013), 657 -- 668. 17. Igor Fulman, Crossed products of von Neumann algebras by equivalence relations and their subalgebras, Mem. Amer. Math. Soc. 126 (1997), no. 602, x+107. MR 1371091 (98d:46073) 18. Masaki Izumi, Roberto Longo, and Sorin Popa, A Galois correspondence for compact groups of automorphisms of von Neumann algebras with a generalization to Kac algebras, J. Funct. Anal. 155 (1998), no. 1, 25 -- 63. MR 1622812 19. E. Christopher Lance, Hilbert C ∗-modules: a toolkit for operator algebraists, London Mathematical Society Lec- ture Note Series, vol. 210, Cambridge University Press, Cambridge; New York, 1995. 20. Hans Lausch, Cohomology of inverse semigroups, J. Algebra 35 (1975), 273 -- 303. MR 0382521 (52 #3404) 21. Mark V. Lawson, Inverse semigroups, World Scientific Publishing Co. Inc., River Edge, NJ, 1998, The theory of partial symmetries. MR 1694900 (2000g:20123) 22. V. M. Manuilov and E. V. Troitsky, Hilbert C ∗- and W ∗-modules and their morphisms, J. Math. Sci. (New York) 98 (2000), no. 2, 137 -- 201, Functional analysis, 6. MR 1755888 23. Richard Mercer, Convergence of Fourier series in discrete crossed products of von Neumann algebras, Proc. Amer. Math. Soc. 94 (1985), no. 2, 254 -- 258. MR 784174 24. , Bimodules over Cartan subalgebras, Rocky Mountain J. Math. 20 (1990), no. 2, 487 -- 502. MR MR1065846 (91f:46081) 25. Paul S. Muhly, Kichi-Suke Saito, and Baruch Solel, Coordinates for triangular operator algebras, Ann. Math. 127 (1988), 245 -- 278. MR 0932297 (89h:46088) 26. David R. Pitts, Structure for regular inclusions. I, J. Operator Theory 78 (2017), no. 2, 357 -- 416, doi: 10.7900/jot.2016sep15.2128. 27. S. Popa and S. Vaes, Unique Cartan decomposition for II1 factors arising from arbitrary actions of free groups, Acta Math. 212 (2014), 141 -- 198. 28. , Unique Cartan decomposition for II1 factors arising from arbitrary actions of hyperbolic groups, J. Reine Angew. Math. 694 (2014), 215 -- 239. 29. N. Sieben, C ∗-crossed products by partial actions and actions of inverse semigroups, J. Austral. Math. Soc. Ser. A 63 (1997), no. 1, 32 -- 46. 30. M. Takesaki, Theory of operator algebras. II, Encyclopaedia of Mathematical Sciences, vol. 125, Springer-Verlag, Berlin, 2003, Operator Algebras and Non-commutative Geometry, 6. MR 1943006 (2004g:46079) 31. Masamichi Takesaki, Theory of operator algebras. I, Springer-Verlag, New York, 1979. MR 81e:46038 32. D. Voiculescu, The analogues of entropy and of fisher's information measure in free probability theory. iii. the absence of cartan subalgebras, Geom. Funct. Anal. 6 (1996), no. 1, 172 -- 199. 36 Dept. of Mathematics, University of Nebraska-Lincoln, Lincoln, NE, 68588-0130 E-mail address: [email protected] Dept. of Mathematics, Ohio University, Athens, OH, 45701 E-mail address: [email protected] Dept. of Mathematics, University of Nebraska-Lincoln, Lincoln, NE, 68588-0130 E-mail address: [email protected] 37
1607.07276	4	1607	2017-10-18T03:57:35	The quantum group fixing a sequence of finite subsets	[ "math.OA", "math.DS", "math.NT" ]	Motivated by generalizing Szemer\'edi's theorem, we the elements in a discrete quantum group fixing a sequence of finite subsets and prove that the set of these elements is a quantum subgroup. Using this we obtain a version of mean ergodic theorem for discrete quantum groups.	math.OA	math	THE QUANTUM GROUP FIXING A SEQUENCE OF FINITE SUBSETS HUICHI HUANG Abstract. Motivated by generalizing Szemer´edi's theorem, we the elements in a discrete quantum group fixing a sequence of finite subsets and prove that the set of these elements is a quantum subgroup. Using this we obtain a version of mean ergodic theorem for discrete quantum groups. 7 1 0 2 t c O 8 1 ] . A O h t a m [ 4 v 6 7 2 7 0 . 7 0 6 1 : v i X r a 1. Introduction In [Sze75], E. Szemer´edi proved the following theorem conjectured by P. Erdos and P. Tur´an [ET36], which generalizes van der Waerden's theorem [vand27]. Theorem 1.1. [Szemer´edi's theorem] A set of positive integers with positive upper density contains arbitrarily long arith- metic progressions. Szemer´edi's original proof is combinatorial and has merits on its own right [Sze69, Sze75]. A good survey is [Tao07]. One may ask the following question: what's the reason behind the fact that a set with positive upper density contains arbitrarily long arithmetic progressions? In this paper, we prove a generalized Szemer´edi's theorem and give a partial answer to this question. Theorem 1.2. Let Σ = {Fn} be a sequence of finite subsets in a discrete group Γ and suppose b in Γ fixes Σ from right (left). If a subset Λ of Γ has positive upper density with respect to Σ, then for any positive integer k, there exist n > 0 and a ∈ Γ such that {bjna}k−1 j=0) is contained in Λ . j=0 ({abjn}k−1 Date: May 12, 2018. 2010 Mathematics Subject Classification. Primary 37A30,37A45, 11B25, 43A05, 43A07, 46L65. Key words and phrases. Discrete quantum group, sequence of finite subsets. The author is partially supported by the Fundamental Research Funds for the Central Univer- sities No. 0208005202045. 1 2 HUICHI HUANG Here we say that b in Γ fixes Σ if lim n→∞ bFn∆Fn Fn = 0 ( lim n→∞ Fnb∆Fn Fn = 0). The upper density DΣ(Λ) of a subset Λ of Γ with respect to Σ is defined by lim sup n→∞ Fn ∩ Λ Fn [BBF10]. It's easy to see that the set ΓΣ of elements in Γ fixing Σ is a subgroup of Γ. For discrete quantum groups, one can still define the subset fixing a sequence of finite sets. Moreover we can prove that it is a discrete quantum subgroup. Theorem 1.3. Given a sequence Σ of finite subsets of bGΣ for a compact quantum group G, the C ∗-algebra C ∗(bGΣ) is a compact quantum group. Mean ergodic theorem for amenable discrete quantum group already appears in [Hua16-2]. Using the concept defined above, we prove a mean ergodic theorem in the setting of arbitrary discrete quantum groups. 1 Theorem 1.4. Fnw Pα∈Fn dαπ(χ(α))}∞ (i) Suppose T is a limit of { 1 n=1 in B(H). Then T P = P T = P where P is the orthogonal projection from H onto HΣ := {x ∈ H π(χ(α))x = dαx for all α ∈ ∪Fn}; (ii) If y in H belongs to Orb(x, bGΣ), then 1 lim n→∞ Fnw X α∈Fn dαπ(χ(α))(x − y) = 0. The article is organized as follows. In section 1, we collect some basic facts about compact quantum groups. In section 2, we prove Theorem 3.6, which uses the concept of the subgroup fixing a sequence of finite subsets in a discrete group. In the remaining sections, we turn to quantum groups. In section 3, we prove Theorem 4.1 which says that in a discrete quantum group, the subset fixing a sequence of finite subsets is a quantum subgroup. Then in section 4, we define the orbit of a vector in a Hilbert space under an action of discrete quantum group and lay down some basic properties. Then we prove a generalized mean ergodic Theorem 4.5. 2. Preliminaries 1A discrete quantum group is the dual of a compact quantum group, hence we state the result in terms of compact quantum groups. THE QUANTUM GROUP FIXING A SEQUENCE OF FINITE SUBSETS 3 2.1. Conventions. Within this paper, we use B(H, K) to denote the space of bounded linear operators from a Hilbert space H to another Hilbert space K, and B(H) stands for B(H, H). A net {Tλ} ⊆ B(H) converges to T ∈ B(H) under strong operator topology (SOT) if Tλx → T x for every x ∈ H, and {Tλ} converges to T ∈ B(H) under weak operator topology (WOT) if hTλx, yi → hT x, yi for all x, y ∈ H. The notation A ⊗ B always means the minimal tensor product of two C ∗-algebras A and B. For a state ϕ on a unital C ∗-algebra A, we use L2(A, ϕ) to denote the Hilbert space of GNS representation of A with respect to ϕ. The image of an a ∈ A in L2(A, ϕ) is denoted by a. In this paper all C ∗-algebras are assumed to be unital and separable. 2.2. Some Facts about Compact Quantum Groups. In this paper, we consider a discrete quantum group, which can be thought of as the dual of a compact quantum group. Compact quantum groups are noncommutative analogues of compact groups [Wor87, BS93, Wor98]. Definition 2.1. A compact quantum group is a pair (A, ∆) consisting of a unital C ∗-algebra A and a unital ∗-homomorphism ∆ : A → A ⊗ A such that (1) (id ⊗ ∆)∆ = (∆ ⊗ id)∆. (2) ∆(A)(1 ⊗ A) and ∆(A)(A ⊗ 1) are dense in A ⊗ A. The ∗-homomorphism ∆ is called the coproduct of G. One may think of A as C(G), the C ∗-algebra of continuous functions on a compact quantum space G with a quantum group structure. In the rest of the paper we write a compact quantum group (A, ∆) as G. There exists a unique state h on A such that (h ⊗ id)∆(a) = (id ⊗ h)∆(a) = h(a)1A for all a in A. The state h is called the Haar measure of G. Throughout this paper, we use h to denote it. For a compact quantum group G, there is a unique dense unital ∗-subalgebra A of A such that 4 HUICHI HUANG (1) ∆ maps from A to A ⊙ A (algebraic tensor product). (2) There exists a unique multiplicative linear functional ε : A → C and a linear map κ : A → A such that (ε ⊗ id)∆(a) = (id ⊗ ε)∆(a) = a and m(κ⊗id)∆(a) = m(id ⊗κ)∆(a) = ε(a)1 for all a ∈ A, where m : A ⊙ A → A is the multiplication map. The functional ε is called counit and κ the coinverse of C(G). Note that ε is only densely defined and not necessarily bounded. If ε is bounded and h is faithful (h(a∗a) = 0 implies a = 0), then G is called coamenable [BMT01]. Examples of coamenable compact quantum groups include C(G) for a compact group G and C ∗(Γ) for a discrete amenable group Γ. A nondegenerate (unitary) representation U of a compact quantum group G is an invertible (unitary) element in M(K(H) ⊗ A) for some Hilbert space H satisfying that U12U13 = (id ⊗ ∆)U. Here K(H) is the C ∗-algebra of compact operators on H and M(K(H) ⊗ A) is the multiplier C ∗-algebra of K(H) ⊗ A. We write U12 and U13 respectively for the images of U by two maps from M(K(H)⊗ A) to M(K(H) ⊗ A ⊗ A) where the first one is obtained by extending the map x 7→ x⊗1 from K(H)⊗A to K(H)⊗A⊗A, and the second one is obtained by composing this map with the flip on the last two factors. The Hilbert space H is called the carrier Hilbert space of U. From now on, we always assume representations are nondegenerate. If the carrier Hilbert space H is of finite dimension, then U is called a finite dimensional representation of G. For two representations U1 and U2 with the carrier Hilbert spaces H1 and H2 re- spectively, the set of intertwiners between U1 and U2, Mor(U1, U2), is defined by Mor(U1, U2) = {T ∈ B(H1, H2)(T ⊗ 1)U1 = U2(T ⊗ 1)}. Two representations U1 and U2 are equivalent if there exists a bijection T in Mor(U1, U2). A representation U is called irreducible if Mor(U, U) ∼= C. Moreover, we have the following well-established facts about representations of com- pact quantum groups: (1) Every finite dimensional representation is equivalent to a unitary represen- tation. (2) Every irreducible representation is finite dimensional. Let bG be the set of equivalence classes of irreducible unitary representations of G. For every γ ∈ bG, let U γ ∈ γ be unitary and Hγ be its carrier Hilbert space with dimension dγ. After fixing an orthonormal basis of Hγ, we can write U γ as THE QUANTUM GROUP FIXING A SEQUENCE OF FINITE SUBSETS 5 (uγ ij)1≤i,j≤dγ with uγ ij ∈ A (uγ ij's are called the matrix entries of γ), and ∆(uγ ij) = dγX k=1 uγ ik ⊗ uγ kj for all 1 ≤ i, j ≤ dγ. The matrix U γ is still an irreducible representation (not necessarily unitary) with the carrier Hilbert space ¯Hγ. It is called the conjugate representation of U γ and the equivalence class of U γ is denoted by ¯γ. Given two finite dimensional representations α, β of G, fix orthonormal basises for α and β and write α, β as U α, U β in matrix forms respectively. Define the direct sum, denoted by α + β as an equivalence class of unitary representations of dimension dα+dβ given by (cid:0) U α 0 0 U β (cid:1), and the tensor product, denoted by αβ, is an equivalence class of unitary representations of dimension dαdβ whose matrix form is given by U αβ = U α 13U β 23. The character χ(α) of a finite dimensional representation α is given by χ(α) = dαX i=1 uα ii. Note that χ(α) is independent of choices of representatives of α. Also kχ(α)k ≤ dα since Pdα ik)∗ = 1 for every 1 ≤ i ≤ dα. Moreover k=1 uα ik(uα χ(α + β) = χ(α) + χ(β), χ(αβ) = χ(α)χ(β) and χ(α)∗ = χ( ¯α) for finite dimensional representations α, β. Every representation of a compact quantum group is a direct sum of irreducible representations. For two finite dimensional representations α and β, denote the number of copies of γ ∈ bG in the decomposition of αβ into sum of irreducible representations by N γ α,β. Hence αβ = X N γ α,βγ. γ∈ bG We have the Frobenius reciprocity law [Wor87, Proposition 3.4.] [Kye08, Example 2.3]. N γ α,β = N α γ, ¯β = N β ¯α,γ, for all α, β, γ ∈ bG. Within the paper, we assume that A = C(G) is a separable C ∗-algebra, which amounts to say, bG is countable. 6 HUICHI HUANG Definition 2.2. boundary of F relative to S, denoted by ∂S(F ), is defined by [Kye08, Definition 3.2] Given two finite subsets S, F of bG, the ∂S(F ) ={α ∈ F ∃ γ ∈ S, β /∈ F, such that N β α,γ > 0 } ∪ {α /∈ F ∃ γ ∈ S, β ∈ F, such that N β α,γ > 0 }. We denote ∂{γ∈ bGγ is contained in α}(F ) by ∂α(F ) for a finite dimensional representation α. We say γ in bG fixes a sequence Σ = {Fn} of finite subsets in bG if lim n→∞ ∂γ(Fn)w Fnw = 0. The weighted cardinality F w of a finite subset F of bG is given by F w = X d2 α. α∈F 3. The Case for Groups Let G be a countable discrete group and Σ = {Fn}∞ of G. n=1 is a sequence of finite subsets Definition 3.1. We say that an element g in G fixes Σ if lim n→∞ gFn∆Fn Fn = 0. Denote by GΣ the set of elements in G fixed by Σ. It's routine to check that GΣ is a subgroup of G. Examples 3.2. [Examples of GΣ] (1) A group G is amenable iff GΣ = G and Σ is a Følner sequence. (2) In Z, let Σ = {Fn}∞ n=1 with Fn = {1, 3, · · · , 2n + 1}. Then ZΣ = 2Z. THE QUANTUM GROUP FIXING A SEQUENCE OF FINITE SUBSETS 7 3.1. Arithmetic Progressions in Discrete Groups. In 1977, H. Furstenberg found that Szemer´edi's theorem is equivalent to a multiple recurrence theorem, which he called "ergodic Szemer´edi theorem". See [Fur77, Thm. 1.4] and [FKO82, Thm. II]. Via proving his ergodic Szemer´edi theorem, Furstenberg gave an ergodic theoretic proof of Szemer´edi's theorem. This is Furstenberg correspondence principle which opens a door for applications of ergodic theory to combinatorial number theory. This is also the main ingredient of the paper. Theorem 3.3. [Ergodic Szemer´edi theorem] Let (X, B, ν, T ) be a dynamical system consisting of a probability measure space (X, B, ν) and a measure-preserving transformation T : X → X. For any positive integer k, there exists n ∈ Z such that µ( k\ j=1 T −jnA) > 0 whenever µ(A) > 0. Along this idea, it appear various generalizations of Szemer´edi's theorem to Zd [FK78, FK91, BL96]. Actually via ergodic Szemer´edi theorem, Furstenberg had proved a theorem stronger than Szemer´edi's theorem [FKO82, Thm. I]. Theorem 3.4. [Furstenberg's version of Szemer´edi's theorem] A set of positive integers with positive upper Banach density contains arbitrarily long arithmetic progressions. A subset Λ of positive integers has positive upper density if lim sup 0 and has positive upper Banach density if lim sup sequence of intervals {[an, bn)} with bn − an → ∞. n→∞ n→∞ Λ ∩ [an, bn) bn − an Λ ∩ [1, n] > n > 0 for a Suppose T is a homeomorphism on a compact metrizable space X. A Borel prob- ability measure ν on X is called T -invariant if ν(T −1A) = ν(A) for every Borel subset A of X and every s in Γ. It's well-known that ν is T -invariant if and only if ν(T −1f ) = ν(f ) for every f in C(X). Here C(X) stands for the set of complex-valued continuous functions on X, ν(f ) = RX f (y) dν(y) and T −1f (x) = f (T (x)) for every x in X. 8 HUICHI HUANG Now we start to prove the first main Theorem, Theorem 3.6, which relies on the multiple recurrence theorem due to Furstenberg. See [Fur77, Thm. 1.4] and [FKO82, Thm. II]. Theorem 3.5. [Ergodic Szemer´edi theorem] Let (X, B, ν, T ) be a dynamical system consisting of a probability measure space (X, B, ν) and a measure-preserving transformation T : X → X. For any positive integer k, there exists n ∈ Z such that µ( k\ j=1 T −jnA) > 0 whenever µ(A) > 0. Theorem 3.6. Let Σ = {Fn} be a sequence of finite subsets in a discrete group Γ and suppose b in Γ fixes Σ from right (left). If a subset Λ of Γ has positive upper density with respect to Σ, then for any positive integer k, there exist n > 0 and a ∈ Γ such that {bjna}k−1 j=0) is contained in Λ . j=0 ({abjn}k−1 Proof. We only need to give a proof for the case that b in Γ fixes Σ from right. If b in Γ fixes Σ from left and Λ is a subset of Γ with positive upper density with n } from left and Λ−1 is a subset with respect to Σ = {Fn}, then b−1 fixes Σ−1 = {F −1 positive upper density with respect to Σ−1 . The existence of {bjna}k j=1 in Λ−1 gives the existence of {abjn}k j=1 in Λ. Let Γ act on {0, 1}Γ by shift, that is, s · x(t) = x(ts) for all s, t in Γ and x in {0, 1}Γ. Define A0 := {x ∈ {0, 1}Γx(eΓ) = 1} and let ω = 1Λ be the characteristic function of Λ. Suppose b in Γ fixes Σ from left. Then for every positive integer k, one has ∃ a ∈ Γ such that {bja}k j=1 ⊆ Λ; ⇐⇒ ∃ a ∈ Γ such that ω(bja) = 1 for all 1 ≤ j ≤ k; ⇐⇒ ∃ a ∈ Γ such that bja · ω(eΓ) = 1 for all 1 ≤ j ≤ k; ⇐⇒ ∃ a ∈ Γ such that {bja · ω}k j=1 ⊆ A0. Let X = Γ · ω be the closure of the orbit of ω in {0, 1}Γ. Let A = A0 ∩ X, which is a closed subset of X. THE QUANTUM GROUP FIXING A SEQUENCE OF FINITE SUBSETS 9 It follows that ∃ a ∈ Γ such that {bja · ω}k j=1 ⊆ A0; ⇐⇒ ∃ a ∈ Γ such that ω ∈ k\ (bja)−1 · A0; j=1 ⇐⇒ ∃ a ∈ Γ such that a · ω ∈ (A0 is open=⇒ Tk j=1 b−j · A0 is open.) k\ j=1 b−j · A0; ⇐⇒ k\ j=1 b−j · A0 ∩ Γ · ω 6= ∅; ⇐⇒ k\ j=1 b−j · A0 ∩ Γ · ω 6= ∅; ⇐⇒ k\ j=1 b−jA 6= ∅. Next we are going to construct a b-invariant Borel probability measure µ on X such that µ(A) > 0. By Theorem 3.5, this will complete the proof. Let δs·ω be the Dirac measure at the point s · ω for s in Γ, and this is a Borel probability measure on X. Fn Pt∈Fn δt·ω. Let µ be a weak-∗ limit of µn. Without loss of generality, Define µn = 1 let µ = lim µn. n→∞ Then the following two claims hold. (1) µ is b-invariant. (2) µ(A) > 0. Proof. [Proof of the first claim] For every continuous function f on X, one has µ(b−1 · f ) = lim n→∞ µn(b−1 · f ) = lim n→∞ 1 Fn X t∈Fn 1 b−1 · f (t · ω) Fn X t∈bFn f (t · ω) (b fixes Σ from right.) = lim n→∞ = lim n→∞ 1 Fn X t∈Fn f (bt · ω) = lim n→∞ 1 Fn X t∈Fn f (t · ω) = µ(f ). Hence µ is b-invariant. Proof. [Proof of the second claim] (cid:3) 10 Note that HUICHI HUANG lim sup n→∞ µn(A) = lim sup n→∞ = lim sup n→∞ = lim sup n→∞ = lim sup n→∞ {t ∈ Fn t · ω ∈ A} Fn {t ∈ Fn t · ω ∈ A0} Fn {t ∈ Fn t · ω(eΓ) = 1} Fn {t ∈ Fn ω(t) = 1} Fn = lim sup n→∞ Fn ∩ Λ Fn = DΣ(Λ) > 0. Since A is a closed subset of X, we have µ(A) ≥ lim sup n→∞ µn(A) > 0 [Wal82, Sec. 6.1, Remarks (3)]. Applying Theorem 3.5 to the dynamical system (X, µ, b) gives the proof. (cid:3) (cid:3) Remark 3.7. A set Λ has positive upper density with respect to Σ iff it has positive density with respect to a subsequence of Σ, hence without loss of generality, we can just assume that Λ has positive density with respect to a sequence. We implicitly use this fact in the proof of Theorem 3.6. A sequence {Fn}∞ left (right) Følner sequence if n=1 of finite subsets in a countable discrete group Γ is called a lim n→∞ sFn∆Fn Fn = 0 ( lim n→∞ Fns∆Fn Fn = 0) for every s in Γ. A group Γ having a Følner sequence is called amenable. Remark 3.8. It might happen that except the neutral element, no other element in Γ fixes Σ for a sequence Σ. For instance no integer except 0 fixes Σ = {Fn}∞ for Fn = {2, 4, · · · , 2n}. So choices of Σ decide the elements fixed by it. n=1 When Γ is amenable, and one can choose Σ to be a left (right) Følner sequence of Γ. Then every b in Γ fixes Σ from right (left). So Theorem 3.6 gives the following application. THE QUANTUM GROUP FIXING A SEQUENCE OF FINITE SUBSETS 11 Corollary 3.9. [Arithmetic progressions in amenable groups] In a subset Λ of an amenable group Γ with positive upper density with respect to a left (right) Følner sequence, for every positive integer k and every b in Γ, there exist a in Γ and a positive integer n such that {bjna}k j=1) is contained in Λ. j=1 ({abjn}k Moreover if Γ contains Z as a subgroup, then a subset of Γ with positive upper den- sity with respect to a left (right) Følner sequence contains arbitrarily long left (right) arithmetic progressions. Proof. If Σ is a left Følner sequence in an amenable group Γ, then every b in Γ fixes Σ from right. By Theorem 3.6, the first statement holds. Since Z is a subgroup of Γ, there exists an element b of infinite order in Γ. By Theorem 3.6, for every positive integer k, there exist a in Γ and a positive integer n such that Λ contains {bjna}k j=1). Since b is of infinite order, the set {bjna}k j=1) has k distinct elements. Hence it is a left (right) arithmetic progression of length k. (cid:3) j=1 ({abjn}k j=1 ({abjn}k 4. The Case for Quantum Groups 4.1. The Quantum Subgroup Fixing A Sequence of Finite Subsets. Denote the elements in bG fixed by a sequence of finite subsets Σ = {Fn}∞ n=1 by bGΣ. In this subsection we are going to prove that bGΣ is a quantum subgroup of bG. This amounts to say that the C ∗-subalgebra C ∗(bGΣ) generated by the matrix entries of all elements in bGΣ is a compact quotient group of G, that is, C ∗(bGΣ) is a compact Theorem 4.1. Given a sequence Σ of finite subsets of bGΣ for a compact quantum group G, the C ∗-algebra C ∗(bGΣ) is a compact quantum group. quantum group. Proof. We are going to verify the following: (1) The trivial representation γ0 is in bGΣ. (2) If γ is in bGΣ, then its conjugate ¯γ is also in bGΣ. (3) If γ1, γ2 are in bGΣ, then every γ ∈ bG contained in γ1γ2 is also in bGΣ. Firstly for every finite subset F of bGΣ, the set ∂γ0(F ) is always empty. So γ0 is in bGΣ. To prove (2) and (3), we need a lemma. 12 HUICHI HUANG Lemma 4.2. For α, β, γ in bGΣ, we have X β∈ bGΣ, N β α,γ >0 β ≤ d2 d2 γ and X αd2 d2 α ≤ α∈ bGΣ, N β α,γ >0 βd2 d2 γ. Proof. Note that dαdγ = X α,γdβ = X N β α,γdβ ≥ X N β dβ. β∈ bG β∈ bG, N β α,γ >0 β∈ bG, N β α,γ>0 So d2 γ ≥ (Pβ∈ bG, N β αd2 α,γ >0 dβ)2 ≥ Pβ∈ bGΣ, N β α,γ >0 d2 β. Furthermore by Frobenius reciprocity law, we have N β β,¯γdα = X N α dβdγ = dβd¯γ = X N α β,¯γdα α,γ = N α β,¯γ. Hence α∈ bG β∈ bG, N α β,¯γ>0 = X α,γdα ≥ X N β β∈ bG, N β α,γ >0 So d2 γ ≥ (Pα∈ bG, N β βd2 α,γ >0 dα)2 ≥ Pα∈ bGΣ, N β α,γ>0 α∈ bG, N β α,γ >0 d2 α. dα. (cid:3) We prove (2) via proving that lim n→∞ ∂¯γ(Fn)w Fnw = 0 provided lim n→∞ ∂γ(Fn)w Fnw = 0. By definition ∂¯γ(Fn) = {α ∈ Fn N β α,¯γ > 0 for some β /∈ Fn} ∪ {α /∈ Fn N β α,¯γ > 0 for some β ∈ Fn} (F robenius reciprocity law) = {α ∈ Fn N α β,γ > 0 for some β /∈ Fn} ∪ {α /∈ Fn N α β,γ > 0 for some β ∈ Fn}. On the other hand ∂γ(Fn) = {α ∈ Fn N β α,γ > 0 for some β /∈ Fn} ∪ {α /∈ Fn N β α,γ > 0 for some β ∈ Fn}. We can define a map ϕ from ∂¯γ(Fn) to ∂γ(Fn) by the following: when α ∈ ∂¯γ(Fn) ∩ Fn, the image ϕ(α) is given by some β in F c α ∈ ∂¯γ(Fn) ∩ F c N β n with N β n, the image ϕ(α) is given by some β in Fn with N β β,γ, we have that ϕ(α) is in ∂γ(Fn). By Lemma 4.2, we have X α,¯γ > 0; when α,¯γ > 0. From d2 α ≤ α,¯γ = N α d2 βd2 γ. Hence ϕ(α)=β {α ∈ Fn N α β,γ > 0 for some β /∈ Fn}w ≤ d2 γ{β /∈ Fn N β α,γ > 0 for some α ∈ Fn}w THE QUANTUM GROUP FIXING A SEQUENCE OF FINITE SUBSETS 13 and {α /∈ Fn N α β,γ > 0 for some β ∈ Fn}w ≤ d2 γ{β ∈ Fn N β α,γ > 0 for some α /∈ Fn}w. Therefore ∂¯γ(Fn)w ≤ d2 γ∂γ(Fn)w and (2) follows immediately. Now we proceed to the proof of (3). Suppose γ1 and γ2 are in bGΣ. We are going to prove that lim n→∞ ∂γ1γ2(Fn)w Fnw = 0. By definition ∂γ1γ2(Fn) = {α ∈ Fn N β α,γ1γ2 > 0 for some β /∈ Fn}∪{α /∈ Fn N β α,γ1γ2 > 0 for some β ∈ Fn}. Since α(γ1γ2) = (αγ1)γ2, we have N β α,γ1γ2 = N β αγ1,γ2 for all α, β in bG. Hence {α ∈ Fn N β Note that N β α,γ1γ2 > 0 for some β /∈ Fn} = {α ∈ Fn N β αγ1,γ2 = Pγ∈ bG N γ Suppose α is in Fn with N β that N γ α,γ1N β α,γ1γ2 > 0 for some β /∈ Fn. Then there exists γ ∈ bG such αγ1,γ2 > 0 for some β /∈ Fn}. γ,γ2 > 0 for some β /∈ Fn. α,γ1 > 0 and N β γ,γ2. If γ is not in Fn, then α is in {η ∈ Fn N ζ η,γ1 > 0 for some ζ /∈ Fn} ⊆ ∂γ1(Fn). If γ is in Fn, then γ is in {η ∈ Fn N ζ can define a map ϕ : {α ∈ Fn N β 0 for some ζ /∈ Fn} by ϕ(α) = γ for some γ with N ζ that X d2 α ≤ d2 α,γ1γ2 > 0 for some β /∈ Fn} → {η ∈ Fn N ζ η,γ2 > 0 for some ζ /∈ Fn} ⊆ ∂γ2(Fn). We η,γ2 > γ,γ2 > 0. By Lemma 4.2, we know γd2 γ1. α∈Fn, ϕ(α)=γ Therefore {α ∈ Fn N β ≤ d2 α,γ1γ2 > 0 for some β /∈ Fn}w γ1({η ∈ Fn N ζ η,γ1 > 0 for some ζ /∈ Fn}w + {η ∈ Fn N ζ η,γ2 > 0 for some ζ /∈ Fn}w). Moreover {α /∈ Fn N β α,γ1γ2 > 0 for some β ∈ Fn} = {α /∈ Fn N β αγ1,γ2 > 0 for some β ∈ Fn}. As before if α is not in Fn with N β such that N γ α,γ1 > 0 and N β γ,γ2 > 0 for some β ∈ Fn. α,γ1γ2 > 0 for some β ∈ Fn, then there exists γ ∈ bG 14 HUICHI HUANG If γ is not in Fn, then γ is in {η /∈ Fn N ζ can define a map ψ : {α /∈ Fn N β 0 for some ζ ∈ Fn} by ψ(α) = γ with N ζ η,γ2 > 0 for some ζ ∈ Fn} ⊆ ∂γ2(Fn). We η,γ2 > d2 α ≤ α,γ1γ2 > 0 for some β ∈ Fn} → {η /∈ Fn N ζ γ,γ2 > 0. By Lemma 4.2, we have X γd2 d2 γ2. If γ is in Fn, then α is in {η /∈ Fn N ζ η,γ1 > 0 for some ζ ∈ Fn} ⊆ ∂γ1(Fn). Hence α /∈Fn, ψ(α)=γ {α /∈ Fn N β ≤ d2 α,γ1γ2 > 0 for some β ∈ Fn}w γ2({η /∈ Fn N ζ η,γ1 > 0 for some ζ ∈ Fn}w + {η /∈ Fn N ζ η,γ2 > 0 for some ζ ∈ Fn}w). Therefore ∂γ1γ2(Fn)w ≤ max {d2 γ1, d2 γ2}(∂γ1(Fn)w + ∂γ2(Fn)w). This proves (3). (cid:3) Give a sequence Σ = {Fn} of finite subsets in bG, we say a subset Λ of bG has positive upper density with respect to Σ if lim sup Λ ∩ Fnw > 0. n→∞ Fnw For α, β ∈ bG, we say a subset Λ of bG contains αjβ if every γ contained in αjβ is also in Λ. Motivated by Theorem 3.6, we give the following conjecture. Conjecture 4.1. Suppose a subset Λ of bG has positive upper density with respect to Σ. Then for every k > 0, there exists α in bGΣ, β ∈ bG and n > 0 such that Λ contains αjnβ for all 0 ≤ j ≤ k − 1. 4.2. Discrete Quantum Group Orbits and a Mean Ergodic Theorem for Discrete Quantum Groups. versely for a discrete quantum group Γ, there is a compact quantum group G such Given a compact quantum group G, the dual bG is a discrete quantum group. Con- that Γ = bG. With this in mind, when we talk about a discrete quantum group, we mean the dual of a compact quantum group [PW90, vanD96, MvD98, KV99, KV00, SW07]. Consider an action of a discrete quantum group bG on a Hilbert space H, that is, a representation π of C(G) on H. We come up a definition of discrete quantum orbits in a Hilbert space. Various definitions of compact quantum group orbits already appear in [Sai09, Hua16-1, DKSS16]. THE QUANTUM GROUP FIXING A SEQUENCE OF FINITE SUBSETS 15 Definition 4.3. For x in H, we define the orbit of x as {y ∈ H ∃ α, β ∈ bG such that π(χ(α)) dα x = π(χ(β)) dβ y}, and denote it by Orb(x, bG). The following are true. Proposition 4.4. [Properties of orbits] (1) x ∈ Orb(x, bG); (2) If y ∈ Orb(x, bG), then x ∈ Orb(y, bG); (3) { π(χ(α)) x}α∈ bG ⊆ Orb(x, bG). dα Proof. (1) and (2) are immediate from the definition of orbit. Note that π(χ(α)) trivial representation of G. This proves (3). x = π(χ(γ0)) ( π(χ(α)) dγ0 dα dα x) for every α in bG. Here γ0 = 1 stands for the (cid:3) We say a vector x in H is fixed by Σ (a sequence of finite subsets in bG) if π(χ(α))x = dαx for every α in ∪Fn. It's easy to see that the set of vectors fixed by Σ is a subspace of H. Denote it by HΣ. In this subsection we prove a mean ergodic theorem for discrete quantum groups, which is a generalization of the mean ergodic theorem for amenable discrete quantum groups [Hua16-2]. Let Σ = {Fn} be a sequence of finite subsets in bG. For a representation π : A = C(G) → B(H), consider the sequence of bounded linear operators { 1 on H. Under weak operator topology, the unit ball of B(H) is compact and contains Fnw Pα∈Fn dαπ(χ(α))}∞ { 1 n=1. Hence there exist limit points for { 1 Fnw Pα∈Fn Fnw Pα∈Fn dαπ(χ(α))}∞ n=1. dαπ(χ(α))}∞ n=1 Theorem 4.5. (i) Suppose T is a limit of { 1 n=1 in B(H). Then T P = P T = P where P is the orthogonal projection from H onto HΣ := {x ∈ H π(χ(α))x = dαx for all α ∈ ∪Fn}; Fnw Pα∈Fn dαπ(χ(α))}∞ 16 HUICHI HUANG (ii) If y in H belongs to Orb(x, bGΣ), then lim n→∞ Fnw dαπ(χ(α))(x − y) = 0. 1 X α∈Fn Proof. (i) For every x in HΣ and every n, we have Hence T P = P . Next we prove that T ∗P = P and this will finish the proof. 1 Fnw Pα∈Fn dαπ(χ(α))x = x. We need a lemma. Lemma 4.6. If π(χ(α))x = dαx, then π(χ( ¯α))x = d ¯αx. Proof. Without loss of generality, we may assume that x is a unit vector in H. Then 0 ≤ kπ(χ( ¯α))x − d ¯αxk2 = hπ(χ( ¯α))x, π(χ( ¯α))xi − hπ(χ( ¯α))x, d ¯αxi − hd ¯αx, π(χ( ¯α))xi + hd ¯αx, d ¯αxi (π(χ( ¯α))x = d ¯αx implies that hπ(χ( ¯α))x, d ¯αxi = hd ¯αx, π(χ( ¯α))xi = d2 α = d2 ¯α.) (kχ( ¯α)k ≤ d ¯α.) = hπ(χ( ¯α))x, π(χ( ¯α))xi − d2 ¯α ≤ 0. Hence π(χ( ¯α))x = d ¯αx. (cid:3) Note that ( have Fnw Pα∈Fn dαπ(χ(α)))∗ = 1 Fnw Pα∈Fn dαπ(χ( ¯α)). By Lemma 4.6, we Fnw Pα∈Fn dαπ(χ( ¯α))x = x for every x in HΣ. This implies that T ∗P = P . 1 1 (ii) We first prove that (4.1) lim n→∞ 1 Fnw X α∈Fn dαπ(χ(α))(π(χ(γ))y − dγy) = 0 Thm. 3.1] for all y in H and γ in bGΣ. The proof is a simplified version of the proof of [Hua16-2, For every y ∈ H and γ ∈ bG, we have X α∈Fn dαπ(χ(α))(π(χ(γ))y − dγy) (χ(α)χ(γ) = χ(αγ) and dγ = d¯γ.) = X dαπ(χ(αγ))y − X α∈Fn α∈Fn dαd¯γπ(χ(α))y THE QUANTUM GROUP FIXING A SEQUENCE OF FINITE SUBSETS 17 = X dαX α,γπ(χ(β))y − X N β α∈Fn β∈ bG α∈Fn X β∈ bG N β α,¯γdβπ(χ(α))y. X dαX α,γπ(χ(β))y = X N β dα(X + X )N β α,γπ(χ(β))y. α∈Fn β∈ bG α∈Fn β∈Fn β /∈Fn α∈Fn β∈ bG X X = X = X = X α∈Fn α∈Fn N β α,¯γdβπ(χ(α))y β∈Fn X X X β∈Fn α,¯γdβπ(χ(α))y + X N β β,γdβπ(χ(α))y + X N α α,γdαπ(χ(β))y + X N β α∈Fn α∈Fn β /∈Fn X X X β /∈Fn N β α,¯γdβπ(χ(α))y N β α,¯γdβπ(χ(α))y N β α,¯γdβπ(χ(α))y. α∈Fn β∈Fn α∈Fn β /∈Fn dαπ(χ(α))(π(χ(γ))y − dγy) X α,γπ(χ(β))y − X dαN β α∈Fn β /∈Fn k 1 Fnw 1 Fnw X X α∈Fn β /∈Fn N β α,¯γdβπ(χ(α))y. Note that Moreover (N β α,¯γ = N α β,γ.) X α∈Fn 1 Fnw Hence 1 Fnw = Note that (kχ(β)k ≤ dβ.) ≤ ≤ α∈Fn X X β /∈Fn X X X α∈Fn β /∈Fn α∈∂γ Fn 1 1 Fnw Fnw dαN β α,γπ(χ(β))yk dαN β α,γdβkyk d2 αdγkyk → 0 as n → ∞ since γ is in bGΣ. Also k 1 Fnw X X X X β /∈Fn α∈Fn α∈Fn β /∈Fn ≤ 1 Fnw N β α,¯γdβπ(χ(α))yk N β α,¯γdβdαkyk 18 HUICHI HUANG ≤ 1 Fnw X α∈∂¯γ Fn d2 αd¯γkyk → 0 as n → ∞ since Theorem 4.1 guarantees that when γ is in bGΣ, so is ¯γ. Therefore lim n→∞ 1 Fnw X α∈Fn dαπ(χ(α))(π(χ(γ))y − dγy) = 0. If y is in Orb(x, bGΣ), then there exist β and γ in bGΣ such that π(χ(β)) dβ From Equation 4.1, we have x = π(χ(γ)) dγ y. lim n→∞ 1 Fnw X α∈Fn dαπ(χ(α))(x − y) ( π(χ(β)) dβ x = π(χ(γ)) dγ y) 1 = lim n→∞ Fnw X α∈Fn dαπ(χ(α))[(x − π(χ(β)) dβ x) + ( π(χ(γ)) dγ y − y)] = 0. Remark 4.7. Theorem 4.5(ii) says that if x and y are in the same orbit of bGΣ, then averages of x and y along Σ coincide. (cid:3) Acknowledgements I thank Hanfeng Li for his illuminating comments. I got familiar with Furstenberg correspondence principle in a 2012 graduate student seminar in SUNY at Buffalo organized by Bingbing Liang, Yongle Jiang, Yongxiao Lin and myself. I thank them for their kind feedback. The latest version of the paper was carried out during a visit to the Research Center for Operator Algebras in East China Normal University in April 2017. I thank Huaxin Lin for his hospitality and Qin Wang for helpful discussions. References [BBF10] M. Beiglbock, V. Bergelson and A. Fish. Sunset phenomenon in countable amenable groups. Adv. Math. 223 (2010), no. 2, 416 -- 432. [BL96] V. Bergelson and A. Leibman. Polynomial extensions of van der Waerden's and Szemer´edi's theorems. J. Amer. Math. Soc. 9 (1996), 725 -- 753. [BMT01] E. B´edos, G. J. Murphy and L. Tuset. Co-amenability of compact quantum groups. J. Geom. Phys. 40 (2001), no. 2, 130 -- 153. THE QUANTUM GROUP FIXING A SEQUENCE OF FINITE SUBSETS 19 [BS93] S. Baaj and G. Skandalis. Unitaires multiplicatifs et dualit´e pour les produits crois´es de C ∗-alg`ebres. Ann. Sci. ´Ecole Norm. Sup. (4) 26 (1993), no. 4, 425 -- 488. [DKSS16] K. De Commer, P. Kasprzak, A.Skalski and P. M.Soltan. Quantum actions on discrete quantum spaces and a generalization of Clifford's theory of representations. arXiv:1611.10341v1. [ET36] P. Erdos and P. Tur´an. On some sequences of integers. J. Lond. Math. Soc. 11 (1936), no. 4, 261 -- 264. [Fur77] H. Furstenberg. Ergodic behavior of diagonal measures and a theorem of Szemer´edi on arithmetic progressions. J. Anal. Math. 31, (1977), 204 -- 256. [Fur81] H. Furstenberg. Recurrence in Ergodic theory and Combinatorial Number Theory. Princeton University Press, Princeton NJ, 1981. [FK78] H. Furstenberg and Y. Katznelson. An ergodic Szemer´edi theorem for commuting transformations. J. Anal. Math. 34 (1978), 275 -- 291. [FK91] H. Furstenberg and Y. Katznelson. A density version of the Hales-Jewett theorem. J. Anal. Math. 57 (1991), 64 -- 119. [FKO82] H. Furstenberg, Y. Katznelson and D. S. Ornstein. The ergodic theoretical proof of Szemer´edi's theorem. Bull. Amer. Math. Soc. 7 (1982), no. 3, 527 -- 552. [GT08] B. Green and T. Tao. The primes contain arbitrarily long arithmetic progressions. Ann. of Math. 167 (2008), no. 2, 481 -- 547. [Hua16-1] H. Huang. Invariant subsets under compact quantum group actions. J. Non- commut. Geom. 10 (2) (2016), 447 -- 469. [Hua16-2] H. Huang. Mean ergodic theorem for amenable discrete quantum groups and a Wiener-type theorem for compact metrizable groups. Anal. PDE 9 (2016), no. 4, 893 -- 906. [Kye08] D. Kyed. L2-Betti numbers of coamenable quantum groups. Munster J. Math. 1 (2008), 143-179. [KV99] J. Kustermans and S. Vaes. A simple definition for locally compact quantum groups. C. R. Acad. Sci. Paris S´er. I Math. 328 (1999), no. 10, 871 -- 876. [KV00] J. Kustermans and S. Vaes. Locally compact quantum groups. Ann. Sci. ´Ecole Norm. Sup. (4) 33 (2000), no. 6, 837 -- 934. [MvD98] M. Maes and A. Van Daele. Notes on compact quantum groups. Nieuw Arch. Wisk. (4) 16 (1998), no. 1-2, 73 -- 112. [PW90] P. Podle´s, P and S. L. Woronowicz. Quantum deformation of Lorentz group. Comm. Math. Phys. 130 (1990), no. 2, 381 -- 431. 20 HUICHI HUANG [Sai09] J. N. Sain. Berezin quantization from ergodic actions of compact quantum groups, and quantum Gromov-Hausdorff distance. Thesis (Ph.D.)University of California, Berkeley. 2009. [Sze69] E. Szemer´edi. On sets of integers containing no four elements in arithmetic pro- gression. Acta Math. Acad. Sci. Hungar. 20 (1969), 89 -- 104. [Sze75] E. Szemer´edi. On sets of integers containing no k elements in arithmetic progres- sion. Acta Arith. 27, (1975), 199 -- 245. [SW07] P. M. So ltan and S. L. Woronowicz. From multiplicative unitaries to quantum groups. II. J. Funct. Anal. 252 (2007), no. 1, 42-67. [Tao07] T. Tao. What is good mathematics? Bull. Amer. Math. Soc. 44 (2007), no. 4, 623 -- 634. [vanD96] A. Van Daele. Discrete quantum groups. J. Algebra 180 (1996), no. 2, 431 -- 444. [vand27] B. L. van der Waerden. Beweis einer Baudetschen Vermutung. Nieuw. Arch. Wisk. 15 (1927), 212 -- 216. [Wal82] P. Walters. An Introduction to Ergodic Theory. Graduate Texts in Mathematics, 79. Springer-Verlag, New York-Berlin, 1982. [Wor87] S. L. Woronowicz. Compact matrix pseudogroups. Comm. Math. Phys. 111 (1987), no. 4, 613 -- 665. [Wor98] S. L. Woronowicz. Compact quantum groups. Sym´etries Quantiques (Les Houches, 1995), 845 -- 884, North-Holland, Amsterdam, 1998. Huichi Huang, College of Mathematics and Statistics, Chongqing University, Chongqing, 401331, PR China E-mail address: [email protected]
1601.07456	1	1601	2016-01-27T17:28:57	An inequality in noncommutative $L_p$-spaces	[ "math.OA", "math.FA" ]	We prove that for any (trace-preserving) conditional expectation $\mathcal E$ on a noncommutative $L_p$ with $p>2$, $Id-\mathcal E$ is a contraction on the positive cone $L_p^+$.	math.OA	math	AN INEQUALITY IN NONCOMMUTATIVE Lp-SPACES ´ERIC RICARD Abstract. We prove that for any (trace-preserving) conditional expectation E on a noncommu- tative Lp with p > 2, Id − E is a contraction on the positive cone L+ p . It is plain that for positive real numbers a, b > 0 and p > 2, one has (a − b)(ap−1 − bp−1) > a − bp. 1. Introduction Integrating the above inequality on some measure space (Ω, µ) implies that for f, g ∈ Lp(Ω, µ)+, ZΩ(cid:0)f (x) − g(x)(cid:1)(cid:0)f p−1(x) − gp−1(x)(cid:1)dµ(x) > (cid:13)(cid:13)f − g(cid:13)(cid:13) p p . In [3], Mustapha Mokhtar-Kharroubi notices that this inequality may be used to get contractivity results on the positive cone of Lp(Ω, µ). This note originates from the question whether its non- commutative analogue remains true. We provide a proof in the next section. We hope that the techniques involved there may be useful for further studies. We end up by making explicit some results from [3] for noncommutative Lp-spaces. We refer the reader to [4] for the definitions of Lp-spaces associated to semifinite von Neumann algebras or more general ones. We also freely use basic results from [1]. 2. Results Let (M, τ ) be a semifinite von Neumann algebra. We denote by fp : M+ → M+, the pth-power map and by M++ the set of positive invertible elements. We will often refer to positivity of the trace for the fact that if a, b ∈ M+ ∩ L1(M), then τ (ab) > 0. Theorem 2.1. Let p > 2 and a, b ∈ Lp(M)+, then τ(cid:0)a − bp(cid:1) 6 τ(cid:0)(a − b)(ap−1 − bp−1)(cid:1). Proof. First, as for any sequence of (finite) projections pi going to 1 strongly in M and any x ∈ Lq(M) (1 6 q < ∞) kpixpi − xkq → 0, we may assume that M is finite. Next by replacing a and b by a + ε1 and a + ε1 for some ε > 0, we may also assume that [a, b] ⊂ M++ to avoid any unnecessary technical complication. We write a = b + δ. To prove the result, we distinguish according to the values of p. Case 1: p ∈ [2, 3] Case 1.a: a > b, i.e. δ > 0. As p − 1 = 1 + θ with θ ∈ [0, 1], we use the well known integral formula 6 1 0 2 n a J 7 2 ] . A O h t a m [ 1 v 6 5 4 7 0 . 1 0 6 1 : v i X r a (1) Hence s1+θ = cθ ZR+ τ(cid:0)δ(a1+θ − b1+θ)(cid:1) = cθ ZR+ tθs2 s + t dt t , s2 s + t = s − t + t2 s + t . tθτ(cid:16)δ(cid:0)δ + t2(b + δ + t)−1 − t2(b + t)−1(cid:1)(cid:17) dt t . Recall the identity (b + δ + t)−1 − (b + t)−1 = −(b + δ + t)−1δ(b + t)−1. Using positivity of the trace with δ(b + δ + t)−1δ 6 δ(δ + t)−1δ and (b + t)−1 6 t−1: τ(cid:16)δ(cid:0)δ + t2(b + δ + t)−1 − t2(b + t)−1(cid:1)(cid:17) > τ(cid:0)δ2 − tδ(δ + t)−1δ(cid:1) = τ(cid:0)δ3(δ + t)−1(cid:1). 2010 Mathematics Subject Classification: 46L51; 47A30. Key words: Noncommutative Lp-spaces. 1 2 ´E. RICARD Integrating, we get the desired inequality τ(cid:0)δ(a1+θ − b1+θ)(cid:1) > τ(cid:0)δ2+θ(cid:1). Case 1.b: δ arbitrary with decomposition δ+ − δ− into positive and negative parts. We reduce it to the previous case by introducing α = a + δ− = b + δ+, so that α > a, b. We have τ(cid:0)(a − b)(ap−1 − bp−1)(cid:1) = τ(cid:0)(a − α)(ap−1 − αp−1)(cid:1) + τ(cid:0)(a − α)(αp−1 − bp−1)(cid:1) +τ(cid:0)(α − b)(αp−1 − bp−1)(cid:1) + τ(cid:0)(α − b)(ap−1 − αp−1)(cid:1). The first and the third terms are bigger than τ(cid:0)δp −(cid:1) and τ(cid:0)δp to check that the two remaining terms are positive. We use again the integral formula (1) and δ+δ− = 0 and positivity of the trace +(cid:1) by Case 1.a. Hence it suffices τ(cid:0) − δ−(αp−1 − bp−1)(cid:1) = −cθ ZR+ = cθ ZR+ tθτ(cid:16)δ−(cid:0)δ+ + t2(b + δ+ + t)−1 − t2(b + t)−1(cid:1)(cid:17) dt tθt2τ(cid:16)δ−(cid:0)(b + t)−1 − (b + δ+ + t)−1(cid:1)(cid:17) dt > 0. t t The last term is handled similarly. Case 2: p > 3. First, for any n ∈ N, n > 1, one easily checks by induction that we have the following identity τ(cid:0)(a − b)(ap−1 − bp−1)(cid:1) = τ(cid:0)δ(cid:0)(b + δ)p−1−n − bp−1−n(cid:1)bn(cid:1) + n Xk=1 τ(cid:0)δ(b + δ)p−1−kδbk−1(cid:1). Let n > 1 be so that p − 1 − n = 1 + θ with θ ∈ [0, 1[. By positivity of the trace, we get τ(cid:0)(a − b)(ap−1 − bp−1)(cid:1) > τ(cid:0)δ(cid:0)(b + δ)1+θ − b1+θ(cid:1)bn(cid:1) + τ(cid:0)δ2(b + δ)p−2). By the same computations as above thanks to (1) τ(cid:0)δ(cid:0)(b + δ)1+θ − b1+θ(cid:1)bn(cid:1) = cθ ZR+ = cθ ZR+ > cθ ZR+ = cθ ZR+ = τ(cid:0)δ2bn+θ(cid:1) = τ(cid:0)δ2bp−2(cid:1), tθτ(cid:16)δ(cid:0)δ + t2(b + δ + t)−1 − t2(b + t)−1(cid:1)bn(cid:17) dt tθτ(cid:16)δ(cid:0)δ − t2(b + δ + t)−1δ(b + t)−1(cid:1)bn(cid:17) dt tθτ(cid:16)δ(cid:0)δ − tδ(b + t)−1(cid:1)bn(cid:17) dt tθτ(cid:16)δ2b(b + t)−1bn(cid:17) dt t t t t where we used again positivity of the trace with (b + δ + t)−1 6 t−1 and 0 6 (b + t)−1bn. To conclude let E to be the conditional expectation onto the subalgebra N = {δ}′′. As N is commutative, the Jensen inequality is valid; for any α > 1 and x ∈ M+: E(xα) > (Ex)α. With α = p − 2 > 1, τ(cid:0)δ2bp−2(cid:1) = τ(cid:0)δ2E(bp−2)(cid:1) > τ(cid:0)δ2(Eb)p−2(cid:1), τ(cid:0)δ2(b + δ)p−2(cid:1) > τ(cid:0)δ2(E(b + δ))p−2(cid:1). But with the usual decomposition δ = δ+ − δ−, as a, b > 0, Eb > δ− and E(b + δ) > δ+. By commutativity of N , we can conclude τ(cid:0)(a − b)(ap−1 − bp−1)(cid:1) > τ(cid:0)δ2(δp−2 − + δp−2 + )(cid:1) = τ(cid:0)δp(cid:1). (cid:3) We provide an alternative proof when p ∈ [3, 4]. Denote by Rx and Lx the right and left multiplication operators by x ∈ M defined on all Lp(M) (1 6 p 6 ∞). When p = 2, for any x ∈ Msa, the C ∗-algebra generated in B(L2(M)) by Lx and Rx is commutative and isomorphic to C(σ(x) × σ(x)) where σ(x) is the spectrum of x. Lemma 2.2. For p > 1, the map fp is Fr´echet differentiable on M++. For M finite, the derivative is given by the formula in L2(M): ∀x ∈ M++, ∀h ∈ Msa, Dxfp(h) = pZ 1 0 (cid:16)tLx + (1 − t)Rx(cid:17)p−1 (h) dt. AN INEQUALITY IN NONCOMMUTATIVE Lp-SPACES 3 Proof. Assume x > 3δ for some δ > 0. Taking h ∈ Msa with khk < δ, we may compute fp(x + h) using the holomorphic functional calculus by choosing a curve γ with index 1 surrounding the spectrum of x with γ ⊂ {z Rez > 0} and dist(γ, σ(x)) > 2δ: (x + h)p = 1 2iπ Zγ zp z − (x + h) dz Hence (x + h)p − xp = 1 2iπ Zγ zp(cid:0)z − (x + h)(cid:1)−1 h(cid:0)z − x(cid:1)−1 dz It follows directly that fp is Fr´echet differentiable with derivative Dxfp(h) = 1 2iπ Zγ zp(cid:0)z − x(cid:1)−1 h(cid:0)z − x(cid:1)−1 dz = 1 2iπ Zγ zpL(z−x)−1R(z−x)−1(h) dz. It then suffices to check that the two formulas coincide when M is finite; as M ⊂ L2(M), we do it for h ∈ L2(M). But in B(L2(M)), this boils down to an equality in C(σ(x) × σ(x)) so that we need only to justify that ∀a, b ∈ R+∗, 1 2iπ Zγ zp (z − a)(z − b) dz = pZ 1 0 (cid:0)ta + (1 − t)b(cid:1)p−1 dt. The above computations yield that the left-hand side is ap−bp a−b clearly coincide with the right-hand side. if a 6= b and pap−1 if a = b which (cid:3) Assuming M finite and a = b + δ, b ∈ M++ as above, the alternative proof when p ∈ [3, 4] relies on Lemma 2.2: τ(cid:0)(a − b)(ap−1 − bp−1)(cid:1) = pZ 1 = pZ 1 0 Z 1 0 Z 1 0 0 τ(cid:16)δ(cid:16)tLb+uδ + (1 − t)Rb+uδ(cid:17)p−2 hδ,(cid:16)tLb+uδ + (1 − t)Rb+uδ(cid:17)p−2 (δ)(cid:17) dt du (δ)iL2(M) dt du onto the subalgebra generated by δ. As p − 2 ∈ [1, 2], fp−2 is operator convex, so that for any m ∈ B(L2(M))+ and any projection E ∈ B(L2(M)), we have Emp−2E > (cid:0)EmE)p−2. We choose E to be the L2-conditional expectation hδ,(cid:16)tLb+uδ + (1 − t)Rb+uδ(cid:17)p−2 (δ)iL2(M) = hδ, E(cid:16)tLb+uδ + (1 − t)Rb+uδ(cid:17)p−2 E(δ)iL2(M) > hδ,(cid:16)tELb+uδE + (1 − t)ERb+uδE(cid:17)p−2 = hδ,(cid:16)tLE(b)+uδE + (1 − t)RE(b)+uδE(cid:17)p−2 = hδ,(cid:16)tLE(b)+uδ + (1 − t)RE(b)+uδ(cid:17)p−2 (δ)iL2(M) (δ)iL2(M) (δ)iL2(M), where in the last equality we have used that Rx and E commute if x ∈ δ′′. Tracking back the equalities, we obtain τ(cid:0)δ((b + δ)p−1 − bp−1)(cid:1) > τ(cid:0)δ((E(b) + δ)p−1 − E(b)p−1)(cid:1) > τ(cid:0)δp(cid:1), where the last inequality comes from the result in the commutative case. Remark 2.3. We point out that, for p ∈]2, 3[, the result cannot be reduced to the commutative case as in the alternative proof. Indeed, t 7→ tp−2 is operator concave and the first inequality right above reverses. Remark 2.4. Let ϕ : Lp(M) → Lp′ (M) be the duality map so that hx, ϕ(x)iLp(M),Lp and kϕ(x)kp′ = kxkp−1 written as: for a, b ∈ Lp(M)+ ′ (M) = kxkp p . When restricted to Lp(M)+, it is exactly fp−1, so the result can be p In this form, the inequality extends to general Lp-spaces in the sense of Haagerup, see [5, 2] for the arguments. ha − b, ϕ(a) − ϕ(b)iLp(M),Lp ′ (M) > (cid:13)(cid:13)a − b(cid:13)(cid:13) p p . 4 ´E. RICARD Corollary 2.5. Let (M, τ ) be a semifinite von Neumann algebra and E : M → M be a τ -preserving conditional expectation, then for all p > 2 and x ∈ Lp(M)+, (2) (cid:13)(cid:13)x − Ex(cid:13)(cid:13)p . 6 (cid:13)(cid:13)x(cid:13)(cid:13)p Proof. Apply the above theorem with a = x and b = Ex, as τ(cid:0)(x − Ex)(Ex)p−1(cid:1) = 0, the Holder inequality gives: p p (cid:13)(cid:13)x − Ex(cid:13)(cid:13) 6 τ(cid:0)(x − Ex)xp−1(cid:1) 6 (cid:13)(cid:13)x − Ex(cid:13)(cid:13)p(cid:13)(cid:13)x(cid:13)(cid:13) p−1 p . (cid:3) Remark 2.6. The inequality (2) does not hold for p < 2; a counterexample with M = ℓ2 ∞ can be found in [3]. There, a slight extension of (2) is given; one can replace E by any positive contractive projection C on Lp(M). As explained in [3], the main inequality applies more generally to semigroups. Corollary 2.7. Let (M, τ ) be a semifinite von Neumann algebra and (Tt)t>0 be a trace preserving 0 e−λtTtdt be its unital positive strongly continuous semigroup on M. For λ > 0, let Rλ = R ∞ resolvent. Then for all p > 2, λ > 0 and x ∈ Lp(M)+, Proof. We proceed as in Corollary 2.5. We apply Theorem 2.1 with a = x and b = λRλx to get (cid:13)(cid:13)x − λRλx(cid:13)(cid:13)p . 6 (cid:13)(cid:13)x(cid:13)(cid:13)p (cid:13)(cid:13)x − λRλx(cid:13)(cid:13) p p 6 τ(cid:0)(x − λRλx)xp−1(cid:1) − τ(cid:0)(x − λRλx)(λRλx)p−1(cid:1) p − τ(cid:0)(x − λRλx)(λRλx)p−1(cid:1). 6 (cid:13)(cid:13)x − λRλx(cid:13)(cid:13)p(cid:13)(cid:13)x(cid:13)(cid:13) p−1 To conclude, it suffices to note that τ(cid:0)(x − λRλx)(Rλx)p−1(cid:1) > 0. It can be checked by approximations thanks to the resolvent formula: x − λRλx = limt→∞ t(1 − tRt)Rλx. Indeed, recall that tRt is positive unital and trace preserving and hence a contraction on Lp so that τ(cid:0)tRt(Rλx).(Rλx)p−1(cid:1) 6 (cid:13)(cid:13)Rλx(cid:13)(cid:13) p p and τ(cid:0)(cid:0)t(1 − tRt)Rλx(cid:1)(Rλx)p−1(cid:1) > 0. (cid:3) References [1] Rajendra Bhatia. Matrix analysis, volume 169 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1997. [2] Uffe Haagerup, Marius Junge, and Quanhua Xu. A reduction method for noncommutative Lp-spaces and appli- cations. Trans. Amer. Math. Soc., 362(4):2125 -- 2165, 2010. [3] Mustapha Mokhtar-Kharroubi. Contractivity theorems in real ordered Banach spaces with applications to relative operator bounds, ergodic projections and conditional expectations. Preprint, https://hal.archives-ouvertes.fr/hal- 01148968. [4] Gilles Pisier and Quanhua Xu. Non-commutative Lp-spaces. In Handbook of the Geometry of Banach Spaces, Vol. 2, pages 1459 -- 1517. North-Holland, Amsterdam, 2003. [5] ´Eric Ricard. Holder estimates for the noncommutative Mazur maps. Arch. Math. (Basel) 104 (2015), no. 1, 37 -- 45. Laboratoire de Math´ematiques Nicolas Oresme, Universit´e de Caen Normandie, 14032 Caen Cedex, France E-mail address: [email protected]
1709.08586	1	1709	2017-09-25T16:41:04	Gelfand-Kirillov dimension of some simple unitarizable modules	[ "math.OA", "math.QA" ]	Let $\mathcal{O}_q(G)$ be the quantized algebra of regular functions on a semisimple simply connected compact Lie group $G$. Simple unitarizable left $\mathcal{O}_q(G)$-module are classified. In this article, we compute their Gelfand-Kirillov dimension where $G$ is of the type $A$, $C$ and $D$.	math.OA	math	Gelfand-Kirillov dimension of some simple unitarizable modules Partha Sarathi Chakraborty, Bipul Saurabh April 8, 2018 Abstract Let Oq(G) be the quantized algebra of regular functions on a semisimple simply con- nected compact Lie group G. Simple unitarizable left Oq(G)-module are classified. In this article we compute their Gelfand-Kirillov dimension where G is of the type A, C and D. AMS Subject Classification No.: 16P90, 17B37, 20G42 Keywords. Weyl group, Gelfand Kirillov dimension, Quantum Groups, Simple Unitarizable Modules. 1 Introduction Earlier in sixties, Gelfand and Kirillov ([6]) introduced a notion of growth of an algebra or a module, namely Gelfand Kirillov dimension. This invariant played a crucial role in the classification of Weyl algebras. It has been computed on many occasions (see [1], [13], [18], [8]). The Gelfand-Kirillov dimension of the universal enveloping algebra of a finite dimensional Lie algebra is same as its dimension (Proposition 8.1.15,(iii) in [14]). Even though this has also been computed for some quantized universal enveloping algebras ([13]) it is generally agreed upon that it is not easy to calculate them for modules in general ([1]). Instead of considering the quantized universal enveloping algebras one can also consider their duals namely the quantized algebra of regular functions Oq(G) and their modules. However in this case most of the literature is concerned with the type A situation only (example (4.8) page 258 in [3]). Here we take up the cases of type C and D as well. More specifically given a semisimple simply connected compact Lie group G, Korogodski and Soibelman ([11]) classified all simple unitarizable modules of the quantized algebra of regular functions on G denoted by Oq(G). Given an element w in the Weyl group and an element in a fixed maximal torus of G, they produced a simple unitarizable Oq(G)-module Vt,w and showed that in each equivalence class there is exactly one such module. Some natural questions arise; what is Gelfand Kirillov dimension of Oq(G)-module Vt,w and how is it related to the parameters t and w. we prove that Gelfand Kirillov dimension of Vt,w is equal to length 1 of the element w of the Weyl group provided G is of type A, C or D. The key idea is to decompose the Weyl word w into smaller parts in a certain way, prove the result for the last part and then use backward induction. This method also applies to the modules of quantized algebra of regular functions on certain homogeneous spaces. The paper is organized as follows. Next section describes all simple unitarizable modules of Hopf ∗-algebra Oq(G) and associates a diagram to each such module. Also, a brief introduction of the Weyl groups of type A, C and D and a reduced expression of an element of the Weyl group are given. In the third section, we prove our main result. In the final section, we take certain homogeneous spaces and prove similar result. Throughout the paper algebras are assumed to be unital and over the field C. Elements of a Weyl group will be called Weyl words. We denote by ℓ(w) the length of the Weyl word w. Let us denote by {en : n ∈ N} the standard basis of the vector space c00(N) where c00(N) is the set of finitely supported sequence of complex numbers. The endomorphism en 7→ en−1 of c00(N) will be denoted by S. The map en 7→ nen will be denoted by N . Given two endomorphisms T and T on V if there exist integers m1, m2, · · · , mk and a nonzero constant C such that of c00(N)⊗k and a subspace V of c00(N)⊗k we say that T ∼ T ′ ′ T = CT ′ (qm1N ⊗ qm2N ⊗ · · · ⊗ qmk N ) on V . Throughout this paper, q will denote a real number in the interval (0, 1) and C is used to denote a generic constant. 2 Simple unitarizable modules of quantized function algebras In this section, we recall the definition of quantized algebra of regular functions on a simply connected semisimple compact Lie group G and the classification of its simple unitarizable modules. For a detailed treatment, we refer the reader to ([10], Chapter 3 in [11]). Let G be a simply connected semisimple compact Lie group, g its complexified Lie algebra of rank n. Fix a nondegenerate symmetric ad-invariant form h·, ·i on g such that its restriction to the real Lie algebra of G is negative definite. Let Π := {α1, α2, · · · , αn} be the set of simple roots. For simplicity, we write the root αi as i and the reflection sαi defined by the root αi as si. The Weyl group Wn of G can be described as the group generated by the reflections {si : 1 ≤ i ≤ n}. Definition 2.1. Let Uq(g) be the quantized universal enveloping algebra equipped with the ∗-structure corresponding to the compact real form of g (see page 161 and 179, [10]). Then the Hopf ∗-subalgebra of the dual Hopf ∗-algebra of Uq(g) consisting of matrix co-efficients of finite dimensional unitarizable Uq(g)-modules is called the quantized algebra of regular functions on G (see page 96 − 97, [11]). We denote it by Oq(G). Let ((ui j,g)) be the defining corepresentation of Oq(G) if G is of type An and Cn and the irreducible corepresentation of Oq(G) corresponding to the highest weight (1, 0, 0, · · · , 0) if G 2 is of type Dn. In first case, entries of the matrix ((ui j,g)) generate the Hopf ∗-algebra Oq(G). In latter case, they genarate a proper Hopf ∗-subalgebra of Oq(Spin(2n)) which we denote as Oq(SO(2n)). The generators of Oq(Spin(2n)) are the matrix entries of the corepresentation ((zi j)) of Oq(Spin(2n)) with highest weight (1/2, 1/2, · · · , 1/2). We denote the dimension of the corepresentation ((ui j,g)) whenever the Lie algebra g is clear from the context. Invoking a result of Korogodski and Soibelman ([11]), we will now describe all simple unitarizable Oq(G)-modules. j,g)) by Nn. We will drop the subscript g in ((ui Elementary simple unitarizable Oq(G)-modules: For 1 ≤ i ≤ n, let di = hαi, αii /2 and qi = qdi. Define φi : Uqi(sl(2)) −→ Uq(g) be a ∗-homomorphism given on the generators of Uqi(sl(2)) by, K 7−→ Ki, E 7−→ Ei, F 7−→ Fi. By duality, it induces an epimorphism φ∗ i : Oq(G) −→ Oqi(SU (2)). We will use this map to get all elementary simple unitarizable modules of Oq(G). Denote by Ψ the following action of Oq(SU (2)) on c00(N) (see Proposition 4.1.1, [11]); Ψ(uk l )ep = p1 − q2pep−1 p1 − q2p+2ep+1 −qp+1ep qpep δklep if k = l = 1, if k = l = 2, if k = 1, l = 2, if k = 2, l = 1, otherwise . (2.1)   si := Ψ◦φ∗ i of Oq(G). Each πn For each 1 ≤ i ≤ n, define an action πn si gives rise to an elementary simple Oq(G)-module Vsi. Also, for each t ∈ Tn, there are one dimensional Oq(G)-module Vt with the action {τ n t }. Given two actions ϕ and ψ of Oq(G), define an action ϕ∗ψ := (ϕ⊗ψ)◦∆. Similarly for any two Oq(G)-module Vϕ and Vφ, define Vϕ ⊗ Vφ as Oq(G)-module with Oq(G) action coming from ϕ ∗ ψ. For w ∈ Wn such that si1si2...sik is a reduced expression for w, define an action πn and the corresponding Oq(G)-module as Vw. Then Vw is an simple unitarizable which is independent of the reduced expression. Moreover, for t ∈ Tn, w ∈ Wn, define an action πn w and denote the corresponding Oq(G)-module by Vt,w. We refer to ([11], page 121) for the following theorem. ∗ · · · ∗ πn sik t,w by τ n t ∗ πn w as πn si1 ∗ πn si2 Theorem 2.2. The set {Vt,w; t ∈ Tn, w ∈ Wn} is a complete set of mutually inequivalent simple unitarizable left Oq(G)-module. 2.1 Diagram representation As we are interested in the Gelfand-Kirillov dimension of certain modules it is essential that we understand the algebra action clearly. A diagram representation of various endomorphisms 3 helps us in this respect. Corresponding to every πn si, we have a diagram and thus a diagram for each πn w is obtained as a concatenation of diagrams of elementary modules. Using equation (2.1), one can write down explicitly the endomorphisms {πn l ) : 1 ≤ k, l ≤ Nn, 1 ≤ i ≤ n} for G of type An, Cn and Dn. We order the indices of the simple roots in Π in the standard way (see Appendix C in [9]). In this set up, these endomorphisms and their diagrams are given explicitly in [4] and [15] for G of type An and Cn respectively. For type Dn, it is given below. For i = 1, 2, · · · , n − 1, si(uk   πn si(uk l )ep = For i = n, p1 − q2nep−1 p1 − q2n+2ep+1 −qp+1ep qpep qp+1ep −qpep if (k, l) = (i, i) or (2n − i, 2n − i), if (k, l) = (i + 1, i + 1) or (2n − i + 1, 2n − i + 1), if (k, l) = (i, i + 1), if (k, l) = (i + 1, i), if (k, l) = (2n − i, 2n − i + 1), if (k, l) = (2n − i + 1, 2n − i), δklep otherwise . πn sn(uk l )ep = p1 − q2nep−1 p1 − q2n+2ep+1 −qp+1ep qpep qp+1ep −qpep if (k, l) = (n, n) or (n − 1, n − 1), if (k, l) = (n + 1, n + 1) or (n + 2, n + 2), if (k, l) = (n − 1, n + 1), if (k, l) = (n + 1, n − 1), if (k, l) = (n, n + 2) if (k, l) = (n + 2, n), δklep otherwise .   For t = (t1, t2, · · · , tn) ∈ Tn, the one dimensional irreducible representation τt maps ui if i ≤ n and to t2n+1−iδij if i > n. j to tiδij Now we will associate a diagram with each of the above action of Oq(G). Let us describe the case of Oq(G) first. For convenience, we use labeled lines and arrows to represent endo- morphisms of a vector space as given in the following table. Arrow type Endomorphism Arrow type Endomorphism I + S∗pI − q2N +2 t − qN −qN 4 Mt pI − q2N +2S −qN +1 qN +1 Note that for t ∈ T, Mt represents the endomorphism on C sending v to tv. For endomor- phism I, the underlying vector space can be either C or c00(N). For other endomorphisms, the vector spaces on which they act are c00(N). Let us describe how to use a diagram to represent the actions πn t where 1 ≤ i ≤ n and t ∈ Tn. si and τ n c00(N) 2n 2n + - + - 2n − i + 1 2n − i i + 1 i 2n − i + 1 2n − i i + 1 i 1 Diagram 1: πn 1 si, i 6= n c00(N) 2n 2n n + 2 n + 1 n n − 1 + + - - n + 2 n + 1 n n − 1 1 1 Diagram 2: πn sn C t1 2n 2n n + 2 n + 1 n n − 1 tn−1 tn tn tn−1 n + 2 n + 1 n n − 1 t1 1 1 Diagram 3: τ n t si(uk l ), πn sn(uk In these diagrams, each path from a node k on the left to a node l on the right stands for an endomorphism given as in the table acting on the vector space given at the top of the diagrams. Now πn l ) are the endomorphisms represented by the paths from k to l in diagram 1, diagram 2 and diagram 3 respectively and are zero if there is no such path. i+1) = −qN +1 if i > 1. Thus, for example, in diagram 1, πn In case of Oq(SU (n + 1)), the table of symbols representing the operators appearing in the 1) is zero and πn 1) is I; πn l ) and τ n t (uk si(u2 si(u1 si(ui diagram of πn si for 1 ≤ i ≤ n are as follows (see [4]). Arrow type Operator Arrow type Operator + I S∗pI − q2N +2 qN t − Mt pI − q2N +2S −qN +1 5 L2(N) n + 1 n + 1 i + 1 i 1 + − i + 1 i 1 C t1 t2 t3 tn t1 · · · tn n + 1 n n − 1 2 1 n + 1 n n − 1 2 1 Diagram 4: πsi Diagram 5: τ n t For Oq(SP (2n)), these data are described below (see [15]). Arrow type Endomorphism Arrow type Endomorphism Arrow type Endomorphism I + S∗pI − q2N +2 t − qN −qN Mt pI − q2N +2S −qN +1 qN +1 c00(N) 2n 2n c00(N) 2n 2n + - + - 2n − i + 1 2n − i i + 1 i 2n − i + 1 2n − i i + 1 i 1 Diagram 6: πn 1 si, i 6= n n + 2 n + 1 n n − 1 + - n + 2 n + 1 n n − 1 1 1 Diagram 7: πn sn −q2N +2 q2N C t1 2n 2n n + 2 n + 1 n n − 1 tn−1 tn tn tn−1 n + 2 n + 1 n n − 1 t1 1 1 Diagram 8: τ n t Next, let us explain how to represent π ∗ρ by a diagram where π and ρ are two actions of Oq(G) acting on the vector spaces V1 and V2 respectively. Simply keep the two diagrams representing π and ρ adjacent to each other. Identify, for each row, the node on the right side of the diagram for π with the corresponding node on the left in the diagram for ρ. Now, (π ∗ ρ)(uk l ) would be an endomorphism of the vector space V1 ⊗ V2 determined by all the paths from the node k 6 on the left to the node l on the right. It would be zero if there is no such path and if there are more than one paths, then it would be the sum of the endomorphisms given by each such path. In this way, we can draw diagrams for each action of Oq(G). The following diagram is for the action π4 t,w of Oq(SO(8)) where w = s1s2s3s4s2 and t = (t1, t2, 1, 1). C ⊗ c00(N) ⊗ c00(N) ⊗ c00(N) ⊗ c00(N) ⊗ c00(N) 8 7 6 5 4 3 2 1 • • • • • • • • t1 t2 t2 t1 π4 t + − + π4 s1 • • • • • • • • ∗ • • • • • • • • ∗ + − + − • • • • • • • • + − + − • • • • • • • • + + − − ∗ π4 s3 π4 s2 Diagram 9: π4 t,w ∗ π4 s4 8 7 6 5 4 3 2 1 • • • • • • • • • • • • • • • • • ∗ + − + − π4 s2 Here there are two paths from the vertex 1 in the left to the vertex 2 in the right and hence the endomorphism π4 1) will be sum of the endomorphisms represented by each path. Therefore we have t,w(u2 π4 t,w(u2 1) = t1qN +1 ⊗ qN +1 ⊗ p1 − q2N +2S∗ ⊗p1 − q2N +2S∗ ⊗ qN −t1qN +1 ⊗ p1 − q2N +2S∗ ⊗ I ⊗ I ⊗ p1 − q2N +2S∗. Let us record few properties of these endomorphisms which are immidiate from the diagram. w(u8 1. The endomorphism π4 6) when applied to vector of the form ei1 ⊗ ei2 ⊗ ei3 ⊗ ei4 ⊗ e0 gives Cei1 ⊗ ei2+1 ⊗ ei3 ⊗ ei4 ⊗ e0. The point is; if a standard basis element has e0 in last co-ordinate then there is only one path from the 8 to 6 that survives as Se0 = 0 and it increases second co-ordinate of the standard basis element by a unit. 7 2. For each vertex i ∈ {2, 3, · · · 7} in the left, there is an unique vertex, say rw(i) in the right r(i))(e0 ⊗ e0 ⊗ · · · ⊗ e0) = Ce0 ⊗ e0 ⊗ · · · ⊗ e0. More explicitly, the map rw such that π4 sends 2 7→ 1, 3 7→ 3, 4 7→ 5, 5 7→ 4, 6 7→ 6 and 7 7→ 8. w(ui 3. Also, there may be many paths from i to rw(i) but when applied to e0 ⊗ e0 ⊗ · · · ⊗ e0, only one path give a nonzero vector i.e Ce0 ⊗ e0 ⊗ · · · ⊗ e0. 4. For any j ∈ {2, 3, · · · 7}, j 6= i, we have π4 w(uj rw(i))(e0 ⊗ e0 ⊗ · · · ⊗ e0) = 0. As we can see that there is nothing special about the above diagram and these observations hold in general. We will make these things more precise in the next section but with this example in mind one can get a better understanding of the results. 2.2 Weyl group We will now give a brief introduction of the Weyl group Wn of type An, Cn and Dn. For details, we refer the reader to [5] or [7]. Let Ei,j denote the n × n matrix having only one non-zero entry 1 at ijth position. The group Wn is isomorphic to a subgroup of GL(n, R) generated by s1, s2, ...sn where Case 1: sl(n + 1) si = I − Ei,i − Ei+1,i+1 + Ei,i+1 + Ei+1,i, for i = 1, 2, ...n. Hence Wn is isomorphic to permutation group of n letters. Case 2: sp(2n) si = I − Ei,i − Ei+1,i+1 + Ei,i+1 + Ei+1,i, sn = I − 2En,n, for for i = 1, 2, ...n − 1, i = n. Hence Wn is the set of n × n matrices having one non-zero entry in each row and each column which is either 1 or −1. Case 3: so(2n) si = I − Ei,i − Ei+1,i+1 + Ei,i+1 + Ei+1,i, sn = I − 2En,n − 2En−1,n−1, for for i = 1, 2, ...n − 1, i = n. Hence Wn is the set of n × n matrices having one non-zero entry in each row and each column which is either 1 or −1 and number of −1's are even. Any element of Wn can be written in the form: ψ(ǫ1) 1,k1 (w) for some choices of ǫ1, ǫ2, · · · , ǫn and k1, k2, · · · kn where ǫr ∈ {0, 1, 2} and n − r + 1 ≤ kr ≤ n with the (w) · · · ψ(ǫn) n,kn (w)ψ(ǫ2) 2,k2 8 convention that, Case 1: sl(n + 1) Case 2: sp(2n) ψǫ r,kr (w) =  srsr−1 · · · sn−kr+1  empty string if ǫ = 1, 2 if ǫ = 0. ψǫ r,kr (w) =   sn−r+1sn−r+2 · · · skr if ǫ = 1, sn−r+1sn−r+2 · · · ...sn−1snsn−1 · · · skr if ǫ = 2, empty string if ǫ = 0. Case 3: so(2n) ψǫ r,kr (w) =   sn−r+1sn−r+2 · · · skr if ǫ = 1, sn−r+1sn−r+2 · · · ...sn−1snsn−2sn−3 · · · skr if ǫ = 2, empty string if ǫ = 0. Also, the above expression is a reduced expression. We will call the word ψ(ǫr) (w) the rth part r,kr wr of w. Accordingly, the simple module Vw decomposes as Vw = Vw1 ⊗ Vw2 ⊗ · · · ⊗ Vwn. We call Vwi the ith part of Vw. 3 Main result In this section, our main aim is to compute the Gelfand-Kirillov dimension of Vt,w for t ∈ Tn and w ∈ Wn. We first recall from [12] the definition of Gelfand Kirillov dimension of a module. Definition 3.1. ([12]) Let A be a unital algebra and M be a left A-module. The Gelfand- Kirillov dimension of M is given by GKdim(M ) = supV,F lim ln dim(V kF ) ln k where the supremum is taken over all finite dimensional subspace V of A containing 1 and all finite dimensional subspaces F of M . If A is a finitely generated unital algebra and M be a finitely generated left A-module then GKdim(M ) = supξ,F lim ln dim(ξkF ) ln k where the supremum is taken over all finite sets ξ containing 1 that generates A and all finite dimensional subspaces F of M that generates M as an left A-module. 9 ln k Remark 3.2. For a finitely generated algebra A and a finitely generated left A-module M , the quantity "lim ln dim(ξkF ) " does not depend on particular choices of ξ and F . Therefore one can choose a fixed (but finite) set of generators of A and a finite dimensional subspaces F of M that generates M as a left A-module. Moreover, if M is a simple left A-module then one can take one dimensional subspace spanned by any vector v ∈ M as a candidate for F (see [12] for details). Let ξn = {ui j : 1 ≤ i, j ≤ Nn} where Nn is the dimension of the representation ((ui j)). Define ξGq =   ξn ∪ {1} for Oq(G) = Oq(SU (n + 1)) or Oq(SP (2n)), ξn ∪ {zi j : 1 ≤ i, j ≤ 2n} ∪ {1} for Oq(G) = Oq(Spin(2n)). Then ξGq is a generating set of Oq(G) containing 1. Throughout the article, we will work with this generating set. Let w ∈ Wn and t ∈ Tn. The following lemma says that GKdim(Vw,t) is less than or equal to the length of the Weyl word. Lemma 3.3. Assume that Oq(G) is one of the Hopf ∗-algebras Oq(SU (n + 1)), Oq(SP (2n)) or Oq(Spin(2n)). For w ∈ Wn and t ∈ Tn, let Vt,w be the associated simple unitarizable left Oq(G)-module. Then one has GKdim(Vt,w) ≤ ℓ(w). Proof : The algebra Oq(SU (2)) has the following vector space basis (see Proposition 4, page 100, [10]). {(u1 1,sl2)b(u2 1,sl2)c(u1 2,sl2)d, (u2 1,sl2)v(u1 2,sl1)w(u2 2,sl2)y : b, c, d, v, w, y ∈ N}. Define the exponent of the element (u1 )d to be b and the exponent of the element (u2 )y to be y. Given a linear combination of these elements, define its exponent to be maximum of the exponent of the terms appearing in the linear combination. Let )w(u2 )v(u1 )c(u1 )b(u2 2,sl2 1,sl2 2,sl1 2,sl2 1,sl2 1,sl2 a = u∈ξGq ,i∈{1,2,··· ,n}(cid:8)exponent of φ∗ max i (u) : u ∈ ξGq(cid:9). Using equation (2.1), we get span πn t,w(ξr Gq )(e0 ⊗ · · · ⊗ e0) ⊂ span{ei1 ⊗ · · · ⊗ eiℓ(w) : 0 ≤ ij ≤ ra for all 1 ≤ j ≤ ℓ(w)}. Therefore one has GKdim(Vt,w) = lim ln dim(ξr Gq (e0 ⊗ · · · ⊗ e0)) ln r ≤ lim ln(ra)ℓ(w) ln r = ℓ(w). This proves the claim. (cid:3) In the following lemma, we give some polynomials which when applied in a certain order give all matrix units of the nth part. More to a fixed vector of the form v ⊗ e0 ⊗ e0 ⊗ · · · ⊗ e0 } nth−part {z precisely, 10 Lemma 3.4. Assume that Oq(G) is one of the Hopf ∗-algebras Oq(SU (n + 1)), Oq(SP (2n)) or Oq(Spin(2n)) and w ∈ Wn. Then there exist polynomials p(w,n) ℓ(wn) with non- commuting variables πn )'s and a permutation σ of {1, 2, · · · , ℓ(wn)} such that for all 1 , rn rn w(uNn ℓ(wn) ∈ N, one has , · · · p(w,n) 2 , · · · , rn , p(w,n) 1 2 j (p(w,n) 1 )rn σ(1)(p(w,n) 2 )rn σ(2) · · · (p(w,n) ℓ(wn))rn σ(ℓ(wn))(v ⊗ e0 ⊗ · · · ⊗ e0) = Cv ⊗ ern 1 ⊗ ern 2 ⊗ · · · ⊗ ern ℓ(wn) where v ∈ c00(N)⊗ Pn−1 i=1 ℓ(wi) and C is a nonzero real number. Proof : Let us consider each case separately. Case 1: Oq(G) = Oq(SU (n + 1)). In this case Nn = n + 1. For 1 ≤ j ≤ ℓ(wn), define σ(j) := j and p(w,n) of Oq(SU (n + 1)), one has w(uNn := πn j Nn−j+1). Using diagram representations p(w,n) j = 1⊗ Pn−1 i=1 ℓ(wi) ⊗ (qN )⊗j−1 ⊗ p1 − q2N S∗ } j th place {z ⊗1⊗ℓ(wn)−j. Therefore (p(w,n) j )r ∼ 1⊗ Pn−1 i=1 ℓ(wi) ⊗ (qrN )⊗j−1 ⊗ (p1 − q2N S∗)r } {z i=1 ℓ(wi) ⊗ 1⊗j−1 ⊗ (p1 − q2N S∗)r } j th place j th place {z = 1⊗ Pn−1 ⊗1⊗ℓ(wn)−j, ⊗1⊗ℓ(wn)−j. on the whole vector space. Using this, one can get the claim. Case 2 : Oq(G) = Oq(SP (2n)). In this case Nn = 2n. If ℓ(wn) ≤ n then w = w1w2 · · · wn−1s1s2 · · · sℓ(wn). In this case, define σ(j) := j and p(w,n) w(uNn πn case, define := Nn−j+1). If ℓ(wn) > n, then w = w1w2 · · · wn−1s1s2 · · · sn−1snsn−1 · · · sNn−ℓ(wn). In this j p(w,n) j = w(uNn πn Nn−j+1) w(uNn πn j+1) w(uNn n+1), πn [πn w(uNn πn ) j w(uNn n )] if 1 ≤ j ≤ Nn − ℓ(wn) − 1, if Nn − ℓ(wn) ≤ j < n, if j = n, if n + 1 ≤ j ≤ ℓ(wn).   Let σ(j) be Nn − j if Nn − ℓ(wn) ≤ j ≤ ℓ(wn) and j otherwise. By a direct computation or using diagram representation, we have p(w,n) j ∼   1⊗ Pn−1 1⊗ Pn−1 i=1 ℓ(wi) ⊗ 1⊗σ(j)−1 ⊗ p1 − q2N S∗ } i=1 ℓ(wi) ⊗ 1⊗σ(n)−1 ⊗ p1 − q4N S∗ } σ(n)th place σ(j)th place {z {z 11 ⊗1⊗ℓ(wn)−σ(j) if j 6= n, ⊗1⊗ℓ(wn)−σ(n) if j = n, (3.1) on the subspace generated by standard basis elements having e0 at σ(k)th place for k < j. Since qNp1 − q2N S∗ = p1 − q2N S∗qN +2 and qNp1 − q4N S∗ = p1 − q4N S∗qN +4, we get (p(w,n) j )r ∼   1⊗ Pn−1 1⊗ Pn−1 i=1 ℓ(wi) ⊗ 1⊗σ(j)−1 ⊗ (p1 − q2N S∗)r } i=1 ℓ(wi) ⊗ 1⊗σ(n)−1 ⊗ (p1 − q4N S∗)r } σ(n)th place σ(j)th place {z {z ⊗1⊗ℓ(wn)−σ(j) if j 6= n, ⊗1⊗ℓ(wn)−σ(n) if j = n, The rest is a straightforward checking. Case 3: Oq(G) = Oq(Spin(2n)). In this case Nn = 2n. If ℓ(wn) < n then w = w1w2 · · · wn−1s1s2 · · · sℓ(wn). In this case, define p(w,n) Nn−j+1) and σ(j) := j. If ℓ(wn) ≥ n, then w = w1w2 · · · wn−1s1s2 · · · sn−2sn−1snsn−2 · · · sNn−ℓ(wn)−1. In this case, define w(uNn := πn j p(w,n) j =   w(uNn πn w(uNn πn Nn−j+1) j+1) if 1 ≤ j ≤ Nn − ℓ(wn) − 2, otherwise. (3.2) Let σ(j) be Nn − j − 1 if Nn − ℓ(wn) − 1 ≤ j ≤ ℓ(wn) and j otherwise. Using diagram representation, we have p(w,n) j ∼ 1⊗ Pn−1 i=1 ℓ(wi) ⊗ 1⊗σ(j)−1 ⊗ p1 − q2N S∗ } σ(j)th place {z ⊗1⊗ℓ(wn)−σ(j) (3.3) on the subspace generated by standard basis elements having e0 at σ(k)th place for k < j. The similar calculations as above will prove the result. (cid:3) Remark 3.5. 1. As we saw in part (1) of the observations made in subsection 2.1 that the order in which we apply these endomorphisms are important just to make sure that the matrix unit e0 remains in certain components and that's why a permutation occurs in this lemma. 2. Note that the polynomials p(w,n) j of the matrix ((ui j)) only. ; 1 ≤ j ≤ ℓ(wn) defined here involve elements of last row To show our main claim, we need to extend these results for all parts of the Weyl word. n = 1 n := Nn − n + i and for type An, let M i n = n − i + 1 and N i For type Cn and Dn, let M i and N i n := i + 1. Lemma 3.6. Assume that Oq(G) is one of the Hopf ∗-algebras Oq(SU (n + 1)), Oq(SP (2n)) or Oq(Spin(2n)). 12 1. Let w ∈ Wn be of the form w = wi+1 and let Vw be the associated Oq(G)-module. Then for each M i n ≤ k ≤ N i n, there exists unique rw(k) ∈ {M i+1 n , · · · , N i+1 n } such that πn w(uk rw(k))(e0 ⊗ e0 ⊗ · · · ⊗ e0) = Ce0 ⊗ e0 ⊗ · · · ⊗ e0 where C is a nonzero real number. Moreover, w(uj πn rw(k))(e0 ⊗ e0 ⊗ · · · ⊗ e0) = 0 for j ∈ {M i n, · · · , N i n}/{k}. (3.4) (3.5) 2. Let w ∈ Wn be of the form w = wi+1wi+2 · · · wl, l ≤ n and let Vw be the associated n, define rw(k) := rwl ◦ rwl−1 ◦ · · · rwi+2 ◦ rwi+1(k) ∈ n ≤ k ≤ N i Oq(G)-module. For each M i {M l n, · · · , N l n}. Then rw satisfies equations (3.4) and (3.5). We have seen these results in a particular case given as part (2), (3) and (4) of the observa- tions in subsection 2.1. This lemma basically generalises these observations to an appropriate set up. Proof : Proof of part (1): To show the claim, we will define the function rwi+1 explicitly in each case in the following way. Case 1: Oq(G) = Oq(SU (n + 1)). In this case, wi+1 = sisi−1 · · · si−ℓ(wi+1)−1 and hence for all 1 ≤ k ≤ n, sk occurs in wi+1 either once or does not occur. rwi+1(j) =   j + 1, if sj occurs in wi+1 once , j otherwise . Case 2 : Oq(G) = Oq(SP (2n)). In this case, wi+1 is either of the form sisi+1 · · · sk where k ≤ n or sisi−1 · · · sn−1snsn−1 · · · sk. Therefore for all 1 ≤ k < n, sk occurs in wi+1 either once or twice or does not occur. For j ≤ n, define For j ≥ n + 1, define j − 1, j j + 1, rwi+1(j) =   rwi+1(j) =   j if sj−1 occurs in wi+1 once , otherwise . if s2n−j occurs in wi+1 once , otherwise . Case 3: Oq(G) = Oq(Spin(2n)). In this case, wi+1 is either of the form sisi−1 · · · sk where k ≤ n − 2 or sisi+1 · · · sn−2sn−1sn or sisi+1 · · · sn−2sn−1snsn−2sn−3 · · · sk. Therefore for all 1 ≤ k < n − 1, sk occurs in wi+1 either 13 once or twice or does not occur. Moreover, sn−1sn can occur either once or does not occur. For j ≤ n − 1, define j − 1, rwi+1(j) =   j if sj−1 occurs in wi+1 once , otherwise . For j > n + 1, define rwi+1(j) =   j + 1, if s2n−j occurs in wi+1 once , j otherwise . Also, rwi+1(n) and rwi+1(n + 1) are n and n + 1 respectively if sn−1sn does not occur in wi+1 and n + 1 and n respectively if sn−1sn occur once in wi+1. Now by a direct verification using diagrams, one can establish the claim. Proof of part (2): Since {M i n, M i n + 1, · · · , N i n} ⊂ {M i+1 n , · · · , N i+1 n } ⊂ · · · ⊂ {M l n, · · · , N l n} n, · · · , N i the map rw : {M i ing diagram representation and possible forms of wi+1 given in subsection 2.2 that πn for k ∈ {M i n} is well defined. To verify the claim, first note us- j ) = 0 n }). Applying this, we get n} → {M l n} and j /∈ {M i+1 n , · · · N i+1 n, · · · , N i n, · · · , N l wi+1(uk rwi+2 ◦rwi+1 (k))(e0 ⊗ e0 ⊗ · · · ⊗ e0) j )(e0 ⊗ e0 ⊗ · · · ⊗ e0) ⊗ πn wi+2(uj rwi+2 ◦rwi+1 (k))(e0 ⊗ e0 ⊗ · · · ⊗ e0) wi+1wi+2(uk πn X wi+1(uk πn = n j=1 N i+1 n X = n j=M i+1 wi+1(uk = πn πn wi+1(uk j )(e0 ⊗ e0 ⊗ · · · ⊗ e0) ⊗ πn wi+2(uj rwi+2 ◦rwi+1 (k))(e0 ⊗ e0 ⊗ · · · ⊗ e0) rwi+1 (k))(e0 ⊗ e0 ⊗ · · · ⊗ e0) ⊗ πn wi+2(u rwi+1 (k) rwi+2 ◦rwi+1 (k))(e0 ⊗ e0 ⊗ · · · ⊗ e0) = C(e0 ⊗ e0 ⊗ · · · ⊗ e0) (by equation (3.5)) (by equation (3.4)) Proceeding in this way, one can verify equation (3.4) for w = wi+1wi+2 · · · wl. Equation (3.5) can be verified similarly. (cid:3) Remark 3.7. Observe that the above lemma is not true if G is of type B. In particular, from the vertex n+1, there is no unique vertex having the property mentioned in Lemma 3.6. That's why we excluded type B case. 14 In the following lemma, we define some endomorphisms for each 1 ≤ i < n that act only on ith part Vwi of the simple module Vw associated with the Weyl word w and keep other parts unaffected. Moreover, on the ith part, they act exactly in the same way as the variables given in Lemma 3.4 assuming that the rank of the Lie algebra g associated with Oq(G) given there is i. Here we are considering two Lie algebras of same type but of different ranks simultaneously. So, instead of g, we will use its rank in the notation ((ui j,g)) to avoid any confusion. Lemma 3.8. Assume that Oq(G) is one of the Hopf ∗-algebras Oq(SU (n + 1)), Oq(SP (2n)) or Oq(Spin(2n)). For w ∈ Wn, let Vw = Vw1 ⊗ Vw2 ⊗ · · · ⊗ Vwn be the associated Oq(G)-module. Then for each 1 ≤ i < n, there exist endomorphisms T i l,n) : 1 ≤ l, k ≤ Nn} such that for all 1 ≤ j ≤ Ni, one has 2, · · · T i Ni ∈ {πn w(uk 1, T i T i j (v1 ⊗ v2 ⊗ e0 ⊗ · · · ⊗ e0) = C(πi w1w2···wi(uNi wi(uNi j,i )(v1 ⊗ v2)) ⊗ e0 ⊗ · · · ⊗ e0 j,i )v2) ⊗ e0 ⊗ · · · ⊗ e0 = Cv1 ⊗ (πi where v1 ∈ Vw1 ⊗ Vw2 ⊗ · · · ⊗ Vwi−1, v2 ∈ Vwi and C is a nonzero constant. Proof : Define the endomorphism T i j = πn w(uN i n rwi+1 wi+2 ···wn (j),n) for 1 ≤ j ≤ Ni. Here rwi+1wi+2···wn : {M i n} → {1, 2, . . . , Nn} be the function defined in Lemma 3.6. First note using diagram representation and possible forms of w1 · · · wi w1···wi(uN i given in subsection 2.2 that πn n}. Take v1 ∈ Vw1 ⊗ Vw2 ⊗ · · · ⊗ Vwi−1 and v2 ∈ Vwi. Then one has k ) = 0 for k /∈ {M i n + 1, . . . , N i n, · · · N i n, M i n n rwi+1 wi+2 ···wn (j),n)(v1 ⊗ v2 ⊗ e0 ⊗ · · · ⊗ e0) w1···wi(uN i πn n k,n)(v1 ⊗ v2)πn wi+1wi+2···wn(uk rwi+1 wi+2 ···wn (j),n)(e0 ⊗ · · · ⊗ e0) w(uN i πn X = n k=1 N i n = X k=M i n = C(πi = Cv1 ⊗ (πi w1···wi(uN i πn n k,n)(v1 ⊗ v2)πn wi+1wi+2···wn(uk rwi+1 wi+2 ···wn (j),n)(e0 ⊗ · · · ⊗ e0) w1w2···wi(uNi wi,(uNi j,i )(v1 ⊗ v2)) ⊗ e0 ⊗ · · · ⊗ e0 j,i )(v2)) ⊗ e0 ⊗ · · · ⊗ e0. ( by Lemma 3.6) This settles the claim. (cid:3) The following lemma is an extension of the result of Lemma 3.4 incorporating all parts of the Weyl word. Lemma 3.9. Assume that Oq(G) is one of the Hopf ∗-algebras Oq(SU (n + 1)), Oq(SP (2n)) or Oq(Spin(2n)). For w ∈ Wn, let Vw be the associated Oq(G)-module. Then for 1 ≤ i ≤ n 15 and 1 ≤ j ≤ ℓ(wi), there exist polynomials p(w,i) permutations σi of {1, 2, · · · , ℓ(wi)} such that j with noncommuting variables πn w(ur s)'s and (p(w,n) )rn σn(1)(p(w,n) 1 2 · · · ⊗ e0) = Cer1 )rn σn(2) · · · (p(w,n) ℓ(wn))rn σn(ℓ(wn)) · · · (p(w,1) )r1 σ1(1) · · · (p(w,1) ℓ(wi))r1 σ1(ℓ(wi))(e0 ⊗ e0 ⊗ · · · ⊗ er1 2 ⊗ · · · ⊗ er1 ℓ(w1) 1 · · · ⊗ ern 1 1 ⊗ ern 2 ⊗ · · · ⊗ ern ℓ(wn) where C is a nonzero constant. j be the polynomial as given in Lemma 3.4 and σn be the Proof : For 1 ≤ j ≤ ℓ(wn), let p(w,n) associated permutation. To define permutations σi and polynomials p(w,i) for 1 ≤ i < n and 1 ≤ j ≤ ℓ(wi), we view w1w2 · · · wi as an element of the Weyl group of G of rank i. So we can define polynomial p(w1w2···wi,i) and the permutation σi from Lemma 3.4. Replace the variables w1w2···wi(uNi πi p(w,i) j for all 1 ≤ j ≤ ℓ(wi). Now the claim follows from Lemma 3.4 and Lemma 3.8. k for 1 ≤ k ≤ Ni in the polynomial p(w1w2···wi,i) to define the polynomial k,i) with T i (cid:3) j j j Theorem 3.10. Assume that Oq(G) is one of the Hopf ∗-algebras Oq(SU (n + 1)), Oq(SP (2n)) or Oq(Spin(2n)). For w ∈ Wn and t ∈ Tn, let Vw,t be the associated simple unitarizable left Oq(G)-module. Then one has GKdim(Vw,t) = ℓ(w). Proof : Since GKdim(Vw,t) = GKdim(Vw), we will consider the Oq(G)-module Vw only. Define M0 := max{degree of p(w,i) j : 1 ≤ i ≤ n, 1 ≤ j ≤ ℓ(wi)}. Then by Lemma 3.9, we have spanπn w(ξM0k)(e0 ⊗ · · · ⊗ e0) ⊃ span{eα1 ⊗ eα2 ⊗ · · · eαℓ(w) : ℓ(w) X i=1 αi = k}. Hence dim span πn w(ξM0k)(e0 ⊗ · · · ⊗ e0) ≥ (cid:0)k+ℓ(w)−1 k (cid:1). Therefore GKdim(Vw) ≥ lim ln dim(ξM0k(e0 ⊗ · · · ⊗ e0)) ln M0k ≥ lim ln(cid:0)k+ℓ(w)−1 (cid:1) ln M0k k = ℓ(w). This together with Lemma 3.3 proves the claim. (cid:3) Consider Vw,t as a simple unitarizable left Oq(SO(2n))-module by restricting the action of Oq(Spin(2n)) to Oq(SO(2n)). Corollary 3.11. For w ∈ Wn and t ∈ Tn, let Vw,t be the associated simple unitarizable left Oq(SO(2n))-module. Then one has GKdim(Vw,t) = ℓ(w). 16 Proof : By Theorem 3.10 and the fact that Oq(SO(2n)) is a subalgebra of Oq(Spin(2n)), it fol- lows that GKdim(Vw,t) ≤ ℓ(w). Further observe that the variables involved in the polynomials p(w,i) j)) and hence in Oq(SO(2n)). Therefore Lemma 3.9 holds j for Oq(SO(2n))-module Vw,t and the same proof as given in Theorem 3.10 will work to show the equality. (cid:3) 's are entries of the matrix ((ui 4 Quotient spaces Fix a subset S ⊂ Π. Let Oq(G/KS ) be the quotient Hopf ∗-subalgebra of Oq(G) (see equation (4.4), [17]) . If S is the empty set φ, define Wφ = {id}. For a nonempty set S, define WS to be the subgroup of Wn generated by the simple reflections sα with α ∈ S. Let W S := {w ∈ Wn : ℓ(sαw) > ℓ(w) ∀α ∈ S}. Dijkhuizen and Stokman ([17]) classified simple unitarizable Oq(G/K S )-modules in terms of elements of W S. To avoid unnecessary complications, let us denote the action of Oq(G/KS ) arising by restricting πn w to Oq(G/KS ) and the associated module Vw by the same notations πn w and Vw respectively. Theorem 4.1. (Theorem 3.5, [16]) The set (cid:8)Vw; w ∈ W S(cid:9) is a complete set of mutually in- equivalent simple unitarizable left Oq(G/KS )-module. Let S1 be the empty subset of Π. For 2 ≤ m ≤ n, define Sm to be the set {1, 2, · · · , m − 1} if Oq(G) = Oq(SU (n + 1)) and {n − m + 2, · · · , n} if Oq(G) = Oq(SP (2n)). Theorem 4.2. Let Oq(G) be one of the Hopf ∗-algebra Oq(SU (n + 1)) or Oq(SP (2n)). For w ∈ W Sn−m+1, let Vw be the associated simple unitarizable left Oq(G/KSn−m+1 )-module. Then one has GKdim(Vw) = ℓ(w). Proof : Take finite sets M ⊂ Oq(G/KSn−m+1 ) and F ⊂ Vw. Following the same arguments as given in Lemma 3.3, one can show that lim ln dim(M rF ) ln r ≤ ℓ(w). Therefore GKdim(Vw) = supF,M lim ln dim(M rF ) ln r ≤ ℓ(w). To show the equality, observe that 1. for an element w ∈ W Sn−m+1 and 1 ≤ r ≤ n − m, the rth-part wr = ψ(ǫr) r,kr (w) is identity element of Wn. Hence w can be written uniquely as w = wn−m+1wn−m+2 · · · wn. 17 2. It follows from the definition that entries of last m rows of ((ui j)) is in the quotient algebra Oq(G/KSn−m+1 ). 3. The polynomials p(w,k) j for 1 ≤ j ≤ ℓ(wk) and n − m + 1 ≤ k ≤ n involve variables consisting of entries of last m rows of ((ui j)). With these facts, the same arguments used in Theorem 3.10 will prove the claim. (cid:3) References [1] Jose L. Bueso, F. J. Castro Jimenez and Pascual Jara. Effective computation of the Gelfand-Kirillov dimension. Proc. Edinburgh Math. Soc. (2) 40 (1997), no. 1, 111-117. [2] Jose L. Bueso, Jose Gomez Torrecillas, Francisco Javier Lobillo and F. J. Castro Jimenez. An introduction to effective calculus in quantum groups. Rings, Hopf algebras, and Brauer groups (Antwerp/Brussels, 1996),55-83, Lecture Notes in Pure and Appl. Math., 197, Dekker, New York, 1998. [3] Jose L. Bueso, Jose Gomez Torrecillas and Alain Verschoren. Algorithmic methods in non- commutative algebra. Applications to quantum groups. Mathematical Modelling: Theory and Applications, 17. Kluwer Academic Publishers, Dordrecht, 2003. [4] Partha Sarathi Chakraborty and Arup Kumar Pal. Characterization of spectral triples: A combinatorial approach. arXiv:math.OA/0305157, 2003. [5] William Fulton and Joe Harris. Representation theory, volume 129 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1991. A first course, Readings in Mathematics. [6] Izrail M. Gelfand and Alexander A. Kirillov. Sur les corps lis aux algbres enveloppantes des algbres de Lie. (French) Inst. Hautes tudes Sci. Publ. Math. No. 31, 1966, 5-19. [7] James E. Humphreys. Reflection groups and Coxeter groups. Cambridge Studies in Ad- vanced Mathematics, 29. Cambridge University Press, Cambridge, 1990. [8] Antony Joseph. Gelfand Kirillov dimension for the annihilators of simple quotients of Verma modules. J. London Math. Soc. (2) 18 (1978), no. 1, 50-60. [9] Alexander Jr. Kirillov. An introduction to Lie groups and Lie algebras. Cambridge Studies in Advanced Mathematics, 113. Cambridge University Press, Cambridge, 2008. [10] Anatoli Klimyk and Konrad Schmudgen. Quantum groups and their representations. Texts and Monographs in Physics. Springer-Verlag, Berlin, 1997. 18 [11] Leonid I. Korogodski and Yan S. Soibelman. Algebras of functions on quantum groups. Part I, volume 56 of Mathematical Surveys and Monographs. American Mathematical Society, 1998. [12] G R Krause and T H Lenagen. Growth of algebras and Gelfand-Kirillov dimension. Revised edition. Graduate Studies in Mathematics, 22. American Mathematical Society, Provi- dence, RI, 2000. [13] John Coulter McConnell. Quantum groups, filtered rings and Gelfand-Kirillov dimension. Noncommutative ring theory (Athens, OH, 1989), Lecture Notes in Math., Vol. 1448, Springer, 1990, pp. 139-149. [14] John Coulter McConnell and J. Chris Robson. Noncommutative Noetherian rings. Revised edition. Graduate Studies in Mathematics, 30. American Mathematical Society, Provi- dence, RI, 2001. [15] Bipul Saurabh. Quantum quaternion spheres. Proc. Indian Acad. Sci. Math. Sci. 127 (2017), no. 1, 133-164. [16] Jasper V. Stokman. The quantum orbit method for generalized flag manifolds. Math. Res. Lett. 10 (2003), no. 4, 469-481. [17] Jasper V. Stokman and Mathijs S. Dijkhuizen. Quantized flag manifolds and irreducible -representations. Comm. Math. Phys. 203 (1999), no. 2, 297-324. [18] Jose Gomez Torrecillas. Gelfand-Kirillov dimension of multi-filtered algebras. Proc. Ed- inburgh Math. Soc. (2) 42 (1999), no. 1, 155-168. [19] S. L. Woronowicz. Compact quantum groups. Symetries quantiques, 845-884, 1998. Partha Sarathi Chakraborty ([email protected]) Institute of Mathematical Sciences (HBNI), CIT Campus, Taramani, Chennai, 600113, INDIA Bipul Saurabh ([email protected]) Institute of Mathematical Sciences (HBNI), CIT Campus, Taramani, Chennai, 600113, INDIA 19
1706.00563	1	1706	2017-06-02T06:02:01	Graded C-algebras, graded K-theory, and twisted P-graph C-algebras	[ "math.OA", "math.KT" ]	We develop methods for computing graded K-theory of C-algebras as defined in terms of Kasparov theory. We establish graded versions of Pimsner's six-term sequences for graded Hilbert bimodules whose left action is injective and by compacts, and a graded Pimsner-Voiculescu sequence. We introduce the notion of a twisted P-graph C-algebra and establish connections with graded C-algebras. Specifically, we show how a functor from a P-graph into the group of order two determines a grading of the associated C-algebra. We apply our graded version of Pimsner's exact sequence to compute the graded K-theory of a graph C*-algebra carrying such a grading.	math.OA	math	GRADED C∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C∗-ALGEBRAS ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS Abstract. We develop methods for computing graded K-theory of C ∗-algebras as de- fined in terms of Kasparov theory. We establish graded versions of Pimsner's six-term sequences for graded Hilbert bimodules whose left action is injective and by compacts, and a graded Pimsner–Voiculescu sequence. We introduce the notion of a twisted P - graph C ∗-algebra and establish connections with graded C ∗-algebras. Specifically, we show how a functor from a P -graph into the group of order two determines a grading of the associated C ∗-algebra. We apply our graded version of Pimsner's exact sequence to compute the graded K-theory of a graph C ∗-algebra carrying such a grading. 1. Introduction This paper has two objectives. The first is to develop techniques for computing graded K-theory of C∗-algebras as defined in terms of Kasparov theory, with a view to expanding on Haag's computation of graded K-theory of Cuntz algebras [13, 14]. The second is to introduce twisted P -graph algebras, generalising [3, 4, 33], and use them to study connections between Z2-gradings of C∗-algebras, and the twisted k-graph C∗-algebras studied in [21, 22]. The idea is that Z2-valued functors on P -graphs determine gradings of the associated C∗-algebras. The twisted C∗-algebras associated to {−1, 1}-valued 2- cocycles on cartesian products of P -graphs can then be used to model graded tensor products of graded C∗-algebras. We begin by discussing graded K-theory for C∗-algebras. Although K-theory for graded Banach algebras has been extensively studied by Karoubi (see, for example, [16]), the modern literature on complex graded C∗-algebras essentially begins with the work of Kasparov [17, 18] on KK-theory. Various definitions of graded K-theory for graded C∗- algebras have been used in the literature (see [5, 6, 12, 13, 14, 35, 39] to name but a few). We take as our definition of K gr 0 (A) the Kasparov group KK(C, A) for the graded C∗-algebra A, and likewise define K gr complex Clifford algebra. We establish that perturbing the grading of a C∗-algebra A by conjugation by an odd self-adjoint unitary in M(A) does not alter the graded K-theory of 1 (A) := KK(C, Ab⊗ Cliff 1) where Cliff 1 is the first Date: June 5, 2017. 2010 Mathematics Subject Classification. Primary 46L05, 19K35. Key words and phrases. KK-theory; graded K-theory; C ∗-algebra; P -graph; twisted C ∗-algebra; graded C ∗-algebra. This research was supported by the Australian Research Council, grant DP150101598. Part of this work was carried out while the third-named author was at the Centre de Recerca Mathematica, Universitat Aut´onoma de Barcelona as part of the Intensive Research Program Operator algebras: dynamics and interactions in 2017. The first author would like to thank his coauthors for their hospitality and support while visiting the University of Wollongong where much of this work was done. He would also like to acknowledge support from Simons Foundation Collaboration grant #353626. 1 2 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS A. In particular, we show that the graded K-theory of the crossed product of a C∗-algebra A by its grading automorphism is identical to the graded K-theory of Ab⊗ Cliff 1. To help compute graded K-theory in examples, we revisit the work of Pimsner in [29] to show that his six-term sequences in KK-theory for Cuntz–Pimsner algebras are also valid, with suitable modifications, for graded C∗-algebras. Unlike Pimsner, we restrict to Hilbert bimodules in which the left action is both compact and injective. We obtain a 6-term exact sequence in graded K-theory, which in turn gives a Pimsner–Voiculescu sequence for graded crossed products by Z. We next develop substantial classes of graded C∗-algebras to which we can apply our theorems. We introduce twisted P -graph C∗-algebras by straightforward generalisation of the notion of a twisted k-graph algebra. We establish a number of fundamental structure results for these C∗-algebras, including a version of the gauge-invariant uniqueness theo- rem, to help us make identifications between these C∗-algebras and key examples later in the paper. We prove that if P has the form Nk × F where F is a countable abelian group, then every P -graph is a crossed product of a k-graph by an action of the group F , in a sense analogous to that studied in [9]. We next discuss how a functor from a P -graph to Z2 induces gradings of the associated twisted C∗-algebras. We show that if Z2 denotes a copy of the order-two group Z2, regarded as a Z2-graph with one vertex, then C∗(Z2), under the grading induced by the degree functor, is isomorphic to Cliff 1. More generally, we consider the situation where P = Nk × Zl 2. Any P -graph Λ carries both a natural functor δΛ taking values in Z2, and a natural {−1, 1}-valued 2-cocycle cΛ. We establish a universal description of the graded twisted C∗-algebra determined by this functor and cocycle. These threads come together when we study graded tensor products in terms of carte- sian products of P -graphs. We prove that if Λ is a P -graph, Γ is a Q-graph, and we consider the associated graded, twisted C∗-algebras for the functors and cocycles de- scribed in the preceding paragraph, then the graded tensor product is isomorphic to the graded twisted C∗-algebra of the (P × Q)-graph Λ × Γ under its own natural cocycle and functor. Combining this with the results of previous sections, we show that the higher complex Clifford algebras can be realised as graded twisted P -graph algebras for appro- priate P , and also that graded tensor products of graded twisted P -graph algebras with Cliff 1 can be realised as graded twisted (P × Z2)-graph algebras. We apply our graded Pimsner sequence to the C∗-algebras of row-finite 1-graphs E with no sources under gradings of the sort discussed above. The result is an elegant generalisation of the well-known formula for the K-theory of a directed graph: if Aδ denotes the E0 × E0 matrix with Aδ(v, w) =Pe∈vE1w(−1)δ(e), then the graded K-groups of C∗(E) are the cokernel and kernel of 1 − At δ. This recovers Haag's formulas for the graded K-theory of Cuntz algebras [14]. If δ(e) = 1 for all e (this corresponds to the grading of C∗(E) coming from the order-two element of the gauge action), then Aδ is the negative of the usual adjacency matrix AE, and so K gr (C∗(E)) is given by the cokernel ∗ and kernel of 1+At E. We also apply our results to compute the graded K-theory of certain crossed-products of graph algebras by Z2. Our examples and results lead us to conjecture that the graded K0-group of a C∗-algebra can be described along the lines of the standard picture of ungraded K0, as a group generated by equivalence classes of graded projective modules. GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 3 We begin by collecting relevant background in Section 2. In Section 3 we introduce graded K-theory in terms of Kasparov theory, and establish some fundamental results about it. In Section 4 we establish graded versions of Pimsner's six-term sequences for Hilbert bimodules with injective left actions by compacts (see Theorem 4.4) and apply them to obtain a graded Pimsner–Voiculescu sequence for grossed products by Z (Corol- lary 4.6). In Section 5 we introduce twisted P -graph C∗-algebras and establish the basic structure theory for them that we will need later in the paper. In Section 6, we discuss gradings of P -graph C∗-algebras induced by functors on the underlying P -graphs. In Section 7, we establish our main results about graded tensor products of graded P -graph algebras: Theorem 7.1 shows that for appropriate gradings and twisting cocycles, we have C∗(Λ, cΛ)b⊗ C∗(Γ, cΓ) ∼= C∗(Λ× Γ, cΛ×Γ). In section 8, we apply our results from Section 4 to calculate graded K-theory for graph C∗-algebras (Lemma 8.2). We apply this lemma and our graded Pimsner–Voiculescu sequence to some illustrative examples. We conclude in Section 9 by formulating our conjecture about the structure of the graded K0 group. 2. Background Notation. We will denote the cyclic group of order n by Zn. We frequently regard Z2 = {0, 1} as a ring, so we always use additive notation for the group operation, and make any identification of Z2 with {−1, 1} ⊆ T explicit. We typically denote the multiplication operation in the ring Z2 by ·. Graded C∗-algebras. Let A be a C∗-algebra. A grading of A is an automorphism α of A such that α2 = 1 (α is sometimes referred to as the grading automorphism). We define A0 := {a ∈ A : α(a) = a} and A1 := {a ∈ A : α(a) = −a}. (2.1) linear self-adjoint subspaces of A and satisfy AiAj ⊂ Ai+j for So A0, A1 are closed, i, j ∈ Z2. We have A = A0 ⊕ A1 as a Banach space. Note that A0 is a C∗-subalgebra of A. Elements of A0 are called even and elements of A1 are called odd. To calculate A0 and A1, it is helpful to note that A0 =(cid:8) a+α(a) 2 : a ∈ A(cid:9) and A1 =(cid:8) a−α(a) 2 : a ∈ A(cid:9). If a ∈ Ai then we say that a is homogeneous of degree i and write ∂a = i. If α is the identity map on A then A0 = A and A1 = {0}. The resulting grading is called the trivial grading. Since a C∗-algebra may admit several different gradings (we discuss explicit examples of this in Examples 8.5 below), we shall frequently write a C∗-algebra A with grading α as the pair (A, α). A graded C∗-algebra is inner-graded if there exists a self-adjoint unitary (called a grading operator) U ∈ M(A) such that α(a) = UaU for all a ∈ A. In [2, §14.1] Blackadar calls an inner-grading even. A graded homomorphism π : (A, α) → (B, β) between graded C∗-algebras is a homomorphism from A to B which intertwines the gradings (i.e. π ◦ α = β ◦ π). Given a graded C∗-algebra (A, αA) and homogeneous elements a, b ∈ A, the graded commutator [a, b]gr is defined as [a, b]gr = ab − (−1)∂a·∂bba. This formula extends to arbitrary a and b by bilinearity. In particular, if a ∈ A1, then (2.2) If A is trivially graded, then [a, b]gr reduces to the usual commutator [a, b] = ab − ba. [a, b]gr = ab − αA(b)a. 4 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS There is a graded tensor product operation for graded C∗-algebras, defined as follows. Let (A, α), (B, β) be graded C∗-algebras, and A ⊙ B be their algebraic tensor product. This becomes a ∗-algebra when endowed with multiplication and involution given by and (ab⊗ b)∗ = (−1)∂a·∂ba∗ b⊗ b∗. (ab⊗ b)(a′ b⊗ b′) = (−1)∂b·∂a′ aa′ b⊗ bb′ and if a and b are both homogeneous. We have for homogeneous elements a, a′ ∈ A and b, b′ ∈ B. We decorate the ⊙ symbol with a hat, norm. (If either A or B is nuclear then this agrees with the maximal norm.) It is straightforward to show that the graded tensor product operation is associative It may aid intuition to observe that for unital graded C∗-algebras A, B, under this (Ab⊗ B)1 = A0b⊗ B1 + A1b⊗ B0. graded C∗-algebras. For this and other basic facts about graded C∗-algebras we refer the reader to [2, §14]. b⊙ to indicate that we are using this ∗-algebra structure on the algebraic tensor product. We write Ab⊗ B for the completion of the ∗-algebra Ab⊙B in the minimal tensor-product The grading automorphism of Ab⊗ B is αb⊗β. So ab⊗ b is homogeneous of degree ∂a+ ∂b (Ab⊗ B)0 = A0 b⊗ B0 + A1 b⊗ B1 (modulo the natural isomorphism (ab⊗ b)b⊗ c 7→ ab⊗ (bb⊗ c)). We have Ab⊗ B ∼= Bb⊗ A as grading we have [ab⊗ 1B, 1A b⊗ b]gr = 0 = [1A b⊗ b, ab⊗ 1B]gr for all a ∈ A, and b ∈ B by the grading operator U ∈ M2n(C) given by Ui,j = (−1)iδij. We often write cM2n(C) to An example of a graded C∗-algebra that we shall use very frequently is M2n(C) with grading automorphism α(θi,j) = (−1)i−jθi,j. This is an inner grading, as it is implemented emphasise that we are using this grading. Clifford algebras over C. Following [2, Examples 14.1.2(b)] the C∗-algebra A = C⊕ C has a grading automorphism α given by α(z, w) = (w, z). So (C ⊕ C)0 = {(z, z) : z ∈ C} and (C ⊕ C)1 = {(z,−z) : z ∈ C}. This graded C∗-algebra is called the first (complex) Clifford algebra [17, Section 2], and we denote it by Cliff 1 with this grading α implicit. As a graded C∗-algebra Cliff 1 is generated by the odd self-adjoint unitary u = (1,−1), because, for (z, w) ∈ Cliff 1, we can write (z, w) = (z + w, z + w) + 1 2 1 2 (z − w, w − z) = 1 2 (z + w)1 + 1 2 (z − w)u. Note that Cliff 1 is not inner-graded (because it is abelian), and is isomorphic to the group C∗-algebra C∗(Z2) with grading given by the dual action ofcZ2 ∼= Z2. The higher complex Clifford algebras are defined inductively: Cliff n+1 = Cliff nb⊗ Cliff 1 for n ≥ 1. It is straightforward to show that Cliff 2 ∼= cM2(C) as graded C∗-algebras (see also Example 6.3(i)). Observe that Cliff n is generated by n mutually anticommuting odd self-adjoint unitaries. For other basic facts about complex Clifford C∗-algebras we refer the reader to [2, §14]. Graded Hilbert modules. Let B be a graded C∗-algebra. A graded (right) Hilbert B-module is a Hilbert B-module X together with a decomposition of X as a direct sum of two closed subspaces X0 and X1 compatible with the grading of B in the sense that XiBj ⊂ Xi+j and hXi, Xji ⊂ Bi+j (the graded components Xi need not be Hilbert submodules). We define the grading operator αX on X on homogeneous elements by αX (x) = (−1)jx if x ∈ Xj. This αX is not necessarily an adjointable operator. Given a graded Hilbert B-module X we write X op for the same Hilbert B-module with the grading GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 5 components switched (so αX op = −αX ). The grading operator on X induces a grading αX on L(X) given by (2.3) αX(T ) = αX ◦ T ◦ αX for all T ∈ L(X). Under this induced grading, T is homogeneous of degree j if and only if T Xk ⊆ Xj+k for j, k ∈ Z2. For ξ, η ∈ X we write θξ,η for the generalised compact operator θξ,η(ζ) = ξ · hη, ζiB. The grading α of L(X) restricts to a grading of K(X) satisfying αX (θξ,η) = θαX (ξ),αX (η). Naturally B may be regarded as a graded Hilbert B-module BB with inner product ha, biB = a∗b, right action given by multiplication, and grading operator αB. We write HB for the graded Hilbert B-module obtained as the direct sum of countably many copies of BB. We define bHB := HB ⊕Hop B . If X is a graded Hilbert B-module, then X ⊕ bHB ∼= bHB by Kasparov's Stabilization Theorem (see [2, Theorem 14.6.1]). A graded Hilbert B-module which is a finitely generated projective module will be called a graded projective B-module throughout the paper. If X is a graded projective B-module, then K(X) = L(X) and so in particular 1X ∈ K(X). Kasparov's Stabilization Theorem implies that every graded projective B-module X is isomorphic to pbHB for some even projection p ∈ K(bHB). Moreover pbHB is a graded projective B-module for any such projection. Given even projections p, q ∈ K(bHB), we have pbHB ∼= qbHB if and only if there is an even partial isometry v ∈ K(bHB) such that p = v∗v and q = vv∗. C∗-correspondences and Cuntz–Pimsner algebras. We briefly recap the notion of a C∗-correspondence and of the associated Cuntz–Pimsner algebra and its Toeplitz exten- sion. For a detailed introduction to C∗-correspondences, see [24, 32]. For more background on Cuntz–Pimsner algebras, see [29] and [30, §8]. Given separable C∗-algebras A and B, an A–B-correspondence X is a pair (φ, X) consisting of a right-Hilbert B-module X and a homomorphism φ : A → L(X). We regard φ as implementing a left action of A on X by adjointable operators, so we often write φ(a)x = a · x for a ∈ A and x ∈ X. We say that X is full if span{hξ, ηiA : ξ, η ∈ X} = A. Any right-Hilbert B-module X may be regarded as C–A correspondence and we write ℓ for the canonical left action of C by scalar multiplication. We say that X is countably generated if there is a sequence (xi)∞i=1 in X such that X = span{xi · b : i ≥ 1, b ∈ B}. Note that BB is a countably generated correspondence if B is σ-unital, so in particular if B is separable. If B is a C∗-algebra then there is an isomorphism of M(B) onto L(BB) that carries a multiplier m to the operator of left-multiplication by m on A. So any homomorphism φ : A → M(B) determines an A–B-correspondence structure on BB. We denote this correspondence by φB. The isomorphism of M(B) onto L(BB) carries B onto K(BB), so the left action of A on φB is by compacts if and only if φ takes values in B. If (φ, X) is an A–B-correspondence and (ψ, Y ) is a B–C-correspondence, then the inter- nal tensor product X ⊗ψ Y is formed as follows: define [·,·]C on the algebraic tensor prod- uct X ⊙ Y by sesquilinear extension of the formula [x⊙ y, x′ ⊙ y′]C := hy, ψ(hx, x′iB)y′iC. Let N = {ξ ∈ X ⊙ Y : [ξ, ξ]C = 0}. Then X ⊗ψ Y is defined to be the completion of (X ⊙ Y )/N in the norm determined by the inner-product hξ + N, η + NiC = [ξ, η]C. For x ∈ X and y ∈ Y , we write x⊗y for (x⊙y)+N ∈ X⊗ψ Y . There is then a homomorphism 6 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS ψ : L(X) → L(X ⊗ψ Y ) given by (2.4) In particular, ψ ◦ φ is a homomorphism of A into L(X ⊗ψ Y ), making X ⊗ψ Y into an A–C-correspondence with for all x ∈ X, y ∈ Y and T ∈ L(X). ψ(T )(x ⊗ y) = (T x) ⊗ y a · (x ⊗ y) = ψ(φ(a))(x ⊗ y) = (a · x) ⊗ y. When the actions φ, ψ are clear from context, we frequently write X ⊗B Y instead of X ⊗ψ Y . If φ : A → B and ψ : B → C are homomorphisms and ψ is nondegenerate, then φB ⊗B ψC ∼= ψ◦φC under an isomorphism taking b ⊗ c to ψ(b)c. If X is an A–A correspondence, then we can form its tensor powers X⊗n given by X⊗0 := A, X⊗1 := X and X⊗(n+1) := X ⊗A X⊗n. The Fock space of X is the completion FX of the algebraic direct sum L∞n=0 X⊗n in the norm coming from the inner product h⊕nxn,⊕nyniA =Pnhxn, yniA. This FX is a C∗-correspondence over A with respect to and such that L1(ξ)ρ = ξ ⊗ ρ for ρ ∈ Sn≥1 X⊗n and L1(ξ)a = ξ · a for a ∈ X⊗0 = A. the pointwise actions. Observe that FX is full even if X is not, but that the left action of A on FX is not by compacts even if the action of A on X is. A representation of X in a C∗-algebra B is a pair (ψ, π) where ψ : X → B is a linear map, π : A → B is a homomorphism, and we have ψ(a·ξ) = π(a)ψ(ξ), ψ(ξ·a) = ψ(x)π(a) and π(hξ, ηiA) = ψ(ξ)ψ(η)∗ for all ξ, η ∈ X and a ∈ A. There is a universal C∗-algebra TX , called the Toeplitz algebra of X, generated by a representation (iX , iA) of X. There is also a representation (L0, L1) of X in L(FX) such that L0(a)ρ = a·ρ for a ∈ A and ξ ∈ FX The universal property of TX gives a homomorphism L1 × L0 : TX → L(FX) satisfying (L1 × L0) ◦ iX = L1 and (L1 × L0) ◦ iA = L0. Pimsner proves [29, Proposition 3.3] that L1 × L0 is injective. To describe the Cuntz–Pimsner algebra of X, we will restrict attention to the situation where the left action of A on X is injective and by compacts. As discussed on [29, Page 202], given a representation (ψ, π) of X in B there is a homomorphism ψ(1) : K(X) → B such that ψ(1)(θξ,η) = ψ(ξ)ψ(η)∗ for all ξ, η ∈ X. We say that the representation (ψ, π) is Cuntz–Pimsner covariant, or just covariant, if ψ(1)(φ(a)) = π(a) for all a ∈ A. The Cuntz–Pimsner algebra OX of X is the universal C∗-algebra generated by a covariant representation (jX , jA) of X; so it coincides with the quotient of TX by the ideal generated by elements of the form iA(a) − i(1) X (φ(a)), a ∈ A. Under our hypotheses, K(FX) ⊆ TX 1 (φ(a)) : a ∈ A}, and so OX ∼= TX/K(FX ). and is generated as an ideal by {L0(a) − L(1) If A is a C∗-algebra and α is an automorphism of A, then there is an isomorphism of the Cuntz–Pimsner algebra of the A–A correspondence X := αA onto the crossed product A ×α Z that intertwines iA : A → OX with the canonical inclusion ι : A ֒→ A ×α Z, and carries iX(a) ∈ OX to Uι(a), where U ∈ A×α Z is the unitary generator of the copy of Z. There is a corresponding isomorphism of TX onto the natural Toeplitz extension of the crossed product (that is, Stacey's endomorphism crossed-product of A by α [40]). Elements of Kasparov theory. We introduce the elements of Kasparov theory needed for our work on graded K-theory later. For more background, see [2]. Let A, B be C∗-algebras, and let X be an A–B-correspondence. Given gradings αA of A and αB of B, a grading operator on X is a map αX : X → X such that α2 X = 1, αX (a · x · b) = αA(a) · αX(a) · αB(b) for all a, x, b, and αB(hx, yiB) = hαX(x), αX(y)iB for GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 7 all x, y. We call the pair (X, αX) (or just X if the grading operator is understood from context) a graded A–B-correspondence (see, for example, [15]). Given graded C∗-algebras A, B, C, a graded A–B-correspondence (X, αX) and a graded homomorphism of C∗-algebras, then φB is a graded Hilbert module. B–C-correspondence (Y, αY ), there is a well-defined grading operator αXb⊗ αY on Xb⊗B Y characterised by (αX b⊗ αY )(xb⊗ y) = αX (x)b⊗ αY (y); note that if φ : A → B is a graded If (A, αA) and (B, αB) are separable graded C∗-algebras, then a Kasparov A–B-module is a quadruple (X, φ, F, αX) where (φ, X) is a countably generated A–B-correspondence, αX is a grading operator on X, and F ∈ L(X) is odd with respect to the grading αX described at (2.3) in the sense that F ◦ αX = −αX ◦ F and satisfies (F − F ∗)φ(a) ∈ K(X), (F 2 − 1)φ(a) ∈ K(X), and [F, φ(a)]gr ∈ K(X) for all a ∈ A. Observe that since F is odd graded, we have [φ(a), F ]gr = φ(a)F − F φ(αA(A)) by (2.2). We say that (X, φ, F, αX) is a degenerate Kasparov module if (F 2 − 1)φ(a) = 0, [F, φ(a)]gr = 0 for all a ∈ A. (F − F ∗)φ(a) = 0, and We say that graded Kasparov modules (X, φ, F, αX) and (Y, ψ, G, αY ) are unitarily equivalent if there is a unitary U ∈ L(X, Y ) of degree zero (in the sense that αY U = UαX ) such that UF = GU and Uφ(a) = ψ(a)U for all a ∈ A. In what follows, C([0, 1]) always has the trivial grading. Fix a C∗-algebra B, and for each t ∈ [0, 1], define ǫt : C([0, 1])b⊗ B → B by ǫt(fb⊗ b) = f (t)b. A homotopy of Kasparov (A, αA)–(B, αB)-modules from (X0, φ0, F0, αX0) to (X1, φ1, F1, αX1) is a Kasparov A–(Bb⊗ C([0, 1]))-module (X, φ, F, αX) such that, for t ∈ {0, 1}, and with ǫt : L(X) → L(Xb⊗BBB) as described in (2.4), the module (X b⊗ǫt BBB, ǫt ◦ φ, ǫt(F ), αX b⊗ αB) is unitarily equivalent to (Xt, φt, Ft, αXt). We write (X0, φ0, F0, αX0) ∼h (X1, φ1, F1, αX1) and say that these two Kasparov modules are homotopy equivalent if there exists a ho- motopy from (X0, φ0, F0, αX0) to (X1, φ1, F1, αX1). We write KK(A, B) for the collection of all homotopy-equivalence classes of Kasparov A–B-modules. This KK(A, B) forms an abelian group with addition given by direct sum: [X, φ, F, αX] + [Y, ψ, G, αY ] = [X ⊕ Y, φ ⊕ ψ, F ⊕ G, αX ⊕ αY ], and identity element equal to the class of the trivial module [BB, 0, 0, idB]; this class coincides with the class of any degenerate Kasparov A–B-module. As detailed in the proof of [2, Proposition 17.3.3], the (additive) inverse of a class in KK(A, B) is given by (2.5) − [X, φ, F, αX] = [X, φ ◦ αA,−F,−αX]. Let (A, αA) and (B, αB) be graded C∗-algebras, let (φ, X) be an A–B-correspondence, If G, H ∈ L(X) are operators for and suppose that αX is a grading operator on X. which (X, φ, G, αX) and (X, φ, H, αX) are both Kasparov A–B-modules, then an operator homotopy between these Kasparov modules is a norm-continuous map t 7→ Ft from [0, 1] to L(X) such that (X, φ, Ft, αX) is a Kasparov module for each t, F0 = G and F1 = H. An operator homotopy is a special case of a homotopy in the following sense: the space X := C([0, 1], X) is a graded A–C([0, 1], B)-correspondence with left action given by 8 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS an operator F ∈ L(C([0, 1], X)) given by F (x)(t) = Ft(x(t)), and then (X, φ, F , αX) is a ical way, we see that (X, φ, F , αX) is a homotopy from (X, φ, G, αX) to (X, φ, H, αX). (cid:0)φ(a)(x)(cid:1)(t) = φ(a)x(t) and grading operator(cid:0)αX (x)(cid:1)(t) = αX(x(t)). Moreover, there is Kasparov A–C([0, 1], B)-module. Identifying C([0, 1], B) with Bb⊗ C([0, 1]) in the canon- in this category is called the Kasparov product, denoted b⊗B : KK(A, B) × KK(B, C) → KK(A, C). The identity morphism for the object (A, αA) is the class of the Kasparov module (AA, id, 0, αA). Given a Kasparov A–B-module (X, φ, F, αX) and a Kasparov B–C-module (Y, ψ, G, αY ), the Kasparov product There is a category whose objects are separable graded C∗-algebras and whose mor- phisms from A to B are homotopy classes of Kasparov A–B-modules. The composition has the form [X, φ, F, αX]b⊗B [Y, ψ, G, αY ] [X b⊗ψ Y, φ, H, αX b⊗ αY ] for an appropriate choice of operator H; the details are formidable in general, but we will not need them here. For us it will suffice to consider Kasparov products in which one of the factors has the form [BB, φ, 0, αB] for some graded homomorphism φ : (A, αA) → (B, αB) of C∗-algebras. In detail, suppose that φ : (A, αA) → (B, αB) is a graded homomorphism of graded C∗- algebras. Since B ∼= K(BB) via the map b 7→ (a 7→ ba), the quadruple (BB, φ, 0, αB) is a Kasparov A–B-module. If (X, ψ, F, αX) is a Kasparov B–C-module, then (X, ψ◦φ, F, αX) is also a Kasparov module, whose class in KK(A, C) we denote by φ∗[X, ψ, F, αX]. Propo- sition 18.7.2(b) of [2] shows that [BB, φ, 0, αB]b⊗B [X, ψ, F, αX] = φ∗[X, ψ, F, αX]. Likewise, if (Y, ψ, G, αY ) is a Kasparov C–A-module, then (Y b⊗φ BB, ψb⊗1, Gb⊗1, αYb⊗αB) is a Kasparov C–B-module whose class we denote by φ∗[Y, ψ, G, αY ], and [2, Proposi- tion 18.7.2(b)] shows that Observe that if (A, α) is a graded C∗-algebra, then the discussion above shows that KK(A, A) is a ring under the Kasparov product, with multiplicative identity [Y, ψ, G, αY ]b⊗A [BB, φ, 0, αB] = φ∗[Y, ψ, G, αY ]. By definition of the additive inverse (see (2.5)), the tuple (A, αA, 0,−αA) is a Kasparov module, and [idA] := [A, idA, 0, αA]. (2.6) [A, αA, 0,−αA] = −[idA] 3. Graded K-theory of C∗-algebras In this section, we consider graded K-theory for C∗-algebras. There does not appear to be a universally-accepted definition of graded K-theory for C∗-algebras in the literature to date. We have chosen to take Kasparov's KK-theory as the basis for our definition (see [2, Definition 17.3.1]). We establish some basic properties of graded K-theory; in particular, that both taking graded tensor products with Cliff 1, and taking crossed products by Z2 with respect to a suitable grading, interchange graded K-groups. The following definition is used implicitly in [13, 14]. GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 9 i (A, α). i (A) for K gr satisfy Bott periodicity (see Remark 3.3 below). So we write K gr Definition 3.1. Let A be a separable C∗-algebra and let α be a grading automorphism of A. We define the graded K-theory of A as follows: K gr 0 (A, α) := KK(C, A) and K gr 1 (A, α) := KK(C, A b⊗ Cliff 1). When α is understood from context we often write K gr Remark 3.2. From the above definition and results from Kasparov theory, we see that K gr j is covariantly functorial, natural, continuous with respect to direct limits, and invariant under Morita equivalence. For j ≥ 2 we define K gr j (A) = KK(C, A b⊗ Cliff j). The functors K gr j (K gr 0 (A), K gr 1 (A)) as we do for ungraded K-theory. Up to isomorphism Definition 3.1 agrees with the usual definition of K-theory if A is σ-unital and trivially graded (see [2, 18.5.4]). Furthermore (see [2, 14.5.1 and 14.5.2]), if A is inner graded, then K gr Remark 3.3. By definition K gr have K gr ∗ (A) ∼= K∗(A). 0 (A b⊗ Cliff 1) = KK(C, A b⊗ Cliff 1) = K gr KK(C, Ab⊗ cM2(C)) ∼= KK(C, A). Using this at the last step, we calculate 1 (Ab⊗ Cliff 1) ∼= K gr 1 (Ab⊗ Cliff 1) = KK(C, Ab⊗ Cliff 1b⊗ Cliff 1) ∼= KK(C, Ab⊗ cM2(C)) = K gr 1 (A). We also 0 (A). To see this, recall that by [2, Corollary 17.8.8], we have ∗ (A) = as claimed. 0 (A) K gr Example 3.4. Since C is trivially graded we have K gr ∗ mark 3.3 we have K gr Hence putting A = C we have i (Ab⊗Cliff 1) = K gr K gr i (Cliff1) = K gr (C) = K∗(C) = (Z, 0). From Re- i+1(A), and it is easy to show that Cb⊗Cliff 1 ∼= Cliff 1. i+1(C) = Ki+1(C) =(0 if i = 0 Z if i = 1. Since Cliff 1b⊗ Cliff 1 = Cliff 2, the preceding paragraph applied with A = Cliff 1 gives i (Cliff 2) ∼= Ki(C). Repeating this procedure we find that K gr K gr i (Cliff n) ∼= Z if i + n is even and it is trivial otherwise. K gr Before moving on to some tools for computing graded K-theory, it is helpful to relate it to our intuition for ordinary K-theory. We think of K0(A) as a group generated by equivalence classes of projections in A ⊗ K so that, in particular, [v∗v] = [vv∗] whenever v is a partial isometry. The following example indicates that in graded K-theory similar relations hold for homogeneous partial isometries in graded C∗-algebras, but with an additional dependence on the parity of the partial isometry in question. We discuss this further in Section 9 i (Cliff n) ∼= Ki+n(C). So Example 3.5. Let A be a graded C∗-algebra with grading automorphism α and suppose follows: let p = v∗v and q = vv∗. By the usual abuse of notation we let α denote the . Recall that ℓ denotes the action of C by scalar multiplication on any Hilbert module. We ). We claim that that v is an odd partial isometry in K(bHA). We obtain graded Kasparov modules as grading operator on bHA and observe that if p ∈ K(bHA) is a homogeneous projection then α restricts to a grading operator on the graded submodule pbHA ⊆ HA denoted αp bHA can form the Kasparov modules (ℓ, pbHA, 0, αp bHA the classes [pbHA]K and [qbHA]K of these Kasparov modules satisfy [pbHA]K = −[qbHA]K in KK(C, A). These are Kasparov modules because K(pbHA) is isomorphic to pK(bHA)p ) and (ℓ, qbHA, 0, αq bHA 10 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS which is unital with unit p, and so all adjointables are compact. In particular the zero operator F = 0 trivially has the property that F 2 − 1, F ∗ − F and [F, a]gr are compact for all a ∈ C. To see that [pbHA]K = −[qbHA]K, observe that Ad v implements an isomorphism pbHA → qbHA,. This isomorphism is odd in the sense that αq bHA = Ad v◦ (−αp bHA ). We claim that (3.1) is operator homotopic to a degenerate Kasparov module. To see this, note that the module (cid:16)pbHA ⊕ qbHA, ℓ, 0, αp bHA ⊕ αq bHA(cid:17) (cid:16)pbHA ⊕ qbHA, ℓ,(cid:16) 0 v∗ v 0(cid:17), αp bHA ⊕ αq bHA(cid:17) (3.2) is a degenerate Kasparov module because Fv :=(cid:0) 0 v∗ v 0(cid:1)2 = id. Since the straight-line path from 0 to Fv implements an operator homotopy from (3.1) to (3.2), we conclude that (3.1) represents 0KK(C,A) as required. Theorem 3.6. With notation as above there is an isomorphism of graded C∗-algebras We now discuss how crossed products by Z2 relate to graded K-theory. Let A be a graded C∗-algebra with grading automorphism α. Let v be an odd self-adjoint unitary in M(A) and define α = α ◦ Ad v. Since α commutes with Ad v, this makes A bi-graded in the sense that it admits two commuting gradings by Z2, or equivalently a grading by Z2 × Z2. Let Ajk denote the bihomogeneous elements of degree j with respect to α and of degree k with respect to Ad v; that is, a ∈ Ajk if and only if α(a) = (−1)ja and Ad v(a) = (−1)ka. So if a ∈ Ajk, then α(a) = (−1)j+ka. Let A denote the C∗-algebra A graded by α. In the following statement and proof, indices in Z2 are denoted j, k, l, and i is reserved for the imaginary number i = √−1. φ : Ab⊗ Cliff 1 → Ab⊗ Cliff 1 such that Proof. First, we check that φ preserves the grading: The element ajkb⊗ ul is (j + l)-graded in Ab⊗ Cliff 1 (with A graded by α). The element ajk(iv)k b⊗ ul is homogeneous of degree (j + k + k) + l = j + l in A b⊗ Cliff 1 with A graded by α. In A b⊗ Cliff 1 we compute (ajk b⊗ ul)(a′j ′k′ b⊗ ul′) = (−1)lj ′ajka′j ′k′ b⊗ ul+l′. Applying φ to both sides, and computing in Ab⊗ Cliff 1 (with A graded by α) we obtain Since va′j ′k′v = (−1)k′a′j ′k′, we have (iv)ka′j ′k′(iv)k′ = (−1)k·k′a′j ′k′(iv)k(iv)k′. Since k, k′ belong to the ring Z2, we have ikik′ = (−1)k·k′ik+k′. So the factors of (−1)k·k′ cancel, and we obtain ) = (ajk(iv)k b⊗ ul)(a′j ′k′(iv)k′ b⊗ ul′ = (−1)l(j ′+k′+k′)ajk(iv)ka′j ′k′(iv)k′ b⊗ ul+l′ where ajk ∈ Ajk for j, k, l ∈ Z2, and u is the odd generator of Cliff 1. φ(ajk b⊗ ul) = ajk(iv)k b⊗ ul, φ(ajk b⊗ ul)φ(a′j ′k′ b⊗ ul′ ) . φ(ajk b⊗ ul)φ(a′j ′k′ b⊗ ul′ ajka′j ′k′(iv)k+k′ b⊗ ul+l′ ) = (−1)lj ′ = φ(cid:0)(−1)lj ′ ajka′j ′k′ b⊗ ul+l′(cid:1) )(cid:1). = φ(cid:0)(ajk b⊗ ul)(a′j ′k′ b⊗ ul′ GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS The result follows because A is spanned by its bihomogeneous elements. 11 (cid:3) ∗ ( A). (A) ∼= K gr ∗ inner grading (so diagonal elements are even and off diagonal elements are odd). Similarly, Recall that cM2(C) denotes the algebra of 2 × 2 complex matrices with the standard let bK denote the C∗-algebra of compact operators with the standard inner grading. We define cM2(A) := Ab⊗cM2(C) ∼= Ab⊗ Cliff 2 ∼= Ab⊗ Cliff 1b⊗ Cliff 1. Corollary 3.7. Continuing with the notation of Theorem 3.6, the isomorphism φ : Ab⊗ Cliff 1 → Ab⊗ Cliff 1 induces an isomorphism cM2(A) ∼= cM2( A), which in turn induces an isomorphism Ab⊗ bK ∼= Ab⊗ bK. In particular, K gr Proof. By Theorem 3.6, we have Ab⊗ Cliff 1 ∼= Ab⊗ Cliff 1. Hence, the canonical isomorphism bK ∼= Kb⊗cM2(C). The final statement follows from the stability we have Bb⊗ Cliff 1 ∼= B ⋊β Z2 as C∗-algebras, and the grading on Bb⊗ Cliff 1 is determined by the automorphism α := (β × 1) ◦ β where β is the grading determined by the dual action on B ⋊β Z2 under this identification. Now let u be the canonical self-adjoint unitary generator of Cliff 1 and let v := 1b⊗ u ∈ M(Bb⊗ Cliff 1). Then v is also an odd self-adjoint cM2(A) ∼= Ab⊗ Cliff 1b⊗ Cliff 1 ∼= Ab⊗ Cliff 1b⊗ Cliff 1 ∼= cM2( A) unitary (with respect to the grading α); moreover, we have Ad v = β × 1. Corollary 3.8. With notation as above, if we endow B ⋊β Z2 with the grading associated to the dual action, then K gr Let B be a graded C∗-algebra with grading automorphism β. By [2, Proposition 14.5.4] by the associativity of the graded tensor product. The second assertion then follows from of Kasparov theory. i+1(B) for i = 0, 1. i (B ⋊β Z2) ∼= K gr (cid:3) Proof. This follows immediately from the final assertion of Corollary 3.7 with A := B b⊗ Cliff 1 and α := (β × 1) ◦ β. Remark 3.9. With notation as in the above corollary, observe that since the canonical embedding B → B ⋊β Z2 may be regarded as a graded homomorphism when B is given the trivial grading and B ⋊β Z2 is given the dual grading, there is a natural homomorphism (cid:3) Ki(B) → K gr i (B ⋊β Z2) ∼= K gr i+1(B) for i = 0, 1. Example 3.10. Let β be the grading automorphism of C(T) such that β(f )(z) = f (z). Then there is a short exact sequence of graded C∗-algebras: where C0(R) and C ⊕ C are trivially graded and ǫ(f ) = f (1) ⊕ f (−1) for all f ∈ C(T). Hence by [39, Theorem 1.1] we have a six-term exact sequence: 0 → C0(R)b⊗ Cliff 1 ı−→ C(T) ǫ−→ C ⊕ C i (B b⊗ Cliff 1) = K gr ı∗ K gr 0 (C(T)) ǫ∗ K gr 0 (C ⊕ C) K gr 0 (C0(R)b⊗ Cliff 1) 0 (C0(R)b⊗Cliff 1) ∼= Z, K gr K gr 1 (C ⊕ C) ǫ∗ Since K gr we obtain K gr 1 (C⊕C) = 0, ∗ (C(T)) = (Z3, 0) (cf. [14, p. 105]). It follows by Corollary 3.8 and Remark 3.3 0 (C⊕C) ∼= Z2 and K gr K gr 1 (C(T)) ı∗ K gr 1 (C0(R)b⊗ Cliff 1) 1 (C0(R)b⊗Cliff 1) = K gr 12 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS that K gr ∗ (C(T) ⋊β Z2) = (0, Z3). Note that C(T) ⋊β Z2 is isomorphic to the C∗-algebra of the infinite dihedral group Z ⋊ Z2 ∼= Z2 ∗ Z2. Under the isomorphism C(T) ⋊β Z2 ∼= C∗(Z2 ∗ Z2), the dual grading β becomes the canonical grading on C∗(Z2 ∗ Z2) determined by requiring that both self-adjoint unitary generators be odd. Note that C∗(Z2 ∗ Z2) is the universal unital C∗-algebra generated by two projections (see [31]). Remark 3.11. Let α be the grading automorphism of C0(R) given by α(f )(x) = f (−x) for f ∈ C0(R). A computation similar to the above shows that K gr ∗ (C0(R)) ∼= (Z2, 0). 4. Pimsner's exact sequences for graded C∗-algebras The main result of this section, Theorem 4.4, shows how to compute the graded K- theory of the Cuntz–Pimsner algebra of a graded C∗-correspondence over a nuclear, sep- arable C∗-algebra. We also obtain a graded version of the Pimsner–Voiculescu 6-term exact sequence for crossed products in Corollary 4.6. To prove our main theorem we follow Pimsner's computation of the KK-theory of the Cuntz–Pimsner algebra OX in [29, §4], keeping track of the gradings. Set up. Throughout this section we fix a graded, separable, nuclear C∗-algebra (A, αA), and a graded A–A-correspondence (X, αX) in the sense of Section 2. We assume that the left action ϕ : A → L(X) is injective, by compacts (i.e. ϕ(A) ⊆ K(X)) and essential in the sense that ϕ(A)X = X. We also assume that X is full, that is span{hξ, ηiA : ξ, η ∈ X} = A. It is not clear that all of these hypotheses are required for our arguments (for example, Pimsner does not require that the left action should be by compacts or injective in [29]), but they simplify the discussion and cover the examples that interest us most. Recall that there is an induced grading αX of L(X) given by αX (T ) = αX ◦ T ◦ αX. Let [X] ∈ KK(A, A) denote the class of the Kasparov module(cid:0)X, ϕ, 0, αX(cid:1). Lemma 4.1. With notation as above, if αA is trivial, then αX ∈ L(X), and it is an even self-adjoint unitary with respect to αX. Let X0 := span{x + αX(x) : x ∈ A} X1 := span{x − αX(x) : x ∈ X}. and Then X ∼= X0⊕X1 as A–A-correspondences, and in KK(A, A), we have [X] = [X0]−[X1]. Proof. Since αX is idempotent and αA is trivial, for all ξ, η ∈ X we have hαX(ξ), ηiA = hαX(ξ), α2 X(η)iA = αA(hξ, αX(η)iA) = hξ, αX(η)iA. Hence αX is a self-adjoint unitary in L(X) and since αX (αX) = αX ◦ αX ◦ αX = αX it follows that αX is even. Since αA is trivial, for a, b ∈ A, we have a · (x ± αX (x)) · b = a · x · b ± a · αX (x) · b = a · x · b ± αX(a · x · b), so A · Xi, Xi · A ⊆ Xi for i = 0, 1. For ξ ∈ X0 and η ∈ X1, we have hξ, ηiA = hαX (ξ), ηiA = hξ, αX(η)iA = hξ,−ηiA = −hξ, ηiA. So X0 ⊥ X1, giving X ∼= X0 ⊕ X1 as right-Hilbert A-modules. Since αA is trivial, if ϕ : A → K(X) is the homomorphism defining the left action, then ϕ(A)Xj ⊆ Xj for GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 13 j = 0, 1. So X ∼= X0 ⊕ X1 as C∗-correspondences. We write ϕj : A → K(Xj) for the homomorphism a 7→ ϕ(a)Xj . We now have (cid:0)X, ϕ, 0, αX(cid:1) ∼=(cid:16)X0 ⊕ X1, ϕ0 ⊕ ϕ1, 0,(cid:0) 1 0 0 −1(cid:1)(cid:17) as graded Kasparov modules. The class of the right-hand side is the Kasparov sum of [X0, ϕ0, 0, id] and [X1, ϕ1, 0,− id]. Since αA = idA we have ϕ1 = ϕ1 ◦ αA, and so [X1, ϕ1, 0,− id] is precisely the inverse of [X1] = [X1, ϕ1, 0, id] in KK(A, A) as described at (2.5), and the result follows. (cid:3) Whether or not A is trivially graded, for ξ, η ∈ X, and a ∈ A, we have αX (ξ · a)b⊗ αX(η) = αX(ξ)b⊗ αX (a · η), b⊗n X on X b⊗n. We regard the X b⊗n as graded A–A-correspondences with respect to these operators. When we want to emphasise this grading, we write X b⊗n for the tensor-product module. and using this we see that there is an isometric idempotent operator αXb⊗αX : Xb⊗A X → Xb⊗AX characterised by ξb⊗η 7→ αX(ξ)b⊗αX (η). So αX induces isometric operators α Under this grading, if ξ1, . . . , ξn are homogeneous, say ξk ∈ Xjk, then ξ1 b⊗A · · ·b⊗A ξn is homogeneous with degreePk jk. When convenient we write X b⊗0 for A. Let FX :=L∞n=0 X b⊗n be the Fock space of X [29]. Then FX is a C∗-correspondence The operator α∞X :=L∞n=0 α over A. We write ϕ∞ for the homomorphism A → L(FX) implementing the diagonal left action. b⊗n X is a grading of FX and the induced grading on L(FX) restricts to gradings αK and αT of K(FX) and TX respectively. These satisfy If A is trivially graded, then Lemma 4.1 shows that each α b⊗n X is a self-adjoint unitary. αK(θξ,η) = θ α b⊗n X (ξ),α b⊗m X (η) and αT (Tξ) = T α b⊗n X (ξ) for ξ ∈ X b⊗n and η ∈ X b⊗m. Since these gradings are compatible with the inclusion K(FX) ֒→ TX, they induce a grading αO on OX ∼= TX/K(FX). If A is trivially graded, then Lemma 4.1 shows that αT and αK are inner gradings; but αO need not be, as we shall see later. Pimsner's six-term exact sequence in KK-theory. Let iA : A → TX denote the canonical inclusion homomorphism. Then iA determines a Kasparov class (4.1) [iA] = [TX , iA, 0, αT ] ∈ KK(A,TX ). Pimsner constructs a class in KK(TX , A) as follows: Let π0 : TX → L(FX) denote the canonical representation determined by π0(Tξ)ρ := ξ b⊗A ρ for ξ ∈ X and ρ ∈ X b⊗n. One checks, using the universal property of TX , that there is a second representation π1 : TX → L(FX) such that for ρ ∈ X b⊗n ⊆ FX, π1(a)ρ =(π0(a)ρ if n ≥ 1 if n = 0. 0 Arguing as in [29, Lemma 4.2], we see that π0(T ) − π1(T ) ∈ K(FX) for all T ∈ TX . The 1 0(cid:1) is odd-graded with respect to the grading operator α∞X :=(cid:0) α∞ operator(cid:0) 0 1 X(cid:1), and 0 0 −α∞ X 14 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS so for T ∈ TX , using the formula (2.2), we compute the graded commutator: h(cid:0) 0 1 1 0(cid:1),(cid:0) π0 0 π1◦αT(cid:1)(T )igr 0 =(cid:16) 0 =(cid:16) π1◦αT (T ) 0 π0(T ) 0 π0(αT (T )) π1◦αT (αT (T )) 0 0 (π1−π0)◦αT (T ) (π0−π1)(T ) (cid:17) (cid:17) −(cid:16) 0 (cid:17), 1 0(cid:1),(cid:0) α∞ X which is compact since (π0−π1)(T ) acts only on the finitely-generated submodule A ⊆ FX. Hence we obtain a Kasparov module M :=(cid:16)FX ⊕ FX, π0 ⊕ π1 ◦ αT ,(cid:0) 0 1 0 0 −α∞ X(cid:1)(cid:17). Since the essential subspace of FX ⊕ FX for π0 ⊕ π1 ◦ αT is complemented, replacing FX ⊕FX with the essential subspace for π0⊕π1◦αT , and adjusting the Fredholm operator accordingly yields a module representing the same class (see [2, Proposition 18.3.6]). Hence, writing P : FX → FX ⊖ A =L∞n=1 X b⊗n for the projection onto the orthogonal complement of the 0th summand, we have [M] =hFX ⊕ (FX ⊖ A), π0 ⊕ π1 ◦ αT ,(cid:0) 0 1 P 0(cid:1),(cid:0) α∞ X 0 0 −α∞ X(cid:1)i ∈ KK(TX , A). In the ungraded case, the classes [M] ∈ KK(TX , A) and [iA] ∈ KK(A,TX ) described at (4.2) and (4.1) are denoted α and β in [29, §4]. Theorem 4.2 (cf. [29, Theorem 4.4]). With notation as above, the pair [M] and [iA] are (4.2) X (4.3) obtain We have 0 0 −α∞ As discussed in Pimsner's proof, [2, Proposition 18.7.2(b)] implies that the product and TX are KK-equivalent as graded C∗-algebras. Proof. Since the class [iA] is induced by a homomorphism of C∗-algebras, we can compute [iA]b⊗TX [M] =hFX ⊕ (FX ⊖ A), (π0 ⊕ π1 ◦ αT ) ◦ iA,(cid:0) 0 1 mutually inverse. That is, [iA]b⊗TX [M] = [idA] and [M]b⊗A [iA] = [idTX ]. In particular A the products [iA]b⊗TX [M] and [M]b⊗A [iA] using [2, Proposition 18.7.2]. [iA] b⊗TX [M] is equal to (iA)∗[M], and hence, using the representative (4.2) of [M], we X(cid:1)i. P 0(cid:1),(cid:0) α∞ (cid:16)FX ⊕ (FX ⊖ A), (π0 ⊕ π1 ◦ αT ) ◦ iA,(cid:0) 0 1 X(cid:1)(cid:17) ⊕(cid:0)A, αA, 0,−αA(cid:1) ∼=(cid:16)FX ⊕ FX, (π0 ⊕ π0 ◦ αT ) ◦ iA,(cid:0) 0 1 1 0(cid:1),(cid:0) α∞ 1 0(cid:1) satisfies F 2 = 1, F = F ∗. For a ∈ A, we have The operator F =(cid:0) 0 1 π0(iA(αA(a)))(cid:17) π0(iA(a))(cid:17)(cid:16)0 1 1 0(cid:17) F(cid:0)π0 ⊕ π0 ◦ αT(cid:1)(iA(a)) =(cid:16)0 1 1 0(cid:17)(cid:16)π0(iA(a)) =(cid:16)π0(iA(αA(a))) =(cid:0)π0 ⊕ π0 ◦ αT(cid:1)(iA(αA(a)))F. is a degenerate Kasparov module, and therefore represents the zero class. By (2.6), So the graded commutator (cid:2)F,(cid:0)π0 ⊕ π0 ◦ αT(cid:1)(iA(a))(cid:3)gr is zero by (2.2). Hence (4.3) we have (cid:2)A, αA, 0,−αA(cid:3) = −[idA], so we have [iA] b⊗TX [M] − [idA] = 0KK(A,A) giving [iA]b⊗TX [M] = [idA]. P 0(cid:1),(cid:0) α∞ X(cid:1)(cid:17). 0 0 −α∞ X 0 0 −α∞ 0 0 0 0 X GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 15 For the reverse composition, [2, Proposition 18.7.2(a)] shows that [M]b⊗A [iA] is equal to (iA)∗[M], which is represented by (cid:16)(FX ⊕ FX)b⊗A TX, (π0 ⊕ π1 ◦ αT )b⊗ 1TX ,(cid:0) 0 1 1 0(cid:1),(cid:0) α∞ X b⊗αT 0 0 −α∞ X b⊗αT(cid:1)(cid:17). We write π′0 and π′1 for π0b⊗1TX and (π1◦αT )b⊗1TX . Since X is essential as a left A-module, we have Ab⊗A TX ∼= TX as graded A–TX -correspondences, so the grading αT implements a left action of TX on Ab⊗A TX . We regard this as an action τ of TX on FX b⊗A TX that acts nontrivially only on the 0th summand. Consider the Kasparov module (cid:16)(FX ⊕ FX)b⊗A TX , (0 ⊕ τ ),(cid:0) 0 1 1 0(cid:1),(cid:0) α∞ X b⊗αT 0 0 −α∞ X b⊗αT(cid:1)(cid:17). X X b⊗αT 0 (4.4) b⊗αT 0 by (2.6). 0 −α∞ X 0 −α∞ X Therefore, b⊗αT(cid:1)i 1 0(cid:1),(cid:0) α∞ this submodule is just αT . Hence The essential subspace of the action 0 ⊕ τ is equal to the copy of Ab⊗A TX in the graded submodule (0 ⊕ FX)b⊗A TX of (FX ⊕ FX)b⊗A TX . Moreover, the restriction of 0 ⊕ τ to (cid:2)(FX ⊕ FX)b⊗A TX, (0 ⊕ τ ),(cid:0) 0 1 b⊗αT(cid:1)(cid:3) = [TX , αT , 0,−αT ] = −[idTX ] (iA)∗[M] − [1TX ] =h(FX ⊕ FX)b⊗A TX , π′0 ⊕ π′1,(cid:0) 0 1 b⊗αT(cid:1)i b⊗αT(cid:1)i. We claim that there is a homotopy of graded homomorphisms π′t : TX → L(FX b⊗A TX ) 1 0(cid:1),(cid:0) α∞ +h(FX ⊕ FX)b⊗A TX , (0 ⊕ τ ),(cid:0) 0 1 1 0(cid:1),(cid:0) α∞ =h(FX ⊕ FX)b⊗A TX , π′0 ⊕ (π′1 + τ ),(cid:0) 0 1 1 0(cid:1),(cid:0) α∞ b⊗αT(cid:1)(cid:17) (cid:16)(FX ⊕ FX)b⊗A TX , π′0 ⊕ π′t,(cid:0) 0 1 1 0(cid:1),(cid:0) α∞ from π′0 ◦ αT to π′1 + τ such that for each t ∈ [0, 1], (4.5) is a Kasparov module. To see this, we invoke the universal property of TX. For each t, ψt(ξ) =(cid:0) cos(πt/2)(π′0(αT (Tξ)) − π′1(Tξ)) + sin(πt/2)τ (ξ)(cid:1) + π′1(Tξ). following Pimsner, define a linear map ψt : X → L(FX b⊗A TX) by of A. We write ϕ∞ for ϕ∞ b⊗A 1TX . We aim to prove that ( ϕ∞ ◦ αA, ψt) is a Toeplitz Recall that ϕ∞ : A → L(FX) denotes the homomorphism given by the diagonal left action representation of X for each t ∈ [0, 1]. Since each ψt is a convex combination of bimodule maps, we see that b⊗αT 0 b⊗αT 0 0 −α∞ 0 −α∞ 0 −α∞ X b⊗αT 0 X X X X X ϕ∞(αA(a))ψt(ξ) = ψt(a · ξ) and ψt(ξ) ϕ∞(αA(a)) = ψt(ξ · a) for all a, ξ, t. Next we check that ψt is compatible with the inner product. Note that for all ξ, η, ζ ∈ X the operators (π′0(αT (Tξ)) − π′1(Tξ)), τ (η) and π′1(Tζ) have mutually the only difference between his operators and ours is post-composition with αT ). Given orthogonal ranges in FXb⊗ATX (the same observation is made in Pimsner's argument, and 16 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS ξ, η ∈ X and t ∈ [0, 1] we have we have ψt(ξ)∗ψt(η) =(cid:0)(cid:0) cos(πt/2)(π′0(αT (Tξ)) − π′1(Tξ)) + sin(πt/2)τ (ξ)(cid:1) + π′1(Tξ)(cid:1)∗ (cid:0) cos(πt/2)(π′0(αT (Tη)) − π′1(Tη)) + sin(πt/2)τ (η)(cid:1) + π′1(Tη)(cid:1) =(cid:0)(cid:0) cos(πt/2)(π′0(αT (Tξ)) − π′1(Tξ)) + sin(πt/2)τ (ξ)(cid:1)(cid:1)∗ (cid:0) cos(πt/2)(π′0(αT (Tη)) − π′1(Tη)) + sin(πt/2)τ (η)(cid:1) + π′1(Tξ)∗π′1(Tη). Write P for the projection onto (FX b⊗A TX) ⊖ (Ab⊗A TX ). Since π′1 is a homomorphism, For ζ, ζ′ ∈ X the range of τ (ζ) is contained in Ab⊗A TX ⊆ FXb⊗A TX , which is orthogonal to the range of ((π′0◦αT )−π′1)(Tζ ′). Also, ((π′0◦αT )−π′1)(Tζ) = π′0(αT (Tζ))(1− P ). Using these two points, and resuming our computation of ψt(ξ)∗ψt(η) from above, we have (cid:0) cos(πt/2)(π′0(αT (Tξ)) − π′1(Tξ)) + sin(πt/2)τ (ξ)(cid:1)∗ π′1(Tξ)∗π′1(Tη) = π′1(T ∗ξ Tη)) = π′1(hξ, ηiA) = P ϕ∞(αA(hξ, ηiA)) P . (cid:0) cos(πt/2)(π′0(αT (Tη)) − π′1(Tη)) + sin(πt/2)τ (η)(cid:1) = cos(πt/2)2(π′0(αT (Tξ))(1 − P ))∗(π′0(αT (Tη))(1 − P )) + sin(πt/2)2τ (ξ)∗τ (η) = (1 − P )(cos(πt/2)2 ϕ∞(αA(hξ, ηiA)) + sin(πt/2)2 ϕ∞(αA(hξ, ηiA)))(1 − P ) = (1 − P ) ϕ∞(αA(hξ, ηiA))(1 − P ). Since P commutes with the range of ϕ∞, we have ψt(ξ)∗ψt(η) = ϕ∞(αA(hξ, ηiA)) and so ( ϕ∞ ◦ αA, ψt) is a Toeplitz representation of X for each t ∈ [0, 1]. Thus the universal property of TX ensures that there exists a homomorphism π′t : TX → L(FX b⊗A TX) such that π′t(Tξ) = ψt(ξ) for ξ ∈ X and π′t(a) = ϕ∞(αA(a)) for a ∈ A. For t ∈ [0, 1] and ξ ∈ X, we have π′t(Tξ) − π′1(Tξ) =(cid:0) cos(πt/2)(π′0(αT (Tξ)) − π′1(Tξ)) + sin(πt/2)τ (ξ)(cid:1). The kernel of this operator contains P (FX b⊗A TX ), and since A acts compactly on (1 − P )(FX b⊗A TX ), Cohen factorisation ensures that π′t(Tξ) − π′1(Tξ) ∈ K(FX b⊗A TX ). So for each t, the homomorphism π′t is a compact perturbation of the homomorphism π′1, which determines a Kasparov module, and therefore (4.5) is a Kasparov module for each t as claimed. The claim shows that the class (4.4) is equal to the class of (cid:16)(FX ⊕ FX)b⊗A TX , π′0 ⊕ π′0 ◦ αT ,(cid:0) 0 1 1 0(cid:1),(cid:0) α∞ X b⊗αT 0 0 −α∞ X b⊗αT(cid:1)(cid:17). This is a degenerate Kasparov module (just calculate directly that F 2 = 1, F ∗ = F = 0 for all a ∈ TX ), so it represents the zero class. Hence and (cid:2)F,(cid:0)π′0 ⊕ π′0 ◦ αT )(a)(cid:3)gr (iA)∗[M] = [idTX ]. (cid:3) Let ι : K(FX) → L(FX) denote the canonical inclusion. Then (FX, ι, 0, α∞X ) is a Kasparov module and we have [FX, ι, 0, α∞X ] ∈ KK(K(FX), A). As Pimsner points out, this is the KK-equivalence given by the equivalence bimodule FX. Let j : K(FX) → TX be the natural inclusion. GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 17 Lemma 4.3. (cf. [29, Lemma 4.7]) With notation as above we have X X 0 0 −α∞ in KK(K(FX), A). tion (4.2) of [M] we therefore obtain Proof. By [2, Proposition 18.7.2(b)] we have [j]b⊗TX [M] = j∗[M]. Using the representa- Since π0 ◦ j(K(FX)) ⊆ K(FX) and similarly for π1, the straight-line path from(cid:0) 0 1 P 0(cid:1) to [j]b⊗TX [M] = [FX, ι, 0, α∞X ]b⊗A ([idA] − [X]) [j]b⊗TX [M] =hFX ⊕ (FX ⊖ A), (π0 ⊕ π1 ◦ αT ) ◦ j,(cid:0) 0 1 X(cid:1)i. P 0(cid:1),(cid:0) α∞ X(cid:1)i [j]b⊗TX [M] =hFX ⊕ (FX ⊖ A), (π0 ⊕ π1 ◦ αT ) ◦ j, 0,(cid:0) α∞ =(cid:2)FX, π0 ◦ j, 0, α∞X(cid:3) +(cid:2)FX ⊖ A, π1 ◦ j ◦ αK, 0,−α∞X(cid:3) =(cid:2)FX, ι, 0, α∞X(cid:3) +(cid:2)FX ⊖ A, π1 ◦ j ◦ αK, 0,−α∞X(cid:3). 0 gives an operator homotopy, so 0 0 −α∞ as claimed. We have FX b⊗A X ∼= FX ⊖ A as right-Hilbert modules, and this isomorphism carries π0 b⊗ 1X to π1, and hence (π0 ◦ j)b⊗ 1X to π1 ◦ j. So (FX ⊖ A, π1 ◦ j ◦ αK, 0,−α∞X ) ∼=(cid:0)FX b⊗A X, (π0 ◦ j ◦ αK)b⊗ 1X, 0,−α∞X b⊗ αX(cid:1). The right-hand side represents [FX, ι◦αK, 0,−α∞X ]b⊗A [X]. We have [FX, ι◦αK, 0,−α∞X ] = [j]b⊗TX [M] =(cid:2)FX, ι, 0, α∞X(cid:3) − ([FX, ι, 0, α∞X ]b⊗A [X]) =(cid:2)FX, ι, 0, α∞X(cid:3)b⊗A ([idA] − [X]) −[FX , ι, 0, α∞X ] by (2.5). Thus Finally, we obtain two six-term exact sequences as in [29, Theorem 4.9]. For the purposes of computing graded K-theory, we are most interested in the first of the two sequences, and in the situation where B = C; but both could be useful in general. We write i : A → OX for the canonical inclusion. Theorem 4.4 (cf. [29, Theorem 4.9]). Let A and B be graded separable C∗-algebras such that A is nuclear and let X be a graded correspondence over A such that the left action is injective and by compacts and X is full. Then with notation as above we have two exact six-term sequences as follows. (cid:3) KK0(B, A) b⊗A([idA]−[X]) KK0(B, A) i∗ KK0(B,OX) KK1(B,OX) i∗ KK1(B, A) b⊗A([idA]−[X]) KK1(B, A) KK0(A, B) ([idA]−[X]) b⊗A KK0(A, B) i∗ KK0(OX, B) (4.6) (4.7) KK1(OX, B) i∗ KK1(A, B) ([idA]−[X]) b⊗A KK1(A, B) These sequences are, respectively, contravariantly and covariantly natural in B. They also natural in the other variable in the following sense: if C is a graded C∗-algebra, and YC is 18 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS a full graded correspondence over C whose left action is injective and by compacts, and if θA : A → C and θX : X → Y constitute a morphism of C∗-correspondences, then θA and the induced homomorphism (θA × θX ) : OX → OY induce morphisms of exact sequences from (4.6) for (A, X) to (4.6) for (C, Y ) and from (4.7) for (C, Y ) to (4.7) for (A, X). Proof. We just prove exactness of the first diagram: the second follows from a similar argument. Since A is nuclear, so is TX (see, for example, [34, Theorem 6.3]) and so the quotient map q : TX → OX has a completely positive splitting. Hence [39, Theorem 1.1] applied to the graded short exact sequence 0 → K(FX) −→ OX → 0 yields homomorphisms δ : KKi(B,OX) → KKi+1(B,K(FX)) for which the following six-term sequence is exact. −→ TX j q KK0(B,K(FX)) δ KK1(B,OX) j∗ q∗ KK0(B,TX) KK1(B,TX) q∗ j∗ KK0(B,OX) δ KK1(B,K(FX)). the following diagram. Define maps δ′ : KK∗(B,OX ) → KK∗+1(B, A) by δ′ = (·b⊗[FX, ι, 0, α∞X ])◦δ, and consider KK0(B, A) b⊗([idA]−[X]) b⊗[FX ,ι,0,α∞ X ] KK0(B, A) (iA)∗ [M ] i∗ KK0(B,OX) id KK0(B,K(FX)) j∗ KK0(B,TX) q∗ KK0(B,OX) δ′ δ δ δ′ KK1(B,OX) q∗ id KK1(B,OX) i∗ KK1(B,TX) [M ] (iA)∗ KK1(B, A) j∗ KK1(B,K(FX)) b⊗[FX ,ι,0,α∞ X ] b⊗([idA]−[X]) KK1(B, A) The left-hand and right-hand squares commute by definition of the maps δ′. Lemma 4.3 implies that the top left and bottom right squares commute. Since q ◦ iA = i as ho- momorphisms, we have q∗ ◦ (iA)∗ = i∗, and so the top right and bottom left squares commute as well. Since all the maps linking the inner rectangle to the outer rectangle are isomorphisms, it follows that the outer rectangle is exact as required. Naturality follows from naturality of Pimsner's exact sequences, which in turn follows (cid:3) from naturality of the KK-functor for graded C∗-algebras [2, Section 17.8]. Corollary 4.5. Let (A, α) be a separable, nuclear graded C∗-algebra and let (X, αX) be a countably generated, graded correspondence over A such that the left action is injective and by compacts and X is full. Then there is a 6-term exact sequence for graded K-theory GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 19 as follows. (4.8) K gr 0 (A, α) b⊗A([idA]−[X]) K gr 0 (A, α) i∗ K gr 0 (OX , αO) K gr 1 (OX , αO) i∗ K gr 1 (A, α) b⊗A([idA]−[X]) K gr 1 (A, α) The Pimsner–Voiculescu exact sequence for graded C∗-algebras. If (A, α) is a graded C∗-algebra, and γ is an automorphism of A that is graded in the sense that it commutes with α, then functoriality of KK shows that γ induces a map γ0 on K gr 0 (A, α) = KK(C, A) and γ1 on K gr 1 (A, α) = KK(C, Ab⊗ Cliff 1). The crossed product A ⋊γ Z has two natural grading automorphisms, which we will denote by β0 and β1. To describe them, write iA : A → A ⋊γ Z and iZ : Z → UM(A ⋊γ Z) for the canonical inclusions. Then for k ∈ Z2, the automorphism βk is characterised by (4.9) So β1 = β0◦ γ−1 where γ is the action of T on the crossed product dual to γ. The inclusion iA : A → A ⋊γ Z is a graded homomorphism with respect to α and βk for each of k = 0, 1. Corollary 4.6 (Graded Pimsner–Voiculescu sequence). Let (A, α) be a separable, nuclear graded C∗-algebra and γ an automorphism of A that commutes with α. Fix k ∈ {0, 1} and let βk be the grading automorphism of A ⋊γ Z described above. Then we obtain a six-term exact sequence in graded K-theory as follows. βk(iA(a)iZ(n)) = (−1)kniA(α(a))iZ(n). K gr 0 (A, α) id−(−α∗)kγ∗ K gr 0 (A, α) i∗ K gr 0 (A ⋊γ Z, βk) (4.10) K gr 1 (A ⋊γ Z, βk) i∗ K gr 1 (A, α) id−(−α∗)kγ∗ K gr 1 (A, α) The sequence is natural in the sense that if (B, κ) is another graded C∗-algebra, θ is an automorphism of B that commutes with κ, and φ : A → B is a graded homomorphism that intertwines γ and θ, then φ and φ× 1 : A ⋊γ Z → B ⋊δ Z induce a morphism of exact sequences from (4.10) for (A, α) to (4.10) for (B, κ). Proof. Let X := γA as a Hilbert module, endowed with the grading (−1)kα. Write (iA, iX ) for the inclusions of A and X in OX, write jA : A → A ⋊γ Z for the canonical inclusion, and write U for the unitary element of M(A ⋊γ Z) implementing γ. Then, as pointed out in [29, Example 3, p.193], there is an isomorphism ρ : OX → A ⋊γ Z such that ρ(iA(a)) = jA(a) and ρ(iX (a)) = UjA(a) for all a ∈ A. It is routine to check that this isomorphism is graded with respect to the grading βk of A ⋊γ Z and the grading αO of OX induced by the grading α of A and the grading (−1)kα of X. Hence Corollary 4.5 gives an exact sequence (4.11) K gr 0 (A, α) b⊗A([idA]−[X]) K gr 0 (A, α) i∗ K gr 0 (A ⋊γ Z, βk) K gr 1 (A ⋊γ Z, βk) i∗ K gr 1 (A, α) b⊗A([idA]−[X]) K gr 1 (A, α) 20 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS By definition, we have αk ∗[X] = (−1)kγ∗. Since α-and hence α∗-has order 2, we deduce that [X] = (−α∗)kγ∗, giving the desired six-term exact sequence. ∗[X] = [AA, γ ◦ αk, 0, (−1)kα]. So (2.6) shows that αk Naturality follows immediately from naturality in Theorem 4.4. (cid:3) 5. Twisted P -graph C∗-algebras and actions by countable groups We now begin our investigation of how to use P -graphs to construct examples of graded C∗-algebras. Let F be a countable (discrete) abelian group, fix k ≥ 0 and let P := Nk× F regarded as a cancellative abelian monoid and let GP denote the Grothendieck group of P . Following [4, Definition 2.1], a P -graph consists of a countable small category Λ equipped with a functor d : Λ → P satisfying the factorisation property: if d(λ) = p + q then there exist unique µ, ν ∈ Λ with d(µ) = p, d(ν) = q and λ = µν. We say that Λ is row-finite if vΛp := {λ ∈ Λ : r(λ) = v, d(λ) = p} is finite for v ∈ Λ0 and p ∈ P , There is a natural pre-order on P given by p ≤ q if q = p + u for some u ∈ P . Note and that it has no sources if vΛp is nonempty for every v ∈ Λ0 and p ∈ P . that ≤ need not be a partial order: if F is nontrivial, then ≤ is not antisymmetric. Examples 5.1. (i) Let ΩP = {(p, q) ∈ P × P : p ≤ q}. Regarding P as a subsemigroup of Zk × F , we can define d : ΩP → P by the expression d(p, q) = q − p. Identify P := d−1(0) with P via (p, p) 7→ p, and define r, s : ΩP → Ω0 Ω0 P by r(p, q) = p and s(p, q) = q. Finally define composition by (p, q)(q, n) = (p, n). Then ΩP is a P -graph. (ii) When P is regarded as a category with one object it becomes a P -graph with degree map the identity map, composition the group operation, and range and source both given by the trivial map p 7→ 0. (iii) Every k-graph is a Nk-graph. (iv) In particular, when P = Nk in Example (ii), we obtain the k-graph with one vertex and one path of each degree in Nk. As is standard [21, 22], we denote this k-graph by Tk; its C∗-algebra is isomorphic to C(Tk). (v) Another example we shall use frequently is the 1-graph Bn with one vertex and n distinct edges, whose C∗-algebra is the Cuntz algebra On. The "B" here stands for "bouquet" and we sometimes refer to Bn as the "bouquet of n loops." The definition of the categorical cohomology of a k-graph given in [22, §3] applies to P -graphs and we use the formalism and notation from there. In detail, let A be an abelian group and Λ∗r the collection of composable r-tuples of elements of Λ. Then Z 2(Λ, A), the group of normalised 2-cocycles on Λ, consists of all functions f : Λ∗2 → A such that f (λ, µ) + f (λµ, ν) = f (µ, ν) + f (λ, µν) for all (λ, µ, ν) ∈ Λ∗3 and f (r(λ), λ) = 0 = f (λ, s(λ)) for all λ ∈ Λ (cf. [22, Lemma 3.8]). Furthermore f1, f2 ∈ Z 2(Λ, A) are cohomologous if they differ by a coboundary: that is there is a map b : Λ → A such that (f1 − f2)(λ, µ) = (δ1b)(λ, µ) = b(λ) − b(λµ) + b(µ) for all (λ, µ) ∈ Λ∗2. As usual, when A = T the group operation is written multiplicatively. The following example of a 2-cocycle on Zl 2 will prove important later. 2 for some l ≥ 1. Let Z 2(A, Z2) be the group of Z2-valued group Example 5.2. Let A = Zl 2-cocycles on A. There is a natural map from Z 2(A, Z2) to Z 2(A, T) given as follows: for GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 21 any κ ∈ Z 2(A, Z2) define cκ ∈ Z 2(A, T) by For example, consider κ : A × A → Z2 given by cκ(m, n) = (−1)κ(m,n) for (m, n) ∈ A × A. κ(m, n) = X1≤j<i≤l mi · nj, where mi, nj ∈ Z2. (5.1) Then κ ∈ Z 2(A, Z2). Indeed, κ is biadditive and on pairs (ei, ej) of generators of Zl satisfies κ(ei, ej) = 1 ∈ Z2 if j < i and κ(ei, ej) = 0 ∈ Z2 if i ≤ j. Let σ be a permutation of {1, . . . , l}. Define (mσ)i = mσ(i) for m ∈ A. Then m → mσ is an automorphism of A, and then for κ ∈ Z 2(A, Z2) we may form the 2-cocycle κσ ∈ Z 2(A, Z2), by κσ(m, n) = κ(mσ, nσ) for (m, n) ∈ A × A. We then have 2, it cκσ(m, n) =Yj<i (−1)mσ(i)nσ(j). If b : A → T is a function, then δ1b is the associated 2-coboundary given by δ1b(m, n) = b(m)b(m + n)−1b(n). Lemma 5.3. With notation as above cκ is cohomologous to cκσ in Z 2(A, T). Hence there is a map b : A → T such that cκσ = cκδ1b. Proof. Let χcκσ : A × A → {1,−1} ⊆ T be the bicharacter of A defined by χcκσ (m, n) = cκσ(m, n)cκσ (n, m)−1 = cκσ(m, n)cκσ (n, m) for all m, n ∈ A. For generators ei, ej ∈ A we have χcκσ (ei, ej) =(1 −1 if i = j otherwise. So χcκσ does not depend on σ, and in particular χcκσ = χcκid = χcκ. Thus [26, Proposi- tion 3.2] implies that cκ and cκσ are cohomologous (see also [19, Lemmata 7.1 and 7.2]). The final statement follows by the definition of a coboundary (see [22, Definition 3.2]). (cid:3) given by taking the residue For P := Nk× Zl 2, there is a natural surjection ρ : P → Zk+l 2 the formula cΛ(λ, µ) = cκ(cid:0)ρ(d(λ))), ρ(d(µ))(cid:1) defines a 2-cocycle cΛ ∈ Z 2(Λ, T) with values of each coordinate modulo 2. Proposition 5.4. Let Λ be a P -graph where P = Nk × Zl in {±1}. Proof. Since κ is biadditive on Zk+l and since ρ is a homomorphism, the map (m, n) 7→ (−1)κ(ρ(m),ρ(n)) is a bicharacter of Nk × Zl It follows immediately from the 2. definition of κ that cΛ(r(λ), λ) = 1 = cΛ(λ, s(λ)). Since d is a functor, for a composable triple λ, µ, ν, 2. With κ defined as in (5.1), 2 × Zk+l 2 cΛ(λµ, ν) = cΛ(λ, ν)cΛ(µ, ν) and cΛ(λ, µν) = cΛ(λ, µ)cΛ(λ, ν). Hence cΛ(λ, µ)cΛ(λµ, ν) = cΛ(λ, µ)cΛ(λ, ν)cΛ(µ, ν) = cΛ(λ, µν)cΛ(µ, ν). (cid:3) Definition 5.5. Let Λ be a row-finite P -graph with no sources, and c ∈ Z 2(Λ, T). A Cuntz–Krieger (Λ, c)-family in a C∗-algebra B is a function t : λ 7→ tλ from Λ to B such that 22 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS (CK1) {tv : v ∈ Λ0} is a collection of mutually orthogonal projections; (CK2) tµtν = c(µ, ν)tµν whenever s(µ) = r(ν); (CK3) t∗λtλ = ts(λ) for all λ ∈ Λ; and (CK4) tv =Pλ∈vΛp tλt∗λ for all v ∈ Λ0 and p ∈ P . The following lemma is more or less standard. Lemma 5.6. Let Λ be a row-finite P -graph with no sources, and take c ∈ Z 2(Λ, T). There exists a universal C∗-algebra C∗(Λ, c) generated by a Cuntz–Krieger (Λ, c)-family {sλ : λ ∈ Λ}. Proof. This follows from a standard argument (see, for example [25, Theorem 2.10]) using that the relations force the tλ to be partial isometries. (cid:3) Let Λ be a row-finite P -graph with no sources. The path space ΛΩ is defined to be the collection of all morphisms x : Ω → Λ where Ω = ΩP is as in Examples 5.1 above. The collection of sets Z(λ) := {x ∈ ΛΩ : λ = x(0, d(λ))}, indexed by λ ∈ Λ, is a basis of compact open sets for a locally compact Hausdorff topology on ΛΩ. For each r ∈ P we define the shift map σr : ΛΩ → ΛΩ by (σrx)(p, q) := x(p + r, q + r). We will need to use the path groupoid GΛ of Λ introduced by Carlsen et al. in [4, §2]: GΛ := {(x, p − q, y) : p, q ∈ P and σp(x) = σq(y)}. We identify the path space ΛΩ with the unit space G(0) µ, ν ∈ Λ with s(µ) = s(ν), let Λ via the map x 7→ (x, 0, x). For Z(µ, ν) := {(x, d(µ) − d(ν), y) : x ∈ Z(µ), y ∈ Z(ν), σd(µ)(x) = σd(ν)(y)}. The topology on GΛ is the one with basis UΛ = {Z(µ, ν) : µ, ν ∈ Λ, s(µ) = s(ν)}. This is a locally compact Hausdorff topology on GΛ and GΛ is ´etale in this topology because s : Z(µ, ν) → Z(µ) is a homeomorphism for each µ. The elements of UΛ are all compact open sets. Given µ, ν ∈ Λ with s(µ) = s(ν) and p ∈ P we have Z(µ, ν) =Fα∈s(µ)Λp Z(µα, να). Proposition 5.7. Let Λ be a row-finite P -graph with no sources and take c ∈ Z 2(Λ, T). There is a continuous normalised T-valued groupoid 2-cocycle ςc on GΛ and a surjective homomorphism π : C∗(Λ, c) → C∗(GΛ, ςc) such that π(sλ) = 1Z(λ,s(λ)) 6= 0 for all λ ∈ Λ. Proof. We follow the argument of [22, §6], which proves the analogous result in the case that Λ is a k-graph (see [22, Theorem 6.7]). The only additional difficulty in our current setting is that the notion of minimal common extension for a pair of elements in Λ does not in general make sense in a P -graph. So we must check that we can still construct a partition of GΛ as in [22, Lemma 6.6] without using minimal common extensions. We must show that there is a partition Q of GΛ consisting of elements of UΛ such that Z(λ, s(λ)) ∈ Q for all λ ∈ Λ. We first observe that for each p ∈ P the set Mp := {(x, p, σp(x)) : x ∈ ΛΩ} is a clopen set and that Mp =Fd(λ)=p Z(λ, s(λ)). Moreover, the Mp are pairwise disjoint, M :=Fp∈P Mp is also open and each Z(µ, ν) is either contained in some Mp or disjoint from M. This implies that M is also closed, so it is a clopen set. It remains to show that the complement of M can be expressed as a disjoint union of elements of UΛ. Since UΛ is a countable basis, we can write the complement of M as a countable union of sets in UΛ. Claim: Let Z(κ, λ), Z(µ, ν) ∈ UΛ. Then both Z(κ, λ) ∩ Z(µ, ν) and Z(κ, λ) \ Z(µ, ν) can be expressed as a disjoint union of elements of UΛ. GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 23 Note that Z(κ, λ) and Z(µ, ν) are disjoint unless d(κ) − d(λ) = d(µ) − d(ν); so we may assume that d(κ) − d(λ) = d(µ) − d(ν). Then there exist p, q ∈ P such that d(κ) + p = d(µ) + q and d(λ) + p = d(ν) + q. Let A := {α ∈ s(λ)Λp : Z(κα, λα) = Z(µβ, νβ) for some β ∈ s(ν)Λq} and let B := s(λ)Λp \ A; then we have Z(κ, λ) ∩ Z(µ, ν) = Gα∈A Z(κα, λα) and Z(κ, λ) \ Z(µ, ν) = Gα∈B This proves the claim. Z(κα, λα). Since intersections and set differences of elements of UΛ can be expressed as disjoint unions of such sets, any countable union of elements of UΛ can also be expressed as a disjoint union of such sets. This gives us the desired partition Q. The groupoid 2-cocycle ςc is constructed as in [22, Lemma 6.3] (there the cocycle is denoted σc and the partition is denoted P). By construction of ςc the map λ 7→ 1Z(λ,s(λ)) constitutes a Cuntz–Krieger (Λ, c)-family. Hence, there is a homomorphism π : C∗(Λ, c) → C∗(GΛ, ςc) such that π(sλ) = 1Z(λ,s(λ)). Moreover, for every Z(µ, ν) ∈ UΛ, we have 1Z(µ,ν) = 1Z(µ,s(µ))1∗Z(ν,s(ν)) = π(sµs∗ν); and since the span of elements of the form 1Z(µ,ν) is dense, π is surjective. (cid:3) If c ∈ Z 2(Λ, T) is the trivial cocycle, then our definition of the twisted C∗-algebra C∗(Λ, c) reduces to the definition of the C∗-algebra C∗(Λ) (see [4, Definition 2.4]). If P = Nk and c is an arbitrary cocycle, then our definition of C∗(Λ, c) agrees with the existing definition of the twisted k-graph algebra (see [22, Definition 5.2]). (i) Let ΩP be the P -graph of Examples 5.1(i). Then C∗(ΩP ) ∼= K(ℓ2(P )). Examples 5.8. (ii) Let P be the P -graph from Examples 5.1(ii). Then C∗(P ) is isomorphic to the group C∗-algebra C∗(GP ) of the Grothendieck group of P . (5.2) γΛ χ (sλ) = χ(d(λ))sλ. Let (Λ, d) be a P -graph and c ∈ Z 2(Λ, T). Following [4, Lemma 2.5] we describe the A standard argument (see [4, Lemma 2.5]) shows that the above formula defines a gauge action of cGP on C∗(Λ, c). For χ ∈ cGP and λ ∈ Λ set strongly continuous action of cGP on C∗(Λ, c). in a C∗-algebra B. Suppose that there is a strongly continuous action β of cGP on B satisfying βχ(tλ) = χ(d(λ))tλ for all λ ∈ Λ and χ ∈ cGP . Then the induced homomorphism Proposition 5.9 (Gauge invariant uniqueness theorem). Let Λ be a row-finite P -graph with no sources, and fix c ∈ Z 2(Λ, T). Let t : Λ → B be a Cuntz–Krieger (Λ, c)-family πt : C∗(Λ, c) → B is injective if and only if tv 6= 0 for all v ∈ Λ0. Proof. The result follows from the same argument as in [4, Proposition 2.7] and the observation that the fixed point algebra for the gauge action on C∗(Λ, c) is identical to the fixed point algebra for the gauge action on C∗(Λ) (cf. [23, Theorem 4.2]). (cid:3) Corollary 5.10. With notation as in Proposition 5.7 the map π : C∗(Λ, c) → C∗(GΛ, ςc) is an isomorphism. 24 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS Proof. We argue as in [22, Corollary 7.8]. The cocycle d : GΛ → GP given by d(x, m, y) = m induces an action β of cGP on C∗(GΛ, ςc) satisfying βχ(f )(x, p− q, y) = χ(p− q)f (x, p− q, y) for f ∈ Cc(GΛ). By construction, π intertwines this action with the gauge action. Proposition 5.7 shows that each π(sv) is nonzero. So the result follows from Proposi- tion 5.9. (cid:3) Lemma 5.11. (cf. [9, Proposition 3.2]) Let F be a countable abelian group and let Λ be a row-finite k-graph with no sources. Suppose that F acts on Λ by automorphisms g 7→ ρg. Let P = Nk × F . There is a unique P -graph Λ ×ρ F such that (1) As a set, Λ ×ρ F = Λ × F with degree map d(λ, g) = (d(λ), g); (2) r(λ, g) = (r(λ), 0), s(λ, g) = (s(ρ−g(λ)), 0); (3) (µ, g)(ν, h) = (µρg(ν), g + h) whenever s(µ) = r(ρgν). Proof. For associativity we compute (cid:0)(λ, g)(µ, h)(cid:1)(ν, k) = (λρg(µ), g + h)(ν, k) = (λρg(µ)ρg+h(ν), g + h + k) = (λ, g)(µρh(ν), h + k) = (λ, g)(cid:0)(µ, h)(ν, k)(cid:1). For the factorisation property we suppose that d(λ, g) = (m + n, h + k). Then µ = λ(0, m) and ν = ρ−h(λ(m, m + n)) satisfy (λ, g) = (µ, h)(ν, k) with d(µ, h) = (m, h) and d(ν, k) = (n, k). To see that this factorisation is unique, suppose that (λ, g) = (µ, h′)(ν, k′) = (µρh′(ν), h′ + k′) with d(µ, h′) = (m, h) and d(ν, k′) = (n, k). Then h′ = h and k′ = k by definition of d. Furthermore, λ = µρh(ν) where d(µ) = m and d(ρh(ν)) = d(ν) = n. So the factorisation property in Λ forces µ = λ(0, m) and ρh(ν) = λ(m, m + n), forcing ν = ρ−h(λ(m, m + n)). (cid:3) Example 5.12. Let F be a countable abelian group. We can regard F as a 0-graph and then there is an action τ of F on this 0-graph by translation. So Lemma 5.11 yields an F -graph F ×τ F . (g, h) 7→ (g, g + h). Proposition 5.13. Continue the notation of Lemma 5.11. Then there is an action ρ : F → Aut(C∗(Λ)) such that ρg(sλ) = sρg (λ). Proof. This follows from the universal property of C∗(Λ) (cf. [9, Proposition 3.1]). It is straightforward to check that F ×τ F ∼= ΩF via the map (cid:3) Theorem 5.14. Continue the notation of Lemma 5.11 and Proposition 5.13. Then (1) there is a unitary representation of u : F → UM(cid:0)C∗(Λ ×ρ F )(cid:1) given by u(g) = Pv∈Λ0 s(v,g); (2) there is a homomorphism φ : C∗(Λ) → C∗(Λ ×ρ F ) given by sλ 7→ s(λ,0); (3) we have u(g)φ(a)u(g)∗ = ρg(a) for all a ∈ C∗(Λ) and g ∈ F ; (4) There is an isomorphism φ×u : C∗(Λ)⋊ρF → C∗(Λ×ρF ) such that (φ×u)(sλ, g) = s(λ,g). Proof. This follows from the proof of [9, Theorem 3.4] mutatis mutandis. (cid:3) Examples 5.15. (i) Let Bn be the 1-graph with a single vertex v and edges f1, . . . , fn (see Example 5.1(v)). Let Zn act on Bn by cyclicly permuting the edges. Then C∗(Bn) × Zn ∼= C∗(Bn × Zn). GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 25 (ii) Let Λ be a k-graph, let F be a countable abelian group, and b : Λ → F be a functor. Then the skew product graph Λ ×b F carries a natural F -action τ (see [20, Remark 5.6]) given by τg(λ, h) = (λ, g + h). We have C∗((Λ ×b F ) ×τ F ) ∼= C∗(Λ ×b F ) ⋊τ F ∼= C∗(Λ)b⊗ K(ℓ2(F )). Theorem 5.16. (cf. [9, Proposition 3.5]) Let P = Nk × F where F is a countable abelian group. Suppose that Λ is a P -graph, and let Γ denote the sub-k-graph d−1(Nk × {0}) of Λ. For each g ∈ F and v ∈ Λ0 the sets vΛ(0,g) and Λ(0,g)v are singletons. Moreover, there is an action ρ of F on Γ such that for all λ ∈ Γ and g ∈ F , the unique elements µ ∈ r(λ)Λ(0,g) and ν ∈ s(λ)Λ(0,g) satisfy µρg(λ) = λν. Furthermore, Λ is isomorphic to the P -graph Γ ×ρ F of Lemma 5.11. Proof. The factorisation property ensures that each vertex v ∈ Λ0 has unique factori- sations v = µν with µ ∈ Λ(0,g) and ν ∈ Λ(0,−g). Hence vΛ(0,g) = {µ}, and similarly Λ(0,g)v = {ν}. The map ρ preserves degree by definition. It remains to show that ρg is a functor for each g and ρg ◦ ρh = ρg+h. Fix λ1, λ2 ∈ Γ such that λ2λ1 ∈ Γ and g ∈ F . Let v0 = s(λ1), v1 = r(λ1) = s(λ2) and v2 = r(λ2). Let µi be the unique element of viΛ(0,g) for i = 0, 1, 2. Then Combining the two equations we get µiρg(λi) = λiµi−1 for i = 1, 2. µ2ρg(λ2)ρg(λ1) = λ2µ1ρg(λ1) = λ2λ1µ0, and hence ρg(λ2λ1) = ρg(λ2)ρg(λ1) by uniqueness of factorisations. A similar argument shows that ρg ◦ ρg = ρg+h for all g, h ∈ F . of P -graphs. element in s(λ)Λ(0,g); then ψ−1(λ, g) = λµ. For λ ∈ Λ(n,g) define ψ(λ) =(cid:0)λ(cid:0)(0, 0), (n, 0)(cid:1), g(cid:1) ∈ Γ ×ρ F . Then ψ is an isomorphism let (λ, g) ∈ Γ ×ρ F and µ be the unique Its inverse is given as follows: (cid:3) Corollary 5.17. Suppose that F is a countable abelian group. Then the only possible F -graphs are disjoint unions of quotients of ΩF by subgroups of F . Proof. Let Λ be a F -graph. Then Theorem 5.16 applied in the case k = 0 shows that Λ ∼= Γ ×ρ F where Γ is a 0-graph, which is just a countable set. The group F then acts on Γ, and each orbit is of the form F/H for some subgroup H of F . Since the orbits of ρ correspond to the connected components of Γ×ρ F each component of Λ is isomorphic to F/H ×τH F where τH is the action of F on F/H induced by translation. The result then follows from the identification of ΩF with F ×τ F given in Example 5.12 (cid:3) 6. Gradings of P -graph algebras induced by functors Let P be a finitely-generated, cancellative abelian monoid of the form P ∼= Nk × Zl 2. We will be particularly interested in gradings of twisted P -graph algebras that arise from functors from the underlying P -graphs into Z2. Lemma 6.1. Let Λ be a P -graph, let δ : Λ → Z2 be a functor and let c be a T-valued 2-cocycle on Λ. Then there is a grading automorphism αδ of C∗(Λ, c) such that for all λ ∈ Λ. (6.1) For i ∈ Z2, we have C∗(Λ, c)i = span{sµs∗ν : δ(µ) − δ(ν) = i}. αδ(sλ) = (−1)δ(λ)sλ 26 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS Proof. The universal property of C∗(Λ, c) yields an automorphism αδ satisfying (6.1). For µ, ν ∈ Λ and j ∈ Z2, we have sµs∗ν + (−1)jα(sµs∗ν) 2 =(sµs∗ν 0 if δ(µ) − δ(ν) = j if δ(µ) − δ(ν) = j + 1, so the description of the C∗(Λ, c)i follows from (2.1). Notation 6.2. Consider nonnegative integers k, l, and let P = Nk × Zl 2, regarded as a finitely generated abelian monoid, with generators EP = {g1, . . . , gk+l}. Let π : P → Z2 be the unique homomorphism such that π(gi) = 1 for all i ≤ k + l. Given a P -graph Λ, there is a functor δΛ : Λ → Z2 such that (6.2) δΛ(λ) = π(d(λ)) (cid:3) for all λ ∈ Λ. The following examples illustrate the connection between gradings of P -graph C∗- algebras and twisted P -graph algebras. Specifically, the graded tensor product of graded P -graph algebras can frequently be realised as a twisted C∗-algebra of the cartesian- product graph. We return to this in Theorem 7.1. Examples 6.3. (i) For k ≥ 1 recall from Examples 5.1(iv) that Tk denotes the k-graph Nk with degree functor given by the identity functor, and C∗-algebra isomorphic to C(Tk). Endow C∗(T1) with the grading automorphism αδT1 induced by δT1 (see Notation 6.2); so the unitary generator is homogeneous of odd degree. Then the unitary generator of C∗(T1). Then graded tensor product C∗(T1)b⊗ C∗(T1) is not abelian. To see this, let s1 denote the In particular, the graded tensor product C∗(T1)b⊗C∗(T1) is not isomorphic to C∗(T2). Instead, we claim that C∗(T1)b⊗ C∗(T1) ∼= C∗(T2, c) with grading automorphism (s1b⊗ 1)(1b⊗ s1) = (s1b⊗ s1) 6= −(s1 b⊗ s1) = (1b⊗ s1)(s1 b⊗ 1). and twisting 2-cocycle c : (T2)∗2 → T with values in {±1} given by αδT2 c(m, n) = (−1)m2n1 for (m, n) ∈ N2 × N2 = (T2)∗2. To see this, for n ∈ N let sn ∈ C∗(T1) denote the corresponding generator. Define calculations using the definition of multiplication and involution in the graded tensor product show that the t(m,n) are a Cuntz–Krieger (T2, c)-family; for example, we can check (CK2) as follows: elements {t(m,n) : (m, n) ∈ N2} ⊆ C∗(T1) b⊗ C∗(T1) by t(m,n) := sm b⊗ sn. Routine t(m,n)t(p,q) = (sm b⊗ sn)(sp b⊗ sq) = (−1)np(smsp b⊗ snsq) = c((m, n), (p, q))tm+p,n+q. The universal property of C∗(T2, c) gives a homomorphism ψ : C∗(T2, c) → C∗(T1)b⊗ C∗(T1), and an application of the gauge-invariant uniqueness theorem (Theorem 5.9) shows that ψ is an isomorphism. Finally, one checks on generators that ψ intertwines the grading automorphisms. (ii) Recall that Z2 denotes Z2 considered as a Z2-graph as in Examples 5.1(ii). Let δ := δZ2 be the identity map Z2 → Z2; so the associated grading automorphism αδ of C∗(Z2) makes the generator u homogeneous of degree one. Then there is an isomorphism (C∗(Z2), αδ) ∼= Cliff 1 that takes s1 to (1,−1) and s0 to (1, 1). GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 27 (iii) Consider the graded tensor product (C∗(Z2), αδ)b⊗ (C∗(Z2), αδ). As above, this is, in general, a nonabelian C∗-algebra. Indeed, let c ∈ Z 2(Z2 × Z2, T) be the cocycle given by c(m, n) = (−1)m2n1. Write δ := δZ2 and recall that δZ2×Z2 : Z2 × Z2 → Z2 satisfies δ(i, j) = i + j. Then (C∗(Z2), αδ)b⊗ (C∗(Z2), αδ) ∼=(cid:0)C∗(Z2 × Z2, c), αδZ2×Z2(cid:1). Since (C∗(Z2), αδ) ∼= Cliff 1 as graded algebras, it follows that there is a graded isomorphism C∗(Z2 × Z2, c) ∼= Cliff 2. Indeed, we will see in Corollary 7.5 that for 2 described at (5.1), then Cliff n ∼= any l ≥ 1, if c is the 2-cocycle on the Zl C∗(Z n 2 , c) carries the grading induced by δZ n as above. 2 , c) as graded C∗-algebras, where C∗(Z n 2-graph Z l 2 (iv) Expanding on (i), let A be a C∗-algebra with grading automorphism α. Then, as at (4.9), A ⋊α Z has a grading β1 given by β1(iA(a)iZ(n)) = (−1)niA(α(a))iZ(n); that is, the copy of A retains its given grading, and the generating unitary in the copy of C∗(Z) is odd. Let T1 denote N regarded as a 1-graph. The universal property of A ⋊α Z and straightforward computations show that the map iA(a)iZ(n) 7→ ab⊗ sn defines a graded isomorphism A ⋊α Z ∼= Ab⊗C∗(T1). We return to this in the context of graph C∗-algebras in Examples 8.9 (i). In Section 7, motivated by Examples 6.3(i) and (iii), we will investigate graded tensor products of graded P -graph algebras, and show that these often coincide with twisted C∗- algebras of cartesian-product graphs. To do this, we first need an alternative description of the graded C∗-algebras of appropriate P -graphs as universal graded C∗-algebras. Let Λ be a P -graph where P ∼= Nk × Zl 2 and let EP be the standard generators {ei : 1 ≤ i ≤ k + l} of P . We call elements of Λ with degree in EP edges. Theorem 6.4. Let P = Nk × Zl 2 and let Λ be a P -graph. Let cΛ be the 2-cocycle of Proposition 5.4. Then for each e ∈ Ep such that 2e = 0 and each λ ∈ Λe, there is a unique λ∗ ∈ s(λ)Λe. This λ∗ satisfies s(λ∗) = r(λ), λλ∗ = r(λ), and λ∗λ = s(λ). Moreover there exists a C∗-algebra D such that (1) D is generated by partial isometries {tλ : d(λ) ∈ EP} and mutually orthogonal projections {pv : v ∈ Λ0} such that (a) tλtµ = −tµ′tλ′ for all λ, λ′ ∈ Λe, µ, µ′ ∈ Λe′ with λµ = µ′λ′, e, e′ ∈ EP and (b) if d(λ) = e ∈ EP and 2e = 0 then t∗λ = tλ∗; (c) t∗λtλ = ps(λ) for all λ ∈ Λe, e ∈ EP ; (d) for all v ∈ Λ0 and e ∈ EP we have e 6= e′; pv = Xλ∈vΛe tλt∗λ. (2) D is universal in the sense that for any other C∗-algebra D′ generated by elements t′λ satisfying (a)–(d), there is a homomorphism D → D′ satisfying tλ 7→ t′λ. The C∗-algebra D carries a grading automorphism α satisfying α(tλ) = −tλ whenever d(λ) ∈ EP , and α(pv) = pv for all v ∈ Λ0. Moreover, if αδΛ is the grading of C∗(Λ, cΛ) obtained from Lemma 6.1 applied to the functor (6.2), then there is a graded isomorphism π : C∗(Λ, cΛ) → D such that π(sv) = pv for all v ∈ Λ0 and π(sλ) = tλ whenever d(λ) ∈ EP . 28 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS Proof. For the first statement, suppose that d(λ) = e with 2e = 0. We have d(s(λ)) = 0 = e + e and so the factorisation property shows that there exist λ∗, ν ∈ Λe such that s(λ) = λ∗ν. Since r(λ∗) = s(λ), the pair (λ, λ∗) is composable and since λλ∗ ∈ Λ2e = Λ0 we have s(λ∗) = λλ∗ = r(λ). Hence, the pair (λ∗, λ) is composable and λ∗λ = r(λ). The factorisation property ensures that this λ∗ is unique. Let {sλ : λ ∈ Λ} be partial isometries generating C∗(Λ, cΛ). Let D be the universal C∗- algebra generated by a family of tλ's and pv's satisfying (a)–(d). The universal property guarantees that D carries a grading automorphism α as described. Define Tλ = sλ and Pv = sv. Then the family {P, T} satisfies conditions (c) and (d) above. Suppose that e ∈ EP satisfies 2e = 0 and that λ ∈ Λe. Using (CK2)–(CK4) and the first paragraph, we have sλ∗ = (s∗λsλ)sλ∗ = s∗λcΛ(λ, λ∗)sλλ∗ = s∗λsr(λ) = s∗λ(sλs∗λ) = s∗λ and hence Tλ∗ = T ∗λ . It remains to check property (a): If i 6= j then cΛ(ei, ej) = −1 if j < i and cΛ(ei, ej) = 1 otherwise. Let λ, λ′ ∈ Λei, µ, µ′ ∈ Λej with λµ = µ′λ′. Suppose that j < i. Then TλTµ = sλsµ = cΛ(λ, µ)sλµ = (−1)1sλµ = −sλµ Tµ′Tλ′ = sµ′sλ′ = cΛ(µ′, λ′)sµ′λ′ = (−1)0sλµ = sλµ. Hence TλTµ = −Tµ′Tλ′. If i < j, the same argument applies (switching the λ's and µ's). Hence by the universal property of D there is a map φ : D → C∗(Λ, cΛ) such that φ(pv) = sv for all v ∈ Λ0 and φ(tλ) = sλ for all edges λ. To show that φ has an inverse, for each v ∈ Λ0 set sv = pv. For λ ∈ Λ with d(λ) = i=1 miei, use the factorisation property to write 2 · · · λm2 k+l · · · λmk+l i ) = ei for j = 1, . . . , mi and i = 1, . . . , k + l, and define Pk+l 1 · · · λm1 where d(λj · · · λ1 λ = λ1 k+l 1 λ1 2 Sλ := tλ1 1 · · · tλ m1 1 2 · · · tλ tλ1 m2 2 · · · tλ1 k+l · · · tλ mk+l k+l . Direct calculation shows that {Sλ : λ ∈ Λ} is a Cuntz–Krieger (Λ, cΛ)-family in D. By the universal property of C∗(Λ, cΛ) there is a map ψ : C∗(Λ, cΛ) → D such that ψ(sλ) = Sλ. By construction the maps ψ and φ are mutually inverse and so D ∼= C∗(Λ, cΛ). The final assertion follows by the universality of D. (cid:3) 7. Graded tensor products of twisted P -graph C∗-algebras Let P = Nk× Za 2 and let Q = Nl× Zb 2. Let Λ be a P -graph and Γ a Q-graph. Then Λ×Γ is a P × Q graph. The functor δΛ×Γ and the 2-cocycle cΛ×Γ defined in Proposition 5.4 via (5.1), still make sense as we are using the map π : P × Q → Z(k+a)+(l+b) to define cΛ×Γ. Theorem 7.1. Let P = Nk × Za Q-graph. Then there is an isomorphism of graded C∗-algebras 2 and let Q = Nl × Zb 2. Let Λ be a P -graph and Γ a 2 C∗(Λ × Γ, cΛ×Γ) ∼= C∗(Λ, cΛ)b⊗ C∗(Γ, cΓ) with respect to the gradings αδΛ×Γ and αδΛ b⊗ αδΓ. GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 29 Proof. By Theorem 6.4, the graded C∗-algebra C∗(Λ × Γ, cΛ×Γ) is generated by families {p(v,w), t(λ,w), t(v,µ)} satisfying (a)–(d) of Theorem 6.4. We define T(λ,w) = sλb⊗ sw for λ ∈ Λei, w ∈ Γ0 T(v,µ) = sv b⊗ sµ for µ ∈ Γej , v ∈ Λ0 P(v,w) = sv b⊗ sw for v ∈ Λ0, w ∈ Γ0. Then {P(v,w) : (v, w) ∈ Λ0 × Γ0} is a family of mutually orthogonal projections. We must check that the P(v,w) and the T(λ,w) and T(v,µ) satisfy relations (a)–(d) for the (P × Q)- graph Λ × Γ and the cocycle cΛ×Γ. Condition (a) is the most difficult, and we present it here; (b)–(d) are routine. Let ¯λ, ¯λ′, ¯µ, ¯µ′ be edges in Λ × Γ such that d(¯λ) = d(¯λ′), d(¯µ) = d(¯µ′), d(¯λ) 6= d(¯µ), s(¯λ) = r(¯µ), and ¯λ¯µ = ¯µ′¯λ′. There are four combinations to check according to whether d(¯λ) = (ei, 0), 1 ≤ i ≤ k or d(¯λ) = (0, ei), 1 ≤ i ≤ l and; d(¯µ) = (ej, 0), 1 ≤ j ≤ k or d(¯µ) = (0, ej), 1 ≤ j ≤ l. First suppose that d(¯λ) = (ei, 0) and d(¯µ) = (ej, 0) where i 6= j. Then ¯λ = (λ, v), ¯µ = (µ, v), ¯λ′ = (λ′, v) and ¯µ′ = (µ′, v) for some v ∈ Γ0 and some λ, µ′ ∈ Λei, µ, λ′ ∈ Λej with λµ = µ′λ′. We then have T¯λT¯µ = T(λ,v)T(µ,v) = (sλb⊗ sv)(sµb⊗ sv) = (−1)∂sv·∂sµ(sλsµ b⊗ sv) = (sλµ b⊗ sv) since ∂sv = 0. On the other hand, T¯µ′T¯λ′ = T(µ′,v)T(λ′,v) = (sµ′ b⊗ sv)(sλ′ b⊗ sv) Now suppose that d(¯λ) = (ei, 0) and d(¯µ) = (0, ej) where i 6= j. Then ¯λ = (λ, r(µ)), = (−1)∂sv·∂sλ′ (sµ′sλ′ b⊗ sv) = sµ′sλ′ b⊗ sv = −(sλµ b⊗ sv). whereas ¯µ = (s(λ), µ), ¯µ′ = (r(λ), µ) and ¯λ′ = (λ, s(µ)) for some λ ∈ Λei and µ ∈ Γej . So = (−1)∂sr(µ)·∂sr(λ)(ss(λ)sλb⊗ sµss(µ)) = (sλb⊗ sµ), = (−1)∂sµ′·∂sλ′ (sr(λ′)sλ′ b⊗ sµ′ss(µ′)) = −(sλ b⊗ sµ). T¯λT¯µ = T(λ,r(µ))T(s(λ),µ) = (sλ b⊗ sr(µ))(ss(λ) b⊗ sµ) T¯µ′T¯λ′ = T(r(λ′),µ′)T(λ′,s(µ′)) = (sr(λ′) b⊗ sµ′)(sλ′ b⊗ ss(µ′)) The remaining two cases are similar. By the universal property of C∗(Λ × Γ, cΛ×Γ) there is a map π : C∗(Λ × Γ, cΛ×Γ) → T(λ,r(µ))T(s(λ),µ). We aim to apply the gauge invariant uniqueness theorem for twisted P -graph algebras given in Proposition 5.9 to show that π is injective. For this, observe C∗(Λ, cΛ)b⊗ C∗(Γ, cΓ) such that π(t(λ,w)) = T(λ,w), π(t(v,µ)) = T(v,µ) and π(p(v,w)) = P(v,w). This π is surjective because C∗(Λ, cΛ)b⊗ C∗(Γ, cΓ) is generated by the elements sλb⊗ sµ = that the projections P(v,w) = sv b⊗ sw are nonzero, so it suffices to show that, identifying bGP×Q with bGP × bGQ in the canonical way, π is equivariant for the gauge-action γΛ×Γ on C∗(Λ × Γ, cΛ×Γ) and the action γΛ b⊗ γΓ on C∗(Λ, cΛ)b⊗ C∗(Γ, cΓ) such that for all (χ, χ′) ∈ bGP × bGQ, λ ∈ Λ and µ ∈ Γ. χ′)(sλb⊗ sµ) = χ(d(λ))χ′(d(µ))(sλb⊗ sµ) χ b⊗ γΓ (γΛ 30 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS Since γΛ×Γ (χ,χ′)s(λ,µ) = χ(d(λ))χ′(d(µ))s(λ,µ) we see that π is equivariant on the generators (cid:3) An interesting special case of Theorem 7.1 occurs when Γ is the Z2-graph Z2, so that p(v,w), t(λ,w), t(v,µ), and therefore on C∗(Λ × Γ, cΛ×Γ). C∗(Γ) ∼= Cliff 1 as graded C∗-algebras. Corollary 7.2. Let P = Nk × Za 2 and let Λ be a P -graph. Then C∗(Λ × Z2, cΛ×Z2) ∼= C∗(Λ, cΛ)b⊗ C∗(Z2) ∼= C∗(Λ, cΛ)b⊗ Cliff 1 . with respect to the gradings αδΛ×Z2 Proof. The first statement follows from Theorem 7.1 because Z2 has trivial second coho- mology, so C∗(Z2, cZ2) ∼= C∗(Z2). The second statement follows from Examples 6.3(ii). and αδΛ b⊗ αδZ2 . (cid:3) Remark 7.3. Since the graded tensor product with Cliff 1 is like a graded suspension operation, Corollary 7.2 has implications for graded K-theory. Let P = Nk × Zl 2 for some k, l, and let Λ be a P -graph. Then K gr i+1(C∗(Λ, cΛ)), and then inductively i (C∗(Λ × Z2, cΛ×Z2)) ∼= K gr 2 , cΛ×Z2)) ∼= K gr i+n(C∗(Λ, cΛ)). i=1 Z2. Then C∗(Λ, cΛ) ∼= Cliff l, the lth complex Corollary 7.4. Let Λ be the Zl Clifford algebra. This isomorphism is a graded isomorphism with respect to the grading δΛ of C∗(Λ, cΛ). Proof. We have C∗(Z2) ∼= Cliff 1 as graded C∗-algebras as discussed in Example 6.3(ii). So the result follows from an induction argument using Corollary 7.2 and the definition (cid:3) K gr i (C∗(Λ × Z n 2-graphQl Cliff l+1 = Cliff lb⊗ Cliff 1. So far we have discussed gradings arising from functors from k-graphs into Z2, but there are other possible gradings including those arising from order two automorphisms of k-graphs. Let θ be an order two automorphism of a row-finite k-graph Λ with no sources. Then θ induces a grading βθ of C∗(Λ) satisfying βθ(sλ) = sθ(λ). With respect to this grading, Proposition 7.5. With notation as above there is a graded isomorphism C∗(Λ)0 = span{sλs∗µ + sθ(λ)s∗θ(µ) : s(λ) = s(µ)}, C∗(Λ)1 = span{sλs∗µ − sθ(λ)s∗θ(µ) : s(λ) = s(µ)}. ρ : C∗(Λ ×θ Z2) → C∗(Λ)b⊗ Cliff 1 and such that ρ(s(λ,i)) = sλ b⊗ ui, where Λ ×θ Z2 is the crossed-product (Nk × Z2)-graph, and C∗(Λ ×θ Z2) is graded by the automorphism βθ such that βθ(s(λ,i)) = (−1)is(θ(λ),i). Proof. Direct calculation shows that the elements t(λ,i) := sλ b⊗ ui constitute a Cuntz– Krieger (Λ ×θ Z2)-family in C∗(Λ) b⊗ Cliff 1. So the universal property of C∗(Λ ×θ Z2) gives a homomorphism ρ : C∗(Λ ×θ Z2) → C∗(Λ)b⊗ Cliff 1 taking s(λ,i) to t(λ,i) = sλ b⊗ ui. is injective; it is surjective because its image contains the generators of C∗(Λ)b⊗ Cliff 1. ρ( βθ(s(λ,i))) = ρ((−1)is(θ(λ),i)) = (−1)isθ(λ) b⊗ ui = (βθ b⊗ α)(sλb⊗ ui) = (βθ b⊗ α)ρ(s(λ,i)); Therefore ρ is an isomorphism. Let α be the grading automorphism of Cliff 1. Then An application of the gauge-invariant uniqueness theorem (Proposition 5.9) shows that ρ hence ρ intertwines the grading automorphisms of the two algebras. (cid:3) GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 31 8. Graded K-theory of graph C∗-algebras In this section we apply the sequence (4.8) to a graph C∗-algebra C∗(E) graded by an automorphism αδ determined by a function δ : E1 → Z2-see Corollary 8.3. Remark 8.1. Following [11] (see also [30, §8]), given a 1-graph E, we can realise C∗(E) as the Cuntz–Pimsner algebra of the graph module X(E) defined as the Hilbert-bimodule completion of Cc(E1), regarded as a C0(E0)–C0(E0)-bimodule under the left and right actions given by (a·x·b)(e) = a(r(e))x(e)b(s(e)), under the inner-product hx, yiC0(E0)(v) = Ps(e)=v x(e)y(e). By [11, Proposition 4.4] this left action is by compacts if E is row-finite, and X(E) is full if E has no sinks. The left action is injective if E has no sources. Observe that C0(E0), is separable and nuclear. It carries the trivial grading αC0(E0) = id. Fix a function δ : E1 → Z2. This δ extends uniquely to a functor δ : E∗ → Z2. There is a grading αX(E) on X(E) determined by αX(E)(1e) = (−1)δ(e)1e. It is straightforward to check that (X, αX ) is a graded C0(E0)–C0(E0)-correspondence. By [10, Proposition 12] (see also [30, Example 8.13]) we have C∗(X(E)) ∼= C∗(E). Hence (4.8) becomes (8.1) 0 (C0(E0)) ։ K gr 0 (C0(E0)) 0 (C∗(E), α) → 0 1−[X(E)] −→ K gr since K gr 1 (C∗(E), α) ֒→ K gr 0 → K gr 1 (A, αA) =Lv∈E0 K1(C) = 0. To apply (8.1) to compute the graded K-theory of the C∗-algebra associated to a 1- 0 (C0(E0)) in graph E we need to examine the central terms more closely. We describe K gr Let E be a row-finite 1-graph with no sources. By definition, we have K gr a way which allows us to compute the map b⊗C0(E0)(1 − [X(E)]). 0 (C0(E0)) = KK(C, CE0). Let Cδv be a copy {zδv : z ∈ C} of C as a vector space with inner-product given by hzδv, z′δviCE0 (u) = δv,uzz′ and right action zδv·a = a(v)zδv. It carries a left action ϕv of C by multiplication. The tuple (Cδv, ϕv, 0, id) is a Kasparov C–CE0. The group KK(C, CE0) is generated by the Kasparov C–CE0 modules [Cδv] := [Cδv, ϕv, 0, id] for v ∈ E0, and there is an isomorphism θ : ZE0 → KK0(C, C0(E0)) such that θ(1v) = [Cδv], where 1v is the generator of ZE0 corresponding to v. Now we describe the map b⊗C0(E0)[X(E)] on ZE0 induced by the isomorphism θ. Let E be the E0 × E0 matrix defined by Aδ (8.2) Aδ E(v, w) = Xe∈vE1w (−1)δ(e) (the empty sum is equal to 0 by convention). If Ej denotes the subgraph (E0, δ−1(j), r, s) of E for j = 0, 1, then Aδ E is just AE0 − AE1. Lemma 8.2. Let E be a row-finite 1-graph with no sources or sinks. Then with notation as above, the following diagram commutes. ZE0 (Aδ E )t ZE0 θ θ KK0(C, CE0) b⊗C0(E0)[X(E)] KK0(C, CE0) 32 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS Proof. It suffices to check the diagram commutes on generators 1v. Fix v ∈ E0. Using Lemma 4.1 at the second equality we calculate: θ(1v)b⊗C0(E0) [X(E)] = [Cδv]b⊗C0(E0) [X(E)] = [Cδv]b⊗C0(E0)(cid:0)[X(E)0] − [X(E)1](cid:1) (−1)δ(g)[Cδw]. [Cδw] − Xf∈vE1w,δ(f )=1 [Cδw] = Xg∈vE1w = Xe∈vE1w,δ(e)=0 (cid:3) This is precisely θ((Aδ E)t1v). We now use Corollary 4.5 and Lemma 8.2 to compute the graded K-theory of graph C∗-algebras for suitable gradings. Corollary 8.3. Let E be a row-finite 1-graph with no sources or sinks, and let δ : E → Z2 be the functor determined by the function δ : E1 → Z2. Let αδ be the associated grading αδ(se) = (−1)δ(e)se of C∗(E). Then with Aδ E as in (8.2) 0 (C∗(E), αδ) ∼= coker(1 − (Aδ K gr 1 (C∗(E), αδ) ∼= ker(1 − (Aδ K gr E)t : ZE0 → ZE0) and E)t : ZE0 → ZE0). Note that if E0 is finite, then K gr 1 (C∗(E), αδ) is a free abelian group with the same rank as K gr 0 (C∗(E), αδ). Remark 8.4. Corollary 8.3 provides a direct parallel with [27, Corollary 4.2.5] (see also [30, Example 7.2]): given δ : E1 → Z2, the graded K0-group K gr 0 (C∗(E), αδ) is generated as an abelian group by the classes of the vertex projections {pv : v ∈ E0} subject only to the relations [pv] =(cid:2) Xe∈vE1 ses∗e(cid:3) = Xe∈vE1 (−1)δ(e)[s∗ese] = Xw∈E0 (Aδ E)t(v, w)[pw] coming from Example 3.5. This motivates, in part, our conjecture in Section 9 below. In particular, taking δ ≡ 0, we recover the well-known formula for the (ungraded) K-theory of a 1-graph C∗-algebra [27, Theorem 4.2.4]. Examples 8.5. (1) As in Examples 5.1, for 1 ≤ n < ∞ let Bn be the 1-graph with one n → Z2, and let p := δ−1(1) and q = δ−1(0), so vertex and n edges. Fix δ : B1 Bn)t is the 1 × 1 matrix (q − p). Since C∗(Bn) ∼= On we that p + q = n. Then (Aδ recover the formula for K gr (On) obtained by Haag in [13, Proposition 4.11]: ∗ (On, αδ) ∼=((cid:0)Z1+p−q, 0(cid:1) (Z, Z) K gr ∗ if 1 + p − q 6= 0, otherwise. 2 → Z2 for which Aδ′ (2) Let K2 be the 1-graph associated to the complete directed graph on two ver- 1 −1(cid:1). Then tices. Endow K2 with the map δ′ : K 1 0 (C∗(K2), αδ′) ∼= Z5. In particular, this and Example (1) above show that al- K gr though C∗(K2) ∼= O2 ∼= C∗(B2), there is no graded isomorphism from (C∗(K2), αδ′) to (C∗(B2), αδ) for any δ : B1 (3) More generally, let Λ be a row-finite k-graph with no sources and fix p ∈ Nk. Recall that the dual graph pΛ := {λ ∈ Λ : d(λ) ≥ p} is a k-graph as follows: dp(λ) = d(λ)− p, and if we use the factorisation property in Λ to write each λ ∈ Λ as λ = λt(λ) = h(λ)λ with d(t(λ)) = d(h(λ)) = p, then the range and source maps on pΛ are h and t respectively, and composition in pΛ is given by λ◦p µ = λµ = λµ K2 = (cid:0) −1 −1 2 → {0, 1}. GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 33 λ λ (sΛ 7→ sΛ whenever t(λ) = h(µ) (cf. [1, Proposition 3.2]). By [1, Theorem 3.5] there is an isomorphism θ : C∗(pΛ) → C∗(Λ) such that spΛ t(λ))∗. So any functor δp : pΛ → Z2 induces a grading αp of C∗(pΛ) and hence a grading α of C∗(Λ). As seen in the preceding example, this grading typically does not arise from a functor from Λ to Z2, but for k = 1, we can still apply Corollary 8.3 (to pΛ) to compute K gr ∗ (C∗(Λ), α). (4) Let F be the 1-graph with vertices {vn : n ∈ Z} and edges {en, fn : n ∈ N} where r(en) = r(fn) = vn and s(en) = s(fn) = vn+1. Then C∗(F ) is Morita equivalent to the UHF-algebra M2∞, and so K∗(C∗(F )) = (Z[ 1 2 ], 0). Define δ : F 1 → Z2 by δ(en) = 0 and δ(fn) = 1 for all n. Then the matrix Aδ F is the zero matrix. Hence (C∗(F ), αδ) = 1 − Aδ (0, 0) by Corollary 8.3. (We can also recover this result by taking a direct-limit decomposition as in Example 8.7 below.) F is the identity map from ZF 0 to ZF 0, and we obtain K gr ∗ (C∗(Λ), αΛ) ∼= K∗(C∗(Λ)). Remark 8.6. Suppose that Λ is a bipartite P -graph. That is, Λ0 = L ⊔ R and for every edge λ ∈ Λ either s(λ) ∈ L and r(λ) ∈ R, or vice versa. Then the gradings αδΛ of C∗(Λ) and C∗(Λ, cΛ) induced by the functors δΛ of (6.2) are inner because the grading automorphism is implemented by the self-adjoint multiplier unitary U = PL − PR. Hence [2, 14.5.2] gives K gr ∗ To see why this observation is useful, observe that the skew-product of a k-graph Λ by the degree functor Λ ×d Zk is bipartite with L = Λ0 × {n ∈ Zk :Pi ni is even} and R = Λ0 × {n ∈ Zk :Pi ni is odd}. If Λ is the 1-graph B2 from (1), then B2 ×d Z ∼= F where F as in (4) above. Hence, as graded algebras C∗(B2 ×d Z) ∼= C∗(F ). Also, let Λ be a k-graph and let δ = δΛ : Λ → Z2 be as in (6.2). Then the skew product graph Λ ×δ Z2 is bipartite (with L = Λ0 × {0} and R = Λ0 × {1}), and so the grading on C∗(Λ ×δ Z2) induced by δΛ×δ Z2 is inner. Example 8.7. Consider again the graph and functor of Examples 8.5(4). We have C∗(F ) = S C∗(Fn) where Fn is the subgraph of F with Fix n ∈ N. We have C∗(Fn) ∼= MF vn via sµs∗ν 7→ θµ,ν. Extend δ to F by setting δ(µ) =Pµi=1 δ(µi), and define F 1 n = {e1, f1, . . . , en−1, fn−1}. F 0 n = {v1, . . . , vn} and U = Xµ∈F vn (−1)δ(µ)sµs∗µ. Then U is a self-adjoint unitary in C∗(Fn) that implements the grading by conjugation. (C∗(Fn), αδ) = K∗(C∗(Fn)) = So the grading on each C∗(Fn) is inner, and therefore K gr ∗ K∗(MF vn) = (Z, 0), with generator [svn]. The inclusion map ιn : C∗(Fn) ֒→ C∗(Fn+1) is given by sµs∗ν 7→ sµens∗νen + sµfns∗νfn. In particular ιn(svn) = sens∗en + sfns∗fn. The partial isometry V := sens∗fn is odd and satisfies V ∗V = sens∗en and V ∗V = sfns∗fn. So ιn(svn) = V ∗V + V V ∗. By Example 3.5, we 0 (C∗(Fn+1), αδ), and it follows that ι∗ : K gr have [V ∗V ] = −[V V ∗] in K gr 0 (C∗(Fn), αδ) → K gr 0 (C∗(Fn+1), αδ) sends [svn] to zero and hence is the zero map. Hence ∗ (C∗(F ), αδ) ∼=(cid:0) lim−→(K gr 0 (C∗(Fn), αδ), ι∗), 0(cid:1) ∼= (lim−→(Z, 0), 0) = (0, 0) as before. K gr 34 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS We turn next to some applications of Corollary 4.6 to the crossed products of the C∗- algebra of a 1-graph E. To do this, we first need to describe the map in graded K-theory induced by an automorphism determined by a function from E1 to Z2. Lemma 8.8. Let E be a row-finite 1-graph with no sources and δ : E1 → Z2 a function. (C∗(E), αδ) induced by the Let δ : E → Z2 be the induced functor. The map α∗ on K gr ∗ automorphism αδ is the identity map. Proof. Let X(E) denote be the graph module described in Remark 8.1. For v ∈ E0, e ∈ E1, the grading operator αδ on X(W ) satisfies αδ(1v · 1e) = δv,r(e)αδ(1e) = δv,r(e)(−1)δ(e)1e = 1v · αδ(1e) and similarly αδ(1e · 1v) = αδ(1e)· 1v. So, by linearity and continuity, αδ : X(E) → X(E) is a bimodule map. Moreover for e, f ∈ E1 we have hαδ(1e), αδ(1f )iC0(E0) = (−1)δ(e)+δ(f )h1e, 1fiC0(E0) =(1s(e) 0 if e = f otherwise, which is precisely h1e, 1fiC0(E0). So αδ is a graded automorphism of X(E). the exact sequence Thus the final statement of Theorem 4.4 implies that αδ induces an automorphism of 0 → K gr 1 (C∗(E), αδ) ֒→ K gr 0 (C0(E0)) 1−[X(E)] −→ K gr 0 (C0(E0)) ։ K gr 0 (C∗(E), αδ) → 0. Since this automorphism is the identity map on the two middle terms in the sequence, we deduce that it is the identity map on K gr (cid:3) ∗ (C∗(E), αδ) as claimed. Examples 8.9. (i) Recall Example 6.3(iv). Let E be a row-finite 1-graph with no sources. Give C∗(E) the grading α induced by the functor δ(λ) = λ (mod 2). Consider the crossed product C∗(E) ⋊α Z under the grading α satisfying α(iA(a)iZ(n)) = (−1)niA(α(a))iZ(n). By applying Corollary 4.6 with k = 1 (so that α = β1), and Lemma 8.8 we obtain the following exact sequence. K gr 0 (C∗(E), αδ) ×2 K gr 0 (C∗(E), αδ) i∗ K gr 0 (C∗(E) ⋊αδ Z, α) (8.3) K gr 1 (C∗(E) ⋊αδ Z, α) i∗ K gr 1 (C∗(E), αδ) ×2 K gr 1 (C∗(E), αδ) By Example 6.3(iv), we have a graded isomorphism We use this to compute K gr (cid:0)C∗(E)b⊗ C∗(T1), αδE b⊗ αδT1(cid:1) ∼= (C∗(E) ⋊α Z, α). ∗ (C∗(E)b⊗ C∗(T1)): Since K gr 1 (C∗(E), αδ) has no torsion, multiplication by 2 is injective on that group, so exactness implies that the right-hand boundary map is zero. There- fore K gr K gr 0 (C∗(E) b⊗ C∗(T1)) is isomorphic to the cokernel of the times-two map on 0 (C∗(E), αδ); that is 0 (C∗(E), αδ)/2K gr 0 (C∗(E), αδ). K gr Exactness of the bottom row gives 0 (C∗(E)b⊗ C∗(T1)) ∼= K gr 1 (C∗(E), αδ)) ∼= K gr i∗(K gr 1 (C∗(E), αδ)/2K gr 1 (C∗(E), αδ), GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 35 so K1(C∗(E)b⊗ C∗(T1)) is an extension of the 2-torsion subgroup 0 (C∗(E), αδ) : 2a = 0} In particular, {a ∈ K gr 1 (C∗(E), αδ). 1 (C∗(E), αδ)/2K gr if K gr 0 (C∗(E), αδ) has no 2- by K gr torsion, then K gr 1 (C∗(E), αδ)/2K gr 1 (C∗(E), αδ). 1 (C∗(E)b⊗ C∗(T1)) ∼= K gr (ii) Recall from Examples 6.3(i) that C∗(T1) b⊗ C∗(T1) ∼= C∗(T2, c) where c is the 2- cocycle c(m, n) = (−1)m2n1 for (m, n) ∈ N2. We can compute K gr ∗ (C∗(T2, c), δT2), by taking E = T1 in (i) above. Since T1 = B1 we have K gr 0 (C∗(T1), δT1) = Z2 by Examples 8.5 (1). Then by (8.3), the times-two map on K gr 0 (C∗(T1), δT1) is the zero map, and so the exact sequence above for K gr ((C∗(T2), c), δT2) collapses to give ∗ ∗ (C∗(T2, c), δT2) ∼= (Z2, Z2). Observe that C∗(T2, c) is the rational rotation algebra K gr , so its ungraded K-theory is (Z2, Z2) (see [7]). A 1 2 Remark 8.10. More generally, by Theorem 7.1, if E is any row-finite 1-graph with no sources endowed with the grading induced by the functor δ(e) = 1 for all e ∈ E1, then ple 8.9(i) computes the graded K-theory of this twisted 2-graph C∗-algebra. C∗(E) b⊗ C∗(T1) ∼= C∗(E × T1, cE×T1) with the grading induced by δE×T1. Thus Exam- We finish with an example describing a 2-graph C∗-algebra C∗(Λ) that is Morita equiva- (C∗(Λ), α) = lent to an irrational-rotation algebra, and a grading α on C∗(Λ) such that K gr ∗ (Z2 × Z2, 0). Example 8.11. Consider the following 2-coloured graph (see [28, Example 6.5]) w1 v1 2 2 3 1 w2 v2 3 3 5 2 w3 v3 5 5 8 3 w4 . . . v4 . . . bluev 6= ∅, write vE1 bluew is even and 1 if vE1 Choose a permutation ρ of E1 where the label on a blue (solid) edge indicates the number of parallel blue edges. The pattern is that the numbers of edges are Fibonacci numbers. The red edges are drawn dashed. Let E = Eblue be the subgraph consisting of blue edges. For v, w ∈ E0 such that wE1 bluew = S0(v, w) ⊔ S1(v, w), where S0(v, w) − S1(v, w) is 0 if vE1 bluew is odd. blue that preserves ranges and sources, and cyclicly per- mutes the elements of each Sj(v, w), for j = 0, 1. For each v ∈ E0, let fv be the dashed loop based at v. Let Λ be the 2-graph with the above skeleton, and with factorisation rules given by fr(e)e = ρ(e)fs(e) for all e ∈ E1 Since the numbers of parallel edges grow exponentially fast, Λ has large-permutation factorisations in the sense of [28, Definition 5.6]. It is also cofinal, and so C∗(Λ) is simple with real-rank zero. Elliott's classification theorem [8] combined with [28, Theorem 4.3] implies that C∗(Λ) is Morita equivalent to the irrational-rotation algebra Aθ where θ = 1+√5 blue → Z2 by δ(e) = k whenever e ∈ Sk(v, w) (see [28, Example 6.5]). Define δ : E1 blue. 2 36 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS for some v, w. This induces a functor δ : E∗blue → Z2. The matrix Aδ Corollary 8.3 has entries in {0, 1}, and corresponds to the 1-graph F with skeleton Eblue defined in w1 v1 w2 v2 w3 v3 w4 . . . v4 . . . 0 (C∗(E), αδ) ∼= lim−→(cid:0)Z2,(cid:0) 1 2 0 1(cid:1)(cid:1). This matrix has determinant 1, and so where the pattern of connecting edges repeats every three levels (note: there are no parallel edges). By telescoping these three levels, and arguing as in Example 8.7, we see that K gr (C∗(E), αδ) ∼= (Z2, 0). K gr ∗ As in [9], the permutation ρ of E1 defining the factorisation rules induces an automor- phism ρ of C∗(E), and C∗(Λ) ∼= C∗(E) ⋊ρ Z by [9, Theorem 3.4]. There are two natural extensions of δ to a Z2-valued functor on Λ: namely δ0, determined by δ0(f ) = 0 for all f ∈ Λe2, and δ1, determined by δ1(f ) = 1 for all f ∈ Λe2. The grading automor- phisms αδk correspond under the identification C∗(Λ) ∼= C∗(E) ⋊ρ Z with the automor- phisms βk of Corollary 4.6, for k = 0, 1. So we can compute the graded K-theory of C∗(Λ, αδk ) by applying that result. The automorphism ρ of C∗(E) permutes equivalent projections in approximating finite-dimensional subalgebras of C∗(E). So the automor- phism ρ∗ of K gr (C∗(E)) induced by αδ ◦ ρ is the identity. By Lemma 8.8 we therefore ∗ have (−(αδ)∗)k ρ∗ = (−1)k id for k = 0, 1. Thus id−(−α∗)0 ρ∗ = 0 and Corollary 4.6 gives K gr ∗ groups). And id−(−(αδ)∗)1 ρ∗ = 2 · id, so Corollary 4.6 gives K gr 2, 0). (C∗(Λ), αδ0) ∼=(cid:0)Z2, Z2) (which is isomorphic to K∗(C∗(Λ)) as a pair of abelian ∗ (C∗(Λ), αδ1) = (Z2 9. Conjecture Our results and examples, particularly Example 3.5, lead us to ask whether the graded K0-group of a unital graded C∗-algebra A consists of equivalence classes of homogeneous projections over A subject to the relation [v∗v] = (−1)∂v[vv∗] for every homogeneous partial isometry v. To make this concrete, let (A, α) be a unital graded C∗-algebra. Let P0(A) denote the collection of homogeneous projections in K(bHA). Let p, q ∈ P0(A); we write p ∼ q if there is an even partial isometry v such that p = v∗v and q = vv∗. If p ⊥ q, then p + q is a projection and we write [p] + [q] = [p + q]. Define V0(A) := P0(A)/∼, which is an abelian monoid under the binary operation induced by orthogonal addition. Given a homotopy t 7→ pt in P0(A), we have [p0] = [p1] (see [36, 2.2.7] or [2, §4]). Note that V0(A) may be identified with the set of isomorphism classes of graded projec- tive modules over A. Indeed, given p ∈ P0(A) we may form the graded projective module Theorem (see [2, Theorem 14.6.1]) every graded projective module is isomorphic to a sum- pbHA (with grading inherited from bHA). Given p, q ∈ P0(A). We have p ∼ q if and only if pbHA ∼= qbHA. Moreover, if p ⊥ q, then (p + q)bHA ∼= pbHA ⊕ qbHA. By the Stabilization mand of bHA and therefore is isomorphic to pbHA for some p ∈ P0(A). Thus we may and do regard V0(A) as the semigroup of isomorphism classes of graded projective modules GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 37 over A with the binary operation given by direct sum, that is, [X] + [Y ] = [X ⊕ Y ] where X and Y are graded projective modules. A graded projective module Z is said to be degenerate if there is a graded projective module X such that Z ∼= X ⊕ X op. For p ∈ P0(A), the graded projective module pbHA is degenerate if and only if there is an odd partial isometry v such that p = v∗v + vv∗. Let D0(A) denote the collection of isomorphism classes of degenerate graded projective modules in P0(A). Observe that D0(A) forms a submonoid of V0(A). Let X, Y be graded projective modules; we write X ≈ Y if there are degenerate graded projective modules Z, W such that X ⊕ Z ∼= Y ⊕ W . Then ≈ forms an equivalence relation on graded projective modules (coarser than isomorphism) and we let L(A, α) denote the collection of equivalence classes. We write [X]L for the equivalence class of the graded projective module X. It is routine to show that the direct sum of graded projective modules yields a well-defined binary operation on L(A, α) which makes it an abelian semigroup. Recall that ℓ : C → L(X) denoted the left action of C by scalar multiplication on X. Proposition 9.1. The semigroup L(A, α) forms an abelian group with inverse given by −[X]L = [X op]L for X a graded projective module and zero element given by the class of the trivial module (or any degenerate module). There is a group homomorphism : L(A, α) → K gr 0 (A, α) such that ([X]L) = [ℓ, X, 0, αX] ∈ KK(C, A) = K gr 0 (A, α) for every graded projective module X. Proof. Let X be a graded projective module. Then [X]L + [X op]L = [X ⊕ X op]L = [0]L. The map is well defined since Example 3.5 shows that the Kasparov element associated to a degenerate graded projective module maps to zero and is clearly additive, (cid:3) Conjecture 9.2. The homomorphism of Proposition 9.1 is an isomorphism. It should not be difficult to show that our conjecture holds when A is trivially graded. In this case L(A, id) ∼= K0(A) since graded projective modules over A are all of the form X ∼= Y ⊕ Z op where Y and Z are trivially graded projective modules over A and αX ∼= (id,−id). Moreover, [Y1⊕ Z op 2 ]L if and only if ([Y1], [Z1]) ∼ ([Y2], [Z2]) in the Grothendieck group K0(A). This is is closely related to the argument that KK0(C, A) ∼= K0(A) for ungraded A-see [2, Proposition 17.5.5]. 1 ]L = [Y2⊕ Z op References [1] S. Allen, D. Pask and A. Sims, A dual graph construction for higher-rank graphs, and K-theory for finite 2-graphs, Proc. Amer. Math. Soc. 134 (2006), 455–464. [2] B. Blackadar, K-theory for operator algebras. MSRI Publications vol. 5, Cambridge University Press, 1998. [3] N. Brownlowe, A. Sims and S.T. Vittadello, Co-universal C*-algebras associated to generalised graphs, Israel J. Math. 193 (2013), 399–440. [4] T. Carlsen, S. Kang, J. Shotwell and A. Sims, The primitive ideals of the Cuntz–Krieger algebra of a row-finite higher-rank graph with no sources, J. Funct. Anal., 266 (2014), 2570–2589. [5] A. van Daele, Graded K-theory for Banach algebras I, Quart. J. Math. Oxford, 39 (1988), 185–199. [6] A. van Daele, Graded K-theory for Banach algebras II, Pacific J. Math., 134 (1988), 377–392. [7] G.A. Elliott, On the K-theory of the C ∗-algebra generated by a projective representation of a torsion- free discrete abelian group, Monogr. Stud. Math., 17, Operator algebras and group representations, Vol. I (Neptun, 1980), 157–184, Pitman, Boston, MA, 1984. 38 ALEX KUMJIAN, DAVID PASK, AND AIDAN SIMS [8] G. A. Elliott, On the classification of C ∗-algebras of real rank zero, J. reine angew. Math. 443 (1993), 179–219. [9] C. Farthing, D. Pask and A. Sims, Crossed products of k-graph C ∗-algebras by Zl, Houston J. Math. 35 (2009), 903–933. [10] N.J. Fowler, M. Laca and I. Raeburn, The C ∗-algebras of infinite graphs, Proc. Amer. Math. Soc. 128 (2000), 2319–2327. [11] N.J. Fowler and I. Raeburn, The Toeplitz algebra of a Hilbert bimodule, Indiana Univ. Math. J. 48 (1999), 155–181. [12] E. Guentner and N. Higson, Group C ∗-algebras and K-theory, in Noncommutative Geometry, A. Connes, J. Cuntz, E. Guentner, N. Higson, J. Kaminker and J.E. Roberts (Eds), Lecture Notes in Mathematics, Springer-Verlag Berlin Heidelberg 2004. [13] U. Haag, On Z/2Z-graded KK-theory and its relation with the graded Ext-functor, J. Operator Theory 42 (1999), 3–36. [14] U. Haag, Some algebraic features of Z2-graded KK-theory, K-Theory 13 (1998), 81–108. [15] G. Hao and C-K. Ng, Crossed products of C ∗-correspondences by amenable group actions , J. Math. Anal. Appl. 345 (2008), 702–707. [16] M. Karoubi, Alg`ebres de Clifford et K-th´eorie, Ann. Sci. ´Ecole Norm. Sup. (4) 1 (1968), 161–270. [17] G. G. Kasparov, The Operator K-Functor and Extensions of C ∗-Algebras, Math. USSR. Izv. 16 (1981), 513–572. [18] G. G. Kasparov, Equivariant KK-theory and the Novikov conjecture, Invent. Math. 91 (1988), 147– 201. [19] A. Kleppner, Multipliers on abelian groups, Math. Ann. 158 (1965), 11–34. [20] A. Kumjian and D. Pask, Higher rank graph C ∗-algebras, New York J. Math, 6 (2000), 1–20. [21] A. Kumjian, D. Pask and A. Sims, Homology for higher-rank graphs and twisted C ∗–algebras, J. Funct. Anal., 263 (2012), 1539–1574. [22] A. Kumjian, D. Pask and A. Sims, On twisted higher-rank graph C ∗-algebras, Trans. Amer. Math. Soc., 367 (2015), 5177–5215. [23] A. Kumjian, D. Pask and A. Sims, On the K theory. of twisted k-graph C ∗-algebras, J. Math. Anal. Appl. 401 (2013), 104–113. [24] E.C. Lance, Hilbert C ∗-modules, A toolkit for operator algebraists, Cambridge University Press, Cambridge, 1995, x+130. [25] T.A. Loring, C ∗-algebra relations, Math. Scand. 107 (2010), 43–72. [26] D. Olesen, G.K. Pedersen and M. Takesaki, Ergodic actions of compact abelian groups, J. Operator Theory 3 (1980), 237–269. [27] D. Pask and I. Raeburn, On the K-theory of Cuntz–Krieger algebras, Publ. RIMS Kyoto Univ., 32 (1996), 415–443. [28] D. Pask, I. Raeburn, M. Rørdam and A. Sims, Rank-two graphs whose C ∗-algebras are direct limits of circle algebras, J. Funct. Anal. 239 (2006), 137–178. [29] M. Pimsner, A class of C ∗-algebras generalising both Cuntz–Krieger algebras and crossed products by Z, Fields Inst. Comm. 12, (1997) 189–212. [30] I. Raeburn, Graph Algebras, CBMS 103, AMS, Providence, RI, 2005. [31] I. Raeburn, A. Sinclair, The C ∗-algebra generated by two projections, Math. Scand. 65 (1989), 278– 290. [32] I. Raeburn and D.P. Williams, Morita equivalence and continuous-trace C ∗-algebras, American Mathematical Society, Providence, RI, 1998, xiv+327. [33] J.N. Renault and D.P. Wiliams, Amenability of groupoids arising from partial semigroup actions and topological higher rank graphs, preprint 2015 (arXiv:1501.03027 [math.OA]). [34] A. Rennie, D. Robertson and A. Sims, Groupoid Fell bundles for product systems over quasi-lattice ordered groups, Math. Scand., to appear (arxiv:1501.05476 [math.OA]). [35] J. Roe, Paschke duality for real and graded C ∗-algebras, Quart J. Math. 55 (2004), 325–331. [36] M. Rørdam, F. Larsen and N. Laustsen, An introduction to K-theory for C ∗-algebras, Cambridge University Press, Cambridge, 2000, xii+242. [37] J. Rosenberg and C. Schochet. The Kunneth theorem and the universal coefficient theorem for Kas- parov's generalized K-functor, Duke Math. J. 55 (1987), 431–474. GRADED C ∗-ALGEBRAS, GRADED K-THEORY, AND TWISTED P -GRAPH C ∗-ALGEBRAS 39 [38] A. Sims, B. Whitehead and M.F. Whittaker, Twisted C ∗-algebras associated to finitely aligned higher- rank graphs, Doc. Math. 19 (2014), 831–866. [39] G. Skandalis, Exact sequences for the Kasparov groups of graded algebras, Canad. J. Math. 37 (1985), 193–216. [40] P.J. Stacey, Crossed products of C ∗-algebras by ∗-endomorphisms, J. Austral. Math. Soc. Ser. A 54 [41] B. Whitehead, Twisted relative Cuntz–Krieger algebras associated to finitely aligned higher-rank (1993), 204–212. graphs, Honours Thesis, University of Wollongong, 2012 (arXiv:1310.7045 [math.OA]). (A. Kumjian) Department of Mathematics (084), University of Nevada, Reno NV 89557- 0084, USA E-mail address: [email protected] (D.Pask and A. Sims) School of Mathematics and Applied Statistics, University of Wol- longong, NSW 2522, AUSTRALIA E-mail address: dpask, [email protected]
1608.00190	1	1608	2016-07-31T06:22:44	On the extendability of some classes of maps on Hilbert $C^*$-modules	[ "math.OA" ]	In this paper, we show that every completely semi-$\phi$-map on a submodule of a Hilbert $C^$-module has a completely semi-$\phi$-map extension on the whole of module. We also investigate the extendability of $\phi$-maps and provide examples of $\phi$-maps which has no $\phi$-map extension. Finally, we introduce a category of Hilbert $C^$-module and determine injective objects in this category.	math.OA	math	ON THE EXTENDABILITY OF SOME CLASSES OF MAPS ON HILBERT C ∗-MODULES Mohammad B. Asadi, Reza Behmani, Ali R. Medghalchi, and Hamed Nikpey Abstract. In this paper, we show that every completely semi-φ-map on a submodule of a Hilbert C ∗-module has a completely semi-φ-map extension on the whole of module. We also investigate the extendability of φ-maps and provide examples of φ-maps which has no φ-map extension. Finally, we introduce a category of Hilbert C ∗-module and determine injective objects in this category. 1. INTRODUCTION One of the most fundamental theorems in the theory of operator spaces is the Wittsock's extension theorem for completely bounded maps which is the noncommutative counterpart of the celebrated Hann-Banach's Extension theorem. The authors in [3], introduced the concept of completely semi-φ-maps as a generalization of φ-maps. Also, they shown that every operator valued completely bounded linear map on a Hilbert C ∗-module is a completely semi-φ-map for some completely positive map φ on the underlying C ∗-algebra of the Hilbert C ∗-module and vice versa [4]. Thus it is natural to seeking for an analogue of Wittsock's extension theorem for completely semi-φ-maps. In this note, we show that every φ-map or completely semi-φ-map on a submodule of a Hilbert C ∗-module has a completely semi-φ-map extension on the whole of the Hilbert C ∗-module. Fur- thermore, we provide examples of some φ-map which has no φ-map extension on the whole of module. However, for some special case of φ-maps, we will show to how construct a completely semi-φ-map extension of a φ-map which is close to being a φ-map. In the last section, we introduce a category of Hilbert C ∗-module and determine injective objects in this category. For every Hilbert spaces H, K, the set of all bounded operators B(H, K) is a right Hilbert B(H)- module, where the module action is the composition of operators and the B(H)-inner product on B(H, K) is given by hT, Si = T ∗S for every S, T ∈ B(H, K). Assume A and B are C ∗-algebras, E and G are right Hilbert C ∗-modules over A and B respec- tively, φ : A → B is a completely positive map and Φ : E → G is a map, we say (1) Φ is a φ-map, if hΦ(x), Φ(y)i = φ(hx, yi), for all x, y ∈ E. (2) Φ is a φ-morphism, if Φ is a φ-map and φ is a ∗-homomorphism. (3) Φ is a completely semi-φ-map, if hΦn(x), Φn(x)i ≤ φn(hx, xi) for every x ∈ Mn(E) and n ∈ N. When G = B(H, K) and A = B(H) for some Hilbert spaces H, K, a φ-morphism is called φ-representation and in this case (4) Φ is non-degenerate, if [Φ(E)H] = K. 2010 Mathematics Subject Classification. Primary: 46L08, Secondary: 46L07. Key words and phrases. Hilbert C ∗-modules, Wittsock's extension theorem, completely positive maps, com- pletely semi-φ-maps. 1 2 M. B. ASADI, R. BEHMANI, A. R. MEDGHALCHI, AND H. NIKPEY Note that a φ-morphism Φ is linear and satisfies Φ(x.a) = Φ(x)φ(a) for every x ∈ E and a ∈ A, therefore Φ is a ternary morphism (triple morphism), that is Φ(xhy, zi) = Φ(x)hΦ(y), Φ(z)i for all x, y, z ∈ E. For more information on representation theory of Hilbert C ∗-modules, φ-maps and their dilation theory refer to [1], [2], [4], [6], [8] and [9]. 2. EXTENDABILITY OF COMPLETELY SEMI-φ-MAPS Throughout this section, we assume that H, H1, H2 are Hilbert spaces, A is a C ∗-algebra and F is a non-trivial closed submodule of a Hilbert A-module E. Also we note that the set F ⊥ = {x ∈ E hx, yi = 0, for all y ∈ F } is a closed submodule of E. In this section, we concentrate on operator valued maps on Hilbert C ∗-modules. In fact, if φ : A → B(H1) is a completely positive map, then an operator valued φ-map on F as Φ, means that Φ is a φ-map from F into B(H1)-module B(H1, H2). We first discuss the extension problem for φ-maps. In the following, we provide an example of some φ-map on a submodule of a Hilbert C ∗-module which can not be extended to any φ-map on the whole of the Hilbert C ∗-module. Example 2.1. Suppose that K(H) is the set of all compact operators on H. Clearly, E = K(H) ⊕ K(H) is a full Hilbert K(H)-module (by K(H)-valued inner product h(T1, S1), (T2, S2)i = 1 T2 + S ∗ T ∗ 1 S2) and F = K(H) ⊕ 0 is a nontrivial Hilbert submodule of E. Consider the inclusion map φ = id : K(H) → K(H) ⊂ B(H). The map Φ : F → B(H, H) defined by Φ((T, 0)) = T is a φ-map which doesn't have any φ-map extension on E. In fact, if Φ′ : E → B(H, H) is a φ-map extension of Φ, then Φ′((T1, S1))∗Φ′(T2, S2) = T ∗ 1 T2 + S ∗ 1 S2 and Φ′((T1, 0)) = T1, for all T1, T2, S1, S2 ∈ K(H). A directly calculation shows that Φ′((0, S)) = 0 for all S ∈ K(H). Consequently, Φ′ = Φ ⊕ 0 and so Φ′ is not a φ-map on E The following lemma provides a necessary condition for a completely positive map φ, such that every φ-map on a submodule has a φ-map extension on the whole of the Hilbert C ∗-module. Lemma 2.2. If φ : A → B(H1) is a completely positive map and there exists a non-degenerate φ-map Φ : F → B(H1, H2) which has a φ-map extension on E, then (i) φ(hF ⊥, Ei) = {0}, (ii) every operator valued φ-map on F has a φ-map extension on E. Proof. (i) Assume there is a φ-map Ψ : E → B(H1, H2), extending Φ. For every h, h′ ∈ H1 and x ∈ F , z ∈ F ⊥ one has hΨ(z)h, Φ(x)h′i = hΨ(z)h, Ψ(x)h′i = hΨ(x)∗Ψ(z)h, h′i = hφ(hx, zi)h, h′i = 0. By the assumption, Φ is a non-degenerate map and therefore Ψ(z)h = 0, thus Ψ(F ⊥) = {0}. On the other hand, Ψ is a φ-map, so for every x ∈ E, y ∈ F ⊥ Then φ(hF ⊥, Ei) = {0}. φ(hy, xi) = Ψ(y)∗Ψ(x) = 0. (ii) Assume Θ : F → B(H1, K) is a φ-map. By [4, Theorem 2.2], there is an isometry S : H2 → K such that SΦ(x) = Θ(x), for every x ∈ F . Define Θ′ : E → B(H1, K) by Θ′(x) := SΨ(x) for each x ∈ E. Since Ψ is a φ-map extension of Φ and S is an isometry, Θ′ is a φ-map extension of Θ. (cid:3) By Kolmogorov's decomposition theorem, for every completely positive map φ : A → B(H1), there is at least a non-degenerate operator valued φ-map on F . Therefore, Corollary 2.3. Assume φ : A → B(H1) is a completely positive map. If every operator valued φ-map on F has a φ-map extension on E, then φ(hF ⊥, Ei) = {0}. ON THE EXTENDABILITY OF SOME CLASSES OF MAPS ON HILBERT C ∗-MODULES 3 Theorem 2.4. Suppose that φ : A → B(H1) is a completely positive map and Φ : F → B(H1, H2) is a completely semi-φ-map. Then Φ has a completely semi-φ-map extension Φ′ : E → B(H1, H2). Furthermore, if Φ is a φ-map and φ(hF ⊥, Ei) = 0, then (i) Φ′(F ⊥) = {0}, (ii) Φ′(x)∗Φ′(y) = φ(hx, yi) and Φ′(y)∗Φ′(x) = φ(hy, xi), for all x ∈ E, y ∈ F ⊕ F ⊥. Proof. Let (ρ, K1, V ) be a minimal Stinespring's dilation triple for φ. There is a triple ((Φφ, Hφ), (Ψρ, Kρ), Wφ) consists of a non-degenerate φ-map Φφ : E → B(H1, Hφ), a non-degenerate ρ-representation Ψρ : E → B(K1, Kρ) and a unitary operator Wφ : Hφ → Kρ such that satisfies WφΦφ(x) = Ψρ(x)V for every x ∈ E by [4, Theorem 2.2 part(i)]. Since Φ is a completely semi- φ-map, we have [Φ(xi)∗Φ(xj )]i,j ≤ [φ(hxi, xj i)]i,j, for every x1, ..., xn ∈ F . Therefore, for every h1, ..., hn ∈ H1 we have k n Xi=1 Φ(xi)hik2 ≤ n n Xi=1 Xj=1 hφ(xj , xi)hi, hji = k n Xi=1 Φφ(xi)hik2. Thus there is a unique contractive linear operator S0 : [Φφ(F )H1] → H2 such that S0Φφ(x)h = Φ(x)h for every x ∈ F and h ∈ H1. Let P ∈ B(Hφ) be the orthogonal projection onto [Φφ(F )H1]. φ Ψρ(x)V for every x ∈ F . Put W := WφS ∗ and Put S := S0P : Hφ → H2. Therefore Φ(x) = SW ∗ define Φ′ : E → B(H1, H2) by Φ′(x) := W ∗Ψρ(x)V for all x ∈ E. Since W is a contraction, Φ′ is a completely semi-φ-map by [4, Theorem 2.2 part (iii)]. Obviously, Φ′ is an extension for Φ. Now, let Φ be a φ-map and φ(hF ⊥, Ei) = 0. In this case, the above inequality becomes equality and so S0 becomes an isometry. Also, 0 ≤ Φ′(x)∗Φ′(x) ≤ φ(hx, xi) = 0 satisfies for every x ∈ F ⊥. Therefore Φ′(F ⊥) = {0}. Finally it must be shown that Φ′ satisfies Φ′(x)∗Φ′(y) = φ(hx, yi) for all x ∈ E , y ∈ F ⊕ F ⊥. For this, assume x ∈ E, y ∈ F and h ∈ H1, we have W W ∗Ψρ(y)V h = WφS ∗SW ∗ φ Ψρ(y)V h = WφP Φφ(y)h = WφΦφ(y)h = Ψρ(y)V h, therefore Φ′(x)∗Φ′(y)h = V ∗Ψρ(x)∗W W ∗Ψρ(y)V h = V ∗Ψρ(x)∗Ψρ(y)V h = V ∗ρ(hx, yi)V h = φ(hx, yi)h. Thus Φ′(x)∗Φ′(y) = φ(hx, yi) for every x ∈ E and y ∈ F . Since Φ′(F ⊥) = {0} and φ(hE, F ⊥i) = {0}, for every x ∈ E, y ∈ F , z ∈ F ⊥ we have Φ′(x)∗Φ′(z) = 0 = φ(hx, zi), and therefore Φ′(x)∗Φ′(y + z) = Φ′(x)∗Φ′(y) + Φ′(x)∗Φ′(z) = φ(hx, yi) + φ(hx, yi) = φ(hx, y + zi). (cid:3) The following corollary says that if a non-degenerate operator valued φ-map has a φ-map extension, then the extension is unique. Corollary 2.5. Let φ : A → B(H1) be a completely positive map, Φ : F → B(H1, H2) a non-degenerate φ-map and Γ : E → B(H1, H2) be a φ-map such that ΓF = Φ. Then Γ = Φ′, where Φ′ is as in the proof of Theorem 2.4. Proof. We use the notions of the proof of Theorem 2.4. Since Γ is an φ-extension of Φ, and Φ is a non-degenerate map, then Γ is non-degenerate and H2 = [Φ(F )H1] = [Γ(E)H1]. By [4, Theorem 2.2], there is a unitary operator W ′ : H2 → Kρ such that W ′Γ(e)h := Ψρ(e)V h for all e ∈ E, h ∈ H1. Thus W ′Φ(f )h = W ′Γ(f )h = Ψρ(f )V h = W Φ′(f )h = W Φ(f )h for all f ∈ F , h ∈ H1. Therefore W ′ = W . 4 M. B. ASADI, R. BEHMANI, A. R. MEDGHALCHI, AND H. NIKPEY Now, if x, y ∈ E and h ∈ H1, then we have Γ(x)∗Γ(y)h = φ(hx, yi)h = V ∗ρ(hx, yi)V h = V ∗Ψρ(x)∗Ψρ(y)V h = V ∗Ψρ(x)∗W ′Γ(y)h = V ∗Ψρ(x)∗W Γ(y)h = Φ′(x)∗Γ(y)h. Since Γ(x)∗ and Φ′(x)∗ are bounded operators and [Γ(E)H1] = H2, we have Γ(x)∗ = Φ′(x)∗. (cid:3) If A is a C ∗-subalgebra of K(H), then it is well known that, every closed submodule F of E satisfies the equations F ⊥⊥ = F and F ⊕ F ⊥ = E. Therefore, by Theorem 2.4 and Corollary 2.5 we can conclude that the necessary condition φ(hF ⊥, Ei) = 0 in Lemma 2.2, is a sufficient condition for existence of φ-map extension, in this case. Corollary 2.6. If A is a C ∗-algebra of compact operators and φ : A → B(H1) is a completely positive map and E is a full Hilbert A-module. Then the following statements are equivalent: (i) φ(hF ⊥, Ei) = 0, (ii) there is a non-degenerate operator valued φ-map on F which has a φ-map extension on E, (iii) every operator valued φ-map on F has a φ-map extension on E. (iv) for every φ-map Φ : F → B(H1, H2), the map Φ′ = Φ ⊕ 0 : E = F ⊕ F ⊥ → B(H1, H2) is a φ-map and also Φ′ is the unique φ-map extension of Φ on E. Since the C ∗-algebra K(H) is simple, every nonzero Hilbert K(H)-module is full. In particular, hF ⊥, Ei = hF ⊥, F ⊥i = K(H). Therefore, φ(hF ⊥, Ei) 6= 0, for every nonzero completely positive map φ : K(H) → B(H1). Hence we have Corollary 2.7. If A = K(H) and φ : A → B(H1) is a nonzero completely positive map, then any operator valued φ-map on F has no φ-map extension on E. 3. Category of Hilbert C ∗-modules and Completely semi-φ-maps In the following, we define the category CH,C ∗ as a category whose objects are pairs (E, A) where A is a C ∗-algebra and E is a right Hilbert A-module and a morphism from (E1, A1) to (E2, A2) is a pair (Φ, φ) consists of a completely positive map φ : A → B and a completely semi-φ-map Φ : E1 → E2 and the composition of two morphisms (Φ, φ) and (Ψ, ψ) is (Φ, φ) ◦ (Ψ, ψ) := (Φ ◦ Ψ, φ ◦ ψ). If we restrict ourselves to the case of full Hilbert C ∗-modules over unital C ∗-algebras and unital completely positive maps, we obtain a subcategory of CH,C ∗ which we denote it by C1 H,C ∗. In the following we generalize some results on the characterization of completely semi-ϕ-maps and use it to better understanding C1 H,C ∗ as a subcategory of operator systems COS , and characterize its injective objects. Finally, we compare this new category with the category of Hilbert C ∗-modules when its morphisms are φ-maps, completely bounded maps or Hilbert modules morphisms. For a Hilbert C ∗-module E over a C ∗-algebra A, the smallest operator system which contains A and E is denoted by SA(E) and is defined as follow SA(E) := (cid:20)CIE E ∗ A(cid:21) = {(cid:20) λ y∗ E x a(cid:21) a ∈ A, λ ∈ C, x, y ∈ E}. The following theorem is a generalization of [3, Lemma 3.2] which is useful in the study of CH,C ∗. Proposition 3.1. Suppose that E and F are right Hilbert C ∗-modules over the C ∗-algebras A, B, respectively, and also φ : A → B is a completely positive map and Φ : E → F is a linear map. Then, Φ is a completely semi-φ-map if and only if (cid:20) id Φ Φ∗ φ(cid:21) : SA(E) → SB(F ) (given by (cid:20) λ y∗ is a completely positive map. x a(cid:21) 7→ (cid:20) λ Φ(y)∗ φ(a)(cid:21)) Φ(x) Proof. The same argument as in the proof of [3, Lemma 3.2], works here. (cid:3) Hence we have the following result. ON THE EXTENDABILITY OF SOME CLASSES OF MAPS ON HILBERT C ∗-MODULES 5 Theorem 3.2. C1 H,C ∗ is (up to isomorphism) a subcategory of COS , the category of operator systems. Proof. Define the map Σ : C1 H,C ∗, the operator system SA(E), and corresponds to every morphism (Φ, φ) between two objects of H,C ∗ → COS which corresponds to every object (E, A) of C1 C1 H,C ∗ such as (E, A) and (F , B), the unital completely positive map (cid:20) id Φ Φ∗ φ(cid:21) : SA(E) → SB(F ). It is easy to check that for (Φ1, φ1) : (E1, A1) → (E2, A2) and (Φ2, φ2) : (E2, A2) → (E3, A3) Σ((Φ2, φ2) ◦ (Φ1, φ1)) = Σ((Φ2, φ2)) ◦ Σ((Φ1, φ1)). Thus Σ is a one-to-one covariant functor. (cid:3) Therefore, we can consider C1 H,C ∗ as a category consists of block-wise operator systems(cid:20)CIE E ∗ A(cid:21) , E where A is a unital C ∗-algebra and E is a full right Hilbert A-module, and morphisms are corner preserving unital completely positive maps. We remark that there is some completely positve map between operator systems SA(E) and SB(F ) which is not corner preserving. Example 3.3. For a given Hilbert space H and every bounded operators T1, T2, T3, T4 on it, by elementary row and column operations we have the following unitary equivalence in B(H4) T1 0 0 T3 0 T4 0 0 0 0 T1 0 T2 0 0 T4   ∼=   T1 T2 T3 T4 0 0 0 0   0 0 T1 0 0 0 0 T4   . Therefore the map ϕ : B(H2) → B(H4) defined by ϕ((cid:20)T1 T2 T3 T4(cid:21)) := T1 0 0 T3   0 T4 0 0 0 0 T1 0 T2 0 0 T4   is a unital completely positive map which is not corner-preserving. Now considering B(Hi) as Hilbert C ∗-module over itself, for i = 2, 4, and restriction of ϕ on SB(H2)(B(H2)) provides an example of a unital completely positive map ϕ : SB(H2)(B(H2)) → SB(H4)(B(H4)) which is not corner preserving, thus it is not a morphism in C1 H,C ∗. Definition 3.4. Let (E, A) and (F , B) be two objects of CH,C ∗ . We say that (E, A) contained in (F , B) (or (F , B) contains (E, A)), and denote it by (E, A) ⊂ (F , B), when A is a C ∗-subalgebra of B and E ⊆ F and hx, yiE = hx, yiF for every x, y ∈ E. Definition 3.5. An object (E, A) ∈ C1 H,C ∗ is an injective object in CH,C ∗ when for every pair of elements of CH,C ∗ such as (F , B) and (G, C) which (G, C) contained in (F , B) if there exists a morphism (Φ, φ) : (G, C) → (E, A), then there exists a morphism (Ψ, ψ) : (F , B) → (E, A) such that ψ is an extension of φ and Ψ is an extension for Φ. We are going to give a characterization of injective objects of C1 H,C ∗. In fact, the next theorem is a generalization of Theorem 2.4. Before proving the theorem, we recall some results on injectivity. For an operator space W, its injective envelope is denoted by I(W ) and is the operator space which contains W such that for every operator space V and every completely bounded map Φ : W → V there exists a completely bounded map Ψ : I(W ) → V such that ΨW = Φ. The Paulson operator system associated to W is (cid:20)Cid W W ∗ Cid(cid:21) and denoted by S(W ). First, we prove the following lemma. Lemma 3.6. Let E be a full right Hilbert C ∗-module over a unital C ∗-algebra A. Then I(S(E)) = I(SA(E)). Proof. There exists a Hilbert space H such that E and A be contained in B(H) and therefore S(E) and SA(E) can be considered as subsets of B(H2). Since B(H) is a unital injective C ∗- algebra, there is an injective envelope I(S(E)) of S(E) such that S(E) ⊂ I(S(E)) ⊂ B(H2) and a 6 M. B. ASADI, R. BEHMANI, A. R. MEDGHALCHI, AND H. NIKPEY completely contractive idempotent Φ : B(H2) → B(H2) which is completely positive and its image is I(S(E)) and act identically on S(E), see [5, 4.2.7]. Since Φ is a unital completely positive map and idempotent, there exist unital completely positive maps ϕi : B(H) → B(H) for 1 ≤ i ≤ 2, and an idempotent ϕ : B(H) → B(H) such that Φ = (cid:20)ϕ1 ϕ 2.6.16] for Φ and projections p = (cid:20)idH 0 ϕ∗ ϕ2(cid:21) , Apply [7, Corollary 5.2.2] or [5, 0(cid:21) and idH2 − p. The injective envelope I(S(E)) is a unital 0 C ∗-algebra by the product ◦Φ defined by u1 ◦Φ u2 := Φ(u1u2) for every u1, u2 ∈ I(S(E)) and it has the following block-wise structure (cid:20)I11(E) I(E) I(E) I22(E)(cid:21) where I(E) = ϕ(B(H)) is the injective envelope of E (by [5, 4.4.3] see [5, 4.4.2]) and Iii(E) = ϕi(B(H)) for i = 1, 2 are injective C ∗-algebras. Since I(S(E)) is a unital C ∗-algebra, I11(E) and I22(E) are unital C ∗-algebras and I(E) is a Hilbert I11(E)-I22(E)-bimodule. By the assumption E is full and for every u1, u2 ∈ E we have (cid:20) 0 u∗ 2 0 0(cid:21)◦Φ(cid:20)0 u1 0 (cid:21) = Φ((cid:20) 0 0 u∗ 2 0 0(cid:21)(cid:20)0 u1 0 (cid:21)) = Φ((cid:20)0 0 hu2, u1i(cid:21)) = (cid:20)0 0 ϕ2(hu2, u1i)(cid:21) ∈ (cid:20)0 0 I22(E)(cid:21) , 0 0 0 0 thus ϕ2(A) ⊂ I22(E). Note that ϕ2 is a unital completely positive map on B(H), which is not necessarily multiplicative, but its restriction on A is an isometric and multiplicative map from A into I22(E). To show this, note that Φ is a completely contractive unital idempotent map which acts identically on S(E), thus for every u ∈ E and a ∈ A by [5, Theorem 4.4.9 (Youngson)] or [7, Lemma 6.1.2] we have Φ((cid:20)0 u 0(cid:21) Φ((cid:20)0 0 a(cid:21))) = Φ(Φ((cid:20)0 u 0(cid:21))(cid:20)0 0 a(cid:21)) = Φ((cid:20)0 u 0(cid:21)(cid:20)0 0 a(cid:21)) = Φ((cid:20)0 ua 0 (cid:21)) = (cid:20)0 ua 0 (cid:21) , 0 0 0 0 0 0 0 0 therefore, if ϕ2(a) = 0 for some a ∈ A, then ua = 0 for every u ∈ E, thus a = 0. Thus ϕ2 is one to one. Let u ∈ E and a, b ∈ A. Put T1 := (cid:20)0 u 0(cid:21) , T2 = (cid:20)0 0 a(cid:21) and T3 = (cid:20)0 0 b(cid:21) . Since T1T2T3 and 0 0 0 T1T2 belongs to S(E) and Φ is identity on S(E) then [5, Theorem 4.4.9 (Youngson)] or [7, Lemma 6.1.2] implies that T1 ◦Φ Φ(T2T3) = Φ(T1Φ(T2T3)) = Φ(Φ(T1)T2T3) = Φ(T1T2T3) = T1T2T3 = Φ((T1T2)T3) = Φ(Φ(T1T2)T3) = Φ(T1T2Φ(T3)) = T1T2 ◦Φ Φ(T3) = Φ(T1T2) ◦Φ Φ(T3) = Φ(Φ(T1)T2) ◦Φ Φ(T3) = Φ(T1Φ(T2)) ◦Φ Φ(T3) = T1 ◦Φ Φ(T1) ◦Φ Φ(T3). Therefore for every u ∈ E we have u ◦Φ (ϕ2(ab) − ϕ2(a) ◦Φ ϕ2(b)) = 0 which implies that ϕ2(ab) = ϕ2(a) ◦Φ ϕ2(b), because E ∗ ◦Φ E is an essential ideal in I22(E). Therefore the restriction of ϕ2 on A is an one to one ∗-homomorphism and therefore it is an isometry from A into I22(E), thus A ⊂ I22(E). Thus SA(E) ⊂ I(S(E)) which implies that I(SA(E)) = I(S(E)). (cid:3) Theorem 3.7. A given object (E, A) ∈ C1 H,C ∗ is injective if and only if A and E are injective objects in the category of operator spaces. Proof. Let (E, A) be an injective element in the category C1 H,C ∗ . By Lemma 3.6, (E, A) is contained in (I(E), I22(E)). Thus the identity morphism (id, id) : (E, A) → (E, A) has an extension to a morphism (Φ, φ) : (I(E), I22(E)) → (E, A). Thus φ : I22(E) → A is a completely positive map which extends the identity map on A thus it is unital and Φ : I(E) → E is a completely semi-φ-map and therefore it is completely contractive. On the other hand, the inclusion of E in I(E) is rigid and ΦE = idE , therefore Φ = idI(E), by [7, Theorem 6.1.2]. Thus E = I(E), and E is injective. Similarly, using the fact that (E, A) is contained in (E, I(A)) we can show that A = I(A) and hence A is injective. Conversely, assume E and A are injective operator spaces and E is a full right Hilbert A- module. We show that (E, A) is an injective object in C1 H,C ∗ and (W, B) contained in (V, C). Assume (Φ, φ) is a morphism from (W, B) into (E, A). Since A is an injective C ∗-algebra, there is a unital completely positive map ψ : C → A extending φ. Note that H,C ∗. Let (W, B), (V, C) ∈ C1 ON THE EXTENDABILITY OF SOME CLASSES OF MAPS ON HILBERT C ∗-MODULES 7 B ⊂ C, thus we can consider W as a Hilbert C-module and Φ is a completely semi-ψ-map. Thus the map Λ := (cid:20) id Φ On the other hand SA(E) ⊂ I(SA(E)) = I(S(E)) = (cid:20)I11(E) extends Λ. It is obvious that Θ has the matrix decomposition form (cid:20) id Ψ Φ∗ ψ(cid:21) : SC(W) → SA(E) is a unital completely positive map (by Proposition 3.1). I22(E)(cid:21) . Note that SC(W) ⊂ SC(V) Ψ∗ ψ(cid:21) for some linear map and I(S(E)) is injective, thus there is a unital completely positive map Θ : SC(V) → I(S(E)) which I(E) I(E) Ψ : V → I(E), but E is injective, thus I(E) = E and, there exists a linear map Ψ : V → E such that (cid:20) id Ψ Ψ∗ ψ(cid:21) : SC(V) → SA(E) is a unital completely positive map. Now Proposition 3.1 implies that Ψ is a completely semi-ψ-map, extending Φ. Therefore (E, A) is an injective object in C1 H,C ∗ . (cid:3) Note that C1 H,C ∗ is different from the category of operator spaces, we show this by an example of an injective object in C1 H,C ∗ which its corresponding object is not a injective operator space. Example 3.8. Assume H be a infinite dimensional Hilbert space. Put E = B(C, H). By H,C ∗. But SC(E) = Theorem 2.4 or theorem 3.7 (B(C, H), C) is an injective object in CH,C ∗ and C1 S(E) and I(S(E)) = (cid:20)B(H) E C(cid:21) . Thus S(E) is not an injective operator system. E ∗ Acknowledgment. The research of the first author was in part supported by a grant from IPM (No. 94470046). References [1] Lj. Arambaic, Irreducible representations of Hilbert C*-modules , Math. Proc. R. Ir. Acad. 2 (2005), 11-24. [2] M. B. Asadi, Stinespring's theorem for Hilbert C ∗-modules, J.Operator Theory,62 (2008),no. 2, 235-238. [3] M. B. Asadi, R. Behmani, A. R. Medghalchi, H. Nikpey, Completely semi-ϕ-maps, Submitted. [4] M. B. Asadi, R. Behmani, A. R. Medghalchi, H. Nikpey, Characterization of completely bounded maps on Hilbert C ∗-modules, Submitted. [5] D. P. Blecher and C. Le Merdy, Operator algebras and their modules-an operator space approach, London Mathematical Society Monographs, New Series 30, Oxford University Press, Oxford, 2004. [6] B. V. R. Bhat and G. Ramesh and K. Sumash, Stinespring's theorem for maps on Hilbert C ∗-modules, J.Operator Theory, 68 (2012), 173-178. [7] E. G. Effros and Z.-J. Ruan, Operator spaces, London Math. Soc. Monographs, New Series 23, Oxford Uni- versity Press, New York, 2000. [8] M. Skeide, Factorization of maps between Hilbert C ∗-modules, J. Operator Theory 68 (2012), 543-547. [9] M.Skeide, K. Sumesh, CP-H-Extendable maps between Hilbert Modules and CPH-semigroups, J. Math. Anal. Appl. 414 (2014), 886-913. School of Mathematics, Statistics and Computer Science, College of Science, University of Tehran, Tehran, Iran, and, School of Mathematics, Institute for Research in Fundamental Sciences (IPM), P.O. Box: 19395-5746, Tehran, Iran E-mail address: [email protected] Department of Mathematics, Kharazmi University, 50, Taleghani Ave.,15618, Tehran Iran E-mail address: [email protected] Department of Mathematics, Kharazmi University, 50, Taleghani Ave.,15618, Tehran IRAN. E-mail address: a [email protected] Department of Mathematics, Shahid Rajaei Teacher Training University, Tehran 16785-136, Iran E-mail address: [email protected]
1703.05634	1	1703	2017-03-15T14:04:02	A construction of inductive limit for operator system	[ "math.OA", "math.FA" ]	In this paper, we show a construction of inductive limit for operator system based on Archimedeanization. This inductive limit may be not a closed operator system. We prove that many nuclearity properties could be preserved by a special case of this inductive limit.	math.OA	math	A CONSTRUCTION OF INDUCTIVE LIMIT FOR OPERATOR SYSTEM JIANZE LI Abstract. In this paper, we show a construction of inductive limit for operator system based on Archimedeanization. This inductive limit may be not a closed operator system. We prove that many nuclearity properties could be preserved by a special case of this inductive limit. 1. Introduction Operator system has played an important role in C ∗-algebra tensor product theory (see [2, 10, 11]) after its abstract characterization (see [1]). Tensor product of operator systems was systematically studied in [5]. Furthermore, nuclearity properties of operator system were studied in [3, 4, 5, 6] and were applied to study the well known Kirchberg conjecture (see [8]). Let S and T be operator systems and S(S, T) be the set of unital completely positive maps. A map ϕ : S → T is called a complete order monomorphism if it is a complete order isomorphism onto its image. Let S ⊗ T be their algebraic tensor product. Denote by S ⊗min T the operator system structure on S ⊗ T determined by the cones Mn(S ⊗min T)+ :={(ui,j) ∈ Mn(S ⊗ T) : ((ϕ ⊗ ψ)(ui,j))i,j ≥ 0 for any k, m ∈ N and ϕ ∈ S(S, Mk), ψ ∈ S(T, Mm)}. Then min is injective in the sense that S1 ⊗min T1 ⊆ S2 ⊗min T2 whenever S1 ⊆ S2 and T1 ⊆ T2. Let Dmax n := {α(P ⊗ Q)α∗ : P ∈ Ml(S)+, Q ∈ Mm(T)+, α ∈ Mn,lm, l, m ∈ N} for all n ∈ N. Denote by S ⊗max T the operator system structure on S ⊗ T determined by the cones Mn(S ⊗max T)+ :={u ∈ Mn(S ⊗ T) : ε(1S ⊗ 1T)n + u ∈ Dmax n for any ε > 0}. This Archimedeanization is a critical step in the construction of max structure (see [5, Definition 5.4] and [12, Proposition 3.16]). Date: October 6, 2018. 2000 Mathematics Subject Classification. 46L06; 46L07. Key words and phrases. operator system; inductive limit; nuclearity properties. This work was supported by the National Natural Science Foundation of China (No.11601371). 1 2 JIANZE LI Furthermore, many other operator system structures on S ⊗ T were introduced, includ- ing e, el, er and c (see [5]). el is left injective in the sense that S1 ⊗el T ⊆ S2 ⊗el T whenever S1 ⊆ S2. Similarly, er is right injective and e is injective. Let α, β ∈ {min, e, el, er, c, max} be two operator system structures. We say β is greater than α, denoted by α ≤ β, if the identity map is completely positive from S ⊗β T to S ⊗α T. These structures can be summarized as follows: min ≤ e ≤ el, er ≤ c ≤ max. S is called (α, β)-nuclear, if α ≤ β and S ⊗α T = S ⊗β T for any operator system T. More properties of these structures can be refer to [4, 5]. The definition of inductive limit for closed operator system was introduced in [7] and was used in [9] to study the nuclearity properties of operator system. In this paper, we study a construction of inductive limit for arbitrary operator system. This inductive limit may be not a closed operator system. In section 2, we show the detailed construction of this inductive limit based on Archimedeanization. We prove the universal properties of this construction. In particular, if an operator system is the union of a sequence of increasing operator subsystems, then it is the inductive limit. In section 3, we prove that all the nuclearity properties in [5] can be preserved by this special case of inductive limit. In section 4, we present two examples. 2. Inductive limit Let {Sk : k ∈ N} be a sequence of operator systems and ϕk ∈ S(Sk, Sk+1) for all k ∈ N. Then {Sk, ϕk : k ∈ N} is called an inductive sequence. Let Πk∈NSk be the direct sum of vector spaces. Denote by S ′ = {(xk), there exists k0 such that xk+1 = ϕk(xk) for all k > k0} ⊆ Πk∈NSk. Then S ′ is a ∗-vector space with the canonical involution map. Let S ′ hermitian elements in S ′. Define h be the set of h : there exists k0 such that xk+1 = ϕk(xk) and xk ∈ S+ k for all k > k0}, D1 :={(xk) ∈ S ′ Dn :={((u(k) (u(k+1) i,j i,j )) ∈ Mn(S ′)h : there exists k0 such that ) = (ϕk(u(k) i,j )) and (u(k) i,j ) ∈ Mn(Sk)+ for all k > k0} ⊆ Mn(S ′)h for all n ∈ N. Let K ⊆ S ′ be the set of (xk) such that there exists k0 for which xk = 0 for all k > k0. Let S := S ′/K be the quotient vector space with q : S ′ → S ′/K as the canonical quotient map. We define S+ := {u ∈ Sh : there exists a ∈ D1 such that u = q(a)}, Mn(S)+ := {(ui,j) ∈ Mn(S)h : there exists (ai,j) ∈ Dn such that (ui,j) = qn((ai,j))} INDUCTIVE LIMIT 3 for all n ∈ N. Note that for any (ui,j) ∈ Mn(S)h, there exists (bi,j) ∈ Mn(S ′)h such that (ui,j) = qn((bi,j)). Then S is a matrix ordered ∗-vector space with (1Sk) + K ∈ S as a matrix ordered unit. Let S = Arch(S) be the Archimedeanization of S. Then S is called the inductive limit of inductive sequence {Sk, ϕk : k ∈ N}, denoted by S = lim Sk or simply Sk → S. −→ Now we would like to show some notations before the universal properties. For any p, q ∈ N with q > p, we denote by ϕp,q the composition of {ϕk : p ≤ k < q}. For any x ∈ Sk, we denote by x the sequence (xk) with xl = 0 for l < k, xk = x and xl = ϕk,l(x) for l > k. Let ϕ(k) : Sk → S be the map which sends x ∈ Sk to x + K. Then ϕ(k) is unital completely positive and the following diagram commutes for all k ∈ N. Sk ❃❃❃❃❃❃❃❃ ϕ(k) ϕk S Sk+1 }⑤⑤⑤⑤⑤⑤⑤⑤ ϕ(k+1) Remark 2.1. (i) It is clear that S is the union of the increasing sequence of operator subsystems ϕ(k)(Sk), that is S = ∪k∈Nϕ(k)(Sk). This is different from the inductive limit for closed operator system in [7, 9], which is the closure of union. (ii) ϕ(k)(x) ∈ S+ if there exist l > k such that ϕk,l(x) ∈ S+ l . (iii) For any v ∈ Mn(S)+, we can find k0 ∈ N such that for any k > k0, there exists u ∈ Mn(Sk)+ such that ϕ(k) n (u) = v. Remark 2.2. Suppose that S is an operator system and {Sk, k ∈ N} is an increasing sequence of operator subsystems. If Sk ⊆ Sk+1 with the inclusion map ik for k ∈ N and S = ∪k∈NSk, then S = lim Sk. For convenience, we denote by Sk i−→ S in this case. −→ Remark 2.3. It is not hard to see that any separable operator system is the inductive limit of an increasing sequence of finite-dimensional operator systems. This shows the importance of finite-dimensional operator system in operator system theory. Proposition 2.4. Let S be the inductive limit of inductive sequence {Sk, ϕk : k ∈ N}. (i) Let x ∈ Sk, y ∈ Sl with k, l ∈ N. Then ϕ(k)(x) = ϕ(l)(y) if and only if there exists p > k, l such that ϕk,p(x) = ϕl,p(y). (ii) If T is an operator system and there exist unital completely positive maps ψ(k) : Sk → T such that the following left diagram commutes for all k ∈ N. Sk ❄❄❄❄❄❄❄❄ ψ(k) ϕk T Sk+1 }④④④④④④④④ ψ(k+1) ϕ(k) Sk ❄❄❄❄❄❄❄❄ ψ(k) T S Ψ Then there exists a unique unital completely positive map Ψ : S → T such that the above right diagram commutes for all k ∈ N. / / } / / } / / 4 JIANZE LI Proof. We only prove (ii). For (xk)+K ∈ S, the sequence {ψ(k)(xk) : k ∈ N} is a sequence in T with constant element except finite positions. Denote by lim ψ(k)(xk) the constant element. It is clear that lim ψ(k)(yk) = lim ψ(k)(xk), if (xk) + K = (yk) + K. Now we define Ψ : S → T sending (xk) + K ∈ S to lim ψ(k)(xk). It is clear that the above right diagram commutes and Ψ is unital. If (xk) + K ∈ S+, then lim ψ(k)(xk) is positive. We can verify that Ψ is completely positive similarly. Note that S is the Archimedeanizaiton of S. We get that Ψ is a unital completely positive map. (cid:3) Proposition 2.5. Let S and T be inductive limits of inductive sequences {Sk, ϕk : k ∈ N} and {Tk, ψk : k ∈ N},respectively. Let ϕ(k) : Sk → S and ψ(k) : Tk → T be the canonical maps. Suppose that there exist unital completely positive maps πk : Sk → Tk such that the following left diagram commutes for all k ∈ N. Sk πk Tk ϕk ψk Sk+1 πk+1 / Tk+1 ϕ(k) ψ(k) Sk πk Tk S π / T Then there exists a unique π ∈ S(S, T) such that the above right diagram commutes for all k ∈ N. Proof. Denote by T = T ′/N such that T = Arch(T ). For any (xk) + K ∈ S, we have that (πk(xk)) + N ∈ T. It is clear that (πk(xk)) + N = (πk(yk)) + N if (xk) + K = (yk) + K. Now we define π : S → T sending (xk)+K to (πk(xk))+N. Then the above right diagram commutes and π is unique. Since π : S → T is completely positive, then π : S → T is also. (cid:3) 3. Nuclearity properties In this section, we prove that all the nuclearity properties in [5] can be preserved by the inductive limit in the case of Remark 2.2. Some similar results were proved in [9] based on different inductive limit definitions and different methods. Proposition 3.1. Let T be an operator system and Sk inductive limit of {Sk ⊗max T, ik ⊗ idT : k ∈ N}. i−→ S. Then S ⊗max T is the Proof. Note that i(k) ⊗ idT : Sk ⊗max T → S ⊗max T is completely positive for any k ∈ N. There exists a unique unital completely positive map Ψ : lim (Sk ⊗max T) → S ⊗max T such that the following diagram commutes. −→ / / / / / / INDUCTIVE LIMIT 5 S ⊗max T Ψ i(k)⊗idT 5❧❧❧❧❧❧❧❧❧❧❧❧❧❧❧❧❧❧ Φ(k) Sk ⊗max T lim −→ (Sk ⊗max T) i(k)⊗idT 7♦♦♦♦♦♦♦♦♦♦♦♦♦ Φ(k) ΠSk K0 ⊗ T Ψ Π(Sk⊗T) K Sk ⊗ T It is clear that Ψ is surjective. Now we show that Ψ is injective. Let v ∈ lim (Sk ⊗max T) with Ψ(v) = 0. There exists u ∈ Sk ⊗max T for large enough k ∈ N such that Φ(k)(u) = v and thus (i(k) ⊗ idT)(u) = 0, that is u=0. Therefore, v=0 and thus Ψ is injective. −→ Now we prove that Ψ is a complete order isomorphism. Note that Mn(S ⊗max T)+ is the Archimedeanization of Dmax n = {α(P ⊗ Q)α∗ : P ∈ Ml(S)+, Q ∈ Mm(T)+, α ∈ Mn,lm, l, m ∈ N} and for any P ∈ Ml(S)+, there exists large enough k ∈ N and P ′ ∈ Ml(Sk)+ such that i(k) (P ′) = P . Then we get that l n ⊆ [ Dmax k∈N (i(k) ⊗ idT)n(Mn(Sk ⊗max T)+) for any n ∈ N. Since the diagram commutes, we have Dmax for any n ∈ N. Since Ψ is completely positive, we get that n ⊆ Ψn(Mn(lim(Sk ⊗max T))+) Mn(S ⊗max T)+ = Ψn(Mn(lim(Sk ⊗max T))+) for any n ∈ N. (cid:3) i−→ S. Then Corollary 3.2. Let T be an operator system and Sk (i) S ⊗er T is the inductive limit of {Sk ⊗er T, ik ⊗ idT : k ∈ N}. (ii) S ⊗c T is the inductive limit of {Sk ⊗c T, ik ⊗ idT : k ∈ N}. Proof. The proof is based on Proposition 3.1 and [5, Theorem 6.4,Theorem 7.5]. (cid:3) Remark 3.3. Let S be an operator system and Tk Proposition 3.1 and Corollary 3.2, we get (i) S ⊗max T is the inductive limit of {S ⊗max Tk, idS ⊗ ik : k ∈ N}. (ii) S ⊗c T is the inductive limit of {S ⊗c Tk, idS ⊗ ik : k ∈ N}. (iii) S ⊗el T is the inductive limit of {S ⊗el Tk, idS ⊗ ik : k ∈ N}. i−→ T. By the similar methods in Remark 3.4. (i) Let α ∈ {min, e, el} and Sk limit of {Sk ⊗α T, ik ⊗ idT : k ∈ N}, since α is left injective. (ii) Let α ∈ {min, e, er} and Tk {S ⊗α Tk, idS ⊗ ik : k ∈ N}, since α is right injective. i−→ T. It is clear that S ⊗α T is the inductive limit of i−→ S. It is clear that S ⊗α T is the inductive / / 5 O O / / 7 O O 6 JIANZE LI Remark 3.5. Let α, β ∈ {min, e, el, er, c, max} and α ≤ β. Note that Remark 3.3 and 3.4. The following are equivalent. (i) S is (α, β)-nuclear. (ii) S ⊗β Tk → S ⊗α T if Tk (iii) S ⊗α Tk → S ⊗β T if Tk i−→ T. i−→ T. i−→ S. Suppose that α, β ∈ {min, e, el, er, c, max} and α ≤ β. Theorem 3.6. Let Sk (i) S is (α, β)-nuclear if there exists k0 such that Sk is (α, β)-nuclear for any k > k0; (ii) If β = e or el, then S is (α, β)-nuclear if and only if there exists k0 such that Sk is (α, β)-nuclear for any k > k0. Proof. (i) If α ∈ {min, e, el}, then α is left injective and thus Sk ⊗α T ⊆ S ⊗α T for any k ∈ N. We only need to prove that idS⊗T in the following left diagram is completely positive. For any u ∈ (S ⊗α T)+, there exists large enough k such that u ∈ (Sk ⊗α T)+ = (Sk ⊗β T)+. Then u ∈ (S ⊗β T)+ and thus idS⊗T is positive. We can prove it is completely positive similarly. If α ∈ {er, c, max}, by Proposition 3.1 and Corollary 3.2, we get that S ⊗α T = S ⊗β T. (ii) is clear from the following right diagram since α and β are both left injective. S ⊗α T idS⊗T / S ⊗β T S ⊗α T idS⊗T / S ⊗β T Sk ⊗α T Sk ⊗β T Sk ⊗α T idS k⊗T Sk ⊗β T i(k)⊗idT 4. Example (cid:3) Example 4.1. Recall that a UHF algebra (uniformly hyperfinite algebra) is a unital C ∗- algebra A such that A = ∪∞ An, where {An : n ∈ N} is an increasing sequence of finite dimensional simple C ∗-subalgebras containing the unit of A (see [10]). Let n, d ∈ N. We denote by n=1 ϕ : Mn → Mdn, x 7→   x 0 · · · 0 x   the canonical map from Mn to Mdn. Note that any finite dimensional simple C ∗-algebra is ∗-isomorphic to some Mk. Then A is in fact the C ∗-algebra inductive limit of a sequence {Mnk, ϕk} with nknk+1 and ϕk is the canonical map from Mnk to Mnk+1. Let γ : N \ {0} → N \ {0} be a function. Define γ! : N \ {0} → N \ {0} by γ!(n) = γ(1)γ(2) . . . γ(n). Let ϕn : Mγ!(n) → Mγ!(n+1) be the canonical map. Then any γ determines a C ∗-algebra inductive sequence {Mγ!(n), ϕn} and thus a UHF algebra Aγ. / / ? O O O O / / ? O O ? O O INDUCTIVE LIMIT 7 −→ Now we consider the operator system inductive limit of this sequence. Let Sγ = lim {Mγ!(n), ϕn}. Then there exists σ : Sγ → Aγ which is a complete order monomor- phism. In fact, if ϕ(n) : Mγ!(n) → Aγ is the canonical ∗-homomorphism, there exists a unique unital completely positive map σ from Sγ to Aγ such that the following diagram commutes by Proposition 2.4. ϕ(n) 7♦♦♦♦♦♦♦♦♦♦♦♦♦ τ (k) Aγ σ Sγ Mγ!(n) By the constructions of C ∗-algebra inductive limit, we see that Sγ is in fact a linear subspace of Aγ and σ is the inclusion map. Then it is injective. Now we prove that σ is a complete order monomorphism. Let l ∈ N and σl(v) ∈ M + l (σ(Sγ)). Since σ(Sγ) = ∪n∈Nϕ(n)(Mγ!(n)), there exists large enough k ∈ N such that σl(v) ∈ M + l (ϕ(k)(Mγ!(k))). Then v ∈ M + l (Sγ) since τ (k) is the inclusion. Example 4.2. Recall that S∞ is the smallest closed operator subsystem in C ∗(F∞) con- taining all the generators (see [6]). Sn is the 2n+1 dimensional operator system in C ∗(Fn) containing the generators. There exist unital completely positive maps i(n) : Sn → S∞ and p(n) : S∞ → Sn such that p(n) ◦ i(n) is the identity map on Sn for any n ∈ N. Then i(n) is a complete order monomorphism and thus Sn ⊆ S∞. Similarly, Sn ⊆ Sn+1 in a canonical way. Now we define i , e, gi : i ∈ N} ⊆ C ∗(F∞). Sε = span{g∗ It is clear that S∞ = ¯Sε and Sε = lim {Sn, in, n ∈ N}, that is, Sε = ∪n∈NSn. Then Sε is (min, er)-nuclear by [6, Proposition 9.9] and Theorem 3.6. By the similar methods for [6, Theorem 9.13], we could get that Kirchberg conjecture has an affirmative answer if and only if Sε is (el, c)-nuclear. −→ References [1] M.D. Choi and E.G. Effros, Injectivity and operator spaces, J. Funct. Anal., 1977, 24: 156-209. [2] E.G. Effros and Z.-J. Ruan, Operator spaces, London Math. Soc. Mono. New Series 23, Oxford Univ. Press (2000). [3] K.H. Han and V.I. Paulsen, An approximation theorem for nuclear operator systems, J. Funct. Anal., 2011, 261: 999-1009. [4] A.S. Kavruk, Nuclearity related properties in Operator systems, Journal of Operator Theory, 2014, 71: 95-156. [5] A.S. Kavruk, V.I. Paulsen, I.G. Todorov and A.M. Tomforde, Tensor products of operator systems, J. Funct. Anal., 2011, 261: 267-299. / / 7 O O 8 JIANZE LI [6] A.S. Kavruk, V.I. Paulsen, I.G. Todorov and A.M. Tomforde, Quotients, exactness and WEP in the operator systems category, Adv. Math., 2013, 235: 321C360. [7] E. Kirchiberg, On Subalgebras of the Car-Algebra, J. Funct. Anal., 1995, 129: 35-63. [8] E. Kirchiberg, On non-semisplit extensions, tensor products and exactness of group C ∗-algebras, Invent. Math., 1993, 112: 449-489. [9] P. Luthra and A. Kumar, Nuclearity properties and C-envelopes of operator system inductive limits, arXiv:1612.01421. [10] G. J. Murphy, C ∗-algebras and Operator theory, San Diego: Academic Press, 1990. [11] V.I. Paulsen, Completely bounded maps and Operator algebras, Cambridge Stud. in Adv. Math. 78, Cambridge Univ. Press, 2002. [12] V.I. Paulsen, I.G. Todorov and A.M. Tomforde, Operator system structures on ordered spaces, Proc. London Math. Soc., 2011, 102: 25-49. (Jianze Li) Department of Mathematics, Tianjin University, Tianjin 300072, China E-mail address: [email protected]
1801.03189	1	1801	2018-01-09T23:38:58	A program for finding all KMS states on the Toeplitz algebra of a higher-rank graph	[ "math.OA" ]	The Toeplitz algebra of a finite graph of rank $k$ carries a natural action of the torus ${\mathbb T}^k$, and composing with an embedding of ${\mathbb R}$ in ${\mathbb T}^k$ gives a dynamics on the Toeplitz algebra. For inverse temperatures larger than a critical value, the KMS states for this dynamics are well-understood, and this analysis is the first step in our program. At the critical inverse temperature, much less is known, and the second step in our program is an analysis of the KMS states at the critical value. This is the main technical contribution of the present paper. The third step shows that the problem of finding the states at inverse temperatures less than the critical value is equivalent to our original problem for a smaller graph. Then we can tackle this new problem using the same three steps, and repeat if necessary. So in principle, modulo some mild connectivity conditions on the graph, our results give a complete description of the simplex of KMS states at all inverse temperatures. We test our program on a wide range of examples, including a very general family of graphs with three strongly connected components.	math.OA	math	A PROGRAM FOR FINDING ALL KMS STATES ON THE TOEPLITZ ALGEBRA OF A HIGHER-RANK GRAPH JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN Abstract. The Toeplitz algebra of a finite graph of rank k carries a natural action of the torus Tk, and composing with an embedding of R in Tk gives a dynamics on the Toeplitz algebra. For inverse temperatures larger than a critical value, the KMS states for this dynamics are well-understood, and this analysis is the first step in our program. At the critical inverse temperature, much less is known, and the second step in our program is an analysis of the KMS states at the critical value. This is the main technical contribution of the present paper. The third step shows that the problem of finding the states at inverse temperatures less than the critical value is equivalent to our original problem for a smaller graph. Then we can tackle this new problem using the same three steps, and repeat if necessary. So in principle, modulo some mild connectivity conditions on the graph, our results give a complete description of the simplex of KMS states at all inverse temperatures. We test our program on a wide range of examples, including a very general family of graphs with three strongly connected components. 1. Introduction The graphs of higher-rank k (the k-graphs) were introduced by Kumjian and Pask [13] as combinatorial models for the higher-rank Cuntz -- Krieger algebras of Robertson and Steger [23]. To each k-graph Λ we associate two C ∗-algebras: a graph C ∗-algebra C ∗(Λ) [13] and a Toeplitz algebra T C ∗(Λ) [20]. Both algebras carry natural gauge actions of the torus Tk. Composing these gauge actions with an embedding t 7→ eitr of R in Tk gives actions αr of R on C ∗(Λ) and T C ∗(Λ) (the dynamics). Then one naturally wonders about the KMS states of these dynamics, and, as usual, the results turn out to be interesting [9, 10, 6, 7]. We assume throughout that Λ is a finite k-graph with no sources or sinks. For inverse temperatures β larger than a critical value βc, the system (T C ∗(Λ), R, αr) has a concretely described simplex of KMSβ states with extreme points parametrised by the vertices in the graph [9, Theorem 6.1]. The critical value βc is determined by the spectral radii {ρ(Ai) : 1 ≤ i ≤ k} of the vertex matrices {Ai : 1 ≤ i ≤ k} of Λ as βc = max{r−1 i ln ρ(Ai) : 1 ≤ i ≤ k}. For the "preferred dynamics" in which ri = ln ρ(Ai) for all i, the critical inverse temperature is βc = 1. Roughly speaking, when C ∗(Λ) is simple and the dynamics is not periodic, there is a unique KMS1 state on T C ∗(Λ) ([9, Theorem 7.2], [6, Theorem 5.1(d)]); when the dynamics is preferred it factors through a KMS1 state of C ∗(Λ). Date: January 2, 2018. 2010 Mathematics Subject Classification. 46L30, 46L55. Key words and phrases. C ∗-algebras, higher-rank graphs, KMS states. This research was supported by the Marsden Fund of the Royal Society of New Zealand. 1 2 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN This interesting topic has been picked up by a number of authors from different points of view. In particular, Yang [26] and Laca -- Larsen -- Neshveyev -- Sims -- Webster [15] have computed the types of the von Neumann algebras associated to these KMS states, finding that the type depends only on the skeleton of the k-graph. Farsi, Gillaspy, Kang and Packer [4] have investigated spatial realisations of these KMS states, and found connections with constructions of wavelets. McNamara [16] has extended the results of [9] to more general dynamics. The shifts on the infinite path spaces of higher-rank graphs provided examples of ∗-commuting local homeomorphisms which informed our analysis with Afsar [2]. And, very recently, the uniqueness of the KMS1 state for the preferred dynamics on the C ∗-algebra of a simple graph has been confirmed by Christensen [3] using work of Neshveyev [17] on KMS states of groupoid C ∗-algebras. Simplicity of a graph algebra C ∗(Λ) is largely determined by two properties: Λ should be aperiodic and irreducible [22]. In [10], we studied the KMS states of C ∗(Λ) for periodic irreducible graphs, and we found that the behaviour of the KMS states at the critical inverse temperature 1 is dictated in a very concrete way by the periodicity [10, Theorem 7.1]. We recently investigated graphs with several irreducible components which are themselves aperiodic [11, 7]. We were pleasantly surprised that we were able to get rather complete descriptions of the KMS states for 1-graphs [11, Theorem 5.3], and that we were then able to extend many of the key arguments of [11] to higher-rank graphs. Our general results were strong enough to describe all the KMS1 states for the preferred dynamics on graphs with one or two components [7, §7]. Unfortunately, even for the preferred dynamics, we ran into new difficulties when the graph had three components. The program of [7] involves reducing problems to smaller graphs by describing what happens when we remove hereditary components. The Toeplitz algebras of these smaller graphs are quotients of the Toeplitz algebra T C ∗(Λ). However, the dynamics on the quotient induced by the preferred dynamics on T C ∗(Λ) need not be the preferred one. For graphs with two components, the quotient is irreducible, and the analysis of [6] was available. For graphs with three components, the quotient graph could have two components. So we were forced to rejig our general results to accommodate a non-preferred dynamics, and that is what we report on here. We always scale our dynamics to ensure that βc = 1, and we are interested primarily in the KMS1 states. And we can then bootstrap our techniques to deal with β < 1. Our approach has been to develop a collection of results which allow us to reduce the number of components. Some of these results were already available in some form in our previous articles, and for them our new contribution has been to finesse the arguments to allow for a non-preferred dynamics. (This is the case, for example, for our new Proposition 3.1 and Theorem 5.2.) But some of our ingredients are quite different, and we have been very pleasantly surprised at how well they work. For example, Theorem 4.2 and the results in §6 have no counterparts in our previous papers. These results are intended to be used in tandem, and in §7 we describe a general procedure for computing all the KMS1 states of a very large class of higher-rank graphs. We do have to place some additional hypotheses on our graphs, but these are mainly to ensure that the components themselves are tractable, and that we do not create sources when we remove components. (We saw in [7, §8] that pretty strange things can happen when we remove a component.) We then devise a program KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 3 that finds all KMS states on T C ∗(Λ) at all inverse temperatures. We are delighted to report that we have not hit any new roadblocks to our program. We have in particular tested our program for a non-preferred dynamics on the Toeplitz algebras of graphs with three components, with very satisfactory results (see §9.2 and §9.4). Outline. We begin in §2 by summarising the necessary background material on higher-rank graphs, their associated C ∗-algebras, and KMS states. We then set about adapting the main general results from [7]. In §3 we show how to identify and remove redundant components from our graphs. The key point is that, provided the graph satisfies some mild connectivity hypotheses, we may replace our original graph with a smaller graph in which every critical component is hereditary, and not lose any KMS1 states in the process. We begin §4 by showing how we may extend the Perron -- Frobenius eigenvector for the vertex matrices of a hereditary component of the graph to an eigenvector of the full vertex matrix. This result is linear-algebraic in nature, and its proof involves recasting the proof of [7, Proposition 6.1] to minimise the use of (and im- plicitly the existence of) inverses of matrices. But we were very surprised to discover Theorem 4.2 and Corollary 4.3, which show (again subject to a mild connectivity hypothesis) how the presence of a hereditary component which is critical for one colour influences the ordering of the spectral radii for the other colours. However, the proof of Theorem 4.2 was sufficiently indirect to make us nervous, and in the end we worked hard to find a direct proof for graphs with only 3 vertices, which can be found in Appendix A. In §5 we extend the main result [7, Theorem 6.5] from our previous paper to the situation where the dynamics are non-preferred. When Λ has a critical heredi- tary component D that dominates all other components -- in the sense that all the spectral radii of ΛD dominate the spectral radii of every other component -- Theo- rem 5.2 completely describes the KMS1 states of T C ∗(Λ). This result enables us to deal with graphs that have one or two components (with potentially non-preferred dynamics), but cannot be used if the graph has two or more critical components. In §6 we prove a completely new theorem which allows us to handle graphs with In §7 we combine all of our several components that are critical and hereditary. results to provide a procedure for finding all of the KMS1 states on the Toeplitz algebra of a reducible higher-rank graph. In §8 we describe our program for finding the KMSβ states at all inverse temperatures β. In §9 we test our program on graphs with two components (generalising [7], where the dynamics was the preferred one) and three components. Then we check that for 1-graphs our results are compatible with what we already know for directed graphs [11], and we explicitly compute the KMS states for the Toeplitz algebras of two specific dumbbell 2-graphs with three single-vertex components. 2. Background 2.1. Higher-rank graphs. A higher-rank graph of rank k (or k-graph) consists of a countable category Λ and a degree functor d : Λ → Nk satisfying the following factorisation property: if λ ∈ Λ and d(λ) = m + n for some m, n ∈ Nk, then there are unique µ, ν ∈ Λ such that d(µ) = m, d(ν) = n and λ = µν. We use the following standard notation when working with k-graphs. For n ∈ Nk, we write Λn := d−1(n), and call elements of Λn paths of degree n; in particular, elements of Λ0 are called vertices. The factorisation property implies that we can 4 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN identify the vertices with the objects in the category. We write s, r : Λ → Λ0 for the domain and codomain maps. For E ⊂ Λ and V ⊂ Λ0 we set V E := {λ ∈ E : r(λ) ∈ V } and vE := {v}E, and similarly on the right. We write {ei : 1 ≤ i ≤ k} for the usual generators of Nk, and m ∨ n for the pointwise maximum of m, n ∈ Nk. In this paper all k-graphs are finite in the sense that Λn is finite for each n ∈ Nk. We also assume that our graphs have no sinks and no sources, in the sense that vΛn and Λnv are both non-empty for all v ∈ Λ0 and n ∈ Nk. For 1 ≤ i ≤ k we write Ai for the Λ0 × Λ0 matrix with entries Ai(v, w) := vΛeiw. i i=1 Ani The factorisation property implies that AiAj = AjAi for each 1 ≤ i, j ≤ k. Then is well defined and has entries given by An(v, w) = vΛnw. For C, D ⊂ Λ0, we write AC,D,i for the C × D matrix with entries AC,D,i(v, w) = Ai(v, w) for v ∈ C, w ∈ D, and we write AC,i := AC,C,i. for each n ∈ Nk, the matrix An := Qk The skeleton is the coloured directed graph (Λ0, Λ1 :=Sk i=1 Λei, r, s) in which the edges of degree ei have been assigned one of k different colours. The relationship between k-graphs and skeletons is discussed in [19, Chapter 10], [21, Section 2], and [5]. When k = 2, we think of the elements of Λe1 as blue/solid edges and elements of Λe2 as red/dashed edges. In any k-graph, the factorisation property gives bijections θi,j of the sets {ef : e ∈ Λei, f ∈ Λej } onto {gh : g ∈ Λej , h ∈ Λei}. However, there are coloured graphs for which there exist such bijections that are not the skeletons of any k-graph (see [14, Example 5.15(ii)]). Fortunately, when k = 2 the situation is simple: a 2-coloured graph is the skeleton of a 2-graph if and only if A1A2 = A2A1 [13, Section 6]. (Though then it is usually the skeleton of many 2-graphs.) 2.2. The C ∗-algebras of higher-rank graphs. We consider a finite k-graph Λ with no sinks and no sources. For λ, µ ∈ Λ, we define Λmin(λ, µ) := {(α, β) ∈ Λ × Λ : λα = µβ ∈ Λd(λ)∨d(µ)}. A collection of partial isometries {Tλ : λ ∈ Λ} in a C ∗-algebra B is a Toeplitz -- Cuntz -- Krieger Λ-family if (T1) {Qv := Tv : v ∈ Λ0} is a collection of mutually orthogonal projections; (T2) TλTµ = Tλµ for each λ, µ ∈ Λ with s(λ) = r(µ); (T3) T ∗ β for all λ, µ ∈ Λ. Relation (T3) implies that T ∗ projections {TλT ∗ λ Tµ =P(α,β)∈Λmin(λ,µ) TαT ∗ (T2) shows that Qv ≥Pλ∈vΛn TλT ∗ C ∗({Tλ : λ ∈ Λ}) = span{TλT ∗ µ : λ, µ ∈ Λ}. λ Tλ = Qs(λ) for λ ∈ Λ, and that for each n ∈ Nk, the λ : λ ∈ Λn} are mutually orthogonal. Combining this with relation λ for each n ∈ Nk. The relations also imply that The Toeplitz C ∗-algebra T C ∗(Λ) is generated by a universal Toeplitz -- Cuntz -- Krieger Λ-family {tλ : λ ∈ Λ} [20, §7]. The Cuntz -- Krieger (or graph) algebra C ∗(Λ) is the quotient of T C ∗(Λ) in which qv =Pλ∈vΛn tλt∗ For z = (z1, . . . , zk) ∈ Tk and n ∈ Nk, we write zn :=Qk continuous gauge action γ : Tk → Aut(T C ∗(Λ)) satisfying γz(tλ) =Qk z ∈ Tk and λ ∈ Λ. Since γz fixes each qv −Pλ∈vΛn tλt∗ on C ∗(Λ), which we also denote by γ. . There is a strongly tλ for λ, this action induces an action λ for all v ∈ Λ0 and n ∈ Nk. i=1 zd(λ)i i i=1 zni i 2.3. Hereditary sets and strongly connected components. A subset H ⊂ Λ0 is hereditary if v ∈ H and vΛw 6= ∅ =⇒ w ∈ H, and forwards hereditary if w ∈ H and vΛw 6= ∅ =⇒ v ∈ H. If H is hereditary then Λ\H := {λ ∈ Λ : s(λ) 6∈ H} is KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 5 a k-graph (the key point is that µν ∈ Λ\H implies µ, ν ∈ Λ\H). If IH is the ideal of T C ∗(Λ) generated by {qv : v ∈ H}, then T C ∗(Λ\H), T C ∗(Λ)/IH, and C ∗({tλ : s(λ) 6∈ H}) ⊂ T C ∗(Λ) are all canonically isomorphic (see [7, Proposition 2.2]). Thus the Toeplitz algebra of the subgraph Λ\H can be realised as both a quotient and a subalgebra of T C ∗(Λ). There is an equivalence relation ∼ on Λ0 such that v ∼ w ⇐⇒ vΛw 6= ∅ and wΛv 6= ∅, and the strongly connected components of Λ are the equivalence classes for this relation. If vΛv = {v} for a vertex v then {v} is a strongly connected component, and we call such classes trivial components. We write C for the set of nontrivial strongly connected components. For C ∈ C we set ΛC := CΛC. It is routine to check that if µν ∈ ΛC, then µ, ν ∈ ΛC, and so ΛC is a k-graph. An n × n matrix A is irreducible if for every 1 ≤ i, j ≤ n there exists m ∈ N such that Am(i, j) 6= 0. If A is the vertex matrix of a directed graph E, then A is irreducible if and only if E is strongly connected. Following [7], we say that the subgraph ΛC is coordinatewise irreducible if all the matrices AC,i are irreducible. Then [9, Lemma 2.1] says that the matrices {AC,i : 1 ≤ i ≤ k} have a common unimodular Perron -- Frobenius eigenvector. If the graph has no sources or sinks, and each ΛC is coordinatewise irreducible, [7, Proposition 3.1] says that we can order Λ0 so that each vertex matrix Ai is block upper triangular, the diagonal blocks are {AC,i : C ∈ C}, and the other blocks are strictly upper triangular matrices. Hence for every 1 ≤ i ≤ k, the spectral radius of Ai is ρ(Ai) = max{ρ(AC,i) : C ∈ C}. 2.4. KMS states. Suppose that α is an action of R on a C ∗-algebra A. An element a ∈ A is analytic for the action α if the map t 7→ αt(a) extends to an analytic function on C. A state φ of A is a KMS state with inverse temperature β ∈ (0, ∞) (or a KMSβ state) if φ(ab) = φ(bαiβ(a)) for all analytic elements a, b ∈ A. By [18, Proposition 8.12.3] it suffices to verify that the state φ satisfies the KMS condition on a set of analytic elements which span a dense invariant subset of A. 2.5. The dynamics on T C ∗(Λ). Let Λ be a k-graph. Choosing r ∈ (0, ∞)k gives a dynamics αr : R → Aut(T C ∗(Λ)) by αr t = γeitr := γ(eitr1 ,...,eitrk ). t (tλt∗ µ) = eitr·(d(λ)−d(µ))tλt∗ For λ, µ ∈ Λ, the map t 7→ αr µ is the restriction of the analytic function z 7→ eizr·(d(λ)−d(µ))tλt∗ µ is an analytic element. Since µ : λ, µ ∈ Λ} is dense in T C ∗(Λ), it suffices to check the KMSβ condition span{tλt∗ on pairs of elements of the form tλt∗ µ. The action αr induces an action, also denoted αr, on the quotient C ∗(Λ). We say that a KMSβ state φ of T C ∗(Λ) factors through C ∗(Λ) if there exists a KMSβ state ψ of C ∗(Λ) such that φ = ψ ◦ q, where q is the quotient map. µ, and hence tλt∗ In Theorem 5.2 and beyond, we assume that the coordinates of the vector r are rationally independent. This implies that t 7→ eitr is an embedding of R in Tk. (To see this, suppose that eitr = 1. Then there are integers ni such that tri = 2πni. But then rin−1 and we have njri − nirj = 0, which contradicts that the coordinates of r are rationally independent.) It then follows that the dynamics αr is not periodic. i = rjn−1 j 6 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN 2.6. Scaling the dynamics. We usually scale the dynamics (that is, multiply r by a scalar, which changes the inverse temperature but does not otherwise affect the dynamics [6, §2.1]) to ensure that βc := max{r−1 ln ρ(Ai) : 1 ≤ i ≤ k} is 1. i 2.7. The preferred dynamics. Suppose that Λ is a coordinatewise-irreducible finite k-graph. Then by [9, Corollary 4.4 and Theorem 7.2], there is a KMS1 state of (C ∗(Λ), αr) if and only if ri = ln ρ(Ai) for 1 ≤ i ≤ k. This means αr is the preferred dynamics, which is characterised by k αr(tλ) = ρ(A)d(λ)tλ := ρ(Ai)d(λ)itλ. Yi=1 For the preferred dynamics, if the coordinates of r are rationally independent, then (T C ∗(Λ), αr) has a unique KMS1 state, and it factors through a state of (C ∗(Λ), αr) [9, Theorem 7.2]. So if we were interested in KMSβ states on (C ∗(Λ), αr), then we would only be interested1 in the preferred dynamics and the inverse temperature β = 1. 3. Removing redundant components We consider a finite k-graph with no sources or sinks. We suppose that r ∈ (0, ∞)k satisfies (3.1) equivalently, max{r−1 i ln ρ(Ai) : 1 ≤ i ≤ k} = 1; ri ≥ ln ρ(Ai) for all i, and K := {i : r−1 i ln ρ(Ai) = 1} 6= ∅. Here it is important that K may be a proper subset of {1, . . . , k}, in which case the dynamics αr on T C ∗(Λ) is not the preferred dynamics studied in [7]. In this section we salvage what we can from the arguments used to prove [7, Theorem 5.1]. We focus on the critical components: a nontrivial strongly connected component C ∈ C such that rj = ln ρ(AC,j) for some j ∈ K and AC,j is irreducible. We then say that C is j-critical. The next result says that KMS1 states do not see many vertices which feed into critical components. Proposition 3.1. Suppose that Λ is a finite k-graph with no sources or sinks, and that r ∈ (0, ∞)k satisfies (3.1). Suppose that C ∈ C is a j-critical component of Λ, and set ΣjC := {w ∈ Λ0 : CΛNej w 6= ∅}. Then for every KMS1 state ψ on (T C ∗(Λ), αr), we have ψ(qw) = 0 for all w ∈ Hj := (ΣjC)\C. Proof. Since j ∈ K and rj = ln ρ(AC,j), we have ρ(AC,j) = ρ(Aj). Thus [9, Propo- sition 4.1(a)] implies that the vector mψ :=(cid:0)ψ(qv)(cid:1) satisfies Ajmψ ≤ erj mψ = ρ(Aj)mψ. (3.2) With respect to the decomposition Λ0 = (Λ0\ΣjC) ∪ C ∪ Hj, the vertex matrix Aj has block form 1Curiously, in her study of k-graphs with a single vertex, Yang also arrived at the preferred dynamics, but for different reasons (see [25, Proposition 5.4]). AΛ0\Σj C,j ⋆ ⋆ 0 0 AC,j AC,Hj ,j AHj ,j 0 Aj =    . KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 7 We fix w ∈ Hj. Then there exits n ≥ 1 and v ∈ C such that vΛnejw 6= ∅. We first suppose that mψC = 0. Then we have 0 ≤ An j (v, w)mψ w = (An j )C,Hj (v, w)mψ w (An j )C,Hj (v, u)mψ u =(cid:0)(An j )C,Hj mψHj(cid:1)v ≤ Xu∈Hj ≤(cid:0)An v j (v, w) > 0, these estimates force mψ j mψ(cid:1)v ≤ ρ(Aj)nmψ w = 0. by (3.2). Since An So we suppose that mψC 6= 0. Now we recall that ρ(AC,j) = ρ(Aj), and read off from the central block of (3.2) that AC,jmψC ≤ (Ajmψ)C ≤ ρ(Aj)mψC = ρ(AC,j)mψC. Since AC,j is irreducible and mψC 6= 0, the subinvariance theorem [24, Theorem 1.6] implies that With n and v as above we have An j (v, w) > 0. Then (3.2) gives AC,jmψC = ρ(AC,j)mψC. ρ(AC,j)nmψ v = ρ(Aj)nmψ v ≥(cid:0)An C,j(v, u)mψ An j mψ(cid:1)v u + Xu∈Hj An C,j(v, u)mψ u + An j (v, w)mψ w An j (v, u)mψ u =Xu∈C ≥Xu∈C = ρ(AC,j)nmψ v + An j (v, w)mψ w. Thus mψ w = 0, as required. (cid:3) Corollary 3.2. Suppose that we have Λ, r, C and j as in Proposition 3.1, and that the set Hj in that proposition is hereditary in Λ. Then every KMS1 state of (T C ∗(Λ), αr) factors through a state of (T C ∗(Λ\Hj), αr). Proof. Suppose that ψ is a KMS1 state of (T C ∗(Λ), αr). Proposition 3.1 implies that ψ vanishes on the set P := {qw : w ∈ Hj}. As at the end of the proof of [7, Theorem 5.1], it follows from [8, Lemma 2.2] that ψ vanishes on the ideal IHj gener- ∼= ated by P . Since Hj is hereditary, [7, Proposition 2.2] implies that T C ∗(Λ)/IHj T C ∗(Λ\Hj), and the result follows. (cid:3) Corollary 3.2 applies in particular if CΛNej w 6= ∅ for all vertices w, and then says that every KMS1 state of (T C ∗(Λ), αr) factors though a state of (T C ∗(ΛC), αr). However, it is possible that Hj = ∅. For example, suppose that Λ is a dumbbell graph as in Example 4.4 with p1 = 0 (and then necessarily with m1 = n1). Then Corollary 3.2 gives no information. When the set Hj of Proposition 3.1 is not hereditary, Λ\Hj may not be a k-graph. Our next goal is to prove that, provided Λ is sufficiently connected, there exists a hereditary subset H such that every KMS1 state of T C ∗(Λ) factors through a KMS1 state of T C ∗(Λ\H). We start with a lemma. Lemma 3.3. Suppose that Λ is a finite k-graph with no sources or sinks. Suppose that S{C : C ∈ C} = Λ0, that all the graphs {ΛC : C ∈ C} are coordinatewise irreducible and that for components C, D ∈ C we have CΛNej D 6= ∅ for some j =⇒ CΛNeiD 6= ∅ for 1 ≤ i ≤ k. (3.3) 8 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN Suppose that v ≤ w, in the sense that vΛw 6= ∅. Then vΛNeiw 6= ∅ for 1 ≤ i ≤ k. Proof. Suppose that λ ∈ vΛw and 1 ≤ i ≤ k. Since Λ0 = S{C : C ∈ C}, we can factor λ = µ1λ1 . . . µmλmµm+1, where each µi has range and source in the same component and each λi is an edge with range and source in different components. For 1 ≤ n ≤ m + 1, let Cn be the component such that µn ∈ ΛCn. Since ΛCn is coordinatewise irreducible, there are paths µ′ n) = r(µn) and s(µ′ n) = s(µn). Let Dn be the components such that λn ∈ CnΛej Dn for some j. Then (3.3) implies that there exists λ′ n ∈ CnΛNeiDn; since both ΛCn and ΛDn are coordinatewise irreducible, we can suppose also that λ′ n ∈ r(λ)ΛNeis(λn). Then 1 . . . µ′ 1λ′ µ′ (cid:3) We consider the partial order on components such that C ≤ D ⇐⇒ CΛD 6= ∅. This restricts to a partial order on the set Ccrit of critical components. Since Ccrit ⊂ C is finite, there are elements of Ccrit which are minimal in this partial order. We denote the set of these minimal elements by Cmincrit. m+1 is a path in vΛNeiw. n ∈ ΛNei Cn such that r(µ′ mλ′ mµ′ Proposition 3.4. Suppose that Λ is a finite k-graph with no sources or sinks, and graphs {ΛC : C ∈ C} are coordinatewise irreducible, and that (3.3) holds. Let H ′ that r ∈ (0, ∞)k satisfies (3.1). Suppose that S{C : C ∈ C} = Λ0, that all the be the hereditary closure of G := S{C : C ∈ Cmincrit}, and set H := H ′\G. Then H is hereditary, Cmincrit is the set of all critical components in the graph Λ\H, and each C ∈ Cmincrit is hereditary in Λ\H. Every KMS1 state of (T C ∗(Λ), αr) factors through a KMS1 state of (T C ∗(Λ\H), αr). Proof. To see that H is hereditary, suppose that v ∈ H and v ≤ w. Since v belongs to the hereditary closure of G, there exists C ∈ Cmincrit such that C ≤ v. But then C ≤ w also, and w ∈ H ′. Suppose, looking for a contradiction, that w ∈ G. Then there exists D ∈ Cmincrit such that w ∈ D. Then CΛD 6= ∅, and minimality forces C = D. Then w ∈ C forces v ∈ C, which is not possible because v ∈ H = H ′\G. So w is not in G, and w ∈ H. Thus H is hereditary. Since we have removed all the critical components that are not minimal, and have not added any new ones, the critical components of Λ\H are those in Cmincrit. To see that C ∈ Cmincrit is hereditary in Λ\H, take v ∈ (Λ\H)0 = Λ0\H such that CΛv 6= ∅. Then v belongs to the hereditary closure H ′, and since H ′ = G ∪ H, we have v ∈ G. So there exists D ∈ Cmincrit such that v ∈ D. Then C ≤ D, and minimality of D forces D = C. Thus v ∈ C, as required. Now we suppose that φ is a KMS1 state of (T C ∗(Λ), αr). Let w ∈ H. Then there exists C ∈ Cmincrit such that CΛw 6= ∅. Then C is j-critical for some j. By Lemma 3.3, CΛNej w 6= ∅. Then w belongs to the set ΣjC of Proposition 3.1, and that proposition implies that φ(qw) = 0. Thus φ(qw) = 0 for all w ∈ H. We deduce from [1, Lemma 6.2] (for example2) that φ vanishes on the ideal generated by {qw : w ∈ H}. This ideal is IH, and [7, Proposition 2.2] shows that the canonical map qH : T C ∗(Λ) → T C ∗(Λ\H) induces an isomorphism of T C ∗(Λ)/IH onto T C ∗(Λ\H). Thus φ factors through a state of T C ∗(Λ\H) which is necessarily a KMS1 state of (T C ∗(Λ\H), αr). (cid:3) Remark 3.5. The hypotheses of Proposition 3.4 imply that the graph Λ\H also has no sources or sinks. In [7, §8], we showed how removing a hereditary component can create sources and hence extra KMS states; Examples 8.4 and 8.5 in [7] show 2We could also use [8, Lemma 2.2] for this, but [1, Lemma 6.2] is easier to use and much more general. KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 9 that quite a variety of things can happen if there are complicated bridges between components. Here, where we are focusing on graphs which have no long bridges between components, Proposition 3.4 says that it suffices to study the KMS1 states for graphs in which every critical component is hereditary. 4. Extending the Perron -- Frobenius eigenvector The following is a strengthening of [7, Proposition 6.1], which is a linear-algebraic result about the vertex matrices of a finite k-graph. We recover [7, Proposition 6.1] when the set LD at (4.1) is {1, . . . , k}. In the applications, the eigenvector x will be the common unimodular Perron -- Frobenius eigenvector of the commuting irreducible matrices {AD,i}. It is important for the application in Theorem 4.2 that the set LD in (4.1) can be a proper subset of {1, . . . , k}. Proposition 4.1. Suppose that Λ is a finite k-graph without sinks or sources. Sup- pose that D ∈ C is hereditary, and set H := {v ∈ Λ0 : vΛD = ∅}. Suppose that (4.1) LD :=(cid:8)i ∈ {1, . . . , k} : ρ(AC,i) < ρ(AD,i) for C ∈ C\{D} with CΛD 6= ∅(cid:9) is nonempty, and that x is a nonnegative eigenvector of AD,i with eigenvalue ρ(AD,i) for 1 ≤ i ≤ k. Write F := Λ0\(D ∪ H). Then with respect to the decomposition Λ0 = F ⊔ D ⊔ H, the vertex matrices have block form Ei Bi 0 AD,i 0 0 Ai =  (a) For 1 ≤ i, j ≤ k we have ⋆ 0 AH,i   for 1 ≤ i ≤ k. (4.2) (ρ(AD,i)1F − Ei)Bjx = (ρ(AD,j)1F − Ej)Bix. (b) For i, j ∈ LD, we have (4.3) (ρ(AD,i)1F − Ei)−1Bix = (ρ(AD,j)1F − Ej)−1Bjx; write y for the common vector (4.3). Then y is nonnegative, and for every i ∈ {1, . . . , k}, (y, x, 0) is an eigenvector of Ai with eigenvalue ρ(AD,i). Proof. Since D is hereditary, we have DΛ(Λ0\D) = ∅, and the block decomposition has the required form. Suppose that i, j ∈ {1, . . . , k}. Since AiAj = AjAi, the block form of the product gives EiEj = EjEi and (4.4) EiBj + BiAD,j = EjBi + BjAD,i. Since x is an eigenvector for both AD,i and AD,j, (4.4) gives (ρ(AD,i)1F − Ei)Bjx = ρ(AD,i)Bjx − EiBjx = ρ(AD,i)Bjx − (EjBix + BjAD,ix − BiAD,jx) = ρ(AD,i)Bjx − (EjBix + Bjρ(AD,i)x − BiAD,jx) = BiAD,jx − EjBix = (ρ(AD,j)1F − Ej)Bix. Thus we have (4.2), and we have proved (a). Next we take i, j ∈ LD. Then ρ(AD,i) > max(cid:8)ρ(AC,i) : C ∈ C\{D} and CΛD 6= ∅(cid:9) = ρ(Ei), 10 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN and similarly for j, so the matrices ρ(AD,i)1F −Ei and ρ(AD,j)1F −Ej are invertible. They also commute, and hence so do their inverses. So we can multiply (4.2) by (ρ(AD,i)1F − Ei)−1(ρ(AD,j)1F − Ej)−1 to get (4.3). Thus we can define y as claimed. Since x ≥ 0 and Bi has nonnegative entries, the expansion (ρ(AD,i)1F − Ei)−1 = ρ(AD,i)−1(1F − ρ(AD,i)−1Ei)−1 ∞ shows that y is nonnegative. To prove that z := (y, x, 0) is an eigenvector of every vertex matrix Aj with eigenvalue ρ(AD,j), we fix j. Then = ρ(AD,i)−1 ρ(AD,i)−nEn i Xn=0 Ajz =  Ei Bi 0 AD,i 0 0 ⋆ 0 AH,i     y x 0  =  Ejy + Bjx AD,jx 0   =  Ejy + Bjx ρ(AD,j)x 0   . We now take i ∈ LD, and work on the top block in Ajz. Since EiEj = EjEi it follows that Ej and (ρ(AD,i)1F − Ei)−1 commute. Thus Ejy + Bjx = (ρ(AD,i)1F − Ei)−1EjBix + Bjx At this point we use (4.2), finding = (ρ(AD,i)1F − Ei)−1(cid:0)ρ(AD,j)1F − (ρ(AD,j)1F − Ej)(cid:1)Bix + Bjx. Ejy + Bjx = (ρ(AD,i)1F − Ei)−1(cid:0)ρ(AD,j)Bix − (ρ(AD,i)1F − Ei)Bjx(cid:1) + Bjx = ρ(AD,j)(ρ(AD,i)1F − Ei)−1Bix − Bjx + Bjx = ρ(AD,j)y. Thus z is an eigenvector of every Aj with eigenvalue ρ(AD,j), and this completes the proof of (b). (cid:3) The next application of Proposition 4.1 has no analogue in [7]. We find it curious that the proof of Theorem 4.2 involves a non-trivial application of the subinvariance theorem from Perron -- Frobenius theory, and wonder whether a more direct linear- algebraic proof is possible. We give such a proof for graphs with three components in Appendix A. (If so, we could then deduce the stronger-looking Proposition 4.1 from [7, Proposition 6.1].) In the statement of Theorem 4.2, we have removed the hereditary set H from the set-up of Proposition 4.1; if there is such a set, we can apply Theorem 4.2 to Λ\H, but then we only get information about ρ(AC,i) for components C ∈ C with C ⊂ Λ0\H. Theorem 4.2. Suppose that Λ is a finite k-graph without sinks or sources, and that the subgraphs {ΛC : C ∈ C} are all coordinatewise irreducible. Suppose that D ∈ C is hereditary, and that CΛeiD 6= ∅ for all C ∈ C and 1 ≤ i ≤ k. If there exists j ∈ {1, . . . , k} such that ρ(AC,j) < ρ(AD,j) for all C ∈ C\{D}, then for all C ∈ C\{D} and 1 ≤ i ≤ k. ρ(AC,i) < ρ(AD,i) Proof. Since ΛD is coordinatewise irreducible, its vertex matrices {AD,i : 1 ≤ i ≤ k} are a commuting family of irreducible matrices, and hence by [9, Lemma 2.1] have a common unimodular Perron -- Frobenius eigenvector x. By assumption, the set LD in Proposition 4.1 is nonempty, and hence that proposition gives a vector KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 11 y ∈ [0, ∞)Λ0\D such that z := (y, x) is an eigenvector of each Ai with eigenvalue ρ(AD,i). Now we take C ∈ C\{D}, and consider the block yC. Fix i ∈ {1, . . . , k}. The AC,iyC ≤ AC,iyC + AC,D,ix ≤ (Aiz)C = ρ(AD,i)zC = ρ(AD,i)yC. C × C block of Ei is AC,i, and the C × D block of Bi is AC,D,i. Thus (4.5) Since CΛeiD 6= ∅, and since x has strictly positive entries, the vector AC,D,ix is nonzero. So yC 6= 0 and the first inequality in (4.5) is strict. So (4.5) implies that yC is a subinvariant vector for the irreducible matrix AC,i, and that it is not an eigenvector. So the subinvariance theorem [24, Theorem 1.6] implies that ρ(AC,i) < ρ(AD,i). (cid:3) The next corollary strengthens Theorem 4.2 by allowing longer singly-coloured bridges between components. Corollary 4.3. Suppose that Λ is a finite k-graph without sinks or sources, and that the subgraphs {ΛC : C ∈ C} are all coordinatewise irreducible. Suppose that D ∈ C is hereditary and satisfies CΛNeiD 6= ∅ for all C ∈ C and 1 ≤ i ≤ k. If there exists j ∈ {1, . . . , k} such that ρ(AC,j) < ρ(AD,j) for all C ∈ C\{D}, then for all C ∈ C\{D} and 1 ≤ i ≤ k. ρ(AC,i) < ρ(AD,i) Proof. For each C ∈ C and i there exists NC,i ∈ N such that CΛNC,ieiD 6= ∅. Since ΛC and ΛD are coordinatewise irreducible, we then have CΛpeiD 6= ∅ for all p ≥ NC,i. Thus there exist Ni ∈ N such that CΛNieiD 6= ∅ for all C. (If some AC,i is a permutation matrix, then also choose Ni coprime to C, so that ANi C,i is also irreducible.) Now we set N := (N1, . . . , Nk) and N Nk :=(cid:8)Nn := (N1n1, . . . , Nknk) : n ∈ Nk(cid:9). Then Λ(N) := {λ ∈ Λ : d(λ) ∈ N Nk} is a k-graph, with the same range and source maps as Λ and degree functor d(N) : Λ(N) → Nk defined by Nd(N)(λ) = d(λ). The choice of the Ni ensures that Λ(N) satisfies the hypotheses of Theorem 4.2. Since the vertex matrices A(N)i of Λ(N) satisfy A(N)C,i = ANi C,i, applying Theorem 4.2 to Λ(N) gives the result. (cid:3) Example 4.4. We consider a 2-graph Λ with skeleton of the form shown in Figure 1 (known as a dumbbell graph). With the vertex set ordered alphabetically, the vertex m2 m1 v p2 p1 w n2 n1 Figure 1. A dumbbell 2-graph with 2 components. matrices of Λ are A1 =(cid:18)m1 p1 n1(cid:19) and A2 =(cid:18)m2 p2 n2(cid:19) . 0 0 The factorisation property implies that A1A2 = A2A1, and (4.6) (A1A2)(v, w) = (A2A1)(v, w) ⇐⇒ m1p2 + p1n2 = m2p1 + p2n1 ⇐⇒ (n2 − m2)p1 = (n1 − m1)p2. 12 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN Thus the graph in Figure 1 is the skeleton of a 2-graph if and only if (4.6) holds. If p1 and p2 are both nonzero, then (4.6) implies that n1 < m1 ⇐⇒ n2 < m2, as predicted by Theorem 4.2. However, if exactly one pi is zero, say p1 = 0, then (4.6) implies that m1 = n1 and imposes no restriction on m2 and n2. So we can- not remove the hypotheses CΛeiD 6= ∅ from Theorem 4.2 or CΛNeiD 6= ∅ from Corollary 4.3 (though we could possibly weaken them if we wanted to allow trivial strongly connected components, as in [7, §8], for example). 5. Dominant components The other main general result in [7] is Theorem 6.5, which describes the KMS1 states of T C ∗(Λ) when there is a hereditary component where all the spectral radii are attained. Here we seek a version of [7, Theorem 6.5] for a non-preferred dynamics. From now on we assume that ρ(AC,i) > 1 for 1 ≤ i ≤ k and C ∈ C. Since we usually assume that the graphs ΛC are coordinatewise irreducible, the assumption ρ(AC,i) > 1 merely removes the possibility that ΛC,i consists of a single cycle, which even for 1-graphs is known to be an exceptional case. Proposition 5.1. Suppose that Λ is a finite k-graph with no sources or sinks, and choose a vector r ∈ (0, ∞)k satisfying (3.1). We suppose that D ∈ C is hereditary, that ΛD is coordinatewise irreducible, and that there exists j ∈ {1, . . . , k} such that (5.1) ρ(AC,j) < ρ(AD,j) for C ∈ C\{D} such that CΛD 6= ∅. We let x be the common Perron -- Frobenius eigenvector of the {AD,i : 1 ≤ i ≤ k}, and take z = (y, x, 0) as in Proposition 4.1. Write b := kzk1. Then there is a KMS1 state ψ of (T C ∗(Λ), αr) such that ψ(tµt∗ (5.2) This state factors through a state of C ∗(Λ) if and only if ri = ln ρ(AD,i) for all i. ν) = δµ,νe−r·d(µ)b−1zs(µ). Outline of proof. Since the argument is very similar to that of [7, Proposition 6.3], we merely outline the argument. Since {AD,i : 1 ≤ i ≤ k} is a commuting family of irreducible nonnegative matrices, Lemma 2.1 of [9] implies that they have a common unimodular Perron -- Frobenius eigenvector x ∈ (0, ∞)D. Let Λ0 = F ⊔ D ⊔ H be the decomposition of Proposition 4.1. The proposition then implies that there is a vector y ∈ [0, ∞)F such that z := (y, x, 0) is a common eigenvector of each Ai with eigenvalue ρ(AD,i). Following the proof of [7, Proposition 6.3], we choose a decreasing sequence {βp} ⊂ (1, ∞) such that βp → 1 as p → ∞. Then βpri ≥ βp ln ρ(Ai) > ln ρ(Ai) ≥ ln ρ(AD,i) for all i, and hence ǫp := k k Yi=1(cid:0)1 − e−βpriAi(cid:1)b−1z = Yi=1(cid:0)1 − e−βpriρ(AD,i)(cid:1)b−1z ≥ 0. Then [9, Theorem 6.1] gives KMSβp states φǫp of (T C ∗(Λ), αr), from which a weak* compactness argument gives a KMS1 state ψ of (T C ∗(Λ), αr) such that ψ(tµt∗ ν) = lim p→∞ δµ,νe−βpr·d(µ)b−1zs(µ) = δµ,νe−r·d(µ)b−1zs(µ). KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 13 By [1, Lemma 6.2], ψ factors through q : T C ∗(Λ) → C ∗(Λ) if and only if ψ(cid:16)qv − Xλ∈vΛn Thus for the last comment, we take v ∈ Λ0, n ∈ Nk and compute: tλtλ(cid:17) = 0 for all v ∈ Λ0 and n ∈ Nk. e−r·nb−1zs(λ) = Xw∈Λ0(cid:18) k Yi=1 Ani i (cid:19)(v, w)e−r·nb−1zw Yi=1 ρ(AD,i)nizv k = e−r·nb−1 k ψ(cid:18) Xλ∈vΛn tλt∗ λ(cid:19) = Xλ∈vΛn = e−r·nb−1(cid:18)(cid:18) k Yi=1 k Ani i (cid:19)z(cid:19)v = b−1 (e−riρ(AD,i))nizv = Yi=1 It follows that ψ factors through q if and only if Qk i=1(e−riρ(AD,i))ni = 1 for every n ∈ Nk. Since each e−riρ(AD,i) ≤ 1, this last condition holds if and only if ri = ln ρ(AD,i) for all i. (cid:3) (e−riρ(AD,i))niψ(qv). Yi=1 We can now state our new version of [7, Theorem 6.5]. Notice that the hypothesis (5.3) is substantially stronger than the hypothesis (5.1) in Proposition 5.1; it implies that there are no other critical components. This hypothesis was crucial in the proof of [7, Theorem 6.5], which we follow. Theorem 5.2. Suppose that Λ is a finite k-graph with no sources or sinks, and choose a vector r ∈ (0, ∞)k satisfying (3.1), and with rationally independent coordi- nates. We suppose that D ∈ C is hereditary, that ΛD is coordinatewise irreducible, that there exists j such that rj = ln ρ(AD,j), and such that (5.3) ρ(AC,i) < ρ(AD,i) for 1 ≤ i ≤ k and C ∈ C\{D}. Write qD for the quotient map of T C ∗(Λ) onto T C ∗(Λ\D). Then every KMS1 state of (T C ∗(Λ), αr) is a convex combination of the state ψ of Proposition 5.1 and a state φ ◦ qD lifted from a KMS1 state φ of (T C ∗(Λ\D), αr). Proof. We begin by observing that we can swap the colours around to ensure that there exists k′ such that ri = ln ρ(Ai) ⇐⇒ k′ ≤ i ≤ k. In particular, we then have rk = ln ρ(Ak) = ln ρ(AD,k). Now suppose that θ is a KMS1 state of (T C ∗(Λ), αr). Then [9, Proposition 4.1] implies that the vector mθ := ( θ(qv) ) ∈ [0, 1]Λ0 satisfies Aimθ ≤ erimθ for all i. Looking at the block structure of Ai for the decomposition Λ0 = (Λ0\D) ⊔ D shows that the vector mθD satisfies and hence AD,i(cid:0)mθD(cid:1) ≤ erimθD for 1 ≤ i ≤ k, AD,i(cid:0)mθD(cid:1) ≤ ρ(Ai)mθD = ρ(AD,i)mθD for k′ ≤ i ≤ k. Since each AD,i is irreducible, the subinvariance theorem [24, Theorem 1.6] implies that AD,i(cid:0)mθD(cid:1) = ρ(AD,i)mθD for k′ ≤ i ≤ k. 14 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN Since the matrices {AD,i : 1 ≤ i ≤ k} have just one common unimodular Perron -- Frobenius eigenvector x, we deduce that there exists a ∈ [0, ∞) such that mθD = ab−1x, for b := kzk1 as in Proposition 5.1. We could have a = 0, in which case it follows from [1, Lemma 6.2] that θ factors through qD. We aim to prove that a ≤ 1, as in the third and fourth paragraphs of the proof of [7, Theorem 6.5], by showing that θ(qv) ≥ aψ(qv) for all v. As in [7], the interesting case is when v belongs to a component C with CΛD 6= ∅, and for this we follow the calculation in the fourth paragraph, working in the coordinate graph Λk. Now we could have a = 1, and then we get θ = ψ, as in [7]. (This uses the direction of [9, Proposition 3.1(b)] which requires rational independence of the ri.) So we are left with the case 0 < a < 1, in which case we have to construct a KMS1 state φǫ of T C ∗(Λ\D) by applying [9, Theorem 6.1] to the graph Λ\D. By (5.3) we have ri ≥ ln ρ(Ai) ≥ ln ρ(AD,i) > ln ρ(AC,i) for 1 ≤ i ≤ k and C ∈ C\{D}, and hence ri > ln ρ(AΛ0\D,i) for all i. Therefore KMS1 states of T C ∗(Λ\D) have the form φǫ as described in [9, Theorem 6.1].3 Define κ := (1 − a)−1(mθ − amψ)Λ0\D, write each Ai in block form with Ei := AΛ0\D,i, and take k ǫ := Yi=1(cid:0)1 − e−riEi(cid:1)κ. The argument in the seventh paragraph of the proof in [7] shows that ǫ ≥ 0. (This is the reason we swapped the colours around at the start: the kth matrix Ak plays a special role in that calculation, and it is crucial that ρ(Ak) = ρ(AD,k).) We have kκk1 = (1 − a)−1(cid:16) Xv∈Λ0\D = (1 − a)−1(cid:16)1 −Xv∈D = (1 − a)−1(cid:16)1 −Xv∈D θ(qv) − a Xv∈Λ0\D θ(qv) − a(cid:0)1 −Xv∈D ab−1xv − a + aXv∈D ψ(qv)(cid:17) ψ(qv)(cid:1)(cid:17) b−1xv(cid:17) = 1. Thus ǫ belongs to the simplex Σ1 of [9, Theorem 6.1]. We then finish off by following the argument in the last two paragraphs of the proof in [7] to see that φ = (1−a)φǫ ◦ qD + aψ. (This is where we use again that the matrices Ai satisfy (5.3), because we need all the matrices 1 − e−riEi to be invertible.) (cid:3) Corollary 5.3. Suppose that we have a k-graph Λ with the properties in Theo- rem 5.2, and suppose in addition that Λ\D does not have sources. Then there is a KMS1 state of (C ∗(Λ), αr) if and only if ri = ln ρ(AD,i) for 1 ≤ i ≤ k. Proof. A convex combination of states of T C ∗(Λ) factors through C ∗(Λ) if and only if all the summands do. As in the proof of Theorem 5.2, we have ri > ln ρ(AΛ0\D,i) for all i. Therefore KMS1 states of T C ∗(Λ\D) have the form φǫ as described in [9, Theorem 6.1]. Since we can recover ǫ from φǫ as (5.4) ǫ = k (1 − e−riAΛ0\D,i)mφǫ, Yi=1 3Since Λ\D could have sources, this application depends on the observation at the start of [7, §8] that [9, Theorem 6.1] applies also to graphs with sources. KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 15 we deduce that the right-hand side of (5.4) is nonzero. Since Λ \ D has no sources, it follows from [9, Proposition 4.1(b)] that φǫ does not factor through a state of C ∗(Λ\D). So the only state of T C ∗(Λ) which could factor through a state of C ∗(Λ) is the state ψ in Proposition 5.1. Thus the corollary follows from the last assertion in Proposition 5.1. (cid:3) Corollary 5.4. In the situation of Theorem 5.2, the existence of a KMS1 state on (C ∗(Λ), αr) implies that αr is the preferred dynamics. Proof. If there is a KMS1 state on (C ∗(Λ), αr), then Corollary 5.3 implies that ri = ln ρ(AD,i) for 1 ≤ i ≤ k. Since each ri ≥ ln ρ(Ai) ≥ ln ρ(AD,i), we must have ri = ln ρ(Ai) for all i. Thus αr is the preferred dynamics. (cid:3) Corollary 5.4 is substantially stronger than [9, Corollary 4.4]: there the graph is coordinatewise irreducible, and hence has only one critical component, namely C = Λ0. 6. Graphs with several hereditary critical components We now consider graphs in which all the critical components are hereditary, but there are more than one of them. Since there can be no paths between distinct hereditary components, none of them dominates in the sense of §5. The following theorem describes the KMS1 states in this situation. Recall that from §5 and beyond we assume that ρ(AC,i) > 1 for 1 ≤ i ≤ k and C ∈ C. Theorem 6.1. Suppose that Λ is a finite k-graph without sinks or sources, and that all the critical components D are hereditary with ΛD coordinatewise irreducible. We consider a dynamics αr given by r ∈ (0, ∞)k that satisfies our standing assumption (3.1) and has rationally independent coordinates. We write Ccrit for the set of critical components, and for each D ∈ Ccrit we denote by ψD the KMS1 state of (T C ∗(Λ), αr) given by applying Proposition 5.1 to the component D. We also set Then every KMS1 state of (T C ∗(Λ), αr) is a convex combination of the states G :=[{D : D ∈ Ccrit}. {ψD : D ∈ Ccrit} and a state φ ◦ qG lifted from a KMS1 state φ of (T C ∗(Λ\G), αr). In the proof, we adapt arguments from the proofs of Theorem 4.3(b) in [11] and Theorem 6.5 in [7]. Proof. Suppose that θ is a KMS1 state of (T C ∗(Λ), αr). By Proposition 4.1 of [9], we have the subinvariance relation Aimθ ≤ erimθ for i ∈ {1, . . . , k}. Let D ∈ Ccrit. By assumption D is hereditary, and thus AD,imθD ≤ erimθD for i ∈ {1, . . . , k}. Since D is critical, there exists jD ∈ {1, . . . , k} such that rjD = ln ρ(AD,jD) = ln ρ(AjD ), 16 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN and then AD,jDmθD ≤ ρ(AD,jD)mθD. Since ΛD is coordinatewise irreducible, AD,jD is irreducible. Hence the subinvariance theorem [24, Theorem 1.6] implies that AD,jDmθD = ρ(AD,jD)mθD, that is, either mθD is a Perron -- Frobenius eigenvector for AD,jD or it is the zero vector. Suppose that C ∈ C\{D} such that CΛD 6= ∅. Then C is not hereditary, and hence is not critical. Thus ρ(AC,jD) < erjD = ρ(AD,jD), which implies that the hypotheses of Propositions 4.1 and 5.1 are satisfied. Set HD := {v ∈ Λ0 : vΛD = ∅} and FD := Λ0\(D ⊔ HD) = {v ∈ Λ0 : vΛD 6= ∅, v 6∈ D}. Let xD be the unimodular Perron -- Frobenius eigenvector of AD,jD and set yD :=(cid:0)ρ(AD,jD)1FD − AFD,jD(cid:1)−1 AFD,D,jDxD. By Proposition 4.1, relative to the decomposition Λ0 = FD ⊔D⊔HD, the vector zD = (yD, xD, 0) is an eigenvector of AjD with eigenvalue ρ(AD,jD). By Proposition 5.1, there exists a KMS1 state ψD characterised by (5.2). We define aD ≥ 0 by Our goal now is to show that mθD = aDkzDk−1 1 xD. To show this, we will prove that (6.1) θ(qv) ≥ XD∈Ccrit aD ≤ 1. XD∈Ccrit aDψD(qv) for each v ∈ Λ0; this suffices since summing over v ∈ Λ0 gives 1 = θ(1) = Xv∈Λ0 θ(qv) ≥ Xv∈Λ0 XD∈Ccrit aDψD(qv) = XD∈Ccrit aD. First, suppose that v ∈ G. Then v ∈ D for some D ∈ Ccrit and aDψD(qv) = aDkzDk−1 1 zD v = aDkzDk−1 1 xD v = (mθD)v = θ(qv). All the critical components are hereditary, so for D′ ∈ Ccrit\{D} we have v ∈ HD′ = {w ∈ Λ0 : wΛD′ = ∅}. Thus aD′ψD′(qv) = aD′kzD′ 1 zD′ k−1 v = 0, and then (6.2) θ(qv) = XD∈Ccrit aDψD(qv) for v ∈ G which verifies (6.1) for v ∈ G. Second, suppose that v 6∈ G. If vΛD = ∅ for all D ∈ Ccrit, then v ∈ HD for all D ∈ Ccrit, and so XD∈Ccrit aDψD(qv) = XD∈Ccrit aDkzDk−1 1 zD v = 0 ≤ θ(qv). So to verify (6.1), it remains to consider v 6∈ G such that vΛD 6= ∅ for at least one D ∈ Ccrit, that is, v ∈ FD for some D. To do this we mimic some calculations from KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 17 [11, page 2545] by looking at the paths which make a "quick exit" from a component D in the colour jD. The argument is long and complicated. We write QEjD(D) for {λ ∈ ΛNejD D : s(λ(0, d(λ) − ejD)) ∈ FD}, which is the set of paths of colour jD which have source in D and such that ranges of all the edges in the path are outside D. We claim that (6.3) {tλt∗ λ : λ ∈ vQEjD (D) for some D ∈ Ccrit} λtµt∗ µ = P(α,β)∈Λmin(λ,µ) tλαt∗ consists of mutually orthogonal projections. To see this, fix C, D ∈ Ccrit and choose distinct λ ∈ vQEjC (C) and µ ∈ vQEjD(D). By the Toeplitz -- Cuntz -- Krieger rela- tion (T3), we have tλt∗ µβ, and so it suffices to show that Λmin(λ, µ) = ∅. Since C and D are hereditary, if there exists (α, β) ∈ Λmin(λ, µ), then the common source of α and β is in both C and D, which means C = D. Thus C 6= D implies that Λmin(λ, µ) = ∅. So suppose that C = D. Since λ, µ ∈ ΛNejD , Λmin(λ, µ) will be empty unless λ ∈ µΛ or µ ∈ λΛ. Looking for a contradic- tion, we assume, without loss of generality, that λ ∈ µΛ, say λ = µη. Since D is hereditary and s(µ) ∈ D, we see that η ∈ ΛD. Since λ 6= µ, we must have d(η) ≥ ejD . Thus s(λ(0, d(λ) − ejD)) = s(η(0, d(η) − ejD)) ∈ D, which is impos- sible since s(λ(0, d(λ) − ejD )) ∈ FD. Thus (6.3) consists of mutually orthogonal projections as claimed, and qv ≥ tλt∗ λ for λ ∈ vΛ gives θ(qv) ≥ XD∈Ccrit Xλ∈vQEjD (D) θ(tλt∗ λ). Using the KMS condition (see [9, Theorem 3.1(a)]) we have (6.4) θ(qv) ≥ XD∈Ccrit Xλ∈vQEjD (D) ρ(AD,jD )−λθ(qs(λ)) = XD∈Ccrit aDkzDk−1 1 (cid:16) Xλ∈vQEjD (D) ρ(AD,jD)−λxD s(λ)(cid:17). We now examine the inner sum. Since we are only looking at paths that make a quick exit from D, we see that (6.5) Xλ∈vQEjD (D) ρ(AD,jD)−λxD s(λ) ∞ = Xw∈D Xn=0 ρ(AD,jD)−(n+1)(cid:0)An FD,jDAFD,D,jD(cid:1)(v, w)xD w . Since all the critical components are hereditary, none of them lie in FD ⊂ Λ0\G. Thus ρ(AD,jD) = erjD > ρ(AFD,jD), and we see that ρ(AD,jD)−1(cid:0)(cid:0)1FD − ρ(AD,jD)−1AFD ,jD(cid:1)−1AFD ,D,jD(cid:1)(v, w)xD w (6.5) = Xw∈D = Xw∈D(cid:0)(ρ(AD,jD)1FD − AFD,jD)−1AFD ,D,jD(cid:1)(v, w)xD =(cid:0)(ρ(AD,jD)1FD − AFD,jD)−1AFD ,D,jDxD(cid:1)v. = yD v . w 18 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN Since (6.5) was the inner sum of the right-hand side of (6.4) and since v ∈ FD, we get θ(qv) ≥ XD∈Ccrit aDkzDk−1 1 yD v = XD∈Ccrit aDψD(qv). We have now established (6.1) for all v ∈ Λ0. As mentioned above, it follows that aD ≤ 1. XD∈Ccrit Suppose that PD∈Ccrit aD = 1. Then Xv∈Λ0 θ(qv) − XD∈Ccrit aDψD(qv)! = Xv∈Λ0 θ(qv) − XD∈Ccrit aD = 0. = 1 − XD∈Ccrit aD Xv∈Λ0 ψD(qv)! It then follows from (6.1) that θ(qv) = PD∈Ccrit aDψD(qv) for v ∈ Λ0. Since both θ and PD∈Ccrit aDψD are KMS1 states and the coordinates of r are rationally inde- pendent, it follows from [9, Proposition 3.1(b)] that they agree on all the spanning elements tµt∗ ν of T C ∗(Λ). Hence by linearity and continuity, we have aDψD θ = XD∈Ccrit and θ is a convex combination, as required. The other possibility is that PD∈Ccrit aD < 1. To handle this case, we adapt the argument of the last four paragraphs in the proof of [7, Theorem 6.5]. For convenience we write E := Λ0\G. Let By (6.2), θ and PD∈Ccrit aDψD agree on the vertex projections {qv : v ∈ G}, and a short calculation using this shows that kκk1 = 1 −PD∈Ccrit aD. We claim that the vector κ :=(cid:16)mθ − XD∈Ccrit . aDmψD(cid:17)(cid:12)(cid:12)(cid:12)E k η := Yi=1(cid:0)1E − e−riAE,i(cid:1)κ belongs to [0, ∞)E. Once we have stablished that η ≥ 0, we will argue that the vector ǫ := kκk−1 1 η gives a KMS1 state φǫ of the quotient T C ∗(Λ \ G) and that θ is a convex combination of φǫ ◦ qG and {ψD : D ∈ Ccrit}. Since θ is a KMS1 state of (T C ∗(Λ), αr), Proposition 4.1(a) of [9] implies that the satisfies the subinvariance relation vector mθ =(cid:0) θ(qv)(cid:1)v∈Λ0 in [0, ∞)Λ0 k (6.6) (1Λ0 − e−riAi)mθ ≥ 0; Yi=1 we will show that the restriction of this vector to E is η. KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 19 Consider the block decomposition of (6.6) relative to Λ0 = E ⊔ G. An induction argument on the number of critical components shows that the top entry of (6.6) is (6.7) k k Yi=1 (1E − e−riAE,i)mθE − XD∈Ccrit Yi=1 i6=jD (1E − e−riAE,i)e−rjD AE,D,jDmθD. Now we restrict attention to the Dth summand of the second term: (1E − e−riAE,i)e−rjD AE,D,jDmθD k Yi=1 i6=jD = aDkzDk−1 1 k Yi=1 i6=jD (1E − e−riAE,i)ρ(AD,jD)−1AE,D,jDxD. Since there can be no paths from D to HD, nor from FD to HD, relative to the decomposition E = FD ⊔ (HD\G), this is given by k i6=jD aDkzDk−1 1 (6.8) 0 (cid:18)1FD − e−riAFD ,i Yi=1 = aDkzDk−1 1 Qk i=1 i6=jD −e−riAFD,HD\G,i 1HD\G − e−riAHD\G,i(cid:19) ρ(AD,jD)−1(cid:18)AFD,D,jD 0 (cid:19) xD (1FD − e−riAFD,i)ρ(AD,jD)−1AFD,D,jDxD 0 ! . Since every critical component is hereditary, FD does not contain any critical com- ponents. Thus it follows from (3.1) that ρ(AD,jD) = erjD > ρ(AFD ,jD), and so the matrix 1FD − ρ(AD,jD)−1AFD,jD is invertible. We can use this inverse to rewrite the top block of (6.8) as k aDkzDk−1 1 (1FD − e−riAFD ,i)(cid:0)1FD − ρ(AD,jD)−1AFD,jD)−1ρ(AD,jD(cid:1)−1AFD ,D,jDxD Yi=1 k (1FD − e−riAFD ,i)(ρ(AD,jD)1FD − AFD,jD)−1AFD ,D,jDxD = aDkzDk−1 1 = aDkzDk−1 1 k Yi=1 Yi=1 = k Yi=1 (1FD − e−riAFD ,i)yD (1FD − e−riAFD,i)(aDmψD FD) 20 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN because mψD mψD HD\G = 0 and E = FD ⊔ (HD\G), we see that (6.8) is given by v = ψD(qv) = kzk−1 for v ∈ Λ0 and zD = (xD, yD, 0). Since 1 zD v i=1(1FD − e−riAFD,i)(aDmψD FD) 0 (cid:18)Qk = = k 0 Yi=1(cid:18)1FD − e−riAFD,i Yi=1 k (1E − e−riAE,i)aDmψD E. (cid:19) 1HD\G − e−riAHD\G,i(cid:19)(cid:18)aDmψD FD −e−riAFD,HD\G,i 0 (cid:19) Summing over D ∈ Ccrit gives us back the second term in (6.7): k k XD∈Ccrit Yi=1 i6=jD (1E − e−riAE,i)e−rjD AE,D,jDmθD = XD∈Ccrit (1E − e−riAE,i)aDmψD E. Yi=1 Now, starting with the definition of η, we trace our way back to (6.7): η = = = k k Yi=1 Yi=1 Yi=1 k k (1E − e−riAE,i)mθE − aDmψD(cid:17)(cid:12)(cid:12)(cid:12)E (1E − e−riAE,i)(cid:16)mθ − XD∈Ccrit Yi=1 (1E − e−riAE,i) XD∈Ccrit (1E − e−riAE,i)mθE − XD∈Ccrit Yi=1 k aDmψD E (1E − e−riAE,i)aDmψD E. Thus η is the top entry of (6.6), and hence η ≥ 0. We now set ǫ := kκk−1 1 η =(cid:16)1 − XD∈Ccrit η. aD(cid:17)−1 Since E does not contain any critical components, by (3.1) we have ri ≥ ln ρ(Ai) > ln ρ(AE,i) for i ∈ {1, . . . , k}, and the inverse temperature β = 1 is in the range for which [9, Theorem 6.1] applies4. Also eri > ρ(AE,i) for i ∈ {1, . . . , k}, and so all the matrices 1E − e−riAE,i are invertible and we can recover κ from η. Thus k Yi=1 (cid:13)(cid:13)(cid:13) (1E − e−riAE,i)−1ǫ(cid:13)(cid:13)(cid:13)1 =(cid:16)1 − XD∈Ccrit =(cid:16)1 − XD∈Ccrit = 1, aD(cid:17)−1(cid:13)(cid:13)(cid:13) aD(cid:17)−1 k (1E − e−riAE,i)−1η(cid:13)(cid:13)(cid:13)1 Yi=1 kκk1 and it follows from [9, Theorem 6.1(a)] that ǫ belongs to the simplex Σ1 of that theorem for the graph ΛE = Λ\G. We deduce that there is a KMS1 state φǫ of 4The graph ΛE could have sources (see [7, Example 8.4]). So this application of the result from [9, Theorem 6.1] depends on the observation at the start of [7, §8] that [9, Theorem 6.1] applies also to graphs with sources. KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 21 (T C ∗(Λ\G), αr) such that for all v ∈ E. Now φǫ(qv) =(cid:16)1 − XD∈Ccrit aD(cid:17)−1 aD(cid:17)(φǫ ◦ qG) + XD∈Ccrit (cid:16)1 − XD∈Ccrit κv aDψD is a KMS1 state of (T C ∗(Λ), αr) which agrees with θ on the vertex projections qv. Since they are both KMS1 states and the coordinates of r are rationally independent, it again follows from [9, Proposition 3.1(b)] that θ =(cid:16)1 − XD∈Ccrit aD(cid:17)(φǫ ◦ qG) + XD∈Ccrit and θ is a convex combination, as required. aDψD, (cid:3) Corollary 6.2. In the situation of Theorem 6.1, the KMS1 simplex of (T C ∗(Λ), αr) has dimension Λ0\G + Ccrit − 1. Proof. Since all the critical components are hereditary and belong to Ccrit, the dy- namics on T C ∗(Λ\G) induced by αr has ri > ln ρ(AΛ0\G,i) for all i, and hence by [9, Theorem 6.1] has a KMS1 simplex of dimension Λ0\G − 1. Thus the result follows from Theorem 6.1. (cid:3) Corollary 6.3. Suppose that we have a k-graph Λ with the properties in Theo- rem 6.1, and suppose in addition that for D ∈ Ccrit the graph Λ\D does not have sources. Then a KMS1 state of T C ∗(Λ) factors through C ∗(Λ) if and only if it is a convex combination of the states {ψD : D ∈ Ccrit, ri = ln ρ(AD,i) for 1 ≤ i ≤ k}. Proof. By Corollary 5.3, each ψD factors through C ∗(Λ) if and only if ri = ln ρ(AD,i) for 1 ≤ i ≤ k. No state φǫ of T C ∗(Λ\G) factors through C ∗(Λ\G) (this follows as in the proof of Corollary 5.3). Since a convex combination of states of T C ∗(Λ) factors through C ∗(Λ) if and only all the summands do, the result follows. (cid:3) 7. Computing the KMS1 states We now describe how to combine our results to find the extreme KMS states at the critical inverse temperature. We begin by making some general assumptions about the graphs we consider. Throughout, we consider a finite k-graph Λ with no sources and no sinks. First we assume that the graph is suitably connected: (A1) We assume that there are no trivial strongly connected components, which forces Λ0 =S{C : C ∈ C}. We assume that there are no isolated subgraphs: there is no decomposition Λ = ΛK ⊔ ΛH with H ∩ K = ∅. We assume that the full results of [9] apply to the reductions ΛC: (A2) For all C ∈ C, the graph ΛC is coordinatewise irreducible, and ρ(AC,i) > 1 for all i. Our next assumption rules out the case in which the only bridge between components consists of edges of a single colour, as in the example in [7, Remark 6.2]. (A3) If C, D ∈ C and CΛej D 6= ∅ for some j, then CΛeiD 6= ∅ for all 1 ≤ i ≤ k. 22 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN Since we know by (A2) that the graphs {ΛC : C ∈ C} are coordinatewise irreducible, (A1) and (A3) imply that Λ satisfies the hypothesis (3.3) in Proposition 3.4. The three assumptions (A1 -- A3) imply that for every hereditary subset H of Λ0, the graph Λ\H obtained by removing H has no sources or sinks. The examples in [7, §8] show that otherwise sources of various types could be created. We now consider a dynamics αr : R → Aut T C ∗(Λ) determined by a vector r ∈ (0, ∞)k satisfying the standing hypothesis (3.1) and having rationally independent coordinates {ri : 1 ≤ i ≤ k}. Then the following procedure will generate the extreme points of the simplex of KMS1 states of the system (T C ∗(Λ), αr). We are thinking of the components as small, even singletons, in which case Λ is one of our favourite dumbbell graphs. (C1) First we calculate the spectral radii of the matrices AC,i, and identify the critical components of Λ. If any of them are not hereditary, we identify the set H of Proposition 3.4. (Recall that H is the the complement of the union of Cmincrit in its hereditary closure.) Then we study the system (T C ∗(Λ\H), αr). (C2) All of the critical components of Λ\H are now hereditary. For each such com- ponent D, we compute the common unimodular Perron -- Frobenius eigenvec- tor x of the matrices AD,i. (Theoretically, it suffices to find the eigenvector of one of them, but there is an opportunity for a reality check here.) Propo- sition 4.1 tells us how to extend x to an eigenvector z of AΛ0\H,i, and then Proposition 5.1 gives us an explicit KMS1 state ψD of (T C ∗(Λ\H), αr). (C3) Now we take G to be the union of the critical components of Λ\H. Then Theorem 6.1 tells us that the states {ψD : D ∈ Ccrit(Λ\H)} = {ψD : D ∈ Cmincrit(Λ)} are extreme points of the KMS1 simplex of (T C ∗(Λ\H), αr), and that the other KMS1 states factor through the quotient map qG : T C ∗(Λ\H) → T C ∗(Λ\(H ∪ G)). Since we have removed all the critical components from Λ, we have ri > ln ρ(AΛ0\(H∪G),i) for 1 ≤ i ≤ k, and Theorem 6.1 of [9] describes an explicit parametrisation ǫ 7→ φǫ of the KMS1 simplex of (T C ∗(Λ\(H ∪G)), αr). The extreme points are those where ǫ is a multiple of a point mass δv; in the notation of [9, Theorem 6.1], the multiple is y−1 v δv. (This vector y is not the one of Proposition 4.1.) Thus the extreme points of the simplex of KMS1 states of (T C ∗(Λ), αr) are (cid:8)ψD : D is critical for Λ\H(cid:9) ∪(cid:8)φy−1 v δv ◦ qG : v ∈ Λ0\(H ∪ G)(cid:9). 8. Finding all the KMS states We now describe our program for finding all the KMS states of (T C ∗(Λ), αr). We suppose that Λ is a finite k-graph with no sources and no sinks, satisfying the assumptions (A1 -- A3) of the previous section. We suppose that r ∈ (0, ∞)k has rationally independent coordinates, and has been normalised (multiplied by a suitable scalar) to ensure that the standing hypothesis (3.1) is satisfied. Our program has three steps, which can be iterated to reduce our problem to the same problem for graphs with fewer strongly connected components. When we get down to a strongly connected graph (one with a single component), [6, Proposition 4.2] implies that there is a unique KMS1 state. The program consists of the following 3 items. KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 23 (P1) For β > 1, we apply [9, Theorem 6.1]. This yields a KMSβ simplex with Λ0 extreme points (cid:8)φy−1 v δv : v ∈ Λ0(cid:9). (P2) At β = 1, we apply the procedure of §7. Let H be the hereditary set described in Proposition 3.4 and let G be the union of the critical components of Λ\H. Then we follow the steps (C1 -- C3), arriving at a KMS1 simplex with extreme points parametrised by the critical components in Λ\H and the vertices v in Λ0\(G ∪ H). (P3) For β < 1, we start with a lemma. Lemma 8.1. Every KMSβ state of (T C ∗(Λ), αr) factors through a KMSβ state of (T C ∗(Λ\(H ∪ G)), αr). Proof. Let φ be a KMSβ state of (T C ∗(Λ), αr). Fix v ∈ H ∪ G. We will show that φ(qv) = 0. Recall that the critical components of Λ \ H are also critical components of Λ (in fact the minimal ones). So there exists a critical component C of Λ such that CΛv 6= ∅. We restrict φ to get a KMSβ state of (T C ∗(ΛC), αr). By [9, Proposition 4.1(a)] AC,imφC ≤ eβrimφC for 1 ≤ i ≤ k. Since each AC,i is irreducible, the subinvariance theorem [24, Theorem 1.6] gives that mφC 6= 0 =⇒ ρ(AC,i) ≤ eβri =⇒ r−1 i ln ρ(AC,i) ≤ β < 1 for all i. But C is j-critical, say, and this gives r−1 impossible. Thus mφC = 0. j ln ρ(AC,j) = 1, which is Now let w ∈ C and choose n ∈ Nk such that wΛnv 6= ∅. Then 0 = φ(qw) ≥ Xλ∈wΛn φ(tλt∗ λ) = Xλ∈wΛn e−βr·nφ(ts(λ)) using that φ is a KMS state. Thus φ(ts(λ)) = 0 for all λ ∈ wΛn. In particular, φ(qv) = 0. It follows from [1, Lemma 6.2] that φ vanishes on the ideal IH∪G, and hence φ factors through a KMSβ state of (T C ∗(Λ\(H ∪ G)), αr). (cid:3) Next we need to check that the graph Σ := Λ\(G∪H) satisfies the assump- tions (A1 -- A3). Since H is hereditary, every component C with C ∩ H 6= ∅ lies entirely inside H. So H is a union of strongly connected components, all of which are nontrivial by (A1). The set G is the union of the critical components in Λ\H. So G ∪ H is a union of strongly connected components of Λ0. Since Λ has no trivial components and all components are contained in one of Λ0\(G ∪ H) or G ∪ H, Λ0\(G ∪ H) is also the union of the components it contains. So Σ has no nontrivial components. But Σ may not satisfy the second part of (A1): there may be disjoint subsets Kj of Σ0 which are unions j=1 Kj, and which do not speak to each other in the sense that KjΣKl = ∅ for j 6= l.5 of components, which satisfy Σ0 =Sn We set Pj =Pv∈Kj qv, and observe that {qv : v ∈ Kj} ∪ {tλ : λ ∈ KjΣKj} 5For example, suppose that Λ has three components arranged as in §9.2, that the hereditary component D belongs to Cmincrit, and and that AC,B,i = 0 for all i. Then Λ\H is the disjoint union of ΛC and ΛB. 24 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN is a Toeplitz -- Cuntz -- Krieger ΣKj -family in PjT C ∗(Σ)Pj which gives an iso- morphism of T C ∗(ΣKj ) onto PjT C ∗(Σ)Pj. Since PjPk = 0 for j 6= k, we have n T C ∗(Σ) = PjT C ∗(Σ)Pj. Mj=1 Thus the KMSβ states of (T C ∗(Σ), αr) are convex combinations of KMSβ states of (T C ∗(ΣKj ), αr). Since each ΣKj is smaller than Λ, we have reduced the problem of computing KMSβ states of (T C ∗(Λ), αr) to the analogous problem for smaller graphs, each of which satisfies the hypotheses (A1 -- A3). We now study one of these smaller graphs, ΣK, say. Since we removed all the critical components in Λ\H when we removed G, βc := max(cid:8)r−1 i ln ρ(AΣK ,i) : 1 ≤ i ≤ k(cid:9) is strictly less than 1. This is another critical inverse temperature for the orig- inal system (T C ∗(Λ), αr). To make the results of §3 -- §7 available verbatim, we consider the dynamics αβcr. This new dynamics satisfies the standing hy- pothesis (3.1) for ΛΣK , and has the same KMS states as (T C ∗(ΣK), αr): the c β states of (T C ∗(ΣK), αβcr) KMSβ states of (T C ∗(ΣK), αr) are the KMSβ−1 [6, Lemma 2.1]. Now the system (T C ∗(ΣK), αβcr) is one to which we can apply our program (P1 -- P3). Since the requirement (3.1) implies that Λ has at least one critical component, the set G contains at least one component. Thus the graphs ΣKj have strictly fewer strongly connected components than Λ, and the iterative process we have described must terminate after finitely many steps. 9. Applications and examples We now discuss implementation of our program. In the various subsections, we focus on graphs with relatively few components (§9.1 and §9.2), graphs with just one colour, where we carry out a reality check by comparing with the results for ordinary graph algebras in [10], and graphs in which the components are singletons, where we can do specific calculations like those for 1-graphs in [12]. We are reassured that our program does not apparently run into new difficulties. 9.1. Graphs with two components. We first apply our program to a graph Λ with exactly two nontrivial strongly connected components. Our assumptions (A1 -- A3) imply that one component C is forwards hereditary, the other component D is hereditary, and the vertex matrices of Λ have the form Ai =(cid:18)AC,i AC,D,i AD,i (cid:19) with AC,D,i 6= 0 for all i. 0 For β > 1, our first step (P1) gives a simplex of KMSβ states with Λ0 = C+D extreme points. At β = 1, (P2) tells us to apply the procedure (C1 -- C3) of §7. First, suppose that C is critical. Then Ccrit is {C} or {C, D}. Either way, only C is minimal. Thus H = D, Λ\H = ΛC and Proposition 3.4 implies that every KMS1 state of (T C ∗(Λ), αr) factors through a KMS1 state of (T C ∗(ΛC), αr). Since ΛC is coor- dinatewise irreducible and {ri : 1 ≤ i ≤ k} are rationally independent, there is a unique KMS1 state on T C ∗(ΛC) by [6, Theorem 4.2]. Thus there is a unique KMS1 state on T C ∗(Λ) as well. KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 25 Second, suppose that C is not critical. Then Ccrit = {D}, H = ∅, and Λ\(G∪H) = ΛC. Thus every KMS1 state of T C ∗(Λ) is a convex combination of ψD and a KMS1 state of (T C ∗(ΛC), αr). Since C is not critical, ri > ln ρ(AC,i) for 1 ≤ i ≤ k, and [9, Theorem 6.1] gives Λ0 C = C extreme KMS1 states of (T C ∗(ΛC), αr). Thus the KMS1-simplex of T C ∗(Λ) has C + 1 extreme points. For 0 < β < 1, we follow (P3). If C is critical, then H = D, G = C and Σ = Λ\(G ∪ H) is empty, and there are no KMSβ states. So we suppose that C is not critical. Then βc := max{r−1 i ln ρ(AC,i) : 1 ≤ i ≤ k} is strictly less than 1, and (P3) tells us to apply the program to ΛC. For βc < β < 1, the KMSβ states are lifted from KMSβ states of (T C ∗(ΛC), αr), and (P1) gives us a KMSβ simplex with C extreme points. Then (P2) gives us a single KMSβc state of (T C ∗(ΛC), αr), and hence also of the original system. Now (P3) tells us to look at C\C = ∅, and the original system has no KMSβ states for β < βc. 9.2. Graphs with three components. We now consider a finite k-graph Λ which satisfies the assumptions (A1 -- A3) of §7, and which has three strongly connected components. Recall from [7, Proposition 3.1] that we can order the components so that the vertex matrices Ai are simultaneously block upper-triangular. The as- sumption of "no trivial components" says that these decompositions have no strictly upper-triangular diagonal blocks. The component C such that the {AC,i} are the top blocks is forwards hereditary, and the component D such that the {AD,i} are the bottom blocks is hereditary. We call the remaining component B. Then each Ai has the form (9.1) AC,i AC,B,i AC,D,i AB,i AB,D,i AD,i 0 0 0 Ai =    . We now consider a dynamics αr : T → Aut T C ∗(Λ) such that r ∈ (0, ∞)k satisfies (3.1) and has rationally independent coordinates. We want to find the KMS1 states of the system (T C ∗(Λ), αr), and we run through the program. The first step (P1) gives a KMSβ simplex with Λ0 extreme points for every β > 1. Next (P2) tells us to look at the KMS1 states using the procedure (C1 -- C3) of §7. Suppose first that C is critical. Since (A1) says that C and D are not isolated, there must be paths from D to C (possibly going through B). Thus the set H in (C1) is either D or B ∪ D. Either way, (C1) reduces the problem of computing the KMS1 states on T C ∗(Λ) to the analogous problem for a graph with one or two components, and the analysis of §9.1 tells us how to do this. We suppose next that C is not critical and B is critical. If there exists j such that the block AB,D,j in (9.1) is nonzero, then the set H in (C1) is D, and again the problem reduces to the same one for the graph ΛC∪B with two components. This is either a disjoint union of two irreducible graphs, in which case T C ∗(ΛC∪B) = T C ∗(ΛC) ⊕ T C ∗(ΛB) and we can study the summands separately, or the analysis of §9.1 applies. So we suppose that AB,D,i = 0 for all i. Then both B and D are hereditary. If D is also critical, then (C2) gives two KMS1 states ψB and ψD, and a (C − 1)-dimensional simplex of KMS1 states lifted from states of (T C ∗(ΛC), αr); thus the KMS1 simplex of (T C ∗(Λ), αr) has dimension C + 1. If D is not critical, the simplex has dimension C + D. Finally, we suppose that neither C nor B is critical. Then D has to be critical, and the KMS1 simplex of (T C ∗(Λ), αr) has dimension C + B. 26 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN Below β = 1, step (P3) tells us to run the program for a smaller graph Σ := Λ\(G ∪ H). Since Σ has one or two components, we have seen in the previous subsection that the program will give us all the KMS states. 9.3. Comparison with previous results for 1-graphs. When k = 1, a k-graph Λ is the path category of a directed graph E, and its Toeplitz algebra is the algebra T C ∗(E) whose KMS states were analysed in [8] and [11]. We write A for the vertex matrix of E, and consider the dynamics α : t 7→ γeit on T C ∗(E) studied in [8, 11]. Then the dynamics in this paper is given by α′ : t 7→ αt ln ρ(A); thus the KMSβ states of (T C ∗(E), α) are the KMSβ(ln ρ(A))−1 states of (T C ∗(E), α′) (see, for example, [6, Lemma 2.1]). In particular, the KMSln ρ(A) simplex of (T C ∗(E), α) should be the KMS1 simplex of our (T C ∗(E), α′). To find the KMSln ρ(A) states of (T C ∗(E), α), we apply the procedure of [11, Theorem 4.3], which focuses on the set mc(E) of "minimal critical components" in E0/∼. Under our hypotheses, E0/∼ is the set C of nontrivial strongly connected components of Λ0 = E0; a component C is critical if ρ(AC) = ρ(A), and minimal if D ≤ C and D critical imply D = C. Part (a) of [11, Theorem 4.3] describes KMSln ρ(A) states {ψC : C ∈ mc(E)}, and part (b) says that every KMSln ρ(A) state is a convex combination of the ψC and a state lifted from the quotient associated to the hereditary closure of C ′ := SC∈mc(E) C. When we carry out our procedure from §7, we take two quotients: first in step (C1) by an ideal IH , and then in step (C3) by an ideal IG. The hereditary closure of C ′ is precisely G ∪ H, and hence [11, Theorem 4.3] merely does both quotients in one hit. Thus the KMSln ρ(A) states of (T C ∗(E), α) and the KMS1 states of our (T C ∗(E), α′) are the same, as expected. (Which is reassuring, because our constructions in §4 -- §6 were based on those of [11, §4].) The other main result in [11] describes the KMSβ states at a fixed inverse tem- perature β, as follows. First, consider the hereditary closure Hβ of the compo- nents C with ln ρ(AC) > β. If Hβ = E0, there are no KMSβ states. Otherwise, Theorem 5.3 of [11] says that all KMSβ states of (T C ∗(E), α) factor through a state of (T C ∗(E\Hβ), α), and are then given by [8, Theorem 3.1] provided that β > ln ρ(AE 0\Hβ ), that is, β is not critical. None of these states factor through C ∗(E). (The subtleties in [11, Theorem 5.3] involving the saturation ΣHβ do not arise here because we are not allowing trivial components.) If β is critical, so that β = ln ρ(AC) for some component C, then [11, Theo- rem 5.3(c)] tells us to look also at the hereditary closure Kβ of {C : ln ρ(AC) ≥ β}, which strictly contains Hβ. Then applying [11, Theorem 4.3] to E\Hβ gives the ex- treme KMSβ states {ψC : C ∈ mc(E\Hβ)} and other states lifted from T C ∗(E\Kβ); this is our step (C3). In our situation (no trivial components), only convex combi- nations of the ψC factor through states of C ∗(E). 9.4. Concrete examples. Our examples are dumbbell graphs with three compo- nents and minimal activity between the components. The most interesting case seems to be graphs with skeleton shown in Figure 2, where as usual the blue loop at w labelled n1 means there are n1 blue loops at w. Thus the vertex matrices of the 2-coloured graph in Figure 2 for the ordering {u, v, w} of Λ0 are l1 p1 0 m1 0 0 A1 =  q1 0 n1   and A2 =  l2 p2 0 m2 0 0 q2 0 n2   . KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 27 l2 l1 u p1 p2 q1 q2 m2 v m1 n2 n1 w Figure 2. A dumbbell graph with 2 hereditary components. Kumjian and Pask proved that the graph in Figure 2 is the skeleton of a 2-graph if and only if the matrices A1 and A2 commute [13, §6] (and is then the skeleton of many such graphs). For the matrices A1 and A2, this is equivalent to (9.2) l1p2 + p1m2 = l2p1 + p2m1, and l1q2 + q1n2 = l2q1 + q2n1. Example 9.1. We consider a 2-graph Λ with skeleton shown in Figure 2, so that the vertex matrices are 2 2 3 0 4 0 A1 = 0 0 5   and A2 = 0 0 4  .  2 1 2 0 3 0 We consider the preferred dynamics given by r = (ln 5, ln 4). Proposition A.1] that ln 4 and ln 5 are rationally independent. It follows from [7, In this graph, there is one critical component D = {w}. The other blocks in the decomposition of Λ0 in Proposition 4.1 are F = {u} and H = {v}, and for the ordering {u, w, v} of Λ0, the vertex matrices become 2 3 2 0 5 0 A1 = 0 0 4  0 0 3  and A2 =  .  2 2 1 0 4 0 In the notation of Proposition 4.1 we have E1 = (2) = E2, B1 = (3) and B2 = (2). The common unimodular Perron -- Frobenius eigenvector x of AD,1 = (5) and AD,2 = (4) is the scalar 1, and the vector y in Proposition 4.1 is the scalar (ρ(AD,1)1F − E1)−1B1x = (5 − 2)−13 = 1. (As a reality check, we confirm that (ρ(AD,2)1F − E2)−1B2x = (4 − 2)−12 is also 1.) Thus the vector z in Proposition 5.1 is (1, 1, 0), and that proposition gives a KMS1 state ψD of (T C ∗(Λ), αr) such that ψD(qu) = 2−1 = ψD(qw) and ψD(qv) = 0. Since ρ(AD,1) = 5 = ρ(A1) and ρ(AD,2) = 4 = ρ(A2), and the spectral radii of the other diagonal blocks are all smaller, Theorem 6.1 says that every KMS1 state of (T C ∗(Λ), αr) is a convex combination aψD + (1 − a)(θ ◦ qD) for some KMS1 state θ of (T C ∗(Λ\D), αr) = (T C ∗(ΛF ∪H), αr). 28 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN So we want to analyse the KMS1 states on T C ∗(ΛF ∪H) for the dynamics induced by the preferred dynamics αr on T C ∗(Λ). The graph ΛF ∪H has vertex matrices AF ∪H,1 =(cid:18)2 2 0 4(cid:19) and AF ∪H,2 =(cid:18)2 1 0 3(cid:19) . Since the dynamics satisfies ri = ln ρ(Ai) > ln ρ(AF ∪H,i) for i = 1, 2, the in- verse temperature β = 1 lies in the range for which Theorem 6.1 of [9] applies to (T C ∗(ΛF ∪H), αr). Thus every KMS1 state of (T C ∗(ΛF ∪H), αr) has the form φǫ for ǫ in a simplex lying in [0, ∞)F ∪H, which is described in [9, Theorem 6.1(c)]. As we did for other dumbbell graphs in [12, §4], we directly solve the subinvariance relation (9.3) ǫ := (1 − e−riAF ∪H,i)m ≥ 0 2 Yi=1 for a vector m ∈ [0, ∞)F ∪H satisfying kmk1 = 1. Thus we seek t ∈ [0, 1] such that m = (mu, mv) = (1 − t, t) satisfies (9.3). Multiplying out the product gives ǫ = 1 20(cid:16)5(cid:18)1 0 0 1(cid:19) − AF ∪H,1(cid:17)(cid:16)4(cid:18)1 0 0 1(cid:19) − AF ∪H,2(cid:17)(cid:18)1 − t t (cid:19) = 1 20(cid:18)6 − 11t t (cid:19) , which is nonnegative if and only if t ∈ [0, 6 11 gives a KMS1 state ψH of (T C ∗(ΛF ∪H), αr) such that ψH (qu) = 5 11 ; taking t = 0 gives a state ψ such that ψ(qv) = 0 and ψ(qu) = 1, which by an application of [1, Lemma 6.2] factors through a state φ of T C ∗(ΛF ). 11 and ψH (qv) = 6 11 ]. Taking t = 6 Thus the KMS1 simplex of (T C ∗(Λ), αr) has extreme points ψD, ψH ◦ qD and φ ◦ qD∪H. We now consider β < 1. The only critical component of Λ is D = {w}, so (P3) tells us to apply (P1) or (P2) to T C ∗(ΛF ∪H), and we get KMSβ states for β < 1 which factor through the quotient map qD. The next critical inverse temperature is βc = maxnln 4 ln 5 , ln 3 ln 4o = ln 4 ln 5 . For βc < β < 1, Theorem 6.1 of [9] gives a KMSβ simplex of dimension 1. Since r has rationally independent coordinates6, Theorem 5.2 says that there is a unique KMSβc state ψ{v} = ψH of (T C ∗(ΛF ∪H, αr) from the critical components, and hence again 1-dimensional simplices of KMSβc states on (T C ∗(Λ\D), αr) and (T C ∗(Λ), αr). For β < βc, all KMSβ states factor through q{v,w} = qD∪H . The dynamics on T C ∗(ΛF ) has another critical inverse temperature at β′ c = maxnln 2 ln 5 , ln 2 ln 4o = ln 2 ln 4 , and a single KMSβ state for β′ there are no KMSβ states. c ≤ β < βc. Thus so does (T C ∗(Λ), αr). For β < β′ c, In the previous example, we had ρ(A{u},i) < ρ(A{v},i) < ρ(A{w},i) for all i. In the next example there are also two hereditary components, but neither dominates the other. This is where our new Theorem 6.1 is useful. 6There is a subtlety here. Strictly speaking, we are applying Theorem 5.2 to the dynamics αβcr ln 5 r. But this vector has rationally independent coordinates if and associated to the vector βcr = ln 4 only if r does, and r1 = ln 5 and r2 = ln 4 are rationally independent by [7, Proposition A.1]. KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 29 Example 9.2. We take a 2-graph Λ with skeleton as described in Figure 2, with p = (1, 2) and q = (1, 1). Then one checks that l = (5, 3), m = (10, 13) and n = (11, 9) satisfy the relations (9.2), and hence for these choices there is a 2-graph Λ with the skeleton in Figure 2. With respect to the ordering {u, v, w} of Λ0, the vertex matrices are 5 1 0 10 0 0 A1 =  1 0 11  and A2 =  2 3 1 0 13 0 0 0 9  . The numbers have been chosen quite carefully: the matrices have to commute, and we have chosen numbers so each of l, m, n has coordinates that are coprime, so the vertex matrices of the subgraphs ΛC have {ln ρ(AC,i) : 1 ≤ i ≤ k} rationally independent for all components C (see [7, Proposition A.1]). We also have ρ(A1) = 11 and ρ(A2) = 13, so ln ρ(A1) and ln ρ(A2) are rationally independent too. Let C := {u}, B := {v} and D := {w}. Then we have ρ(AB,1) < ρ(A1) = 11 = ρ(AD,1), ρ(AD,2) < ρ(A2) = 13 = ρ(AB,2), and We consider the preferred dynamics αr on T C ∗(Λ), so that r = (ln 11, ln 13). ρ(AC,i) ≤ 5 < 9 ≤ min(cid:8)ρ(AB,i), ρ(AD,i)(cid:9) for i = 1, 2. Proposition 5.1 gives two KMS1 states ψB and ψD of (T C ∗(Λ), αr) such that mψB =  ψB(qu) ψB(qv) ψB(qw)  = 5 6  1/5 1 0   and mψD = 6 7  1/6 0 1   . For all e ∈ vΛe1 we have s(e) = v, and hence Proposition 5.1 implies that ψB(tet∗ e) = ρ(A1)−1ψB(qv) = (11)−1 · 5 6 = 5 66 . Thus ψB(cid:16) Xe∈vΛe1 tet∗ e(cid:17) = 10 · 5 66 < 5 6 = ψB(qv), and the state ψB does not factor through a state of C ∗(Λ). A similar argument (using red edges instead of blue ones) shows that ψD does not factor through C ∗(Λ) either. The dynamics on the quotient qB∪D(T C ∗(Λ)) = T C ∗(ΛC) induced by the pre- ferred dynamics αr on T C ∗(Λ) falls in the range covered by [9, Theorem 6.1]. Hence there is a unique KMS1 state φ on (T C ∗(ΛC), αr). Since both components B and D are critical, Theorem 6.1 implies that the KMS1 simplex of (T C ∗(Λ), αr) has extreme points ψB, ψD and φ ◦ qB∪D. For β < 1, Lemma 8.1 implies that all KMSβ states of T C ∗(Λ) factor through T C ∗(ΛC). The induced dynamics αr on T C ∗(ΛC) has critical inverse temperature βc = maxn ln 5 ln 11 , ln 3 ln 13o = ln 5 ln 11 . Since 11 and 13 are coprime, the vector r has rationally independent coordinates, and it follows as in the previous example from [9, Theorem 6.1] and Theorem 5.2 (or [6, Proposition 4.2]) that the system (T C ∗(Λ), αr) has a single KMSβ state for βc ≤ β < 1. Also, there are no KMSβ states of T C ∗(Λ) for β < βc. 30 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN Appendix A. Dumbbell graphs with three components As we pointed out earlier, while Theorem 4.2 is essentially a result about commut- ing integer matrices, our proof is indirect and makes heavy use of Perron -- Frobenius theory. So, as a reality check, we tried to prove it directly. For 2-graphs with two components, it was quite easy to see why it works (see Example 4.4). It was a little harder to see what was going on for three components, but the exercise was instruc- tive -- it led us, for example, to the strengthening of Theorem 4.2 in Corollary 4.3 (see Remark A.1). We suppose throughout this appendix that Λ is a dumbbell 2-graph with three strongly connected components. Thus Λ has three vertices u, v and w, and vertex matrices of the form qi mi ri 0 ni si 0 pi 0 Ai =    for i = 1, 2. We write C = {u}, B = {v} and D = {w}. The hypothesis CΛeiD 6= ∅ and BΛeiD 6= ∅ in Theorem 4.2 says that all the ri and si are nonzero, and we assume until further notice that this holds. We now aim to prove Theorem 4.2 directly. We may as well suppose that the j in the hypothesis of Theorem 4.2 is j = 1 (otherwise swap colours). So we assume that (A.1) p1 = ρ(AD,1) > max{ρ(AC,1), ρ(AB,1)} = max{n1, m1}, and aim to prove that p2 > n2 and p2 > m2. The factorisation property implies that the vertex matrices Ai commute. Looking at the super-diagonal entries in A1A2 and A2A1, and rearranging, shows that A1A2 = A2A1 if and only if (D1) q2(n1 − m1) = q1(n2 − m2), (D2) s2(p1 − n1) = s1(p2 − n2), and (D3) r2(p1 − m1) + q2s1 = r1(p2 − m2) + q1s2. The hypothesis (A.1) implies that p1 > n1, and since si 6= 0, (D2) forces p2 > n2. Thus it remains to show that p2 > m2. The first case we consider is when q1 and q2 are both nonzero. From (D1) we deduce that the numbers mi − ni have the same sign. If n1 ≥ m1, then we have n2 ≥ m2 and p2 > n2 ≥ m2. So we assume that m1 > n1 and m2 > n2, and aim to prove that p2 > m2. We break up each side of (D2) using pi = (pi − mi) + mi, multiply the resulting equation by q2, and apply (D1) to get (A.2) q2s1(p2 − m2) + q2s1(m2 − n2) = q2s2(p1 − m1) + q2s2(m1 − n1) = q2s2(p1 − m1) + q1s2(m2 − n2). Next, we swap sides in (D3) and multiply by m2 − n2 to get (A.3) r1(m2 − n2)(p2 − m2) + q1s2(m2 − n2) = r2(m2 − n2)(p1 − m1) + q2s1(m2 − n2). Adding both sides of (A.2) and (A.3) gives an equation in which each of q1s2(m2−n2) and q2s1(m2 − n2) appears on both sides; cancelling them gives r1(m2 − n2)(p2 − m2)+q2s1(p2 − m2) = r2(m2 − n2)(p1 − m1) + q2s2(p1 − m1). KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 31 Equivalently, we have (cid:0)r1(m2 − n2) + q2s1(cid:1)(p2 − m2) =(cid:0)r2(m2 − n2) + q2s2(cid:1)(p1 − m1). Since the coefficients of (p2 − m2) and (p1 − m1) are both positive, we deduce that p2 − m2 and p1 − m1 have the same sign. Thus the hypothesis p1 > m1 implies that p2 > m2, as required. The second case we consider is when q1 = 0 = q2. Then (D1) gives no information and (D3) collapses to r2(p1 − m1) = r1(p2 − m2). Since ri 6= 0, we deduce that the numbers pi − mi have the same same sign. Since p1 > max{m1, n1}, we deduce that p2 > m2, as required. The remaining case to consider is when exactly one qi is zero. If q1 = 0 and q2 6= 0, then (D3) gives r1(p2 − m2) = r2(p1 − m1) + q2s1 > r2(p1 − m1) ≥ 0. Thus p2 > m2. On the other hand if q1 6= 0 and q2 = 0, then (D1) reduces to q1(n2 − m2) = 0, and so n2 = m2. Hence, p2 > n2 = m2 as required. This completes the direct proof of Theorem 4.2 for dumbbell graphs with three components. The next remark shows that we can relax the hypotheses of Theorem 4.2 slightly. Graphs of this type served as motivation for our development of Corollary 4.3. Remark A.1. Again consider a 2-graph Λ with three vertices, and keep the above notation. Suppose that r1 = r2 = 0, but all qi and si are nonzero. Notice that whilst CΛeiD = ∅, we still have CΛNeiD 6= ∅ since CΛeiB = ∅ and BΛeiD = ∅. As before, our goal is to show that if p1 = ρ(AD,1) > max{ρ(AC,1), ρ(AB,1)} = max{n1, m1}, then p2 > n2 and p2 > m2. We again derive (D1 -- D3). Notice that (D3) reduces to (D3') q2s1 = q1s2 ⇐⇒ q2 q1 As before, since p1 > n1 and si 6= 0, (D2) forces p2 > n2, and it remains to show that p2 > m2. = s2 s1 . First we consider the case where n1 = m1. Since qi 6= 0, (D1) shows that n2 = m2. Thus p2 > n2 = m2 as required. Secondly, we consider the situation where n1 6= m1. Combining (D1) and (D2) with (D3'), we get that n2 − m2 n1 − m1 = q2 q1 = s2 s1 = p2 − n2 p1 − n1 . Since p1 > n1 and p2 > n2, we must have that n2 6= m2, and n2−m2 and n1−m1 must have the same sign. If n1 > m1, then n2 > m2, and so p2 > n2 > m2 as required. Alternatively, n1 < m1. Since p1 > m1 by assumption, we have 0 < m1−n1 < p1−n1. Thus p2 − n2 p1 − n1 = m2 − n2 m1 − n1 > m2 − n2 p1 − n1 . Hence, p2 − n2 > m2 − n2, and so p2 > m2. Remark A.2. Now we consider the case where si = 0 for i = 1, 2. These graphs do not satisfy the hypotheses of either Theorem 4.2 or Corollary 4.3, and we shall see that they are examples of 2-graphs in which the conclusions of Theorem 4.2 or Corollary 4.3 do not hold. 32 JAMES FLETCHER, ASTRID AN HUEF, AND IAIN RAEBURN When si = 0, (D2) says nothing, and (D3) reduces to r2(p1 − m1) = r1(p2 − m2). Thus p1 > max{n1, m1} implies p2 > m2. But there is no relation relating p2 to n2. Indeed, consider the matrices 1 2 2 0 3 0 A1 = 0 0 5  0 0 3  and A2 =  .  1 3 1 0 4 0 Then there are 2-graphs Λ with these vertex matrices, and in fact many: see [13, §6] (or [5] for a concrete description of these graphs). For such Λ, we have ρ(AD,1) = 5 > max{ρ(AB,1), ρ(AC,1)} = max{3, 1} = 3, ρ(AD,2) = 3 < ρ(AB,2) = 4. but So there is no version of Theorem 4.2 for such Λ. References [1] Z. Afsar, A. an Huef and I. Raeburn, KMS states on C ∗-algebras associated to local homeo- morphisms, Internat. J. Math. 25 (2014), article no. 1450066, 1 -- 28. [2] Z. Afsar, A. an Huef and I. Raeburn, KMS states on C ∗-algebras associated to a family of ∗-commuting local homeomorphisms, arXiv:1701.07183. [3] J. Christensen, Symmetries of the KMS simplex, arXiv:1710:04412. [4] C. Farsi, E. Gillaspy, S. Kang and J.A. Packer, Separable representations, KMS states, and wavelets for higher-rank graphs, J. Math. Anal. Appl. 434 (2016), 241 -- 270. [5] R. Hazlewood, I. Raeburn, A. Sims and S.B.G. Webster, Remarks on some fundamental results about higher-rank graphs and their C ∗-algebras, Proc. Edinburgh Math. Soc. 56 (2013), 575 -- 597. [6] A. an Huef, S. Kang and I. Raeburn, Spatial realisations of KMS states on the C ∗-algebras of higher-rank graphs, J. Math. Anal. Appl. 427 (2015), 977 -- 1003. [7] A. an Huef, S. Kang and I. Raeburn, KMS states on the operator algebras of reducible higher- rank graphs, Integral Equations & Operator Theory 88 (2017), 91 -- 126. [8] A. an Huef, M. Laca, I. Raeburn and A. Sims, KMS states on the C ∗-algebras of finite graphs, J. Math. Anal. Appl. 405 (2013), 388 -- 399. [9] A. an Huef, M. Laca, I. Raeburn and A. Sims, KMS states on C ∗-algebras associated to higher-rank graphs, J. Funct. Anal. 266 (2014), 265 -- 283. [10] A. an Huef, M. Laca, I. Raeburn and A. Sims, KMS states on the C ∗-algebra of a higher-rank graph and periodicity in the path space, J. Funct. Anal. 268 (2015), 1840 -- 1875. [11] A. an Huef, M. Laca, I. Raeburn and A. Sims, KMS states on the C ∗-algebras of reducible graphs, Ergodic Theory Dynam. Systems 35 (2015), 2535 -- 2558. [12] A. an Huef and I. Raeburn, Equilibrium states on graph algebras, in Operator Algebras and Applications, Proc. 2015 Abel Symposium, Springer, 2016, pp. 171 -- 183. [13] A. Kumjian and D. Pask, Higher-rank graph C ∗-algebras, New York J. Math. 6 (2000), 1 -- 20. [14] A. Kumjian, D. Pask and A. Sims, Generalised morphisms of k-graphs: k-morphs, Trans. Amer. Math. Soc. 363 (2011), 2599 -- 2626. [15] M. Laca, N.S. Larsen, S. Neshveyev, A. Sims and S.B.G. Webster, Von Neumann algebras of strongly connected higher-rank graphs, Math. Ann. 363 (2015), 657 -- 678. [16] R. McNamara, KMS states of graph algebras with a generalised gauge dynamics, PhD thesis, Univ. of Otago, 2016. [17] S. Neshveyev, KMS states on the C ∗-algebras of non-principal groupoids, J. Operator Theory 70 (2013), 513 -- 530. [18] G.K. Pedersen, C ∗-Algebras and their Automorphism Groups, London Math. Soc. Mono- graphs, vol. 14, Academic Press, London, 1979. [19] I. Raeburn, Graph Algebras, CBMS Regional Conference Series in Math., vol. 103, Amer. Math. Soc., Providence, 2005. [20] I. Raeburn and A. Sims, Product systems of graphs and the Toeplitz algebras of higher-rank graphs, J. Operator Theory 53 (2005), 399 -- 429. KMS STATES ON HIGHER-RANK GRAPH ALGEBRAS 33 [21] I. Raeburn, A. Sims and T. Yeend, Higher-rank graphs and their C ∗-algebras, Proc. Edinburgh Math. Soc. 46 (2003), 99 -- 115. [22] D. Robertson and A. Sims, Simplicity of C ∗-algebras associated to row-finite locally convex higher-rank graphs, Israel J. Math. 172 (2009), 171 -- 192. [23] G. Robertson and T. Steger, Affine buildings, tiling systems and higher-rank Cuntz -- Krieger algebras, J. Reine Angew. Math. 513 (1999), 115 -- 144. [24] E. Seneta, Non-Negative Matrices and Markov Chains, second edition, Springer-Verlag, New York, 1981. [25] D. Yang, Endomorphisms and modular theory of 2-graph C ∗-algebras, Indiana Univ. Math. J. 59 (2010), 495 -- 520. [26] D. Yang, Factoriality and type classification of k-graph von Neumann algebras, Proc. Edin- burgh Math. Soc. 60 (2017), 499 -- 518. E-mail address: james.fletcher, astrid.anhuef, [email protected] School of Mathematics and Statistics, Victoria University of Wellington, P.O. Box 600, Wellington 6140, New Zealand.
0906.1267	2	0906	2010-07-20T16:30:09	A View on Optimal Transport from Noncommutative Geometry	[ "math.OA", "hep-th", "math-ph", "math-ph" ]	We discuss the relation between the Wasserstein distance of order 1 between probability distributions on a metric space, arising in the study of Monge-Kantorovich transport problem, and the spectral distance of noncommutative geometry. Starting from a remark of Rieffel on compact manifolds, we first show that on any - i.e. non-necessary compact - complete Riemannian spin manifolds, the two distances coincide. Then, on convex manifolds in the sense of Nash embedding, we provide some natural upper and lower bounds to the distance between any two probability distributions. Specializing to the Euclidean space $R^n$, we explicitly compute the distance for a particular class of distributions generalizing Gaussian wave packet. Finally we explore the analogy between the spectral and the Wasserstein distances in the noncommutative case, focusing on the standard model and the Moyal plane. In particular we point out that in the two-sheet space of the standard model, an optimal-transport interpretation of the metric requires a cost function that does not vanish on the diagonal. The latest is similar to the cost function occurring in the relativistic heat equation.	math.OA	math	Symmetry, Integrability and Geometry: Methods and Applications SIGMA 6 (2010), 057, 24 pages A View on Optimal Transport from Noncommutative Geometry(cid:63) Francesco D'ANDREA † and Pierre MARTINETTI ‡ † Ecole de Math´ematique, Univ. Catholique de Louvain, Chemin du Cyclotron 2, 1348 Louvain-La-Neuve, Belgium E-mail: [email protected] ‡ Institut fur Theoretische Physik, Universitat Gottingen, Friedrich-Hund-Platz 1, 37077 Gottingen, Germany E-mail: [email protected] Received April 14, 2010, in final form July 08, 2010; Published online July 20, 2010 doi:10.3842/SIGMA.2010.057 Abstract. We discuss the relation between the Wasserstein distance of order 1 between probability distributions on a metric space, arising in the study of Monge -- Kantorovich transport problem, and the spectral distance of noncommutative geometry. Starting from a remark of Rieffel on compact manifolds, we first show that on any -- i.e. non-necessary compact -- complete Riemannian spin manifolds, the two distances coincide. Then, on convex manifolds in the sense of Nash embedding, we provide some natural upper and lower bounds to the distance between any two probability distributions. Specializing to the Euclidean space Rn, we explicitly compute the distance for a particular class of distributions genera- lizing Gaussian wave packet. Finally we explore the analogy between the spectral and the Wasserstein distances in the noncommutative case, focusing on the standard model and the Moyal plane. In particular we point out that in the two-sheet space of the standard model, an optimal-transport interpretation of the metric requires a cost function that does not va- nish on the diagonal. The latest is similar to the cost function occurring in the relativistic heat equation. Key words: noncommutative geometry; spectral triples; transport theory 2010 Mathematics Subject Classification: 58B34; 82C70 1 Introduction The idea at the core of Noncommutative Geometry [11] is the observation that, in many in- teresting cases, the description of a space as a set of points is inadequate. Think for example of quantum mechanics, where position and momentum are replaced by non-commuting opera- tors: as a consequence, Heisenberg uncertainty relations impose limitations in the precision of their simultaneous measurement so that the notion of "point in phase space" loses any opera- tional meaning. Taking into account General Relativity, one can show by simple arguments that not only phase-space coordinates but also space-time coordinates should be non-commutative (cf. [17, 18] and references therein) making the concept of "points in space-time" also problem- atic. Noncommutative Geometry provides efficient tools to study these "spaces" that are no longer described by a commutative algebra of coordinate functions, but by some noncommutative operator algebra A. Losing the notion of points, one also loses the notion of distance between points. However one can still define a distance between states of the algebra A, for example with the help of (cid:63)This paper is a contribution to the Special Issue "Noncommutative Spaces and Fields". The full collection is available at http://www.emis.de/journals/SIGMA/noncommutative.html 2 F. D'Andrea and P. Martinetti a generalized Dirac operator D. The latter is the starting point of Connes theory of spectral triples [12], which is the datum (A,H, D) of an involutive (non necessarily commutative) algeb- ra A, a representation π of A as bounded operators on a Hilbert space H and a self-adjoint operator D, such that [D, a] is bounded and a(D−λ)−1 is compact for any a ∈ A and λ /∈ Sp(D) (where the symbol π is omitted). The spectral distance between two states ϕ1, ϕ2 of A is defined as [10, 13] (cid:8)ϕ1(a) − ϕ2(a); [D, a]op ≤ 1(cid:9), dD(ϕ1, ϕ2) . = sup a∈A (1.1) where the norm is the operator norm coming from the representation of A on H. It is easy to check that (1.1) defines a distance in a strict mathematical sense except that it may be infinite. Recall that a state is by definition a positive linear application ¯A → C with norm 1, where ¯A is the C∗-algebra completion of A. When A is the algebra of observables of a physical system, this notion of state coincides with physicist's intuition, namely density matrices or Gibbs states in statistical physics, state vectors or wave functions in quantum mechanics. For the commutative 0 (M) -- i.e. complex smooth functions vanishing at infinity on some spin manifold M algebra C∞ of dimension1 m -- a canonical spectral triple is (cid:0)C∞ 0 (M), L2(M,S), D(cid:1), m(cid:88) D = −i γµ∇µ, where L2(M,S) is the Hilbert space of square integrable spinors on which C∞ wise multiplication and D is the Dirac operator. The latest is given in local coordinates by (1.2) 0 (M) acts by point- µ=1 where ∇µ = ∂µ + ωµ is the covariant derivative associated to the spin connection 1-form ωµ and {γµ}µ=1,2,...,m are the self-adjoint Dirac gamma matrices satisfying γµγν + γνγµ = 2gµνI (1.3) with g the Riemannian metric of M and I the identity matrix of dimension 2n if m = 2n or 2n + 1. In this case ¯A = C0(M) and the spectral distance between pure states (i.e. states that cannot be written as a convex sum of other states), which by Gelfand transform are in one-to-one correspondence with the points of the manifold ∀ f ∈ C0(M), x (cid:55)→ δx : δx(f ) . = f (x) (1.4) coincides [15] with the geodesic distance d of M, ∀ x, y ∈ M. dD(δx, δy) = d(x, y) (1.5) In a noncommutative framework it is tempting -- inspired by (1.4) -- to take the set P(A) of pure states of A as the noncommutative analogue of points and dD as natural generalization of the geodesic distance. This idea has been tested in several examples inspired by physics: finite-dimensional algebras [5, 26], functions on M with value in a matrix algebra [35, 32, 33] encoding the inner structure of the space-time of the standard model of particles physics [8], 0 (M) [7]. Most often the computation of the supremum non-commutative deformations of C∞ in (1.1) is quite involved, however several explicit results have been obtained. They all indicate that as soon as A is noncommutative one looses an important feature of the geodesic distance, 1Unless otherwise specified, in all the paper we assume that "spin manifolds" are Riemannian, finite- dimensional, connected, complete, without boundary. A View on Optimal Transport from Noncommutative Geometry 3 namely P(A) is no longer a path metric space [25]. Explicitly there is no curve [0, 1] (cid:51) t (cid:55)→ ϕt ∈ P(A) such that dD(ϕs, ϕt) = t − sdD(ϕ0, ϕ1), (1.6) not even a sequence of such curves ϕn such that dD(ϕ0, ϕ1) = Inf n {length of ϕn between ϕ0 and ϕ1} . This can be seen on a simple noncommutative examples studied in [26], based on A = M2(C) acting on C2 with D = D∗ ∈ M2(C). P(M2(C)) is weak* homeomorphic to the Euclidean 2- sphere. The image under this homeomorphism of the pure states ωi = (ψi, .ψi), i = 1, 2, defined by the eigenvectors ψi of D are antipodal, and determine a distinguished vertical axis on S2. The spectral distance is invariant by rotation around this axis and the connected components (see (2.16) below) are circles parallel to the horizontal plane. Explicitly, on the circle with radius r ∈ [0, 1] one finds dD(θ1, θ2) = 2r D1 − D2 (cid:12)(cid:12)(cid:12)(cid:12) sin (cid:12)(cid:12)(cid:12)(cid:12) , θ1 − θ2 2 (1.7) where θ ∈ [0, 2π[ is the azimuth and Di are the eigenvalues of D. One then checks2 that (1.6) has no solution in P(M2(C)) [34]. The lack of geodesic curves (1.6) within P(A) is cured by conside- ring non-pure states. Indeed (1.1) not only generalizes the geodesic distance to noncommutative algebras, it also extends the distance to objects that are not equivalent to points, namely non- pure states. Noticing that (1.7) is the geodesic distance within the Euclidean disk of radius D1−D2 and that the set S(M2(C)) of states of M2(C) is homeomorphic to the 2-dimensional Euclidean ball (see Section 4.1), one easily obtains a curve in S(M2(C)) satisfying (1.6), namely 2r ϕt = (1 − t)ϕ0 + tϕ1. This remains true in full generality since, whatever algebra A and operator D, (cid:8)(s − t)(ϕ0 − ϕ1)(a); [D, a]op ≤ 1(cid:9) = s − tdD(ϕ0, ϕ1). dD(ϕs, ϕt) = sup a∈A (1.9) In other terms, in view of (1.9) and (1.5), the spectral distance is a natural generalization of the geodesic distance to the noncommutative setting as soon as one takes into account the whole space of states, and not only its extremal points. This motivates the study of the spectral distance between non-pure states that we undertake in this paper. We begin by giving a detailed proof of Rieffel's remark [39] (also mentioned in [4]), according to which in the commutative case A = 0 (M) the spectral distance coincides with a distance well known in optimal transport theory, C∞ namely the Wasserstein distance W of order 1 (see the bibliographic notes of [44, Chapter 6]). We stress in particular that M has to be complete for that result to hold besides the compact case. Then we present few explicit calculations of dD between non-pure states, and question on simple examples -- including the standard model -- the pertinence of the optimal transport interpretation of the spectral distance in a noncommutative framework. These are preliminaries 2A curve t (cid:55)→ θ(t) satisfies (1.6) if and only if (cid:12)(cid:12)sin θ(t)−θ(s) 2 (cid:12)(cid:12) = Kt − s (1.8) for any t, s ∈ [0, 1] and K a constant. The right hand side of (1.8) being a function of t−s, there exists a function f such that the left hand side is f (t − s). Putting s = 0, one obtains f (t) = sin θ(t)−θ(0) . Reinserted in (1.8), this yields Cauchy's functional equation θ(t)− θ(s) = θ(t− s)− θ(0), whose continuous solutions are linear. Since (1.8) has no linear solutions, this proves the claim. 2 4 F. D'Andrea and P. Martinetti results, intending to bring the attention of the transport theory community on the metric aspect of noncommutative geometry, and vice-versa. Notice that some properties of the spectral distance between non-pure states have been investigated in [39]: considering instead of [D, a]op an arbitrary semi-norm on A, it is shown that the knowledge of the distance between pure states of a noncommutative A may not be enough to recover the semi-norm on A. One also needs the distance between non-pure states. This suggests that the metric information encoded in formula (1.1) is not exhausted once one knows the distance between pure states. The plan of the paper is the following. In Section 2 we recall some basics of transport theory and noncommutative geometry in order to establish -- in Proposition 2.1 -- the equality between dD and W for any complete spin manifold M. We also discuss various definitions of the spectral distance, characterize its connected components and emphasize the importance of the completeness condition. In Section 3 we provide some lower and upper bounds for the distance. Specializing to M = Rn, we explicitly compute the distance between a class of states generalizing Gaussian wave-packets. Section 4 deals with noncommutative examples. We show that on the truncations of the Moyal plane introduced in [7], the Wasserstein distance WD with cost dD defined on P(A) does not coincide with the spectral distance on S(A). But on almost- commutative geometries, including the standard model of elementary particles, the spectral distance between certain classes of states may be recovered as a Wasserstein distance W (cid:48) with cost d(cid:48) defined on a subset of S(A) containing P(A). We also point out a reformulation of the spectral distance on P(A) in term of the minimal work WI associated to a cost cI non-vanishing on the diagonal (i.e. cI (x, x) (cid:54)= 0). 2 Spectral distance as Wasserstein distance of order 1 2.1 Spectral distance from Kantorovich duality For any locally compact Hausdorff topological space X , states ϕ ∈ S(C0(X )) are given by Borel probability measures µ on X , via the rule ϕ(f ) . = f dµ ∀ f ∈ A. (2.1) (cid:90) X This is a simple application of Riesz representation and Hahn -- Banach theorems together with the assumption that X is σ-compact in order to avoid regularity problems. Any such µ defines a state since f vanishing at infinity (hence being bounded) guarantees that (2.1) is finite. Pure states correspond to Dirac-delta measures. To provide some physical intuition, one can view ϕ as a wave-packet and imagine that it describes the probability distribution of a bunch of particles. Strictly speaking a wave-packet is a square root of the Radon -- Nikodym derivative of dµ with respect to some fixed σ-finite positive measure dx on X (it is a square-integrable function, almost everywhere defined and unique modulo a phase, whenever dµ is absolutely continuous with respect to dx). For instance in quantum mechanics, with X = Rn and dx the Lebesgue measure, a wave-packet is a function φ ∈ L2(Rn) and the corresponding measure is dµ = φ(x)2dx. Assuming X is a metric space with distance function d, there is a natural way to measure how much two states ϕ1 and ϕ2 differ, which is the expectation value of the distance between the two corresponding distributions (cid:90) E(d; µ1 × µ2) = d(x, y)dµ1(x)dµ2(y), X×X where µi is associated to ϕi via (2.1). Other ways are suggested by transport theory (all material here is taken from [1, 43, 44, 6], where an extensive bibliography can be found. Following [43] A View on Optimal Transport from Noncommutative Geometry 5 we assume from now on that X has a countable basis, so that it is a Polish space). Assume there exists a positive real function c(x, y) -- the "cost function" -- that represents the work needed to move from x to y. A good measure on how much the ϕi's differ is given by the minimal work W required to move the bunch of particles from the configuration ϕ1 to the configuration ϕ2, namely W (ϕ1, ϕ2) . = inf π X×X c(x, y)dπ, (2.2) where the infimum is over all measures π on X × X with marginals µ1, µ2 (i.e. the push- forwards of π through the projections X, Y : X × X → X , X(x, y) . = y, are X∗(π) = µ1 and Y∗(π) = µ2). Such measures are called transportation plans. Finding the optimal transportation plan, that is the one which minimizes W , is a non-trivial question known as the Monge -- Kantorovich problem. This is a generalization of Monge [36] "d´eblais et remblais" problem, where one considers only those transportation plans that are supported on the graph of a transportation map, i.e. a map T : X → X such that T∗µ1 = µ2. Namely, . = x, Y(x, y) c(x, T (x))dµ1(x). (2.3) (cid:90) (cid:90) X WMonge(ϕ1, ϕ2) . = inf T (cid:90) X One of the interests of Kantorovich's generalization [28] is that the infimum in (2.3) is not always a minimum: an optimal transportation map may not exist. On the contrary the infimum in (2.2) is a minimum and always coincides with Monge infimum, even when the optimal transportation map does not exist. Moreover when the cost function c is a distance d, (2.2) is in fact a distance on the space of states -- with the infinite value allowed -- called the Kantorovich -- Rubinstein distance (this case was first studied in [29]). To be sure it remains finite (see (2.13) below), it is convenient to restrict to the set S1(C0(X )) of states whose moment of order 1 is finite, that is those distributions µ such that E(d(x0,◦); µ) = d(x0, x)dµ(x) < +∞, (2.4) where x0 is an arbitrary but fixed point in X . Note that as soon as E(d(x0,◦); µ) is finite for x0, then by the triangle inequality E(d(x,◦); µ) ≤ E(d(x0,◦); µ) + d(x0, x) is finite for any x ∈ X so that S1(C0(X )) is independent on the choice of x0. The Kantorovich -- Rubinstein distance is also known as the Wasserstein distance of order 1. The distance of order p is given by a similar formula with E(dp; µ1 ⊗ µ2)1/p, 1 ≤ p < ∞, but in this paper we are interested only in the distance of order one and we shall simply call W the Wasserstein distance. The link with non-commutative geometry, which seems to have been first noted for M com- pact in [39], is the following: when X is a spin manifold M, the Wasserstein distance with cost function the geodesic distance d is nothing but the spectral distance (1.1) associated to the canonical spectral triple (1.2). This is a priori not difficult to see: On one side a central result of transport theory, Kantorovich duality, provides a dual formulation of the Wasserstein distance as a supremum instead of an infimum, namely (cf. Theorem 5.10 and equation (5.11) of [44]) W (ϕ1, ϕ2) = sup fLip≤1, f∈L1(µ1)∩L1(µ2) X f dµ2 (2.5) for any pair of states in S(C0(X )) such that the right-hand side in the above expression is finite. The supremum is on all real µi=1,2-integrable functions f that are 1-Lipschitz, that is to say f (x) − f (y) ≤ d(x, y) ∀ x, y ∈ X . (cid:18)(cid:90) (cid:90) X f dµ1 − (cid:19) 6 F. D'Andrea and P. Martinetti On the other side, the commutator −i[γµ∂µ, f ] (where we use Einstein summation convention and sum over repeated indices) acts on L2(M,S) as multiplication by −iγµ∂µf . Moreover the 0 (M) simply supremum in (1.1) can be searched on self-adjoint elements [26], that for A = C∞ means real functions f . Thus op = γµ∂µf2 op = (γµ∂µf )(γν∂νf )op = 1 2 (γµγν + γνγµ)∂µf ∂νfop [D, f ]2 Lip, = Igµν∂µf ∂νfop = gµν∂µf ∂νf∞ = f2 ces, equation (1.3), that (cid:112)gµν∂µf ∂νf evaluated at x is the norm of the gradient ∇f (x) and (2.6) where we used: the C∗-property of the norm, that ∂µf commutes with ∂νf and the γ's matri- supx∈M ∇f (x)TxM = fLip by Cauchy's mean value theorem. Consequently the commu- tator-norm condition in the spectral distance formula yields on f the condition required in Kantorovich's dual formula. However one has to be careful that, although the ϕi's on the l.h.s. 0 (M), the supremum on the r.h.s. includes functions non-vanishing of (2.5) denote states of C∞ at infinity. Therefore (1.1) equals (2.5) if and only if the supremum on 1-Lipschitz smooth functions vanishing at infinity is the same as the supremum on 1-Lipschitz continuous functions 0 (M) is dense within C0(M) is well known, but non-necessarily vanishing at infinity. That C∞ it might be less known that continuous K-Lipschitz functions can be approximated by smooth K-Lipschitz functions. In the following we use this result -- proved for finite-dimensional man- ifolds in [24] and extended to infinite dimension in [2] -- in order to to prove Rieffel's remark, generalized to complete locally compact manifolds (e.g. complete finite-dimensional manifolds). 0 (M)) with M a (complete, Riemannian, finite di- Proposition 2.1. For any ϕ1, ϕ2 ∈ S(C∞ mensional, connected, without boundary) spin manifold, one has W (ϕ1, ϕ2) = dD(ϕ1, ϕ2). Proof . i) It is well known [2] that on Rn K-Lipschitz functions are the uniform limit of smooth K-Lipschitz functions (for any K ≥ 0). It may be less known that the same is true for any (finite- dimensional) Riemannian manifold M, and for K-Lipschitz functions vanishing at infinity. Let us give a short proof of this result. According to Theorem 1 in [2] (that is valid for separable Riemannian manifolds, so in particular for finite-dimensional manifolds), given a Lipschitz function f , for any continuous function ε : M → R+ and for any r > 0 there exists a smooth function gε,r : M → R such that gε,rLip ≤ fLip + r and f (x) − gε,r(x) ≤ ε(x) ∀ x ∈ M. As a corollary, if f is a K-Lipschitz function vanishing at infinity, we can fix a sequence εn of continuous functions vanishing at infinity and uniformly converging to zero, and a sequence of positive numbers rn converging to zero in order to get a sequence of smooth functions gn : M → R such that gnLip ≤ K + rn and f (x) − gn(x) ≤ εn(x) ∀ x ∈ M. Obviously gn vanishes at infinity (since both f and εn vanish at infinity). Let us call fn := K(K + rn)−1gn. The fn's are the required smooth K-Lipschitz functions vanishing at infinity that converge uniformly to f . Indeed f − fn = K K + rn (f − gn) + rn K + rn f implies f (x) − fn(x) ≤ K K + rn εn(x) + rn K + rn sup x f (x), sup x sup x A View on Optimal Transport from Noncommutative Geometry 7 and the right hand side goes to zero for n → ∞ since supx εn(x) → 0 and rn → 0 by assumption, while supx f (x) is finite since f is continuous vanishing at infinity. ii) By (2.6) the supremum in (1.1) is on 1-Lipschitz smooth functions vanishing at infinity. By i) above any 1-Lipschitz function f in C0(M) can be uniformly approximated by smooth 0 (M). Since any state ϕ of C0(M) is continuous [3] with respect 1-Lipschitz functions fn in C∞ to the sup-norm, namely the supremum in (1.1) can be equivalently searched on continuous functions and the spectral distance writes n→+∞ ϕ(fn) = ϕ(f ), lim n→+∞f − fn∞ = 0 =⇒ lim (cid:18)(cid:90) dD(ϕ1, ϕ2) = sup f∈C0(M,R), fLip≤1 M (cid:90) M f dµ1 − (cid:19) f dµ2 . (2.7) iii) In case M is compact C0(M, R) coincides with C(M, R) and (2.7) equals (2.5), hence In case M is only locally compact, C0(M, R) ⊂ C(M, R) so that dD(ϕ1, ϕ2) ≤ the result. W (ϕ1, ϕ2). To get the equality, it is sufficient to show that to any 1-Lipschitz µi-integrable function f ∈ C(M, R) is associated a sequence of functions fn ∈ C0(M, R) such that fnLip ≤ 1 and lim n→+∞ ∆(fn) = ∆(f ), (2.8) (2.9) where ∆(f ) = ϕ1(f ) − ϕ2(f ). We claim that such a sequence is given by . = f (x)e−d(x0,x)/n, . n ∈ N, fn(x) (2.10) where x0 is any fixed point. Indeed since M is complete, d(x0, x) diverges at infinity as explained in Section 2.3 below; by the 1-Lipschitz condition f (x) ≤ f (x0) + d(x0, x), and this proves To obtain (2.8), one first notices that ∆(f +C) = ∆(f ) for any C ∈ R, so that we can assume that fn(x) ≤(cid:0)f (x0) + d(x0, x)(cid:1)e−d(x0,x)/n vanishes at infinity. without loss of generality that f (x0) = 0, that is to say f (x) = f (x) − f (x0) ≤ d(x0, x). Second, from ∇fn = (cid:0)∇f − n−1f∇d(x0,◦ )(cid:1) e−d(x0,◦)/n and remembering that both f and d(x0,◦) are 1-Lipschitz, one gets ∇fn ≤(cid:0)1 + n−1d(x0,◦)(cid:1)e−n−1d(x0,◦). The inequality (2.8) then follows from (1 + ξ)e−ξ ≤ 1 ∀ ξ > 0. (2.9) comes from Lebesgue's dominated convergence theorem: fn(x) ≤ f (x) ∀ x, n and f is µi integrable by hypothesis, (cid:4) so that lim n→∞ (cid:82) M fndµi =(cid:82) M f dµi. 2.2 Alternative def initions There exist several equivalent formulations of the spectral distance. First one may consider continuous instead of smooth functions. Indeed, as explained in [11], for any measurable bounded function f one can view [D, f ] as the bilinear form ξ, η → (cid:104)Dξ, f η(cid:105) − (cid:104)f∗ξ, Dη(cid:105), 8 F. D'Andrea and P. Martinetti well defined on the domain of D (a dense subset of L2(M,S)). Therefore [D, f ] makes sense also when f is not smooth and one can define [12] the spectral distance as the supremum on all continuous functions f ∈ C0(M ) with [D, f ] ≤ 1, that is the set of 1-Lipschitz functions, ob- taining thus directly (2.7). In the literature one finds both definitions: supremum on continuous functions [12, 23] or on smooth functions [13, 16]. Second, one may be puzzled by the use of the spin structure to recover the Riemannian metric. In fact instead of the Dirac operator one can equivalently use, as explained in [14], the signature operator d + d† acting on the Hilbert space L2(M,∧) of square-integrable differential forms, or the de-Rham Laplacian ∆ = dd† + d†d acting on L2(M). Here d is the exterior derivative and d† its adjoint with respect to the inner product [23] (cid:90) M (cid:104)ω, ω(cid:48)(cid:105) = (ω, ω(cid:48))νg ∀ ω, ω(cid:48) ∈ L2(M,∧) with νg the volume form and the inner product on k-form given by (cid:0)dxα1 ∧ ··· ∧ dxαk , dxβ1 ∧ ··· ∧ dxβk(cid:48)(cid:1) = δkk(cid:48) det(cid:0)gαiβj(cid:1) 1 ≤ i, j ≤ k. The action of both operators only depends on the Riemannian structure and suitable commu- 0 (M) the same semi-norm as the commutator tators yield on self-adjoint elements f = f∗ in C∞ with D. Explicitly3 [d + d†, f ]2op = 1 2[[∆, f ], f ]op = [D, f ]2 op. (2.11) To show these equalities, let us note that on L2(M,∧), [d, f ] = (df ) where denotes the wedge multiplication, (df )ω = df ∧ ω. . Therefore [d, f ] commutes with 0-form, and the same is true for its adjoint [d, f ]†. Assuming f = f∗, that is [d†, f ] = −[d, f ]†, few manipulations with commutators yield [[∆, f ], f ] = [[d, f ]d†, f ] + [d[d†, f ], f ] + [[d†, f ]d, f ] + [d†[d, f ], f ] = 2[d, f ][d†, f ] + 2[d†, f ][d, f ] = 2[d + d†, f ]2, (2.12a) (2.12b) (2.12c) where [d, f ]2 = [d†, f ]2 = 0 due to the graded commutativity of the wedge product. Remembering that (dxµ)† = gµνι(∂ν) where ι is the contraction (see e.g. [23]), one gets [d, f ]† = (df )† = ((∂µf )(dxµ))† = (gµν∂νf )ι(∂µ) so that (2.12b) becomes [[∆, f ], f ] = −2∂ρf (gµν∂νf ) ((dxρ)ι(∂µ) + ι(∂µ)(dxρ)) = −2gνρ∂ρf ∂νf, where we used the fact that (dxρ)ι(∂µ) + ι(∂µ)(dxρ) = δρ [d + d†, f ]2 = 1 whose operator norm is gµν∂µf ∂νf∞. (2.11) then follows from (2.6). µ. With (2.12c) this shows that 2 [[∆, f ], f ] is the operator of point-wise multiplication by the function gµν∂µf ∂νf , Notice that instead of the Laplacian one could use any other 2nd order differential operator with the same principal symbol. 3Although we use the same symbol, one should keep in mind that the three operator norms in (2.11) correspond to actions on different Hilbert spaces: L2(M,∧), L2(M), L2(M,S). A View on Optimal Transport from Noncommutative Geometry 9 2.3 On the importance of being complete At point iii in the proof of Proposition 2.1 it is crucial that M is complete. By the Hopf -- Rinow theorem a complete finite-dimensional Riemannian manifold is a proper metric space, hence the geodesic distance from any fixed point x0, x (cid:55)→ d(x0, x), is a proper map [42]. In particular for non-compact M this means that d(x0, x) diverges at infinity, so that the functions fn in (2.10) vanish at infinity. When M is not complete, not only the fn do not vanish at infinity which spoils the proof, but also the definition of the Wasserstein distance requires more attention. Indeed in [43] Kantorovich duality is proved assuming X is complete. It is not clear to the authors whether the duality holds in the non-complete case. However one can still take (2.5) as a definition of W , letting aside whether this is the same quantity as (2.2) or not. Then it is easy to see that on non-complete M the spectral distance and W are not necessarily equal. Suppose indeed that N is compact, and M is obtained from N by removing a point x0, so that M is locally D computed on M and N are equal. compact and not complete. The spectral distances dM Indeed by (2.7), which holds also in the non-complete case, we can compute dD as supremum over continuous functions instead of smooth ones; in the computation of the spectral distance on N , it is equivalent to take the supremum over 1-Lipschitz f ∈ C(N ) or over 1-Lipschitz f(cid:48) = f − f (x0) ∈ C(N ) vanishing at x0 (since ϕ1(f )− ϕ2(f ) = ϕ1(f(cid:48))− ϕ2(f(cid:48)) for any two states ϕ1, ϕ2); therefore D and dN (cid:8)ϕ1(f ) − ϕ2(f ); fLip ≤ 1(cid:9) (cid:8)ϕ1(f ) − ϕ2(f ); fLip ≤ 1(cid:9) = dM D (ϕ1, ϕ2), dN D (ϕ1, ϕ2) = sup f∈C(N ) = sup f∈C(N ),f (x0)=0 where in last equality we used C0(M) = {f ∈ C(N ), f (x0) = 0}. On the contrary, W computed between pure states coincide with the geodesic distance, and the latest may or may not be the same on N and M. For instance one does not modify the geodesic distance by removing a point from the two sphere. But taking for N the circle, thought of as the closed interval [0, 1] with 0 and 1 identified, and M the open interval (0, 1), one gets WN (x, y) = min{x − y, 1 − x − y} whereas WM(x, y) = x − y. Thus on M = (0, 1) dM D (x, y) = dN D (x, y) = WN (x, y) (cid:54)= WM(x, y). 2.4 On the hypothesis of f inite moment of order 1 Restricting to the states with finite moment of order one (cf. (2.4)) guarantees that Wasserstein distance is finite, since by (2.5) W (ϕ1, ϕ2) = sup fLip≤1 (cid:18)(cid:90) (cid:90) (cid:90) (cid:0)f (x) − f (x0)(cid:1)dµ1(x) − (cid:12)(cid:12)f (x) − f (x0)(cid:12)(cid:12)dµ1(x) + sup X X (cid:19) (cid:0)f (x) − f (x0)(cid:1)dµ2(x) (cid:90) (cid:12)(cid:12)f (x) − f (x0)(cid:12)(cid:12)dµ2(x) ≤ sup fLip≤1 X (cid:90) X ≤ d(x, x0)dµ1(x) + (cid:90) X fLip≤1 d(x, x0)dµ2(x) < ∞. X An obvious upper-bound is then obtained by choosing π = µ1 × µ2 in (2.2): dD(ϕ1, ϕ2) = W (ϕ1, ϕ2) ≤ E(d; µ1 × µ2). When at least one of the states is pure, this upper bound is an exact result, even outside S1(C∞ 0 (M)). (2.13) (2.14) 10 Proposition 2.2. For any x ∈ M and any state ϕ ∈ S(C∞ dD(ϕ, δx) = W (ϕ, δx) = E(cid:0)d(x,◦); µ(cid:1). F. D'Andrea and P. Martinetti 0 (M)), (cid:19) (cid:18)(cid:90) (cid:18)(cid:90) (cid:90) M Proof . By Proposition 2.1, dD(ϕ, δx) = W (ϕ, δx) = = ≤ ≤ sup fLip≤1,f∈L1(µ) sup fLip≤1,f∈L1(µ) M f dµ − f (x) (cid:19) (f (y) − f (x)(cid:1) dµ(y) M f (x) − f (y) dµ(y) sup fLip≤1,f∈L1(µ) d(x, y) dµ(y) = E(cid:0)d(x,◦); µ(cid:1). (cid:90) M 0 (M)), This supremum is attained on the 1-Lipschitz functions f (y) or is obtained by the sequence fn as defined in (2.10) in case µ has not a finite moment of (cid:4) order 1. since E(cid:0)d(x,◦); µ(cid:1) is either infinite or convergent, the integrand in (2.4) being a positive function. Notice that this proposition does not rely on the finiteness of W nor dD, and makes sense = d(x, y) in case µ ∈ S1(C∞ . When the distributions are both localized around two points, x, y, transportation maps are simply paths from x to y and the minimal work coincides with the cost c(x, y) to move from x to y, i.e. with the geodesic distance. In other words dD(δx, δy) = W (δx, δy) = d(x, y) (2.15) and one retrieves (1.5). Proposition 2.2 also yields an alternative definition of S1(C∞ Corollary 2.1. ϕ ∈ S1(C∞ any pure state. 0 (M)) if and only if ϕ is at f inite spectral-Wasserstein distance from 0 (M)). Proof . By the triangle inequality, the moment of order one of ϕ is finite for a fixed x = x0 if and only if it is finite for all x ∈ M. (cid:4) Let us conclude this section with some topological remarks. Definition 2.1. Given an arbitrary spectral triple (A,H, D), we call [33] Con(ϕ) = {ϕ(cid:48) ∈ S(A), dD(ϕ, ϕ(cid:48)) < +∞}. . Notice that connected components in S(A) for the topology induced by dD coincide with sets of states at finite distance from each other, thus justifying the name Con(ϕ). Indeed Con(ϕ) is path-connected since for any ϕ0, ϕ1 ∈ Con(ϕ) the map [0, 1] (cid:51) t (cid:55)→ ϕt = (1 − t)ϕ0 + tϕ1 ∈ Con(ϕ) is continuous (for all > 0 called δ = /d(ϕ0, ϕ1) from (1.9) we get t−s < δ ⇒ dD(ϕt, ϕs) < ). That Con(ϕ) is maximal -- i.e. there is no connected component containing it properly -- can be easily seen: for any ϕ(cid:48) ∈ Con(ϕ) any open ball centered at ϕ(cid:48) is contained in Con(ϕ), so that Con(ϕ) is open; by the triangle inequality the same is true for the complementary set, proving that Con(ϕ) is also closed. Therefore, any set containing properly Con(ϕ) is not connected since it contains a subset that is both open and closed. For the Wasserstein distance, the set of states with finite moment of order one is a connected component. A View on Optimal Transport from Noncommutative Geometry Corollary 2.2. For any ϕ ∈ S1(C∞ Proof . Let ϕ ∈ S1(C∞ 0 (M)). By the triangle inequality, 0 (M)), Con(ϕ) = S1(C∞ 0 (M)). dD(ϕ, ϕ(cid:48)) ≤ dD(ϕ, δx) + dD(δx, ϕ(cid:48)) is finite by Corollary 2.1 as soon as ϕ(cid:48) ∈ S1(C∞ 0 (M)). Similarly dD(δx, ϕ(cid:48)) ≤ dD(δx, ϕ) + dD(ϕ, ϕ(cid:48)) implies that dD(ϕ, ϕ(cid:48)) is infinite when ϕ(cid:48) /∈ S1(C∞ 0 (M)). 11 (cid:4) (cid:103)Con(ϕ) Notice that considering only pure states, the set = Con(ϕ) ∩ P(A), . ϕ ∈ P(A) is not necessarily (path)-connected in P(A). In the example (1.7) (cid:103)Con(ϕ) is indeed a connected component of P(A); but in the standard model (see Section 4.2) (cid:103)Con(ϕ) = P(A) and contains (2.16) two disjoint connected components (M, δC), (M, δH). 3 Bounds for the distance and (partial) explicit results Having in mind that noncommutative geometry furnishes a description of the full standard model of the electro-weak and strong interactions minimally coupled to Euclidean general relativity [8], computing the spectral distance could be a way to obtain a "picture" of spacetime at the scale of unification. Regarding pure states, this picture has been worked out in [35] and is recalled in Section 4: one finds that the connected component of dD is the disjoint union of two copies of a spin manifold M, with distance between the copies coming from the Higgs field. Extending this picture to non-pure states is far from trivial since already on Euclidean space explicit computations of the Wasserstein distance are very few. This does not seem to be the most interesting issue in optimal transport, where one is rather interesting in determining the optimal plan than computing W . On the contrary from our perspective computing dD is of most interest, while finding the optimal plan (i.e. the element that reaches the supremum in the distance formula) is not an aim in itself. In this section, we collect various results on the Wasserstein-spectral distance in the commutative case A = C∞ 0 (M): upper and lower bounds for the distance on any spin manifold M, and explicit result for a certain class of states in case M = Rn. Some of these results might be known from optimal transport theory, but we believe it is still interesting to present them from our perspective. We postpone to the next section a discussion of the noncommutative case. In all this section, dD = W and to avoid repetition we simply call it the spectral distance. 3.1 Upper and lower bounds on any spin manifold On a spin manifold, a lower bound on the spectral distance between states with finite moment of order 1 is given by the distance between their mean points. For M = Rm the mean point of ϕ ∈ S1(C∞ (cid:90) 0 (M)) is the barycenter of µ, namely with ¯xα = E(xα; µ) = . = (¯xα) xαdµ, Rm (3.1) ¯x where xα : Rm → R, x (cid:55)→ xα(x), 12 F. D'Andrea and P. Martinetti with α = 1, . . . , m denote a set of Cartesian coordinate functions on Rm. For M that is not the Euclidean space, the mean point can be defined through Nash embedding, that is an isometric embedding of M onto a subset (cid:102)M of the Euclidean space Rn for some n ≥ m [37, 38]. We say that the embedding N : M → (cid:102)M is convex if (cid:102)M is a convex subset of Rn. In that case, the = N∗µ is in (cid:102)M, and . barycenter x of µ = N−1(x) ∈ M . ¯x is a well defined generalization of (3.1) to M. Before showing in Proposition 3.1 that the distance between mean points bounds from below the spectral distance, let us recall two properties of Nash embedding that will be useful in the following. Lemma 3.1. Let M be a Riemannian manifold admitting a convex isometric embedding N : M → (cid:102)M ⊂ Rn. Then d(x, y) = N (x) − N (y), . (3.2) where · is the Euclidean distance on Rn. Moreover, if f is a 1-Lipschitz function on the metric space (M, d) then f = f ◦ N−1 is 1-Lipschitz on ((cid:102)M, · ). Euclidean metric of Rn, geodesics are straight lines and the distance is the Euclidean one. This proves (3.2). Proof . An isometry N : M → (cid:102)M preserves the distance function and sends geodesics to geodesics (cf. e.g. [9, page 61]). If (cid:102)M is convex, since the metric on (cid:102)M is the restriction of the As a consequence, if f is a 1-Lipschitz function on M, for any x, y ∈ (cid:102)M we have where x = N−1(x) and y = N−1(y). This means that f is 1-Lipschitz on (cid:102)M. (cid:4) Proposition 3.1. Let M be a Riemannian spin manifold that admits a convex isometric em- bedding M (cid:44)→ Rn. For any states ϕ1, ϕ2 in S(C0(M)) with mean points ¯x1, ¯x2, f (x) − f (y) = f (x) − f (y) ≤ d(x, y) = x − y, Proof . (2.14) holds true, so we need to prove only the lower bound. Let us fix a basis of the vector space Rn such that (3.3) xβdµ2 = 0 for β ∈ [2, n] . (cid:90) (cid:102)M (cid:12)(cid:12)(cid:12) f dµ2 (cid:1). d(¯x1, ¯x2) ≤ dD(ϕ1, ϕ2) ≤ E(cid:0)d; µ1 × µ2 x1 − x2 =(cid:0)x1 − x2, 0, . . . , 0(cid:1), (cid:90) (cid:102)M (cid:90) (cid:102)M x1dµ1 − β = 1, . . . , n. xβ d µi, x1dµ2 = x1 − x2 and (3.3) is equivalent to where xi = N (¯xi), i = 1, 2, has Cartesian coordinates (cid:90) (cid:102)M (cid:12)(cid:12)(cid:12)(cid:90) (cid:102)M Therefore x1 − x2 = x1dµ1 − x1dµ2 (cid:90) (cid:102)M (cid:90) (cid:102)M (cid:12)(cid:12)(cid:12) ≤ sup fLip≤1 (cid:90) (cid:102)M xβdµ1 − (cid:12)(cid:12)(cid:12)(cid:90) (cid:102)M f dµ1 − A View on Optimal Transport from Noncommutative Geometry where we noticed that x1 is Lipschitz in (cid:102)M with constant 1 since Using(cid:82)(cid:102)M f dµ =(cid:82) (x1(x) − x1(y))2 + M f dµ and Lemma 3.1, x1(x) − x1(y) ≤ (cid:114) α=2 (cid:88)n (cid:12)(cid:12)(cid:12)(cid:12) = dD(ϕ1, ϕ2). (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) M x1 − x2 ≤ sup fLip≤1 f dµ1 − f dµ2 M (cid:90) (xα(x) − yα(x))2 = x − y. 13 Proposition follows from (3.2), i.e. x1 − x2 = d(¯x1, ¯x2). (cid:4) Note that when M does not admit a convex embedding, Proposition 3.1 still holds with x1 − x2 instead of d(¯x1, ¯x2). However this might not be the most interesting lower bound since it involves a distance on Rn that is not the push-forward of the one on M (the points xi may of the geodesic distance on M). not be in (cid:102)M, and the Euclidean distance, even when restricted to (cid:102)M, is not the push-forward 3.2 Spectral distance in the Euclidean space In this section we study the spectral distance between states of S1(C∞ 0 (M)) in the case M = Rm. To a given density probability ψ ∈ L1(Rm) with finite moment of order 1, e.g. a Gaussian ψ(x) = π− m 2 e−x2, one can associate a state Ψσ,x given by Ψσ,x(f ) . = 1 σm Rm f (ξ)ψ( ξ−x σ )dmξ, (cid:90) (cid:90) for any x ∈ Rm and σ ∈ R+. This becomes the pure state (the point) δx in the σ → 0+ limit ψσ,x(f ) = f (x). In this sense Ψσ,x can be viewed as a "fuzzy" point, that is to say since lim σ→0+ a wave-packet -- characterized by a shape ψ and a width σ -- describing the uncertainty in the localization around the point x. The spectral distance between wave packets with the same shape is easily calculated. Proposition 3.2. The distance between two states Ψσ,x and Ψσ(cid:48),y is dD(Ψσ,x, Ψσ(cid:48),y) = x − y + (σ − σ(cid:48))ξψ(ξ)dmξ. (3.4) In particular, for σ = σ(cid:48) the distance does not depend on the shape ψ: dD(Ψσ,x, Ψσ,y) = x − y. Proof . Since f (z) − f (w) ≤ z − w for any 1-Lipschitz function f , we have Ψσ,x(f ) − Ψσ(cid:48),y(f ) = ≤ (cid:90) (cid:0)f (σξ + x) − f (σ(cid:48)ξ + y)(cid:1) ψ(ξ)dmξ (cid:90) x − y + (σ − σ(cid:48))ξ ψ(ξ) dmξ. When σ = σ(cid:48) the upper bound is attained by the function h(z) = z · x − y x − y , 14 F. D'Andrea and P. Martinetti Figure 1. The function h that attains the supremum, in case σ > σ(cid:48) and σ = σ(cid:48). Dot lines are tangent to the gradient of h. while for σ (cid:54)= σ(cid:48) it is attained by the function h(z) = z − α , where α . = σ(cid:48)x − σy σ(cid:48) − σ as shown in Fig. 1. (3.5) (cid:4) In case σ (cid:54)= σ(cid:48) the function h that attains the supremum measures the geodesic distance between z ∈ Rm and the point α defined in (3.5). Geometrically the latest is the intersection of (x, y) with (x + σξ, y + σ(cid:48)ξ), and is independent of ξ. In case σ = σ(cid:48), α is send to infinity and h measures the length of the projection of z on the (x, y) axis (see Fig. 1). The picture is still valid for pure states, i.e. σ = σ(cid:48) = 0: h can be taken either as the geodesic distance to any point on (x, y) outside the segment [x, y], or as the distance to any axis perpendicular to (x, y) that does not intersect [x, y]. For Gaussian shape, Proposition 3.2 can be confronted with the Wasserstein distance of order 2 between Gaussians computed in [21]. The final expression for the distance of order 2 is quite simpler, since it is an algebraic function of the means and covariance matrices. Here, computing the integral in (3.4) for ψ a Gaussian with the help of a symbolic computation software, one finds complicated expressions involving Bessel functions. The distance between two arbitrary states on Rm is less easily computable. Let us first recall what happens in the one dimensional case (cf. e.g. [19]). Proposition 3.3. For any two states ϕ1 and ϕ2, whose corresponding probability measures µ1 and µ2 have no singular continuous part, (cid:0)dµ1(ξ) − dµ2(ξ)(cid:1)(cid:12)(cid:12)(cid:12)(cid:12) dz. (cid:90) +∞ (cid:12)(cid:12)(cid:12)(cid:12)(cid:90) z (cid:90) R dD(ϕ1, ϕ2) = −∞ −∞ Proof . Integrating by parts one finds ϕ1(f ) − ϕ2(f ) = − f(cid:48)(z)∆(z)dz, A View on Optimal Transport from Noncommutative Geometry where ∆(z) . = (cid:90) z −∞ (cid:0)dµ1(ξ) − dµ2(ξ)(cid:1) 15 (3.6) is the cumulative distribution of the measure µ1 − µ2. We are assuming that µi are the sum of a pure point part and a part that is absolutely continuous with respect to the Lebesgue measure. In this case ∆(z) is piecewise continuous, its sign is a piecewise continuous function, and the primitive h(z) of the sign of ∆(z) is a Lipschitz continuous function. Now, the 1-Lipschitz condition says that a.e. f(cid:48)(z) ≤ 1 on R, hence (cid:90) R dD(ϕ1, ϕ2) ≤ ∆(z)dz. This upper bound is attained by any 1-Lipschitz function h such that h(cid:48)(z) = 1 when ∆(z) ≥ 0, h(cid:48)(z) = −1 otherwise. By the above consideration, such a 1-Lipschitz function (at least one) always exists. In a quantum context, one can view the cumulative distributions ci = (cid:82) z∞ dµi(ξ) as the probability to find the particle on the half-line (−∞, z] before the transport (i = 1) or after the transport (i = 2). ∆(z) in (3.6) measures the probability flow across z, and the Wasserstein distance is the integral on R of the modulus of this probability flow. On Rm with m > 1 there is no such an explicit result. If ψ1, ψ2 are bounded Lipschitz functions with compact support, we know from [19] that there exist two (a.e. unique) bounded measurable Lipschitz functions a, u : Rm → R such that a ≥ 0, (cid:4) −∇(a∇u) = ψ1 − ψ2 in the weak sense, and ∇u = 1 almost everywhere on the set where a > 0. We have then Rm (cid:90) (cid:90) (cid:90) dD(ϕ1, ϕ2) = a(x)dx. Indeed integrating by parts, we can write ϕ1(f ) − ϕ2(f ) = (∇f ) · (a∇u)dx Rm and since a∇u = a we have dD(ϕ1, ϕ2) ≤ a(x)dx . Rm The sup is attained by the function f = u. 4 Noncommutative examples At the light of Proposition 2.1 one may wonder if the analogy between the spectral and the Wasserstein distances still makes sense in a non-commutative framework. In other terms, for A noncommutative is the distance on S(A) computed by (1.1) related to some Wasserstein dis- tance? The most obvious answer, based on Gelfand's identification (1.4) between points and pure states, would be to consider the Wasserstein distance WD on the metric space (P(A), dD), and 16 F. D'Andrea and P. Martinetti question whether WD on the set Prob(P(A)) of probabilities distributions on P(A) coincides with dD on S(A). This is obviously true for pure states since by (2.15) WD(ω1, ω2) = c(ω1, ω2) = dD(ω1, ω2) ∀ ω1, ω2 ∈ P(A). For non-pure states however this is usually not true. The reason is that even if S(A) ⊂ Prob(P(A)) for commutative and almost commutative C∗-algebras (see the definition in the next paragraph), there is not a 1-to-1 correspondence between the two sets (except in the commutative case). For instance, as recalled in Section 4.1, S(M2(C)) is a non-trivial quotient of Prob(P(M2(C))), mea- ning that two probability distributions φ1 (cid:54)= φ2 may give the same state ϕ1 = ϕ2. Thus, given a spectral triple on M2(C) as the one studied in [7], one has dD(ϕ1, ϕ2) = 0 while WD(φ1, φ2) (cid:54)= 0 since the cost being a distance (and assuming P(A) is a polish space), W is a distance and vanishes if and only if φ1 = φ2. However this does not mean that the optimal transport interpretation of the spectral dis- tance loses all interest in the noncommutative framework. As explained in Section 4.2, there are interesting analogies between the spectral distance and some Wasserstein distances other than WD in almost-commutative geometries. The latest are spectral triples (A,H, D) obtained as the product of a spin manifold by a finite-dimensional spectral triple (AI ,HI , DI ), namely A = C∞ 0 (M) ⊗ AI , H = L2(M,S) ⊗ HI , D = −iγµ∂µ ⊗ II + γm+1 ⊗ DI , (4.1) where II is the identity of HI and γm+1 is the product of the gamma matrices in case m is even, 0 (M) being or the identity in case m is odd. AI being finite-dimensional (or, equivalently, C∞ Abelian) one has [27]: P(A) = P(C∞ (4.2) For simplicity we discuss here the case AI = C2 acting on HI = C2, while in Section 4.2 we will focus on the finite-dimensional spectral triple describing the internal degrees of freedom (hence the subscript I) of the standard model of particle physics. Since C2 has two pure states -- δ0(z0, z1) = z0, δ1(z0, z1) = z1 ∀ (z0, z1) ∈ C2 -- from (4.2) 0 (M)) × P(AI ). P(A) = M × {0, 1} This is Connes' idea of "product of the continuum by the discrete" [11] seen at the level of pure states: through the product by a finite-dimensional spectral triple, the points of the manifold acquire a Z2 internal discrete structure, namely to a point x = δx in the commutative case corresponds two pure states in P(A), x0 = (δx, δ0), x1 = (δx, δ1). Moreover, although P(A) is the disjoint union of two copies of M, points on distinct copies need not to be at infinite spectral distance from one another. Indeed one shows that [35] dD(x0, x1) = dDI (δ0, δ1), (4.3) where dDI denotes the spectral distance associated to (AI ,HI , DI ). AI is represented on HI by diagonal matrices but DI ∈ M2(C) needs not to be diagonal. Especially, if DI has non-zero off-diagonal terms then [D, a] ≤ 1 is a non-trivial constraint which guarantees that (4.3) is finite. Note that if one takes the direct sum of spectral triples instead of a product, namely 0 (M) acting on L2(M,S)⊕ L2(M,S), one gets the same pure states as above but 0 (M)⊕ C∞ C∞ A View on Optimal Transport from Noncommutative Geometry 17 D = −iγµ∂µ⊕−iγµ∂µ does not have off-diagonal terms and the two copies of M are at infinite distance from one another (points have not enough space to "talk to each other" through the off-diagonal terms of D). From an optimal transport point of view, the non-vanishing of (4.3) can be interpreted as the fact that "staying at a point", which is costless in the commutative case since c(x, x) = dD(x, x) = 0, may have a cost dD(x0, x1) (cid:54)= 0 (4.4) in almost-commutative geometries, corresponding to the "internal jump" from x0 to x1. We investigate this idea in Section 4.2, showing in Proposition 4.2 that the spectral distance between pure states is the minimal work WI associated to a cost cI . 4.1 The Moyal plane The Moyal algebra Aθ is the noncommutative deformation of the non-unital Schwartz algebra S(R2) with point-wise product 1 d2yd2z f (x + y)g(x + z)e−i2yΘ−1z ∀ f, g ∈ S(R2), (cid:90) (f (cid:63) g)(x) = (πθ)2 where yΘ−1z ≡ yµΘ−1 µν zν and (cid:19) (cid:18) 0 1 −1 0 Θµν = θ with θ ∈ R, θ (cid:54)= 0. In [7] we studied the spectral distance associated to the spectral triple built in [20] around the action of Aθ on L2(R2) and the usual Dirac operator on R2. We found that the topology induced by the spectral distance on S(Aθ) is not the weak*-topology (a condition required by Rieffel [39, 40, 41] in the unital case in order to define compact quantum metric spaces, and adapted by [30] to the non-unital case). However by viewing Aθ as an algebra of infinite-dimensional matrices, we proposed some finite-dimensional truncations of the Moyal spectral triple, based on the algebra Mn(C), n ∈ N, that makes it a quantum metric space. Explicitly, for n = 2 P(M2(C)) is homeomorphic to the Euclidean 2-sphere, ξ ∈ P(M2(C)) −→ xξ = (xξ, yξ, zξ) ∈ S2, so that a non-pure state ϕ is determined by a probability distribution φ on S2. Its evaluation ϕ(a) = φ(xξ) Tr(sξa) dxξ = Tr φ(xξ)sξ dxξ a ∀ a ∈ M2(C), (cid:19) (cid:19) (cid:90) S2 (cid:18)(cid:18)(cid:90) S2 (cid:90) S2 with sξ ∈ M2(C) the support of the pure state µ−1(xξ) and dxξ the SU (2) invariant measure on S2, only depends on the barycenter of φ, ¯xφ = (¯xφ, ¯yφ, ¯yφ) with ¯xφ := φ(xξ)xξdxξ and similar notation for ¯yφ, ¯zφ. With the equivalence relation φ ∼ φ(cid:48) ⇐⇒ ¯xφ = ¯xφ(cid:48) one gets that S(M2(C)) = S(C(S2))/ ∼ is homeomorphic to the Euclidean 2-ball: µ−→ ¯xφ ∈ B2. ϕ 18 F. D'Andrea and P. Martinetti Figure 2. The vertical plane containing ¯xφ, ¯xφ(cid:48). In [7] we computed dD(¯xφ, ¯xφ(cid:48)) = (cid:114) × θ 2 cos α dEc(xφ, xφ(cid:48)) 1 2 sin α when α ≤ π 4 dEc(xφ, xφ(cid:48)) when α ≥ π 4 , , where dEc(¯xφ, ¯xφ(cid:48)) = ¯xφ− ¯xφ(cid:48) is the Euclidean distance and α is the angle between the segment [¯xφ, ¯xφ(cid:48)] and the horizontal plane zξ = const (see Fig. 2). The state space of the truncated Moyal algebra is not the space of distributions on S2, but a quotient of it. Therefore for φ ∼ φ(cid:48), φ (cid:54)= φ(cid:48), one has dD(ϕ, ϕ(cid:48)) = 0 while WD(ϕ, ϕ(cid:48)) (cid:54)= 0. In this example there seems to be no Wasserstein distance naturally associated to the spectral distance. In the next section, we exhibit another noncommutative example for which there exists a Wasserstein distance W (cid:48), with cost d(cid:48) defined on a set bigger than P(A), that coincides with dD for some non-pure states. 4.2 The standard model The spectral triple describing the standard model of elementary particles is the product (4.1) of a manifold with the finite-dimensional algebra C⊕H⊕M3(C) acting on HI = C96. DI is a 96×96 matrix with entries the masses of the elementary fermions, the Cabibbo -- Kobayashi -- Maskawa matrix and the neutrino mixing-angles. The choice of the algebra is dictated by physics [8] (its unitaries group gives back the gauge group of the standard model) and 96 is the number of elementary fermions4. The pure states of M3(C) turn out to be at infinite distance from one another and from any other pure state [35] so that, from the metric point of view, the interesting part is the product (A,H, D) of a manifold by (AI = C ⊕ H,HI , DI ). By (4.2) the pure-states . space is P(A) = M × {0, 1} since both C and the algebra of quaternion H have one single pure state5 (cid:18) α (cid:19) δC(z) = (cid:60)(z) ∀ z ∈ C, δH(h) = Tr h ∀ h = 1 2 β − ¯β ¯α ∈ H, α, β ∈ C. 4(6 leptons + 6 quarks × 3 colors) × 2 chiralities = 48, to which are added 48 antiparticles. 5H is a real, but not a complex algebra. Real C∗-algebras are defined similarly to complex ones [22], except that one imposes that 1 + a∗a is invertible -- which comes as a consequence in the complex case. A state ϕ is then defined as a real, real-linear, positive form such that ϕ(I) = 1 and ϕ(a∗) = ϕ(a). Hence H has only one state, 0 (M) as real C∗-algebras, their states are obtained by taking the real given by half the trace. Viewing C and C∞ part of their "usual" complex states. See [31, page 42]. A View on Optimal Transport from Noncommutative Geometry 19 The space underlying the standard model is thus the disjoint union of two copies of the manifold, and with some computation one shows [35] that the spectral distance is finite on P(A) and coincides with the geodesic distance d(cid:48) on the manifold M(cid:48) = M×I, with I := [0, 1] the closed unit interval, with metric (cid:18) gµν(x) (cid:19) 0 , DI2 0 (4.5) where g is the Riemannian metric on M. Assuming that M is complete, and writing t ∈ I the extra-coordinate, this can be restated as: the spectral distance between pure states in the standard model coincides with the Wasserstein distance W (cid:48) on the metric space (M(cid:48), d(cid:48)) restricted to the two hyperplanes t = 0, t = 1. This reformulation of the main result of [35] allows to determine the spectral distance between a certain class of non-pure states, namely those localized on one of the copies of M. Explicitly, S(A) is the set of couples of measures ϕ = (µ, ν) on M, normalized to (cid:90) (cid:90) M dµ + dν = 1, whose evaluation on A (cid:51) a = f ⊕ M (cid:18) g where f , g, b are in C∞ , b −¯b ¯g 0 (M), is given by (cid:60)(g) dν. (cid:60)(f ) dµ + (cid:90) (cid:19) (cid:90) M ϕ(a) = M Two states ϕ1, ϕ2 are localized on the same copy of M if ϕ1 = (0, ν1), ϕ2 = (0, ν2); or ϕ1 = (µ1, 0), ϕ2 = (µ2, 0). Such states can be viewed as elements of S(A), S(C∞ 0 (M(cid:48))) and S(C∞ Proposition 4.1. For two states ϕ1, ϕ2 localized on the same copy of M, 0 (M)). dD(ϕ1, ϕ2) = W (ϕ1, ϕ2) = W (cid:48)(ϕ1, ϕ2). (4.6) Proof . To fix notation we assume that ϕ1 = (0, ν1), ϕ2 = (0, ν2) are localized on the δH copy of M, that we write MH and associate to the value t = 1 of the extra-parameter t ∈ I. The evaluation of ϕ1 − ϕ2 on a ∈ A only depends on Re(g). Moreover for a = a∗, [D, a] ≤ 1 implies that g is 1-Lipschitz [35, equation (16)], so that dD(ϕ1, ϕ2) ≤ W (ϕ1, ϕ2). (4.7) The equality is attained by considering a = gw ⊗ I3 where gw is the real 1-Lipschitz function that attains the supremum in the computation of W . Hence the first equation in (4.6). To show that W (ϕ1, ϕ2) = W (cid:48)(ϕ1, ϕ2), one first notices that (4.5) being block-diagonal = g(cid:48)(y, 1) on MH of any implies that d(cid:48)((x, 1), (y, 1)) = d(x, y). Hence the restriction g(y) . 1-Lipschitz function g(cid:48) in C∞ 0 (M(cid:48)) is 1-Lipschitz, g(cid:48)(x, t0) − g(cid:48)(y, t1) ≤ d(cid:48)((x, t0), (y, t1)) ⇒ g(x) − g(y) ≤ d(x, y). Conversely to any 1-Lipschitz function g ∈ C∞ 0 (M) one associates g(cid:48)(x, t) . = g(x) which is 1-Lipschitz in C∞ sup g(cid:48)∈C∞ 0 (M(cid:48)),g(cid:48)Lip=1 ∀ x ∈ M, t ∈ I 0 (M(cid:48)) and takes the same value as g on MH. Therefore ϕ1(g(cid:48)) − ϕ2(g(cid:48)) = ϕ1(g) − ϕ2(g) sup g∈C∞ 0 (M),gLip=1 which, together with (4.7), yields the result. (cid:4) 20 F. D'Andrea and P. Martinetti It is tempting to postulate that dD(ϕ1, ϕ2) = W (cid:48)(ϕ1, ϕ2) for states localized on different copies, or states that are not localized on any copy (i.e. ϕ = (µ, ν) with µ, ν both non-zero). This point is under investigation. A disturbing point in the computation of the spectral distance in the standard model is the appearance of a compact extra-dimension I while the internal structure is discrete. In [35] I came out more as a computational artifact than a requirement of the model. From the Wasserstein distance point of view, the introduction of the extra-dimension can be seen in the following proper inclusions: P(A) ⊂ M(cid:48) ⊂ S(A). Namely W (cid:48) is associated to the metric space (M(cid:48), d(cid:48)) which is bigger than P(A) (the points between the sheets are not pure states of A) and smaller than S(A) (non-pure states localized on a copy are not in M(cid:48)). To get rid of the extra-dimension, one could consider the metric space (P(A), dD) with associated Wasserstein distance WD, but this would be of poor interest since the definition of dD = d(cid:48) requires the knowledge of M(cid:48). Alternatively one could look for a cost defined solely on M. For states (pure or not) localized on the same copy, this cost is simply the geodesic distance on M, as shown in Proposition 4.1. For pure states on distinct copies such a cost cI also exists and is given by (cid:115) cI (x, y) . = d(x, y)2 + 1 DI2 . (4.8) Note that cI is not a distance since it does not vanish on the diagonal, cI (x, x) = 1 DI ∀ x ∈ M, but gives back the jump-cost (4.4) (by (4.5), d(x0, x1) = 1DI ). However cI satisfies the condi- tion required to proved Kantorovich's duality (namely [43] it is lower semicontinuous and satisfies cI (x, y) ≥ a(x) + b(y) for some real-valued upper semicontinous µi-integrable functions). Hence it makes sense to consider the minimal work WI associated to cI , whose formula is given by (2.5) with the Lipschitz condition replaced by f (x) − f (y) ≤ cI (x, y). Proposition 4.2. The spectral distance between pure states x0 distinct copies is . = (δx, δC), y1 . = (δy, δH) on dD(x0, y1) = WI (x, y). Proof . For pure states the minimal work is the cost itself: WI (x, y) = cI (x, y) for any x, y ∈ M. Since dD(x0, y1) = d(cid:48)(x0, y1) as recalled above (4.5), the result follows if one proves that d(cid:48)2(x0, y1) = d2(x, y) + 1 DI2 . (4.9) This has been shown in [35] but we briefly restate the argument here for sake of completeness. Let us write xa = (xµ ∈ M, t ∈ [0, 1]) a point of M(cid:48) and gtt = g−1 tt = DI2 the extra-metric component. The Christoffel symbols involving t are tµ = Γt Γt µt = 1 2 gt∂µgt, t = − 1 Γµ 2 gµν∂νgtt, The geodesic equation xa + Γa bc xb xc writes xt + gtt(∂µgtt) xt xµ = 0, 0ν = Γµ Γµ νt = Γt tt = Γt µν = 0. (4.10a) 21 (4.10b) (4.11a) A View on Optimal Transport from Noncommutative Geometry xµ − 1 2 gµν(∂νgtt)( t)2 + Γµ λρ xλ xρ = 0 and, because gtt is a constant, they simplify to t = const = K, xµ + Γµ λρ xλ xρ = 0. (4.11b) The first geodesic equation indicates that the proper length τ of a geodesic G(cid:48) = x(τ ) in M(cid:48) between x0 and y1 is proportional to t, dt = Kdτ, as well as to the line element ds = 1 = x = gab xa xb = gµν dxµ dτ √ gννdxµdxν of M since dxν dτ + gttK2 = ds2 dτ 2 + gttK2, (4.12) so that, assuming gttK2 (cid:54)= 1 (which from (4.12) amounts to take x (cid:54)= y), dτ = . ds(cid:112)1 − gttK2 (cid:90) G(cid:48) The second geodesic equation (4.11b) shows that the projection on M of G(cid:48) is a geodesic G of M. Therefore d(cid:48)(x0, y1) = d(x, y), dτ = (cid:90) (cid:90) ds = 1(cid:112)1 − gttK2 ds = G 1(cid:112)1 − gttK2 dτ ds G that is to say d(cid:48)2(x0, y1) = d2(x, y) + gttK2d(cid:48)2(x0, y1). Writing K2d(cid:48)2(x0, y1) as (cid:19)2 (cid:18)(cid:90) (cid:18)(cid:90) Kdτ = G(cid:48) G(cid:48) (cid:19)2 dt = (t((y, 1) − t((x, 0)))2 = 1 one obtains (4.9), and the result. (cid:4) Let us underline an interesting feature of this proposition, namely that a cost which is not a distance M can be seen as a distance on M × M. Proposition 4.2 can also be rewritten using a cost vanishing on the diagonal, namely (cid:115) c(cid:48) I (x, y) = cI (x, y) − 1 . DI2 = d(x, y)2 + 1 DI2 − 1 DI . Writing W (cid:48) I the associated minimal work, one has dD(x0, y1) = WI (x, y) + 1 DI2 . Quite remarkably, the cost (4.8) is similar to the cost (32) introduced in [6] in the framework of the relativistic heat equation. Proposition 4.2 relies on the fact that the jump-cost (4.4) is constant. From a physics point of view, this means that one does not take into account the Higgs field. In almost commutative 22 F. D'Andrea and P. Martinetti geometries, the latest is obtained by inner fluctuation of the metric [13] that substitute D with a covariant Dirac operator. From the metric point of view this amounts to replacing (see [35] for details) (4.5) by (cid:19) (cid:18) gµν(x) H(x) =(cid:0)1 + h1(x)2 + h2(x)2(cid:1) m2 DI + H(x)2 0 where 0 , t where (h1, h2) is the (complex) Higgs doublet and mt is the mass of the quark top. Instead of (4.3) one has dD(x0, x1) = dDI +H(x)(δC, δH) (the jump-cost is no longer constant). In analogy with (4.8), one could define the cost (cid:113) cI (x, y) = d(x, y)2 + dDI +H(x)(δC, δH)2 which allows to avoid the introduction of the extra-dimension. However the projection of a geodesic in M(cid:48) is no longer a geodesic of M ((4.10b) does not simplify to (4.11b)) so there is no way to express d(cid:48)(x0, y1) as a function of d(x, y), and dD(x0, y1) = d(cid:48)(x0, y1) no longer equals WI (x0, y1). 5 Conclusion The spectral distance between states on a complete Riemannian spin manifold coincides with the Wasserstein distance of order 1. In the noncommutative case the analogies between the spectral distance and various Wasserstein distances, although still not fully understood, shed an interesting light on the interpretation of the distance formula in noncommutative geometry. In physics, defining the distance as a supremum rather than an infimum is useful since, at small scale, quantum mechanics indicates that notions as "paths between points" no longer make sense, so that the classical definition of distance as the length of the shortest path loses any operational meaning. An interesting feature of noncommutative geometry is to provide a notion of distance overcoming this difficulty. In transport theory the interpretation of Kantorovich formula has an interpretation in economics rather than in quantum mechanics: while Monge formulation corresponds to the minimization of a cost, Kantorovich dual formula corresponds to the maximization of a profit. To repeat a classical example found in the literature: assume that the distribution of flour-producers on a given territory M is given by µ1 and the distribution of bakeries by µ2. Consider a transport-consortium whose job consists in buying the flour at factories and selling it to bakers. The consortium fixes the value f (x) of the flour at the point x (it buys the flour at the price f (x) if there is a factory, or it sells it at a price f (x) if there is a bakery). The Wasserstein distance is the maximum profit the consortium may hope, under the constraint of staying competitive, that means not selling the flour to a price higher than the bakeries would pay if they were doing the transport by themselves (i.e. f (x) ≤ f (y) + c(x, y) in a quantum context, what is the physical for all x, y). This raises an interesting question: meaning of this "profit" that one is maximizing while computing the distance? If one view the states on M as wave functions, is the distance related to the minimum work required to transform one wave into the second? More specifically, for M = Rm what physical quantity represents (3.4), and what is the meaning of the point α? Acknowledgments We would like to thank Hanfeng Li and anonymous referees for their valuable comments. A View on Optimal Transport from Noncommutative Geometry 23 References [1] Ambrosio L., Lecture notes on optimal transport problems, in Mathematical Aspects of Evolving Interfaces (Funchal, 2000), Lecture Notes in Math., Vol. 1812, Springer, Berlin, 2003, 1 -- 52. [2] Azagra D., Ferrera J., L´opez-Mesas F., Rangel Y., Smooth approximation of Lipschitz functions on Rie- mannian manifolds, J. Math. Anal. Appl. 326 (2007), 1370 -- 1378, math.DG/0602051. [3] Bratteli O., Robinson D.W., Operator algebras and quantum statistical mechanics. 1. C∗- and W ∗-algebras, symmetry groups, decomposition of states, 2nd ed., Texts and Monographs in Physics, Springer-Verlag, New York, 1987. [4] Biane P., Voiculescu D., A free probability analogue of the Wasserstein metric on the trace-state space, Geom. Funct. Anal. 11 (2001), 1125 -- 1138, math.OA/0006044. [5] Bimonte G., Lizzi F., Sparano G., Distances on a lattice from non-commutative geometry, Phys. Lett. B 341 (1994), 139 -- 146, hep-lat/9404007. [6] Brenier Y., Extended Monge -- Kantorovich theory, in Optimal Transportation and Applications (Martina Franca, 2001), Lecture Notes in Math., Vol. 1813, Springer, Berlin, 2003, 91 -- 121. [7] Cagnache E., D'Andrea F., Martinetti P., Wallet J.C., The Spectral distance in the Moyal plane, arXiv:0912.0906. [8] Chamseddine A.H., Connes A., Marcolli M., Gravity and the standard model with neutrino mixing, Adv. Theor. Math. Phys. 11 (2007), 991 -- 1089, hep-th/0610241. [9] Chavel I., Riemannian geometry -- a modern introduction, Cambridge Tracts in Mathematics, Vol. 108, Cambridge University Press, Cambridge, 1993. [10] Connes A., Compact metric spaces, Fredholm modules, and hyperfiniteness, Ergodic Theory Dynam. Systems 9 (1989), 207 -- 220. [11] Connes A., Noncommutative geometry, Academic Press, Inc., San Diego, CA, 1994. [12] Connes A., Noncommutative geometry and reality, J. Math. Phys. 36 (1995), 6194 -- 6231. [13] Connes A., Gravity coupled with matter and the foundation of non-commutative geometry, Comm. Math. Phys. 182 (1996), 155 -- 176. [14] Connes A., A unitary invariant in Riemannian geometry, Int. J. Geom. Methods Mod. Phys. 5 (2008), 1215 -- 1242, arXiv:0810.2091. [15] Connes A., Lott J., The metric aspect of noncommutative geometry, in New symmetry principles in quantum field theory (Carg`ese, 1991), NATO Adv. Sci. Inst. Ser. B Phys., Vol. 295, Plenum, New York, 1992, 53 -- 93. [16] Connes A., Marcolli M., Noncommutative geometry, quantum fields and motives, American Mathematical Society Colloquium Publications, Vol. 55, American Mathematical Society, Providence, RI; Hindustan Book Agency, New Delhi, 2008. [17] Doplicher S., Fredenhagen K., Roberts J.E., Spacetime quantization induced by classical gravity, Phys. Lett. B 331 (1994), 39 -- 44. [18] Doplicher S., Fredenhagen K., Roberts J.E., The quantum structure of spacetime at the Planck scale and quantum fields, Comm. Math. Phys. 172 (1995), 187 -- 220, hep-th/0303037. [19] Evans L.C., Gangbo W., Differential equations methods for the Monge -- Kantorevich mass transfer problem, Mem. Amer. Math. Soc. 137 (1999), no. 653. [20] Gayral V., Gracia-Bond´ıa J.M., Iochum B., Schucker T., Varilly J.C., Moyal planes are spectral triples, Comm. Math. Phys. 246 (2004), 569 -- 623, hep-th/0307241. [21] Givens C.R., Shortt R.M., A class of Wasserstein metrics for probability distributions, Michigan Math. J. 31 (1984), 231 -- 240. [22] Goodearl K.R., Notes on real and complex C∗-algebras, Shiva Mathematics Series, Vol. 5, Shiva Publishing Ltd., Nantwich, 1982. [23] Gracia-Bond´ıa J.M., Varilly J.C., Figueroa H., Elements of noncommutative geometry, Birkhauser Advanced Texts: Basler Lehrbucher, Birkhauser Boston, Inc., Boston, MA, 2001. [24] Greene R.E., Wu H., C∞ approximations of convex, subharmonic, and plurisubharmonic functions, Ann. Sci. ´Ecole Norm. Sup. (4) 12 (1979), 47 -- 84. [25] Gromov M., Metric structures for Riemannian and non-Riemannian spaces, Progress in Mathematics, Vol. 152, Birkhauser Boston, Inc., Boston, MA, 1999. 24 F. D'Andrea and P. Martinetti [26] Iochum B., Krajewski T., Martinetti P., Distances in finite spaces from noncommutative geometry, J. Geom. Phys. 37 (2001), 100 -- 125, hep-th/9912217. [27] Kadison R.V., Ringrose J.R., Fundamentals of the theory of operator algebras. Vol. II. Advanced theory, Pure and Applied Mathematics, Vol. 100, Academic Press, Inc., Orlando, FL, 1986. [28] Kantorovich L.V., On the transfer of masses, Dokl. Akad. Nauk. SSSR 37 (1942), 227 -- 229. [29] Kantorovich L.V., Rubinstein G.S., On a space of totally additive functions, Vestnik Leningrad. Univ. 13 (1958), no. 7, 52 -- 59. [30] Latr´emoli`ere F., Bounded-Lipschitz distances on the state space of a C∗-algebra, Taiwanese J. Math. 11 (2007), 447 -- 469. [31] Martinetti P., Distances en g´eom´etrie non-commutative, PhD Thesis, math-ph/0112038. [32] Martinetti P., Carnot -- Carath´eodory metric and gauge fluctuations in noncommutative geometry, Comm. Math. Phys. 265 (2006), 585 -- 616, hep-th/0506147. [33] Martinetti P., Spectral distance on the circle, J. Funct. Anal. 255 (2008), 1575 -- 1612, math.OA/0703586. [34] Martinetti P., Smoother than a circle or how noncommutative geometry provides the torus with an ego- centred metric, in Modern Trends in Geometry and Topology (Deva, 2005), Cluj Univ. Press, Cluj-Napoca, 2006, 283 -- 293, hep-th/0603051. [35] Martinetti P., Wulkenhaar R., Discrete Kaluza -- Klein from scalar fluctuations in noncommutative geometry, J. Math. Phys. 43 (2002), 182 -- 204, hep-th/0104108. [36] Monge G., M´emoire sur la Th´eorie des D´eblais et des Remblais, Histoire de l'Acad. des Sciences de Paris, 1781. [37] Nash J., C 1-isometric imbeddings, Ann. of Math. (2) 60 (1954), 383 -- 396. [38] Nash J., The imbedding problem for Riemannian manifolds, Ann. of Math. (2) 63 (1956), 20 -- 63. [39] Rieffel M.A., Metric on state spaces, Doc. Math. 4 (1999), 559 -- 600, math.OA/9906151. [40] Rieffel M.A., Compact quantum metric spaces, in Operator Algebras, Quantization, and Noncommutative Geometry, Contemp. Math., Vol. 365, Amer. Math. Soc., Providence, RI, 2004, 315 -- 330, math.OA/0308207. [41] Rieffel M.A., Gromov -- Hausdorff distance for quantum metric spaces, Mem. Amer. Math. Soc. 168 (2004), no. 796, 1 -- 65, math.OA/0011063. [42] Roe J., Index theory, coarse geometry, and topology of manifolds, CBMS Regional Conference Series in Mathematics, Vol. 90, American Mathematical Society, Providence, RI, 1996. [43] Villani C., Topics in optimal transportation, Graduate Studies in Mathematics, Vol. 58, American Mathe- matical Society, Providence, RI, 2003. [44] Villani C., Optimal transport. Old and new, Grundlehren der Mathematischen Wissenschaften, Vol. 338, Springer-Verlag, Berlin, 2009.
1510.01826	1	1510	2015-10-07T05:27:58	The strong Morita equivalence for coactions of a finite dimensional $C^$-Hopf algebra on unital $C^$-algebras	[ "math.OA" ]	Following Jansen and Waldmann, and Kajiwara and Watatani, we shall introduce notions of coactions of a finite dimensional $C^$-Hopf algebra on a Hilbert $C^$-bimodule of finite type in the sense of Kajiwara and Watatani and define their crossed product. We shall investigate their basic properties and show that the strong Morita equivalence for coactions preserves the Rohlin property for coactions of a finite dimensional $C^$-Hopf algebra on unital $C^$-algebras.	math.OA	math	THE STRONG MORITA EQUIVALENCE FOR COACTIONS OF A FINITE DIMENSIONAL C ∗-HOPF ALGEBRA ON UNITAL C ∗-ALGEBRAS 5 1 0 2 t c O 7 ] KAZUNORI KODAKA AND TAMOTSU TERUYA Abstract. Following Jansen and Waldmann, and Kajiwara and Wata- tani, we shall introduce notions of coactions of a finite dimensional C ∗- Hopf algebra on a Hilbert C ∗-bimodule of finite type in the sense of Kajiwara and Watatani and define their crossed product. We shall in- vestigate their basic properties and show that the strong Morita equiva- lence for coactions preserves the Rohlin property for coactions of a finite dimensional C ∗-Hopf algebra on unital C ∗-algebras. . A O h t a m [ 1 v 6 2 8 1 0 . 0 1 5 1 : v i X r a 1. Introduction Let A and B be unital C ∗-algebras and X a Hilbert A − B-bimodule of finite type in the sense of Kajiwara and Watatani [8]. Let H be a finite dimensional C ∗-Hopf algebra with its dual C ∗-Hopf algebra H 0. In this paper, following Jansen and Waldmann [7], we shall introduce the notion of coactions of H 0 on X and define their crossed product. That is, for coactions ρ and σ of H 0 on A and B, respectively, we introduce a linear map λ from X to X ⊗ H 0, which is compatible with the coactions ρ, σ and the left A- module action, the right B-module action and the left A-valued and right B-valued inner products. Then we can define the crossed product X ⋊λ H, which is a Hilbert A ⋊ρ H − B ⋊σ H-bimodule of finite type. Furthermore, we shall give a duality theorem similar to the ordinary one. This theorem in the case of countably discrete group actions and Kac systems are found in Kajiwara and Watatani [9] and Guo and Zhang [5], respectively. The latter result is almost a generalization of our duality theorem. But our approach to coactions of a finite dimensional C ∗-Hopf algebra on a unital C ∗-algebra is a useful addition to our paper, especially the main result on preservation of the Rohlin property under the strongly Morita equivalence. So, in Section 5, we give a duality theorem, a version of crossed product duality for coactions of finite dimensional C ∗-Hopf algebras on Hilbert C ∗-bimodules of finite type. Also, we see that if X is an A − B-equivalence bimodule, we can show that 2010 Mathematics Subject Classification. Primary 46L05; Secondary 46L08. Key words and phrases. C ∗-algebras, C ∗-Hopf algebras, Hilbert C ∗-bimodules, the Rohlin property, strong Morita equivalence. 1 2 KAZUNORI KODAKA AND TAMOTSU TERUYA X ⋊λ H is an A ⋊ρ H − B ⋊σ H-equivalence bimodule. Hence A ⋊ρ H is strongly Morita equivalent to B ⋊σ H. Finally, if X is an A − B-equivalence bimodule and ρ has the Rohlin property, then σ has also the Rohlin property. As an application of the result, we can obtain the following: Under a certain condition, if a unital C ∗-algebra A has a finite dimensional C ∗-Hopf algebra coaction of H 0 with the Rohlin property, then any unital C ∗-algebra that is strongly Morita equivalent to A has also a finite dimensional C ∗-Hopf algebra coaction of H 0 with the Rohlin property. In [13, Section 4], we gave an incorrect example of an action of a finite dimensional C ∗-Hopf algebra on a unital C ∗-algebra with the Rohlin property. But applying the above result to [11, Section 7], we can give several examples of them. For an algebra A, we denote by 1A and idA the unit element in A and the identity map on A, respectively. If no confusion arises, we denote them by 1 and id, respectively. For each n ∈ N, we denote by Mn(C) the n × n-matrix algebra over C and In denotes the unit element in Mn(C). For projections p, q in a C ∗-algebra A, we write p ∼ q in A if p and q are Murray-von Neumann equivalent in A. 2. Preliminaries Let H be a finite dimensional C ∗-Hopf algebra. We denote its comul- tiplication, counit and antipode by ∆, ǫ and S, respectively. We shall use Sweedler's notation ∆(h) = h(1) ⊗ h(2) for any h ∈ H which suppresses a possible summation when we write comultiplications. We denote by N the dimension of H. Let H 0 be the dual C ∗-Hopf algebra of H. We denote its co- multiplication, counit and antipode by ∆0, ǫ0 and S0, respectively. There is the distinguished projection e in H. We note that e is the Haar trace on H 0. Also, there is the distinguished projection τ in H 0 which is the Haar trace on H. Since H is finite dimensional, H ∼= ⊕L k=1Mdk (C) as C ∗-algebras. Let { vk ij k = 1, 2, . . . , L, i, j = 1, 2, . . . , fk } be a system of matrix units of H. Let { wk ij k = 1, 2, . . . , K, i, j = 1, 2, . . . , dk} be a basis of H satisfying Szyma´nski and Peligrad's [17, Theorem 2.2,2], which is called a system of comatrix units of H, that is, the dual basis of a system of matrix units of H 0. Also let { φk ij k = 1, 2, . . . , K, i, j = 1, 2, . . . , dk} and { ωk ij k = 1, 2, . . . , L, i, j = 1, 2, . . . , fk} be systems of matrix units and comatrix units of H 0, respectively. k=1Mfk(C) and H 0 ∼= ⊕K Let A and B be unital C ∗-algebras and X a Hilbert A − B-bimodule of finite type in the sense of [8]. We regard a C ∗-Hopf algebra H 0 as an H 0 − H 0-equivalence bimodule in the usual way. nXi=1 nXi=1 THE STRONG MORITA EQUIVALENCE FOR COACTIONS 3 Let X ⊗ H 0 be an exterior tensor product of Hilbert C ∗-bimodules X and H 0, which is a Hilbert A ⊗ H 0 − B ⊗ H 0-bimodule. Lemma 2.1. With the above notations, X ⊗H 0 is a Hilbert A⊗H 0−B⊗H 0- bimodule of finite type. In particular, if X is an A−B-equivalence bimodule, X ⊗ H 0 is an A ⊗ H 0 − B ⊗ H 0- equivalence bimodule. Proof. Since X is of finite type, there is a right B-basis {ui}n for any x ∈ X, φ ∈ H 0, i=1 of X. Then (ui ⊗ 10)hui ⊗ 10 , x ⊗ φiB⊗H 0 = (ui ⊗ 10)(hui , xiB ⊗ φ) = x ⊗ φ. Thus a family {ui ⊗ 10}n i=1 is a right B ⊗ H 0-basis of X ⊗ H 0. In the same way as above, we can see that there is a left A ⊗ H 0-basis of X ⊗ H 0. Hence by [8, Proposition 1.12] or [9, Lemma 1.3], X ⊗ H 0 is a Hilbert A ⊗ H 0 − B ⊗ H 0-bimodule of finite type. Furthermore, we suppose that X is an A − B-equivalence bimodule. Since X is full with the both-sided inner products, by the definitions of the left and right inner products of X ⊗ H 0, so is X ⊗ H 0. Moreover, the associativity condition of the left and right inner products of X ⊗ H 0 holds since the associativity condition of the left and right inner products of X holds. Hence X ⊗ H 0 is an A ⊗ H 0 − B ⊗ H 0- equivalence bimodule. (cid:3) Let Hom(H, X) be the vector space of all linear maps from H to X. Then X ⊗ H 0 is isomorphic to Hom(H, X) as vector spaces. Sometimes, we identify X ⊗ H 0 with Hom(H, X). 3. Coactions of a finite dimensional C ∗-Hopf algebra on a Hilbert C ∗-bimodule of finite type and strong Morita equivalence Let A and B be unital C ∗ algebras and X a Hilbert A − B-bimodule of finite type. Let H be a finite dimensional C ∗-Hopf algebra with its dual C ∗-Hopf algebra H 0. Let ρ be a weak coaction of H 0 on A and λ a linear map from X to X ⊗ H 0. Following [7], [9], we shall introduce the several definitions. Definition 3.1. With the above notations, we say that (A, X, ρ, λ, H 0) is a weak left covariant system if the following conditions hold: (1) λ(ax) = ρ(a)λ(x) for any a ∈ A, x ∈ X, (2) ρ(Ahx , yi) = A⊗H 0hλ(x) , λ(y)i for any x, y ∈ X, (3) (idX ⊗ ǫ0) ◦ λ = idX. We call λ a weak left coaction of H 0 on X with respect to (A, ρ). 4 KAZUNORI KODAKA AND TAMOTSU TERUYA We define the weak action of H on A induced by ρ as follows: For any a ∈ A, h ∈ H, h ·ρ a = (id ⊗ h)(ρ(a)), where we regard H as the dual space of H 0. In the same way as above, we can define the action of H on X induced by λ as follows: For any x ∈ X, h ∈ H, h ·λ x = (id ⊗ h)(λ(x)) = λ(x)b(h), where λ(x)b is the element in Hom(H, X) induced by λ(x) in X ⊗ H 0. Then we obtain the following conditions which are equivalent to Conditions (1)- (3) in Definition 3.1, respectively: (1)' h ·λ ax = [h(1) ·ρ a][h(2) ·λ x] for any a ∈ A, x ∈ X, h ∈ H, (2)' h ·ρ Ahx, yi = Ah[h(1) ·λ x], [S(h∗ (3)' 1H ·λ x = x for any x ∈ X. (2)) ·λ y]i for any x, y ∈ X, h ∈ H, Let σ be a weak coaction of H 0 on B. Definition 3.2. With the above notations, we say that (B, X, σ, λ, H 0) is a weak right covariant system if the following conditions hold: (4) λ(xb) = λ(x)σ(b) for any b ∈ B, x ∈ X, (5) σ(hx, yiB) = hλ(x), λ(y)iB⊗H 0 for any x, y ∈ X, (6) (idX ⊗ ǫ0) ◦ λ = idX. We call λ a weak right coaction of H 0 on X with respect to (B, σ). We can also define the weak action of H on X induced by λ satisfying the similar conditions to (1)'-(3)'. That is, we have the following conditions which are equivalent to Conditions (4)-(6), respectively: (4)' h ·λ xb = [h(1) ·λ x][h(2) ·σ b] for any b ∈ B, x ∈ X, h ∈ H, (5)' h ·σ hx, yiB = h[S(h∗ (6)' 1H ·λ x = x for any x ∈ X. (1)) ·λ x] , [h(2) ·λ y]iB for any x, y ∈ X, h ∈ H, Let ρ and σ be weak coactions of H 0 on A and B, respectively. Let X be a Hilbert A − B-bimodule of finite type. Definition 3.3. We say that (A, B, X, ρ, σ, λ, H 0) is a weak covariant sys- tem if the following conditions hold: (1) λ(ax) = ρ(a)λ(x) for any a ∈ A, x ∈ X, (2) λ(xb) = λ(x)σ(b) for any b ∈ B, x ∈ X, (3) ρ(Ahx , yi) = A⊗H 0hλ(x) , λ(y)i for any x, y ∈ X, (4) σ(hx, yiB) = hλ(x), λ(y)iB⊗H 0 for any x, y ∈ X, (5) (idX ⊗ ǫ0) ◦ λ = idX. We call λ a weak coaction of H 0 on X with respect to (A, B, ρ, σ). We note that the above conditions are equivalent to the following conditions, THE STRONG MORITA EQUIVALENCE FOR COACTIONS 5 respectively: (1)' h ·λ ax = [h(1) ·ρ a][h(2) ·λ x] for any a ∈ A, x ∈ X, h ∈ H, (2)' h ·λ xb = [h(1) ·λ x][h(2) ·σ b] for any b ∈ B, x ∈ X, h ∈ H, (3)' h ·ρ Ahx, yi = Ah[h(1) ·λ x], [S(h∗ (4)' h ·σ hx, yiB = h[S(h∗ (5)' 1H ·λ x = x for any x ∈ X. (2)) ·λ y]i for any x, y ∈ X, h ∈ H, (1)) ·λ x] , [h(2) ·λ y]iB for any x, y ∈ X, h ∈ H, We extend the above notions to coactions of a finite dimensional C ∗-Hopf algebra on unital C ∗-algebras. Definition 3.4. Let A, B and H, H 0 be as above. Let ρ and σ be coactions of H 0 on A and B, respectively and let X be a Hilbert A − B-bimodule of finite type. (i) We say that (A, X, ρ, λ, H 0) is a left covariant system if it is a weak left covariant system and a weak left coaction λ of H 0 on X with respect to (A, ρ) satisfies that (∗) (λ ⊗ id) ◦ λ = (id ⊗ ∆0) ◦ λ, which is equivalent to the condition that (∗)' h ·λ [l ·λ x] = hl ·λ x for any x ∈ X , h, l ∈ H. We call λ a left coaction of H 0 on X with respect to (A, ρ). (ii) We say that (B, X, σ, λ, H 0) is a right covariant system if it is a weak right covariant system and a weak right coaction λ of H 0 on X with respect to (B, σ) satisfies (∗) or (∗)'. We call λ a right coaction of H 0 on X with respect to (B, σ). (iii) We say that (A, B, X, ρ, σ, λ, H 0) is a covariant system if it is a weak covariant system and a weak coaction λ with respect to (A, B, ρ, σ) satisfies (∗) or (∗)'. We call λ a coaction of H 0 on X with respect to (A, B, ρ, σ). Furthermore, we extend the notion of the covariant system to twisted coactions of a finite dimensional C ∗-Hopf algebra on unital C ∗-algebras. We recall the definition of a twisted coaction (ρ, u) of a C ∗-Hopf algebra H 0 on a unital C ∗-algebra A (See [9], [10]). Let ρ be a weak coaction of H 0 on A and u a unitary element in A ⊗ H 0 ⊗ H 0. Then we say that (ρ, u) is a twisted coaction of H 0 on A if the following conditions hold: (1) (ρ ⊗ id) ◦ ρ = Ad(u) ◦ (id ⊗ ∆0) ◦ ρ, (2) (u ⊗ 10)(id ⊗ ∆0 ⊗ id)(u) = (ρ ⊗ id ⊗ id)(u)(id ⊗ id ⊗ ∆0)(u), (3) (id ⊗ h ⊗ ǫ0)(u) = (id ⊗ ǫ0 ⊗ h)(u) = ǫ0(h)10 for any h ∈ H. The above conditions are equivalent to the following conditions, respec- tively: (1)' h ·ρ [l ·ρ a] =bu(h(1), l(1))[h(2)l(2) ·ρ a]bu∗(h(3), l(3)), for any a ∈ A, h, l ∈ H, 6 KAZUNORI KODAKA AND TAMOTSU TERUYA h, l, m ∈ H, (2)' bu(h(1), l(1))bu(h(2)l(2), m) = [h(1) ·ρ bu(l(1), m(1))]bu(h(2), l(2)m(2)), for any (3)'bu(h, 1) =bu(1, h) = ǫ(h)10 for any h ∈ H. Definition 3.5. Let A, B and H, H 0 be as above. Let (ρ, u) and (σ, v) be twisted coactions of H 0 on A and B, respectively and let X be a Hilbert A − B-bimodule of finite type. We say that (A, B, X, ρ, u, σ, v, λ, H 0) is a twisted covariant system if it is a weak covariant system and a weak coaction λ of H 0 with respect to (A, B, ρ, σ) satisfies that (∗∗) (λ ⊗ id)(λ(x)) = u(id ⊗ ∆0)(λ(x))v∗ for any x ∈ X, which is equivalent to the condition that (∗∗)' h·λ [l ·λ x] =bu(h(1), l(1))[h(2)l(2) ·λ x]bv∗(h(3), l(3)) for any x ∈ X, h, l ∈ H, wherebu andbv are elements in Hom(H × H, A) and Hom(H × H, B) induced by u ∈ A ⊗ H 0 ⊗ H 0 and v ∈ B ⊗ H 0 ⊗ H 0, respectively. We call λ a twisted coaction of H 0 on X with respect to (A, B, ρ, u, σ, v). Next, we introduce the notion of strong Morita equivalence for coactions of a finite dimensional C ∗-Hopf algebra on unital C ∗-algebras. Definition 3.6. Let A, B and H, H 0 be as above. (i) Let ρ and σ be weak coactions of H 0 on A and B, respectively. We say that ρ is strongly Morita equivalent to σ if there are an A−B-equivalence bi- module X and a weak coaction λ of H 0 on X such that (A, B, X, ρ, σ, λ, H 0) is a weak covariant system. (ii) Let ρ and σ be coactions of H 0 on A and B, respectively. We say that ρ is strongly Morita equivalent to σ if there are an A − B-equivalence bi- module X and a coaction λ of H 0 on X such that (A, B, X, ρ, σ, λ, H 0) is a covariant system. (iii) Let (ρ, u) and (σ, v) be twisted coactions of H 0 on A and B, respec- tively. We say that (ρ, u) is strongly Morita equivalent to (σ, v) if there are an A − B-equivalence bimodule X and a twisted coaction λ of H 0 on X such that (A, B, X, ρ, u, σ, v, λ, H 0) is a twisted covariant system. We shall show that the above strong Morita equivalences are equivalence relations. Proposition 3.7. The strong Morita equivalence of weak coactions of a finite dimensional C ∗-Hopf algebra on a unital C ∗-algebra is an equivalence relation. Proof. It suffices to show the transitivity since the other conditions clearly hold. Let A, B, C be unital C ∗-algebras and let X and Y be an A − B- equivalence bimodule and a B − C-equivalence bimodule, respectively. Let THE STRONG MORITA EQUIVALENCE FOR COACTIONS 7 ρ, σ and γ be weak coactions of H 0 on A, B and C, respectively. We sup- pose that ρ is strongly Morita equivalent to σ and that σ is strongly Morita equivalent to γ. Let λ and µ be weak coactions of H 0 on X and Y satis- fying Definition 3.6(i), respectively. Then X ⊗B Y is an A − C-equivalence bimodule. We define a bilinear map " ·λ⊗µ" from H × (X ⊗B Y ) to X ⊗B Y as follows: For any x ∈ X, y ∈ Y , h ∈ H, h ·λ⊗µ (x ⊗ y) = [h(1) ·λ x] ⊗ [h(2) ·µ y]. Then we can show that the above map " ·λ⊗µ" satisfies Conditions (1)'-(5)' in Definition 3.3 by routine computations. (cid:3) Corollary 3.8. The strong Morita equivalence of twisted coactions of a finite dimensional C ∗-Hopf algebra on a unital C ∗-algebra is an equivalence relation. Proof. By Proposition 3.7, we have only to show that Condition (∗∗)' in Definition 3.5. Let (ρ, u), (σ, v) and (γ, w) be twisted coactions of H 0 on unital C ∗-algebras A, B and C, respectively. Let the other notations be as in the proof of Proposition 3.7. For any x ∈ X, y ∈ Y , h, l ∈ H, h ·λ⊗µ [l ·λ⊗µ x ⊗ y] = [h(1) ·λ [l(1) ·λ x]] ⊗ [h(2) ·µ [l(2) ·µ y]] =bu(h(1), l(1))[h(2)l(2) ·λ x] ⊗ [h(3)l(3) ·µ y]bw∗(h(4), l(4)) =bu(h(1), l(1))[h(2)l(2) ·λ⊗µ (x ⊗ y)]bw∗(h(3), l(3)). Therefore, we obtain the conclusion. (cid:3) Of course, the notion of the strong Morita equivalence of coactions of a finite dimensional C ∗-Hopf algebra on unital C ∗-algebras is an extension of that of actions of a finite group on unital C ∗-algebras. We shall show it. Let G be a finite group and α an action of G on a unital C ∗-algebra A. We consider the coaction of C(G) on A induced by the action α of G on A. We denote it by the same symbol α. That is, α : A −→ A ⊗ C(G), αt(a) ⊗ δt a 7−→Xt∈G δt(s) =(0 if s 6= t 1 if s = t . for any a ∈ A, where for any t ∈ G, δt is a projection in C(G) defined by Let B be a unital C ∗-algebra and β an action of G on B. We denote by the same symbol β the coaction of C(G) on B induced by the action β. 8 KAZUNORI KODAKA AND TAMOTSU TERUYA Proposition 3.9. With the above notations, the following conditions are equivalent: (1) The actions α and β of G on A and B are strongly Morita equivalent, (2) The coactions α and β of C(G) on A and B are strongly Morita equiv- alent. Proof. We suppose Condition (1). Then by Raeburn and Williams [15, Def- inition 7.2], there are an A − B-equivalence bimodule X and an action u of G by linear isomorphisms of X such that αt( Ahx, yi) = Ahut(x) , ut(y)i, βt(hx, yiB) = hut(x) , ut(y)iB for any x, y ∈ X, t ∈ G. We note that by [15, Remark 7.3], we have the following equations: ut(ax) = αt(a)ut(x), ut(xb) = ut(x)βt(b) for any a ∈ A, b ∈ B, x ∈ X, t ∈ G. Let λ be a linear map from X to X ⊗ C(G) defined by for any x ∈ X, λ(x) =Xt∈G ut(x) ⊗ δt. Then by routine computations, we can see that λ is a coaction of C(G) on X with respect to (A, B, α, β). Hence we obtain Condition (2). Next we suppose Condition (2). Then there are an A − B-equivalence bimodule X and a coaction λ of C(G) on X with respect to (A, B, α, β). We regard G as a subset of C ∗(G). For any t ∈ G, we define a linear map ut on X as follows: For any x ∈ X, ut(x) = t ·λ x. Then for any t, s ∈ G, x ∈ X, ut(us(x)) = t ·λ [s ·λ x] = ts ·λ x = uts(x). Thus we can see that u is an action of G by linear isomorphisms of X, which satisfies the desired conditions by easy computations. Thus we obtain (1). (cid:3) Modifying [15, Example 7.4(b)], we shall obtain the following lemma, which can give examples of the strong Morita equivalence of coactions of a finite dimensional C ∗-Hopf algebra on a unital C ∗-algebra. Before it, we introduce the following definition: Definition 3.10. Let ρ and σ be weak coactions of H 0 on A. We say that ρ is exterior equivalent to σ if there is a unitary element w ∈ A⊗H 0 satisfying that σ = Ad(w) ◦ ρ, (id ⊗ ǫ0)(w) = 1. THE STRONG MORITA EQUIVALENCE FOR COACTIONS 9 Lemma 3.11. Let ρ and σ be weak coactions of H 0 on A. Then the following conditions are equivalent: (1) The weak coactions ρ and σ are exterior equivalent, (2) The weak coactions ρ and σ are strongly Morita equivalent by a weak coaction λ from an A−A-equivalence bimodule AAA to an A⊗H 0 −A⊗H 0- equivalence bimodule A⊗H 0A⊗H 0 A⊗H 0, where we regard A and A⊗H 0 as an A − A-equivalence bimodule and an A ⊗ H 0 − A ⊗ H 0-equivalence bimodule in the usual way, respectively. Proof. We suppose Condition (1). Then there is a unitary element w ∈ A ⊗ H 0 satisfing that σ = Ad(w) ◦ ρ and that (id ⊗ ǫ0)(w) = 1. Let λ be a linear map from AAA to A⊗H 0A ⊗ H 0 A⊗H 0 defined by λ(x) = ρ(x)w∗ for any x ∈ AAA. By routine computations, we can see that λ is a weak coaction of H 0 on AAA with respect to (A, A, ρ, σ). Next, we suppose Condition (2). We note that λ is a weak coaction of H 0 on AAA with respect to (A, A, ρ, σ). We identify A ⊗ H 0 with EndA⊗H 0(A ⊗ H 0 A⊗H 0) is a C ∗-algebra of all right A ⊗ H 0-module maps on A ⊗ H 0 A⊗H 0. Let w = θλ(1)∗ ,1⊗10 be a rank-one operator on A ⊗ H 0 A⊗H 0 induced by λ(1)∗ and 1 ⊗ 10. Then w is a unitary element in EndA⊗H 0(A ⊗ H 0 A⊗H 0). Indeed, for any x ∈ A ⊗ H 0 A⊗H 0), where EndA⊗H 0(A ⊗ H 0 A⊗H 0 ww∗(x) = λ(1)∗(1 ⊗ 10)λ(1)x = hλ(1) , λ(1)iA⊗H 0 x = σ(1)x = x, w∗w(x) = λ(1)λ(1)∗x = A⊗H 0hλ(1) , λ(1)ix = ρ(1)x = x. Also, for any a ∈ A, x ∈ A ⊗ H 0 A⊗H 0 (wρ(a)w∗)(x) = w(ρ(a)λ(1)x) = λ(1)∗λ(a)x = hλ(1) , λ(a)iA⊗H 0 x = σ(a)x. Thus w is a unitary element in A ⊗ H 0 and σ = Ad(w) ◦ ρ. Furthermore, let z = (id ⊗ ǫ0)(w). Then z is a unitary element in A such that az = za for any a ∈ A since σ = Ad(w) ◦ ρ. Let w1 = w(z∗ ⊗ 10). Then w1 is a unitary element in A⊗H 0 satisfying that σ = Ad(w1) ◦ ρ and that (id⊗ǫ0)(w1) = 1. Therefore we obtain Condition (1). (cid:3) Lemma 3.12. Let (ρ, u) and (σ, v) be twisted coactions of H 0 on A. Then the following conditions are equivalent: (1) The twisted coactions (ρ, u) and (σ, v) are exterior equivalent, (2) The twisted coactions (ρ, u) and (σ, v) are strongly Morita equivalent by a twisted coaction λ from an A − A-equivalence bimodule AAA to an A ⊗ H 0 − A ⊗ H 0-equivalence bimodule A⊗H 0A ⊗ H 0 A⊗H 0, where we regard 10 KAZUNORI KODAKA AND TAMOTSU TERUYA A and A ⊗ H 0 as an A − A-equivalence bimodule and an A ⊗ H 0 − A ⊗ H 0- equivalence bimodule in the usual way, respectively. Proof. We suppose Condition (1). Then there is a unitary element w ∈ A ⊗ H 0 such that σ = Ad(w) ◦ ρ , v = (w ⊗ 10)(ρ ⊗ id)(w)u(id ⊗ ∆0)(w∗). Let λ be as in the proof of Lemma 3.11. Then for any x ∈ AAA, ((λ ⊗ id) ◦ λ)(x) = u(id ⊗ ∆0)(ρ(x))u∗(ρ ⊗ id)(w∗)(w∗ ⊗ 10) = u(id ⊗ ∆0)(ρ(x))(id ⊗ ∆0)(w∗)v∗ = u(id ⊗ ∆0)(λ(x))v∗. Thus by Lemma 3.11, λ is a twisted coaction of H 0 on AAA with respect to (A, A, ρ, u, σ, v). Next, we suppose Condition (2). We note that λ is a twisted coaction of H 0 on AAA with respect to (A, A, ρ, u, σ, v). We identify A ⊗ H 0 with EndA⊗H 0(A ⊗ H 0 A⊗H 0) is a C ∗- algebra of all right A ⊗ H 0-module maps on A ⊗ H 0 A⊗H 0. Let w = θλ(1)∗ ,1⊗10 A⊗H 0 induced by λ(1)∗ and 1 ⊗ 10. Then be a rank-one operator on A ⊗ H 0 w is a unitary element in EndA⊗H 0(A ⊗ H 0 A⊗H 0) such that σ = Ad(w) ◦ ρ by Lemma 3.11. We note that w∗ = A⊗H 0hλ(1) , 1 ⊗ 10i. Indeed, for any x ∈ A ⊗ H 0 A⊗H 0), where EndA⊗H 0(A ⊗ H 0 A⊗H 0 w∗x = (1 ⊗ 10)hλ(1)∗ , xiA⊗H 0 = λ(1)x = A⊗H 0hλ(1) , 1 ⊗ 10ix. Hence w∗ = A⊗H 0hλ(1) , 1 ⊗ 10i. Thus (ρ ⊗ id)(w∗) = (ρ ⊗ id)(A⊗H 0hλ(1) , 1 ⊗ 10i) = A⊗H 0⊗H 0h((λ ⊗ id) ◦ λ)(1) , λ(1) ⊗ 10i = A⊗H 0⊗H 0hu((id ⊗ ∆0) ◦ λ)(1)v∗ , λ(1) ⊗ 10i = u((id ⊗ ∆0) ◦ λ)(1)v∗(λ(1)∗ ⊗ 10). It follows that (ρ ⊗ id)(w∗)(w∗ ⊗ 10) = u((id ⊗ ∆0) ◦ λ)(1)v∗(λ(1)∗ ⊗ 10)(A⊗H 0hλ(1) , 1 ⊗ 10i ⊗ 10) = u((id ⊗ ∆0) ◦ λ)(1)v∗(hλ(1) , λ(1)iA⊗H 0 ⊗ 10) = u((id ⊗ ∆0) ◦ λ)(1)v∗(σ(h1 , 1iA) ⊗ 10) = u(id ⊗ ∆0)(λ(1))v∗ = u(id ⊗ ∆0)(A⊗H 0hλ(1) , 1 ⊗ 10i)v∗ = u(id ⊗ ∆0)(w∗)v∗. Thus v = (w ⊗ 10)(ρ ⊗ id)(w)u(id⊗ ∆0)(w∗). Therefore we obtain Condition (1). (cid:3) THE STRONG MORITA EQUIVALENCE FOR COACTIONS 11 Next, we discuss on relations between innerness, outerness and strong Morita equivalence. Let ρA H 0 be the trivial coaction of H 0 on A. Lemma 3.13. (i) Let ρ be a weak coaction of H 0 on A. Then the following conditions are equivalent: (1) The weak coaction ρ is inner, (2) The weak coaction ρ is strongly Morita equivalent to ρA (ii) Let ρ be a coaction of H 0 on A. Then the following conditions are equivalent: (1) The coaction ρ is strongly inner, (2) The coaction ρ is strongly Morita equivalent to ρA H 0. H 0. Proof. (i) We suppose that ρ is inner. Then we can see that there is a unitary H 0 and that (id⊗ǫ0)(w) = 1 element w ∈ A⊗H 0 satisfying that ρ = Ad(w)◦ρA in the same way as in the proof that Condition (2) implies Condition (1) in Lemma 3.11. Thus ρ is exterior equivalent to ρA H 0. Hence by Lemma 3.11, ρ is strongly Morita equivalent to ρA H 0. Next we suppose that ρ is strongly Morita equivalent to ρA H 0. Then there are an A − A-equivalence bimodule X and a weak coaction λ of H 0 on X with respect to (A, A, ρ, ρA H 0). We note that for any a ∈ A, x ∈ X, λ(xa) = λ(x)ρA H 0(a) = λ(x)(a ⊗ 10). bw(h)x = h ·λ x. algebra of all right A-module maps on X. Since X is an A − A-equivalence For any h ∈ H, let bw(h) be a linear map on X defined by for any x ∈ X, Then by the above discussion, bw(h) ∈ EndA(X), where EndA(X) is a C ∗- bimodule, we can identify EndA(X) with A and we can regard bw(h) as an element in A for any h ∈ H. Furthermore, since the map h 7→ bw(h) is linear, bw ∈ Hom(H, A). Let w be the element in A ⊗ H 0 induced by bw ∈ Hom(H, A). By the definition of w, clearly bw(1) = 1. We show that w is a unitary element in A ⊗ H 0 such that ρ = Ad(w) ◦ ρA h ∈ H, H 0. For any x, y ∈ X, h(cw∗bw)(h)x , yiA = hbw(h(2))x , bw(S(h∗ = h[h(2) ·λ x] , [S(h∗ (1)) ·λ y]iA = S(h∗) ·ρA H 0 (1)))yiA Thus w∗w = 1 ⊗ 10. Also, for any x, y ∈ X, h ∈ H, hx, yiA = hǫ(h)x , yiA. h ·ρ Ahx, yi = Ah[h(1) ·λ x] , [S(h∗ = bw(h(1))Ahx, yicw∗(h(2)). (2)) ·λ y]i = Ahbw(h(1))x , bw(S(h∗ (2)))yi 12 KAZUNORI KODAKA AND TAMOTSU TERUYA Hence ρ = Ad(w) ◦ ρA H 0 since X is an A − A-equivalence bimodule. Thus H 0(1)w∗ = ρ(1) = 1 ⊗ 10. Therefore, the weak coaction ρ is inner. ww∗ = wρA (ii) We suppose that a coaction ρ is strongly inner. Then ρ is exterior equiv- alent to ρA H 0. Next, we suppose that ρ is strongly Morita equivalent to ρA H 0. Then there are an A − A-equivalence bimodule X and a coaction λ of H 0 on X with respect to (A, A, ρ, ρA H 0). Let w be as in (i). It suffices to show that for any H 0. Hence by Lemma 3.12, ρ is strongly Morita equivalent to ρA h, l ∈ H, bw(hl) = bw(h)bw(l). For any x ∈ X, h, l ∈ H, bw(h)bw(l)x = h ·λ [l ·λ x] = hl ·λ x = bw(hl). Therefore, ρ is strongly inner. (cid:3) Let ρA H 0 and ρB H 0 be the trivial coactions of H 0 on A and B, respectively. We suppose that A and B are strongly Morita equivalent and let X be an A − B-equivalence bimodule. Then ρA H 0 are strongly Morita equivalent. If a linear map λX H 0(x) = H 0 is a coaction of H 0 on X with respect x ⊗ 10 for any x ∈ X, then the λX to (A, B, ρA H 0 from X to X ⊗ H 0 is defined by λX H 0 and ρB H 0, ρB H 0). Corollary 3.14. (i) Let ρ and σ be weak coactions of H 0 on A and B, respectively. If ρ is strongly Morita equivalent to σ, then ρ is inner if and only if so is σ. (ii) Let ρ and σ be coactions of H 0 on A and B, respectively. If ρ is strongly Morita equivalent to σ, then ρ is strongly inner if and only if so is σ. Proof. (i) We suppose that ρ is inner. Then σ is strongly Morita equivalent to ρB H 0 by Lemma 3.13(i), Proposition 3.7 and the above discussion. There- fore, σ is inner by Lemma 3.13(i). (ii) We suppose that ρ is strongly inner. Then σ is strongly Morita equiv- alent to ρB H 0 by Lemma 3.13(ii), Corollary 3.8 and the above discussion. Therefore, σ is strongly inner by Lemma 3.13(ii). (cid:3) Proposition 3.15. We suppose that H 0 is not trivial. Let ρ and σ be coac- tions of H 0 on A and B, respectively. If ρ is strongly Morita equivalent to σ, then ρ is outer if and only if so is σ. Proof. We suppose that ρ is outer. We show that σ is outer. Let π be a surjective C ∗-Hopf algebra homomorphism of H 0 onto a non-trivial C ∗- Hopf algebra K 0. We suppose that (id ⊗ π) ◦ σ is inner. Then (id ⊗ π) ◦ σ is strongly Morita equivalent to (id ⊗ π) ◦ ρ by easy computations. Thus by Corollary 3.14(i), (id ⊗ π) ◦ ρ is inner. This is a contradiction. Therefore, we obtain the conclusion. (cid:3) THE STRONG MORITA EQUIVALENCE FOR COACTIONS 13 Furthermore, we have also the following easy lemma: Lemma 3.16. Let (ρ, u) be a twisted coaction of H 0 on A and let (ρ ⊗ id, u ⊗ In) be a twisted coaction of H 0 on A ⊗ Mn(C), where n is any positive integer and we identify A ⊗ H 0 ⊗ Mn(C) with A ⊗ Mn(C) ⊗ H 0. Then (ρ, u) is strongly Morita equivalent to (ρ ⊗ id, u ⊗ In). Proof. Let f be a minimal projection in Mn(C) and let X = (1 ⊗ f )(A ⊗ Mn(C)). We regard it as an A − A ⊗ Mn(C)-equivalence bimodule in the usual way. Let λ be a linear map from X to X ⊗ H 0 defined by λ((1 ⊗ f )x) = (1 ⊗ f ⊗ 10)(ρ ⊗ id)(x) for any x ∈ A⊗Mn(C), where we identify A⊗H 0⊗Mn(C) with A⊗Mn(C)⊗ H 0. By routine computations, we can see that λ satisfies Conditions (1)-(5) in Definition 3.3 and Condition (∗∗). (cid:3) 4. Crossed products of Hilbert C ∗-bimodules of finite type by finite dimensional C ∗-Hopf algebras In this section, we extend the notion of crossed products of Hilbert C ∗- bimodules of finite type defined in [7], [9] to (twisted) coactions of finite dimensional C ∗-Hopf algebras. Let H be a finite dimensional C ∗-Hopf algebra with its dual C ∗-Hopf algebra H 0. Let A and B be unital C ∗-algebras and X a Hilbert A − B- bimodule of finite type. Let (A, B, X, ρ, u, σ, v, λ, H 0) be a twisted covariant system. Under certain conditions, we define X ⋊λ H, a Hilbert A ⋊ρ,u H − B ⋊σ,v H-bimodule of finite type as follows: X ⋊λ H is just X ⊗ H (the algebraic tensor product) as vector spaces. Its left action and right action are given by (a ⋊ρ,u h)(x ⋊λ l) = a[h(1) ·λ x]bv(h(2), l(1)) ⋊λ h(3)l(2), (x ⋊λ l)(b ⋊σ,v m) = x[l(1) ·σ,v b]bv(l(2), m(1)) ⋊λ l(3)m(2) for any a ∈ A, b ∈ B, x ∈ X and h, l, m ∈ H. Then for any a1, a2 ∈ A, x ∈ X, h, l, m ∈ H, ((a1 ⋊ρ,u h)(a2 ⋊ρ,u l))(x ⋊λ m) = a1[h(1) ·ρ,u a2]bu(h(2), l(1))[h(3)l(2) ·λ x]bv(h(4)l(3), m(1)) ⋊λ h(5)l(4)m(2) = a1[h(1) ·λ a2[l(1) ·λ x]]bv(h(2), l(2))bv(h(3)l(3), m(1)) ⋊λ h(4)l(4)m(2) = a1[h(1) ·λ a2[l(1) ·λ x]bv(l(2), m(1))]bv(h(2), l(3)m(2)) ⋊λ h(3)l(4)m(3) = (a1 ⋊ρ,u h)((a2 ⋊ρ,u l)(x ⋊λ m)). 14 KAZUNORI KODAKA AND TAMOTSU TERUYA Also, for any b1, b2 ∈ B, x ∈ X, h, l, m ∈ H, (x ⋊λ h)((b1 ⋊σ,v l)(b2 ⋊σ,v m)) ⋊λ h(5)l(4)m(2) = x[h(1) ·σ,v b1][h(2) ·σ,v [l(1) ·σ,v b2]]bv(h(3), l(2))bv(h(4)l(3), m(1)) = x[h(1) ·σ,v b1]bv(h(2), l(1))[h(3)l(2) ·σ,v b2]bv(h(4)l(3), m(1)) ⋊λ h(5)l(4)m(2) = ((x ⋊λ h)(b1 ⋊σ,v l))(b2 ⋊σ,v m). Thus X ⋊λ H is a left A ⋊ρ,u H and right B ⋊σ,v H-bimodule. Also, its left A ⋊ρ,u H-valued inner product and right B ⋊σ,v H-valued inner product are given by A⋊ρ,uHhx ⋊λ h , y ⋊λ li = Ahx , [S(h(2)l∗ hx ⋊λ h , y ⋊λ liB⋊σ,vH (2), S(h(1))∗)[h∗ (3))∗ ·λ y]bv(S(h(1)l∗ (2))∗ , l(1))i ⋊ρ,u h(3)l∗ (4), = bv∗(h∗ (4), l(1)) ⋊σ,v h∗ (5)l(2) (3) ·σ,v hx , yiB]bv(h∗ for any x, y ∈ X and h, l ∈ H. We shall show that X ⋊λ H is a Hilbert A ⋊ρ,u H − B ⋊σ,v H-bimodule of finite type proving that X ⋊λ H satisfies Conditions (1)-(10) in [9, Lemma 1.3]. Clearly X ⋊λ H is a left A ⋊ρ H- and right B ⋊σ H-bimodule. Thus Conditions (1), (4) in [9, Lemma 1.3] are satisfied. For any a, b ∈ A, x, y ∈ X and h, l, m ∈ H, (a ⋊ρ,u h) A⋊ρ,uH hx ⋊λ l , y ⋊λ mi = a[h(1) ·ρ,u Ahx , [S(l(2)m∗ ⋊ρ,u h(3)l(4)m∗ (5) (3))∗ ·λ y]bv(S(l(1)m∗ = a Ah[h(1) ·λ x] , [S(h∗ (3) ·λ [S(l(2)m∗ (3))∗ ·λ y][S(h∗ (2))∗, m(1))]i (4)) (2))∗, m(1))i]bu(h(2), l(3)m∗ (2)) ·σ,vbv(S(l(1)m∗ (4)), S(l(3)m∗ (3))∗) (5)), S(l(4)m∗ (4))∗)[S(h(4)l(3)m∗ (3))∗ ·λ y] (4))∗ ·λ y]]bv(S(h∗ (1)))ibu(h(6), l(5)m∗ (5)) (2)), S(l∗ (5)) ·λ [S(l(4)m∗ ⋊ρ,u h(7)l(6)m∗ (6) (4)) ⋊ρ,u h(5)l(4)m∗ (5) = a Ah[h(1) ·λ x] , [S(h∗ ×bu(h(4), l(3)m∗ ×bv(S(h(3)l(2)m∗ = a Ah[h(1) ·λ x] ,bu(S(h∗ ×bv(S(h(3)l(2)m∗ = a Ah[h(1) ·λ x]bv(h(2), l(1)) , [S(h(4)l(3)m∗ × bu∗(h(5), l(4)m∗ = a Ah[h(1) ·λ x]bv(h(2), l(1)) , [S(h(4)l(3)m∗ = A⋊ρ,uHha[h(1) ·λ x]bv(h(2), l(1)) ⋊λ h(3)l(2) , y ⋊λ mi (2))∗, m(1))bv∗(S(h∗ (2))∗, m(1))bv(h(2), l(1))∗ibu(h(6), l(5)m∗ (4))bu(h(6), l(5)m∗ (5)) ⋊ρ,u h(7)l(6)m∗ (6) ⋊ρ,u h(5)l(4)m∗ (4) (5)) ⋊ρ,u h(7)l(6)m∗ (6) (3))∗ ·λ y]bv(S(h(3)l(2)m∗ (3))∗ ·λ y]bv(S(h(3)l(2)m∗ (2))∗, m(1))i (2))∗, m(1))i THE STRONG MORITA EQUIVALENCE FOR COACTIONS 15 = A⋊ρ,uH h(a ⋊ρ,u h)(x ⋊λ l) , y ⋊λ mi. Also, hx ⋊λ h , y ⋊λ liB⋊σ,vH(b ⋊σ,v m) (1)))[h∗ (1)))[h∗ (3) ·σ,v hx, yiB]bv(h∗ (3) ·σ,v hx, yiB][h∗ (4), l(1))[h∗ (6)l(3), m(1)) (5)l(2) ·σ,v b]bv(h∗ (6)l(3), m(1)) (1)))[h∗ (3) ·σ,v hx, yiB[l(1) ·σ,v b]]bv(h∗ (3) ·σ,v hx, yiB[l(1) ·σ,v b]][h∗ (1)))[h∗ (5), l(2))bv(h∗ (4) ·σ,v [l(1) ·σ,v b]]bv(h∗ (4), l(2))bv(h∗ (4) ·σ,vbv(l(2), m(1))]bv(h∗ (5)l(3), m(1)) (5), l(3)m(2)) ⋊σ,v h∗ ⋊σ,v h∗ = bv∗(h∗ = bv∗(h∗ = bv∗(h∗ = bv∗(h∗ = bv∗(h∗ (2), S(h∗ (7)l(4)m(2) (2), S(h∗ (7)l(4)m(2) (2)S(h∗ (6)l(4)m(2) (2), S(h∗ (6)l(4)m(3) (2), S(h∗ (5)l(4)m(3) ⋊σ,v h∗ ⋊σ,v h∗ ⋊σ,v h∗ (1)))[h∗ (3) ·σ,v hx, yiB[l(1) ·σ,v b]bv(l(2), m(1))]bv(h∗ (4), l(3)m(2)) = hx ⋊λ h , (y ⋊λ l)(b ⋊σ,v m)iB⋊σ,vH. Thus Conditions (3), (6) in [9, Lemma 1.3] are satisfied. For any x, y ∈ X and h, l ∈ H, A⋊ρ,uHhx ⋊λ h , y ⋊λ li∗ (3)))[l(6)h∗ (5) ·ρ,u Ah[S(l(3)h∗ (1)), l(1)) , xi] (2)) ·λ y]bv(S(l(2)h∗ (4), S(l(4)h∗ ⋊ρ,u l(7)h∗ (6) (4), S(l(4)h∗ (3))) (5) ·λ [S(l(3)h∗ = bu∗(l(5)h∗ = bu∗(l(5)h∗ = bu∗(l(6)h∗ × bv∗(l(10)h∗ = bu∗(l(6)h∗ × Ahbu(l(7)h∗ × bv∗(l(10)h∗ × Ah[l(6)h∗ ⋊ρ,u l(9)h∗ (8) (6), S(l(5)h∗ (5))) (7) ·λ [S(l(4)h∗ (10), S(h∗ (6), S(l(5)h∗ (7), S(l(4)h∗ (10), S(h∗ = Ahy , [S(l(2)h∗ × Ah[l(7)h∗ = A⋊ρ,uHhy ⋊λ l , x ⋊λ hi. (2)) ·λ y]][l(7)h∗ (1)), l(1))] , [S(l(8)h∗ (7))∗ ·λ x]i (6) ·σ,vbv(S(l(2)h∗ (9)S(l(2)h∗ (2)), l(1)) (9)S(l(2)h∗ (2)), l(1)) (4)) ·λ y]]bv(l(8)h∗ (1))) , [S(l(11)h∗ (5))) (4)))[l(8)h∗ (8)S(l(3)h∗ (11))∗ ·λ x]i ⋊ρ,u l(12)h∗ (8), S(l(3)h∗ (12) (3)))bv(l(9)h∗ (3)) ·λ y]bv(l(9)h∗ (11))∗ ·λ x]i ⋊ρ,u l(8)h∗ (8) (2))∗, h(1))i ⋊ρ,u l(3)h∗ (4) (1))) , [S(l(11)h∗ (3))∗ ·λ x]bv(S(l(1)h∗ 16 KAZUNORI KODAKA AND TAMOTSU TERUYA Similarly hx ⋊λ h , y ⋊λ li∗ (3)h(6), S(l∗ (5)h(8) ⋊σ,v l∗ B⋊σ,vH (2)h(5)))[l∗ × [l∗ × [l∗ = [[S(h∗ (7)l(3)), h∗ (3)h(6))∗, (l∗ (2)h(5))∗)∗[S(h∗ (3)h(6), S(l(2)h(5)))[l(4)∗ h(7) ·σ,v bv∗(S(h(4)), S(l∗ (5)h(8) ·σ,v [S(h(3)) ·σ,v hy, xiB]][l∗ (5)h(8) ·σ,v [S(h(3)) ·σ,v hy, xiB]][l∗ (4)h(7) ·σbv(h∗ = bv∗(l∗ = bv∗(l∗ =bv(S(l∗ (4), l(1))]bv(S(h∗ (7)l(4)) ·σ,vbv(h∗ (6)h(9) ·σ,vbv(S(h(2)), h(1))] ⋊σ,v l∗ (4))bv(S(h∗ = [bv(S(h∗ (5)h(9) ·σ,vbv(S(h(2)), h(1))] ⋊σ,v l∗ =bv(S(l(2)), l(1))∗bv∗(l∗ (5)h(7) ·σ,vbv(S(h(2)), h(1))] ⋊σ,v l∗ (3)h(5)S(h(4)) ·σ,v hy, xiB]bv∗(l∗ =bv(S(l(2)), l(1))∗[l∗ (5)h(7) ·σ,vbv(S(h(2)), h(1))[⋊σ,vl∗ (3) ·σ,v hy, xiB]bv∗(l∗ =bv(S(l(2)), l(1))∗[l∗ (3) ·σ,v hy, xiB]bv(l∗ =bv(S(l(2)), l(1))∗[l∗ (6)l(3)), h∗ (7)h(10) (5), l(1))]∗[l∗ (6)h(10) (3)h(5), S(h(4)))[l∗ (6)l(2))h∗ (6)h(6) (6)h(8) ⋊σ,v l∗ (6)h(7) ⋊σ,v l∗ (6)h(8) = hy ⋊λ l , x ⋊λ hiB⋊σ,vH. × [l∗ × [l∗ × [l∗ × [l∗ (4)h(4), S(h(3)))[l∗ (4)h(5)S(h(4)), h(1))bv∗(l∗ (4)h(6) ·σ,v [S(h(3)) ·σ,v hy, xiB]] (4)h(6), S(h(3))) (1)))] (4), l(1))∗[S(h(3)) ·σ,v hy, xiB]bv(S(h(2)), h(1))] (6)h(9) ·σ,vbv(S(h(2)), h(1))] ⋊σ,v l∗ (7)l(4)) ·σbv(h∗ (6)h(9) ·σ,vbv(S(h(2)), h(1))] ⋊σ,v l∗ (5)l(2))]∗[l∗ (4), l(1))]∗ (7)h(10) (7)h(10) (5)h(8) ·σ,v [S(h(3)) ·σ,v hy, xiB]] (4)h(8) ·σ,v [S(h(3)) ·σ,v hy, xiB]] (5)h(5) ·σ,vbv(S(h(2)), h(1))] (5)h(6), S(h(3))h(2)) Thus Conditions (2), (5) in [9, Lemma 1.3] are satisfied. Moreover, for any b ∈ B, x, y ∈ X, l, m ∈ H, A⋊ρ,uHhx ⋊λ l , (y ⋊λ m)(b ⋊σ,v 1)∗i = Ahx , [S(l(3)m∗ ⋊ρ,u l(4)m∗ (6) = Ahx[l(1) ·σ,v b] , [S(l(3)m∗ (5))∗ ·λ y]bv(S(l(2)m∗ = A⋊ρ,uHh(x ⋊λ l)(b ⋊σ,v 1) , y ⋊λ mi. (3))∗ ·λ y]bv(S(l(2)m∗ (4))∗, m(1))[S(l(1)m∗ (3))∗m(2) ·σ,v b∗]i (2))∗, m(1))i ⋊ρ,u l(4)m∗ (4) THE STRONG MORITA EQUIVALENCE FOR COACTIONS 17 Also, for any x, y ∈ X, h, l, m ∈ H, A⋊ρ,uHhx ⋊λ l , (y ⋊λ m)(1 ⋊σ,v h)∗i (4))∗, m(3)h∗ (3)) ⋊λ m(3)h∗ (3))] (4)i (4), S(h∗ (1)))] (2), S(h∗ (1)))] = Ahx , [S(l(3)h(6)m∗ = Ahx , [S(l(2)h(7)m∗ = Ahx , [S(l(2)h(6)m∗ (4))∗, m(3)h∗ (5))i ⋊ρ,u l(3)h(8)m∗ (6) (4)) ⋊ρ,u l(3)h(7)m∗ (6) (3)S(h∗ (5))∗ ·λ y[m(1) ·σ,v bv∗(h∗ (5))∗ ·λ ybv(m(1), h∗ = A⋊ρ,uHhx ⋊λ l , y[m(1) ·σ,vbv(S(h(2)), h(1))∗]bv(m(2), h∗ (2), S(h(1))∗)]bv(m(2), h∗ ×bv(S(l(1)h(5)m∗ (2)))bv∗(m(2)h∗ ×bv(S(l(1)h(6)m∗ (4))∗ ·σ,v bv∗(m(1)h∗ ×bv(S(l(1)h(4)m∗ × bv∗(S(l(1)h(5)m∗ = Ahxbv(l(1), h(1)) , [S(l(3)h(3)m∗ (3))∗ ·λ y]bv(S(l(2)h(2)m∗ (5))∗ ·λ y][S(l(2)h(5)m∗ (3))i ⋊ρ,u l(4)h(7)m∗ (6) (5))∗ ·λ y]bv(S(l(2)h(6)m∗ (4))∗, m(1)h∗ (1)))i ⋊ρ,u l(4)h(8)m∗ (6) (2))∗, m(1))i (4), S(h∗ (3)S(h∗ (2))) = Ahx , [S(l(3)h(7)m∗ (3))∗, m(2)h∗ (3))∗m(2)h∗ ⋊ρ,u l(4)h(4)m∗ (4) = A⋊ρ,uHh(x ⋊λ l)(1 ⋊σ,v h) , y ⋊λ mi. Thus we obtain that for any b ∈ B, x, y ∈ X, h, l, m ∈ H, A⋊ρ,uHh(x ⋊λ l)(b ⋊σ,v h) , y ⋊λ mi = A⋊ρ,uHhx ⋊λ l , (y ⋊λ m)(b ⋊σ,v h)∗i. We note that for any a ∈ A, x, y ∈ X, h, l, m ∈ H, h(a ⋊ρ,u h)(x ⋊λ l) , y ⋊λ miB⋊σ,vH = (1 ⋊σ,v l)∗h(a ⋊ρ,u h)(x ⋊λ 1) , y ⋊λ 1iB⋊σ,vH(1 ⋊σ,v m). Hence in order to show that for any a ∈ A, x, y ∈ X, h, l, m ∈ H, h(a ⋊ρ,u h)(x ⋊λ l) , y ⋊λ miB⋊σ,vH = hx ⋊λ l , (y ⋊λ m)(a ⋊ρ,u h)∗iB⋊σ,vH, we have only to show that for any a ∈ A, x, y ∈ X, h ∈ H, h(a ⋊ρ,u h)(x ⋊λ 1) , y ⋊λ 1iB⋊σ,vH = hx ⋊λ 1 , (a ⋊ρ,u h)∗(y ⋊λ 1)iB⋊σ,vH. For any a ∈ A, x, y ∈ X, h(a ⋊ρ,u 1)(x ⋊λ 1) , y ⋊λ 1iB⋊σ,vH = hax ⋊λ 1 , y ⋊λ 1iB⋊σ,vH = hax, yiB = hx ⋊λ 1 , (a ⋊ρ,u 1)∗(y ⋊λ 1)iB⋊σ,vH. 18 KAZUNORI KODAKA AND TAMOTSU TERUYA Also, for any x, y ∈ X, h ∈ H, h(1 ⋊ρ,u h)(x ⋊λ 1) , y ⋊λ 1iB⋊σ,vH = h[S(h(4)) ·λ [h(1) ·λ x]]bv(S(h(3)), h(2)) , [h∗ = hbu(S(h(2)), h(1))x , [h∗ = hx ⋊λ 1 ,bu(S(h(2)), h(1))∗[h∗ (3) ·λ y]iB ⋊σ,v h∗ (4) (3) ·λ y] ⋊λ h∗ = hx ⋊λ 1, (1 ⋊ρ,u h)∗(y ⋊λ 1)iB⋊σ,vH . (5) ·λ y]iB ⋊σ,v h∗ (6) (4)iB⋊σ,vH Thus Condition (8) in [9, Lemma 1.3] is satisfied. Moreover, for any a ∈ A, b ∈ B, x ∈ X, h, l, m ∈ H, (a ⋊ρ,u h)[(x ⋊λ l)(b ⋊σ,v m)] = a[h(1) ·λ x][h(2) ·σ,v [l(1) ·σ,v b]]bv(h(3), l(2))bv(h(4)l(3), m(1)) ⋊λ h(5)l(4)m(2) = a[h(1) ·λ x]bv(h(2), l(1))[h(3)l(2) ·σ,v b]bv(h(4)l(3), m(1)) ⋊λ h(5)l(4)m(2) = [(a ⋊ρ,u h)(x ⋊λ l)](b ⋊σ,v m). Thus Condition (7) in [9, Lemma 1.3] is satisfied. Since X is of finite type, there are finite subsets {wi}n j=1 in X such that i=1 and {zj}m x = nXi=1 wihwi, xiB = Ahx, zjizj mXj=1 for any x ∈ X. Then we have the following lemma: Lemma 4.1. With the above notations, if (A, B, X, ρ, σ, λ, H 0) is a covari- ant system, then for any x ∈ X, h ∈ H, x ⋊λ h = = nXi=1 mXj=1 (wi ⋊λ 1)hwi ⋊λ 1 , x ⋊λ hiB⋊σ H A⋊ρH hx ⋊λ h , zj ⋊λ 1i(zj ⋊λ 1). Proof. For any x ∈ X, h ∈ H, (wi ⋊λ 1)hwi ⋊λ 1 , x ⋊λ hiB⋊σ H = nXi=1 wihwi , xiB ⋊λ h = x ⋊λ h. nXi=1 THE STRONG MORITA EQUIVALENCE FOR COACTIONS 19 Also, A⋊ρHhx ⋊λ h , zj ⋊λ 1i(zj ⋊λ 1) mXj=1 mXj=1 mXj=1 = = Ah[h(2)S(h(1)) ·λ x] , [S(h(3))∗ ·λ zj]i[h(4) ·λ zj] ⋊λ h(5) [h(2) ·λ Ah[S(h(1)) ·λ x] , zjizj] ⋊λ h(3) = [h(2) ·λ [S(h(1)) ·λ x]] ⋊λ h(3) = x ⋊λ h. Therefore, we obtain the conclusion. (cid:3) For any Hilbert C ∗-bimodule Y , l − Ind[Y ] and r − Ind[Y ] denote its left index and right index, respectively. Corollary 4.2. With the above notations and assumptions, l − Ind[X ⋊λ H] = l − Ind[X] ⋊σ 1, r − Ind[X ⋊λ H] = r − Ind[X] ⋊ρ 1. Proof. By the definitions of the left index and the right index of a Hilbert C ∗-bimodule, l − Ind[X ⋊λ H] = r − Ind[X ⋊λ H] = mXj=1 nXi=1 hzj , zjiB ⋊σ 1 = l − Ind[X] ⋊σ 1, Ahwi , wii ⋊ρ 1 = r − Ind[X] ⋊ρ 1. (cid:3) Proposition 4.3. With the above notations and assupmtions, X ⋊λ H is a Hilbert A ⋊ρ H − B ⋊σ H-bimodule of finite type with l − Ind[X ⋊λ H] = l − Ind[X] ⋊σ 1, r − Ind[X ⋊λ H] = r − Ind[X] ⋊ρ 1. Proof. This is immediate by Lemma 4.1, Corollary 4.2 and [9, Lemma 1.3]. (cid:3) Lemma 4.4. With the above notations, if (A, B, X, ρ, u, σ, v, λ, H 0) is a twisted covariant system and X is an A − B-equivalence bimodule, then for any x ∈ X, h ∈ H, x ⋊λ h = = nXi=1 mXj=1 (wi ⋊λ 1)hwi ⋊λ 1 , x ⋊λ hiB⋊σ,zH A⋊ρ,wHhx ⋊λ h , zj ⋊λ 1i(zj ⋊λ 1). 20 KAZUNORI KODAKA AND TAMOTSU TERUYA Proof. For any x ∈ X, h ∈ H, (wi ⋊λ 1)hwi ⋊λ 1 , x ⋊λ hiB⋊σ,vH = wihwi , xiB ⋊λ h = x ⋊λ h. nXi=1 nXi=1 Also, A⋊ρ,uH hx ⋊λ h , zj ⋊λ 1i(zj ⋊λ 1) Ahx , [S(h(1))∗ ·λ zj]i[h(2) ·λ zj] ⋊λ h(3) xh[S(h(1))∗ ·λ zj] , [h(2) ·λ zj]iB ⋊λ h(3) x[h(1) ·σ hzj, zjiB] ⋊λ h(2) = x ⋊λ h. = = mXj=1 mXj=1 mXj=1 mXj=1 = Therefore, we obtain the conclusion. (cid:3) Lemma 4.5. With the above notations and assumptions, if X is an A − B- equivalence bimodule, then the Hilbert A ⋊ρ,u H − B ⋊σ,v H-bimodule is full with the both-sided inner products. Proof. For any x, y ∈ X, A⋊ρ,uH hx ⋊λ 1 , y ⋊λ 1i = Ahx , yi ⋊ρ,u 1. Since A⋊ρ,uHhX ⋊λ H , X ⋊λ Hi is a closed ideal of A ⋊ρ,u H, for any x, y ∈ X, h ∈ H, (Ahx , yi ⋊ρ,u 1)(1 ⋊ρ,u h) = Ahx , yi ⋊ρ,u h ∈ A⋊ρ,uHhX ⋊λ H , X ⋊λ Hi. Since AhX , Xi = A, we obtain that A⋊ρ,uHhX ⋊λ H , X ⋊λ Hi = A ⋊ρ,u H. Also, for any x, y ∈ X, h ∈ H, hx ⋊λ 1 , y ⋊λ hiB⋊σ,vH = hx , yiB ⋊σ,v h ∈ hX ⋊λ H , X ⋊λ HiB⋊σ,vH. Since hX , XiB = B, we obtain that hX ⋊λ H , X ⋊λ HiB⋊σ,vH = B ⋊σ,v H. (cid:3) Corollary 4.6. With the above notations and assumptions, we suppose that X is an A − B-equivalence bimodule. Then the X ⋊λ H is an A ⋊ρ,u H − B ⋊σ,v H-equivalence bimodule. THE STRONG MORITA EQUIVALENCE FOR COACTIONS 21 Proof. By Lemma 4.5, it suffices to show that A⋊ρ,uHhx ⋊λ h , y ⋊λ li(z ⋊λ m) = (x ⋊λ h)hy ⋊λ l , z ⋊λ miB⋊σ,vH for any x, y, z ∈ X, h, l, m ∈ H. Since X is an A − B-equivalence bimodule, (3) ·σ,v hy, ziB]bv(l∗ (4), m(1))]bv(h(2), l∗ (5)m(2)) (x ⋊λ h)hy ⋊λ l , z ⋊λ miB⋊σ,vH (2), S(l(1))∗)[l∗ (6)m(3) ⋊λ h(3)l∗ = x[h(1) ·σ,vbv∗(l∗ = x[h(1) ·σ,vbv∗(l∗ × [h(3) ·σ,vbv(l∗ = x[h(1) ·σ,vbv∗(l∗ ×bv(h(4)l∗ = xbv∗(h(1)l∗ On the other hand, (2), S(l(1))∗)][h(2) ·σ,v [l∗ (3) ·σ,v hy, ziB]] (6)m(3) (5)m(2)) ⋊λ h(5)l∗ (4), m(1))]bv(h(4), l∗ (2), S(l∗ (5), m(1)) ⋊λ h(5)l∗ (1)))[h(2)l∗ (2), S(l∗ (6)m(2) (3))[h(3)l∗ (1)))]bv(h(2), l∗ (3) ·σ,v hy, ziB]bv(h(3)l∗ (4) ·σ,v hy, ziB] (4), m(1)) ⋊λ h(4)l∗ (5)m(2). A⋊ρ,uHhx ⋊λ h , y ⋊λ li(z ⋊λ m) = Ahx, [S(h(2)l∗ ⋊λ h(5)l∗ (3))∗ ·λ y]bv(S(h(1)l∗ (3))∗ ·λ y]bv(S(h(1)l∗ (2), S(l∗ (1)))[h(2)l∗ (6)m(2) = xh[S(h(2)l∗ ⋊λ h(5)l∗ (6)m(2) = xbv∗(h(1)l∗ Therefore, we obtain the conclusion. (2))∗, l(1))i[h(3)l∗ (5), m(1)) (2))∗, l(1)) , [h(3)l∗ (5), m(1)) (4) ·λ z]bv(h(4)l∗ (4) ·λ z]iBbv(h(4)l∗ (3) ·σ,v hy, ziB]bv(h(3)l∗ (4), m(1)) ⋊λ h(4)l∗ (5)m(2). (cid:3) By the above discussions, we obtain the following: Corollary 4.7. (1) Let (A, B, X, ρ, u, σ, v, λ, H 0) be a twisted covariant sys- tem. We suppose that X is an A − B-equivalent bimodule. Then the crossed product X ⋊λ H is an A ⋊ρ,u H − B ⋊σ,v H-equivalence bimodule. (2) Let (A, B, X, ρ, σ, λ, H 0) be a covariant system. Then the crossed X ⋊λH is a Hilbert A ⋊ρ H − B ⋊σ H-bimodule of finite type. In the situation of Corollary 4.7(1), let X ⋊λ H be the crossed product associated to a twisted covariant system (A, B, X, ρ, u, σ, v, λ, H 0), where X is an A−B-equivalence bimodule. Then we define the dual covariant system with X ⋊λ H as follows: Letbρ andbσ be the dual coactions of H on A ⋊ρ,u H and B ⋊σ,v H of (ρ, u) and (σ, v), respectively. Letbλ be the dual coaction of H on X ⋊λ H defined by bλ(x ⋊λ h) = (x ⋊λ h(1)) ⊗ h(2) 22 KAZUNORI KODAKA AND TAMOTSU TERUYA for any x ∈ X, h ∈ H. Then by easy computations, we can see that is a covariant system. Hence we obtain the following: (A ⋊ρ,u H , B ⋊σ,v H , X ⋊λ H , bρ ,bσ ,bλ , H) Corollary 4.8. Let (ρ, u) and (σ, v) be twisted coactions of H 0 on A and B, respectively. Then the following conditions are equivalent: (1) The twisted coaction (ρ, u) is strongly Morita equivalent to the twisted coaction (σ, v), Proof. By the above discussion, it is clear that Condition (1) implies Condi- tion (2). We suppose Condition (2). Then by Condition (2), we can see that (2) The dual coaction bρ of (ρ, u) is strongly Morita equivalent to the dual coactionbσ of (σ, v). bbρ is strongly Morita equivalent tobbσ, wherebbρ andbbσ are the dual coactions of bρ and bσ, respectively. By [11, Theorem 3.3], there is an isomorphism Ψ of MN (A) onto A ⋊ρ,u H ⋊bρ H 0 such that bbρ is exterior equivalent to the Hence by Lemma 3.12,bbρ is strongly Morita equivalent to (ρ ⊗ id, u ⊗ IN ). Thus, by Lemma 3.16 and Corollary 3.8, bbρ is strongly Morita equivalent to (ρ, u). Similarly bbσ is strongly Morita equivalent to (σ, v). Therefore, by ((Ψ ⊗ id) ◦ (ρ ⊗ id) ◦ Ψ−1 , (Ψ ⊗ idH 0 ⊗ idH 0)(u ⊗ IN )). Corollary 3.8, (ρ, u) is strongly Morita equivalent to (σ, v). (cid:3) twisted coaction Also, in the situation of Corollary 4.7(2), we can see that is a covariant system in the same way as above. (A ⋊ρ H, B ⋊σ H, X ⋊λ H, bρ,bσ,bλ, H) 5. Duality In this section, we present a duality theorem for a crossed product of a Hilbert C ∗-bimodule of finite type by a (twisted) coaction of a finite dimensional C ∗-Hopf algebra in the same way as in [11]. As mentioned in Section 1, Guo and Zhang have already obtained a duality result using the language of multiplicative unitary elements and Kac systems in [5]. But we give our duality result because our approach to coactions of a finite dimensional C ∗-Hopf algebra on a unital C ∗-algebra is a useful addition to the main result in Section 6. First, we suppose Condition (1) or Condition (2) in Corollary 4.7. In the both cases, we can consider the dual covariant systems (A ⋊ρ,u H , B ⋊σ,v H , X ⋊λ H , bρ ,bσ ,bλ , H), THE STRONG MORITA EQUIVALENCE FOR COACTIONS 23 (A ⋊ρ H, B ⋊σ H, X ⋊λ H, bρ,bσ,bλ, H). Let Λ be the set of all triplets (i, j, k) where i, j = 1, 2, . . . , dk and k = I be k = N. For each I = (i, j, k) ∈ Λ, let W ρ I , V ρ k=1 d2 1, 2, . . . , K and PK elements in A ⋊ρ,u H ⋊bρ H 0 defined by W ρ I =pdk ⋊ρ,u wk ij, V ρ I = (1 ⋊ρ,u 1 ⋊bρ τ )(W ρ I ⋊bρ 10). Similarly for each I = (i, j, k) ∈ Λ, we define elements W σ I =pdk ⋊σ,v wk ij, V σ I = (1 ⋊σ,v 1 ⋊bσ τ )(W σ I ⋊bσ 10) in B ⋊σ,v H ⋊bσ H 0. We regard MN (C) as a Hilbert MN (C) − MN (C)- bimodule in the usual way. Let X ⊗ MN (C) be an exterior tensor product of X and MN (C), which is a Hilbert A ⊗ MN (C) − B ⊗ MN (C)-bimodule. In the same way as Lemma 2.1, we can see that X ⊗ MN (C) is of finite type. Let {fIJ }I,J ∈Λ be a system of matrix units of MN (C). Let ΨX be a linear map from X ⊗ MN (C) to X ⋊λ H ⋊bλ H 0 defined by ΨX(XI,J xIJ ⊗ fIJ) =XI,J V ρ∗ I (xIJ ⋊λ 1 ⋊bλ 10)V σ J for any xIJ ∈ X. Let ΨA and ΨB be isomorphisms of A ⊗ MN (C) and B ⊗ MN (C) onto A ⋊ρ,u H ⋊bρ H 0 and B ⋊σ,v H ⋊bσ H 0 defined by ΨA(XI,J ΨB(XI,J aIJ ⊗ fIJ ) =XI,J bIJ ⊗ fIJ ) =XI,J I (aIJ ⋊ρ,u 1 ×bρ 10)V ρ V ρ∗ J , I (aIJ ⋊σ,v 1 ×bσ 10)V σ V σ∗ J for any aIJ ∈ A, bIJ ∈ B, respectively (see [11]). 24 KAZUNORI KODAKA AND TAMOTSU TERUYA Lemma 5.1. With the above notations, xIJ ⊗ fIJ)) xIJ ⊗ fIJ), bIJ ⊗ fIJ )) aIJ ⊗ fIJ )(XI,J aIJ ⊗ fIJ )ΨX(XI,J xIJ ⊗ fIJ)(XI,J xIJ ⊗ fIJ )ΨB(XI,J (1) ΨX((XI,J = ΨA(XI,J (2) ΨX((XI,J = ΨX (XI,J (3) A⋊ρ,uH⋊ bρH 0hΨX(XI,J = ΨA(A⊗MN (C)hXI,J (4) hΨX(XI,J = ΨB(hXI,J xIJ ⊗ fIJ ) , ΨX(XI,J xIJ ⊗ fIJ ,XI,J bIJ ⊗ fIJ), xIJ ⊗ fIJ ) , ΨX(XI,J xIJ ⊗ fIJ ,XI,J yIJ ⊗ fIJ)i yIJ ⊗ fIJi), yIJ ⊗ fIJ )iB⋊σ,vH⋊ bσH 0 yIJ ⊗ fIJ iB⊗MN (C)) for any aIJ ∈ A, bIJ ∈ B, xIJ , yIJ ∈ X, I, J ∈ Λ. Proof. This is immediate by routine computations. Indeed, ΨX ((XI,J aIJ ⊗ fIJ )(XI,J xIJ ⊗ fIJ )) =XI,J,L V ρ∗ I (aILxLJ ⋊λ 1 ⋊bλ 10)V σ J . On the other hand, by [11, Lemma 3.1] aIJ ⊗ fIJ)ΨX (XL,M xLM ⊗ fLM ) V ρ∗ I (aIJ ⋊ρ,u 1 ⋊bρ 10)(1 ⋊ρ,u 1 ⋊bρ τ )(xJM ⋊λ 1 ⋊bλ 10)V σ M V ρ∗ I (aIJ xJM ⋊λ 1 ⋊bλ 10)V σ M . ΨA(XI,J = XI,J,M = XI,J,M THE STRONG MORITA EQUIVALENCE FOR COACTIONS 25 Thus we obtain Equation (1). Similarly we can obtain the Equation (2). Also, by [11, Lemma 3.1] xIJ ⊗ fIJ ) , ΨX(XI,J yIJ ⊗ fIJ )i A⋊ρ,uH⋊ bρH 0hV ρ∗ I (xIJ ⋊λ 1 ⋊bλ 10)V σ J , V ρ∗ I1 (yI1J1 ⋊λ 1 ⋊bλ 10)V σ J1i A⋊ρ,uH⋊ bρH 0hΨX(XI,J = XI,J,I1,J1 = XI,J,I1 = XI,J,I1 On the other hand, V ρ∗ I (A⋊ρ,uHhxIJ ⋊λ 1 , yI1J ⋊λ 1i ⋊bρ τ )V ρ I1 I (AhxIJ , yI1Ji ⋊ρ,u ⋊bρ10)V ρ V ρ∗ I1. xIJ ⊗ fIJ , XI1,J1 yI1,J1 ⊗ fI1,J1i) ΨA(AhxIJ , yI1J i ⊗ fII1) V ρ∗ I (AhxIJ , yI1J i ⋊ρ,u 1 ⋊bρ 10)V ρ I1. ΨA( A⊗MN (C)hXI,J = XI,J,I1 = XI,J,I1 Thus we obtain Equation (3). Furthermore, hΨX(XI,J = XI,J,J1 = ΨB(hXI,J xIJ ⊗ fIJ ) , ΨX(XI1,J1 J (hxIJ , yIJ1iB ⋊σ,v 1 ⋊bσ 10)V σ V σ∗ J1 yI1J1 ⊗ fI1J1)iB⋊σ,vH⋊ bσH 0 xIJ ⊗ fIJ , XI1,J1 yI1J1 ⊗ fIiJiiB⊗MN (C)). Thus we obtain Equation (4). (cid:3) By the above lemma, we can see that ΨX is injective. Next, we show that ΨX is surjective. Lemma 5.2. With the above notations, (X ⋊λ H ⋊bλ 10)(1 ⋊σ,v 1 ⋊bσ τ )(B ⋊σ,v H ⋊bσ 10) = X ⋊λ H ⋊bλ H 0. Proof. Let x ∈ X, h ∈ H, φ ∈ H 0. Since Xi,j,k (pdk ⋊σ,v wk ij ⋊bσ 10)∗(1 ⋊σ,v 1 ⋊bσ τ )(pdk ⋊σ,v wk ij ⋊bσ 10) = 1 ⋊σ,v 1 ⋊bσ 10 26 KAZUNORI KODAKA AND TAMOTSU TERUYA by [10, Proposition 3.18], x ⋊λ h ⋊bλ φ ij ij ⋊bσ 10) =Xi,j,k × (pdk ⋊σ,v wk = Xi,j,k,j1,j2 = Xi,j,k,j1,j2,j3 (x ⋊λ h ⋊bλ φ)(pdk ⋊σ,v wk dk((x ⋊λ h)[φ ·bσ (bv(S(wk j3j)((x ⋊λ h)(bv(S(wk × (1 ⋊σ,v wk ij dkφ(wk∗ × (1 ⋊σ,v 1 ⋊bσ τ )(1 ⋊σ,v wk ij ⋊bσ 10). ⋊bσ 10) ⋊bσ 10)∗(1 ⋊σ,v 1 ⋊bσ τ ) j1j2), wk ij1)∗ ⋊σ,v wk∗ j2j)] ⋊bλ τ ) j1j2), wk ij1)∗ ⋊σ,v wk∗ j2j3) ⋊bλ 10) Therefore we obtain the conclusion. (cid:3) Let Eσ 1 be the canonical conditional expectation from B ⋊σ,v H onto B 1 be a linear 1 (b ⋊σ,v h) = τ (h)b for any b ∈ B, h ∈ H. Let Eλ defined by Eσ map from X ⋊λ H onto X defined by for any x ∈ X, h ∈ H. Eλ 1 (x ⋊λ h) = τ (h)x Lemma 5.3. With the above notations, for any x ∈ X, h ∈ H, Xi,j,k (pdk ⋊ρ,u wk ij)∗Eλ 1 ((pdk ⋊ρ,u wk ij)(x ⋊λ h)) = x ⋊λ h. Proof. This is also immediate by routine computations. Indeed, for any x ∈ X, h ∈ H, by [17, Theorem 2.2], ij)(x ⋊λ h)) s3j ij)∗Eλ ss1, wk ⋊λ τ (wk j2jh(2))wk∗ Xi,j,k (pdk ⋊ρ,u wk = Xi,j,k,j1,j2,s,s1,s2,s3 = Xi,j,k,j1,j2,s2,s3 = Xi,j,k,j1,j2,s2,s3 = Xj,k,s2 1 ((pdk ⋊ρ,u wk dkbu∗(wk∗ dkxbv∗(wk∗ dkxbv(wk∗ s2jh) ⋊λ S(wk ij1, h(1))bv∗(wk∗ js2) = Xi,k,s2 dkxτ (wk ij1)[wk∗ is2, wk = x ⋊λ S(τ (Neh(1))S(h(2))) = x ⋊λ h. is2wk s2s3 ·σ,vbv(wk s2s3, wk si)[wk∗ s1s2 ·λ [wk ij1 ·λ x]][wk∗ j1j2, h(1))] s2s3 ·σ,vbv(wk j1j2, h(1))] ⋊λ τ (wk j2jh(2))wk∗ s3j j1j2h(2))τ (wk j2jh(3)) ⋊λ wk∗ s3j dkxτ (wk s2jh(1)) ⋊λ S(wk js2h(2)S(h(3))) Therefore, we obtain the conclusion. (cid:3) THE STRONG MORITA EQUIVALENCE FOR COACTIONS 27 Lemma 5.4. With the above notations, (1 ⋊ρ,u 1 ⋊bρ φ)(x ⋊λ 1 ⋊bλ 10) = x ⋊λ 1 ⋊bλ φ = (x ⋊λ 1 ⋊bλ 10)(1 ⋊σ,v 1 ⋊bσ φ) for any x ∈ X, φ ∈ H 0. Proof. For any x ∈ X, φ ∈ H 0, (1 ⋊ρ,u 1 ⋊bρ φ)(x ⋊λ 1 ⋊bλ 10) = [φ(1) ·bλ (x ⋊λ 1)] ⋊bλ φ(2) = x ⋊λ 1 ⋊bλ φ = (x ⋊λ 1 ⋊bλ 10)(1 ⋊σ,v 1 ⋊bσ φ). (cid:3) Lemma 5.5. With the above notations, ΨX is surjective. Proof. By Lemma 5.2, it suffices to show that for any b ∈ B, x ∈ X, h, l ∈ H, there is an element y ∈ X ⊗ MN (C) such that ΨX(y) = (x ⋊λ h ⋊bλ 10)(1 ⋊σ,v 1 ⋊bσ τ )(b ⋊σ,v l ⋊bσ 10). By Lemma 5.3 and [10, Proposition 3.18] x ⋊λ h =XI b ⋊σ,v l =XI W ρ∗ I (Eλ 1 (W ρ I (x ⋊λ h)) ⋊λ 1), (Eσ 1 ((b ⋊σ,v l)W σ∗ I ) ⋊σ,v 1)W σ I . (x ⋊λ h ⋊bλ 10)(1 ⋊σ,v 1 ⋊bσ τ )(b ⋊σ,v l ⋊bσ 10) (W ρ∗ I ⋊bρ 10)(Eλ 1 (W ρ I (x ⋊λ h)) ⋊λ 1 ⋊bλ τ ) 1 ((b ⋊σ,v l)W σ∗ J ) ⋊σ,v 1 ⋊bσ τ )(W σ J ⋊bσ 10). =XI,J × (Eσ Thus Since =XI,J 1 (W ρ Eλ 1 (W ρ I (x ⋊λ h)) ⋊λ 1 ⋊bλ τ = (1 ⋊ρ,u 1 ⋊bρ τ )(Eλ 1 (W ρ I (x ⋊λ h)) ⋊λ 1 ⋊bλ 10) by Lemma 5.4, (x ⋊λ h ⋊bλ 10)(1 ⋊σ,v 1 ⋊bσ τ )(b ⋊σ,v l ⋊bσ 10) 1 ((b ⋊σ,v l)W σ∗ I (x ⋊λ h))Eσ 1 (W ρ V ρ∗ I [Eλ J ) ⋊λ 1 ⋊bλ 10]V σ J . Since Eλ I (x ⋊λ h))Eσ 1 ((b ⋊σ,v l)W σ∗ J ) ∈ X, we obtain the conclusion. (cid:3) Let cV ρ be a linear map from H to A ⋊ρ,u H defined by cV ρ(h) = 1 ⋊ρ,u h for any h ∈ H. By [10],cV ρ is a unitary element in Hom(H, A⋊ρ,u H). Let V ρ be the unitary element in (A ⋊ρ,u H) ⊗ H 0 induced by cV ρ. Similarly, we also define unitary elements cV σ ∈ Hom(H, B ⋊σ,v H) and V σ ∈ (B ⋊σ,v H) ⊗ H 0. 28 KAZUNORI KODAKA AND TAMOTSU TERUYA Lemma 5.6. With the above notations, for any x ∈ X, h ∈ H, [h ·λ x] ⋊λ 1 = cV ρ(h(1))(x ⋊λ 1)dV σ∗(h(2)). Proof. This is also immediate by routine computations. Indeed, for any x ∈ X, h ∈ H, cV ρ(h(1))(x ⋊λ 1)dV σ∗(h(2)) = [h(1) ·λ x][h(2) ·σ,v bv∗(S(h(7)), h(8))]bv(h(3), S(h(6))) ⋊λ h(4)S(h(5)) = [h(1) ·λ x]bv(h(2), S(h(5))h(6))bv∗(h(3)S(h(4)), h(7)) ⋊λ 1 = [h ·λ x] ⋊λ 1. (cid:3) Theorem 5.7. (Cf. Guo and Zhang [5, Theorem 2.7]) Let A, B be unital C ∗-algebras and H a finite dimensional C ∗-Hopf algebra with its dual C ∗- Hopf algebra H 0. Then the following hold: (1) If X is an A − B-equivalence bimodule and (A, B, X, ρ, u, σ, v, λ, H 0) is a twisted covariant system, then there is a linear isomorphism ΨX from X ⊗MN (C) onto X ⋊λH ⋊bλH 0 which satisfies Conditions (1)-(4) in Lemma 5.1, where X ⋊λH ⋊bλH 0 is an A⋊ρ,uH ⋊bρH 0−B⋊σ,vH ⋊bσH 0 -equivalence bi- module and X⊗MN (C) is an exterior tensor product of an A−B-equivalence bimodule X and an MN (C) − MN (C)-equivalence bimodule MN (C). Fur- thermore, there are unitary elements U ∈ (A ⋊ρ,u H ⋊bρ H 0) ⊗ H 0 and V ∈ (B ⋊σ,v H ⋊bσ H 0) ⊗ H 0 such that Ubbλ(x)V ∗ = ((ΨX ⊗ id) ◦ (λ ⊗ idMN (C)) ◦ Ψ−1 X )(x) for any x ∈ X ⋊λ H ⋊bλ H 0. (2) If X is a Hilbert A−B-bimodule of finite type and (A, B, X, ρ, σ, λ, H 0) is a covariant system, then there is a linear isomorphism ΨX from X ⊗MN (C) onto X ⋊λ H ⋊bλ H 0 which satisfies Conditions (1)-(4) in Lemma 5.1, where X ⋊λ H ⋊bλ H 0 is a Hilbert A ⋊ρ H ⋊bρ H 0 − B ⋊σ H ⋊bσ H 0-bimodule of finite type and X ⊗ MN (C) is an exterior tensor product of a Hilbert A − B- bimodule X of finite type and an MN (C) − MN (C)-equivalence bimodule MN (C). Furthermore, there are unitary elements U ∈ (A ⋊ρ H ⋊bρ H 0) ⊗ H 0 and V ∈ (B ⋊σ H ⋊bσ H 0) ⊗ H 0 such that Ubbλ(x)V ∗ = ((ΨX ⊗ id) ◦ (λ ⊗ idMN (C)) ◦ Ψ−1 X )(x) for any x ∈ X ⋊λ H ⋊bλ H 0. Proof. (1) Let ΨX be as in Lemma 5.1. By Lemmas 5.1 and 5.5, we can see that ΨX is a linear isomorphism from X⊗MN (C) onto X ⋊λH ⋊bλH 0. By [11, THE STRONG MORITA EQUIVALENCE FOR COACTIONS 29 Theorem 3.3], there are U and V , unitary elements in (A ⋊ρ,u H ⋊bρ H 0)⊗H 0 and (B ⋊σ,v H ⋊bσ H 0) ⊗ H 0 such that A , B , MN (C), xIJ ⊗ fIJ ))V ∗ Ad(U) ◦bbρ = (ΨA ⊗ id) ◦ (ρ ⊗ idMN (C)) ◦ Ψ−1 Ad(V ) ◦bbσ = (ΨB ⊗ id) ◦ (σ ⊗ idMN (C)) ◦ Ψ−1 respectively. Let V ρ and V σ be as above. For any PI,J xIJ ⊗ fIJ ∈ X ⊗ Ubbλ(ΨX(XI,J =XI,J I ⊗ 10)V ρbbλ((1 ⋊ρ,u 1 ⋊bρ τ )(xIJ ⋊λ 1 ⋊bλ 10)(1 ⋊σ,v 1 ⋊bσ τ )) U =XI V =XI by [11, Lemma 3.1] since × V σ∗(V σ J ⊗ 10) (V ρ∗ (V σ∗ I ). Since I ), (V ρ∗ I ⊗ 10)V ρbbρ(V ρ I ⊗ 10)V σbbσ(V σ bbρ(1 ⋊ρ,u 1 ⋊bρ τ ) = V ρ∗((1 ⋊ρ,u 1 ⋊bρ τ ) ⊗ 10)V ρ, bbσ(1 ⋊σ,v 1 ⋊bσ τ ) = V σ∗((1 ⋊σ,v 1 ⋊bσ τ ) ⊗ 10)V σ Ubbλ(ΨX (XI,J =XI,J =XI,J I ⊗ 10)V ρ((xIJ ⋊λ 1 ⋊bλ 10) ⊗ 10)V σ∗(V σ I ⊗ 10)λ(xIJ ⋊λ 1 ⋊bλ 10)(V σ xIJ ⊗ fIJ ))V ∗ J ⊗ 10) J ⊗ 10) (V ρ∗ (V ρ∗ by the proof of [10, Proposition 3.19], by Lemma 5.6, where we identify X with X ⋊λ 1 and X ⋊λ 1 ⋊bλ 10. On the other hand, ((ΨX ⊗ id) ◦ (λ ⊗ id))(xIJ ⊗ fIJ ) = (ΨX ⊗ id)(λ(xIJ ) ⊗ fIJ ). I, J, i. Then We write that λ(xIJ ) = Pi yIJ i ⊗ φi, where φi ∈ H 0, yIJ i ∈ X for any ((ΨX ⊗ id)◦(λ ⊗ id))(xIJ ⊗ fIJ ) =XI,J,i V ρ∗ I (yIJ i ⋊λ 1 ⋊bλ 10)V σ J ⊗ φi (V ρ∗ I ⊗ 10)λ(xIJ ⋊λ ⋊bλ10)(V σ J ⊗ 10). =XI,J Therefore, we obtain the conclusion. (2) We can prove (2) in the same way as (1). (cid:3) 30 KAZUNORI KODAKA AND TAMOTSU TERUYA 6. The strong Morita equivalence for coactions and the Rohlin property For a unital C ∗-algebra A, we set c0(A) = { (an) ∈ l∞(N, A) lim n→∞ an = 0 }, A∞ = l∞(N, A)/c0(A). We denote an element in A∞ by the symbol [an] for an element (an) ∈ l∞(N, A). We identify A with the C ∗-subalgebra of A∞ consisting of the equivalence classes of constant sequences and set A∞ = A∞ ∩ A′. Let X be a Hilbert A − B- bimodule of finite type, where B is a unital C ∗-algebra. We define X ∞ in the same way as above. We set c0(X) = { (xn) ∈ l∞(N, X) lim n→∞ xn = 0 }, X ∞ = l∞(N, X)/c0(X). We denote an element in X ∞ by the symbol [xn] for an element (xn) ∈ l∞(N, X). We regard X ∞ as an A∞ − B∞-bimodule as follows: For any [an] ∈ A∞, [bn] ∈ B∞, [xn] ∈ X ∞, [an][xn] = [anxn], [xn][bn] = [xnbn]. Also, we define the left A∞-valued inner product and the right B∞-valued inner product as follows: For any [xn], [yn] ∈ X ∞, A∞h[xn] , [yn]i = [Ahxn , yni], h[xn] , [yn]iB∞ = [hxn , yniB]. By [15, Lemma 2.5] and easy computations, the above definitions are well- defined. We identify X with the Hilbert A∞ − B∞-subbimodule of X ∞ con- sisting of the equivalence classes of constant sequences. Also, we can see that X ∞ is a complex vector space satisfying Conditions (1)-(8) in [9, Lemma 1.3]. Since X is of finite type, there are finite subsets {ui}n j=1 ⊂ X such that for any x ∈ X, i=1, {vj}m uihui, xiB = x = nXi=1 Ahx, vjivj. mXj=1 Then we can regard ui, vj ∈ X as elements in X ∞ for i = 1, . . . , n, j = 1, . . . , m. Thus X ∞ is a Hilbert A∞ − B∞-bimodule of finite type by [9, Lemma 1.3]. Furthermore, if X is an A − B-equivalence bimodule, then X ∞ is an A∞ − B∞-equivalence bimodule. THE STRONG MORITA EQUIVALENCE FOR COACTIONS 31 Lemma 6.1. With the above notations, we suppose that X is an A − B- equivalence bimodule. Let b ∈ B∞. If xb = 0 for any x ∈ X, then b = 0, where we regard X as the Hilbert A∞ − B∞-subbimodule of X ∞. Proof. Since b ∈ B∞, we write that b = [bm], where bm ∈ B for any m ∈ N. Since xb = 0, xbm → 0 (m → ∞). For any y ∈ X, hy, xiB bm = hy, xbmiB ≤ y xbm → 0 (m → ∞) by [15, Lemma 2.5]. On the other hand, there are x1, . . . , xn, y1, . . . , yn ∈ X i=1hyi, xiiB = 1 since X is full with the right B-valued inner such that Pn product. Hence bm = Therefore b = 0. hyi, xiiB bm ≤ nXi=1 nXi=1 hyi, xiiB bm → 0. (cid:3) We are in position to present the main result in this paper. Before doing it, we give the definitions of the approximate representability and the Rohlin property for a coaction of a finite dimensional C ∗-Hopf algebra on a unital C ∗-algebra and a remark on the definitions. Definition 6.2. (Cf. [11, Definitions 4.3 and 5.1]) Let (ρ, u) be a twisted coaction of a finite dimensional C ∗-Hopf algebra H 0 on a unital C ∗-algebra A. We say that (ρ, u) is approximately representable if there is a unitary element w ∈ A∞ ⊗ H 0 satisfying the following conditions: (1) ρ(a) = (Ad(w) ◦ ρA H 0 ⊗ id)(w)(id ⊗ ∆0)(w∗), (2) u = (w ⊗ 10)(ρA∞ (3) u = (ρ∞ ⊗ id)(w)(w ⊗ 10)(id ⊗ ∆0)(w∗). H 0)(a) for any a ∈ A, Also, we say that (ρ, u) has the Rohlin property if its dual coaction bρ of H on A ⋊ρ H is approximately representable. By [11, Corollary 6.4], we can see that a coaction ρ of H 0 on A has the Rohlin property if and only if there is a projection p ∈ A∞ such that e ·ρ∞ p = 1 N , where N = dim(H). Theorem 6.3. Let H be a finite dimensional C ∗-Hopf algebra with its dual C ∗-Hopf algebra H 0. Let ρ and σ be coactions of H 0 on unital C∗-algebras A and B, respectively. We suppose that ρ is strongly Morita equivalent to σ. Then ρ has the Rohlin property if and only if σ has the Rohlin property. Proof. Since ρ and σ are strongly Morita equivalent, there are an A − B- equivalence bimodule X and a coaction λ of H 0 on X with respect to (A, B, ρ, σ). According to the proof of Rieffel [16, Proposition 2.1], we obtain 32 KAZUNORI KODAKA AND TAMOTSU TERUYA i=1, y = (yi)n are elements x1, . . . , xn, y1, . . . , yn ∈ X such that Pn the following: Since X is full with the right B-valued inner product, there i=1hxi, yiiB = 1. Let E = A ⊗ Mn(C) and we consider X n as an E − B-equivalence bimodule in the usual way. Let x = (xi)n 2 x and let q = Ehz, zi ∈ E. Then q is a projection in E. Let π be a map from B to E defined by π(b) = Ehzb, zi for any b ∈ B. Then π is an isomorphism of B onto qEq. We suppose that ρ has the Rohlin property. Then by [11, Corollary 6.4] there is a projection p ∈ A∞ such that e ·ρ∞ p = 1 N . We regard (X ∞)n as an E∞ − B∞-equivalence bimodule in the usual way. Since p ⊗ In ∈ E∞, there are elements i=i ∈ X n. Let z = Ehy, yi 1 u1, . . . , um, v1, . . . , vm ∈ (X ∞)n such that p ⊗ In =Pm k=1 E∞huk, vki. We write that uk = (uk1, . . . , ukn), vk = (vk1, . . . , vkn), where uki, vki ∈ X ∞ for k = 1, 2, . . . , m, i = 1, 2, . . . , n. Thus p ⊗ In = i,j=1. [A∞huki , vkji]n mXk=1 A∞huki , vkji =(p 0 i = j i 6= j . Hence (∗ ∗ ∗) mXk=1 We note that since p ∈ A∞, q(p ⊗ In)q = q(p ⊗ In) ∈ (qMn(A)q)∞ ∩ (qMn(A)q)′. Let π∞ be the isomorphism of B∞ onto (qMn(A)q)∞ induced by π. Let p1 = (π∞)−1(q(p ⊗ In)q). Then p1 is a projection in B∞ since π(B) = qMn(A)q. We show that e ·σ∞ p1 = 1 N . Since q = Ehz, zi, q(p ⊗ In)q = = mXk=1 mXk=1 E∞hEhz, ziuk , Ehz, zivki E∞hzhz, ukiB∞ hvk, ziB∞ , zi = π∞( hz, ukiB∞ hvk, ziB∞). mXk=1 Thus p1 = mXk=1 hz, ukiB∞ hvk, ziB∞ = hz , E∞huk , vkiziB∞. mXk=1 THE STRONG MORITA EQUIVALENCE FOR COACTIONS 33 Since z ∈ X n, we write z = (zi)n by Equation (∗ ∗ ∗), i=1, where zi ∈ X for i = 1, 2, . . . , n. Hence [ A∞huki , vkji]n i,j=1 z1 ... zn iB∞ hzi , A∞huki , vkjizjiB∞  , mXk=1 mXk=1 = hzi , pziiB∞. For any w ∈ X, w[e ·σ∞ p1] = z1 ... zn = p1 = h nXi,j=1 nXi=1 nXi=1 nXi=1 nXi=1 nXi=1 nXi=1 = = = = wh[S(e∗ (1)) ·λ zi] , [e(2) ·λ∞ pzi]iB∞ Ahw , [S(e∗ (1)) ·λ zi]i[e(2) ·λ∞ pzi] Ah[e(2)S(e(1)) ·λ w] , [S(e∗ (3)) ·λ zi]i [e(4) ·λ∞ pzi] [e(2) ·ρ Ah[S(e(1)) ·λ w] , zii][e(3) ·λ∞ pzi] [e(2) ·λ∞ p[S(e(1)) ·λ w]hzi, ziiB] = [e(2) ·ρ∞ p][e(3)S(e(1)) ·λ w]. Since e =Pi,k dk N wk ii, w[e ·σ∞ p1] = Xi,j,k,j1 =Xj,k dk N [wk jj1 ·ρ∞ p][wk j1iS(wk ij) ·λ w] dk N [wk jj ·ρ∞ p]w = [e ·ρ∞ p]w = 1 N w. Thus e ·σ∞ p1 = 1 [11, Corollary 6.4.]. N by Lemma 6.1. Therefore we obtain the conclusion by (cid:3) Corollary 6.4. Let (ρ, u) and (σ, v) be twisted coactions of H 0 on A and B, respectively. We suppose that they are strongly Morita equivalent. Then the following hold: (1) The twisted coaction (ρ, u) has the Rohlin property if and only if so does (σ, v), (2) The twisted coaction (ρ, u) is approximately representable if and only if so is (σ, v). 34 KAZUNORI KODAKA AND TAMOTSU TERUYA the Rohlin property by [11, Proposition 5.5]. Also, since (ρ, u) and (σ, v) Corollary 4.8. Thus (σ, v) has the Rohlin property by Theorem 6.3 and [11, Proposition 5.5]. Proof. (1) We suppose that (ρ, u) has the Rohlin property. Then bbρ has are strongly Morita equivalent,bbρ and bbσ are strongly Morita equivalent by (2) We suppose that (ρ, u) is approximately representable. Then bρ has the tion 4.6]. Since (ρ, u) and (σ, v) are strongly Morita equivalent, bρ andbσ are strongly Morita equivalent by Corollary 4.8. Thus by Theorem 6.3, bσ has the Rohlin property. Hence by the definition of the Rohlin property and [11, Proposition 4.6], (σ, v) is approximately representable. (cid:3) Rohlin property by the definition of the Rohlin property and [11, Proposi- 7. Application Let A and B be unital C ∗-algebras and H a finite dimensional C ∗-Hopf algebra with its dual C ∗-Hopf algebra H 0. We suppose that A is strongly Morita equivalent to B. Let ρ be a coaction of H 0 on A. By [16, Proposition 2.1], there are an n ∈ N and a full projection q ∈ Mn(A) such that B is isomorphic to qMn(A)q. We identify B with qMn(A)q. We suppose that (ρ ⊗ id)(q) ∼ q ⊗ 10 in Mn(A) ⊗ H 0. Hence there is a partial isometry w ∈ Mn(A) ⊗ H 0 such that w∗w = (ρ ⊗ id)(q), ww∗ = q ⊗ 10. Lemma 7.1. With the above notations, there is a partial isometry z ∈ q ⊗ 10, Mn(A) ⊗ H 0 such that z∗z = (ρ ⊗ id)(q), zz∗ = q ⊗ 10 and that bz(1) = q. Proof. We note that cw∗(1) = bw(1)∗. Since w∗w = (ρ ⊗ id)(q) and ww∗ = bw(1)cw∗(1) = q. cw∗(1)bw(1) = (id ⊗ ǫ0)((ρ ⊗ id)(q)) = q, Let z = (cw∗(1) ⊗ 10)w. Thenbz(1) =cw∗(1)bw(1) = q. Also, z∗z = w∗(bw(1) ⊗ 10)(cw∗(1) ⊗ 10)w = (ρ ⊗ id)(q), zz∗ = (cw∗(1) ⊗ 10)ww∗(bw(1) ⊗ 10) = q ⊗ 10. Therefore, we obtain the conclusion. (cid:3) Let σ = Ad(z) ◦ (ρ ⊗ idMn(C)), u = (z ⊗ 10)(ρ ⊗ idMn(C) ⊗ idH 0)(z)(idMn(A) ⊗ ∆0)(z∗). We note that u ∈ B⊗H 0⊗H 0. We shall show that (σ, u) is a twisted coaction of H 0 on B, which is strongly Morita equivalent to ρ. We sometimes identify A ⊗ H 0 ⊗ Mn(C) with A ⊗ Mn(C) ⊗ H 0. THE STRONG MORITA EQUIVALENCE FOR COACTIONS 35 Lemma 7.2. With the above notations, σ is a weak coaction of H 0 on B. Proof. For any x ∈ Mn(A), σ(qxq) = z(ρ ⊗ id)(qxq)z∗ = (q ⊗ 10)z(ρ ⊗ id)(x)z∗(q ⊗ 10). Hence σ is a map from B to B ⊗ H 0. Also, by routine computations, we can see that σ is a homomorphism of B to B ⊗ H 0 with σ(q) = q ⊗ 10. Furthermore, sincebz(1) = q, for any x ∈ Mn(A), (id ⊗ ǫ0)(σ(qxq)) = (id ⊗ ǫ0)((q ⊗ 10)z(ρ ⊗ id)(x)z∗(q ⊗ 10)) Thus σ is a weak coaction of H 0 on B. = qbz(1)(id ⊗ ǫ0)((ρ ⊗ id)(x))bz∗(1)q = qxq. (cid:3) Lemma 7.3. With the above notations, (σ, u) is a twisted coaction of H 0 on B. Proof. By routine computations, we can see that uu∗ = u∗u = q ⊗ 10 ⊗ 10. Thus u is a unitary element in B ⊗ H 0 ⊗ H 0. For any x ∈ Mn(A), ((σ ⊗ idH 0) ◦ σ)(qxq) = (z ⊗ 10)(ρ ⊗ id ⊗ idH 0)(z)((ρ ⊗ id ⊗ idH 0) ◦ (ρ ⊗ id))(qxq) × (ρ ⊗ id ⊗ idH 0)(z∗)(z∗ ⊗ 10). On the other hand, (Ad(u) ◦ (id ⊗ ∆0) ◦ σ)(qxq) = (z ⊗ 10)(ρ ⊗ id ⊗ idH 0)(z)((id ⊗ ∆0) ◦ (ρ ⊗ id))(qxq) × (ρ ⊗ id ⊗ idH 0)(z∗)(z∗ ⊗ 10). Since (ρ ⊗ id ⊗ idH 0) ◦ (ρ ⊗ id) = (id ⊗ ∆0) ◦ (ρ ⊗ id), we obtain that (σ ⊗ idH 0) ◦ σ = Ad(u) ◦ (id ⊗ ∆0) ◦ σ. Also, (u ⊗ 10)(id ⊗ ∆0 ⊗ idH 0)(u) = (z ⊗ 10 ⊗ 10)(ρ ⊗ id ⊗ idH 0 ⊗ idH 0)(z ⊗ 10) × (id ⊗ ∆0 ⊗ idH 0)((ρ ⊗ id ⊗ idH 0)(z)(id ⊗ ∆0)(z∗)). 36 KAZUNORI KODAKA AND TAMOTSU TERUYA On the other hand, since (ρ ⊗ id ⊗ idH 0) ◦ (ρ ⊗ id) = (id ⊗ ∆0) ◦ (ρ ⊗ id), (σ ⊗ idH 0 ⊗ idH 0)(u)(id ⊗ idH 0 ⊗ ∆0)(u) = (z ⊗ 10 ⊗ 10)(ρ ⊗ id ⊗ idH 0 ⊗ idH 0)(z ⊗ 10) × (id ⊗ ∆0 ⊗ idH 0)((ρ ⊗ id ⊗ idH 0)(z)) × (ρ ⊗ id ⊗ idH 0 ⊗ idH 0)((id ⊗ ∆0)(z∗)) × (id ⊗ idH 0 ⊗ ∆0)((ρ ⊗ id ⊗ idH 0)(z))(id ⊗ ∆0 ⊗ idH 0)((id ⊗ ∆0)(z∗)). We can see that (ρ ⊗ id ⊗ idH 0 ⊗ idH 0) ◦ (id ⊗ ∆0) = (id ⊗ idH 0 ⊗ ∆0) ◦ (ρ ⊗ id ⊗ idH 0) by easy computations. Furthermore, we note that Thus since (id ⊗ idH 0 ⊗ ∆0) ◦ (id ⊗ ∆0) ◦ (ρ ⊗ id) = (id ⊗ ∆0 ⊗ idH 0) ◦ (id ⊗ ∆0) ◦ (ρ ⊗ id) = (id ⊗ ∆0 ⊗ idH 0) ◦ (ρ ⊗ id ⊗ idH 0) ◦ (ρ ⊗ id). (id ⊗ idH 0 ⊗ ∆0)((ρ ⊗ id ⊗ idH 0)((ρ ⊗ id)(q))) = (id ⊗ idH 0 ⊗ ∆0)((id ⊗ ∆0)((ρ ⊗ id)(q))), (σ ⊗ idH 0 ⊗ idH 0)(u)(id ⊗ idH 0 ⊗ ∆0)(u) = (z ⊗ 10 ⊗ 10)(ρ ⊗ id ⊗ idH 0 ⊗ idH 0)(z ⊗ 10) × (id ⊗ ∆0 ⊗ idH 0)((ρ ⊗ id ⊗ idH 0)(z)) × (id ⊗ idH 0 ⊗ ∆0)((ρ ⊗ id ⊗ idH 0)((ρ ⊗ id)(q))) × (id ⊗ ∆0 ⊗ idH 0)((id ⊗ ∆0)(z∗)) = (z ⊗ 10 ⊗ 10)(ρ ⊗ id ⊗ idH 0 ⊗ idH 0)(z ⊗ 10) × (id ⊗ ∆0 ⊗ idH 0)((ρ ⊗ id ⊗ idH 0)(z)(id ⊗ ∆0)(z∗)). Hence we obtain that (u ⊗ 10)(id ⊗ ∆0 ⊗ idH 0)(u) = (σ ⊗ idH 0 ⊗ idH 0)(u)(id ⊗ idH 0 ⊗ ∆0)(u). Furthermore, sincebz(1) = q, for any h ∈ H, (id ⊗ h ⊗ ǫ0)(u) =bz(h(1))[h(2) ·ρ⊗id q]bz∗(h(3)) = (id ⊗ h)(σ(q)) = ǫ(h)q, (id ⊗ ǫ0 ⊗ h)(u) =bz(1)[1 ·ρ⊗idbz(h(1))]bz∗(h(2)) =bz(1)ǫ(h) = ǫ(h)q. Therefore, (σ, u) is a twisted coaction of H 0 on B. (cid:3) Let f be a minimal projection in Mn(C) and let p be a full projection in Mn(A) defined by p = 1A ⊗ f . Let X = pMn(A)q. We regard X as an A − B-equivalence bimodule in the usual way, where we identify A and THE STRONG MORITA EQUIVALENCE FOR COACTIONS 37 B with pMn(A)p and qMn(A)q, respectively. Then we can regard X as a set {[a1, . . . , an]q ai ∈ A, i = 1, . . . , n}. Let λ be a linear map from X to X ⊗ H 0 defined by λ([a1, . . . , an]q) = [ρ(a1), . . . , ρ(an)](ρ ⊗ id)(q)z∗ = [ρ(a1), . . . , ρ(an)]z∗(q ⊗ 10) for any [a1, . . . , an]q ∈ X. Lemma 7.4. With the above notations, λ is a twisted coaction of H 0 on X with respect to (A, B, ρ, σ, u). Proof. By routine computations, we can see that λ is a weak coaction of H 0 on X with respect to (A, B, ρ, σ, u). For any [a1, . . . , an]q ∈ X, ((λ ⊗ idH 0) ◦ λ)([a1, . . . , an]q) = [((ρ ⊗ idH 0) ◦ ρ)(a1), . . . , ((ρ ⊗ idH 0) ◦ ρ)(an)] × (ρ ⊗ id ⊗ idH 0)(z∗)(z∗ ⊗ 10). On the other hand, since (ρ ⊗ idH 0) ◦ ρ = (id ⊗ ∆0) ◦ ρ, ((id ⊗ ∆0) ◦ λ)([a1, . . . , an]q)u∗ = [((id ⊗ ∆0) ◦ ρ)(a1), . . . , ((id ⊗ ∆0) ◦ ρ)(an)] × (ρ ⊗ id ⊗ idH 0)(z∗)(z∗ ⊗ 10). Hence for any [a1, . . . , an]q ∈ X, ((λ ⊗ idH 0) ◦ λ)([a1, . . . , an]q) = ((id ⊗ ∆0) ◦ λ)([a1, . . . , an]q)u∗. Thus λ is a twisted coaction of H 0 on X with respect to (A, B, ρ, σ, u). (cid:3) Theorem 7.5. Let A be a unital C ∗-algebra and H a finite dimensional C ∗-Hopf algebra with its dual C ∗-Hopf algebra H 0. Let ρ be a coaction of H 0 on A with the Rohlin property. Let q be a full projection in a C ∗-algebra Mn(A) such that (ρ ⊗ idMn(C))(q) ∼ q ⊗ 10 in Mn(A) ⊗ H 0. Let B = qMn(A)q. Then there is a coaction of H 0 on B with the Rohlin property. Proof. By Lemmas 7.3 and 7.4, there is a twisted coaction (σ, u) such that (σ, u) is strongly Morita equivalent to ρ. By Corollary 6.4, (σ, u) has the Rohlin property. Furthermore, by [11, Theorem 9.6], there is a unitary ele- ment y ∈ B ⊗ H 0 such that (y ⊗ 10)(σ ⊗ idH 0)u(id ⊗ ∆0)(y∗) = 1B ⊗ 10 ⊗ 10. 38 KAZUNORI KODAKA AND TAMOTSU TERUYA Let σ1 = Ad(y) ◦ σ. Then σ1 is a coaction of H 0 on B with the Rohlin property by easy computations since σ1 is exterior equivalent to (σ, u). (cid:3) Let A be a UHF-algebra of type N ∞, where N is the dimension of a finite dimensional C ∗-Hopf algebra H. In [11], we showed that there is a coaction ρ of H 0 on A with the Rohlin property. Corollary 7.6. With the above notations, for any unital C ∗-algebra B, that is strongly Morita equivalent to A, there is a coaction σ of H 0 on B with the Rohlin property. Proof. By [16, Proposition 2.1] there are n ∈ N and a full projection q ∈ Mn(A) such that B is isomorphic to qMn(A)q. We identify B with qMn(A)q. Let ρ be a coaction of H 0 on A with the Rohlin property. Then [11, Lemma 10.10], (ρ ⊗ idMn(C))(q) ∼ q ⊗ 10 in Mn(A) ⊗ H 0 since A has cancellation. Therefore, by Theorem 7.5 we obtain the conclusion. (cid:3) Acknowledgement. The authors wish to thank the referee for valuable suggestions for improvement of the manuscript. References [1] B. Blackadar, K-theory for operator algebras, M. S. R. I. Publications 5, 2nd Edition, Cambridge Univ. Press, Cambridge, 1998. [2] R. J. Blattner, M. Cohen and S. Montgomery, Crossed products and inner actions of Hopf algebras, Trans. Amer. Math. Soc., 298 (1986), 671 -- 711. [3] F. Combes, Crossed products and Morita equivalence, Proc. London Math. Soc., 49 (1984), 289 -- 306. [4] R. E. Curto, P. S. Muhly and D. P. Williams, Cross products of strong Morita equivalent C ∗-algebras,, Proc. Amer. Math. Soc., 90 (1984), 528 -- 530. [5] M. Z. Guo and X. X. Zhang, Takesaki-Takai duality theorem in Hilbert modules, Acta Mth. Sin., Engl. Ser. 20 (2004) 1079-1088. [6] M. Izumi, Finite group actions on C ∗-algebras with the Rohlin property- I, Duke Math. J., 122 (2004) 233-280. [7] S. Jansen and S. Waldmann, The H-covariant strong Picard groupoid, J. Pure Appl. Algebra, 205 (2006), 542 -- 598. [8] T. Kajiwara and Y. Watatani, Jones index theory by Hilbert C ∗- bimodules and K-Thorey, Trans. Amer. Math. Soc., 352 (2000), 3429 -- 3472. THE STRONG MORITA EQUIVALENCE FOR COACTIONS 39 [9] T. Kajiwara and Y. Watatani, Crossed products of Hilbert C ∗- bimodules by countable discrete groups, Proc. Amer. Math. Soc., 126 (1998), 841 -- 851. [10] K. Kodaka and T. Teruya, Inclusions of unital C ∗-algebras of index- finite type with depth 2 induced by saturated actions of finite dimen- sional C ∗-Hopf algebras, Math. Scand., 104 (2009), 221 -- 248. [11] K. Kodaka and T. Teruya, The Rohlin property for coactions of fi- nite dimensional C ∗-Hopf algebras on unital C ∗-algebras, J. Operator Theory, 74 (2015), 101 -- 142, to appear. [12] E. C. Lance, Hilbert C ∗-modules, A toolkit for operatros algebraists, London Math. Soc. Lecture Note Series, 210, Cambridge Univ. Press, Cambridge, 1995. [13] H. Osaka, K. Kodaka and T. Teruya, The Rohlin property for inclusions of C ∗-algebras with a finite Watatani index, Operator structures and dynamical systems, 177-185, Contemp. Math., 503 Amer. Math. Soc., Providence, RI, 2009. [14] J. A. Packer, C ∗-algebras generated by projective representations of the discrete Heisenberg group, J. Operator Theory, 18 (1987), 41 -- 66. [15] I. Raeburn and D. P. Williams, Morita equivalence and continuous - trace C ∗-algebras, Mathematical Surveys and Monographs, 60, Amer. Math. Soc., 1998. [16] M. A. Rieffel, C ∗-algebras associated with irrational rotations, Pacific J. Math., 93 (1981), 415 -- 429. [17] W. Szyma´nski and C. Peligrad, Saturated actions of finite dimensional Hopf *-algebras on C ∗-algebras, Math. Scand., 75 (1994), 217 -- 239. [18] Y. Watatani, Index for C ∗-subalgebras, Mem. Amer. Math. Soc., 424 (1990). Department of Mathematical Sciences, Faculty of Science, Ryukyu Uni- versity, Nishihara-cho, Okinawa, 903-0213, Japan E-mail address: [email protected] Faculty of Education, Gunma University, 4-2 Aramaki-machi, Maebashi City, Gunma, 371-8510, Japan E-mail address: [email protected]
1608.03229	3	1608	2018-01-10T15:40:56	A Non-Commutative Unitary Analogue of Kirchberg's Conjecture	[ "math.OA" ]	The $C^{\ast}$-algebra $\mathcal{U}_{nc}(n)$ is the universal $C^{\ast}$-algebra generated by $n^2$ generators $u_{ij}$ that make up a unitary matrix. We prove that Kirchberg's formulation of Connes' embedding problem has a positive answer if and only if $\mathcal{U}_{nc}(2) \otimes_{\min} \mathcal{U}_{nc}(2)=\mathcal{U}_{nc}(2) \otimes_{\max} \mathcal{U}_{nc}(2)$. Our results follow from properties of the finite-dimensional operator system $\mathcal{V}_n$ spanned by $1$ and the generators of $\mathcal{U}_{nc}(n)$. We show that $\mathcal{V}_n$ is an operator system quotient of $M_{2n}$ and has the OSLLP. We obtain necessary and sufficient conditions on $\mathcal{V}_n$ for there to be a positive answer to Kirchberg's problem. Finally, in analogy with recent results of Ozawa, we show that a form of Tsirelson's problem related to $\mathcal{V}_n$ is equivalent to Connes' Embedding problem.	math.OA	math	A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE SAMUEL J. HARRIS Abstract. The C ∗-algebra Unc(n) is the universal C ∗-algebra generated by n2 generators uij that make up a unitary matrix. We prove that Kirchberg's formulation of Connes' embedding problem has a positive answer if and only if Unc(2) ⊗min Unc(2) = Unc(2) ⊗max Unc(2). Our results follow from properties of the finite-dimensional operator system Vn spanned by 1 and the generators of Unc(n). We show that Vn is an operator system quotient of M2n and has the OSLLP. We obtain necessary and sufficient conditions on Vn for there to be a positive answer to Kirchberg's problem. Finally, in analogy with recent results of Ozawa, we show that a form of Tsirelson's problem related to Vn is equivalent to Connes' embedding problem. Contents Introduction 1. Operator Systems and their Tensor Products 2. The OSLLP, WEP AND DCEP for Operator Systems 3. Complete Quotient Maps 4. The C ∗-algebra Unc(n) 5. Relating Vn to Kirchberg's conjecture 6. Unitary Correlation Sets References 1 2 8 9 10 20 21 28 Introduction One of the most significant open problems in the field of operator algebras is Connes' embedding problem [8], which asks whether every finite von Neumann algebra with separable predual can be embedded into the ultrapower of the hy- perfinite II1 factor in a trace-preserving way. One of the simpler formulations of the problem is known as Kirchberg's problem [19, Proposition 8], which asks whether or not C ∗(Fn) ⊗min C ∗(Fn) = C ∗(Fn) ⊗max C ∗(Fn) for some (equivalently all) n ≥ 2, where Fn is the free group on n generators. Due to recent work in [12], [15], and [21], another equivalent statement of the embedding problem is in terms of certain sets of quantum bipartite correlations (see [12], [15], [21] and [27] for more information on these correlations). Key words and phrases. Connes' Embedding Problem; Kirchberg's Conjecture; Operator Sys- tems; Unitary Correlation Sets. The author was supported in part by NSERC (Canada). 1 2 SAMUEL J. HARRIS One of our main results is that Kirchberg's problem, stated in terms of C ∗(Fn), is equivalent to the same problem when C ∗(Fn) is replaced by Brown's C ∗-algebra Unc(n), defined in [3] as the universal C ∗-algebra whose generators make up a unitary n × n matrix. In other words, Connes' embedding problem has a positive answer if and only if Unc(2) ⊗min Unc(2) = Unc(2) ⊗max Unc(2) (see Theorem 5.2). On the way to proving this equivalence, we obtain a new proof of Kirchberg's theorem [20, Corollary 1.2] that C ∗(Fn) ⊗min B(H) = C ∗(Fn) ⊗ B(H) for every Hilbert space H. Both of these results follow from properties of the finite- dimensional operator system Vn spanned by the generators of Unc(n). The signif- icance of stating Kirchberg's conjecture in terms of Unc(n) lies in recent advances in quantum information theory. Indeed, the phenomenon of embezzling entangle- ment in a bipartite scenario can be modelled using states on tensor products of Unc(n) (see [7] for more information on embezzlement of entanglement). Our study of the C ∗-algebra Unc(n) and states on the tensor products Unc(n)⊗min Unc(n) and Unc(n)⊗max Unc(n) allow for a theory of so-called "unitary correlation sets", which is motivated by their connection to quantum information theory. A problem in- volving unitary correlation sets that is analogous to Tsirelson's problem is shown to also be equivalent to Kirchberg's problem. The methods in this paper draw on many results in the recent theory of operator system tensor products and operator system quotients. In Section §1 we review some basic results in the theory of operator system tensor products, and in Section §2 we recall some nuclearity-related properties of operator systems that arise in equivalent formulations of Kirchberg's problem. Section §3 gives a short introduction to the theory of operator system quotients. In Section §4, we explore properties of the operator system Vn and the C ∗-algebra Unc(n) while giving al- ternate characterizations of the WEP and DCEP in terms of tensor products with Vn. We link both Vn and Unc(n) to Kirchberg's problem in Section §5. Finally, Section §6 draws on some results in quantum bipartite correlations in the field of quantum information theory, and an analogous theory of such correlations is developed in terms of the C ∗-algebra Unc(n). 1. Operator Systems and their Tensor Products In this section, we include a short introduction to operator systems and their tensor theory. The interested reader can see [17] for a thorough introduction to the subject. First, we give the abstract definition of an operator system, bearing in mind that concrete operator systems are always self-adjoint vector subspaces of B(H), for some Hilbert space H, which contain the identity element. ∗ : S → S such that for all x, y ∈ S and α ∈ C, Assume that S is a complex vector space. An involution on S is a map • (αx + y)∗ = αx∗ + y∗, and • (x∗)∗ = x. We denote by Sh the set of all x ∈ S with x = x∗. We call S a ∗-vector space if it is a complex vector space equipped with an involution. Whenever S is a ∗-vector space, there is a natural way to make Mn(S) into a ∗-vector space, where Mn(S) is A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 3 the space of all n× n matrices with entries in S; indeed, one may let (xij)∗ = (x∗ ji) for each (xij) ∈ Mn(S). Given a ∗-vector space S, a matrix ordering on S is a set of cones {Cn}∞ n=1, where Cn ⊆ (Mn(S))h, satisfying the following conditions: • Cn + Cn ⊆ Cn and tCn ⊆ Cn for all t ≥ 0 and n ∈ N, • A∗CnA ⊆ Cm for all n, m ∈ N and A ∈ Mn,m(C), and • Cn ∩ (−Cn) = {0} for all n ∈ N. A ∗-vector space S is said to be a matrix-ordered ∗-vector space if there is a matrix order {Cn}∞ We say that an element e ∈ Sh is an order unit if for every x ∈ Sh, there is r > 0 such that x + re ∈ C1. We call e an Archimedean order unit if whenever x ∈ Sh is such that x + re ∈ C1 for all r > 0, we have x ∈ C1. We call e a matrix n=1 on S. order unit if for each n ∈ N, In =  e . . .  ∈ Mn(S) is an order unit for e n=1 n=1, e) and (T ,{Dn}∞ Mn(S). Note that e is a matrix order unit if and only if it is an order unit for S with respect to the cone C1. We say that e is an Archimedean matrix order unit provided that each In is an Archimedean order unit. With this terminology in hand, we can define the abstract version of operator systems. An (abstract) operator system is a triple (S,{Cn}∞ n=1, e) where S is a ∗-vector space, {Cn}∞ is a matrix ordering on S, and e is an Archimedean matrix order unit for S. We will usually refer to S as an operator system, allowing the context to dictate which matrix ordering is being used. If (S,{Cn}∞ n=1, e) are operator systems with T ⊆ S, we will say that T is an operator subsystem of S provided that T has the same ∗-vector space structure as S and Dn = Cn ∩T for all n ∈ N. Given operator systems S and T and a number k ∈ N, we say that a linear map ϕ : S → T is k-positive provided that for all n ≤ k, (ϕ(xij)) ∈ Mn(T )+ whenever (xij) ∈ Mn(S)+. We say that ϕ is completely positive if it is k- positive for every k ∈ N. We use the abbreviaton "ucp" for "unital and completely positive". A ucp map ϕ : S → T is a complete order isomorphism provided that ϕ is a bijection and ϕ−1 : T → S is also ucp. We will also need the notion of an order isomorphism, which is a linear map ϕ : S → T between operator systems that is a bijection, such that ϕ and ϕ−1 are 1-positive. Finally, given a linear map ϕ : S → T between operator systems, we say that ϕ is a complete order embedding (or complete order injection) if there is an operator subsystem T1 ⊆ T such that ϕ(S) = T1 and ϕ : S → T1 is a complete order isomorphism. The following celebrated result of Choi and Effros ensures that there is no difference between considering abstract operator systems and concrete operator systems. Theorem 1.1. (Choi-Effros, [6]) Let S be an abstract operator system equipped with Archimedean matrix order unit e. Then there is a Hilbert space H and a com- plete order embedding ϕ : S → B(H) such that ϕ(e) = IH. Conversely, any oper- ator system contained in B(H) is an abstract operator system with Archimedean matrix order unit e = IH. 4 SAMUEL J. HARRIS into Mn, and we let S∞(S) =Sn∈N Sn(S). We will often use the term state to It will be useful to consider matricial states on an operator system S. By way of notation, for each n ∈ N, we let Sn(S) be the set of all ucp maps from S refer to elements of S1(S). pleted, give S an operator space structure. Indeed, if (S,{Cn}∞ ator system structure on S and X ∈ Mn(S), the norms Any operator system S has a sequence of matrix norms which, when com- n=1, e) is the oper- kXkn = inf(cid:26)r > 0 :(cid:18)rIn X rIn(cid:19) ∈ C2n(cid:27) X ∗ make S into a matricially normed space (see, for example, [22, Chapter 13]). Before moving on to tensor products, we require some information about duals of operator systems. The Banach space dual of an operator system S can be given the structure of a matrix-ordered ∗-vector space [6, Lemma 4.2, Lemma let S d denote the Banach space dual of 4.3]. This structure is given as follows: S. Given f ∈ S d, we define f ∗ ∈ S d by f ∗(x) = f (x∗) for x ∈ S, which makes S d into a ∗-vector space. We say an element (fij) ∈ Mn(S d) is positive provided that the map F : S → Mn given by F (x) = (fij(x)) is completely positive. If we also assume that S is finite-dimensional, then S d becomes an operator system with order unit given by a faithful state on S [6]. In fact, if S is finite-dimensional, then S dd and S are completely order isomorphic via the canonical map i : S → S dd. Definition 1.2. (Kavruk-Paulsen-Todorov-Tomforde, [17]) Let S and T be oper- ator systems. A collection of matricial cones τ = {Cn}∞ n=1 with Cn ⊆ Mn(S ⊗T )h is said to be an operator system structure on S ⊗ T if n=1, 1S ⊗ 1T ) is an operator system, (1) (S ⊗ T ,{Cn}∞ (2) (sij ⊗tkℓ) ∈ Cnm whenever (sij) ∈ Mn(S)+, (tkℓ) ∈ Mm(T )+ and n, m ∈ N, (3) whenever n, k ∈ N and ϕ ∈ Sn(S) and ψ ∈ Sk(T ), then ϕ⊗ψ ∈ Snk(S⊗T ) with respect to the collection of matricial cones τ = {Cn}∞ n=1. and We denote by S ⊗τ T the resulting operator system. Let O be the category of operator systems with ucp maps as the morphisms. Following the definitions in [17], we say that a mapping τ : O × O → O given by (S,T ) 7→ S ⊗τ T is an operator system tensor product if for all operator systems S and T , the matrix ordering on τ (S,T ) is an operator system structure on S ⊗ T . We say that an operator system tensor product τ is functorial if it satisfies the following property: • If S1 and T1 are operator systems and ϕ : S → S1 and ψ : T → T1 are ucp maps, then ϕ ⊗ ψ : S ⊗τ S1 → T ⊗τ T1 is ucp. Definition 1.3. (Kavruk-Paulsen-Todorov-Tomforde, [17]) Let S and T be op- erator systems. The minimal tensor product of S and T is the vector space S ⊗ T , with order unit 1 ⊗ 1, equipped with positive cones in Mn(S ⊗ T ) given by (S,T ) of all X ∈ Mn(S ⊗ T ) for which X = X ∗ and ϕ ⊗ ψ(X) ≥ 0 the set C min whenever ϕ ∈ S∞(S) and ψ ∈ S∞(T ). n A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 5 Definition 1.4. (Kavruk-Paulsen-Todorov-Tomforde, [17]) Given operator sys- tems S,T , the commuting tensor product of S and T is the vector space S ⊗ T with order unit 1 ⊗ 1, with positive cones C comm (S,T ) given by the follow- ing property: X ∈ Mn(S ⊗ T )h is in C comm (S,T ) if and only if whenever H is a Hilbert space and ϕ : S → B(H) and ψ : T → B(H) are ucp maps with commuting ranges, then ϕ · ψ(X) ≥ 0, where ϕ · ψ(s ⊗ t) := ϕ(s)ψ(t) for all s ∈ S and t ∈ T . Definition 1.5. (Kavruk-Paulsen-Todorov-Tomforde, [17]) Let S,T be operator systems. For each n ∈ N, define the set Dmax (S,T ) to be the set of all X ∈ Mn(S ⊗ T )h for which there exist S ∈ Mk(S)+, T ∈ Mm(T )+ and a linear map A : Ck ⊗ Cm → Cn such that n n n X = A(S ⊗ T )A∗. Then the maximal tensor product of S and T is defined to be the operator system (S ⊗ T , 1 ⊗ 1, C max (S,T )), where n C max n (S,T ) = {X ∈ Mn(S ⊗ T )h : ∀ε > 0, X + ε1 ∈ Dmax n (S,T )}. Each of min, c and max are functorial operator system tensor products [17]. For finite-dimensional operator systems, the min and max tensor products are dual to each other. Proposition 1.6. (Farenick-Paulsen, [11]) If S and T are finite-dimensional op- erator systems, then (S ⊗min T )d is completely order isomorphic to S d ⊗max T d, and (S ⊗max T )d is completely order isomorphic to S d ⊗min T d. Two more tensor products are of interest: the essential left and essential right tensor products. Definition 1.7. (Kavruk-Paulsen-Todorov-Tomforde, [17]) Let S,T be operator systems. Define the operator system S ⊗el T to be the operator system structure arising from the inclusion S ⊗ T ⊆ I(S) ⊗max T , where I(S) is the injective envelope of S. (See [14] or [22, Chapter 15] for more on injective envelopes.) Similarly, define S ⊗er T to be the operator system structure arising from the inclusion S ⊗ T ⊆ S ⊗max I(T ). The tensor products er and el are examples of asymmetric tensor products; in fact, the map s⊗t 7→ t⊗s induces a complete order isomorphism S⊗erT ≃ T ⊗elS and S ⊗el T ≃ T ⊗er S. Given two operator system tensor products α, β, we will write α ≤ β to mean that whenever S,T are operator systems, the identity map id : S ⊗β T → S ⊗α T is ucp. For example, we have the following (see [17]): min ≤ er, el ≤ c ≤ max . For operator system tensor products α, β, we say that an operator system S is (α, β)-nuclear if S ⊗α T = S ⊗β T for all operator systems T . Equivalently, S is (α, β)-nuclear if the identity map id : S ⊗α T → S ⊗β T is a complete order isomorphism for every operator system T . Frequently, the theory of operator system tensor products has been moti- vated by the theory of operator space tensor products. In particular, suppose that 6 SAMUEL J. HARRIS x y∗ X is an operator space contained in B(H). Then there is a canonical operator system that contains a completely isometric copy of X; namely, one may define which is equipped with the complete isometry X ֒→ SX given by sending x ∈ X µIH(cid:19) ∈ M2(B(H)) : x, y ∈ X, λ, µ ∈ C(cid:27) , SX =(cid:26)(cid:18)λIH 0 0(cid:19). See [22, Chapter 8] for more information on SX . It is left to the matrix(cid:18)0 x to the reader to check that SX does not depend on the embedding X ⊆ B(H), up to unital complete order isomorphism. For operator spaces X and Y , considering the copy of X ⊗ Y inside of SX ⊗ SY gives rise to operator space tensor products. For any operator spaces X, Y and operator system tensor product τ , there is a natural operator space tensor product, denoted by X ⊗τ Y , given by the inclusion of X ⊗ Y in SX ⊗τ SY . We will refer to the tensor product X ⊗τ Y as the induced operator space tensor product of X and Y (see [17]). Two important operator space tensor products are the injective tensor product and the projective tensor product, which we will denote by X ⊗Y and Xb⊗Y , respectively (see [2]). The relation between tensor products of operator systems of the form SX and operator spaces is outlined in the following theorem. Theorem 1.8. (Kavruk-Paulsen-Todorov-Tomforde, [17]) Let X, Y be operator spaces. The following are true: (1) X ⊗min Y = X ⊗Y completely isometrically. (2) X ⊗max Y = Xb⊗Y completely isometrically. erator system of an operator space X by letting Similar to the operator system SX , one may define another "canonical" op- X =(cid:26)(cid:18)λIH S 0 y∗ x λIH(cid:19) ∈ M2(B(H)) : x, y ∈ X, λ ∈ C(cid:27) . If τ is an operator system tensor product and X, Y are operator spaces, then we shall denote by X ⊗τ 0 Y the operator space structure on X ⊗ Y induced by the inclusion X ⊗ Y ⊆ S 0 X ⊗τ S 0 Y . The analogous result to Theorem 1.8 holds for S 0 as well. For completeness, we include the proofs. X ⊗ S 0 Lemma 1.9. Let X, Y be operator spaces. Then the inclusion X ⊗ Y ⊆ S 0 Y gives rise to an operator space tensor product of X and Y ; that is, the following conditions hold (in the sense of [2]): X (1) If x ∈ Mn(X) and y ∈ Mm(Y ), then kx ⊗ ykMnm(X⊗τ 0 Y ) ≤ kxkMn(X)kykMm(Y ). φ⊗ψ : X⊗τ (2) If φ : X → Mn and ψ : Y → Mm are completely bounded maps, then 0 Y → Mmn is completely bounded, with kφ⊗ψkcb ≤ kφkcbkφkcb. Y as an operator space yields Y . Since the Y are complete isometries, condition (1) follows Proof. By [17, Proposition 3.4], treating S 0 an operator space tensor product of the operator spaces S 0 X and S 0 inclusions X ⊆ S 0 since it holds for the operator space tensor product S 0 X ⊗τ S 0 Y . X and Y ⊆ S 0 X ⊗τ S 0 A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 7 To show that condition (2) holds, we may assume without loss of generality that φ and ψ are completely contractive. Let Φ : S 0 Φ(cid:18)(cid:18) λ x1 2 λ(cid:19)(cid:19) =(cid:18) λIn x∗ φ(x2)∗ λIn (cid:19) . X → M2(Mn) be defined by φ(x1) The proof that Φ is ucp is standard and analogous to [22, Lemma 8.1]. Similarly, the map Ψ : S 0 Y → M2(Mm) given by Ψ(cid:18)(cid:18) λ y1 2 λ(cid:19)(cid:19) =(cid:18) λIm ψ(y2)∗ y∗ ψ(y1) λIm(cid:19) is unital and completely positive. By Property (3) of operator system tensor products, Φ⊗ Ψ : S 0 Y → M4mn is ucp. Compressing to a corner block yields φ ⊗ ψ, so that kφ ⊗ ψkcb ≤ 1. X ⊗τ S 0 (cid:3) Since the minimal operator system tensor product and the spatial operator space tensor product are injective (and equal for operator systems), the following result is immediate. Theorem 1.10. Let X and Y be operator spaces. Then X ⊗min isometric to X ⊗Y . 0 Y is completely The analogous result also holds for the maximal operator system tensor Y is completely product. Theorem 1.11. Let X and Y be operator spaces. Then X ⊗max 0 0 0 0 isometric to Xb⊗Y . Y , and define kUk to be the norm of U in Xb⊗Y . First, suppose that kUk < let e =(cid:18)e1 Proof. Let U ∈ Mp(X ⊗ Y ). Define kUkmax to be the norm of U in S 0 X ⊗max S 0 1. Every operator system tensor product is an operator space tensor product [17, Proposition 3.4], so that k · kmax is smaller than the projective tensor product norm [2]. Hence, kUkmax < 1. Conversely, suppose that kUkmax < 1. As in the proof of [17, Theorem 5.9], Y , respectively; X ⊗max S 0 Y . Write U = (ur,s) ∈ Mp(X ⊗ Y ). Then U 0 then e ⊗ f is the identity of S 0 e2(cid:19) and f =(cid:18)f1 (cid:18)kUkmax(e ⊗ f )p By adding (1 − kUkmax)(cid:18)(e ⊗ f )p (cid:18)(e ⊗ f )p f2(cid:19) be the identities of S 0 kUkmax(e ⊗ f )p(cid:19) ∈ C max 2p (S 0 X ,S 0 Y ). (e ⊗ f )p(cid:19), it follows that (e ⊗ f )p(cid:19) ∈ Dmax X,S 0 2p (S 0 Y ). X )+, Q = (Qkℓ) ∈ Mm(S 0 This implies that there are matrices P = (Pij) ∈ Mn(S 0 B(cid:19) where A = (ar,(i,k)) and B = (br,(i,k)) are p×mn matrices of scalars, and T =(cid:18)A X and S 0 Y )+ U ∗ U ∗ U 0 0 8 such that SAMUEL J. HARRIS (cid:18)(e ⊗ f )p U ∗ U (e ⊗ f )p(cid:19) = T (P ⊗ Q)T ∗. Comparing blocks yields the equations (e⊗ f )p = A(P ⊗ Q)A∗, U = A(P ⊗ Q)B∗, U ∗ = B(P⊗Q)A∗ and (e⊗f )p = B(P⊗Q)B∗. We may write Pij =(cid:18)αije1 and Qkℓ =(cid:18)γkℓf1 αije2(cid:19) γkℓf2(cid:19) where αij, γkℓ ∈ C, xij, wij ∈ X, and ykℓ, zkℓ ∈ Y . Now w∗ ij z∗ kℓ ykℓ xij The fact that P, Q are positive implies that R and S are positive, (w∗ set R = (αij), S = (γkℓ), X = (xij) and Y = (ykℓ). ij) = X ∗, (z∗ kℓ) = Y ∗, and that for every r > 0, we have k(R + rIn)−1/2X (R + rIn)−1/2k ≤ 1 in Mn(X) and k(S + rIm)−1/2Y(S + rIm)−1/2k ≤ 1 in Mn(Y ) [22, p. 99]. Let Re1 := (αije1), and similarly define Re2, Sf1 and Sf2. Looking at the blocks of the 4 × 4 block matrix equation (e ⊗ f )p = A(P ⊗ Q)A∗, we obtain the equations (ei⊗fj)p = A(Rei⊗Sfj)A∗ for i, j = 1, 2. Therefore, Ip = A(R⊗S)A∗ = B(R ⊗ S)B∗. The element U is only present in the (1, 4)-block of the 4 × 4 block matrix, with the other blocks being equal to zero. Hence, the equation U = A(P ⊗ Q)B∗ in S 0 If R, S are invertible, then let A0 = A(R ⊗ S) U = A0(R⊗S)−1/2(X ⊗Y)(R⊗S)−1/2B∗ We know that A0A∗ S−1/2YS−1/2, we have U = A0(X0 ⊗ Y0)B∗ factorization having norm at most one. Therefore, kUk ≤ 1. If either of R or S are not invertible, then add rIn and rIm to R and S, respectively, for r > 0, and define new matrices A0 = A[(R + rIn) ⊗ (S + rIm)]1/2 and B0 = B[(R + rIn) ⊗ (S + rIm)]−1/2. The corresponding factorization yields kUk ≤ 1 + Cr for some C that is independent of r. Since this is possible for all r > 0, we obtain kUk ≤ 1. 2 )]B∗ 0. 0 = Ip. Thus, letting X0 = R−1/2X R−1/2 and Y0 = 0 with all the matrices appearing in this Y gives U = A(X ⊗ Y)B∗ in X ⊗ Y . 2 . Then 0 = A0[(R−1/2X R−1/2)⊗(S−1/2YS− 1 2 and let B0 = B(R ⊗ S) X ⊗ S 0 0 = B0B∗ (cid:3) 1 1 2. The OSLLP, WEP AND DCEP for Operator Systems There are two C ∗-algebraic properties that play a crucial role in Kirchberg's conjecture: the local lifting property (LLP) and the weak expectation property (WEP) (see [19]). Here we outline these properties for operator systems, as well as the double commutant expectation property (DCEP). For the convenience of the reader, we give some known characterizations of these properties in terms of tensor products with B(H) and tensor products with C ∗(F∞). See [18] for more information and proofs of the theorems in this section. Let S be an operator system. We say that S has the operator system local lifting property (OSLLP) if whenever A is a unital C ∗-algebra, I ⊆ A is a two-sided ideal with π : A → A/I the canonical quotient map, and ϕ : S → A/I is a ucp map, then for every finite-dimensional operator system T ⊆ S, there is a ucp map ψT : T → A such that π ◦ ψT = ϕT . This property for operator systems was initially defined in [18]. The OSLLP can be characterized in a few ways. A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 9 Theorem 2.1. (Kavruk-Paulsen-Todorov-Tomforde, [18]) Let S be an operator system. The following are equivalent. (1) S has the OSLLP. (2) S is (min, er)-nuclear. (3) S ⊗min B(H) = S ⊗max B(H) for all Hilbert spaces H. Also defined in [18], we say that an operator system S has the weak ex- pectation property (WEP) if the canonical inclusion i : S ֒→ S dd extends to a ucp map φ : I(S) → S dd, where I(S) is the injective envelope of S. The WEP is in fact a form of nuclearity. Theorem 2.2. (Kavruk-Paulsen-Todorov-Tomforde, [18]) An operator system S has the WEP if and only if S is (el, max)-nuclear. Finally, we say that an operator system S has the double commutant expectation property [18] (DCEP), if for any unital complete order embedding ι : S → B(H), there is a ucp extension ϕ : I(S) → ι(S)′′ of ι, where I(S) is the injective envelope of S and ι(S)′′ is the double commutant of ι(S) in B(H). The DCEP is a weaker form of nuclearity than the WEP. Theorem 2.3. (Kavruk-Paulsen-Todorov-Tomforde, [18]) Let S be an operator system. The following are equivalent. (1) S has the DCEP. (2) S is (el, c)-nuclear. (3) S ⊗min C ∗(F∞) = S ⊗max C ∗(F∞). All unital C ∗-algebras are (c, max)-nuclear [17, Theorem 6.7], so that the WEP and the DCEP are the same for unital C ∗-algebras. However, the DCEP is a weaker property in general. Indeed, the operator system Tn of tridiagonal matrices in Mn is (min, c)-nuclear but Tn ⊗min T d n for n ≥ 3 (see [17] for more information on Tn). Thus, Tn has the DCEP but not the WEP. n 6= Tn ⊗max T d 3. Complete Quotient Maps The theory of operator system quotients is not as straightforward as in other categories. Here, we give a very brief introduction to this quotient theory. Much more information on operator system quotients can be found in [18]. Suppose that S,T are operator systems and ϕ : S → T is a ucp map with kernel J . We endow the quotient vector space S/J with an operator system structure as follows. If q : S → S/J denotes the canonical (vector space) quotient map and we denote by x the image q(x) of a vector x ∈ S, then the order unit of S/J is 1, while the adjoint of x is simply x∗. We define the sets Dn(S,J ) = { X ∈ Mn(S/J )h : ∃Y ∈ Mn(S)+ such that q(n)(Y ) = X}, where q(n) is the n-fold amplification of q. To ensure that the positive cones on S/J satisfy the Archimedean property, we define the cones to be Cn(S,J ) = { X ∈ Mn(S/J )h : ∀ε > 0, X + ε1 ∈ Dn(S,J )}. 10 SAMUEL J. HARRIS In general, the norm induced by the above operator system structure on S/J can differ greatly from the usual quotient norm on S/J (see [18, Example 4.4] or [11, Lemma 2.1] for examples). Given operator systems S,T and a u.c.p. map ϕ : S → T with kernel J , we say that J is completely order proximinal if Dn(S,J ) = Cn(S,J ) for all n. A surjective ucp map ϕ : S → T with kernel J is said to be a complete quotient map if the induced map ϕ : S/J → T given by ϕ( x) = ϕ(x) is a complete order isomorphism. There is a relation between complete quotient maps and complete order injections via adjoint maps. For a linear map ϕ : S → T , we define the adjoint map ϕd : T d → S d by [ϕd(ψ)](s) = ψ(ϕ(s)) for all ψ ∈ T d and s ∈ S. Proposition 3.1. (Farenick-Paulsen, [11]) Let S and T be finite-dimensional op- erator systems. Then a linear map ϕ : S → T is a complete quotient map if and only if ϕd : T d → S d is a complete order injection. 4. The C ∗-algebra Unc(n) For any n ∈ N, denote by Unc(n) the universal C ∗-algebra of n2 generators i,j=1 with the restriction that the matrix U = (uij) is unitary. This algebra {uij}n was first defined by L. Brown in [3]. We may define the operator subsystem Vn = span ({1} ∪ {uij}n i,j=1 ∪ {u∗ ij}n i,j=1). The operator system possesses an important universal property. Tij (√I − T ∗T )ij (√I − T T ∗)ij Proposition 4.1. Let (Tij) ∈ Mn(B(H)) be a contraction. Then there is a unique ucp map ψ : Vn → B(H) such that ψ(uij) = Tij for all 1 ≤ i, j ≤ n. The ucp map ψ dilates to a unital ∗-homomorphism π : Unc(n) → M2(B(H)) such that π(uij) =(cid:18) (cid:19) for all 1 ≤ i, j ≤ n. Proof. Let T = (Tij); then kTk ≤ 1. The operator V =(cid:18) (√I − T T ∗)ij √I − T T ∗ −T ∗ (cid:19) is unitary in M2(Mn(B(H))). Performing a canonical shuffle (see [22, p. 97]) yields a unitary W = (Wij) ∈ Mn(M2(B(H))) such that −T ∗ Tij ji T √I − T ∗T (cid:19) . Wij =(cid:18) (√I − T ∗T )ij −T ∗ ji By the universal property of Unc(n), there is a unital ∗-homomorphism π : Unc(n) → M2(B(H)) with π(uij) = Wij for all i, j. Compressing to the (1, 1)-corner in M2(B(H)) yields the ucp map ψ, as desired. (cid:3) We remark that Vn is the image of a unital, completely positive map on M2n. Indeed, define ϕ : M2n → Vn by ϕ(Eij) = 1 2n 1 if i = j 1 2n ui,j−n if i ≤ n and j ≥ n + 1 2n u∗ j,i−n if i ≥ n + 1 and j ≤ n 0 otherwise. 1 A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 11 1 2n U 1 n ∪Λ− 2n I 1 2n U ∗ Then the Choi matrix of ϕ is (ϕ(Eij)) = (cid:18) 1 Then J2n = (cid:26)(cid:18)A 0 2n I(cid:19). Since U ∗U = I, 0 B(cid:19) ∈ M2(Mn) : tr(A ⊕ B) = 0(cid:27). It is readily checked that (ϕ(Eij)) ∈ M2(Mn(Vn))+, so that ϕ is unital and completely positive by a the- orem of Choi (see [22, Theorem 3.14]). We will denote by J2n the kernel of ϕ. To simplify notation, we define for each n ≥ 2 the index sets Λ+ J2n contains no positive element except 0. It follows by [16, Proposition 2.4] that J2n is completely order proximinal. n = {(i, j) ∈ {1, ..., 2n}2 : i ≤ n, j ≥ n + 1} and Λ− n = {(i, j) ∈ {1, ..., 2n}2 : i ≥ n + 1, j ≤ n}. We define Λn = Λ+ n . Using the notation of [11], whenever i, j ∈ {1, ..., 2n} are such that ϕ(Eij) 6= 0, we define eij = Eij ∈ M2n/J2n. We let q : M2n → M2n/J2n be the canonical quotient map. To show that Vn is a complete quotient of M2n, we need some equivalent characterizations of positivity in the quotient operator system M2n/J2n. This result and its proof are analogous to [11, Proposition 2.3]. The key difference is the use of the universal property of Vn in the proof that (6) implies (5) below. Lemma 4.2. Let A11, Aij ∈ Mp for every (i, j) ∈ Λn. The following are equivalent. eij ⊗ Aij is positive in (M2n/J2n) ⊗ Mp. ψ(eij)⊗ Aij is positive in Mr ⊗ Mp whenever r ∈ N and ψ(Eij) ⊗ Aij is positive in Mr ⊗ Mp whenever r ∈ N n and the matrix B = (Bi,j+n) ∈ Mn(Mr) (1) 1 ⊗ A11 +P(i,j)∈Λn (2) 1⊗ A11 +P(i,j)∈Λn ψ : M2n/J2n → Mr is ucp. (3) 1 ⊗ A11 +P(i,j)∈Λn and ψ : M2n → Mr is ucp with ψ(J2n) = {0}. are such that (4) Whenever Bij ∈ Mr for (i, j) ∈ Λ+ B∗ ji ⊗ Aij ∈ (Mr ⊗ Mp)+. n and the matrix C = (Ci,j+n) ∈ Mn(Mr) n n is such that (cid:18) 1 Ir ⊗ A11 + X(i,j)∈Λ+ (5) Whenever Cij ∈ Mr for (i, j) ∈ Λ+ 2n Ir B 1 B∗ is positive in M2(Mn(Mr)), then 2n Ir(cid:19) Bij ⊗ Aij + X(i,j)∈Λ− (cid:18) Ir C Ir(cid:19) Cij ⊗ Aij + X(i,j)∈Λ− Ir ⊗ A11 + X(i,j)∈Λ+ (6) Ir ⊗ A11 +P(i,j)∈Λ+ 2n ui,j−n ⊗ Aij +P(i,j)∈Λ− i=1 Eii ⊗ Bi +P(i,j)∈Λn (7) Ir ⊗ (2nA11) +P2n for some matrices B1, ..., B2n ∈ Mp such thatP2n C ∗ is positive in M2(Mn(Mr)), then Vn ⊗ Mp. 1 2n 1 2n 1 n n C ∗ ji ⊗ Aij ∈ (Mr ⊗ Mp)+. n 1 n 2n u∗ j,i−n ⊗ Aij is positive in Eij ⊗ Aij is positive in M2n ⊗ Mp i=1 Bi = (2n − 4n2)A11. 12 SAMUEL J. HARRIS Eij ⊗ Rij is positive in M2n ⊗ Mp for some ma- trix R = (Rij) ∈ (M2n ⊗ Mp)+ such that Rij = Aij for (i, j) ∈ Λn and i=1 Eii ⊗ Rii +P(i,j)∈Λn (8) P2n P2n i=1 Rii = 2nA11. Proof. Suppose that (1) holds. Since J2n is completely order proximinal, there i,j=1 Eij ⊗ Rij ∈ (M2n ⊗ Mp)+ such that q ⊗ idp(R) = eij ⊗ Aij. It follows that eij⊗Aij = q⊗idp 2nXi,j=1 exists a matrix R = P2n 1 ⊗ A11 +P(i,j)∈Λn Eij ⊗ Rij! = 2nXi=1 1⊗A11+ X(i,j)∈Λn Therefore, Rij = Aij whenever (i, j) ∈ Λn, and P2n Rii = 2nA11 + Bi where Bi = −Pj6=i Rjj. Then 2nA11 = Rii = 4n2A11 + 2nXi=1 Hence,P2n i=1 Bi = (2n − 4n2)A11 and (7) follows. If (7) is true, then applying ϕ ⊗ idp shows that that (8) is true. Assume that (8) is true, and let R = (Rij) ∈ (M2n ⊗ Mp)+ be as given. Write 1 2n eii⊗Rii+ X(i,j)∈Λn eij⊗Rij. i=1 Rii = 2nA11, which shows 2nXi=1 Bi. u∗ j,i−n⊗ Aij ∈ (Vn⊗ Mp)+. that n n 1 2n 1 i=1 2nXi=1 1⊗ (2nA11) + 1⊗ Bi + X(i,j)∈Λ+ ui,j−n⊗ Aij + X(i,j)∈Λ− 2n 1 ⊗(cid:0)P2n ui,j−n ⊗ Aij + X(i,j)∈Λ− 2n 1 ⊗ Bi = 1 i=1 Bi(cid:1) = (1 − 2n)1 ⊗ A11 shows Using the fact that P2n 1 ⊗ A11 + X(i,j)∈Λ+ Ir(cid:19) ∈ (M2(Mn(Mr)))+, where C = (Ci,j+n) ∈ Mn(Mr). Then C is a C ∗ contraction in Mn(Mr). By Proposition 4.1, there is a ucp map ψ : Vn → Mr such that ψ(uij) = Ci,j+n for all 1 ≤ i, j ≤ n. Applying ψ ⊗ idp to the positive element in (6) shows that Suppose that (6) is true, and let Cij ∈ Mr for (i, j) ∈ Λ+ u∗ j,i−n ⊗ Aij ∈ (Vn ⊗ Mp)+. (cid:18) Ir C Thus, (7) implies (6). n be such that n n Ir ⊗ A11 + X(i,j)∈Λ+ n 1 2n Cij ⊗ Aij + X(i,j)∈Λ− n 1 2n C ∗ ji ⊗ Aij ∈ (Mr ⊗ Mp)+. Therefore, (5) is true. such that(cid:18) 1 Assume (5). Let Bij ∈ Mr for (i, j) ∈ Λ+ 2n Ir(cid:19) is in (M2(Mn(Mr)))+. Let Cij = 2nBij, so that B = 1 n and B = (Bi,j+n) ∈ Mn(Mr) be 2n C, 2n Ir B 1 B∗ where C = (Ci,j+n). Then (5) immediately implies (4). A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 13 Suppose that (4) holds, and let ψ : M2n → Mr be a ucp map such that 2n Ir for n , then the Choi matrix of ψ is If Bij = ψ(Eij) for (i, j) ∈ Λ+ (cid:18) 1 ψ(J2n) = {0}g. Since Eii − Ejj ∈ J2n for all i 6= j, we have ψ(Eii) = 1 all 1 ≤ i ≤ 2n. 2n Ir B 1 B∗ 2n Ir(cid:19), and hence must be positive. Then (3) follows from (4). If (3) is true and ψ : M2n/J2n → Mr is ucp, then ψ := ψ ◦ q : M2n → Mr is ucp and annihilates J2n. This shows that (2) holds. Finally, suppose that (2) is true. Let h = 1 ⊗ A11 +P(i,j)∈Λn eij ⊗ Aij. Note that an element x of (M2n/J2n) ⊗ Mp is positive if and only if whenever r ∈ N and γ : (M2n/J2n) ⊗ Mp → Mr is ucp, then γ(x) ∈ (Mr)+ (see [6]). Now, if γ : (M2n/J2n) ⊗ Mp → Mr is ucp, then we may find linear maps V1, ..., Vm : Cp → Cr ⊗ Cp and ucp maps ψ1, ..., ψm : M2n/J2n → Mr such that γ = mXi=1 i ( ψi ⊗ idp(·))Vi. V ∗ Applying (2), we see that γ(h) ∈ (Mr)+ for each ucp map γ : (M2n/J2n) ⊗ Mp → Mr and for each r ∈ N. Thus, h is positive in (M2n/J2n) ⊗ Mp. Therefore, (1) follows from (2), as desired. (cid:3) Theorem 4.3. For ϕ : M2n → Vn and for J2n as above, the following are true: (1) The map ϕ : M2n → Vn is a complete quotient map; i.e., M2n/J2n is completely order isomorphic to Vn. (2) The C ∗-envelope of Vn is Unc(n). Proof. The proof is similar to the proof of [11, Theorem 2.4]. Since ϕ is a sur- jection, the map ϕ : M2n/J2n → Vn given by ϕ( x) = ϕ(x) is ucp and a linear bijection. Using the fact that statements (1) and (6) are equivalent in Lemma 4.2, we see that ϕ is a complete order isomorphism, which proves the first statement. For the second statement, we will show that Unc(n) satisfies the univer- e (Vn) (see, for example, [13]). Let A be any unital C ∗-algebra sal property of C ∗ equipped with a unital complete order embedding ι : Vn → A such that C ∗(ι(Vn)) = A. We assume that Unc(n) is represented faithfully on some Hilbert space H. The identity map id : Vn → Vn ⊆ Unc(n) can be written as id = κ ◦ ι, where κ : ι(Vn) → Vn is the ucp inverse of ι. We extend κ to a ucp map ρ : A → B(H) by Arveson's extension theorem [1]. Let ρ = V ∗π(·)V be a minimal Stinespring representation of ρ on some Hilbert space Hπ = ran(V ) ⊕ ran(V )⊥. With respect to this decomposition, for all 1 ≤ i, j ≤ n, we have π(ι(uij)) =(cid:18)uij ∗ ∗(cid:19) . ∗ The matrix π(n) ◦ ι(n)(U) = (π ◦ ι(uij))n looks like i,j=1, after applying the canonical shuffle, (cid:18)U ∗ ∗ ∗(cid:19) . 14 SAMUEL J. HARRIS Since U is unitary and π ◦ ι is completely contractive, the (1, 2) and (2, 1) blocks must be 0. By applying the inverse shuffle, it follows that for all i, j, we have π(ι(uij)) =(cid:18)uij 0 ∗(cid:19) . 0 Equivalently, S has property Vn if and only if the above holds when replacing the above positive element of Unc(n) ⊗min Mp(S) with 1 ⊗ S11 + X(i,j)∈Λn eij ⊗ Sij ∈ (M2n/J2n ⊗min Mp(S))+. We will say that S has property V if it has property Vn for every n ∈ N. These properties were inspired by the similar notion of operator systems having property Thus, ρ is multiplicative on the generators {ι(uij)}n i,j=1 of A, so that ρ is a ∗- homomorphism with ρ(ι(uij)) = uij for all i, j. This shows that ρ is surjective from A onto Unc(n). By the universal property of C ∗-envelopes, we conclude that C ∗ e (Vn) = Unc(n). (cid:3) Mn. n → M d i,j=1 is the dual basis for M d i,j=1 for M2n, then M2n is completely order isomorphic to M d Using the fact that M2n/J2n ≃ Vn allows for a description of the dual of Vn. n of Vn is completely order isomorphic Corollary 4.4. The operator system dual V d to S 0 Proof. We use the same argument as in [11, Proposition 2.7]. Since ϕ : M2n → Vn is a complete quotient map, ϕd : V d 2n is a complete order embedding by Proposition 3.1. If {δij}2n 2n of the canonical basis {Eij}2n 2n via the map- ping Eij 7→ δij [23, Theorem 6.2]. Taking the unit of M d 2n to be the canonical normalized trace, this mapping is a unital complete order isomorphism. It follows that the vector space dual of M2n/J2n, equipped with the operator system struc- ture inherited from M2n, is the operator system dual of Vn. It is not hard to see that the vector space dual of M2n/J2n is the annihilator of J2n in M d 2n. Therefore, V d n ≃ S 0 (cid:3) Mn. We will now move towards an analogue of Kirchberg's Theorem for Unc(n). Kirchberg's famous result on the full group C ∗-algebra of the free group Fn, for n ≥ 2, is that C ∗(Fn) ⊗min B(H) = C ∗(Fn) ⊗max B(H) for every Hilbert space H. We will show that a similar result is true when replacing C ∗(Fn) by Unc(n). First, we adopt some terminology using Lemma 4.2. We say that an operator system S has property Vn if whenever p ∈ N and S11, Sij ∈ Mp(S) for (i, j) ∈ Λn are such that 1 2n u∗ j,i−n ⊗ Sij ∈ (Unc(n) ⊗min Mp(S))+, 1 ⊗ S11 + X(i,j)∈Λ+ n 1 2n ui,j−n ⊗ Sij + X(i,j)∈Λ− n then for each ε > 0 there exist Rε • The matrix Rε = (Rε • Rε ij = Sij for all (i, j) ∈ Λn. i=1 Rε ii = 2n(S11 + ε1Mp(S)). • P2n ij ∈ Mp(S) for 1 ≤ i, j ≤ 2n such that ij) is positive in M2n(Mp(S)). A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 15 Wn+1 with regards to the operator system Wn+1 ⊆ C ∗(Fn) given by Wn+1 = span {wiw∗ j : 1 ≤ i, j ≤ n + 1}, where w2, ..., wn+1 are the generators of Fn and w1 = 1 (see [11]). Lemma 4.2 shows that Mp has property V for every p ∈ N. The operator systems satisfying property Vn are characterized in the following proposition. Proposition 4.5. Let S be an operator system. Then S has property Vn if and only if Vn ⊗min S = Vn ⊗max S. In particular, if S is (min, max)-nuclear, then S has property V. Proof. We proceed as in the proof of [11, Proposition 3.3]. Let X = (xkℓ) ∈ Mr(Vn ⊗ S) for r > 1; then xkℓ = 1 ⊗ s(kℓ) 11 + X(i,j)∈Λ+ n 1 2n ui,j−n ⊗ s(kℓ) ij + X(i,j)∈Λ− n 1 2n j,i−n ⊗ s(kℓ) u∗ ij , so if S11 = (s(kℓ) 11 ) and Sij = (s(kℓ) ij ), then we obtain X = 1 ⊗ S11 + X(i,j)∈Λ+ n 1 2n ui,j−n ⊗ Sij + X(i,j)∈Λ− n 1 2n u∗ j,i−n ⊗ Sij, where S11, Sij ∈ Mr. Now, S satisfies the definition of property Vn when p = r if and only if Mr(S) satisfies property Vn when p = 1. Moreover, we have Mr(Vn ⊗min S) = Vn ⊗min Mr(S) and Mr(Vn ⊗max S) = Vn ⊗max Mr(S). By replacing S with Mr(S) if necessary, in order to check that S has property Vn, it suffices to show that S satisfies the definition of property Vn when p = 1. Suppose that Vn ⊗min S = Vn ⊗max S, and suppose that x := 1 ⊗ S11 + X(i,j)∈Λ+ n 1 2n ui,j−n ⊗ Sij + X(i,j)∈Λ− n 1 2n u∗ j,i−n ⊗ Sij ∈ (Unc(n) ⊗min S)+. Then x ∈ (Vn⊗maxS)+. Hence, for every ε > 0, x + ε(1⊗ 1S ) ∈ Dmax (Vn,S). This means that there is V ∈ Mk(Vn)+, S ∈ Mm(S)+ and a linear map A : Ck⊗Cm → C such that 1 Since ϕ : M2n → Vn is a complete quotient map, there is R ∈ Mk(M2n)+ such that x + ε(1 ⊗ 1S) = A(V ⊗ S)A∗. x + ε(1 ⊗ 1S) = A(ϕ(R) ⊗ S)A∗. So, with Rε = A(R ⊗ S)A∗ ∈ M2n(S)+, we have ϕ ⊗ idS(Rε) = x + ε(1 ⊗ 1S). That is to say, for each ε > 0, there is Rε ∈ M2n(S)+ such that x + ε(1 ⊗ 1S) = 1 ⊗ Rε ii + X(i,j)∈Λ+ n 1 2n ui,j−n ⊗ Rε ij + X(i,j)∈Λ− n 1 2n u∗ j,i−n ⊗ Rε ij. Comparing coefficients with the coefficients of x + ε(1 ⊗ 1S) shows that Rε for (i, j) ∈ Λn and 1 that Cmin ij = Sij ii = S11 + ε1. This shows that S has property Vn. Conversely, suppose that S has property Vn and let p ∈ N; we must show (Vn,S). As before, by replacing S with Mr(S) if necessary, (Vn,S) ⊆ Cmax i=1 Rε p p 2nXi=1 2nP2n 16 SAMUEL J. HARRIS we may assume that p = 1. Let x ∈ (Vn ⊗min S)+. Then there are s11, sij ∈ S for (i, j) ∈ Λn such that x = 1 ⊗ s11 + X(i,j)∈Λ+ n 1 2n ui,j−n ⊗ sij + X(i,j)∈Λ− n 1 2n u∗ j,i−n ⊗ sij. Since S has property Vn, given ε > 0, there are Rε ij ∈ S for 1 ≤ i, j ≤ 2n such that Rε = (Rε ii = 2n(s11 + ε1S). If ϕ : M2n → Vn is the complete quotient map given as before, then the map ϕ ⊗ idS : M2n ⊗max S → Vn ⊗max S is ucp, and ij = sij for all (i, j) ∈ Λn andP2n ij) ∈ Mn(S)+, Rε i=1 Rε ϕ ⊗ idS(Rε) = x + ε(1 ⊗ 1S) ∈ (Vn ⊗max S)+. Thus, x + ε(1 ⊗ 1S) ∈ Dmax 1 which completes the proof. (Vn,S) for all ε > 0. Therefore, x ∈ (Vn ⊗max S)+, (cid:3) The next fact about tensor products of Vn is very useful. p Proposition 4.6. Let S be any operator system. For all n ≥ 2, the inclusion Vn ⊗c S ⊆ Unc(n) ⊗max S is a complete order embedding. Proof. By [17, Theorem 6.7], A ⊗c S = A ⊗max S for every unital C ∗-algebra A. We must show that C comm (Unc(n),S) ∩ Mp(Vn ⊗ S) for each p ∈ N. The inclusion map ιn : Vn → Unc(n) is ucp, so by functoriality of the commuting tensor product, ιn ⊗ idS : Vn ⊗c S → Unc(n) ⊗max S is ucp. Therefore, C comm (Unc(n),S)∩ Mp(Vn ⊗S). Let ψ : Vn → B(H) and γ : S → B(H) be ucp maps with commuting ranges. Let T = (Tij) where Tij = ψ(uij). By Proposition 4.1, the map ψ dilates to a unital ∗-homomorphism π : Unc(n) → M2(B(H)) such that, for all 1 ≤ i, j ≤ n, (Vn,S) ⊆ C comm Conversely, suppose that X ∈ C comm (Unc(n),S) ∩ Mp(Vn ⊗ S) for each p. (Vn,S) = C comm p p p p π(uij) =(cid:18) Tij (√I − T ∗T )ij (√I − T T ∗)ij −T ∗ ji (cid:19) . 0 0 γ(s)(cid:19) . We extend ψ to a ucp map on all of Unc(n) by letting ψ(x) be given by the (1, 1) Since γ(s) commutes with each Tij and T ∗ eγ(s) =(cid:18)γ(s) corner of π(x), for each x ∈ Unc(n). Defineeγ : S → M2(B(H)) by setting ij, we see that γ(s)⊗IH commutes with T and T ∗. Hence, γ(s)⊗ IH commutes with C ∗(IH, T, T ∗), which contains √I − T ∗T and √I − T T ∗. Thus, γ(s) ⊗ IH commutes with each block of π(uij). It follows that the range of eγ commutes with each π(uij). Since π is a ∗-homomorphism, the ucp maps π and eγ must have commuting ranges. Therefore, the map π ·eγ : Unc(n) ⊗c S → M2(B(H)) is ucp. Compressing to the (1, 1) corner in M2(B(H)) yields the map ψ · γ. It follows that (ψ · γ)Vn⊗S is ucp on the inclusion of Vn ⊗ S into Unc(n) ⊗c S. Hence, (ψ · γ)(p)(X) ∈ Mp(B(H))+. As ψ and γ were arbitrary, we conclude that X ∈ C comm (Vn,S). (cid:3) p A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 17 Lemma 4.7. Let S be any operator system. Then Vn ⊗min S = Vn ⊗c S if and only if Unc(n) ⊗min S = Unc(n) ⊗max S. In particular, if S has property Vn, then Unc(n) ⊗min S = Unc(n) ⊗max S. Proof. Suppose that Unc(n)⊗minS = Unc(n)⊗max S. Since the min tensor product is injective, Vn ⊗min S is completely order isomorphic to the image of Vn ⊗ S in Unc(n) ⊗min S. By Proposition 4.6, Vn ⊗c S is completely order isomorphic to its image in Unc(n) ⊗max S. It follows that Vn ⊗min S = Vn ⊗c S. Conversely, suppose that Vn ⊗min S = Vn ⊗c S. We employ an argument analogous to the proof of [11, Proposition 3.6]. Let X ∈ Mp(Unc(n) ⊗min S)+, and let ψ : Unc(n) → B(H) and γ : S → B(H) be ucp maps with commuting ranges. Let ψ = V ∗π(·)V be a minimal Stinespring representation for ψ on some Hilbert space Hπ. By Arveson's commutant lifting theorem [1, Theorem 1.3.1], there is a unital ∗-homomorphism ρ : (ψ(Unc(n)))′ → (π(Unc(n)))′ such that ρ(T )V = V T for all T ∈ (ψ(Unc(n)))′. Since ψ and γ have commuting ranges, we see that eγ = ρ ◦ γ : S → (π(Unc(n)))′ ⊆ B(K) is ucp and its range commutes with the range of π. Since Vn ⊗min S = Vn ⊗c S, the map (π ·eγ)Vn⊗minS is ucp. of (π ·eγ)Vn⊗minS. For any 1 ≤ i, j ≤ n, we see that Since the min tensor product is injective, Vn ⊗min S is completely order e (S). Arveson's extension e (S) → B(K) isomorphic to the image of Vn ⊗ S in Unc(n) ⊗min C ∗ theorem [1] guarantees existence of a ucp extension η : Unc(n)⊗min C ∗ Thus, {uij : 1 ≤ i, j ≤ n} is in the multiplicative domain Mη of η. It follows that Unc(n) ⊗ 1 ⊆ Mη. Hence whenever a ∈ Unc(n) and s ∈ S, we obtain η(a ⊗ s) = η((a ⊗ 1) η(uij ⊗ 1) = π ·eγ(uij ⊗ 1) = π(uij). {z } (1 ⊗ s)) = η(a ⊗ 1)η(1 ⊗ s) = π(a)eγ(s). V ∗π(a)eγ(s)V = V ∗π(a)ρ(γ(s))V = V ∗π(a)V γ(s) = ψ(a)γ(s). ψ · γ(a ⊗ s) = ψ(a)γ(s) = V ∗π(a)eγ(s)V = V ∗η(a ⊗ s)V. Now, the upper-left corner of π(a)eγ(s) is Using this fact, we have So, for all z ∈ Unc(n)⊗min S we have ψ · γ(z) = V ∗η(z)V , so that (ψ · γ)Unc(n)⊗minS is ucp. Therefore, (ψ · γ)(n)(X) ∈ Mp(Mm)+ so that Unc(n) ⊗min S = Unc(n) ⊗c S. ∈Mη (cid:3) Lemma 4.8. If H is any Hilbert space, then B(H) has property V. Equivalently, Vn has the OSLLP for every n ≥ 2. Proof. The proof proceeds in a similar manner to the proof of [11, Proposition 3.5]. Let n ∈ N and let S11, Sij ∈ B(H) for (i, j) ∈ Λn. Suppose that 1 ⊗ S11 + X(i,j)∈Λ+ n 1 2n ui,j−n ⊗ Sij + X(i,j)∈Λ− n 1 2n u∗ j,i−n ⊗ Sij ∈ (Unc(n) ⊗min B(H))+. As in the proof of Proposition 4.5, we may assume that p = 1. The matrix 0 ∈ Mn is a contraction, so by Proposition 4.1, the map α : Vn → C given by 18 SAMUEL J. HARRIS α(uij) = 0 and α(1) = 1 extends to a ucp map on Unc(n). Hence, α ⊗ idB(H) is ucp on Unc(n) ⊗min B(H), which forces S11 ≥ 0. Fix ε > 0. For any finite- dimensional subspace M of H, we know that B(M) has property V. Let PM denote the orthogonal projection onto M. Replacing S11 with PMS11PM and Sij with PMSijPM, we may find Rε,M ij ∈ B(M) such that ij ) is in (M2n(B(M)))+, ij = PMSijPM for (i, j) ∈ Λn, and i=1 Rε,M • Rε,M := (Rε,M • Rε,M • P2n ClearlyP2n ii ii = 2n(PMS11PM + εIM) = 2n(PMS11PM + εPM). i=1 Rε,M ii ≤ 2n(S11 + εIH), so since each Rε,M is positive, the di- agonal blocks of Rε,M are bounded. Since M is finite-dimensional and Rε,M ≥ 0, the norm of Rε,M is given by the largest eigenvalue. Therefore, indexing finite- dimensional subspaces of H by inclusion, the net (Rε,M)M≤H, dim(M)<∞ is uni- formly bounded. Let Rε be a w∗-limit point of the net (Rε,M)M. Then the It follows that if corresponding subnet of (PM)M converges strongly to IH. ij) ∈ Mn(B(H)), then Rε ≥ 0, while Rε Rε = (Rε ij = Sij for all (i, j) ∈ Λn and ii = 2n(S11 + εIH). Therefore, B(H) has property Vn for every n ∈ N. (cid:3) Theorem 4.9. Unc(n) has the LLP; i.e., Unc(n) ⊗min B(H) = Unc(n) ⊗max B(H). Proof. By Lemma 4.8 and Proposition 4.5, Vn ⊗min B(H) = Vn ⊗max B(H). Ap- plying Lemma 4.7 gives the desired result. (cid:3) i=1 Rε P2n It should be noted that Kirchberg's Theorem for C ∗(Fn) follows from Theo- rem 4.9. To show this fact, we will need the notion of a retract of operator systems. We will say that an operator system S is a retract of an operator system T if there are ucp maps ψ : S → T and χ : T → S such that χ ◦ ψ = idS. Lemma 4.10. Let S1,S2,T1,T2 be operator systems, and let τ1, τ2 ∈ {min, c, max}. For i = 1, 2, suppose that Si is a retract of Ti. If T1 ⊗τ1 T2 = T1 ⊗τ2 T2 completely order isomorphically (respectively, order isomorphically), then S1⊗τ1S2 = S1⊗τ2S2 completely order isomorphically (respectively, order isomorphically). Proof. Since min ≤ c ≤ max as operator system tensor products, we may assume that τ1 ≤ τ2. Since Si is a retract of Ti, there are ucp maps ϕi : Si → Ti and ψi : Ti → Si such that ψi ◦ ϕi = idSi. For each j = 1, 2, by functoriality of τj, the maps ϕ1 ⊗ ϕ2 : S1 ⊗τj S2 → T1 ⊗τj T2 and ψ1 ⊗ ψ2 : T1 ⊗τj T2 → S1 ⊗τj S2 are ucp. Moreover, the following diagram commutes: idT1 ⊗ idT2 T1 ⊗τ1 T2 T1 ⊗τ2 T2 ϕ1 ⊗ ϕ2 S1 ⊗τ1 S2 idS1 ⊗ idS2 ψ1 ⊗ ψ2 S1 ⊗τ2 S2 A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 19 By assumption, the map id : T1 ⊗τ1 T2 → T1 ⊗τ2 T2 is completely positive (respectively, positive). Thus, id : S1 ⊗τ1 S2 → S1 ⊗τ2 S2 is completely positive (respectively, positive). The result follows. (cid:3) For the next lemma, we define the operator system Sn ⊆ C ∗(Fn) to be 1, ..., w∗ n}, where w1, ..., wn are the generators of Fn. Sn = span {1, w1, ..., wn, w∗ Lemma 4.11. Let n ≥ 2. (1) C ∗(Fn) is a retract of Unc(n). (2) Sn is a retract of Vn. Proof. To prove (1), we note that  w1 . . . wn  ∈ Mn(C ∗(Fn)) is unitary. Hence, there is a unital ∗-homomorphism π : Unc(n) → C ∗(Fn) such that π(uij) = 0 for i 6= j and π(uii) = wi. Then π⊗ idS : Unc(n)⊗max S → C ∗(Fn)⊗max S is ucp, while id : Unc(n)⊗minS → Unc(n)⊗maxS is ucp by Proposition 4.5 and Lemma 4.7. We let U1 = U := (uij) ∈ Mn(Unc(n)). For 2 ≤ i ≤ n, let Ui be the conjugation of U by a permutation matrix such that the (1, 1)-entry of Ui is uii. Then each Ui ∈ Mn(Unc(n)) is unitary, so by the universal property for C ∗(Fn), there is a unital ∗-homomorphism ρ : C ∗(Fn) → Mn(Unc(n)) such that ρ(wi) = Ui for all 1 ≤ i ≤ n. Compressing to the (1, 1)-entry in Mn(Unc(n)) gives rise to a ucp map ψ : C ∗(Fn) → Unc(n) such that ψ(wi) = uii for all 1 ≤ i ≤ n. Since π ◦ ψ(wi) = wi and since wi is unitary, it follows that wi lies in the multiplicative domain of π◦ ψ. Since C ∗(Fn) is generated by {w1, ..., wn}, it follows that π ◦ ψ is multiplicative on C ∗(Fn). The fact that π ◦ ψ(wi) = wi for all i forces π ◦ ψ = idC ∗(Fn). Thus, (1) holds. For (2), since ψ(wi) = uii ∈ Vn for all 1 ≤ i ≤ n, we have ψ(Sn) ⊆ Vn. Clearly π(Vn) = Sn. Since π ◦ ψ = idC ∗(Fn), it follows that Sn is a retract of Vn via the maps ψSn : Sn → Vn and πVn : Vn → Sn. Theorem 4.12. Whenever S is an operator system with property Vn, we have C ∗(Fn) ⊗min S = C ∗(Fn) ⊗max S. Proof. By Lemma 4.11, C ∗(Fn) is a retract of Unc(n). Applying Lemma 4.10, since Unc(n) ⊗min S = Unc(n) ⊗max S, it follows that C ∗(Fn) ⊗min S = C ∗(Fn) ⊗max S, which completes the proof. Corollary 4.13. (Kirchberg's Theorem, [20]) Let n ≥ 2. Then C ∗(Fn) has the LLP. In other words, C ∗(Fn) ⊗min B(H) = C ∗(Fn) ⊗max B(H). Using Theorem 4.12, it is possible to characterize unital C ∗-algebras having the WEP and operator systems having the DCEP in terms of tensor products with V2. Theorem 4.14. Let A be a unital C ∗-algebra. The following are equivalent. (cid:3) (cid:3) (1) A has the WEP. (2) A has property V. (3) A has property V2. 20 SAMUEL J. HARRIS (4) A ⊗min V2 = A ⊗max V2. Proof. Clearly (2) implies (3), while (3) implies (4) by Proposition 4.5. Suppose that A has the WEP. By Theorem 2.2, A is (el, max)-nuclear. By Theorem 2.1, each Vn having the OSLLP implies that each Vn is (min, er)-nuclear. Hence, Vn ⊗min A = Vn ⊗er A = A ⊗el Vn = A ⊗max Vn = Vn ⊗max A. By Proposition 4.5 and the fact that n ≥ 2 was arbitrary, we conclude that A has property V. This shows that (1) implies (2). Finally, we prove that (4) implies (1). Suppose that A⊗min V2 = A ⊗max V2. Then by Lemma 4.7, we have Unc(2)⊗minA = Unc(2)⊗maxA. Using Theorem 4.12, we have C ∗(F2)⊗minA = C ∗(F2)⊗maxA. As F∞ embeds as a subgroup into F2, by [24, Proposition 8.8] it follows that there are ucp maps Φ : C ∗(F∞) → C ∗(F2) and Ψ : C ∗(F2) → C ∗(F∞) with Ψ◦Φ = id. By Lemma 4.10, we have C ∗(F∞)⊗minA = C ∗(F∞) ⊗max A. By [19, Proposition 1.1(iii)], A has the WEP. (cid:3) There is a similar characterization for operator systems with the DCEP. Theorem 4.15. Let S be an operator system. The following are equivalent. (1) S has the DCEP. (2) S ⊗min Vn = S ⊗c Vn for all n ≥ 2. (3) S ⊗min V2 = S ⊗c V2. Proof. Assume that S has the DCEP. By Theorem 2.3, S is (el, c)-nuclear, while Vn is (min, er)-nuclear. It follows that S ⊗min Vn = S ⊗c Vn for all n ≥ 2. Hence, (1) implies (2). Clearly (2) implies (3). If (3) is true, then by Lemma 4.7 and by Theorem 4.12, we must have S ⊗min C ∗(F2) = S ⊗max C ∗(F2). Since C ∗(F∞) is a retract of C ∗(F2), using Lemma 4.10 gives S ⊗min C ∗(F∞) = S ⊗max C ∗(F∞). Applying Theorem 2.3 shows that S has the DCEP, so that (1) is true. (cid:3) 5. Relating Vn to Kirchberg's conjecture The proof of Theorem 4.12 shows that C ∗(Fn) is a retract of Unc(n) via ucp maps. Using this trick allows for a connection between Unc(n) and Kirchberg's conjecture. Theorem 5.1. If Unc(n) ⊗min Unc(n) = Unc(n) ⊗max Unc(n) for some n ≥ 2, then Kirchberg's conjecture is valid. Proof. It is well known that Kirchberg's conjecture is true if and only if it holds for some n ∈ N with n ≥ 2. Now, if Unc(n) ⊗min Unc(n) = Unc(n) ⊗max Unc(n), then combining Lemmas 4.10 and 4.11 yields the complete order isomorphism C ∗(Fn) ⊗min C ∗(Fn) = C ∗(Fn) ⊗max C ∗(Fn). (cid:3) The link between Kirchberg's conjecture and the WEP allows us to prove the converse of Theorem 5.1. In other words, while the assumption that Unc(n) ⊗min Unc(n) = Unc(n) ⊗max Unc(n) for some n ≥ 2 appears to be slightly stronger than Kirchberg's conjecture, it is in fact equivalent to Kirchberg's conjecture. Theorem 5.2. The following statements are equivalent. (1) V2 ⊗min V2 = V2 ⊗c V2. A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 21 (2) Unc(2) ⊗min Unc(2) = Unc(2) ⊗max Unc(2). (3) C ∗(F2) ⊗min C ∗(F2) = C ∗(F2) ⊗max C ∗(F2). (4) C ∗(F∞) ⊗min C ∗(F∞) = C ∗(F∞) ⊗max C ∗(F∞). (5) Connes' embedding problem has a positive answer. Proof. Note that if (3) holds, then since C ∗(F∞) is a retract of C ∗(F2), (4) also holds by the same argument as in the proof of Theorem 4.14. Clearly F2 embeds into F∞ so that, by [24, Proposition 8.8], C ∗(F2) is a retract of C ∗(F∞). Hence, (4) implies (3). Using Lemma 4.7 shows that (1) implies (2), while Theorem 5.1 shows that (2) implies (3). Assuming (4) is true, it follows that C ∗(F∞) has the WEP [19]. Then [18, Theorem 9.1] shows that any operator system S that is (min, er)-nuclear satisfies S ⊗min S = S ⊗c S. By Lemma 4.8 and Theorem 2.1, V2 is (min, er)-nuclear. Therefore, V2 ⊗min V2 = V2 ⊗c V2, as required. (cid:3) e (Vn ⊗max Vm) = C ∗ e (Vn) ⊗max C ∗ e (Vn ⊗max Vm) 6= C ∗ Because Sn is a retract of Vn, we can prove the following. Proposition 5.3. For all n, m ≥ 2, Vn ⊗c Vm 6= Vn ⊗max Vm. Proof. By [10, Theorem 3.8], we have Sn ⊗c Sm 6= Sn ⊗max Sm for all n, m ≥ 2. if Vn ⊗c Vm = Vn ⊗max Vm, then by Lemmas 4.10 and 4.11, we have Hence, Sn ⊗c Sm = Sn ⊗max Sm, which is a contradiction. Corollary 5.4. For all n, m ≥ 2, C ∗ e (Vm). The latter C ∗- Proof. Suppose that C ∗ algebra is Unc(n) ⊗max Unc(m). Applying Proposition 4.6, Vn ⊗c Vm is completely order isomorphic to its inclusion in Unc(n) ⊗max Unc(m). The operator system Vn ⊗max Vm is completely order isomorphic to its inclusion in C ∗ e (Vn ⊗max Vm) [14]. Thus, Vn ⊗c Vm = Vn ⊗max Vm, contradicting Proposition 5.3. Corollary 5.5. Let U = (uij) ∈ Mn(Unc(n)) and V = (vkℓ) ∈ Mm(Unc(m)) be the matrices of generators of Unc(n) and Unc(m), respectively, where n, m ≥ 2. Then U0 = (uij ⊗ 1) ∈ Mn(C ∗ e (Vn ⊗max Vm)) fail to be unitary. Proof. Suppose that U0 and V0 were unitary. The entries of U0 ∗-commute with the entries of V0, so there are unital ∗-homomorphisms π : Unc(n) → C ∗ e (Vn ⊗max Vm) and ρ : Unc(m) → C ∗ e (Vn ⊗max Vm) with π(uij) = uij ⊗ 1 and ρ(vkℓ) = 1⊗ vkℓ. The ranges of π and ρ commute, so that π · ρ : Unc(n) ⊗max Unc(m) → C ∗ e (Vn ⊗max Vm) is a ∗-homomorphism. In particular, π · ρ is ucp and π · ρ is the identity map when restricted to Vn ⊗ Vm. By Proposition 4.6, the inclusion of Vn ⊗ Vm into Unc(n) ⊗max Unc(m) is Vn ⊗c Vm. Therefore, id : Vn ⊗c Vm → Vn ⊗max Vm is ucp, contradicting Proposition 5.3. e (Vn ⊗max Vm)) and V0 = (1 ⊗ vkℓ) ∈ Mm(C ∗ e (Vn) ⊗max C ∗ e (Vm). (cid:3) (cid:3) (cid:3) 6. Unitary Correlation Sets There is a way of stating Connes' embedding Problem in terms of what are known as quantum correlation matrices. We outline the definition of these sets, as they will motivate the definition of a new collection of correlation sets. The background for Tsirelson's problem in quantum correlations is motivated by a question in bipartite quantum information theory. Essentially, there are two 22 SAMUEL J. HARRIS models often used for nonlocal quantum correlations: one is a tensor product model, and one is a commuting model. Tsirelson's problem asks whether these models are the same, up to approximation. We will formulate the sets of nonlocal quantum correlations in each model below. Pm If H is a Hilbert space, we say that a set of operators (Pi)m i=1 is a posi- tive operator-valued measure with m outputs (POVM) if each Pi ≥ 0 and i=1 Pi = I. If the Pi's are also orthogonal projections, then we say that (Pi)m i=1 is a projection-valued measure with m outputs (PVM). Note that if (Pi)m i=1 is a PVM on H, then it necessarily follows that Pi ⊥ Pj for i 6= j. For each choice of n, m ∈ N, the set of quantum commuting correlation probabilities of two separated systems of n POVM's with m outputs is given by Cqc(n, m) = {(hPa,xQb,yξ, ξi)a,b,x,y}, where for each a, b ∈ {1, ..., n}, (Pa,x)m y=1 are POVM's with m out- puts, H is some Hilbert space, ξ ∈ H is a unit vector, and Pa,xQb,y = Qb,yPa,x for all choices of a, b, x, y. Similarly, we define x=1 and (Qb,y)m Cq(n, m) = {(h(Pa,x ⊗ Qb,y)ξ, ξi)a,b,x,y}, where each (Pa,x)m x=1 is a POVM with m outputs on a Hilbert space HA, each (Qb,y)m y=1 is a POVM with m-outputs on a Hilbert space HB, ξ ∈ HA ⊗ HB is a unit vector, and dim(HA), dim(HB) < ∞. We also define the possibly larger set Cqs(n, m) to be the set of all correlations with the same form as for Cq(n, m), except that we allow the Hilbert spaces to be infinite-dimensional. For convenience, we denote by Cqa(n, m) the closure of Cq(n, m). It is known that Cq(n, m) ⊆ Cqs(n, m) ⊆ Cqa(n, m) ⊆ Cqc(n, m), ∀n, m, and Cqc(n, m) is closed. Moreover, each of these sets is convex. One form of Tsirelson's problem, then, is determining whether Cqa(n, m) = Cqc(n, m) for all n, m ≥ 2. More information on these correlation sets can be found in [27], [12] and [15]. It is shown in [12] and [15] that, in the definitions of Cq(n, m) and Cqc(n, m), one may take the POVM's to simply be PVM's. There is a natural link between the sets Cqa(n, m), Cqc(n, m) and states on tensor products of the C ∗-algebra C ∗(∗nZm), where ∗nZm denotes the free product of n copies of the finite cyclic group Zm (see [12, 15, 21]). One key fact is that C ∗(∗nZm) is isomorphic to ∗nℓ∞ m [12, 15, 21]. If gx denotes the generator of the x-th copy of Zm in C ∗(∗nZm) and ea,x denotes the generator of the a-th coordinate in the x-th copy of ℓ∞ m , then the isomorphism is implemented via m , the free product of n copies of ℓ∞ It follows (see [12, 15, 21]) that Cqa(n, m) is the set of coordinates in Rn2m2 given by the images of states s ∈ S(C ∗(∗nZm) ⊗min C ∗(∗nZm)) on the generating set {ea,x ⊗ eb,y : 1 ≤ a, b ≤ m, 1 ≤ x, y ≤ n}, while Cqc(n, m) corresponds to the images of states on C ∗(∗nZm) ⊗max C ∗(∗nZm). For our purposes, we may consider the special case of m = 2, which involves the C ∗-algebra C ∗(∗nZ2). Following the notation in [10], we let hi be the generator gx 7→ mXa=1 exp(cid:18)2πai m (cid:19) ea,x. A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 23 of the i-th copy of Z2 inside of C ∗(∗nZ2). Each hi is a self-adjoint unitary. We let NC(n) be the operator system generated by {h1, ..., hn} inside of C ∗(∗nZ2). Proposition 6.1. (Farenick-Kavruk-Paulsen-Todorov, [10]) If X1, ..., Xn ∈ B(H) are hermitian contractions, then there is a unique ucp map γ : NC(n) → B(H) given by γ(hi) = Xi for all 1 ≤ i ≤ n. The isomorphism C ∗(∗nZ2) ≃ ∗nℓ∞ 2 is implemented by the mapping hi 7→ pi − qi, where pi is the element (1, 0) in the i-th copy of ℓ∞ 2 , and qi is the element (0, 1) in the i-th copy of ℓ∞ 2 . By [10, Lemma 6.2], the operator system NC(n)⊗c NC(n) is completely order isomorphic to its image inside of C ∗(∗nZ2)⊗max C ∗(∗nZ2). With this information in hand, we easily obtain the following: Proposition 6.2. For n ≥ 2, Cqa(n, 2) = Cqc(n, 2) if and only if the identity map id : NC(n) ⊗min NC(n) → NC(n) ⊗c NC(n) is an order isomorphism. Proposition 6.3. For any n ≥ 2, NC(n) is a retract of Vn. Proof. By [10, Proposition 5.7], there are ucp maps η : NC(n) → Sn and θ : Sn → NC(n) such that θ ◦ η = idN C(n). By lemma 4.11, there are ucp maps ψ : Sn → Vn and π : Vn → Sn with π ◦ ψ = idSn. Then ψ ◦ η : NC(n) → Vn and θ ◦ π : Vn → NC(n) are ucp maps satisfying (θ ◦ π) ◦ (ψ ◦ η) = idN C(n). We conclude that NC(n) is a retract of Vn. (cid:3) We wish to define correlation matrices with respect to Unc(n) that are similar in nature to Tsirelson's correlation sets. We recall that a C ∗-algebra A is said to be residually finite-dimensional if there is a family (πi)i∈I of ∗-homomorphisms key component in linking the usual quantum correlation matrices with Kirchberg's conjecture is the fact that C ∗(Fn) is RFD for every n. Here, we show that Unc(n) also enjoys this property. πi : A → B(Hi) with each dim(Hi) < ∞ such that π := Li∈I πi is faithful. A Theorem 6.4. For any n ≥ 2, Unc(n) is RFD. Proof. The proof mimics the proof that C ∗(Fn) is RFD (see [5, Theorem 7]). It is not hard to see that Unc(n) is a separable C ∗-algebra, so we may assume that Unc(n) ⊆ B(H) is faithfully represented on a separable infinite-dimensional Hilbert space H. Hence, there are operators Uij ∈ B(H) for 1 ≤ i, j ≤ n such that Unc(n) ≃ C ∗({Uij}i,j) via the mapping uij 7→ Uij. Let (Pm)∞ m=1 be a sequence of increasing projections with rank(Pm) = m and SOT -limm→∞ Pm = I. Define Vm,ij = PmUijPm and let Vm = (Vm,ij). Since rank(Pm) = m, we may identify Vm,ij ∈ Mm for each i, j and hence Vm ∈ Mn(Mm). Observe that Pm Vm = . . . Pm  U Pm . . . Pm  , 24 SAMUEL J. HARRIS where U = (Uij). Therefore, each Vm is a contraction. By Proposition 4.1, there exist unital ∗-homomorphisms πm : Unc(n) → M2(Mm) for each m ∈ N such that for all i, j. Since V ∗ U ∗. Hence, every entry of Vm converges in SOT, so that m,ij = PmU ∗ ijPm, SOT -limm→∞ Vm = U and SOT -limm→∞ V ∗ m = Xm,ij := πm(uij) =(cid:18) mVm)ij Vm,ij (pI − V ∗ Xm =(cid:18)Uij 0 0 −U ∗ SOT - lim m→∞ (pI − VmV ∗ −V ∗ m,ji m)ij (cid:19) ji(cid:19) . ji,−Uji})(cid:19) , Let F be any word in the generators of Unc(n). We similarly obtain SOT - lim m→∞ πm(F ) =(cid:18)F 0 0 F ({−U ∗ m,ij})k ≥ 1 − ε. Hence, π := Lm∈N πm is isometric on the dense where F ({−U ∗ ji,−Uji}) is the word obtained by replacing every occurrence of ji, and every occurrence of U ∗ Uij with −U ∗ ij with −Uji. Assume that F is norm 1. Then given ε > 0, there is m0 ∈ N such that for all m ≥ m0, we have kF ({Xm,ij, X ∗ subspace of linear combinations of words in the generators of Unc(n). Since π must be continuous, π is isometric on Unc(n). This shows that π is faithful and Unc(n) is RFD. Remark 6.5. It is not hard to see that whenever A and B are RFD C ∗-algebras, then A ⊗min B is also RFD. Hence, Unc(n) ⊗min Unc(k) is RFD for every n, k ≥ 2. As with C ∗(Fn), we can reformulate Kirchberg's conjecture in terms of whether or not Unc(n) ⊗max Unc(n) is RFD. The proof is identical to the C ∗(Fn) case [4, Proposition 7.4.4] and is omitted. (cid:3) Theorem 6.6. The following statements are equivalent. all/some n ≥ 2. (1) (Kirchberg's Conjecture) C ∗(Fn) ⊗min C ∗(Fn) = C ∗(Fn) ⊗max C ∗(Fn) for (2) Unc(n) ⊗min Unc(n) = Unc(n) ⊗max Unc(n) for all/some n ≥ 2. (3) Unc(n) ⊗max Unc(n) is RFD for all/some n ≥ 2. We will show below that (3) holds if we weaken the assumption of resid- ual finite-dimensionality to being quasidiagonal. Recall that a C ∗-algebra A is quasidiagonal, or QD, if there is a net of ucp maps ϕλ : A → Mk(λ) such that limλ kϕλ(a)k = kak and limλ kϕλ(ab) − ϕλ(a)ϕλ(b)k = 0 for all a, b ∈ A. It is easy to see that every RFD C ∗-algebra is QD. Theorem 6.7. For every n ≥ 2, Unc(n) ⊗max Unc(n) is QD. Proof. The proof is similar to the proof for C ∗(Fn)⊗max C ∗(Fn) (see [4, Proposition 7.4.5]). Let π : Unc(n) ⊗max Unc(n) be a faithful representation on a Hilbert space H. Let U = (Uij) be the matrix of generators of Unc(n) ⊗ 1, and let V = (Vij) be the matrix of generators of 1⊗Unc(n), so that each Uij, Vij ∈ B(H). The nature of the max tensor product forces the Uij's and Vkℓ's to ∗-commute. The unitary group of B(H(n)) is path connected by the Borel functional calculus. Hence, there are A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 25 norm-continuous functions u, v : [0, 1] → B(H(n)) such that u(0) = IH(n) = v(0), u(1) = U and v(1) = V . Since UV = V U and UV ∗ = V ∗U, the von Neumann algebras W ∗(U) and W ∗(V ) must commute with each other. Using the Borel functional calculus, we can arrange to have u(t) ∈ W ∗(U) and v(t) ∈ W ∗(V ) for all t ∈ [0, 1]. The entries of U and V ∗-commute, so this must also hold for the entries of p(U, U ∗) and q(V, V ∗) for any ∗-polynomials p, q. Taking limits, one sees that the entries of u(t) ∈ B(H(n)) must ∗-commute with the entries of v(t) ∈ B(H(n)) for all t ∈ [0, 1]. Since u(t) and v(t) are unitary with ∗-commuting entries, there is a unique ∗-homomorphism πt : Unc(n)⊗maxUnc(n) → B(H) with πt(uij⊗1) = (u(t))ij and πt(1⊗ vij) = (v(t))ij. As π0 is the trivial representation onto CIH and π1 = π, we see that π is homotopic to the trivial representation. Since π is injective and C is obviously QD, by [4, Proposition 7.3.5], Unc(n) ⊗max Unc(n) is QD. (cid:3) To obtain a unitary version of Tsirelson's problem that is equivalent to Kirch- berg's conjecture, it is helpful to have a characterization of RFD C ∗-algebras in terms of their state spaces. By way of notation, we denote by S(A) the set of all states on a unital C ∗-algebra A. We define Fin(A) to be the set of all states on A whose GNS representations act on finite-dimensional Hilbert spaces. While a number of characterizations for residual finite-dimensionality are given in [9], we only require the following one. Theorem 6.8. (Exel-Loring [9]) A unital C ∗-algebra A is RFD if and only if Fin(A) is w∗-dense in S(A). We are now in a position to define our unitary correlation sets. As in the usual setting, we will consider a tensor product model as well as a commuting model. For n ≥ 2 and a unitary U = (Uij) ∈ Mn(B(H)) for some Hilbert space H, we let Bn(U) = {IH} ∪ {Uij, U ∗ i,j=1. We define UCq(n1, n2) to be the set of all (2n2 2 + 1)-tuples of the form 1 + 1)(2n2 ij}n (hX ⊗ Y )ξ, ξi)X∈Bn1 (U ), Y ∈Bn2 (V ), where U ∈ Mn1(B(HA)) and V ∈ Mn2(B(HB)), HA and HB are finite-dimensional Hilbert spaces, and ξ ∈ HA⊗HB is a unit vector. We define the possibly larger set UCqs(n1, n2) to be the set of all correlations of the same form as for UCq(n1, n2), except that we allow the Hilbert spaces to be infinite-dimensional. For conve- nience, we will also define UCqa(n1, n2) to be the closure of UCq(n1, n2). For the commuting unitary correlation sets, we define UCqc(n1, n2) to be the set of all (2n2 2 + 1)-tuples of the form 1 + 1)(2n2 (hXY ξ, ξi)X∈Bn1(U ), Y ∈Bn2 (V ), where U ∈ Mn1(B(H)) and V ∈ Mn2(B(H)) are unitaries, H is a Hilbert space, ξ ∈ H is a unit vector, and XY = Y X for all X ∈ Bn1(U) and Y ∈ Bn2(V ). Since U and V are commuting unitaries, it follows that the Uij's and Vkℓ's ∗-commute. For convenience, we denote by Gn1,n2 the set of generators of Vn1 ⊗ Vn2 of the form x ⊗ y, where x ∈ {1} ∪ {uij, u∗ k,ℓ=1. By the correspondence between GNS representations and states, i,j=1 and y ∈ {1} ∪ {vkℓ, v∗ kℓ}n2 ij}n1 UCqc(n1, n2) = {(s(x))x∈Gn1,n2 : s ∈ S(Unc(n1) ⊗max Unc(n2))}. 26 SAMUEL J. HARRIS By Proposition 4.6, the inclusion Vn1 ⊗c Vn2 ⊆ Unc(n1) ⊗max Unc(n2) is a complete order embedding. Therefore, we may also write UCqc(n1, n2) = {(s(x))x∈Gn1 ,n2 : s ∈ S(Vn1 ⊗c Vn2)}. It is not hard to see that UCq(n1, n2) = {(s(x))x∈Gn1,n2 : s ∈ Fin(Unc(n1) ⊗min Unc(n2))}. These unitary correlation sets have similar properties to the quantum corre- lation sets. Proposition 6.9. For every n1, n2 ≥ 2, UCq(n1, n2) ⊆ UCqs(n1, n2) ⊆ UCqa(n1, n2) ⊆ UCqc(n1, n2), and each of these sets is convex. Moreover, UCqc(n1, n2) is closed. Proof. Since the state space of any operator system is convex, it is easy to see that each set above is convex. Clearly UCq(n1, n2) ⊆ UCqs(n1, n2). Every element of UCqs(n1, n2) corresponds to a state on Vn1 ⊗min Vn2, which extends to a state on Unc(n1) ⊗min Unc(n2) by the Hahn-Banach theorem. By Theorem 6.8, the set Fin(Unc(n1) ⊗min Unc(n2)) is w∗-dense in S(Unc(n1) ⊗min Unc(n2)), so that each element of UCqs(n1, n2) is also in UCqa(n1, n2). To show that UCqa(n1, n2) ⊆ UCqc(n1, n2), it suffices to show that UCqc(n1, n2) is closed. To that end, let p=1 ⊆ ((sp(x)x∈Gn1 ,n2 S(Vn1⊗cVn2). The mapping s : Vn1⊗cVn2 → C given by s(x) = limp→∞ sp(x) for all x ∈ Gn1,n2 extends to a linear functional. It follows that s = w∗-limp→∞ sp. Since the state space on an operator system is w∗-closed, we see that s ∈ S(Vn1 ⊗c Vn2) so that UCqc(n1, n2) is closed. )∞ p=1 be a sequence in UCqc(n1, n2) that converges, where (sp)∞ (cid:3) Before we link these unitary correlation sets to Connes' embedding problem, it will be helpful to have a better description of UCqa(n1, n2). Lemma 6.10. For each n1, n2 ≥ 2, UCqa(n1, n2) = {(s(x))x∈Gn1 ,n2 : s ∈ S(Vn1 ⊗min Vn2)}. Proof. Note that Vn1 ⊗min Vn2 is completely order isomorphic to its inclusion in Unc(n1) ⊗min Unc(n2). Since S(Vn1 ⊗min Vn2) is w∗-closed, the proof of Proposition 6.9 shows that UCqa(n1, n2) ⊆ {(s(x))x∈Gn1,n2 : s ∈ S(Vn1 ⊗min Vn2)}. Conversely, let s ∈ S(Vn1 ⊗min Vn2). By the Hahn-Banach theorem we may extend s to a state on Unc(n1)⊗minUnc(n2). Since Unc(n1)⊗minUnc(n2) is RFD, by Theorem 6.8, s can be approximated pointwise by elements of Fin(Unc(n1) ⊗min Unc(n2)). Restricting to the set Gn1,n2 yields a net of states whose images on the set Gn1,n2 are elements of UCq(n1, n2). Since this net of states converges pointwise to s, we see that (s(x))x∈Gn1 ,n2 ∈ UCqa(n1, n2), as required. (cid:3) Using Lemma 6.10 allows us to formulate the problem of deciding whether UCqa(n1, n2) = UCqc(n1, n2) in terms of Vn1 ⊗min Vn2 and Vn1 ⊗c Vn2. A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 27 Lemma 6.11. Let n1, n2 ≥ 2. Then UCqa(n1, n2) = UCqc(n1, n2) if and only if id : Vn1 ⊗min Vn2 → Vn1 ⊗c Vn2 is an order isomorphism. Proof. If UCqa(n1, n2) = UCqc(n1, n2), then by linearity the states on Vn1 ⊗min Vn2 and Vn1 ⊗c Vn2 are the same. An element in an operator system is positive if and only if its image under each state is positive (see, for example, [22, Chapter 13]), so we conclude that C min (Vn1,Vn2). Therefore, Vn1⊗minVn2 and Vn1⊗cVn2 must be order isomorphic. Conversely, if id : Vn1⊗minVn2 → Vn1⊗cVn2 is an order isomorphism, then the positive elements are the same in the two operator systems, so the state spaces are identical. Restricting to the set Gn1,n2, we obtain the equality UCqa(n1, n2) = UCqc(n1, n2). (Vn1,Vn2) = C comm (cid:3) 1 1 We are now ready for the main result of this section. Theorem 6.12. The following are equivalent. (1) Connes' embedding problem has a positive answer. (2) UCqa(n1, n2) = UCqc(n1, n2) for all n1, n2 ≥ 2. (3) UCqa(n, n) = UCqc(n, n) for all n ≥ 2. (4) Cqa(n, m) = Cqc(n, m) for all n, m ≥ 2. (5) Cqa(n, 2) = Cqc(n, 2) for all n ≥ 2. Proof. Suppose (1) holds. By [18, Theorem 9.1], Kirchberg's conjecture is equiva- lent to every (min, er)-nuclear operator system being (el, c)-nuclear. As each Vn is (min, er)-nuclear, it follows that Vn1⊗minVn2 = Vn1⊗cVn2 for all n1, n2 ≥ 2. Hence, these operator systems are order isomorphic, so that UCqa(n1, n2) = UCqc(n1, n2) for all n1, n2 ≥ 2. Clearly (2) implies (3) and (4) implies (5). The implication (5) =⇒ (1) was obtained by Ozawa [21, Theorem 36]. Hence, we need only show that (3) implies (5). By Lemma 6.11, condition (3) implies that Vn ⊗min Vn and Vn⊗cVn are order isomorphic. By Proposition 6.3, NC(n) is a retract of Vn. Using Lemma 4.10, the identity map id : NC(n) ⊗min NC(n) → NC(n) ⊗c NC(n) is 1- positive. Since min ≤ c, we see that NC(n)⊗min NC(n) and NC(n)⊗c NC(n) are order isomorphic for all n ≥ 2. Applying Proposition 6.2, we obtain the equality Cqa(n, 2) = Cqc(n, 2), as desired. (cid:3) In contrast, it is now known that UCqs(2, 2) ( UCqc(2, 2). Some striking differences arise between the quantum correlation sets and the unitary correlation sets. It is known that Cqa(2, 2) = Cqc(2, 2) (see, for example, [21]). The question of whether Cq(n, m) = Cqc(n, m) for all n, m ≥ 2 was open until Slofstra [25] recently proved that there are large n, m for which Cqs(n, m) 6= Cqc(n, m). Similarly, it was unknown whether Cqs(n, m) is closed for all n, m ≥ 2, until Slofstra recently provided a counterexample [26] for large n, m. Indeed, in [7] it is shown that there is a state s : Unc(2) ⊗min Unc(2) → C that cannot arise from a finite-dimensional representation of Unc(2) ⊗min Unc(2). In fact, it is shown that this state cannot arise from a spatial representation of Unc(2) ⊗min Unc(2) on a tensor product of Hilbert spaces, even if the Hilbert spaces are infinite- dimensional. Since C comm (Unc(2),Unc(2)), s is also a state on Unc(2) ⊗max Unc(2). Hence we obtain an element of UCqc(2, 2) that cannot be in UCqs(2, 2). This shows that UCqs(2, 2) ( UCqc(2, 2). Moreover, it is shown in (Unc(2),Unc(2)) ⊆ C min n n 28 SAMUEL J. HARRIS [7] that s can be approximated in the w∗-topology by states on Unc(2)⊗min Unc(2) corresponding to elements of UCq(2, 2). Therefore, UCq(2, 2) and UCqs(2, 2) are not even closed. The methods in [7] can be adapted in a natural way to show that UCqs(n, n) ( UCqc(n, n) for all n ≥ 2, and that UCq(n, n) and UCqs(n, n) are not closed. References [1] W.B. Arveson, Subalgebras of C ∗-algebras, Acta. Math. 123 (1969), 141 -- 224. [2] D.P. Blecher and V.I. Paulsen, Tensor products of operator Spaces, J. Funct. Anal. 99 (1991), 262 -- 292, DOI 10.1016/0022-1236(91)90042-4. [3] L. Brown, Ext of certain free product C ∗-algebras, J. Operator Theory 6 (1981), no. 1, 135 -- 141. [4] N. Brown and N. Ozawa, C ∗-algebras and finite-dimensional approximations, Graduate Studies in Mathematics, American Mathematical Society, Providence, RI, 2008. [5] M.D. Choi, The full C ∗-algebra of the free group on two generators, Pacific J. Math. 87 (1980), no. 1, 41 -- 48. [6] M.D. Choi and E.G. Effros, Injectivity and Operator Spaces, J. Funct. Anal. 24 (1977), no. 2, 156 -- 209. [7] R. Cleve, L. Liu, and V.I. Paulsen, Perfect embezzlement of entanglement, J. Math. Phys. 58 (2017), no. 1, 012204, 18pp., DOI 10.1063/1.4974818. [8] A. Connes, Classification of injective factors. Cases II1, II∞, IIIλ, λ 6= 1, Ann. Math. (2) [9] R. Exel and T.A. Loring, Finite-dimensional representations of free product C ∗-algebras, 104 (1976), no. 1, 73 -- 115. Internat. J. Math. 3 (1992), no. 4, 469 -- 476, DOI 10.1142/S0129167X92000217. [10] D. Farenick, A.S. Kavruk, V.I. Paulsen, and I.G. Todorov, Operator systems from discrete groups, Comm. Math. Phys. 329 (2014), no. 1, 207 -- 238, DOI 10.1007/s00220-014-2037-6. [11] D. Farenick and V.I. Paulsen, Operator system quotients of matrix algebras and their tensor products, Math. Scan. 111 (2012), no. 2, 210 -- 243. [12] T. Fritz, Tsirelson's problem and Kirchberg's conjecture, Rev. Math. Phys. 24 (2012), no. 5, 1250012, 67 pp., DOI 10.1142/S0129055X12500122. [13] V.P. Gupta and P. Luthra, Operator system nuclearity via C ∗-envelopes, J. Aust. Math. Soc. 101 (2016), no. 3, 356 -- 375, DOI 10.1017/S1446788716000082. [14] M. Hamana, Injective envelopes of operator systems, Publ. Res. Inst. Math. Sci. 15 (1979), 773 -- 785, DOI 10.2977/prims/1195187876. [15] M. Junge, M. Navascues, C. Palazuelos, D. Perez-Garcia, V.B. Scholz, and R.F. Werner, Connes embedding problem and Tsirelson's problem, J. Math. Phys. 52 (2011), no. 1, 012102, 12 pp., DOI 10.1063/1.3514538. [16] A.S. Kavruk, Nuclearity related properties in operator systems, J. Operator Theory 71 (2014), no. 1, 95 -- 156, DOI 10.7900/jot.2011nov16.1977. [17] A.S. Kavruk, V.I. Paulsen, I.G. Todorov, and M. Tomforde, Tensor products of operator systems, J. Funct. Anal. 261 (2011), no. 2, 267 -- 299, DOI 10.1016/j.jfa.2011.03.014. [18] , Quotients, exactness, and nuclearity in the operator system category, Adv. Math. 235 (2013), 321 -- 360, DOI 10.1016/j.aim.2012.05.025. [19] E. Kirchberg, On nonsemisplit extensions, tensor products and exactness of group C ∗- algebras, Invent. Math. 112 (1993), no. 3, 449 -- 489, DOI 10.1007/BF01232444. [20] , Commutants of unitaries in UHF algebras and functorial properties of exactness, J. Reine Angew. Math. 452 (1994), 39 -- 77, DOI 10.1515/crll.1994.452.39. [21] N. Ozawa, About the Connes embedding conjecture: algebraic approaches, Jpn. J. Math. 8 (2013), 147 -- 183, DOI 10.1007/s11537-013-1280-5. [22] V.I. Paulsen, Completely bounded maps and operator algebras, Cambridge Studies in Ad- vanced Mathematics, vol. 78, Cambridge University Press, Cambridge, 2002. A NON-COMMUTATIVE UNITARY ANALOGUE OF KIRCHBERG'S CONJECTURE 29 [23] V.I. Paulsen, I.G. Todorov, and M. Tomforde, Operator system structures on ordered spaces, Proc. Lond. Math. Soc. (3) 102 (2011), no. 1, 25 -- 49, DOI 10.1112/plms/pdq011. [24] G. Pisier, Introduction to operator space theory, London Mathematical Society Lecture Note Series, vol. 294, Cambridge University Press, Cambridge, 2003. [25] W. Slofstra, Tsirelson's problem and an embedding theorem for groups arising from non-local games, arXiv:1606.03140 [quant-ph] (2016). [26] , The set of quantum correlations is not closed, arXiv:1703.08618 [quant-ph] (2017). [27] B.S. Tsirelson, Some results and problems on quantum Bell-type inequalities, Hadronic J. Suppl. 8 (1993), no. 4, 329 -- 345. University of Waterloo, Department of Pure Mathematics, 200 University Ave. W., Waterloo, Ontario, Canada N2L 3G1, E-mail address: [email protected]
1007.1730	2	1007	2011-02-23T02:43:54	Subfactors of index less than 5, part 1: the principal graph odometer	[ "math.OA", "math.QA" ]	In this series of papers we show that there are exactly ten subfactors, other than $A_\infty$ subfactors, of index between 4 and 5. Previously this classification was known up to index $3+\sqrt{3}$. In the first paper we give an analogue of Haagerup's initial classification of subfactors of index less than $3+\sqrt{3}$, showing that any subfactor of index less than 5 must appear in one of a large list of families. These families will be considered separately in the three subsequent papers in this series.	math.OA	math	Subfactors of index less than 5, part 1: the principal graph odometer Scott Morrison Noah Snyder URLs: http://tqft.net/ and http://math.columbia.edu/~nsnyder Email: [email protected] [email protected] Abstract In this series of papers we show that there are exactly ten subfactors, other than A∞ subfactors, of index between 4 and 5. Previously this classification was known up to index 3+ 3. In the first paper we give an analogue of Haagerup's initial classification of subfactors of index less than 3 + 3, showing that any subfactor of index less than 5 must appear in one of a large list of families. These families will be considered separately in the three subsequent papers in this series. √ √ AMS Classification 46L37; 18D10 Keywords Subfactors 1 Introduction A subfactor is an inclusion N ⊂ M of von Neumann algebras with trivial center. The theory of subfactors can be thought of as a nonabelian version of Galois theory, and has had many applications in operator algebras, quantum algebra, and knot theory. For example, given a finite depth subfactor, you can produce two fusion categories (by taking the even parts) and a 3-dimensional TQFT (via the Ocneanu-Turaev-Viro construction [TV92, Ocn94, BW99]). The fundamental example of a subfactor, which illustrates the analogy with Galois theory, is when G is a finite group with an outer action on a factor M and the fixed point factor M G is a subfactor of M . A subfactor N ⊂ M has three key invariants. From strongest to weakest, they are: the standard invariant (which captures all information about "basic" bimodules over M and N ), the principal and dual principal graphs (which together describe the fusion rules for these basic bimodules), and the index (which is a real number measuring the "size" of the basic bimodules [Jon83]). For the case of the fixed point subfactor the standard invariant recovers the structure of G itself, the principal graph and dual principal graph recover the size of the group and the dimensions of each of its irreducible representations, and the index is the size of G. Thus studying possible standard invariants (called paragroups [Ocn88], λ-lattices [Pop95], or subfactor planar algebras[Jon]) of a fixed index is a generalization of studying all groups of a given size. In turn, studying subfactors with a fixed standard invariant is a generalization of studying all outer actions of a group on a factor. In particular, just as any finite (or more generally, amenable group) has only one action on the hyperfinite II1 factor [Jon80, Ocn80], any finite depth (or more generally amenable) subfactor planar algebra can be realized in a unique way as a subfactor of the hyperfinite II1 factor [Pop90, Pop94]. Much early work in the theory of subfactors concerned the classification of subfactors of index up to 4 [GdlHJ89, Ocn88, Izu91, KO02]. One reason for concentrating on 1 index below 4 was that for every real number greater than 4 there is a subfactor with that index and principal graph A∞ [Pop93]. The study of subfactors of small index larger than 4 was initiated by Haagerup who gave an exhaustive list of possible principal graphs other than A∞ of subfactors of index less than 3+ 3 [Haa94]. Most of the graphs on this list were excluded by Bisch [Bis98] and Asaeda-Yasuda [AY09], while the remaining 3 graphs and their duals were shown to come from unique subfactor planar algebras/λ-lattices/paragroups (and thus by Popa [Pop90, Pop94] unique subfactors of the hyperfinite II1 factor) by Asaeda-Haagerup [AH99] and Bigelow-Morrison-Peters-Snyder [BMPS09]. The goal of this series of papers is to give the following classification of irreducible subfactors of index less than 5. √ Theorem 1.1 There are exactly ten subfactor planar algebras other than Temperley- Lieb with index between 4 and 5. √ • The Haagerup planar algebra [AH99], with index 5+ 2 13 and principal bigraph • The extended Haagerup planar algebra [BMPS09], with index 8 3 + 2 3 Re 3 and principal bigraph pair (cid:113) 13 2 3(cid:1) √ (cid:0)−5 − 3i (cid:17) (cid:17) (cid:17) (cid:16) pair and its dual. (cid:16) (cid:16) , , , , and its dual. √ • The Asaeda-Haagerup planar algebra [AH99], with index 5+ 2 17 bigraph pair and principal (cid:17) • The 3311 Goodman-de la Harpe-Jones planar algebra [GdlHJ89], with index and its dual. √ (cid:16) 3 + 3 and principal bigraph pair and its dual (since it is not self-dual despite having the same principal and dual principal graphs [Kaw95]). • Izumi's self-dual 2221 planar algebra [Izu01] and its complex conjugate, with index 5+ 21 and principal bigraph pair √ 2 (cid:16) (cid:17) . As an immediate corollary (using Popa's results [Pop90, Pop94]) this theorem gives a classification of all amenable subfactors of the hyperfinite II1 factor of index less than 5. However, the classification of non-amenable subfactors coming from Temperley-Lieb is still wide open (see [Pop91, Bis94]). The choice of 5 as an upper limit is convenient, but somewhat arbitrary. In particular, classifying the principal graphs of subfactors at index 5 does not require a significant 2 amount of new work. However, at index 5 there are a large number of possible graphs. Furthermore, although most of these graphs can be realized via group/subgroup subfactors, they may also be realized in other ways. This will be addressed in work in joint progress with Izumi. In this paper, we begin by proving a weaker classification result inspired by Haagerup's original argument. In particular, we only use the combinatorics of the principal graphs of the subfactors. We describe several known obstructions to being principal graphs, and a method of enumerating principal graphs satisfying certain size bounds. Combining these, we obtain various classification results. Much of the subtlety in this paper lies in finding the right balance between looking for obstructions and extending the enumeration, in order to produce classification results that are both true and useful! In order to describe our classification (as well as Haagerup's) it's helpful to define two terms. We use the term translation of a graph pair to indicate a graph pair obtained by increasing the supertransitivity (that is, adding a longer chain of edges at the left) by an even integer, and the term extension of a graph pair to indicate a graph pair obtained by extending the graphs in any way at greater depths (i.e. adding vertices and edges at the right). For example, the D2n principal graph pairs are all translations of D4 , while the Haagerup principal graph pair is an extension of D6 , an extension of A4 , a translated extension of D4 , and a translated extension of A2 . See §2.1 for more details. Haagerup's initial classification is given in the following theorem. Theorem 1.2 [Haa94] Any subfactor of index between 4 and 3 + 3 is non- amenable with standard invariant the Temperley-Lieb algebra, or its principal graph pair is a translate of one of the following three graph pairs. √ (cid:110)(cid:16) (cid:16) (cid:16) (cid:17) (cid:17) , , (cid:17)(cid:111) , , , Haagerup eliminated all but one graph in the second family using an unpublished connection-based approach, Bisch eliminated the third family completely using fusion algebra techniques [Bis98], and Asaeda -- Yasuda [AY09] eliminated all but two graphs from the first family using arithmetic techniques. Number theory techniques of Calegari-Morrison-Snyder [CMS10] (generalizing earlier work of Asaeda-Yasuda [AY09]) give an effective bound on how large a translate of a fixed graph can be a principal graph. Thus any classification along the lines of Haagerup's can now be reduced to a finite list by applying this result. Unfortunately, in our case the techniques used by Haagerup do not suffice to restrict to the transla- tions of a finite list of graph pairs. Thus, in addition to a long list of families which can be eliminated using number theoretic techniques, we also have a short list of bad cases. 3 Theorem 1.3 The principal graphs of any subfactor of index between 4 and 5 is a translate of one of an explicit finite list of graph pairs (see Theorem 6.1), or is a translated extension of one of the following graphs. (cid:17) , (cid:17) (cid:17) , , (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) C = F = B = Q = Q(cid:48) = , , , , , (cid:17) (cid:17) , For these remaining cases we need new techniques. Cases C , F , and B (which have initial triple points) will be eliminated in joint work with Penneys and Peters [MPPS10], while Q and Q(cid:48) (which have initial quadruple points) will be eliminated in work in progress joint with Izumi and Jones [IJMS]. The uniqueness of 2221 up to conjugation was proved by Han [Han10], while the uniqueness of 3331 up to duality will be proved in [IJMS]. Finally, in [PT10] Penneys and Tener apply the number theoretic test of [CMS10] to eliminate the remaining cases. Classifications of subfactors of small index may be of interest for several reasons. We would like to understand whether the appearance of "exotic" subfactors (like the Haagerup, extended Haagerup, and Asaeda-Haagerup subfactors) is a common phenomenon, or whether these small examples are truly exceptional. We would like to understand where the boundary takes place between "small index" and "large index." For example, the smallest possible index other than 4 for an infinite depth subfactor whose standard invariant is not Temperley-Lieb is unknown. Similarly, we'd like to know the smallest index for which the standard invariant fails to classify subfactors of the hyperfinite factor [BNP07]. In order to answer these questions, eventually we would like to extend our classification up to index 3 + 5 where a Fuss-Catalan infinite depth subfactor [BJ97] appears which may allow for some of the "large index" behavior found at index 6. √ We would like to thank Richard Burstein, Vaughan Jones, Dave Penneys, and Emily Peters for helpful conversations and careful reading of the manuscript. We also thank all the attendees of the Bodega Bay "Planar algebra programming camps" for their interest in this problem: several attendees are writing sequels to this paper. Scott Morrison was at Microsoft Station Q at UC Santa Barbara and at the Miller Institute for Basic Research at UC Berkeley during this work, and Noah Snyder was supported in part by RTG grant DMS-0354321 and in part by an NSF Postdoctoral Fellowship at Columbia University. We would also like to acknowledge support from the DARPA HR0011-11-1-0001 grant. 2 Enumerating principal graphs 2.1 Notation and background Throughout, we use the following definitions. 4 Definition 2.1 A bigraph is a bipartite graph with a specified root vertex. We allow graphs with infinitely many vertices, but require that every vertex have finite degree. The depth of a vertex is the geodesic distance from the root. A bigraph is called finite depth if it has finitely many vertices, in which case its depth is the maximum of the depths of the vertices. When we draw bigraphs, the root vertex always appears on the left, and the depth of a vertex is always given by its horizontal distance from the root. Definition 2.2 The supertransitivity of a bigraph Γ is the greatest integer n so that up to vertices at depth n the graph Γ is just An . Equivalently, it is the number of edges from the root vertex before the first branch point or multiple edge. Definition 2.3 A bigraph with dual data is a bigraph together with an involution, called duality, of the vertices at each even depth. In diagrams, duality is represented by red arcs joining dual pairs of vertices. Self-dual vertices have a small red dash above them. Definition 2.4 A bigraph pair is a pair of bigraphs with dual data, with depths differing by at most one and the same supertransitivity, together with a bijection, also called duality, between the vertices at each odd depth of one graph with the corresponding vertices at the same odd depth on the other graph. In diagrams, we do not explicitly indicate the bijections between odd vertices, but rather use the convention that the vertical ordering of vertices at a given depth determines the bijection: the lowest vertices in each graph at each odd depth are dual to each other, etc. Two bigraph pairs are isomorphic if there is a graph isomorphism between the underlying graphs which preserves the duality structure and the base vertex. For example, the following two bigraph pairs are isomorphic: We say a bigraph pair is equal if both graphs have the same depth. Note that if a bigraph pair is unequal then the two depths differ by one, and the longer bigraph needs to have even depth, because of the duality bijection between the odd vertices. The principal graphs of a subfactor naturally carry the structure of a bigraph pair. For example, the principal graphs of the Haagerup subfactor are the following pair of bigraphs with dual data. We say that a bigraph pair Γ is a translate of Γ0 if there is an even integer k such that up to depth k , the bigraphs in Γ look like the Dynkin diagram Ak , and such that Γ0 is the induced bigraphs with dual data produced by removing the first k − 1 vertices from Γ and declaring the unique vertex at depth k the new base vertex. 5 (cid:16) (cid:16) (cid:110)(cid:16) (cid:17) (cid:17) (cid:17)(cid:111) , , , (The integer k is required to be even because duality is a different kind of structure for odd depths and even depths.) We say a bigraph pair Γ is an extension of a equal bigraph pair Γ0 if Γ is the same as Γ0 up to the depth of Γ0 . We will use the phrase "Γ starts like Γ0 " as an abbreviation for "Γ is an extension of a translate of Γ0 ". For example, the D2n principal graph pairs are all translations of the D4 pair. The Haagerup principal graphs are an extension of D6 , an extension of A4 , a translated extension of D4 , and a translated extension of A2 . Since we are often looking not just at a particular bigraph pair, but instead at all ways of extending and translating it, we do not insist that the Frobenius-Perron eigenvalues of the two graphs in a bigraph pair are equal (although they are certainly equal for principal bigraph pairs of finite depth subfactors). Similarly, although most graphs we discuss are finite depth, since we are considering all translated extensions of these graphs our results apply to infinite depth subfactors. We also define a sequence of numbers associated to any bigraph. Definition 2.5 The annular multiplicities {an}n∈N of a bigraph Γ are defined by the formula n(cid:88) r=0 an = (−1)r−n 2n n + r (cid:19) (cid:18)n + r n − r wr where wr is the number of length 2r loops on Γ based at the initial vertex. Note that the annular multiplicity an only depends on the bigraph Γ out to depth n. Since the annular multiplicities for the Al graphs are {1, 0, 0, 0, . . .}, the annular mul- tiplicities an for any k -supertransitive graph are a1 = 0 and an = 1 for 1 ≤ n ≤ k . Thus we will often describe the annular multiplicities by dropping this initial string, and listing the sequence of annular multiplicities starting from the first non-trivial entry. If a bigraph pair is the principal bigraph pair of a subfactor, then the corre- sponding planar algebra is a representation of the annular Temperley-Lieb category, and these numbers are in fact the multiplicities of the irreducible representations (c.f. [Jon01, GL98, JR06]). More specifically, the irreducible unitary representations of the annular Temperley-Lieb category at a parameter δ are parametrized by the set {(0, µ) µ ∈ [0, δ]} ∪ {(n, ω) n ∈ N, n ≥ 1, ω ∈ C, ωn = 1} , and the annular multiplcity an is the sum of the multiplicities of all irreducible representations of the form (n, x) for some x. In this paper, we only use the annular multiplicities to divide up large collections of graphs, but other papers in this series will make further use of the representation theory of the annular Temperley-Lieb category. 2.2 Classification statements Most of the results of this paper are of the following form: Every subfactor whose principal bigraph pair is not A∞ , which starts like a fixed bigraph pair Γ0 , and which has index strictly between 4 and Λ ∈ R has principal bigraph pair which either 6 (1) is a translate (but not an extension!) of one of a certain set of bigraph pairs V , called "vines", or (2) is a translate of an extension of one of a certain set of bigraph pairs W , called "weeds". We will abbreviate such a result by saying that the data (Γ0, Λ,V,W) is a "classi- fication statement". The idea behind the names is that the 'vines' can only grow taller (i.e. increase supertransitivity) but the 'weeds' can still grow out of control. There are of course many boring classification statements, in particular we trivially have (Γ0,∞,∅,{Γ0}) for any bigraph pair Γ0 . The interesting classification state- ments we produce will have weeds which are much larger than the bigraph pair Γ0 . In ideal circumstances, we even find classification statements in which the set of weeds is empty. Example 1 In [Haa94], Haagerup proves the classification statement √ 3,VH ,∅) with VH = (cid:110)(cid:16) (cid:16) (cid:16) (( , ), 3 + , , , (cid:17) (cid:17)(cid:111) (cid:17) Remark. These families of graphs are called the Haagerup family, the Asaeda- Haagerup family, and the hexagon family, respectively. In Haagerup's paper he claims that that in the Asaeda-Haagerup family only supertransitivity 5 is possible. Since Haagerup's paper the hexagon family has been entirely ruled out [Bis98] and the Haagerup family at supertransitivities 11 and above have been ruled out [Asa07, AY09]. A uniform argument for all three cases (and indeed excluding all but finitely many examples coming from any vine) can be given using [CMS10]. The 3-supertransitive bigraph pair in the Haagerup family is realised as the principal bigraph pair of the Haagerup subfactor [AH99], and the 7-supertransitive bigraph pair is realised by the extended Haagerup subfactor [BMPS09]. The Asaeda-Haagerup bigraph pair is also realised [AH99]. In this sense, the classification statement referred to above in terms of vines (and no weeds) has since been turned into a complete classification. The goal of this paper is to prove a classification statement that can then lead to a complete classification. Unlike in Haagerup's case, we will still have several weeds in our classification. 2.3 The odometer Given a bigraph Γ, we can readily enumerate all its extensions with depth one greater and with Frobenius-Perron eigenvalue less than some limit Λ. Indeed, the bound on the size of the Frobenius-Perron eigenvalue gives a bound on the valency of each vertex. In practice, one must be a little careful in order to do this enumeration 7 efficiently. Suppose we've already enumerated all extensions of Γ where the last depth is has rank k , and we next want to find all extensions of Γ with k + 1 new vertices. In terms of the inclusion matrix between the final two depths we are looking for ways of adding a new row to a fixed matrix such that the entries aren't too large. These can be enumerated via an odometer process: increment the first entry of the row until the graph norm is too large, then reset the first entry and increment the second entry, and so on. In order to further increase efficiency we may assume that the rows of the adjacency matrix between the top two depths are in lexicographic order, as permutations of the rows correspond to permutations of the new vertices, giving isomorphic graphs. Again with a fixed limit Λ on the Frobenius-Perron eigenvalue, we can enumerate all depth one extensions of a bigraph with dual data Γ in much the same way. We first forget about dual data and extend the bigraph, then, if the new depth is even, for each extension we consider all possible involutions of the new vertices. We denote the resulting set OΛ(Γ). Given a bigraph pair which is equal, we can enumerate all depth one extensions so that both bigraphs have Frobenius-Perron dimension at most Λ. When we are adding an odd depth, we are only interested in pairs where we add the same number of new vertices; in this case, because the duality involution between the new odd vertices is implicitly determined by the order in which they appear, we can insist the rows of the new inclusion matrix are in lexicographic order when extending one of the graphs, but not both. On the other hand, when we are adding an even depth, there may be different numbers of new vertices on the two graphs, and we can ask that both new inclusion matrices have rows in lexicographic order. This process frequently results in duplicate bigraph pairs which are related by a nontrivial bigraph isomorphism (which may permute vertices at earlier depths as well), and for efficiency these should be removed. Reusing notation, we denote by OΛ(W ) the set of all depth one extensions of a equal bigraph pair W . Alternatively, if W has odd depth we can extend one bigraph but not the other, and we call the resulting set LΛ(W ). When W has an even depth we can't do this (because the numbers of new odd vertices on the new graphs must agree), so LΛ(W ) = ∅. The following 'meta-theorem' explains how to use these enumeration techniques, which we collectively call 'the odometer', to derive new classification statements from old ones. Theorem 2.6 (The odometer) Suppose we have a classification statement (Γ0, Λ,V,{W} ∪ W). Then there is another classification statement (Γ0, Λ,V ∪ {W} ∪ LΛ(W ),W ∪ OΛ(W )). Proof This follows immediately from the definitions. Note that by repeatedly applying Theorem 2.6 we can arbitrarily increase the mini- mum depth of the weeds in a classification statement, at the expense of dramatically 8 increasing the number of vines and weeds. When we say that we "run the odometer" what we mean is that we apply Theorem 2.6, remove all weeds and vines which do not pass the associativity test (explained in the next section), and then repeat. See Section 4 for details. Next, we turn to another class of techniques for modifying classification statements: finding obstructions that can rule out entire families of bigraph pairs, coming from either vines or weeds. 3 Obstructions Suppose that Γ is a bigraph pair which starts like Γ0 . In this section we give several obstructions to Γ being the principal bigraph pair of a subfactor. Furthermore, all of these obstructions are local in the sense that they can be computed only from Γ0 . Locality in this sense is crucial because it means that these tests can be applied to remove vines and weeds from classification statements. 3.1 Associativity Recall that the principal graph gives the multiplicities for tensoring on the right with ABB and BBA between A − A bimodules and A − B bimodules. The dual principal graph gives the multiplicities for tensoring on the left with ABB and BBA between B − B bimodules and B − A bimodules. However, since taking duals interchanges the order of tensor product, using the dual data we can also recover from a bigraph pair the multiplicities for fusion on the left by the basic bimodules. (The multiplicities for tensoring with the basic bimodules on both the left and right are often encoded, following Ocneanu, as a 4-partite graph. We find bigraph pairs easier to deal with combinatorially and far more compact to display. However, one can easily go between the two descriptions.) By associativity, we can first tensor on the left and then on the right, or we can first tensor on the right and then on the left. This gives a combinatorial obstruction for potential principal graph pairs. These associativity conditions were first observed by Ocneanu. In the language of paragroups, associativity becomes the condition that biunitary matrices are square. In the language of bigraph pairs, associativity becomes the following. Lemma 3.1 Suppose that (Γ, Γ(cid:48)) is the principal bigraph pair of a subfactor and that Γ and Γ(cid:48) are simply laced. Consider two vertices V and W of the same parity in the principal and dual principal graphs respectively. The following two numbers are equal: (1) The number of vertices Z in the principal graph which are adjacent to V , such that Z∗ is adjacent to W ∗ . (2) The number of vertices U in the dual principal graph which are adjacent to W such that U∗ is adjacent to V ∗ . 9 Remark. If the graphs are not simply laced, the same theorem holds if you count the intermediate vertices with multiplicities corresponding to the product of the multiplicities of the two adjacencies. The obstruction given by the above theorem is local in the following sense. If V and W are not both the largest depth nor both at the smallest depth, then the dimensions computed in either of these fashions are the same for a bigraph pair Γ0 as they are for any bigraph pair which starts like Γ0 . Furthermore, if V and W are both at the largest depth, then the dimensions computed in either of these fashions are the same for all translates of the given graph. For example, it is easy to see using the associativity test that if a graph begins like Dn then its dual graph must also begin like Dn . While running the odometer, we can apply a trick which saves some time in checking the associativity obstruction. Recall that when we are adding an even depth, we construct two sets LΛ(W ) (in which we have extended only one of the two graphs) and OΛ(W ) (in which both graphs have been extended). It is easy to see that the extended graph in any unequal extension in LΛ(W ) which passes the (global, in the sense of the paragraph of locality above) associativity test can not also pass the (local) associativity test with any other extended graph in a level extension in OΛ(W ). Suppose W = (W1, W2), and W (cid:48) 2 is an extension of W2 . Thus if either (W (cid:48) 2) satisfies the global associativity test, we can immediately rule out the pair (W (cid:48) 2) as an element of OΛ(W ), because it cannot satisfies the local associativity test. 1 is an extension of W1 , W (cid:48) 1, W2) or (W1, W (cid:48) 1, W (cid:48) 3.2 The triple point obstruction Versions of the following results are proved in [Haa94], where they are attributed to Ocneanu. Lemma 3.2 Suppose (Γ, Γ(cid:48)) is the principal bigraph pair of a subfactor of index 4 or larger, and V ∈ Γ and V ∗ ∈ Γ(cid:48) are a pair of dual triple points at an odd depth. Denote by K the three neighbours of V on Γ, and by L the three neighbours of V ∗ on Γ(cid:48) . Consider φ : K → L, a dimension preserving bijection. Then there exists K ∈ K and L ∈ L so φ(K) (cid:54)= L, and an odd vertex Z (cid:54)= V on Γ so Z is adjacent to K and Z∗ is adjacent to L. Lemma 3.3 Suppose (Γ, Γ(cid:48)) is the principal bigraph pair of a subfactor of index 4 or larger, and V ∈ Γ and W ∈ Γ(cid:48) are a pair of triple points at an even depth, such that the neighbours of V coincide with the duals of the neighbours of W ∗ , and the duals of the neighbours of V ∗ coincide with the neighbours of W . Denote by K the three odd neighbours of V on Γ, and by L the three odd neighbours of W on Γ(cid:48) . Consider φ : K → L, a dimension preserving bijection. Then there exists K ∈ K and L ∈ L so φ(K) (cid:54)= L, and an even vertex Z (cid:54)= V on Γ so Z is adjacent to K and Z∗ is adjacent to L∗ . Both of these results are proved in the same way. Assume for the sake of contradiction that the last condition (the existence of vertex adjacent to K and L) does not hold. 10 By considering the paragroup (or equivalently the 6j-symbols) there is a 3-by-3 unitary matrix each of whose matrix entries have specific absolute values. A short argument in linear algebra then shows that the index must be less than 4. See the proof of Proposition 3.5 from [Haa94] for more details. (cid:83) A2 = A(cid:48) 1 (cid:83) A(cid:48) 2 , or A1 = A(cid:48) 2 . We call (S, A1, A2, A(cid:48) In order for these lemmas to give us a local condition, we need to be able to determine which bijections are dimension preserving. However, under most circumstances it is not possible to compute all the dimensions just from local information. However, if the triple point in question is at depth n for an n-supertransitive subfactor, and if the principal graphs are sufficiently simple through depth n + 2, then it is possible to apply the triple point obstruction without knowing anything about the graph at higher depths. First let us fix some terminology. Suppose that S is a finite set, and that A1 , A2 , A(cid:48) 1 , 2 are subsets of S with S = A1 and A(cid:48) 1, A(cid:48) 2) 1 and A(cid:48) forbidden if A1 and A2 are disjoint, A(cid:48) 2 are disjoint and either A1 = A(cid:48) 1 2 and A2 = A(cid:48) and A2 = A(cid:48) 1 . Corollary 3.4 Suppose that a bigraph pair (Γ, Γ(cid:48)) has dual initial triple points at an even depth n. Let α and β be the vertices of Γ at depth n + 1. Let S be the set of vertices in the principal graph at depth n + 2. Let A1 be the set of vertices at depth n + 2 which are adjacent to α and let A2 be the set of vertices at depth n + 2 which are adjacent to β . Let A(cid:48) i be the set of duals of vertices in Ai . If (S, A1, A2, A(cid:48) 2) is forbidden, then (Γ, Γ(cid:48)) is not the principal bigraph pair of a subfactor of index 4 or more. Corollary 3.5 Suppose that a bigraph pair (Γ, Γ(cid:48)) has dual initial triple points at an odd depth n. Let α and β be the two vertices of Γ at depth n + 1 and α(cid:48) , β(cid:48) the two vertices of Γ(cid:48) at depth n + 1. Let S be the set of vertices of Γ at depth n + 2. Let A1 be the set of vertices at depth n + 2 which are adjacent to α and let A2 be the set of vertices at depth n + 2 which are adjacent to β . Let A(cid:48) 2 be the duals of the vertices at depth n + 2 which are adjacent to α(cid:48) and β(cid:48) respectively. If (S, A1, A2, A(cid:48) 2) is forbidden, then (Γ, Γ(cid:48)) is not the principal bigraph pair of a subfactor of index 4 or more. 1 and A(cid:48) 1, A(cid:48) 1, A(cid:48) Proof Both of these results follow quickly from the above by using the fact that if two vertices have the same set of adjacent vertices, then they must have the same dimension. 3.3 Duals at depth n + 1 Suppose that Γ is an n-supertransitive bigraph with duals for n odd. Suppose that a of the vertices at depth n + 1 are self-dual and that 2b of them are not self-dual. Considering the 180-degree rotation acting on the quotient of n + 1-box space by the ideal of elements which factor through the n − 1-box space, we see that the dimension of the −1-eigenspace is b and the dimension of the +1-eigenspace is a + b. Lemma 3.6 Suppose that (Γ, Γ(cid:48)) is the principal bigraph pair of a subfactor, and that a, b, a(cid:48), b(cid:48) are defined as above. Then a = a(cid:48) and b = b(cid:48) . 11 Proof Because the half-click rotation gives isomorphisms between the corresponding eigenspaces of the shaded and unshaded n + 1 box space, we must have b = b(cid:48) and a + b = a(cid:48) + b(cid:48) , and hence (a, b) = (a(cid:48), b(cid:48)). 3.4 Even quadruple points The following obstruction to quadruple points is proved by Jones using quadratic tangles techniques. A more sophisticated version of this argument will rule out Q entirely. Theorem 3.7 The principal graph pair of a subfactor does not begin like (cid:16) , (cid:17) . Proof [Jon03, Theorem 5.2.2.] 4 Running the odometer By running the odometer on a classification statement we mean repeating the following two procedures: (1) Apply Theorem 2.6 to the classification statement, extending all the weeds by one depth. (2) Apply the associativity test and remove all vines and all weeds which fail it. We have written a package of computer programs which automates the odometer, and we describe its use at the end of this section. It is easy (but very tedious) to check the output of this program by hand. We've included figures which summarize the output of the odometer as it runs. These figures are trees, with the initial graph on the left, and each successive equal shows the new weeds that arise that pass the associativity tests. See below for more details. The example that we consider reproduces a small part of Haagerup's classification of subfactors out to index 3 + 3 by finding a complete set of vines for principal graph pairs starting like √ (cid:17) (Γ, Γ(cid:48)) = (cid:16) √ with index less that 3 + 3. (cid:0)(Γ, Γ(cid:48)), 3 + 3,∅,{(Γ, Γ(cid:48))}(cid:1). √ , 12 We now "run the odometer" beginning with the tautological classification statement We first enumerate all the depth 1 extensions of Γ and Γ(cid:48) with norms less than 3 + 3. Using the notation of §2.3, we have √ (cid:110) √ O3+ 3(Γ) = , , , , , (cid:111) , , , , , , , , (cid:111) 3(Γ(cid:48)) =(cid:8) √ O3+ Now, for the depth one extensions of the bigraph pair (Γ, Γ(cid:48)), we take one graph from each list, subject to the condition that the two graphs have the same number of vertices at the new (odd) depth. Since the depth we're adding is an odd depth, there is no need to choose an involution of the new vertices specifying dual data. On the other hand, it is important that in at least one of the lists O3+ 3(Γ) and 3(Γ(cid:48)) we include as distinct elements graphs which differ by a permutation of O3+ the vertices at the new depth, since the duality for odd vertices is given by their vertical ordering. In the lists above, it is convenient to remove such redundancies in the first list, but not in the second, as there are no redundancies there anyway. √ √ Next, we apply the associativity test of Lemma 3.1, remembering to only use the local version where at least one of the vertices V and W are not at the newly introduced depth and at least one is not the root vertex. This cuts down the previous large list to just two bigraph pairs, (cid:110)(cid:16) (cid:16) W1 = , , (cid:17) (cid:17)(cid:111) , . As an example of how we rule out all the others, the bigraph pair (cid:32) (cid:33) , fails the associativity test with the marked vertices: there is one vertex satisfying the conditions for Z in Lemma 3.1, but none satisfying the conditions for U . Having found all equal extensions, we now look for unequal extensions of (Γ, Γ(cid:48)) to add to the list of vines and apply the associativity test for vines (where we allow V and W to both be at the largest depth). Since the new depth we're adding is odd, there can be no unequal extensions. (However, notice that unequal extensions are easy to enumerate, they're of the form (Γ,G(cid:48)) for G(cid:48) ∈ O3+ 3(Γ(cid:48)) or of the form √ (G, Γ(cid:48)) for G ∈ O3+ √ 3(Γ)). 13 VWFinally, the old weed (Γ, Γ(cid:48)) is added to the list of vines. So we can apply the associativity test to this as a vine and see that it fails. classification statement (cid:0)(Γ, Γ(cid:48)), 3 + √ (cid:1). The progress of the odometer so far 3,∅,W1 At this point we've successfully run the odometer one step. We now have the is summarized by the following figure. The current weeds are highlighted in the figure in red. The old weed (Γ, Γ(cid:48)) is left in the picture to explain what happened in the odometer at earlier steps (in this case the zeroth step) and is necessary for recovering the vines (see below). We now run the odometer another step. This means we go through the above process for each of the two weeds. One of the weeds has no extensions which pass the associativity test. The other weed has two extensions. Furthermore there are two vines which pass, one of which is unequal. (cid:16) At this point we have proved the classification statement √ (Γ, Γ(cid:48)), 3 + (cid:110)(cid:16) (cid:16) (cid:110)(cid:16) (cid:16) 3, , , , , (cid:17) (cid:17)(cid:111) , (cid:17) (cid:17)(cid:111)(cid:17) and the progress of the odometer is summarized by the following figure. 14 Running the odometer the next step gets us down to a single weed, summarized by the figure On the next step of the odometer, there are no weeds that pass the associativity test. At this point the odometer stops, with nothing more to do. So our final result is the classification statement (Γ, Γ(cid:48)), 3 + (cid:16) √ 3, (cid:110)(cid:16) (cid:16) (cid:16) (cid:16) ∅(cid:17) , , , , (cid:17) (cid:17) , , (cid:17) (cid:17)(cid:111) , whose proof is summarized by the following figure. To read a classification statement off from a figure do the following: • The weeds are the graph pairs highlighted in red. • The equal vines are the graph pairs not highlighted which in addition pass the associativity test for pairs of vertices at the largest depth (associativity for other pairs of vertices has already been checked). • The unequal vines are found by looking at unequal extensions of the non- highlighted pairs and checking associativity. Note that if a graph pair passes the full associativity test, then any unequal extension of it automatically fails the associativity test. Furthermore, it is often easy to quickly deduce by hand the unequal extensions which pass the associativity test rather than enumerating then testing all of them. Finally, we include a brief tutorial on using the Mathematica package called FusionAtlas' (written by the authors along with Dave Penneys, Emily Peters and James Tener) to perform these calculations. First, you'll need a copy of the package. The best way to obtain this is to first install 'Subversion', then type at the command line 15 svn checkout http://tqft.net/svn/FusionAtlas/ This will create a FusionAtlas directory in your current directory. Now, in Mathematica, we need to load the package. First, we add it to the path, with a command like In[1]:= AppendTo[$Path, " /FusionAtlas/"] (if you downloaded the package somewhere outside your home directory, you'll need to adjust this path). Next, we load the package, with In[2]:= <<FusionAtlas' (note the backtick at the end of the line). You should see a message saying the package has been successfully loaded. The most powerful command for running the odometer is FindBigraphPairExtensionsUpToDepth[L][g1,g2,k] √ Here L is the graph norm limit we're working with (the square root of the index), which should be some real number strictly larger than the limit we're really interested 5 + 10−3 throughout. Using a slightly higher limit means that in. We've used sometimes spurious results will be returned with index above 5, that have to be removed. The parameters g1 and g2 are the bigraphs with dual data that we want to extend. You can find a description of the syntax for specifying bigraphs on the FusionAtlas' web page, at http://tqft.net/wiki/Atlas_of_subfactors. All bigraphs have a representation as a string (you can find many examples by looking at the LATEX source of the article, from the arXiv: all the diagrams are generated automatically from these strings), and you can generate the appropriate expressions in Mathematica from these strings using the function GraphFromString, and display bigraphs using DisplayBigraph, for example In[3]:= g1 = GraphFromString["bwd1v1v1v1p1v1x0p0x1v1x0p0x1p0x1duals1v1v1x2v2x1x3"] Out[3]= BigraphWithDuals[ GradedBigraph[{{1}}, {{1}}, {{1}}, {{1}, {1}}, {{1, 0}, {0, 1}}, {{1, 0}, {0, 1}, {0, 1}}], DualData[{1}, {1}, {1, 2}, {2, 1, 3}]] In[4]:= g2 = GraphFromString["bwd1v1v1v1p1v0x1p0x1v0x1duals1v1v1x2v1"] Out[4]= BigraphWithDuals[ GradedBigraph[{{1}}, {{1}}, {{1}}, {{1}, {1}}, {{0, 1}, {0, 1}}, {{0, 1}}], DualData[{1}, {1}, {1, 2}, {1}]] In[5]:= DisplayBigraph[g2] Out[5]= The final parameter k specifies the maximum number of times to run the odometer; ∞ is an allowed value, although in that case there is no guarantee of termination. The output of FindBigraphPairExtensionsUpToDepth is a list of three lists. The first list consists of all the vines produced, the second list consists of all the weeds produced, and in this mode the third list is always empty. Thus we can run 16 In[6]:= {vines, weeds, {}} = FindBigraphPairExtensionsUpToDepth[Sqrt[3+Sqrt[3]]+10−3 ][g1,g2,∞]; In[7]:= DisplayBigraph /@ vines Out[7]=(cid:110)(cid:16) (cid:16) (cid:16) (cid:16) (cid:17) (cid:17) , , (cid:17) (cid:17)(cid:111) , , , , In[8]:= weeds Out[8]= {} exactly reproducing the classification described above. Finally, it is possible to tell FindBigraphPairExtensionsUpToDepth to not run the odometer further on some specific weeds, by supplying a fourth argument "Weeds" -> { {h1, h2}, {h3, h4}, ... }, where {h1, h2} are a bigraph pair, etc. Now the third list in the output repeats back these weeds which were not run further. Since this paper was finished, we have written an independent and much faster reimplementation of the odometer in the language Scala. Happily, it produces all the same results as those described here (and does so in under a minute on a MacBook Pro). This reimplementation is not yet in a state that others can readily use, but we are happy to share the code and expect that it will soon be more accessible. 5 1-supertransitive subfactors In this section we use ad hoc methods to eliminate all 1-supertransitive subfactors of index between 4 and 5. This reduces the size of the odometer output drastically, simply because there are far fewer 3-supertransitive graphs below index 5 than there are 1-supertransitive graphs below index 5. The crux of the proof is that no subfactor of index between 4 and 5 can have an intermediate subfactor. Theorem 5.1 There are no 1-supertransitive subfactors with index strictly between 4 and 5. Proof Suppose that all of the objects at depth 2 have dimension 1. Then the index is an integer, and so does not lie between 4 and 5. Suppose that some object at depth 2 has dimension 1, but that not all objects at depth 2 have dimension 1. Then there must exist an intermediate subfactor. However, there are no numbers between 4 and 5 which are the product of two allowed indices. 17 Let X be the fundamental object. Suppose that there is an object V at depth 2 with dimension bigger than 1 but less than 2. Consider the connected component of X in the fusion graph for tensoring on the left with V . (Be careful here: usually we talk about principal graphs for tensoring on the right with an object.) The Frobenius-Perron eigenvalue for this graph is dim V < 2, hence this graph must be an ADET type graph. Furthermore, the Frobenius-Perron eigenvector is, up to scaling, given by the dimensions of the fundamental object and the other objects which come up in the fusion graph. From the principal graph we see that there's a non-zero map from X ⊗ X∗ → V . Hence by Frobenius reciprocity, there's a nonzero map X → V ⊗ X . Therefore in the fusion graph, X must have a self-loop. Hence X must be at the loop end of a type T graph. Let Y0, Y1, . . . be the other objects in this graph with Y0 all the way at the non-loop end. Then the normalized Frobenius-Perron eigenvector is (1, dim Y1/ dim Y0, . . . , dim X/ dim Y0). If the fusion 5)/2. If dim Y0 = 1, then the index, graph is T2 we see that dim X/ dim Y0 = (1 + (dim X)2 , is less than 4, while if dim Y0 is at least 2 then dim X is larger than 5. If the fusion graph is Tk for k > 2, then (dim X)2 > (dim X/ dim Y0)2 > 5. √ √ Thus there are at least two objects at depth 2 both of which have dimension at least 2, hence the index is at least 1 + 2 + 2 = 5. We note that in order to extend our work here to include index equal to 5, one would have to classify by hand the 1-supertransitive graphs at index exactly 5, by extending the methods of the proof here. 6 Main result We now use the techniques described in the three previous sections to develop classification statements for all subfactors with index between 4 and 5. In particular, we obtain the following classification statement with a 'manageable' set of weeds. Theorem 6.1 Subfactors with index between 4 and 5 are described by the classifi- cation statement ), 5,V∞,W∞) with V∞ = (cid:110)(cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (( , (cid:17) , , , , , , , , (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) , , , , , , 18 (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) , , , , , (cid:17) (cid:17) , , (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) , , , , , (cid:17) (cid:17) (cid:17) , , , , , , , , , , , , , , , , , , , , , , , , , , , , , , 19 (cid:17) (cid:17) (cid:17) , , , (cid:17) (cid:17) , , (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) , , , , , , , (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:110)C = F = B = Q = Q(cid:48) = (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) and W∞ = , , , , , , , , , , , , , , (cid:17) (cid:17) (cid:17) (cid:17) , , , , (cid:17) (cid:17) , , (cid:17) (cid:17) (cid:17)(cid:111) , , (cid:17) , (cid:17) (cid:17) , , (cid:17) (cid:17)(cid:111) The vines can all be eliminated (with four exceptions: Haagerup, Asaeda-Haagerup, extended Haagerup, and 2221) by showing that the indices are non-cyclotomic by applying the results of [CMS10]. This will be done in an upcoming paper [PT10] by Dave Penneys and James Tener. Furthermore, a forthcoming joint paper[MPPS10] with Dave Penneys and Emily Peters shows that there are no subfactors with principal graphs starting like B , C or F , at any index. A forthcoming paper [IJMS] joint with Masaki Izumi and Vaughan Jones shows that there are no subfactors with principal graphs starting like Q or Q(cid:48) at any index other than the 3331 principal graph. The rest of this section is dedicated to the proof of Theorem 6.1. First, in §6.1 we produce an initial list of weeds, which begin with either a triple point or a quadruple point. In §6.2, we then run the odometer for a single step extending all of the triple point weeds by one depth. We then apply the triple point obstruction from §3.2 to rule out many of the resulting weeds. We then run the odometer on all the surviving triple points weeds. In most cases we can iterate this until no more weeds survive, but for one weed with annular multiplicities 10 (the one responsible for the Haagerup, extended Haagerup, and Asaeda-Haagerup subfactors) we have to stop the odometer by hand, leaving a set of three more complicated weeds. In §6.3 we run the odometer on all the quadruple point weeds, stopping again with a more complicated set of weeds. In §6.3.2 we rule out certain of these weeds by hand. The full list of vines and weeds in Theorem 6.1 above is assembled out of all the vines and weeds produced in the various subsections. 20 6.1 Initial seeds Lemma 6.2 We have the classification statement ), 5,∅,W1) (( , with W1 = (cid:110) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) Γo1,a = Γo1,b = Γo1,c = Γe1 = Γo2,a = Γo2,b = Γo2,c = Γe2 = (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17)(cid:111) , , , , , , , , Proof By Theorem 5.1, we can assume that any odd-supertransitive bigraphs with index less than 5 are at least 3-supertransitive. Since the 5-pointed star and the vertex with a single and double edge coming into it both have norm 5, it follows that there are only two possibilities at the first non-trivial depth: we have either a vertex of valence 3 or a vertex of valence 4. By the associativity test both graphs must have the same valency. The naming scheme Γxn,m indicates whether the branch point is at an odd or even depth, according to whether x = o or x = e, the number n denotes the annular multiplicity at the depth one past the branch point, and then m is an arbitrary alphabetic index. We immediately rule out the weeds Γo1,b and Γo2,b on the basis of Lemma 3.6. The proof of Theorem 6.1 will follow by successively applying the odometer and removing graphs which fail the obstructions. For ease of exposition we will split up into cases based on the form of the initial branch point. 6.2 Triple points We now run the odometer for one step on each of Γo1,a , Γo1,c and Γe1 , obtaining Lemma 6.3 There are classification statements (cid:16) Γo1,a, 5,{Γo1,a} , (cid:110)(cid:16) , 21 (cid:17) , (6.1) , , , , , , , (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17)(cid:111)(cid:17) (cid:17) (cid:17) (cid:17) (cid:17)(cid:111)(cid:17) , , , , , , (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:110)(cid:16) (cid:16) (cid:16) (cid:16) Γo1,c, 5,{Γo1,c} , (cid:16) and (cid:16) Γe1, 5,{Γe1} , (cid:110)(cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) , , , , , , , , , , , , , , , , , , , , , , , , , 22 (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) , , , , , , , , , , , (6.2) (6.3) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) , , , , , , , , , , , , , , , , , , (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17)(cid:111)(cid:17) , , , , , , , We now apply the triple point obstruction to rule out many of these weeds. Lemma 6.4 There are no subfactors with principal graph pairs starting like any of the last 6 weeds in Equation (6.1), any of the last 2 weeds in Equation (6.2) or any of the last 20 weeds in Equation (6.3). Proof This is an immediate consequence of Corollary 3.5 (for the weeds in Equations (6.1) and (6.2)) and of Corollary 3.4 (for the weeds in Equation (6.3)). The standard example is the second last pair from Equation (6.1), which was treated by Haagerup. In our notation, A1 = A(cid:48) 2 , so the pair is forbidden. At this point, it is convenient to partition the remaining weeds which begin with a triple point according to their annular multiplicities. Thus we define W12 = 1 and A2 = A(cid:48) (cid:110)(cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:17) (cid:17) (cid:17) (cid:17) , , , , , , , , , , 23 (cid:17) (cid:17) , , (cid:16) (cid:110)(cid:16) (cid:16) (cid:16) (cid:110)(cid:16) W11 = W10 = (cid:111) (cid:17) , (cid:17)(cid:111) , (cid:17) (cid:17) , , (cid:17)(cid:111) , , , , , and We deal with these sets in the next three subsections. 6.2.1 Annular multiplicities 12 For each of the weeds in W12 , we run the odometer one more step and find that there are no remaining weeds. None of the bigraph pairs in W12 have a equal or unequal extensions, and only the bigraph pairs (cid:16)(cid:16) (cid:16) (cid:16) (cid:16) (cid:17) (cid:17) , , , , , , (cid:17) (cid:17)(cid:17) , survive the associativity test, so these are the only vines with annular multiplicities 12. 6.2.2 Annular multiplicities 11 For each of the weeds in W11 , we can run the odometer, eventually removing all weeds. We have the following classification statements: (cid:16)(cid:16) (cid:16)(cid:16) (cid:16)(cid:16) and where V11,a = V11,b = (cid:110)(cid:16) (cid:110)(cid:16) (cid:16) , , , 5,V11,a,∅(cid:17) (cid:17) , 5,V11,b,∅(cid:17) (cid:17) , 5,V11,c,∅(cid:17) (cid:17) (cid:17)(cid:111) (cid:17) , , , , , , 24 , , (cid:17)(cid:111) , (cid:110)(cid:16) and V11,c = , (cid:17)(cid:111) . See Figures 1, 2 and 3 for the detailed output of the odometer in each of these three cases. Figure 1: The odometer, running on Γ11,a . Figure 2: The odometer, running on Γ11,b . Figure 3: The odometer, running on Γ11,c . 6.2.3 Annular multiplicities 10 The single weed in W10 causes more difficulty. It appears that running the odometer never terminates -- weeds with arbitrarily high depth appear. Thus, we choose a 25 certain set of weeds to terminate at, balancing the desire for a small list of weeds which are as deep as possible with the desire for a small list of vines. The particular choices we've made were influenced by our expectation of various methods eliminating vines or weeds in subsequent papers, in particular the quadratic tangles methods which we will use to eliminate certain weeds with annular multiplicities 10. Theorem 6.5 There is a classification statement (cid:16)(cid:16) (cid:17) (cid:17) , 5,V10,W10 , (cid:17) (cid:17) (cid:17) , , , (cid:17) , (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) , , , , , , , (cid:17) , (cid:17) (cid:17) (cid:17) (cid:17) , , , , (cid:17) (cid:17)(cid:111) , (cid:17) (cid:17)(cid:111) , where V10 = and W10 = (cid:110)(cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:110)(cid:16) (cid:16) (cid:16) (cid:17) , , , , , , , , , , , , , , , , , , , , , , 26 Proof See Figure 4. 6.3 Quadruple points In the next subsection we run the odometer on all of the quadruple point weeds. In each case, the odometer runs forever, so we carefully choose a convenient set of stopping points, thus exchanging the current list of weeds for another slightly longer list of more complicated weeds, along with several vines. In §6.3.2 we then rule out several of these more complicated weeds by specialized methods, again producing several more vines. 6.3.1 Running the odometer Theorem 6.6 There's a classification statement with and (cid:110)(cid:16) (cid:16) (cid:110)(cid:16) (cid:16) Vo2,a = Wo2,a = Proof See Figure 5. (Γo2,a, 5,Vo2,a,Wo2,a) , , , , (cid:17) , (cid:17)(cid:111) (cid:17) , (cid:17)(cid:111) Theorem 6.7 There's a classification statement with and (cid:110)(cid:16) (cid:16) (cid:110)(cid:16) Vo2,c = Wo2,c = Proof See Figure 6. (cid:17) , (cid:17)(cid:111) (cid:17)(cid:111) (Γo2,c, 5,Vo2,c,Wo2,c) , , , 27 Theorem 6.8 There's a classification statement (Γe2, 5,Ve2,We2) with Ve2 = and We2 = (cid:110)(cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:16) (cid:110)(cid:16) (cid:16) (cid:16) Proof See Figure 7. (cid:17) , (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) (cid:17) , , , , , , , , , , , , , , , , , 6.3.2 Killing quadruple point weeds Finally, we kill some of the remaining weeds, in particular , from Wo2,a and all three weeds from We2 : (cid:16) (cid:16) (cid:16) (cid:16) The second weed on that list is exactly what is ruled out by the even quadruple point obstruction in Theorem 3.7. The others will take a bit more work. Theorem 6.9 Any subfactor with index below 5 with principal graph pair starting like must be a translate of one of the following graph pairs: , (cid:17) (cid:16) (cid:110)(cid:16) (cid:17) , (cid:17)(cid:111) (cid:17) , (cid:17) (cid:17)(cid:111) , (cid:17) (cid:17) , (cid:17) (cid:17) , . , , , , 28 (cid:16) , (cid:17)(cid:111) Proof We'll do two cases, first without any translation, and then with any non- trivial translation. Without any translation, the univalent vertex at depth 2 past the quadruple point has dimension [4] 5 ) < [4] [3] < 2 cos( π Next, we find that we can just run the odometer on √ [3] . Now q must lie between 1.59 and 1+ 2 6 ) < 2, which is not an allowed dimension. and we see 1 < 2 cos( π 5 (cid:16) Γ5321 = (cid:17) , and obtain the classification statement (Γ5321, 5,VΓ5321,∅) where VΓ5321 is the list of graph pairs appearing in the statement of the theorem. The output of the odometer appears in Figure 8. Theorem 6.10 Any subfactor with index below 5 with principal graph pair starting like must be a translate of one of the following graph pairs: , , (cid:17) (cid:16) (cid:110)(cid:16) (cid:16) (cid:16) (cid:17) , (cid:17) , (cid:17)(cid:111) , , , (cid:17) (cid:16) Proof First, observe that without any translation, the univalent vertex at depth 3 has dimension [3] [2] . We know that for any extension of Γ, q must lie between 1.56 √ and 1+ 7 ) < 2, which is not an allowed 2 dimension. . In this range, 1 < 2 cos( π [2] < 2 cos( π 6 ) < [3] 5 Next, we increase the supertransitivity by two and run the odometer on Γ4321 = , but only for a few addi- , tional depths, obtaining the classification statement (Γ4321, 5,VΓ4321,{Γ4621}) (cid:16) with VΓ4321 the list of graph pairs appearing in the statement of the theorem and Γ4621 = , (cid:17) The output of the odometer appears in Figure 9. For the new weed we've produced, Γ4621 , we again split into cases. Without any translation, we look at the vertex at depth 10 and see that it has dimension p(q) = q −10 − q −8 − 2q −6 − 3q −4 − 4q −2 − 6 − 4q2 − 3q4 − 2q6 − q8 + q10. 29 The largest real root of p is at approximately 1.61501, so for any q in the relevant √ range, namely 1.6161 < q < 1+ ) = 1, 2 so in fact the dimension of this vertex is always strictly less than 1, which is not allowed. √ , p is strictly increasing. Moreover, p( 1+ 2 5 5 Translating Γ4621 by two, we find that the index of the graph is approximately 5.0062 > 5. Theorem 6.11 No subfactor with index below 5 has principal graph which starts like . Proof If you translate this graph by 1 then its graph norm is 5 (and in fact the graph is the principal graph for the group-subgroup factor A4 ⊂ A5 ), so we need only consider extensions (rather than translates of extensions). √ We need only consider q between 1.59438 and 1+ . In this range, at least one of 2 the two vertices at depth 5 will have dimension less than 1, which is a contradiction. 5 To see this, notice that the sum of the dimensions of the two vertices at depth 5 is 1 − 3q4 − 5q6 − 3q8 + q12 q5 + q7 . This is equal to 2 near q = 0.61492 and q = 1.62623 and smaller than 2 between those two values. In particular, it is smaller than 2 in the range that we are considering. 7 Future directions As we explained in the introduction, this paper is the first step towards classifying subfactors of index less than 5. We complete this classification in a series of subse- quent papers. This project was developed at several Planar Algebra Programming Camps organized by the authors and Emily Peters, and hosted by Vaughan Jones. Further progress was made during a visit by the current authors with Masaki Izumi at Kyoto University. As a result the subsequent papers in the project have a variety of different authors. All translates of the vines in our classification can all be eliminated (with four exceptions, corresponding to subfactors that actually exist: Haagerup, Asaeda- Haagerup, extended Haagerup, and 2221) by showing that the indices are non- cyclotomic by applying the results of [CMS10]. This will be done in a forthcoming paper [PT10] by David Penneys and James Tener. There's not yet a uniform approach to eliminating the weeds, which we instead deal with separately. In a joint paper with David Penneys and Emily Peters, we will prove that there are no subfactors (of any index) whose principal graphs begin like B , C , or F . This paper uses several different "triple point obstructions" coming from the theory of connections and from an identity proved by Vaughan Jones in [Jon03]. The quadratic tangles identity does not eliminate the weed B as there the 30 rotational eigenvalue is −1 and the identity is automatically satisfied. However, we are able to apply an ad hoc connections argument to remove that case. In a forthcoming paper with Masaki Izumi and Vaughan Jones we will show that the only subfactor with principal graphs starting like either Q or Q(cid:48) is the 3311 subfactor, which is unique up to taking duals. The quadratic tangles technique from [Jon03] can also be applied to graphs which begin with a quadruple point, and this approach readily rules out subfactors with principal graphs beginning like Q. A connections argument due to Izumi, followed by a number theoretic argument along the lines of [CMS10] shows that any principal graph starting like Q(cid:48) must in fact be the 3331 graph. An involved argument then establishes that any subfactor with this principal graph must be (up to taking dual) the GHJ subfactor. The final piece of the classification is the uniqueness of the 2221 subfactor. This has recently been established in the Ph.D. thesis of Richard Han [Han10], who was able to derive a set of generators and relations for the corresponding planar algebra directing from the principal graphs. A natural related question is to consider subfactors of index exactly 5. Our tech- niques generalize easily to understanding possible principal graphs at index equal to 5, however, the uniqueness problem becomes more difficult. The techniques for constructing subfactors of integer index are somewhat different from those of non-integer index (in particular, Hopf algebraic techniques, and group cohomology) so we have avoided dealing with the index 5 case in detail. However, as was pointed out to us by Izumi, using the results from [Izu97] simplifies the situation immensely. We now expect to be able to extend our classification to index equal to 5. References [AH99] Marta Asaeda and Uffe Haagerup. Exotic subfactors of finite depth with Jones 17)/2. Comm. Math. Phys., 202(1):1 -- 63, 1999. indices (5 + MR1686551 DOI:10.1007/s002200050574 arXiv:math.OA/9803044. 13)/2 and (5 + √ √ [Asa07] Marta Asaeda. subfactors. of DOI:10.1142/S0129167X07003996 arXiv:math.OA/0605318. Internat. J. Math., Galois groups and an obstruction to principal graphs MR2307421 18(2):191 -- 202, 2007. [AY09] Marta Asaeda and Seidai Yasuda. On Haagerup's list of potential principal graphs of subfactors. Comm. Math. Phys., 286(3):1141 -- 1157, 2009. MR2472028 DOI:10.1007/s00220-008-0588-0 arXiv:0711.4144. [Bis94] Dietmar Bisch. An example of an irreducible subfactor of the hyperfinite II1 factor with rational, noninteger index. J. Reine Angew. Math., 455:21 -- 34, 1994. arXiv:MR1293872. [Bis98] Dietmar Bisch. Jones DOI:http://dx.doi.org/10.1007/s002080050185. index. Principal Math. Ann., graphs 311(2):223 -- 231, of subfactors 1998. with small MR1625762 [BJ97] Dietmar Bisch and Vaughan Jones. Algebras associated to intermediate subfactors. Invent. Math., 128(1):89 -- 157, 1997. MR1437496. [BMPS09] Stephen Bigelow, Scott Morrison, Emily Peters, and Noah Snyder. Constructing the extended Haagerup planar algebra, 2009. arXiv:0909.4099, to appear Acta Mathematica. 31 [BNP07] Dietmar Bisch, Remus Nicoara, and Sorin Popa. Continuous families of hyperfinite subfactors with the same standard invariant. Internat. J. Math., 18(3):255 -- 267, 2007. MR2314611 arXiv:math.OA/0604460 DOI:10.1142/S0129167X07004011. [BW99] John W. Barrett and Bruce W. Westbury. Adv. Math., 143(2):357 -- 375, DOI:10.1006/aima.1998.1800. 1999. categories. MR1686423 arXiv:hep-th/9310164 Spherical [CMS10] Frank Calegari, Scott Morrison, and Noah Snyder. Cyclotomic integers, fusion categories, and subfactors, 2010. With an appendix by Victor Ostrik. To appear in Communications in Mathematical Physics. arXiv:1004.0665. [GdlHJ89] Frederick M. Goodman, Pierre de la Harpe, and Vaughan F. R. Jones. Coxeter graphs and towers of algebras, volume 14 of Mathematical Sciences Research Institute Publications. Springer-Verlag, New York, 1989. MR999799. [GL98] John J. Graham and Gus I. Lehrer. The representation theory of affine Temperley- Lieb algebras. Enseign. Math. (2), 44(3-4):173 -- 218, 1998. MR1659204. √ [Haa94] Uffe Haagerup. Principal graphs of subfactors in the index range 4 < [M : N ] < 3 + 2. In Subfactors (Kyuzeso, 1993), pages 1 -- 38. World Sci. Publ., River Edge, NJ, 1994. MR1317352 available at http://tqft.net/other-papers/subfactors/ haagerup.pdf. [Han10] Richard Han. A Construction of the 2221 Planar Algebra. PhD thesis, University of California, Riverside, 2010. arXiv:1102.2052. [IJMS] Masaki Izumi, Vaughan F. R. Jones, Scott Morrison, and Noah Snyder. Classification of subfactors of index less than 5, part 3: quadruple points. In preparation. [Izu91] Masaki Izumi. Application of fusion rules to classification of subfactors. Publ. Res. Inst. Math. Sci., 27(6):953 -- 994, 1991. MR1145672 DOI:10.2977/prims/1195169007. [Izu97] Masaki Izumi. Goldman's type theorems in index theory. In Operator algebras and quantum field theory (Rome, 1996), pages 249 -- 269. Int. Press, Cambridge, MA, 1997. MR1491121. [Izu01] Masaki Izumi. The structure of sectors associated with Longo-Rehren in- MR1832764 clusions. II. Examples. Rev. Math. Phys., 13(5):603 -- 674, 2001. DOI:10.1142/S0129055X01000818. [Jon] Vaughan F. R. Jones. Planar algebras, I. arXiv:math.QA/9909027. [Jon80] Vaughan F. R. Jones. Actions of finite groups on the hyperfinite type II1 factor. Mem. Amer. Math. Soc., 28(237):v+70, 1980. MR715556. [Jon83] Vaughan F. R. Jones. Index for subfactors. Invent. Math., 72(1):1 -- 25, 1983. MR696688 DOI:10.1007/BF01389127. [Jon01] Vaughan F. R. Jones. The annular structure of subfactors. In Essays on geometry and related topics, Vol. 1, 2, volume 38 of Monogr. Enseign. Math., pages 401 -- 463. Enseignement Math., Geneva, 2001. MR1929335. [Jon03] Vaughan F. R. Jones. Quadratic tangles in planar algebras, 2003. arXiv:1007.1158. [JR06] Vaughan F. R. Jones and Sarah A. Reznikoff. Hilbert space representations of the annular Temperley-Lieb algebra. Pacific J. Math., 228(2):219 -- 249, 2006. MR2274519 DOI:10.2140/pjm.2006.228.219. [Kaw95] Yasuyuki Kawahigashi. Classification of paragroup actions in subfactors. Publ. Res. Inst. Math. Sci., 31(3):481 -- 517, 1995. MR1355948. [KO02] Alexander Kirillov, Jr. and Viktor Ostrik. the McKay correspondence and the ADE classification of sl2 conformal field theo- ries. Adv. Math., 171(2):183 -- 227, 2002. MR1936496 arXiv:math.QA/0101219 DOI:10.1006/aima.2002.2072. On a q -analogue of 32 [MPPS10] Scott Morrison, David Penneys, Emily Peters, and Noah Snyder. Classification of subfactors of index less than 5, part 2: triple points, 2010. arXiv:1007.2240. [Ocn80] Adrian Ocneanu. Actions des groupes moyennables sur les alg`ebres de von Neumann. C. R. Acad. Sci. Paris S´er. A-B, 291(6):A399 -- A401, 1980. MR596082. [Ocn88] Adrian Ocneanu. Quantized groups, string algebras and Galois theory for alge- bras. In Operator algebras and applications, Vol. 2, volume 136 of London Math. Soc. Lecture Note Ser., pages 119 -- 172. Cambridge Univ. Press, Cambridge, 1988. MR996454. [Ocn94] Adrian Ocneanu. Chirality for operator algebras. In Subfactors (Kyuzeso, 1993), pages 39 -- 63. World Sci. Publ., River Edge, NJ, 1994. MR1317353. [Pop90] Sorin Popa. Classification of subfactors: the reduction to commuting squares. Invent. Math., 101(1):19 -- 43, 1990. MR1055708 DOI:10.1007/BF01231494. [Pop91] Sorin Popa. Subfactors and classification in von Neumann algebras. In Proceedings of the International Congress of Mathematicians, Vol. I, II (Kyoto, 1990), pages 987 -- 996, Tokyo, 1991. Math. Soc. Japan. MR1159284. [Pop93] Sorin Popa. Markov traces on universal Jones algebras and subfactors of finite index. Invent. Math., 111(2):375 -- 405, 1993. MR1198815 DOI:10.1007/BF01231293. [Pop94] Sorin Popa. Classification of amenable subfactors of type II. Acta Math., 172(2):163 -- 255, 1994. MR1278111 DOI:10.1007/BF02392646. [Pop95] Sorin Popa. An axiomatization of the lattice of higher relative commutants of a sub- factor. Invent. Math., 120(3):427 -- 445, 1995. MR1334479 DOI:10.1007/BF01241137. [PT10] David Penneys and James Tener. Classification of subfactors of index less than 5, part 4: cyclotomicity, 2010. arXiv:1010.3797. [TV92] Vladimir G. Turaev and Oleg Ya. Viro. State sum invariants of 3-manifolds and quantum 6j -symbols. Topology, 31(4):865 -- 902, 1992. MR1191386. This paper is available online at arXiv:1007.1730, and at http://tqft.net/ index5-part1. 33 Figure 4: The odometer, running on Γ10 . 34 Figure 5: The odometer, running on Γo2,a . Figure 6: The odometer, running on Γo2,c . 35 Figure 7: The odometer, running on Γe2 . Figure 8: The odometer, running on Γ5321 . Figure 9: The odometer, running on Γ4321 . 36
1210.6768	2	1210	2013-11-18T11:56:07	Symmetries of L\'evy processes on compact quantum groups, their Markov semigroups and potential theory	[ "math.OA", "math.PR" ]	Strongly continuous semigroups of unital completely positive maps (i.e. quantum Markov semigroups or quantum dynamical semigroups) on compact quantum groups are studied. We show that quantum Markov semigroups on the universal or reduced C${}^$-algebra of a compact quantum group which are translation invariant (w.r.t. to the coproduct) are in one-to-one correspondence with L\'evy processes on its $$-Hopf algebra. We use the theory of L\'evy processes on involutive bialgebras to characterize symmetry properties of the associated quantum Markov semigroup. It turns out that the quantum Markov semigroup is GNS-symmetric (resp. KMS-symmetric) if and only if the generating functional of the L\'evy process is invariant under the antipode (resp. the unitary antipode). Furthermore, we study L\'evy processes whose marginal states are invariant under the adjoint action. In particular, we give a complete description of generating functionals on the free orthogonal quantum group $O_n^+$ that are invariant under the adjoint action. Finally, some aspects of the potential theory are investigated. We describe how the Dirichlet form and a derivation can be recovered from a quantum Markov semigroup and its L\'evy process and we show how, under the assumption of GNS-symmetry and using the associated Sch\"urmann triple, this gives rise to spectral triples. We discuss in details how the above results apply to compact groups, group C$^*$-algebras of countable discrete groups, free orthogonal quantum groups $O_n^+$ and the twisted $SU_q (2)$ quantum group.	math.OA	math	SYMMETRIES OF L´EVY PROCESSES ON COMPACT QUANTUM GROUPS, THEIR MARKOV SEMIGROUPS AND POTENTIAL THEORY FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA Abstract. Strongly continuous semigroups of unital completely positive maps (i.e. quantum Markov semigroups or quantum dynamical semigroups) on com- pact quantum groups are studied. We show that quantum Markov semigroups on the universal or reduced C∗-algebra of a compact quantum group which are translation invariant (w.r.t. to the coproduct) are in one-to-one correspon- dence with L´evy processes on its ∗-Hopf algebra. We use the theory of L´evy processes on involutive bialgebras to characterize symmetry properties of the associated quantum Markov semigroup. It turns out that the quantum Markov semigroup is GNS-symmetric (resp. KMS-symmetric) if and only if the gen- erating functional of the L´evy process is invariant under the antipode (resp. the unitary antipode). Furthermore, we study L´evy processes whose marginal states are invariant under the adjoint action. In particular, we give a complete description of generating functionals on the free orthogonal quantum group O+ n that are invariant under the adjoint action. Finally, some aspects of the potential theory are investigated. We describe how the Dirichlet form and a derivation can be recovered from a quantum Markov semigroup and its L´evy process and we show how, under the assumption of GNS-symmetry and using the associated Schurmann triple, this gives rise to spectral triples. We discuss in details how the above results apply to compact groups, group C∗-algebras of countable discrete groups, free orthogonal quantum groups O+ n and the twisted SUq(2) quantum group. Contents 1. Introduction 2. Preliminaries 3. Translation invariant Markov semigroups 4. GNS-Symmetry and KMS-symmetry of convolution operators 5. Schurmann triples corresponding to KMS-symmetric generators 6. Generating functionals invariant under adjoint action 7. Dirichlet forms 8. Derivations, cocycles and Spectral Triples 2 3 9 13 18 20 26 30 2010 Mathematics Subject Classification. 20G42,43A05,46L57,60G51,81R50. Key words and phrases. Compact quantum group, Dirichlet form, L´evy process, quantum Markov semigroup, spectral triple. 1 2 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA 9. Two classical examples: commutative and cocommutative CQGs 10. Example: free orthogonal quantum groups O+ N 11. Example: Woronowicz quantum group SUq(2) Acknowledgements References 35 38 43 51 51 1. Introduction Quantum Markov semigroups, i.e. semigroups of contractive, completely posi- tive maps, are mathematical models for open quantum systems. In this paper we study a class of such semigroups on the C∗-algebras of compact quantum groups (also called Woronowicz C∗-algebras). A compact quantum group G is a unital C∗-algebra C(G) equipped with ad- ditional structure, that generalizes the C∗-algebra of continuous functions on a compact group (see Section 2 for the precise definition). In particular, any com- mutative compact quantum group is isomorphic to the C∗-algebra of continuous functions on a compact group. The quantum group structure thus plays two roles: on one hand, positive functionals on C(G) replace the states of a classical Markov process and, on the other hand, the actions of C(G) on itself, allow us to formulate important symmetry properties for quantum Markov semigroups. We show that quantum Markov semigroups on the (reduced or universal) C∗- algebra of a compact quantum group that are translation invariant w.r.t. the coproduct are in one-to-one correspondence with L´evy processes on its ∗-Hopf algebra A, see Theorems 3.2 and 3.4. This shows that the characterisation of L´evy processes in topological groups as the Markov processes which are invariant under time and space translations extends to compact quantum groups. In particular, if the compact quantum group is commutative, the associated stochastic processes reduce to L´evy processes with values in a compact group, i.e., stochastic processes with stationary and independent increments. In general, a KMS-symmetry property of a quantum Markov semigroup on a C∗-algebra with respect to given KMS state on it, allows to study the semigroup on the scale of associated noncommutative Lp-spaces. On compact quantum groups the natural state to refer to is the unique translation invariant state: it is called the Haar state because it reduces to the Haar measure of compact group when C(G) is commutative. In Section 4 we show that the quantum Markov semigroup is KMS-symmetric (with respect to the Haar state) if and only if the generating functional of its associated L´evy process is invariant under the unitary antipode, and that the quantum Markov semigroup satisfies the stronger condi- tion of GNS-symmetry if and only if the generating functional of its associated L´evy process is invariant under the antipode. In Section 5 we characterize the Schurmann triples of KMS-symmetric L´evy processes. SYMMETRIES 3 In the classical literature on Brownian motion or L´evy processes on (simple) Lie groups, the analysis of their invariance under the adjoint action of the group on itself has been particularly intense. We formulate this invariance property for compact quantum groups and show that it imposes a very strong restriction. In Section 6 we develop a method that allows to determine ad-invariant generating functionals on compact quantum groups of Kac type, i.e. when the Haar state is a trace. Using this method we find a complete classification of the ad-invariant L´evy processes on the free orthogonal quantum groups in Section 10. In Section 7 we give a complete description of the Dirichlet form associated to KMS symmetric L´evy processes and, in the GNS symmetric case, a characteri- zation of the associated quadratic forms on the Hopf algebra A, arising in this way, in terms of their translation invariance. In the framework of Alain Connes' noncommutative geometry [Co94], efforts have been directed towards the construction and investigation of Dirac operators and spectral triples on the Woronowicz quantum groups SUq(2) and related ho- mogeneous noncommutative spaces (see for example [CP03a], [CP03b], [Co04a], [Co04b], [DLSSV05]). The relevance of this point of view relies in the fact that a spectral triple allows to construct topological invariants as cyclic cocycles in cyclic cohomology and local couplings with the K-theory of the C∗-algebra (Connes- Chern character). In Section 8 we construct a Hilbert bimodule derivation, giving rise to differen- tial calculus on the compact quantum group C∗-algebra C(G), which, in the GNS symmetric case, allows to represent the Dirichlet form as a generalized Dirichlet integral E[a] = 1 2kdak2 . Using the derivation, we then construct a Dirac operator D whose spectrum is explicitly determined by the spectrum of the Dirichlet form on the GNS Hilbert space of the Haar state. Later we show that the Dirac operator D is part of spectral triple with respect to which the elements of the Hopf algebra A are Lipschitz. In the last three Sections 9, 10, 11, we discuss in detail examples of the above constructions on compact Lie groups, group C∗-algebras of countable discrete groups, the free orthogonal quantum groups O+ N and the Woronowicz quantum groups SUq(2). 2. Preliminaries Sesquilinear forms will be linear in the right entry. For an algebra A, A′ will denote the algebraic dual of A, i.e. the space of all linear functionals from A to C. For a C∗-algebra A, by A′ we will mean the dual space of all linear continuous functionals on A. The symbol ⊗ will denote the spatial tensor product of C∗- algebras and ⊙ the algebraic tensor product, see, e.g., [Ped79] for tensor products and other facts about C∗-algebras. 4 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA 2.1. Compact Quantum Groups. The notion of compact quantum groups has been introduced in [Wor87a]. Here we adopt the definition from [Wor98] (Definition 1.1 of that paper). Definition 2.1. A C∗-bialgebra (a compact quantum semigroup) is a pair (A, ∆), where A is a unital C∗-algebra, ∆ : A → A ⊗ A is a unital, ∗-homomorphic map which is coassociative, i.e. If the quantum cancellation properties (∆ ⊗ idA) ◦ ∆ = (idA ⊗ ∆) ◦ ∆. Lin((1 ⊗ A)∆(A)) = Lin((A ⊗ 1)∆(A)) = A ⊗ A, are satisfied, then the pair (A, ∆) is called a compact quantum group (CQG). If the algebra A of a compact quantum group is commutative, then A is iso- morphic to the algebra C(G) of continuous functions on a compact group G. To emphasis that for an arbitrary (i.e. not necessarily non-commutative) compact quantum group (A, ∆) the algebra A replaces the algebra of continuous func- tions on an (abstract) quantum analog of a group, the notation G = (A, ∆) and A = C(G) is also frequently used. The map ∆ is called the coproduct of A and it induces the convolution product of functionals The following fact is of fundamental importance, cf. [Wor98, Theorem 2.3]. λ ⋆ µ := (λ ⊗ µ) ◦ ∆, λ, µ ∈ A′. Proposition 2.2. Let A be a compact quantum group. There exists a unique state h ∈ A′ (called the Haar state of A) such that for all a ∈ A (h ⊗ idA) ◦ ∆(a) = h(a)1 = (idA ⊗ h) ◦ ∆(a). The left (resp. right) part of the equation above are usually referred to as left- (resp. right-) invariance property of the Haar state. In general, the Haar state of a compact quantum group need not be faithful or tracial. ∆(ujk) =Pn 2.2. Corepresentations. An element u = (ujk)1≤j,k≤n ∈ Mn(A) is called an n-dimensional corepresentation of G = (A, ∆) if for all j, k = 1, . . . , n we have p=1 ujp ⊗ upk. All corepresentations considered in this paper are sup- posed to be finite-dimensional. A corepresentation u is said to be non-degenerate, if u is invertible, unitary, if u is unitary, and irreducible, if the only matrices T ∈ Mn(C) with T u = uT are multiples of the identity matrix. Two corepre- sentations u, v ∈ Mn(A) are called equivalent, if there exists an invertible matrix U ∈ Mn(C) such that U u = vU. An important feature of compact quantum groups is the existence of the dense ∗-subalgebra A (the algebra of the polynomials of A), which is in fact a Hopf ∗-algebra -- so for example ∆ : A → A ⊙ A. With the notation G = (A, ∆), the ∗-algebra A is often denoted in the literature as Pol(G). SYMMETRIES 5 Fix a complete family (u(s))s∈I of mutually inequivalent irreducible unitary corepresentations of A, then {u(s) kℓ ; s ∈ I, 1 ≤ k, ℓ ≤ ns} (where ns denotes the dimension of u(s)) is a linear basis of A, cf. [Wor98, Proposition 5.1]. We shall reserve the index s = 0 for the trivial corepresentation u(0) = 1. The Hopf algebra structure on A is defined by ε(u(s) jk ) = δjk, S(u(s) jk ) = (u(s) kj )∗ for s ∈ I, j, k = 1, . . . , ns, (id ⊗ ε) ◦ ∆ = id = (ε ⊗ id) ◦ ∆, mA ◦ (id ⊗ S) ◦ ∆ = ε(a)1 = mA ◦ (S ⊗ id) ◦ ∆, where ε : A → C is the counit and S : A → A is the antipode. They satisfy (2.1) (2.2) (2.3) for all a ∈ A. Let us also remind that the Haar state is always faithful on A. jk ; 1 ≤ j, k ≤ ns} for s ∈ I. By [Wor98, Proposition 5.2], there exists an irreducible unitary corepresentation u(sc), called the contragredient representation of u(s), such that V ∗s = Vsc. Clearly (sc)c = s. (cid:0)S(a∗)∗(cid:1) = a Set Vs = span {u(s) We shall frequently use Sweedler notation for the coproduct of an element a ∈ A, i.e. omit the summation and the index in the formula ∆(a) =Pi a(1),i ⊗ a(2),i and write simply ∆(a) = a(1) ⊗ a(2). 2.3. The dual discrete quantum group. To every compact quantum group G = (A, ∆) there exists a dual discrete quantum group G, cf. [PW90]. For our purposes it will be most convenient to introduce G in the setting of Van Daele's algebraic quantum groups, cf. [VD98, VD03]. However, the reader should be aware that we adopt a slightly different convention for the Fourier transform. A pair (A, ∆), consisting of a ∗-algebra A (with or without identity) and a coassociative comultiplication ∆ : A → M(A⊙ A), is called an algebraic quantum group if the product is non-degenerate (i.e. ab = 0 for all a implies b = 0), if the two operators T1 : A ⊙ A ∋ a ⊗ b 7→ ∆(a)(b ⊗ 1) ∈ A ⊙ A and T2 : A ⊙ A ∋ a ⊗ b 7→ ∆(a)(1 ⊗ b) ∈ A ⊙ A are well-defined bijections and if there exists a nonzero left-invariant positive functional on A. Here, M(B) denotes the set of multipliers on B. We refer the reader to [VD98] for proofs and further details and will just recall a few facts here that we shall need later. If (A, ∆) is a compact quantum group then (A, ∆A) is an algebraic quantum group ("of compact type") and the Haar state is a faithful left- and right-invariant functional. For a ∈ A we can define ha ∈ A′ by the formula ha(b) = h(ab) for b ∈ A, where h is the Haar state, and we denote by A the space of linear functionals on A of the form ha for a ∈ A. 6 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA The set A becomes an associative ∗-algebra with the convolution of functionals as the multiplication: λ ⋆ µ = (λ ⊗ µ) ◦ ∆, and the involution λ∗(x) = λ(S(x)∗) (λ, µ ∈ A). Note that A is closed under the convolution by [VD98, Proposition 4.2]. The Hopf structure is given as follows: the coproduct ∆ is the dual of the product on A, the antipode S is the dual to S and the counit ε is the evaluation in 1. In particular, we have S(λ)(x) = λ(Sx) for λ ∈ A, x ∈ A and if ∆(λ) ∈ A⊙ A then ∆(λ)(x ⊗ y) = λ(1)(x) ⊗ λ(2)(y) = λ(xy), x, y ∈ A. The pair G = ( A, ∆) is an algebraic quantum group, called the dual of G. The linear map which associates to a ∈ A the functional ha ∈ A is called the Fourier transform. Let us note that, due to the faithfulness of the Haar state h, A separates the points of A. 2.4. Woronowicz characters and modular automorphism group. A nice introduction to this part can be found in [Wor87a, Wor98], [KS97] or [Tim08]. For a ∈ A, λ ∈ A′ we define λ ⋆ a = (id ⊗ λ)∆(a), a ⋆ λ = (λ ⊗ id)∆(a). If a ∈ A and λ ∈ A′, then λ ⋆ a, a ⋆ λ ∈ A. For a compact quantum group A with dense ∗-Hopf algebra A, there exists a unique family (fz)z∈C of linear multiplicative functionals on A, called Woronowicz characters (cf. [Wor98, Theorem 1.4]), such that a ∈ A, (1) fz(1) = 1 for z ∈ C, (2) the mapping C ∋ z 7→ fz(a) ∈ C is an entire holomorphic function for all (3) f0 = ε and fz1 ⋆ fz2 = fz1+z2 for any z1, z2 ∈ C, (4) fz(S(a)) = f−z(a) and f¯z(a∗) = f−z(a) for any z ∈ C, a ∈ A, (5) S2(a) = f−1 ⋆ a ⋆ f1 for a ∈ A, (6) the Haar state h satisfies: The formulas h(ab) = h(b(f1 ⋆ a ⋆ f1)), a, b ∈ A. ρz,z ′(a) = fz ⋆ a ⋆ fz ′, σz = ρiz,iz (2.4) define automorphisms of A, in terms of which σt = ρit,it and τt = ρit,−it, t ∈ R, define one parameter groups of automorphisms of A. The former is known as modular automorphism group. Moreover, h is the (σ,−1)-KMS state, which means that it satisfies and τz = ρiz,−iz (2.5) h(ab) = h(bσ−i(a)), a, b ∈ A, SYMMETRIES 7 cf. [BR97, Definition 5.3.1] or [Ped79, Section 8.12]. For z, z′ ∈ C, h(ρz,z ′(a)) = h(a), so (2.6) h(σz(a)) = h(τz(a)) = h(a), a ∈ A. The matrix elements of the irreducible unitary corepresentations satisfy the famous generalized Peter-Weyl orthogonality relations (2.7) h(cid:16)(cid:0)u(s) kℓ(cid:17) = ij (cid:1)∗u(t) where f1 : A → C is the Woronowicz character and kℓ(cid:1)∗(cid:17) = ij (cid:0)u(t) ℓj(cid:1) δstδikf1(cid:0)u(s) Ds , , Ds ki(cid:1) δstδjℓf−1(cid:0)u(s) nsXℓ=1 Ds = h(cid:16)u(s) ℓℓ(cid:17) f1(cid:16)u(s) is the quantum dimension of u(s), cf. [Wor87a, Theorem 5.7.4]. Note that unitarity implies that the matrix (cid:16)f1(cid:0)(u(s) is invertible, with inverse(cid:0)f1(u(s) jk )∗(cid:1)(cid:17)1≤j,k≤ns ∈ Mns(C) jk )(cid:1)jk ∈ Mns(C), cf. [Wor87a, Equation (5.24)]. Remark 2.3. The Haar state on a compact quantum group is a trace if and only if the antipode is involutive, i.e. we have S2(a) = a for all a ∈ A. In this case we say that (A, ∆) is of Kac type. This is also equivalent to the following equivalent conditions, cf. [Wor98, Theorem 1.5], (1) fz = ε for all z ∈ C, (2) σt = id for all t ∈ R. The antipode S and the automorphism τ i 2 closure S admits the polar decomposition are closable operators on A, and the (2.8) S = R ◦ T, and R : A → A is a linear antimultiplicative norm where T is the closure of τ i preserving involution that commutes with hermitian conjugation and with the semigroup (τt)t∈R, i.e. τt ◦ R = R◦ τt for all t ∈ R, see [Wor98, Theorem 1.6]. The operator R is called the unitary antipode and is related to Woronowicz characters through the formula 2 (2.9) R(a) = S(f 1 2 ⋆ a ⋆ f − 1 2 ) for a ∈ A. 2.5. L´evy processes on involutive bialgebras. We recall the definition of L´evy processes on ∗-bialgebras, cf. [Sch93]. An introduction to this topic can also be found in [Fra06]. Lindsay and Skalski have developped an analytic theory of L´evy processes on C∗-bialgebras, see [LS12] and the references therein. 8 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA Definition 2.4. A family of unital ∗-homomorphisms (jst)0≤s≤t defined on a ∗- bialgebra A with values in a unital ∗-algebra B with some fixed state Φ : B → C is called a L´evy process on A (w.r.t. Φ), if the following conditions are satisfied: i.e. [jst(A), js′t′(A)] = {0} for 0 ≤ s ≤ t ≤ s′ ≤ t′, and expectations cor- responding to disjoint time intervals factorize, i.e. (i) the images corresponding to disjoint time intervals commute, Φ(js1t1(a1) · · · jsntn(an)) = Φ(js1t1(a1)) · · · Φ(jsntn(an)), for all n ∈ N, a1, . . . , an ∈ A and 0 ≤ s1 ≤ t1 ≤ . . . ≤ tn; multiplication of B and ∆ is the comultiplication on A; (ii) mB ◦ (jst ⊗ jtu) ◦ ∆ = jsu for all 0 ≤ s ≤ t ≤ u, where mB denotes the (iii) the functionals ϕst = Φ ◦ jst : A → C depend only on t − s; (iv) lim tցs jst(a) = jss(a) = ε(a)1B for all a ∈ A, where 1B denotes the unit of B. We do not distinguish two L´evy processes on the same ∗-bialgebra A which are equivalent. By this we mean that two processes (jst)0≤s≤t and (kst)0≤s≤t with values in unital ∗-algebras (B, Φ) and (B′, Φ′), respectively, agree on all their finite joint moments, i.e. Φ(cid:0)js1t1(b1) · · · jsntn(bn)(cid:1) = Φ′(cid:0)ks1t1(b1) · · · ksntn(bn)(cid:1), for all n ∈ N, s1 ≤ t1, . . . , sn ≤ tn and b1, . . . , bn ∈ A. form a convolution semigroup of states, i.e. If (jst)0≤s≤t is a L´evy process, then the functionals ϕt := ϕ0,t = ϕs,t+s (t ≥ 0) • ϕ0 = ε, ϕs ⋆ ϕt = ϕs+t, limt→0 ϕt(a) = ε(a) for all a ∈ A, • ϕt(1) = 1, ϕt(a∗a) ≥ 0 for all a ∈ A and t ≥ 0. For such a semigroup there exists a linear functional φ which is hermitian (i.e. φ(a∗) = φ(a) for a ∈ A), conditionally positive (φ(a∗a) ≥ 0 when a ∈ ker ε), vanishes on 1, and is such that (2.10) ϕt = exp⋆ tφ = ε + tφ + t2 2 φ ⋆ φ + · · · + tn n! φ⋆n + · · · . Conversely, by the Schoenberg correspondence (cf. [Fra06]), for every hermitian conditionally positive linear functional φ : A → C with φ(1) = 0 there exists a unique convolution semigroup of states (ϕt)t≥0 which satisfies (2.10) and a unique (up to equivalence) L´evy process (jst)0≤s≤t. The functional φ will be called the generating functional of the L´evy process (jst)0≤s≤t. Given the convolution semigroup of states (ϕt)t≥0, we can also define the semigroup of operators on A (called the Markov semigroup on A associated to (jst)0≤s≤t) The infinitesimal generator of this semigroup is an operator L : A → A, which is related to φ by the relations Tt = (id ⊗ ϕt) ◦ ∆, t ≥ 0. L(a) = (id ⊗ φ) ◦ ∆(a) = φ ⋆ a and φ(a) = ε ◦ L(a). SYMMETRIES 9 In this case we write L = Lφ. As usual the formula to recover the semigroup from the generator is Tt = exp(tL) for t ≥ 0. The fundamental theorem of coalgebra ensures that all this makes sense in the bialgebra A. Let L(A) denotes the algebra of linear operators from A to A. Operators L ∈ L(A) of the form L = Lφ = (id ⊗ φ) ◦ ∆ for some linear functional φ ∈ A′ will play an important role in the paper and we will refer to them as convolution operators. An operator L ∈ L(A) is a convolution operator if and only if it is translation invariant on A, i.e. ∆ ◦ L = (id ⊗ L) ◦ ∆, and, if this is the case, the linear functional φ can be recovered from L using the formula The map A′ ∋ φ → Lφ ∈ L(A) is also called the dual right representation. It is a unital algebra homomorphism for the convolution product, i.e. we have φ = ε ◦ L. Lε = id, Lφ ◦ Lψ = Lφ⋆ψ, for φ, ψ ∈ A′. Moreover, Lφ is hermitian, i.e. Lφ(a∗) = (Lφa)∗ iff φ is hermitian, i.e. φ(a∗) = φ(a), a ∈ A. for a ∈ A 3. Translation invariant Markov semigroups Our goal is to construct Markov semigroups on compact quantum groups that reflect the structure of the quantum group. In this section we show that it is exactly the translation invariant Markovian semigroups that can be obtained from L´evy processes on the algebra of smooth functions A = Pol(G) of the quantum group G = (A, ∆). For this purpose we first prove that the Markov semigroup (Tt)t≥0 of a L´evy process on A has a unique extension to a strongly continuous Markov semigroup on both its reduced and its universal C∗-algebra. We then show that the charac- terisation of L´evy processes in topological groups as the Markov processes which are invariant under time and space translations extends to compact quantum groups. Definition 3.1. A strongly continuous semigroup of operators (Tt)t≥0 on a C∗- algebra A is called a quantum Markov semigroup on A if every Tt is a unital, completely positive contraction. If (jst)0≤s≤t is a L´evy process on a ∗-bialgebra A with the convolution semigroup of states (ϕt)t≥0 on A and the Markov semigroup (Tt)t≥0 on A, then, by a result of B´edos, Murphy and Tuset [BMT01, Theorem 3.3], each ϕt extends to a continuous 10 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA functional on Au, the universal C∗-algebra generated by A. Then the formula Tt = (id ⊗ ϕt) ◦ ∆ makes sense on Au (where ∆ : Au → Au ⊗ Au denotes the unique unital ∗-homomorphism that extends ∆ : A → A ⊗ A) and one easily shows (in the same way as in Proposition below) that (Tt)t becomes a Markov semigroup on Au (in the sense of Definition 3.1). For us, however, it will be more natural to consider the reduced C∗-algebra generated by A. This is the C∗-algebra Ar obtained by taking the norm closure of the GNS representation of A with respect to the Haar state h. The Haar state h is by construction faithful on Ar. The coproduct on A extends to a unique unital ∗-homomorphism ∆ : Ar → Ar ⊗ Ar which makes the pair (Ar, ∆) a compact quantum group. The following result shows that, even though ϕt : A → C can be unbounded with respect to the reduced C∗-norm and therefore may not extend to Ar, (Tt)t≥0 always extends to a Markov semigroup on Ar. Michael Brannan showed that states on any C∗-algebraic version C(G) of G define a continuous convolution operator on the reduced version Cr(G), cf. [Bra11, Lemma 3.4]. We will need a similar result for convolution semigroups of states on Pol(G). Theorem 3.2. Each L´evy process (jst)0≤s≤t on the Hopf ∗-algebra A gives rise to a unique strongly continuous Markov semigroup (Tt)t≥0 on Ar, the reduced C∗-algebra generated by A. Proof. Let (λ,H, ξ) be the GNS representation of A for the Haar state h, thus h(a) = hξ, λ(a)ξi for a ∈ A. We denote by k.kr the norm in Ar, that is kakr = kλ(a)k, where k.k denotes the operator norm. Similarly, let (ρt,Ht, ξt) be the GNS representation of A for the state ϕt = Φ ◦ j0t, so that ϕt(a) = hξt, ρt(a)ξti for a ∈ A. We define the operators it πt Et : H ∋ v → v ⊗ ξt ∈ H ⊗ Ht : H ⊗ Ht ∋ v ⊗ w → hξt, wiHt v ∈ H : B(H ⊗ Ht) ∋ X → πt ◦ X ◦ it ∈ B(H). Since for each t, it is an isometry and πt is contractive, Et is contractive too: kEt(X)k = kπt ◦ X ◦ itk ≤ kXk. Next we define U : λ(A)ξ ⊗ ρt(A)ξt ∋ λ(a)ξ ⊗ ρt(b)ξt 7→ λ(a(1))ξ ⊗ ρt(a(2)b)ξt ∈ H ⊗ Ht and we check that it is an isometry with adjoint given by U∗(λ(a)ξ ⊗ ρt(b)ξt) = λ(a(1))ξ ⊗ ρt(S(a(2))b)ξt. Indeed, using the invariance of the Haar measure, we show that U is isometric SYMMETRIES 11 (cid:10)U(λ(a)ξ ⊗ ρt(b)ξt), U(λ(c)ξ ⊗ ρt(d)ξt)(cid:11) = (cid:10)λ(a(1))ξ ⊗ ρt(a(2)b)ξt, λ(c(1))ξ ⊗ ρt(c(2)d)ξt(cid:11) = h(a∗(1)c(1))ϕt(b∗a∗(2)c(2)d) = (h ⊗ ϕb,d = h(a∗c)ϕb,d t )(a∗(1)c(1) ⊗ a∗(2)c(2)) = (h ⋆ ϕb,d t (1) = h(a∗c)ϕt(b∗d) =(cid:10)λ(a)ξ ⊗ ρt(b)ξt, λ(c)ξ ⊗ ρt(d)ξt(cid:11), t )(a∗c) t (x) := ϕt(b∗xd). Moreover, by the antipode property (2.2) we have where ϕb,d U U∗(λ(a)ξ ⊗ ρt(b)ξt) = U(λ(a(1))ξ ⊗ ρt(S(a(2))b)ξt) = λ(a(1))ξ ⊗ ρt(a(2)S(a(3))b)ξt = λ(a(1)ε(a(2)))ξ ⊗ ρt(b)ξt = λ(a)ξ ⊗ ρt(b)ξt, which implies that U is an isometry with dense image and therefore extends to a unique unitary operator denoted again by U. Now the fact that the Markov semigroup (Tt)t is bounded on Ar, i.e. kTt(a)kr = kλ(Tt(a))kB(H) ≤ kλ(a)kB(H) = kakr, follows immediately from the relation (3.1) since λ(Tt(a)) = Et(cid:0)U(λ(a) ⊗ idHt)U∗(cid:1), kλ(Tt(a))k = kEt(U(λ(a) ⊗ idHt)U∗)k ≤ kU(λ(a) ⊗ idHt)U∗k = kλ(a) ⊗ idHtk = kλ(a)k. To see that (3.1) holds, let us fix v ∈ H and b ∈ A such that v = λ(b)ξ. Then Et(U(λ(a) ⊗ idHt)U∗)v = (πt ◦ U ◦ (λ(a) ⊗ idHt) ◦ U∗ ◦ it)(λ(b)ξ) = (πt ◦ U ◦ (λ(a) ⊗ idHt) ◦ U∗) (λ(b)ξ ⊗ ξt) = πt ◦ U ◦ (λ(a) ⊗ idHt)(cid:0)λ(b(1))ξ ⊗ ρt(S(b(2)))ξt(cid:1) = πt ◦ U (cid:0)λ(ab(1))ξ ⊗ ρt(S(b(2)))ξt(cid:1) = πt (cid:0)λ(a(1)b(1))ξ ⊗ ρt(a(2)b(2)S(b(3)))ξt(cid:1) = πt (cid:0)λ(a(1)b)ξ ⊗ ρt(a(2))ξt(cid:1) = hξt, ρt(a(2))ξti λ(a(1)b)ξ = λ(a(1)ϕt(a(2)))λ(b)ξ = λ(Tt(a))v. This way we showed that each Tt extends to a contraction on Ar. The exten- sions again form a semigroup and since both ∆ and ϕt are completely positive and unital, Tt is too. Let us now check that (Tt)t forms a strongly continuous semigroup on Ar. For a given a ∈ Ar we choose, by density, an element b ∈ A such ka − bkr < ǫ. By definition for b ∈ A, Tt(b) = ϕt ⋆ b = (id ⊗ ϕt) ◦ ∆(b), where (ϕt)t is the 12 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA convolution semigroup of states on A (cf. Section 2.5). Thus kTt(a) − akr ≤ kTt(a) − Tt(b)kr + kTt(b) − bkr + kb − akr ≤ 2ka − bkr + k(ϕt ⋆ b) − bkr ≤ 2ǫ +Xkb(1)ϕt(b(2)) − b(1)ε(b(2))kr = 2ǫ +X ϕt(b(2)) − ε(b(2))kb(1)kr. Since limt→0+ ϕt(b) = ε(b) for any b ∈ A and the sum is finite, we conclude that lim t→0+kTt(a) − akr = 0 for each a ∈ Ar. (cid:3) The next results give a characterisation of Markov semigroups which are related to L´evy processes on compact quantum groups. Lemma 3.3. Let (A, ∆) be a compact quantum group and let T : A → A be a completely bounded linear map. If T is translation invariant, i.e. satisfies ∆ ◦ T = (id ⊗ T ) ◦ ∆ then T (Vs) ⊆ Vs for all s ∈ I and therefore T also leaves the ∗-Hopf algebra A invariant. Proof. Let s, s′ ∈ I, s 6= s′, and 1 ≤ j, k ≤ ns, 1 ≤ p, q ≤ ns′. Since the Haar state is idempotent, we have pq (cid:17)∗ h(cid:16)(cid:16)u(s′) T(cid:16)u(s) jk(cid:17)(cid:17) pq (cid:17)∗ jk(cid:17)(cid:17) = (h ⋆ h)(cid:16)(cid:16)u(s′) T(cid:16)u(s) (h ⊗ h)(cid:16)(cid:16)(cid:16)u(s′) pr (cid:17)∗ jk(cid:17)(cid:17)(cid:17) rq (cid:17)∗(cid:17) ∆(cid:16)T(cid:16)u(s) ⊗(cid:16)u(s′) (h ⊗ h)(cid:16)(cid:16)u(s′) pr (cid:17)∗ ℓk(cid:17)(cid:17)(cid:17) jℓ ⊗ T(cid:16)u(s) rq (cid:17)∗(cid:16)u(s) ⊗(cid:16)u(s′) nsXℓ=1 ℓk )(cid:17) , h(cid:16)(u(s′) jk(cid:1)(cid:17) = 0 for all s, s′ ∈ I, with s 6= s′, and all 1 ≤ j, k ≤ ns, f1((u(s) Ds ℓq )∗T (u(s) jp )∗) (cid:3) = = ns′Xr=1 ns′Xr=1 nsXℓ=1 i.e. h(cid:16)(cid:0)u(s′) jk(cid:1) ∈ Vs. 1 ≤ p, q ≤ ns′. Therefore T(cid:0)u(s) pq (cid:1)∗T(cid:0)u(s) δss′ = Theorem 3.4. Let (A, ∆) be a compact quantum group and (Tt)t≥0 a quantum Markov semigroup on A. Then (Tt)t≥0 is the quantum Markov semigroup of a (uniquely determined) L´evy process on A if and only if Tt is translation invariant for all t ≥ 0. SYMMETRIES 13 Proof. If (Tt)t comes from a L´evy process on A, then, on A, Tt = (id ⊗ ϕt) ◦ ∆ and so ∆ ◦ Tt = (id ⊗ id ⊗ ϕt) ◦ (∆ ⊗ id) ◦ ∆ = (id ⊗(cid:0)(id ⊗ ϕt) ◦ ∆(cid:1)) ◦ ∆ = (id ⊗ Tt) ◦ ∆. Hence Tt is translation invariant on A, and therefore also on A by continuity. Conversely, if every Tt is translation invariant, then Lemma 3.3 implies that, for all a ∈ Vs, Tta ∈ Vs and so, since Vs is finite dimensional, ε(Tta) → ε(a) as t → 0. It now follows easily that ϕt := ε◦ TtA defines a convolution semigroups of states whose generating functional defines a L´evy process whose Markov semigroup is (Tt)t. (cid:3) The corresponding result, for counital multiplier C∗-bialgebras satisfying a residual vanishing at infinity condition, was proved by Lindsay and Skalski [LS11, Proposition 3.2]). Their result covers coamenable compact quantum groups (where the counit extends continuously to the C∗-algebra). The above proof, for all compact quantum groups, is simpler. 4. GNS-Symmetry and KMS-symmetry of convolution operators In this section we study symmetry properties of convolution operators Lφ(a) = φ ⋆ a on A and we show that they can be translated into invariance properties of the corresponding generating functional φ. We will use two antilinear involutions # and ⋆ on A′, defined by φ#(a) = φ(a∗), φ⋆(a) = φ#(cid:0)S(a)(cid:1), for a ∈ A. A functional φ ∈ A′ is hermitian if and only if φ# = φ. Furthermore, we have ε# = ε⋆ = ε and h# = h⋆ = h. Note that # is multiplicative whereas ⋆ is anti-multiplicative with respect to the convolution of functionals: (φ ⋆ ψ)# = φ# ⋆ ψ#, (φ ⋆ ψ)⋆ = ψ⋆ ⋆ φ⋆ for φ, ψ ∈ A′. Let us denote by L2(A, h) the GNS Hilbert space of (A, h), by ξh = 1A ∈ L2(A, h) the cyclic vector representing the Haar state: h(a) = hξh, aξhi and let us assume that we are given an embedding, i.e. an injective linear map i : A → L2(A, h) with a dense range. We say that a linear operator L : A → A admits an i-adjoint if there exists L† : A → A such that (cid:10)i(a), i(Lb)(cid:11) =(cid:10)i(L†a), i(b)(cid:11) for any a, b ∈ A. Since h is faithful on A and since i has a dense range, the adjoint is unique if it exists. Then, an operator L ∈ L(A) is called i-symmetric if L equals to its i-adjoint. 14 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA In this paper we shall consider two embeddings. The first one is the natural inclusion coming from the GNS construction ih : A ∋ a → aξh ∈ L2(A, h). Definition 4.1. A map L⋆ ∈ L(A) such that (4.1) for all a, b,∈ A will be called a GNS-adjoint, or simply adjoint of L w.r.t. h. A map L will be called GNS-symmetric if L = L⋆. h(cid:0)a∗L(b)(cid:1) = h(cid:0)L⋆(a)∗b(cid:1) Let us observe that a convolution operator always admits a GNS-adjoint. Proposition 4.2. Let φ ∈ A′. Then there exists a unique convolution operator L⋆ φ that is adjoint to Lφ w.r.t. the Haar state, i.e. that satisfies for all a, b ∈ A. The adjoint of Lφ is given by φ = Lφ⋆. L⋆ h(cid:0)a∗Lφ(b)(cid:1) = h(cid:0)L⋆ φ(a)∗b(cid:1) Therefore Lφ is GNS-symmetric if and only if φ⋆ = φ. Proof. This is simply the fact that the dual right representation is a ∗-represen- tation w.r.t. to the involution ⋆ and the inner product defined by the Haar state as A × A ∋ (a, b) 7→ ha, bi = h(a∗b) ∈ C. The proof is the same as in the finite-dimensional case, see [VD97, Proposition 2.3]. See also [FS08, Proposition 3.4]. (cid:3) The second embedding we can consider is the symmetric embedding is : A ∋ a 7→ is(a) = σ − i 4 (a)ξh ∈ L2(A, h) and the related notion of symmetry is the following. Definition 4.3. We shall call a map L♭ ∈ L(A) the KMS-adjoint of L ∈ L(A), if we have (4.2) for all a, b,∈ A. An operator L ∈ L(A) is called KMS-symmetric if L♭ = L. (a)∗L(b)(cid:1) = h(cid:0)L♭(a)∗σ (b)(cid:1) h(cid:0)σ − i − i 2 2 Let us note here that a definition of KMS-symmetric operator on a von Neu- mann algebra was introduced by Goldstein and Lindsay ([GL95]) in the frame- work of (noncommutative) Haagerup Lp-spaces and by Cipriani in his PhD thesis (cf. [Cip97]) in the context of the standard form of von Neumann algebras. Later, in [Cip08, Definition 2.31], a definition of a KMS-symmetric operator on a C∗- algebra was provided. In the sequel, we shall also need the definition of KMS-symmetric operators on the whole C∗-algebra. It is stated as follows. SYMMETRIES 15 Definition 4.4. A linear map L : A → A is called KMS-symmetric w.r.t. h with modular automorphism group (σt)t∈R, if (4.3) for all a, b in a dense σ-invariant ∗-subalgebra B of the C∗-algebra A. h(cid:0)aL(b)(cid:1) = h(cid:0)σ i (a))(cid:1) (b)L(σ − i 2 2 Note that a continuous map L : A → A is KMS-symmetric in the sense of Definition 4.4 if it is A-invariant (i.e. L(A) ⊂ A), hermitian and (σ,−1)-KMS- symmetric in the sense of [Cip08, Definition 2.31]. The temperature β = −1 is chosen according to the KMS-property of the Haar state h(ab) = h(bσ−i(a)), see Equation (2.5). The analogue of Proposition 4.2 for KMS-symmetric operators on A is now the following. Theorem 4.5. Let φ ∈ A′. Then there exists a unique convolution operator L♭ that is the KMS-adjoint to Lφ w.r.t. the Haar state, i.e. that satisfies (a)∗Lφ(b)(cid:1) = h(cid:0)L♭ φ(a)∗σ − i for all a, b ∈ A. The KMS-adjoint of Lφ is given by L♭ φ = Lφ#◦R, where R denotes the unitary antipode. Proof. Let us observe first that a linear map L ∈ L(A) admits a KMS-adjoint if and only if it admits a GNS-adjoint, and that the two adjoints are related by (b)(cid:1) h(cid:0)σ − i φ 2 2 (4.4) L♭ = σ i 2 ◦ L⋆ ◦ σ − i 2 . Indeed, if the GNS-adjoint exists then, by (4.1) and the σ-invariance of h, we have h(cid:0)σ − i 2 (a)∗L(b)(cid:1) = h(cid:0)L⋆(σ − i 2 (a))∗b(cid:1) = h(cid:0)(σ i 2 ◦ L⋆ ◦ σ − i 2 )(a)∗σ − i 2 (b)(cid:1). Comparing with Equation (4.2) and using the faithfulness of the Haar state, we deduce that L♭ exists and satisfies (4.4). Conversely, if the KMS-adjoint exists then, using similar arguments, we show that the GNS-adjoint exists and L⋆ = σ . 2 ◦ L♭ ◦ σ i − i 2 Now, it follows from Proposition 4.2 that L♭ φ exists and for all a ∈ A we have ⋆ a ⋆ f 1 since R(a) = S(f 1 2 ⋆ a ⋆ f (cid:3) L♭ φ(a) = f − 1 2 2 2 2 2 2 2 ⋆ a = f − 1 − 1 ⋆(cid:0)Lφ#◦S(f 1 )(cid:1) ⋆ f ⋆ (φ# ◦ S) ⋆ f 1 = a(1)(cid:0)f 2(cid:1)(a(2)) ⋆ (φ# ◦ S) ⋆ f 1 − 1 (a(2))(φ# ◦ S)(a(3))f 1 − 1 = a(1)(cid:0)(φ# ◦ S)(f 1 )(cid:1) − 1 = a(1)(φ# ◦ R)(a(2)) = Lφ#◦R(a), − 1 ), see Equation (2.9). = a(1)f ⋆ a(2) ⋆ f 2 2 2 2 2 (a(4)) 16 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA Corollary 4.6. Suppose that φ ∈ A′. Then (1) Lφ is GNS-symmetric if and only if φ satisfies φ# ◦ S = φ. (2) Lφ is KMS-symmetric if and only if φ satisfies φ# ◦ R = φ. The generating functionals φ of L´evy processes are necessarily hermitian (φ# = φ). We call a hermitian functional φ on A φ GNS-symmetric if it is invariant under the antipode: φ ◦ S = φ, and KMS-symmetric if it is invariant under the unitary antipode: φ ◦ R = φ. Remark 4.7. A hermitian φ is GNS-symmetric if and only if each matrix φ(s) = [φ(u(s) jk )]j,k is hermitian: φ(u(s) jk ) = (φ ◦ S)(u(s) jk ) = φ((u(s) kj )∗) = φ(u(s) kj ). We shall show now that invariance under the phase in the polar decomposition of the antipode has also an influence on the properties of Lφ. Proposition 4.8. Let φ ∈ A′. Then the following conditions are equivalent: (1) Lφ commutes with the modular automorphism group σ, (2) φ commutes with the Woronowicz characters: φ ⋆ fz = fz ⋆ φ for z ∈ C, (3) φ ◦ τ i = φ. 2 Proof. By Equation (2.4), we have Lφ ◦ σt = σt ◦ Lφ if and only if φ ⋆ fit ⋆ a ⋆ fit = fit ⋆ φ ⋆ a ⋆ fit for all a ∈ A. Convolving by f−it from the right and applying the counit, we see that Lφ commutes with the modular automorphism group, if and only if φ ⋆ fit = fit ⋆ φ for all t ∈ R, which is equivalent to (4.5) for all z ∈ C by uniqueness of analytic continuation. We have shown this way that (1) ⇔ (2). φ ⋆ fz = fz ⋆ φ From Equation (4.5) we deduce immediately that φ ◦ τz(a) = φ(fiz ⋆ a ⋆ f−iz) = (f−iz ⋆ φ ⋆ fiz)(a) = φ(a), so (2) implies (3). Finally, let us see that (3) implies (2). For that we adopt the matrix notation from [Wor87a]: F (s) = [f−1(u(s) jk )]s j,k=−s and φ(s) = [φ(u(s) jk )]s j,k=−s. From therein we know that F (s) is invertible and positive and that fz(u(s)) = ⋆ φ = φ, then by the definition of τz we have φ ⋆ f = f (F (s))−z. If φ ◦ τ i and also φ ⋆ f−1 = f−1 ⋆ φ. This means that 2 φ(s)F (s) = F (s)φ(s) − 1 2 − 1 2 SYMMETRIES 17 and by the functional calculus φ(s) must commute with all (F (s))z for z ∈ C. This translates into φ ⋆ fz = fz ⋆ φ for all z ∈ C. It is known that on von Neumann algebras GNS-symmetry is a stronger condi- tion than the KMS-one (cf. [Cip08, Remarks after Definition 2.31]). The previous observation allows to provide a simple proof of this fact in our setting. (cid:3) Corollary 4.9. If φ is GNS-symmetric, then φ commutes with all Woronowicz characters and is KMS-symmetric. Proof. For GNS-symmetric φ we have φ = φ ◦ S2 = φ ◦ τi, which translates into φ⋆f−1 = f−1 ⋆φ. From the proof of Proposition 4.8 we see that this implies that φ is invariant under all τz (z ∈ C) or, equivalently, commutes with all Woronowicz characters. In particular φ = φ ◦ τ i φ = φ ◦ τ i = (φ ◦ S) ◦ τ i = φ ◦ R. 2 and 2 2 (cid:3) Remark 4.10. If the algebra A is of Kac type (S2 = id), then R = S and the notions of GNS-symmetry and KMS-symmetry coincide. However, Example 11.5 shows that in general KMS-symmetry is a weaker condition than GNS-symmetry. We end this Section with an observation linking the symmetries of the gener- ators and the related Markov semigroups. Theorem 4.11. Let (Tt)t≥0 be the Markov semigroup of a L´evy process on A with generating functional φ. (a) The following three conditions are equivalent: (b) The following four conditions are equivalent: (a1) φ is KMS-symmetric. (a2) Lφ is KMS-symmetric. (a3) for each t ≥ 0, Tt is KMS-symmetric on A (see Definition 4.4). (b1) φ is GNS-symmetric. (b2) Lφ is GNS-symmetric. (b2′) Lφ satisfies the quantum detailed balance condition, i.e. we have (b3) (Tt)t≥0 satisfies the quantum detailed balance condition, i.e. (4.6) holds for all Tt, t ≥ 0. Proof. The equivalences (x1) ⇔ (x2) follow from Corollary 4.6. The KMS-symmetry as well as the GNS-symmetry of φ is preserved under the convolution powers (for example, if φ(Sa) = φ(a), then (φ ⋆ φ)(Sa) = (φ ⊗ φ)(S(a(2)) ⊗ S(a(1))) = φ(a(1))φ(a(2)) = (φ ⋆ φ)(a).) Since Ln φ(a) = φ⋆n ⋆ a, we see that both kinds of symmetry are also preserved for the powers of Lφ. This implies that for (Tt)t≥0, being of the form Tt = exp tLφ, the KMS-symmetry or condition (4.6) of (Tt)t≥0 is equivalent to KMS-symmetry or (4.6) of Lφ. (4.6) h(cid:0)aLφ(b)(cid:1) = h(cid:0)Lφ(a)b(cid:1) for a, b ∈ A. 18 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA Finally we need to check that (b2′) ⇔ (b1). Assume that Lφ satisfies (4.6). Then, by Proposition 4.2, Lφ satisfies which implies φ ◦ S = φ. Conversely, if φ ◦ S = φ, then by the same calculation we see that Lφ(a) = L⋆(a∗)∗ = (φ⋆ ⋆ a∗)∗ = a(1)φ(cid:0)S(a∗(2))∗(cid:1) = Lφ◦S−1(a), φ(a∗)∗b(cid:1) = h(cid:0)aLφ(b)(cid:1). h(cid:0)Lφ(a)b(cid:1) = h(cid:0)Lφ◦S−1(a)b(cid:1) = h(cid:0)L⋆ (cid:3) 5. Schurmann triples corresponding to KMS-symmetric generators In this Section we give a method to produce KMS-symmetric generating func- tionals. To this aim, we recall the notion of a Schurmann triple and describe its behavior under the composition of an arbitrary generator with the unitary antipode. Our steps are motivated by the following easy observation. Proposition 5.1. Let φ be a generating functional of a L´evy process. Then φ + φ ◦ R is a KMS-symmetric generating functional of a L´evy process. Proof. Since R(a∗) = R(a)∗ and ε(R(a)) = ε(a), we easily check that the Schoen- berg criteria for a generating functional are satisfied for φ + φ ◦ R. Moreover, R2 = id implies that φ + φ ◦ R is invariant under the unitary antipode. (cid:3) Note that the same procedure cannot be applied to the GNS-symmetric case, since S does not preserve the positivity and is not involutive. For a pre-Hilbert space D we denote by L#(D) the set of all operators from D to D which admit an adjoint. Definition 5.2. A Schurmann triple on a ∗-bialgebra A with counit ε is a triple ((π, D), η, φ) consisting of: (1) a unital ∗-representation π : A → L#(D) of A on some pre-Hilbert space (2) a linear map η : A → D, called cocyle, such that D, η(ab) = π(a)η(b) + η(a)ε(b) for all a, b ∈ A, (3) a hermitian linear functional φ : A → C satisfying φ(ab) = hη(a∗), η(b)i for a, b ∈ ker ε. Schurmann proved (cf. [Sch93]) that for any generating functional φ of a L´evy process there exists a Schurmann triple ((π, D), η, φ) (such that the generating functional is the last ingredient of the triple). Moreover, the Schurmann triple is uniquely determined (modulo unitary equivalence) provided that η is surjective. SYMMETRIES 19 Definition 5.3. Given a pre-Hilbert space D, the opposite space Dop is defined as Dop = {v : v ∈ D} (the set of the same elements as D) with the same addition v + w = v + w, but with the scalar multiplication given by λ · v = λv and with the scalar product h¯v, ¯wiop = hw, vi. Given a unital ∗-representation π : A → L#(D) we define πop : A → L#(Dop) by the formula πop(a)¯v = (π ◦ R)(a∗)v, ¯v ∈ Dop. We check directly that πop is unital, multiplicative, and ∗-preserving, so it is a ∗-representation of A on Dop. We shall call it the opposite representation. Theorem 5.4. If φ is a generating functional of a L´evy process with the Schurmann triple ((π, D), η, φ) on A, then φ ◦ R is a generating functional of a L´evy process with the Schurmann triple ((πop, Dop), ηop, φ ◦ R) on A where πop is the opposite representation with the representation space Dop and ηop : A → Dop is defined by ηop(a) = η(R(a∗)). Proof. Let φ be a generating functional of a L´evy process. Then it follows from the properties of R, mentioned after formula (2.8), that φ ◦ R is hermitian, con- ditionally positive and vanishes at 1. By the Schoenberg correspondence, φ ◦ R is a generating functional of a L´evy process. Now we want to check that ((πop, Dop), ηop, φ ◦ R) is a Schurmann triple. For that, note that ηop is linear and by the cocycle property of η we have ηop(ab) = η(R((ab)∗)) = η(R(a∗)R(b∗)) = π(R(a∗))η(R(b∗)) + η(R(a∗))ε(R(b∗)) = πop(a)η(R(b∗)) + η(R(a∗))ε(b) = πop(a)ηop(b) + ηop(a)ε(b). Moreover, φ ◦ R is linear and hermitian, and for a, b ∈ ker ε we have hηop(a∗), ηop(b)iop = hη(R(a)), η(R(b∗))iop = hη(R(b∗)), η(R(a))i = φ(cid:0)R(b)R(a)(cid:1) = (φ ◦ R)(ab). (cid:3) Corollary 5.5. If φ is invariant under R and ((π, D), η, φ) is the related surjec- tive Schurmann triple, then π is equivalent to its opposite representation πop. Corollary 5.6. If φ is a generating functional of a L´evy process with surjective is a Schurmann triple of a KMS symmetric generator φ + φ ◦ R. Schurmann triple ((π, D), η, φ) on A, then(cid:0)(π⊕ πop, D⊕ Dop), η⊕ ηop, φ + φ◦ R(cid:1) Note that the Schurmann triple (cid:0)(π ⊕ πop, D ⊕ Dop), η ⊕ ηop, φ + φ ◦ R(cid:1) in the Corollary 5.6 is not necessarily surjective, even if the triple ((π, D), η, φ) is surjective. This is for example the case if φ is already KMS symmetric -- then the range of η ⊕ ηop is the diagonal of D ⊕ Dop. 20 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA Remark 5.7. Let φ be a generating functional of a L´evy process on A with the associated Schurmann triple ((π, D), η, φ), i.e. ((π, D), η, φ) is the unique Schurmann triple for φ with a surjective cocycle. If A is an algebraic quantum group "of compact type", i.e. is the ∗-subalgebra of polynomials of a compact quantum group G = (A, ∆), then A is linearly spanned by the coefficients of unitary corepresentations and thus for every a ∈ A, π(a) is a bounded operator in D. In this case the space D can be completed to a Hilbert space H, η : A → D turns into a cocycle η : A → H with dense image, and π maps A to B(H). 6. Generating functionals invariant under adjoint action On classical Lie groups, central measures play an important role in harmonic analysis and the study of L´evy processes. A measure µ on a topological group G is called central, if it commutes with all other measures (w.r.t. to the convolution). This is the case if ZG f (gxg−1)dµ(x) =ZG f (x)dµ(x) for all g ∈ G and f ∈ C(G), or, equivalently, if δg ⋆ µ ⋆ δg−1 = µ for all g ∈ G. On compact quantum groups we don't have Dirac measures, but we can translate this condition to ψ(1) ⋆ µ ⋆ S(ψ(2)) = ψ(1)µ for all functionals ψ : A → C, for which (id ⊗ S) ◦ ∆(ψ) = ψ(1) ⊗ S(ψ(2)) can be defined, i.e. for functionals which belong to the algebra of smooth functions A on the dual discrete quantum group. This condition is equivalent to invariance of the functional µ under the adjoint action, see below. In this Section we will study ad-invariance for functionals on compact quantum groups. On cocommutative compact quantum groups (i.e. such that τ ◦ ∆ = ∆, where τ is the flip operator τ (x⊗ y) = y ⊗ x) all functionals are ad-invariant, but on non-cocommutative compact quantum groups, ad-invariance characterizes an interesting class of functionals that share many similar properties with central measures. After reviewing several characterizations and showing that the ad- invariant functionals are exactly those that belong to the center of A′, we show that it is possible to construct from a given functional an ad-invariant one. But this construction does not preserve positivity. Recall that the adjoint action of a Hopf algebra is defined by ad : A → A⊗A, for a ∈ A, see, e.g., [Maj95], [KS97, Section 1.3.4]. The adjoint action is a left coaction, i.e. we have ad(a) = a(1)S(a(3)) ⊗ a(2) (id ⊗ ad) ◦ ad = (∆ ⊗ id) ◦ ad, (ε ⊗ id) ◦ ad = id. But note that ad is not an algebra homomorphism. Definition 6.1. We call a linear functional φ ∈ A′ ad-invariant, if it satisfies SYMMETRIES 21 Similarly, a linear map L ∈ L(A) is called ad-invariant, if it satisfies (id ⊗ φ) ◦ ad = φ1A. (id ⊗ L) ◦ ad = ad ◦ L. If the quantum group is cocommutative, then the adjoint action is the trivial coaction ad(a) = 1 ⊗ a. Therefore in this case all functionals are ad-invariant. invariant. It is straightforward to verify that the counit ε and the Haar state h are ad- The following characterisations show that the ad-invariant functionals are a natural generalisation of central measures. Proposition 6.2. Let φ ∈ A′. The following conditions are equivalent. (a): φ is ad-invariant. (b): We have ψ(1) ⋆ φ ⋆ S(ψ(2)) = ψ(1)φ for all ψ ∈ A. (c): φ commutes with all elements of A: φ ⋆ ψ = ψ ⋆ φ for all ψ ∈ A. (d): φ belongs to the center of A′: φ ⋆ ψ = ψ ⋆ φ for all ψ ∈ A′. Proof. (a)⇔(b): If φ : A → C is ad-invariant, then we have a(1)S(a(3))φ(a(2)) = φ(a)1. Applying the functional ψ = hb ∈ A with b ∈ A to this, we get ψ(1)φ(a) = ψ(cid:0)a(1)S(a(3))(cid:1)φ(a(2)) = ψ(1)(a(1))ψ(2)(cid:0)S(a(3))(cid:1)φ(a(2)) = ψ(1)(a(1))φ(a(2)) S(ψ(2))(a(3)) =(cid:0)ψ(1) ⋆ φ ⋆ S(ψ(2))(cid:1)(a) for all a ∈ A and all ψ ∈ A. The converse follows, because by the faithfulness of the Haar state on A we have ∀b ∈ A, ψ(a) = h(ba) = 0 ⇒ a = 0. (c)⇒(b): This follows directly from the antipode axiom, ψ(1) ⋆ φ ⋆ S(ψ(2)) = ψ(1) ⋆ S(ψ(2)) ⋆ φ = ε(ψ)1 ⋆ φ = ψ(1)φ, (a)⇒(c): Suppose that φ is ad-invariant and apply ψ ◦ m ◦ (id ⊗ φ ⊗ id) ◦ where 1 = ε is the unit of A′. (ad ⊗ id) to ∆(a), then this gives which is equal to ψ(a(1))φ(a(2)) = (ψ ⋆ φ)(a) by the antipode axiom. On the other hand, using the ad-invariance of φ, the same expression becomes ψ(cid:0)a(1)S(a(3))a(4)(cid:1)φ(a(2)) ψ(1a(2))φ(a(1)) = (φ ⋆ ψ)(a). 22 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA (c)⇔(d): This follows, because A′ embeds into the multiplier algebra M( A) of A, since ψ ⋆ ha = hc, ha ⋆ ψ = hd with c = ψ(S(a(1)))a(2), d = ψ(S−1(a(2)))a(1). (cid:3) Corollary 6.3. The ad-invariant functionals form a unital subalgebra of A′ with respect to the convolution. The following formula shows that the coproduct ∆ : A → A⊗A is ad-invariant, if we define the adjoint action of A ⊗ A by ad⊗ = (m ⊗ id ⊗ id) ◦ (id ⊗ τ ⊗ id) ◦ (ad ⊗ ad). Lemma 6.4. The adjoint action satisfies the relation (m ⊗ id ⊗ id) ◦ (id ⊗ τ ⊗ id) ◦ (ad ⊗ ad) ◦ ∆ = (id ⊗ ∆) ◦ ad. Proof. Using Sweedler notation, we get (m ⊗ id ⊗ id) ◦ (id ⊗ τ ⊗ id) ◦ (ad ⊗ ad) ◦ ∆(a) = = a(1)S(a(3))a(4)S(a(6)) ⊗ a(2) ⊗ a(5) = a(1)ε(a(3))1AS(a(5)) ⊗ a(2) ⊗ a(4) for a ∈ A, where we used the antipode property (2.2). After further simplifica- tion, using the counit property (2.1), we get = a(1)S(a(4)) ⊗ a(2) ⊗ a(3) = (id ⊗ ∆) ◦ ad(a). (cid:3) Lemma 6.5. Let φ ∈ A′. Then φ is ad-invariant if and only if Lφ is ad-invariant. Proof. Let us observe that (id ⊗ Lφ) ◦ ad(a) = (id ⊗ id ⊗ φ)(a(1)S(a(4)) ⊗ a(2) ⊗ a(3)) = (id ⊗ id ⊗ φ)(a(1)S(a(3))a(4)S(a(6)) ⊗ a(2) ⊗ a(5)) = a(1)S(a(3)) a(4)S(a(6))φ(a(5)) ⊗ a(2). (cf. Lemma 6.4) If we assume that φ is ad-invariant, then (id ⊗ Lφ) ◦ ad(a) = a(1)S(a(3))φ(a(4)) ⊗ a(2) = φ(a(2))ad(a(1)) = ad ◦ Lφ(a). On the other hand, if we suppose that Lφ is ad-invariant, then the application of (id ⊗ ε) to both sides of the equation φ(a(4))a(1)S(a(3)) ⊗ a(2) = a(1)S(a(3)) a(4)S(a(6))φ(a(5)) ⊗ a(2) gives the ad-invariance of φ. (cid:3) We can use the Haar state to produce ad-invariant functionals. Apply the invariance of the Haar measure (Proposition 2.2) to the element under the Haar state, and after the appropriate renumbering, we get adh ◦ adh(a) = h(cid:0)h[a(1)S(a(5))]a(2)S(a(4))(cid:1)a(3). adh ◦ adh(a) = h(cid:0)h(a(1)S(a(3)))1(cid:1)a(2) = adh(a). SYMMETRIES 23 Proposition 6.6. Denote by adh ∈ L(A) the linear map given by adh = (h ⊗ id) ◦ ad. Then φad := φ ◦ adh is ad-invariant for all φ ∈ A′. Proof. Observe that by definition we have φad = φ ◦ adh = (h ⊗ φ) ◦ ad. Using the invariance of the Haar measure (Proposition 2.2) we check that φad(a)1 = h(cid:0)a(1)S(a(3))(cid:1)φ(a(2))1 = a(1)S(a(5))h(cid:0)a(2)S(a(4))(cid:1)φ(a(3)) = a(1)S(a(3))φad(a(2)) = (id ⊗ φad) ◦ ad(a). (cid:3) Proposition 6.7. Let us collect the basic properties of adh. (a) adh ◦ adh = adh. (b) (φ ◦ adh)⋆ = φ⋆ ◦ adh for all φ ∈ A′. (c) A linear functional φ ∈ A′ is ad-invariant if and only if φ = φ ◦ adh. Proof. Ad (a). Explicit calculations give Ad (b). Recall that φ⋆ = φ# ◦ S, where S is the antipode and φ#(a) = φ(a∗). Then the assertion will follow if we show that [adh ◦ S(a)∗]∗ = S ◦ adh(a). Using the properties that S ◦ ∗ ◦ S ◦ ∗ = id and ∆(S(a)) = τ ◦ (S ⊗ S) ◦ ∆(a) we check that ad(S(a)∗) = S(a(3))∗a∗(1) ⊗ S(a(2))∗. Then since h is hermitian, we have [adh ◦ S(a)∗]∗ = [(h ⊗ id) ◦ (ad ◦ S)(a)∗]∗ = h(cid:0)S(a(3))∗a∗(1)(cid:1)S(a(2)) = h(cid:0)a(1)S(a(3))(cid:1)S(a(2)) = S(adh(a)). Ad (c). First we check that for an ad-invariant functional φ we have φ = φ◦adh: φ ◦ adh(a) = φ ◦ (h ⊗ id) ◦ ad(a) = h ◦ (id ⊗ φ) ◦ ad(a) = h(φ(a)1) = φ(a). The converse follows immediately from Proposition 6.6. (cid:3) Applying Lemma 6.5 and Corollary 6.3, we get an analogue of Theorem 4.11 for ad-invariance. Corollary 6.8. Let (Tt)t≥0 be the Markov semigroup of a L´evy process on A with generating functional φ. The following three conditions are equivalent: (a1) φ is ad-invariant. (a2) Lφ is ad-invariant. nXp,r=1 nXp,r=1 24 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA (a3) for each t ≥ 0, Tt is ad-invariant. In the next proposition we show that ad-invariance of functionals can be char- acterized by the form of their characteristic matrices. jk )(cid:1)1≤j,k≤ns Proposition 6.9. A functional φ is ad-invariant if and only if its characteristic are multiples of the identity matrix for all s ∈ I, i.e. if jk ) = csδjk for all s ∈ I and matrices(cid:0)φ(u(s) there exist complex numbers cs, s ∈ I, such that φ(u(s) all 1 ≤ j, k ≤ ns. Proof. We use the orthogonality relation for the Haar measure (2.7) to show that for the ad-invariant functional φ we have φ(u(s) jk ) = φad(u(s) jk ) = h(u(s) jp (u(s) kr )∗)φ(u(s) pr ) = 1 Ds f1(u(s) rp )φ(u(s) pr ) · δjk, and we observe that the constant 1 pr ) does not depend on j or k. Reciprocally, if φ is of this form, then we check that φ = φad and, by (c) in Proposition 6.7, φ is ad-invariant. (cid:3) p,r=1 f1(u(s) rp )φ(u(s) DsPn In general, the mapping ad∗h : φ 7→ φad in Proposition 6.6 preserves neither her- miticity nor positivity, see Example 11.7. But [Voi11, Lemma 4.1] and [BMT03, Theorem 4.5] suggest that some properties of adh can be improved if we replace the antipode by the twisted antipode defined by eS(a) = f1 ⋆ S(a) for a ∈ A. A = Pol(G). Denote by eS the twisted antipode defined by eS(a) = f1 ⋆ S(a) = f−1(a(1))S(a(2)) and denote byfad the twisted adjoint actionfad(a) = a(1)eS(a(3))⊗ Theorem 6.10. Let G be a compact quantum group with dense ∗-Hopf algebra a(2), a ∈ A. satisfies (a): The mapfadh : A → A defined by fadh(a) = (h ⊗ id) ◦fad(a) = h(cid:0)a(1)eS(a(3))(cid:1)a(2) fadh(a∗a) = (h ⊗ id)(cid:16)(cid:0)fad(a)(cid:1)∗fad(a)(cid:17) for a ∈ A and therefore preserves positivity. (b): If G is of Kac-type, then we have fadh ◦fadh =fadh. fadh = adh. phism, we have Proof. (b) was already shown in Proposition 6.7, since in the Kac case we have Let us now prove (a). Since the twisted antipode is an algebra anti-homomor- fad(a∗a) = a∗(1),j a(1),keS(a(3),k)eS(a∗(3),j) ⊗ a∗(2),ja(2),k, SYMMETRIES 25 where we put back summation indices to distinguish the sums coming from the first and the second factor. Therefore Now for any b ∈ A, (h ⊗ id)(cid:16)fad(a∗a)(cid:17) = h(cid:16)a∗(1),ja(1),keS(a(3),k)eS(a∗(3),j)(cid:17) a∗(2),ja(2),k = h(cid:16)σi(cid:16)eS(a∗(3),j)(cid:17) a∗(1),ja(1),keS(a(3),k)(cid:17) a∗(2),j a(2),k. σi(eS(b∗)) = f−1 ⋆eS(b∗) ⋆ f−1 = S(b∗) ⋆ f−1 = S−1(b)∗ ⋆ f−1 = (S−1(b) ⋆ f1)∗ =(cid:0)f1 ⋆ S(b)(cid:1)∗ = eS(b)∗, (h ⊗ id)(cid:16)fad(a∗a)(cid:17) = h(cid:16)(cid:0)a∗(1),jeS(a(3),j)(cid:1)∗a(1),keS(a(3),k)(cid:17) a∗(2),j a(2),k and so we get Recall that the linear span of the characters of the irreducible unitary corep- resentations of a compact quantum group is an algebra (cid:3) called the algebra of central functions on G. corepresentations as Note thatfadh(A) ⊆ A0. Indeed,fadh acts on coefficients of irreducible unitary We see that we can usefad∗ h : φ 7→ φ ◦fadh to produce ad-invariant functionals. = (h ⊗ id)(cid:16)(cid:0)fad(a)(cid:1)∗fad(a)(cid:17). : s ∈ I) , A0 = span(χs = u(s) jj nsXj=1 jk ) = fadh(u(s) pq jp S(u(s) h(cid:16)u(s) qk )(cid:17)u(s) nsXp,q=1 jpeS(u(s) h(cid:16)u(s) ℓk )(cid:17)f−1(u(s) nsXp,q,ℓ=1 kℓ(cid:1)∗(cid:17)f−1(u(s) h(cid:16)u(s) nsXp,q,ℓ=1 jp(cid:0)u(s) nsXp,q,ℓ=1 nsXp=1 ℓp )f−1(u(s) δjkf1(u(s) 1 Ds δjk Ds u(s) pp qℓ )u(s) pq qℓ )u(s) pq qℓ )u(s) pq = = = = (6.1) (6.2) 26 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA h : (A0)′ → A′, fad∗ h(φ) = φ ◦fadh, maps h(φ) = φ ◦fadh is ad-invariant, since functionals on A0 to ad-invariant functionals on A. It maps states on A0 to states on A. h defines bijections between states on A0 and ad- invariant states on A, and between generating functionals on A0 and ad-invariant generating functionals on A. Proof. It follows immediately from Equation (6.2) and Proposition 6.9 that for Corollary 6.11. The linear map fad∗ If G is of Kac type, then fad∗ any φ ∈ A′ the functionalfad∗ By Theorem 6.10,fad∗ have alsofad(1) = 1 and it follows thatfad∗ In the Kac case we have furthermore ε ◦fadh = ε ◦ adh = ε, so in this case fadh maps the kernel of the counit onto itself and thereforefad∗ state or generating functional ψ = ψ ◦fadh on A. By Proposition 6.9 it is clear functionals on A onto ad-invariant generating functionals on A. Conversely, any state or generating functional ψ on A0 can be extended to a h maps states on A onto ad-invariant states on A. h maps positive functionals to positive functionals. Since we φ ◦fadh(1) = φ(1), that ψ is the unique ad-invariant extension of ψ. jk ) = φ ◦fadh(u(s) δjk Ds φ nsXℓ=1 ℓℓ! . u(s) h maps generating (cid:3) In Section 10, we will show that this Corollary allows to completely characterize the ad-invariant generating functionals on the free orthogonal quantum group O+ n . 7. Dirichlet forms In this Section we determine explicitly the structure of the Dirichlet forms as- sociated to KMS-symmetric generating functionals on compact quantum groups. In case of GNS symmetry, we also characterize the invariance under translation of generators on the algebra A, in terms of an associated quadratic form on A. Recall that L2(A, h) denotes the GNS Hilbert space of (A, h) and that the cyclic vector ξh = 1A ∈ L2(A, h) represents the Haar state as h(a) = hξh, aξhi. From now on and until the end of Section 8, we assume that the Haar state is faithful on the C∗-algebra A so that we can identify A with an involutive subalgebra of the von Neumann algebra L∞(A, h) of bounded operators on L2(A, h), generated by A by the GNS representation. As a consequence, the vector ξ is cyclic for the von Neumann algebra L∞(A, h) too. Notice that, as the Haar state h is a (σ,−1)-KMS state for the modular auto- morphism group σ (see Section 2.4), it follows, by the KMS theory (in particular [BR97, Corollary 5.3.9]), that the vector ξ is also separating for the von Neumann SYMMETRIES 27 algebra L∞(A, h). This fact allows to apply the Tomita-Takesaki modular theory to the Haar state h on the C∗-algebra A of the compact quantum group. Recall also that the symmetric embedding is defined by is : A → L2(A, h), is(a) = ∆ 1 4 aξh, where ∆ denotes (exceptionally) the Tomita-Takesaki modular operator. This definition agrees with the one from Section 4 (page 14), since for a ∈ A we have is(a) = ∆ 1 4 a∆− 1 4 ξh = σ (a)ξh. − i 4 For a given KMS-symmetric generating functional φ of a L´evy process on A, and the related convolution operator Lφ(a) = φ ⋆ a, we define the sesquilinear and the quadratic forms Eφ(cid:0)is(a), is(b)(cid:1) = (cid:10)is(a), is(−Lφ(b))(cid:11) = −h(cid:0)σ Eφ[is(a)] = Eφ(cid:0)is(a), is(a)(cid:1), D(Eφ) = {is(a) ∈ L2(A, h) : a ∈ D(Lφ) and Eφ[is(a)] < ∞} . (a)∗(σ 4 ◦ Lφ)(b)(cid:1), − i − i 4 on the domain The explicit values of the sesquilinear form Eφ on the basis of the coefficients of the unitary corepresentations are the following jk ), is(u(t) Eφ(cid:0)is(u(s) jk ), is(u(t) lr )(cid:11)φ(u(t) rm) 4 − i (u(t) (u(s) jk )∗σ lm)(cid:1) = (cid:10)is(u(s) jk ), is(−Lφ(u(t) = Xr h(cid:0)σ = Xr,p,p′ h(cid:0)(u(s) DsXp lm))(cid:11) =Xr (cid:10)is(u(s) lr )(cid:1)φ(u(t) pp′(cid:1)f 1 pj )(cid:1)f 1 lp )Xr f−1(u(t) jk )∗u(t) lp )f 1 (u(t) (u(t) rm) − i δst = f 1 2 2 2 4 2 (u(t) p′r)φ(u(t) rm) (u(t) kr )φ(u(t) rm) = δst Ds f − 1 2 (u(t) lj )(f 1 2 ⋆ φ)(u(t) km). Since σz leaves the subspaces Vs invariant, hence is(Vs) = σ − i 4 (Vs)ξh = Vsξh = ih(Vs) and the operator defined by Hφis(a) := is(−Lφa), a ∈ D(Hφ) := is(A) ⊂ L2(A, h) leaves invariant the subspaces Es = Vsξh = Span{u(s) jk ξh : j, k = 1,· · · , ns} ⊂ L2(A, h), s ∈ I . 28 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA Therefore, since L2(A, h) =Ls∈I Hφ =Ms∈I H s φ Es, the operator Hφ decomposes as a direct sum of its restrictions H s φ on each finite dimensional subspace Es. Theorem 7.1. Let φ be a KMS-symmetric generating functional of a L´evy process on A. Then the operator Hφ is essentially self-adjoint, the quadratic form Eφ is closable and its closure is a Dirichlet form. Proof. The operator Hφ is a direct sum of bounded operators and is symmetric as Lφ is KMS symmetric. It follows that Hφ is essentially self-adjoint and its closure is given by D(Hφ) = {ξ = ⊕s∈I ξs ∈ L2(A, h) :Xs∈I kHφξsk2 < +∞} . Hφ(⊕s∈I ξs) = ⊕s∈I Hφξs , ⊕s∈I ξs ∈ D(Hφ) . As, by definition, Eφ[ξ] = hξ, Hφξi for ξ ∈ D(Hφ), we have that Eφ is closable and its closure is given by D(Eφ) = {ξ = ⊕s∈I ξs ∈ L2(A, h) :Xs∈I Eφ[⊕s∈I ξs] =Xs∈I hξs, Hφξsi , hξs, Hφξsi < +∞} . ⊕s∈I ξs ∈ D(Eφ) . Now, the quantum Markov semigroup Tt on the C∗-algebra A, generated by Lφ, is KMS symmetric, i.e. is (σ,−1)-KMS symmetric in the sense of Definition 2.1 in[Cip98] (see also Definition 2.31 in [Cip08]). By Theorem 2.3 and Theorem 2.4 in [Cip98] (see also Theorem 2.39 and Theorem 2.44 in [Cip08]) the semigroup e−tH φ on L2(A, h) is Markovian so that the quadratic form Eφ is a Dirichlet form by Theorem 4.11 in [Cip97] (see also Theorem 2.52 in [Cip08]). Remark 7.2. Using the embedding ih : A → L2(A, h), we can identify the Dirichlet form on L2(A, h), associated to a KMS-symmetric generating functional φ, with the following quadratic form on the C∗-algebra A (cid:3) Qφ[a] = Eφ[ih(a)] = −h(cid:0)a∗(σ 4 ◦ Lφ ◦ σ i − i 4 )(b)(cid:1) defined on dom (Qφ) := {a ∈ A : ih(a) ∈ dom (Eφ)}. If furthermore, φ is GNS- symmetric, then Lφ commutes with the modular group (σz)z (Prop. 4.8 and Cor. 4.9) and one has Qφ[a] = −h(cid:0)a∗Lφ(a)(cid:1). The next theorem shows that the Dirichlet forms associated to GNS-symmetric L´evy processes admit an additional invariance. Theorem 7.3. Let L be a GNS symmetric operator on A ⊂ L2(A, h). Then the following conditions are equivalent: SYMMETRIES 29 (1) There exists a functional φ ∈ A′ such that L = Lφ, where Lφ = (id⊗φ)◦∆; (2) L is translation invariant on A; (3) The semigroup (Tt)t≥0 on A (or Ar or Au) associated to L by the formula (4) the sesquilinear form Q defined by Q(a, b) = −h(a∗L(b)) on A satisfies TtA = exp⋆ tL is translation invariant. (7.1) Q(a, b)1 = (m∗ ⊗ Q)(∆(a), ∆(b)), a, b ∈ A, where m∗ denotes the sesquilinear map obtained from the multiplication, namely, m∗(a, b) = a∗b. Proof. We already observed in Subsection 2.5 the equivalence (1) ⇔ (2) whereas the equivalence (2) ⇔ (3) follows from Theorem 3.4, so that we need to prove only (1) ⇔ (4). Let us assume that L satisfies (id ⊗ L) ◦ ∆ = ∆ ◦ L. Then, using Sweedler notation and the invariance of the Haar state, On the other hand, if we assume that Equation 7.1 holds, then (m∗ ⊗ Q)(∆(a), ∆(b)) = a∗(1)b(1)Q(a(2), b(2)) = −a∗(1)b(1)h(cid:0)a∗(2)L(b(2))(cid:1) = −(id ⊗ h)(cid:0)(a∗(1) ⊗ a∗(2))(id ⊗ L)∆(b)(cid:1) = −(id ⊗ h)(cid:0)(a∗(1) ⊗ a∗(2))(L(b)(1) ⊗ L(b)(2))(cid:1) = −(id ⊗ h)∆(cid:0)a∗L(b)(cid:1) = −h(a∗L(b))1 = Q(a, b)1. (h ⊗ h)(a∗ ⊗ 1)∆(b∗)(cid:0)(id ⊗ L)∆(c)(cid:1) = (h ⊗ h)(cid:0)a∗b∗(1)c(1) ⊗ b∗(2)L(c(2))(cid:1) = −h(cid:0)a∗b∗(1)c(1)(cid:1)Q(cid:0)b(2), c(2)(cid:1) = −h(cid:0)a∗(m∗ ⊗ Q)(∆(b), ∆(c))(cid:1) = −h(a∗)Q(b, c) (h ⊗ h)(a∗ ⊗ 1)∆(b∗)(∆ ◦ L)(c) = (h ⊗ h)(cid:0)a∗b∗(1)(Lc)(1) ⊗ b∗(2)(Lc)(2)(cid:1) = h(cid:0)a∗(id ⊗ h)∆(b∗L(c))(cid:1) = h(a∗)h(b∗L(c)) = −h(a∗)Q(b, c). and Since A ⊙ A is the linear span of (A ⊗ 1)∆(A) and h ⊗ h is faithful on A ⊙ A, we conclude that L is translation invariant. Corollary 7.4. Let φ ∈ A′ be the generating functional of a GNS-symmetric L´evy process and Eφ the associated Dirichlet form. Then the sesquilinear form Q on A defined by (cid:3) Q(a, b) := Eφ(ih(a), ih(b)) a, b ∈ A satisfies Eq. (7.1). Conversely, let L is a GNS-symmetric operator on A such that L(1) = 0, L is hermitian and positive on ker ε, and the sesquilinear form on A defined by Q(a, b) := −h(a∗Lb) a, b ∈ A 30 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA satisfies Eq. symmetric L´evy process. (7.1). Then L = Lφ for a generating functional φ of a GNS- 8. Derivations, cocycles and Spectral Triples In this section we associate to any L´evy process on a CQG G = (A, ∆), a natural derivation on its Hopf ∗-subalgebra A, with values in a Hilbert bimodule over the C∗-algebra A = C(G). This gives rise, on the same bimodule, to a self- adjoint operator D, with respect to which we prove that the elements of A are "Lipschitz" in a natural, suitable sense. The construction makes essential use of the Schurmann triple associated to the generator of the process. In case the GNS symmetry holds true, we will show that the derivation is, essentially, a differential square root of the generator Hφ. Moreover, if the spec- trum of Hφ on L2(A, h) is discrete, then the Hilbert bimodule and the operator D form a spectral triple in the sense of the noncommutative geometry of A. Connes [Co94]. This fact suggests to refer to D as the Dirac operator associated to the process. We remark that the role of GNS symmetry of the process is to provide a suitable closability property of the derivation, needed to prove that the Dirac operator D is self-adjoint and that the spectrum of the Dirac Laplacian D2 coincides with that of the generator Hφ, away from zero. We will show in Section 9 that in case the CQG is a compact Lie group and the L´evy process is the Brownian motion associated to a given Riemannian metric, the differential calculus illustrated above reduces to the familiar one: the derivation coincides with the gradient operator and the Lipschitz property has the usual meaning. Consider on the Hopf ∗-subalgebra A of a compact quantum group G = (A, ∆), the generating functional φ ∈ A′ of a L´evy process and its associated Schurmann triple ((π, Hπ), η, φ) on a Hilbert space Hπ (see Remark 5.7). Denote by λL, λR : A → B(L2(A, h)) the left and right actions of A on the Hilbert space L2(A, h) λL(a)(bξh) := abξh, λR(a)(bξh) := baξh, a, b ∈ A, where ξh ∈ L2(A, h) denotes the cyclic vector representing the Haar state. Recall now that ∆ : A → A ⊗ A is a morphism of C∗-algebras and λL, π are representa- tions of the C∗-algebra A, so that λL⊗π is a representation of the C∗-algebra A⊗A. Correspondingly, consider the left and right actions ρL, ρR : A → B(L2(A, h)⊗Hπ) of A on the Hilbert space L2(A, h) ⊗ Hπ defined by ρL := (λL ⊗ π) ◦ ∆ ρR := λR ⊗ idHπ SYMMETRIES 31 or, more explicitly, by ρL(a)(bξh ⊗ v) = ((λL ⊗ π) ◦ ∆(a))(bξh ⊗ v) =X a(1)bξh ⊗ π(a(2))v ρR(a)(bξh ⊗ v) = λR(a)(bξh) ⊗ v = baξh ⊗ v , for a, b ∈ A and v ∈ Hπ. The actions λL , ρL are continuous and form representa- tions of the C∗-algebra A. Likewise, also the actions λR , ρR are continuous and form antirepresentations of the C∗-algebra A or representations of the opposite C∗-algebra Aop. Moreover, as λL , λR (resp. ρL , ρR) commute, they provide a A-bimodule structure on the Hilbert space L2(A, h) (resp. L2(A, h) ⊗ Hπ). In the following we shall adopt the simplified notations: L2(A, h) ⊗ Hπ we write for a ∈ A and ξ ∈ a · ξ := ρL(a)ξ , ξ · a := ρR(a)ξ . Recall that we denote by ih : A → L2(A, h) the GNS embedding (cf. page 14) ih(a) := aξh a ∈ A. Proposition 8.1. Consider on the Hopf ∗-subalgebra A of a compact quantum group G = (A, ∆), the generating functional φ ∈ A′ of a L´evy process, its asso- ciated Schurmann triple ((π, Hπ), η, φ) and the induced A-bimodule structure on L2(A, h) ⊗ Hπ. Then the linear map defined by ∂ : A → L2(A, h) ⊗ Hπ ∂ := (ih ⊗ η) ◦ ∆ or, more explicitly, by ∂a = (ih ⊗ η)(∆a) =X a(1)ξh ⊗ η(a(2)) a ∈ A , is a derivation in the sense that it satisfies the Leibniz rule ∂(ab) = (∂a) · b + a · (∂b) a, b ∈ A . Proof. The map is well defined because ∆(A) ⊆ A⊗A (where, forcing notation a little bit, we denoted by A⊗A the image in A⊗ A of the subspace A⊙A ⊆ A⊙ A under the canonical quotient map from A ⊙ A to A ⊗ A). 32 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA As the map η : A → Hπ is a 1-cocycle and the counit ε : A → C satisfies the identity (id ⊗ ε) ◦ ∆ = id, i.e. P b(1)ε(b(2)) = b, we have ∂(ab) = (ih ⊗ η)(∆(ab)) = (ih ⊗ η)(∆(a)∆(b)) a(1),j b(1),kξh ⊗ [π(a(2),j)η(b(2),k) + η(a(2),j)ε(b(2),k)] a(1),j b(1),kξh ⊗ η(a(2),j b(2),k) = Xj,k = Xj,k = Xj + Xk = ρL(a)(∂(b)) + λR(Xk λL(a(1),j) ⊗ π(a(2),j)! Xk λR(b(1),k) ⊗ ε(b(2),k)idHπ! Xj = ρL(a)(∂(b)) + (λR(b) ⊗ idHπ) (∂(a)) = ρL(a)(∂(b)) + ρR(b)(∂(a))a · ∂(b) + ∂(a) · b . b(1),kξh ⊗ η(b(2),k)! a(1)j ξh ⊗ η(a(2),j)! b(1),kε(b(2),k)) ⊗ idHπ! (∂(a)) (cid:3) Proposition 8.2. Let φ ∈ A′ be a GNS-symmetric generating functional and consider the hermitian convolution generator Lφ : A → A, its Hilbert-space ex- tension (Hφ, D(Hφ)) as well as the Dirichlet form (Eφ, D(Eφ)) (see Section 7). Then the operator d : D(d) → L2(A, h) ⊗ Hπ defined as D(d) := ih(A) = Aξh ⊂ L2(A, h) , d(ih(a)) := ∂a, a ∈ A, is closable and a ∈ A . L2(A,h)⊗Hπ = 2hih(a), Hφih(a)iL2(A,h) = 2 Eφ[ih(a)], kd(ih(a))k2 Proof. We have a(1),j ⊗ η(a(2),j),Xk kd(ih(a))k2 = hXj = Xj,k h(cid:0)a∗(1),j a(1),k(cid:1)(cid:2)φ(a∗(2),j a(2),k) − ε(a∗(2),j)φ(a(2),k) − φ(a∗(2),j)ε(a(2),k)(cid:3) = Xj,k h(cid:0)a∗(1),j a(1),k(cid:1)φ(a∗(2),ja(2),k) −Xj,k h(cid:0)a∗(1),ja(1),k(cid:1)ε(a∗(2),j)φ(a(2),k) −Xj,k h(cid:0)a∗(1),j a(1),k(cid:1)φ(a∗(2),j)ε(a(2),k). a(1),k ⊗ η(a(2),k)i SYMMETRIES 33 The first term vanishes because, by the GNS symmetry of Lφ, we have The second term becomes Xj,k h(cid:0)a∗(1),j a(1),k(cid:1)φ(a∗(2),ja(2),k) = h(cid:0)1 · Lφ(a∗a)(cid:1) = h(Lφ(1) a∗a) = 0. Xj,k h(cid:0)a∗(1),ja(1),k(cid:1)ε(a∗(2),j)φ(a(2),k) = h(cid:0)(Xj a(1),j ε(a(2),j))∗(Xk a(1),kφ(a(2),k))(cid:1) = h(cid:0)a∗Lφ(a)(cid:1) . The third term is the complex conjugate of the second term and, since it is real, they are the same. Hence the identity kdξk2 = 2hξ, HφξiL2(A,h) = 2Eφ[ξ] holds for ξ ∈ D(d). Since φ is also KMS symmetric, Lemma 7.1 implies that the quadratic form is closable. It follows that the operator d is closable, too. From now on we will denote by the same symbol (cid:0)d, D(d)(cid:1) the closure of the closable operator considered in the previous result. One of the conclusion of the above result reads H φ = 1 the L2- generator has the aspect of a "generalized Laplacian" composed of a "generalized divergence" operator d∗ and a "generalized gradient" operator d. In other words, the operator d (essentially the derivation ∂) is a differential square root of the L2-generator. 2 d∗ ◦ d, i.e. (cid:3) The next result shows that in the noncommutative space C(G), the elements of the dense subalgebra A have a noncommutative Lipschitz property. Below Hilbert space Hπ which is the completion of the algebraic tensor product of A and Hπ with respect to the norm we denote by Ab⊗Hπ the projective tensor product of the C∗-algebra A and the kxkA b⊗Hπ := inf{XkaikAkξikHπ , where x =X ai ⊗ ξi}, see [Gro55] or [Ry02]. Proposition 8.3. Let φ ∈ A′ be a GNS-symmetric generating functional with associated Schurmann triple ((π, Hπ), η, φ). Let us consider the Hilbert space Hφ := (L2(A, h) ⊗ Hπ) ⊕ L2(A, h) as a A-bimodule under the commuting left and right actions πL := ρL ⊕ λL , πR := ρR ⊕ λR . Consider also on Hφ the self-adjoint operator d D :=(cid:18) 0 d∗ 0 (cid:19) . Then the commutator [D, πL(a)] is bounded for all a ∈ A with norm bounded by k[D, πL(a)]k ≤ k∂akA b⊗Hπ . 34 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA Proof. It follows from [D, πL(a)] = D ◦ πL(a) − πL(a) ◦ D = (cid:18) 0 = (cid:18) = (cid:18) d d∗ 0 (cid:19)(cid:18) ρL(a) 0 0 λL(a) (cid:19) −(cid:18) ρL(a) 0 0 d ◦ λL(a) − ρL(a) ◦ d 0 d∗ ◦ ρL(a) − λL(a) ◦ d∗ 0 −(cid:0)d ◦ λL(a∗) − ρL(a∗) ◦ d(cid:1)∗ d 0 λL(a) (cid:19)(cid:18) 0 d∗ 0 (cid:19) (cid:19) (cid:19) 0 d ◦ λL(a) − ρL(a) ◦ d that [D, πL(a)] is bounded for all a ∈ A if and only if d ◦ λL(a) − ρL(a) ◦ d is bounded for all a ∈ A. To check that the latter is actually the case, let us observe that, for b ∈ A, we have (cid:0)d ◦ λL(a) − ρL(a) ◦ d(cid:1)ih(b) = d(ih(ab)) − ρL(a)(∂b) = ∂(ab) − ρL(a)(∂b) = ρR(b)(∂a) . For any presentation ∂a =Pn k=1 ak ⊗ ξk ∈ A ⊗ Hπ we then have kρR(b)(∂a)kL2(A,h)⊗Hπ = k(λR ⊗ idHπ)(∂a)kL2(A,h)⊗Hπ akbξh ⊗ ξkkL2(A,h)⊗Hπ = k nXk=1 nXk=1 ≤ kakbξhkL2(A,h)kξkkHπ ≤ kih(b)kL2(A,h) kakkA · kξkkHπ . nXk=1 Optimizing among all presentations ∂a =Pn kρR(b)(∂a)kL2(A,h)⊗Hπ ≤ kih(b)kL2(A,h) · k∂akA b⊗Hπ k=1 ak ⊗ ξk ∈ A ⊗ Hπ we get a, b ∈ A kd ◦ λL(a) − ρL(a) ◦ dk ≤ k∂akA b⊗Hπ a ∈ A . Finally notice that, setting Ta := d◦ λL(a)− ρL(a)◦ d, we have kTak ≤ k∂akA b⊗Hπ, [D, πL(a)] =(cid:18) 0 [D, πL(a)]2 =(cid:18) TaT ∗a −T ∗a 0 Ta 0 (cid:19) T ∗a Ta (cid:19) , 0 k[D, πL(a)]k = kTak , a ∈ A . (cid:3) so that and so that SYMMETRIES 35 Theorem 8.4. Consider, on the Hopf ∗-subalgebra A of a compact quantum group G = (A, ∆), the GNS-symmetric generating functional φ ∈ A′ with Schurmann triple ((π, Hπ), η, φ). Consider also the GNS-symmetric, hermitian convolution generator Lφ : A → A and its closed extension (Hφ, D(Hφ)) on the space L2(A, h), characterized by Hφ(ih(a)) := −ih(Lφa) on its core ih(A) = Aξh ⊂ D(Hφ). If the spectrum of (Hφ, D(Hφ)) is discrete and considering the representation of A = C(G) constructed above πL := ρL ⊕ λL : A → B(Hφ) Hφ := (L2(A, h) ⊗ Hπ) ⊕ L2(A, h) , we have that (A, D, (πL,Hφ)) is a (possibly kernel-degenerate) spectral triple in the sense that • [D, πL(a)] is a bounded operator for all a ∈ A, • D has discrete spectrum on the orthogonal complement of its kernel. Proof. By construction D2 =(cid:18) dd∗ 0 0 d∗d (cid:19) , so that the spectrum of D2 is the union of the spectra of dd∗ and d∗d. Since these two operators are unitarily equivalent on the orthogonal complement of their kernels and zero belongs to the spectrum of d∗d, the spectrum of D2 coincides with the spectrum of 2Hφ, by Proposition 8.2. Since, by assumption, the spectrum of Hφ is discrete we have that the spectrum of D2, hence the one of D, are discrete too on the orthogonal complement of their kernels. This result, together with Theorem 8.3 allows us to conclude the proof. (cid:3) The fact that the kernel of the Dirac operator D may be infinite dimensional is a variation with respect to the original definition of spectral triple given in [Co94], due to the definition of D as an antidiagonal matrix. To construct the associated K-homology invariants this fact has to be taken into account, for example using the methods developed in Section 3 of [CGIS12]. 9. Two classical examples: commutative and cocommutative CQGs 9.1. Algebras of functions on compact groups. Let G be a compact Lie group and let C(G) denotes the commutative C∗-algebra of all continuous func- tions on G. Then C(G) is a compact quantum group with the comultiplication defined by ∆ : C(G) → C(G) ⊗ C(G) ∼= C(G × G), ∆(f )(s, t) = f (st), f ∈ C(G), s, t ∈ G. 36 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA The counit and the antipode are defined on the dense ∗-subalgebra Cc(G) gener- ated by the coefficients of arbitrary continuous finite-dimensional representation π, i.e. functions πij : G → C, and they are given by ε(f ) = f (e), S(f )(x) = f (x−1), f ∈ Cc(G). This is a general example of a commutative compact quantum group, in the sense that if A is the algebra of continuous functions on a compact quantum group which is commutative as a C∗-algebra, then there exists a unique compact group G such that A is isomorphic to C(G) with coproduct corresponding to the classical one given above (cf. [Wor87a, Theorem 1.5]). The quantum group C(G) is cocommutative if and only if the group G is com- mutative. It is always of Kac type, i.e. S2 = id. This implies that the modular automorphism group is trivial and that the Haar state is tracial (see Remark 2.3), and so the notions of GNS- and KMS-symmetry coincide (see Remark 4.10). The generating functionals of L´evy processes in G are classified by Hunt's [Li04]). Let {X1, X2, . . . , Xd} be a fixed basis of the Lie formula as follows (cf. algebra g associated to the Lie group G and let x1, x2, . . . xd ∈ C∞c (G) be the local coordinates associated to this basis, i.e. Xi = ∂ at the neutral element e. ∂xi Then an arbitrary generating functional φ is of the form φ(f ) = ajkXjXkf (e) 1 2 ciXif (e) + dXj,k=1 dXi=1 + ZG\{e} f (g) − f (e) − xi(g)Xif (e)! ν(dg) dXi=1 for twice differentiable f . Here ci, ajk are real constants, (ajk)d definite symmetric matrix and the measure ν on G satisfies j,k=1 is a positive ν({e}) = 0, ZU dXi=1 x2 i dν < ∞, ν(G \ U) < ∞ for any neighborhood U of e in G. The first term in the decomposition above is called the drift, whereas the second one is called the diffusion. The measure ν is called L´evy measure. The GNS-symmetric processes correspond to functionals with no drift part and symmetric L´evy measures, i.e. ν(E) = ν(E−1) for measurable E (see [Li04, Proposition 4.3], where such processes are called invariant under the inverse map). The characterisation of ad-invariant processes (called conjugate invariant in [Li04]) depends on the particular group structure. The two extreme cases are abelian Lie groups and simple Lie groups. In the first case, as observed in Section 6, the adjoint action is trivial and all functionals are ad-invariant. If the Lie group is simple and connected, then the adjoint action ad(f )(x, y) = f (xyx−1) SYMMETRIES 37 has trivial kernel. Then the L´evy measure of an ad-invariant process must be [App10]), that is ν(gEg−1) = ν(E) for all conjugate-invariant (or central, cf. measurable E. Moreover, the drift part vanishes and the diffusion part is (up to a constant) the Beltrami-Laplace operator on G (see [Li04, Propositions 4.4, 4.5]). In the case the Dirichlet form reduces to the Dirichlet integral on G E[a] =ZG ∇a(g)2 dg defined on the Sobolev space H 1,2(G) of functions having square integrable gradi- ent and the derivation is just the gradient operator. We refer to [Li04] for details on this topic. 9.2. C∗-algebra of a countable discrete group. Let Γ be a countable discrete group and let ℓ2(Γ) denote the Hilbert space of all square-summable functions on Γ. The space ℓ2(Γ) is spanned by the orthonormal basis {δg : g ∈ Γ}, where as usual δg(h) = 1 if g = h and δg(h) = 0 otherwise. Then each element g ∈ Γ defines the linear operator λg : ℓ2(Γ) → ℓ2(Γ) by the formula λg(δh) = δgh, h ∈ Γ. Each λg is a unitary operator and the mapping g → λg is called the left regular unitary representation of the Hilbert space Γ on ℓ2(Γ). The closure of the ∗-algebra generated by {λg : g ∈ Γ} in B(ℓ2(Γ)) is denoted by C∗r (Γ) and called the reduced C∗-algebra or the group algebra of Γ. One can also define the universal C∗-algebra of the group, denoted by C∗u(Γ), by taking the direct sum of all cyclic representations of Γ (universal representation) instead of the left regular one. The two algebras are isomorphic if and only if Γ is amenable, cf. [Ped79]. The mapping ∆ defined by ∆(λg) = λg ⊗ λg extends (in a unique way) to a ∗-homomorphism from C∗r (Γ) to C∗r (Γ) ⊗ C∗r (Γ) which preserves the unit. The pair (C∗r (Γ), ∆) is a compact quantum group. The linear span A of {λg : g ∈ Γ} in B(ℓ2(Γ)) is a ∗-Hopf algebra on which counit and antipode are defined by ε(λg) = 1 and S(λg) = λg−1 respectively, for g ∈ Γ. The quantum group C∗r (Γ) is always cocommutative (i.e. the comultiplication is invariant under the flip). Moreover, each algebra of continuous functions on a compact quantum group which is cocommutative is essentially of this form (there exists a unique discrete group Γ and ∗-homomorphisms C∗u(Γ) → A → C∗r (Γ)), see [Wor87a, Theorem 1.7]. Cocommutativity implies that the adjoint action is trivial: ad(a) = 1⊗ a, adh(a) = a and all functionals are ad-invariant φ◦ adh = φ. The algebra C∗r (Γ) is of Kac type so that the modular automorphism group is trivial. The Haar state is a trace and, on generators, it is explicitly given by h(δg) = 0 for g 6= e and h(δe) = 1. The GNS Hilbert space L2(C∗r (Γ), h) can then be identified with l2(Γ). 38 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA In this case the notions of GNS and KMS symmetry coincide, and φ is symmet- ric iff φ(λg) = φ(λg−1) for any g ∈ Γ. Moreover, symmetric generating functionals of L´evy processes are in one-to-one correspondence with (obviously continuous) positive, conditionally negative-type functions d : Γ → [0,∞), d(g) = −φ(λg), g ∈ Γ (cf. [CS03, Example 10.2]). The associated Dirichlet form is given by E[a] =Xg∈Γ d(g)a(g)2 a ∈ l2(Γ) and the generator of the Markovian semigroup on l2(Γ) is just the multiplication operator (Hφa)(g) = d(g)a(g), defined for those a ∈ l2(Γ) such that the right hand side in square integrable. The derivation associated to the KMS symmetric generating functional φ (recall Section 8) is given by ∂(λg) = λg ⊗ η(λg) for g ∈ Γ, where η is the 1-cocycle corresponding to φ in the Schurmann triple ((π, D), η, φ). Composing the 1- cocycle η on the C∗-algebra C∗r (Γ) with the left regular representation, one obtains the 1-cocycle on the group Γ. In terms of this, the negative type function is given by c : Γ → D, c(g) = η(λg) Identifying l2(Γ)⊗ D with l2(Γ, D), one obtains that the derivation above reduces to the multiplication operator d(g) = kc(g)k2 D . (∂a)(g) = c(g)a(g) g ∈ Γ , defined for all a in the domain of the Dirichlet form. The spectrum of the generator Hφ is discrete if and only if the negative-type function d is proper on Γ (a condition which is met, for example, for some length functions of finitely generated groups, see for example [CCJJV01]). In these situations the construction of a spectral triple shown in Theorem 8.4 applies. 10. Example: free orthogonal quantum groups O+ N Let N ≥ 2. The compact quantum group (Cu(O+ N ), ∆) is the universal unital C∗-algebra generated by N 2 self-adjoint elements vjk, 1 ≤ j, k ≤ N subject to the condition that the matrix V = (vjk) ∈ MN⊗Cu(O+ N ) is a unitary corepresentation, i.e. that vℓjvℓk = δjk = NXℓ=1 vjℓvkℓ NXℓ=1 SYMMETRIES 39 and ∆(vjk) = NXℓ=1 vjℓ ⊗ vℓk N ) associated to O+ N is not faithful on Cu(O+ for all 1 ≤ j, k ≤ N, see [VDW96, Ban96]. The equivalence classes of the irre- ducible unitary corepresentations of this compact quantum group can be indexed by N, with u(0) = 1 the trivial corepresentation and u(1) = (vjk)1≤j,k≤N the corep- resentation whose coefficients are exactly the N 2 generators of Cu(O+ N ) (this is also called the fundamental corepresentation of O+ N ). The dense -Hopf alge- bra Pol(O+ N , also called the -algebra of polynomial on O+ N , is the *-algebra generated by vjk, 1 ≤ j, k ≤ N. The compact quantum group O+ N is called the free orthogonal compact quantum group. For N > 2 it is not co-amenable, i.e. the Haar state of O+ N ), therefore we will study the Markov semigroups of L´evy processes on Pol(O+ N ) on the reduced C∗-algebraic version Cr(O+ N ) of O+ N . The compact quantum group O+ N is of Kac type, and therefore a generating functional φ is KMS-symmetric if and only if it is GNS-symmetric, which is the case if the characteristic matrices are symmetric, i.e. if φ(u(s) kj ) for all s ∈ N and j, k running from 1 up to the dimension of the sth corepresentation. Corollary 6.11 reduced the problem of classifying ad-invariant generating func- tionals on a compact quantum group to the classification of generating functionals on the subalgebra of central functions. For the free orthogonal quantum group O+ N the algebra of central functions is isomorphic the the C∗-algebra of contin- uous functions on the interval [−N, N], cf. [Bra11, Corollary 4.3]. Furthermore, the restriction of the counit to this subalgebra is the evaluation of a function in a boundary point. jk ) = φ(u(s) Let us begin by describing linear functionals which are positive on a given interval and vanish in a given point. Proposition 10.1. Denote by τx : C([0, 1]) → C the evaluation of a function in x ∈ [0, 1]. (a) Suppose 0 < x < 1. A linear functional ϕ : C[x] → C with ϕ(1) = 0 is positive on the cone Kx([0, 1]) = C[x] ∩ C([0, 1])+ ∩ ker(τx) if and only if there exist real numbers a, b with a ≥ 0 and a finite measure ν on [0, 1] with ν({x}) = 0 such that ϕ(f ) = bf′(x) + af′′(x) +Z 1 0 (cid:0)f (y) − f (x) − yf′(x)(cid:1) ν(dy) (y − x)2 for all polynomials f ∈ C([0, 1]). the characteristic triple of the linear functional ϕ. The triple (a, b, ν) is uniquely determined by ϕ. We will call (a, b, ν) 40 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA (b) Suppose x ∈ {0, 1}. Then a linear functional ϕ : C[x] → C with ϕ(1) = 0 is positive on the cone Kx([0, 1]) = C[x] ∩ C([0, 1])+ ∩ ker(τx) if and only if there exist a real number d with d ≥ 0 if x = 0, and d ≤ 0 if x = 1, and a finite measure µ on [0, 1] with µ({0}) = 0 such that ϕ(f ) = df′(x) +Z 1 0 (cid:0)f (y) − f (x)(cid:1) µ(dy) y for all polynomials f ∈ C([0, 1]). characteristic pair of the linear functional ϕ. The pair (d, µ) is uniquely determined by ϕ. We will call (b, ν) the Proof. (a) This is actually the classical L´evy-Khinchin formula for L´evy pro- cesses on R, see, e.g., [Sat99, Theorem 8.1], which can be viewed as a special case of Hunt's formula [Hun56]. Skeide [Ske99] has given a C∗-algebraic proof which doesn't use the group structure, but works for the C∗-algebra of continuous functions on a compact set, with a character given by evaluation in a fixed point which has neighborhood with Euclidean coordinates (i.e. smooth functions admit a Taylor expansion around the fixed point). (b) This is actually the classical L´evy-Khinchin formula for subordinators, cf. [Sat99, Theorem 21.5]. We prove the formula for x = 0, the case x = 1 follows easily by a change of variable t 7→ 1 − t. By (a), since ϕ has to be positive also on the smaller cone given by polynomials that vanish in x = 0 and which are positive on [−ε, 1] for any ε > 0, there exists a unique triple (a, b, ν) with a, b ∈ R with a ≥ 0 and ν a finite measure on [0, 1], such that ϕ(f ) = bf′(0) + af′′(0) +Z 1 For n ∈ N we set gn(y) = 1 and g′′n(0) = −n, therefore n+1 yPn 0 (cid:0)f (y) − f (0) − yf′(0)(cid:1) ν(dy) y2 . k=0(1− y)k. We have gn ∈ K0([0, 1]), g′n(0) = 1, y2 0 (cid:0)gn(y) − y(cid:1) ν(dy) n→∞−→ Z 1 y 0 ν(dy) 0 ≤ ϕ(gn) = b − na +Z 1 b ≥Z 1 0 (cid:0)y − gn(y)(cid:1) ν(dy) ϕ(f ) = df′(0) +Z 1 y2 0 (cid:0)f (y) − f (0)(cid:1) µ(dy) y . for all n ∈ N. The sequence (gn)n∈N is decreasing and therefore we must have a = 0. By monotone convergence we get which proves that the measure 1 0, we get the desired formula y ν is finite. Putting µ = 1 y ν and d = b−R 1 0 ν(dy) y ≥ SYMMETRIES 41 Conversely, since a polynomial f which vanishes in x = 0 and is positive on [0, 1] has a positive derivative at x = 0, it is clear that any such functional is positive on K0([0, 1]). Uniqueness follows from (a). (cid:3) This result allows us to describe all ad-invariant generating functionals on N ). This result can be considered as Hunt's formula for ad-invariant L´evy Pol(O+ processes on the free orthogonal quantum group O+ N . Let us denote by Pol0(O+ N ) the algebra of central polynomial functions on O+ N , see Eq. (6.1). We will use the same isomorphism between Pol0(O+ N ) and poly- nomials Pol([−N, N]) as Brannan [Bra11]. Recall that Banica [Ban96] showed that the equivalence classes of irreducible unitary corepresentations of O+ N can be labelled by non-negative integers and that they satisfy the "fusion rules" u(s) ⊗ u(t) ∼= u(s−t) ⊕ u(s−t+2) ⊕ · · · ⊕ u(s+t) for s, t ∈ N. Since the trivial corepresentation u(0) = 1 has dimension 1 and the fundamental corepresentation u(1) = (vjk)1≤j,k≤N has dimension N, one can show by induction that the dimensions of the irreducible unitary corepresentations are given by Chebyshev polynomials of the second kind, Ds = Us(N). The N ) onto the algebra of central N ) → Pol0(O+ functions is therefore given by conditional expectation fadh : Pol(O+ fadh(u(s) jk ) = j=1 u(s) jj denotes the trace of u(s). 1 Us(N) δjkχs, where χs =PDs relation The fusion rules imply that the characters satisfy the three-term recurrence χ1χs = χs+1 + χs−1 N )0 ∼= Pol([−N, N]) by setting for s ≥ 1, we get the desired isomorphism Pol(O+ χs 7→ Us for s ∈ N, where Us denotes the sth Chebyshev polynomial of the second kind, defined by U0(x) = 1, U1(x) = x, and Us+1(x) = xUs(x) − Us−1(x). Theorem 10.2. The ad-invariant generating functional on Pol(O+ form N ) are of the with L defined on Pol(O+ L = L ◦fadh N )0 ∼= Pol([−N, N]) by Lf = −bf′(N) +Z N −N f (x) − f (N) N − x dν(x) where b ≥ 0 is a real number and ν is a finite measure on [−N, N] with ν({N}) = 0.. Proof. This follows from Theorem 6.10 and Proposition 10.1. (cid:3) 42 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA Using the discussion above, we can give a formula for the values of ad-invariant generating functionals on the coefficients of the irreducible unitary corepresenta- tions of O+ N . Corollary 10.3. The ad-invariant generating functional on Pol(O+ N ) given in Theorem 10.2 with characteristic pair (b, ν) acts on the coefficients of unitary irreducible corepresentations of O+ N as jk(cid:1) = L(cid:0)u(s) δjk Us(N)(cid:18)−bU′s(N) +Z N −N Us(x) − Us(N) N − x ν(dx)(cid:19) for s ∈ N, where Us denotes the sth Chebyshev polynomial of the second kind. Remark 10.4. Since the characteristic matrices of L are diagonal, we can read off the eigenvalues of TL from Corollary 10.3. Assume for simplicity b = 1, ν = 0. Then the eigenvalues of TL are given by λs = − U′s(N) Us(N) , s ∈ N, with multiplicities given by the square of the dimension ms = D2 u(s). Recall that the "spectral dimension" dD of the associated spectral triple is, by definition (see [Co04a], [Co04b]), the abscissa of convergence of the zeta function z 7→ ZD(z) := Tr (D−z), initially defined for z ∈ C with Re z > 0. It coincides present situation of Corollary 10.3 and assuming N = 2, b = 1, ν = 0, we have Us(2) = s + 1, U′s(2) = s(s+1)(s+2) with the infimum of all d > 0 such that the sumPs ms(−λs)−d/2 is finite. In the , 6 s =(cid:0)Us(N)(cid:1)2 of λs = − s(s + 2) 6 and finally dD = 3. This value of the spectral dimension agrees nicely with the known fact that O+ 2 is isomorphic to SU−1(2), see [Ban96], and that C(SU−1(2)) can be realized by matrix-valued functions on the three-dimensional Lie group SU(2), cf. [Zak91]. On the other hand, for N > 2, we have Us(N) = U′s(N) = = q(N)s+1 − q(N)−s−1 q(N) − q(N)−1 , q′(N) q(N) q′(N) q(N) s s(cid:0)q(N)s+2 − q(N)−s−2(cid:1) − (s + 2)(cid:0)q(N)s − q(N)−s(cid:1) (cid:0)q(N) − q(N)−1(cid:1)2! , q(N)s+1 − q(N)−s−1 q(N) − q(N)−1 − 2 (cid:0)q(N) − q(N)−1(cid:1)2 q(N)s − q(N)−s SYMMETRIES 43 with q(N) = 1 2(N + √N 2 − 4) > 1, q′(N) = 1 q(N) s − 2 2(cid:16)1 + N√N 2−4(cid:17) > 0, and (cid:0)q(N)s+1 − q(N)−s−1(cid:1)(cid:0)q(N) − q(N)−1(cid:1)! . q(N)s − q(N)−s q′(N) λs = − Since q(N) is bigger then 1 (and fixed), the term q(N)s − q(N)−s (cid:0)q(N)s+1 − q(N)−s−1(cid:1)(cid:0)q(N) − q(N)−1(cid:1) → 0 as s → ∞. q(N ) s, while the multiplicities ms = Us(N)2 ∼= q(N )2s This implies that the growth of the eigenvalues λs (as a function of s) is asymp- totically linear λs ∼= − q′(N ) grow exponentially, therefore the sum Xs ms(−λs)−d/2 ∼=Xs (1−q(N )−2)2 q(N)2s sd/2 can never converge, which means that dD = +∞. 11. Example: Woronowicz quantum group SUq(2) Let us fix q ∈ (0, 1). The compact quantum group C(SUq(2)) is the universal unital C∗-algebra generated by α and γ subject to the following relations α∗α + γ∗γ = 1, αα∗ + q2γγ∗ = 1, γ∗γ = γγ∗, αγ = qγα, αγ∗ = qγ∗α with the comultiplication extended uniquely to a unit-preserving ∗-homomorphism from the formulas ∆(α) = α ⊗ α − qγ∗ ⊗ γ, ∆(γ) = γ ⊗ α + α∗ ⊗ γ. For C(SUq(2)) the equivalence classes of irreducible unitary corepresentations N and are of dimension ns = jk )j,k the indices j, k run over the set {−s,−s + [PW00] for the detailed description of u(s)). Moreover, are indexed by non-negative half-integers s ∈ 1 2s + 1. For each u(s) = (u(s) 1, . . . , s − 1, s} (see eg. every corepresentation is equivalent to its contragredient one and we have 2 (11.1) S(ukj) = (u(s) jk )∗ = (−q)k−ju(s) −j,−k. The quantum group is neither commutative nor cocommutative. The Woronow- icz characters, the modular automorphism group, the unitary antipode and the 44 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA quantum dimension are the following (cf. [Wor87a, Appendix A1]: (11.2) (11.3) (11.4) (11.5) fz(u(s) σz(u(s) R(u(s) jk ) = q2jzδjk, jk ) = fiz ⋆ u(s) jk ) = S(f 1 2 jk ⋆ fiz = q2iz(j+k)u(s) jk , ) = qk−j(u(s) ⋆ u(s) kj )∗, jk ⋆ f f1(u(s) kk ) = q2k = q−2s[2s + 1]q2. Ds = sXk=−s − 1 2 sXk=−s The following example describes the irreducible representations of C(SUq(2)) and the related opposite representations (cf. Section 5). Example 11.1. On A = C(SUq(2)) we have two families of irreducible ∗- representation indexed by θ ∈ [0, 2π): (1) the 1-dimensional representations δθ : A → C: δθ(γ) = 0; δθ(α) = eiθ, (2) the infinitely-dimensional representations on a Hilbert space ρθ : A → B(ℓ2): ρθ(α)en = W en, ρθ(γ)en = eiθqnen, where (en)n∈N is the standard orthonormal basis of ℓ2 and W is the weighted shift defined by W e0 = 0 and W en =p1 − q2nen−1 for n ≥ 1. We check directly that δop θ = δ−θ and ρop θ = ρπ+θ. Indeed, we note first that R(α) = α∗ and R(γ) = −γ. So for δθ we have δop θ (α)¯1 = δθ(R(α∗))1 = δθ(α)1 = eiθ1 = e−iθ¯1 = δ−θ(α)¯1 and δop θ (γ)¯1 = δθ(R(γ∗))1 = −δθ(γ∗)1 = 0 = δ−θ(γ)¯1. Similarly, for ρθ we compute ρop θ (γ)¯en = −ρθ(γ∗)en = −e−iθqnen = ei(π+θ)qn¯en = ρπ+θ(γ)¯en. ϕ (α)¯en = ρϕ(α)en = W en =p1 − q2n¯en = ρπ+θ(α)¯en. ρop SYMMETRIES 45 11.1. GNS-symmetric generators. We first describe a generic GNS-symmetric functional on SUq(2) and provide an example of an unbounded generating func- tional. Proposition 11.2. A hermitian functionals φ on SUq(2) defined by (11.6) φ(u(s) jk ) = cs,jδjk with real constants (cs,j)s∈ 1 2 N,−s≤j≤s is GNS-symmetric. Reciprocally, any GNS-symmetric generator φ on SUq(2) must be of the form (11.6) and it is hermitian if and only if the constants satisfy the supplementary symmetry condition: cs,j = cs,−j for all s and j. Proof. We calculate explicitly that φ of the form (11.6) is invariant under the antipode: φ ◦ S(u(s) jk ) = φ((u(s) kj )∗) = φ(u(s) kj )) = cs,jδjk = cs,jδjk = φ(u(s) jk ). Conversely, suppose that φ ◦ S = φ, then also φ ◦ S2 = φ. On SUq(2) we have jk ) and φ can have S2(u(s) non-zero values only on the diagonal. By Remark 4.7, all cs,j must be real. jk ) = q2(j−k)φ(u(s) jk ) = q2(j−k)u(s) jk , thus φ(u(s) jk ) = φ ◦ S2(u(s) jj ) and cs,−j = φ(u(s) Finally, cs,j = ¯cs,j = φ(u(s) last part follows. −j,−j) = φ((u(s) jj )∗) from which the (cid:3) 11.2. Unbounded GNS-symmetric generator. Let π be the ∗-representation of SUq(2) on ℓ2(N × Z) given by π(α)ek,n = p1 − q2kek−1,n (k ≥ 1), π(α∗)ek,n = p1 − q2k+2ek+1,n (k ≥ 0), π(γ)ek,n = qkek,n−1, π(γ∗)ek,n = qkek,n+1, π(α)e0,n = 0, where {ek,n; k ≥ 0, n ∈ Z} is the standard orthonormal basis of ℓ2(N × Z). For a fixed 0 < λ < 1 let us consider a Poisson type generator φλ(a) = hvλ, (π − ε)(a)vλi with vλ = λkek,0. ∞Xk=0 46 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA The related cocycle ηλ(a) = (π − ε)(a)vλ is uniquely determined by the value on α∗ (see [SS98]), where it equals λk(cid:0)p1 − q2k+2ek+1,0 − ek,0(cid:1) ηλ(α∗) = = ∞Xk=0 ∞Xk=1 = −e0,0 + ∞Xk=0 ∞Xk=0 λk(π − ε)(α∗)ek,0 = λk−1p1 − q2kek,0 − ∞Xk=1(cid:0)λk−1p1 − q2k − λk(cid:1)ek,0. λkek,0 Note that ηλ(α∗) ∈ H since kηλ(α∗)k2 = 1 + Define also a cocyle η∞ by its value on α∗: η∞(α∗) = −e0,0 + ∞Xk=1(cid:12)(cid:12)λk−1p1 − q2k − λk(cid:12)(cid:12)2 ∞Xk=1 (p1 − q2k − 1)ek,0 = − ≤ 1 + 4 ∞Xk=1 λ2(k−1) < +∞. ∞Xk=0 (1 −p1 − q2k)ek,0. We shall show that η∞(α∗) ∈ H and that ηλ(α∗) → η∞(α∗) in H when λ → 1−. For the first part, we check directly that kη∞(α∗)k2 = ∞Xk=0(cid:0)1 −p1 − q2k(cid:1)2 = ∞Xk=0 Next, we show the convergence: q4k (1 +p1 − q2k)2 < 1 1 − q4 < +∞. kηλ(α∗) − η∞(α∗)k2 = k ∞Xk=1(cid:0)λk−1p1 − q2k − λk −p1 − q2k + 1)ek,0k2 = = ∞Xk=1(cid:12)(cid:12)(1 −p1 − q2k)(1 − λk−1) + λk−1(1 − λ)(cid:12)(cid:12)2 ∞Xk=1 ∞Xk=1 (1 −p1 − q2k)2(1 − λk−1)2 + (1 − λ)2 ∞Xk=1 (1 −p1 − q2k)(1 − λk−1)λk−1. + 2(1 − λ) λ2(k−1) kηλ(α∗) − η∞(α∗)k2 = ≤ (1 − λ)2 + q4k ∞Xk=1 (k − 1)2 ∞Xk=1 q2k 1 +p1 − q2k ∞Xk=1 (k − 1)2q4k + (1 +p1 − q2k)2 (k − 1)λk−1 1 − λ 1 + λ + 2(1 − λ)2 ≤ (1 − λ)2 (1 − λ)2 1 − λ2 + 2(1 − λ)2 ∞Xk=1 q2k(k − 1) Note that 1 − λk−1 = (1 − λ)(λk−2 + . . . + 1) ≤ (k − 1)(1 − λ). This implies SYMMETRIES 47 and we see that each term tends to 0 when λ → 1−. Let us now define by φ∞ the functional related to the cocycle η∞ by the formula φ∞(ab) = hη∞(a∗), η∞(b)i, a, b ∈ ker ε with the additional conditions that φ∞(1) = 0 and that the 'drift' part is zero (which remains to say that φ∞(α) = φ∞(α∗) ∈ R). This way φ∞ is uniquely determined on the whole of A = Lin{1, α − α∗, K2}, where K2 is the linear [SS98]). By the Schoenberg span of products of two elements from ker ε (cf. correspondence, if well-defined, φ∞ is a generating functional of a L´evy process. To see that φ∞ is well-defined, we can check that on K2 the functional is just a limit of functionals related to ηλ. Indeed, if a, b ∈ ker ε then ηλ(a∗), lim µ→1− φ∞(ab) = hη∞(a∗), η∞(b)i = h lim λ→1− ηµ(b)i and since both limits exist we have φ∞(ab) = lim We conclude that λ→1−hηλ(a∗), ηλ(b)i = lim λ→1− φλ(ab) = lim λ→1−hvλ, (π − ε)(ab)vλi. (11.7) φ∞(a) = lim λ→1−hvλ, (π − ε)(a)vλi for a ∈ K2. Our aim now is to show that φ∞ is GNS-symmetric and unbounded. Proposition 11.3. The functional φ∞ is GNS-symmetric. Proof. By Proposition 11.2, it is enough to show that φ∞ vanishes on the non- diagonal coefficients of the corepresentations u(s), s ∈ 1 N. These coefficients are of the form bm,n := αmp(γ∗γ)γn for m, n ∈ Z, (with the notation a−n = (a∗)n for n > 0 and p denoting a polynomial, see [PW00]) and they are off diagonal iff n 6= 0. So it is enough to check that φ∞ vanishes on αm(γ∗γ)kγn (n 6= 0). γ, α − 1 ∈ ker ε imply that q 1 − q We first observe that γ, γ∗ ∈ K2. Indeed, the relation αγ = qγα together with (α − 1)γ ∈ K2. γ(α − 1) − 1 1 − q γ = 2 48 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA Therefore an element αm(γ∗γ)kγn belongs to K2 provided k 6= 0 or n 6= 0. So the formula (11.7) can be applied φ∞(αm(γ∗γ)kγn) = lim λ→1− ∞Xp,r=0 λp+rhep,0, π(α)mπ(γ∗γ)kπ(γ)n)er,0i Since π(γ) and π(γ∗) move (down and up, respectively) the second index of the basis vectors ek,n and since none of π(α), π(α∗) and π(γ∗γ) move the second index, we immediately see that if n 6= 0 then π(α)mπ(γ∗γ)kπ(γ)ner,0 ∈ C· er+m,n, which is orthogonal to ep,0 for any m, k and p, q. So the sum under the limit equals to 0 and thus φ∞ is of the form (11.6). (cid:3) Proposition 11.4. The functional φ∞ is unbounded. Proof. We shall show that φ∞(α∗mαm) → +∞ when m → +∞. Since kα∗mαmkA ≤ kαk2m A ≤ 1, this will imply that φ∞ is unbounded. Observe first that α∗mαm = α∗(m−1)(α∗α)αm−1 = α∗(m−1)(1 − γ∗γ)αm−1 = α∗(m−1)αm−1(1 − q−2(m−1)γ∗γ) and by induction α∗mαm = (1 − γ∗γ)(1 − q−2γ∗γ) . . . (1 − q−2(m−1)γ∗γ), m ≥ 1. Applying the standard formula from the q-calculus (cf. [KS94, Equation (0.3.5)]: (a; q)n = nXk=0 (q2; q2)n (q2; q2)k(q2; q2)n−k qk(k−1)(−a)k we arrive at α∗mαm − 1 = (1 − γ∗γ) . . . (1 − q−2(m−1)γ∗γ) − 1 = mXk=1 (−1)k (q2; q2)m (q2; q2)k(q2; q2)m−k qk(k−1)(γ∗γ)k. We see that each term under the sum contains γ∗γ and thus belongs to K2. SYMMETRIES 49 Now that we have proved that α∗mαm − 1 ∈ K2, we can apply the formula (11.7) to calculate the value of φ∞ on α∗mαm. Namely, φ∞(α∗mαm) = φ∞(α∗mαm − 1) = lim λ→1− λj+khej,0, (π − ε)(α∗mαm − 1)ek,0i λj+khej,0,(cid:2)(IH − π(γ∗γ)) . . . (IH − q−2(m−1)π(γ∗γ)) − IH(cid:3)ek,0i ∞Xj,k=0 ∞Xj,k=0 = lim λ→1− = = − ∞Xk=0(cid:0)(1 − q2k) . . . (1 − q2k−2m+2) − 1(cid:1) m−1Xk=0 ∞Xk=m(cid:0)1 − (1 − q2k) . . . (1 − q2k−2m+2)(cid:1). 1 − We finally note that the infinite sum is non-negative, and so φ∞(α∗mαm) = m − 1 + ∞Xk=m(cid:0)1 − (1 − q2k) . . . (1 − q2k−2m+2)(cid:1) ≥ m − 1. (cid:3) 11.3. KMS-symmetry. In case of C(SUq(2)) it is easy to check that a hermitian φ is KMS-symmetric iff for each s ∈ 1 jk)] is hermitian. Moreover, if a functional φ is hermitian and KMS-symmetric, then the values of φ on the corepresentation matrix u(s) are determined by the values φ(u(s) jk ) for k ≤ j and the conditions N the matrix φ(s) q = [qjφ(us 2 φ(u(s) −j,−k) = (−q)j−kφ(u(s) jk ) and φ(u(s) kj ) = qj−kφ(u(s) j,k). Below we provide an example of a KMS-symmetric generator which is not GNS- symmetric. Example 11.5 (KMS-symmetric generator which is not GNS-symmetric). Let us consider the Poisson type generating functional on SUq(2) φ(a) = hek, (ρθ − ε)(a)eki, If (and only if) θ = π where ρθ is the infinite-dimensional representation of C(SUq(2)) on ℓ2, described in Example 11.1, θ ∈ [0, 2π), and ek is the kth standard orthonormal basis vector of ℓ2. Indeed, by Theorem 5.4 and Example 11.1 the condition for the generating functional to be KMS-symmetric, φ(a) = φ ◦ R(a), reduces to (11.8) 2 , then φ is KMS-symmetric. 2 or θ = 3π hek, (ρθ(a) − ρθ+π(a∗))eki = ε(a) − ε(a∗) 50 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA for any a ∈ A. For a = γ the left hand side of (11.8) is hek,(cid:0)ρθ(γ) − ρθ+π(γ∗)(cid:1)eki = hek, (eiθ − e−i(θ+π))qkeki = (eiθ + e−iθ)qk, which equals ε(γ)2 = 0 only when θ = π 2 . For such θ we check by a direct calculation that the equation (11.8) holds true for each element of the form a = (α∗)lγm(γ∗)n (such elements form a linear basis of A). The Poisson generator related to ρθ with θ ∈ { π since S(γ) = −qγ and q 6= 1 imply φ(γ) 6= φ ◦ S(γ). 2 } is not GNS-symmetric 2 or θ = 3π For k = 0 and θ = π 2 we can calculate explicitly the values of the generating functional (11.8). Namely, using the explicit formula for the coefficient of the corepresentations (cf. [PW00, (B.19)]), the Vandermonde summation formula (cf. (0.5.9) in [KS94]) and the standard q-transformation (cf. (0.2.14) therein), we get 2 , 3π φ(u(s) jk ) = −1, i2sq(s−j)(s+j+1) − δ0,j, i−2sq(s−j)(s+j+1)−2j, 0, j = k j ≥ 0, k = −j j < 0, k = −j otherwise. In particular, it has non-zero entries only on the diagonal and the anti-diagonal. 2 − ε)(a)e0i are: The eigenvalues of the matrix φ(s) = [φ(u(s) jk )], φ(a) = he0, (ρ π λ−j = −(1 − q(s−j)(s+j+1)+j) λ+ j = −(1 + q(s−j)(s+j+1)+j), 2 , . . . , k 2 when s = k 2 , 3 2 (k ∈ 2N + 1) or λ+ j = −(1+q(s−j)(s+j+1)+j), with j = 1 λ0 = −1+(−1)sqs(s+1), for j = 1, 3, . . . , k when s = k (k ∈ N). 11.4. ad-invariance. We already noted that ad-invariance is a strong constraint on the functional. Namely, it is necessarily a multiple of the identity on each of the corepresentation matrices, in particular of diagonal form. A comparison of this notion with that of GNS-symmetry (suggested by Proposition 11.2), shows that ad-invariant generating functionals of a L´evy process on C(SUq(2)) are necessarily GNS-symmetric. λ−j = −(1−q(s−j)(s+j+1)+j) Corollary 11.6. Let φ be a functional on SUq(2). hermitian, then φ is GNS-symmetric. If φ is ad-invariant and Proof. By Proposition 6.9, φ is of the form φ(u(s) cs = φ(u(s) 11.2 that φ is GNS-symmetric. jk ) = csδjk. By hermiticity, −j,−j) = ¯cs and we conclude by Proposition jj ) = φ(cid:0)(u(s) −j,−j)∗(cid:1) = φ(u(s) (cid:3) The following example shows that the map φ → φad preserves neither hermitic- ity nor positivity. SYMMETRIES 51 Example 11.7. Let φ be the functional on SUq(2) defined by φ(α) = eit, φ(α∗) = e−it and zero otherwise, where t 6∈ 2πZ. Then φad is ad-invariant and φad(α) = φad(α∗) = (1 − q2)−1(eit + q2e−it), so it is not hermitian. Acknowledgements U.F. and A.K. would like to thank the Politecnico di Milano for the hospitality during their visits. They also thank the GDRE GREFI GENCO for the financial support for these visits. The research was supported by the MIUR project PRIN 2009 N. 2009KNZ5FK-004 "Equazioni alle derivate parziali degeneri o singolari: metodi metrici". The research was mainly conducted during A.K.'s post-doctoral stay in Labora- toire de Math´ematiques de Besancon, supported by the Region Franche-Comt´e. A.K. was also partially supported by the Polish National Research Center (NCN) post-doctoral internship no. DEC-2012/04/S/ST1/00102. U.F. was supported by the ANR Project OSQPI (ANR-11-BS01-0008). We are also indebted to Makoto Yamashita for bringing the reference [Voi11, Lemma 4.1] to our attention. We wish to thank the referee for his/her scrupulous work. References [App10] D. Applebaum, Infinitely Divisible Central Probability Measures on Compact Lie Groups - Regularity, Semigroups and Transition Kernels. Ann. Probab. 39 (2011), no. 6, 24742496. [Ban96] T. Banica. Th´eorie des repr´esentations du groupe quantique compact libre O(n). C. R. Acad. Sci. Paris S´er. I Math., 322:241 -- 244, 1996. [BMT01] E. B´edos, G. J. Murphy, and L. Tuset, Co-amenability of compact quantum groups. J. Geom. Phys., 40(2):130 -- 153, 2001. [BMT03] E. B´edos, G. J. Murphy, and L. Tuset. Amenability and co-amenability of algebraic quantum groups. II. J. Funct. Anal., 201(2):303 -- 340, 2003. [BR97] O. Bratteli and D.W. Robinson, Operator algebras and quantum statistical mechanics. 2: Equilibrium states. Models in quantum statistical mechanics. 2nd ed. Texts and Monographs in Physics. Springer, Berlin, 1997. [Bra11] M. Brannan. Approximation properties for free orthogonal and free unitary quantum groups. Journal fur die reine und angewandte Mathematik 672, 223 -- 251, 2012. [CCJJV01] P.-A. Cherix, M. Cowling, P. Jolissant, P. Julg, A. Valette, Groups with the Haagerup Property Progress in Mathematics, vol. 197, Birkhauser Verlag, Basel- Boston-Berlin, 2001. [CP03a] P. S. Chakraborty, A. Pal, Equivariant spectral triples on the quantum SU (2) group. K-Theory 28 (2003), no. 2, 107-126. [CP03b] P. S. Chakraborty, A. Pal, Spectral triples and associated Connes-de Rham complex for the quantum SU (2) and the quantum sphere. Comm. Math. Phys. 240 (2003), no. 3, 447-456. [CGS96] F. Cipriani, D. Guido, S. Scarlatti, A remark on trace properties of K-cycles. J. [Cip97] Operator Theory 35 (1996), no. 1, 179-189. F. Cipriani, Dirichlet forms and Markovian semigroups on standard forms of von Neumann algebras. J. Funct. Anal., 147(2):259 -- 300, 1997. 52 [Cip98] [Cip08] [CS03] FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA F. Cipriani, The variational approach to the Dirichlet problem in C∗-algebras. Banach Center Publ., 43:135 -- 146, Polish Acad. Sci. Warsaw, 1998. F. Cipriani, Dirichlet forms on noncommutative spaces. In Quantum potential theory, Lecture Notes in Math., vol. 1954, 161 -- 276. Springer, Berlin, 2008. F. Cipriani and J.-L. Sauvageot, Derivations as square roots of Dirichlet forms. J. Funct. Anal., 201(1):78 -- 120, 2003. [CGIS12] F. Cipriani, D. Guido, T. Isola, J.L. Sauvageot Spectral triples for the Sierpi´nski gasket. arXiv:1112.6401, 40 pages. [Co94] A. Connes, Noncommutative geometry. Academic Press, Inc., San Diego, CA, 1994. [Co04a] A. Connes, Cyclic cohomology, noncommutative geometry and quantum group sym- metries. Noncommutative geometry, 1-71, Lecture Notes in Math., 1831, Springer, Berlin, 2004. [Co04b] A. Connes, Cyclic cohomology, quantum group symmetries and the local index for- mula for SUq(2). J. Inst. Math. Jussieu 3 (2004), no. 1, 17-68. [DLSSV05] L. D¸abrowski, G. Landi, A. Sitarz, W. van Suijlekom, J. C. V´arilly, The Dirac operator on SUq(2). Comm. Math. Phys. 259 (2005), no. 3, 729-759. [Fra06] U. Franz, L´evy processes on quantum groups and dual groups. In: Quantum indepen- dent increment processes. II, Lecture Notes in Math., vol. 1866, 181-257. Springer, Berlin, 2006. U. Franz and A. Skalski, On ergodic properties of convolution operators associated with compact quantum groups. Colloq. Math., 113(1):13 -- 23, 2008. S. Goldstein and J.M. Lindsay, KMS-symmetric Markov semigroups. Math. Z., 219(4): 591 -- 608, 1995. [GL95] [FS08] [Gro55] A. Grothendieck, Produits Tensoriel Topologiques et Espaces Nucleaires. Memoires of the American mathematical Society. A.M.S. ed., Providence, Rhodes Island, 1955. [Hun56] G.A. Hunt. Semi-groups of measures on Lie groups. Trans. Amer. Math. Soc., 81: 264 -- 293, 1956. [Koe91] H.T. Koelink, On ∗-representations of the Hopf ∗-algebra associated with the quan- tum group Uq(n). Compositio Math., 77(2): 199 -- 231, 1991. R. Koekoek and R.F. Swarttouw, The Askey-scheme of hypergeometric orthogonal [KS94] polynomials and its q-analogue. Technical Report 94-05, Technical University of Delft, 1994. Available from math.CA/9602214. [Li04] [KS97] [KK89] H.T. Koelink and T.H. Koornwinder, The Clebsch-Gordan coefficients for the quan- tum group SµU(2) and q-Hahn polynomials. Nederl. Akad. Wetensch. Indag. Math. 51 (1989), no. 4, 443456. A. Klimyk and K. Schmudgen, Quantum groups and their representations. Texts and Monographs in Physics. Springer-Verlag, Berlin, 1997. M. Liao, L´evy processes in Lie groups. Cambridge Tracts in Mathematics, vol. 162. Cambridge University Press, Cambridge, 2004. J.M. Lindsay and A.G. Skalski, Convolution semigroups of states. Math. Z., 267(1-2) (2011), 325-339. J.M. Lindsay, A. Skalski, Quantum stochastic convolution cocycles III. Math. Ann. 352 (2012), no. 4, 779804. [LS11] [LS12] [Maj95] S. Majid, Foundations of quantum group theory. Cambridge University Press, 1995. [Ped79] G.K. Pedersen. C ∗-algebras and their automorphism groups. London Mathematical Society Monographs, 14. Academic Press, Inc. [Harcourt Brace Jovanovich, Publish- ers], London-New York, 1979. [PW90] P. Podle´s and S.L. Woronowicz, Quantum deformation of Lorentz group. Commun. Math. Phys., 130:381 -- 431, 1990. SYMMETRIES 53 [PW00] W. Pusz and S.L. Woronowicz, Representations of quantum Lorentz group on Gelfand [Ry02] spaces. Rev. Math. Phys. 12 (2000), no. 12, 15511625. R.A. Ryan, Introduction to Tensor Products of Banach Spaces. Series: Springer Monographs in Mathematics, New York: Springer 2002. [Sat99] K.-I. Sato. L´evy processes and infinitely divisible distributions. Cambridge University Press, Cambridge, 1999. Translated from the 1990 Japanese original, Revised by the author. [Sch93] M. Schurmann, White Noise on Bialgebras, volume 1544 of Lecture Notes in Math. Springer-Verlag, Berlin, 1993. [Ske99] M. Skeide. Hunt's formula for SUq(2) -- a unified view. Open Syst. Inf. Dyn., 6(1):1 -- 27, 1999. [SS98] M. Schurmann and M. Skeide, Infinitesimal generators of the quantum group SUq(2). Inf. Dim. Anal., Quantum Prob. and Rel. Topics, 1(4):573 -- 598, 1998. [Tim08] T. Timmermann, An invitation to Quantum Groups and Duality. EMS Textbooks in Mathematics, 2008. [VD97] A. Van Daele, The Haar measure on finite quantum groups. Proc. Amer. Math. Soc., 125(12):3489 -- 3500, 1997. [VD98] A. Van Daele, An algebraic framework for group duality. Adv. Math., 140(2):323 -- 366, 1998. [VD03] A. Van Daele, Multiplier Hopf ∗-algebras with positive integrals: a laboratory for locally compact quantum groups. Vainerman, Leonid (ed.), Locally compact quan- tum groups and groupoids. Proceedings of the 69th meeting of theoretical physicists and mathematicians, Strasbourg, France, February 21 -- 23, 2002. Berlin: Walter de Gruyter. IRMA Lect. Math. Theor. Phys. 2, 229-247, 2003. [VD07] A. Van Daele, The Fourier transform in quantum group theory. In New techniques in Hopf algebras and graded ring theory, 187 -- 196. K. Vlaam. Acad. Belgie Wet. Kunsten (KVAB), Brussels, 2007. [VDW96] A. Van Daele and Shuzhou Wang. Universal quantum groups. Internat. J. Math., 7(2):255 -- 263, 1996. [Voi11] C. Voigt. The Baum-Connes conjecture for free orthogonal quantum groups. Adv. Math., 227(5):1873 -- 1913, 2011. [Wor87a] S.L. Woronowicz, Compact matrix pseudogroups. Commun. Math. Phys., 111:613 -- 665, 1987. [Wor87b] S.L. Woronowicz, Twisted SU(2) group. An example of a noncommutative differential calculus. Publ. Res. Inst. Math. Sci., 23(1):117 -- 181, 1987. [Wor88] S.L. Woronowicz, Tannaka-Kreın duality for compact matrix pseudogroups. Twisted SU(N ) groups. Invent. Math., 93(1):35 -- 76, 1988. [Wor98] S.L. Woronowicz, Compact quantum groups. In A. Connes, K. Gawedzki, and J. Zinn- Justin, editors, Sym´etries Quantiques, Les Houches, Session LXIV, 1995, pages 845 -- 884. Elsevier Science, 1998. S. Zakrzewski. Matrix pseudogroups associated with anti-commutative plane. Lett. Math. Phys., 21(4):309 -- 321, 1991. [Zak91] 54 FABIO CIPRIANI, UWE FRANZ, AND ANNA KULA FC: Dipartimento di Matematica, Politecnico di Milano, Piazza Leonardo da Vinci 32, 20133 Milan, Italy E-mail address: [email protected] UF: D´epartement de math´ematiques de Besanc¸on, Universit´e de Franche-Comt´e 16, route de Gray, 25 030 Besanc¸on cedex, France E-mail address: [email protected] URL: http://lmb.univ-fcomte.fr/uwe-franz AK: Instytut Matematyczny, Uniwersytet Wroc lawski, pl. Grunwaldzki 2/4, 50-384 Wrocaw, Poland and Instytut Matematyki, Jagiellonian University, ul. Lojasiewicza 6, 30 348 Krak´ow, Poland E-mail address: [email protected]
1010.5872	1	1010	2010-10-28T07:05:33	$\zeta-$function and heat kernel formulae	[ "math.OA" ]	We present a systematic study of asymptotic behavior of (generalised) $\zeta-$functions and heat kernels used in noncommutative geometry and clarify their connections with Dixmier traces. We strengthen and complete a number of results from the recent literature and answer (in the affirmative) the question raised by M. Benameur and T. Fack \cite{BF}.	math.OA	math	ζ−function and heat kernel formulae Fedor Sukocheva,∗, Dmitrii Zaninb,1 aSchool of Mathematics and Statistics, University of New South Wales, Sydney, 2052, bSchool of Computer Science, Engineering and Mathematics, Flinders University, Bedford Australia. Park, 5042, Australia. Abstract We present a systematic study of asymptotic behavior of (generalised) ζ−func- tions and heat kernels used in noncommutative geometry and clarify their con- nections with Dixmier traces. We strengthen and complete a number of results from the recent literature and answer (in the affirmative) the question raised by M. Benameur and T. Fack [1]. Keywords: Zeta function, Heat kernel formulae, Dixmier trace, Noncommutative geometry. 2000 MSC: Primary: 58B34, 46L51, 46L52, 58J42 1. Introduction The interplay between Dixmier traces, ζ−functions and heat kernel formulae is a cornerstone of noncommutative geometry [8]. These formulae are widely used in physical applications. To define these objects, let us fix a Hilbert space H and let B(H) be the algebra of all bounded operators on H with its standard trace Tr. Let A and B be positive operators from B(H). Consider the following [0, ∞]-valued functions t → 1 t Tr(A1+1/t), t → and, for fixed 0 < q < ∞ t → 1 t Tr(exp(−(tA)−q)), t → 1 t 1 t Tr(A1+1/tB)) Tr(exp(−(tA)−q)B). (1) (2) When these functions are finitely valued, they are frequently referred to as ζ−functions and heat kernel functions associated with the operators A and B. ∗Corresponding Author Email addresses: [email protected] (Fedor Sukochev), [email protected] (Dmitrii Zanin ) 1Research supported by the Australian Research Council Preprint submitted to Elsevier May 31, 2021 When these functions are bounded, a particular interest is attached to their asymptotic behavior when t → ∞, which is usually measured with the help of some generalised limit γ : L∞(0, ∞) → R yielding the following functionals ζγ(A) := γ( 1 t Tr(A1+1/t)), ζγ,B(A) := γ( Tr(A1+1/tB)) (3) 1 t and, ϕγ(A) := γ( 1 t Tr(exp(−(tA)−q))), ϕγ,B(A) := γ( Tr(exp(−(tA)−q))B). (4) 1 t A natural class of operators for which the formulae (1) and (3) are well defined (respectively, (2) and (4)) is given by the set M1,∞ (respectively, L1,∞) of compact operators from B(H). More precisely, denote by µn(T ), n ∈ N, the singular values of a compact operator T (the singular values are the eigenvalues of the operator T = (T ∗T )1/2 arranged with multiplicity in decreasing order, ([23, §1]). Then M1,∞ := M1,∞(H) = {T : sup n∈N 1 log(n + 1) n Xk=1 µk(T ) < ∞}. (5) defines a Banach ideal of compact operators. We set L1,∞ := {T ∈ M1,∞ : ∃C > 0 such that µn(A) ≤ C/n, n ≥ 1}. It is important to observe that the subset L1,∞ is not dense in M1,∞ (see e.g. [18]). It should also be pointed out that our notation here differs from that used in [8]. It follows from [6, Theorem 4.5] that the functions defined in (1) are bounded if and only if A ∈ M1,∞. It also follows from [6] and [4] that the functions defined in (2) are bounded if and only if A ∈ L1,∞. In fact the last result is a strong motivation to consider the following modification of formulae (2). Let us consider a Cesaro operator on L∞(0, ∞) given by (M x)(t) = 1 log(t)Z t 1 x(s) ds s , t ∈ (0, ∞). It follows from [6] and [4] that the functions M (t → 1 t Tr(exp(−(tA)−q))), M (t → 1 t Tr(exp(−(tA)−q))B) (6) are bounded if and only if A ∈ M1,∞. Therefore, for a given generalised limit ω, let us set ω′ := ω ◦ M (7) and instead of the functions given in (4) consider the functions ξω(A) := ω′( 1 t Tr(exp(−(tA)−q))), ξω,B(A) := ω′( 1 t Tr(exp(−(tA)−q))B). (8) 2 The class of dilation invariant states ω′ as above was introduced by A. Connes (see [8]) and it is natural to refer to this class as "Connes states". We prove in section 5 that if ω in (7) is dilation invariant, then ξω is a linear functional on M1,∞. In fact, we also show in Proposition 18 that if ω in (7) is such that ξω is linear on M1,∞, then necessarily there exists a dilation invariant generalised limit ω0 such that ξω = ξω0 . There is a deep reason to require that the functionals ξω and ζγ be defined on M1,∞ and be linear (and thus, by implication, to consider Connes states). Important formulae in noncommutative geometry [8] and its semifinite coun- terpart [5, 7, 1, 6, 4] then connect these functionals with Dixmier traces on M1,∞. Recall that in [9], J. Dixmier constructed a non-normal semifinite trace (a Dixmier trace) on B(H) using the weight Trω(T ) := ω ( 1 log(1 + n) µk(T ))∞ n=1! T > 0, n Xk=1 (9) where ω is a dilation invariant state on L∞(0, ∞). The interplay between positive functionals Trω, ζγ and ξω on M1,∞ makes an important chapter in noncommutative geometry and has been treated (among many other papers) in [8, 5, 7, 1, 6, 22, 4, 24]. We now list a few most important known results concerning this interplay and explain our contribution to this topic. In [5], the equality Trω(AB) = (ω ◦ log)( 1 t τ (A1+1/tB)) = ζω◦log,B(A), 0 ≤ A ∈ M1,∞ (10) was established for every B ∈ B(H) under very restrictive conditions on ω. These conditions are dilation invariance for both ω and ω◦log and M −invariance of ω. In [6], for the special case B = 1, the assumption that ω is M −invariant has been removed. However, the case of an arbitrary B appears to be inaccessible by the methods in that article. In Section 4, we prove the general result which im- plies, in particular, that the equality (10) holds without requiring M −invariance of ω. In [5], the equality ω( 1 t τ (exp(−(tA)−q)B)) = Γ(1 + 1 q )τω(AB) (11) was established under the same conditions on ω and ω ◦ log as above. In [24], in the special case B = 1 the equality (11) was established under the assumption that ω is M −invariant. However, again the case of an arbitrary B appears to be inaccessible by the methods in that article. Here, we are able to treat the case of a general operator B. In [1] a more general approach to the heat kernel formulae is suggested. It consists of replacing the function t → exp(t−q) with an arbitrary function f 3 from the Schwartz class. The following equality was proved in [1] ω( 1 t τ (f (tA)B)) =Z ∞ 0 f ( 1 s )ds · τω(AB) (12) for A ∈ L1,∞ and M −invariant ω. In [1, p.51], M. Benameur and T. Fack have asked whether the result above continues to stand without the M −invariance assumption on ω. In Theorem 49 below, we answer this question affirmatively for a much larger class of functions than the Schwartz class and for any A ∈ M1,∞. Finally, it is important to emphasize the connection between our results with the theory of fully symmetric functionals. Recall that a linear positive functional ϕ : M1,∞ → C is called fully symmetric if ϕ(B) ≤ ϕ(A) for every positive A, B ∈ M1,∞ such that B ≺≺ A. The latter symbol means that µk(B) ≤ n Xk=1 n Xk=1 µk(A), ∀n ∈ N. It is obvious that every Dixmier trace Trω is a fully symmetric functional. However, the fact that every fully symmetric functional coincides with a Dixmier trace is far from being trivial (see [19] and Theorem 1 below). It is therefore quite natural to ask whether a similar result holds for the sets of all linear positive functionals on M1,∞ formed by the ξω and ζγ respectively. To this end, we establish results somewhat similar to those of [19]. Firstly, in Theorem 22 we prove that if ω in (7) is dilation invariant, then the functional ξω extends to a fully symmetric functional on M1,∞. Secondly, in Theorem 31 we show that in fact every normalized fully symmetric functional on M1,∞ coincides with some ξω, where ω is dilation invariant. Thus, in view of [19], we can conclude that the set {Trω : ω is a dilation invariant generalised limit} coincides with the set {ξω : ω is a dilation invariant generalised limit} (up to a norming constant). At the same time, a natural question, namely, whether the equality ξω = Γ(1 + 1 q )Trω holds for every dilation invariant generalised limit ω is answered in the negative in Theorem 37. Finally, we note that the question on the relationship between the sets {Trω : ω is a dilation invariant generalised limit}, {ζγ : γ is a generalised limit} and {ζω : ω is a dilation invariant generalised limit} remains open. 2. Definitions and notations The theory of singular traces on operator ideals rests on some classical anal- ysis which we now review for completeness. As usual, L∞(0, ∞) is the set of all bounded Lebesgue measurable func- tions on the semi-axis equipped with the uniform norm k · k. Given a function 4 x ∈ L∞(0, ∞), one defines its decreasing rearrangement µ(x) = µ(·, x) by the formula (see e.g. [17]) µ(t, x) = inf{s ≥ 0 : m({x > s}) ≤ t}. Let H be a Hilbert space and let B(H) be the algebra of all bounded opera- tors on H equipped with the uniform norm k · k. Let N ⊂ B(H) be a semi-finite von Neumann algebra with a fixed faithful and normal semi-finite trace τ. For every A ∈ N , the generalised singular value function µ(A) = µ(·, A) is defined by the formula (see e.g. [14]) µ(t, A) := inf{kApk : τ (1 − p) ≤ t}. If, in particular, N = B(H), then µ(A) is a step function and, therefore, can be identified with the sequence {µ(n, A)}n≥0 of singular numbers of the operators A (the singular values are the eigenvalues of the operator A = (A∗A)1/2 arranged with multiplicity in decreasing order). Equivalently, µ(A) can be defined in terms of the distribution function dA of A. That is, setting dA(s) := τ (eA(s, ∞)), s ≥ 0, we obtain µ(t, A) = inf{s : dA(s) ≥ t}, t > 0. Here, eA denotes the spectral measure of the operator A. The following formula follows directly from the von Neumann definition of trace (see the definition at [20, Definition 15.1.1]) τ (f (A)) = −Z ∞ 0 f (λ)ddA(λ). (13) Using the Jordan decomposition, every operator A ∈ B(H) can be uniquely written as A = (ℜ(A)+ − ℜ(A)−) + i(ℑ(A)+ − ℑ(A)−). 2 (A + A∗) (respectively, ℑ(A) := 1 Here, ℜ(A) := 1 2i (A − A∗)) for any operator A ∈ B(H) and B+ = BeB(0, ∞) (respectively, B− = BeB(−∞, 0)) for any self-adjoint operator B ∈ B(H). Recall that ℜA, ℑA ∈ N for every A ∈ N and B+, B− ∈ N for every self-adjoint B ∈ N . Let ψ : R+ → R+ be an increasing concave function such that ψ(t) = O(t) [17]) consists of all as t → 0. The Marcinkiewicz function space Mψ (see e.g. x ∈ L∞(0, ∞) satisfying kxkMψ := sup t>0 1 ψ(t)Z t 0 µ(s, x)ds < ∞. The Marcinkiewicz operator space Mψ := Mψ(N , τ ) (see e.g. of all A ∈ N satisfying [7, 6]) consists kAkMψ := sup t>0 1 ψ(t)Z t 0 5 µ(s, A)ds < ∞. We are especially interested in Marcinkiewicz spaces M1,∞ and M1,∞ that arise when ψ(t) = log(1 + t), t ≥ 0. In the literature, the ideal M1,∞ is sometimes referred to as the Dixmier ideal. We recommend the recent paper of A. Pietsch, [21], discussing the origin of M1,∞ in mathematics. For s > 0, dilation operators σs : L∞ → L∞ are defined by the formula (σsx)(t) = x(t/s). Clearly, σs : M1,∞ → M1,∞ (see also [17, Theorem II.4.4]). Further, we need to recall the important notion of Hardy-Littlewood ma- jorization. Let A, B ∈ N . B is said to be majorized by A and written B ≺≺ A if and only if Z t 0 µ(s, B)ds ≤Z t 0 µ(s, A)ds, t ≥ 0. We have (see [14]) (14) (15) A + B ≺≺ µ(A) + µ(B) ≺≺ 2σ1/2µ(A + B). One of the most widely used ideals in von Neumann algebras is Lp := Lp(N , τ ) = {A ∈ N : kAkp := τ (Ap)1/p < ∞}, p ≥ 1, usually called the Schatten-von Neumann ideal of p-summable operators. Using Hardy-Littlewood majorization, it is very easy to see (e.g. [5, Lemma 2.1]) that M1,∞ ⊂ Lp for all p > 1. A linear functional ϕ : M1,∞ → C is said to be symmetric if ϕ(B) = ϕ(A) for every positive A, B ∈ M1,∞ such that µ(B) = µ(A). A linear functional ϕ : M1,∞ → C is said to be fully symmetric if ϕ(B) ≤ ϕ(A) for all A, B ∈ M+ 1,∞ such that B ≺≺ A [10, 11, 12]. Every fully symmetric functional is symmetric and bounded. The converse fails [18]. A positive normalised linear functional γ : L∞(0, ∞) → R is called a gen- eralised limit if γ(z) = 0 for every z ∈ L∞(0, ∞) such that limt→∞ z(t) = 0. A linear functional γ : L∞(0, ∞) → R is called dilation invariant if γ(σsz) = γ(z) for every z ∈ L∞(0, ∞) and every s > 0. Let S ⊆ B(H). We denote by S+ the set of all positive operators from S. Let ω : L∞(0, ∞) → R be a dilation invariant generalised limit. Define a functional τω on M+ 1,∞ by the formula τω(A) = ω( 1 log(1 + t)Z t 0 µ(s, A)ds). The functional τω is additive and unitarily invariant on M+ 1,∞. Thus, τω extends to a fully symmetric functional on M1,∞. One usually refers to it as to a Dixmier trace. We refer the reader to [9, 8, 5, 7, 6, 19] for details. Further, we use the following properties of Dixmier traces. Let A ∈ M1,∞ and let B ∈ N . We have (see [8, 5]) Suppose that B > 0. It follows from (16) that τω(AB) = τω(BA). τω(AB) = τω(B1/2AB1/2). 6 (16) (17) Suppose that the trace τ on the von Neumann algebra N is infinite and the algebra N is either diffuse (that is with no minimal projections) or else is B(H). Given any finite sequence {An} of operators, we can construct a sequence of operators {Bn} such that µ(An) = µ(Bn) for all n's and BnBm = 0 for all n 6= m. Further, we refer to any such sequence {Bn} as a "sequence of disjoint copies of {An}". Cesaro operator M is defined on L∞(0, ∞) by the formula (M x)(t) = 1 log(t)Z t 1 x(s) ds s , t ∈ (0, ∞). 3. Preliminary important results In this section, for the reader's convenience, we collect a number of key known results, which will be used throughout this paper. The following important theorem is proved in [19, Theorem 11] for general Marcinkiewicz spaces. Theorem 1. Every fully symmetric functional on M1,∞ is a Dixmier trace. The following theorem is an analog of Lidskii formula (see [23]) for Dixmier traces. It is proved in [24, Theorem 33] for a large subclass of Marcinkiewicz spaces which contains M1,∞. Theorem 2. Let A ∈ M1,∞ and let τω be an arbitrary Dixmier trace on M1,∞. We have The following ω-variant of the classical Karamata theorem is established in [5]. Theorem 3. Let β be a continuous increasing function. Set We have h(t) =Z ∞ 0 e−(u/t)q dβ(u). ω( h(t) t ) = Γ(1 + 1 q )ω( β(t) t ) for any dilation invariant generalised limit ω. Consider the ideal KN of τ -compact operators in N (that is the norm closed ideal generated by the projections E ∈ N with τ (E) < ∞). The following result is not new (see [15, Chapter II, Lemma 3.4]). We present a short proof for convenience of the reader. 7 τω(A) = ω  1 log(t) Xλ>log(t)/t,λ∈σ(A) λ  . Theorem 4. Let A, B ∈ N be positive τ −compact operators. We have B ≺≺ A if and only if τ ((B − t)eB(t, ∞)) ≤ τ ((A − t)eA(t, ∞)), ∀t > 0. (18) Proof. Fix t > 0. It follows from the definition of generalised singular value function that µ(AeA(t, ∞)) = µ(A)χ[0,dA(t)]. Applying [14, Proposition 2.7] to the operator AeA(t, ∞), we have τ (AeA(t, ∞)) =Z dA(t) τ ((A − t)eA(t, ∞)) =Z dA(t) 0 0 µ(s, A)ds, (µ(s, A) − t)ds. (19) and hence The function 0 attains its maximum at u = dA(t). u →Z u (µ(s, A) − t)ds If B ≺≺ A, then Z dB (t) 0 (µ(s, B) − t)ds ≤Z dB (t) 0 (µ(s, A) − t)ds ≤Z dA(t) 0 (µ(s, A) − t)ds. Inequality (18) follows now from (19). Suppose now that (18) holds. Fix u > 0 and set t = µ(u, A). It follows that Z u 0 (µ(s, B) − t)ds ≤Z dB(t) 0 (µ(s, B) − t)ds = τ ((B − t)eB(t, ∞)) ≤ Hence, ≤ τ ((A − t)eA(t, ∞)) =Z u µ(s, B)ds ≤Z u Z u 0 0 0 (µ(s, A) − t)ds. µ(s, A)ds. Since u is arbitrary, we have B ≺≺ A. 4. ζ−function formulae We begin by showing that the functionals given in (3) are well defined on 1,∞. M+ Lemma 5. If γ : L∞(0, ∞) → R is a generalised limit, then ζγ(A) < ∞ and ζγ,B(A) < ∞ for any A ∈ M+ 1,∞. 8 Proof. It is clear that µ(s, A) ≺≺ (1 + s)−1kAk1,∞. Therefore, τ (A1+1/t) ≤ kAk1+1/t 1,∞ Z ∞ 0 dt (1 + s)1+1/t = tkAk1+1/t 1,∞ . Hence, ζγ(A) ≤ kAk1,∞. It follows from τ (A1+1/tB) ≤ kBkτ (A1+1/t) that ζγ,B(A) ≤ kBkζγ(A). Remark 6. Let x, y ∈ L∞(0, ∞). For any generalised limit γ such that γ(x − 1) = 0, we have γ(xy) = γ(y). Indeed, γ(xy − y) ≤ γ(x − 1)kyk = 0. Lemma 7. For any A, C ∈ M+ 1,∞ we have τ (A1+s + C1+s) ≤ τ ((A + C)1+s) ≤ 2sτ (A1+s + C1+s), s > 0. Proof. In the special case when N = B(H), the first inequality can be found in [16, (2.9)]. In the general case, it follows directly from Proposition 4.6(ii) of [14] when f (u) = u1+s, u > 0. The second inequality follows from the same proposition by setting there a = a∗ = b = b∗ = 2−1/2. Let A ∈ M1,∞. For a functional ζγ defined on M+ 1,∞ by (3) (see Lemma 5), we set ζγ(A) := (ζγ (ℜ(A)+) − ζγ(ℜ(A)−)) + i(ζγ(ℑ(A)+) − ζγ(ℑ(A)−)). (20) The following theorem shows that functionals ζγ defined by (20) are fully symmetric on M1,∞. Theorem 8. If γ : L∞(0, ∞) → R is a generalised limit, then ζγ is a fully symmetric linear functional on M1,∞. Proof. To verify that ζγ is linear, it is sufficient to check that ζγ(A + C) = ζγ(A)+ ζγ(C) for any A, C ∈ M+ 1,∞. It follows from the left hand side inequality of Lemma 7 that ζγ(A + C) ≥ ζγ(A) + ζγ(C). Noting that γ(21/t − 1) = 0, it follows from the right hand side inequality of Lemma 7 and Remark 6 that Therefore, we have ζγ(A + C) ≤ ζγ(A) + ζγ(C). ζγ(A + C) = ζγ(A) + ζγ(C). The homogeneity of ζγ follows from Remark 6. Finally, if 0 ≤ C ≺≺ A ∈ M+ then C, A ∈ L1+s and τ (C1+s) ≤ τ (A1+s). Hence, 1 and so ζγ(C) ≤ ζγ(A). t τ (C1+1/t) ≤ 1 1,∞, t τ (A1+1/t) 9 Let B ∈ N . We extend the functional ζγ,B on M1,∞, similarly to (20). Observe that ζγ,B1+B2 (A) = ζγ,B1(A) + ζγ,B2(A), B1, B2 ∈ N , A ∈ M1,∞. Lemma 9. If A ∈ M1,∞ and Bn → B in N , then ζγ,Bn (A) → ζγ,B(A). Proof. It is sufficient to prove the assertion for A ∈ M+ 1,∞. Since τ (A1+sB) − τ (A1+sBn) ≤ τ (A1+s)kB − Bnk, we obtain ζγ,B(A) − ζγ,Bn (A) ≤ ζγ(A)kB − Bnk. The following lemma follows immediately from [5, Lemma 3.3]. Lemma 10. Let A, B ∈ B+(H) and let s > 0. We have i) (B1/2AB1/2)1+s ≤ B1/2A1+sB1/2 if 0 ≤ B ≤ 1. ii) (B1/2AB1/2)1+s ≥ B1/2A1+sB1/2 if B ≥ 1. The result below significantly strengthens [5, Proposition 3.6] by removing all extra assumptions on the generalised limit γ. Proposition 11. If γ : L∞(0, ∞) → R is a generalised limit, then ζγ,B(A) = ζγ(B1/2AB1/2), ∀A ∈ M1,∞, B ∈ N +. Proof. It is sufficient to prove the assertion for A ∈ M+ 1,∞. Suppose first that there are constants 0 < m ≤ M < ∞ such that m ≤ B ≤ M. Applying Lemma 10 to the operators A and M −1B (respectively, m−1B), we have msB1/2A1+sB1/2 ≤ (B1/2AB1/2)1+s ≤ M sB1/2A1+sB1/2. Therefore, 1 t m1/tτ (A1+1/tB) ≤ 1 t τ ((B1/2AB1/2)1+1/t) ≤ 1 t M 1/tτ (A1+1/tB). Since γ(m1/t − 1) = 0 and γ(M 1/t − 1) = 0, it follows from Remark 6 that ζγ,B(A) = ζγ(B1/2AB1/2). For an arbitrary B ∈ N +, we set Bn := BeB(1/n, ∞)+1/neB[0, 1/n], n ≥ 1. From the first part of the proof, we have ζγ,Bn (A) = ζγ(B1/2 n AB1/2 n ). Since B1/2 n AB1/2 n → B1/2AB1/2 in M1,∞, we have by Theorem 8 ζγ(B1/2 n AB1/2 n ) → ζγ(B1/2AB1/2). On the other hand, by Lemma 9 we have ζγ,Bn (A) → ζγ,B(A). 10 The following is our main result on the ζ−function. Theorem 12. If γ : L∞(0, ∞) → R is a generalised limit, then ζγ,B(A) = ζγ(AB), ∀A ∈ M1,∞, B ∈ N . Proof. It is sufficient to prove the assertion for B ∈ N +. By Theorems 8 and 1, we know that ζγ is a Dixmier trace on M1,∞. Hence, by (17), we have ζγ(B1/2AB1/2) = ζγ(AB). The assertion follows now from Proposition 11. Our remaining objective in this section is to provide strengthening of several formulae linking Dixmier traces and ζ-functions from [5, 6]. Lemma 13. Let A ∈ M+ therefore, continuous. 1,∞. The mapping s → s−1ζγ◦σs (A) is convex and, Proof. For all t, s > 0, we have s−1σs( 1 t τ (A1+1/t)) = 1 t τ (A1+s/t). Therefore, for every s > 0 s−1ζγ◦σs = γ( 1 t τ (A1+s/t)). Let λi > 0 and let λ1 + λ2 = 1. Since the mapping t → a1+t is convex for every a > 0, it follows from the spectral theorem that the map s → As is also convex. Therefore, for all positive real numbers s1, s2 and t, we have A1+(λ1s1+λ2s2)/t ≤ λ1A1+s1/t + λ2A1+s2/t. The assertion follows immediately. Let γ be a generalised limit on L∞(0, ∞). Below, we will formally apply the notation ζγ,B(A) introduced in (3) to some unbounded positive operators B on H. Lemma 14. Let A ∈ N be a positive τ −compact operator and let B ≥ 1 be an unbounded operator commuting with A. If (the closure of ) the product AB ∈ M1,∞ and ABn ∈ N for every n ∈ N, then ζγ(AB) = ζγ,B(A). Proof. It follows from AB = BA and B ≥ 1 that A1+sB ≤ (AB)1+s. The inequality ζγ,B(A) ≤ ζγ(AB) follows immediately. Set cn := kAB2nk, n ≥ 1 and observe that BA1/2n ≤ c1/2n n BeA[0, c−1 n ], we obtain . Setting Bn = BnA1/n = BA1/2n · A1/2neA[0, c−1 n ] ≤ (cnA)1/2neA[0, c−1 n ] ≤ 1. (21) It follows from (21) that A1+1/tBn ≥ (ABn)1+n/t(n−1). Thus, γ( 1 t τ (A1+1/tBn)) ≥ γ( 1 t τ ((ABn)1+n/t(n−1))) = n − 1 n ζγ◦σn/(n−1) (ABn). 11 Since A is τ −compact, then B − Bn is bounded operator with finite support. Due to the linearity with respect to B, we have ζγ,B(A) = ζγ,Bn (A) ≥ n − 1 n ζγ◦σn/(n−1) (ABn) = n − 1 n ζγ◦σn/(n−1)(AB). The assertion follows now from Lemma 13. The following result is mainly known (see [5, 6]). Our proof is however much simpler than the arguments used there. Theorem 15. If ω is a dilation invariant generalised limit such that the gen- eralised limit ω ◦ log is still dilation invariant, then τω = ζω◦log. Proof. It is sufficient to verify the equality τω = ζω◦log on positive operators A ∈ M+ 1,∞ such that A ≤ e−1. Define a continuously increasing function β : (0, ∞) → (0, ∞) by β(u) := −Z ∞ ue−u λddA(λ). Let h be as in Theorem 3 as applied to the above β. Define an operator B ≥ 1 by the formula A = Be−B and set C = e−B. We have h(t) =Z ∞ 0 e−u/tdβ(u) = −Z ∞ 0 e−u(1+1/t)uddA(ue−u) (13) = τ (C1+1/tB). (22) The conditions of Lemma 14 are valid for B and C. Indeed, B commutes with C, BC = A ∈ M1,∞ and Bne−B ∈ N for every n ∈ N. By Lemma 14, we have ζω◦log(A) = ζω◦log,B(C) = (ω ◦ log)( h(t) ). t By Theorem 2, we have τω(A) = ω( We can now conclude −1 log(t)Z ∞ log(t)/t λddA(λ)) = (ω ◦ log)( β(t) t ). (23) ζω◦log(A) (22) = (ω ◦ log)( h(t) t ) (Thm 3) = (ω ◦ log)( β(t) t (23) = τω(A). ) The following corollary strengthens and extends the results of [6, Theorem 4.11] and [5, Theorem 3.8]. It follows immediately from Theorems 15 and 12. Corollary 16. If ω is a dilation invariant generalised limit such that the gen- eralised limit ω ◦ log is still dilation invariant, then τω(AB) = (ω ◦ log)( 1 t τ (A1+1/tB)), ∀A ∈ M+ 1,∞, B ∈ N . 12 5. The linearity criterion for functionals ξγ In this section we focus on functionals ξγ(·) defined in (8). It was implicitly proved in [6, Theorem 5.2] that M(cid:18)t → 1 t τ (exp(−(tA)−q))(cid:19) ∈ L∞(0, ∞), ∀A ∈ M+ 1,∞ and therefore, ξγ(A) := (γ ◦ M )(cid:18)t → 1 t τ (exp(−(tA)−q))(cid:19) (24) is finite for every A ∈ M+ 1,∞ and every generalised limit γ on L∞(0, ∞). We note, in passing that a stronger result than [6, Theorem 5.2] is established in Theorem 40 below. Let A ∈ M1,∞. For a functional ξγ, we set ξγ(A) := (ξγ (ℜ(A)+) − ξγ(ℜ(A)−)) + i(ξγ(ℑ(A)+) − ξγ(ℑ(A)−)). (25) It is probably a difficult task to describe the set of all generalised limits γ for which (25) yields a linear functional ξγ. However, the class of linear func- tionals ξγ is an easier object. Below in Proposition 18, we show that the sets of linear functionals {ξγ : γ is a generalised limit} and linear function- als {ξω : ω is a dilation invariant generalised limit} coincide. Lemma 17. For every locally integrable z with M z ∈ L∞(0, ∞), we have (M ◦ σs−1 − σs−1 ◦ M )(z) ∈ Cb 0(0, ∞), ∀s > 0. Here, Cb ∞. 0(0, ∞) is the space of all bounded continuous functions tending to 0 at Proof. Fix s > 0. The assertion follows by writing (M ◦ σs−1 − σs−1 ◦ M )(z) = 1 log(t)Z st s z(u) du u − 1 log(st)Z st 1 z(u) du u and noting that the assumption M z ∈ L∞(0, ∞) easily implies that 1 log(st)Z st 1 z(u) du u − 1 log(t)Z st 1 z(u) du u ∈ C0 b (0, ∞). Proposition 18. Suppose that a generalised limit γ on L∞(0, ∞) is such that ξγ is a linear functional on M1,∞. Then, there exists a dilation invariant gen- eralised limit ω on L∞(0, ∞) such that ξγ = ξω. 13 Proof. Fix s > 0 and observe that (cid:18)t → 1 t τ (exp(−(tsA)−q))(cid:19) = sσs−1(cid:18)t → 1 t τ (exp(−(tA))−q)(cid:19) . (26) Therefore, ξγ(sA) = s(γ ◦ M ◦ σs−1 )( τ (exp(−(tA)−q))). 1 t By the assumption, we have ξγ(sA) = sξγ(A) and appealing to Lemma 17, we obtain ξγ(A) = (γ ◦ σs−1 ◦ M )( τ (exp(−(tA)−q))), ∀s > 0. (27) 1 t Let E be the linear span of the functions t → M ( 1 t τ (exp(−(tA)−q))), A ∈ M+ 1,∞ and let F := E + Cb Indeed, it follows from Lemma 17 and (26) that every function 0(0, ∞). We claim that the space F is dilation invariant. σs−1(cid:18)t → M ( 1 t τ (exp(−(tA))−q))(cid:19) belongs to the set s−1(cid:18)t → M ( 1 t τ (exp(−(tsA)−q)))(cid:19) + Cb 0(0, ∞). It follows from (27) that γ ◦ σs−1 = γ on F. By the invariant form of the Hahn- Banach theorem (see [13, p. 157]) applied to the group of dilations {σs}s>0, we see that γF can be extended to a dilation invariant generalised limit ω on L∞(0, ∞). The following lemma can be found in [24]. We present a shorter proof for convenience of the reader. Lemma 19. If ω is a dilation invariant generalised limit on L∞(0, ∞), then ξω(A) = Γ(1 + 1 q )(ω ◦ M )( 1 t dA( 1 t )), ∀A ∈ M+ 1,∞. Proof. It follows from (13) that τ (exp(−(tA)−q)) =Z ∞ 0 e−(u/t)q ddA( 1 u ). (28) (29) Setting β(u) = dA(1/u), multiplying both sides of (29) by 1/t and applying Theorem 3 to ω ◦ M (which is dilation invariant, see [8]), we obtain (28). 14 Lemma 20. Let A ∈ M+ on L∞(0, ∞). We have 1,∞ and let ω be a dilation invariant generalised limit ξω(A) = Γ(1 + 1 q )ω( 1 log(1 + t) τ ((A − 1 t )eA( 1 t , ∞))). (30) Proof. In view of Lemma 19, it is sufficient to show that right hand sides of (28) and (30) coincide. This easily follows from the following computation, where we use integration by parts M ( 1 t dA( 1 t )) = dA( 1 s ) ds s2 = 1 log(t)Z 1 1/t dA(u)du = 1 1 log(t)Z t log(t)Z 1 1/t 1 = 1 log(t) udA(u)1 1/t − uddA(u) = 1 log(t) τ ((A − 1 t )eA( 1 t , ∞)) + o(1). Lemma 21. Let ω be a dilation invariant generalised limit on L∞(0, ∞) and let A, B ∈ M+ 1,∞ be such that B ≺≺ A. We have ξω(B) ≤ ξω(A). Proof. The assertion follows from Lemma 20 and Theorem 4. The following is the main result of this section. Theorem 22. For any dilation invariant generalised limit ω on L∞(0, ∞), the functional ξω given by (25) is linear and fully symmetric on M1,∞. Proof. The assertion follows from Lemma 21 provided we have shown that ξω(A + B) = ξω(A) + ξω(B), ∀A, B ∈ M+ 1,∞. (31) To this end, we observe first that since ω and ω ◦ M are dilation invariant, it follows from Lemma 21 and (15) that ξω(A + B) = ξω(µ(A) + µ(B)), ∀A, B ∈ M+ 1,∞. Now, let C and D be disjoint copies of A and B (see Section 2). Thus, we have ξω(C + D) = ξω(µ(C) + µ(D)) = ξω(µ(A) + µ(B)) = ξω(A + B). However, the equality ξω(C + D) = ξω(C) + ξω(D) for positive operators C and D such that CD = 0 follows immediately from the definition (24). Since the equalities ξω(A) = ξω(C), ξω(B) = ξω(D) are obvious, we arrive at (31). 15 6. Every fully symmetric functional has form ξω It follows from Theorem 22 and Theorem 1, that the functional ξω is a fully symmetric functional on M1,∞ whenever ω is a dilation invariant generalised limit ω on L∞(0, ∞). In this section, we show the converse. Define a (non-linear) operator T : M+ 1,∞ → L∞(0, ∞) by the formula (T A)(t) = 1 log(1 + t) τ ((A − 1 t )eA( 1 t , ∞)), t > 0. (32) We need some properties of the operator T. Firstly, we show that it is additive on certain pairs of A, B ∈ M+ 1,∞. Lemma 23. Let A, B ∈ M+ T (A + B) = T A + T B. 1,∞ be such that AB = BA = 0. It follows that Proof. It follows immediately from the assumption that (A + B − 1 t )eA+B( 1 t , ∞) = (A − 1 t )eA( 1 t , ∞) + (B − 1 t )eB( 1 t , ∞). Next, we explain the connection of the operator T with fully symmetric functionals on M1,∞. Lemma 24. Let the operators A, B ∈ M+ fully symmetric functional ϕ on M1,∞, we have ϕ(B) ≤ ϕ(A). 1,∞ be such that T B ≤ T A. For every Proof. It follows immediately from the definition (32) that τ ((B − 1 t )eB( 1 t , ∞)) ≤ τ ((A − 1 t )eA( 1 t , ∞)), ∀t > 0. Applying Theorem 4 we obtain B ≺≺ A and so ϕ(B) ≤ ϕ(A). Lemma 25. Let A, B ∈ M+ M1,∞, we have 1,∞. For every fully symmetric functional ϕ on ϕ(B) − ϕ(A) ≤ kϕkM∗ 1,∞ lim sup (T B − T A)(t). t→∞ Proof. Without loss of generality, kϕkM∗ 1,∞ = 1. Denote the right hand side by c and suppose that c ≥ 0 (the case when c < 0 is treated similarly). Fix ε > 0. We have (T B − T A)(t) ≤ c + ε for all sufficiently large t. Let C be an operator with µ(t, C) = (c + 2ε)/(1 + t). We have T B ≤ T A + T C for all sufficiently large t. Let A1 and C1 be disjoint copies of A and C, respectively. It follows from Lemma 23 that T B(t) ≤ T (A1 + C1)(t) for all sufficiently large t. Choose 0 < δ small enough to guarantee T B1(t) ≤ T (A1 + C1)(t) for all t > 0, where B1 := min{B, δ}. By Corollary 24, we have ϕ(B1) ≤ ϕ(A1) + ϕ(C1), or equivalently ϕ(B) ≤ ϕ(A) + c + 2ε. Since ε is arbitrarily small, we are done. 16 Lemma 26. Let A1, · · · , An ∈ M+ For every fully symmetric functional ϕ on M1,∞ we have 1,∞ and let λ1, · · · , λn ∈ R for some n ≥ 1. n Xk=1 λkϕ(Ak) ≤ lim sup t→∞ n Xk=1 λk(T Ak)(t). (33) Proof. Both sides of the inequality (33) depend continuously on the λk's. With- out loss of generality, we may assume that all λk ∈ Q. Multiplying both sides by the common denominator, we may assume that all λk ∈ Z. Writing λkAk = λk Xk=1 sgn(λk)Ak we see that it is sufficient to prove (33) only for the case when λk = ±1 for every k. Let {Bk} be a disjoint copy sequence of {Ak}. Both sides of the inequality (33) do not change if we replace Ak with Bk. Without loss of generality, the operators AkAj = 0, k 6= j. By Lemma 25 we have n Xk=1 λkϕ(Ak) = ϕ(Xλk=1 (T (Xλk=1 Ak) − ϕ( Xλk=−1 Ak) − T ( Xλk=−1 t→∞ ≤ lim sup Ak) ≤ Ak))(t). Since AkAj = 0 for all k 6= j, we have by Lemma 23 that T (Xλk=1 Ak) − T ( Xλk=−1 Ak) = λkT Ak n Xk=1 and the assertion follows. Lemma 27. Let E be the linear span of T M+ we have σsE = E. 1,∞ and Cb 0(0, ∞). For every s > 0 Proof. It follows from the definition (32) that for every s > 0, we have σsT A ∈ sT (s−1A) + Cb 0(0, ∞), ∀A ∈ M+ 1,∞. (34) Let ϕ be a normalised fully symmetric functional on M1,∞. We need the following linear functional on E. Definition 28. For every z ∈ E such that z ∈ n Xk=1 λkT Ak + C∞ 0 (0, ∞) 17 we set ρ(z) = n Xk=1 λkϕ(Ak). That ρ is well-defined is proved below. Lemma 29. The linear functional ρ : E → R is well-defined. For every z ∈ E, we have Proof. Let z ∈ E be such that ρ(z) ≤ lim sup t→∞ z(t). z ∈ n Xk=1 λkT Ak + Cb 0(0, ∞), z ∈ µkT Bk + Cb 0(0, ∞). m Xk=1 We have n λkT Ak − Xk=1 It follows from Lemma 26 that µkT Bk ∈ Cb 0(0, ∞). m Xk=1 λkϕ(Ak) = n Xk=1 µkϕ(Bk), m Xk=1 so that ρ is well-defined. The second assertion directly follows from Lemma 26. Lemma 30. Let ϕ be a normalised fully symmetric functional on M1,∞. There exists a dilation invariant generalised limit ω on L∞(0, ∞) such that ϕ(A) = ω(T A) for every A ∈ M+ 1,∞. Proof. For every A ∈ M+ 1,∞, we have ρ(σsT A) (34) = ρ(sT (s−1A)) Def.28 = sϕ(s−1A) = ρ(T A). Therefore, ρ is σs−invariant on E. It follows from Lemma 29 that ρ(z) ≤ lim sup t→∞ z(t), z ∈ E. By the invariant form of the Hahn-Banach theorem (see [13, p. 157]) applied to the group of dilations {σs}s>0, we can extend ρ to a dilation invariant generalised limit on L∞(0, ∞). The following assertion is the main result of this section. It permits repre- sentation of a fully symmetric functional ϕ via heat kernel formulae. Theorem 31. Let ϕ be a fully symmetric functional on M1,∞. There exists dilation invariant generalised limit ω on L∞(0, ∞) such that ϕ = const · ξω. 18 Proof. It follows from Lemma 30 that there exists a dilation invariant gener- alised limit ω such that ϕ(A) = ω( 1 log(1 + t) τ ((A − 1 t )eA( 1 t , ∞))). The assertion follows now from Lemma 20. 7. A counterexample It is known (see [24, Theorem 33] and the more general result in Corollary 51 below) that the equality ξω(A) = Γ(1 + 1 q )τω(A), A ∈ M+ 1,∞ holds for every M −invariant generalised limit ω on L∞(0, ∞) (see also earlier results with more restrictive assumptions on ω in [5, Theorem 4.1] and [6, The- orem 5.2]). In view of Theorem 31 and Theorem 1, it is quite natural to ask whether the equality above holds for every dilation invariant generalised limit ω. In this section we prove that this is not the case. Lemma 32. Let ω be a dilation invariant generalised limit on L∞(0, ∞). For every s > 1, we have χ[eek ,seek )) = 0. χ(ek+ek /s,ek+ek ]) = 0. ω(Xk ω(Xk (35) (36) Proof. Denote the left hand side of (35) by f (s). Due to the dilation invariance of ω, we have f (s) = ω(Xk χ[teek ,steek )) = f (st) − f (t), s, t > 1. Since f is monotone and bounded, we have f = 0. Denote the left hand side of (36) by g(s). Due to the dilation invariance of ω, we have g(s) = ω(Xk χ(ek+ek /st,ek+ek /t]) = g(st) − g(t), s, t > 1. Since g is monotone and bounded, we have g = 0. Lemma 33. Let ω be a dilation invariant generalised limit on L∞(0, ∞). We have i) t log(t) e−ek ω(Xk χ[ek−1+ek−1 ,ek+ek ](t)) = 0. 19 ii) 1 t log(t) ek+ek ω(Xk χ[eek ,eek+1 ](t)) = 0. Proof. We only prove the first assertion. Proof of the second one is similar. Fix s > 1. We have t log(t) e−ek ≤ 2 s + 2e−ek/2, ∀t ≤ ek+ek /s, ∀k ≥ 1 and, therefore, t log(t) e−ek Xk χ[ek−1+ek−1 ,ek+ek ](t) ≤ 2 s +Xk χ[ek+ek /s,ek+ek ](t)+ Clearly, +2Xk ω(Xk e−ek/2χ[ek−1+ek−1 ,ek+ek ](t). e−ek/2χ[ek−1+ek−1 ,ek+ek ](t)) = 0. It follows from the Lemma 32 that t log(t) e−ek ω(Xk χ[ek−1+ek−1 ,ek+ek ](t)) ≤ 2 s . Since s is arbitrarily large, we have t log(t) e−ek ω(Xk χ[ek−1+ek−1 ,ek+ek ](t)) = 0. Lemma 34. There exists a dilation invariant generalised limit ω on L∞(0, ∞) such that ω(Xk χ[eek ,ek+ek )) = 1, ω(Xk χ[ek+ek ,eek+1 )) = 0. Proof. Define a positive, homogeneous functional π on L∞(0, ∞) by the formula π(x) = lim sup N→∞ 1 log(log(N ))Z N log(N ) N x(s) ds s . It is verified in [24, Lemma 4] that every ω ∈ L∞(0, ∞)∗ satisfying ω ≤ π is dilation invariant. Observing that χ[eek ,ek+ek )) = 1, π(Xk 20 let us select ω ∈ L∞(0, ∞)∗ satisfying ω ≤ π and such that χ[eek ,ek+ek )) = 1. ω(Xk Therefore, ω(Xk χ[ek+ek ,eek+1 )) = 1 − ω(Xk χ[eek ,ek+ek )) = 0. Define a function x by the formula x = sup k∈N e−ek χ[0,ek+ek ]. (37) Fix k ≥ 1. For every t ∈ [ek−1+ek−1 , ek+ek ], we have 1 log(1 + t)Z t 0 x(s)ds ≤ e1−kZ ek+ek 0 x(s)ds ≤ e1−k e−en · en+en ≤ e2 e − 1 , k Xn=1 which guarantees x ∈ M1,∞. Lemma 35. Let x be as in (37) and let ω be as in Lemma 34. We have τω(x) = (e − 1)−1. Proof. Fix t ∈ [ek−1+ek−1 , ek+ek ]. We have x(u)du = ek e − 1 Z t 0 + te−ek + O(1). It follows that τω(x) = (e − 1)−1ω(Xk ek log(t) χ[ek−1+ek−1 ,ek+ek ](t))+ t log(t) e−ek χ[ek−1+ek−1 ,ek+ek ](t)). +ω(Xk By Lemma 33, the second generalised limit above vanishes. We claim that the first generalised limit above is 1. Indeed, ek log(t) Xk χ[ek−1+ek−1 ,ek+ek ](t) ≥ (1 + o(1))Xk χ[eek ,ek+ek ](t) and ek log(t) Xk χ[ek−1+ek−1 ,ek+ek ](t) ≤Xk χ[eek ,ek+ek ](t) + eXk χ[ek−1+ek−1 ,eek ]. The claim follows from Lemma 34. 21 Lemma 36. Let x be as in (37) and let ω be as in Lemma 34. We have ξω(x) = e e − 1 Γ(1 + 1 q ). Proof. Fix t ∈ [eek , eek+1 ). We have Zx>1/t (x(u) − 1 t )du = ek+1 e − 1 − 1 t ek+ek + O(1). This estimate and Lemma 20 yield 1 Γ(1 + 1/q) ξω(x) = e e − 1 ω(Xk ek log(t) χ[eek ,eek+1 ](t))− 1 t log(t) ek+ek χ[eek ,eek+1 ](t)). −ω(Xk It follows from Lemma 33 that the second generalised limit is 0. We claim that the first generalised limit is 1. Indeed, ek log(t) Xk χ[eek ,eek+1 ](t) ≥ (1 + o(1))Xk χ[eek ,ek+ek ] and ek log(t) χ[eek ,eek+1 ](t) ≤ 1. Xk The claim follows from Lemma 34. The following theorem delivers the promised counterexample. Theorem 37. There exists A ∈ M1,∞ and dilation invariant generalised limit ω on L∞(0, ∞) such that Γ(1 + 1 q )τω(A) < ξω(A). Proof. For brevity, we assume that the von Neumann algebra N is of type II (the argument can be easily adjusted when N is of type I). Let x be as in (37) and let A ∈ M+ 1,∞ be such that x = µ(A). The assertion follows from Lemmas 35 and 36. 8. Correctness of the definition for generalised heat kernel formulae Let ω be a dilation invariant generalised limit on L∞(0, ∞) and let B ∈ N . Following [1], we consider the functionals on M+ 1,∞ defined by the formula ξω,B,f (A) = (ω ◦ M )(t → 1 t τ (f (tA)B)). (38) 22 The main result of this section, Theorem 40, shows that the function M(cid:18)t → 1 t τ (f (tA)B)(cid:19) is bounded, and so the formula (38) is well-defined. Lemma 38. Let A ∈ M+ ∞. 1,∞. We have τ (A2eA[0, 1/t]) = O(t−1 log(t)) as t → Proof. Let c := kAk1,∞. We have µ(s, A) ≺≺ c(1 + s)−1. Fix t > 0. Define decreasing function xt ∈ M1,∞(0, ∞) by setting xt(s) =( log(1+ct log(t)) c 1+s , t log(t) , 0 ≤ s ≤ ct log(t) s > ct log(t). Define a decreasing function yt ∈ M1,∞(0, ∞) by setting yt(s) = µ(A)χ{µ(A)≤1/t}(s) + 1 t χ{µ(A)≥1/t}(s), s > 0. We claim that yt ≺≺ xt. Indeed, yt(s) ≤ 1/t ≤ xt(s) for s ≤ ct log(t) and Z s 0 yt(u)du ≤ cZ s 0 du 1 + u =Z s 0 xt(u)du for s > ct log(t). It follows that τ (A2eA[0, 1 t ]) ≤Z ∞ 0 We have y2 t (s)ds ≤Z ∞ 0 x2 t (s)ds. Z ∞ 0 x2 t (s)ds = c log2(1 + ct log(t)) t log(t) +Z ∞ ct log(t) c2 (1 + s)2 ds ≤ 5c log(t) t . Lemma 39. Let f (t) = t2χ[0,1](t) and let A ∈ M+ 1,∞. We have t → M ( 1 t τ (f (tA))) ∈ L∞(0, ∞). Proof. For fixed t > 0, we have M ( 1 t τ (f (tA))) = 1 log(t)Z t 1 τ (A2eA[0, Integrating by parts, we obtain Z t 1 eA[0, 1 s ]ds = seA[0, 1 s ]t 1 −Z t 1 sdeA[0, 23 1 s 1 s ])ds = 1 log(t) τ (A2Z t 1 eA[0, 1 s ]ds). ] = seA[0, 1 s ]t 1 +Z 1 1/t u−1deA[ 1 t , u] = = O(1) + A−1eA[ , ∞] + teA[0, 1 t 1 t ]. Therefore, M ( 1 t τ (f (tA))) = 1 log(t) τ (AeA( 1 t , ∞)) + t log(t) τ (A2eA[0, 1 t ]) + O( 1 log(t) ). It follows from the definitions of k · k1,∞ and dA(·) that for every A ∈ M1,∞ and every t > 0, we have dA( 1 t ) ≤ max{1, kAk1,∞} log(1 + t). Clearly, 1 log(t) τ (AeA[0, 1 t ]) = 1 log(t)Z dA(1/t) 0 µ(s, A)ds ≤ log(dA(1/t)) log(t) kAk1,∞ ∈ L∞. The assertion follows now from the Lemma 38. Theorem 40. Let a bounded function f ∈ C2[0, ∞) be such that f (0) = f ′(0) = 0. Let A ∈ M+ 1,∞ and let B ∈ N . We have M(cid:18)t → 1 t τ (f (tA)B)(cid:19) ∈ L∞(0, ∞). Proof. Due to the well known inequality τ (CB) ≤ τ (C)kBk, it suffices to prove the theorem only when B = 1. In this case, for the function f (t) := t2χ[0,1](t), the assertion follows from Lemma 39. If f (t) := χ(1,∞)(t) then it holds trivially. Thus, it holds for the function f (t) := min{1, t2}. Finally, observe that the assumptions on f guarantee that there exists a constant c > 0 such that f (t) ≤ c min{1, t2}. Since the function t → exp(−t−q) satisfies the assumptions of Theorem 40 we obtain the following corollary, which was implicitly proved in [6, Theorem 5.2]. Corollary 41. For every q > 0 and every A ∈ M+ 1,∞, we have M(cid:18)t → 1 t τ (exp(−(tA)−q))(cid:19) ∈ L∞(0, ∞). 9. Reduction theorem for generalised heat kernel formulae The results of this section extend and generalise those of [5, Theorem 4.1] and [6, Theorem 5.2]. We also give an answer to the question asked in [1, page 52]. We explicitly prove that the functional ξω,B,f (extended to M1,∞ as in (25)) is linear on M1,∞. 24 Lemma 42. Let f ∈ C2[0, ∞) be such that f (0) = f ′(0) = 0. Let A ∈ M+ 1,∞ and let B ∈ N . For every dilation invariant generalised limit ω on L∞(0, ∞), we have (ω ◦ M )( lim ε→0 1 t τ (f (tAeA[0, ])B)) = 0. ε t Proof. Since f (t) ≤ const · t2 for t ∈ [0, 1], it is sufficient to prove the assertion for f (t) = t2. As in the proof of Theorem 40, it is sufficient to assume that B = 1. By Theorem 40, for every ε > 0 we have M(cid:18)t → 1 t τ ((tAeA[0, ε t ])2)(cid:19) ∈ L∞(0, ∞). Since ω is dilation invariant, we conclude (ω ◦ M )( 1 t τ ((tAeA[0, ε t ])2)) = ε(ω ◦ M )( 1 t τ ((tAeA[0, 1 t ])2)). The assertion follows immediately. Lemma 43. Let f ∈ L∞(0, ∞) be such that f (0) = 0. Let A ∈ M+ 1,∞ and let B ∈ N . For every dilation invariant generalised limit ω on L∞(0, ∞), we have (ω ◦ M )( lim ε→0 1 t τ (f (tAeA( 1 εt , ∞))B)) = 0. Proof. As before, we may assume that B = 1. It is clear that f (tAeA( 1 εt , ∞)) ≤ kf keA( 1 εt , ∞). Since ω ◦ M is dilation invariant, we obtain (ω ◦ M )( 1 t τ (eA( 1 εt , ∞))) = ε(ω ◦ M )( 1 t dA( 1 t )). The assertion follows immediately. Lemma 44. Let f : R+ → R be monotone on [a, b] and such that f (0) = 0. Let A ∈ M+ 1,∞ and let B ∈ N . For every dilation invariant generalised limit ω on L∞(0, ∞) we have (ω ◦ M )( 1 t τ (f (tAeA[ a t , b t ))B)) = (Z b a f (s) ds s2 ) · (ω ◦ M )( 1 t τ (eA[ 1 t , ∞)B)). Proof. Without loss of generality, we may assume that f is increasing on [a, b] and that B ≥ 0. Let a = a0 ≤ a1 ≤ a2 ≤ · · · ≤ an = b. For every given t > 0, we have eA[ a t , b t ) = n−1 Xk=0 eA[ ak t , ak+1 t ). 25 Since f is increasing on [a, b] and f (0) = 0, we have f (ak)eA[ ak t , ak+1 t ) ≤ f (tAeA[ ak t , ak+1 t )) ≤ f (ak+1)eA[ ak t , ak+1 t ). Therefore, (ω ◦ M )( 1 t τ (f (tAeA[ a t , b t ))B)) ≤ n−1 Xk=0 and f (ak+1)(ω ◦ M )( 1 t τ (eA[ ak t , ak+1 t )B)) (ω ◦ M )( 1 t τ (f (tAeA[ a t , b t ))B)) ≥ n−1 Xk=0 f (ak)(ω ◦ M )( 1 t τ (eA[ ak t , ak+1 t )B)). We have eA[ ak t , ak+1 t ) = eA[ ak t , ∞) − eA[ ak+1 t , ∞). For all c > 0, we have (ω ◦ M )( 1 t τ (eA( c t , ∞)B)) = c−1(ω ◦ M )( 1 t τ (eA( 1 t , ∞)B)). Therefore, (ω ◦ M )( 1 t τ (eA[ ak t , ak+1 t )B)) = ( 1 ak − 1 ak+1 )(ω ◦ M )( 1 t τ (eA( 1 t , ∞)B)). Hence, n−1 ( Xk=0 f (ak)( 1 ak − 1 ak+1 ))(ω ◦ M )( 1 t τ (eA( 1 t , ∞)B)) ≤ ≤ (ω ◦ M )( 1 t τ (f (tAeA[ a t , b t ))B)) ≤ f (ak+1)( 1 ak − 1 ak+1 ))(ω ◦ M )( 1 t τ (eA( 1 t , ∞)B)). ≤ ( n−1 Xk=0 Both coefficients in the latter formula tend to R b Lemma 45. Let a bounded function f ∈ C2[0, ∞) be such that f (0) = f ′(0) = 0. Let A ∈ M+ 1,∞ and let B ∈ N . For every dilation invariant generalised limit ω on L∞(0, ∞) we have a f (s)s−2ds. ξω,B,f (A) = (Z ∞ 0 f (s) ds s2 )(ω ◦ M )( 1 t τ (eA( 1 t , ∞)B)). Proof. Let f satisfy the assumptions above. Observe that the assertion of Lemma 44 holds for the function f [a,b], where 0 < a < b < ∞. Indeed, every such function is a function of bounded variation and therefore may be written as a difference of two monotone functions. Now the assertion follows from Lemmas 42,43,44 by setting a := ε and b := ε−1 and letting ε → 0. 26 Corollary 46. Let a bounded function f ∈ C2[0, ∞) be such that f (0) = f ′(0) = 0. Let A ∈ M+ 1,∞ and let B ∈ N +. For every dilation invariant generalised limit ω on L∞(0, ∞) we have ξω,B,f (A) = (Z ∞ 0 f (s) ds s2 )ω( 1 log(1 + t) τ ((A − 1 t )eA( 1 t , ∞)B)). Proof. It follows from the definition of Cesaro operator M that τ (eA( 1 s , ∞)B) ds s2 . τ (eA(u, ∞)B)du = udτ (eA(u, ∞)B) = M(cid:18)t → 1 t τ (eA( 1 t , ∞)B)(cid:19) = Integrating by parts, we obtain 1 log(t)Z t 1 1 log(t)Z t 1 τ (eA( 1 s , ∞)B) ds s2 = = 1 log(t) · uτ (eA(u, ∞)B)1 1/t − 1 1 1/t log(t)Z 1 log(t)Z 1 τ (Z ∞ 1/t 1/t = −1 t log(t) · τ (eA( 1 t , ∞)B) + −1 log(t) udeA(u, ∞)B) + o(1). Evidently, Therefore, −τ (Z ∞ 1/t udeA(u, ∞)B) = τ (AeA( 1 t , ∞)B). M(cid:18)t → 1 t τ (eA( 1 t , ∞)B)(cid:19) = 1 log(t) τ ((A − 1 t )eA( 1 t , ∞)B) + o(1). The assertion follows now from Lemma 45. The first assertion in lemma below can be found in [3, Theorem 11]. For the second assertion we refer to [2, Theorem 3.5]. Lemma 47. Let A, B ∈ B+(H) and let f be convex continuous function such that f (0) = 0. We have i) τ (B1/2f (A)B1/2) ≥ τ (f (B1/2AB1/2)) if B ≤ 1. ii) τ (B1/2f (A)B1/2) ≤ τ (f (B1/2AB1/2)) if B ≥ 1. We show in the following lemma that ξω,B,f depends continuously on B. Lemma 48. If A ∈ M+ 1,∞ and let Bn, B ∈ N , n ≥ 1, then kξω,Bn (A) − ξω,B(A)k ≤ ξω(A) · kBn − Bk. 27 Proof. The assertion follows from the inequality τ (f (tA)Bn) − τ (f (tA)B) ≤ τ (f (tA)) · kBn − Bk. The following theorem extends the results of [5, 6] and gives an affirmative answer to the question stated in [1]. It also shows that the functionals ξω,B,f (·) are linear functionals on M1,∞ for a wide class of functions f. Theorem 49. Let a bounded function f ∈ C2[0, ∞) be such that f (0) = f ′(0) = 0. Let A ∈ M1,∞ and let B ∈ N . For every dilation invariant generalised limit ω on L∞(0, ∞) we have ξω,B,f (A) = 1 Γ(1 + 1/q) (Z ∞ 0 f (s) ds s2 )ξω(AB). (39) Proof. It follows from Theorem 22 that ξω is linear and fully symmetric. By Theorem 1 and (17)), we have ξω(B1/2AB1/2) = ξω(AB). Recall that function u → (u − 1/t)+ is convex. It follows from Lemma 47 that i) τ ((A − 1 ii) τ ((A − 1 t )+B) ≥ τ ((B1/2AB1/2 − 1 t )+B) ≤ τ ((B1/2AB1/2 − 1 t )+) if B ≤ 1. t )+) if B ≥ 1. It follows from Corollary 46 that for 0 ≤ B ≤ 1 we have ξω,B,f (A) ≥ 1 Γ(1 + 1/q) (Z ∞ 0 f (s) ds s2 )ξω(B1/2AB1/2). Since both sides are homogeneous, the inequality (40) is valid for every B. It follows from 46 that for B ≥ 1 we have ξω,B,f (A) ≤ 1 Γ(1 + 1/q) (Z ∞ 0 f (s) ds s2 )ξω(B1/2AB1/2). (40) (41) Since both sides are homogeneous, the inequality (41) is valid if B is bounded from below by a strictly positive constant. Thus, we have the equality (39) valid for every B bounded from below by a strictly positive constant. Set Bn = BeB(1/n, ∞) + 1/neB[0, 1/n]. It follows that equality (39) holds with B replaced with Bn throughout. By Lemma 48, we have ξω,Bn,f (A) → ξω,B,f (A). Since ABn → AB in M1,∞ and since ξω is bounded on M1,∞, we have ξω(ABn) → ξω(AB). The assertion follows immediately. The following corollary treats the case of classical heat kernel formulae. We use the notation ξω,B(A) = (ω ◦ M )( τ (exp(−(tA)−q)B)). 1 t 28 Corollary 50. Let A ∈ M+ generalised limit ω on L∞(0, ∞) we have ξω,B(A) = ξω(AB). 1,∞ and let B ∈ N . For every dilation invariant Proof. Use f (t) = exp(−t−q) in Theorem 49 and observe that Z ∞ 0 f (s) ds s2 = Γ(1 + 1 q ). The following assertion extends [24, Theorem 33]. Corollary 51. Let A ∈ M+ generalised limit ω on L∞(0, ∞) such that ω = ω ◦ M, we have 1,∞ and let B ∈ N . For every dilation invariant ξω,B(A) = Γ(1 + 1 q )τω(AB). References [1] M. Benameur, T. Fack, Type II noncommutative geometry. I. Dixmier trace in von Neumann algebras, Adv.Math. 199 (2006), 29-87 [2] J. Bourin, Convexity or concavity inequalities for Hermitian operators, Math. Inequal. Appl. 7 (2004), no. 4, 607 -- 620. [3] L. Brown, H. Kosaki, Jensen's inequality in semi-finite von Neumann al- gebras, J. Operator Theory 23 (1990), no. 1, 3 -- 19. [4] A. L. Carey, V. Gayral, A. Rennie, F. Sukochev, Integration on locally compact noncommutative spaces, arXiv:0912.2817v1. [5] A. Carey, J. Phillips, F. Sukochev, Spectral flow and Dixmier traces, Adv.Math. 173 (2003), no. 1, 68 -- 113. [6] A. Carey, A. Rennie, A. Sedaev and F. Sukochev, The Dixmier trace and asymptotics of zeta functions, J.Funct.Anal. 249 (2007), no. 2, 253 -- 283. [7] A. Carey, F. Sukochev, Dixmier traces and some applications to noncom- mutative geometry, (Russian) Uspekhi Mat. Nauk 61 (2006), no. 6 (372), 45 -- 110; translation in Russian Math. Surveys 61 (2006), no. 6, 1039 -- 1099 [8] A. Connes, Noncommutative geometry, Academic Press, San Diego 1994. [9] J. Dixmier, Existence de traces non normales, C. R. Acad. Sci. Paris 262 (1966), A1107-A1108. [10] P. Dodds, B. de Pagter, A. Sedaev, E. Semenov and F. Sukochev, Singular symmetric functionals, Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov. (POMI) 290 (2002) Issled. po Linein. Oper. i Teor. Funkts. 30,42- 71 (Russian). English translation in J. Math. Sci. (N. Y.) 124 (2) (2004), 4867 -- 4885. 29 [11] P. Dodds, B. de Pagter, A. Sedaev, E. Semenov and F. Sukochev, Singu- lar symmetric functionals with additional invariance properties, (Russian) Izv. Ross. Akad. Nauk Ser. Mat. 67 (6) (2003), 111 -- 136. English transla- tion in Izvestiya: Mathematics 67 (2003), 1187-1213. [12] P. Dodds, B. de Pagter, E. Semenov and F. Sukochev, Symmetric func- tionals and singular traces, Positivity 2 (1998), no. 1, 4775. [13] R. Edwards, Functional Analysis, Holt, Rinehart and Winston, New York, 1965. [14] T. Fack, H. Kosaki, Generalized s-numbers of τ -measurable operators, Pa- cific J. Math. 123 (1986), no. 2, 269 -- 300. [15] I. Gohberg, M. Krein, Introduction to the theory of linear nonselfadjoint operators, Translations of Mathematical Monographs, Vol. 18 American Mathematical Society, Providence, R.I. 1969 [16] L.S. Koplienko, Trace formula for nontrace-class perturbations, Sibirsk. Mat. Zh. 25 (1984), no. 5, 62 -- 71 (Russian). English translation in Sib. Math. J. 25 (1984), no. 5, 735 -- 743. [17] S. Krein, Ju. Petunin and E. Semenov, Interpolation of linear operators, Nauka, Moscow, 1978 (Russian). English translation in Translation of Mathematical Monographs, Amer. Math. Soc. 54 (1982). [18] N. Kalton and F. Sukochev, Rearrangement-invariant functionals with applications to traces on symmetrically normed ideals, Canad. Math. Bull. 51 (2008), 67 -- 80. [19] N. Kalton, A. Sedaev, F. Sukochev, Fully symmetric functionals on a Marcinkiewicz space are Dixmier traces, submitted. [20] F.J. Murray, J. von Neumann, On rings of operators, Ann. Math. 37 (1936), no. 1, 116-229. [21] Pietsch A. About the Banach Envelope of l1,∞, Rev. Mat. Complut. 22 (1) (2009) 209 -- 226. [22] Sedaev A. Generalized limits and related asymptotic formulas, Math. Notes. 86:4 (2009), 612-627. [23] B. Simon, Trace ideals and their applications, AMS 2005. [24] A. Sedaev, F. Sukochev, D. Zanin, Lidskii-type formulae for Dixmier tra- ces, Int.Eq.Oper.Th. (to appear) http://arxiv.org/pdf/1003.1817 30
1101.3009	2	1101	2011-01-20T10:34:43	On $q$-normal operators and quantum complex plane	[ "math.OA", "math.QA", "math.RT" ]	For $q>0$ let $\cA$ denote the unital $\ast$-algebra with generator $x$ and defining relation $xx^\ast=qxx^\ast$. Based on this algebra we study $q$-normal operators, the complex $q$-moment problem, positive elements and sums of squares.	math.OA	math	ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN Abstract. For q > 0 let A denote the unital ∗-algebra with generator x and defining relation xx∗ = qxx∗. Based on this algebra we study q-normal operators, the complex q-moment problem, positive elements and sums of squares. 1. Introduction Suppose that q is a positive real number. A densely defined closed linear operator X on a Hilbert space is called q-normal if (1) XX ∗ = qX ∗X. This and other classes of q-deformed operators have been introduced and investigated by S. Ota [Ota], see e.g. [OS2]. In this paper we continue the study of q-normal operators. Further, let A denote the unital complex ∗-algebra with single generator x and defining relation (2) xx∗ = qx∗x. The algebra A appears in the theory of quantum groups where it is considered as the coordinate algebra of the q-deformed complex plane, or briefly, of the complex q-plane. Let us set for a moment q = 1. Then the q-normal operators are precisely the normal operators and A is a complex polynomial algebra C[x, x]. It is well known that there is a close relationship between various important algebraic and analytic problems: Hilbert space), • the complex moment problem, • the extension of formally normal operators to normal operators (possibly in a larger • the characterization of well-behaved representations of the ∗-algebra C[x, x], • the representation of positive polynomials as sums of squares (motivated by 17-th Hilbert problem) The aim of the present paper is to begin a study of these problems and their interplay in the q-deformed case. Date: August 31, 2018. 2000 Mathematics Subject Classification. Primary 14P99, 47L60; Secondary 14A22, 46L52, 11E25. Key words and phrases. q-normal operator, quantum complex plane, q-moments problem. 1 2 JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN We now discuss the contents of this paper. Section 2 deals with q-normal operators. After giving some equivalent characterizations of q-normality we prove a structure theorem for q- normal operators (Theorem 1) which is the counter-part of the spectral theorem for unbounded normal operators. Section 3 deals with positive q-polynomials and their possible representations as sums of squares. We define the cone A+ of positive elements to be the set of elements which are mapped into positive symmetric operators by all well-behaved ∗-representations of the ∗-algebra A. A ∗-representation π of A with domain D(π) is called well-behaved if there exists a q-normal operator X such that D(π) = ∩∞ n=1D(X n) and π(x) = X⌈D(π). Theorem 2 states that for each positive q6=1 there exists a polynomial pq ∈ R[t] of degree four such that the element f := pq(x + x∗) is in A+, but f is not a sum of squares in A. In contrast we prove that if p ∈ R[t] and p(x∗x) ∈ A+, then p(x∗x) is always a sum of squares. In Section 4 we study a generalization of the complex moment problem to the ∗-algebra A. Let F be a linear functional on A. We say that F is a q-moment functional if there exists a well-behaved ∗-representation π of A and a vector ϕ ∈ D(π) such that F (a) = hπ(a)ϕ, ϕi for a ∈ A. (In Section 4 we use Theorem 1 to give a formulation of q-moment functionals in terms of measures.) Further, F is called positive if F (f ∗f ) ≥ 0 for all f ∈ A and strongly positive if F (f ) ≥ 0 for all f ∈ A+. Then, by Theorem 3, a linear functional on A is a q-moment functional if and only if it is strongly positive. This result can be considered as the counter- part of Haviland's theorem. In contrast, by Theorems 4 and 5, there exists a positive linear functional F on A which is not a q-moment functional and a formally q-normal operator which has no q-normal extensions. In the commutative case q = 1 the solution of Hilbert's 17-th problem (see e.g. [M]) implies that positive polynomials of C[x, x] are sums of squares of rational functions. In Section 5 we prove a Positivstellensatz (Theorem 6) which states, roughly speaking, that strictly positive elements of A+ can be represented as sums of squares by allowing "nice" denominators. Operators in Hilbert space. For a linear operator A on a Hilbert space operator we denote by D(A), RanA, A and A∗ denote its domain, its range, its closure and its adjoint, respectively, and we set D∞(A) := ∩n=1D(An). A core of a closed operator A is a linear subset D0 ⊆ D(A) such that the closure of A ↾ D0 coincides with A. If A is self-adjoint, then D∞(A) is a core of A. A number λ ∈ C is called a regular point for an operator A if there exists cλ > 0 such that If Ai, i ∈ I, are linear operators on a Hilbert space Hi, the direct sum ⊕i∈I Ai denotes the k(A − λI)ϕk ≥ cλ kϕk for all ϕ ∈ D(A). operator on H := ⊕i∈IHi defined by (⊕i∈I Ai)(ϕi)i∈I := (Aiϕi)i∈I for (ϕ)i∈I in We close this introduction by collecting some definitions and notations. D(⊕i∈IAi) := {(ϕi)i∈I : ϕi ∈ D(Ai), (ϕi)i∈I ∈ H and (Aiϕi)i∈I ∈ H}. ∗-Algebras and ∗-representations. By a ∗-algebra we mean a complex associative algebra A equipped with a mapping a 7→ a∗ of A into itself, called the involution of A, such that ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. 3 j=1 a∗ (λa + µb)∗ = ¯λa∗ + ¯µb∗, (ab)∗ = b∗a∗ and (a∗)∗ = a for a, b ∈ A and λ, µ ∈ C. In this paper each ∗-algebra A has an identity element denoted by 1A or 1. An element of the formPn j aj, where a1, . . . , an ∈ A is called a sum of squares in A. The set of all sums of squares is denoted by PA2. We use some terminology and results from unbounded representation theory in Hilbert space (see e.g. in [S4]). Let D be a dense linear subspace of a Hilbert space H with scalar product h·,·i. A ∗-representation of a ∗-algebra A on D is an algebra homomorphism π of A into the algebra L(D) of linear operators on D such that π(1) = ID and hπ(a)ϕ, ψi = hϕ, π(a∗)ψi for all ϕ, ψ ∈ D and a ∈ A. We call D(π) := D the domain of π and write H(π) := H. A ∗-representation is faithful if π(a) = 0 implies a = 0. We say that an element a = a∗ ∈ A is positive in a ∗-representation π if hπ(a)ϕ, ϕi ≥ 0 for all ϕ ∈ D(π). Suppose that π is a ∗-representation of A. The graph topology of π is the locally convex topology on the vector space D(π) defined by the norms ϕ 7→ kϕk + kπ(a)ϕk , where a ∈ A. Then π is closed if and only if D(π) = ∩a∈AD(π(a)). We say that π is strongly cyclic if there exists a vector ξ ∈ D(π) such that π(A)ξ is dense in D(π) in the graph topology of π. A linear functional F : A → C on a ∗-algebra A is positive if F (PA2) ≥ 0. Every positive [S4]), that is, there exists a ∗-representation functional F has a GNS representation (see e.g. πF and with cyclic vector ϕ such that F (a) = hπF (a)ϕ, ϕi for a ∈ A. In what follows q is a positive real number. Recall the definition of a q-normal operator, see e.g. [Ota]. 2. q-Normal operators Definition 1. A densely defined operator X on a Hilbert space H is a q-normal operator if D(X) = D(X ∗) and kX ∗fk = √q kXfk , f ∈ D(X). since kX ∗fk = √q kXfk , the graph norms of X and X ∗ are equivalent. Therefore, since X ∗ is A q-normal operator with q = 1 is normal. Each q-normal operator X is closed. Indeed, closed, X is also closed. It also implies that ker X = ker X ∗. The following proposition collects different characterizing properties of q-normal operators, cf. Chapter 2 in [OS]. Proposition 1. Let X be a closed operator on a Hilbert space H and let X = UC be its polar decomposition. The following statements are equivalent: (i) X is q-normal, (ii) XX ∗ = qX ∗X, UC 2U ∗ = qC 2, (iii) UCU ∗ = q1/2C, (iv) UEC(∆)U ∗ = EC(q−1/2∆) for each Borel ∆ ⊆ R+, (v) Uf (C)U ∗ = f (q1/2C) for every Borel function f on C. (vi) 4 JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN Proof. (i) ⇒ (ii) : We use some basic properties of quadratic forms associated with positive operators, see e.g. [RS]. Introduce the quadratic forms t1[ϕ, ϕ] := hX ∗ϕ, X ∗ϕi and t2[ϕ, ϕ] := qhXϕ, Xϕi, D[t1] = D(X ∗), D[t2] = D(X). Let X be q-normal. Then X is closed and t1[ϕ, ϕ] = kX ∗ϕk2 = q kXϕk2 = t2[ϕ, ϕ]. Thus, we get t1 = t2. Since X and X ∗ are closed, t1 and t2 are closed. The operators associated with t1 and t2 are XX ∗ and qX ∗X respectively. Hence XX ∗ = qX ∗X. (ii) ⇒ (iii) : Since X = UC is a polar decomposition, we have X ∗ = CU ∗ and ker U = ker C. Then equation (ii) implies UC 2U ∗ = qCU ∗UC = qC 2. (iii) ⇒ (iv) : Equation (iii) implies that for ϕ ∈ D(C 2) holds q kCϕk2 = qhC 2ϕ, ϕi = hUC 2U ∗ϕ, ϕi = hCU ∗ϕ, CU ∗ϕi = kCU ∗ϕk2 . It implies ker U ∗ ⊆ ker C. On the other hand, if ϕ ∈ ker C, then by the last equation U ∗ϕ ∈ ker C = ker U. That is UU ∗ϕ = 0, which implies U ∗ϕ = 0. Hence ker U ∗ = ker C = ker U. Restricting U, U ∗, C onto (ker C)⊥ we can assume that U is unitary. Then relation (iii) defines a unitary equivalence of C 2 and qC 2. Hence, the square roots C and q1/2C are unitary equivalent and we get (iv). (iv) ⇒ (v) : As in the previous case we can assume that U is unitary. Then C and q1/2C are unitary equivalent and for every Borel ∆ ⊆ R we get UEC(∆)U ∗ = Eq1/2C(∆) = EC(q−1/2∆). (v) ⇒ (i) : Note that (v) implies that U and U ∗ commute with EC({0}), that is ker C is invariant under U and U ∗. Considering the restriction of X (resp. U and C) onto (ker C)⊥ we can assume that U is unitary. Then (v) means that C and q1/2C are unitarily equivalent, namely UCU ∗ = q1/2C, which implies CU ∗ = q1/2U ∗C. Using the latter we get D(X) = D(UC) = D(U ∗C) = D(CU ∗) = D(X ∗) and hCU ∗ϕ, CU ∗ϕi = qhU ∗Cϕ, U ∗Cϕi = qhUCϕ, UCϕi, ϕ ∈ D(X) which implies kX ∗ϕk = q1/2 kXϕk . (v) ⇒ (vi) : Follows from the computation Uf (C)U ∗ = U(cid:18)Z f (λ)dEC(λ)(cid:19) U ∗ =Z f (λ)dEC(q−1/2λ) =Z f (λ)dEq1/2C(λ) = f (q1/2C). (vi) ⇒ (v) : Follows by setting f (λ) = 1∆(λ). (cid:3) We provide a basic example of a q-normal operator for q 6= 1. Put ∆q =(cid:26) [1, q1/2), (q1/2, 1], if q > 1, if q < 1. (3) ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. 5 Let µ be a Borel measure on R+ = [0, +∞) such that (4) Since (0, +∞) = ∪k∈Zqk∆q, the measure µ is uniquely defined by its restriction onto the subspace ∆q ∪ {0} ⊂ R+. Define an operator Xµ, on the Hilbert space Hµ := L2(R+, dµ) as follows. µ(∆) = µ(q1/2∆) for all Borel subsets ∆ ⊆ R+. (5) Let Hµ,0 = {f ∈ Hµ f ≡ 0, a.e. on (0, +∞)} . Then µ({0}) = 0 if and only if Hµ,0 = {0} . We prove the following (Xµϕ)(t) := q1/2tϕ(q1/2t), D(Xµ) = {ϕ(t) ∈ Hµ tϕ(t) ∈ Hµ} . Proposition 2. (i) The operator Xµ in (5) is a well-defined q-normal operator with ker Xµ = Hµ,0. (ii) The adjoint operator X ∗ µ is defined by µϕ)(t) = tϕ(q−1/2t), D(X ∗ (X ∗ µ) = D(Xµ). (iii) Let Xµ = UµCµ be the polar decomposition of Xµ. Then (Cµψ)(t) := tψ(t), (Uµϕ)(t) :=(cid:26) ϕ(q1/2t), 0, for t 6= 0, for t = 0. (6) where ϕ ∈ H, ψ ∈ D(Cµ) = D(Xµ). Proof. It follows from (4) that Uµ is a well-defined partial isometry. Further, we have Xµ = UµCµ and ker Cµ = ker Uµ = Hµ,0. Since Cµ is a positive self-adjoint operator, Xµ = UµCµ is a polar decomposition of Xµ. Since Uµ is bounded, we have X ∗ µ. Together with (4) it implies D(X ∗ µ = CµU ∗ µ) = D(Cµ) = D(Xµ). For ϕ ∈ D(Xµ) we calculate using (4): Thus, Xµ is q-normal. kXµϕk = ktϕ(t)k and (cid:13)(cid:13)X ∗ µϕ(cid:13)(cid:13) = q1/2 ktϕ(t)k . (cid:3) The following theorem can be viewed as an analogue of the spectral theorem for normal operators (see e.g. Theorem VII.3 in [RS]). Theorem 1. Let X be a q-normal operator on a Hilbert space H. Then there exists a family of Borel measures µi, i ∈ I on R+ satisfying µi(x) = µi(q1/2x) such that X is unitarily equivalent to the direct sum of operators Xµi defined by (5). Proof. Let H0 ⊆ H denote ker X = ker X ∗. The restriction X ↾ H0 is a direct sum of copies of Xµ0 where µ0(R+) = µ({0}) = 1. The restriction X1 = X ↾ H⊥ 0 is again a q-normal operator with X ∗ 0 . Without loss of generality, we can assume that ker X = ker X ∗ = {0} . Let X = UC be the polar decomposition of X. Since ker X = {0} , U is unitary. Let K = RanEC(∆q). Then K is invariant under C and we denote by D the restriction C ↾ K. Further, let K = ⊕i∈IKi an arbitrary orthogonal sum decomposition such that Ki is invariant under D and let Di = D ↾ Ki, i ∈ I. We show that there is a corresponding orthogonal sum decomposition of X = ⊕i∈I Xi. 1 = X ∗ ↾ H⊥ 6 JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN It follows from Proposition 1,(v) that U k (RanEC(∆q)) = RanEC(q−k/2∆q), k ∈ Z. Since (0, +∞) is a disjoint union of qk/2∆q, k ∈ Z, we get the following direct sum decom- (7) position RanEC(qk/2∆q) = H = RanEC((0, +∞)) =Mk∈Z U ∗kRanEC(∆q) =Mk∈ZMi∈I =Mk∈Z Hi, where Ki,k = U ∗kKi and Hi =Lk∈Z Ki,k. For i ∈ I, k ∈ Z we define the operators (8) Then for each ϕ ∈ Ki,k holds U kϕ ∈ Ki and we calculate using Proposition 1, (iv) Di,k = U ∗k(qk/2Di)U k : Ki,k → Ki,k. U ∗kKi =Mi∈I Mk∈Z Ki,k =Mi∈I Di,kϕ = qk/2U ∗kDi(U kϕ) = qk/2U ∗kC(U kϕ) = CU ∗k(U kϕ) = Cϕ. It implies that Ki,k is invariant under C and C ↾ Ki,k = Di,k for all i ∈ I, k ∈ Z. Put Ci = Lk∈Z Di,k, i ∈ I. Then Ci, i ∈ I are self-adjoint with RanCi ⊆ Hi and Li∈I Ci = C in particular, Ci = C ↾ Hi. The subspaces Hi, i ∈ I are invariant under U by definition of Hi. We denote by Ui the restriction of U onto Hi, so that U = ⊕i∈IUi. Thus we get X = UC = ⊕i∈I UiCi. Using Zorn's Lemma we can choose Di to be cyclic with cyclic vectors ψi ∈ Ki. Since σ(Di) ⊆ ∆q, there exist unitary operators Vi : Ki → L2(∆q, dµi) such that (9) where µi(·) = hEDi(·)ψi, ψii, see e.g. Chapter VII in [RS]. It follows from (8) that operators Di,k are cyclic on Ki,k, with cyclic vectors ψi,k := U ∗kψi, i ∈ I, k ∈ Z. By (8) we also have σ(Di,k) ⊆ qk/2∆q. We calculate the corresponding measures µi,k(·) = hEDi,k(·)ψi,k, ψi,ki using the unitary equivalence (8): i f )(t) = tf (t), f (t) ∈ L2(∆q, dµi), (ViDiV ∗ µi,k(qk/2∆) = hEDi,k(qk/2∆)U ∗kψi, U ∗kψii = hU kEDi,k(qk/2∆)U ∗kψi, ψii = (10) for each Borel set ∆ ⊆ ∆q. For every k ∈ Z we define an operator: = hEqk/2Di(qk/2∆)ψi, ψii = hEDi(∆)ψi, ψii = µi(∆), Wk : L2(∆q, dµi) → L2(qk/2∆q, dµi,k), (W f )(t) = f (q−k/2t). It follows from (10) that Wk are unitary. Further, for each i ∈ I, k ∈ Z we define unitary operators Vi,k = WkViU k, Vi,k : Ki,k → L2(qk/2∆q, dµi,k). Equations (8),(9) imply that (Vi,kDi,kV ∗ i,kf )(t) = tf (t) for f ∈ L2(qk/2∆q, dµi,k). Since suppµi,k ⊆ qk/2∆q are disjoint for different k ∈ Z, we can define a Borel measure eµi := Pk∈Z µi,k on R+ which satisfies (4). Then eVi = ⊕k∈ZVi,k is a unitary operator from Hi to L2(R+,eµi) such that eViCieVi = Ceµi, where Ceµi, i ∈ I are defined by (6). Using (8) and (9) ∗ ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. 7 ∗ = Ueµi, where Ueµi, i ∈ I are defined by (6). Hence, X = ⊕i∈IXi, where every (cid:3) we get eViUieVi Xi = UiCi is unitary equivalent to Xeµi, i ∈ I. Definition 2. We say that a q-normal operator X is reducible if D(X) = D1 ⊕D2, Di 6= 0 and D1,D2 are invariant under X. Otherwise we say that X is irreducible. For irreducible q-normal operators we obtain the following description, see also [OS], p.71. Proposition 3. Let X be a non-zero irreducible q-normal operator on a Hilbert space H. Then there exists a unique λ ∈ ∆q, and unique orthonormal base {ek}k∈Z in H such that (11) Proof. Since X, X 6= 0 is irreducible, by Theorem 1 we get X = Xµ for some Borel measure µ on (0, +∞) satisfying (4). It follows from the proof of Theorem 1 that operator D = CµECµ(∆q) is irreducible, i.e. one-dimensional. In particular suppµ∩∆q consists of a singular point λ ∈ ∆q. Put Xek = λq−k/2ek+1, X ∗ek = λq−(k−1)/2ek−1, k ∈ Z. ek = 1{λq−k/2}(µ((cid:8)λq−k/2(cid:9)))−1, k ∈ Z. Direct computations show that (11) is satisfied. Further, one can check that every bounded operator which commutes with X and X ∗ is a multiple of identity. Hence X is irreducible. (cid:3) Below we will often use the following Lemma 1. Let q > 0, X be a q-normal operator on a Hilbert space H and let X = UC be its polar decomposition. (i) For all m, n ∈ N0 (12) X ∗mX n = q(m2+m−n2+n−2mn)/4U n−mC m+n, Where U −k denotes U ∗k, k ∈ N. In particular, D∞(X) = ∩m,n∈ND(X ∗mX n) = D∞(C) is dense in H. (ii) The set D∞(X) is a core of X ∗mX n for all m, n ∈ N0. (iii) The set D∞(X) is invariant under U and U ∗. Proof. (i) : By Proposition 1, (iv) we have UC = q1/2CU, which implies X ∗mX n = (CU ∗)m(UC)n = q[(m(m+1))/4−(n(n−1))/4]U ∗mC mU nC n = = q(m2+m−n2+n−2mn)/4U n−mC m+n. (ii) : Since D∞(X) = D∞(C) and C is self-adjoint, D∞(X) is a core of C m, m ∈ N. It follows from (12) that kX ∗mX nϕk = q(m2+m−n2+n−2mn)/4 kC m+nϕk , ϕ ∈ D∞(C), which implies the assertion. (iii) : By Proposition 1, (iv), we have UC k = qk/2C kU, k ∈ Z. Hence D(C k) is invariant for U for all k ∈ Z, i.e. C∞ is invariant for U. In the same way one shows that D∞(X) is invariant for U ∗. (cid:3) 8 JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN Recall that 3. Positive q-polynomials A = Chx, x∗ xx∗ = qx∗xi, where q is a positive real number. Since the set {x∗mxn; m, n ∈ N0} is a vector space basis of A, each element f ∈ A can be written uniquely as f = Pm,n αmnx∗mxn, where amn ∈ C. We define the degree of f by deg f := max{m+n αmn 6= 0} . We will also refer to an element f of A as a q-polynomial and write f = f (x, x∗). If X is a q-normal operator, then f (X, X ∗) denotes the operator Pm,n αmnX ∗mX n. Definition 3. An element f = f ∗ ∈ A is called positive if hf (X, X ∗)ϕ, ϕi ≥ 0 for ϕ ∈ D∞(X) for every q-normal operator X and every ϕ ∈ D∞(X). The set of positive elements of A is denoted by A+. With this notion of positivity one can develop a non-commutative real algebraic geometry on the complex q-plane. In this section we investigate positive elements and sum of squares in A. In Section 5 we prove a strict Positivstellensatz for A. Definition 4. A ∗-representation π of A is called well-behaved if there is a q-normal operator X such that D∞(X) = D(π) and π(x) = X ↾ D∞(X). By Lemma 1 the domain D∞(X) is a core of a q-normal operator X, so there is a one-to-one correspondence between q-normal operators and well-behaved representations of A. Further, an element f = f ∗ ∈ A is in A+ if and only if π(f ) ≥ 0 for every well-behaved ∗-representation π. Thus our definition of positive elements fits into the definition of positivity via ∗-representations proposed in [S2]. Remarks. 1. If µ is a positive Borel measure on R+ satisfying (4) and Xµ is the q-normal operators defined by (5), we denote the corresponding well-behaved ∗-representation of A by πµ. That is, πµ(x) = Xµ ↾ D∞(Xµ), D∞(Xµ) = D(πµ). (13) 2. The ∗-algebra A has a natural Z-grading given by deg x = 1 and deg x∗ = −1. In [SS] a notion of well-behaved ∗-representations was introduced for class of group graded ∗-algebras which contains A. It can be shown that Definition 4 is equivalent to corresponding definition of well-behavedness in [SS], see Definition 11 therein. Suppose that p ∈ R[t]. We consider the following questions: When p(x + x∗) ∈ A+? When p(x + x∗) ∈PA2? First we consider the case when deg p = 2. Below we will need the following ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. 9 N CwN . vector of monomials wN = (1 x x∗ x2 x∗x x∗2 . . . x∗N )T , Lemma 2. Let f = Pm,n αmnX ∗mX n ∈ A, deg f = 2N, N ∈ N. Denote by wN the column N = (1 x x∗ x∗2 x∗x x2 . . . ). Then f is a sum of squares in A if and only if and let w∗ there exists a positive semidefinite (N + 1)(N + 2)/2 × (N + 1)(N + 2)/2 complex matrix C such that f = w∗ Proof. Asume f = Pi f ∗ C(N +1)(N +2)/2 such that fi = aiwN . It implies that f = Pi(aiwN )∗aiwN = w∗ C =Pi a∗ matrices. That is there exist row vectors ai ∈ C(N +1)(N +2)/2 such that C = Pi a∗ implies that f =Pi(aiwN )∗aiwN ∈PA2. Proposition 4. Let a, b ∈ R. The element L := (x + x∗)2 − 2a(x + x∗) + b is in PA2 if and only if b ≥ Proof. First suppose that L ∈PA2. Then by Lemma 2 there is a positive semi-definite matrix N CwN , C ≥ 0, then C is a sum of rank one positive semidefinite i ai, which (cid:3) i fi, f ∈ A. Then deg fi ≤ N, and there exist row vectors ai ∈ N CwN , where 1Cw1. Comparing coefficients at x∗mxn, m + n ≤ 2 yields C = [ci,j]i,j=1,3 such that L = w∗ On the other hand, if f = w∗ i ai is positive semidefinite. 4a2q (q + 1)2 . b = c11 −2a = c13 + c21 = c31 + c12 1 = c23 = c32 q + 1 = c22 + qc33 By multiplying these equations with α1 = 1, α2 = 2aq and adding them we derive   . the product of two positive semidefinite matrices is nonnegative, b − 4a2q (q+1)2 ≥ 0. (q + 1)2 = tr  α1 α2 α2 α2 α4 α3 α2 α3 qα4 (1+q)2 , α3 = 0 and α4 = 4a2q2   c11 c21 c31 c13 c23 c33 c12 c22 c32   b − 4a2q Both matrices in the preceding equation are positive semidefinite. Therefore, since the trace of (1+q)3 , respectively, Conversely, suppose that b ≥ 4a2q(q + 1)−2. Setting f = q−1/2(−2aq(1 + q)−1 + qx + x∗), we compute (14) (x + x∗)2 − 2a(x + x∗) + 4a2q (q + 1)2 = f ∗f. Hence L = f ∗f + ((b−4a2q(q + 1)−2)1/21)2. The preceding result has the following interesting application. (cid:3) 10 JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN Proposition 5. Let X 6= 0 be a q-normal operator. Then each non-zero real number is a regular point for the symmetric operator X + X ∗. In particular, X + X ∗ is not essentially self-adjoint. Proof. Suppose that a ∈ R \ {0}. It follows from (14) that the element (x + x∗)2 − 2a(x + x∗) + 4a2q q + 1(cid:19)2 (q + 1)2 = ((x + x∗) − a)2 − a2(cid:18)q − 1 is a square in A. This implies that (15) k(X + X ∗ − a)ϕk ≥ a(cid:12)(cid:12)(cid:12)(cid:12) q − 1 q + 1(cid:12)(cid:12)(cid:12)(cid:12)kϕk , for ϕ ∈ D∞(X). Since D∞(X) is a core for X + X ∗ by Lemma 1 (ii), the inequality (15) holds for all ϕ ∈ D(X+X ∗). This shows that a is a regular point for X + X ∗. Let T denote the closure of X + X ∗ and assume to the contrary that T is self-adjoint. The numbers of R \ {0} are regular points for X + X ∗ and hence for the selfadjoint operator T . Therefore, R \ {0} ⊆ ρ(T ), so that σ(T ) = {0}. The latter implies that X + X ∗ = 0 which is impossible for X 6= 0. (cid:3) For a positive Borel measure µ on R+ satisfying (4) we define a linear functional Fµ on A by (16) Fµ(f ) = hπµ(f )1∆q, 1∆qi, f ∈ A, where πµ is defined by (13). Lemma 3. Let πµ be the ∗-representation of A defined by (13) and let f = f ∗ ∈ A. Then we have πµ(f ) ≥ 0 if and only if (17) Fµ(g∗f g) ≥ 0 for all g ∈ A. Proof. Let Cµ be as in (6). It follows from relation (12) that the graph topology on D(πµ) = define D∞(Cµ) is generated by the family of seminorms (cid:13)(cid:13)C n µ (·)(cid:13)(cid:13) , n ∈ N0. Set ϕ0 = 1∆q. For k ∈ Z ϕk :=(cid:26) (X k (X ∗k µϕ0)(t), µ ϕ0)(t), if k ≥ 0; if k < 0. with respect to the graph topology. Then ϕk = qk/2tk1{q−k/2∆q}(t) if k ≥ 0 and ϕk = tk1{qk/2∆q}(t) if k < 0. Since X ∗ for each k ∈ Z the set(cid:8)p(X ∗ µXµ)ϕk p ∈ C[t](cid:9) is dense in the subspace L2(q−k/2∆q, dµ) ⊆ D(πµ) Consider the case µ({0}) = 0. Then D(πµ) is a direct sum of L2(qk/2∆q, dµ), k ∈ Z. Hence ϕ0 is cyclic for πµ. Therefore, since Fµ(g∗f g) = hπµ(f )πµ(g)ϕ0, πµ(g)ϕ0i and πµ(A)ϕ0 is dense in D(πµ) in the graph topology, it follows that πµ(f ) ≥ 0 if and only if condition (17) is satisfied. If suppµ = {0} , then the statement is trivial. Consider the case suppµ 6= {0} , µ({0}) 6= 0 and let µ1 be the Borel measure on R+ defined by µ1(∆) = µ(∆ \ {0}). Then πµ(x) is a direct µXµ = C 2 µ, ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. 11 sum of 0 and πµ1(x). Let f = f ∗ ∈ A(q) satisfy (17). Since Fµ = Fµ1, f is positive in π1. It suffices to prove f (0, 0) ≥ 0. For let f =Pm,n αmnx∗mxn and q > 1. Then for k ∈ N we have q−kFµ(x∗kf xk) = q−kXm,n = q−kXm,n αmnhX k+n µ 1∆q, X k+m µ αmnhπµ(x∗(k+m)xk+n)ϕ0, ϕ0i = 1∆qi =Xn qnαnnZ t2n1 q−(k+n)/2∆qdt which converges to α00 = f (0, 0), for k → ∞. In the case q < 1 we have qkFµ(xkf x∗k) → f (0, 0), k → ∞. It implies f (0, 0) ≥ 0. Remark. In the case µ({0}) = 0 we have seen in the preceding proof that the vector ϕ0 = 1∆q is cyclic for πµ. Therefore, by uniqueness of GNS-representation (see Theorem 8.6.4. in [S4]), πµ is unitarily equivalent to the GNS-representation of the positive functional Fµ. (cid:3) Denote by B = C[x∗x] the unital ∗-subalgebra of A generated by the single element x∗x. For every element g ∈ B there exists a unique polynomial, denoted by g(t) ∈ C[t], such that g = g(x∗x). For f =Pm,n αmnx∗mxn ∈ A(q) we define p(f ) :=Xn αnnx∗nxn =Xn αnnqn(n−1)/2(x∗x)n ∈ B. Then the mapping p : A → B is a conditional expectation as introduced in [SS]. We collect some properties of p in a lemma. We omit its simple proof. Lemma 4. Let q > 0. (i) For every positive functional Fµ defined by (16) and every f ∈ A holds Fµ(f ) = Fµ(p(f )). In particular, Fµ(f ) =Z∆q (ii) For f ∈ A and g1, g2 ∈ B, we have p(g∗ (p(f ))(t)dµ(t1/2). 1f g2) = g∗ 1p(f )g2. Proposition 6. Let H = L2(R, dλ), where dλ is the Lebesgue measure on R. We define operators U0 and C0 on H by (18) where D(U0) = H, D(C0) = {ψ tψ(t) ∈ L2(R, dλ)}. Then X0 := U0C0 is a q-normal operator and an element f = f ∗ ∈ A is in A+ if and only if (19) (U0ϕ)(t) = ϕ(t + 1), (C0ψ)(t) = qt/2ψ(t), hf (X0, X ∗ 0 )ψ, ψi ≥ 0 for all ψ ∈ D∞(X0). Proof. Let µ0 be a measure on R+ satisfying (4) such that µ0 ↾ ∆q coincides with the Lebesgue measure dλ. Then the unitary operator U : L2(R, dλ) → L2(∆q, dµ0), (Uψ)(t) = ψ(qt/2) defines a unitary equivalence of X0 and Xµ0. Hence X0 is q-normal and for each f ∈ A+ condition (19) is satisfied. 12 JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN Conversely, suppose that (19) holds. Let πµ0 and Fµ0 be the ∗-representation and positive functional defined by (13) and (16) respectively. Since (19) holds, we have Fµ0(g∗f g) ≥ 0 for all g ∈ A and hence Fµ0(h∗g∗f gh) ≥ 0 for all h ∈ B, g ∈ A(q). Using Lemma 4 we obtain for a fixed g ∈ A and every h ∈ B, Fµ0(h∗g∗f gh) = pµ0(p(h∗g∗f gh)) =Z∆q h∗(t)(p(g∗f g))(t)h(t)dλ(t1/2) = =Z∆q (p(g∗f g))(t)h(t)2dλ(t1/2) ≥ 0. Since the polynomials h ∈ C[t] are dense in L2(∆q, dλ) it follows that (p(g∗f g))(t) ≥ 0 for t ∈ ∆q, g ∈ A. Let µ be another measure on R+ satisfying (4) and Fµ be the corresponding positive func- tional. We assume first that µ({0}) = 0. Then Fµ(g∗f g) =Z∆q (p(g∗f g))(t)dλ(t1/2) ≥ 0 for all g ∈ A which implies that πµ(f ) ≥ 0. Theorem 2. For c ∈ R, set Lc := (x + x∗)4 − 2(x + x∗)2 + c. Then: (i) Lc ∈PA2 if and only if c ≥ q(q+1)2 (ii) there exists ε > 0 such that L := (x + x∗)4 − 2(x + x∗)2 + q(q+1)2 Proof. (i): First suppose that Lc ∈PA2. By Lemma 2 there exists a positive semidefinite 6×6- 2Cw2. Comparing coefficients at x∗mxn in the equation L = w∗ (q2+1)2 − ε ∈ A+. (q2+1)2 , 2Cw2, matrix C such that L = w∗ we obtain the following equations (cid:3) c = c1,1, 0 = c1,2 + c3,1 = c2,1 + c1,3, −2 = c1,4 + c3,2 + c6,1 = c4,1 + c2,3 + c1,6, −2q − 2 = c1,5 + c2,2 + qc3,3 + c5,1, 0 = c2,6 + c4,3 = c6,2 + c3,4, 0 = qc3,5 + c2,4 + c5,2 + q2c6,3 = qc5,3 + c4,2 + c2,5 + q2c3,6, 1 = c4,6 = c6,4, (1 + q)(1 + q2) = q2c5,6 + c4,5 = q2c6,5 + c5,4, (1 + q2)(1 + q + q2) = c4,4 + qc5,5 + q4c6,6, ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. 13 Multiplying each line with the respective αi, (1+q2)2 , α1 = 1, α2 = 0, α3 = 0, α4 = q+q2 α5 = 0, α6 = 0, α7 = α8 = − α9 = q3(1+q)2 (−1+q)2(1+q2)2(1+q+q2), q2(1+q)3 (−1+q)2(1+q2)3(1+q+q2) , q(1+q)2 (−1+q)2(1+q2)2(1+q+q2), and adding them we get q(1 + q)2 (1 + q2)2 = Tr c −     α1 α2 α2 α4 α2 α3 α3 α6 α4 α6 α3 α5 α3 α2 α6 α3 qα4 α5 α5 α9 qα6 α8 q2α6 α7 α4 α6 qα6 α8 qα9 q2α8 α3 α5 q2α6 α7 q2α8 q4α9   C   By some simple computations one checks that the matrix containing the αi is positive semide- finite. Since C is also positive semidefinite, it follows from the preceding that (20) q(1 + q)2 (1 + q2)2 . c ≥ Conversely, suppose that (20) is satisfied. Setting u1(x, x∗) = − q 1 + q2 + x∗2 q(1 + q) + (1 + q2)x∗x 1 + q + q2x2 1 + q u2(x, x∗) = − q 1 + q2 + x∗2 1 + q + (1 + q2)x∗x 1 + q + qx2 1 + q , and we compute 1 + q + q2 q(1 + q)2 (1 + q2)2 = u∗ (21) 1u1 + (x + x∗)4 − 2(x + x∗)2 + This implies that L ∈PA2 if (20) holds. (ii): Since the element L remains invariant if we replace x by x∗ and q by q−1, it suffices to treat the case q > 1. Assume to the contrary that no such ε > 0 exists. Let X0 be the q-normal operator from Proposition 6. Then there exists a sequence of unit vectors ϕn ∈ D∞(X0) such that hL(X0, X ∗ (22) 0 )ϕn, ϕni → 0 as n → ∞. It follows from (21) that u∗ 2u2. 0 )ϕn → 0 and u2(X0, X ∗ 0 )ϕn → 0 as n → ∞. u1(X0, X ∗ q 14 JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN Put X := X 2 0 . Then X is a q4-normal operator, X ∗ = X ∗2 0 and X ∗ 0 X0 = (qX ∗2 0 X 2 0 )1/2 = q1/2X. Define operators q(q + 1) (q − 1) 1 q − 1 F1 = F2 = (u1(X0, X ∗ 0 ) − u2(X0, X ∗ 0 )) = X ∗2 (qu1(X0, X ∗ 0 ) − u2(X0, X ∗ 0 )) = qX + 0 − q2X 2 1 + q2 1 + q 0 = X ∗ − q2X, q q1/2X − 1 + q2 . Then F1ϕn → 0 and F2ϕn → 0. Let X = UC be the polar decomposition of X. From Proposition 1 we get X ∗ − q2X = q2U ∗(I − U 2)C which implies that (23) Put α = q, β = q1/2 1+q2 1+q , γ = − q Combined with (23) it follows that 1+q2 , so that F2 = αUC + βC + γ. Then (I − U 2)F2ϕn → 0. (I − U 2)Cϕn → 0. (24) (I − U 2)ϕn → 0. Let T+ =(cid:8)eit, t ∈ [−π/2, π/2)(cid:9) , T− =(cid:8)eit, t ∈ [π/2, 3π/2)(cid:9) . We put ξn = EU (T+)ϕn, ψn = EU (T−)ϕn. Then ϕn = ψn + ξn and (24) yields (25) Since C is positive, C + 1 is invertible, so that (C + 1)−1F2ϕn → 0. By UC = q2CU, (U − I)ξn → 0 and (U + I)ψn → 0. Using (25) we obtain which implies that (C + 1)−1(αq2CU + βC + γ)ϕn → 0. αq2C C + 1 (ξn − ψn) + βC + γ C + 1 (ξn + ψn) → 0 (αq2 + β)C + γ C + 1 (αq2 − β)C − γ C + 1 ξn − ψn → 0. (26) Since q > 1, αq2 − β > 0 and γ < 0. Hence the operator (αq2 − β)C − γ has a bounded inverse. Applying this inverse to the preceding equation we get (αq2 + β)C + γ (αq2 − β)C − γ Applying U and using Proposition 1 and (25) we derive (αq2 + β)q2C + γ (αq2 − β)q2C − γ ξn → 0. ξn → 0. −ψn + ψn + (27) Adding (26) and (27) we obtain (28) α1C 2 + β1C + γ1 ((αq2 − β)q2C − γ)((αq2 − β)C − γ) ξn → 0, n → ∞, where α1 = 2q2(α2q4 − β2), β1 = −2βγ(1 + q2), γ1 = −2γ2. The polynomial α1c2 + β1c + γ1 has two real roots c1 < 0 < c2. ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. 15 Since X = (U0C0)2 = q−1/2U 2 0 C 2 0 , we get U = U 2 0 , C = q−1/2C 2 0 . By (18), (Uϕ)(t) = ϕ(t + 2), (Cψ)(t) = qt−1/2ψ(t) for ϕ ∈ L2(R, dλ), ψ ∈ D(C 2 0 ). Set t2 = logq c2 + 1/2. Then (28) implies that kξnk2 −Z t2+1/2 t2−1/2 ξn2dt → 0. We end up the section with the following Hence hUξn, ξni = RR ξn(t + 2)ξn(t)dt → 0. Combined with (25) this yields ξn → 0. Further, (26) implies ψn → 0 which contradicts kξn + ψnk = kϕnk = 1. Proposition 7. Let f ∈ R[t]. If f (x∗x) ∈ A(q)+, then f (x∗x) ∈PA(q)2. Proof. Let us choose a measure µ satisfying (4) such that supp µ = R+ and let Xµ be the opera- µXµ is equal to R+. Therefore, since f (X ∗ tor defined by (5). Then the spectrum of X ∗ µXµ) ≥ 0, we have f ≥ 0 on R+. Hence (see e.g. [M]) there exist polynomials g1, g2 ∈ C[t] such that f (t) = g1(t)∗g1(t) + t · g2(t)∗g2(t). Then (cid:3) f (x∗x) = g1(x∗x)∗g1(x∗x) + x∗x · g2(x∗x)∗g2(x∗x) = = g1(x∗x)∗g1(x∗x) + x∗ · g2(qx∗x)∗g2(qx∗x) · x ∈XA(q)2. (cid:3) 4. The complex q-moments problem and formally q-normal operators Definition 5. A linear functional F on A is called a q-moment functional if there exists a well-behaved ∗-representation π of A and a vector ξ ∈ D(π) such that (29) F (a) = hπ(a)ξ, ξi for all a ∈ A. Then the q-moment problem asks: When is a given functional F on A a q-moment functional? In this formulation the q-moment problem is a generalized moment problem in the sense of [S5]. Next we give two reformulations of the q-moment problem. Since {x∗kxl; k, l ∈ N0} is a vector space basis of A, there is a one-to-one-correspondence between complex 2-sequences and linear functionals on A given by Fa(x∗kxl) = akl, k, l ∈ N0, where a = (akl)k,l∈N0 is a 2-sequence. The definition of a well-hehaved representation (Definition 4) yields the following equivalent formulation of the q-moment problem: Given a 2-sequence (akl)k,l∈N0, does there exist a q-normal operator X and a vector ξ ∈ D∞(X) such that akl = hX ∗kX lξ, ξi for all k, l ∈ N0? (30) 16 JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN Before we turn to the second reformulation we consider an example. Example. Suppose that µ is a positive Borel measure on ∆q. Denote byeµ the unique extension of µ to a measure on R+ satisfying (4). Let Xµ be the q-normal operator defined by (5) and ξ ∈ D∞(Xµ). Then there is a q-moment functional defined by Fµ,ξ(f (x, x∗)) := hf (Xµ, X ∗ µ)ξ, ξi. Since (X ∗k µ X l µξ)(t) = q(l2+l−k2+k−2kl)/4tk+lξ(q(l−k)/2t), k, l ∈ N0, by Lemma 1, the corresponding q-moments are akl = Fµ,ξ(x∗kxl) = hX ∗k µ X l µξ, ξi =ZR+ q(l2+l−k2+k−2kl)/4tk+lξ(q(l−k)/2t)ξ(t)dµ(t) (31) (qk/2t)k+lξ(ql/2t)ξ(qk/2t)dµ(qk/2t) = = q(l2+l−k2+k−2kl)/4ZR+ = q(l2+l+k2+k)/4ZR+ tk+lξ(ql/2t)ξ(qk/2t)dµ(t). Using Theorem 1 and formula (31) we obtain another equivalent formulation of the q-moment problem in terms of measures and integrals: Given a 2-sequence (akl)k,l∈N0, does there exist a family µi, i ∈ I, of positive Borel measures on ∆q and a vector ξ = (ξi) ∈Li D∞(Xµi) in the Hilbert space Li L2(R+,eµi) such that akl =Xi q(l2+l+k2+k)/4ZR+ tk+lξi(ql/2t)ξi(qk/2t)deµi(t) for k, l ∈ N0? The next theorem is the counter-part of Haviland's theorem from the classical moment prob- lem. For this we need the following Definition 6. A linear functional F on A is said to be positive if F (a∗a) ≥ 0 for all a ∈ A and it strongly positive if F (a) ≥ 0 for all a ∈ A+. Each strongly positive functional is positive, but Proposition 4 below shows that the converse is not true. Theorem 3. A linear functional F on A is a q-moment functional if and only if F is strongly positive. Proof. From the definition of the cone A+ (Definition 3) it is obvious that q-moment functionals are strongly positive. Suppose that F is strongly positive. To prove that F is a q-moment functional we need some preparations. First we define some auxiliary algebras. Let F be the ∗-algebra of all Borel functions f (t) on R+ which are polynomially bounded (that is, there exists a polynomial p ∈ C[t] such that f (t) ≤ p(t) for t ∈ R+). We denote by X the ∗-algebra generated by an element u and the ∗-algebra F with defining relations (32) u∗u = uu∗ = 1, uf (t) = f (q1/2t), f (t)u∗ = f (q1/2t), ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. 17 for f ∈ F . Clearly, X has a vector space basis {xnck; k ∈ N0, n ∈ Z}, where c2 = x∗x and x−n := x∗n for n < 0, n ∈ Z. Hence there is an injective ∗-homomorphism J of A into X given by J(x) = uf0, where f0(t) = t. We identify J(a) and a for a ∈ A and consider A as a ∗-subalgebra of X. With a slight abuse of notation we shall write x = ut, where t means the function f0(t) = t on R+. Let µ be a Borel measure on R+ satisfying (4). Then (πµ(f )ϕ)(t) = f (q1/2t)ϕ(q1/2t), (uϕ)(t) = ϕ(q1/2t), ϕ ∈ D(πµ) =(cid:8)ϕ ∈ L2(R+, µ) tnϕ ∈ L2(R+, µ) for all n ∈ N(cid:9) , defines a ∗-representation of X on H = L2(R+, dµ) and πµ(ut) = Xµ is the q-normal operator given by (5). Setting X+ := {x ∈ X πµ(x) ≥ 0 for all measures µ satisfying (4)} , we clearly have A+ = X+ ∩ A. Let Xb be the ∗-subalgebra of X generated by u and the subset Fb of all f ∈ F of compact support and consider the ∗-subalgebra Y = A + Xb of X. Clearly, A+ is cofinal in Y+ := Y ∩X , that is, for each y ∈ Y+ there exists a ∈ A+ such that a−y ∈ Y+. Therefore, since A+ = X+∩A, F extends to a linear functional, denoted again by F, such that F (y) ≥ 0 for all y ∈ Y+. Let πF denote the ∗-representation of Y with cyclic vector ϕ obtained by the GNS construction from the functional F (see e.g. [S4], Section 8.6). Then, by the GNS-construction, (33) F (y) = hπF (y)ϕ, ϕi for y ∈ Y. Let 1∆ be the characteristic function of a Borel subset ∆ ⊆ R+ and define (E(∆)f )(t) = πF (1∆)f (t), f ∈ H(πF ). Then E defines a spectral measure on R+. Let U = πF (u). From (32) it follows that UE(∆)U ∗ = E(q−1/2∆). Let C = R ∞ 0 λdE(λ). Then X := UC is a q- normal operator on H(πF ) by Proposition 1. The proof is complete once we have shown that πF (x) ⊆ X, or equivalently, (34) Indeed, because X is q-normal, by Lemma 1 and Definition 4 there is a well-behaved ∗- representation π of A on D(π) := ∩nD(X n) such that π(x) = X⌈D(π). The relation πF (x) ⊆ X implies that πF ⊆ π. Therefore, by (33), F (a) = hπF (a)ϕ, ϕi = hπ(a)ϕ, ϕi for a ∈ A, so F is a q-moment functional. Let f ∈ Fb and k ∈ Z. Then the operator πF (f ) is bounded and we have f (C) = R ∞ 0 f (λ)dE(λ) = πF (f ) by the spectral calculus. Therefore, since ukf (t) = f (qk/2t)uk by (32), we have πF (a)ϕ ∈ D(X) and πF (x)πF (a)ϕ = XπF (a)ϕ for a ∈ A. CπF (ukf (t)) = CπF (f (gk/2t))πF (uk) = Cf (gk/2t)πF (uk) (35) We prove (34) for a = uτ ntn, where τ = ±1 and n ∈ N0. Let ε > 0 be fixed. We choose αε > 0 such that t2n ≤ ε(1 + t2n+2) for t > αε and denote the characteristic function of the = πF (tf (gk/2t))πF (uk) = πF (tukf (t)). 18 JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN interval [0, αε] by 1ε. Setting gε(t) = 1ε(t)tn and fε(t) = tn − gε(t), we have gε ∈ Fb and fε(t)2 ≤ ε(1 + t2n+2) for t ∈ R+. Hence (36) ε(t2 + t2n+4) − t2fε(t)2 ∈ Y+, ε(1 + t2n+2) − fε(t)2 ∈ Y+. Now we compute kπF (x)πF (uτ ntn)ϕ − XπF (uτ ngε)ϕk2 = kπF (ut)πF (uτ ntn)ϕ − πF (u)CπF (uτ ngε)ϕk2 = = kπF (u)(πF (tuτ ntn)ϕ − πF (tuτ ngε))ϕk2 = kπF (q−n/2uτ (n+1))t(tn − gε(t)))ϕk2 = = q−n kπF (tfε(t))ϕk2 = q−nhπF (t2fε(t)2)ϕ, ϕi = q−nF (t2fε(t)2) ≤ εq−nF (t2 + t2n+4). (37) Here we used first equations (35) and (32), then the fact that πF (uτ (n+1)) preserves the norm and equation (33) for y = t2fε(t)2. Since F is Y+-positive, we have F (ε(t2+t2n+4)−t2fε(t)2) ≥ 0 by (36) which gives the inequality in the last line. Using now the fact that ε(1 + t2n+2) − fε(t)2 ∈ Y+ by (35) we derive kπF (uτ ntn)ϕ − πF (uτ ngε)ϕk2 = kπF (uτ nfε)ϕk2 = kπF (fε(t)ϕk2 = F (fε(t)2) ≤ εF (1 + t2n+2). (38) Letting ε → 0, (4) and (4) imply that πF (uτ ngε)ϕ → πF (uτ ntn)ϕ and XπF (uτ ngε)ϕ → πF (x)πF (uτ ntn)ϕ. Therefore, since X is closed, we have πF (x)πF (uτ ntn)ϕ ∈ D(X) and πF (x)πF (uτ ntn)ϕ = XπF (uτ ntn)ϕ. This proves (34) for a = uτ ntn. Since these elements span A, (34) holds for all a ∈ A which completes the proof. (cid:3) Theorem 4. There exists a positive linear functional on A which is not a q-moment functional. Before we prove this we state two technical lemmas. The first one is taken from [S3], Lemma 2. Lemma 5. Let A be a unital ∗-algebra which has a faithful ∗-representation π and is the union of a sequence of finite dimensional subspaces En, n ∈ N. Assume that for each n ∈ N there exists a number kn ∈ N such that the following is satisfied: if a ∈PA2 is in En, then we can write a as a finite sum Pj a∗ Then the cone PA2 is closed in A with respect to the finest locally convex topology on A. Lemma 6. Suppose that π is a well-behaved representation such that π(x) 6≡ 0. Then π is faithful. Proof. Since π is well-behaved, there is a q-normal operator X such that π(x) = X ↾ D∞(X). By Theorem 1, X is a direct sum of operators Xµi. Since π(x) 6= 0, Xµi 6= 0 for one i. Suppose that f (x, x∗) ∈ A, f 6= 0. It suffices to prove that there exists a vector ϕ ∈ D∞(Xµi) such that f (Xµi, X ∗ The polar decomposition Xµi = UµiCµi is given by (6). From Proposition 1(iv) it follows µifk(Cµi). qm/2∆q (t) and choose m ∈ Z such that the that there are polynomials fk ∈ C[t], k = −n, . . . , n sich that f (Xµi, X ∗ Since f 6= 0, there is a j such that fj 6= 0. Put ϕ = 1 j aj such that all aj are in Ekn. µi) =Pn µi)ϕ 6= 0. k=−n U k ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. 19 interval qm/2∆q contains no zero of fj(t). Then, since µi 6= 0, we have µi(qm/2∆q) 6= 0 by (4) and hence ϕ 6= 0. Using (6) we calculate µ fk(t)1 q(m−k)/2∆q(t). qm/2∆q(t) = fk(qk/2t)1 f (X ∗ µ, Xµ)ϕ = nXk=−n U k nXk=−n µ, Xµ)ϕ 6= 0. q(m−k)/2∆q(t) ≡ 0 in L2(R+, dµ) for all k, in particular If the latter would be zero, then fk(qk/2t)1 for k = j which is a contradiction. Thus f (X ∗ (cid:3) Proof of Theorem 4: We denote by Ek the subspace of elements f ∈ A, deg f ≤ k. Obviously, P g∗ i gi ∈ E2k implies that gi ∈ Ek for all i. By Lemma 6 A has a faithful representation. Therefore, Lemma 5 applies, so the conePA(q)2 is closed in the finest locally convex topoloogy. By Theorem 2 there exists an element L ∈ A+ such that L /∈PA2. SincePA2 is closed, by the separation theorem for convex sets there is a linear functional F on A such that F (L) < 0 and F (PA(q)2) ≥ 0. By the latter condition, F is a positive linear functional. Since F is not strongly positive (by F (L) < 0), it is not a moment functional by Theorem 3. (cid:3) operator if D(X) ⊆ D(X ∗) and kXfk = √q kX ∗fk for f ∈ D(X). Definition 7. A densely defined operator X on a Hilbert space H is a formally q-normal It is well-known [C] that there exist formally normal operators which have no normal exten- sions in larger Hilbert spaces. The next theorem shows that a similar result holds for formally q-normal operators. Theorem 5. There exists a formally q-normal operator X which has no q-normal extension in a possibly larger Hilbert space. Proof. We retain the notation from the proof of Theorem 4. Let πF denote the GNS represen- tation of F with cyclic vector ϕ, see [S4]. Then F (a) = hπF (a)ϕ, ϕi for a ∈ A. sion. D(πF (x)∗) = D(X ∗). For ψ ∈ D(X) we have We show that X := πF (x) is a formally q-normal operator which has no q-normal exten- Indeed, since πF is a ∗-representation of A, we have D(X) = D(πF ) = D(πF (x∗)) ⊆ kXψk2 = hπF (x)ψ, πF (x)ψi = hπF (x∗x)ψ, ψi = q−1hπF (xx∗)ψ, ψi = = hX ∗ψ, X ∗ψi = q−1 kX ∗ψk2 . Assume that Y is a q-normal operator on a (possible larger) Hilbert space such that X ⊆ Y . Then L(X, X ∗) ⊆ L(Y , Y ∗) and hence hL(Y , Y ∗)ϕ, ϕi = hL(X, X ∗)ϕ, ϕi = hπF (L)ϕ, ϕi = F (L) < 0. (cid:3) Since L ∈ A+, this is a contradiction. 5. A strict Positivstellensatz for q-polynomials The strict Positivstellensatz (Theorem 6) proved in this section can be viewed as a q-analogue of the Reznick's Positivstellensatz [R]. 20 JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN Let f =Pi,j aijx∗ixj ∈ A and deg f = m. We denote by fm =P{i+j=m} aijx∗ixj the highest order degree part of f. We write fm as fm = ⌊m/2⌋Xr=0 br(x∗x)rxm−2r + ⌊m/2⌋−1Xr=0 b−rx∗(m−2r)(x∗x)r, br ∈ C. The symbol of f is the function σf (ω, ω) on C \ {0} defined by (39) σf (ω, ω) := brω−r(m−2r)ω m 2 −r + b−rω m 2 −rω−r(m−2r). ⌊m/2⌋Xr=0 ⌊m/2⌋−1Xr=0 Let N denote the set consisting of 1 and all finite products of elements qkx∗x + 1, where k ∈ Z. Theorem 6. Let f = f ∗ ∈ A, deg f = 4m, m ∈ N. Suppose that: (i) For every q-normal operator X there exists a εX > 0 such that hf (X, X ∗)ϕ, ϕi ≥ εXhϕ, ϕi, ϕ ∈ D∞(X), (ii) σf (ω, ω) > 0 for all ω ∈ T :=(cid:8)z ∈ C : z = q1/2(cid:9) . Then there exists an element b ∈ N such that b∗f b ∈PA(q)2. The proof of this theorem follows a similar pattern as the proof of the strict Positivstellensatz for the Weyl algebra given in [S1]. We first recall a basic definition and a result from [S1], see e.g. [S2]. A unital ∗-algebra Y is called algebraically bounded if for each element y ∈ B there exists a λy > 0 such that λy · 1 − y∗y ∈XY 2. hπ(y)ϕ, ϕi > 0 for all ϕ ∈ Hπ, ϕ 6= 0, (40) Lemma 7. Let Y be an algebraically bounded ∗-algebra and y = y∗ ∈ B. If (41) for ∗-representation π of Y, then y ∈PY 2. Proof. Assume to the contrary that y /∈PX 2. Since Y is algebraically bounded, 1 is an internal point of the wedge PY 2. Therefore, by the Eidelheit separation theorem for convex sets [K], there exists a linear functional F on Y such that F (y) ≤ 0, F (1) > 0, and F (PY 2) ≥ 0. If πF denotes the GNS-representation of F with cyclic vector ϕ, then F (y) = hπF (y)ϕ, ϕi ≤ 0. Since F (1) = kϕk2 > 0, the latter contradicts (41). (cid:3) The proof of the theorem will be divided into three steps. I. Let ρ be a fixed well-behaved representation of A such that ρ(x) 6= 0. By Lemma 6, ρ is faithful. For notational simplicity we identify a ∈ A with ρ(a). Then x is a q-normal operator and A becomes a ∗-algebra of operators acting on the invariant dense domain D(ρ) = D∞(x). Define the following operators yk := x2(qkx∗x + 1)−1, vk := x(qkx∗x + 1)−1, zk := (qkx∗x + 1)−1, k ∈ Z. ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. 21 Is is easily checked that yk, y∗ itself. Let X be the ∗-algebra of operators on D∞(x) generated by x, x∗, yk, y∗ Then the follwing relations hold in X : k are bounded operators which map the domain D∞(x) into k, zk, k ∈ Z. k, vk, v∗ k, vk, v∗ q2k+1y∗ kyk + qkv∗ y∗ myk = y∗ kym, y∗ kyk = q−4yk−2y∗ xzk = vk, zkx∗ = v∗ kvk + zk = 1, k, x2zk = yk, zkx∗2 = y∗ k xzk = zk+1x, x∗zk = zk−1x∗ zkzm = zmzk, z∗ k = zk, kvk + z2 qkv∗ k = zk, k = qv∗ k−2, vkv∗ k+1vk+1, kzm = zm−2y∗ y∗ k, kzm = zm−1v∗ v∗ k, vkzm = vmzk, k = y∗ m+2, qkzk(1 − zm) = qmzm(1 − zk), kym = (1 − zk)(1 − zm), k, m ∈ Z. ykzm = zm+2yk, vkzm = zm+1vk, ymy∗ k+1ym+1, ykym = ym+2yk−2, y∗ k = qy∗ k−2y∗ my∗ (42) (43) (44) (45) (46) (47) (48) (49) (50) (51) (52) qk+m+1y∗ Let Y denote the subalgebra of X generated by 1, yk, y∗ follows that condition (40) holds for the algebra generators y = yk, y∗ is an algebraically bounded ∗-algebra by Lemma 2.1 in [S1]. Our next aim is to study representations of Y. k, vk, v∗ k, zk, where k ∈ Z. From (45) it k, zk of Y. Hence Y k, vk, v∗ II. Suppose that π is a non-zero ∗-representation of Y on a Hilbert space Hπ. Since Y is algebraically bounded, all operators π(y), y ∈ Y, are bounded, so we can assume that D(π) = Hπ. Let H0 = π(z0), H1 = ker π(1−z0). By (51) we have ker π(zk) = H0 and ker π(1−zk) = H1, k ∈ Z. From (45) it follows that π(vk) ↾ H0 = 0. The third relation in (46) yields π(v∗ k), k ∈ Z. It follows from (45) that π(vk), π(v∗ k) restricted onto H1 are 0. The preceding implies that H0 and H1 are invariant subspaces of the representation π. Let π = π0 ⊕ π1 ⊕ π2 be the corresponding decomposition of π on Hπ = H0 ⊕ H1 ⊕ H2. k) ↾ H0 = 0. Further, (48) implies that H0 is invariant under π(yk), π(y∗ k), π(yk), π(y∗ kyk) = 1/q2k+1. By (46), π0(yky∗ Now we analyze the three subrepresentations π0, π1 and π2. We begin with π0. By con- struction of π0 we have π0(zk) = π0(vk) = π0(v∗ k) = 0 for k ∈ Z. From (45) we obtain π0(y∗ k) = q4π0(y∗ k+2yk+2) = 1/q2k+1, so that π0(qk+1/2yk) is unitary. Finally, it follows from (52) that all operators π0(qk+1/2yk) coincide. Hence there exists a unitary operator Y on H0 such that (53) π0(yk) = q−k− 1 2 Y, k ∈ Z. Next we consider π1. As noted above, π1(1) = π1(zk) = IH1 and π1(vk) = π1(v∗ k) = π1(yk) = π1(y∗ k) = 0. 22 JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN Finally, we turn to π2. It is convenient to introduce the notation Zk := π2(zk), Vk := π2(vk), Yk := π2(yk). Then ker Zk = ker(I − Zk) = {0} by the construction of π2. Combined with (45) we conclude that 0 < Zk < I. Further, (52) implies that all operators q−k(Z −1 k − 1), k ∈ Z, are equal and positive. Set C := q−k/2(Z −1 k − 1)1/2. Then Zk = qkC 2 + 1)−1 and using (45) we get Vk = C(qkC 2 +1)−1. Let Vk = UkVk be the polar decomposition of Vk. Note that ker C = {0} . From (49) we derive UkC(qkC 2 + 1)−1(qmC 2 + 1)−1 = UmC(qmC 2 + 1)−1(qkC 2 + 1)−1, k, m ∈ Z. This implies that all operators Uk, k ∈ Z, are equal. Set U := Uk. Since ker Vk = ker V ∗ U is unitary. By (48) we have k = {0} , UC(qkC 2 + 1)−1(qmC 2 + 1)−1 = (qm+1C 2 + 1)−1UC(qkC 2 + 1)−1, k, m ∈ Z, and since C(qkC 2 + 1)−1 is invertible, U(qmC 2 + 1)−1 = (qm+1C 2 + 1)−1U. The latter is equivalent to UC 2U ∗ = qC 2. Therefore, by Proposition 1, X := UC = VkZ −1 is k a q-normal operator. Further, π2 leaves D∞(X) = D∞(C) invariant. Indeed, by construction this is true for the generators and hence for all elements of Y. Let h ∈ X. Using the relations (43) and (qkx∗x + 1)zk = 1X it follows that h is of the form h = h1zk1zk2 . . . zkm, where h1 ∈ A and k1, . . . , km ∈ Z. Since the operators zk1, . . . , zkm map D∞(x) bijectively onto itself, h = 0 if and only if h1 = 0. Therefore, the ∗-representation π2 gives rise to a unique ∗-representation eπ2 of X on D∞(X) defined by eπ2(y) = π2(y) ↾ D∞(X) for y ∈ Y, III. Now let f be as in Theorem 6 and let f4m =Pi+j=4m aijx∗ixj be its highest degree part. eπ2(x) = X ↾ D∞(X), and eπ2(zk) = (qkX ∗X + 1)−1 ↾ D∞(X), k ∈ Z. From (42) and (43) it follows that y := zm 0 f (x, x∗)zm 0 ∈ Y. Our next aim is to apply Lemma 7 in order to conclude that y ∈PY 2. Let π = π0 ⊕ π1 ⊕ π2 be a representation of Y as analyzed above. First we determine π0(y). Suppose i, j ∈ N0 and i + j < 4m. Applying the relations (42) 0 = w1w2 . . . ws, where each wl is equal to one of the elements k, vk, v∗ k, z0. Since i + j = 4m, not all elements wl can be equal to some yk. Therefore, since 0 ) = 0 ) = 0. Hence π0(y) = π0(zm k) = 0, we obtain π0(zm 0 x∗ixjzm and (43) it follows that zm yk, y∗ π0(zk) = π0(vk) = π0(v∗ π0(zm 0 ). Now we write 0 f4mzm 0 x∗ixjzm 0 f zm (54) f4m = 2mXk=0 bk(x∗x)2m−kx2k + 2mXk=1 b−kx∗2k(x∗x)2m−k. ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. 23 For the monomial (x∗x)2m−kx2k we treat the cases k ≤ m and k > m. First suppose that k ≤ m. Using relations (42)-(44) we compute zm 0 (x∗x)2m−kx2kzm 0 = zm = (z0x∗x)m(x∗xz2k)m−kx2kzk 0 (x∗x)2m−kzm−k 2k x2kzk 0 = 0 = = (1 − z0)m(q−2k(1 − z2k))m−k(x2z2k−2)(x2z2k−4) . . . (x2z0) = = (1 − z0)m(q−2k(1 − z2k))m−ky2k−2y2k−4 . . . y0. Applying π0 to both sides and using (53) we derive π0(zm 0 ) = q−2k(m−k)π0(y2k−2y2k−4 . . . y0) = 0 (x∗x)2m−kx2kzm = q−2k(m−k)q−k2+k/2Y k = q1/2−2k(2m−k)(q1/2Y )k. In the case k > m we obtain zm 0 (x∗x)2m−kx2kzm 0 = z2m−k 0 (x∗x)2m−kzk−m 0 x2kzm 0 = 0 (cid:1) = x2(k−m)(cid:1)(cid:0)x2mzm = (z0x∗x)2m−k(cid:0)zk−m = (1 − z0)2m−k(x2z−2)(x2z−4) . . . (x2z−2(k−m))× ×(x2z2(m−1))(x2z2(m−2)) . . . (x2z0) = 0 = (1 − z0)2m−ky−2y−4 . . . y−2(k−m) · y2(m−1)y2(m−2) . . . y0. Again, applying π0 and using (53) we get π0(zm 0 (x∗x)2m−kx2kzm 0 ) = = π0(y−2y−4 . . . y−2(k−m))π0(y2(m−1)y2(m−2) . . . y0) = = q(k−m)(k−m+1)− k−m 2 Y k−mq−m(m−1)− m 2 Y m = q1/2−2k(2m−k)(q1/2Y )k. Proceeding in a similar manner we derive π0(zm 0 x∗2k(x∗x)2m−kzm 0 ) = q1/2−2k(2m−k)(q1/2Y ∗)k. Let Y =RT ω dE(ω) be the spectral decomposition of the unitary operator Y . Comparing the preceding computations with the definition of σf we get π0(y) = RT σf (q1/2ω, q1/2ω) dE(ω). From assumption (ii) it follows that there exists ε > 0 such that σf (q1/2ω, q1/2ω) ≥ ε for ω ∈ T. Hence hπ0(y)ψ, ψi =ZT σf (q1/2ω, q1/2ω) dhE(ω)ψ, ψi ≥ εkψk2, ψ ∈ H0. Applying assumption (i) to the q-normal operator X = 0 yields a00 = f (0, 0) > 0. By (42) and (43) we have π1(y) = π(a00z2m 0 ) = a00 · I. Finally we turn to π2. As shown above, there exists a ∗-representation eπ2 of the ∗-algebra X on D∞(X) such that eπ2 ↾ A is well-behaved and eπ2(y) = π2(y) ↾ D∞(X) for y ∈ Y. Using 24 JAKA CIMPRI C, YURII SAVCHUK, AND KONRAD SCHM UDGEN assumption (i) of the theorem we obtain for ζ ∈ D∞(X), 0 )ζi ≥ εX keπ2(zm = heπ2(f )eπ2(zm 0 )ζ, ζi = heπ2(zm 0 f zm 0 f zm 0 )ζ, ζi = 0 )ζk2 = εX kπ2(zm hπ2(y)ζ, ζi = hπ2(zm 0 )ζ,eπ2(zm Since π2(y) and π2(z0) are bounded operators and ker π2(z0) = {0} by construction, we conclude that hπ2(y)ζ, ζi > 0 for all ζ ∈ H2, ζ 6= 0. Since π = π0 ⊕ π1 ⊕ π2, it follows from the preceding analysis that hπ(y)ψ, ψi > 0 for all ψ ∈ Hπ, ψ 6= 0. Therefore, by Lemma 7, rXi=1 g = i gi ∈X Y2. The relations (43) imply that in the algebra X each gi ∈ Y can be written gi = fihi, where fi ∈ A and hi is a finite product of elements zj. That is, h−1 i ∈ N ⊆ A. Multiplying both sides of the equation 0 )ζk2 . g∗ g = zm 0 f zm (fihi)∗fihi by (h1h2 . . . hrzm 0 )−1 from the left and from the right we obtain 0 = rXi=1 rXi=1 ef ∗ bf b = i efi ∈ A(q)2, 0 )−1 ∈ A and b = (h1h2 . . . hr)−1 ∈ N . where efi = fihi(h1h2 . . . hrzm gous to the proof of Theorem 2. (cid:3) The next example illustrates the assertion of the strict Positivstellensatz. Its proof is analo- Proposition 8. Suppose that q = 1/2. Then: (i) L := (xx∗)2 − (x + x∗)2 + 3.7 /∈PA2, (ii) (1 + qx∗x)L(1 + qx∗x) ∈PA2. (iii) L satisfies the assumptions (i) and (ii) in Theorem 6, References [C] E.A. Coddington, Formally normal operators having no normal extensions, Canad. J. Math. 17 (1965), 1030 -- 1040. [K] G. Kothe. Topological vector spaces. I. Springer-Verlag New York Inc., New York 1969. [M] M. Marshall. Positive polynomials and sums of squares. Mathematical Surveys and Monographs, 146. Providence, RI, 2008. [OS] V. Ostrovskyı and Yu. Samoılenko. Introduction to the Theory of Representations of Finitely Presented ∗-Algebras. I. Gordon and Breach, London (1999). [Ota] S. Ota, Some classes of q-deformed operators. J. Operator Theory 48 (2002), no. 1, 151186. [OS1] S. Ota, Szafraniec, F.H. q-positive definitness and related topics. J. Math. Anal. Appl. 329 (2007), no. 2, 987 -- 997. [OS2] S. Ota, Szafraniec, F.H. Notes on q-deformed operators. Studia Math. 165 (2004), no. 3, 295 -- 301. ON q-NORMAL OPERATORS AND QUANTUM COMPLEX PLANE. 25 [R] B. Reznick, Uniform denominators in Hilbert's seventeenth problem. Math. Z. 220 (1995), no. 1, 75 -- 97. [RS] M. Reed and B. Simon Methods of modern mathematical physics. I Academic Press Inc., New York, 1980. [Ru] W. Rudin. Functional analysis. McGraw-Hill Book Co., 1973. [S1] Schmudgen K. A strict Positivstellensatz for the Weyl algebra. Math. Ann. 331 (2005), 779 -- 794. [S2] K. Schmudgen, Non-commutative Real Algebraic Geometry - Some Basic Concepts and First Ideas, arXiv:0709.3170. Math. Pures et Appl. 29 (1984), 89 -- 96. [S3] Schmudgen K., Graded and filtrated topological ∗-algebras II. The closure of the positive cone, Rev. Roum. [S4] Schmudgen K. Unbounded operator algebras and representation theory, Birkhauser Verlag, Basel, 1990. [S5] Schmudgen K. Non-commutative moment problems, Math. Z., 206 (1991), 623 -- 650. [SS] Yu. Savchuk, K. Schmudgen. Unbounded induced ∗-representations, arXiv:0806.2428. [W] S. L. Woronowicz, The quantum problem of moments. I. Rep. Mathematical Phys. 1 1970/1971 no. 2, 135 -- 145. University of Ljubljana, Department of Mathematics, Jadranska 19, SI-1000 Ljubljana, Slovenia E-mail address: [email protected] Universitat Leipzig, Mathematisches Institut, Johannisgasse 26, 04103 Leipzig, Germany E-mail address: [email protected] Universitat Leipzig, Mathematisches Institut, Johannisgasse 26, 04103 Leipzig, Germany E-mail address: [email protected]
1805.10936	1	1805	2018-05-28T14:20:47	On irreducible operators in factor von Neumann algebras	[ "math.OA" ]	Let $\mathcal M$ be a factor von Neumann algebra with separable predual and let $T\in \mathcal M$. We call $T$ an irreducible operator (relative to $\mathcal M$) if $W^(T)$ is an irreducible subfactor of $\mathcal M$, i.e., $W^(T)'\cap \mathcal M={\mathbb C} I$. In this note, we show that the set of irreducible operators in $\mathcal M$ is a dense $G_\delta$ subset of $\mathcal M$ in the operator norm. This is a natural generalization of a theorem of Halmos.	math.OA	math	On irreducible operators in factor von Neumann algebras Junsheng Fang, Rui Shi, and Shilin Wen Abstract. Let M be a factor von Neumann algebra with separable predual and let T ∈ M. We call T an irreducible operator (relative to M) if W ∗(T ) is an irreducible subfactor of M, i.e., W ∗(T )′ ∩ M = CI. In this note, we show that the set of irreducible operators in M is a dense Gδ subset of M in the operator norm. This is a natural generalization of a theorem of Halmos. 1. Introduction In [1], Halmos proved the following theorem. Let H be a separable (finite or infinite- dimensional) complex Hilbert space. Then the set of irreducible operators on H is a dense Gδ subset of B(H) in the operator norm. Recall that an operator T ∈ B(H) is irreducible if T has no nontrivial reducing subspaces, i.e., if P is a projection such that P T = T P then P = 0 or P = I. We refer to [3] for a beautiful short proof. In this note, we generalize the above theorem to arbitrary factor von Neumann algebras with separable predual. Let M be a factor von Neumann algebra with separable predual and let T ∈ M. We call T an irreducible operator (relative to M) if W ∗(T ) is an irreducible subfactor of M, i.e., W ∗(T )′ ∩ M = CI. We show that the set of irreducible operators in M is a dense Gδ subset of M in the operator norm. Let A ∈ B(H) and B ∈ B(K), where H, K are Hilbert spaces. For every operator X ∈ B(K, H), we define an operator τA,B is called a Rosenblum operator [5, 4]. τA,B(X) = AX − XB. Lemma 1.1 (Corollary 0.13 of [4]). If σ(A) ∩ σ(B) = ∅, then AX = XB implies X = 0. Another ingredient in the proof of our main result is related to the generator problem of factor von Neumann algebras with separable predual. Precisely, we need the following lemma. Lemma 1.2. Let M be a factor von Neumann algebra with separable predual. Then there exists a singly generated irreducible subfactor N in M. Proof. It is well-known that if M is type I, II∞ or III, then M is singly generated. So the lemma is clear in these cases. When M is a type II1 factor with separable predual, by [2], there exists a hyperfinite irreducible subfactor in M which is singly generated. (cid:3) 2010 Mathematics Subject Classification. Primary 47C15. Key words and phrases. factor von Neumann alegbras, irreducible operators. Junsheng Fang was partly supported by NSFC(Grant No.11431011) and a start up funding from Hebei Normal University. Rui Shi was partly supported by NSFC(Grant No.11401071) and the Fundamental Research Funds for the Central Universities (Grant No.DUT18LK23). 1 2 JUNSHENG FANG, RUI SHI, AND SHILIN WEN 2. Main result Theorem 2.1. Let M be a factor von Neumann algebra with separable predual. Then the set of irreducible operators in M is a dense Gδ subset of M in the operator norm. Proof. Let T ∈ M and ǫ > 0. We need to show that there exists an operator S ∈ M such that kT − Sk < ǫ and S is irreducible relative to M, i.e., if P ∈ M is a projection such that P S = SP , then P = 0 or P = I. Write T = A + iB, where both A and B in M are self-adjoint operators. By the spec- tral theorem for self-adjoint operators, there exist λ1 < λ2 < · · · < λn ∈ σ(A) and projec- tions E1, E2, . . . , En ∈ M such that Pn 4 ǫ. Let A1 = Pn i=1 λiEi and Bij = EiBEj. By the spectral theorem for self-adjoint operators again, there exist ηi1, ηi2, . . . , ηimi ∈ σ(Bii) and projections Fi1, Fi2, . . . , Fimi ∈ M such that Pmi j=1 Fij = Ei and kBii − Pmi j=1 Ej = I and kA − Pn Define T1 = A1 + iB1, where B1 is self-adjoint, defined in the form j=1 λjEjk < 1 j=1 ηijFijk < 1 ii = Pmi 4 ǫ. Let B′ j=1 ηijFij. E2 E1 E1 B′ 11 B12 E2 B21 B′ ... ... En Bn1 Bn2   ... 22 · · · En · · · B1n · · · B2n ... . . . · · · B′ nn   . B1 = Then T1 can be expressed in the form E1 E2 0 λ1 0 λ2 ... ... 0 0 E1 E2 ... En   · · · En · · · 0 · · · 0 ... . . . · · · λn   + T1 = A1 + iB1 = Note that E2 E1 E1 B′ 11 B12 E2 B21 B′ ... ... En Bn1 Bn2   ... 22 · · · En · · · B1n · · · B2n ... . . . · · · B′ nn   . kT − T1k = k(A + iB) − (A1 + iB1)k ≤ kA − A1k + kB − B1k < 1 2 ǫ. (2.1) For 1 ≤ i ≤ n, we can choose real numbers λi1, λi2, . . . , λimi such that (1) the inequality kλiEi − Pmi (2) λ11 < λ12 < · · · < λ1m1 < λ21 < · · · < λ2m2 < · · · < λn1 < · · · < λnmn. Define A2 = Pn j=1 λijFij. Then kA1 − A2k < 1 8 ǫ. Now we make a small self-adjoint perturbation B2 of B1 such that each off-diagonal entry of B2, with respect to the decomposition 8 ǫ holds for every i; j=1 λijFijk < 1 i=1 Pmi I = n mi X i=1 X j=1 Fij, ON IRREDUCIBLE OPERATORS IN FACTOR VON NEUMANN ALGEBRAS 3 is nonzero. That is we can construct a self-adjoint operator B2 in M such that kB2 − B1k < 1 8 ǫ and FijB2Fi′j ′ 6= 0 if i 6= i′ or j 6= j′. Let T2 be defined in the form T2 = A2 + iB2, for F11 F12 0 λ11 0 λ12 ... ... 0 0 F11 F12 ... Fnmn   · · · Fnmn · · · · · · . . . · · · 0 0 ... λnmn   A2 = and B2 = F11 F12 ∗ η11 ∗ η12 ... ... ∗ ∗ F11 F12 ... Fnmn   · · · Fnmn · · · · · · . . . · · · ∗ ∗ ... ηnmn   , where each ∗-entry is nonzero. By applying (2.1), it follows that 3 4 kT − T2k ≤ kT − T1k + kT1 − T2k < ǫ. (2.2) Since M is a separable factor, FijMFij is also a separable factor. By Lemma 1.2, we can find positive elements Xij, Yij ∈ FijMFij such that {Xij, Yij}′′ is an irreducible subfactor of FijMFij. Now we can choose δ > 0 sufficiently small such that the spectra of λijFij + δXij are pairwise disjoint, for 1 ≤ i ≤ n and 1 ≤ j ≤ mi. Let T3 be defined in the form T3 = A3 + iB3, for A3 = F11 F12 ... Fnmn   F11 λ11 + δX11 F12 0 0 ... 0 λ12 + δX12 ... 0 Fnmn · · · · · · · · · . . . · · · λnmn + δXnmn 0 0 ...   and F11 η11 + δY11 F12 ∗ ∗ ... ∗ η12 + δY12 ... ∗ F11 F12 ... Fnmn   · · · · · · · · · . . . · · · B3 = Fnmn ∗ ∗ ... ηnmn + δYnmn   , where each ∗-entry is the same as in T2. Then, clearly, if δ > 0 is small enough, we have kT2 − T3k < 1 4 ǫ. (2.3) Hence, the inequalities (2.2) and (2.3) entail that kT − T3k < ǫ. We assert that T3 is irreducible relative to M. Let P ∈ M be a projection commuting with T3. Then P A3 = A3P and P B3 = B3P . Write P = (Pab)1≤a,b≤k with respect to the decomposition I = Pn i=1 mi. That P A3 = A3P implies that j=1 Fij, where k = Pn (λ11F11 + δX11)P12 = P12(λ12F12 + δX12). i=1 Pmi Since σ(λ11F11 + δX11) ∩ σ(λ12F12 + δX12) = ∅, we have P12 = 0 by Lemma 1.1. Similarly, we have Pab = 0 for 1 ≤ a 6= b ≤ k. Thus P = Pk a=1 Paa is diagonal with respect to the decomposition I = Pn i=1 Pmi j=1 Fij. 4 JUNSHENG FANG, RUI SHI, AND SHILIN WEN By the construction that {Xij, Yij}′′ is an irreducible subfactor of FijMFij, it follows that Paa is either 0 or IFijMFij for each a. Since the off-diagonal entries of B3 are nonzero, an easy calculation shows that, if P11 = 0, then Paa = 0 for 1 ≤ a ≤ k. Therefore, P = 0 or P = I. This proves that T3 is irreducible relative to M. The remainder is to prove the set of irreducible operators relative to M is a Gδ subset of M in the operator norm. The proof is similar to that provided by Halmos. For the sake of completeness, we include the details. Let P be the set of all those selfadjoint operators P in M for which 0 ≤ P ≤ I. Let P0 be the subset of those elements of P that are not scalar multiples of the identity. Since P is a weakly closed subset of the unit ball of M, it is weakly compact, and hence the weak topology for P is metrizable. Since the set of scalars is weakly closed, it follows that P0 is weakly locally compact. Since the weak topology for P has a countable base, the same is true for P0, and therefore P0 is weakly σ-compact. Let P1, P2, . . . be weakly compact subsets of P0 such that ∪∞ n=1Pn = P0. It is to be proved that the set of reducible operators relative to M, denoted by R(M), is an Fσ set in the operator norm topology. Let Pn be the set of all those operators T in M for which there exists a P ∈ Pn such that T P = P T . Then ∪∞ Pn = R(M). The proof can be completed by showing that each Pn is closed in the operator norm. Suppose that Tk ∈ Pn and limk→∞ Tk = T in the operator norm. For each k, find a Pk ∈ Pn such that TkPk = PkTk. Since Pn is weakly compact and metrizable, we may assume that Pk is weakly convergent to P in Pn. Then limk→∞ TkPk = T P and limk→∞ PkTk = P T in the weak operator topology. Hence, T P = P T and T ∈ Pn. This implies that Pn is closed in the operator norm and R(M) is an Fσ set in the operator norm. (cid:3) n=1 References [1] P.R. Halmos. Irreducible operators, Michigan Math J. 15, 1968, 215–223. [2] S. Popa. On a problem of R. V. Kadison on maximal abelian ∗-subalgebras in factors. Invent. Math. 65, 1981/82, no. 2, 269–281. [3] Heydar Radjavi and Peter Rosenthal. Shorter Notes: The Set of Irreducible Operators is Dense. Proceedings of the American Mathematical Society. 21, No. 1 1969, p. 256 [4] Heydar Radjavi and Peter Rosenthal. Invariant Subspaces (second edition). Dover Publications, Inc., Mine- ola, NY, 2003. [5] M.Rosenblum. On the operator equation BX − XA = Q. Duke Math. J. 23, 1956, 263–269. School of Mathematical Sciences, Dalian University of Technology, Dalian, 116024, China Current address: College of Mathematics and Information Science, Hebei Normal University, Shijiazhuang, 050024, China E-mail address: [email protected] School of Mathematical Sciences, Dalian University of Technology, Dalian, 116024, China E-mail address: [email protected], [email protected] School of Mathematical Sciences, Dalian University of Technology, Dalian, 116024, China E-mail address: [email protected]
1709.04597	1	1709	2017-09-14T02:48:05	About the Cuntz Comparison for Non-simple C*-algebras	[ "math.OA", "math.FA" ]	We study the Cuntz semigroup for non-simple $\text{C}^$-algebras in this paper. In particular, we use the extended Elliott invariant to characterize the Cuntz comparison for $\text{C}^$-algebras with the projection property which have only one ideal.	math.OA	math	ABOUT THE CUNTZ COMPARISON FOR NON-SIMPLE C∗-ALGEBRAS GEORGE ELLIOTT AND KUN WANG Abstract. We study the Cuntz semigroup for non-simple C∗-algebras in this paper. In particular, we use the extended Elliott invariant to characterize the Cuntz comparison for C∗- algebras with the projection property which have only one ideal. . A O h t a m [ 1 v 7 9 5 4 0 . 9 0 7 1 : v i X r a 1. Introduction Classification of C∗-algebras by using the so-called Elliott invari- ant has been shown to be very successful in many cases. Recent examples due first to Rørdam and later Toms have shown the cur- rently proposed invariants to be insufficient for the classification of all simple, separable, and nuclear C∗-algebras. There are simple, separable, and nuclear C∗-algebras that can be distinguished by their Cuntz semigroups but not by their Elliott Invariant. So the Cuntz semigroup has become popular and important. In [3], Brown, Perera and Toms recover the Cuntz semigroup for a class of simple C∗-algebras by using the ingredient of Elliott invariant-the Murry-von Neumann semigroup and lower semi- continuous dimension functions. In this paper, we give a charac- terization of Cuntz comparability for a class of C∗-algebras with only one ideal by using the Murry-von Neumann semigroup and lower semi-continuous dimension functions on the C∗-algebra and on its ideal. We will give a example to see that though the ex- tended valued traces can reflect the ideal structure of C∗-algebras, the traces on ideals still have their own meaning. 2000 Mathematics Subject Classification. 46L05. 1 2 GEORGE ELLIOTT AND KUN WANG 2. Cuntz Comparability Let A be a C∗-algebra, and let Mn(A) denote the n × n matrices whose entries are elements of A. Let M∞(A) denote the alge- braic limit of the direct system (Mn(A), φn), where φn : Mn(A) → Mn+1(A) is given by a 7−→(cid:18)a 0 0 0(cid:19) . Let M∞(A)+ (resp. Mn(A)) denote the positive elements in M∞(A) (resp. Mn(A)). For positive elements a and b in M∞(A), write a⊕b to denote the element (cid:18)a 0 0 b(cid:19) , which is also positive in M∞(A). Definition 2.1. Given a, b ∈ M∞(A)+, we say that a is Cuntz subequivalent to b (written a - b) if there is a sequence (vn)∞ n=1 of elements of M∞(A) such that kvnbv ∗ n − ak n→∞ −−−→ 0. We say that a and b are Cuntz equivalent (written a ∼ b) if a - b and b - a. This relation is an equivalence relation, and write hai for the equivalence class of a. The set W(A) := M∞(A)+/ ∼ becomes a positively ordered Abelian monoid when equipped with the operation and the partial order hai + hbi = ha ⊕ bi hai ≤ hbi ⇐⇒ a - b. In the sequel, we refer to this object as the Cuntz semigroup of A. Proposition 2.2. ([[12], Proposition 2.4]) Let A be a C∗-algebra, and let a, b be positive elements in A. The following are equivalent: (i) a - b, (ii) for all ε > 0, (a − ε)+ - b, (iii) for all ε > 0, there exists δ > 0 such that (a − ε)+ - (b − δ)+, ABOUT THE CUNTZ COMPARISON FOR NON-SIMPLE C∗-ALGEBRAS 3 (iv) for all ε > 0, there exists δ > 0 and x in A such that (a − ε)+ = x∗(b − δ)+x. Proposition 2.3. ([15], Proposition 2.2) Let A be a C ∗-algebra. Let a, p ∈ M∞(A)+ be such that p is a projection and p - a. Then there exists b ∈ M∞(A)+ such that p ⊕ b ∼ a. Definition 2.4. Let A be a local C ∗-algebra. A dimension func- tion on A is a mapping D : A → [0, ∞) such that: (i) sup{D(a)a ∈ A} = 1 (normalization). (ii) If a⊥b (i.e., ab = ab∗ = a∗b = a∗b∗ = 0), then D(a + b) = D(a) + D(b). (iii) For all a, D(a) = D(aa∗) = D(a∗a) = D(a∗). (iv) If 0 ≤ a ≤ b, then D(a) ≤ D(b). (v) If a - b (i.e., there exists xn, yn with {xnbyn} converging to a in norm), then D(a) ≤ D(b). Proposition 2.5. ([13], Corollary 4.7) Let A be a C ∗-algebra for which W (A) is almost unperforated (in particular, A could be a Z-absorbing C ∗-algebra). Let a, b be positive elements in A. Sup- pose that a belongs to AbA and such that d(a) < d(b) for every dimension function d on A with d(b) = 1. Then a - b. Lemma 2.6. (see also in [7]) Let A be an exact C ∗-algebra for which W (A) is almost unperforated (in particular, A could be a Z-absorbing C ∗-algebra). Let a, b be positive elements in A. Sup- pose that a belongs to AbA and such that dτ (a) < dτ (b) for every dimension function dτ on A with dτ (b) = 1. Then a - b. Proof. Suppose d is any dimension function on A with d(b) = 1. Let ¯d(hxi) = lim d(hfε(x)i), ε→0 t−ε 0, ε , 1, t ≤ ε ε ≤ t ≤ 2ε t ≥ 2ε. where fε(t) =  Then ¯d is a lower semi-continuous dimension function on A. There- fore, ¯d = dτ for some τ ∈ TA. (This follows from Blackadar and Handelman, [[1], Theorem II.2.2].) 4 GEORGE ELLIOTT AND KUN WANG If ¯d(b) = 0, then a ∈ AbA implies a - b ⊗ 1k for some integer k. Therefore, ¯d(a) = 0 since 0 ≤ ¯d(a) ≤ k ¯d(b) = 0. Hence d(hfε(a)i) ≤ ¯dτ (a) = 0 for any ε > 0. If ¯d(b) 6= 0, then let l , ¯d/ ¯d(b), which is a lower semi-continuous dimension function with l(b) = 1. Then by the assumption d(hfε(a)i)/ ¯d(b) ≤ l(hai) < l(hbi) ≤ d(hbi)/ ¯d(b). Hence d(hfε(a)i) ≤ ¯d(hai) < ¯d(hbi) ≤ d(hbi). In either case, we have d(hfε(a)i) < d(hbi) for ∀ε > 0. Since a ∈ AbA, fε(a) ∈ AbA. Therefore, by Proposition 2.5 fε(a) - b for ∀ε > 0. Hence a - b. (cid:3) Definition 2.7. Let A be a C ∗-algebra with the ideal property. Let A++ be the set of A+ consisting of all positive elements which are not Cuntz equivalent to a non-zero projection in any quotient of A. Lemma 2.8. Let A be an exact C ∗-algebra for which W (A) is almost unperforated (in particular, A could be a Z-absorbing C ∗- algebra). Let a ∈ A++ and p be a projection in A. Then p - a if and only if p ∈ AaA and dτ (p) < dτ (a) for each τ ∈ TA with dτ (a) = 1. Proof. The reverse direction immediately follows from Lemma 2.6. Now suppose p - a, then by Proposition 2.3, there exists a positive element c with p ⊕ c ∼ a. Considering the quotient A/(c), where (c) stands for the ideal generated by c, we have a ∼ p in A/(c). Therefore, 0 = a = p in A/(c). Hence p ∈ (c). If dτ (c) = 0, then dτ (p) = 0, which implies dτ (a) = 0. Therefore, if dτ (a) 6= 0, then dτ (c) 6= 0, hence dτ (p) < dτ (a). (cid:3) ABOUT THE CUNTZ COMPARISON FOR NON-SIMPLE C∗-ALGEBRAS 5 3. Lemma 3.1. Let A be a C∗-algebra such that 0 → I −→ A/I → 0 is a short exact sequence. Let q ∈ A satisfying π(q) is a projection in A/I but q is not a projection in A. Suppose a ∈ A++. Then q - a if and only if (1) dτ (q) < dτ (a) for every τ ∈ TA with ker τ = I, and (2) dτ (q) ≤ dτ (a) for every τ ∈ TA with ker(τ ) = {0}. −→ A ι π Proof. Suppose q - a. Then dτ (q) ≤ dτ (a) for all τ ∈ TA. Let [x] = π(x) be the equivalent class of x in A/I for any x ∈ A. Since [q] - [a] in A/I and [q] is a projection, by Proposition 2.3, there exists b ∈ M∞(A) such that [q] ⊕ [b] ∼ [a]. Since a ∈ A++, [a] is not a projection. So [b] 6= 0. If τ ∈ TA with ker(τ ) = I, then τ is a lower semi-continuous trace on A/I. Therefore, dτ ([q]) + dτ ([b]) = dτ ([a]). Thus, dτ (q) < dτ (a) for τ ∈ TA with ker τ = I. Now suppose the condition (1) and (2) hold. If τ ∈ TA satis- fying ker(τ ) = 0, then τ is a faithful trace on A. Since q is not a projection in A, 0 ∈ σ(q) is not an isolated point. For any ε > 0, there exists δ ∈ σ(q) such that dτ ((q − ε)+) < dτ ((q − δ)+). Thus dτ ((q − ε)+) < dτ ((q) ≤ dτ (a) for all τ ∈ TA with ker τ = {0}. Therefore, dτ ((q − ε)+) < dτ (a) for all τ ∈ TA. By Lemma 2.6, Thus, q - a. (q − ε)+ - a for all ε > 0. Remark 3.2. For the second part of above proof, a can be any positive element. That is, we do not require a that belong to A++. (cid:3) 6 GEORGE ELLIOTT AND KUN WANG Lemma 3.3. Let A be an exact, Z-stable C∗-algebra. Suppose A has only one non-trivial ideal I, which is also exact and Z-stable. Let a ∈ A++ and b ∈ A+. Then a - b if and only if dτ (a) ≤ dτ (b) for every τ ∈ TA, Proof. The sufficiency is clear. If a = 0, then the conclusion is true. So we assume a 6= 0. If a, b ∈ I, then for any τ ∈ TI, τ can be extended to an element in TA. Thus, dτ (a) ≤ dτ (b) for every τ ∈ TA implies dτ (a) ≤ dτ (b) for every τ ∈ TI. Since I is simple, a - b follows from the Proposition 2.6 of [15]. Since a ∈ A++ and a 6= 0, there is a strictly decreasing sequence εn of positive real numbers in σ(a) converge to 0. If a ∈ I, b /∈ I, then dτ ((a − εn)+) ≤ dτ (a) = 0 < dτ (b) for all τ ∈ TA with ker τ = I and all n ∈ N. For τ ∈ TA with ker τ = {0}, we have dτ ((a − εn)+) < dτ (a) ≤ dτ (b) for all n ∈ N. Therefore, in this case we can get dτ ((a − εn)+) < dτ (b) for all τ ∈ TA, and all n ∈ N. By Lemma 2.6, (a − εn)+ - b for all n ∈ N. Since the set {x ∈ A+x - b} is closed, and since (a − εn) → a in norm, we have a - b. If a /∈ I, and b /∈ I, for τ ∈ TA with ker τ = {0}, we have dτ ((a − εn)+) < dτ (a) ≤ dτ (b) for all n ∈ N. For any τ ∈ TA with ker τ = I, then τ is a lower semi-continuous trace on A/I and dτ ([a]) ≤ dτ ([b]). Since a ∈ A++ and [a] 6= 0, there is a strictly decreasing sequence δm of positive real numbers in σ([a]) converge to 0. Since A/I is simple, every trace on A/I is faithful. dτ (([a] − δn)+) < dτ ([a]) ≤ dτ ([b]) for all τ ∈ T(A) with ker τ = I and all m ∈ N. Therefore, dτ ((a − δn)+) < dτ (b) for all τ ∈ T(A) with ker τ = I and all m ∈ N. Since both {εn} and {δm} are sequences converging to 0, there exist subsequence ABOUT THE CUNTZ COMPARISON FOR NON-SIMPLE C∗-ALGEBRAS 7 nk and mk such that εn1 ≥ δm1 ≥ εn2 ≥ δm2 · · · . Without loss of generality, we can assume nk = mk = k. Therefore, dτ ((a − εn)+) < dτ (b) for all τ ∈ TA and all n ∈ N. Thus, (a − εn)+ - b for all n ∈ N. So a - b. (cid:3) 4. Lemma 4.1. Let A be a C∗-algebra and x, y ∈ A be two positive elements. For any ε > 0, there exists δ > 0 such that: if kx 1 2 y 1 2 k < min{δ, δ kx + yk , 1}, Proof. Since (cid:18)x2 0 δ > 0 such that 0 y(cid:19) , for any ε > 0, there exists then Let Then since 0 0 0 0 [(cid:18)x 0 0 y(cid:19) − ε]+ -(cid:18)x + y 0 0(cid:19) . y2(cid:19) ∼ (cid:18)x 0 [(cid:18)x 0 0 y(cid:19) − ε]+ - [(cid:18)x2 ω =(cid:18)x 2(cid:19)(cid:18)x + y 0 =(cid:18) x2 + x kw −(cid:18)x2 =k(cid:18) x y2(cid:19) − δ]+. 2(cid:19) 0(cid:19)(cid:18)x 2 + y2 (cid:19) . 2 + x 2 y 2 xy y 2 y x y 2 xy 0 y2(cid:19) k 2 yx 2 + y y 2 x 2 y 2 x 1 2 yx 2 x 3 2 + x 2 + y 2 x 2 x 2 x 2 y 2 y x 1 2 1 1 2 1 1 2 1 0 0 1 1 1 1 2 3 3 1 1 3 1 1 1 2 1 2 y y y y 1 3 3 2 1 3 1 2 1 1 2 1 1 3 2 (cid:19) k < δ, kx 1 2 yx 1 2 k = k(x 1 2 y 1 2 )(x 1 2 y 1 2 )∗k = kx 1 2 y 1 2 k2 < δ, 8 GEORGE ELLIOTT AND KUN WANG kx 3 2 y 1 2 + x 1 2 y 3 2 k = kx 1 2 (x + y)y 1 2 k = k(x + y) 1 2 y 1 2 x 1 2 (x + y) 1 2 k ≤k(x + y) 1 2 k2kx 1 2 y 1 2 k = kx + ykkx 1 2 y 1 2 k < δ. Therefore, [(cid:18)x 0 0 y(cid:19) − ε]+ - [(cid:18)x2 0 0 y2(cid:19) − δ]+ - ω -(cid:18)x + y 0 0(cid:19) . 0 (cid:3) Lemma 4.2. Let A be a C∗-algebra such that 0 → I ι −→ A π −→ A/I → 0 is a short exact sequence, where I is a closed two-sided ideal of A. Suppose {pn} is a quasi-central approximate unit of I consisting of projections. Let a, b be two positive elements in A. Then [π(a)] - [π(b)] in A/I if and only if, for any ε > 0, there exists an integer N > 0 such that [(1 − pn)a(1 − pn) − ε]+ - (1 − pn)b(1 − pn) for all n ≥ N. Proof. Necessity. If for any ε > 0, there exists an integer N > 0 such that [(1 − pn)a(1 − pn) − ε]+ - (1 − pn)b(1 − pn) for all n ≥ N. Then (π(a) − ε)+ = (π((1 − pN )a(1 − pN )) − ε)+ = π((1 − pN )a(1 − pN ) − ε)+ - π((1 − pN )b(1 − pN )) = π(b). Since the above relation holds for all ε > 0, by Proposition 2.2, π(a) - π(b). Sufficiency. if [π(a)] - [π(b)] in A/I, then there exist {vk}∞ k=1 in A, such that [π(a)] = lim [π(vk)][π(b)][π(vk)]∗. k→∞ ABOUT THE CUNTZ COMPARISON FOR NON-SIMPLE C∗-ALGEBRAS 9 Given ε > 0, we can find k0 such that k[π(a)] − [π(vk0)][π(b)][π(vk0)]∗k < ε/2. Thus there exists d ∈ A such that ([π(a)] − ε/2)+ = [π(d)][π(b)][π(d)]∗. Since [π(a)] = [π((1 − pn)a(1 − pn))] and [π(b)] = [π((1 − pn)b(1 − pn))] for all n ∈ N, we can find cn ∈ I such that ((1 − pn)a(1 − pn) − ε/2)+ = d(1 − pn)b(1 − pn)d∗ + cn. Multiplying pn from both left and right sides of the above equation, we get pn((1−pn)a(1−pn)−ε/2)+pn = pnd(1−pn)b(1−pn)d∗pn +pncnpn. Since pn is a quasi central approximate unit of I, lim n kpn((1 − pn)a(1 − pn) − ε/2)+pnk = 0, lim n kpnd(1 − pn)b(1 − pn)d∗pnk = 0. Therefore, lim a natural number N such that for all n ≥ N, kpncnpnk = 0. Thus, lim kcnk = 0. Then we can find n n k((1 − pn)a(1 − pn) − ε/2)+ − d(1 − pn)b(1 − pn)d∗k < ε/2. Thus (((1 − pn)a(1 − pn) − ε/2)+ − ε/2)+ - d(1 − pn)b(1 − pn)d∗ for all n ≥ N. That is, ((1 − pn)a(1 − pn) − ε)+ - d(1 − pn)b(1 − pn)d∗ for all n ≥ N. (cid:3) Lemma 4.3. Let A be a separable C∗-algebra and I be an ideal of A. Suppose {pn}∞ n=1 is a quasi-central approximate unit of I. If a, b are two positive elements in A satisfying π(a) - π(b) and pnapn - pnbpn for all sufficiently large n. Then a - b. 10 GEORGE ELLIOTT AND KUN WANG Proof. By assumption, we know that a = (a − pnapn) + pnapn 0 -(cid:18)a − pnapn -(cid:18)b − pnbpn (a − ε)+ - [(cid:18)b − pnbpn pnapn(cid:19) pnbpn(cid:19) pnbpn(cid:19) − ε′]+. 0 0 0 0 0 (∗) For any ε > 0, there exists ε′ > 0 such that Since {pn} is a quasi-central approximate unit for I, (b − pnbpn)pnbpn = (1 − pn)bpnbpn → 0. For ε′ > 0, applying Lemma 4.1, there exists δ and N ∈ N such that: for all n ≥ N we have Therefore, k(b − pnbpn)pnbpnk < min{δ, δ kbk , 1}. [(cid:18)b − pnbpn 0 0 pnbpn(cid:19) − ε′]+ -(cid:18)b 0 0 0(cid:19) . for all n ≥ N. Combining with (∗), we get (a − ε)+ - b for all ε > 0. Thus, a - b. (cid:3) 5. For a C∗-algebra A with one ideal I which has the projection property, we define fW (A) = V (A) ⊔ LAf f (T (A))++ ⊔ LAf f (T (I))++ and a map ι : W (A) →fW (A) ABOUT THE CUNTZ COMPARISON FOR NON-SIMPLE C∗-ALGEBRAS11 by ι(a) =  [a] d·(a) d·(a) if a ∈ V (A) if a ∈ A++ if a ∈ P(A/I) Theorem 5.1. ι is an isomorphism. References [1] B. Blackadar and D. Handelman, Dimension functions and traces on C ∗- algebras, J. Funct. Anal. 45(1982), 297-340. [2] J. Bosa, G. Tornetta and J. Zacharias, Open projections and suprema in the Cuntz semigroup. [3] N. P. Brown, F. Perera, A. S. Toms, The Cuntz semigroup, the Elliott conjecture, and dimension functions on C∗- algebras, J. reine angew. Math. 621 (2008), 191211. [4] G. A. Elliott, A classification of certain simple C ∗-algebras, Quantum and Non- Commutative Analysis (editors, H. Araki et al.), Kluwer, Dordrecht, (1993), pp. 373-385. [5] G. A. Elliott and G. Gong, On the classification of C ∗-algebras of real rank zero. II. Ann. of Math. 144 (1996), 497-610. [6] G. A. Elliott, G. Gong, and L. Li, On the classification of simple inductive limit C ∗- algebra II. The isomorphism theorem. Invent. Math. 168 (2007), 249-320. [7] G. A. Elliott, L. Robert and L. Santiago, The cone of lower semicontin- uous traces on a C ∗-algebra, American Journal of Mathematics, Vol 133, No. 4, Aug. 2011, 969-1005. [8] K. Ji and C.L. Jiang, A complete classification of AI algebras with the ideal property, Canad. J. Math. 63(2011), 381-412. [9] C. Jiang and K. Wang, A complete classification of limits of splitting interval algebras with the ideal property, J. Ramanujan Math. Soc. 27, No.3 (2012) 305-354. [10] E. Kirchberg and M. Rordam, Infinite non-simple C*-algebras: absorbing the Cuntz algebra O∞, Adv. Math. 167(2002), 195-264. [11] Huaxin Lin, An Introduction to the classification of amenable C ∗- algebras, World Scientific. [12] M. Rørdam, On the structure of simple C ∗-algebras tensored with a UHF- algebra. II. J. Funct. Anal. 107(1992), 255-269. [13] M. Rørdam, The stable and the real rank of Z-absorbing C ∗-algebras, Int. J. Math. 15(2004), 1065-1084. [14] G. K. Pedersen, C ∗-algebras and their automorphism groups, Academic Press INC. (LONDON) LTD. 12 GEORGE ELLIOTT AND KUN WANG [15] F. Perera and A. S. Toms, Recasting the Elliott Conjecture, Math. Ann. 338, (2007), pp. 669-702. [16] K. H. Stevens, The classification of certain non-simple approximate in- terval algebras, Fields Institute Communications 20 (1998), 105-148.
1603.06084	1	1603	2016-03-19T11:45:53	Integrable actions and quantum subgroups	[ "math.OA", "math.QA" ]	We study homomorphisms of locally compact quantum groups from the point of view of integrability of the associated action. For a given homomorphism of quantum groups $\Pi\colon\mathbb{H}\to\mathbb{G}$ we introduce quantum groups $\mathbb{H}/\!\ker{\Pi}$ and $\overline{\mathrm{im}\,\Pi}$ corresponding to the classical quotient by kernel and closure of image. We show that if the action of $\mathbb{H}$ on $\mathbb{G}$ associated to $\Pi$ is integrable then $\mathbb{H}/\!\ker\Pi\cong\overline{\mathrm{im}\,\Pi}$ and characterize such $\Pi$. As a particular case we consider an injective continuous homomorphism $\Pi\colon{H}\to{G}$ between locally compact groups $H$ and $G$. Then $\Pi$ yields an integrable action of $H$ on $L^\infty\;\!\!(G)$ if and only if its image is closed and $\Pi$ is a homeomorphism of $H$ onto $\mathrm{im}\,\Pi$. We also give characterizations of open quantum subgroups and of compact quantum subgroups in terms of integrability and show that a closed quantum subgroup always gives rise to an integrable action. Moreover we prove that quantum subgroups closed in the sense of Woronowicz whose associated homomorphism of quantum groups yields an integrable action are closed in the sense of Vaes.	math.OA	math	INTEGRABLE ACTIONS AND QUANTUM SUBGROUPS PAWE L KASPRZAK, FATEMEH KHOSRAVI, AND PIOTR M. SO LTAN Abstract. We study homomorphisms of locally compact quantum groups from the point of view of integrability of the associated action. For a given homomorphism of quantum groups Π : H → G we introduce quantum groups H/ker Π and im Π corresponding to the classical quotient by kernel and closure of image. We show that if the action of H on G associated to Π is integrable then H/ker Π ∼= im Π and characterize such Π. As a particular case we consider an injective continuous homomorphism Π : H → G between locally compact groups H and G. Then Π yields an integrable action of H on L∞(G) if and only if its image is closed and Π is a homeomorphism of H onto im Π. We also give characterizations of open quantum subgroups and of compact quantum sub- groups in terms of integrability and show that a closed quantum subgroup always gives rise to an integrable action. Moreover we prove that quantum subgroups closed in the sense of Woronow- icz whose associated homomorphism of quantum groups yields an integrable action are closed in the sense of Vaes. 1. Introduction The theory of locally compact quantum groups is by now a well established branch of mathe- matical research. Once the details of the definition of a locally compact quantum group have been worked out in a series of papers (cf. e.g. [16, 17, 18]) and a lot of work has been done on examples of such objects ([20, 32, 36, 21, 26, 31, 30]), a number of deep results of the theory of locally compact groups have been generalized and analyzed from the point of view of the new theory of quantum groups ([28, 15, 29]). These papers pointed the way to a more thorough analysis of such basic concepts as subgroups ([4]) or homomorphisms ([19]) as well as actions ([28], [27, Chapter 2], see also [22]). In this paper we continue this line of research by analyzing the concept of an integrable action in the context of homomorphisms of quantum groups. We show that integrability of actions associated to homomorphisms is deeply connected with the notion of a closed quantum subgroup. For a homomorphism of quantum groups we introduce and study quantum groups which are non- commutative analogs of the closure of the image of the homomorphism and the quotient by the kernel of the homomorphism. Integrability of the associated action is then equivalent to a condition which can be interpreted as compactness of the kernel and closeness of the image together with a natural isomorphism of the image with quotient by kernel. These considerations provide a way to show that the difference between the two notions of a closed quantum subgroup analyzed in [4] lies precisely in the integrability of the associated action. In particular a Woronowicz-closed quantum subgroup (see the end of Section 6 or [4, Section 3]) with its associated action integrable is closed (in the sense of Vaes, see [4]). These results have an overlap with [5, Proposition 3.12] with also practically the same technique of proof. We also provide characterization of open and compact quantum subgroups in terms of integra- bility of certain natural actions associated with them (but not with the homomorphism mapping the quantum subgroup into the ambient quantum group). The former is then used in other re- sults on integrability, while the latter is of decidedly different nature and is hence placed in the Appendix. Our results are also of interest in the case the considered quantum groups are in fact classical In particular, if H and G are locally compact groups and locally compact quantum groups. 2010 Mathematics Subject Classification. Primary: 46L89 Secondary: 46L85, 46L52. Key words and phrases. Locally compact quantum group, quantum subgroup, homomorphism of quantum groups, integrable action. 1 2 PAWE L KASPRZAK, FATEMEH KHOSRAVI, AND PIOTR M. SO LTAN Π : H → G is an injective continuous homomorphism then the integrability of the associated action of H on L∞(G) is equivalent to Π being a homeomorphism onto its closed image. Let us give a more detailed account of the contents of the paper. In Section 2 we collect the main definitions and preliminary results about operator algebras and locally compact quantum groups. We discuss in detail the notions of actions, homomorphisms and closed quantum subgroups of locally compact quantum groups. We also discuss the notion of the canonical implementation of an action ([28]) which is one of our crucial tools. Section 3 is devoted to the proof of an integrability criterion for ergodic actions. This is very much in the spirit of some of the results of [15], but the type of actions considered is different. In Section 4 we introduce the quantum groups which play the role of closure of the image and quotient by kernel of a homomorphism of locally compact quantum groups. We also generalize the latter construction to introduce quotient by the kernel of an action (an example of this construction already appeared in literature). The short Section 5 sheds light on the canonical implementation of the action associated to a homomorphism of quantum groups and recalls an integrability criterion from [28]. Then, in Section 6 all of our main results are obtained. The Appendix (Section 7) contains discussion of consequences of our results for classical groups and a theorem characterizing compact quantum subgroups of locally compact quantum groups in terms of integrability of the canonical action on the homogeneous space. We also recall several facts about convolutions of various classes of functionals on C∗-algebras and von Neumann alge- bras associated to a quantum group and prove a few results necessary for the above mentioned characterization of compact quantum subgroups. Some of these results are also used earlier (in Section 5). 2. Preliminaries We will use the language and notation of operator algebras for locally compact quantum groups. The operator algebraic prerequisites are contained e.g. in [24]. Some notions related to C∗-algebras like morphisms or multipliers may be found e.g. in [33, 34]. We will also use the concept of an operator valued weight ([7]) in the context of invariant weights on locally compact quantum groups ([27, 28, 17]). The symbol σ will always denote the flip morphism between tensor product of operator algebras. Almost all tensor products will be denoted by the symbol ⊗ with the precise meaning depending on the context. The only exception will be the tensor product of von Neumann algebras which we will denote by ¯⊗ . 2.1. Locally compact quantum groups. We refer to [16, 17, 18] for fundamentals of the theory of locally compact quantum groups. Our conventions are those introduced in e.g. [1, 35, 23, 4]. Thus a locally compact quantum group G is described by a von Neumann algebra L∞(G) with comultiplication ∆G : L∞(G) → L∞(G) ¯⊗ L∞(G). The right Haar weight on L∞(G) will be denoted by h with corresponding GNS Hilbert space denoted by L2(G) and associated GNS map η ([18, Appendix B]). The modular operator and conjugation for h will be denoted by ∇ and J respectively. We will also denote by L1(G) the space L∞(G)∗ of σ-weakly continuous functionals on L∞(G). The left Haar weight will be denoted by hL. The symbols C0(G) and Cu 0(G) will denote the reduced and universal C∗-algebra related to G. Since C0(G) ⊂ L∞(G) we will use ∆G to also denote the comultiplication on C0(G), so we have ∆G ∈ Mor(C0(G), C0(G) ⊗ C0(G)), while the comultiplication on Cu 0(G) will be denoted by 0(G) are endowed with much more natural structure. In ∆u particular there is the scaling group (τt)t∈R and unitary antipode R ([35, 16, 23]). G. The algebras L∞(G), C0(G) and Cu The Kac-Takesaki operator or the right regular representation of G is the unitary operator WG which extends the map to the level of L2(G) ⊗ L2(G). η(a) ⊗ η(b) 7−→ (η ⊗ η)(cid:0)∆(a)(1 ⊗ b)(cid:1) It turns out that WG is a multiplicative unitary operator on L2(G) ⊗ L2(G) and G is the object related to this multiplicative unitary via the theory developed in [1, 35]. Thus, in particular ◮ C0(G) =(cid:8)(ω ⊗ id)WG ω ∈ B(L2(G))∗(cid:9)norm closure , INTEGRABLE ACTIONS AND QUANTUM SUBGROUPS 3 , ◮ for any x ∈ L∞(G) we have ∆G(x) = WG(x ⊗ 1)WG∗ ◮ L∞(G) =(cid:8)(ω ⊗ id)WG ω ∈ B(L2(G))∗(cid:9)σ-weak closure The multiplicative unitary gives rise to the dual bG of G which turns out to also be a locally compact quantum group. In particular there is a right Haar weight bh on L∞(bG), where The GNS Hilbert space for bh can and will be naturally identified with L2(G) and in fact W σ(cid:0)WG ∗(cid:1) and WG ∈ M(C0(bG) ⊗ C0(G)). Moreover, denoting by b∇ and bJ the modular operator and conjugation related to bh we have L∞(bG) =(cid:8)(id ⊗ ω)WG ω ∈ B(L2(G))∗(cid:9)σ-weak closure x ∈ L∞(G), bG = . . R(x) = bJx∗bJ, τt(x) = b∇itxb∇−it, x ∈ L∞(G), t ∈ R. 0(G)) such that if Λ bG ∈ Mor(Cu M(Cu reducing morphisms ([23, Definition 35]) then WG = (Λ bG ⊗ ΛG)VV VVG are WG = (id⊗ΛG)VVG ∈ M(Cu The results of [14, 23, 19] provide a lift of WG ∈ M(C0(bG) ⊗ C0(G)) to an element VVG ∈ 0(bG) ⊗ Cu 0(G), C0(G)) are G. The "half-lifted" versions of 0(G)). Let G be a locally compact quantum group and let L ⊂ L∞(G) be a von Neumann subalgebra. We say that L is a left coideal if ∆G(L) ⊂ L∞(G) ¯⊗ L. If, moreover, L satisfies ∆G(L) ⊂ L ¯⊗ L then we say that L is an invariant subalgebra (this terminology was introduced in [25]). Finally an invariant subalgebra L such that R(L) = L and τt(L) = L for all t ∈ R is called a Baaj-Vaes subalgebra. The important Baaj-Vaes theorem ([2, Proposition 10.5]) says that if L ⊂ L∞(G) is a Baaj-Vaes subalgebra then there exists a locally compact quantum group K such that L = L∞(K) 0(bG))⊗C0(G)) and WG = (Λ bG ⊗id)VVG ∈ M(C0(bG)⊗Cu 0(bG), C0(bG)) and ΛG ∈ Mor(Cu and ∆K = ∆G(cid:12)(cid:12)L. If L ⊂ L∞(G) is a left coideal then by [12, Proposition 3.5] the relative commutant L′ ∩ L∞(bG) is a left coideal in L∞(bG) called the co-dual of L and is denoted by eL. By [12, Theorem 3.9] we have eeL = L for any left coideal L ⊂ L∞(G). 2.2. Actions of quantum groups. Let H be a locally compact quantum group. A right action of H on a von Neumann algebra N is an injective normal unital ∗-homomorphism α : N → N ¯⊗ L∞(H) such that (α ⊗ id) ◦ α = (id ⊗ ∆H) ◦ α. Since most of the time we will only use right actions, the term "action" will always refer to a right action. Similarly to a right action we define a left action of H on N as an injective normal unital ∗-homomorphism β : N → L∞(H) ¯⊗ N such that (id ⊗ β) ◦ β = (∆H ⊗ id) ◦ β. Let α : N → N ¯⊗ L∞(H) be a (right) action of H on N. The crossed product of N by (the action of) H is defined as the von Neumann subalgebra of N ¯⊗ B(L2(H)) generated by α(N) and 1 ¯⊗ L∞(bH). The crossed product will be denoted by the symbol N ⋊α H. Let us note that there is a simple passage from right to left actions (and conversely). Indeed, if α : N → N ¯⊗ L∞(H) is a right action of H on N then β = σ ◦ α : N → L∞(H) ¯⊗ N is a left action of Hop, i.e. the locally compact quantum group H with reversed comultiplication ∆Hop = σ ◦ ∆H. This allows simple translation of many results about right actions to left actions. The central notion of this paper is that of an integrable action. The concept of integrability of an action of a locally compact group was introduced in [3, Definition 2.6] for locally compact groups. The definition can easily be generalized to Kac algebras and locally compact quantum groups as was done e.g. in [27, Definition 2.3.4], [28, Definition 1.4]. Thus an action α : N → N ¯⊗ L∞(H) is integrable if the set (2.1) is σ-weakly dense in N+. Elements of (2.1) are called integrable for α. Similarly an element x ∈ N is square integrable for α if x∗x is integrable for α. Note that integrability of a right action is with respect to the left Haar weight. Similarly we say that a left action β : N → L∞(H) ¯⊗ N is integrable if the set (cid:8)x ∈ N+ (id ⊗ hL)α(x) ∈ N+(cid:9) (cid:8)x ∈ N+ (h ⊗ id)β(x) ∈ N+(cid:9) 4 PAWE L KASPRZAK, FATEMEH KHOSRAVI, AND PIOTR M. SO LTAN is σ-weakly dense in N+. Let us recall briefly some of the results of [28]. If α : N → N ¯⊗ L∞(H) is an action of H on N then choosing a n.s.f. weight θ on N one can define a weight eθ on N ⋊α H ([28, Definition 3.4 and following remarks]) such that the corresponding GNS Hilbert space is Hθ ⊗ L2(H). Let eJ be the modular conjugation related to this weight. Then the canonical implementation of α is by definition the unitary Some of the major results of [28] are that ◮ U ∈ B(Hθ) ¯⊗ L∞(H), ◮ (id ⊗ ∆H)U = U12U13, ◮ for x ∈ N we have α(x) = U (x ⊗ 1)U ∗. U = eJ(Jθ ⊗ bJ). The last formula says that U implements the action α. Finally let us introduce two more important classes of actions: an action α : N → N ¯⊗ L∞(H) of a locally compact quantum group H on a von Neumann algebra N is ergodic if the condition α(x) = x ⊗ 1 implies that x ∈ C1. The action α is free if the set generates the von Neumann algebra L∞(H). (cid:8)(ω ⊗ id)α(x) x ∈ N, ω ∈ N∗(cid:9) 2.3. Homomorphism of quantum groups. Let H and G be locally compact quantum groups. The notion of a homomorphism from H to G can be described in one of three ways. More precisely there are natural bijections between the following three sets ◮ the set of bicharacters from H to G, i.e. unitaries V ∈ L∞(bG) ¯⊗ L∞(H) such that and (∆bG ⊗ id)V = V23V13, (id ⊗ ∆H)V = V12V13 ◮ the set of actions α : L∞(G) → L∞(G) ¯⊗ L∞(H) such that (∆G ⊗ id) ◦ α = (id ⊗ α) ◦ ∆G (such actions are called right quantum group homomorphisms, cf. [19, Section 1]), ◮ the set of those π ∈ Mor(Cu 0(G), Cu 0(H)) which satisfy ∆u H ◦ π = (π ⊗ π) ◦ ∆u G (such morphisms are referred to as Hopf ∗-homomorphisms). Each element of either of the three sets described above represents a homomorphism from H to G. Thus from now on we will write Π : H → G to denote a homomorphism from H to G and freely use its three "incarnations", namely a bicharacter V ∈ L∞(bG) ¯⊗ L∞(H), a right quantum group homo- morphism α : L∞(G) → L∞(G) ¯⊗ L∞(H) and a Hopf ∗-homomorphism π ∈ Mor(Cu The relationships between V , α and π representing the same Π are 0(G), Cu 0(H)). V =(cid:0)id ⊗ (ΛH ◦ π)(cid:1)W G, α ◦ ΛG = (ΛG ⊗ ΛH) ◦ (id ⊗ π) ◦ ∆u G and α(x) = V (x ⊗ 1)V ∗, x ∈ L∞(G). (2.2) 2.4. Closed quantum subgroups and open quantum subgroups. Let G and H be locally compact quantum groups. Following [29, Definition 2.5] (cf. [4] and the discussion therein) we say that a homomorphism of quantum groups Π : H → G identifies H with a closed quantum subgroup of G if there exists an injective normal ∗-homomorphism γ : L∞(bH) → L∞(bG) such that the corresponding bicharacter V is given by (γ ⊗ id)WH. Equivalently one can only demand that γ satisfy ∆bG ◦ γ = (γ ⊗ γ) ◦ ∆bH ([4, Theorem 3.3]). It follows from [16, Proposition 5.45] that if H is a closed subgroup of G then the image of L∞(bH) under the injection γ is a Baaj-Vaes subalgebra of L∞(bG). In particular this shows that there is a bijection between closed quantum subgroups of G and Baaj-Vaes subalgebras of L∞(bG). INTEGRABLE ACTIONS AND QUANTUM SUBGROUPS 5 Now let G and H be locally compact quantum groups. According to [9, Definition 2.2] we say that H is an open quantum subgroup of G if there is a surjective normal ∗-homomorphism Θ : L∞(G) → L∞(H) such that ∆H ◦ Θ = (Θ ⊗ Θ) ◦ ∆G. In this case the algebra L∞(H) can be identified with a corner of L∞(G) defined by the central support of Θ which we denote by the symbol 1H. Let us denote the map L∞(H) → 1H L∞(G) ⊂ L∞(G) by ι. The composition ι ◦ Θ is the conditional expectation L∞(G) ∋ x 7−→ 1Hx ∈ L∞(G). By [9, Theorem 3.6] an open quantum subgroup of G is a closed quantum subgroup of G. In fact the corresponding bicharacter V is (id ⊗ Θ)WG (cf. [9, Remark 3.7]) and the image of L∞(bH) in L∞(bG) under the corresponding injection is the σ-weak closure of (cid:8)(id ⊗ ω)(id ⊗ Θ)WG ω ∈ B(L2(H))∗(cid:9). 3. Integrable ergodic actions and open quantum subgroups In this section we will prove a result on integrability of an action (Proposition 3.2) which will enable us to give a characterization of open subgroups in terms of existence of certain integrable elements (Theorem 3.3). This result will in turn become useful in Section 6. The following lemma is a generalization of [16, Lemma 6.4], cf. also [19, Theorem 2.1]. The technique of proof is virtually identical to that of proof of [16, Lemma 6.4]. Lemma 3.1. Let α : N → N ¯⊗ L∞(G) be an ergodic action of a locally compact quantum group G on a von Neumann algebra N. If P ∈ N is a projection such that α(P ) ≤ P ⊗ 1 then P = 0 or P = 1. Proof. Since α is an injective map, N can be embedded into N ¯⊗ L∞(G) via this mapping. Viewing α as this embedding we shall denote it by ι. Using this notation we have (id ⊗ ∆G) ◦ ι = (id ⊗ ∆G) ◦ α = (α ⊗ id) ◦ α = (ι ⊗ id) ⊗ α which shows that under the identification of N with ι(N) the action α is given by id ⊗ ∆G. In particular, as α is ergodic, id ⊗ ∆G is ergodic on ι(N). Therefore it is enough to prove the lemma for (id ⊗ ∆G)(cid:12)(cid:12)ι(N). Denote ι(P ) by Q. The condition that (id ⊗ ∆G)(Q) ≤ Q ⊗ 1 means (id ⊗ ∆G)(Q)Q12 = (id ⊗ ∆G)(Q) or 23Q12WG 23 WG ∗ Q12 = WG 23Q12WG 23 ∗ , which can be rewritten as WG 23Q12 = Q12WG 23Q12 (3.1) (3.2) Let b∇ and bJ be the modular operator and modular conjugation for bh -- the right Haar weight of 2(cid:1) and all ζ ∈ Dom(cid:0)b∇ 1 2(cid:1) bG. By [18, Lemma 2.1 & Theorem 3.5] applied to bG, for all ξ ∈ Dom(cid:0)b∇− 1 we have = (id ⊗ ω bJ b∇− 1 2 ξ, bJ b∇ 1 2 ζ )WG. (3.3) (cid:0)(id ⊗ ωξ,ζ)WG(cid:1)∗ Therefore for such ξ and ζ we can apply id ⊗ id ⊗ ω bJ b∇− 1 2 ξ, bJ b∇ 1 2 ζ to both sides of (3.2) to get (cid:0)1 ⊗ (id ⊗ ω bJ b∇− 1 2 ξ, bJ b∇ 1 2 ζ )(WG)(cid:1)Q = Q(cid:0)1 ⊗ (id ⊗ ω bJ b∇− 1 2 ξ, bJ b∇ )(WG)(cid:1)Q 1 2 ζ which by (3.3) is equivalent to It follows that Q12WG 23 = Q12WG 23Q12 which means that Q12 = Q12(id ⊗ ∆G)(Q), i.e. Q ⊗ 1 ≤ (id ⊗ ∆G)(Q). (3.4) or Q = Q(cid:0)1 ⊗ (id ⊗ ωξ,ζ)(WG)(cid:1)∗ (cid:0)1 ⊗ (id ⊗ ωξ,ζ)(WG)(cid:1)∗ Q(cid:0)1 ⊗ (id ⊗ ωξ,ζ)(WG)(cid:1) = Q(cid:0)1 ⊗ (id ⊗ ωξ,ζ)(WG)(cid:1)Q. Q i.e. (id ⊗ ω)α(x) is integrable. Now for any y ∈ N and ω ∈ L1(G) as above the Kadison inequality gives 1 (id ⊗ hL)(cid:16)α(cid:0)(id ⊗ ω)α(x)(cid:1)(cid:17) =(cid:13)(cid:13)δ (cid:0)(id ⊗ ω)α(y)(cid:1)∗(cid:0)(id ⊗ ω)α(y)(cid:1) ≤ kωk(id ⊗ ω)α(y ∗ y) 2 ξ(cid:13)(cid:13)2 (id ⊗ hL)(α(x)). 6 PAWE L KASPRZAK, FATEMEH KHOSRAVI, AND PIOTR M. SO LTAN Combining (3.1) and (3.4) we find that (id ⊗ ∆G)(Q) = Q ⊗ 1, which by ergodicity of (id ⊗ ∆G) (cid:3) on N implies that Q = 0 or Q = 1. The next proposition is very similar to [15, Proposition 6.2]. The difference is that the action considered in [15] is the action of a closed quantum subgroup on the ambient quantum group. In particular it is not ergodic, but has a particularly simple form. Proposition 3.2. Let α : N → N ¯⊗ L∞(G) be an ergodic action of a locally compact quantum group G on a von Neumann algebra N+. Then α is integrable if and only if there exists a non-zero x ∈ N integrable for α. Proof. Clearly if α is integrable then there are plenty of integrable elements. Assume now that there exists a non-zero integrable element. σ-weak closure J is a principal left ideal, i.e. there exists a non-zero projection P ∈ N such that J = NP . Let J = (cid:8)x ∈ N x∗x is integrable for α(cid:9). Then J is a non-zero left ideal in N and so its 2(cid:1). Let x ∈ N+ be integrable for α. Let δ be the modular element for G and let ξ ∈ Dom(cid:0)δ 1 Consider ω ∈ L1(G)+ given by ω(y) = hξ yξi for all y ∈ L∞(G). Using right δ-invariance of hL we get and thus it follows that for any y ∈ J we have (id⊗ω)α(y) ∈ J . By continuity for any y ∈ J we have (id ⊗ ω)α(y) ∈ J. In particular (id ⊗ ω)α(P ) ∈ J and consequently (id ⊗ ω)α(P ) = (id ⊗ ω)α(P )P . As this is true for separating set of functionals ω, we conclude that α(P )(P ⊗1) = α(P ), i.e. α(P ) ≤ P ⊗ 1. By lemma 3.1 P = 1, so J is σ-weakly dense in N. (cid:3) Proposition 3.2 has its obvious analog for left actions (cf. Section 2.2). Now let G and H be locally compact quantum groups and assume that H is an open subgroup of G. As mentioned in Section 2.4 H is then also a closed quantum subgroup and the image of L∞(bH) in L∞(bG) under the corresponding embedding is the σ-weak closure of (cid:8)(id ⊗ ω)(id ⊗ Θ)WG ω ∈ B(L2(H))∗(cid:9), where Θ : L∞(G) → L∞(H) is the surjection corresponding to the embedding H ⊂ G. (3.5) Let ι : L∞(H) ֒→ L∞(G) be the (non-unital) embedding as in Section 2.4. Now take ω in (3.5) to be of the form ϑ ◦ ι, where ϑ ∈ B(L2(G))∗. Then (id ⊗ ω)(cid:0)(id ⊗ Θ)(WG)(cid:1) = (id ⊗ ϑ)(cid:0)(1 ⊗ 1H)WG(cid:1) = (id ⊗ ϑ1H)WG, where 1H is the central support of Θ and ϑ1H denotes the functional x 7→ ϑ(1Hx). Now assume that ϑ is a L2-bounded functional on L∞(G). This means that there exists a vector ξϑ ∈ L2(G) such that ϑ(x) = hξϑ η(x)i for all x ∈ Dom(η). It is easy to see that for any c ∈ L∞(G) the functional ϑc is still L2-bounded: (ϑc)(x) = ϑ(cx) = hξϑ η(cx)i = hξϑ cη(x)i = hc∗ξϑ η(x)i for all x ∈ Dom(η). It is known ([18, Section 6]) that the set of L2-bounded functionals is norm dense in L1(G). Theorem 3.3. Let H be a closed quantum subgroup of G via γ : L∞(bH) ֒→ L∞(bG). Then H is open in G if and only if there exists a non-zero element x ∈ L∞(bH) such that γ(x) is square- integrable with respect to the right Haar measure of bG. INTEGRABLE ACTIONS AND QUANTUM SUBGROUPS 7 Proof. We can identify L∞(bH) with its image under γ, so L∞(bH) ⊂ L∞(bG). Assume that x ∈ L∞(bH) is non-zero and square integrable for bh. The map ∆bG restricted to L∞(bH) can be viewed as an ergodic left action α of bG on bH: β = ∆bG(cid:12)(cid:12)L∞( bH) : L∞(bH) −→ L∞(bG) ¯⊗ L∞(bH) and the element x∗x is non-zero and integrable for β. By a left analog of Proposition 3.2 β is guarantees that H is open. Assume now that H is open. From the discussion preceding the statement of the theorem we integrable, and by right invariance, bh restricted to L∞(bH) is semifinite. Now [9, Theorem 7.5] know that the image of L∞(bH) in L∞(bG) is the closure of the set (cid:8)(id ⊗ ϑ1H)WG ϑ ∈ L1(G)(cid:9). Since the set of L2-bounded functionals is dense in L1(G), there must exist an L2-bounded ϑ such that x = (id ⊗ ϑ1H)WG is non-zero. Moreover, we also know that ϑ1H is L2-bounded, so (cid:3) Proof. This has practically been proved in Theorem 3.3. To make things more precise, we note Corollary 3.4. Let G and H be locally compact quantum groups and let H be a closed quantum x∗ ∈ Dom(bη) by [18, Theorem 6.4(4)]. subgroup of G with corresponding injection γ : L∞(bH) ֒→ L∞(bG). Let bh be the right Haar weight on L∞(bG). Then H is open in G if and only if the right Haar measure of bH is proportional to bh ◦ γ. that the condition that the Haar weight on L∞(bH) is proportional to bh ◦ γ is equivalent to bh ◦ γ being semifinite on L∞(bH). Thus if bh ◦ γ is semifinite then H is open in G by Theorem 3.3. Conversely if H is open in G then Theorem 3.3 says that there is a non-zero element of γ(cid:0)L∞(bH)(cid:1) integrable for bh. By right invariance this element is integrable for the ergodic action of bG on γ(cid:0)L∞(bH)(cid:1) given by ∆bH. Thus by the left analog of Proposition 3.2 this action is integrable, which again by right invariance means that bh ◦ γ is semifinite. In terms of integrability of actions the characterization of open quantum subgroups is given by (cid:3) the following corollary: Corollary 3.5. Let G and H be locally compact quantum groups and let H be a closed subgroup of G with corresponding injection γ : L∞(bH) ֒→ L∞(bG). Then H is open in G if and only if the left action of bG on γ(cid:0)L∞(bH)(cid:1) given by ∆bG is integrable. 4. Image and kernel of a homomorphism Let G and H be locally compact quantum groups throughout this section we will be focused on a homomorphism Π : H → G with corresponding bicharacter V ∈ L∞(bG) ¯⊗ L∞(H) and a right quantum group homomorphism α : L∞(G) → L∞(G) ¯⊗ L∞(H). We will introduce a closed quantum subgroup of G and a "quotient" quantum group of H which correspond to the closure of the image and quotient by the kernel of Π. 4.1. Closure of image of Π. The next proposition is a slightly stronger version of [9, Lemma 1.2]. Proposition 4.1. Let G and H be locally compact quantum groups and let Π : H → G be a homomorphism of quantum groups with corresponding bicharacter V ∈ L∞(bG) ¯⊗ L∞(H). Let Then L is a Baaj-Vaes subalgebra of L∞(bG). L =(cid:8)(id ⊗ ζ)V ζ ∈ L1(H)(cid:9)σ−weak closure . 8 PAWE L KASPRZAK, FATEMEH KHOSRAVI, AND PIOTR M. SO LTAN Proof. First we note that L is a von Neumann algebra. Indeed this follows directly from [35, Theorem 1.6(1)]. To check that L is an invariant subalgebra it is enough to see that for any ζ ∈ L1(H) and any φ ∈ L1(bG) we have This we infer from the fact that (id ⊗ φ)(cid:0)∆bG(cid:0)(id ⊗ ζ)V(cid:1)(cid:1), (φ ⊗ id)(cid:0)∆bG(cid:0)(id ⊗ ζ)V(cid:1)(cid:1) ∈ L. (id ⊗ φ)(cid:0)∆bG(cid:0)(id ⊗ ζ)V(cid:1)(cid:1) = (id ⊗ φ ⊗ ζ)(V23V13) = (id ⊗ eφ)V, (φ ⊗ id)(cid:0)∆bG(cid:0)(id ⊗ ζ)V(cid:1)(cid:1) = (φ ⊗ id ⊗ ζ)(V23V13) = (id ⊗ φ)V, where eφ(y) = (φ ⊗ ζ)(cid:0)V (1 ⊗ y)(cid:1) for all y ∈ L∞(H). Similarly with φ(y) = (φ ⊗ ζ)(cid:0)(1 ⊗ y)V(cid:1) for all y ∈ L∞(H). bG(cid:0)(id ⊗ ζ)V(cid:1) =(cid:0)id ⊗ [ζ ◦ RH](cid:1)(R −t](cid:1)(τ t (cid:0)(id ⊗ ζ)V(cid:1) =(cid:0)id ⊗ [ζ ◦ τ H by [19, Proposition 3.10] (cf. [23, Remark 41]). bG t ⊗ τ H R τ bG bG ⊗ RH)V =(cid:0)id ⊗ [ζ ◦ RH](cid:1)V ∈ L −t](cid:1)V ∈ L t )V =(cid:0)id ⊗ [ζ ◦ τ H What is left is to establish invariance of L under the unitary antipode RH and the scaling group τ H of H. This follows from the fact that for each ζ ∈ L1(H) and t ∈ R we have (cid:3) As mentioned in Section 2.4, a Baaj-Vaes subalgebra L of L∞(bG) defines a quantum subgroup K of G by setting bK to be the quantum group associated with L. Thus, using Proposition 4.1 we can introduce the following definition: Definition 4.2. Let G and H be locally compact quantum groups and let Π : H → G be a homomorphism of quantum groups. We define the closed quantum subgroup im Π of G as the quantum subgroup related to the Baaj-Vaes subalgebra L of L∞(bG) described in Proposition 4.1. In particular L∞(cid:0)[ im Π(cid:1) =(cid:8)(id ⊗ ζ)V ζ ∈ L1(H)(cid:9)σ−weak closure . 4.2. Quotient by kernel of Π. In a manner completely analogous to the one used in the proof of Proposition 4.1 we obtain also the following result: Proposition 4.3. Let G and H be locally compact quantum groups and let Π : H → G be a homomorphism of quantum groups with corresponding bicharacter V ∈ L∞(bG) ¯⊗ L∞(H). Let R =(cid:8)(φ ⊗ id)V φ ∈ L1(bG)(cid:9)σ−weak closure . Then R is a Baaj-Vaes subalgebra of L∞(H). Definition 4.4. Let G and H be locally compact quantum groups. Given a homomorphism group H/ker Π by setting Π : H → G described by a bicharacter V ∈ L∞(bG) ¯⊗ L∞(H) we define a locally compact quantum and ∆H/ker Π = ∆H(cid:12)(cid:12)L∞(H/ker Π). L∞(H/ker Π) =(cid:8)(φ ⊗ id)V φ ∈ L1(bG)(cid:9)σ−weak closure Given G, H and Π as in Definition 4.4 it may happen that the subalgebra L∞(H/ker Π) is all of L∞(H). In this case we say that H/ker Π = H. Using this language we can rewrite implication (1) ⇒ (3) of [4, Theorem 3.4] in the following way: Corollary 4.5. Let G and H be locally compact quantum groups and let Π : H → G be a homo- morphism of quantum groups. If Π identifies H with a closed subgroup of G then H/ker Π = H. Remark 4.6. If G and H are locally compact groups and Π : H → G is a continuous homomor- phism then it is not difficult to see that Definitions 4.2 and 4.4 yield quantum groups corresponding to the closure of the image of Π and kernel of Π respectively. INTEGRABLE ACTIONS AND QUANTUM SUBGROUPS 9 4.3. Quotient by kernel of a quantum group action. Let G and H be locally compact quantum groups. A homomorphism of quantum groups Π : H → G defines the quantum group H/ker Π as in Definition 4.4. The next theorem shows that this quantum group can be defined directly from the action α : L∞(G) → L∞(G) ¯⊗ L∞(H) associated to Π. Theorem 4.7. Let H and G be locally compact quantum groups and let Π : H → G be a homomor- phism of quantum groups with corresponding bicharacter V ∈ L∞(bG) ¯⊗ L∞(H) and right quantum group homomorphism α : L∞(G) → L∞(G) ¯⊗ L∞(H). Then (cid:8)(ω ⊗ id)α(x) x ∈ L∞(G), ω ∈ L1(G)(cid:9)′′ = L∞(H/ker Π). (4.1) Proof. Let us denote by M the left hand side of (4.1) and let R be a shorthand for L∞(H/ker Π). Clearly M is a von Neumann algebra. The next thing to see is that M ⊂ L∞(H) is a left coideal. To check this take x ∈ L∞(G), ω ∈ L1(G) and ζ ∈ L1(H). We have (ζ ⊗ id)∆H(cid:0)(ω ⊗ id)α(x)(cid:1) =(cid:0)(cid:2)(ω ⊗ ζ) ◦ α(cid:3) ⊗ id(cid:1)α(x) ∈ M, so ∆H(cid:0)M(cid:1) ⊂ L∞(H) ¯⊗ M. Once we know that M is a left coideal in L∞(G), we can consider its co-dual eM (cf. the end of The co-dual of M consists of those y ∈ L∞(bH) which satisfy uy = yu for all u ∈ M. Thus y ∈ eM Section 2.1 and [12, Section 3]). if and only if V (x ⊗ 1)V ∗(1 ⊗ y) = (1 ⊗ y)V (x ⊗ 1)V ∗, x ∈ L∞(G) or, equivalently, V ∗(1 ⊗ y)V (x ⊗ 1) = (x ⊗ 1)V ∗(1 ⊗ y)V, x ∈ L∞(G). This means that V ∗(1 ⊗ y)V ∈ L∞(G)′ ¯⊗ B(L2(H)), but at the same time V ∗(1 ⊗ y)V ∈ L∞(bG) ¯⊗ B(L2(H)). Since L∞(bG) ∩ L∞(G)′ = C1 (this is [18, Proposition 4.7(3)] applied to bG), we find that V ∗(1 ⊗ y)V = 1 ⊗ z for some z ∈ B(L2(H)). This is equivalent to (1 ⊗ y)V = V (1 ⊗ z) and slicing the left leg of this equality yields yu = uz, u ∈ R. Since 1 ∈ R, we conclude that y = z, so for y to belong to eM is equivalent to y ∈ L∞(bH) ∩ R′ = eR. In other words eM = eR and so M = eeM = eeR = R. In analogy with the way L∞(H/ker Π) is defined in terms of the action of H on G given by the homomorphism Π in Theorem 4.7 we can define a similar object for an arbitrary action α of H on a von Neumann algebra M. (cid:3) Proposition 4.8. Let M be a von Neumann algebra and let H be a locally compact quantum group. . Let α : M → M ¯⊗ L∞(H) be an action of H on M. Let N0 = (cid:8)(ω ⊗ id)α(x) ω ∈ M∗, x ∈ M(cid:9)′′ Then N0 is an invariant subalgebra of L∞(H). Proof. The proof is completely analogous to the first stage of the proof of Theorem 4.7. (cid:3) Definition 4.9. Let M be a von Neumann algebra and let H be a locally compact quantum group. Let α : M → M ¯⊗ L∞(H) be an action of H on M. Let N be the smallest Baaj-Vaes subalgebra of L∞(H) containing the set We define the quantum group H/ker α as the locally compact quantum group such that (cid:8)(ω ⊗ id)α(x) ω ∈ M∗, x ∈ M(cid:9). and ∆H/ker α = ∆H(cid:12)(cid:12)L∞(H/ker α). Remark 4.10. L∞(H/ker α) = N 10 PAWE L KASPRZAK, FATEMEH KHOSRAVI, AND PIOTR M. SO LTAN (1) Theorem 4.7 makes it clear that if H and G are locally compact quantum groups and Π : H → G is a homomorphism of quantum groups with associated right quantum group morphism (action) α : L∞(G) → L∞(G) ¯⊗ L∞(H) then H/ker α coincides with H/ker Π. (2) The quantum group of inner automorphisms of a locally compact quantum group defined in [11, Definition 3.3] is a special case of the group of the form H/ker α for a certain action α. (3) If H is a locally compact group acting on a von Neumann algebra N with corresponding "quantum action" α : N → N ¯⊗ L∞(H) then the quantum group H/ker α defined in Def- inition 4.9 corresponds to the quotient of H by the subgroup of all elements which act trivially on N (cf. Remark 4.6). The next result shows that if an action α : N → N ¯⊗ L∞(H) of a locally compact quantum group H on a von Neumann algebra N is integrable, the kernel of α is in some sense small, or rather compact (cf. [9, Theorem 7.2]). Proposition 4.11. Let N be a von Neumann algebra and let H be a locally compact quantum group. Let α : N → N ¯⊗ L∞(H) be an action of H on N. Assume that α is integrable. Then of N we then have Proof. Since α is integrable, there exists x ∈ N+ such that (id ⊗ hL)α(x) ∈ N+ and there is a \H/ker α is an open quantum subgroup of bH. normal state ω on N such that ω(cid:0)(id ⊗ hL)α(x)(cid:1) 6= 0. By definition of the extended positive part 0 6= hL(cid:0)(ω ⊗ id)α(x)(cid:1) < +∞. 2(cid:1) of L∞(H/ker α) is non-zero and square-integrable with Thus the element R(cid:0)(cid:0)(ω ⊗ id)α(x)(cid:1) 1 respect to h, so by Theorem 3.3, \H/ker α is an open subgroup of bH. 5. Canonical implementation of right quantum group homomorphism (cid:3) In this section we address the problem of determining the canonical implementation ([28] and Section 2.2) of the action associated to a homomorphism of quantum groups. Let G and H be locally compact quantum groups and let Π : H → G be a homomorphism of quantum groups. Let α : L∞(G) → L∞(G) ¯⊗ L∞(H) be the corresponding action of H. In order to define the canonical implementation of α one must choose a n.s.f. weight on L∞(G). We choose for this the right Haar weight h of G. A way to determine the canonical implementation is then provided by the proof of [28, Propo- sition 4.3]. In this proposition S. Vaes proves that in case the chosen weight is δ−1-invariant (see [6, D´efinition 2.7], [28, Definition 2.3]) the unitary implementation U coincides with a unitary Vθ defined in [28, Proposition 2.4] (cf. [6, Th´eor`eme 2.9]). A careful examination of the proof of [28, Proposition 4.3] shows that the claim is established on the basis of several properties of Vθ listed in [28, Proposition 2.4]. In particular δ−1-invariance of the chosen weight does not play a crucial role. In our situation we choose an invariant weight, so [28, Proposition 4.3] is not applicable directly. Nevertheless we already know that the action α : L∞(G) → L∞(G) ¯⊗ L∞(H) is implemented by the corresponding bicharacter (see (2.2)). Moreover V has the following properties: ◮ (cid:0)J G ⊗ J ◮ (cid:0)∇G ⊗ ∇ ◮ for any ω ∈ L1(H) and any x ∈ Dom(η) we have(cid:0)(id ⊗ ω)V(cid:1)η(x) = η(cid:0)(id ⊗ ω)α(x)(cid:1), where bH(cid:1)V = V ∗(cid:0)J G ⊗ J bH(cid:1), bH(cid:1)V = V(cid:0)∇G ⊗ ∇ bH(cid:1), η is the GNS map for the right Haar weight of G. bG ⊗ τ H which The first two properties are consequences of invariance of V under R we already used in the proof of Proposition 4.1 (cf. [17, Proposition 2.1]). The last property follows from (7.1) and Proposition 7.4 from the Appendix. Indeed, using notation introduced at the beginning of Section 7.2, for ω ∈ L1(H) we set µ = ω ◦ ΛH ◦ π, where π is the Hopf ∗-homomorphism corresponding to Π and ΛH is the reducing morphism for H. Then by (7.1) bG ⊗ RH and τ µ ∗ x = (id ⊗ µ)(cid:0)W G(x ⊗ 1)W G∗(cid:1) = (id ⊗ ω)(cid:0)V (x ⊗ 1)V ∗(cid:1) = (id ⊗ ω)(cid:0)α(x)(cid:1) INTEGRABLE ACTIONS AND QUANTUM SUBGROUPS 11 (because (cid:0)id ⊗ (ΛH ◦ π)(cid:1)WG = V ) and by Proposition 7.4. (cid:0)(id ⊗ ω)V(cid:1)η(x) =(cid:0)(id ⊗ µ)W G(cid:1)η(x) = η(µ ∗ x) Repeating the steps of [28, Proposition 4.3] with 1 instead of δ−1 we obtain the following: Theorem 5.1. Let G and H be locally compact quantum groups and let Π : H → G be a homomor- phisms of quantum groups with corresponding bicharacter V ∈ L∞(bG) ¯⊗ L∞(H) and right quantum group homomorphism α : L∞(G) → L∞(G) ¯⊗ L∞(H). Then V is the canonical implementation of α. Let us also remark that Theorem 5.1 was actually stated and used already in [5, Proof of Proposition 3.12]. We are now ready to use [28, Theorem 5.3] which characterizes integrability of actions in the following terms: let α : N → N ¯⊗ L∞(H) be an action of a locally compact quantum group H on a von Neumann algebra N. Choose a n.s.f. weight θ on N and identify N with its image in B(Hθ). Furthermore let U ∈ B(Hθ) ¯⊗ L∞(H) be the canonical implementation of α and let N2 be the von Neumann algebra Now [28, Theorem 5.3] says that α is integrable if and only if there exists a normal surjective ∗-homomorphism ρ : N ⋊α H → N2 such that N2 =(cid:0)N ∪(cid:8)(id ⊗ ω)U ω ∈ L1(H)(cid:9)(cid:1)′′ . ◮ ρ(cid:0)α(x)(cid:1) = x for all x ∈ N, ◮ ρ(cid:0)1 ⊗ [(id ⊗ ω)WH](cid:1) = (id ⊗ ω)U for all ω ∈ L1(H). 6. Integrability and quantum subgroups In this section we will study homomorphisms of quantum groups with the property that the as- sociated action is integrable. The first consequence will be that the kernel of such a homomorphism is in some sense small and the other that the quotient by the kernel can be canonically identified with the closure of the image. To make this more precise let Π : H → G be a homomorphism of quantum groups with associated bicharacter V ∈ L∞(bG) ¯⊗ L∞(H). We will say that H/ker Π can be canonically identified with im Π if there exists an isomorphism χ : L∞(cid:0)im Π(cid:1) → L∞(H/ker Π) such that (id ⊗ χ)(Wim Π) = V. (6.1) Note that indeed the right leg of V belongs to L∞(H/ker Π), in fact we have We will write simply H/ker Π ∼= im Π to denote this situation. This will not lead to confusion, as an isomorphism χ satisfying (6.1) is necessarily unique. V ∈ L∞(cid:0)[ im Π(cid:1) ¯⊗ L∞(H/ker Π). Theorem 6.1. Let G and H be locally compact quantum groups and let Π : H → G be a homomor- phism of quantum groups. Assume that the corresponding right quantum group homomorphism α is integrable. Then (1) \H/ker Π is an open subgroup of bH, (2) H/ker Π ∼= im Π. Proof. Ad (1). This follows from Proposition 4.11 and Remark 4.10(2). Ad (2). Let us denote L∞(G) by N, thus α : N → N ¯⊗ L∞(H) and let us, as usual, denote by V the bicharacter corresponding to Π. The assumption of integrability means that, by [28, Theorem 5.3] and Theorem 5.1, we have a normal surjective ∗-homomorphism ρ : N ⋊α H → N2, where N2 is the von Neumann algebra generated by N and (cid:8)(id ⊗ ω)V ω ∈ L1(H)(cid:9) such that ◮ ρ(cid:0)α(x)(cid:1) = x for all x ∈ N, ◮ ρ(cid:0)1 ⊗ [(id ⊗ ω)WH](cid:1) = (id ⊗ ω)V for all ω ∈ L1(H). 12 PAWE L KASPRZAK, FATEMEH KHOSRAVI, AND PIOTR M. SO LTAN [ Let eρ denote the map L∞(bH) ∋ y 7→ ρ(1 ⊗ y) ∈ N2. By the second point above and the definition of im Π the map eρ has image L∞(cid:0)[ im Π(cid:1). Since it also preserves comultiplications, we find that im Π is an open subgroup of bH and, as open subgroups are closed, there exists a unital injective normal ∗-homomorphism χ : L∞(cid:0)im Π(cid:1) ֒→ L∞(H) such that (eρ ⊗ id)WH = (id ⊗ χ)Wim Π (cf. [9, Remark 3.7]). But the left hand side of (6.2) is V , due to properties of ρ, which proves the claim. (cid:3) (6.2) Theorem 6.2. Let G and H be locally compact quantum groups and let H be a closed quantum subgroup of G with right quantum group homomorphism α : L∞(G) → L∞(G) ¯⊗ L∞(H) corre- sponding to the embedding H ֒→ G. Then α is integrable. As in the proof of Theorem 6.1, we will use the criterion of integrability from [28] mentioned at the end of Section 5. We will in fact show that if H is a closed subgroup of G and we write N for the algebra L∞(G) then the algebras N ⋊α H and N2 are isomorphic with the isomorphism satisfying the requirements of [28, Theorem 5.3]. Theorem 6.2 was proved in [5] by the same technique. Proof of Theorem 6.2. Let us begin with the remark that if Y = (J ⊗ J)W bG(J ⊗ J) then Y ∈ Y (u ⊗ 1)Y ∗ = ∆bGop (u), L∞(G)′ ¯⊗ L∞(bG) and Y implements the opposite comultiplication on L∞(bG): Indeed, writing bR for the unitary antipode of bG we have bG(cid:0)bR(u∗) ⊗ 1(cid:1)W = (J ⊗ J)∆bG(cid:0)bR(u∗)(cid:1)(J ⊗ J) = (J ⊗ J)(bR ⊗ bR)(cid:0)∆bGop (u∗)(cid:1)(J ⊗ J) = ∆bGop (u). u ∈ L∞(bG). bG(J ⊗ J)(u ⊗ 1)(J ⊗ J)W Y (u ⊗ 1)Y ∗ = (J ⊗ J)W = (J ⊗ J)W (J ⊗ J) ∗ bG (J ⊗ J) ∗ bG Let us write N for the algebra L∞(G). We can now define a map from N ⋊α H to the algebra N2 Recall that since H is a closed subgroup of G, we have a normal injective unital ∗-homomorphism which is the von Neumann algebra generated inside B(L2(G)) by N and (cid:8)(id ⊗ ω)V ω ∈ L1(H)(cid:9). γ : L∞(bH) ֒→ L∞(bG) such that (γ ⊗ id)WH = V. For a ∈ N ⋊α H we will consider the element Y ∗(cid:2)(id ⊗ γ)(V ∗aV )(cid:3)Y. (6.3) It is not immediately clear that V ∗aV belongs to the domain of id ⊗ γ, but it follows from the following computation. We know that elements of the form α(x)(1 ⊗ y) = V (x ⊗ 1)V ∗(1 ⊗ y) with x ∈ N and y ∈ L∞(bH) span a dense subspace in N ⋊α H ([13, Proposition 2.3]). Taking a = V (x ⊗ 1)V ∗(1 ⊗ y) in (6.3) yields Y ∗(cid:2)(id ⊗ γ)(cid:0)(x ⊗ 1)V ∗(1 ⊗ y)V(cid:1)(cid:3)Y = Y ∗(cid:2)(cid:0)(x ⊗ 1)(γ ⊗ γ)(cid:0)WH∗ = Y ∗(cid:2)(x ⊗ 1)(γ ⊗ γ)(cid:0)∆bHop (y)(cid:1)(cid:3)Y = Y ∗(cid:2)(x ⊗ 1)∆bGop(cid:0)γ(y)(cid:1)(cid:3)Y = Y ∗(cid:2)(x ⊗ 1)Y(cid:0)γ(y) ⊗ 1(cid:1)Y ∗(cid:3)Y = Y ∗(cid:2)Y (x ⊗ 1)(cid:0)γ(y) ⊗ 1(cid:1)Y ∗(cid:3)Y = (x ⊗ 1)(cid:0)γ(y) ⊗ 1(cid:1) = xγ(y) ⊗ 1. (1 ⊗ y)WH(cid:1)(cid:1)(cid:3)Y Thus there exists a normal injective unital ∗-homomorphism ρ : N ⋊α H → N2 such that In particular ρ(cid:0)α(x)(1 ⊗ y)(cid:1) = xγ(y), x ∈ N, y ∈ L∞(bH). INTEGRABLE ACTIONS AND QUANTUM SUBGROUPS ◮ ρ(cid:0)α(x)(cid:1) = x for x ∈ N, ◮ ρ(cid:0)1 ⊗ [(id ⊗ ω)WH](cid:1) = (id ⊗ ω)V for ω ∈ L1(H). Clearly N and γ(cid:0)L∞(bH)(cid:1) are both contained in the range of ρ, so ρ must be surjective. Theorem 6.3. Let G and H be locally compact quantum groups and let Π : H → G be a homomor- phism of quantum groups with corresponding right quantum group homomorphism α : L∞(G) → L∞(G) ¯⊗ L∞(H). Assume that (1) \H/ker Π is an open subgroup of bH, (2) H/ker Π ∼= im Π. Then α is integrable. Proof. As we already noticed in the first part of the proof of Theorem 6.1, the image of α is contained in L∞(G) ¯⊗ L∞(H/ker Π). Now let χ : L∞(cid:0)im Π(cid:1) → L∞(H/ker Π) be as in (6.1) and let eα = (id ⊗ χ−1) ◦ α. Then eα is an action of im Π on L∞(G) implemented by the bicharacter (id ⊗ χ−1)V ∈ L∞(bG) ¯⊗ L∞(cid:0)im Π(cid:1). But we know that im Π is a closed subgroup of G via the corresponding homomorphism of quantum groups, so by Theorem 6.2 the action eα is integrable. To conclude that α is integrable we note that due to assumption (1) the Haar measure of H/ker Π is (proportional to) the restriction of the Haar measure of H to L∞(H/ker Π) by Corollary 3.4. (cid:3) Remark 6.4. Theorems 6.1 and 6.3 provide the following characterization of those homomor- phisms of quantum groups Π : H → G for which the corresponding right quantum group homo- morphism α : L∞(G) → L∞(G) ¯⊗ L∞(H) is an integrable action: α is integrable if and only if the conditions (1) and (2) from Theorem 6.1 hold. Corollary 6.5. Let G and H be locally compact quantum groups and let Π : H → G be a homomor- phism of quantum groups with corresponding right quantum group homomorphism α : L∞(G) → L∞(G) ¯⊗ L∞(H). Assume that α is integrable and free. Then Π identifies H with a closed quantum subgroup of G. is all of L∞(H). In view of Theorem 4.7, this means that L∞(H/ker Π) = L∞(H) (cf. discussion after Definition 4.4). Since α is assumed to be integrable, by Theorem 6.1(2), we have H/ker Π ∼= im Π in the sense that Proof. The freeness of α means that (cid:8)(ω ⊗ id)α(x) x ∈ L∞(G), ω ∈ L1(G)(cid:9)′′ there exists an isomorphism χ : L∞(cid:0)im Π(cid:1) → L∞(H/ker Π) such that (id ⊗ χ)Wim Π = V (6.4) (cf. remarks before Theorem 6.1). (id ⊗ χ)Wim Π, cf. [19], [4, Theorem 1.10]. In other words we have Thus χ can be regarded as an isomorphism L∞(cid:0)im Π(cid:1) → L∞(H) with the property (6.4). Let im Π(cid:1) such that (bχ ⊗ id)WH = bχ be the dual of χ. Then bχ is an isomorphism L∞(bH) → L∞(cid:0)[ where bχ is a normal, unital injective ∗-homomorphism. It follows that the bicharacter V identifies V = (bχ ⊗ id)WH, H with a closed quantum subgroup of G. (cid:3) We now turn our attention to quantum subgroups closed in the sense of Woronowicz ([4, Def- inition 3.2]). Let us recall that if H and G are locally compact quantum groups and Π : H → G identifies H with a closed quantum subgroup of G in the sense of Woronowicz if the bicharacter V associated with Π generates the C∗-algebra C0(H). One of the equivalent formulations of this condition is that 13 (cid:3) is norm dense in C0(H) ([4, Theorem 3.6]). In particular it is then σ-weakly dense in L∞(H). In what follows we will use the phrase "Woronowicz-closed quantum subgroup" instead of "closed quantum subgroup in the sense of Woronowicz" in order to make the statements more transparent. (cid:8)(ω ⊗ id)V ω ∈ B(L2(G))∗(cid:9) 14 PAWE L KASPRZAK, FATEMEH KHOSRAVI, AND PIOTR M. SO LTAN Corollary 6.6. Let G and H be locally compact quantum groups and let Π : H → G be a homo- morphism of quantum groups which identifies H with a Woronowicz-closed quantum subgroup of G. If the corresponding right quantum group homomorphism α : L∞(G) → L∞(G) ¯⊗ L∞(H) is integrable then Π identifies H with a closed quantum subgroup of G. Proof. It is clear from the remark preceding the corollary that if H is a Woronowicz-closed quantum subgroup of G then α is free. As it is assumed to be integrable, the fact that it is a closed quantum subgroup as defined in Section 2.4 follows from Corollary 6.5. (cid:3) Remark 6.7. (1) Corollary 6.6 together with Theorem 6.2 show that closed quantum subgroups are pre- cisely Woronowicz-closed quantum subgroups with integrable corresponding right quantum group homomorphism. (2) This gives as an easy consequence an immediate proof of [4, Theorem 6.1] as any action of a compact quantum group is integrable. (3) Finally, Corollary 6.6 is really a characterization of those homomorphisms of quantum groups Π : H → G which identify H with a closed quantum subgroup of G. Indeed, if H is a closed quantum subgroup then the action corresponding to Π is free (because a closed quantum subgroup of G is a Woronowicz-closed quantum subgroup of G, [4, Theorem 3.5]) and it is integrable by Theorem 6.2. 7. Appendix 7.1. Classical case. Let us discuss the results we obtained for locally compact quantum groups in the special case when the groups involved are in fact classical. In what follows the symbol vN(K) will denote the (right) group von Neumann algebra of a locally compact group K. Thus let G and H be locally compact groups and let Π : H → G be a continuous homomorphism. The three equivalent descriptions of Π used when dealing with quantum groups take the following form: ◮ the bicharacter V ∈ L∞(bG) ¯⊗ L∞(H) = vN(G) ¯⊗ L∞(H) ∼= L∞(H, vN(G)) is the repre- sentation of H on L2(G) by right shifts along the image of Π: ◮ The action α : L∞(G) → L∞(G) ¯⊗ L∞(H) ∼= L∞(G × H) is the pull-back of the natural right action of H on L∞(G) arising from Π: so that for f ∈ L∞(G) the function α(f ) ∈ L∞(G × H) is α(f )(g, h) = f(cid:0)gΠ(h)(cid:1). ◮ The Hopf ∗-homomorphism π ∈ Mor(C0(G), C0(H)) is the pre-composition with Π: g · h = gΠ(h), g ∈ G, h ∈ H, Let us fix G, H and Π (so we also have V , α and π). We leave to the reader the verification π(f ) = f ◦ Π, f ∈ C0(G). (Vhψ)(g) = ψ(cid:0)gΠ(h)(cid:1), ψ ∈ L2(G), g ∈ G, h ∈ H. that and (cid:8)(id ⊗ ζ)V ζ ∈ L1(H)(cid:9)σ−weak closure (cid:8)(φ ⊗ id)V φ ∈ vN(G)∗(cid:9)σ−weak closure coincide respectively with vN(cid:0)im Π(cid:1) canonically embedded in vN(G) and L∞(H/ker Π) viewed as a subalgebra of L∞(H) (functions constant on cosets of ker Π). It is easy to see that the action α is integrable if and only if the set of positive f ∈ L∞(G) such that the integral Z H f(cid:0) · Π(h)(cid:1)dhL(h) exists in L∞(G) is σ-weakly dense in L∞(G)+. Rewriting the results of Section 6 for classical groups we find the following: INTEGRABLE ACTIONS AND QUANTUM SUBGROUPS 15 Theorem 7.1. Let G and H be locally compact groups and let Π : H → G be a continuous homomorphism. Let α be the action of H on L∞(G) corresponding to Π. Then α is integrable if and only if (1) ker Π is compact, (2) the image of Π is closed and H/ker Π is homeomorphic to im Π. We leave the details to the reader, but let us only mention that condition (1) of Theorem 7.1 is equivalent to \H/ker Π being an open quantum subgroup of bH by [9, Theorem 7.2]. In the special case of an injective morphism we obtain the following corollary: Corollary 7.2. Let G and H be locally compact groups and let Π : H → G be a continuous injective homomorphism. Then the image of Π is closed and Π is a homeomorphism onto im Π if and only if the associated action of H on L∞(G) is integrable. 7.2. Characterization of compact quantum subgroups. In this section we will present a characterization of compact quantum subgroups somewhat analogous to that of open quantum subgroups given in Theorem 3.3 (cf. also Corollary 7.5). To that end let us recall from [28, Definition 4.1] (see also [12, Lemma 3.11]) that if H and G are locally compact quantum groups homogeneous space G/H is described by the left coideal L∞(G/H) ⊂ L∞(G) defined as the co- and H is identified with a closed subgroup of G via γ : L∞(bH) ֒→ L∞(bG) then the measured dual of γ(cid:0)L∞(bH)(cid:1) ⊂ L∞(bG) (see Section 2.1). Theorem 7.3. Let G and H be locally compact quantum groups and let H be a closed quantum subgroup of G. Then H is compact if and only if there exists a non-zero element x ∈ L∞(G/H) which is square integrable with respect to the Haar measure of G. For the proof of Theorem 7.3 we need a few results about convolutions. Let G be a locally compact quantum group. Then we have Banach algebra structures on spaces of functionals like L1(G) or Cu 0(G)∗. For the purposes of this section we will denote the (convolution) products on these spaces respectively by ∗ and ∗ : ω1 ∗ ω2 = (ω1 ⊗ ω2) ◦ ∆G, µ1 ∗ µ2 = (µ1 ⊗ µ2) ◦ ∆u G, ω1, ω2 ∈ L1(G), 0(G)∗. µ1, µ2 ∈ Cu Moreover the adjoint of the reducing morphism ΛG : Cu isometry of ι of L1(G) into Cu Proposition 5.3]) and denoting by Rµ the operator 0(G)∗. The image of ι is a closed two sided ideal in Cu 0(G) ։ C0(G) ⊂ L∞(G) provides an 0(G)∗ ([14, we obtain a bounded map L1(G) → L1(G). The adjoint of this map, denoted by L∞(G) ∋ x 7→ µ ∗ x ∈ L∞(G), is σ-weakly continuous and is in fact given by ω 7−→ ι−1(cid:0)ι(ω) ∗ µ(cid:1) µ ∗ x = (id ⊗ µ)(cid:0)WG(x ⊗ 1)W G∗(cid:1) (7.1) (this follows e.g. from [14, Proof of Proposition 5.3]). In particular for any ω ∈ L1(G) we have where ω ∗ x is by definition (id ⊗ ω)∆G(x). Also it is not difficult to check that ι(ω) ∗ x = ω ∗ x, ω ∗ (µ ∗ x) = Rµ(ω) ∗ x (7.2) for all ω ∈ L1(G), µ ∈ Cu 0(G)∗ and x ∈ L∞(G). By [10, Lemma 3.4], for µ ∈ Cu 0(G)∗ and x ∈ L∞(G)+ we have In particular, it follows from the Kadison inequality that if x ∈ L∞(G) is square integrable for h and µ ∈ Cu + then µ ∗ x is also square integrable for h. 0(G)∗ h(µ ∗ x) = µ(1)h(x). 16 PAWE L KASPRZAK, FATEMEH KHOSRAVI, AND PIOTR M. SO LTAN Proposition 7.4. Let G be a locally compact quantum group. Then for µ ∈ Cu Dom(η) 0(G)∗ and x ∈ η(µ ∗ x) =(cid:0)(id ⊗ µ)(WG)(cid:1)η(x) Proof. It is enough to prove this for positive µ because both sides of (7.3) are linear in this variable. Take ω ∈ L1(G). Using [18, Proposition 1.10, formula (1.7)] and [18, Theorem 3.5, formula (7.3) (7.4) (7.5) (7.6) (3.6)] we find that on one hand and on the other hand, by (7.2), In order to analyze the operator (id ⊗ Rµω)(WG) we compute: η(cid:0)(Rµω) ∗ x(cid:1) = (id ⊗ Rµω)(WG)η(x) η(cid:0)(Rµω) ∗ x)(cid:1) = η(cid:0)ω ∗ (µ ∗ x)(cid:1) = (id ⊗ ω)(WG)η(µ ∗ x). (id ⊗ Rµω)(WG) =(cid:0)id ⊗ ι−1(ι(ω) ∗ µ)(cid:1)WG =(cid:0)id ⊗ ι(ω) ∗ µ(cid:1)W G =(cid:0)id ⊗ ι(ω) ⊗ µ(cid:1)(id ⊗ ∆u =(cid:0)id ⊗ ι(ω) ⊗ µ(cid:1)WG =(cid:0)id ⊗ ω ⊗ µ(cid:1)WG 12 W G 13 G)W G WG 13 = (id ⊗ ω)(WG)(id ⊗ µ)(WG). 12 Comparing (7.4) and (7.5) we find using (7.6) that (id ⊗ ω)(WG)η(µ ∗ x) = (id ⊗ ω)(WG)(id ⊗ µ)(W G)η(x). As operators of the form (id ⊗ ω)(WG) can strongly approximate the identity operator, we have proved (7.3). (cid:3) Proof of Theorem 7.3. Assume first that H is a closed quantum subgroup of G with a non-zero x ∈ L∞(G/H) square integrable for h. Thus we have a normal unital and injective ∗-homomorphism γ : L∞(bH) ֒→ L∞(bG), the cor- responding bicharacter V = (γ ⊗ id)WH and the surjective Hopf ∗-homomorphism π : Cu Cu 0(G) → 0(H) such that V = (id ⊗ ΛH)(id ⊗ π)(W G). Moreover L∞(bH) (now identified with a subalgebra of L∞(bG)) is the σ-weak closure of For the purposes of this proof we will identify L∞(bH) with its image in L∞(bG) under γ. (cid:8)(id ⊗ ω)V ω ∈ B(L2(H))∗(cid:9) =(cid:8)(id ⊗ ω)V ω ∈ L1(H)(cid:9). Let P ∈ B(L2(G)) be the projection onto the one dimensional subspace spanned by η(x). We will first show that for any ω ∈ L1(H) and y = (id ⊗ ω)V we have Indeed, let µ = ω ◦ ΛH ◦ π. Then y = (id ⊗ µ)WG, so yP = ω(1)P. yη(x) = (id ⊗ µ)WGη(x) = η(µ ∗ x) by Proposition 7.4. (7.7) (7.8) Now x ∈ L∞(G/H) = L∞(bH)′ ∩ L∞(G), so µ ∗ x = (id ⊗ µ)(cid:0)WG(x ⊗ 1)W G∗(cid:1) = (id ⊗ ω)(cid:0)V (x ⊗ 1)V ∗(cid:1) = (id ⊗ ω)(x ⊗ 1) = ω(1)x, which in view of (7.8) proves (7.7). INTEGRABLE ACTIONS AND QUANTUM SUBGROUPS 17 Equality (7.7) for all ω shows that for any y ∈ L∞(bH) the operator yP is proportional to P . Thus we can define a σ-weakly continuous functional bε : L∞(bH) → C by yP = bε(y)P (it is not difficult to check that bε is in fact a ∗-homomorphism and, in particular, a state). For y = (id ⊗ ω)V we get bε(y) = ω(1) or, in other words (bε ⊗ ω)V = ω(1). (bε ⊗ id)V = 1. It follows that 1 belongs to C0(H), so H is compact. Assume now that H is a compact subgroup of G. In [22, Theorem 5.1] a conditional expectation 0(G/H) was defined (see [12, Example 5.3(2)]), but in fact the arguments used there 0(G) → Cu Cu show that we also have the conditional expectation on the reduced level: E : C0(G) → C0(G/H) As this holds for all ω, we find that E(a) = (id ⊗ hH)(cid:0)V (a ⊗ 1)V ∗(cid:1) where hH is the Haar measure of H. Note that E(a) = µ ∗ a, where µ = hH ◦ ΛH ◦ π (cf. the first part of the proof). By remarks preceding the proof of the theorem we find that there are many non-zero elements of L∞(G/H) which are square integrable for the right Haar measure of G. (cid:3) One can remark that the character bε constructed in the proof of Theorem 7.3 can easily be shown to satisfy (bε ⊗ id) ◦ ∆bH = id, so that bε is the counit on L∞(bH). We end with a re-statement of Theorem 7.3 in terms of integrability (cf. Corollary 3.5). Corollary 7.5. Let G and H be locally compact quantum groups and let H be a closed quantum subgroup of G. Then H is compact if and only if the left action of G on G/H given by ∆G is integrable. Acknowledgements. The second author was partially supported by the Ministry of Science of Iran and wishes to thank the Department of Mathematical Methods in Physics, Faculty of Physics, University of Warsaw for warm hospitality. The first and third authors were partially supported by the National Science Center (NCN) grant no. 2015/17/B/ST1/00085. References [1] S. Baaj & G. Skandalis: Unitaires multiplicatifs et dualit´e pour les produits crois´es de C∗-alg`ebres. Ann. Sci- ent. ´Ec. Norm. Sup., 4e s´erie, t. 26 (1993), 425 -- 488. [2] S. Baaj & S. Vaes: Double crossed products of locally compact quantum groups, J. Inst. Math. Jussieu 4 (2005), 135 -- 173. [3] A. Connes & M. Takesaki: The flow of weights on factors of type III, Tohoku Math. J. 2 (1977), 473 -- 575. [4] M. Daws, P. Kasprzak, A. Skalski & P.M. So ltan: Closed quantum subgroups of locally compact quantum groups, Adv. Math. 231 (2012), 3473 -- 3501. [5] K. De Commer: Galois Objects and Cocycle Twisting for Locally Compact Quantum Groups. J. Operator Theory 66 (2011), 59 -- 106. [6] M. Enock: Sous-facteurs interm´ediaries et groupes quantiques mesur´es, J. Op. Theory 42 (1999), 305 -- 330. [7] U. Haagerup: Operator valued weights in von Neumann algebras, I, J. Funct. Anal. 32 (1979), 175 -- 206. [8] Z. Hu, M. Neufang & Z.-J. Ruan: Module maps over locally compact quantum groups, Studia Math. 211 (2012), 111 -- 145. [9] M. Kalantar, P. Kasprzak & A. Skalski: Open quantum subgroups of locally compact quantum groups, Preprint arXiv:1511.03952 [math.OA]. [10] M. Kalantar, M. Neufang & Z.-J. Ruan: Poisson Boundaries over Locally Compact Quantum Groups, Int. J. Math. 24 (2013), 1350023-1 -- 1350023-21. [11] P. Kasprzak, A. Skalski & P.M. So ltan: Short exact sequence {e} / Z (G) / G / Inn(G) / {e} for locally compact quantum groups, Preprint arXiv:1508.02943 [math.OA]. [12] P. Kasprzak & P.M. So ltan: Embeddable quantum homogeneous spaces, J. Math. Anal. Appl. 411 (2014), 574-591. [13] P. Kasprzak & P.M. So ltan: Quantum groups with projection on von Neumann algebra level, J. Math. Anal. Appl. 427 (2015), 289 -- 306. [14] J. Kustermans: Locally compact quantum groups in the universal setting. Int. J. Math. 12 (2001) 289 -- 338. [15] J. Kustermans: Induced corepresentations of locally compact quantum groups, J. Funct. Anal. 194 (2002), 410 -- 459. / / / / 18 PAWE L KASPRZAK, FATEMEH KHOSRAVI, AND PIOTR M. SO LTAN [16] J. Kustermans & S. Vaes: Locally compact quantum groups, Ann. Scient. ´Ec. Norm. Sup. 4e s´erie, t. 33 (2000), 837 -- 934. [17] J. Kustermans & S. Vaes: Locally compact quantum groups in the von Neumann algebraic setting. Math. Scand. 92 (2003), 68 -- 92. [18] T. Masuda, Y. Nakagami & S.L. Woronowicz: A C∗-algebraic framework for the quantum groups, Int. J. Math. 14 (2003), 903 -- 1001. [19] R. Meyer, S. Roy & S. L. Woronowicz: Homomorphisms of quantum groups, Munster J. Math 5 (2012), 1 -- 24. [20] P. Podle´s & S.L. Woronowicz: Quantum deformation of Lorentz group. Comm. Math. Phys. 130 (1990), 381 -- 431. [21] P.M. So ltan: New quantum "az + b" groups, Rev. Math. Phys. 17 (2005), 313 -- 364. [22] P.M. So ltan: Examples of non-compact quantum group actions. J. Math. Anal. Appl. 372 (2010), 224 -- 236. [23] P.M. So ltan & S.L. Woronowicz: From multiplicative unitaries to quantum groups II, J. Funct. Anal. 252 (2007), 42 -- 67. [24] S, . Stratila & L. Zsid´o: Lectures on von Neumann algebras. Abacus Press 1979. [25] M. Takesaki & N. Tatsuuma: Duality and subgroups. Ann. Math. 93 (1971), 344 -- 364. [26] S. Vaes: Examples of locally compact quantum groups through the bicrossed product construction. In Pro- ceedings of the XIIIth International Congress of Mathematical Physics, Imperial College, London (2000), International Press of Boston, Somerville MA (2001), pp. 341 -- 348. [27] S. Vaes: Locally compact quantum groups. PhD thesis, Katholieke Universiteit Leuven, Faculteit Wetenschap- pen, Departement Wiskunde (2001). Available at http://perswww.kuleuven.be/~u0018768/PhD.html. [28] S. Vaes: The unitary implementation of a locally compact quantum group action, J. Funct. Anal. 180 (2001), 426 -- 480. [29] S. Vaes: A new approach to induction and imprimitivity results. J. Funct. Anal. 229 (2005), 317 -- 374. [30] L. Vainerman & S. Vaes: Extensions of locally compact quantum groups and the bicrossed product construction, Adv. Math. 175 (2003), 1 -- 101. [31] L. Vainerman & S. Vaes: On low-dimensional locally compact quantum groups. In Locally Compact Quantum Groups and Groupoids. Proceedings of the Meeting of Theoretical Physicists and Mathematicians, Strasbourg, February 21-23, 2002., Ed. L. Vainerman, IRMA Lectures on Mathematics and Mathematical Physics, Walter de Gruyter, Berlin, New York (2003), pp. 127 -- 187. [32] S.L. Woronowicz: Quantum E(2)-group and its Pontryagin dual, Let. Math. Phys. 23 (1991), 251 -- 263. [33] S.L. Woronowicz: Unbounded elements affiliated with C∗-algebras and non-compact quantum groups. Com- mun. Math. Phys. 136 (1991), 399 -- 432. [34] S.L. Woronowicz: C∗-algebras generated by unbounded elements. Rev. Math. Phys. 7, (1995), 481 -- 521. [35] S.L. Woronowicz: From multiplicative unitaries to quantum groups. Int. J. Math. 7, (1996), 127 -- 149. [36] S.L. Woronowicz: Quantum 'az + b' group on complex plane. Int. J. Math. 12, (2001), 461 -- 503. Department of Mathematical Methods in Physics, Faculty of Physics, University of Warsaw, Poland E-mail address: [email protected] Department of Pure Mathematics, Ferdowsi University of Mashhad, Iran E-mail address: [email protected] Department of Mathematical Methods in Physics, Faculty of Physics, University of Warsaw, Poland E-mail address: [email protected]
1403.1625	5	1403	2015-04-08T23:34:53	Remarks on multi-dimensional noncommutative generalized Brownian motions	[ "math.OA" ]	We consider certain questions pertaining to noncommutative generalized Brownian motions with multiple processes. We establish a framework for generalized Brownian motion with multiple processes similar to that defined by Guta and prove multi-dimensional analogs of some results of Guta and Maassen. We then consider examples of processes indexed by a two-element set and characterize the function on I-indexed pair partitions associated via the I-indexed generalized Brownian motion construction to certain pairs of representations connected to certain spherical representations of infinite symmetric groups. In doing so, we generalize the notion (introduced by Bozejko and Guta) of the cycle decomposition of a pair partition. We then generalize Guta's q-product of generalized Brownian motions to a product corresponding to a (possibly infinite) matrix $(q_{ij})$ and show that this $q_{ij}$-product satisfies a central limit theorem.	math.OA	math	REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS ADAM MERBERG Abstract. We consider certain questions pertaining to noncommutative generalized Brow- nian motions with multiple processes. We establish a framework for generalized Brownian motion with multiple processes similar to that defined by Gut¸a and prove multi-dimensional analogs of some results of Gut¸a and Maassen. We then consider examples of processes indexed by a two-element set and characterize the function on I-indexed pair partitions associated via the I-indexed generalized Brownian motion construction to certain pairs of representations connected to certain spherical representations of infinite symmetric groups. In doing so, we generalize the notion (introduced by Bozejko and Gut¸a) of the cycle decom- position of a pair partition. We then generalize Gut¸a's q-product of generalized Brownian motions to a product corresponding to a (possibly infinite) matrix (qij ) and show that this qij-product satisfies a central limit theorem. 1. Introduction Bozejko and Speicher initiated the study of noncommutative generalized Brownian mo- tions, introducing operators satisfying an interpolation between Fermionic and Bosonic com- mutation relations [BS91]. Specifically, for q ∈ [−1, 1] and a complex Hilbert space H, they constructed a q-twisted Fock space Fq(H) with creation operators c∗(f ) and annihilation operators c(f ) for f ∈ H satisfying the relations (1) c(f )c∗(g) − qc∗(g)c(f ) = hf, gi · 1. Subsequently, they developed a broader framework of generalized Brownian motion which incorporated this example [BS96]. In this general framework, one considers the algebra obtained by applying the GNS construction to the free tensor algebra of a real Hilbert space H with certain states, called Gaussian states, associated to a class of functions, called positive definite, on pair partitions via a pairing prescription. Gut¸a and Maassen further explored this notion of generalized Brownian motion [GM02a, GM02b]. They showed that Gaussian states ρt can be alternatively characterized by se- quences of complex Hilbert spaces (Vn)∞ n=1 with densely defined maps jn : Vn → Vn+1 and representations Un of the symmetric group Sn on Vn satisfying jn · U(π) = U(in(π)) · jn where in : Sn → Sn+1 is the inclusion arising from the map {1, . . . , n} ֒→ {1, . . . , n + 1}, data which give rise to a symmetric Fock space with creation and annihilation operators. They also provided an algebraic characterization of the notion of positive definiteness for a function t on pair partitions and characterized the functions t which give rise to analogs of the Gaussian functor. Separately [GM02b], they examined a class of Brownian motions arising from the combinatorial notion of species of structure. Bozejko and Gut¸a [BG02] considered a special case of the generalized Brownian motion of Gut¸a and Maassen arising from II1-factor representations of the infinite symmetric group S∞ constructed by Vershik and Kerov [VK82]. Lehner [Leh05] considered these generalized 1 2 ADAM MERBERG Brownian motions in the context of exchangeability systems, which generalize various notions of independence and give rise to cumulants analogous to the well-known free and classical cumulants. Recent work of Avsec and Junge [Avs13] offers another point of view on the subject of noncommutative Brownian motion. In [Gut¸03] Gut¸a extended the notion of generalized Brownian motion to multiple processes indexed by some set I. He went on to define for −1 ≤ q ≤ 1 a q-product of generalized Brownian motions interpolating between the graded tensor product previously considered by Mingo and Nica [MN97] (q = −1), the reduced free product [Voi85] (q = 0) and the usual tensor product (q = 1). He also showed that this q-product obeys a central limit theorem as the size of the index set I grows. In this paper, we explore certain additional questions pertaining to the I-indexed gener- alized Brownian motions. As a warmup, we begin with the very simple case of a generalized Brownian motion arising from a tensor product of representations of the infinite symmetric group S∞. We compute the functions on pair partitions associated to the Gaussian states in this context. We then proceed to consider the generalized Brownian motions associated to spherical representations of the Gelfand pair (S∞ × S∞, S∞). Here again we give a com- binatorial formula for the function on pair partitions arising from the associated Gaussian states, and in the course of doing so we generalize the notion of a cycle decomposition of a pair partition introduced by Bozejko and Gut¸a [BG02]. We also generalize Gut¸a's q-product of generalized Brownian motions to a qij product, where i, j ∈ I and show that a central limit theorem holds when qij = qji and the qij are periodic in both i and j. The paper has four sections, excluding this introduction. In Section 2, we expand upon the notion of generalized Brownian motion with multiple processes established by Gut¸a [Gut¸03], proving analogs of some results of Gut¸a and Maassen [GM02a]. We also review Vershik and Kerov's factor representations of symmetric groups [VK82] and Bozejko and Gut¸a's work on generalized Brownian motions with one process associated to the infinite symmetric group [BG02]. In Section 3, we move on to consider generalized Brownian motions indexed by a two-element set associated to tensor products of factor representations of the infinite symmetric group S∞. In Section 4, we consider generalized Brownian motions associated to spherical representations of (S∞ × S∞, S∞). In Section 6, we generalize Gut¸a's q-product to a qij product, where i, j ∈ I. Acknowledgments. While working on this paper, the author was supported in part by a National Science Foundation (NSF) Graduate Research Fellowship. He was also supported in part by funds from NSF grant DMS-1001881. The author also benefited from attending the Focus Program on Noncommutative Distributions in Free Probability Theory at the Fields Institute at the University of Toronto in July of 2013. The author's travel expenses for this conference were funded by NSF grant DMS-1302713. The author would like to thank Dan-Virgil Voiculescu for suggesting the problems and Stephen Avsec and Natasha Blitvi´c for a number of thoughtful discussions. 2. Preliminaries 2.1. Generalized Brownian motion. We begin by establishing the notion of a noncom- mutative generalized Brownian motion with multiple processes. Our framework for general- ized Brownian motion with multiple processes is slightly more general than that defined in [Gut¸03]. However everything in this section is in the spirit of results found in [Gut¸03] and [GM02a]. REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 3 1 2 3 4 5 6 -1 1 -1 Figure 1. The {−1, 1}-indexed pair partition (V, c) where V is the pair parti- tion {(1, 4), (2, 5), (3, 6)} and c(1, 4) = −1, c(3, 6) = −1 and c(2, 5) = 1. Solid lines represent the color −1 and dotted lines denote the color 1. Throughout the section, we assume that I is some fixed index set. In later sections we will specialize to specific index sets. Notation 1. We will make extensive use of notations for integer intervals, which appear frequently in the combinatorial literature: (2) for m, n ∈ Z. [m, n] := {m, m + 1, . . . , n − 1, n}; [n] := [1, n] = {1, 2, . . . , n − 1, n}. Notation 2. For a real Hilbert space K, AI(K) will denote the quotient of the free unital ∗-algebra with generators ωi(h) for h ∈ K and i ∈ I by the relations (3) ωi(cf + dg) = cωi(f ) + dωi(g), ωi(f ) = ωi(f )∗ for all f, g ∈ K, i ∈ I and c, d ∈ R. Notation 3. If H is a complex Hilbert space, CI(H) denotes the free unital ∗-algebra with generators ai(h) and a∗ i (h) for all h ∈ H and i ∈ I divided by the relations i (λf + µg) = λa∗ a∗ i (f ) = ai(f )∗, a∗ i (f ) + µa∗ i (g), (4) for all f, g ∈ H, i ∈ I, and λ, µ ∈ C. We will also use the notations a1 i (h) := a∗ a2 Consistent with the notation used in [GM02a], we will assume that the inner product on i (h) := ai(h) and i (h). a complex Hilbert space is linear in the second variable and conjugate linear in the first. Definition 1. If P is a finite ordered set, let P2(P ) be the set of pair partitions of P . That is, (5) P2(P ) :=({(l1, r1), . . . , (lm, rm)} : lk < rk, n {lk, rk} = P, {lp, rp} ∩ {lq, rq} = ∅ if p 6= q) . [k=1 The set of I-indexed pair partitions, P I 2 (P ) is the set of pairs (V, c) with V ∈ P2(P ) and c : V → I. We will sometimes refer to the elements of I as colors and the function c as the coloring function. If P ′ is another finite ordered set and α : P → P ′ is an order-preserving 2 (P ′). Considering all order-preserving bijection, then α induces a bijection P I bijections gives an equivalence relation on the union of P I 2 (P ) over sets of cardinality 2m. Let P I 2 (2m). 2 (2m) be the set of equivalence classes under this relation, and let P I 2 (P ) → P I n=1 P I 2 (∞) :=S∞ 4 ADAM MERBERG We can create a visual representation of an I-colored pair partition by connecting the pairs (lj, rj) by a path and labeling that path with the color c(lj, rj). When the number of colors is small, we may find it convenient to use different line styles to indicate colors, instead of an explicit label. Figure 1 gives the diagram for a simple example with I = {−1, 1}. It is clear that the coloring function c : V → I defines a function c : [2m] → I. It should not create confusion to refer to this function by the same name c. Note that c(l) = c(r) when (l, r) ∈ V. We will use these two descriptions of the coloring function c interchangeably. Definition 2. A Fock state on the algebra CI(H) is a positive unital linear functional ρt : CI(H) → C which for some t : P I 2 (∞) → C satisfies (6) ρt m Yk=1 aek ik (fk)! = X(V ,c)∈P I 2 (m) t(V, c) Y(l,r)∈V hfl, fri δilc(l,r)δir,c(l,r)Beler, where for k ∈ [m], fk ∈ H the ek are chosen from {1, 2} and ik ∈ I. (7) B :=(cid:18)0 1 0 0(cid:19) . Definition 3. A Gaussian state on AI(K) is a positive unital linear function ρt : AI(H) → C which for some t : P I 2 (∞) → C satisfies (8) ρt m Yk=1 ωik(fk)! = X(V ,c)∈P I 2 (m) t(V, c) Y(l,r)∈V any fk ∈ H (k ∈ [m]) and ik ∈ I. hfl, fri δil,c(l,r)δir,c(l,r), Remark 1. If K is a real Hilbert space, then there is a canonical injection AI(K) → CI(KC) given by (9) ωi(h) 7→ ai(h) + a∗ i (h), where KC denotes the complexification of K. Considering CI(KC) as a subalgebra of AI(K), the restriction of a Fock state ρt on CI(KC) to AI(K) is a Gaussian state. While we can use (8) and (6) to define linear functionals ρt and ρt for any choice of 2 (∞) → C, these linear functionals are not always positive. This leads to the following t : P I definition. Definition 4. A function t : P I functional on CI(K) for any complex Hilbert space K. 2 (∞) → C is positive definite if ρt is a positive linear Remark 2. Our definition of a positive definite function t : P I 2 (∞) → C is modeled after the definition made for a single process in [GM02a] and is not obviously the same as the definition in [Gut¸03]. However, we will see in Theorem 3 that the definitions are equivalent. Suppose that for each n : I → N ∪ {0}, Vn is a complex Hilbert space with unitary representation Un of the direct product group (10) Sn(b), Sn :=Yb∈I REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 5 If H is a complex Hilbert space, then the Fock-like space is given by (11) Here H⊗n means (12) FV (H) :=Mn 1 n! Vn ⊗s H⊗n. H⊗n :=Ob∈I H⊗n(b), and n! means Qb∈I n(b)!, and the factor 1 n! modifies the inner product. Also, we set H⊗0 = CΩ for some distinguished unit vector Ω. The notation ⊗s denotes the subspace of vectors which are invariant under the action of Sn given by Un ⊗ Un, where Un is the natural action of Sn on H⊗n. That is, Un(π) permutes the vectors according to π. The projection onto the subspace 1 n!Vn ⊗s H⊗n is given by (13) Pn := 1 n! Xσ∈Sn U(σ) ⊗ U (σ). For v ∈ Vn and f ∈ Hn we denote by v ⊗s f the vector Pn(v ⊗ f) To define creation and annihilation operators on the Fock-like space, we will also require densely defined operators jb : Vn → Vn+δb (where δb(b′) = δb,b′) satisfying the following intertwining relations: (14) jbUn(σ) = Un+δb(ι(b) n (σ))jb, where ι(b) n is the natural embedding of Sn into Sn+δb. Note that we have used the same notation for the maps Vn → Vn+δb for different n, but the choice of n should be clear from context so confusion should not result. We will call the maps jb the transition maps for our Hilbert spaces Vn. Given transition maps, we can define creation and annihilation operators on the Fock-like b (cid:17)∗ space FV (H) for each b ∈ I and each h ∈ H. Let (cid:16)r(n) (15) (h) : H⊗n → H⊗n+δb (h) be the operator b (cid:17)∗ (cid:16)r(n) which acts as right creation operator on H⊗n(b) and identity on H⊗n(b′) for b′ 6= b. Let r(n) b (cid:17)∗ be the adjoint of (cid:16)r(n) on the level n component of the Fock-like space by (h). If n(b) 6= 0, then the annihilation operator aV,j (h) b (f ) is defined b (16) aV,j b (f ) : Vn ⊗s H⊗n → Vn−δb ⊗s H⊗(n−δb) aV,j 1 ⊗ · · · ⊗ h(b) h(b) n(b) b (f ) : v ⊗sOb∈I Xk=1 7→ n(b) hf, hki n(b) j∗ b v ⊗s Ob′∈I\{b}(cid:16)h(b′) 1 ⊗ · · · ⊗ h(b′) n(b′)(cid:17) ⊗ h(b) 1 ⊗ · · · ⊗ h(b) k ⊗ · · · ⊗ h(b) n(b) . If n(b) 6= 0, then aV,j b (f ) (Vn ⊗s H⊗n) = 0. 6 ADAM MERBERG The creation operator (aV,j b )∗(h) is the adjoint of aV,j b (h), and its action on a vector v ⊗s f is given by (17) (aV,j b )∗(h) (vn ⊗s f) = (n(b) + 1)(jbvn) ⊗s (rn b )∗ (h)f. We denote by CV,j(H) the ∗-algebra generated by the operators aV,j f ∈ H, and b ∈ I. b (f ) and (aV,j b )∗(f ) for There is a representation νV,j of CI(H) on the Fock-like space FV (H) satisfying (18) νV,j : ab(f ) 7→ aV,j b (f ) and a∗ b(f ) 7→ (aV,j b )∗(f ) for all b ∈ I and f ∈ H. We will usually identify X ∈ CI(H) with its image νV,j(X). We will sometimes write ab(f ) and a∗ b(f ) for the annihilation and creation operators aV,j b (cid:17)∗ and(cid:16)aV,j aV,j,2 b (f ) or simply a2 b(f ) for (aV,j b )∗(f ). (f ). We will also use the notation aV,j,1 b (f ) or simply a1 b(f ) for aV,j The following is an I-indexed generalization of Theorem 2.6 of [GM02a]. b (f ) b (f ) and likewise Theorem 1. Let I be an index set and H a complex Hilbert space. Let (Un, Vn) be repre- sentations of Sn with maps jb : Vn → Vn+δb satisfying the intertwining relation (14). Let ξV ∈ V0 be a unit vector. The state ρV,j on CI(H) defined by (19) ρV,j(X) = hξV ⊗s Ω, X(ξV ⊗s Ω)i is a Fock state. That is, there is a positive definite function t such that ρV,j = ρt. The proof is very similar to the proof of Theorem 2.6 of [GM02a], but we include it for completeness. Proof. Let H be an infinite-dimensional complex Hilbert space, and let {fk}∞ k=1 be an or- thonormal basis for H. Also let V = {(lk, rk) : k ∈ [n]} with lk < rk for 1 ≤ k ≤ n and lk < lk′ for k < k′. Define (20) aek bk t((V, c)) = ρV,j n Yk=1 (fk)! =ξV ⊗s Ω, n Yk=1 aek bk (fk)! (ξV ⊗s Ω)+ where bk and ek are chosen as follows. Each k is an element of one pair (li, ri) ∈ V for some i. If k = li, we let ek = 1, and if k = ri we let ek = 2. In either case, we let bk = c(li, ri). For A ∈ B(H) and b ∈ I, define the operator dΓb V (A) on FV (H) by (21) (22) dΓb V (A) :v ⊗sOb′∈I hb′,1 ⊗ · · · ⊗ hb′,m′ b 7→ mb Xk=1 v ⊗s Ob′6=b V (A)(cid:3) = δb,b′ab′(A∗f ) (cid:2)ab′(f ), dΓb The operators dΓb V (A) satisfies the commutation relations and V (A), a∗ (cid:2)dΓb b′(f )(cid:3) = δb,b′a∗ b′(Af ). hb′,1 ⊗ · · · ⊗ hb′,mb′! ⊗ (hb,1 ⊗ · · · ⊗ Ahb,k ⊗ · · · ⊗ hb,mb) . REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 7 b,k for aV,j,e We write ae (fk), and denote by fi0ihfi the rank-one operator X on H which is 0 on the orthogonal complement of fi and such that Xfi = fi0. Applying (22) with A = fi0ihfi, we have when ik 6= i0 b (23) Consider a vector of the form φ = ae1 so that ab,i0φ = 0 for any b ∈ I. Applying (23), V (fi0ihfi), aek b1,i1 · · · aen (cid:2)dΓb b,ik(cid:3) = δik,i · δek,2 · a∗ b,i0 bn,in(ξV ⊗s Ω). Choose i0 different from i1, . . . , in, (24) Applying (23) repeatedly now yields V (fi0ihfi), aek ab,iφ =(cid:2)dΓb δi,ikδek,2δb,bk · ab,i0 k−1 Yr=1 Xk=1 b,ik(cid:3) φ = ab,i0dΓb b,i0 n br,ir! · a∗ Yr=k+1 aer n V (fi0ihfi)φ aer br ,ir! (ξV ⊗s Ω) (25) ab,iφ = Considering a monomial in creation and annihilation operators Qn , the theorem follows by applying (25) for each annihilation operator in the product with a new index i0. (cid:3) k=1 aek bk,ik There is also the following partial converse. Theorem 2. Let H be a separable, infinite-dimensional complex Hilbert space, and let I be a countable index set. Let t be a positive definite function on I-indexed pair partitions. Then there exist Hilbert spaces Vn with representations Un of Sn and densely defined transition maps jb : Vn → Vn+δb for b ∈ I satisfying (14) and a unit vector ξV ∈ V0 such that the GNS representation of (CI(H), ρt) is unitarily equivalent to (FV (H), CV,j(H), ξV ⊗s Ω). Remark 3. If we are given complex Hilbert spaces Vn with representations Un of Sn and maps jb : Vn → Vn+δb, then Theorem 1 gives a corresponding positive definite function t : P I n with representations U ′ . We will see in Example 1 that the Hilbert spaces V ′ 2 (∞) → C. Applying Theorem 2 gives complex Hilbert spaces V ′ n need not be the same as the original Hilbert spaces Vn. n of Sn and transition maps j′ n → V ′ b : V ′ n+δb The proof of Theorem 2 is very similar to the proof of Theorem 2.7 in [GM02a] (though we are only able to prove it for infinite-dimensional H). Regardless, we will include the proof of the I-indexed theorem here for the sake of completeness. Proof of Theorem 2. Choose an orthonormal basis {fk,b}k∈N,b∈I for H. Let Ft(H), Ct(H), and Ωt be the complex Hilbert space, operator algebra, and distinguished unit vector of the GNS construction of CI(H) with respect to the state ρt. We denote the image of ab(f ) in b(f ) in Ct(H) by (at Ct(H) by at b (f ) to mean (at)∗ b(f ) for e = 1. We will construct the complex Hilbert spaces Vn as subspaces of Ft(H). b(f ) and the image of a∗ b (f ) for e = 2 and at b)∗(f ). We will use the notation at,e Suppose that for each function n : I → N∪{0} which is 0 at all but finitely many points in I and each i ∈ I, we have an injective function αn,i : [n(i)] → N and that αn,i(j) = αn′,i(j) when j < n(i), n′(i). Denote by Rα n the set of the vectors of the form (26) at,e1 b1 (fb1,i1) · · · a t,e2p+n b2p+n (fb2p+n,i2p+n)Ωt, where n =Pb∈I n(b) satisfying the following conditions: 8 ADAM MERBERG (1) In the product at,e1 b1 (fb1,i1) · · · a t,e2p+n b2p+n (fb2p+n,i2p+n), a creation operator at,2 b (fb,αb(j)) appears exactly once provided that 1 ≤ j ≤ n(b). bq (fbq ,lq))p q=1 and p annihilation operators (at,1 (2) Among the remaining 2p operators in the product, there are p creation operators (at,2 q=1. Moreover, each anni- hilation operator appears to the left of the corresponding creation operator in the product. We also let V α restricting the creation operator at,2 immediately from the definition of V α n+δb′ by n of Ft(H). It follows n that the image of this restriction lies in V α b′ (fb′,αn,b′ (n(b′))) to the subspace V α bq (fbq ,lq))p n be the span of the vectors in Rα n. We define the map jα b′ : V α n → V α We define a unitary representation U α n of Sn on V α n . Since ρt is a Fock state, it is invariant n+δb′ . under unitary transformations U on H in the sense that (27) ρt n Yk=1 aek bk (fbk,ik)! = ρt n Yk=1 aek bk (Ufbk,ik)! . Therefore, there is a unitary map Ft(U) given by (28) Ft(U) : at,ek bk (fbk,ik)Ωt 7→ aek bk (Ufbk,ik)Ωt. n n Yk=1 Yk=1 The map Ft(U) induces an automorphism on the algebra of creation and annihilation oper- ators by (29) Ft(U)at,ek bk (h)Ft(U ∗) = at,ek bk (Uh). For σ ∈ Sn, let U α {fb,αn,b(1), . . . , fb,αn,b(n(b))} according to σ and fixes fb,r when r > n(b). The map U α is a unitary representation of Sn on V α n . σ be the unitary operator on H which for each b ∈ I acts by permuting n : σ 7→ U α σ Define ιn,i : [n(i)] → N by ιn,i(j) = j, and let Rn := Rι b, and let n . It follows from the definitions that these data satisfy the intertwining property n, Vn := V ι n, jb := jι Un := U α (14). We also define ξV := Ωt ∈ V0. We now show that (FV (H), CV,j(H), ξV ⊗s Ω) is unitarily equivalent to the GNS represen- tation of (CI(H), ρt). We will begin by showing that ρt = ρV,j. By Theorem 1, ρV,j is a Fock state associated to some positive definite function t′ : P I 2 (∞) → C, so it will suffice to show that t′ = t. For the proof, we will also need to define for a unitary map U on H, FV (U) : FV (H) → FV (H) (30) v ⊗s Ob∈I hb,1 ⊗ · · · ⊗ hb,n(b)! 7→ v ⊗s Ob∈I Uhb,1 ⊗ · · · ⊗ Uhb,n(b)!, for all v ∈ Vn. This induces an action on the creation and annihilation operators satisfying (31) For αn,i as before, define FV (U)ae V,j(f )FV (U ∗) = ae V,j(Uf ) (32) V α n := span{v ⊗sOb∈I fb,α(1) ⊗ · · · ⊗ fb,α(n(b))}. REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 9 Define an isometry Tn : Vn → FV,j(H) by v 7→ v ⊗sOb∈I fb,1 ⊗ · · · ⊗ fb,n(b). Let Uα,n be a unitary map on H which permutes the basis vectors fb,j such that Uαfn,j = fn,α(j) whenever 1 ≤ j ≤ n(b). Define a map T α (34) n := FV(Uα,n)TnFt(U ∗ T α n → V α α,n). n : V α n by This map does not depend on the choice of Uα,n permuting the basis vectors according to α. It follows immediately from the definitions that the diagram Vn Tn−−−→ Vn is commutative, whence the diagram (36) Vn+δb Tn+δb−−−→ Vn+δb V α n T α n−−−→ Vnα (aV,j b (aV,j b )∗(fb,n(b)+1) )∗(fb,αn,b(n(b)+1)) jb y jα b y V α n+δb T α n+δb−−−→ Vn+δb y y also commutes. A similar argument gives a corresponding commutative diagram for the annihilation operators, and this implies the equality of the states ρt and ρV,j, which implies t = t. Finally, we must prove that the vacuum vector ΩV := ξV ⊗ Ω is cyclic for CI(H). It will suffice to show that for any n, any v ∈ Rn and any vectors h1, . . . , hn ∈ H, there is some X ∈ CI(H) with (33) (35) (37) (38) By the definition of Rn, we can write XΩV = v ⊗sOb∈I hb,1 ⊗ · · · ⊗ hb,n(b). v =  2p+n Yk=1 at,ek bk (fbk,ik)  Ωt where a creation operator at,2 b (fb,j) appears exactly once for 1 ≤ j ≤ n(b), and among the remaining 2p operators in the product, there are p creation operators (at,2 q=1 and p annihilation operators at,1 q=1, with each annihilation operator appearing to the left of the corresponding creation operator in the product. We need simply choose X of the form bq (fbq ,lq))p bq (fbq ,lq)p (39) where the gk satisfy: (40) X := 1 n! · 2p+n Yk=1 aek bk (gk), gk :=(hbk,rk, h′ lk , if 1 ≤ ik ≤ n(bk) otherwise , 10 ADAM MERBERG where rk is defined so that k is the rk-th smallest element of the set {u : bu = bk, 1 ≤ u ≤ n(bk)}, the (h′ i=1 are an orthonormal sequence of vectors which are orthogonal to each hb,k, and lk = l′ k if and only if bk = bk′ and ik = ik′. It follows from the definitions that this X satisfies (37), so the proof is complete. (cid:3) i)∞ We now pursue an algebraic characterization of positive definiteness for functions on I- indexed pair partitions. This will involve Gut¸a's ∗-semigroup of I-indexed broken pair partitions [Gut¸03]. Definition 5. Let X be an arbitrary finite ordered set and (La, Pa, Ra)a∈I a disjoint partition of X into triples of subsets indexed by elements of I. Suppose that for each a ∈ I, we have a triple (Va, f (l) (41) a , f (r) a ) where Va ∈ P2(Pa) and f (l) a : La → {1, . . . , La} and f (r) : Ra → {1, . . . , Ra} a are bijective. An order preserving bijection α : X → Y induces a map a ◦ α−1) a ) → (α ◦ Va, f (l) a ◦ α−1, f (r) αa : (Va, f (l) a , f (r) (42) where α ◦ V := {(α(i), α(j)) : (i, j) ∈ V}. This determines an equivalence relation on the set of such I-indexed triples, and we call an equivalence class under this relation an I-indexed broken pair partition. We denote the set of all I-indexed broken pair partition by BP I 2 (∞). There is a convenient diagrammatic representation of a broken pair partition. Given a broken pair partition as just defined, we write the elements of the base set X in order. For each pair (x, x′) ∈ Va, we connect x and x′ by a piecewise-linear path and label that path with the index a. For each index a ∈ I such that La 6= ∅, we write the numbers 1, . . . , La in order on the left side and connect each y ∈ La to the number f (l) a (y). Likewise, for each color a ∈ I such that Ra 6= ∅, we write the numbers 1, . . . , Ra in order on the left side and connect each y ∈ Ra to the number f (l) a (y). When I is small, we may also use different line styles (e.g. dotted and solid lines) to indicate the different colors a ∈ I. Figure 2 gives two examples of these diagrams. a a and f (r) to the numbers f (l) The diagrammatic representations of the I-colored broken pair partitions inspires some additional terminology. Namely, we call the functions f (l) the left and right leg functions for the color a. Moreover, we call the piecewise-linear paths from the domains of f (l) a and f (r) a (y) the left and right legs of the I-colored broken pair partitions. This terminology will be useful in describing the semigroup structure on BP I 2 (∞). In the case that I = 1, we recover the (uncolored) broken pair partitions of Gut¸a and Maassen [GM02a]. Moreover, each d ∈ BP I 2 (∞) gives for each a ∈ I a broken pair partition da in the sense of [GM02a]. However, all but finitely many of the da are the unique broken pair partition on the empty set. a (y) and f (r) a The space BP I 2 (∞) can be given the structure of a semigroup with involution, similar to the ∗-semigroup of broken pair partitions of [GM02a]. In terms of the diagrams, multiplication of two I-colored broken pair partitions corresponds to concatenation of diagrams. Right legs of the first diagram are joined with left legs of the second diagram of the same color to form pairs. In the event that the second diagram has more left legs of some color a than the first diagram has right legs of color a, we join the right legs of the first diagram with the largest-numbered left legs of the second diagram, and the remaining left legs of the second diagram are extended to become low-numbered left legs in the product. An analogous rule REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 11 1 2 3 4 1 2 3 4 5 6 d = 1 1 ¯d = 1 2 1 1 a , f (r) a , ¯f (r) a }a∈{−1,1} and ¯d = { ¯Va, ¯f (l) Figure 2. The diagram of the {−1, 1}-colored broken pair partitions d and ¯d. Here d = {Va, f (l) a }a∈{−1,1}, with the following definitions. For d, V−1 = V1 is the unique pair partition on the −1 : ∅ → ∅,f (l) empty set and the right and left leg functions are defined by f (l) : {2} → {1} given by f (l) −1 : {1, 3} → {1, 2} given by f (r) −1 (1) = 2 and f (r) −1 (3) = 1, and f (r) 1 (4) = 1. For ¯d, ¯V2 = {(1, 4)}, ¯V1 = {(3, 6)} and the right and left leg functions are defined by ¯f (l) −1 : {2} → {1} with ¯f (l) : {5} → {1} given by ¯f (r) 1 (5) = 1. The solid lines represent the "color" −1 and the dotted lines represent the "color" 1. 1 (2) = 1, f (r) : {4} → {1} given by f (r) : ∅ → ∅, and ¯f (r) −1(2) = 1, ¯f (l) : ∅ → ∅, ¯f (r) 1 1 1 1 1 is used when the first diagram has more right legs of some color a than the second diagram has left legs of color a. The precise definition of the product on BP I 2 (∞) is as follows. For i = 1, 2, let di = a,i )a∈I be an I-colored broken pair partition on the ordered base set Xi. The a,i , f (r) (Va,i, f (l) product is a broken pair partition on the base set X := X1` X2 with the order relation x < x′ if either x < x′ in Xi or x ∈ X1 and x ∈ X2. For each a ∈ I, define Ma = min(Ra,1, La,2). Following [Gut¸03], we define (43) where (44) and f (l) a (45) d1 · d2 = (Va, f (l) a , f (r) a )a∈I, Va = Va,1 ∪ Va,2 ∪n(cid:16)(f (r) a,1 )−1([Ra,1 − j]), (f (l) is defined on the disjoint union of La,1 and (f (l) a,2)−1([La,2 − j])(cid:17) : j ∈ [Ma]o a,2)−1([La,2 − Ma]) by f (l) a (i) =(f (l) a,1(i), f (l) a,2(i) + La,1 − Ma, if i ∈ La,1 if i ∈ (f (l) a,2)−1([La,2 − Ma]) . The function of right legs, f (r) Ma]) by a is defined on the disjoint union of Ra,2 and (f (r) a,1 )−1([Ra,1 − (46) f (r) a (i) =(f (r) a,2 (i), f (r) a,1 (i) + Ra,2 − Ma, if i ∈ Ra,2 if i ∈ (f (r) a,1 )−1([Ra,1 − Ma]) . An example of multiplication of I-colored broken pair partitions is illustrated in Figure 3.1 The involution is given by mirror reflection of the I-colored broken pair partitions. For- a )a∈I is an a )a∈I with underlying set X then d∗ = (V ∗ mally, if d = (Va, f (l) a , f (l) a , f (r) a, f (r) 1The multiplication for BP I 2 (∞) stated here differs slightly from that stated in [Gut¸03]. We believe that the rule stated here is the one intended by the author of that work, as it ensures that condition (14) is satisfied. However, we do not believe that this discrepancy is consequential for Gut¸a's results. 12 1 1 2 3 4 1 2 3 4 65 1 2 3 4 5 6 87 9 10 ADAM MERBERG · 1 1 2 1 = 1 1 1 1 2 Figure 3. The multiplication d · ¯d of the {−1, 1}-colored broken pair parti- tions defined in Figure 2. 1 2 3 4 5 6 1 ¯d∗ = 1 Figure 4. The involution of the {−1, 1}-colored broken pair partition ¯d de- picted in Figure 1. 1 2 3 4 5 6 7 8 9 10 Figure 5. The standard form of the pair partition (V, c) ∈ P I 2 (10) with I = {−1, 1} and V = {(1, 5), (2, 8), (3, 6), (4, 10), (7, 9)} and c((1, 5)) = c((2, 8)) = c((7, 9)) = −1 and c((3, 6)) = c((4, 10)) = 1. The solid lines represent the "color" −1 and the dotted lines represent the "color" 1. I-colored broken pair partition with underlying set X ∗, the same as X but with the order reversed, where V ∗ a = {(i, j) : (j, i) ∈ Va}. The involution is illustrated in Figure 4. For each function n : I → N which is zero except on finitely many elements of I, let 2 (n, 0) be the subset of BP I 2 (∞) consisting of elements having exactly n(a) left legs of BP I color a and no right legs. Let da be the unique element of BP I 2 (∞) with no right legs, no pairs, and only one left leg, colored a ∈ I. We call da the a-colored left hook and d∗ a the a-colored right hook. An I-colored broken-pair partition can be written as a sequence of left hooks, followed by permutations acting on the legs of the same color, followed by right hooks connecting with left legs of the same index, possibly followed by additional sequences of left hooks, permutations, and right hooks. The "standard form" of an element of P I 2 (∞) is the sequence of this form such that if two like-colored pairs cross, they do so in the rightmost permutation possible. Figure 5 depicts the standard form of one example. As in [Gut¸03], we can use the standard form of (V, c) for V ∈ P2(2m) to compute the value of t((V, c)) as follows. Consider c as a function [2m] → I taking the same value on points belonging to the same pair of V. Partition [2m] into 2t blocks B(r) for i = 1, . . . , t such that the B(l) contain right legs of the pairs of contain left legs of the pairs of V and the B(r) i , B(l) i i i REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 13 V. Write B(r) there are permutations πj such that := {ki−1, . . . , pi} and B(l) j i := {pi+1, . . . , ki} with k0 = 1 and kr = 2m. Then (47) (V, c) = dc(l) · · · Unr (πr) p1 k1 d∗ c(l)Un1(π1) Yl=p1+1 Yl=1 2m Yl=pt+1 dc(l) The function on pair partitions can then be calculated as (48) tV,j((V, c)) =ξV , Yl=1 p1 k1 2m j∗ c(l)Un1(π1) jc(l) · · · Unr (πr) Yl=p1+1 jc(l)ξV+ . Yl=pt+1 An I-colored pair partition V can be considered as an element of BP I 2 (∞) having no left 2 (∞) → C thus extends to a function or right legs in the obvious way. A function t : P I t : BP I 2 (∞) → C by (49) t(d) =(t(d), 0, if d ∈ P I otherwise 2 (∞), . The following is an I-indexed generalization of Theorem 3.2 of [GM02a]. 2 (∞) → C is positive definite if t : BP I Theorem 3. A function t : P I 2 (∞) → C is posi- tive definite in the usual sense of positive definiteness for a function on a semigroup with involution. The proof is very similar to the proof of Theorem 3.2 of [GM02a], but we include it for completeness. Proof. Suppose that t is positive definite on the ∗-semigroup BP I representation χ of BP I that 2 (∞). Then there is a 2 (∞) on a complex Hilbert space V having cyclic vector ξ ∈ V such (50) for all d ∈ BP I 2 (∞). (51) The complex Hilbert space V is expressible as a direct sum hξ, χ(d)ξi = t(d) Vn, V =Mn where the sum is over functions n : I → N with only finitely many nonzero values and Vn = span{χt(d)ξ : d ∈ BP I (52) The action of Sn on BP I 2 (n, 0) (by permutation of the left legs) gives a unitary representation Un of Sn on Vn. Restriction of jb := χ(db) (where, as before db is the b-colored left hook) gives a map jb : Vn → Vn+δb satisfying (14). Choose a unit vector ξV ∈ V0. 2 (n, 0)}. Let H be an infinite-dimensional complex Hilbert space. Using the Un, Vn and jb, we can construct the Fock space FV (H) and the algebra CV,j(H) with vacuum vector ΩV . By Theorem 1, the vacuum state is a Fock state arising from some positive definite function t′ : BP I 2 (∞) → C. It will suffice to show that t′ = t, whence it will follow that t is positive definite. In fact, this follows from Theorem 2.3 of [Gut¸03], but we also provide a proof for completeness. 14 ADAM MERBERG Given an I-indexed pair partition (V, c) ∈ P I 2 (2m) with V = {(l1, r1), . . . , (lm, rm)}, let the standard form of (V, c) be p1 k1 (53) (V, c) = d∗ c(l)Un1(π1) dc(l) · · · Unr (πr) Yl=1 2m Yl=pt+1 dc(l). Let H = ℓ2(Z) have an orthonormal basis (fk)∞ k=1, and choose a monomial (54) M := aek bk (fik), Yl=p1+1 2m Yk=1 where ik is chosen such that either k = lik or k = rik, and bk = c(lik, rik), and ek = 1 if k = lik and ek = 2 if k = rik. From the definition of a Fock state, (55) t′(V, c) = hΩV , MΩV i . Using the definition of the creation operator and (14), we get (56) t′(V, c) = d∗ c(l)Un1(π1) p1 Yl=1 = t(V, c) = t(V, c). k1 Yl=p1+1 dc(l) · · · Unr (πr) 2m Yl=pt+1 dc(l) This completes the proof that t : P I 2 (∞) → C is positive definite. Suppose now that t : P I 2 (∞) → C is positive definite. Applying Theorem 2 we get complex Hilbert spaces Vn with representations Un of Sn and densely defined maps jb : Vn → Vn+δb satisfying the intertwining relation (14). Let Vn. V =Mn (57) (58) (59) Since the left hooks {db : b ∈ I} and the actions λn of the symmetric groups Sn on BP I generate BP I hξ, χ(d)ξi = t(d) for any unit vector ξ ∈ V0 and any d ∈ BP I definite. 2 (n, 0) 2 (∞), and it is easily verified that 2 (∞), whence t is positive (cid:3) 2 (∞), we have a representation χ of BP I Gut¸a [Gut¸03] considered the case in which the Vn and the ja are defined as follows. Let 2 (n, 0) the set 2 (∞) having Ra = 0 and La = n(a) for each a ∈ I. Consider the GNS 2 (∞) → C be a positive definite function. As before, denote by BP I t : P I of d ∈ BP I representation (χt, V, ξt) of BP I 2 (n, 0) with respect to t, characterized by hχt(d1)ξt, χt(d2)ξtiV = t(d∗ 1d2). The complex Hilbert space V is given by Vn where Vn = span{χt(d)ξt : d ∈ BP I 2 (n, 0)}. V :=Mn Each Vn has a representation of Sn with Sn(a) acting by permuting the left legs of BP I of color a. That is for π = (πa)a∈I ∈ Sn and (Va, f (l) a )a∈I ∈ BP I a , f (r) 2 (n, 0), 2 (n, 0) (60) Un(π)(Va, f (l) a , f (r) a )a∈I = (Va, π−1 a ◦ f (l) a , f (r) a )a∈I. REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 15 We denote by Ft(H) the Fock-like space arising from using these Vn in (11) (61) Ft(H) :=Mn 1 n! Vn ⊗sO H⊗n. The creation and annihilation operators on Ft(H) will be assumed to be those associated to the following operators ja. For a ∈ I, denote by ja the operator χt(da), where da is the broken pair partition with no pairs, no right legs, and one a-colored left leg. We can now state the following theorem of Gut¸a [Gut¸03]. Theorem 4. Let f1, . . . , fn be vectors in a complex Hilbert space H. Then the expectation values with respect to the vacuum state ρt of the monomials in creation and annihilation operators on the Fock space Ft(H) have the expression (62) ρt m Yi=1 aei bi (fi)! = X(V ,c)∈P I 2 (n) t(V, c) Y(i,j)∈V hfi, fji δbi,bj Beiej , where the ei are chosen from {1, 2} and (63) B :=(cid:18)0 1 0 0(cid:19) . 2.2. Factor representations of S∞. In Sections 3 and 4, we will take an interest in non- commutative generalized Brownian motions which are related to factor representations of the group S∞ of permutations of N which fix all but finitely many points. Here we briefly recall some relevant background information pertaining to those representations. The finite factor representations of a group are determined by the group's characters, that is, the positive, normalized indecomposable functions which are constant on conjugacy classes. In the case of S∞, the characters are given by the following famous result. Theorem 5 (Thoma's Theorem [Tho64]). The normalized finite characters of S∞ are given by the formula (64) φα,β(σ) = Ym≥2 ∞ Xi=1 i + (−1)m+1 αm βm i !ρm(σ) ∞ Xi=1 where ρm(σ) is the number of cycles of length m in the permutation σ, and (αi)∞ i=1 and (βi)∞ i=1 are decreasing sequences of positive real numbers such that Pi αi +Pi βi ≤ 1. The pairs of sequences (αi)∞ i=1 and (βi)∞ i=1 satisfying the conditions in Theorem 5 are commonly called Thoma parameters. We now recall Vershik and Kerov's representation of the symmetric group Sn (for n ∈ {0, 1, 2, . . . , ∞}) [VK82]. i=1 and (βi)∞ Notation 4. Fix sequences (αi)∞ N− be two copies of the set N = {1, 2, . . .}. Let Q := N+ ∪N− ∪[0, γ], and define a measure µ on Q to be the Lebesgue measure on [0, γ] and such that µ(i) = αi for i ∈ N+ and µ(j) = βj for j ∈ N−. Let Xn denote the n-fold Cartesian product of Q with the product measure 1 µ, and let Sn act on Xn by σ(x1, . . . , xn) = (xσ−1(1), . . . , xσ−1(n)). For x, y ∈ Xn, i=1, and let γ = 1−Pi αi −Pi βi and let N+ and mn = Qn 16 ADAM MERBERG say that x ∼ y if there exists σ ∈ Sn such that x = σy. Let Xn = {(x, y) ∈ Xn × Xn : x ∼ y}. The complex Hilbert space V (α,β) defined by n (65) V (α,β) n :=(f : Xn → C∞ > kf k2 =ZXnXy∼x f (x, y)2dm(α,β) n (x)) carries a unitary representation U (α,β) n of S(n) given by (66) (U (α,β) n (σ)h)(x, y) = (−1)i(σ,x)h(σ−1x, y), where i(σ, x) is the number of inversions in the sequence (σi1(x), σi2(x), . . .) of indices ir(x) for which σxi ∈ N−. Denote by 1n the indicator function of the diagonal {(x, x)} ⊂ Xn. Vershik and Kerov showed the following. Theorem 6 ([VK82]). On V (α,β) n , (67) n (cid:10)U (α,β) (σ)1n, 1n(cid:11) = φα,β(σ). For n = ∞ we get the representation of S∞ associated to φα,β in the convex hull of 1∞. There is an isometry jn : V (α,β) n → V (α,β) n+1 defined by (68) (jnh)(x, y) = δxn+1,yn+1h((x1, . . . , xn), (y1, . . . , yn)) 2.3. Generalized Brownian motions arising from factor representations of S∞. Before considering multi-dimensional noncommutative generalized Brownian motions asso- ciated to representations associated to infinite symmetric groups, we review some of the work of [BG02] on Brownian motions connected with representations of S∞ with one pro- cess. The Vershik-Kerov factor representations of the symmetric groups Sn give all the data needed for a 1-dimensional generalized Brownian motion. Bozejko and Gut¸a [BG02] were able to characterize the function on pair partitions arising from Theorem 2.6 of [GM02a] (the one-dimensional version of Theorem 1). Their result depends on the following terminology. Definition 6 ([BG02]). Let V ∈ P2(2m), and denote by V the unique noncrossing pair partition such that the set of left points of V and V coincide. A cycle in V is a sequence of pairs ((l1, r1), . . . , (lm, rm)) of V such that the pairs (l1, r2), (l2, r3), . . . , (lm, r1) belong to V. (In the case that m = 1 we interpret this condition as (l1, r1) ∈ V.) The number m is called the length of the cycle. Denote by ρm(V) the number of cycles of length m in the pair partition V. Bozejko and Gut¸a's formula is as follows. i=1 and (βi)∞ Theorem 7 ([BG02]). Let (αi)∞ such that Pi αi +Pi βi ≤ 1. Let V (α,β) i=1 be decreasing sequences of positive real numbers be the complex Hilbert space of the Vershik-Kerov representation of Sn, and let jn : V (α,β) be the natural isometry. Let ξV (α,β) = 10. Denote by tα,β the function on P2(∞) associated to these representations by Theorem 1. Then n → V (α,β) n+1 n (69) tα,β(V) = Ym≥2 ∞ Xi=1 i + (−1)m+1 αm βm i !ρm(V) . ∞ Xi=1 REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 17 Remark 4. There is another equivalent characterization of the cycle decomposition of a pair partition. As in Definition 6, let V = {(a1, z1), . . . , (an, zn)} ∈ P2(2m) and let V be the noncrossing pair partition whose left points coincide with those of V. Let σ ∈ Sn be the permutation such that V = {(a1, zσ−1(1)), . . . , (an, zσ−1(n))}. If the cycles of σ are τi = (bi1 · · · biri) ∈ Sn (1 ≤ i ≤ m), then the cycles of V are {(abi1, zbi1), . . . , (abiri )}. Moreover, Theorem 7 says that tα,β(V) = φα,β(σ). , zbiri We are now in a position to show that the framework for multi-dimensional generalized Brownian motion presented here is more general than that presented in [Gut¸03]. More precisely, we will exhibit complex Hilbert spaces Vn with representations Un of Sn and transition maps jb : Vn → Vn+δb and V ′ n → V ′ such that both sets of data give rise to the same function on pair partitions according n+δb to Theorem 1. n of Sn and maps jb : V ′ n with representations U ′ Example 1. We work with the index set I = {1}, which places us in the setting of the generalized Brownian motion with only one process, developed by Gut¸a and Maassen in [GM02a]. Fix an integer N with N > 1 and let H be a complex Hilbert space. We will consider generalized Brownian motions associated to the character φN of S∞ given by the sequences (70) αn =( 1 N , 0, if 1 ≤ n ≤ N otherwise and βn =( 1 N , 0, if 1 ≤ n ≤ −N otherwise n For each n, let V (N ) be the complex Hilbert space of the Vershik-Kerov representation of Sn, and let j(N ) : V (N ) n+1 be the natural isometry. Let ξV (N) = 10 be the indicator function of the diagonal. Denote by tN the function on P2(∞) associated to these representations by Theorem 1. It was shown in [BG02] that n → V (N ) (71) tN (V) =(cid:18) 1 N(cid:19)n−ρ(V) . We will exhibit another sequence of complex Hilbert spaces V (N ) n with unitary represen- tations U (N ) n which gives rise to the same positive function on pair partitions. Since the character φN : S∞ → C restricts to a positive definite function on Sn, there is a representa- tion U (N ) of Sn on a complex Hilbert space V (N ) n with a cyclic vector ξn such that n (72) Dξn, U (N ) n (π)ξnE = φN (π). for every π ∈ Sn. There is also a natural inclusion j(N ) : V (N ) n → V (N ) n+1 satisfying (73) j( U (N ) n (π)ξn) = U (N ) n+1(ιnπ)ξn+1, where ιn is the inclusion Sn → Sn+1 induced by the natural inclusion [n] ⊂ [n + 1]. By con- struction, the maps j(N ) and representations V (N ) satisfy the intertwining relation (14), so we can construct the Fock space F V (N),j(N) with creation and annihilation operators a∗ V (N),j(N)(f ) and a V (N),j(N)(f ). n The action of the algebra C V (N),j(N)(H) on F V (N),j(N)(H) is unitarily equivalent to the action of the algebra creation and annihilation operators on the following deformed Fock space. Let 18 ADAM MERBERG F (alg)(H) =Ln H⊗n. Define a sesquilinear form on F (alg)(H) by sesquilinear extension of (74) hf1 ⊗ · · · fn, g1 ⊗ · · · gmiN = δmn Xπ∈Sn φN (π)(cid:10)f1, gπ(1)(cid:11) · · ·(cid:10)fn, gπ(n)(cid:11) . This form is positive definite and thus gives an inner product on F (alg)(H). Let FN (H) be the completion of F (alg)(H) with respect to the inner product h·, ·iN . Let DN be the operator in F (alg)(H) whose restriction to H⊗n is given by (75) D(n) N :=(1 + 1 1, k=2 N Pn Un(τ1,k), if n > 0 otherwise, where τi,k ∈ Sn is the permutation transposing i and k and fixing all other elements of [n] and Un is the representation of Sn such that Un(π) permutes the tensors in H⊗n according to π. For f ∈ H let l(f ) and l∗(f ) denote the left annihilation and creation (respectively) operators on the free Fock space over H. We define annihilation and creation operators on F (alg)(H) by (76) aN (f ) = l(f )DN a∗ N (f ) = l∗(f ). It was shown in [BG02] that these operators are bounded with respect to h·, ·iN and thus extend to bounded linear operators on FN (H). The map (77) FN (H) → F VN (H) v1 ⊗ · · · vn 7→ ξn ⊗s vn ⊗ · · · ⊗ v1 is unitary. It was shown in [BG02] that the vacuum state on the algebra of creation and anni- . hilation operators on FN (H) is the Fock state associated to the function tN (V) =(cid:0) 1 This shows that tV (N),j(N) = t V (N),j(N) even though dim V (N ) N(cid:1)n−ρ(V) n < dim V (N ) for n ≥ 1. n 3. Generalized Brownian motions associated to tensor products of representations of S∞ In this section , we are interested in the case where I = {−1, 1} and the Vn arise from unitary representations of the group S∞ of permutations of N fixing all but finitely many points. (78) fn(x) =(1, 0, otherwise if x = qi for 1 ≤ i ≤ n Notation 5. When I = {1, −1}, we will represent a function n : I → N by the pair n(−1), n(1), so we write Vr,s for Vn where n(−1) = r and n(1) = s. We will also represent an element (V, c) of P I 2 (∞) as (V−1, V1), where Vb = c−1(b). One of the simplest such cases is that arising from the tensor product of two unitary representations of S∞. In this setting, we can prove the following. Proposition 1. Let (U (i), V (i)) be unitary representations of S∞ for i ∈ I. Suppose that each V (i) n → V (i) n+1 is a subspace of V (i) carrying a unitary representation U (i) n of Sn with j(i) n : V (i) n REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 19 an isometry. Assume that we have distinguished unit vectors ξV (i) ∈ V (i) ξV (−1) ⊗ ξV (1). Let Vm,n = V (−1) n , j−1 = j(−1) ⊗ 1, and j1 = 1 ⊗ j(1). Then m ⊗ V (1) 0 and let ξV = (79) tV,j(V, c) = tV (−1),j(−1)(V−1) · tV (1),j(1)(V1). Proof. From the definitions, it is clear that for any v1 ∈ V (−1) m and v2 ∈ V (1) n , (80) −1(v1 ⊗ v2) = j∗ j1j−1(v1 ⊗ v2) = j−1j1(v1 ⊗ v2) −1j∗ 1j∗ j∗ 1(v1 ⊗ v2) 1j−1(v1 ⊗ v2) = j−1j∗ j∗ 1(v1 ⊗ v2) j1j∗ −1j1(v1 ⊗ v2). −1(v1 ⊗ v2) = j∗ Consequently, if b 6= b′ the operators aV,j,e and all f, f ′ ∈ H. b (f ) and aV,j,e′ b′ (f ′) commute for all e, e′ ∈ {1, 2} Assume that V := {(l1, r1), . . . , (lm, rm)} with lk < rk and lk < lk+1 for all k. Let H be ℓ2(N) with orthonormal basis (hk)∞ k=1 We can compute tV,j(V, c) as (81) tV,j(V, c) =* 2m Yp=1 aV,j,ep c(p) (hkp)! ξV ⊗s Ω, ξV ⊗s Ω+ where kp is the unique k ∈ [n] such that p is an element of the k-th pair of V and ep = 2 if p is a right point in V and ep = 1 if p is a right point. c(p)(hkp) commutes with a ep′ c(p′)(hkp′ ) when c(p) 6= c(p′), we have Now, using the fact that aer tV,j(V, c) =  Yc(p)=−1 =Yb∈I Yc(p)=b aV,j,ep −1 aV,j,ep 1  (hkp)  Yc(p)=1  (hkp)ξV (b) ⊗s Ω, ξV (b) ⊗s Ω+ (hkp)  ξV ⊗s Ω, ξV ⊗s Ω+ aV (b),j(b),ep b (82) = tV (−1),j(−1)(V1) · tV (1),j(1)(V2). This completes the proof. (cid:3) Combining Theorem 7 with our Proposition 1 immediately gives the following. Corollary 1. Let I = {1, 2}. Fix (αi)∞ i=1 and (βi)∞ numbers such that Pi αi +Pi βi ≤ 1 and let V (1) representation of Sn. For i ∈ {1, 2}, let j(i) : V (i) n → V (i) ξV (i) = 10 and let ξV = ξV (1) ⊗ ξV (2). Let Vm,n = V (−1) j1 = 1 ⊗ j(1). Then for (V, c) ∈ P I n = V (2) 2 (∞), n = V (α,β) i=1 decreasing sequences of positive real with the Vershik-Kerov n+1 be the natural isometry. Let n , j−1 = j(−1) ⊗ 1, and m ⊗ V (1) n (83) tV,j(V, c) = Ym≥2 ∞ Xi=1 i + (−1)m+1 αm βm i !ρm(V1)+ρm(V2) ∞ Xi=1 20 ADAM MERBERG 4. Generalized Brownian motions associated to spherical representations of (S∞ × S∞, S∞) In this section, we again use the index set I = {−1, 1}. Of course, we could have taken I to be any two-element set, but we have chosen {−1, 1} for the reason that if b ∈ I then we can concisely refer to the other index as −b ∈ I. Notation 6. With I = {−1, 1}, we will represent a function n : I → N by the pair n(−1), n(1), so we write Vr,s for Vn where n(−1) = r and n(1) = s. G. Olshanski initiated the study of a broad class of representations of infinite symmetric groups [Ols90], and this study has been further developed by Okounkov [Oko97]. In this framework, one considers unitary representations of a pair of groups K ⊂ G forming a Gelfand pair. Two groups (G, K) form a Gelfand pair if for every unitary representation (T, H) of G, the operators PKT (g)PK commute with each other as g ranges over G. Here PK denotes the orthogonal projection of H onto the subspace of K-invariant vectors for the representation T . Of interest to us are the spherical representations, which are defined as irreducible unitary representations of G with a nonzero K-fixed vector ξ. If T is such a representation of the pair (G, K), then the function g 7→ hξ, T (g)ξi is called a spherical function of (G, K). Here we consider the case where G = S∞ × S∞ and K = S∞ is the diagonal subgroup. It is well-known (c.f. [Ols90]) that the finite factor representations of a discrete group G are in bijective correspondence with the spherical representations of the Gelfand pair (G × G, G), where G is a subgroup of G × G by the diagonal embedding. In light of Thoma's Theorem (Theorem 5) this means that the spherical functions of (S∞ × S∞, S∞) are parametrized by the Thoma parameters and that the spherical function associated to the pairs (αi)∞ i=1 is given by the formula i=1 and (βi)∞ (84) χα,β (π, π′) = φα,β(cid:0)π′π−1(cid:1) = Ym≥2 ∞ Xi=1 i + (−1)m+1 αm i !ρm(π′π−1) βm . ∞ Xi=1 For the generalized Brownian motion construction, we can consider the following data. i=1 be decreasing i=1 be a Thoma parameter. That is, let (αi)∞ i=1 and (βi)∞ i=1 and (βi)∞ Let (αi)∞ sequences of positive real numbers such thatPi αi +Pi βi ≤ 1. Given n−1, n1 ∈ N ∪ {0} let n = max(n−1, n1) and define (85) where V (α,β) by n Vn−1,n1 = V (α,β) n , is as in (65). Then Vn−1,n1 carries a natural representation of Sn × Sn defined (U (α,β) (σ, π)h)(x, y) = (−1)i(σ,x)+i(π,y)h(σ−1x, π−1y), (86) and thus a representation of Sn−1 × Sn1 considering Sn−1 × Sn1 as a subgroup of Sn × Sn. For n−1 = n1, it is easy to see that the indicator function of the diagonal is fixed by the diagonal subgroup. n Moreover, we define the map j−1 : Vn−1,n1 → Vn−1+1,n1 to be the natural embedding. When P αi +P βi = 1, this means that j−1 is given by (87) j−1δ(x(−1),x(1)) =(δ(x(−1),x(1)), Pz∈Q δ((x (−1) 1 ,...,x (−1) n ,z),(x (1) 1 ,...,x (1) n ,z)), otherwise. if n1 > n−1, REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 21 Likewise, we define the map j1 : Vn−1,n1 → Vn−1,n1+1 to be the natural embedding. We will also need to make use of the maps j∗ b for b ∈ I. The map j∗ −1 is given by (88) j∗ δ(x(−1),x(1)), −1δ(x(−1),x(1)) = µ(cid:16)x(−1)  n (cid:17) δx (−1) n ,x (1) n δ(cid:16)(cid:16)x (−1) 1 ,...,x (−1) n−1 (cid:17),(cid:16)x (1) 1 ,...,x if n1 ≥ n−1, n−1(cid:17)(cid:17), otherwise. (1) Here x(n−1) refers to the first n − 1 terms of the n-tuple (x1, . . . , xn−1−1) and the measure µ is as in Notation 4. The maps j∗ 1 are defined analogously. To motivate our results in the 2-colored case, we will consider another interpretation of the cycle decomposition of a pair partition. This interpretation involves some graph theory. We assume that a reader is familiar with the notion of a directed graph, a subgraph of a directed graph, and a cycle in a directed graph. These definitions can be found, for instance, in [BJG09]. For a subgraph H of G, we write V (H) to mean the vertex set of H and A(H) to mean the arc set of H. Given a pair partition V ∈ P2(2m), we define a directed graph GV with vertex set [2m]. For each (l, r) ∈ V with l < r, we add an arc (l, r) to GV. For each (l′, r′) ∈ V (as defined in Definition 6) with l′ < r′, we add an arc (r′, l′) to GV. The directed graph GV is the union of vertex-disjoint cycles, and the cycles of the graph GV give the cycles of the pair partition V. More precisely, if C is a cycle of GV then A(C)∩V is a cycle of V. In particular, this means that ρm(V) is the number of cycles of GV of length 2m. Definition 7. For a directed graph G whose vertex set V has a total order <, an increasing path P in G is a sequence of arcs (s1, s2), (s2, s3), . . . , (sr, sr+1) of G such that s1 < s2 < · · · < sr+1. We call r the length of P . A maximal increasing path in G is an increasing path which is not contained in any increasing path in G of greater length. We define the notions of decreasing paths and maximal decreasing paths in G analogously. A monotone path is a path which is either increasing or decreasing, and a maximal monotone path is a monotone path which is not contained in any longer monotone path. In the directed graph GV, each arc is a maximal monotone path, so the length of a cycle is the same as the number of maximal increasing paths in that cycle. Combining with Theorem 7, (89) tα,β(V) = Ym≥2 ∞ Xi=1 i + (−1)m+1 αm βm i !γm(GV ) . ∞ Xi=1 where γm(GV) denotes the number of cycles of GV having m maximal increasing paths. The case of a 2-colored pair partition is naturally more complicated. As in the uncolored case, our function on 2-colored pair partitions (V, c) will be calculated with the aid of the cycle decomposition of a directed graph (denoted GV ,c), but the construction of a graph from a 2-colored pair partition will be rather more involved. However, in the case that the coloring function is the constant function c(l, r) = 1, the graph GV ,c will be identical to the graph GV just described. Before defining the graph GV ,c we fix some notation. Notation 7. For (V, c) ∈ P I 2 (∞) with V = {(l1, r1), · · · , (lm, rm)}, let LV := {l1, . . . , ln} be the set of left points and RV := {r1, . . . , rn} denote the set of right points. If c : V → {−1, 1} 22 ADAM MERBERG is a coloring function, define for b ∈ I the functions (90) Also define (91) pb V ,c : [0, 2m + 1] → N ∪ {0} pb V ,c(u) = {j ∈ [n] : lj ≤ u ≤ rj, c(lj, rj) = b} pV ,c(u) = max{p−1 V ,c(u), p1 V ,c(u)} and rV ,c(u) = pc(u) V ,c (u). Remark 5. In terms of the diagrams (e.g. Figure 1), pb V ,c(m) is the number of b-colored paths intersecting the vertical line drawn through m (provided that the diagrams are drawn so as to minimize this quantity). Furthermore, rV ,c(m) is the number of paths of the same color as m which intersect the vertical line drawn through m (provided that the diagrams are drawn so as to minimize this quantity). The following properties are immediate consequences of the definitions and will be used frequently. Proposition 2. Suppose that (V, c) ∈ BP I {−1, 1}. 2 (∞), V = {(l1, r1), . . . , (lm, rm)} and b ∈ I = V ,c(k − 1) then c(k) = b and k ∈ LV. V ,c(k) then c(k) = b and k ∈ RV. V ,c(k − 1) ∈ {−1, 0, 1}. V ,c(k) > pb V ,c(k + 1) < pb V ,c(k) − pb (1) If k ∈ [2m] and pb (2) If k ∈ [2m] and pb (3) If k ∈ [2m + 1] then pb (4) If k, k′ ∈ [0, 2m + 1] with k < k′, pb there is some l ∈ [k, k′] such that pb (5) If k, k′ ∈ [0, 2m + 1] with k < k′, pb there is some l ∈ [k, k′] such that pb V ,c(k) < pb V ,c(l) = u. V ,c(k) > pb V ,c(l) = u. V ,c(k′) and u ∈ [pb V ,c(k), pb V ,c(k′)], then V ,c(k′) and u ∈ [pb V ,c(k′), pb V ,c(k)], then (6) If k ∈ LV and k ∈ [2m] then rV ,c(k) ≥ pc(k) (7) If k ∈ RV and k ∈ [2m] then rV ,c(k) ≥ pc(k) We need some additional notation. V ,c (k − 1) and pb V ,c (k + 1) and pb V ,c(k) ≥ pb V ,c(k) ≥ pb V ,c(k + 1) for b ∈ I. V ,c(k − 1) for b ∈ I. Notation 8. For an I-indexed pair partition (V, c), define (92) DV ,c = {k ∈ [2m] : rV ,c(k) > p−c(k) SV ,c = [2m] \ DV ,c. V ,c (k)} The next remark should clarify the importance of these terms. Remark 6. If (V, c) ∈ P I 2 (2n) then by (6), tα,β(V, c) can be computed by evaluating the vacuum state at a word in creation and annihilation operators. More precisely, we can write (93) tα,β(V, c) = hξV α,β ⊗ Ω, A1 · · · A2n (ξV α,β ⊗ Ω)i where Ak is a creation operator if k ∈ RV ,c and an annihilation operator if k ∈ LV ,c. In either case, the color of the operator Ak is c(k). One can characterize these operators more precisely, but we will not need to do so at this point. For any k ∈ [2n], the vector (Ak+1 · · · · · A2n) ξV α,β ⊗ Ω lies in the space V α,β some function nk : I → Z. A c(k)-colored creation operator maps the space V α,β V α,β ⊗ H⊗nk+δc(k). If k ∈ RV , then k ∈ SV ,c if and only if the transition map nk+δc(k) nk ⊗ H⊗nk for nk ⊗ H⊗nk to (94) jc(k) : V α,β nk → V α,β nk+δc(k) REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 23 is the identity map on V α,β holds for k ∈ LV . nk , where nk = max{nk(b) : b ∈ I}. The analogous statement also We will make use of the following equivalence relation on [2m]. Notation 9. For k, k′ ∈ [2m], say that k V ,c ∼ k′ if rV ,c(k) = rV ,c(k′). Proposition 3. If k ∈ LV then {k′ > k : k′ V ,c∼ k} 6= ∅. If k ∈ RV then {k′ < k : k′ V ,c∼ k} 6= ∅. Proof. We will consider the case in which k ∈ LV. By Proposition 2 (item 6), pc(k) pc(k) V ,c (k). Let V ,c (k + 1) ≥ (95) V ,c (k′) ≥ rV ,c(k)}, V ,c (r + 1) < rV ,c(k) By Proposition 2 (item 3), pc(k) r = max{k′ > k : pc(k) so that pc(k) must be the case that pc(k) V ,c∼ r. rV ,c(r) = rV ,c(k) and k V ,c (r + 1) = 1, so it V ,c (r) = rV ,c(k). By Proposition 2 (item 2), c(r) = c(k) whence (cid:3) V ,c (r) − pc(k) In defining the graph GV ,c, the following function will be very important. Definition 8. Define a map ZV ,c(k) : [2m] → [2m] by (96) ZV ,c(k) := min{k′ > k : k′ V ,c ∼ k}, max{k′ < k : k′ V ,c∼ k}, max{k′ < k : k′ V ,c ∼ k}, min{k′ > k : k′ V ,c   IV ,c(k) :=([k + 1, ZV ,c(k) − 1], [ZV ,c(k) + 1, k − 1], ∼ k}, if k ∈ LV and k ∈ DV ,c; if k ∈ RV and k ∈ DV ,c; if k ∈ LV and k ∈ SV ,c; if k ∈ RV and k ∈ SV ,c. if ZV ,c(k) > k if ZV ,c(k) < k. Notation 10. For k ∈ [2m], let IV ,c(k) be the interval (97) Proposition 4. Suppose that k ∈ DV ,c and k′ ∈ IV ,c(k). Then (98) p−c(k) V ,c (k′) < rV ,c(k) ≤ pc(k) V ,c (k′) Proof. We will assume that k ∈ LV since the case of k ∈ RV is similar. Suppose that k′ ∈ IV ,c(k) satisfies pc(k) V ,c (k). We can assume that k′ is the smallest element of IV ,c(k) satisfying this inequality, so that by Proposition 2 , pc(k) V ,c (k) (by item V ,c∼ k, which contradicts the definition of 3) and c(k′ − 1) = c(k) (by item 2). Thus, k′ − 1 IV ,c(k). V ,c (k′ − 1) = pc(k) V ,c (k′) < pc(k) V ,c Now suppose that p−c(k) (k′) ≥ pc(k) satisfying this condition so that p−c(k) out the possibility that k′ = k + 1). By Proposition 2, p−c(k) item 3) and c(k′ − 1) = −c(k) (by item 6). Thus, p−c(k) contradicting the definition of IV ,c(k). V ,c (k). Assume that k′ is the smallest element of IV ,c(k) V ,c (k) (again it is straightforward to rule (k′) = −1 (by V ,c∼ k, (cid:3) V ,c (k) whence k′ − 1 (k′ − 1) − p−c(k) (k′ − 1) < pc(k) (k′ − 1) = pc(k) V ,c V ,c V ,c V ,c 24 ADAM MERBERG Proposition 5. If k ∈ SV ,c and k′ ∈ IV ,c(k) then (99) pc(k) V ,c (k′) < p−c(k) V ,c (k) ≤ p−c(k) V ,c (k′). V ,c (k′) ≥ p−c(k) Proof. We will again assume that k ∈ LV as the case of k ∈ RV is similar. Suppose that k′ ∈ IV ,c(k) is such that pc(k) (k). Assume that k′ is the largest element of IV ,c(k) V ,c satisfying this inequality. One can check that if k′ = k − 1 then k − 1 ∼ k and IV ,c(k) = ∅, so we assume that k′ + 1 ∈ IV ,c(k) and thus pc(k) (k). By Proposition 2, it follows that pc(k) contradicts the definition of IV ,c(k). V ,c (k′) = −1 and c(k′ + 1) = c(k). Thus, k′ + 1 V ,c (k′ + 1) < p−c(k) V ,c (k′ + 1) − pc(k) V ,c ∼ k, which V ,c V ,c Now suppose that there is some k′ ∈ IV ,c(k) such that p−c(k) V ,c (k′) < p−c(k) V ,c (k), we must have k 6= k − 1, so that k′ ∈ IV ,c(k) and p−c(k) that k′ is the largest element of IV ,c(k) satisfying this inequality. Since p−c(k) p−c(k) V ,c Proposition 2, p−c(k) pc(k) V ,c (k) whence k′ + 1 V ,c (k) = 1 and c(k′ + 1) = −c(k). But also p−c(k) V ,c∼ k, contradicting the definition of IV ,c(k). (k′ + 1) − p−c(k) (k′ + 1) > p−c(k) V ,c V ,c V ,c V ,c (k). Again assume (k − 1) ≥ (k). By (k′ + 1) = V ,c (cid:3) Proposition 6. Suppose that (V, c) ∈ P I 2 (2m) and k ∈ [2m]. Then the following hold: (1) If k ∈ DV ,c then ZV ,c(k) ∈ DV ,c if and only if c(k) = c(ZV ,c(k)); (2) If k ∈ SV ,c then ZV ,c(k) ∈ SV ,c if and only if c(k) = c(ZV ,c(k)); (3) If k ∈ LV then ZV ,c(k) ∈ LV if and only if c(ZV ,c(k)) = −c(k); (4) If k ∈ RV then ZV ,c(k) ∈ RV if and only if c(ZV ,c(k)) = −c(k); (5) ZV ,c(ZV ,c(k)) = k. Proof. We fix an I-indexed pair partition (V, c). For compactness and readability, we will abbreviate ZV ,c(k) by k∗. If k ∈ LV ∩ DV ,c then by definition k∗ > k. By Proposition 4, (100) p−c(k) V ,c (k∗ − 1) < rV ,c(k) < pc(k) V ,c (k∗ − 1). Using the fact that k∗ V ,c k∗ ∈ RV and k∗ ∈ DV ,c whereas if c(k∗) = c(−k) then k∗ ∈ LV and k∗ ∈ SV, c. ∼ k and applying Proposition 2, it follows that if c(k∗) = c(k) then If k ∈ LV ∩ SV ,c then by definition k∗ < k. By Proposition 5, (101) pc(k) V ,c (k∗ + 1) < p−c(k) V ,c (k) < p−c(k) V ,c (k∗ + 1). Using the fact that k∗ V ,c k∗ ∈ RV and k∗ ∈ SV ,c whereas if c(k∗) = c(−k) then k∗ ∈ LV and k∗ ∈ DV ,c. ∼ k and applying Proposition 2, it follows that if c(k∗) = c(k) then Collecting all of these cases, as well as the analogous statements for k ∈ RV gives items 1, 2, 3, and 4. Making use of the definition of ZV ,c, it follows that (k∗)∗ < k∗ if and only if k < k∗, whence k = (k∗)∗. (cid:3) Corollary 2. If (V, c) ∈ P I partition ¯V (c) ∈ P I 2 (2m) by 2 (2m) then the function ZV ,c : [2m] → [2m] defines a pair (102) ¯V (c) := {(k, ZV ,c(k)) : k < ZV ,c(k)} . REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 25 1 2 3 4 5 6 7 8 9 10 11 12 Figure 6. The {−1, 1}-indexed pair partition (V, c) given in (106). We define a coloring function on ¯V (c) by: (103) ¯c (k, ZV ,c(k)) =(c(k), −c(k), if k ∈ SV ,c if k ∈ DV ,c. Remark 7. This definition of ¯c (k, ZV ,c(k)) does not depend on which point of a pair is chosen as k because c(k) = c(ZV ,c(k)) if and only if k and ZV ,c(k) are either both in SV ,c or both in DV ,c. Notation 11. Define a map (·, ·)(−1) on [2m] × [2m] by (u, v)(−1) = (v, u). Also let (·, ·)(1) be the identity map on [2m] × [2m]. We are now ready to define the graph GV ,c. Definition 9. If (V, c) ∈ P I arcs defined as follows. Let 2 (2m), then GV ,c is the directed graph with vertices [2m] and (104) FV ,c = {(l, r)(c(l,r)) : (l, r) ∈ V} ¯FV ,c = {(k, k′)(¯c(k,k′)) : (k, k′) ∈ ¯V (c)}. The graph GV ,c has arc set (105) Proposition 7. The sets FV ,c and ¯FV ,c have empty intersection. AV ,c := FV ,c ∪ ¯FV ,c Proof. We need only show that if (l, r) ∈ V with ZV ,c(l) = r then c(l, r) = −¯c(l, r). By the definition of ¯c, it will suffice to show that l ∈ DV ,c. Since l < ZV ,c(l), ZV ,c(l) = r ∈ RV and c(l) = c(ZV ,c(l)), this follows from Proposition 6 and the definition of ZV ,c. (cid:3) Example 2. We consider the example of the graph GV ,c for the I-indexed pair partition (V, c) with (106) V = {(1, 5), (2, 10), (3, 8), (6, 7), (4, 12), (9, 11)}; c(1, 5) = c(2, 9) = c(8, 10) = c(6, 7) = −1; c(3, 8) = c(4, 12) = 1. The I-indexed pair partition (V, c) is depicted in Figure 6. We can determine the values of p−1 V ,c(k) and p1 V ,c(k) by drawing a vertical line through the diagram at k and counting the intersections with paths of the respective colors. For instance, a vertical line through k = 3 intersects 2 solid paths and 1 dotted path (at the endpoint), so p(1) V ,c (3) = −1. Since c(3) = 1, this means that 3 ∈ SV ,c. V ,c(3) = 1 and p(−1) 26 ADAM MERBERG 1 2 3 4 5 6 7 8 9 10 11 12 Figure 7. The I-indexed pair partition ( ¯V (c), ¯c) where (V, c) is the pair par- tition depicted in the Figure 6. (V ,c) (k) DV ,c or SV ,c Z(V ,c)(k) DV ,c DV ,c SV ,c SV ,c SV ,c SV ,c SV ,c DV ,c DV ,c DV ,c SV ,c DV ,c 3 4 1 2 6 7 8 7 10 9 12 11 k 1 2 3 4 5 6 7 8 9 10 11 12 c(k) -1 -1 1 1 -1 -1 -1 1 -1 -1 -1 -1 r(V ,c)(k) p−c(k) 0 0 2 2 2 2 2 1 1 1 1 1 1 2 1 2 2 2 2 2 2 2 1 1 Table 1. Data pertaining to the I-indexed pair partition depicted in Figure 5. Repeating the process for the other elements of [12] we can fill in the first few rows of V ,c ∼ is {1, 3, 11, 12} Table 1. From the data, one sees that the equivalence class of 3 under whence (107) ZV ,c(3) = 1. Continuing in this way for other elements of 2m, the pair partition ¯V (c) is given by (108) and the color function ¯c : ¯V (c) → I is given by ¯V (c) := {(1, 3), (2, 4), (5, 6), (7, 8), (9, 10), (11, 12)} (109) ¯c(5, 6) = ¯c(7, 8)) = ¯c(11, 12) = −1; ¯c(1, 3) = ¯c(2, 4)) = ¯c(9, 10) = 1. This I-indexed broken pair partition is depicted in Figure 7. Accordingly, the sets FV ,c and ¯FV ,c are given by (110) FV ,c = {(5, 1), (10, 2), (3, 8), (4, 12), (7, 6), (11, 9)} ¯FV ,c = {(6, 5), (12, 11), (1, 3), (2, 4), (7, 8), (9, 10)} The directed graph GV ,c is depicted in Figure 8. Proposition 8. Let (V, c) ∈ P I GV ,c is the union of vertex-disjoint directed cycles. 2 (∞) be a {−1, 1}-colored pair partition. Then the graph Proof. Let m = V. It is an immediate consequence of the definition of GV ,c that for each k ∈ [2m] there are exactly two arcs having k as either the start point or the end point. We must show that each vertex k ∈ [2m] is the starting point of one edge and the end point of another edge. REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 27 1 2 3 4 5 6 7 8 9 10 11 12 Figure 8. The directed graph GV ,c for the {−1, 1}-colored pair partition depicted in Figure 6. The graph GV ,c has 2 cycles, one with vertex set {1, 3, 8, 7, 6, 5} and one with vertex set {2, 4, 12, 11, 9, 10}. The former cy- cle has one maximal increasing path, and the latter has 2 maximal increasing paths. We will assume k ∈ LV, so that (k, r) ∈ V for some r ∈ [2m]. If k ∈ DV ,c then ZV ,c(k) > k, so (k, ZV ,c(k)) ∈ ¯V (c). Since c(k, r) = −¯c(k, ZV ,c(k)), k is the starting point of exactly If k ∈ SV ,c then ZV ,c(k) < k, so (ZV ,c(k), k) ∈ ¯V (c). Furthermore, one of these arcs. c(k, r) = ¯c(ZV ,c(k), k), whence k is again the starting point of exactly one of the arcs. (cid:3) Remark 8. If c(d) = 1 ∈ I for every d ∈ V then every DV ,c = [2m] and ¯V (c) = V (where V is as in Definition 6) with ¯c(d) = −1 for all d ∈ ¯V (c). Thus GV ,c = GV . A cycle C in the graph GV ,c can be decomposed into maximal monotone paths (Definition 7). That is, there are maximal monotone paths P1, . . . , Ps such that each arc of C lies along exactly one of the Pj. Notation 12. For a directed graph G on a totally ordered vertex set, we denote by γm(G) the number of cycles of G having exactly m maximal increasing paths (equivalently, m maximal decreasing paths). With this established, we are able to state the main result of this section. Theorem 8. Let (αi)∞ i=1 and (βi)∞ that Pi αi + Pi βi ≤ 1. Let V (α,β) representation of Smax(n−1,n1) endowed with the representation of (86), and let (111) j−1 n : V (α,β) n−1,n1 → V (α,β) n−1+1,n1 and j1 n : V (α,β) n−1,n1 → V (α,β) n−1,n1+1 i=1 be decreasing sequences of positive real numbers such n−1,n1 be the complex Hilbert space of the Vershik-Kerov be the natural embedding. Let ξV (α,β) = 10 be the indicator function of the diagonal. Denote by tα,β the function on P I 2 (∞) associated to this sequence of representations by Theorem 1. Let (V, c) be a {−1, 1}-colored pair partition. Then (112) tα,β(V, c) = Ym≥2 ∞ Xi=1 i + (−1)m+1 αm βm i !γm(GV,c) . ∞ Xi=1 We will present the proof of Theorem 8 in the case thatPn αn = 1. This case will contain the key ideas of the more general argument, but will simplify notation considerably and allow us to consider discrete sums instead of integrals. Notation 13. For (V, c) ∈ P I 2 (2m), define (113) RD LD V ,c := {k ∈ RV : k ∈ DV ,c} V ,c := {k ∈ LV : k ∈ DV ,c}. 28 ADAM MERBERG Proposition 9. Let I = {−1, 1} and let (V, c) ∈ P I (V, c) has its starting point in LD path in (V, c) has its starting point in RD V ,c and its terminal point in RD 2 (2m). Every maximal increasing path in V ,c. Every maximal decreasing V ,c and its terminal point in LD V ,c. Proof. This follows immediately from the definitions of the map ZV ,c and the graph GV ,c. A left point l ∈ LD V ,c, is adjacent to vertices u, v with l < u and l < v whereas if l ∈ LV ∈ SV ,c, l is adjacent to one vertex u′ with u′ < l and another vertex v′ with l < v′. (cid:3) The proof of Theorem 8 will be aided by some additional terminology. Definition 10. Let H be an infinite-dimensional complex Hilbert space with orthonormal basis B := {hi}∞ i=1. We will say that a vector η ∈ FV (H) is B-elementary if there is a constant C, and some n−1, n1 such that there is a pair of tuples x(−1), x(1) ∈ Xn of length n := max(n−1, n1) with x(−1) ∼ x(1) (where Xn and the relation ∼ are as in Notation 4) and indices i(−1) (114) n1 ∈ N all distinct such that 1 , . . . , i(1) , . . . , i(−1) n−1 ∈ N and i(1) . 1 η = C · δ(x(−1),x(1)) ⊗s hi (−1) 1 ⊗ · · · ⊗ hi (−1) n−1 ⊗ hi (1) 1 ⊗ · · · ⊗ hi (1) n1 Denote by E B V (H) the set of all B-elementary vectors in FV (H). Notation 14. For the remainder of this section, we assume that H is an infinite-dimensional complex Hilbert space with orthonormal basis B := {hi}∞ i=1. For a nonzero B-elementary vector η ∈ Vn−1,n1 ⊗s H⊗n−1 ⊗ H⊗n1 ⊂ FV (H), define nη : I → N by nη(b) = nb. The next proposition says that, up to permutations, a nonzero B-elementary vector can be expressed uniquely in the form (114). It follows immediately from the definition of the symmetric tensor product ⊗s. Proposition 10. Suppose that a nonzero B-elementary vector η has two expressions of the form (114), (115) η = C · δ(x(−1),x(1)) ⊗s hi = C · δ(x(−1),x(1)) ⊗s hi (−1) 1 (−1) 1 ⊗ · · · ⊗ hi (−1) n−1 ⊗ hi (1) 1 ⊗ · · · ⊗ hi (1) n1 ⊗ · · · ⊗ hi (−1) n−1 ⊗ hi (1) 1 ⊗ · · · ⊗ hi (1) n1 where for b ∈ I, x(b), x(b) ∈ Xn and n := max(n−1, n1). Then there is some (π−1, π1) ∈ n πb)x(b), Sn−1 × Sn1 such that for each b ∈ I = {−1, 1}, is = iπ−1 where ιnb n denotes the natural inclusion from Snb → Sn. b s for s ∈ [nb] and x(b) = (ιnb Proposition 10 invites the following. Corollary 3. Let H be an infinite-dimensional complex Hilbert space with orthonormal basis B := {hi}∞ i=1. Let η ∈ FV (H) be a B-elementary vector. Fix some expression for η of the form (114). (1) If η 6= 0 then the sets J (b)(η) := {i(b) (116) 1 , . . . , i(b) nη(b)} (b ∈ I) and J (η) := J (−1)(η) ∪ J (1)(η) do not depend on the choice of the expression for η in the form (114). (2) If η 6= 0 and b ∈ I, the function Sb η : J (η) → Q (with Q as in Notation 4) given by u ) := x(b) u Sη(i(b) (117) does not depend on the choice of the expression for η in the form (114). REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 29 (3) If η 6= 0 and b ∈ I is such that nη(b) > nη(−b) then the function (118) (119) given by xη : [nη(−b) + 1, nη(b)] → Q xη(k) = x(−b) k does not depend on the choice of the expression for η in the form (114). (4) Moreover, (120) D(η) =(0, C, if η = 0 if η 6= 0 is expressed in the form (114) does not depend on the choice of the expression for η in the form (114). We can now characterize the effect of an annihilation operator on a B-elementary vector. In doing so, it will be convenient to fix the Hilbert spaces Vn and the transition maps jb and omit the superscripts V and j. The next two propositions follow immediately from the definition of the annihilation operators. Proposition 11. If η ∈ FV (H) is a B-elementary vector and b ∈ I is such that nη(b) > nη(−b) and i ∈ J (b) , then ab(hi)η is also a B-elementary vector and η (121) D(ab(hi)η) = δxη(nη (b)),Sη (i) · µ(Sη(i)) nη(b) · D(η). Furthermore if ab(hi)η 6= 0, then the following hold: (1) J (b)(ab(hi)η) = J (b)(η) \ {i}; (2) J (−b)(ab(hi)η) = J (−b)(η); (3) For i′ ∈ [nη(−b) + 1, nη(b) − 1], xab(hi)η(nη(b)) = xη(nη(b)); (4) For any i′ ∈ J (ab(hi)η) = J (η) \ {i}, Sab(hi)η(i′) = Sη(i′). Proposition 12. Let η ∈ FV (H) be a nonzero B-elementary vector. Fix i ∈ Jη, and let b ∈ I be such that nη(b) ≤ nη(−b). Then ab(hi)η is a B-elementary vector and (122) D(ab(hi)η) = 1 nη(b) · D(η). If ab(hi)η 6= 0 then the following also hold: (1) J (b)(ab(hi)η) = J (b)(η) \ {i}; (2) J (−b)(ab(hi)η) = J (−b)(η); (3) For i′ ∈ [nη(b) + 1, nη(b)], xab(hi)η(nη(b)) = xη(nη(b)); (4) xab(hi)η(nη(b)) = Sη(i); (5) For any i′ ∈ J (ab(hi)η) = J (η) \ {i}, Sab(hi)η(i′) = Sη(i′). A creation operator need not take a B-elementary vector to another B-elementary vector. However, we can express a creation operator as a sum of operators which preserve the B- elementary property. Notation 15. For z ∈ Q = N, let ǫz : Vn → Vn+1 to be the linear map such that ǫzδ(x,y) = δ((x,z),(y,z)). If h ∈ H and n−1 ≥ n1 define an operator (123) −1,z(hr) : Vn−1 ⊗s H⊗n−1 ⊗ H⊗n1 → Vn−1+1 ⊗s H⊗n−1+1 ⊗ H⊗n1 a∗ 30 by (124) Define a∗ ADAM MERBERG a∗ −1,z(hr)η = (nη(b)ǫz ⊗ r−1(h)) η. 1,z(hr) similarly for n1 ≥ n−1. Remark 9. In the case that P αi = 1 (and thus βi = 0 and γ = 0), we have a∗ b,z(h)η = a∗ b(h)η (125) Xz∈Q for a B-elementary vector η ∈ F (V ) with nη(b) > nη(−b). The following two propositions follow directly from the definition of the respective opera- tors. Proposition 13. If η ∈ FV (H) is a B-elementary vector, b ∈ I is such that nη(b) ≥ nη(−b), and i ∈ N \ J (η) then for any z ∈ Q, the vector a∗ b,z(hi)η is also B-elementary. Furthermore, the following hold: (1) If η 6= 0 then J (b)(a∗ (2) If η 6= 0 then J (−b)(a∗ (3) For any i′ ∈ J (η), Sa∗ (4) Sa∗ (5) D(a∗ b,z(hi)η(i) = z; b,z(hi)η) = (nη(b) + 1) · D(η). b,z(hi)η) = J (b)(η) ∪ {i}; −b,z(hi)η) = J (−b)(η); b,z(hi)η(i′) = Sη(i′); Proposition 14. If η ∈ FV (H) is a B-elementary vector, b ∈ B is such that nη(−b) > nη(b), and i ∈ N \ Jη then the vector a∗ b(hi)η is also B-elementary. Furthermore, (1) If η 6= 0 then J (b)(a∗ (2) If η 6= 0 then J (−b)(a∗ (3) For any i′ ∈ J (η), Sa∗ (4) If η 6= 0 then Sa∗ (5) D(a∗ b (hi)η(i) = xη(nη(b)). b(hi)η) = (nη(b) + 1)D(η). b(hi)η) = J (b)(η) ∪ {i}; −b(hi)η) = J (−b)(η); b,z(hi)η(i′) = Sη(i′); We are now ready to prove the main theorem. Proof of Theorem 8. We reiterate that we are focusing on the case P αi = 1 for simplicity. The key ideas of the more general case P αi +P βi ≤ 1 are in this case, but this case is slightly more straightforward in that it allows us to work with discrete sums instead of integrals. In this case, we can assume that Q = N. Let n = V and let H be an infinite-dimensional complex Hilbert space with an orthonor- mal basis B := {hi}∞ i=1. By Theorem 1, we can compute tV,j(V, c) as (126) tV,j(V, c) =Dae1 c(1)(hk1) · · · aen c(n)(hkn) (10 ⊗s Ω) , 10 ⊗s ΩE , where i is an element of the ki-th pair of V, and ei = 2 if i ∈ RV and ei = 1 if i ∈ LV. For each k ∈ [2m] define (127) so that (128) Ai :=(a∗ c(i)(hki), ac(i)(hki), if i ∈ RV if i ∈ LV , tV,j(V, c) = h(A1 · · · A2m) 10 ⊗s Ω, 10 ⊗s Ωi , REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 31 Denote by ΛV ,c the space of functions λ : RD V ,c → Q. For λ ∈ ΛV ,c and i ∈ [2m], define (129) We define (130) Aλ,i :=(a∗ c(i),λ(i)(hki), Ai, if i ∈ RD V ,c otherwise. A(k) λ := Aλ,k · · · Aλ,2m for 1 ≤ k ≤ 2m. For convenience, we also set A(2m+1) definitions and (128) that λ = 1. It follows immediately from the (131) Also define (132) tV,j(V, c) = Xλ∈ΛV,cDA(1) λ 10 ⊗s Ω, 10 ⊗s ΩE ηλ,k := A(k) λ 10 ⊗s Ω. It can be seen from the definition of pb V ,c that if ηλ,k 6= 0 then (133) for b ∈ I. nηλ,k (b) =(p(b) V ,c(k) − 1, p(b) V ,c(k), if k ∈ LV and c(k) = b, otherwise. It follows from Propositions 11, 12, 13 and 14 that ηλ,k is B-elementary for 1 ≤ k ≤ 2m+1. Since ηλ,1 ∈ C10 ⊗s Ω, it is determined by the constant D(ηλ,1). For k ∈ [2m], define (134) so that (135) Bλ,k :=(δxηλ,k+1(nηλ,k+1 (b)),Sηλ,k+1 (i) · µ(Sηλ,k+1(i)) 1, if k ∈ LD V ,c otherwise, D(ηλ,1) = 2m Yk=1 Bλ,k = Yk∈LD V,c Bλ,k. For a maximal increasing path P of GV ,c from ls to rt with ls < rt, define a function HP,λ : [ls + 1, rt] → Q as follows. If u ∈ [ls + 1, rt], denote by fu = (au, zu) the unique arc along the path P such that au < u ≤ zu. Then we define (136) HP,λ(u) :=(xηλ,u(rV ,c(au)), Sηλ,u(i), if fu = (au, zu) ∈ ¯V c if fu = (li, ri) ∈ V. It can be verified using (133) that if b ∈ I is such that pc(b)(au) > pc(−b)(au) and fu ∈ ¯FV ,c then nηλ,u(b) < rV ,c(au) ≤ nηλ,u(−b), so that rV ,c(au) is indeed in the domain of xηλ,u. We will show that HP,λ(u) = λ(ls) for all u ∈ [ls + 1, rt]. By definition, (137) ηλ,rt = Aλ,rtηλ,rt+1 = a∗ c(rt),λ(rt)(ht)ηλ,rt+1. As an endpoint of a maximal monotone path, rt ∈ DV ,c, and pc(rt) that nηλ,rt−1(c(rt)) > nηλ,rt+1(−c(rt)). If frt ∈ ¯FV ,c then by Proposition 13, (138) c(rt ),λ(rt)(ht)ηλ,rt+1(rV ,c(c(rt))) = λ(rt). HP,λ(rt) = xa∗ V ,c (rt) > p−c(rt) V ,c (rt) implies 32 ADAM MERBERG If, on the other hand, fu = (lt, rt) ∈ V then (139) HP,λ(rt) = Sa∗ c(rt),λ(rt)(ht)ηλ,rt+1(t) = λ(rt), where we have again applied Proposition 13. To show that HP,λ(u) = λ(rt) for all u, it will now suffice to show that HP,λ(u) = HP,λ(u+1) for u ∈ [ls + 1, rt − 1]. For u such that fu = fu+1 this is an immediate consequence of the definition of HP,λ(u). Otherwise, u is a vertex along the path P and u ∈ SV ,c. If u ∈ LV , say u = li, then (au, zu) = (u, zu) ∈ ¯FV ,c, thus (140) HP,λ(u) = xηλ,u(rV ,c(c(au))) = xac(u)(hi)ηλ,u+1(rV ,c(c(au))) = Sηλ,u+1(i) = HP,λ(u + 1). Here we have made use of Proposition 12. If instead u ∈ RV, say u = ri, then (au, zu) = (li, ri) ∈ V, so (141) HP,λ(u) = Sηλ,u(i) = Sa∗ c(u)(hi)ηλ,u+1(i) = xηλ,u+1(nηλ,u+1(c(u))) = xηλ,u+1(rV ,c(ri)) = HP,λ(u + 1), where we have used Proposition 14. Thus, HP,λ(u) = λ(rt) for all u ∈ [ls + 1, rt]. In particular, taking u = ls + 1 shows that xηλ,ls+1(rV ,c(ls)) = λ(rt) if the initial arc of P is in ¯FV ,c and Sηλ,ls+1(i) = λ(rt) if the initial arc is (li, ri) ∈ V. By a similar argument, if P ′ is a maximal decreasing path from some rt′ ∈ RD V ,c to ls then xηλ,ls+1(rV ,c(ls)) = λ(rt) if the final arc of P ′ is in ¯FV ,c and Sηλ,ls+1(i′) = λ(rt) if the final arc of P ′ is (li′, ri′) ∈ V. Thus, (142) Bλ,ls = δxηλ,k+1 (nηλ,k+1 (b)),Sηλ,k+1 (i) · µ(Sηλ,k+1(i)) = δλ(rt),λ(rt′ ) · µ(λ(rt)). Therefore, if A : LD maximal decreasing path terminating at l and D : LD endpoint of the maximal increasing path starting at l then V ,c is the function taking l ∈ LD V ,c → RD V ,c to the starting point of the V ,c is the function l to the V ,c → RD (143) D(ηλ,1) = Yk∈LD V,c Bλ,k = Yl∈LD V,c δλ(D(l)),λ(A(l)) · µ(D(l)). This means that D(ηλ,1) = 0 unless λ(r) = λ(r′) whenever there is some l ∈ LD V ,c such that there are maximal monotone paths from l to r and r to l′. But this condition holds only if λ(r) = λ(r′) whenever r and r′ lie along the same cycle of GV ,c. Denote by ΛC V ,c the set of functions in ΛV ,c satisfying this condition. For λ ∈ ΛC V ,c and a cycle K of GV ,c we write λ(K) for the common value λ(r) for any r ∈ RD V ,c along the cycle K. REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 33 Denoting the set of cycles of GV ,c by C(GV ,c) and the number of maximal increasing paths of a cycle K by M(K) we deduce that (144) Finally, (145) D(ηλ,1) =(QK∈C(GV,c) µ(λ(K))M (K), 0, if λ ∈ ΛC V ,c, otherwise. tα,β(V, c) = Xλ∈ΛV,cDA(1) λ 10 ⊗s Ω, 10 ⊗s ΩE µ(λ(K))M (K) D(ηλ,1) = Xλ∈ΛV,c = Xλ∈ΛC V,c YK∈C(GV,c) = Ym≥2 ∞ Xi=1 αm i !γm(GV,c) . (cid:3) 5. A special case of the generalized Brownian motions associated to spherical representations of (S∞ × S∞, S∞) In this section, we specialize the investigation begun in Section 4 to a countable class of Thoma parameters which were also considered in [BG02]. Namely, for N ∈ Z \ {0} we will consider the spherical function ϕN of the Gelfand pair (S∞ × S∞, S∞) arising from the Thoma parameters with (146) αn =(1/N, N > 0 and 1 ≤ n ≤ N; otherwise. 0, and βn =(1/N, N < 0 and 1 ≤ n ≤ N; otherwise. 0, The character ϕN on S∞ is given by (147) ϕN (π) =(cid:18) 1 N(cid:19)m−γ(m)(π) where m is large enough so that σ(k) = k for k > m and γ(m) (σ) is the number of cycles in the permutation σ ∈ S∞ when σ is considered as an element of Sm. Although γ(m) (σ) depends on the choice of m, the quantity m − γ(m) (σ) does not. We denote by ψN the associated spherical function on the Gelfand pair (S∞ × S∞, S∞). That is, (148) ψN (π−1, π1) = ϕN (π−1 1 π−1) =(cid:18) 1 N(cid:19)m−γ(m)(π−1 1 π−1) , where (π−1, π1) ∈ S∞ × S∞. 34 ADAM MERBERG The function on {−1, 1}-indexed pair partitions associated to ψN by Theorem 8 is given by (149) tN (V, c) =(cid:18) 1 N(cid:19)m(GV,c)−γ(GV,c) , n=1 wb(n). P∞ (153) (154) where m(GV ,c) and γ(GV ,c) denote the number of maximal increasing paths and number of cycles of the graph GV ,c defined in Section 4, respectively. For a complex Hilbert space H, the function tN gives rise to a Fock state ρN on the algebra CI(H). We denote by F I N (H) the Hilbert space, distinguished cyclic vector, and algebra of operators of the GNS construction for this pair. We will see that for N < 0, the field operators on FN (H) are bounded operators which generate a von Neumann algebra containing the projection onto vacuum vector. N (H), ΩN , and CI b,i the creation operator a∗ Notation 16. Fix an integer N 6= 0 and an infinite-dimensional complex Hilbert space H with orthonormal basis B := {hn : n ∈ N}. Denote by a∗ tN ,b(hi) and by ab,i the annihilation operator atN ,b(hi). As we have done before, we will write ae b,i for either a creation (e = 2) or annihilation operator (e = 1) and let ωb,i = ab,i + a∗ b,i. Let ΓI N (H) be the von Neumann algebra generated by the spectral projections of the ωb,i (b ∈ I, i ∈ N). b,i. Each word in P(H, B) can be considered a (possibly unbounded) operator on F I N (H) simply by regarding it as a product of the creation and annihilation operators that comprise the word. The set P(H, B) inherits the involution ∗ from CI(H). For A ∈ P(H, B), b ∈ I and i ∈ N, let cb,i(A) be the number of occurrences of the creation operator a∗ b,i in the word A and ab,i(A) the number of occurrences of the annihilation operator ab,i in the word A. For b ∈ I, define wA Let P(H, B) be the set of finite words in the ae b : N → Z by (150) wA b (n) = cb,n(A) − ab,n(A). Given a function wb : N → Z for each b ∈ I taking only finitely many nonzero values, define (151) Denote by Hw(H, B) the space Pw(H, B) :=(cid:8)A ∈ P(H, B) : wA = w(cid:9) . (152) Hw(H, B) := span{AΩN : A ∈ Pw(H, B)}. For a function wb : N → Z which is 0 at all but finitely many points, we define wb = Remark 10. If wb(n) < 0 for some b ∈ I and some n ∈ N then Hw(H, B) = 0. Definition 11. Suppose that A ∈ P(H, B) and the terms in the product A are indexed by some ordered set S, We will say that (V, c) ∈ P I 2 (S) is compatible with A if aek bk,ik . A = Yk∈S δel,1δer,2δc(l,r),blδc(l,r),br δil,ir = 1. Y(l,r)∈V Denote by C(A) the set of all (V, c) ∈ P I 2 (S) which are compatible with A. REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 35 Remark 11. The motivation for Definition 11 is that (6), for A ∈ P(H, B), (155) ρN (A) = X(V ,c)∈C(A) tN (V, c). Remark 12. The condition (V, c) ∈ C(A) uniquely determines ( ¯V (c), ¯c). This is because once we know that (V, c) ∈ C(A), we can immediately determine the color function c and which points are left points of V and which are right points of V. This, in turn, determines the index function ¯V (c) and the involution ZV ,c, which gives ¯V (c). Notation 17. For T ⊆ S, we denote by AT the product of those elements in the word A which are indexed by elements of T . The next proposition is an immediate consequence of the definitions. Proposition 15. Suppose that A ∈ Pw(H, B) is given by r (156) A = aek bk,ik . Yk=1 k=s+1 aek bk,ik . Then s ∈ DV ,c if and only if one of the following holds: and let (V, c) ∈ C(A). For each s ∈ [r] denote by As the product As = Qr (1) s ∈ RV and (cid:12)(cid:12)(cid:12) (2) s ∈ LV and (cid:12)(cid:12)(cid:12) The definition (6) of the Fock state ρN and the definition of wA −c(s)(cid:12)(cid:12)(cid:12) −c(s)(cid:12)(cid:12)(cid:12) c(s)(cid:12)(cid:12)(cid:12) c(s)(cid:12)(cid:12)(cid:12) ≥(cid:12)(cid:12)(cid:12) >(cid:12)(cid:12)(cid:12) wAs wAs wAs wAs ; . b give the following. Proposition 16. Suppose that w−1, w1, w′ points. The spaces Hw(H, B) and Hw′(H, B) are orthogonal unless wb = w′ 1 : N → Z are zero except at finitely many −1, w′ b for b ∈ I. This enables us to make the following definition. Definition 12. Let Nb,n be the operator defined on the dense subspace ⊕wHw(H, B) of F I N (H) by linear extension of (157) η 7→ wb(n) · η for η ∈ Hw(H, B). The next proposition is an exclusion principle analogous to that proven in [BG02]. Lemma 1. If N < 0 and there is some b ∈ I and some n ∈ N such that if wb(n) > N, then Hw(H, B) = 0. Corollary 4. If N < 0 then Nb,n is bounded for all b ∈ I and n ∈ N. Moreover, kNb,nk = N. Our proof of Lemma 1 will make use of some basic combinatorics, which we now recall. Notation 18. Denote by s(n, k) the number of permutations in the symmetric group Sn which can be written as the product of k disjoint cycles. The numbers s(n, k) are known as the unsigned Stirling number of the first kind. It [Sta12]) that the unsigned Stirling numbers of the first kind satisfy the is well-known (c.f. relation (158) x(x + 1) · · · (x + n − 1) = s(n, k)xk. n Xk=0 36 ADAM MERBERG Proposition 17. If N ∈ N and N < 0 then (159) N c(σ) = 0 Xσ∈SN+1 Proof. This follows immediately from (158): (160) Xσ∈SN+1 N c(σ) = N +1 Xk=0 s(N + 1, k)N k = N(N + 1) · · · (N + (N + 1) − 1) = 0. To prove Lemma 1, we will need the following proposition, which is proven by applying the definitions. Proposition 18. Suppose that A, B ∈ P(H, B) are words of length ℓA and ℓB, respectively. Assume that A can be expressed as a product (cid:3) (161) and that B∗ can be expressed as (162) A = ae1 b1,i1 · · · a eℓA bℓA ,iℓA B∗ = a e−ℓB b−ℓB ,i−ℓB · · · ab−1,i e−1 −1 If (V, c) ∈ C(B∗A) then for any i ∈ N, there is a unique subset Y +,b,i unique subset Y −,b,i A,B (V, c) ⊂ −[ℓB] such that the following conditions are satisfied: A,B (V, c) ⊂ [ℓA] and a A,B (V, c) = Y −,b,i A,B (V, c) = wA b (i); (1) Y +,b,i (2) If k ∈ Y +,b,i (3) If k ∈ Y −,b,i (4) If k ∈ Y +,b,i (5) If k ∈ Y −,b,i A,B (V, c) then ek = 2, c(k) = b, and ik = i; A,B (V, c) then ek = 1, c(k) = b, and ik = i; A,B (V, c) then πV (k) ∈ Y −,b,i A,B (V, c) then πV (k) ∈ Y +,b,i A,B (V, c); A,B (V, c). Here πV is the permutation obtained by regarding the pairs of V as transpositions. Remark 13. The sets Y +,b,i C(B∗A), but they are uniquely determined by this choice. A,B (V, c) and Y −,b,i A,B (V, c) may depend on the choice of (V, c) ∈ Proof of Lemma 1. It will suffice to consider the case wb(i) = N + 1, and we will further assume that b = 1. Consider a word A ∈ P(H, B) containing r ≥ N + 1 creation operators a∗ b,i and r − N − 1 annihilation operators ab,i, say (163) We need to show that ρN (A∗A) = 0. It will be convenient to view the word A∗A as a product of operators whose terms are indexed by the set J := {−s, −s + 1, . . . , −1, 1, . . . , s − 1, s}. Thus bk = b−k, ek 6= e−k, and ik = i−k for k ∈ J. By (155), b1,i1 · · · aes A = ae1 bs,is (164) ρN (A∗A) = X(V ,c)∈C(A∗A) tN (V, c) For (V, C) ∈ C(A∗A), consider the sets Y +,b,i A,A (V, c) provided by Propo- sition 18, and denote these simply by Y +(V, c) and Y −(V, c). These sets have cardinal- ity wA b (i) = N + 1. The I-indexed pair partition (V, c) can be seen as a pair partition WV ,c ∈ C(A∗AJ\(Y +(V ,c)∪Y −(V ,c))) together with a bijection ιV ,c : Y +(V, c) → Y −(V, c). A,A (V, c) and Y −,b,i REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 37 It will suffice to show that for any F +, F − ⊂ J with F + = F − = N + 1 and any W ∈ C(A∗AJ\(F +∪F −)), (165) tN (V, c) = 0. X(V ,c)∈C(A∗A) WV,c=W If the sum is empty, there is nothing to show. Otherwise, define a directed graph GF −,F + whose vertices are the elements of the index set J, and whose arc set is (166) (cid:8)(k, k′)cA∗A(k) : (k, k′) ∈ W(cid:9) ∪n(k, k′)¯cA∗A(k,k′) : (k, k′) ∈ ¯VA∗Ao , where (k, k′)(b) is as in Notation 11. A vertex k ∈ J \ (F + ∪ F −) of GF −,F + is the start point and end point of exactly one arc. A vertex k ∈ F + is the start point of 1 arc and is not the end point of any arc. A vertex k ∈ F − is the end point of 1 arc and is not the start point of any arc. Therefore, each vertex k ∈ F + of GF −,F + is the starting point of a maximal path which ends at some vertex k′ ∈ F −. This gives a bijection ǫ : F + → F −. (To clarify these notions, we consider a specific case, including diagrams in Example 3.) A term in the sum in (165) can be characterized by the bijection ιV ,c : F + → F −. The graph GV ,c can be formed from GF −,F + by adding the edges arising from ιV ,c. The number of maximal increasing paths m(GV ,c) does not depend on the choice of ιV ,c, and we denote the common value by m. Furthermore γ(GV ,c) = γ( F )+γ(ιV ,cǫ−1) where γ(ιV ,cǫ−1) is the number of cycles of ιV ,cǫ−1 as a permutation on F −. As ιV ,c ranges over all bijections F + → F −, the permutation ιV ,cǫ−1 ranges over the symmetric group, whence X(V ,c)∈C(A∗A) WV,c=W G±(V ,c)=F ± (167) N(cid:19)m−(γ( F )+γ(σ)) tN (V, c) = Xσ∈SN+1(cid:18) 1 N(cid:19)m−γ( F ) =(cid:18) 1 = 0. N γ(σ) Xσ∈SN+1 (cid:3) Example 3. We consider a simple example, with a diagram, to clarify some of the ideas 1,1 and define A := a∗aa∗a∗ in the proof of Lemma 1. Let N = −1, let a = a1,1 and a∗ = a∗ so that wA(1) = 3 − 1 = 2 > N = 1. If (V, c) ∈ C(A∗A) then the sets Y +,1,1 A,A (V, c) and Y −,1,1 A,A (V, c) have cardinality wA(1) = 2. As usual, we refer to these sets by Y +(V, c) and Y −(V, c). Indexing the product on [−4, 4] \ {0}, we will consider the (V, c) ∈ C(A∗A) having Y +(V, c) = {1, 4} and Y −(V, c) = {−3, −1}. There are two such (V, c), with pair partitions (168) V1 = {(−4, −2), (−3, −1), (−1, 4), (2, 3)} and V2 = {(−4, −2), (−3, 4), (−1, 1), (2, 3)} and with color functions c1 and c2 defined to be 1 on all pairs of their respective pair partitions. The graph G{−3,−1},{1,4} is depicted in Figure 9. This graph, whose name we abbreviate by G, can be completed to either of the graphs G(V1,c1) or G(V2,c2) by adding the appropriate 38 ADAM MERBERG -4 -3 -2 -1 1 2 3 4 Figure 9. The directed graph G{−1,−3},{1,4} for the word considered in Ex- ample 3. arcs. The former arises from the map ι1 : {1, 4} → {−3, −1} given by 1 7→ −3 and 4 7→ −1 and the latter arises from ι2 with 4 7→ −3 and 1 7→ −1. Following the maximal paths of the graph, one sees that the bijection ǫ is given by 4 7→ −3 and 1 7→ −1. Thus, the bijection ι2ǫ−1 is the identity permutation on {−3, −1} and ι1ǫ−1 is the 2-cycle on the set {−3, −1}. Both of the graphs GV1,c1 and GV2,c2 have 4 maximal increasing paths, and these graphs have 2 and 3 cycles, respectively. The graph G has exactly 1 cycle. The following is a partial analog of Lemma 5.1 of [BG02]. Proposition 19. Suppose that A ∈ Pw(H, B) and b ∈ I with wb ≥ w−b. Then for i ∈ N, (169) ab,ia∗ b,iAΩN =(cid:18)1 + 1 N Nb,i(cid:19) AΩN . Proof. It will suffice to show that if B ∈ Pw(H, B) then (170) ρN (B∗ab,ia∗ b,iA) =(cid:18)1 + wA b (i) N (cid:19) ρN (B∗A). To save space, we define X := B∗ab,ia∗ b,iA. We will assume that b = 1 as the case b = −1 is similar. Let ℓA be the length of the word A and ℓB the length of the word B, so that X is a word of length ℓA + ℓB + 2. It will be convenient to choose J = [ℓA + ℓB + 2] as our index set for the product X = B∗ab,ia∗ b,iA and K = J \ {ℓB + 1, ℓB + 2} as the index set for the product B∗A. These choices allow us to write the products X and B∗A as (171) aek bk,ik X =Yk∈J and B∗A = Yk∈K aek bk,ik for some choices of bk, ik, and ek. Using the assumption that wb ≥ w−b, if (V, c) ∈ C(X) then Proposition 15 implies that ℓB + 1 ∈ DV ,c, which means that ZV ,c(ℓB + 1) = ℓB + 2, whence (ℓB + 1, ℓB + 2) ∈ ¯V (c). Then (172) ρN (X) = X(V ,c)∈C(X) (ℓB +1,ℓB+2)∈V tN (V, c) + X(V ,c)∈C(X) (ℓB +1,ℓB+2)6∈V tN (V, c). If (V, c) ∈ C(X) with (ℓB + 1, ℓB + 2) ∈ V then (W, d) ∈ C(B∗A), where W = V \ {(ℓB + 1, ℓB + 2)} and d is the restriction cW. Furthermore, ¯W (d) = ¯V (c) \ {(ℓB + 1, ℓB + 2)} and ¯d = ¯c ¯W . Thus the graph GV ,c can be formed from the graph GW,d by adding the two vertices (ℓB + 1, ℓB + 2) and 1 arc between these vertices in each direction. This means that REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 39 γ(GV ,c) = γ(GW,d) + 1 and m(GV ,c) = m(GW,d) + 1, whence tN (V, c) = tN (W, d), and (173) X(V ,c)∈C(X) (ℓB +1,ℓB+2)∈V tN (V, c) = X(W,d)∈C(B∗A) tN (W, d) = ρN (B∗A). b,iA,a∗ a∗ b,iA,a∗ a∗ b,iB(V, c) and Y −,b,i Now suppose that (V, c) ∈ C(X) with (ℓB + 1, ℓB + 2) 6∈ V. We will need the sets Y +,b,i b,iB(V, c) given by Proposition 18, and we denote them simply by Y +(V, c) and Y −(V, c). Since we have chosen a different index set than in the statement of Proposition 18, we have in this case Y +(V, c) ⊂ [ℓB + 2, ℓA + ℓB + 2] and Y −(V, c) ⊂ [ℓB + 1]. The assumption (ℓB + 1, ℓB + 2) 6∈ V implies that πV (ℓB + 2) ∈ Y +(V, c) and πV(ℓB + 1) ∈ Y −(V, c). In particular, πV (ℓB + 2) ∈ Y −(V, c) and πV (ℓB + 1) ∈ Y +(V, c). If W = V \ {(πV(ℓB + 2), ℓB + 2), (ℓB + 1, πV(ℓB + 1))} ∪ {(πV(ℓB + 2), πV(ℓB + 1))} and d : W → I is given by d(p) = c(p) for p ∈ V and d(πV(ℓB + 2), πV(ℓB + 1)) = c(ℓB + 2) = c(ℓB + 1) then (W, d) ∈ C(B∗A). The graphs GV ,c and GW,d have the same number of cycles, but GV ,c has one more maximal increasing path than GW,d, whence tN (W, c) = 1 N tN (W, d). The correspondence (V, c) 7→ (W, d) is not injective. Given (W, d) ∈ C(B∗A), one can (i) pairs (k, k′) with k ∈ Y −(V, c) and k′ ∈ Y +(V, c) with the pairs replace any of the w(A) (k, ℓB + 2) and (ℓA + 2, k′) to get an element of C(X). We have thus shown that 1 (174) tN (V, c) = w(A) 1 X(V ,c)∈C(X) (ℓB +1,ℓB+2)6∈V (i) X(W,d)∈C(B∗A) 1 N tN (W, d) = (i) w(A) 1 N ρN (B∗A). This proves (170). (cid:3) Lemma 2. For all b ∈ I and all i ∈ N, the creation operator a∗ b,i is bounded. Proof. Let (175) F< = Mwb<w−b Hw(H, B) and F≤ = Mwb≤w−b Hw(H, B), and define F≥ and F> analogously. Then we have two decompositions of F I N (H): (176) F I N (H) = F≤ ⊕ F> = F< ⊕ F≥. Using these decompositions, we can write a creation operator a∗ b,i as (177) a∗ b,i = a∗ b,i,< ⊕ a∗ b,i,≥ b,i,< : F< → F≤ and a∗ b,i,< : F≥ → F> are the restrictions of a∗ b,i. It will suffice to b,i,< and a∗ b,i,≥ are bounded operators. Boundedness of the operator a∗ need only show that a∗ which we denote by ab,i,<. We will find an upper bound for b,i,≥ is an immediate consequence of Proposition 19, so we b,i,< is bounded. This will follow from boundedness of its adjoint, where a∗ show that a∗ (178) with A, B ∈ Pw(H, B) having norm 1 and wb ≤ w−b. By Lemma 1, we can assume that 0 ≤ wb(i) ≤ N. Denote by ℓA the length of the word A and by ℓB the length of the word B. Let X = b,iA so that X has length ℓA +ℓB +2. Index the product X on the set J = [ℓA +ℓB +2] B∗ab,ia∗ ρN(cid:0)B∗a∗ b,iab,iA(cid:1) 40 ADAM MERBERG and let K = J \ {ℓB + 1, ℓB + 2} as the index set for the product B∗A. These choices permit us to write the products X and B∗A as (179) X =Yk∈J aek bk,ik and B∗A = Yk∈K aek bk,ik for some choices of bk, ik, and ek. Suppose that (V, c) ∈ C(X). We will assume here that b = 1, as the case b = −1 is very similar. By the assumption wb ≤ w−b and Proposition 15, ℓB + 1, ℓB + 2 ∈ SV ,c. Thus (ℓB + 1, ℓB + 2) ∈ ( ¯V, ¯c). Letting πV be the permutation arrived at by treating V as a product of transpositions, πV (ℓB + 1) ∈ [ℓB] and πV (ℓB + 2) ∈ [ℓB + 3, ℓA + ℓB + 2],. Thus, there exists an increasing path in GV ,c, πV(ℓB + 1) → ℓB + 1 → ℓB + 2 → πV (ℓB + 2). Replacing this path with a single edge, (πV(ℓB + 1), πV(ℓB + 2)) gives the graph GW,d for W = V \ {(ℓB + 1, ℓB + 2)} ∪ (πV (ℓB + 1), πV(ℓB + 2)) and d(p) = c(p) for p ∈ V and d(ℓB + 1, ℓB + 2) = c(πV (ℓB + 1), ℓB + 1). The graphs GW,d and GV ,c have the same numbers of cycles and maximal increasing paths, and the correspondence (V, c) 7→ (W, d) is wb(i)-to-1. Since 0 ≤ wb(i) < N, it follows that ρN(cid:0)B∗a∗ Proposition 20. For M ∈ B(F I N (H)), b ∈ I and n ∈ N define b,iab,iA(cid:1) ≤ N, whence (cid:13)(cid:13)a∗ b,i,<(cid:13)(cid:13) <pN. (cid:3) (180) Φb,n(M) := ωb,2n · · · ωb,n+1Mωb,n+1 · · · ωb,2n. The following limiting relations hold: b,ia∗ (1) w-limn→∞ Φb,n(a∗ b,i) = 0; (2) w-limn→∞ Φb,n(ab,iab,i) = 0; (3) w-limn→∞ Φb,n(ab,ia∗ (4) w-limn→∞ Φb,n(a∗ (5) w-limn→∞ Φb,n(1) = 1; (6) w-limn→∞ Φb,n(Nb,i) = Nb,i. b,iab,i) = 1 b,i) = 1 + 1 N Nb,i; N 2 Nb,i; Proof. Items 5 and 6 are straightforward, and item 3 follows from Proposition 19 together with items 5 and 6. Item 2 will follow immediately from 1 and continuity of the map X 7→ X ∗ in the weak operator topology. It will therefore suffice to prove items 1 and 4. We will assume that b = 1 as the case b = −1 is similar. For the proof of 1, fix words A, B ∈ P(H, B), and let ℓA be the length of the word A and ℓB the length of the word B. We index the terms in b,ia∗ Xn := B∗ab,2n · · · ab,n+1a∗ (181) b,ia∗ b,n+1 · · · a∗ b,2nA, considered as a product of the operators ae b,k, with the set (182) Jn := [−ℓB − n − 1, ℓA + n + 1] \ {0}. We write the word Xn as a product (183) Xn := Yk∈Jn aek bk,ik . By the definition of Xn, bk = 1 for k ∈ [−n − 1, n + 1] \ {0}, ek = 1 for k ∈ −[2, n + 1], ek = 2 for k ∈ [n + 1] ∪ {−1} and ik = n + k − 1 for k ∈ [−n + 1, n + 1] \ {−1, 0, 1}. Moreover i1 = i−1 = i. Since for k ∈ [2, n+1], k and −k are the only indices for which an operator ae n+k−1 appears in the product Xn, if (V, c) ∈ C(Xn) then (k, −k) ∈ V for k ∈ [2, n]. Assume that n is large REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 41 -8 -7 -6 -5 -4 -3 -2 -1 1 2 3 4 5 6 7 Figure 10. Part of the directed graph GV ,c considered in the proof of item 1 of 20. enough that n + wA condition ensures that k ∈ DV ,c for all k ∈ [2r]. One can check that −1 > 2r for some fixed r ∈ N. Together with Proposition 15, this 1 − wA (184) − 1 V ,c ∼ −2, 1 V ,c ∼ −3, 2 V ,c ∼ −4, · · · , 2r V ,c ∼ −2r − 2, whence (−2, −1), (−3, 1), (−4, 2), · · · , (−r − 2, r) ∈ V. This means that the graph GV ,c has a path (185) − 2 → 2 → −4 → 4 → −6 → 6 → · · · → −2r → 2r. Each arc (−k, k) for k ∈ [r] is a maximal increasing path, so the cycle containing this path has at least r maximal increasing paths, whence tN (V, c) ≤ N 1−r. Since the cardinality C(Xn) does not depend on n, it follows that (186) ρN (Xn) = X(V ,c)∈C(Xn) tN (V, c) ≤ CN 1−r for some constant C, whence Φn(a∗ b,ia∗ b,i) → 0 weakly. We now move on to proving item 4. Fix words A, B ∈ P(H, B) of length ℓA and ℓB with wA = wB. It will suffice to show that (187) ρN (B∗ab,2n · · · ab,n+1a∗ b,iab,ia∗ b,n+1 · · · a∗ b,2nA) = wA 1 (i) N 2 ρN (B∗A) for n sufficiently large. We assume that n is large enough that n + wA 1 > wA −1 and that no ae b,m with m > n appears in either word A or B. We let (188) Xn := B∗ab,2n · · · ab,n+1a∗ b,iab,ia∗ b,n+1 · · · a∗ b,2nA, 42 ADAM MERBERG and write this product as (189) where Kn :=Sl∈{−1,0,1} K (l) n = −[ℓB], K 0 K (−1) (190) n with Xn := Yk∈Kn aek bk,ik . n =(cid:8)−n − 1, . . . , −1, 1, · · · , n + 1(cid:9) , K (1) n = [ℓA], where for k denotes a distinct copy of the integer k ∈ Z. We order Kn by imposing the usual order on Z on each K (l) n with l < l′ set k < k′. Setting K ′ n and for k ∈ K (l) n and k′ ∈ K (l′) ∪ K (1) n , n := K (−1) n (191) B∗A = Yk∈K ′ n aek bk,ik . If (V, c) ∈ C(Xn) then as in the proof of item 1, (−k, k) ∈ V for k ∈ [2, n + 1]. Let k+, k− ∈ K ′ n be such that (k−, −1), (1, k+) ∈ V. Let (192) W := V \ {(−k, k) : k ∈ [2, n]} ∪ {(k−, k+)} V ,c∼ −2 and 1 and define d : W → I by d(p) = c(p) for c ∈ V\{(−k, k) : k ∈ [2, n]} and d(k−, k+) = c(k−) = c(k+). Then (W, d) ∈ C(B∗A), and rV ,c(k) = rW,d(k) for any k ∈ K ′ n. By Proposition 15, V ,c∼ 2 so that (−2, −1), (1, 2) ∈ ¯V (c). If k ∈ [2, n] and k ∈ −1, 1 ∈ DV ,c with −1 DV ,c then (−k, k) ∈ ¯V (c) so that −k and k are on a cycle with exactly 1 maximal increasing path. If instead k ∈ [2, n] with k ∈ SV ,c then ZV ,c(k) = ZW,d(−k) and ZV ,c(−k) = ZW,d(k). This shows that the correspondence (V, c) 7→ (W, d) preserves the cycle structure of the corresponding graph except that it removes some number of cycles with exactly 1 maximal increasing path and reduces by 2 the number of maximal increasing paths in the cycle through k+ and k−. As such, tN (V, c) = 1 N 2 tN (W, d). Similar to the proof of Proposition 19, there are w(A) 1 pairs (V, c) mapped to of each (W, d) by the correspondence just described, so (187) follows. (cid:3) Proposition 21. Let 0b : N → Z be the constant function 0b(n) = 0 for b ∈ I. Then H0(H) = CΩN . Proof. It is sufficient to show that for any A ∈ P0, (193) This follows from the fact that (V, c) ∈ C(A∗A) cannot have any pairs (k, k′) with k ≤ ℓA and k′ > ℓA, where ℓA is the length of the word A. (cid:3) kρN (A)ΩN − AΩN kN = 0. The following is a partial Proposition 22. If N < 0 and H is an infinite dimensional real Hilbert space, then ΓI contains the projection onto the vacuum vector ΩN . N (H) Proof. By Proposition 20, (194) w-limn→∞ Φn(ω2 b,i) = 1 + N + 1 N 2 Nb,i. In particular, Nb,i ∈ ΓI N (H). REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 43 It is a consequence of the definitions of the operators Nb,i that (195) ker Nb,i = Mwb(i)=0 Hw(H, B). If Pb,i is the projection onto ker Nb,i and PΩN is the projection onto the vacuum vector then by Proposition 21, (196) whence PΩN ∈ ΓI N (H). PΩN = inf{Pb,i : b ∈ I, i ∈ N}, (cid:3) In [BG02], Bozejko and Gut¸a used a result analogous to Proposition 22 to show that a von Neumann algebra under consideration was in fact the whole space of bounded operators. However, they worked in a setting with a cyclic vacuum vector. We do not yet know whether the vacuum vector is cyclic in our context. Accordingly, the best that we can prove is the following. N (H) := ΓI N (H) = B( F I N (H)). N (H)ΩN , and define ΓI Proposition 23. Let F I Then ΓI Proof. Suppose that Q is a nonzero projection in ΓI (197) In arriving at the second equality, we use the assumption that Q commutes with PΩN ∈ ΓI N (H), the projection onto the vacuum. Since ΩN is cyclic for the action of ΓI N (H) on N (H), it follows that Q = 1, whence ΓI F I (cid:3) N (H) =nX F I QXΩN = X(QΩN ) = XΩN . N (H)′. For X ∈ ΓI N (H), N (H) : X ∈ ΓI N (H)o. N (H)′ = C. 6. The qij-product of generalized Brownian motions In this section, we present a generalization of Gut¸a's q-product of noncommutative gen- eralized Brownian motions. cr(V) = {((a1, z1), (a2, z2)) ∈ V × V : a1 < a2 < z1 < z2}. Definition 13. For V ∈ P2(∞), define the set of crossings of V by (198) If (V, c) ∈ P I 2 (∞), we also define cr(V, c) = cr(V). Suppose that for each i ∈ I, a positive- definite function ti : P2(∞) → C is given, and that we have a (possibly infinite) matrix Q = (qij)i,j∈I with qij = qji and qij ∈ [−1, 1]. Then we define the Q-product of the ti to be the function on P I 2 (∞) given by (199) (cid:16)∗Q bnItb(cid:17) (V, c) := Y(p,p′)∈cr(V) qc(p),c(p′)Yb∈I tb(c−1(b)). Definition 14. A function t : P I with 1 ≤ k ≤ l ≤ n and any I-colored pair partitions V1 ∈ P I V2 ∈ P I 2 (∞)({k + 1, . . . , l − 1}), we have t(V1 ∪ V2) = t(V1) · t(V2). 2 (∞) → C is said to be multiplicative if for every k, l, n ∈ N 2 (∞)({1, . . . , k, l, . . . , n}) and Proposition 24. Suppose that tb : P2(∞) → C (b ∈ I) are multiplicative positive definite functions such that t(V, c) = 1 whenever V is the element of P2(∞) with only one pair. Suppose also that for each i, j ∈ I, some symmetric Q = (qij)i,j∈I with qij ∈ [−1, 1] is given. Then (cid:16)∗Q a∈Itb(cid:17) (V, c) is a positive definite function on P I 2 (∞). 44 ADAM MERBERG Remark 14. In [Gut¸03], the number of crossings between pairs of different colors was used instead of all crossings. Of course, if we wish to impose this restriction in our framework, we can assume qbb = 1 for all b ∈ I. The proof is essentially the same as the proof of positive definiteness of the q-product in [Gut¸03] but we present the argument again here for completeness. Proof. As a first step, we show that for each n : I → N, the kernel kn defined on BP I by 2 (n, 0) (200) (201) is positive definite. Using the definition of the Q product, 1 · d2). kn(d1, d2) =(cid:16)∗Q kn(d1, d2) = Y(p,p′)∈cr(d∗ b∈I tb(cid:17) (d∗ qb,b′Yb∈I 1·d2) p∈(d∗ p′∈(d∗ 1·d2)b 1·d2)b′ tb((d∗ 1 · d2)b), where the subscript b refers to the b-colored pair partition. Since the tb are positive definite and the pointwise product of positive definite kernels is positive definite, if we can show that (202) k′ n(d1, d2) = Y(p,p′)∈cr(d∗ 1·d2) p∈(d∗ p′∈(d∗ 1·d2)b 1·d2)b′ qb,b′ then positive definiteness of kn will follow. However, positive definiteness of k′ n follows from positivity of the vacuum state on a ∗-algebra generated by annihilation operators ab,i for i = 1, . . . , n(b) satisfying the commutation relation (203) ab,ia∗ c,j − qb,ca∗ c,jab,i = δb,cδi,j. Positivity of that state has already been proven by Bozejko and Speicher in [BS94]. For each n denote the complex Hilbert space generated by the positive definite kernel kn 2 (n, 0) → Vn be the Gelfand map, i.e. hλn(d1), λn(d2)i = kn(d1, d2). 2 (n, 0) preserves kn, and thus gives by Vn and let λn : BP I The natural action of the symmetric group S(n) on BP I rise to a unitary representation Un on Vn. On V := Ln Vn, define the operators jb (for a ∈ I) by jbλn(d1) = λn+δb(db,0 · d1). By multiplicativity of tb (b ∈ I), (204) kn(db,0 · d1, db,0 · d2) = kn(d1, d2), which shows that the definition of jb makes sense. Since jb also satisfies the requisite inter- twining property, we have a representation of the ∗-semigroup BP I 2 (∞) on V with respect (cid:3) to the extension of (cid:16)∗Q b∈I tb(cid:17) to the broken pair partitions. As in the case of [Gut¸03], we can use this construction to define new positive definite functions on pair partitions provided that our index set I is finite. Assume that I is finite and tb is a multiplicative positive definite function for each b ∈ I. On the Fock-like space F(∗Q b∈I tb)(K), we can define creation operators 1 (205) a∗(f ) := a∗ b(f ) pIXb∈I REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 45 for f ∈ K. The restriction of the vacuum state to the ∗-algebra generated by the a∗(f ) is a Fock state, and we denote the associated positive definite function on BP 2(∞) by(cid:16)∗Q Explicitly, this function is given by b∈I tb(cid:17)(r) . (206) b∈Itb(cid:17)(r) (cid:16)∗Q (V) = 1 IV Xc:V→I Y(p,p′)∈cr(V) qc(p),c(p′)Yb∈I tb(c−1(b)). In the case that the functions tb are all the same, tb = t for all b ∈ I, we write t∗I Q for b∈I t(cid:17)(r) (cid:16)∗Q , and in the case I = [n] := {1, . . . , n}, we write t∗n Q . Remark 15. One can ask the question of whether a central limit theorem may be found in this context, similar to the Central Limit Theorem of [Gut¸03]. The Central Limit Theorem of Gut¸a concerns the q-product of n copies of a function t : P2(∞) → C. By definition, the q-product is a function on P I 2 (∞) for some set I of cardinality n. However, we can define a function on (uncolored) pair partitions by taking a normalized sum over all colorings. The content of Gut¸a's Central Limit Theorem is that this normalized sum converges to the positive definite function arising from the algebra of q-commutation relations, tq(V) = qcr(V). An analogous result in the qij setting will at least require additional assumptions on the qij. One can assume, for instance, that qij → q as i, j → ∞. In this case, the argument of Gut¸a [Gut¸03] can be extended to show an analogous result. We prefer to pursue a different line of inquiry, namely the case in which the qij are periodic in the indices i and j. Theorem 9 (Central Limit Theorem). Let Q ∈ MN (R) be an N × N real symmetric matrix. Let Qn be the n × n symmetric matrix with entries qij where qij = q¯i¯j for ¯i, ¯j such that 1 ≤ ¯i, ¯j ≤ n and i ≡ ¯i (mod N), j ≡ ¯j (mod N). Let t : P2 → C be a positive definite multiplicative function such that t(V1) = 1 where V1 is the pair partition consisting of a single pair. Then t∗n Qn converges pointwise to tQ, where (207) tQ(V) = N −V Xd:V→[N ] Y(p,p′)∈cr(V) qd(p)d(p′). Proof. Fix a pair partition V ∈ P2(∞). For a function c : V → [N] denote by P (c) the partition of V such that two pairs p and p′ are in the same block if and only if c(p) = c(p′). Then (208) t∗n Qn(V) = n−V Xc:V→[N ] Y(p,p′)∈cr(V) n−V Xc:V→[n] = Xπ∈Π(V) P (c)=π Y(p,p′)∈cr(V) qc(p),c(p′) Yb∈[n] qc(p),c(p′) Yb∈[n] t(c−1(b)) t(c−1(b)), where Π(V) is the set of all partitions of V. We will consider the contribution of the various π ∈ Π(V) to the sum as n → ∞. First consider the partition π1 of V into V blocks of size 1, corresponding (for fixed n) to injective functions c : V → [n]. For each such c and a ∈ [n], the pair partition c−1(b) is empty or a single pair, whence Qb∈[n] t(c−1(b)) = 1. Furthermore, for each c, (209) qc(p),c(p′), Y(p,p′)∈cr(V) qc(p),c(p′) = Y(p,p′)∈cr(V) 46 ADAM MERBERG where c : V → [N] is the map such that c(p) ≡ c(p) (mod N) for all p ∈ V. If M is the natural number such that MN < n ≤ (M + 1)N, then for each function d : V → [N], the number mn of injective maps c : V → [n] such that c = d is between M(M −1) · · · (M −V+1) and (M + 1)(M) · · · (M − V). In particular mn/nV → N −V as n → ∞. Thus, as n → ∞ the contribution to the sum in (208) by the term corresponding to P1 converges to (210) 1 N V Xd:V→[N ] Y(p,p′)∈cr(V) qd(p),d(p′) = tQ(V). Now we will show that any other partition π 6= π1 contributes 0 to the sum in (208) in the limit as n → ∞. Such a partition π has at most V − 1 blocks. For a given n, the number of maps c : V → [n] with P (c) = π is (211) n(n − 1) · · · (n − P + 1) ≤ n(n − 1) · · · (n − V + 2) < nV−1. Thus, the contribution of the term indexed by P is indeed 0 in the limit. (cid:3) References [Avs13] Stephen Avsec, New examples of exchangeable noncommutative brownian motions, Workshop on Analytic, Stochastic, and Operator Algebraic Aspects of Noncommutative Distributions and Free Probability (Fields Institute), July 2013. [BG02] Marek Bozejko and Madalin Gut¸a, Functors of white noise associated to characters of the infinite symmetric group, Comm. Math. Phys. 229 (2002), no. 2, 209 -- 227. MR 1923173 (2003g:81110) [BJG09] Jørgen Bang-Jensen and Gregory Gutin, Digraphs, second ed., Springer Monographs in Math- ematics, Springer-Verlag London, Ltd., London, 2009, Theory, algorithms and applications. MR 2472389 (2009k:05001) [BS91] Marek Bozejko and Roland Speicher, An example of a generalized Brownian motion, Comm. Math. Phys. 137 (1991), no. 3, 519 -- 531. MR 1105428 (92m:46096) [BS94] [BS96] , Completely positive maps on Coxeter groups, deformed commutation relations, and oper- ator spaces, Math. Ann. 300 (1994), no. 1, 97 -- 120. MR 1289833 (95g:46105) , Interpolations between bosonic and fermionic relations given by generalized Brownian mo- tions, Math. Z. 222 (1996), no. 1, 135 -- 159. MR 1388006 (97g:81031) [GM02a] Madalin Gut¸a and Hans Maassen, Generalised Brownian motion and second quantisation, J. Funct. Anal. 191 (2002), no. 2, 241 -- 275. MR 1911186 (2003d:46084) [GM02b] , Symmetric Hilbert spaces arising from species of structures, Math. Z. 239 (2002), no. 3, 477 -- 513. MR 1893849 (2003g:46085) [Gut¸03] Madalin Gut¸a, The q-product of generalised Brownian motions, Quantum probability and infinite dimensional analysis (Burg, 2001), QP -- PQ: Quantum Probab. White Noise Anal., vol. 15, World Sci. Publ., River Edge, NJ, 2003, pp. 121 -- 134. MR 2010602 (2004k:46108) [Leh05] Franz Lehner, Cumulants in noncommutative probability theory. III. Creation and annihilation operators on Fock spaces, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 8 (2005), no. 3, 407 -- 437. MR 2172307 (2007g:46096) James A. Mingo and Alexandru Nica, Crossings of set-partitions and addition of graded- independent random variables, Internat. J. Math. 8 (1997), no. 5, 645 -- 664. MR 1468355 (98h:46071) [MN97] [Oko97] A. Okoun′kov, On representations of the infinite symmetric group, Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov. (POMI) 240 (1997), no. Teor. Predst. Din. Sist. Komb. i Algoritm. Metody. 2, 166 -- 228, 294. MR 1691646 (2000c:20027) [Ols90] Grigori Olshanski, Unitary representations of (g, k)-pairs that are connected with the infinite sym- metric group s(∞), Leningrad Math. J 1 (1990), no. 4, 983 -- 1014. [Sta12] Richard P. Stanley, Enumerative combinatorics. Volume 1, second ed., Cambridge Studies in Ad- vanced Mathematics, vol. 49, Cambridge University Press, Cambridge, 2012. MR 2868112 REMARKS ON MULTI-DIMENSIONAL NONCOMMUTATIVE GENERALIZED BROWNIAN MOTIONS 47 [Tho64] Elmar Thoma, Die unzerlegbaren, positiv-definiten Klassenfunktionen der abzahlbar unendlichen, symmetrischen Gruppe, Math. Z. 85 (1964), 40 -- 61. MR 0173169 (30 #3382) [VK82] A. M. Vershik and S. V. Kerov, Characters and factor-representations of the infinite unitary group, Dokl. Akad. Nauk SSSR 267 (1982), no. 2, 272 -- 276. MR 681202 (84g:22032) [Voi85] Dan Voiculescu, Symmetries of some reduced free product C ∗-algebras, Operator algebras and their connections with topology and ergodic theory (Bu¸steni, 1983), Lecture Notes in Math., vol. 1132, Springer, Berlin, 1985, pp. 556 -- 588. MR 799593 (87d:46075) Department of Mathematics, University of California, Berkeley, CA, USA 94720 E-mail address: [email protected]
1304.8091	1	1304	2013-04-30T17:49:59	Deformation of involution and multiplication in a C*-algebra	[ "math.OA", "math.FA" ]	We investigate the deformation of involution and multiplication in a unital $C^$-algebra when its norm is fixed. Our main result is to present all multiplications and involutions on a given $C^$-algebra $\mathcal{A}$ under which $\mathcal{A}$ is still a $C^$-algebra whereas we keep the norm unchanged. For each invertible element $a\in\mathcal{A}$ we also introduce an involution and a multiplication making $\mathcal{A}$ into a $C^$-algebra in which $a$ becomes a positive element. Further, we give a necessary and sufficient condition for that the center of a unital $C^*$-algebra $\mathcal{A}$ is trivial.	math.OA	math	DEFORMATION OF INVOLUTION AND MULTIPLICATION IN A C∗-ALGEBRA H. NAJAFI AND M. S. MOSLEHIAN Abstract. We investigate the deformation of involution and multiplication in a unital C ∗-algebra when its norm is fixed. Our main result is to present all multiplications and involutions on a given C ∗-algebra A under which A is still a C ∗-algebra whereas we keep the norm unchanged. For each invertible element a ∈ A we also introduce an involution and a multiplication making A into a C ∗- algebra in which a becomes a positive element. Further, we give a necessary and sufficient condition for that the center of a unital C ∗-algebra A is trivial. 1. Introduction A C∗-algebra is a complex Banach ∗-algebra A satisfying ka∗ak = kak2 (a ∈ A). By the Gelfand -- Naimark theorem, a C∗-algebra is a norm closed ∗-subalgebra of B(H) for some Hilbert space H. A strongly closed ∗-subalgebra of B(H) containing the identity operator is called a von Neumann algebra. By the double commutant theorem a unital ∗-subalgebra A of B(H) is a von Neumann algebra if and only if A is equal to its double commutant Acc, where Ac = {B ∈ B(H) : AB = BA for all A ∈ A}. By Sakai's characterization of von Neumann algebras, A is a von Neumann algebra if and only if it is a W ∗-algebra, i.e. it is a C∗-algebra being the norm dual of a Banach space A . Throughout the paper A denotes an arbitrary C∗-algebra and Z(A) stands for its center. For a self adjoint element a ∈ A, it holds that r(a) = kak, where r(a) denotes the spectral radius of A. This implies that the norm of a C∗-algebra is unique when we fix the involution and the multiplication. Indeed, if A is a C∗-algebra under two norms k.k1 and k.k2, then kak1 = ka∗ak 2 = kak2 for all a ∈ A. Bohnenblust and Karlin [BK] showed that there is at most one involution on a Banach algebra with the unit 1 making it into a C∗-algebra (see also [R]): Let ∗ and # be two involutions on a unital Banach algebra A making it into C∗-algebras. 1 = r(a∗a) 2 = ka∗ak ∗ 1 2 1 2 1 2010 Mathematics Subject Classification. Primary 46L05; Secondary 46L10. Key words and phrases. C ∗-algebra; von Neumann algebra; involution; multiplication; positive element, center, double commutant. 1 2 H. NAJAFI, M.S. MOSLEHIAN Let x ∈ A. It follows from the fact "an element x of a unital C∗-algebra is self- adjoint if and only if τ (x) is real for every bounded linear functional τ on A with k τ k= τ (1) = 1 ([KR 1, Proposition 4.3.3])" that x is self-adjoint with respect to ∗ if and only if it is self-adjoint with respect to #. Now let a ∈ A be arbitrary and a = a1 + ia2 with self-adjoint parts a1, a2 with respect to ∗. Then a∗1 = a# 1 and a∗2 = a# 2 and a∗ = a1 − ia2 = a#. There is another way to show the uniqueness of the involution. Indeed if ∗ and # be two involutions on a unital Banach algebra A making it into C∗-algebras, then the identity map from (A, ∗) onto (A, #) is positive (see [P, Proposition 2.11]) and so a∗ = a# for all a ∈ A. There are several characterizations of C∗-algebras among involutive Banach alge- bras, see [DT] in which the authors start with a C∗-algebra and modify its structure. We however investigate a different problem in the same setting. In fact we investi- gate the deformation of involution and multiplication in a unital C∗-algebra when its norm is fixed. Our main result is to present all multiplications ◦ and involutions ⋆ on a given C∗-algebra A under which A is still a C∗-algebra whereas we keep the norm unchanged. As an application, for each invertible element a ∈ A we introduce an involution and a multiplication making A into a C∗-algebra in which a becomes a positive element. Further, we give a necessary and sufficient condition for that the center of a unital C∗-algebra A is trivial. Recall that a Jordan ∗-homomorphism is a self-adjoint map preserving squares of self-adjoint operators. Jacobson and Rickart [JR] showed that for every Jordan ∗-homomorphism ρ of a C∗-algebra A with the unit 1 into a von Neumann algebra B there exist central projections p1, p2 ∈ B such that ρ = ρ1 + ρ2, ρ(1) = p1 + p2, ρ1(a) = ρ(a)p1 is a ∗-homomorphism and ρ2(a) = ρ(a)p2 is a ∗-antihomomorphism. Kadison [K] showed that an isometry of a unital C∗-algebra onto another C∗-algebra is a Jordan ∗-isomorphism. We start our work with the following lemma. 2. Results Lemma 2.1. Let A be a unital C∗-algebra of operators acting on a Hilbert space H. Let p ∈ A be a central projection and u ∈ A be a unitary. Let ◦ be the multiplication and ⋆ be the involution defined on A by a ◦ b = paub + (1 − p)bua and a⋆ = u∗a∗u∗ (2.1) for a, b ∈ A, respectively. Then A equipped with the multiplication ◦ and the invo- lution ⋆ is a unital C∗-algebra. Proof. It is easy to check that A is a complex Banach algebra under the mul- tiplication ◦ and u∗ is the unit for this multiplication. By the decomposition DEFORMATION OF INVOLUTION AND MULTIPLICATION IN A C ∗-ALGEBRA 3 H = pH ⊕ (1 − p)H, we can represent any element a ∈ A by the 2 × 2 ma- trix   . For a, b ∈ A, therefore pa + (1 − p)b can be identified by   0 (1 − p)a pa 0   pa 0 0 (1 − p)b  , whence pa + (1 − p)b = max(kpak, k(1 − p)bk). Hence  ka⋆ ◦ ak = kp u∗a∗a + (1 − p) aa∗u∗k = max(kpa∗ak, kaa∗(1 − p)k) = max(kpak2, k(1 − p)ak2) = max(kpak, k(1 − p)ak)2 = kak2 for all a ∈ A. (cid:3) The unital C∗-algebra A equipped with the multiplication ◦ and the involution ⋆ is denoted by A(◦, ⋆). Next we establish a converse of Lemma 2.1. Theorem 2.2. Let A be a unital C∗-algebra of operators acting on a Hilbert space H and there exist a multiplication ◦ and an involution ⋆ on the normed space A making it into a C∗-algebra. Then there exists a unitary element u ∈ A and a central projection p in the double commutant of Acc of A such that both equalities (2.1) hold. Proof. Since A is unital, the closed unit ball of A has an extreme point, hence the C∗- algebra A(◦, ⋆) is unital. Since ι(x) = x is an isometric linear map of A onto A(◦, ⋆), the unitary elements of A(◦, ⋆) and those of A coincide [KR 2, Exercise 7.6.17]. Thus if u∗ is the unit of A(◦, ⋆), then u is a unitary of A. Define ρ : A → A(◦, ⋆) by ρ(a) = u∗a. Clearly ρ is a unital isometric linear map of A onto A(◦, ⋆). Hence ρ is a positive map. This implies that u∗a∗ = ρ(a∗) = (u∗a)⋆ and so a⋆ = u∗a∗u∗. For determining the multiplication, define a multiplication ⋄ on Acc (with respect to the original multiplication) by (2.1) as p = 1. Then Acc with the multiplication ⋄ is a C∗-algebra. The space Acc as a Banach space is already the dual of a Banach space, so Acc with the new product and the new involution is a von Neumann algebra. Then the map ρ(x) = x is a unital isometric linear map of A(◦, ⋆) into the von Neumann algebra Acc(⋄, ⋆). By the result of Kadison [K] it is a Jordan ∗-isomorphism and by the Jacobson and Rickart theorem [JR] there exists a central ′ ⋄ ρ(x) is a ∗-homomorphism and projection p ′) ⋄ ρ(x) is a ∗-antihomomorphism. Therefore for each a, b ∈ A we ρ2(x) = (u∗ − p ′ in Acc(⋄, ⋆) such that ρ1(x) = p 4 have H. NAJAFI, M.S. MOSLEHIAN a ◦ b = ρ(a ◦ b) = ρ1(a ◦ b) + ρ2(a ◦ b) ′ ′ ′ ′ = p = p = p = p ⋄ ρ1(a) ⋄ ρ1(b) + (u∗ − p ′ ) ⋄ ρ2(b) ⋄ ρ2(a) ⋄ a ⋄ b + (u∗ − p ′ ) ⋄ b ⋄ a. uaub + (u∗ − p ′ )ubua uaub + (1 − p ′ u)bua. ′ ′ ′ ′ ′ ′ ′ up u)2 = p u1 = p ′ ⋆u = p u. Since (p ∗u∗u = p ′ u)∗ = Let p = p u, so p is a projection in Acc. A similar argument shows ∗ = u∗p u∗p that θ : Acc(⋄, ⋆) → Acc defined by θ(a) = au is a Jordan ∗-isomorphism. So, by ′)θ(au∗) = [JR, Corollary 1] , θ(Z(Acc(⋄, ⋆))) = Z(θ(Acc(⋄, ⋆))). Therefore pa = θ(p θ(au∗)θ(p (cid:3) ′) = ap for each a ∈ A. Hence p is a central projection in Acc. u and (p ′ ⋄ (p ′ ⋄ 1) = p ′ ⋄ 1 = p ′ Remark 2.3. Note that in general case, a C∗-algebra A has many representations. However the proof of Theorem 2.2 shows that for any representation of A, we can present all multiplications and involutions on A which keep it still a C*-algebra with the same norm by a unitary and a central projection in the double commutant with respect to the same representation. Further, since p in Theorem 2.2 is in Acc ⊆ B(H), it depends on H. If A is a von Neumann algebra, then p ∈ Acc = A. Corollary 2.4. Let I be an ideal of a von Neumann algebra A. Then I is also an ideal of the C∗-algebra A(◦, ⋆) for any multiplication ◦ and any involution ⋆. Proof. It is sufficient to note that paub and (1−p)bua belong to I when a ∈ A, b ∈ I and so a ◦ b = paub + (1 − p)bua ∈ I. (cid:3) It is easy to see that a ◦ b = b ◦ a if and only if aub = bua. We therefore have Corollary 2.5. Suppose that A is a unital C∗-algebra and the normed space A is equipped with a multiplication ◦ and an involution ⋆ is a C∗-algebra with the unit u∗, where u ∈ A is a unitary. Then (i) A is commutative if and only if so is A(◦, ⋆). (ii) Z(A) = C1 if and only if Z(A(◦, ⋆)) = Cu∗. Proof. (i) Let A be commutative. By Theorem 2.2 there exist a unitary element u ∈ A and a central projection p in Acc such that a ◦ b = paub + (1 − p)bua (a, b ∈ A). DEFORMATION OF INVOLUTION AND MULTIPLICATION IN A C ∗-ALGEBRA 5 Hence a ◦ b = paub + (1 − p)bua = pbua + (1 − p)aub = b ◦ a . Therefore A(◦, ⋆) is commutative. Changing the role of A by A(◦, ⋆), we reach the reverse assertion. (ii) Let Z(A) = C1. If a ∈ Z(A(◦, ⋆)), then for any b ∈ A we have a◦b = b◦a. As in the proof of Theorem 2.2 we observe that θ : Acc(◦, ⋆) → Acc defined by θ(a) = au is a Jordan ∗-isomorphism. Hence, by [JR, Corollary 1] , θ(b)θ(a) = θ(a)θ(b), so aubu = buau. Since each element of A is of the form bu for some b ∈ A, it follows that au ∈ Z(A). Hence au = λ1 for some λ ∈ C. Therefore Z(A(◦, ⋆)) = Cu∗. Similarly we can deduce the converse. (cid:3) Remark 2.6. The Arens product on (c0)∗∗ = l∞ coincide with the usual product in l∞ [D, Example 2.6.22]. It was extended to arbitrary C∗-algebras in [BD]. We reprove the fact in our own way: Let A be a C∗-algebra and its second dual A∗∗ be also a C∗-algebra under a multiplication (a, b) 7−→ a · b whose restriction to A × A is the same multiplication of A. We shall show that the Arens product (denoted by ⋄) on A∗∗ is the same as the multiplication · on A∗∗. It is known that A∗∗ is a von Neumann algebra under the Arens multiplication [D, Theorem 3.2.37]. By the Kaplansky density theorem, A is dense in A∗∗ in the weak∗-topology, so there exists a net uα in A such that uα → 1 in the weak∗-topology in which 1 denotes the unit of A∗∗. So b = w∗ − lim α uαb = w∗ − lim α uα ⋄ b = 1 ⋄ b for each b ∈ A. The Kaplansky density theorem implies that 1 ⋄ x = x for each x ∈ A∗∗. Therefore the units of both multiplications · and ⋄ are same. By Theorem 2.2 there exist a central projection p ∈ A such that x ⋄ y = pxy + (1 − p)yx, for each x, y ∈ A∗∗. On the other hand for each a, b ∈ A, we have a ⋄ b = ab. So (1 − p)ab = (1 − p)ba. Since A is dense in A∗∗ in the weak∗-topology, we have (1 − p)xy = (1 − p)yx for each x, y ∈ A∗∗. Therefore x ⋄ y = pxy + (1 − p)yx = pxy + (1 − p)xy = xy for each x, y ∈ A∗∗. For instance, we deduce that the Arens product on K(H)∗∗ = B(H) is equal to the operator multiplication on B(H). Theorem 2.7. Let A be a unital C∗-algebra. Then the following assertions are equivalent: (i) Z(A) = C1 (ii) If for invertible operators a, b ∈ A, axb = x holds for each x ∈ A, then there exists λ > 0 such that both λa and 1 λ b are unitary. 6 H. NAJAFI, M.S. MOSLEHIAN Proof. (i) ⇒ (ii) Note that if a−1xa ≤ x for each x ∈ A, then map ϕ(x) = a−1xa is a contractive unital linear map on A . It follows from [P, Proposition 2.11] that ϕ is positive. Therefore (a−1xa)∗ = a−1x∗a and so aa∗x∗ = x∗aa∗ for each x ∈ A. Hence aa∗ ∈ Z(A) = C1. So a∗a = λ1 for some λ > 0. Therefore 1 √λ a is unitary. First, assume that axb = x for positive invertible operators a, b and each x ∈ A. Then b−1a−1 = a−1b−1 = aa−1b−1b = 1, whence a−1xa ≤ axb b−1a−1 ≤ x. ′ λ λ that λ = 1 λ a is unitary. Since 1 Therefore there exists λ > 0 such that 1 λ a is positive and ′. unitary we have a = λ. A similar argument shows that b = λ It follows from 1 = 1 = ab = λ ′ . Second, assume that axb = x for invertible operators a, b and each x ∈ A. Utilizing the polar decompositions of a and b∗, there exist unitary operators u, v such that a = ua and b = b∗v. Hence a x b∗ = u a x b∗ v = axb = x for each x ∈ A. The above argument shows that a = λ and b∗ = 1 (ii) ⇒ (i) Note that each central invertible element a of A is a scalar multiple of a unitary element. In fact, we have a−1xa = a−1ax = x for all x ∈ A, so λa is unitary for some λ > 0. Let a ∈ Z(A) be a positive element and λ1, λ2 ∈ sp(a) are distinct. Then there exists an invertible continuous function f on sp(a) such that f (λ1) = 1 2 and f (λ2) = 1. Hence f (a), which is a central invertible element should be a scalar multiple of a unitary. On the other hand, 1 2 , 1 ∈ sp(f (a)), which is impossible. Hence the spectrum of a is singleton, so a = a. Since Z(A) is a C∗-algebra, any one of its elements is a linear combination of four positive elements. Therefore Z(A) = C1. (cid:3) λ for some λ > 0, so a = λu and b = 1 λ v. Let A(u, p) denote the C∗-algebra given via Lemma 2.1 corresponding to a unitary u and a central projection p in A. The self-adjoint elements of A(u, p) are all elements a such that au = u∗a∗, a fact which is independent of the choice of p. Also a self-adjoint element a is positive in A(u, p) if and only if a = b◦b = pbub+(1−p)bub = bub for some self-adjoint element b ∈ A(u, p) and this occurs if and only if a is positive in A(u, 1). Theorem 2.8. Let A be a C∗-algebra and a ∈ A be invertible. Then there exists a unique unitary u ∈ A such that a is a positive element of the C∗-algebra A(u∗, p) for any central projection p ∈ A. 1 1 2 a Proof. Let a = ua be the polar decomposition of a. Then u = aa−1 ∈ A. So a = ua 2 , where ◦ is defined in A(u∗, 1) by (2.1). So a is positive in A(u∗, p) for every central projection p ∈ A. To see the uniqueness, note that if a is invertible and a, wa are positive for a unitary w, then a = w∗(wa). By the 2 = a ⋆ ◦ a 1 2 1 DEFORMATION OF INVOLUTION AND MULTIPLICATION IN A C ∗-ALGEBRA 7 uniqueness of polar decomposition, we have w = 1. Now if a is positive in A(v∗, 1), then a = b⋆ ◦ b = vb∗b. Hence v∗ua = v∗a = b∗b is positive. Therefore v∗ua and a are positive and so v = u according to what we just proved. (cid:3) Remark 2.9. The invertibility condition in Proposition 2.8 is essential. For example let A = C[−1, 1] and f (t) = t. If f is positive in C[−1, 1](u, 1) for a unitary function u, then there exist g ∈ C[−1, 1] such that t = f (t) = u(t)g(t)2 for each t ∈ [−1, 1]. So u(t) = 1 for each t ∈ (0, 1] and u(t) = −1 for each t ∈ [−1, 0), which is impossible. Acknowledgements. The authors would like to sincerely thank Professor Marcel de Jeu and Professor Jun Tomiyama for their valuable comments. References [BK] H.F. Bohnenblust and S. Karlin, Geometrical properties of the unit sphere of Banach algebras, Ann. of Math. (2) 62, (1955), 217-229. [BD] F.F. Bonsall and J. Duncan, Complete Normed Algebras, Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 80. Springer-Verlag, New York- Heidelberg, 1973. [D] H.G. Dales, Banach Algebras and Automatic Continuity, London Math. Soc. Monogr. Ser., vol. 24, Clarendon Press, Oxford, 2000. [DT] M. de Jeu and J. Tomiyama, A characterisation of C∗-algebras through posi- tivity of functionals, Ann. Funct. Anal. 4 (2013), no. 1, 61 -- 63. [JR] N. Jacobson and C. Rickart,Homorphisms of Jordan rings, Trans. Amer. Math. Soc. 69 (1950), 479 -- 502. [K] R.V. Kadison,Isometries of operator algebras, Ann. Math. 54 (1951), 325 -- 338. [KR 1] R.V. Kadison and J.R. Ringrose,Fundamentals of the Theory of Operator Algebras, Vol. I. Elementary theory. Pure and Applied Mathematics, 100. Aca- demic Press, Inc. [Harcourt Brace Jovanovich, Publishers], New York, 1983. [KR 2] R.V. Kadison and J.R. Ringrose,Fundamentals of the Theory of Operator Algebras, Volume II. Advanced theory. Pure and Applied Mathematics, 100. Academic Press, Inc., Orlando, FL, 1986. [P] V.I. Paulsen, Completely Bounded Maps and Operator Algebras, Cambridge Studies in Advanced Mathematics 78, Cambridge University Press, Cambridge, 2002. [R] C.E. Rickart, General Theory of Banach Algebras The University Series in Higher Mathematics D. van Nostrand Co., Inc., Princeton, N.J.-Toronto- London-New York, 1960. 8 H. NAJAFI, M.S. MOSLEHIAN Department of Pure Mathematics, Center of Excellence in Analysis on Alge- braic Structures (CEAAS), Ferdowsi University of Mashhad, P. O. Box 1159, Mash- had 91775, Iran E-mail address: [email protected] E-mail address: [email protected] and [email protected]
1110.4026	2	1110	2012-10-15T20:54:17	Aperiodicity Conditions in Topological $k$-Graphs	[ "math.OA" ]	We give two new conditions on topological $k$-graphs that are equivalent to the Yeend's aperiodicity Condition (A). Each of the new conditions concerns finite paths rather than infinite. We use a specific example, resulting from a new construction of a twisted topological $k$-graph, to demonstrate the improvements achieved by the new conditions. Reducing this proof of equivalence to the discrete case also gives a new direct proof of the corresponding conditions in discrete $k$-graphs, where previous proofs depended on simplicity of the corresponding C$^*$-algebra.	math.OA	math	APERIODICITY CONDITIONS IN TOPOLOGICAL k-GRAPHS SARAH WRIGHT Abstract. We give two new conditions on topological k-graphs that are equivalent to the Yeend's aperiodicity Condition (A). Each of the new conditions concerns finite paths rather than infinite. We use a specific example, resulting from a new construction of a twisted topological k- graph, to demonstrate the improvements achieved by the new conditions. Reducing this proof of equivalence to the discrete case also gives a new direct proof of the corresponding conditions in discrete k-graphs, where previous proofs depended on simplicity of the corresponding C∗-algebra. 1. Introduction The theory of graph C∗-algebras began with the work of Cuntz and Kreiger [1] and the later work of Enomoto and Watatani [2]. Since then, there have been many contributions by a variety of researchers resulting in an extensive collection of literature. The main idea is to associate to each directed graph a C∗-algebra and use the combinatorics of the graph to an- swer questions about the structure of the C∗-algebra. The study of graph algebras had been particularly fruitful in its provision of a rich class of easily accessible examples of C∗-algebras with various properties. To this end, the idea of a graph has been generalized in a few ways including the k-graphs of Kumjian and Pask [6] and the topological graphs of Katsura [5]. Most recently, the work of Yeend provides a generalization of both higher-rank graphs and topological graphs, with the unifying theory of topological k-graphs [13]. An important outcome of the study of each type of graph is the rela- tionship between the periodic paths (or lack there of) in the graph and the simplicity of the associated C∗-algebra. The first such aperiodicity condition for directed graphs appears in [7], there referred to as Condition (L), and states that every cycle of the graph must have an entry. This is one of the necessary conditions for simplicity and that relationship demonstrates the beauty of the subject; it is easy to look at picture of a directed graph and check if each cycle has an entry. The key to cycles having or not having entries is the idea of being able to "back out of a cycle" and build an aperiodic infinite path. As the infinite 2000 Mathematics Subject Classification. Primary 46L05; Secondary 22A22, 05C99. Key words and phrases. topological k-graph, higher-rank graph, graph algebra, aperiodicity. 1 2 SARAH WRIGHT paths play an important role in the groupoid construction of the C∗-algebra, in the case of higher-rank graphs the original aperiodicity condition, called Condition (A), was stated in terms of infinite paths. Lewin, Robertson, and Sims have developed conditions on higher-rank graphs, which are equivalent to aperiodicity [8, 10], that involve only finite paths. The work of this paper gives extensions of these aperiodicity conditions to the case of topological k-graphs and proves their equivalence. In the first section we'll discuss the background on topological k-graphs that is needed. In the middle section we state the two new aperiodicity con- ditions and prove the main result of equivalence with Yeend's Condition (A). We also address a new proof of the equivalence of the corresponding discrete conditions. The final section gives a new method of constructing a topolog- ical k-graph from a discrete k-graph. The construction of these twisted topological k-graphs shares the flavor of Yeend's skew product graphs [13, Definition 8.1] as well as the topological dynamical systems defined by Far- thing, Patani, and Willis, [4]. An advantage over the skew product graphs is the fact that we need not begin with a topological k-graph, but can use a discrete graph with desired properties. Also, we need only a suitable topo- logical space and a continuous functor rather than the k commuting local homeomorphisms {Ti} of topological dynamical systems. We give a specific example of this construction and use one of the new conditions to show the topological k-graph is aperiodic, demonstrating the improvements gained by considering finite paths. Acknowledgements: I would like to give special thanks to Aidan Sims for many helpful conversations, ideas, and motivation. I would also like to thank an anonymous referee for pointing out the errors in the original statement of the main theorem as well as many insightful comments. 2. Background The basics (and considerably more) on topological k-graphs and their C∗-algebras can be found in Yeend's original papers [12, 13, 14]. We review some of the basics here for convenience. We will consider N to contain 0 and regard Nk as the category with a single object and composition given by addition. We use {ei}k i=1 to represent the standard basis of Nk and for m ∈ Nk denote the ith component by mi. For m, n ∈ Nk, we say m ≥ n if mi ≥ ni for every i ∈ {1, 2, . . . , k}, write m ∨ n for the coordinate maximum, and m ∧ n for the coordinate minimum. For a natural number k, a topological k-graph is a pair (Λ, d) consisting of a small category Λ = (Obj(Λ), Mor(Λ), r, s) and a functor d : Λ → Nk which satisfy the following: (1) Obj(Λ) and Mor(Λ) are second-countable, locally compact, Haus- dorff spaces; (2) r, s : Mor(Λ) → Obj(Λ) are continuous and s is a local homeomor- phism; APERIODICITY CONDITIONS IN TOPOLOGICAL k-GRAPHS 3 (3) Composition ◦ : Λ ×c Λ → Λ is continuous and open, where Λ ×c Λ has the relative topology inherited from the product topology on Λ × Λ; (4) d : Λ → Nk is continuous, where Nk has the discrete topology; and (5) For all λ ∈ Λ and all m, n ∈ Nk such that d(λ) = m + n, there exist unique µ, ν ∈ Λ such that d(µ) = m, d(ν) = n, and λ = µν. Condition (5) is called the unique factorization property and is exactly the same as that for discrete k-graphs. The functor d is called the degree map, or the shape map. For n ∈ Nk, we denote by Λn the open set d−1(n) in Mor(Λ), and refer to its elements as paths of shape n. It is often necessary to deal with the range and source maps as they relate to paths of a specific shape, so the notation rn refers to the restriction of the range map r to the set Λn and similarly sn := sΛn. A consequence of the unique factorization property is that Λ0 is the set of identity morphisms of Λ, and these are referred to as the vertices of Λ. For any sets X, Y ⊂ Λ we use XY for the set {µν µ ∈ X, ν ∈ Y, s(µ) = r(ν)}. This notation is of particular use when the first set is a set of vertices and the second set is all the paths. So, for V ⊂ Λ0, V Λ = {λ ∈ Λ r(λ) ∈ V } and ΛV = {λ ∈ Λ s(λ) ∈ V }. A vertex v is a source in Λ if for some n ∈ Nk vΛn is empty. The factorization property implies that for λ ∈ Λ and m < n < d(λ) ∈ Nk there are unique paths denoted λ(0, m) ∈ Λm, λ(m, n) ∈ Λn−m, and λ (n, d(λ)) ∈ Λd(λ)−n such that λ(0, m)λ(m, n)λ(n, d(λ)) = λ. We think of λ(p, q) as the portion of the path λ that runs from q to p. An important class of examples of k-graphs are the grid graphs, Ωk,m. For a fixed k ≥ 1 and m ∈ (N ∪ {∞})k the discrete k-graph Ωk,m has morphisms {(p, q) ∈ Nk × Nk p ≤ q ≤ m}, range and source maps given by r(p, q) = (p, p) and s(p, q) = (q, q), composition defined as (p, q)(q, r) = (p, r) and degree map given by d(p, q) = q − p. We visualize a discrete k-graph by its 1-skeleton, a directed graph whose vertices are Λ0 and whose edges are paths from the sets Λei and are colored k different colors depending on the shape. Frequently dashed or dotted arrows are also (or instead) used to depict edges of different shape. For consistency, in this paper green edges will be solid and of shape e1, blue edges dotted and of shape e2, and red edges dashed and of shape e3. In some cases, such as any Ωk,m, we can construct the entire k-graph from the 1-skeleton. In other cases we need more information. Consider the path bd in Figure 1(b) of shape (0, 1, 1). By the unique factorization property, bd must factor uniquely as a "red-blue" path. So, bd = µν with d(µ) = (0, 0, 1) and d(ν) = (0, 1, 0). It cannot be determined from the 1-skeleton alone if the unique factorization we desire is cb or cb′. We need to specify factorization rules for each bi-colored path: da = a′c bf = eb′ f a = a′e f d = df bd = cb′ ce = ec. 4 SARAH WRIGHT v e c a′ a b b′ w d f (a) The 1-Skeleton of Ω3,(3,2,1) (b) A 1-skeleton of a 3-graph requiring factorization rules Figure 1. 1-Skeletons of two different 3-graphs For topological k-graphs, (Λ1, d1) and (Λ2, d2), a topological k-graph mor- phism is a continuous degree preserving functor x : Λ1 → Λ2. That is, d2(x(λ)) = d1(λ). We are particularly concerned with the graph morphisms from the grid graphs. The path space of a topological k-graph Λ XΛ := [ m∈(N∪{∞})k {x : Ωk,m → Λ x is a graph morphism} can be thought to include Λ since any finite path λ uniquely determines a graph morphism xλ : Ωk,d(λ) → Λ such that x(m, n) = λ(m, n) for any m ≤ n ≤ d(λ). We extend the idea of range and degree so that r(x) = x(0, 0) and d(x) = m for x : Ωk,m → Λ. In practice, we think of taking a particular grid graph and labeling the vertices and edges while following the structure of and factorization rules associated to the 1-skeleton of Λ. So, a path of shape (∞, ∞, ∞) with range v in the 3-graph of Figure 1(b) with the factorization rules given could look like that of Figure 2. . . . e . . . v e a . . . f w f b′ v e . . . e c . . . v a d w b′ c v c . . . v a′ d w w c b v . . . e . . . v e a′ . . . f w f b v e . . . e c . . . v a′ d w c b v . . . e . . . v . . . f a . . . . . . . . . . . v c . . . . v a′ . . . . . . w d . . . w b . . . b′ . . . v c . . . v e a w f b′ v e . . . e Figure 2. An Infinite Path in a 3-Graph In order to define aperiodic paths in a topological k-graph we need the idea of the shift map. The shift map σm on XΛ assigns to x the unique path σmx in XΛ which satisfies d(σmx) = d(x) − m and σmx(0, n) = x(m, m + n) for all 0 6 n 6 d(x) − m. For m ∈ Nk, the shift map of degree m removes a APERIODICITY CONDITIONS IN TOPOLOGICAL k-GRAPHS 5 path of shape m from the range end of x. We say a path x ∈ XΛ is aperiodic if for any m, n ∈ Nk σmx = σnx implies that m = n. The path of Figure 2 is periodic, as σ(2,0,1)x = x. The idea of a minimal common extension of two paths will be important in one of our formulations of aperiodicity and it is also necessary in our definition of a boundary path. For paths µ, ν ∈ Λ, we say λ is a common extension of µ and ν if we can factor λ = µµ′ and λ = νν′ for some µ′, ν′ ∈ Λ. We consider λ to be a minimal common extension if it also satisfies d(λ) = d(µ) ∨ d(ν). We denote by MCE(µ, ν) the set of minimal common extensions of µ and ν and for subsets X, Y ⊂ Λ define MCE(X, Y ) := [ µ∈X,ν∈Y MCE(µ, ν). A topological k-graph Λ is compactly aligned if whenever X, Y ⊂ Λ are compact, then MCE(X, Y ) is also compact. For a vertex v ∈ Λ0, we say a subset E ⊂ Λ is compact exhaustive for v if E is compact, r(E) is a neighborhood of v, and for all λ ∈ r(E)Λ there exists a µ ∈ E such that MCE(λ, µ) 6= ∅. A path x ∈ XΛ is a boundary path if for any m ∈ Nk with m ≤ d(x) and any set E which is compact exhaustive set for x(m, m), there exists a λ ∈ E such that x(m, m + d(λ)) = λ. We write ∂Λ for the set of all boundary paths in XΛ. As in discrete k-graphs the boundary paths are a generalization of the infinite paths of directed graphs, [11, Lemma 4.22]. Under the condition that Λ is compactly aligned, we can build a topo- logical groupoid GΛ called the boundary path groupoid which has ∂Λ as its unit space and can be endowed with a locally compact Hausdorff topology. The C∗-algebra of Λ, C ∗(Λ), is the full groupoid C∗-algebra of the boundary path groupoid, C ∗(GΛ). As we will not need the details of the construction of the C∗-algebra to understand the results of this paper, we refer the reader to [13] for the formulation of this topology as well as a more in depth discus- sion. However, to demonstrate the importance of aperiodicity, we conclude the background information with a result of Yeend. Proposition 2.1. [13, Theorem 5.2] For a compactly aligned topological k- graph (Λ, d), the boundary path groupoid GΛ is topologically principal if and only if for every open set V ⊂ Λ0 there exists an x ∈ V ∂Λ which is aperiodic. See [9], for structure and simplicity arguments relying on the groupoid being topologically principal and amenable. The authors use groupoid tech- niques to prove the equivalence of (A) and (C) of the main theorem of this paper. Theorem 3.1 remains the cleaner and more accessible proof of the equivalence of the two notions of aperiodicity 3. Results The condition that every open set of vertices must have an aperiodic path that terminates in that set is the natural extension of aperiodicity conditions 6 SARAH WRIGHT for topological graphs and discrete k-graphs. Yeend introduced the condition and referred to it as Condition (A). Our main result, Theorem 3.1, gives two new equivalent conditions to Yeend's Condition (A). Theorem 3.1. Let (Λ, d) be a compactly aligned topological k-graph with no sources. Then the following conditions are equivalent. (A) For any open set V ⊂ Λ0 there exists an aperiodic path x ∈ V ∂Λ. (B) For any open set V ⊂ Λ0 and any pair m 6= n ∈ Nk there exists a path λV,m,n ∈ V Λ such that d(λ) ≥ m ∨ n and (⋆) λ(m, m + d(λ) − (m ∨ n)) 6= λ(n, n + d(λ) − (m ∨ n)). (C) For any distinct pair of open sets X, Y ⊂ Λ such that s(X) = s(Y ) and both sX and sY are homeomorphisms there exists τ ∈ s(X)Λ such that MCE(Xτ, Y τ ) = ∅. It is not difficult to see that in the discrete case, (if Obj(Λ) and Mor(Λ) both have the discrete topology) then Conditions (A) -- (C) above reduce to [6, Definition 4.3], Condition (iv) of [10, Lemma 3.2], and [8, Definition 3.1] respectively. This special case gives the following corollary. The equiv- alence that appears in Corollary 3.2 is already known, as each condition is a necessary condition for simplicity of the associated C∗-algebra. The direct proof of this equivalence that follows from the proof of Theorem 3.1 is new, and an important outcome of this work. Corollary 3.2. Let (Λ, d) be a discrete finitely aligned k-graph with no sources. Then the following are equivalent. (i) For each v ∈ Λ0 there exists and aperiodic path x ∈ v∂Λ. (ii) For each v ∈ Λ0 and each m 6= n ∈ Nk there exists a path λ ∈ vΛ such that d(λ) ≥ m ∨ n and which satisfies (⋆). (iii) For every pair of distinct paths α, β ∈ Λ with s(α) = s(β) there exists a path τ ∈ s(α)Λ such that MCE(ατ, βτ ) = ∅. The proof of the theorem requires an important lemma. Lemma 3.3 is an extension of Lemma 5.6 of [5] which shows that near a path which satisfies (⋆) are a lot of other paths which also satisfy (⋆). This allows us to extend the proof of the equivalence of (i) and (iv) in Lemma 3.2 of [10] to topological k-graphs. Lemma 3.3. Let V be a nonempty open subset of Λ0, m 6= n ∈ Nk, and λ ∈ V Λ. If λ satisfies (⋆) for m and n, then there exists a compact neighborhood E ⊂ V Λd(λ)of λ such that every µ ∈ E satisfies (⋆). Proof. Suppose λ ∈ V Λ satisfies (⋆). We can factor λ in two ways: λ = λ(0, m)λ(m, m + d(λ) − (m ∨ n))λ(m + d(λ) − (m ∨ n), d(λ)) and λ = λ(0, n)λ(n, n + d(λ) − (m ∨ n))λ(n + d(λ) − (m ∨ n)), d(λ)) Let Em and En be disjoint compact neighborhoods of λ(m, m + d(λ) − (m ∨ n)) and λ(n, n + d(λ) − (m ∨ n)). Also, let E1, E2, E3, and E4 be APERIODICITY CONDITIONS IN TOPOLOGICAL k-GRAPHS 7 compact neighborhoods of λ(0, m), λ(m + d(λ) − (m ∨ n), d(λ)), λ(0, n), and λ(n + d(λ) − (m ∨ n), d(λ)) respectively. We define the set E′ to be the paths in Λd(λ) that can be factored as paths in E1, Em and E2 as well as E3, En, and E4. So, E′ := {αβγ : (α, β, γ) ∈ E1 ×c Em ×c E2} ∩ {µνξ : (µ, ν, ξ) ∈ E3 ×c En ×c E4} ⊂ Λd(λ). Now, let F ⊂ V be a compact neighborhood of r(λ) and E = E′ ∩ r−1 d(λ)(F ). Then E ⊂ V Λd(λ) is a compact neighborhood of λ with nonempty interior in which every element satisfies (⋆). (cid:3) Proof of Theorem 3.1. We will show Condition (B) is equivalent to each of (A) and (C). (A) =⇒ (B). Fix and open set V ⊂ Λ0, and a pair m 6= n ∈ Nk, and suppose x ∈ V ∂Λ is aperiodic. Then, σmx 6= σnx and so for a sufficiently large p ∈ Nk Let λ = x(0, p + m ∨ n), then σmx(0, p) 6= σnx(0, p). λ(m, m + d(λ) − m ∨ n) = x(m, m + (p + m ∨ n) − m ∨ n) = σmx(0, p) 6= σnx(0, p) = x(n, n + p + m ∨ n − m ∨ n) = λ(n, n + d(λ) − m ∨ n). So, d(λ) ≥ m ∨ n and λ satisfies (⋆). (B) =⇒ (A). Fix an open set V ⊆ Λ0 and let {(mi, ni)}∞ i=1 be a listing of the elements in the set (cid:8)(m, n) ∈ Nk × Nk : m 6= n(cid:9). Let V1 = V , choose λ1 ∈ V1Λ such that d(λ1) ≥ m1 ∨ n1 and (⋆) is satisfied for m1 and n1; choose a compact neighborhood F1 of λ1 by Lemma 3.3; and let E1 := F1∂Λ = {x ∈ ∂Λ : x (0, d(λ1)) ∈ F1}. Now, proceed inductively: Vi := interior of s(Fi−1), λi := λVi,mi,ni satisfy (⋆) for mi and ni, Fi := a compact neighborhood of λi given by Lemma 3.3, and Ei := F1 . . . Fi∂Λ. Notice that each Ei is compact in ∂Λ by Lemma 3.8 of [12]: Let p = q = Pi j=1 d(λj ). Then Ei is exactly the set Z(U ∗s V, p − q), which is compact in ∂Λ. Also note that the Ei's are j=1 d(λj ) and U = V = F1 . . . Fi ⊂ ΛPi 8 SARAH WRIGHT nested; E1 ⊃ E2 ⊃ . . . . Now, let E := ∞ \ i=1 Ei ⊂ ∂Λ. The set E is non-empty, so take any x ∈ E. Let µi = x (d(λ1λ2 . . . λi−1), d(λi)) for each i ∈ N. Then by Lemma 4.1.3 of [14], x is the unique element in XΛ such that d(x) = limj→∞ d(µ1µ2 . . . µj) and x(0, d(µ1µ2 . . . µj)) = µ1µ2 . . . µj, so it makes sense to write with each µi ∈ Fi. Fix m 6= n ∈ Nk. Then for some i ∈ N, (m, n) = (mi, ni) in the listing above. So, x = µ1µ2µ3 . . . , d(λj) − (m ∨ n) i−1 σmx X  j=1 d(λj), i X j=1 Similarly, j=1 d(λj )x(0, d(λi) − (m ∨ n)) = σm+Pi−1 = σPi−1 = µi(m, m + d(µi) − (m ∨ n)). j=1 d(λj )x(m, m + d(λi)−(m ∨ n)) i−1 σnx X  j=1 d(λj), i X j=1 d(λj) − (m ∨ n)  = µi(n, n + d(µi) − (m ∨ n)). By the definition of Fi, we know that µi(m, m + d(µi) − (m ∨ n)) 6= µi(n, n + d(µi) − (m ∨ n)), and thus σmx 6= σnx and x is an aperiodic boundary path. (B) =⇒ (C). Fix distinct open sets X, Y ⊂ Λ such that s(X) = s(Y ) and sX and sY are both homeomorphisms. Similar to [8, Remark 3.2], we may assume r(X) = r(Y ) and further that for µ ∈ X and ν ∈ Y if s(µ) = s(ν), then r(µ) = r(ν). To the contrary, suppose there exist µ ∈ X and ν ∈ Y with s(µ) = s(ν) = v but r(µ) 6= r(ν). Then MCE(Xv, Y v) = MCE(µ, ν) = ∅. Also, we may assume that there exist m 6= n ∈ Nk such that s(X m) ∩ s(Y n) 6= ∅. If this were not the case, then for any µ ∈ X the unique ν ∈ Y s(µ) must also have degree m. Thus, either MCE(Xs(µ), Y s(µ)) = MCE(µ, ν) = ∅ or µ = ν. Since X 6= Y , there must exist a pair with no minimal common extensions. Now, fix m 6= n ∈ Nk such that s(X m) ∩ s(Y n) 6= ∅ as above. Let V be the nonempty open set s(X m) ∩ s(Y n). Choose λV,m,n by (B). Suppose that α ∈ MCE(Xλ, Y λ). Then, α = µλµ′ = νλν′ where µ ∈ X m, ν ∈ Y n, and µ′, ν′ ∈ s(λ)Λ. Notice that λ(m, m + d(λ) − (m ∨ n)) = α(m + n, m + n + d(λ) − (m ∨ n)) = λ(n, n + d(λ) − (m ∨ n)), APERIODICITY CONDITIONS IN TOPOLOGICAL k-GRAPHS 9 but this contradicts the choice of λ. Thus, MCE(Xλ, Y λ) = ∅. (C) =⇒ (B). Fix an open set V ⊂ Λ0 and a pair m 6= n ∈ Nk. Notice r−1 m∨n(V ) ⊂ Λm∨n is open and let U ⊂ r−1 m∨n(V ) such that sU is a homeomor- phism. Define X := {µ(m, m ∨ n) µ ∈ U } and Y := {µ(n, m ∨ n) µ ∈ U }. Using a restriction of the continuous and open composition map, it can be shown that X and Y are both open. As s(X) = s(U ) = s(Y ) and both sX and sY are homeomorphisms, by (C) choose τ ∈ s(X)Λ such that MCE(Xτ, Y τ ) = ∅. Since sU is a homeomorphism, there is a unique µ ∈ U such that s(µ) = r(τ ). Let ξ = µ(m, m ∨ n) ∈ X and υ = µ(n, m ∨ n) ∈ Y and notice Xτ = {ξτ } and Y τ = {υτ }. Let ω ∈ s(τ )Λm∨n and define λ = µτ ω. Now, λ(m, m + d(λ) − (m ∨ n)) = (µτ ω)(m, m + d(τ ) + (m ∨ n)) = ξ(τ ω)(0, d(τ ) + m) = ξτ ω(0, m) Similarly, λ(n, n + d(λ) − (m ∨ n)) = υτ ω(0, n). If ξτ ω(0, m) = υτ ω(0, n), then λ ∈ MCE(ξτ, υτ ) = MCE(Xτ, Y τ ), a con- tradiction. (cid:3) To get a better intuition for Conditions (B) and (C), it is helpful to consider the special case of a directed graph. If a directed graph E satisfies Condition (B), then for any two natural numbers m 6= n we can find a path in E such that the mth and nth edges are different. There is also an enlightening diagram of Condition (B) in the discrete case in Appendix A of [10]. Consider two paths µ and ν in a directed graph E where s(µ) = s(ν), r(µ) = r(ν) and d(µ)∧d(ν) = 0 as described in [8, Remark 3.2]. This implies that one path, say µ, is just a vertex and ν must be a loop based at that vertex. Condition (C) then says that the loop µ must have an entry. 4. Examples To demonstrate the importance of these formulations of aperiodicity we give an example of a topological k-graph in which Condition (C) is straight- forward to verify whereas Condition (A) would be quite difficult. First we define what we call a twisted product topological k-graph. This twisting con- struction takes a discrete k-graph and puts a copy of an appropriate topo- logical space X at each vertex, and twists the edges according to a functor τ . Proposition 4.1. Let (Λ, d) be a finitely aligned k-graph with no sources, X be a second countable, locally compact, Hausdorff space, and τ : Λ → {φ : X → X φ is a local homeomorphism} 10 SARAH WRIGHT a continuous functor. Then the pair (cid:0)Λ ×τ X, d(cid:1), with object and morphism sets Obj(Λ ×τ X) := Obj(Λ) × X and Mor(Λ ×τ X) := Mor(Λ) × X, range and source maps r(λ, x) := (r(λ), τλ(x)) and s(λ, x) := (s(λ), x) , and composition (λ, τµ(x)) ◦ (µ, x) = (λµ, x), whenever s(λ) = r(µ) in (Λ, d), and degree functor d(λ, x) = d(λ) is a topological k-graph. The proof that (cid:0)Λ ×τ X, d(cid:1) satisfies all the requirements of a topological k-graph is straightforward and the details unenlightening, so we omit the proof here. Consider the discrete 2-graph of Evans and Sims [3] whose 1-skeleton is pictured in Figure 3. α1 0 β1 0 v1 v0 α2 1 α2 0 β2 0 β2 1 v2 α3 2 α3 1 α3 0 β3 0 β3 1 β3 2 · · · vn v3 n-1 αn ... αn 1 αn 0 βn 0 βn ... 1 βn n-1 · · · vn+1 Figure 3. The 1-skeleton of a discrete k-graph Λ. with factorization rules given by i βn+1 αn j = βn ξn(i,j)αn+1 ζn(i,j) where (i, j) 7→ (ξn(i, j), ζn(i, j)) is a permutation of Z/nZ × Z/(n + 1)Z. For an example of a twisted product we take the 1-skeleton of Figure 3, with ξn and ζn being "plus 1" in the group Z/nZ, the topological space T (z) := zn. with the usual topology, and the functor τ given by ταn We show the twisted topological k-graph above satisfies Condition (C). We start with the appropriate open sets X, Y ∈ Λ ×τ T. As the topology on Λ is discrete, we may assume that X and Y have the form X = {µ} × U and Y = {ν}×V , for paths µ, ν ∈ Λ and open sets U, V ⊂ T. Since s(X) = s(Y ), it must be true that s(µ) = s(ν) and U = V . As in [8, Remark 3.2], we (z) = τβn j i APERIODICITY CONDITIONS IN TOPOLOGICAL k-GRAPHS 11 may assume r(µ) = r(ν) and d(µ) ∧ d(ν) = 0. We assume µ consists of only green edges and ν only blue. We can see that µ must have the form , zm+n(cid:17)(cid:16)αm+n i0 , zk(cid:17) . . . (cid:16)αm+n−2 ,(cid:0)zm+n(cid:1)m+n−1 (cid:17)(cid:16)αm+n−1 µ = (cid:16)αm , z(cid:17) in−2 in−1 in m+n where z ∈ U and k is given by k = i . Similarly, Y i=m+1 ν = (cid:16)βm j0 , zk(cid:17) . . . (cid:16)βm+n jn , z(cid:17) , z1/m+n+1(cid:17). Notice the path ντ is built We consider the path τ = (cid:16)βm+n+1 of all blue edges, so it cannot be factored to appear as a different path. We follow the factorization rules to rewrite the path µτ such that the first edge is a blue edge, and then compare to ντ . j0−n i1 in i0 , zk(cid:17)(cid:16)αm+1 j0+1, zk(cid:17)(cid:16)αm+1 j0 , zk(cid:17) 6= (cid:16)βm , zk/m+n(cid:17) . . .(cid:16)αm+n µτ = (cid:16)αm i0+1, zk/m+n(cid:17) . . .(cid:16)αm+n = (cid:16)βm j0+1, zk(cid:17) and the paths µτ and ντ will have no If m 6= 1, (cid:16)βm common extensions. If m = 1, then we'd amend a similar path of shape (0, 2) so that µτ and ντ would be guaranteed to differ in the second blue edge by a similar calculation. Thus, MCE(Xτ, Y τ ) = ∅ and Λ ×τ T is aperiodic. , z(cid:17)(cid:16)βm+n+1 in+1, z1/m+n+1(cid:17) , z1/m+n+1(cid:17) j0−n In contrast, if we were to check for aperiodicity using Condition (A) we 0 0, z(cid:1)(cid:16)βi+1 , z1/i+1(cid:17)(cid:16)αi+2 would need to consider infinite paths of the form , z1/(i+1)(i+2)(i+3)(cid:17) . . . x = (cid:0)αi and investigate σmx for any m ∈ Nk. A precise calculation requires a variety of formulas for σmx to cover different cases and then checking the various factorizations of the result for periodic behavior. It is at least three long and somewhat complicated calculations. , z1/(i+1)(i+2)(cid:17)(cid:16)βi+3 0 0 It is worth noting that this work covers only topological k-graphs with- out sources. This was a common hypothesis in the early works on direct graphs, topological graphs, and k-graphs that was eventually replaced with less restrictive hypotheses. It is the hope of the author to develop similar conditions and results for topological k-graphs with sources. References 1. Joachim Cuntz and Wolfgang Krieger, A class of C ∗-algebras and topological Markov chains, Invent. Math. 56 (1980), no. 3, 251 -- 268. MR MR561974 (82f:46073a) 2. Masatoshi Enomoto and Yasuo Watatani, A graph theory for C ∗-algebras, Math. Japon. 25 (1980), no. 4, 435 -- 442. MR MR594544 (83d:46069a) 3. D. Gwion Evans and Aidan Sims, When is the cuntz-krieger algebra of a higher-rank graph approximately finite-dimensional?, preprint, arXiv:1112.4549v1. 4. Cynthia Farthing, Nura Patani, and Paulette Willis, The crossed-product structure of C∗-algebras arising from topological dynamical systems, to appear. 12 SARAH WRIGHT 5. Takeshi Katsura, A class of C ∗-algebras generalizing both graph algebras and homeo- morphism C ∗-algebras. I. Fundamental results, Trans. Amer. Math. Soc. 356 (2004), no. 11, 4287 -- 4322 (electronic). 6. Alex Kumjian and David Pask, Higher rank graph C ∗-algebras, New York J. Math. 6 (2000), 1 -- 20 (electronic). MR MR1745529 (2001b:46102) 7. Alex Kumjian, David Pask, Iain Raeburn, and Jean Renault, Graphs, groupoids, and Cuntz-Krieger algebras, J. Funct. Anal. 144 (1997), no. 2, 505 -- 541. MR MR1432596 (98g:46083) 8. Peter Lewin and Aidan Sims, Aperiodicity and cofinality for finitely aligned higher-rank graphs, Math. Proc. Cambridge Philos. Soc. 149 (2010), no. 2, 333 -- 350. MR 2670219 9. Jean Ranault, Aidan Sims, Dana P. Williams, and Trent Yeend, Uniqueness theorems for topological higher-rank graph C ∗-algebras, preprint. 10. David I. Robertson and Aidan Sims, Simplicity of C ∗-algebras associated to higher- rank graphs, Bull. Lond. Math. Soc. 39 (2007), no. 2, 337 -- 344. MR MR2323468 (2008g:46099) 11. Sarah Wright, Aperiodicity in topological k-graphs, Ph.D. thesis, May 2010. 12. Trent Yeend, Topological higher-rank graphs and the C ∗-algebras of topological 1- graphs, Operator theory, operator algebras, and applications, Contemp. Math., vol. 414, Amer. Math. Soc., Providence, RI, 2006, pp. 231 -- 244. MR MR2277214 (2007j:46100) 13. 14. , Groupoid models for the C ∗-algebras of topological higher-rank graphs, J. Op- erator Theory 57 (2007), no. 1, 95 -- 120. MR MR2301938 (2008f:46074) , Topological higher-rank graphs, their groupoids and operator algebras, Ph.D. thesis, March 2005. College of the Holy Cross, Department of Mathematics and Computer Science, One College Street, Worcester, MA, 01610 E-mail address: [email protected]
1109.2962	1	1109	2011-09-14T01:26:52	C$^{*}$-bialgebra defined as the direct sum of UHF algebras	[ "math.OA", "math.QA" ]	Let ${\cal A}_{0}()$ denote the direct sum of a certain set of UHF algebras and let ${\cal A}()\equiv {\bf C}\oplus {\cal A}_{0}()$. We introduce a non-cocommutative comultiplication $\Delta_{\phi}$ on ${\cal A}()$, and give an example of comodule-C$^{}$-algebra of the C$^{}$-bialgebra $({\cal A}(),\Delta_{\phi})$. With respect to $\Delta_{\phi}$, we define a non-symmetric tensor product of -representations of UHF algebras and show tensor product formulas of GNS representations by product states.	math.OA	math	C∗-bialgebra defined as the direct sum of UHF algebras Katsunori Kawamura∗ College of Science and Engineering, Ritsumeikan University, 1-1-1 Noji Higashi, Kusatsu, Shiga 525-8577, Japan Abstract Let A0(∗) denote the direct sum of a certain set of UHF alge- bras and let A(∗) ≡ C ⊕ A0(∗). We introduce a non-cocommutative comultiplication ∆ϕ on A(∗), and give an example of comodule-C∗- algebra of the C∗-bialgebra (A(∗), ∆ϕ). With respect to ∆ϕ, we define a non-symmetric tensor product of ∗-representations of UHF algebras and show tensor product formulas of GNS representations by product states. Mathematics Subject Classifications (2010). 16T10; 46K10 Key words. C∗-bialgebra; comodule-C∗-algebra; UHF algebra; tensor product; Kronecker product. 1 Introduction A C∗-bialgebra is a generalization of bialgebra in the theory of C∗-algebras, which was introduced in C∗-algebraic framework for quantum groups [26, 27]. We have studied C∗-bialgebras and their construction method, and computed non-symmetric tensor products of ∗-representations with respect to non-cocommutative comultiplications [17, 18, 20, 21, 24, 22, 25]. In this paper, we introduce a non-cocommutative C∗-bialgebra defined as the di- rect sum of a certain set of uniformly hyperfinite (=UHF) algebras. With respect to the comultiplication, we define a non-symmetric tensor product of ∗-representations of UHF algebras, and show tensor product formulas of Gel'fand-Naımark-Segal (=GNS) representations by product states. The ∗e-mail: [email protected]. 1 part of tensor product formulas has been given in the previous paper [23] without C∗-bialgebra. The present version is reorganized such that the ten- sor product is given by the comultiplication of a C∗-bialgebra. In this sec- tion, we show our motivation, definitions and construct the C∗-bialgebra. The main feature of this paper is as follows: • A new non-commutative and non-cocommutative C∗-bialgebra is ob- tained. In our previous research, we treated only C∗-bialgebras (A, ∆) which satisfy ∆(A) ⊂ A ⊗ A. In this paper, this property does not holds. The bialgebra structure does not appear unless one takes the direct sum of all UHF algebras. Until now, there is no theory which treat all UHF algebras at once. • The C∗-bialgebra is naturally constructed by using a well-known struc- ture of UHF algebras. The standard parametrization of GNS repre- sentations of UHF algebras by product states is compatible with the tensor product formulas. • Tensor product formulas of non-type I representations are obtained for the first time except [22]. • A construction method of C∗-bialgebra is a little bit generalized. 1.1 Motivation In this subsection, we explain our motivation and the background of this study. Explicit mathematical definitions will be shown after § 1.2. According to [12, 14, 30], given two representations of a group G, their tensor product (or Kronecker product [30]) is a new representation of G, which decomposes into a direct sum of indecomposable representations. The problem of finding this decomposition is called the Clebsch-Gordan problem and the resulting formula for the decomposition is called the tensor product formula (or Clebsch-Gordan formula [14]). Furthermore, the tensor prod- uct is important to describe the duality of G [33]. A generalization of the Clebsch-Gordan problem for groups is to consider modules over associative algebras instead of group algebras. However, there lies an obvious obstruc- tion in that there is no known way to define the tensor product of two left modules over an arbitrary associative algebra. For group algebras, the extra structure coming from the group yields the tensor product. For a bialgebra A, the associative tensor product of representations of A can be defined by using the comultiplication. In this way, one of most important motivations of the study of bialgebras is the tensor product of their representations. 2 In [17], we introduced a non-symmetric tensor product among all ∗- representations of Cuntz algebras and determined tensor product formulas of all permutative representations completely, in spite of the unknown of any comultiplication of Cuntz algebras. In [18], we generalized this construction of tensor product to a system of C∗-algebras and ∗-homomorphisms indexed by a monoid. For example, we constructed a non-symmetric tensor prod- uct of all ∗-representations of Cuntz-Krieger algebras by using Kronecker products of matrices [19, 20, 22]. On the other hand, UHF algebras and their ∗-representations are well studied [2, 3, 4, 5, 8, 13, 29, 28]. For example, GNS representations of product states of UHF algebras were completely classified by [5]. This class contains ∗-representations of UHF algebras of all Murray-von Neumann's types I, II, III ([6], § III.5). Our interests are to construct a C∗-bialgebra from UHF algebras, and to define a tensor product of ∗-representations of UHF algebras with respect to the comultiplication. Since one knows neither cocommutative nor non- cocommutative comultiplication of UHF algebras, the tensor product is new if one can find it. 1.2 C∗-bialgebra In this subsection, we recall terminology about C∗-bialgebra according to [11, 26, 27]. For two C∗-algebras A and B, we write Hom(A, B) as the set of all ∗-homomorphisms from A to B, and let M(A) denote the multiplier algebra of A. We assume that every tensor product ⊗ as below means the minimal C∗-tensor product. Definition 1.1 A pair (A, ∆) is a C∗-bialgebra if A is a C∗-algebra and ∆ ∈ Hom(A, M(A ⊗ A)) such that the linear span of {∆(a)(b ⊗ c) : a, b, c ∈ A} is norm dense in A ⊗ A and (∆ ⊗ id) ◦ ∆ = (id ⊗ ∆) ◦ ∆. We call ∆ the comultiplication of A. We say that a C∗-bialgebra (A, ∆) is unital if A is unital and ∆ is unital; (A, ∆) is counital if there exists ε ∈ Hom(A, C) which satisfies (ε ⊗ id) ◦ ∆ = id = (id ⊗ ε) ◦ ∆. We call ε the counit of A and write (A, ∆, ε) as the counital C∗-bialgebra (A, ∆) with the counit ε. Remark that Definition 1.1 does not mean ∆(A) ⊂ A ⊗ A. A bialgebra in the purely algebraic theory [1, 16] means a unital counital bialgebra with the unital counit with respect to the algebraic tensor product, which does not need to have an involution. Hence a C∗-bialgebra is not a bialgebra in general. 3 According to [18], we recall several notions of C∗-bialgebra. A ∗- homomorphism f from A to M(B) is nondegenerate if f (A)B is dense in B. A pair (B, Γ) is a right comodule-C∗-algebra of a C∗-bialgebra (A, ∆) if B is a C∗-algebra and Γ is a nondegenerate ∗-homomorphism from B to M(B ⊗ A) which satisfies (Γ ⊗ id) ◦ Γ = (id ⊗ ∆) ◦ Γ where both Γ ⊗ id and id ⊗ ∆ are extended to unital ∗-homomorphisms from M(B ⊗ A) to M(B ⊗ A ⊗ A). In this case, the map Γ is called the right coaction of A on B. A C∗-bialgebra (A, ∆) is proper if ∆(a)(I ⊗ b), (b ⊗ I)∆(a) ∈ A ⊗ A for any a, b ∈ A where I denotes the unit of M(A). A proper C∗-bialgebra (A, ∆) satisfies the cancellation law if ∆(A)(I ⊗ A) and ∆(A)(A ⊗ I) are dense in A ⊗ A where ∆(A)(I ⊗ A) and ∆(A)(A ⊗ I) denote the linear spans of sets {∆(a)(I ⊗ b) : a, b ∈ A} and {∆(a)(b ⊗ I) : a, b ∈ A}, respectively. 1.3 UHF algebras and ∗-isomorphisms among their tensor products In this subsection, we recall UHF algebras [13] and introduce a set of ∗- isomorphisms among UHF algebras and their tensor products. Let N ≡ {1, 2, 3, . . .}, N≥2 ≡ {2, 3, 4, . . .} and let N∞ ≥2 denote the set of all sequences in N≥2. For n ∈ N, let Mn denote the (finite-dimensional) C∗-algebra of all n × n-complex matrices. For a = (a1, a2, . . .) ∈ N∞ ≥2, the sequence {Man }n≥1 of C∗-algebras defines the tensor product An(a) ≡ j=1 Maj . With respect to the embedding Nn ψ(n) a : An(a) ∋ A 7→ A ⊗ I ∈ An(a) ⊗ Man+1 = An+1(a), (1.1) we regard An(a) as a C∗-subalgebra of An+1(a) and let A(a) denote the inductive limit of the system {(An(a), ψ(n) a ) : n ≥ 1}: A(a) ≡ lim−→(An(a), ψ(n) a ). (1.2) By definition, A(a) is a UHF algebra of Glimm's type {a1, a1a2, a1a2a3, . . .} which was classified by [13]. On the contrary, any UHF algebra is isomorphic to A(a) for some a ∈ N∞ ≥2. Hence we call A(a) a UHF algebra in this paper. : i, j = 1, . . . , n} denote the set of standard matrix units of ≥2. Then ≥2, define the (standard) ≥2, ·) is a commutative semigroup. For a, b ∈ N∞ ≥2, let a · b ≡ (a1b1, a2b2, . . .) ∈ N∞ Mn. For a = (an), b = (bn) ∈ N∞ (N∞ ∗-isomorphism ϕ(n) Let {E(n) i,j a,b from An(a · b) onto An(a) ⊗ An(b) by ) ⊗ (E(b1) ′′ 1 ,k ⊗ · · · ⊗ E(an) j′ n,k′ ) ≡ (E(a1) ′ 1,k ⊗ · · ·⊗ E(anbn) jn,kn ′ 1 j n j ϕ(n) a,b(E(a1 b1) j1,k1 ′′ 1 ⊗ · · ·⊗ E(bn) j′′ n ,k′′ (1.3) ) n 4 ′ ′′ ′′ ′ ′′ ′′ n, k 1 , . . . , k i ∈ {1, . . . , ai}, j n are defined as ji = bi(j 1, . . . , k i−1)+k i ∈ {1, . . . , bi} for each i = 1, . . . , n. For ψ(n) for each ji, ki ∈ {1, . . . , aibi}, i = 1, . . . , n where j n, j 1 , . . . , j i , a in i , k j i, k a ⊗ ψ(n) (1.1), we see that (ψ(n) a·b for each a, b and n. From this, we can define a unique ∗-isomorphism ϕa,b from A(a · b) onto A(a) ⊗ A(b) such that n, k i and ki = bi(k a,b = ϕ(n+1) a,b ◦ ψ(n) b ) ◦ ϕ(n) i −1)+j 1, . . . , j ′′ ′′ ′′ ′ ′ ′ ′ ′ ′ ′′ (ϕa,b)An(a·b) = ϕ(n) a,b (n ≥ 1) (1.4) where we identify A(a)⊗A(b) with the inductive limit of the system {(An(a)⊗ An(b), ψ(n) a ⊗ ψ(n) b ) : n ≥ 1}. We add the unit 1 ≡ (1, 1, 1, . . .) for the semigroup N∞ ≥2 and write ≥2 ≡ N∞ ≥2 ∪ {1}, eN∞ (1.5) which is a subsemigroup of N∞. For convenience, define the 1-dimensional C∗-algebra A(1) ≡ C. (1.6) Remark that A(1) is not a UHF algebra by definition [13]. In addition, for a ∈ N∞ ≥2, define ϕ1,1, ϕ1,a and ϕa,1 by ϕ1,1 = idA(1), ϕ1,a(x) ≡ 1 ⊗ x, ϕa,1(x) ≡ x ⊗ 1, (x ∈ A(a)) (1.7) where we identify A(1) ⊗ A(1) with A(1). Remark 1.2 We consider the meaning of (1.3). For two matrices A ∈ Mn and B ∈ Mm, define the matrix A ⊠ B ∈ Mnm by (A ⊠ B)m(i−1)+i′ ,m(j−1)+j′ ≡ Ai,jBi′ ,j′ (1.8) for i, j ∈ {1, . . . , n} and i the Kronecker product of A and B [10, 32]. For a, b ∈ N∞ ′ ∈ {1, . . . , m}. The new matrix A ⊠ B is called , j ≥2, we see that ′ ϕ(1) a,b(A ⊠ B) = A ⊗ B (A ∈ Ma1 , B ∈ Mb1). (1.9) Hence ϕ(1) be called the Kronecker coproduct. a,b is the inverse operation of the Kronecker product, which should 1.4 Main theorems In this subsection, we show our main theorems. 5 1.4.1 Construction of C∗-bialgebra In this subsubsection, we construct a C∗-bialgebra. For the uncountable set ≥2} of C∗-algebras in (1.2) and (1.6), define the direct sum {A(a) : a ∈ eN∞ For a ∈ eN∞ ≥2, define and A(∗) ≡M{A(a) : a ∈ eN∞ ≥2}. ≥2 : b · c = a}, ≥2 × eN∞ Na ≡ {(b, c) ∈ eN∞ A(2)(a) ≡M{A(b) ⊗ A(c) : (b, c) ∈ Na}. (1.10) (1.11) (1.12) We see that A(∗)⊗ A(∗) =L{A(2)(a) : a ∈ eN∞ in (1.4), define the ∗-homomorphism ∆(a) ≥2}. For {ϕa,b : a, b ∈ eN∞ from A(a) to M(A(2)(a)) by ϕ ≥2} ∆(a) ϕ (x) ≡ Y(b,c)∈Na ϕb,c(x) (x ∈ A(a)) (1.13) where we identify M(A(2)(a)) with the direct product Q(b,c)∈Na A(b) ⊗ A(c). Define the ∗-homomorphism ∆ϕ from A(∗) to M(A(∗) ⊗ A(∗)) by ∆ϕ ≡M{∆(a) : a ∈ eN∞ where we also identify ⊕{M(A(2)(a)) : a ∈ eN∞ M(A(∗) ⊗ A(∗)) ∼=Q{M(A(2)(a)) : a ∈ eN∞ ≥2}. ϕ ≥2} (1.14) ≥2} with a C∗-subalgebra of Theorem 1.3 Let (A(∗), ∆ϕ) be as in (1.10) and (1.14), and let ε denote the projection from A(∗) onto A(1). Then the following holds: (i) (A(∗), ∆ϕ) is a non-cocommutative proper C∗-bialgebra with counit ε. (ii) (A(∗), ∆ϕ) satisfies the cancellation law. Remark 1.4 (i) The idea of the definition of {ϕa,b} in (1.4) is an anal- ogy of the set of embeddings of Cuntz algebras in § 1.2 of [17]. In § 6.1 of [18], we also defined a C∗-bialgebra defined as the direct sum of a countably infinite set of UHF algebras: U HF∗ ≡ C ⊕ U HF2 ⊕ U HF3 ⊕ U HF4 ⊕ · · · (1.15) 6 where U HFn is defined as the fixed point subalgebra of the Cuntz al- gebra On with respect to the U (1)-gauge action, which is ∗-isomorphic n . Clearly, U HF∗ is a C∗-subalgebra onto the inductive limit lim−→k of A(∗) in (1.10), but not a C∗-subbialgebra. The reason is as follows: The comultiplication ∆ of U HF∗ in [18] satisfies ∆(U HF∗) ⊂ U HF∗ ⊗ U HF∗. On the other hand, the restriction ∆ϕU HF∗ of the comultipli- cation ∆ϕ of A(∗) in (1.14) satisfies ∆ϕ(U HF∗) 6⊂ U HF∗ ⊗ U HF∗. M ⊗k (ii) In order to help reader's understanding, we demonstrate the image of comultiplication ∆ϕ according to the definition. Let 6 = (6, 6, . . .) ∈ ≥2. Then eN∞ N6 =(cid:26)(6, 1), (1, 6), (a, ¯a) : a = (a1, a2, . . .), ai ∈ {2, 3}, i ≥ 1 ¯a = (6/a1, 6/a2, . . .) (cid:27) . (1.16) Remark that N6 is also a uncountable set. When (b, c) ∈ N6 and b = (b1, b2, b3, . . .), b1 is 1 or 2 or 3 or 6. Recall A(6) = lim−→ An(6) and A1(6) = M6 ⊃ {E(6) i,j : i, j = 1, . . . , 6}. For x ∈ A1(6), ∆ϕ(x) = ∆(6) ϕ (x) = Y(b,c)∈N6 ϕb,c(x) = Y(b,c)∈N6 ϕ(1) b,c(x). (1.17) Since ∆ϕ(x) ∈ M(A(2)(a)) =Q(b,c)∈N6 as an element (xb,c)(b,c)∈N6 Especially, we compute the case x = E(6) E(3) 1,1 in Q(b,c)∈N6 ⊠ E(2) 2,2 and (1.9), components of ∆ϕ(x) are given as follows: A(b) ⊗ A(c), we write ∆ϕ(x) A(b) ⊗ A(c) (see § 2.1). 2,2 . Since E(6) 2,2 = E(2) 1,1 ⊠ E(3) 2,2 = xb,c =   1 ⊗ E(6) 2,2 E(2) 1,1 ⊗ E(3) 2,2 E(3) 1,1 ⊗ E(2) 2,2 E(6) 2,2 ⊗ 1 (when b1 = 1), (when b1 = 2), (when b1 = 3), (when b1 = 6). (1.18) This shows that the flip of the element ∆ϕ(E(6) is not equal to ∆ϕ(E(6) 2,2 ). Hence ∆ϕ is non-cocommutative. 2,2 ) in M(A(∗) ⊗ A(∗)) 7 1.4.2 Construction of comodule-C∗-algebra Next, we introduce an example of comodule-C∗-algebra of (A(∗), ∆ϕ). Let O∞ denote the Cuntz algebra generated by the canonical generators {si : i ∈ N} [9]. For two sequences J = (j1, . . . , jn), K = (k1, . . . , kn) ∈ Nn, define E(∞) ∈ O∞ and define U HF∞ as the unital J,K ≡ sj1 · · · sjns∗ kn · · · s∗ k1 C∗-subalgebra of O∞ generated by Sn≥1{E(∞) J,K : J, K ∈ Nn}: U HF∞ ≡ C ∗h[n≥1 {E(∞) J,K : J, K ∈ Nn}i ⊂ O∞. (1.19) We prepare some new notations. For a = (a1, a2, . . .) ∈ eN∞ define the finite subset Sn(a) of Nn by ≥2 and n ≥ 1, Sn(a) ≡ {1, . . . , a1} × · · · × {1, . . . , an}. (1.20) j,k } be as in § 1.3. For J = (ji), K = (ki) ∈ Sn(a), define E(a) J,K ∈ Let {E(n) An(a) by E(a) J,K ≡ E(a1) j1,k1 ⊗ · · · ⊗ E(an) jn,kn . (1.21) Define the ∗-homomorphism ϕ∞,a from U HF∞ to U HF∞ ⊗ A(a) by ϕ∞,a(E(∞) J,K ) ≡ E(∞) J ′ ,K ′ ⊗ E(a) J ′′ ,K ′′ (J, K ∈ Nn, n ≥ 1) (1.22) ′ ′ where J i), K defined as ji = ai(j = (j ′ ′ = (k ′ i) ∈ Nn and J ′′ i − 1) + j i and ki = ai(k ′′ ′′ = (j i ), K i − 1) + k ′ ′′ ′′ ′′ = (k i ) ∈ Sn(a) are i for i = 1, . . . , n. Theorem 1.5 Let U HF∞ and {ϕ∞,a : a ∈ eN∞ (1.22), respectively. Define the ∗-homomorphism Γϕ from U HF∞ to M(U HF∞⊗ A(∗)) by ≥2} be as in (1.19) and ϕ∞,a(x) (x ∈ U HF∞) (1.23) Γϕ(x) ≡ Ya∈ eN∞ ≥2 where we identify M(U HF∞ ⊗ A(∗)) with Q{U HF∞ ⊗ A(a) : a ∈ eN∞ Then (U HF∞, Γϕ) is a right comodule-C∗-algebra of (A(∗), ∆ϕ). ≥2}. In § 2, we will prove theorems in § 1.4. In § 3, we will show tensor product formulas among ∗-representations of UHF algebras, and show more concrete tensor product formulas for special UHF algebras. 8 2 Proofs of main theorems In this section, we show a general method to construct C∗-bialgebras and prove theorems in § 1.4. 2.1 C∗-weakly coassociative system In this subsection, we review a general method to construct C∗-bialgebras [18, 25], and generalize it in order to prove theorems. and the direct sum of Ai's, respectively. It is known that M(⊕i∈ΩAi) ∼= First, we recall basic facts of the direct product and the direct sum of general C∗-algebras [6]. We define two C∗-algebras Qi∈Ω Ai and Li∈Ω Ai as follows: Qi∈Ω Ai ≡ {(ai) : k(ai)k ≡ supi kaik < ∞}, Li∈Ω Ai ≡ {(ai) : kaik → 0 as i → ∞}. We call Qi∈Ω Ai and Li∈Ω Ai the direct product Qi∈Ω M(Ai). If Ai is unital for each i, then M(⊕i∈ΩAi) ∼=Qi∈Ω Ai. Let {Bi : i ∈ Ω} be another set of C∗-algebras and let {fi : i ∈ Ω} be a set of ∗-homomorphisms such that fi ∈ Hom(Ai, Bi) for each i ∈ Ω. Then we obtain ⊕i∈Ωfi ∈ Hom(⊕i∈ΩAi, ⊕i∈ΩBi). If fi is nondegenerate for each i, then ⊕i∈Ωfi is also nondegenerate. If both Ai and Bi are unital and fi is unital for each i ∈ Ω, then ⊕i∈Ωfi is nondegenerate. A monoid is a set M equipped with a binary associative operation M × M ∋ (a, b) 7→ ab ∈ M, and a unit with respect to the operation. We recall the definition of C∗-weakly coassociative system in [25]. Definition 2.1 Let M be a monoid with the unit e. A data {(Aa, ϕa,b) : a, b ∈ M} is a C∗-weakly coassociative system (= C∗-WCS) over M if Aa is a unital C∗-algebra for a ∈ M and ϕa,b is a unital ∗-homomorphism from Aab to Aa ⊗ Ab for a, b ∈ M such that (i) for all a, b, c ∈ M, the following holds: (ida ⊗ ϕb,c) ◦ ϕa,bc = (ϕa,b ⊗ idc) ◦ ϕab,c (2.1) where idx denotes the identity map on Ax for x = a, c, (ii) there exists a counit εe of Ae such that (Ae, ϕe,e, εe) is a counital C∗- bialgebra, (iii) for each a ∈ M, the following holds: (εe ⊗ ida) ◦ ϕe,a = ida = (ida ⊗ εe) ◦ ϕa,e. (2.2) We slightly generalize Theorem 2.2 in [25] as follows. 9 Theorem 2.2 Let {(Aa, ϕa,b) : a, b ∈ M} be a C∗-WCS over a monoid M. Define C∗-algebras A∗ ≡ ⊕{Aa : a ∈ M}, Ca ≡ ⊕{Ab ⊗ Ac : (b, c) ∈ Na} (a ∈ M) (2.3) where Na ≡ {(b, c) ∈ M × M : bc = a}. Define ∆(a) ∆ϕ ∈ Hom(A∗, M(A∗ ⊗ A∗)) and ε ∈ Hom(A∗, C) by ϕ ∈ Hom(Aa, M(Ca)), ∆(a) ϕ (x) ≡ Y(b,c)∈Na ϕb,c(x) (x ∈ Aa), ∆ϕ ≡ ⊕{∆(a) ϕ : a ∈ M}, ε ≡ εe ◦ Ee (2.4)   where we identify M(A∗ ⊗ A∗) with Q{M(Ca) : a ∈ M}, and Ee denotes the projection from A∗ onto Ae. Then the following holds: (i) (A∗, ∆ϕ, ε) is a proper counital C∗-bialgebra. (ii) In addition, if #Na < ∞ for each a ∈ M, then ∆ϕ(A∗) ⊂ A∗ ⊗ A∗ where we naturally identify A∗ ⊗ A∗ with a C∗-subalgebra of M(A∗ ⊗ A∗). Proof. (i) From (2.3), M(Ca) = Q{Ab ⊗ Ac : (b, c) ∈ Na}. Hence ∆(a) well-defined. Since M(A∗ ⊗ A∗) =Qa,b∈M Aa ⊗ Ab, ∆ϕ is also well-defined. We show the coassociativity of ∆ϕ. Let a, b, c ∈ M and let Y ∈ Aa ⊗ Ab ⊗ Ac and x ∈ Aabc. From (2.1) and (2.4), we can verify that ϕ is {(∆ϕ ⊗ id) ◦ ∆ϕ}(x)Y = {(id ⊗ ∆ϕ) ◦ ∆ϕ}(x)Y. (2.5) This implies {(∆ϕ ⊗ id) ◦ ∆ϕ}(x) = {(id ⊗ ∆ϕ) ◦ ∆ϕ}(x) on A∗ ⊗ A∗ ⊗ A∗ for each x ∈ A∗. Hence the coassociativity is verified. As the same token, we can verify that ε is a counit and (A∗, ∆ϕ) is proper. (ii) This follows from Theorem 2.2 of [25]. We call (A∗, ∆ϕ, ε) in Theorem 2.2 by a (counital) C∗-bialgebra associated with {(Aa, ϕa,b) : a, b ∈ M}. We prepare lemmas as follows. Lemma 2.3 (i) Let {(Aa, ϕa,b) : a, b ∈ M} be a C∗-WCS over a monoid M and let (A∗, ∆ϕ) be as in Theorem 2.2 associated with {(Aa, ϕa,b) : 10 a, b ∈ M}. For a, b ∈ M, let Ia denote the unit of Aa for a ∈ M, and define Xa,b ≡ ϕa,b(Aab)(Aa ⊗ Ib), Ya,b ≡ ϕa,b(Aab)(Ia ⊗ Ab) (2.6) where ϕa,b(Aab)(Aa ⊗ Ib) and ϕa,b(Aab)(Ia ⊗ Ab) mean the linear spans of {ϕa,b(x)(y ⊗ Ib) : x ∈ Aab, y ∈ Aa} and {ϕa,b(x)(Ia ⊗ y) : x ∈ Aab, y ∈ Ab}, respectively. If both Xa,b and Ya,b are dense in Aa ⊗ Ab for each a, b ∈ M, then (A∗, ∆ϕ) satisfies the cancellation law. (ii) For a C∗-WCS {(Aa, ϕa,b) : a, b ∈ M} over a monoid M, assume that B is a unital C∗-algebra and a set {ϕB,a : a ∈ M} of unital ∗-homomorphisms such that ϕB,a ∈ Hom(B, B ⊗ Aa) for each a ∈ M and the following holds: (ϕB,a ⊗ idb) ◦ ϕB,b = (idB ⊗ ϕa,b) ◦ ϕB,ab (a, b ∈ M). (2.7) Then B is a right comodule-C∗-algebra of the C∗-bialgebra (A∗, ∆ϕ) with the unital coaction Γϕ ≡Qa∈M ϕB,a. (i) By definition, the algebraic direct sum Q ≡ ⊕alg{∆ϕ(Ac)(Aa ⊗ Proof. I) : c, a ∈ M} is dense in ∆ϕ(A∗)(A∗ ⊗ I). On the other hand, ∆ϕ(Ac)(Aa ⊗ I) =  From this, ϕa,b(Aab)(Aa ⊗ Ib) (∃b ∈ M s.t. ab = c), (2.8) 0 (otherwise). Q = ⊕alg{ϕa,b(Aab)(Aa ⊗ Ib) : a, b ∈ M} = ⊕alg{Xa,b : a, b ∈ M}. (2.9) By assumption, ⊕alg{Xa,b : a, b ∈ M} is dense in ⊕alg{Aa ⊗ Ab : a, b ∈ M}. Hence ∆ϕ(A∗)(A∗ ⊗ I) is dense in A∗ ⊗ A∗. In a similar fashion, we see that ∆ϕ(A∗)(I ⊗ A∗) is also dense in A∗ ⊗ A∗. Hence the statement holds. (ii) By identifying M(B ⊗ A∗) with Q{B ⊗ Aa : a ∈ M}, Γϕ is well-defined. From (2.7), we can verify that (Γϕ ⊗ idA∗ ) ◦ Γϕ = (idB ⊗ ∆ϕ) ◦ Γϕ. Hence the statement holds. 2.2 Proofs of theorems In this subsection, we prove theorems in § 1.4. 11 Proof of Theorem 1.3. (i) From Theorem 2.2(i), it is sufficient to show that ≥2} in (1.2) and (1.4) is a C∗-WCS over the monoid {(A(a), ϕa,b) : a, b ∈ eN∞ eN∞ ≥2. By definition, the following holds: (ϕa,b ⊗ idc) ◦ ϕa·b, c = (ida ⊗ ϕb,c) ◦ ϕa, b·c (a, b, c ∈ eN∞ ≥2) (2.10) where idx denotes the identity map on A(x) for x = a, c. Equivalently, the following diagram is commutative: Figure 2.4 ✑✑✸ ϕa,b·c ✑ ✑ ✑ ◗ A(a) ⊗ A(b · c) ◗ ◗ ida ⊗ ϕb,c ◗ ◗◗s A(a · b · c) A(a) ⊗ A(b) ⊗ A(c). ◗ ϕa·b,c ◗ ◗◗s ✑✑✸ ✑ ϕa,b ⊗ idc ✑ ✑ A(a · b) ⊗ A(c) of (A(∗), ∆ϕ) has been shown in Remark 1.4(ii). Especially, (A(1), ϕ1,1) is a one-dimensional C∗-bialgebra with counit idA(1). ≥2} is a C∗-WCS. The non-cocommutativity Hence {(A(a), ϕa,b) : a, b ∈ eN∞ (ii) For a, b ∈ eN∞ ≥2, let Xa,b ≡ ϕa,b(A(a · b))(A(a) ⊗ Ib) and Ya,b ≡ ϕa,b(A(a · b))(Ia ⊗ A(b)). From Lemma 2.3(i), it is sufficient to show that both Xa,b and Ya,b are dense in A(a) ⊗ A(b). By definition, Xa,b is the linear span of {ϕa,b(x)(y ⊗ Ib) : x ∈ A(a · J,K be as in (1.20) and (1.21), respectively. b), y ∈ A(a)}. Let Sn(a) and E(a) We see that A(a) ⊗ A(b) is linearly spanned by the set [n,m≥1 {E(a) J ′ ,K ′ ⊗ E(b) J ′′ ,K ′′ : J ′ ′ , K ∈ Sn(a), J ′′ ′′ , K ∈ Sm(b)} (2.11) as a Banach space. For n, m ≥ 1, fix J ′ , K ′ ∈ Sn(a) and J ′′ , K ′′ ∈ Sm(b). (a) If n = m, then we can choose J and K in Sn(a · b) such that E(a·b) J,K = J ′′ ,K ′′ where ⊠ means the componentwise Kronecker prod- J ′ ,K ′ ⊗E(b) J ′ ,K ′ ⊗ J ′ ,K ′ ⊗Ib), we see E(a) J ′′ ,K ′′ = ∆ϕ(E(a·b) J,K )(E(a) J ′ ,K ′ ⊠ E(b) E(a) uct. Since E(a) E(b) J ′′ ,K ′′ ∈ ∆ϕ(A(a · b))(A(a) ⊗ Ib). 12 (b) If n − m = k > 0, then we can write as J ′ ,K ′ ⊗ E(b) E(a) J ′′ ,K ′′ = XL∈Sm,k(b) J ′ ,K ′ ⊗ E(b) E(a) (J ′′ L),(K ′′ L) (2.12) ′′ L) denotes the concatenation of two sequences J where Sm,k(b) ≡ {1, . . . , bm+1}×· · ·×{1, . . . , bm+k} for b = (b1, b2, . . .) and (J and L. The right hand side of (2.12) is contained in ∆ϕ(A(a · b))(A(a) ⊗ Ib) from (a). ′′ (c) If n − m = −k < 0, then this follows from (b) by the same token. From (a),(b),(c), ∆ϕ(A(a · b))(A(a) ⊗ Ib) is dense in A(a) ⊗ A(b). Just the same, we see that ∆ϕ(A(a · b))(Ia ⊗ A(b)) is dense in A(a) ⊗ A(b). Hence the statement holds. Proof of Theorem 1.5. Let a, b ∈ eN∞ x ∈ U HF∞. By definition, we see that ≥2, Y ∈ U HF∞ ⊗ A(a) ⊗ A(b) and {(ϕ∞,a ⊗ idb) ◦ ϕ∞, b}(x)Y = {(id∞ ⊗ ϕa,b) ◦ ϕ∞, a·b}(x)Y (2.13) where id∞ denotes the identity map on U HF∞. From this and Lemma 2.3(ii) for B = U HF∞ and ϕB,a ≡ ϕ∞,a, the statement holds. 3 Tensor product formulas In this section, we show tensor product formulas of ∗-representations of the C∗-algebra A(∗) with respect to the comultiplication ∆ϕ in § 1.4. 3.1 Basic properties In this subsection, we introduce a tensor product of ∗-representations and that of states of UHF algebras, and show its basic properties. For a C∗- algebra A, let RepA and S(A) denote the class of all ∗-representations and the set of all states of A, respectively. By using the set {ϕa,b : a, b ∈ N∞ ≥2} in (1.4), define the operation ⊗ϕ from RepA(a) × RepA(b) to RepA(a · b) by π1 ⊗ϕ π2 ≡ (π1 ⊗ π2) ◦ ϕa,b (3.1) 13 ′ ′ ′ for (π1, π2) ∈ RepA(a) × RepA(b). We see that if πi and π i are unitarily equivalent for i = 1, 2, then π1 ⊗ϕ π2 and π 2 are also unitarily equiv- alent. Furthermore, define the operation ⊗ϕ from S(A(a)) × S(A(b)) to S(A(a · b)) by 1 ⊗ϕ π ρ1 ⊗ϕ ρ2 ≡ (ρ1 ⊗ ρ2) ◦ ϕa,b (3.2) for (ρ1, ρ2) ∈ S(A(a)) × S(A(b)). From (2.10), we see that (π1⊗ϕπ2)⊗ϕπ3 = π1⊗ϕ(π2⊗ϕπ3), (ρ1⊗ϕρ2)⊗ϕρ3 = ρ1⊗ϕ(ρ2⊗ϕρ3) (3.3) for each (π1, π2, π3) ∈ RepA(a) × RepA(b) × RepA(c) and for (ρ1, ρ2, ρ3) ∈ S(A(a)) × S(A(b)) × S(A(c)) and a, b, c ∈ N∞ ≥2. The following fact is a paraphrase of well-known results of tensor prod- ucts of factors. Fact 3.1 Let π1 and π2 be ∗-representations of A(a) and A(b), respectively. Then the following holds: (i) If both π1 and π2 are factor representations, then so is π1 ⊗ϕ π2. (ii) The type of π1 ⊗ϕ π2 coincides with that of π1 ⊗ π2 where the type of a representation π of a C∗-algebra A means the type of the von Neumann algebra π(A) ([6], Theorem III.2.5.27). ′′ (iii) If both π1 and π2 are irreducible, then so is π1 ⊗ϕ π2. Proof. By the definition of ⊗ϕ, (π1 ⊗ϕ π2)(A(a · b)) = (π1 ⊗ π2)(A(a) ⊗ A(b)). (3.4) (i) Since π1 ⊗ π2 is also a factor representation, the statement holds from (3.4). (ii) By definition, the type of π1 ⊗ϕ π2 is that of {(π1 ⊗ϕ π2)(A(a · b))}′′ From this and (3.4), the statement holds. (iii) By assumption, π1 ⊗ π2 is also irreducible. From this and (3.4), the statement holds. . By definition, the essential part of the tensor product ⊗ϕ is given by the set {ϕa,b} of isomorphisms in (1.3). This type of tensor product is known yet in neither operator algebras nor the purely algebraic theory of quantum groups [16]. 14 Remark 3.2 (i) Our terminology "tensor product of representations" is ′ ∈ RepA(a), ′ ∈ RepA(a · a) because a · a 6= a for different from usual sense [12]. Remark that, for π, π π ⊗ϕ π any a ∈ N∞ ≥2. 6∈ RepA(a) but π ⊗ϕ π ′ (ii) From Fact 3.1(iii), there is no nontrivial branch of the irreducible de- composition of the tensor product of any two irreducibles. In general, such a tensor product of the other algebra is decomposed into more than one irreducible component. For example, see Theorem 1.6 of [17]. 3.2 GNS representations by product states and their tensor product formulas We recall well-known GNS representations by product states of UHF alge- bras [3, 4, 28]. Let Mn be as in § 1.3 and let Mn,+,1 denote the set of all positive elements in Mn whose traces are 1. Then a linear functional ω on Mn is a state of Mn if and only if ω is equal to the state ωT which is defined as ωT (x) ≡ tr(T x) (x ∈ Mn) for some T ∈ Mn,+,1 where tr denotes the trace of Mn. For a = (a1, a2, . . .) ∈ N∞ sequence T = (T (n))n≥1 ∈ T (a), define the state ωT of A(a) by ≥2, let T (a) ≡ Qn≥1 Man,+,1. For a ωT(E(a1) j1,k1 ⊗ · · · ⊗ E(an) jn,kn ) ≡ T (1) k1,j1 · · · T (n) kn,jn (3.5) for each j1, . . . , jn, k1, . . . , kn and n ≥ 1 where T (n) j,k 's denote matrix elements of the matrix T (n). Then ωT coincides with the product state Nn≥1 ωT (n). Theorem 3.3 ([15], Remark 11.4.16) For each T ∈ T (a), the state ωT in (3.5) is a factor state, that is, if (HT, πT, ΩT) is the GNS triplet of A(a) by ωT, then πT(A(a)) is a factor. ′′ ′′ The factor MT ≡ πT(A(a)) is called an Araki-Woods factor (or infinite tensor product of finite dimensional type I (=ITPFI) factor) [3, 4]. Prop- erties of MT and (HT, πT, ΩT) are closely studied in [3, 4, 31] and [5], respectively. Next, we show tensor product formulas of πT's in Theorem 3.3 as follows. Theorem 3.4 Let a, b ∈ N∞ representation πT. ≥2 and let ωT be as in (3.5) with the GNS (i) For each T ∈ T (a) and R ∈ T (b), ωT ⊗ϕ ωR = ωT⊠R (3.6) 15 where T ⊠ R ∈ T (a · b) is defined as T ⊠ R ≡ (T (1) ⊠ R(1), T (2) ⊠ R(2), T (3) ⊠ R(3), . . .) (3.7) for T = (T (n)) and R = (R(n)). (ii) For each T ∈ T (a) and R ∈ T (b), πT ⊗ϕ πR is unitarily equivalent to πT⊠R. (i) By definition, the statement holds from direct computation. Proof. (ii) Let a ∈ N∞ ≥2. For T ∈ T (a), let (HT, πT, ΩT) denote the GNS triplet by the state ωT. Define the GNS map ΛT [26, 27] from A(a) to HT by ΛT(x) ≡ πT(x)ΩT for x ∈ A(T). Let a, b ∈ N∞ ≥2. For T ∈ T (a) and R ∈ T (b), define the unitary U (T,R) from HT⊠R to HT ⊗ HR by U (T,R)ΛT⊠R(x) ≡ (ΛT ⊗ ΛR)(ϕa,b(x)) (x ∈ A(a · b)). (3.8) Since ϕa,b is bijective, U (T,R) is well-defined as a unitary, and we see that U (T,R)πT⊠R(x)(U (T,R))∗ = (πT ⊗ϕ πR)(x) (x ∈ A(a · b)). (3.9) Hence two representations πT⊠R and πT ⊗ϕ πR are unitarily equivalent. From Theorem 3.4, the tensor product ⊗ϕ is compatible with product states and their GNS representations. More precisely, for the following two semi- groups (T , ⊠) and (S, ⊗ϕ), the map T ≡ [a∈N∞ ≥2 T (a) ∋ T 7→ ωT ∈ S ≡ [a∈N∞ ≥2 S(A(a)) (3.10) is a semigroup homomorphism. Let Ra denote the set of all unitary equiv- alence classes in RepA(a). Then T ∋ T 7→ [πT] ∈ R ≡ [a∈N∞ ≥2 Ra (3.11) is also a semigroup homomorphism from (T , ⊠) to (R, ⊗ϕ) where [π] denotes the unitary equivalence class of a representation π. 16 3.3 Examples In this subsection, we show examples of Theorem 3.4 for special UHF alge- bras. Let A(a) and T (a) be as in § 3.2. For n ≥ 1, let n ≡ (n, n, n, . . .) ∈ N∞ and let U HFn ≡ A(n) = (Mn)⊗∞ (n ≥ 2). Then U HFn is the UHF algebra of Glimm's type {nl}l≥1. (3.12) (3.13) Let {E(n) i,j } be as in § 1.3. For j ∈ {1, . . . , n}, define F (n) j ≡ E(n) j,j . j ∈ Mn,+,1 for each j. For J = (j1, j2, . . .) ∈ {1, . . . , n}∞, define , . . .) ∈ T (n). From (3.5), we see that Then F (n) T(J) ≡ (F (n) j1 , F (n) j2 ⊗ · · · ⊗ E(n) ωT(J)(E(n) l1,k1 lm,km ) ≡ δl1,j1 · · · δlm,jm δk1,j1 · · · δkm,jm (3.14) ′ for each l1, . . . , lm, k1, . . . , km ∈ {1, . . . , n} and m ≥ 1. For J = (jn)n∈N, J (j such that jr = j = if there exists an integer n0 ≥ 1 n)n∈N ∈ {1, . . . , n}∞, we write J ≈ J ′ ′ r for each r ≥ n0. ′ Proposition 3.5 For J ∈ {1, . . . , n}∞, let πT(J) denote the GNS repre- sentation of U HFn by the state ωT(J) in (3.14), and let Pn[J] denote the unitary equivalence class of πT(J). Then the following holds: (i) For each J ∈ {1, . . . , n}∞, πT(J) is irreducible. (ii) For J, J ′ ∈ {1, . . . , n}∞, Pn[J] = Pn[J ′ ] if and only if J ≈ J ′ . (iii) For each J = (jl) ∈ {1, . . . , n}∞, K = (kl) ∈ {1, . . . , m}∞ and n, m ≥ 2, Pn[J] ⊗ϕ Pm[K] = Pnm[J ⋆ K] (3.15) where J ⋆ K ∈ T (n · m) is defined by J ⋆ K ≡ (m(j1 − 1) + k1, m(j2 − 1) + k2, m(j3 − 1) + k3, . . .). (3.16) (i) Since the state ωT(J) is the product state of pure states, ωT(J) Proof. is also pure. Hence the statement holds. (ii) From (2.1) in [2] (see also [7]), the statement holds. (iii) Recall ⊠ in (3.7). Then we can verify that T(J) ⊠ T(K) = T(J ⋆ K) for each J ∈ {1, . . . , n}∞ and K ∈ {1, . . . , m}∞. From this and Theorem 3.4(ii), the statement holds. In Theorem 1.6 of [17], "J ⋆ K" is written as the different notation "J · K." 17 Example 3.6 From Proposition 3.5, P ≡ Sn≥2{Pn[J] : J ∈ {1, . . . , n}∞} is a semigroup of unitary equivalence classes of irreducible representations with respect to the product ⊗ϕ. For n ≥ 1, let n be as in (3.12). Then P2[1] ⊗ϕ P2[2] = P4[2], P2[2] ⊗ϕ P2[1] = P4[3] (3.17) because 1 ⋆ 2 = 2 and 2 ⋆ 1 = 3. Since 2 6≈ 3, P4[2] 6= P4[3]. Hence (P, ⊗ϕ) is non-commutative. The class Pn[J] in Proposition 3.5(ii) coincides with the restriction of a permutative representation of the Cuntz algebra On on the UHF subalgebra of U (1)-gauge invariant elements in On, which is called an atom [2, 8]. This class contains only type I representations of U HFn. Relations with representations of Cuntz algebras and quantum field theory are well studied [2, 8]. From Example 3.6 and pure states associated with P2[1] and P2[2], the following holds. Corollary 3.7 Let S(A(a)) and Ra be as in § 3.1. (i) There exist a, b ∈ N∞ ′ 6= ω such that ω ⊗ϕ ω ≥2, and states ω ∈ S(A(a)) and ω ′ ⊗ϕ ω. ′ ∈ S(A(b)) (ii) There exist a, b ∈ N∞ ≥2, and classes [π] ∈ Ra and [π [π] ⊗ϕ [π ′ ′ ] 6= [π ] ⊗ϕ [π]. ′ ] ∈ R b such that From Corollary 3.7(i) and (ii), we say that ⊗ϕ is non-symmetric. In other words, two semigroups (S, ⊗ϕ) and (R, ⊗ϕ) are non-commutative. These non-commutativities come from the non-commutativity of the Kronecker product of matrices in Remark 1.2. Problem 3.8 (i) Generalize the tensor product in (3.1) to that of repre- sentations of approximately finite dimensional (=AF) algebras which are not always UHF algebras. (ii) Reconstruct UHF algebras from (S, ⊗ϕ) and (R, ⊗ϕ), and show a Tat- suuma duality type theorem for UHF algebras [33]. References [1] Abe, E. (1977). Hopf algebras. Cambridge University Press. 18 [2] Abe, M., Kawamura, K. (2008). Branching laws for endomorphisms of fermions and the Cuntz algebra O2. J. Math. Phys. 49:043501-01 -- 043501-10. [3] Araki, H., Woods, E. J. (1968). A classification of factors. Publ. Res. Inst. Math. Sci. 3:51 -- 130. [4] Araki, H. (1969). A classification of factors, II. Publ. Res. Inst. Math. Sci. 4:585 -- 593. [5] Araki, H., Nakagami, Y. (1972/73). A remark on an infinite tensor product of von Neumann algebras. Publ. Res. Inst. Math. Sci. 8:363 -- 374. [6] Blackadar, B. (2006). Operator algebras. Theory of C∗-algebras and von Neumann algebras. Springer-Verlag Berlin Heidelberg New York. [7] Bergmann, W. R., Conti, R. (2003). Induced product representation of extended Cuntz algebras. Annali Mat. Pure Appl. 182:271-286. [8] Bratteli, O., Jorgensen, P. E. T. (1999). Iterated function systems and permutation representations of the Cuntz algebra. Memoirs Am. Math. Soc. 139(663):1 -- 89. [9] Cuntz, J. (1977). Simple C ∗-algebras generated by isometries. Commun. Math. Phys. 57:173 -- 185. [10] Dief, A. S. (1982). Advanced matrix theory for scientists and engineers. Abacus Press, Turnbridge Wells & London. [11] Enock, M., Schwartz, J. M. (1992). Kac algebras and duality of locally compact groups. Springer-Verlag. [12] Fulton, W., Harris, J. (1991). Representation theory. Springer-Verlag. [13] Glimm, J. (1960). On a certain class of operator algebras. Trans. Amer. Math. Soc. 95:318 -- 340. [14] Herschend, M. (2009). On the Representation Ring of a Quiver. Algebr. Represent. Theor. 12:513 -- 541. [15] Kadison, R. V., Ringrose, J. R. (1983). Fundamentals of the theory of operator algebras II. Academic Press. [16] Kassel, C. (1995). Quantum groups. Springer-Verlag. 19 [17] Kawamura, K. (2007). A tensor product of representations of Cuntz algebras. Lett. Math. Phys. 82:91 -- 104. [18] Kawamura, K. (2008). C∗-bialgebra defined by the direct sum of Cuntz algebras. J. Algebra 319:3935 -- 3959. [19] Kawamura, K. (2009). Classification and realizations of type III fac- tor representations of Cuntz-Krieger algebras associated with quasi-free states. Lett. Math. Phys. 87:199 -- 207. [20] Kawamura, K. (2009). C∗-bialgebra defined as the direct sum of Cuntz- Krieger algebras. Comm. Algebra 37:4065 -- 4078. [21] Kawamura, K. Triangular C∗-bialgebra defined as the direct sum of matrix algebras. Linear Alg. Appli., to appear. [22] Kawamura, K. Tensor products of type III factor representations of Cuntz-Krieger algebras. math.OA/0805.0667v1. [23] Kawamura, K. A tensor product of representations of UHF algebras arising from Kronecker products. math.OA/0910.1420v1. [24] Kawamura, K. Inductive limit violates quasi-cocommutativity. math.OA/1003.3739v1. [25] Kawamura, K. Non-cocommutative C∗-bialgebra defined as the direct sum of free group C∗-algebras. math.OA/0153631. [26] Kustermans, J., Vaes, S. (2000). The operator algebra approach to quantum groups. Proc. Natl. Acad. Sci. USA 97(2):547 -- 552. [27] Masuda, T., Nakagami, Y., Woronowicz, S. L. (2003). A C ∗-algebraic framework for quantum groups. Int. J. Math. 14:903 -- 1001. [28] Von Neumann, J. (1939). On infinite direct products. Compositio Math. 6:1 -- 77. [29] Powers, R. T. (1967). Representations of uniformly hyperfinite algebras and their associated von Neumann rings. Ann. Math. 86:138 -- 171. [30] Puk´anszky, L. (1961). On the Kronecker product of irreducible repre- sentations of the 2 × 2 real unimodular group, I. Trans. Amer. Math. Soc. 100:116 -- 152. 20 [31] Shlyakhtenko, D. (2004). On the classification of full factors of type III. Trans. Amer. Math. Soc. 356(10):4143 -- 4159. [32] Steeb, W.-H., Shi, T. K. (1997). Matrix calculus and Kronecker product with applications and C++ programs. World Scientific. [33] Tatsuuma, N. (1967). A duality theorem for locally compact groups. J. Math. Kyoto Univ. 6:187 -- 293. 21
1503.01344	1	1503	2015-03-04T15:37:39	Quadratic Conorm and extremally rich JB*-triples	[ "math.OA" ]	We introduce and study the class of extremally rich JB$^$-triples. We establish new results to determine the distance from an element $a$ in an extremally rich JB$^$-triple $E$ to the set $\partial_{e} (E_1)$ of all extreme points of the closed unit ball of $E$. More concretely, we prove that $$\hbox{dist} (a,\partial_e (E_1)) =\max \{ 1, \\|a\\|-1\},$$ for every $a\in E$ which is not Brown-Pedersen quasi-invertible. As a consequence, we determine the form of the $\lambda$-function of Aron and Lohman on the open unit ball of an extremally rich JB$^$-triple $E$, by showing that $\lambda (a)= \frac12$ for every non-BP quasi-invertible element $a$ in the open unit ball of $E$. We also prove that for an extremally rich JB$^$-triple $E$, the quadratic connorm $\gamma^{q}(.)$ is continuous at a point $a\in E$ if, and only if, either $a$ is not von Neumann regular {\rm(}i.e. $\gamma^{q}(a)=0${\rm)} or $a$ is Brown-Pedersen quasi-invertible.	math.OA	math	QUADRATIC CONORM AND EXTREMALLY RICH JB∗-TRIPLES FATMAH B. JAMJOOM, ANTONIO M. PERALTA, AKHLAQ A. SIDDIQUI, AND HAIFA M. TAHLAWI Abstract. We introduce and study the class of extremally rich JB∗- triples. We establish new results to determine the distance from an element a in an extremally rich JB∗-triple E to the set ∂e(E1) of all extreme points of the closed unit ball of E. More concretely, we prove that dist(a, ∂e(E1)) = max {1, kak − 1} , for every a ∈ E which is not Brown-Pedersen quasi-invertible. As a con- sequence, we determine the form of the λ-function of Aron and Lohman on the open unit ball of an extremally rich JB∗-triple E, by showing that λ(a) = 1 2 for every non-BP quasi-invertible element a in the open unit ball of E. We also prove that for an extremally rich JB∗-triple E, the quadratic connorm γ q(.) is continuous at a point a ∈ E if, and only if, either a is not von Neumann regular (i.e. γ q(a) = 0) or a is Brown-Pedersen quasi-invertible. 1. Introduction This paper presents new investigations which provide some answers to problems concerning the geometric structure of those complex Banach spaces included in the class of JB∗-triples. W. Kaup proved in 1983 that the open unit ball of a complex Banach space X is a bounded symmetric domain (a pure holomorphic property) if, and only, if X is a JB∗-triple (cf. [19]). That is, the holomorphic properties of the open unit ball of X determine when X satisfies certain algebraic-geometric axioms listed below. L. Harris proved in [13] that the open unit ball of a C∗-algebra A is a bounded symmetric domain; actually A is a JB∗-triple with triple product defined by (1.1) {x, y, z} := 1 2 (xy∗z + zy∗x). 1991 Mathematics Subject Classification. Primary 46L70; 17C65; 46L05 Secondary 15A09 47A05 47D25 . Key words and phrases. Extremally rich JB∗-triple, Brown-Pedersen quasi-invertibility, reduced minimum modulus, conorm, quadratic conorm. The authors extend their appreciation to the Deanship of Scientific Research at King Saud University for funding this work through research group no RG-1435-020. Second author also partially supported by the Spanish Ministry of Economy and Competitiveness, D.G.I. project no. MTM2011-23843. 1 2 F.B. JAMJOOM, A.M. PERALTA, A.A. SIDDIQUI, AND H.M. TAHLAWI JB∗-triples have been intensively studied during the last three decades, and special attention has been paid to the geometric properties of these spaces. In several cases the studies determine whether JB∗-triples satisfy certain properties fulfilled by C∗-algebras. For example, the quasi-invertible elements of C∗-algebras studied by L.G. Brown and G.K. Pedersen in [4] gave rise to the introduction of Brown-Pedersen quasi-invertible elements in a JB∗-triple in [23] and [24]. Brown and Pedersen show in [5] that quasi- invertible elements in C∗-algebras play a crucial role to determine the form of the λ-function of R. Aron and R. Lohman [1]. The precise description of the λ-function is determined in [5]. A C∗-algebra A is said to be extremally rich if the set A−1 of quasi-invertible elements in A is norm-dense in A (cf. [4, §3]). The class of extremally rich C∗-algebras is strictly bigger than the class of von Neumann algebras. From the geometric point of view, a unital C∗-algebra A is extremally rich if and only if it has the (uniform) λ-property, that is, the infimum of the values of the λ-function on the closed unit ball of A is bigger than zero (cf. [4, §3] and [5, Theorem 3.7]). q In a recent paper, we proved that every JBW∗-triple (i.e. a JB∗-triple which is also a dual Banach space) satisfies the uniform λ-property (cf. [15]). In the same paper we determine the λ-function of the closed unit ball of a JBW∗-triple, and on the set E−1 of all Brown-Pedersen quasi-invertible elements in the closed unit ball of a general JB∗-triple E. If we also assume that the set ∂e(E1) of all extreme points in the closed unit ball E1 of E is non-empty, we can only prove that the inequality q (1.2) λ(a) ≤ 1 2 (1 − αq(a)), holds for every a ∈ E1\E−1 [15, Corollary 3.7]). q , where αq(a) is the distance from a to E−1 q (cf. The question whether in (1.2) the inequality sign can be replaced with an equality symbol, is one of the main open problems in the setting of JB∗- triples. This question is deeply related to the problem of determining the distance from an element a to the set ∂e(E1). The best estimation follows from Theorem 3.6 in [15], where it is established that for every JB∗-triple E with ∂e(E1) 6= ∅, the inequalities 1 + kak ≥ dist(a, ∂e(E1)) ≥ max {1 + αq(a), kak − 1} , hold for every a in E\E−1 q . We introduce, in this paper, the notion of extremally rich JB∗-triple with the aim of studying the above problems more in deep. We say that a JB∗- triple E is extremally rich if the set of Brown-Pedersen quasi-invertible ele- ments in E is norm dense in E. Several characterizations of extremally rich JB∗-triples are provided in Proposition 2.4. Among the new results in this paper, we prove that for an extremally rich JB∗-triple we have dist(a, ∂e(E1)) = max {1, kak − 1} , QUADRATIC CONORM AND EXTREMALLY RICH JB∗-TRIPLES 3 for every a ∈ E\E−1 λ(a) = 1 of an extremally rich JB∗-triple (see Corollary 2.7). (see Theorem 2.6). As a consequence, we show that 2 for every non-BP quasi-invertible element a in the open unit ball q We also deal with another related, open question. We recall that the reduced minimum modulus of a non-zero bounded linear or conjugate linear operator T between two normed spaces X and Y is defined by (1.3) γ(T ) := inf{kT (x)k : dist(x, ker(T )) ≥ 1}. Following [17], we set γ(0) = ∞. When X and Y are Banach spaces, we have γ(T ) > 0 ⇔ T (X) is norm closed, (cf. [17, Theorem IV.5.2]). The quadratic-conorm, γq(a), of an element a in a JB∗-triple E is de- fined as the reduced minimum modulus of the conjugate linear operator Q(a) : E → E, x 7→ Q(a)(x) := {a, x, a}, that is, γq(a) = γ(Q(a)) (see [7]). Theorem 3.13 in [7] proves that γq(.) is upper semi-continuous on E\{0}. It is also remarked, in the just quoted reference, that the continuity points of γq(.) are, in general, unknown. In this paper we throw some light into the question of determining the continuity points of the quadratic conorm, by showing that for an extremally rich JB∗-triple E, the quadratic connorm γq(.) is continuous at a point a ∈ E if, and only if, either a is not von Neumann regular (i.e. γq(a) = 0) or a is BP quasi-invertible (see Theorem 3.3). We also explore the applications of this result to determine the con- tinuity points of the conorm of an extremally rich C∗-algebra in the sense introduced by R. Harte and M. Mbekhta in [14]. 1.1. Preliminaries. A complex Banach space E is a JB∗-triple if it can be equipped with a triple product {., ., .} : E × E × E → E, (x, y, z) 7→ {x, y, z}, which is linear and symmetric in x and z and conjugate linear in y, and satisfies the following axioms: (a) (Jordan identity) {x, y, {a, b, c}} = {{x, y, a}, b, c} − {a, {y, x, b}, c} + {a, b, {xyc}}, for every a, b, c ∈ E; (b) For each a ∈ E, the operator x 7→ L(a, a)(x) := {a, a, x} is hermitian with non-negative spectrum; (c) k{x, x, x}k = kxk3, for all x ∈ E. The class of JB∗-triples includes all C∗-algebras, all complex Hilbert It is further known that every JB∗-algebra spaces, and all spin factors. is a JB∗-triple with the triple product {x, y, z} := (x◦y∗)◦z − (x◦z)◦y∗ + (y∗◦z)◦x. A JBW∗-triple is a JB∗-triple which is also a dual Banach space. The the [10]). Every JBW∗- second dual E∗∗ of a JB∗-triple E is JBW∗-triple (cf. triple W admits a unique isometric predual W∗, and its triple product is separately σ(W, W∗)-continuous (cf. [2]). 4 F.B. JAMJOOM, A.M. PERALTA, A.A. SIDDIQUI, AND H.M. TAHLAWI JB∗-triples are stable by ℓ∞-sums (cf. [19, page 523]), that is, if (Ej) is j Ej is a JB∗-triple with respect a family of JB∗-triples, then the ℓ∞-sum ⊕∞ to the product {(aj), (bj ), (cj )} := ({aj , bj, cj}). Following standard notation, given two elements x, y in a JB∗-triple E, the conjugate linear operator Q(x, y) : E → E is defined by Q(x, y)z := {x, z, y} for all z ∈ E; we usually write Q(x) instead of Q(x, x). The symbol L(x, y) will stand for the linear operator on E given by L(x, y)(z) = {x, y, z}. We recall that an element e in a JB∗-triple E is said to be a tripotent if {e, e, e} = e. It is known that, for each tripotent e in E, the eigenvalues of the operator L(e, e) are contained in the set {0, 1 2 , 1}, and E decomposes in the form E = E2(e) ⊕ E1(e) ⊕ E0(e), where, for i = 0, 1, 2, Ei(e) is the i 2 -eigenspace of L(e, e). This decomposition is called the Peirce decomposition of E with respect to e. The Peirce sub- spaces appearing in the above decomposition satisfy certain multiplication rules (called Peirce rules), which can be stated as follows: {Ei(e), Ej (e), Ek(e)} ⊆ Ei−j+k(e) if i − j + k ∈ {0, 1, 2} and is zero otherwise. In addition, {E2(e), E0(e), E} = 0. The projection of E onto Ek(e) is denoted by Pk(e), and it is called the Peirce k-projection. Peirce projections are contractive (cf. [12]) and satisfy that P2(e) = Q(e)2, P1(e) = 2(L(e, e)−Q(e)2), and P0(e) = IdE −2L(e, e)+ Q(e)2. A tripotent e in E is said to be unitary if L(e, e) coincides with the If E0(e) = {0} we say that e is identity map on E, that is, E2(e) = E. complete. The Peirce space E2(e) is a unital JB∗-algebra with unit e, product x ◦e y := {x, e, y} and involution x∗e := {e, x, e}, respectively. Furthermore, the triple product on E2(e) is given by {a, b, c} = (a ◦e b∗e) ◦e c + (c ◦e b∗e) ◦e a − (a ◦e c) ◦e b∗e (a, b, c ∈ E2(e)). Let a be an element in a JB∗-triple E, and let Ea denote the JB∗-subtriple of E generated by a. That is, Ea coincides with the closed linear span of the elements a, a[3] = {a, a, a}, a[2n+1] := {a, a, a[2n−1]} (n ≥ 2). It follows from the commutative Gelfand theory that there exist a locally compact Hausdorff space La ⊆ (0, kak], with La ∪ {0} compact, and a triple isomorphism Ψa : Ea → C0(La), where C0(Ωx) denotes the Banach space of all complex-valued continuous functions vanishing at 0, such that Ψa(a)(t) = t, ∀t ∈ La (cf. [19, §1, Corollary 1.15]). Therefore, for each natural n, there exists a unique a[1/(2n−1)] ∈ Ea satisfying (a[1/(2n−1)])[2n−1] = a. The sequence (a[1/(2n−1)]) need not be convergent in the norm topology of E. However, when a is an element in a JBW∗-triple W , the sequence (a[1/(2n−1)]) converges in the weak∗ topology of W to a tripotent in W , which is denoted by r(a) and is QUADRATIC CONORM AND EXTREMALLY RICH JB∗-TRIPLES 5 called the range tripotent of a. The tripotent r(a) is the smallest tripotent e in W satisfying that a is positive in the JBW∗-algebra W2(e) (cf. [11, §3, Lemma 3.2]). The deep geometric-algebraic connections appearing in the setting of JB∗- triples materialize in many important properties, one of them ensures that complete tripotents in a JB∗-triple E coincide with the extreme points of its closed unit ball (cf. [3, Lemma 4.1] and [18, Proposition 3.5] or [9, Theorem 3.2.3]). Throughout this paper, the set of all the extreme points of the closed unit ball of a Banach space X is denoted by ∂(X1). The Bergman operator, B(x, y), associated with a couple of elements x, y ∈ E is the mapping defined by B(x, y) := IE − 2L(x, y) + Q(x)Q(y), where IE is the identity operator on E. We observe that, for a tripotent e ∈ E, B(e, e) = P0(e). According to [24], an element x in a JB∗-triple E is called Brown-Pedersen quasi-invertible (BP quasi-invertible for short) if there exists y ∈ E satisfying B(x, y) = 0. We say that y is a BP quasi-inverse of x. It is known that B(x, y) = 0 ⇒ B(y, x) = 0. A BP quasi-invertible element need not admit a unique BP quasi-inverse. It is established in [24] that an element x in E is BP quasi-invertible if and only if there exits a complete tripotent v ∈ E (v ∈ ∂e(E1)) such that x is positive and invertible in the Peirce 2-space E2(v). Therefore, the set E−1 of all BP quasi-invertible elements in E contains all extreme points of the closed unit ball of E. When E = J is a JB∗-algebra, J −1 contains the set J −1 of all invertible elements in E. q q When a C∗-algebra A is regarded as a JB∗-triple, BP quasi-invertible elements in A are precisely the quasi-invertible elements of A in the sense defined by Brown and Pedersen in [4]. We also recall that an element a in a JB∗-triple E is called von Neumann regular if and only if there exists b ∈ E such that Q(a)b = a, Q(b)a = b and [Q(a), Q(b)] := Q(a) Q(b) − Q(b) Q(a) = 0 (cf. [20, Lemma 4.1]). For a von Neumann regular element a, there might exist many elements c in E such that Q(a)c = a. However, there exists a unique element b ∈ E (denoted by a†) satisfying Q(a)b = a, Q(b)a = b and [Q(a), Q(b)] = 0, this unique element b is called the generalized inverse of a in E. For an element a in a JB∗-triple E, we can consider the range tripotent, r(a), of a in E∗∗. It is known that a is von Neumann regular if, and only if, r(a) ∈ E and a is positive and invertible in E2(r(a)) (cf. [7, §§2, pages 191 and 192]). 2. Extremally Rich JB∗-triples In [4, §3] L. Brown and G.K. Pedersen introduced and studied the class of extremally rich C∗-algebras. We recall that a unital C∗-algebra A is extremally rich if the set A−1 of Brown-Pedersen quasi-invertible elements in A is (norm) dense in A. When A is non-unital, it is extremally rich if its unitization enjoys this property. Every von Neumann algebra and every purely infinite (simple) C∗-algebra is extremally rich (cf. [4, §3]). From the q 6 F.B. JAMJOOM, A.M. PERALTA, A.A. SIDDIQUI, AND H.M. TAHLAWI point of view of Banach space theory, a unital C∗-algebra is extremally rich if and only if it has the (uniform) λ-property defined by R. Aron and R. Lohman in [1] (cf. [4, §3] and [5, Theorem 3.7]). We recall that, given a normed space X, x, y ∈ X, with kyk ≤ 1, e ∈ ∂e(X1), and 0 < λ ≤ 1, the ordered 3-tuple (e, y, λ) is said to be amenable to x if x = λe + (1 − λ)y. The λ-function is defined by λ(x) := sup{λ : (e, y, λ) is a 3-tuple amenable to x}. The space X is said to have the λ-property if each element in its closed unit ball admits an amenable 3-tuple (cf. [1]). The notion of Brown-Pedersen quasi-invertibility in JB∗-triples have been recently studied in [23], [24] and [25]. The study of the λ-function in JB∗- triples is developed in [25] and [15], where we proved that every JBW∗-triple (i.e. a JB∗-triple which is a dual Banach space) satisfies the (uniform) λ- property. We introduce the following definition with the aim of determining those JB∗-triples satisfying the (uniform) λ-property. Definition 2.1. A JB∗-triple E is called extremally rich if the set E−1 BP quasi-invertible elements in E is norm dense in E. q of Recall that an element u in a unital JB∗-algebra J is called unitary if u∗ = u−1 (where u−1 denotes the inverse of u), or equivalently, if {u, u, z} = z, ∀z ∈ J (cf. [26, Definition 19.12] or [8, Definition 4.1.53, Proposition 4.1.54 4.1.55]), that is, L(u, u) = IJ (the identity operator over J ). Remark 2.2. (a) We recall that a unital C∗-algebra A is of topological stable rank 1 (tsr 1) if the subgroup A−1 of invertible elements in A is norm dense in A (see [21]). A similar definition is introduced in the category of JB∗- algebras in [22]. If J is a JB∗-algebra of tsr 1, then J = J −1 ⊆ J −1 . This shows that every JB∗-algebra J of tsr 1 is extremally rich. There exist examples of extremally rich C∗-algebras which are not of tsr 1. For example, suppose A is a von Neumann algebra that contains a non-unitary, maximal partial isometry, say v, which is a non-unitary extreme point of A1.Then, v ∈ ∂e(A1) 6= U (A), which implies that A is not of tsr 1 (cf. [22, Corollary 6.10]). On the other hand, every von Neumann algebra is extremally rich (see [4, p.126]). q (b) It should be also noted that the von Neumann envelope of a JB∗- algebra of tsr 1 need not be, in general, of tsr 1 (cf. [22, Theorems 3.1 and 3.2]). (c) Let A be a C∗-algebra. Then A is extremally rich as a C∗-algebra if and only if A is extremally rich when it is regarded as a JB∗-triple with the product in (1.1). Since the class of extremally rich C∗-algebras is strictly bigger than the class of von Neumann algebras, we can immediately confirm that the class of QUADRATIC CONORM AND EXTREMALLY RICH JB∗-TRIPLES 7 extremally rich JB∗-triples is agreeably large, strictly bigger than the class of JBW∗-triples. In our next result we shall establish some characterizations of extremally rich JB∗-triples in the line set down by Brown and Pedersen for C∗-algebras in [4, Theorem 3.3]. To this end, we shall recall a result taken from [15, Proposition 4.4]. First, we recall that for each element a in a JB∗-triple E, mq(a) := dist(a, E\E−1 q ) coincides with the square root of the quadratic conorm of a, whenever a is in E−1 (see [15, Theorem 3.1]). q Proposition 2.3. [15, Proposition 4.4] Let a and b be elements in a JB∗- triple E. Suppose ka − bk < β and b ∈ E−1 and the inequality . Then a + βr(b) ∈ E−1 q q mq(a + βr(b)) ≥ β − kb − ak, holds. Furthermore, under the above conditions, the element P2(r(b))(a) + βr(b) is invertible in the JB∗-algebra E2(r(b)). (cid:3) The promised characterization of extremally rich JB∗-triples reads as fol- lows: Proposition 2.4. For a JB∗-triple E with ∂e(E1) 6= ∅, the following condi- tions are equivalent: (i) E is extremally rich; (ii) For every a ∈ E and any β > 0, there is an element b ∈ E−1 q range tripotent r(b) ∈ ∂e(E1), satisfying that a + βr(b) ∈ E−1 . such that q (iii) For every a ∈ E and any β > 0, there is an element b ∈ E−1 P2(r(b))(a) + βr(b) is invertible in the JB∗-algebra E2(r(b)). , with q Proof. The implications (i) ⇒ (ii) and (i) ⇒ (iii) follow from the above Proposition 2.3 (see [15, Proposition 4.4]). The implication (ii) ⇒ (i) is clear from the definition of extremal richness and the arbitrariness of β. (iii) ⇒ (ii) Fix a ∈ E and β > 0. By assumptions, there exists b ∈ E−1 such that P2(r(b))(a) + βr(b) ∈ E2(r(b)) is invertible in the JB∗-algebra E2(r(b)). Since r(b) is an extreme point of E and P2(r(b))(a + βr(b)) = P2(r(b))(a) + βr(b)) is invertible in the JB∗-algebra E2(r(b)), it follows from [15, Lemma 2.2] that a+βr(b) ∈ E−1 , and clearly ka − (a + βr(b))k = β. (cid:3) q q We shall explore next the stability of the property of being extremally rich under ℓ∞-sums, ideals, and quotients. We recall that a (closed) subtriple I of a JB∗-triple E is said to be an ideal of E if {E, E, I} + {E, I, E} ⊆ I. It is known that I is an ideal if and only if {E, E, I} ⊆ I or {E, I, E} ⊆ I (compare [6]). Theorem 2.5. Every quotient of an extremally rich JB∗-triple is extremally rich. Let (Ej) be a family of JB∗-triples. Then an element a = (aj) ∈ E is BP quasi-invertible if, and only if, aj is BP quasi-invertible in Ej for every j. Consequently, the ℓ∞-sum E = ⊕∞ j Ej is an extremally rich JB∗-triple if and only if each Ej is extremally rich. 8 F.B. JAMJOOM, A.M. PERALTA, A.A. SIDDIQUI, AND H.M. TAHLAWI Proof. Let E be an extremally rich JB∗-triple, and let J be a closed ideal of E. Let π : E → E/J, π(x) = x + J, denote the canonical projection of E onto E/J. Since π is a surjective triple homomorphism, it follows that π(E−1 q ) ⊆ (E/J)−1 [24, Theorem 3]). By hypothesis, there exist a (cf. q sequence (xn) in E−1 converging to x in the norm topology of E. Since π is q continuous, π(xn) → π(x) in norm, which proves that π(x) ∈ (E/J)−1 q hence (E/J)−1 q = E/J. , and Now, let (Ej)j∈I be an indexed family of JB∗-triples. We set E = ⊕∞ j Ej. Suppose a = (aj) ∈ E is BP quasi-invertible. Since, for each j0, the canoni- cal projection πj0 : ⊕∞ j∈IEj → Ej0 is a surjective triple homomorphism, we deduce that aj0 = πj0(a) ∈ (Ej0)−1 . Suppose now, that aj is BP quasi- invertible in Ej for every j. Consider bj ∈ Ej satisfying B(aj, bj) = 0 on Ej. Then B((aj), (bj)) = 0 on E, which shows that a = (aj) is BP quasi- invertible in E. The final statement follows from the above. (cid:3) q Theorem 3.6 in [15] establishes that for every JB∗-triple E with ∂e(E1) 6= ∅, the inequalities 1 + kak ≥ dist(a, ∂e(E1)) ≥ max {1 + αq(a), kak − 1} , hold for every a in E\E−1 . Under the additional hypothesis of E being extremally rich we can obtain an optimal computation of the distance from an element in E to the set ∂e(E1) of extreme points of E1. q Theorem 2.6. Let E be an extremally rich JB∗-triple and x ∈ E\E−1 , then dist(x, ∂e(E1)) = max{1, kxk − 1}. In particular, if x ∈ E1 then dist(x, ∂e(E1)) = 1. Consequently, for each x in E we have q dist(x, ∂e(E1)) =   max {1 − mq(x), kxk − 1} , if x ∈ E−1 q max {1, kxk − 1} , if x /∈ E−1 q ; , where mq(x) = dist(x, E\E−1 q ). Proof. Since the JB∗-triple E is extremally rich, αq(x) = dist(x, E−1 for all x ∈ E. Theorem 3.6 in [15] implies that q ) = 0 dist(x, ∂e(E1) ≥ max{1 + αq(x), kxk − 1} = max{1, kxk − 1} = 1. Applying [24, Theorem 27] we obtain dist(x, ∂e(E1) ≤ max{1, kxk − 1} for all x ∈ E−1 q = E. Combining the above inequalities we have dist(x, ∂e(E1) = max{1, kxk − 1}, . The second statement of the theorem follows immediately for all x ∈ E\E−1 when kxk ≤ 1. q The final statement follows form the first estimation and from [15, The- (cid:3) orem 3.1 and Proposition 3.2]. QUADRATIC CONORM AND EXTREMALLY RICH JB∗-TRIPLES 9 We have already commented that from the geometric point of view of Banach space theory, a C∗-algebra is extremally rich if and only if it has [4, §3] and [5, Theorem 3.7]). We do not the (uniform) λ-property (cf. know if this statement remains true for JB∗-triples. We know that every JBW∗-triple satisfies the uniform λ-property (cf. [15]). For an element a in a JB∗-triple E we further known that λ(a) = 1+mq(a) , whenever a is BP quasi-invertible element in E1 ([15, Theorem 3.4]). If we also assume that ∂e(E1) 6= ∅, the inequalities 2 1 + kak ≥ dist(a, ∂e(E1)) ≥ max {1 + αq(a), kak − 1} , hold for every a in E\E−1 q ([15, Theorem 3.6]), and hence (2.1) λ(a) ≤ 1 2 (1 − αq(a)), for every a ∈ E1\E−1 values bigger or equal than 1 JB∗-triple. q . We shall prove now that the λ-function takes only 2 on the open unit ball of an extremally rich Corollary 2.7. Let a be an element in the open unit ball of an extremally rich JB∗-triple. Suppose a is not BP quasi-invertible. Then λ(a) = 1 2 . Proof. Let us pick a real number t < 1 and a ∈ E\E−1 , we deduce, via Theorem 2.6, that 2 . Clearly β = 1 t > 2. Since kak < 1 q dist(βa, ∂e(E1)) = max{1, βkak − 1} < β − 1. Therefore, there exists e ∈ ∂e(E1) satisfying kβa − ek < β − 1. The element y = 1 β−1 (βa − e) lies in the open unit ball of E, and we can write βa = e + (β − 1)y, and hence a = 1 β y = te + (1 − t)y, which proves that λ(a) ≥ t. The arbitrariness of t shows that 1 2 ≤ λ(a). The final statement follows from [15, Theorem 3.6]. (cid:3) β e + β−1 Remark 2.8. Let E be a JB∗-triple satisfying the uniform λ-property of Aron and Lohman with 1 2 ≤ inf{λ(x) : x ∈ E1}. We can assure, via (2.1), that αq(a) = 0, for every a ∈ E1\E−1 . This shows that E is extremally rich. q In [25, §4] the authors introduce the so-called Λ-condition in JB∗-triples. A JB∗-triple E satisfies the Λ-condition if for every v ∈ ∂e(E1), and every y ∈ (E2(v))1\E−1 q with λ(y) = 0 we have αq(y) = 1. We shall consider the following stronger variant: A JB∗-triple E satisfies the strong-Λ-condition q with λ(y) = 0 we have αq(y) = 1. Every C∗-algebra if for each y ∈ E1\E−1 A satisfies that λ(a) = 1 (cf. [5, Theorem 3.7]). Therefore every C∗-algebra fulfills the strong-Λ-condition. A similar identity and statement is also valid for every JBW∗-triple ([15, Theorem 4.2]). 2 (1 − αq(a)) for every a ∈ A1\A−1 q Clearly, if E satisfies the strong-Λ-condition, then λ(a) > 0 for every q with αq(a) < 1. The following result is a consequence of this a ∈ E1\E−1 10 F.B. JAMJOOM, A.M. PERALTA, A.A. SIDDIQUI, AND H.M. TAHLAWI fact, Corollary 2.7, and the comments preceding it (cf. and 3.6]). [15, Theorems 3.4 Corollary 2.9. Every extremally rich JB∗-triple satisfying the strong-Λ- condition satisfies the Λ-property of Aron and Lohman. (cid:3) 3. Quadratic conorm in extremally rich JB∗-triples As mentioned in the introduction, the quadratic conorm, γq(a), of an element a in a JB∗-triple E is defined as the reduced minimum modulus of the conjugate linear operator Q(a) (see [7, Definition 3.1]), that is, γq(a) := γ(Q(a)) = inf{kQ(a)(x)k : dist(x, ker(Q(a))) ≥ 1}. The authors in [7] established that γq(a) = 1 ka†k2 , whenever a is von Neu- mann regular (where a† is the unique generalized inverse of a), and γq(a) = 0 otherwise (see [7, Theorem 3.4 and its proof]). Theorem 8 in [24] asserts that the set E−1 of all BP quasi-invertible elements in a JB∗-triple E is open in the norm topology. A more explicit measure of this fact is given in the next result. q Lemma 3.1. Let a be a BP quasi-invertible element in a JB∗-triple E. Suppose b is an element in E satisfying ka − bk < γq(a) 2 . Then b ∈ E−1 . 1 q Proof. We recall that a being BP quasi-invertible implies that e = r(a) ∈ ∂e(E1), and a is positive and invertible in the JB∗-algebra (E2(e), ◦e, ∗e). We further know that its generalized inverse a† ∈ E2(e) coincides with its inverse in this JB∗-algebra. 1 Let c = a 2 denote the square root of a in E2(e). We observe that a is positive and invertible in E2(e). Moreover, the inverse of a coincides with (a E2(e). 2 )−1, 2 , where the latter is the square root of a† in 2 )† = (a†) 2 , (a 1 1 1 1 1 2 Since ka − bk < γq(a) 1 2 , then e − Q(c†)(P2(e)(b))(cid:13)(cid:13)(cid:13) (cid:13)(cid:13)(cid:13) Q(c†)(a − b)(cid:13)(cid:13)(cid:13) = (cid:13)(cid:13)(cid:13) ≤ kQ(c†)kka − bk < kc†k2(γq(a)) 1 2 = ka†k γq(a) 1 2 = (see [7, Theorem 3.4 and its proof]) = 1. Since e is the unit of E2(e), we deduce that Q(c†)(P2(e)(b)) is invertible in E2(e). It is well known from the theory of invertible elements in JB∗-algebras that Q(c†)E2(e) : E2(e) → E2(e) is invertible as a mapping from E2(e) into itself with inverse Q(c)E2(e) : E2(e) → E2(e) (cf. [8, §4.1.1]). Since Q(c†)(P2(e)(b)) is invertible, we deduce that P2(e)(b) = Q(c)Q(c†)(P2(e)(b)) is invertible in E2(e) (cf. [8, Theorem 4.1.3]). Finally, Lemma 2.2 in [15] implies that b ∈ E−1 (cid:3) , as we desired. q QUADRATIC CONORM AND EXTREMALLY RICH JB∗-TRIPLES 11 Proposition 3.2. Let a be a BP quasi-invertible element in a JB∗-triple E. Suppose b is an element in E such that ka − bk < γq(a) 2 . Then b ∈ E−1 and q 1 kγq(a) − γq(b)k < γq(a) 2 ka − bk. 1 Proof. Lemma 3.1 implies that b is BP quasi-invertible. It is known that γq(x) = mq(x)2 ≤ kxk2, and mq(x) − mq(y) ≤ kx − yk for every x, y ∈ E (cf. [15, §3]). Therefore, γq(a) − γq(b) = mq(a)2 − mq(b)2 = mq(a) − mq(b) mq(a) + mq(b) < ka − bk(kak + kbk) < γq(a) 1 2 ka − bk. (cid:3) It is proved in [7, Theorem 3.13] that the quadratic conorm, γq(.), in a JB∗-triple E is upper semi-continuous on E\{0}. In the setting of extremally rich JB∗-triples we can characterize now the precise points at which γq(.) is continuous. Theorem 3.3. Let E be an extremally rich JB∗-triple. Then the quadratic conorm γq(.) is continuous at a point a ∈ E if, and only if, either a is not von Neumann regular (i.e. γq(a) = 0) or a is BP quasi-invertible. Proof. The upper semi-continuity of γq(.) implies that it is continuous at every point a ∈ E which is not von Neumann regular. , the continuity of γq(.) at a follows from Proposition 3.2. If a ∈ E−1 q Suppose that γq(.) is continuous at a, and a is von Neumann regular (i.e. γq(a) > 0). In this case, Q(a)(E) is norm closed, or equivalently, γq(a) = γ(Q(a)) > 0 (see [7, Corollary 2.4 and proof of Theorem 3.4]). The mapping x 7→ γq(x) 2 is continuous at a. So, there exists δ > 0 such that 1 ka − bk < δ ⇒ γq(a) 1 2 − γq(b) 1 2 < γq(a) 1 2 2 , that is, γq(b) 1 2 1 2 > γ q(a) 2 implies that E−1 q = E. Thus, there is c ∈ E−1 q with ka−ck < min{δ, γ q(a) 2 1 2 }. , whenever ka − bk < δ. Extremally richness of E, In particular ka − ck < δ, that is, γq(c) above proves that a ∈ E−1 . q 1 2 > γ q(a) 1 2 2 > ka − ck. Lemma 3.1 (cid:3) Remark 3.4. In [7, Remark 3.18] it is shown that the quadratic conorm γq(.) of a JB∗-triple E is continuous at every element a ∈ E for which Q(a) is left or right invertible in B(E). It is also asked, in the just quoted re- mark, whether these points are the only non-trivial continuity points of γq(.). Theorem 3.3 characterizes the continuity points of the quadratic conorm in the class of extremally rich JB∗-triples (a class including all JBW∗-triples). Proposition 3.2 shows the existence of points x satisfying that the quadratic 12 F.B. JAMJOOM, A.M. PERALTA, A.A. SIDDIQUI, AND H.M. TAHLAWI conorm is continuous at x, but Q(x) is not left nor right invertible. For ex- ample, when E is an extremally rich JB∗-triple and e is a complete tripotent with E1(e) 6= {0}, then the quadratic conorm is continuous at e, but Q(e) is not left nor right invertible. The arguments in the second part of the proof of Theorem 3.3 are also valid to prove the following: Proposition 3.5. Let (an) be a sequence of BP quasi-invertible elements in a JB∗-triple E. Suppose that (an) converges in norm to some element a in E, and γq(an) → γq(a) > 0. Then a is BP quasi-invertible. (cid:3) Our next result is a consequence of [7, Theorem 3.16 and Corollary 3.17] and the previous Proposition 3.5. n) is a bounded sequence in E; Corollary 3.6. Let (an) be a sequence of BP quasi-invertible elements in a JB∗-triple E. Suppose that (an) converges in norm to some element a in E. Then the following assertions are equivalent: (a) (a† (b) γq(an) → γq(a) > 0; Furthermore, if any of the above statements holds, then a is BP quasi- invertible and ka† n − a†k → 0. Proof. (a) ⇒ (b) Suppose (a† n) is a bounded sequence in E. Corollary 3.17 in [7] implies that a is von Neumann regular (i.e. γq(a) > 0). It follows from [7, Theorem 3.16](d) ⇒ (c) that γq(an) → γq(a) > 0. (b) ⇒ (a) Suppose γq(an) → γq(a) > 0. In particular, a is von Neumann regular. The desired statement follows from [7, Theorem 3.16](c) ⇒ (d). The final statement is a consequence of Proposition 3.5 and [7, Theorem (cid:3) 3.16]. The result in Theorem 3.3 is new, even in the case of C∗-algebras. R. Harte and M. Mbekhta introduce the notions of left and right conorms for C∗-algebras in [14]. According to their notation, the left conorm, γ(a), of an element a in a C∗-algebra A is given by γ(a) = γlef t(a) = γ(La) = inf (cid:26) kaxk d(x, ker(La)) : x 6∈ ker(La)(cid:27) , where La is the left multiplication mapping by a, that is, La(x) = ax (x ∈ A). The right conorm is similarly defined. Theorem 4 in [14] shows that γ(a)2 = γ(aa∗) = γ(a∗a) = γright(a)2 = inf{t : t ∈ σ(aa∗)\{0}}, where σ(aa∗) denotes the spectrum of aa∗. Harte and Mbekhta established that the conorm γ(.) of a C∗-algebra is upper semi-continuous [14, Theorem 7], they also show in [14, Theorem 9] that the conorm is always continuous at semi-invertible elements, and, by the upper semi-continuity of γ(.), the conorm is continuous at elements with QUADRATIC CONORM AND EXTREMALLY RICH JB∗-TRIPLES 13 no generalized inverses (i.e. at elements a with γ(a) = 0). When A = B(H), the C∗-algebra of all bounded linear operators on a complex Hilbert space H, then these results cover all continuity points. The general case is left as open problem. For a general C∗-algebra A, Corollary 4.1 in [7] proves that γq(a) = γ(a)2, for all a ∈ A. Theorem 3.3 particularizes in the following result, which provides an additional information to the problem left open by Harte and Mbekhta. Corollary 3.7. Let A be an extremally rich C∗-algebra. Then the conorm of A is continuous at a point a ∈ A if, and only if, either a is not von Neumann regular (i.e. γ(a) = 0) or a is quasi-invertible. (cid:3) References [1] R.M. Aron, R. H. Lohman, A geometric function determined by extreme points of the unit ball of a normed space, Pac. J. Math., 127, 209-231 (1987). [2] T.J. Barton, R.M. Timoney, Weak∗-continuity of Jordan triple products and its ap- plications, Math. Scand. 59, 177-191 (1986). [3] R.B. Braun, W. Kaup, H. Upmeier, A holomorphic characterization of Jordan-C∗- algebras, Math. Z. 161, 277-290 (1978). [4] L.G. Brown, G.K. Pedersen, On the geometry of the unit ball of a C∗-algebra, J. Reine Angew. Math., 469, 113-147 (1995). [5] L.G. Brown, G.K. Pedersen, Approximation and convex decomposition by extremals in C∗-algebras, Math. Scand., 81, 69-85 (1997). [6] L.J. Bunce, Structure of representations and ideals of homogeneous type in Jordan algebras, Quart. J. Math. Oxford Ser. (2) 37, no. 145, 1-10 (1986). [7] M. Burgos, A. Kaidi, A. Morales Campoy, A.M. Peralta, M. Ram´ırez, Von Neumann regularity and quadratic conorms in JB∗-triples and C∗-algebras, Acta Math. Sinica, vol. 24, no. 2, 185-200 (2008). [8] M. Cabrera Garc´ıa, A. Rodr´ıguez Palacios, Non-Associative Normed Algebras, Vol- ume 1. The Vidav-Palmer and Gelfand-Naimark Theorems, Part of Encyclopedia of Mathematics and its Applications, Cambridge University Press 2014. [9] Ch.-H. Chu, Jordan Structures in Geometry and Analysis, Cambridge Tracts in Math. 190, Cambridge. Univ. Press, Cambridge, 2012. [10] S. Dineen, The second dual of a JB∗-triple system, n: J. Mujica (Ed.), Complex analysis, Functional Analysis and Approximation Theory, North-Holland, Amster- dam, 1986. [11] C.M. Edwards, F.J. Fern´andez-Polo, C.S. Hoskin, A.M. Peralta, On the facial struc- ture of the unit ball in a JB∗-triple, J. Reine Angew. Math. 641, 123-144 (2010). [12] Y. Friedman, B. Russo, Structure of the predual of a JBW∗-triple, J. Reine Angew. Math. 356, 67-89 (1985). [13] L.A. Harris, Bounded symmetric homogeneous domains in infinite dimensional spaces. In: Proceedings on infinite dimensional Holomorphy (Kentucky 1973); pp. 13-40. Lecture Notes in Math. 364. Berlin-Heidelberg-New York: Springer 1974. [14] R. Harte, M. Mbekhta, Generalized inverses in C∗-algebras II, Studia Math. 106, no. 2, 129-138 (1993). [15] F.B. Jamjoom, A.M. Peralta, A.A. Siddiqui, H.M. Tahlawi, Approximation and convex decomposition by extremals, Quart. J. Math. (Oxford) 00, 1-21 (2015). doi:10.1093/qmath/hau036 [16] R. V. Kadisonn, Isometries of operator algebras, Ann. of Math., vol. 54. no. 2, 325-338 (1951). 14 F.B. JAMJOOM, A.M. PERALTA, A.A. SIDDIQUI, AND H.M. TAHLAWI [17] T. Kato, Perturbation theory for linear operators,Springer-Verlag New York, Inc., New York 1966. [18] W. Kaup, H. Upmeier, Jordan algebras and symmetric Siegel domains in Banach spaces, Math. Z. 157, 179-200 (1977). [19] W. Kaup, A Riemann mapping theorem for bounded symmetric domains in complex Banach spaces, Math. Z. 183, 503-529 (1983). [20] W. Kaup, On spectral and singular values in JB∗-triples, Proc. Roy. Irish Acad. Sect. A 96, no. 1, 95-103 (1996). [21] M.A. Rieffel, Dimension and stable rank in the K-theory of C∗-algebras, Proc. London Math. Soc. (3) 46, no. 2, 301-333 (1983). [22] A.A. Siddiqui, JB∗-algebras of topological stable rank 1, Int. J. Math. and Mathe- matical Sci., (2007), Article ID 37186, 24 pages, 2007. doi:10.1155/2007/37186 [23] H.M. Tahlawi, A.A. Siddiqui, Analogue of Brown-Pedersen' quasi invertibility for JB∗-triples, Int. Journal of Math. Analysis, 30, (5), 1469-1476 (2011). [24] H. M. Tahlawi, A. A. Siddiqui and F. B. Jamjoom, On the geometry of the unit ball of a JB∗-triple, Abstract and Applied Analysis, vol. 2013, Article ID 891249, 8 pages, 2013. doi:10.1155/2013/891249 [25] H.M. Tahlawi, A.A. Siddiqui, F.B. Jamjoom, The λ-function in JB∗-triples, J. Math. Anal. Appl., 414, 734-741 (2014). [26] H. Upmeier, Symmetric Banach Manifolds and Jordan C*-algebras, Elsevier Science Publishers B.V., 1985. Department of Mathematics, College of Science, King Saud University, P.O.Box 2455-5, Riyadh-11451, Kingdom of Saudi Arabia. E-mail address: [email protected] Departamento de An´alisis Matem´atico, Universidad de Granada,, Facultad de Ciencias 18071, Granada, Spain Current address: Visiting Professor at Department of Mathematics, College of Science, King Saud University, P.O.Box 2455-5, Riyadh-11451, Kingdom of Saudi Arabia. E-mail address: [email protected] Department of Mathematics, College of Science, King Saud University, P.O.Box 2455-5, Riyadh-11451, Kingdom of Saudi Arabia. E-mail address: [email protected] Department of Mathematics, College of Science, King Saud University, P.O.Box 2455-5, Riyadh-11451, Kingdom of Saudi Arabia. E-mail address: [email protected]
1510.07534	2	1510	2016-08-25T18:20:04	Topological construction of $C^$-correspondences for groupoid $C^$-algebras	[ "math.OA", "math.FA" ]	Let $(G,\alpha)$ and $(H,\beta)$ be locally compact groupoids with Haar systems. We define a topological correspondence from $(G,\alpha)$ to $(H,\beta)$ to be a $G$-$H$-bispace $X$ on which $H$ acts properly and $X$ carries a continuous family of measures which is $H$-invariant and each measure in the family is $G$-quasi invariant. We show that a topological correspondence produces a $C^$-correspondence from $C^(G,\alpha)$ to $C^*(H,\beta)$. We give many examples of topological correspondences.	math.OA	math	TOPOLOGICAL CONSTRUCTION OF C∗-CORRESPONDENCES FOR GROUPOID C∗-ALGEBRAS ROHIT DILIP HOLKAR Abstract. Let (G, α) and (H, β) be locally compact groupoids with Haar systems. We define a topological correspondence from (G, α) to (H, β) to be a G-H-bispace X on which H acts properly, and X carries a continuous family of measures which is H-invariant and each measure in the family is G-quasi invariant. We show that a topological correspondence produces a C∗-correspondence from C∗(G, α) to C∗(H, β). We give many examples of topological correspondences. Contents Introduction 1. Preliminaries 1.1. Groupoids 1.2. Proper actions and families of measures 1.3. Cohomology for groupoids 1.4. Quasi-invariant measures 1.5. C∗-correspondences 2. Topological correspondences 2.1. The right action-construction of the Hilbert module 2.2. The left action and construction of the C∗-correspondence 3. Examples References 1 3 3 4 6 8 9 9 13 18 19 24 Introduction A C∗-algebraic correspondence H from a C∗-algebra A to B is an A-B-bimodule which is a Hilbert B-module and A acts on H via the adjointable operators in a non-degenerate fashion. Let A = C∗(G, α) and B = C∗(H, β) where the ordered pairs (G, α) and (H, β) consist of a locally compact groupoid and a Haar system for it. Given a G-H-bispace X carrying an H-invariant family of measures such that each measure in the family is G-quasi-invariant, we show that if the H-action is proper, then Cc(X) can be completed into a C∗-correspondence from C∗(G, α) to C∗(H, β). This work is an extension of some part of my thesis [9], where we worked with Hausdorff typologies with certain countability assumptions. In the present work we get rid of the Hausdorffness and the countability hypotheses. Morita equivalence of C∗-algebras is defined by the existence of an imprimitivity bimodule, a special kind of C∗-correspondence. In the well-known result ([7]) that a Morita equivalence between two locally compact groupoids with Haar systems induces a Morita equivalence between the groupoid C∗-algebras the imprimitivity module is constructed directly from a bispace giving the Morita equivalence of the two groupoids. The Hausdorff case of Morita equivalence between locally compact groupoids is discussed in [7], and a much general and non-Hausdorff situation is 1 2 ROHIT DILIP HOLKAR studied in [12, Definition 5.3]. Which extra structure or conditions are needed for a bispace to give only a C∗-algebraic correspondence instead of a Morita equivalence? In general, we need a family of measures on the bispace as an extra structure to get started. In the Morita equivalence case, a family of measures on the bispace appears automatically (See Example 3.9). The the family of measures must be invariant for the right action and each measure of the family must be quasi-invariant for the left action. We also need that the right action is proper. We use the -category of a locally compact groupoids introduced in [5, 12] to prove that certain actions and a bilinear form are well-defined (Equations 2.8 and 2.9). The process of constructing a C∗-correspondence from a topological correspondence is divided into two main parts: the first part is to construct the Hil- bert module and the second one is to define the representation of the left groupoid C∗-algebra on this Hilbert module. For the first part, we use the representation the- ory of groupoids and the transverse measure theory introduced by Renault in [12]. In the second part, our motivation and techniques are derived from the theory of quasi-invariant measures for locally compact groups ([4, Section 2.6]). As a C∗-correspondence from a C∗-algebra A to B induce a representation of B to that of A, a topological correspondence from a locally compact groupoid with a Haar system (G, α) to (H, β) induce a representation of H to that of G. Renault proves this in [13]. A locally compact, Hausdorff space is a locally compact groupoid with a Haar system, and so is a locally compact group. A well-known fact about groupoid equivalence is that two spaces are equivalent if and only if they are homeomorphic and two groups are equivalent if and only if they are isomorphic. But since any continuous map between spaces gives a topological correspondence and so does a group homomorphism, a topological correspondence is far more general than an equivalence. We give many examples of topological correspondences, most of which are the analogues of the standard examples of C∗-correspondences. In Example 3.1, we show that a continuous map f : X → Y between spaces gives a topological correspondence from Y to X. Example 3.4 shows that a continuous group homomorphism φ : G → H gives a topological correspondences from G to H. Theorem 2.31 gives that the topological correspondences, the one from Y to X and the one from G to H, produce C∗-correspondences from C0(Y ) to C0(X) and C∗(G) to C∗(H), respectively. It is easy to see that the C∗-correspondence from C0(Y ) to C0(X) is exactly the one give by the -homomorphism f ∗ : C0(Y ) → M(C0(X)) as in the theory of commutative C∗-algebras. However, it is not equally easy to see that the C∗-correspondence given by the group homomorphism φ agrees with the one which is induced by the -homomorphism φ∗ : C∗(G) → C∗(H). map, then we get a topological correspondence from G to H. Example 3.5 shows that if the group homomorphism in Example 3.4 is a proper Let E0 and E1 be locally compact, Hausdorff and second countable spaces, and let s, r : E1 → E0 be continuous maps. Let λ = {λe}e∈E0 be a continuous family of measures along s. By applying the definition of a topological correspondence it is straightforward to check that s, r and λ give a topological correspondence from E0 to itself. Muhly and Tomforde ([8, Definition 3.1]) call this correspondence a topological quiver. They construct a C∗-correspondence associated to a topological quiver in [8, Section 3.1] and the construction in [8] is exactly the construction of a C∗-correspondence from a topological correspondence. Muhly and Tomforde define the C∗-algebra associated to a topological quiver ([8, Definition 3.17]) which includes a vast class of C∗-algebras: graph C∗-algebras, C∗-algebras of topological TOPOLOGICAL CORRESPONDENCES 3 graphs, C∗-algebras of branched coverings, C∗-algebras associated with topological relations are all associated to a topological quiver [8, Section 3.3]. Topological quivers justify our use of families of measures in the definition of a topological correspondence. At a first glance, the families of measures and their quasi-invariance for the left action might look artificial. However, as discussed on page 10, the quasi-invariance of families of measures is natural to ask for. We show that the notion of correspondences introduced by Macho Stadler and O'uchi ([15]), the generalised morphisms introduced by Buneci and Stachura ([3]) are topological correspondences. See Examples 3.7 and 3.11, respectively. In [16], Tu defines locally proper generalised homomorphism for locally compact groupoids with Hausdorff space of units. Example 3.7, which shows that a corres- pondence in the sense of Macho Stadler and O'uchi is a topological, also shows that a locally proper generalised homomorphism is a topological correspondence. However, Tu proves that a locally proper generalised homomorphism induces a C∗-correspondence between the reduced C∗-algebras of the groupoids ([16, Propos- ition 2.28]). Whereas our result involves the full C∗-algebras of the groupoids. We know that the result of Tu holds for topological correspondences when the bispace involved in the topological correspondence is Hausdorff and second countable, and the right action is amenable [9, Proposition 2.3.4]. Following is the sectionwise description of the contents. Section 1 (Preliminaries): In this section we mention our conventions, notation, and rewrite some standard definitions and results. A notion of cohomology for Borel groupoids is introduced in [17] by Westman. In [11, Chapter 1, page 14], Renault discusses a continuous version of the same cohomology. We need a groupoid equivariant continuous version of this cohomology. For this purpose we define action of a groupoid on another groupoid and then define the equivariant cohomology for Borel and continuous groupoids. Section 2 (Topological correspondences): This section contains the main construc- tion. Immediately after the definition of topological correspondence (Definition 2.1) we discuss the role of the adjoining function. Let (X, λ) be a topological correspondence from (G, α) to (H, β). Then we write the formulae of the actions of Cc(G) and Cc(H) on Cc(X). These actions make Cc(X) into a Cc(G)-Cc(H)-bimodule. We also define the formula of a Cc(H)-valued bilinear map on Cc(X) which, we latter prove, is a positive bilinear map. We com- plete this setup to get a C∗-correspondence. The process, as mentioned earlier, is divided into two parts: constructing a C∗(H, β)-Hilbert module H(X) and defining a representation of C∗(G, α) on this Hilbert module. We advise the reader to jump to Section 3 after the discussion that follows Definition 2.1 to have a look at some examples. Section 4 Examples: This section contains examples of topological correspondences. 1. Preliminaries 1.1. Groupoids. The reader should be familiar with the theory of locally compact groupoids ([1, 10–12] and [16]). A groupoid G is a small category in which every arrow is invertible. Except a few instances, we denote the space of arrows of groupoid G by the letter G itself, rather than the more precise symbol G(1). We denote the space of units of G by G(0). The source and range maps are denoted by sG and rG, respectively. The inverse of an element γ ∈ G is denoted by γ−1. There are instances when we need to write γ 7→ γ−1 as a function G → G and then we denote the function by invG. A pair (γ, γ′) ∈ G × G is called composable if sG(γ) = rG(γ′). Sometimes we abuse the language by saying 'γ, γ′ ∈ G are composable' by which we mean that 4 ROHIT DILIP HOLKAR the pair (γ, γ′) is composable. For A, B ⊆ G(0) define GA = {γ ∈ G : rG(γ) ∈ A} = r−1 GA = {γ ∈ G : sG(γ) ∈ A} = s−1 GA B = GA ∩ GB = {γ ∈ G : rG(γ) ∈ A and sG(γ) ∈ B}. G (A), G (A) and When A = {u} and B = {v} are singletons, we write Gu, Gv and Gu G{u}, G{v} and G{u} isotropy group at u. {v} , respectively. For u ∈ G(0), Gu v instead of u is a group. It is called the A topological groupoid and measurable groupoid have their standard meanings ([1]). We call a subset A ⊆ X of a topological space X quasi-compact if every open cover of A has a finite subcover. And A is called compact if it is quasi-compact and Hausdorff. The space X is called locally compact if every point x ∈ X has a compact neighbourhood. All the spaces considered in this article are locally compact. For a locally compact space X by Cc(X)0 we denote the set of functions f on X such that f ∈ Cc(V ) where V ⊆ X is open Hausdorff, and f is extended outside V by 0 (See [16, Section 4]). By Cc(X) we denote the linear span of functions in Cc(X)0. The functions in Cc(X) need not be continuous on X but they are Borel. Furthermore, if µ is a positive σ-finite Radon measure on X, then Cc(X) ⊆ L2(X, µ) is dense. As in [12], we call a topological groupoid G locally compact if G is a locally compact topological space and G(0) ⊆ G is Hausdorff. In this case r−1 G (u) ⊆ G (and equivalently, s−1 If X and Y are topological spaces, X ≈ Y means X and Y are homeomorphic. If G and H are group(oids)s, then G ≃ H means G and H are isomorphic via a group(oid) homomorphism. In group case, this means G and H are isomorphic. All the measures we work with are positive, Radon and σ-finite. G (u) ⊆ G) are Hausdorff for all u ∈ G(0). For groupoid actions we do not assume that the momentum maps are open or surjective. However, it is well-known that for a locally compact groupoid with a Haar system the source map (equivalently the range map) is automatically open. We need that each measure in a family of measures along a continuous open map f : X → Y is non-zero, but it need not have full support. 1.2. Proper actions and families of measures. Since we shall not come across any case where there are more than one different left (or right) action of a group- oid G on a space X, we denote the momentum map by rX (respectively, sX ). When we write 'X is a left (or right) G-space' without specifying the momentum map, the above convention will be tacitly assumed and then in such instances the momentum map is rX (respectively, sX ). and Yu similar to GA, Gu, GA and Gu. Let A ⊆ G(0). For a left G-space X and a right G-space Y we define X A, X u, YA Let X, Y and Z be spaces, and let f : X → Z and g : Y → Z be maps. We denote the fibre product of X and Y over Z by X ×f,Z,g Y . When there is no confusion about the maps f and g, we simply write X ×Z Y instead of X ×f,Z,g Y . Let G be a groupoid and X a left G-space. By G ⋉ X we denote the transforma- tion groupoid. Its space of arrows is G ×sG,G(0),rX X, which we prefer denoting by G ×G(0) X. Recall that a Haar system on G induces a Haar system on G ⋉ X. Let G be a locally compact groupoid with open range map and X a locally compact G-space. Then the quotient map X → X/G is open. If the action of G is proper, then X/G is locally compact. If X is Hausdorff (or Hausdorff and second countable) the X/G is also Hausdorff (Hausdorff and second countable, respectively). TOPOLOGICAL CORRESPONDENCES 5 Let X be a left G-space. For subsets K ⊆ G and A ⊆ X with sG(K)∩rX (A) 6= ∅, define KA = {γx : γ ∈ K, x ∈ A and (γ, x) ∈ G ×G(0) X}. If sG(K) ∩ rX (A) = ∅, then we define AK = ∅. By an abuse of notation, for x ∈ X we write Kx instead of K{x}. The meaning of γA for γ ∈ G is similar. For a right action, we define AK, xK and Aγ similarly. Let G and H be groupoids and X be a space on which G acts from the left and H from the right. If for all γ ∈ G , x ∈ X and η ∈ H with sG(γ) = rX (x) and sX (x) = rH (η) we have sX (γx) = sX (x), rX (xη) = rX (x), and then we call X a G-H-bispace. (γx)η = γ(xη), λyη = λyη holds; on Y is continuous. Definition 1.1 (Invariant continuous family of measures). Let H be a locally compact groupoid, and let X and Y be locally compact right H-spaces. Let π : X → Y be an H-equivariant continuous map. An H-invariant continuous family of measures along π is a family of Radon measures λ = {λy}y∈Y such that: i) each λy is defined on π−1(y); ii) (invariance) for all composable pairs (y, η) ∈ Y ×H(0) H, the condition iii) (continuity condition) for f ∈ Cc(X) the function Λ(f )(y) :=Rπ−1(y) f dλy We clarify that in the above definition the measure λyη is given by R f dλyη = R f (xη) dλy(x) for f ∈ Cc(X). If for each y ∈ Y , supp(λy) = π−1(y), we say the family of measures λ has full support. If there is a continuous function f on X with Λ(f ) = 1 on π(X), we say that λ is proper. Lemma 1.1.2 in [1] says that, in the continuous case, λ is proper if and only if λy 6= 0 for all y ∈ Y . Thus if λ is continuous and has full support, then λ is proper. In the rest of the article we assume that every family of measures we work with is proper. Let Pt be the trivial point group(oid). If X and Y are spaces and π : X → Y is a continuous map, then π is a Pt-equivariant map between Pt-spaces. A continuous Pt-invariant family of measures along π is simply called a continuous family of measures along π. We denote families of measures by small Greek letters. For a given family of measures, the corresponding integration function that appears in the continuity condition in Definition 1.1 will be denoted by the Greek upper case letter used to denote the family of measures. For α, β and µ it will be A, B and M , respectively. Definition 1.2. (1) Let H be a groupoid, X a left H-space. An H-invariant continuous family of measures along the momentum map rX is called a left H-invariant continuous family of measures on X. A right H-invariant continuous family of measures on X is defined analogously. (2) For a groupoid H, a Haar system on H is a left H-invariant continuous family of measures with full support on H for the left multiplication action of H on itself. Lemma 1.3. Let H be a groupoid, let π : X → Y be a continuous H-map between the H-spaces X and Y and let λ be a continuous family of measures along π. If λy has full support for all y ∈ π(X), then π is an open map onto its image. Proof. Consider the map π : X → π(X) and then the proof is similar to the one of Proposition 2.2.1 in [10]. (cid:3) Corollary 1.4. If (H, β) is a groupoid with a Haar system, then the range and source maps are open. 6 ROHIT DILIP HOLKAR Proof. Lemma 1.3 implies that the range map rH is open. Since sH = rH ◦ invH and invH is a homeomorphism, sH is open. (cid:3) We need the following two lemmas from [12]. Lemma 1.5 (Lemme 1.1, [12]). Let X and Y be spaces, let π : X → Y be an open surjection and let λ be a family of measures with full support along π. For every open U ⊆ X and for a non-negative function g ∈ Cc(π(U )), there is a non-negative function f ∈ Cc(U ) with Λ(f ) = g. Lemma 1.6 (Lemme 1.2, [12]). Let X, Y and Z be spaces, let π and τ be open surjections from X and Y to Z, respectively. Let π2 denote the projection from the fibre product X×Z Y onto the second factor Y . Assume that for each z ∈ Z, there is a measure λz on π−1(z). For each y ∈ Y define the measure λ2y = λτ (y)×δy, where δy is the point-mass at y. Then λ is continuous if and only if λ2 is continuous. Let H be a locally compact groupoid, and let X and Y be locally compact right H-spaces. For x ∈ X the equivalence class of x in X/H is denoted by [x]. The map π induces a map from X/H to Y /H, which we denote by [π]. Let π : X → Y be an H-equivariant map and λ an H-invariant continuous family of measures along π. Then λ induces a continuous family of measures [λ] along [π] : X/H → Y /H. The proof is similar to the proof of Proposition 2.15 in [9]. For f ∈ Cc(X/H) and [y] ∈ Y /H the measure [λ][y] is defined as f d[λ][y] :=ZX (1.7) ZX/H f ([x]) dλy(x). Example 1.8. Let X be a proper right H-space. Let β be a Haar system on H. Then Lemma 1.6 says that β1 = {δx×βsX (x)}x∈X is a continuous family of measures along the projection map π1 : X ×H(0) H → X. Take the quotient by the action of H to get a continuous family of measure [π1] along the map [π1] : (X×H(0) H)/H → X/H. Identifying (X ×H(0) H)/H = X gives that [π1] is the quotient map, and a small computation gives that for f ∈ Cc(X) and for [x] ∈ X/H, ZX f d[β1][x] =ZHsX (x) f (xη) dβsX (x)(η). In the rest of the article we write βX instead of [β1]. 1.3. Cohomology for groupoids. In this section, the groupoids and maps are assumed to be Borel. The whole discussion goes through when the Borel properties are replaced by the continuous properties. Definition 1.9 (Action of a groupoid on another groupoid). A left action of a groupoid G on another groupoid H is given by maps rH,G : H → G(0) and a : G ×sG,G(0),rH,G H → H which satisfy the following conditions: i) if η, η′ ∈ H are composable, γ ∈ G with sG(γ) = rH,G(η) = rH,G(η′), then a(γ, η), a(γ, η′) ∈ H are composable and a(γ, η)a(γ, η′) = a(γ, ηη′); ii) if u ∈ G(0), then a(u, η) = η for all η ∈ r−1 iii) if γ, γ′ ∈ G are composable, then (γ, a(γ′, η)) ∈ G ×sG,G(0),rH,G H and H,G(u) ⊆ H; a(γγ′, η) = a(γ, a(γ′, η)). To simplify the notation, we write γ · η or simply γη for a(γ, η), and G ×G(0) H for G ×sG,G(0),rH,G H. Then (i) and (ii) above read γ · (ηη′) = (γ · η)(γ · η′) and (γγ′) · η = γ · (γ′ · η), respectively. We call the map rH,G the momentum map TOPOLOGICAL CORRESPONDENCES 7 for the action and a the action map. When the momentum and action maps are continuous (or Borel) the action is called continuous (or Borel, respectively). It is not hard to see that Definition 1.9 gives an action of G on H via invertible functors. When G is a group, our definition matches Definition 1.7 in [11, Chapter 1], which is the action of a group on a groupoid by invertible functors. A proof of this fact is below. Lemma 1.10. When G is a group, an action of G on H as in Definition 1.9 above is the same as the action in [11, Definition 1.7, Chapter 1], that is, there is a homomorphism φ : G → Aut(H) which gives the action. Here Aut(H) is the set of all invertible functors from H to itself. Proof. Definition 1.9 implies [11, Definition 1.7, Chapter 1]: For γ ∈ G define φ(γ)(η) = γ · η. We first prove that each φ(γ) is a functor from H to itself. Note that an element u in a groupoid is a unit if and only if u is composable with itself and u2 = u. If u ∈ G(0), then φ(γ)(u) = φ(γ)(uu) = φ(γ)(u)φ(γ)(u) = (φ(γ)(u))2. Hence for each unit u ∈ H (0), φ(γ)(u) ∈ H is a unit. Condition (i) of Definition 1.9 gives that for each γ ∈ G, φ(γ)(ηη′) = φ(γ)(η)φ(γ)(η′). This proves that φ(γ) is functor for each γ ∈ G. Now we show that each of the φ(γ) is invertible. Condition (iii) of Definition 1.9 gives that γ 7→ φ(γ) is a homomorphism. Use (iii) of Definition 1.9 to see that φ(γ) is invertible: φ(γ)φ(γ−1)(η) = φ(γγ−1)(η) = φ(rG(γ))(η) = η = IdH (η). Similarly, φ(γ−1)φ(γ) = IdH . Thus φ(γ)−1 = φ(γ−1). Hence φ(γ) ∈ Aut(H). [11, Definition 1.7, Chapter 1] implies Definition 1.9 implies: Proof of this part (cid:3) is routine checking of the conditions in Definition 1.9. A continuous (and Borel) version of Lemma 1.10 can be proved along same lines merely by adding continuity (or Borelness) of the action map and the momentum map and the continuity (Borelness) of the group homomorphism φ. Example 1.11. Let G be a groupoid and H a space. Then an action of G on H is the same as an action of G on H viewed as a groupoid. In this case, condition (i) in Definition 1.9 is irrelevant and then the definition matches the usual one. Example 1.12. Let G and H be groupoids and let X be a G-H-bispace. Define an action of G on the transformation groupoid X ⋊ H by γ(x, η) := (γx, η). The momentum map for this action is (x, η) 7→ rX (x) ∈ G(0). Let (xη, η′), (x, η) ∈ X ×H(0) H be composable elements then (γxη, η′)(γx, η) are composable and γ · (x, η) · γ (xη, η′) = (γx, η)(γxη, η′) = (γx, ηη′) = γ(x, ηη′). This verifies (i) of Definition 1.9. The other conditions are easy to check. Thus H acts on the groupoid G ⋉ X in our sense. This is an important example for us. Let H be a Borel groupoid and assume that H acts on a Borel groupoid G. Let G(0) and G(1) have the usual meaning. For n = 2, 3, . . . define G(n) = {(γ0, . . . , γn−1) ∈ G × G × ··· × G } {z n-times : sG(γi) = rG(γi+1) for 0 ≤ i < n − 2}. Definition 1.13. Let G, H be Borel groupoids, let A be an abelian Borel group and H (G; A), d•) let H act on G. The A-valued H-invariant Borel cochain complex (BC• is defined as follows: i) The abelian groups BCn H are: (a) BC0 H (G; A) := {f : G(0) → A : f is an H-invariant Borel map}; 8 ROHIT DILIP HOLKAR ii) the coboundary map d is: (b) for n > 0 BCn H (G; A) := {f : G(n) → A : f is an H-invariant Borel map and f (γ0, . . . , γn−1) = 0 if γi ∈ G(0) for some 0 ≤ i ≤ n − 1}, H (G; A) is d0(f )(γ) = f (sG(γ)) − f (rG(γ)), H (G; A) → BC1 (a) d0 : BC0 (b) for n > 0, dn : BCn H (G; A) → BCn+1 H (G; A) is dn(f )((γ0, . . . , γn)) = f (γ1, . . . , γn) + n Xi=1 (−1)if (γ0, . . . , γi−1γi, . . . , γn) + (−1)n+1f (γ0, . . . , γn−1). The n-th cohomology group of this complex is the n-th H-invariant Borel co- homology of G for n ≥ 0, and it is denoted by Hn Bor,H (G; A). By adding the action of H to all the maps and spaces, one can make sense of the machinery and the results in [17, §1 and §2]. Remark 1.14. Any H-invariant Borel function f on G(0) is a 0-cochain. A cochain H (G; A) is a cocycle iff d0(f ) = 0 which is true iff f is constant on the f ∈ BC0 orbits of G(0). A cochain k ∈ BC1 H (G; A) is a cocycle iff k(γ0)−k(γ0γ1)+k(γ1) = 0 for all composable γ0 and γ1, which is equivalent to saying that k is an H-invariant Borel groupoid homomorphism. We drop the suffixes B and Bor and write merely C0 H (G; A) and Hn H (G; A). H(G; A) with d0(b) = d0(b′). Then c = b − b′ is a Remark 1.15. Let b, b′ ∈ C0 0-cocycle since d0(c) = d0(b)− d0(b′) = 0. Remark 1.14 now gives that c is constant on the orbits of G(0). Thus c is a function on G(0)/G. 1.4. Quasi-invariant measures. Let (G, α) be a locally compact groupoid with a Haar system. Using α we get a right invariant family of measures α−1 along the u (γ) =R f (γ−1) dαu(γ). Let X be a left source map. For f ∈ Cc(G) setR f (γ) dα−1 G-space and let µ be a measure on X. We define a measure µ ◦ α−1 on the space G ×G(0) X by f d(µ ◦ α−1) =ZXZGrX (x) ZG× f (γ−1, x) dαrX (x)(γ) dµ(x) G(0) X for f ∈ Cc(G ×G(0) X). Similarly we define the measure µ ◦ α. Definition 1.16 (Quasi-invariant measure). Let (G, α) be a groupoid with a Haar system and X a G-space. A measure µ on X is called (G, α)-quasi-invariant if µ◦ α and (µ ◦ α) ◦ invG⋉X are equivalent. In the above definition, invG⋉X is the inverse function on the transformation groupoid G ⋉ X. Thus for f ∈ Cc(G ×G(0) X) (µ ◦ α) ◦ invG⋉X (f ) =ZXZGrX (x) f (γ, γ−1x) dαrX (x)(γ)dµ(x). When the groupoid (G, α) with a Haar measure in the discussion is fixed and there is no possibility of confusion, we write 'µ is a quasi-invariant measure' Remark 1.17. As in [1] or [3], it can be shown that the Radon-Nikodym derivative d(µ ◦ α)/d(µ ◦ α−1) is a µ ◦ α-almost everywhere a groupoid homomorphism from the transformation groupoid G ⋉ X to R∗ +. Remark 1.18. A (G, α)-quasi-invariant measure µ is G-invariant if and only if the Radon-Nikodym derivative d(µ ◦ α)/d(µ ◦ α−1) = 1 µ ◦ α-almost everywhere on G ×G(0) X. TOPOLOGICAL CORRESPONDENCES 9 A measured groupoid is a triple (G, α, µ) where (G, α) is a locally compact group- oid with a Haar system and µ is a (G, α)-quasi-invariant measure on G(0). Definition 1.19 (Modular function). The modular function of a measured group- oid (G, α, µ) is the Radon-Nikodym derivative d(µ ◦ α)/d(µ ◦ α−1). The reason to use the article the for the modular function is that if ∆ and ∆′ are two modular, then ∆ = ∆′ µ ◦ α-almost everywhere. 1.5. C∗-correspondences. We shall use the theory of Hilbert modules and assume that the reader is familiar with the basics of the theory (For example, [6]). Definition 1.20. Let A and B be C∗ algebras. A C∗-correspondence from A to B is a Hilbert B-module H with a non-degenerate -representation A → BB(H). Here BB(H) denotes the C∗-algebra of adjointable operators on H. A Hilbert B-module H is full if the linear span of the image of H × H under the inner product map is dense in B. We call a C∗-correspondence H from A to B a proper correspondence if A acts on H by compact operators, that is, the action of A is given by a non-degenerate -representation A → KB(H). Definition 1.21. An imprimitivity bimodule from A to B is an A-B-bimodule H such that i) H is a full left Hilbert A-module with an inner product ∗h,i; ii) H is a full right Hilbert B-module with an inner product h,i∗; iii) (H, ∗h,i) is a correspondence from B to A; iv) (H,h,i∗) is a correspondence from A to B; v) for a, b, c ∈ H ahb , ci∗ = ∗ha , bic. Our notion of C∗-correspondence (Definition 1.20) is wider, in the sense that many authors demand that the Hilbert module involved in a C∗-correspondence is full, or for some authors a C∗-correspondence is what we call a proper correspond- ence. In [14], Rieffel shows that an A-B-imprimitivity bimodule induces an isomorph- In general, if H is a ism between the representation categories of B and A. C∗-correspondence from A to B, then H induces a functor from Rep(B), the rep- resentation category of B, to Rep(A). 2. Topological correspondences Definition 2.1 (Topological correspondence). A topological correspondence from a locally compact groupoid G with a Haar system α to a locally compact groupoid H equipped with a Haar system β is a pair (X, λ) where: is an H-invariant proper continuous family of measures i) X is a locally compact G-H-bispace, ii) the action of H is proper, iii) λ = {λu}u∈H(0) iv) ∆ is a continuous function ∆ : G×G(0) X → R+ such that for each u ∈ H (0) ZXuZGrX (x) along the momentum map sX : X → H (0), and F ∈ Cc(G ×G(0) X), F (γ−1, x) dαrX (x)(γ) dλu(x) =ZXuZGrX (x) F (γ, γ−1x) ∆(γ, γ−1x) dαrX (x)(γ) dλu(x). 10 ROHIT DILIP HOLKAR If ∆′ is another function that satisfies condition (iv) in Definition 2.1, then ∆ = ∆′ λu ◦ α-almost everywhere for each u ∈ H (0). As both ∆ and ∆′ are continuous, we get ∆ = ∆′. We call the function ∆ the adjoining function of the correspondence (X, λ). Remark 2.2. Note that we do not need that the momentum maps sX and rX are open surjections. We also do not demand that the family of measures λ has full support, hence the Hilbert module in the resulting C∗-correspondence need not be full. The resulting C∗-correspondence need not be proper, as well. Remark 2.3. Referring to Definition 1.16, we can see that Condition (iv) in Defini- tion 2.1 says that the measure α×λu on G×G(0) Xu is (G, α)-quasi-invariant for each u ∈ H (0) where the measure α × λu is defined as follows: for f ∈ Cc(G ×G(0) Xu), ZG× G(0) Xu f d(α × λu) =ZXuZGrX (x) f (γ−1, x) dαrX (x)(γ) dλu(x). In short, "A topological correspondence from G to H is a pair (X, λ) where X is a G-H-bispace and λ is an H-invariant family of measures on X indexed by H (0) and each measure in λ is G-quasi-invariant." Remark 2.4. As in [1] or [3], it can be shown that ∆ restricted to G ×G(0) Xu is α × λu-almost everywhere a groupoid homomorphism for all u ∈ H (0). So the function ∆ in (iv) of Definition 2.1 is a continuous 1-cocycle on the groupoid G ⋉ X. We shall use this fact in many computations. Remark 2.5. Example 1.12 gives a right action of H on G ⋉ X. In [3], Buneci and Stachura use an adjoining function exactly like us. The topological correspondence Buneci and Stachura define is a special case of our construction (Example 3.11). They show that the adjoining function in their case is H-invariant (see [3, Lemma 11]). In the similar fashion we may prove that ∆ is H-invariant under the right action of H, that is, ∆(γ, xη) = ∆(γ, x) for all composable triples (γ, x, η) ∈ G ×sG,G(0),rX X ×sX ,H(0),rH H. Thus, in fact, ∆ : G ⋉ X/H → R∗ +. Now Remark 2.4 can be made finer by saying that ∆ is an H-invariant continuous 1-cocycle on the groupoid G ⋉ X. Use of the family of measures and the adjoining function: In the following discussion we explain the role of the adjoining function. Let (X, λ) be a topological corres- pondence from (G, α) to (H, β) with ∆ as the adjoining function. We make Cc(X) into a Cc(H)-module using the same formula as in [7] or [15]. To make Cc(X) into a C∗(H, β)-pre-Hilbert module, we need to define a Cc(H)-valued inner product on Cc(X). The formula for this inner product cannot be copied directly from either [7] or [15]. This formula has to be modified, and it uses the family of measures λ. Talking about the left action, for φ ∈ Cc(G) and f ∈ Cc(X) [7] and [15] define φ · f ∈ Cc(X) by (2.6) (φ · f )(x) =ZG φ(γ)f (γ−1x) dαrX (x)(γ). For our definition of topological correspondence, the action of Cc(G) on the C∗(H, β)-pre-Hilbert module Cc(X) defined by Formula 2.6 is not necessarily an action by adjointable operators. For φ and f as above we define the left action by (2.7) (φ · f )(x) :=ZG φ(γ)f (γ−1x) ∆1/2(γ, γ−1x) dαrX (x)(γ). TOPOLOGICAL CORRESPONDENCES 11 We shall see that the adjoining function gives a nice scaling factor for the action of Cc(G) ⊆ C∗(G, α) and makes this action a -homomorphism. This is the reason we call ∆ the adjoining function. Two examples of topological correspondences are: an equivalence between group- oids (See [7] or Definition 3.10) and the correspondence of Marta Stadler and O'uchi (See [15] or Example 3.7). For equivalences and Macho Stadler-O'uchi correspond- ences the adjoining function is the constant function 1, and then formulae (2.6) and (2.7) match. To understand the role of ∆ the reader may have a look at Lemma 2.30. To support the necessity of the adjoining function, consider a toy example: Let G be a locally compact group and let it act on a locally compact (possibly Haus- dorff) space X carrying a measure λ. The left multiplication action of Cc(G) on Cc(X) ⊆ L2(X, λ) defined by Equation (2.6) is not necessarily bounded. To make this action bounded, it is sufficient that λ is G-quasi-invariant, which brings the adjoining function into the picture. Then the left action of Cc(G) given by Equa- tion (2.7) becomes a -representation. This motivated us to introduce Condition (iv) in Definition 2.1. Buneci and Stachura [3] also use the adjoining function. For the left multiplication action of G on G/K, where K is a closed subgroup of G, the space G/K always carries a G-quasi-invariant measure λ. Hence there is a representation of G on L2(G/K, λ). Quasi-invariant measures and the correspond- ing adjoining functions are studied very well in the group case, for example, see Section 2.6 of [4]. At this point, readers may peep into Section 3 to see some examples of adjoining functions. We start with the main construction now. For φ ∈ Cc(G), f ∈ Cc(X) and ψ ∈ Cc(H) define functions φ · f and f · ψ on X as follows: (φ · f )(x) :=ZGrX (x) (f · ψ)(x) :=ZHsX (x)   (2.8) (2.9) φ(γ)f (γ−1x) ∆1/2(γ, γ−1x) dαrX (x)(γ), f (xη)ψ(η−1) dβsX (x)(η). For f, g ∈ Cc(X) define the function hf, gi on H by hf, gi(η) :=ZXrH (η) f (x)g(xη) dλrH (η)(x). Most of the times we write φf and f ψ instead of φ · f and f · ψ. Lemma 2.10. Among the functions φf , f ψ and hf , gi defined above, the first two are in Cc(X) and the last one is in Cc(H). Proof. The proof follows from [12, Lemme 3.1] and the -category Cc(H) used in [12] or [5]. Detail computations can be found in [9, Section 3.3.1] in which Tables 3.1 and 3.2 list the equivalence of the operations in Definitions 2.8 and 2.9 and those in the -category Cc(H). We prove that hf , gi ∈ Cc(H) and the remaining claims can be proved similarly. Let Cc(H) denote the -category for H. Then (H, β) and (X, λ) are objects in Cc(H). We identify X with (X ×H(0) H)/H and then think of f, g ∈ Cc((X ×H(0) H)/H) as arrows from (X, λ) to (H, β). Then hf , gi = f ∗ ∗λ g, where the latter is in Cc((H ×rH ,H(0),rH H)/H). Now identify Cc((H ×rH ,H(0),rH H)/H) with Cc(H) to conclude the proof. (cid:3) Both Cc(G) and Cc(H) are -algebras. Denote the convolution product on them by ∗. 12 ROHIT DILIP HOLKAR Lemma 2.11. Let φ, φ′ ∈ Cc(G), ψ, ψ′ ∈ Cc(H) and f, g, g′ ∈ Cc(X). Then (2.12) (2.13) (2.14) (2.15) (2.16) (2.17) (2.18) (φ ∗ φ′)f = φ(φ′f ), f (ψ ∗ ψ′) = (f ψ)ψ′, (φf )ψ = φ(f ψ), hf, g + g′i = hf, gi + hf, g′i, hf , gi∗ = hg , fi, hf , gi ∗ ψ = hf , gψi, hφf , gi = hf , φ∗gi. Proof. Checking most of the equalities above are straightforward computations in- volving obvious change of variables, using the invariances of the families of measures, Fubini's theorem and some properties of the adjoining function. We write two detail computations and hints for computing the others. Let γ ∈ G, x ∈ X and η ∈ H. Equation (2.12): φ(ζ)φ′(ζ−1γ)f (γ−1x)∆(γ, γ−1x)1/2 dαrG(γ)(ζ) dαrX (x)(γ). (φ ∗ φ′)(γ)f (γ−1x)∆(γ, γ−1x)1/2 dαrX (x)(γ) ((φ ∗ φ′)f )(x) =ZGrX (x) =ZGrX (x)ZGrG(γ) First apply Fubini's theorem and then change the variable γ 7→ ζγ and use the invariance of α to see that the last term equals ZGrG(γ)ZGrX (x) φ(ζ)φ′(γ)f (γ−1ζ−1x)∆(ζγ, γ−1ζ−1x)1/2 dαrX (x)(γ) dαrG(γ)(ζ). We observe that (ζγ, γ−1ζ−1x) = (ζ, ζ−1x)(γ, γ−1ζ−1x) in the transformation groupoid G ⋉ X. This relation, Remark 2.4 and the associativity of the left action together allow us to write the previous term as φ(ζ)φ′(γ)f (γ−1ζ−1x) ∆(ζ, ζ−1x)1/2∆(γ, γ−1ζ−1x)1/2 dαrX (x)(γ) dαrG(γ)(ζ) φ(ζ)(cid:18)ZGrX (x) φ′(γ)f (γ−1ζ−1x)∆(γ, γ−1ζ−1x)1/2 dαrX (x)(γ)(cid:19) ∆(ζ, ζ−1x)1/2 dαrG(γ)(ζ) φ(ζ) (φ′f )(ζ−1x) ∆(ζ, ζ−1x)1/2 dαrG(γ)(ζ) ZGrG(γ)ZGrX (x) =ZGrG(γ) =ZGrG(γ) = (φ(φ′f ))(x). Equation (2.13): This computation is similar to the above computation for Equa- tion (2.12) except that there is no adjoining function here. Equation (2.14): This uses Fubini's theorem and the H-invariance of ∆. Equation (2.15): This is a direct computation. Equation (2.16): This involves a change of variable and the right invariance of the family of measures λ. Equation (2.17): This uses a change of variable, the left invariance of β and Fubini's theorem. TOPOLOGICAL CORRESPONDENCES 13 Equation 2.18: (2.19) hφf, gi(η) =ZX =ZXZG =ZXZG (φf )(x)g(xη) dλrH (η)(x) φ(γ) f (γ−1x)g(xη) ∆1/2(γ, γ−1x) dαrX (x)(γ) dλrH (η)(x) f (γ−1x) φ(γ)g(xη) ∆1/2(γ, γ−1x) dαrX (x)(γ) dλrH (η)(x). Make a change of variables (γ, γ−1x) 7→ (γ−1, x). Then we use the fact that ∆ is an almost everywhere groupoid homomorphism (see Remark 2.4). Due to Remark 2.5, we know that ∆ is H-invariant. Taking into account these facts we see that f (x) φ(γ−1)g(γ−1xη) ∆1/2(γ, γ−1xη) dαr(x)(γ) dλrH (η)(x) hφf, gi(η) =ZXZG =ZXZG = hf, φ∗gi(η). f (x)(φ∗g)(xη) dλrH (η)(x) (cid:3) in the inductive limit topology. It can be seen that the left and the right actions and the map h ,i are continuous Equations (2.12), (2.13) and (2.14) show that Cc(X) is a Cc(G)-Cc(H)-bimodule. Equations (2.15), (2.16) and (2.17) show that the map h , i : Cc(X) × Cc(X) → Cc(H) is a Cc(H)-conjugate bilinear map. And Equation (2.18) says that Cc(G) acts on Cc(X) by Cc(H)-adjointable operators. 2.1. The right action-construction of the Hilbert module. In this subsec- tion, we describe how to construct a C∗(H, β)-Hilbert module H(X), where X is Indeed, writing a proper H-space and λ is an H-invariant family of measures. H(X, λ) is more precise than H(X). But we shall not come across any case which involves the same space with different families of measures. Hence we write H(X). Note that to get H(X) from Cc(X), we only need to prove that the bilinear map is positive. The other required properties of h,i are clear from Lemma 2.11. The main result of this section, namely, Theorem 2.29 is a special case of the well- know theorem of Renault [12, Corrolaire 5.2]. In [12], Renault uses a -category of measures free and proper H-spaces to prove Corollaire 5.2. A remark following the corollaire says that the freeness of the action is not a necessary hypothesis; the result holds even if this hypothesis is removed. Based on this remark, a similar category that uses measured proper H-spaces is constructed in [5]. In [12] and [5], the main idea of the proof is that the representations of (H, β) induce representations of this category. These induced representations are used to complete Cc(X) into a Hilbert C∗(H, β)-module. Thus Theorem 2.29 follows directly from the result concerning the completion of the -category. The proof below is written for the sake of completeness; the proof uses elementary techniques. Proposition 2.20. Let (H, β) be a locally compact groupoid equipped with a Haar system, X a proper left H-space and λ an invariant family of measures on X. Then the bilinear map defined by Equation (2.9) is a Cc(H)-valued inner product on Cc(X). To prove the positivity of h,i we prove the following fact: for every (non-degenerate) representation π : C∗(H, β) → B(K) on a Hilbert space K, π(hf , fi) ∈ B(K) is pos- itive. However, instead of working with the algebraic representations of C∗(H, β) we work with representation of the groupoid (H, β) and prove the positivity result. Since the integration and disintegration of representations establish an equivalence 14 ROHIT DILIP HOLKAR between the representations of C∗(H, β) and (H, β), working with the representa- tions of (H, β) is enough. Before proceeding to the proof we set up few notions and notation. We draw X ×H(0) H βX π1 X λ sX λ1 π2 H rH β H (0). Figure 1 Figure 1 in which πi for i = 1, 2 are the projections on the ith component, λ1 and βX are families of measures as in Lemma 1.6, that is, for f ∈ Cc(X ×H(0) H) and x ∈ X the measure βx is ZX× (2.21) f d βx =ZH f (x, η) dβsX (x)(η). H(0) H And λ1 is defined similarly. Clearly, β ◦ λ1 = λ ◦ βX . Let m be a transverse measure class on H. We take the quotient of each space in Figure 1 and the corresponding induced maps and families of measures. We do the following identifications: (X ×H(0) H)/H ≈ X, H/H ≈ H (0), [π2] = sX and [π1] is the quotient map X → X/H. i) The coherence of m gives m(β) ◦ [λ1] = m(λ) ◦ [ βX ], where m(β) and m(λ) ii) A straight forward computation gives that [λ1] = λ. iii) A computation as in (ii) above gives that [ βX ] = βX , where βX is as in are the measure classes induced by m on H (0) and X/H, respectively. Example 1.8. iv) Observations (i), (ii) and (iii) above together say that m(β)◦ λ = m(λ)◦ βX . f 7→ZX/HZHsX (x)ZHsH (η) f (xη, h) dβsH (η)(h) dβsX (x)(η) dν[x] Hence if µ ∈ m(β) and ν ∈ m(λ), then (2.22) The symbol ∼ denotes the equivalence of measures. µ ◦ λ ∼ ν ◦ βX . Reader may refer to [9, Section 1.3] for the computations in (ii) and (iii) above. Now Proposition 2.20 follows from Lemma 2.27 and Lemma 2.28 below. In the following discussion, we shall write hf , fiCc(H) instead of hf , fi for f ∈ Cc(X). Let (m,H, pH, π) be a representation of (H, β) where m is a transverse measure class for H, pH : H → H (0) is a measurable H-Hilbert bundle which has separable fibres and π is the action of H on fibres of H. By definition, the transverse measure class m induces a measure class m(λ) on X/H. We fix µ ∈ m(β) and ν ∈ m(λ), that is, µ is a measure on H 0 and ν is a measure on X/H. Furthermore, let Λ : Cc(X) → Cc(H 0) and BX : Cc(X) → Cc(X/H) be the integration operators for the families of measures {λu}u∈H0 and {β[x] X }[x]∈X/H. Equation 2.22 shows that ν ◦ βX and µ ◦ λ are equivalent measures on X. Let βX be the family of measures along the projection map π1 : X×H(0) H → X, π1(x, h) = x, as in Equation (2.21) Lemma 2.23. The measure ν ◦ βX on X is H-invariant. Proof. We must prove that the measure ν ◦ βX ◦ βX on X ×H(0) H defined by TOPOLOGICAL CORRESPONDENCES 15 for f ∈ Cc(X ×H(0) H) is invariant under the inversion map (x, h) 7→ (xh, h−1). For this, we substitute η−1h for h and write ν ◦ βX ◦ βX (f ) =ZX/HZHsX (x)ZHsX (x) now replacing f by f ◦ invX⋊H replaces f (xη, η−1h) by f (xηη−1h, (η−1h)−1) = f (xh, h−1η). The substitution that switches h ↔ η shows that the integrals over f and f ◦ invX⋊H are the same. f (xη, η−1h) dβsX (x)(h) dβsX (x)(η) dν[x]; (cid:3) Since µ◦λ is equivalent to ν◦βX , this measure on X must also be quasi-invariant. We compute its Radon-Nikodym derivative. Let f ∈ Cc(X ×H(0) H), then we get µ ◦ λ ◦ βX (f ) =ZH0ZXuZHu =ZH0ZHuZXu f (x, h) dβu(h) dλu(x) dµ(u) f (x, h) dλu(x) dβu(h) dµ(u) by Fubini's Theorem. When we replace f by f ◦ invX⋊H and use the H-invariance of λ, we get µ ◦ λ ◦ βX (f ◦ invX⋊H ) =ZH0ZHuZXu =ZH0ZHuZXsH (h) =ZH0ZHuZXrH (h) f (x, h−1) dλsH (h)(x) dβu(h) dµ(u) f (xh, h−1) dλu(x) dβu(h) dµ(u) f (x, h) dλrH (h)(x) dβ−1 u (h) dµ(u), where the last step uses the substitution h 7→ h−1. In terms of the integration oper- ator Λ1 : Cc(X ×H(0) H) → Cc(H) along π2, we may rewrite this as µ◦ β−1(Λ1(f )), whereas µ ◦ λ ◦ βX (f ) = µ ◦ β(Λ1(f )). Thus the Radon-Nikodym derivative is d inv∗ X⋊H(µ ◦ λ ◦ βX ) d(µ ◦ λ ◦ βX ) (x, h) = d(µ ◦ β−1) d(µ ◦ β) (h). Now let M (x) = d(µ ◦ λ) d(ν ◦ β) . Lemma 2.24. Let x ∈ X and h ∈ H satisfy sX (x) = rH (h). Then M (xh) = M (x) d(µ ◦ β−1) d(µ ◦ β) (h). Proof. Let g ∈ Cc(X×H(0) H) and let f = βX (g), that is, f (x) =RHsX (x) f (x, h) dβsX (x)(h). By definition of the Radon-Nikodym derivative, we have f (x)M (x)−1 d(µ ◦ λ)(x). ZX f (x) d(ν ◦ β)(x) =ZX Thus ZX× H(0) H g(x, h) d(ν ◦ β ◦ βX )(x, h) =ZX× H(0) H g(x, h)M (x)−1 d(µ ◦ λ ◦ βX )(x, h). 16 ROHIT DILIP HOLKAR Since the measure ν◦β is H-invariant, the left hand side is invariant under replacing g by g ◦ invX⋊H . Hence so is the right-hand side, that is, ZX g(x, h)M (x)−1 d(µ ◦ λ ◦ βX )(x, h) =ZX =ZX g(xh, h−1)M (x)−1 d(µ ◦ λ ◦ βX )(x, h) g(xh, h−1)M (xh)−1 M (xh) M (x) d(µ ◦ λ ◦ βX )(x, h). Letting g′(x, h) = g(x, h)M (x)−1, we see that M (xh)/M (x) has to be the Radon- Nikodym derivative M (xh) M (x) = d(inv∗ X⋊H µ ◦ λ ◦ βX ) d(µ ◦ λ ◦ βX ) (x, h) = d(µ ◦ β−1) d(µ ◦ β) (h). (cid:3) Hu dµ(u) < ∞. Recall that H = (Hx)x∈X is a µ-measurable field of Hilbert spaces over H (0) equipped with a representation π of H. The Hilbert space L2(H (0), µ,H) consists of all µ-measurable sections ξ : H (0) → H such that We pull back H to a field s∗ ZH(0)kξ(u)k2 This norm comes from an inner product on L2(H (0), µ,H), of course. XH of Hilbert spaces over X along sX : X → H (0). Then we take the induced field of Hilbert spaces HX over X/H whose µ ◦ λ- measurable sections are those sections ζ of s∗H that satisfy πh(ζ(xh)) = ζ(x) for all x ∈ X, h ∈ H with sX (x) = rH (h). For ν as above, we define the Hilbert space L2(X/H, ν,HX) to consist of those sections ζ of HX with dν[x] < ∞. HsX (x) ZX/Hkζ(x)k2 HsX (x) The function kζ(x)k2 is constant on H-orbits and thus descends to X/H be- cause πh(ζ(xh)) = ζ(x) and the operators πh are unitary. The norm defining L2(X/H, ν,HX) comes from an obvious inner product. Notice that an element of L2(X/H, ν,HX) is not a function on X/H. Now we define the operator fii from L2(H (0), µ,H) to L2(X/H, ν,HX) and its adjoint hhf. Let ξ ∈ L2(H (0), µ,H) and ζ ∈ L2(X/H, ν,HX). Computations by Renault which are discussed in [5] or [9, Section 3.3.1] lead to the following formulae for hhf and fii: (2.25) f (xη)πη(ξ(sH (η)))pM (xη) dβsX (x)(η), 1 f (x)ζ(x) dλu(x). (fiiξ)([x]) =ZHsX (x) (hhfζ)(u) =ZXu (2.26) Notice that pM (x) πh(fiiξ(xh)) =ZHsX (x) f (xhη)πhη(ξ(sH (η)))pM (xhη) dβsX (x)(η) = fiiξ(x) by the substitution hη 7→ η because β is left-invariant. Thus fiiξ is a section of HX . Let Cc(X;H)0 and Cc(X;H) have similar meanings as Cc(X)0 and Cc(X), If we pick ξ, ζ ∈ Cc(X;H), then fiiξ ∈ Cc(X/H) respectively, as on page 4. and hhfζ ∈ Cc(H (0)). Hence our operators fii and hhf are well-defined on dense subspaces. TOPOLOGICAL CORRESPONDENCES 17 Lemma 2.27. Let ξ, ζ ∈ Cc(X;H). Then hζ,fiiξi = hhhfζ, ξi, that is, hhf is formally adjoint to fii. Proof. On the one hand, hζ,fiiξi =ZX/Hhζ(x),fiiξ(x)i dν(x) =ZX/HZHsX (x)hζ(x), f (xη)πη ξ(sH (η))ipM (xη) dβsX (x)(η) dν[x] =ZX/HZHsX (x)hζ(xη), ξ(sH (η))if (xη)pM (xη) dβsX (x)(η) dν[x] =ZX/HZHsX (x)hζ(x), ξ(sX (x))if (x)pM (x) d(ν ◦ β)(x), where we used πhζ(xh) = ζ(x), the unitarity of πh, and the definition of the measure ν ◦ β on X. On the other hand, hhhfζ, ξi =ZH(0)hhhfζ(u), ξ(u)i dµ(u) =ZH(0)ZXuhf (x)ζ(x)pM (x) =ZXhζ(x), ξ(sX (x))if (x)pM (x) =ZXhζ(x), ξ(sX (x))if (x)pM (x) −1 , ξ(u)i dλu(x) dµ(u) −1 d(µ ◦ λ)(x) −1 d(µ ◦ λ) d(ν ◦ β) (x) d(ν ◦ β)(x). Now the definition of M shows that this is the same as the previous integral. (cid:3) The convolution algebra Cc(H) acts on L2(H (0), µ,H) by f (η)πηξ(sH (η))s d(µ ◦ β−1) d(µ ◦ β) L(f )ξ(u) =ZHu (η) dβu(η). This is a ∗-representation. Lemma 2.28. Let ξ ∈ Cc(X;H). Then hhf ◦ fii(ξ) = L(hf, fiCc(H))(ξ). Hence hhf ◦ fii extends to a bounded operator with norm at most khf, fikC ∗(H). It follows that fii and hhf extend to bounded operators between the Hilbert spaces L2(H (0), µ,H) and L2(X/H, ν,HX) which are adjoints of one another. Proof. We compute hhf ◦ fii(ξ)(u) =ZXu f (x)fii(ξ)(x)√M (x) dλu(x) −1 f (x)f (xη)πη(ξ(sH (η)))√M (xη)√M −1 (x) dβu(η) dλu(x) =ZXuZHu Now we use Lemma 2.24 to identify M (xη)/M (x) with the function Then we use Fubini's Theorem and continue the computation: δ(η) = d(µ ◦ β−1) d(µ ◦ β) (η). hhf ◦ fii(ξ)(u) =ZHuZXu f (x)f (xη)πη(ξ(sH (η)))pδ(η) dλu(x) dβu(η) =ZHuhf, fiCc(H)(η)πη(ξ(sH (η)))pδ(η) dβu(η) = L(hf, fiCc(H))(ξ). 18 ROHIT DILIP HOLKAR Since L(hf, fiCc(H)) is bounded, it follows that hhf ◦ fii extends to a bounded operator on L2(H (0), µ,H). Let C > 0 be its norm. Then kfiiξk2 = hξ,hhf ◦ fiiξi ≤ Ckξk2 by Lemma 2.27 for all ξ ∈ Cc(X,H). Hence fii extends to a bounded operator from L2(H (0), µ,H) to L2(X/H, ν,HX). A similar estimate shows that hhf extends to a bounded operator from L2(X/H, ν,HX) to L2(H (0), µ,H). Proof of Proposition 2.20. Follows from Lemma 2.28. (cid:3) (cid:3) The last proposition shows that Cc(X) is a C∗(H, β)-pre-Hilbert module. Let H(X) denote the C∗(H, β)-Hilbert module obtained by completing Cc(X). Theorem 2.29 (See [12, Corrolaire 5.2], or the discussion above). Let (H, β) be a locally compact groupoid with a Haar system and let X be a locally compact proper right H-space carrying an H-invariant continuous family of measures λ. Then using Formulae (2.8) and (2.9) the right inner product Cc(H)-module over the pre- C∗-algebra Cc(X) can be completed to a C∗(H, β)-Hilbert module H(X). 2.2. The left action and construction of the C∗-correspondence. Now we turn our attention to the left action. We wish to extend the action of Cc(G) on Cc(X) to an action of C∗(G) on H(X). For a groupoid equivalence the adjoining function vanishes (see Example 3.9), that is, it becomes the constant function 1, and the formulae for the left actions in Definition 2.8 and [7] match. In this case, Cc(G) acts on Cc(X) by C∗(H, β)-adjointable operators. Our proof for the non-free case runs along the same lines as in [7]. Lemma 2.30. The action of Cc(G) on Cc(X) defined by Equation 2.8 extends to a non-degenerate -homomorphism from C∗(G, α) to BC∗(H)(H(X)). Proof. Equation 2.18 in Lemma 2.11 shows that Cc(G) acts by on Cc(X) by ad- jointable operators. We need to prove that this representation is non-degenerate. For this purpose we show that there is continuous family of measures with full sup- port α = { αx}x∈X along the projection map π2 : G ⋉ X → X. Then Lemma 1.5 gives that A : Cc(G×G(0) X) → Cc(X) is a surjection. Since, due to the theorem of Stone-Weierstass the set {g ⊗ h : g ∈ Cc(G), h ∈ Cc(X)} ⊆ Cc(G ×G(0) X) is dense, the claim of the current lemma will be proved. 2 (x) = Let f ∈ Cc(G ×G(0) X) is given. For x ∈ X define the measure αx on π−1 sG× Since αrX (x) has full support and ∆ is non-zero, α G(0) X has full support. Using an argument similar to Lemma 1.6 we may infer that α := { αx}x∈X is continuous. Finally, we check that the action is bounded. Once we prove this, then the action extends to C∗(G, α). Let ǫ be a state on C∗(H, β). Then ǫ(h,i) makes H(X) into a Hilbert space, say H(X)ǫ. Let Vǫ ⊆ H(X)ǫ be the dense subspace space generated by {ζf : ζ ∈ Cc(G), f ∈ Cc(X)}. Define a representation L of Cc(G) on Vǫ by L(ζ)f = ζf . i) The representation L is a non-degenerate representation of Cc(G) on Vǫ. Non-degenerate means that the set {ζf : ζ ∈ Cc(G), f ∈ Cc(X)} is dense in Vǫ. ii) The continuity of the operations in Lemma 2.11 in the inductive limit topo- logy implies that L is continuous: for f, g ∈ Cc(X), Lf,g(ζ) = hf, L(ζ)gi is a continuous functional on Cc(G) when Cc(G) is given the inductive limit topology. GrX (x) × {x} by Zπ−1 2 (x) f d αx =ZGrX (x) f (η−1, x)∆1/2(η−1, x) dαrX (x)(γ). TOPOLOGICAL CORRESPONDENCES 19 iii) L preserves the involution, that is, hζf, gi = hf, ζ∗gi. This is proved in Equation 2.18 in Lemma (2.11). Proposition 4.2 of [12] says that L is a representation of G on Vǫ. Hence L is bounded with respect to the norm on C∗(G). Thus ǫ(hζf , ζfi = ǫ(hL(ζ)f , L(ζ)fi ≤ ζC∗(G) ǫ(h f, fi) for all f ∈ Cc(X) and ζ ∈ Cc(G). The state ǫ was arbitrary. Hence for all f ∈ Cc(X) and ζ ∈ Cc(G) we get h ζf, ζfi ≤ ζC∗(G) h f, fi. This shows that the action of Cc(G) on Cc(X) is bounded in the topology induced by the norm of the inner product h ,i. Hence the action can be extended to C∗(G). (cid:3) Now we are ready to state the main theorem. Theorem 2.31. Let (G, α) and (H, β) be locally compact groupoids with Haar systems. Then a topological correspondence (X, λ) from (G, α) to (H, β) produces a C∗-correspondence H(X) from C∗(G, α) to C∗(H, β). Proof. Follows by putting Proposition 2.29 and Lemma 2.30 together. (cid:3) 3. Examples Example 3.1. Let X and Y be locally compact, Hausdorff spaces, and let f : X → Y be a continuous function. We view X and Y as groupoids with Haar systems consisting of Dirac measures on X and Y , δX = {δx}x∈X and δY = {δy}y∈Y , respectively. We write X ′ for the space X. We use this notation to avoid confusing the space and the groupoid structures. The function f is the momentum map for the trivial left action of Y on X ′, that is, for (f (x), x) ∈ Y ×IdY ,Y,f X ′, f (x) · x = x. In fact, this is the only possible action of Y on X ′. In a similar way X acts on itself trivially via the momentum map IdX . The family of Dirac measures δX mentioned above is an X-invariant family of measures on X ′. Both the actions are proper. If h ∈ Cc(Y ×IdY ,Y,f X ′), then ZX ′ZY h(y, x) d(δY )f (a)(y) d(δX )a(x) = h(f (a), a) =ZX ′ZY h(y−1, yx) d(δY )f (a)(y) d(δX )a(x). Therefore δX is Y -invariant. Thus (X ′, δX ) is a topological correspondence from Y to X with the constant function 1 as the adjoining function. The action of Cc(X) on Cc(X ′) as well as the Cc(X)-valued inner product on Cc(X ′) are the pointwise multiplication of two functions. For h ∈ Cc(Y ), k ∈ Cc(X ′), (h · k)(x) = h(f (x))k(x). The C∗-correspondence H(X ′) : C0(Y ) → C0(X) is the C∗-correspondence asso- ciated with the -homomorphism f ∗ : C0(Y ) → M(C0(X)) produced by the Gel- fand transform. Example 3.2. Let X, Y , X ′ and f be as in Example 3.1. Let λ = {λy}y∈Y be a continuous family of measures along f . It follows from the discussion in Example 3.1 that X is a proper X-Y -bispace. For h ∈ Cc(X ×IdX ,X,IdX X ′), ZX ′ZX h(x−1, xz) d(δX )x(z) dλy(x) =ZX ′ZX h(x, z) d(δX )x(z) dλy(x) =ZX ′ h(f (x), x) dλy(x). 20 ROHIT DILIP HOLKAR The first equality above is due to the triviality of the action and the second one follows from the definition of the measures δX × λy as in Remark 2.3. Thus λ is X-invariant and the modular function is the constant function 1. Hence (X ′, λ) is a correspondence from X to Y . Example 3.3. Let X and Y be as in Example 3.1. Let f, b : X → Y be continuous maps and let λ be a continuous family of measures along f . Make X a Y -Y -bispace using actions similar to those in Example 3.1. For f : X → Y use the family of measures λ and the formulae in Example 3.2 to define a right action of Cc(Y ) on Cc(X). It is straightforward to check that (X, λ) is a topological correspondence from Y to itself. When the spaces are second countable, the quintuple (Y, X, s, r, λ) is called a topological quiver [8]. In general, assume that X, Y and Z are locally compact Hausdorff spaces, f : X → Y and b : X → Z are maps. Let λ be a continuous family of measures along f . Then (X, λ) is a topological correspondence from Z to Y . Example 3.4. Let G and H be locally compact groups, φ : H → G a continuous group homomorphism, and α and β the Haar measures on G and H, respectively. The right multiplication action is a proper action of G on itself. The measure α−1 is invariant under this action. Using φ define an action of H on G as ηγ = φ(η)γ for (η, γ) ∈ H × G. We claim that α−1 is H-quasi-invariant for this H-action. Let δG and δH be the modular functions of G and H, respectively. The modular functions allow to switch between the left and right invariant Haar measures α and α−1, and β and β−1. The relations are α−1 = δ−1 H β. If Rγ : G → G is the right multiplication operator, then G α and β−1 = δ−1 ZG Rγf dα = δG(γ)−1ZG f dα f (η, φ(η)−1γ) for f ∈ Cc(G). A similar equality holds for g ∈ Cc(H). Let f ∈ Cc(H × G), ZGZH =ZGZH =ZGZH (by removing φ(η)−1 in G). f (η−1, γ) dβ(η) dα−1(γ) dβ(η) dα−1(γ) δH (η) δG(φ(η)) 1 f (η−1, φ(η)γ) dβ(η) dα−1(γ) (by sending η to η−1 in H) δG(φ(η)) If one compares the first term of the above computation with the equation in (iv) of Definition 2.1, and uses the fact that that the adjoining function is a groupoid homomorphism, then it can be seen that ∆(η, η−1γ) = δH (η) δG◦φ(η) . Hence ∆(η−1, γ) = ∆(η, η−1γ)−1 = δG◦φ(η) δH (η) . Thus a group homomorphism φ : H → G gives a topological correspondence (G, α−1) from (H, β) to (G, α) and δG◦φ is the δH adjoining function. Example 3.5. Let G, H, α, β, δH and φ be as in Example 3.4. Additionally, assume that φ : H → G is a proper function. For the time being, assume that the action of H on G given by γη := γφ(η) for (γ, η) ∈ G × H is proper, which is a fact and we prove it towards the end of this example. With this action of H and the left multiplication action of G on itself, G is a proper G-H-bispace. α−1 is an H-invariant measure. The adjoining function of this action is the constant function TOPOLOGICAL CORRESPONDENCES 21 1. To see this, let f ∈ Cc(G × G), then ZZ f (γ−1, η) dα(γ)dα−1(η) =ZZ f (γ, η) δG(γ)−1dα(γ)dα−1(η) =ZZ f (γ, γ−1η) dα(γ)dα−1(η) (because α−1 = δ−1α) (because Lγα−1 = δ(γ)α−1). Now we prove that the action of G on H is proper, that is, the map Ψ : H × G → H × H sending (γ, η) 7→ (γ, γφ(η)) is proper. The maps IdH × φ : H × G → H × H and m : (η, η′) 7→ (η, ηη′) from H × H → H × H are proper, and Ψ = m ◦ (IdH × φ). Hence Ψ is proper. Example 3.6. Let G be a locally compact group and α the Haar measure on G. Let X be a locally compact proper left G-space. Let λ be a strongly G-quasi-invariant measure on X, that is, there is a continuous function ∆ : G × X → R+ such that d(gλ)(x) = ∆(g, x)dλ(x) for every g in G. In this setting, (X, λ) is a correspondence from (G, α) to (Pt, δPt), with ∆ as the adjoining function. The C∗-algebra for Pt is C, the Hilbert module H(X) is the Hilbert space L2(X, λ) and the action of C∗(G) on this Hilbert module is the representation of C∗(G) obtained from the representation of G on Cc(X). An example of this situation is: when X is a homogeneous space for G, X carries a G-strongly quasi-invariant measure. For details, see [4, Section 2.6]. Example 3.7 (Macho Stadler and O'uchi's correspondences). In [15], Macho Stadler and O'uchi present a notion of groupoid correspondences. We change the direction of correspondence in their definition to fit our construction and reproduce the defin- ition here: Definition 3.8. A correspondence from a locally compact, Hausdorff groupoid with Haar system (G, α) to a groupoid with Haar system (H, β) is a G-H-bispace X such that: i) the action of H is proper and the momentum map for the right action sX is open, ii) the action of G is proper, iii) the actions of G and H commute, iv) the right momentum map induces a bijection from G\X to H (0). Macho Stadler and O'uchi do not assume that the left momentum map is open. We do the same. Condition (iv) above is equivalent to saying that G\X and H (0) are homeomorphic. Macho Stadler and O'uchi do not require a family of measures on the G-H- bispace X. We show that a correspondence of Macho Stadler and O'uchi carries a canonical H-invariant continuous family of measures λ which is given by ZXu f dλu :=ZG f (γ−1x) dαrX (x)(γ) for f ∈ Cc(X), where u = sX (x). Note that this family of measures is the family of measures α−1 X along the quotient map X → G\X as in Example 1.8. Condition (iv) in the above definition identifies G\X ≈ H (0) to give the desired result. Since λ is invariant for the G-actions, we get ∆ = 1. Thus (X, λ) is a topological correspondence from (G, α) to (H, β) in our sense. 22 ROHIT DILIP HOLKAR Macho Stadler and O'uchi prove that such a correspondence from (G, α) to (H, β) induces a C∗-correspondence from C∗ r (G, α) to C∗ r (H, β). Example 3.9 (Equivalence of groupoids). Definition 3.10 (Equivalence of groupoids, a slight modification of Definition 2.1 [7]). Let G and H be locally compact groupoids. A locally compact space X is a G-H- equivalence if i) X is a left free and proper G-space; ii) X is a right free and proper H-space; iii) the momentum maps rX and sX are open; iv) the actions of G and H commute, that is, X is a G-H-bispace; v) the left momentum map rX : X → G(0) induces a bijection of X/H onto vi) the right momentum map sX : X → H (0) induces a bijection of G\X onto G(0); H (0). Equivalences of Hausdorff groupoids ([7]) are a special case of Macho Stadler- O'uchi correspondences. Hence equivalences of groupoids are topological corres- pondences as well. Similarly, one can check that an equivalence of locally compact groupoids defined in [12] is also a topological correspondence. Furthermore, an equivalence of groupoids is an invertible correspondence. Example 3.11 (Generalised morphisms of Buneci and Stachura). Buneci and Stach- ura define generalised morphisms in [3]. We modify this definition to fit our con- ventions and repeat it here: Definition 3.12. A generalised morphism from (G, α) to (H, β) is a left action Θ of G on the space H with rGH as the anchor map, the action commutes with the right multiplication action of H on itself and there is a continuous positive function ∆Θ on G ×sG,H(0),rGH H such that ZZ f (γ, γ−1η) ∆Θ(γ, γ−1η) dαrGH (η)(γ) dβ−1 for all f ∈ Cc(G ×sG,H(0),rGH H) and u ∈ H (0). u (η) =ZZ f (γ−1, η) dαrGH (η)(γ)dβ−1 u (η) If Θ is a generalised morphism from (G, α) to (H, β) then (H, β−1) is a topological correspondence from (G, α) to (H, β), where β−1 is the family of measures Z f d(β−1)u =Z f (η−1) dβu(η) for f ∈ Cc(H) and u ∈ H (0). It is obvious from the definition itself that ∆Θ is the adjoining function for this correspondence. In [3], Buneci and Stachura prove that a generalised morphism induces a -homomorphism from C∗(G, α) to M(C∗(H, β)). This is a C∗-correspondence from C∗(G, α) to C∗(H, β) with the underlying Hilbert module C∗(H, β). Example 3.13. Let X be a locally compact right G-space for a locally compact group G and let λ be the Haar measure on G. Let H and K be subgroups of G. Assume that K is closed and let α and β be the Haar measures on H and K, respectively. Then X ⋊H and X ⋊K are subgroupoids of X ⋊G. Denote these three transformation groupoids by H, K and G, respectively. Then G is an H-K-bispace for the left and the right multiplication actions, respectively. We bestow H and K with the Haar systems {αy}y∈X and {βz}z∈X , respectively, where αy = α and βz = β for each y, z ∈ X. If λ−1 x = λ−1 for all x ∈ X, then the family of measures {λ−1 x }x∈X on G is K-invariant. We show that this family is H-quasi-invariant with the adjoining function δG/δH. TOPOLOGICAL CORRESPONDENCES 23 For ((x, γ), (y, κ)) ∈ G ⋊ K we have y = xγ and the map (x, γ, y, κ) 7→ (x, γ, κ) gives an isomorphism between the groupoids G ⋊ K and X ⋊ (G × K). Using this identification, it can be checked that the right action of K on G is proper, which is implied by the fact that K ⊂ G is closed. Another quicker way to see this is to observe that K ⊆ G is a closed subgroupoid. G (u) ⊆ G. If a = (uγ−1, η) ∈ H, then a−1 = (uγ−1, η)−1 = (uγ−1η, η−1) is composable with b and a−1b = (uγ−1, η)−1(uγ−1, γ) = (uγ−1η, η−1γ). Now a computation similar to that in Example 3.4 shows that Let f ∈ Cc(H ⋉ G), u ∈ K(0) = G(0) ≈ X. Let b = (uγ−1, γ) ∈ s−1 f (a, a−1b) δH (η) δG(η) ZGZH =ZGZH =ZGZH =ZGZH (by changing η 7→ η−1). dαrG(b)(a) dλ−1 u (b) δH (η) δG(η) δH (η) δG(η) 1 δG(η) f ((uγ−1, η), (uγ−1η, η−1γ)) f ((uγ−1, η), (uγ−1η, η−1γ)) dαrG(uγ−1,γ)(uγ−1, η) dλ−1 u (uγ−1, γ) dα(η) dλ−1(γ) f ((uγ−1, η−1), (uγ−1η−1, ηγ)) dα(η) dλ−1(γ) δG(η) to see that the previous term equals f ((uγ−1η, η−1), (uγ−1, γ)) dα(η) dλ−1(γ) Now we change γ 7→ η−1γ. Then we use the relation dλ−1(η−1γ) = dλ−1(γ) ZGZH =ZGZH =ZGZH Thus {λ−1 the adjoining function. f ((uγ−1, η)−1, (uγ−1, γ)) dα(uγ−1) dαrG(uγ−1,γ)(uγ−1, η) dλ−1 x }x∈X is an H-quasi-invariant family of measures on G with δG/δH as u (uγ−1, γ) dαrG(b)(a) dλ−1 u (b). f (a−1, b) δH (η) δG(η) Let f ′ : X → Y be a proper map between locally compact spaces. Let B ⊆ Y , A ⊆ f ′−1(B) and f : A → B be the map obtained by restricting the domain and the codomain of f ′. Assume that A ⊆ X is closed. Then we claim that f is proper. Let K ⊆ B be compact, then K is compact in Y also. Thus it is enough to consider the case when B = Y , in which case f is the restriction of f ′ to the closed subspace A. Since the inclusion map iA : A ֒→ X is closed, [2, Chapter I, §10.1, Proposition 2] shows that iA is proper. Hence f = f ′ ◦ iA is proper. Example 3.14 (The induction correspondence). Let (G, α) be a locally compact groupoid with a Haar system, H a closed subgroupoid. Let β be a Haar system for H. Note that GH(0) is a G-H-bispace where the left and right actions are multiplication from the left and right, respectively. Both actions are free. We claim that the actions of G and H are proper. Let ι : G(0) → G be the inclusion map which is continuous. Then H (0) = ι−1(H) ⊆ G(0) is closed. Let Ψ : GH(0) ×H(0) H → GH(0) × GH(0) be the map Ψ(x, η) = (x, xη). Note that Ψ is obtained from the proper map (x, η) 7→ (x, xη), G ×sG,G(0),rG G → G × If we prove that GH(0) ×H(0) H ⊆ G, by restricting the domain and codomain. G ×sG,G(0),rG G is closed, then, from the discussion preceding this example, it will follow that Ψ is proper. 24 ROHIT DILIP HOLKAR To see that the left action is proper, first note that GH(0) = s−1 Since H (0) ⊆ G(0) is closed, (sG×rG)−1(H (0)) = G×sG,H(0),rG G ⊆ G×sG,G(0),rG G is closed where sG × rG : G ×sG,G(0),rG G → G(0) is the map (γ, η) 7→ sG(γ) = rG(η). Now the projection on the second factor π2 : G ×sG,H(0),rG G → G is a continuous, due to which GH(0) ×H(0) H = π−1 G (H (0)) ⊆ G is closed. Thus π−1 GH(0) ⊆ G ×sG,G(0),rG G is closed. And then arguing same as the right action shows that the left action is also proper. Now it is not hard to see that G\X ≈ H (0). By Example 3.7, X produces a Though both actions are free and proper, this correspondence need not be a 2 (H) ⊆ G ×sG,H(0),rG G is closed. topological correspondence from (G, α) to (H, β). 2 (GH(0) ) = G ×sG,G(0),rG H(0) groupoid equivalence as it might fail to satisfy Condition (v) of Definition 3.10. Acknowledgement: I thank Jean Renault and Ralf Meyer for many fruitful dis- cussions. I thank Ralf Meyer specially; for proofreading the document carefully. I am thankful to the EuroIndia Project of Erasmus Mundus and CNPq, Brasil, that is, the Brazilian National Council for Scientific and Technological Development, for their supports. References [1] C. Anantharaman-Delaroche and J. Renault, Amenable groupoids, Monographies de L'Enseignement Mathématique [Monographs of L'Enseignement Mathématique], vol. 36, L'Enseignement Mathématique, Geneva, 2000. With a foreword by Georges Skandalis and Appendix B by E. Germain. MR1799683 (2001m:22005) [2] Nicolas Bourbaki, Elements of mathematics. General topology. Part 1, Hermann, Paris; Addison-Wesley Publishing Co., Reading, Mass.-London-Don Mills, Ont., 1966. MR0205210 (34 #5044a) [3] Mădălina Roxana Buneci and Piotr Stachura, Morphisms of locally compact groupoids en- dowed with Haar systems (2005), eprint. arXiv 0511613. [4] Gerald B. Folland, A course in abstract harmonic analysis, Studies in Advanced Mathematics, CRC Press, Boca Raton, FL, 1995. MR1397028 (98c:43001) [5] Rohit Dilip Holkar and Jean Renault, Hypergroupoids and C ∗-algebras, C. R. Math. Acad. Sci. Paris 351 (2013), no. 23-24, 911–914, DOI 10.1016/j.crma.2013.11.003 (English, with English and French summaries). MR3133603 [6] E. C. Lance, Hilbert C ∗-modules, London Mathematical Society Lecture Note Series, vol. 210, Cambridge University Press, Cambridge, 1995. A toolkit for operator algebraists. MR1325694 (96k:46100) [7] Paul S. Muhly, Jean N. Renault, and Dana P. Williams, Equivalence and isomorphism for groupoid C ∗-algebras, J. Operator Theory 17 (1987), no. 1, 3–22. MR873460 (88h:46123) [8] Paul S. Muhly and Mark Tomforde, Topological quivers, Internat. J. Math. 16 (2005), no. 7, 693–755, DOI 10.1142/S0129167X05003077. MR2158956 (2006i:46099) [9] Rohit Dilip Holkar, Topological construction of C∗-correspondences for groupoid C ∗-algebras, Georg-August-Universität Göttingen, 2014. [10] Alan L. T. Paterson, Groupoids, inverse semigroups, and their operator algebras, Pro- gress in Mathematics, vol. 170, Birkhäuser Boston, Inc., Boston, MA, 1999. MR1724106 (2001a:22003) [11] Jean Renault, A groupoid approach to C ∗-algebras, Lecture Notes in Mathematics, vol. 793, Springer, Berlin, 1980. MR584266 (82h:46075) [12] [13] , Représentation des produits croisés d'algèbres de groupoïdes, J. Operator Theory 18 (1987), no. 1, 67–97 (French). MR912813 (89g:46108) , Induced representations and hypergroupoids, SIGMA Symmetry Integrability Geom. Methods Appl. 10 (2014), Paper 057, 18, DOI 10.3842/SIGMA.2014.057. MR3226993 [14] Marc A. Rieffel, Induced representations of C ∗-algebras, Advances in Math. 13 (1974), 176– 257. MR0353003 (50 #5489) [15] Marta Macho Stadler and Moto O'uchi, Correspondence algebras, www.mathjournals.org/jot/1999-042-001/1999-042-001-005.pdf. J. Operator Theory (1999), no. 42 1, of 103–119, groupoid C ∗- at available [16] Jean-Louis Tu, Non-Hausdorff groupoids, proper actions and K-theory, Doc. Math. 9 (2004), 565–597 (electronic). MR2117427 (2005h:22004) TOPOLOGICAL CORRESPONDENCES 25 [17] Joel J. Westman, Cohomology for ergodic groupoids, Trans. Amer. Math. Soc. 146 (1969), 465–471. MR0255771 (41 #431) E-mail address: [email protected] Department of Mathematics, Federal University of Santa Catarina, 88. 040-900, Flori- anopólis, SC, Brazil
1005.4561	1	1005	2010-05-25T13:10:35	Automatic continuity and $C_0(\Omega)$-linearity of linear maps between $C_0(\Omega)$-modules	[ "math.OA", "math.FA" ]	Let $\Omega$ be a locally compact Hausdorff space. We show that any local $\mathbb{C}$-linear map (where "local" is a weaker notion than $C_0(\Omega)$-linearity) between Banach $C_0(\Omega)$-modules are "nearly $C_0(\Omega)$-linear" and "nearly bounded". As an application, a local $\mathbb{C}$-linear map $\theta$ between Hilbert $C_0(\Omega)$-modules is automatically $C_0(\Omega)$-linear. If, in addition, $\Omega$ contains no isolated point, then any $C_0(\Omega)$-linear map between Hilbert $C_0(\Omega)$-modules is automatically bounded. Another application is that if a sequence of maps $\{\theta_n\}$ between two Banach spaces "preserve $c_0$-sequences" (or "preserve ultra-$c_0$-sequences"), then $\theta_n$ is bounded for large enough $n$ and they have a common bound. Moreover, we will show that if $\theta$ is a bijective "biseparating" linear map from a "full" essential Banach $C_0(\Omega)$-module $E$ into a "full" Hilbert $C_0(\Delta)$-module $F$ (where $\Delta$ is another locally compact Hausdorff space), then $\theta$ is "nearly bounded" (in fact, it is automatically bounded if $\Delta$ or $\Omega$ contains no isolated point) and there exists a homeomorphism $\sigma: \Delta \rightarrow \Omega$ such that $\theta(e\cdot \varphi) = \theta(e)\cdot \varphi\circ \sigma$ ($e\in E, \varphi\in C_0(\Omega)$).	math.OA	math	AUTOMATIC CONTINUITY AND C0(Ω)-LINEARITY OF LINEAR MAPS BETWEEN C0(Ω)-MODULES CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG Abstract. Let Ω be a locally compact Hausdorff space. We show that any local C- linear map (where "local" is a weaker notion than C0(Ω)-linearity) between Banach C0(Ω)-modules are "nearly C0(Ω)-linear" and "nearly bounded". As an application, a local C-linear map θ between Hilbert C0(Ω)-modules is automatically C0(Ω)-linear. If, in addition, Ω contains no isolated point, then any C0(Ω)-linear map between Hilbert C0(Ω)-modules is automatically bounded. Another application is that if a sequence of maps {θn} between two Banach spaces "preserve c0-sequences" (or "preserve ultra-c0- sequences"), then θn is bounded for large enough n and they have a common bound. Moreover, we will show that if θ is a bijective "biseparating" linear map from a "full" essential Banach C0(Ω)-module E into a "full" Hilbert C0(∆)-module F (where ∆ is another locally compact Hausdorff space), then θ is "nearly bounded" (in fact, it is automatically bounded if ∆ or Ω contains no isolated point) and there exists a homeo- morphism σ : ∆ → Ω such that θ(e · ϕ) = θ(e) · ϕ ◦ σ (e ∈ E, ϕ ∈ C0(Ω)). 2000 Mathematics Subject Classification: 46H40, 46L08, 46H25. Keywords: Banach modules, Banach bundles, local mappings, separating mappings, automatic continuity, C0(Ω)-linearity 1. Introduction A linear map θ between the spaces of continuous sections of two bundle spaces over the same locally compact Hausdorff base space Ω is said to be local if for any continuous section f , one has supp θ(f ) ⊆ supp f , or equivalently, for each g ∈ C0(Ω), f g = 0 =⇒ θ(f )g = 0. Consequently, local property is weaker than C0(Ω)-linearity. In the case when the domain and the range bundles are over different base spaces, a more general notion is defined; namely, disjointness preserving, or separating (see Section 5). Local and disjointness preserving linear maps are found in many researches in anal- ysis. For example, a theorem of Peetre [19] states that local linear maps of the space of smooth functions defined on a manifold modelled on Rn are exactly linear differential Date: November 9, 2018. The authors are supported by National Natural Science Foundation of China (10771106) and Taiwan NSC grant (NSC96-2115-M-110-004-MY3). 1 0 1 0 2 y a M 5 2 ] . A O h t a m [ 1 v 1 6 5 4 . 5 0 0 1 : v i X r a 2 CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG operators (see, e.g., [17]). This is further extended to the case of vector-valued differen- tiable functions defined on a finite dimensional manifold by Kantrowitz and Neumann [16] and Araujo [3]. In the topological setting, similar results have been obtained. Local linear maps of the space of continuous functions over a locally compact Hausdorff space are multiplication operators, while disjointness preserving (separating) linear maps between two such spaces over possibly different base spaces are weighted composition operators (see, e.g., [1, 5, 18, 14, 12, 15]). Among many interesting questions arising from these two notions, quite a few efforts has been put on the automatic continuity of such maps. See, e.g., [2, 7, 14, 15] for the scalar case, and [13, 4, 3, 6] for the vector-valued case. In this paper, we extend this context to local or separating linear maps between spaces of continuous sections of vector bundles. Note that similar to the correspondence de- veloped by Swan [20] between finite dimensional vector bundles over a locally compact Hausdorff space Ω and certain C0(Ω)-modules, the spaces of continuous sections of "Ba- nach bundles" are certain Banach C0(Ω)-modules (see, e.g., [10], and Section 2 below). One of the original motivation behind this work is to investigate up to what extent will a local linear map between two Banach C0(Ω)-modules be C0(Ω)-linear. Surprisingly, on top of finding that such maps are "nearly C0(Ω)-linear", we find that they are also "nearly bounded". In fact, it is well known that there are many unbounded C-linear maps from an infinite dimensional Banach space to another Banach space and so, if S is a finite set, there are many unbounded C(S)-module maps from certain Banach C(S)-module to another Banach C(S)-module. The interesting thing we discovered is that the above is, in many cases, the "only obstruction" to the automatic boundedness of C0(Ω)-module maps (see Proposition 3.5 as well as Theorems 3.7 and 4.2). More precisely, if θ is a local C-linear map (not assumed to be bounded) from an essential Banach C0(Ω)-module E to another such module F , then θ is "nearly C0(Ω)- linear", in the sense that the induced map θ : E → F is a C0(Ω)-module map (where F is the image of F in the space of C0-sections on the canonical "(H)-Banach bundle" associated with F ; see Section 2). Moreover, θ is "nearly bounded" in the sense that there exists a finite subset S ⊆ Ω such that sup ω∈Ω\S sup e∈E; kek≤1(cid:13)(cid:13)(cid:13)θ(e)(ω)(cid:13)(cid:13)(cid:13) < ∞. θ = θ0 ⊕Mω∈S θω Furthermore, if F is "C0(Ω)-normed" (in particular, if F is a Hilbert C0(Ω)-module), then the finite set S consists of isolated points in Ω, and where θ0 : EΩ\S → FΩ\S is a bounded C0(Ω \ S)-linear map (where EΩ\S and FΩ\S are the canonical essential Banach C0(Ω \ S)-modules induced from E and F respectively) AUTOMATIC CONTINUITY AND C0(Ω)-LINEARITY 3 and θω are (unbounded) C-linear maps (see Theorems 4.2 and 3.7). Consequently, if Ω contains no isolated point and F is C0(Ω)-normed, then θ is automatically bounded. As another application, if X and Y are two Banach spaces and if θk : X → Y is a sequence of C-linear maps (not assumed to be bounded) such that for any (xn) ∈ c0(X), we have (θn(xn)) ∈ c0(Y ), then there exists n0 with kθnk < ∞. sup n≥n0 On the other hand, we will also study C-linear maps between two Banach modules over two different base spaces. In this case, we will consider "separating" maps instead of local maps. More precisely, if Ω and ∆ are two locally compact Hausdorff spaces, E is a "full" essential Banach C0(Ω)-module (see Remark 3.2(b)), and F is a "full" Banach C0(∆)-normed module, then for any bijective linear map θ : E → F (not assumed to be bounded) with both θ and θ−1 being separating, there exists a homeomorphism σ : ∆ → Ω such that θ(e · ϕ) = θ(e) · ϕ ◦ σ (e ∈ E, ϕ ∈ C0(Ω)), and there exists a finite set S consisting of isolated points of ∆ such that the restriction of θ from EΩ\σ(S) to F∆\S is bounded. This paper is organised as follows. In Section 2, we will first collect some basic facts about the correspondence between Banach bundles and Banach C0(Ω)-modules. In Sec- tion 3, we will show two technical lemmas concerning "near C0(Ω)-linearity" and "near boundedness" of certain mappings. Section 4 is devoted to automatic C0(Ω)-linearity and automatic boundedness of local linear mappings, while Section 5 is devoted to the au- tomatic boundedness of bijective biseparating linear mappings between Banach modules over different base spaces. Finally, as an attempt to a further generalisation, we show in the Appendix that for an arbitrary C-algebra A, every bounded local linear map from a Banach A-module into a Hilbert A-module is A-linear. The boundedness assumption can be removed in the case when A is finite dimension (Corollary 4.9). 2. Preliminaries and Notations Let us first recall (mainly from [10]) some basic terminologies and results concerning Banach modules and Banach bundles. Notation 2.1. In this article, Ω and ∆ are two locally compact Hausdorff spaces, E is an essential Banach C0(Ω)-module, F is an essential Banach C0(∆)-module, and θ : E → F is a C-linear map (not assumed to be bounded). Furthermore, Ω∞ and ∆∞ are the one- point compactifications of Ω and ∆ respectively. We denote by NΩ(ω) the set of all compact neighbourhoods of an element ω in Ω, and by IntΩ(S) the set of all interior points of a subset S in Ω. Moreover, if U, V ⊆ Ω such that the closure of V is a compact subset of IntΩ(U), we denote by UΩ(V, U) the collection of all λ ∈ Cc(Ω) with 0 ≤ λ ≤ 1, λ ≡ 1 on V and the support of λ lies inside IntΩ(U). 4 CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG Definition 2.2. Let Ξ be a Hausdorff space and p : Ξ → Ω be a surjective continuous open map. Suppose that for each ω ∈ Ω, 1). there exists a complex Banach space structure on Ξω := p−1(ω) such that its norm topology coincides with the topology on Ξω (as a topological subspace of Ξ); 2). {W (ǫ, U) : ǫ > 0, U ∈ NΩ(ω)} forms a neighbourhood basis for the zero element 0ω ∈ Ξω where W (ǫ, U) := {ξ ∈ p−1(U) : kξk < ǫ}; 3). the maps C × Ξ → Ξ and {(ξ, η) ∈ Ξ × Ξ : p(ξ) = p(η)} → Ξ given respectively, by the scalar multiplications and the additions are continuous. Then (Ξ, Ω, p) (or simply, Ξ) is called an (H)-Banach bundle (respectively, an (F)-Banach bundle) over Ω if ξ 7→ kξk is an upper-semicontinuous (respectively, continuous) map from Ξ to R+. In this case, Ω is called the base space of Ξ, the map p is called the canonical projection and Ξω is called the fibre over ω ∈ Ω. If Ξ is an (H)-Banach bundle over Ω and Ω0 ⊆ Ω is an open set, then ΞΩ0 := p−1(Ω0) is an (H)-Banach bundle over Ω0 and is called the restriction of Ξ to Ω0. (F)-Banach bundle, then so is ΞΩ0. If Ξ is an Definition 2.3. If Λ is an (H)-Banach bundle over ∆, a map ρ : Ξ → Λ is called a fibrewise linear map covering a map σ : Ω → ∆ if ρ(Ξω) ⊆ Λσ(ω) and the restriction ρω : Ξω → Λσ(ω) is linear. Moreover, a fibrewise linear map ρ covering a continuous map σ : Ω → ∆ is called a Banach bundle map if ρ is continuous. A Banach bundle map ρ is said to be bounded if sup ξ∈Ξ; kρ(ξ)k < ∞. kξk≤1 For any map e : Ω → Ξ, we denote e(ω) := ke(ω)k (ω ∈ Ω). Such an e is called a C0-section on Ξ if e is continuous, p(e(ω)) = ω (ω ∈ Ω), and for any ǫ > 0, there exists a compact set C ⊆ Ω such that e(ω) < ǫ (ω ∈ Ω \ C). We put Γ0(Ξ) := {e : Ω → Ξ e is a C0−section on Ξ}. Note that e is always upper semi-continuous for every e ∈ Γ0(Ξ) and Ξ is an (F)-Banach bundle if and only if all such e are continuous. Next, we recall some terminologies and properties concerning an essential Banach (right) C0(Ω)-module E (regarded as a unital Banach C(Ω∞)-module). For any ω ∈ Ω∞ and S ⊆ Ω∞, we denote KS := {ϕ ∈ C(Ω∞) : ϕ(S) = {0}}, K E S := E · KS and I E ω := [V ∈NΩ∞ (ω) K E V . AUTOMATIC CONTINUITY AND C0(Ω)-LINEARITY 5 ω := K E For simplicity, we set K E ∞ = E because E is an essential Banach C0(Ω)-module. By [10, p.37], there exists an (H)-Banach bundle ΞE over Ω∞ ∞ = {0}, if we set ΞE := p−1(Ω), then Γ0(ΞE) ∼= Γ0(ΞE) with ΞE under the canonical identification. Furthermore, there exists a contraction {ω}. Note that K E ω . Since ΞE ω = E/K E such that e(ω) = e + K E ω . We put E to be the closure of the image of ∼. On the other hand, if θ is as in Notation 2.1, we define ∼ : E −→ Γ0(ΞE) θ : E → F by (e ∈ E). θ(e) = gθ(e) Definition 2.4. Let E be an essential Banach C0(Ω)-module. (a) E is called a Banach C0(Ω)-convex module if for any ϕ, ψ ∈ C(Ω∞)+ with ϕ + ψ = 1, one has kxϕ + yψk ≤ max{kxk, kyk}. (b) E is called a Banach C0(Ω)-normed module if there exists a map · : E → C0(Ω)+ such that for any x, y ∈ X and a ∈ A, i). x + y ≤ x + y; ii). xa = xa; iii). kxk = kxk. Recall that every Hilbert C0(Ω)-module is a Banach C0(Ω)-normed module, and every Banach C0(Ω)-normed module is C0(Ω)-convex. On the other hand, an essential Banach C0(Ω)-module E is C0(Ω)-convex if and only if ∼ is an isometric isomorphism onto Γ0(ΞE) (see e.g. [10, Theorem 2.5]). In this case, we will not distinguish E and Γ0(ΞE). Furthermore, E is C0(Ω)-normed if and only if E is C0(Ω)-convex and ΞE is an (F)- Banach bundle (see e.g. [10, p.48]). For any open subset Ω0 ⊆ Ω, we set EΩ0 := K E Ω0). One can regard as an essential Banach C0(Ω0)-module under the identification C0(Ω0) ∼= KΩ\Ω0. and EΩ0 := Γ0(ΞE Ω\Ω0 Ω\Ω0 K E Note that if E is C0(Ω)-convex, then EΩ0 = EΩ0. Remark 2.5. (a) Let E be a Banach C0(Ω)-convex module and 0ω is the zero element in the fibre ΞE ω (ω ∈ Ω). It is well-known that ω 7→ 0ω is a continuous map from Ω into ωi, then e ∈ K E ω0. ΞE. Thus, if {ωi}i∈I is a net in Ω converging to ω0 ∈ Ω and e ∈Ti∈I K E ω , there exists U ∈ NΩ(ω) such that e /∈ K E Consequently, if e /∈ K E (b) For any ω ∈ Ω and e ∈ K E ke − eV k → 0. (c) Let Ω = {ω1, ω2, ...} be a countable compact Hausdorff space and E be a Banach C(Ω)-module. Then α for any α ∈ U. ω , there exists a net {eV }V ∈NΩ(ω) such that eV ∈ K E V and K E ω = {0}, \ω∈Ω 6 CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG ω and any ǫ > 0. For k ∈ N, there exists ¯ϕk ∈ K{ωk} with ke−e ¯ϕkk < ǫ/2k+1. Thus, there exists ϕk ∈ C(Ω) with ϕk vanishing on an open neighbourhood Vk of ωk and ke − eϕkk < ǫ/2k. Now, consider a finite subcover {V1, ..., Vn} for Ω and a continuous partition of unity {ψ1, ...ψn} or equivalently, the map ∼ is injective. In fact, consider any e ∈Tω∈Ω K E subordinated to {V1, ..., Vn}. Then kek = ke − ePn k=1 ϕkψkk ≤Pn k=1 ke − eϕkk < ǫ. 3. Some technical results In this section, we will give two technical lemmas (3.3 and 3.6) which are the crucial ingredients for all the results in this paper. Before presenting them, let us give another automatic continuity type lemma that is needed for those two essential lemmas. Lemma 3.1. Zθ := {ν ∈ ∆ : θ(e)(ν) = 0 for all e ∈ E} is a closed subset (where θ is as in Section 2). Moreover, if σ : ∆θ → Ω∞ (where ∆θ := ∆ \ Zθ) is a map satisfying θ(I E ν (ν ∈ ∆θ), then σ is continuous. σ(ν)) ⊆ K F It follows from Remark 2.5(a) that Zθ is closed. Suppose on the contrary, that Proof: there exists a net {νi}i∈I in ∆θ that converges to ν0 ∈ ∆θ but σ(νi) 9 σ(ν0). Then there are U, W ∈ NΩ∞(σ(ν0)) with U ⊆ IntΩ∞(W ) and {i ∈ I : σ(νi) /∈ IntΩ(W )} being cofinal. As Ω∞ is compact, by passing to a subnet if necessary, we can assume that {σ(νi)} converges to an element ω ∈ Ω∞, and there exists V ∈ NΩ∞(ω) with V ∩ U = ∅. Pick any e ∈ E and ϕ ∈ UΩ∞(V, Ω∞ \ U). Since σ(νi) → ω, we see that e(1 − ϕ) ∈ I E when i is large enough and so eventually, σ(νi) θ(e(1 − ϕ))(νi) = 0 (by the hypothesis). By Remark 2.5(a), we see that θ(e(1 − ϕ))(ν0) = 0. On the other hand, we have θ(eϕ) ∈ K F ν0, which gives the contradiction that ν0 ∈ Zθ. (cid:3) σ(ν0)) and θ(e) ∈ K F ν0 (because eϕ ∈ I E Remark 3.2. (a) Note that for any ν ∈ Zθ, one has (3.1) θ(I E ω ) ⊆ K F ν (ω ∈ Ω). σ(ν)) ⊆ K F ν (ν ∈ ∆) but one should not expect such σ to be continuous. Consequently, if we extend σ in Lemma 3.1 by setting σ(ν) arbitrarily for each ν ∈ Zθ, then θ(I E (b) θ is said to be full if Zθ = ∅. Moreover, E is said to be full if id : E → E is full (or equivalently, E 6= K E (c) One can use our proof for Lemma 3.1 to give the following (probably known) result: ω for any ω ∈ Ω). Suppose that σ : ∆ → Ω is a map and Φ : C0(Ω) → Cb(∆) is a C-linear map satisfying Φ(λ · ψ) = Φ(λ) · (ψ ◦ σ) (λ, ψ ∈ C0(Ω)), and for any ν ∈ ∆, there exists λ ∈ C0(Ω) with Φ(λ)(ν) 6= 0. Then σ is continuous. AUTOMATIC CONTINUITY AND C0(Ω)-LINEARITY 7 Lemma 3.3. Let σ : ∆θ → Ω be a map satisfying θ(I E (a) If Uθ :=nν ∈ ∆ : supkek≤1 kθ(e)(ν)k = ∞o, then Uθ ⊆ ∆θ, σ(ν)) ⊆ K F sup kθ(e)(ν)k < ∞ ν (ν ∈ ∆θ). ν∈∆\Uθ;kek≤1 (we use the convention that sup ∅ = 0) and σ(Uθ) is a finite set. (b) If Nθ,σ :=nν ∈ ∆θ : θ(K E σ(ν)) K F νo, then Nθ,σ ⊆ Uθ and σ(Nθ,σ) consists of non- isolated points in Ω. (c) If, in addition, σ is an injection sending isolated points in ∆θ to isolated points in Ω, then θ(e · ϕ) = θ(e) · ϕ ◦ σ (e ∈ E, ϕ ∈ C0(Ω)). ν∈∆ ΞF ν . Proof: For every ν ∈ ∆ \ Uθ, one can regard e 7→ θ(e)(ν) as a bounded C-linear map from E (a) The first conclusion is clear. We put Y to be the c0-direct sumLc0 second conclusion. Assume now that σ(Uθ) is infinite. For n = 1, we can find ν1 ∈ ∆ as into Y (note that(cid:13)(cid:13)(cid:13)θ(e)(ν)(cid:13)(cid:13)(cid:13) ≤ kθ(e)k), the uniform boundedness principle will give the well as e1 ∈ E with ke1k ≤ 1 and(cid:13)(cid:13)(cid:13)θ(e1)(ν1)(cid:13)(cid:13)(cid:13) > 1. Inductively, we can find νn ∈ ∆ and en ∈ E such that kenk ≤ 1 and kθ(en)(νn)k > n3. σ(νn) 6= σ(νk) (k = 1, ..., n − 1), There exist n1 ∈ N and U1 ∈ NΩ(σ(νn1)) such that {n ∈ N : n > n1 and σ(νn) /∈ U1} is infinite. Inductively, we can find a subsequence {νnk} and Uk ∈ NΩ(σ(νnk)) (k ∈ N) such that Uk ∩ Ul = ∅ for distinct k, l ∈ N. Without loss of generality, we assume that nk = k. Pick Vn ∈ NΩ(σ(νn)) such that Vn is subset of IntΩ(Un). Consider λn ∈ UΩ(Vn, Un) ekλ2 (n ∈ N) and notice that kenλ2 k2 ∈ E and take n ∈ N. Since k k=1 nk ≤ 1. Define e :=P∞ k2 ! Xk6=n n = n2 Xk6=n ekλk λk! ∈ K E Un, n2e − enλ2 (cid:13)(cid:13)(cid:13)θ(e)(ν)(cid:13)(cid:13)(cid:13) = (cid:13)(cid:13)(cid:13)θ(e − eV )(ν)(cid:13)(cid:13)(cid:13) ≤ κ ke − eV k , we have n2 θ(e)(νn) = θ(enλ2 en(1 − λ2 Vn, we have, n) ∈ K E n)(νn) (by the hypothesis). On the other hand, as en −enλ2 n = (cid:13)(cid:13)(cid:13)θ(e)(cid:13)(cid:13)(cid:13) ≥ (cid:13)(cid:13)(cid:13)θ(e)(νn)(cid:13)(cid:13)(cid:13) = 1 n2(cid:13)(cid:13)(cid:13)θ(enλ2 n)(νn)(cid:13)(cid:13)(cid:13) = which contradicts the finiteness of kθ(e)k. 1 n2(cid:13)(cid:13)(cid:13)θ(en)(νn)(cid:13)(cid:13)(cid:13) > n, (b) Consider ν ∈ ∆ \ Uθ and denote κ := supkek≤1(cid:13)(cid:13)(cid:13)θ(e)(ν)(cid:13)(cid:13)(cid:13) < ∞. Pick any e ∈ K E V (V ∈ NΩ(σ(ν))) with keV − ek → 0 (Remark 2.5(b)). As θ(eV ) ∈ K F and eV ∈ K E has σ(ν) ν , one 8 CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG which shows that ν ∈ ∆ \ Nθ,σ. Now, if σ(ν) is an isolated point in Ω, then {σ(ν)} ∈ NΩ(σ(ν)), and we have the contradiction that θ(K E ν . This gives second state- ment. (c) For any ν ∈ ∆ \ Nθ,σ and e ∈ E, we have eϕ − eϕ(σ(ν)) = e(ϕ − ϕ(σ(ν))1) ∈ K E Thus, σ(ν)) ⊆ K F σ(ν). (3.2) θ(eϕ)(ν) = θ(e)(ν)ϕ(σ(ν)) (e ∈ E, ν ∈ ∆ \ Nθ,σ). In particular, (3.2) is true when ν ∈ ∆ \ Uθ (by part (b)) or when ν ∈ Uθ is an isolated point of ∆θ (by the hypothesis as well as part (b)). Suppose that ν ∈ Uθ is a non-isolated point of ∆θ. As σ is injective, part (a) implies that Uθ is a finite set. Hence, there exists a net {νi} in ∆θ \ Uθ converging to ν. Now, by Lemma 3.1, θ(eϕ)(ν) = lim θ(eϕ)(νi) = lim θ(e)(νi)ϕ(σ(νi)) = θ(e)(ν)ϕ(σ(ν)). (cid:3) Remark 3.4. Note that since Zθ is closed, isolated points in ∆θ are the same as isolated points of ∆. Moreover, for any ν ∈ Zθ, we have supkek≤1 kθ(e)(ν)k = 0, and (3.1) holds. Therefore, Lemma 3.3 remains valid if we replace all the ∆θ with ∆ (in fact, the current form is stronger as any injection on ∆ restricted to an injection on ∆θ). The same is true for all the remaining results in this section. If σ is injective, then Uθ is finite and we have our first nearly automatically boundedness result which states that if θ is a "module map through an injection σ : ∆ → Ω" (one can relax this slightly to an injection on ∆θ), then θ is "bounded after taking away finite number of points from ∆". Proposition 3.5. Let Ω and ∆ be two locally compact Hausdorff spaces. Let E and F be an essential Banach C0(Ω)-module and an essential Banach C0(∆)-module respectively, and let θ : E → F be a C-linear map (not assumed to be bounded). Suppose that σ : ∆θ → Ω is an injection satisfying θ(e · ϕ)(ν) = θ(e)(ν)ϕ(σ(ν)) (e ∈ E, ϕ ∈ C0(Ω), ν ∈ ∆θ). Then there exist a finite subset T ⊆ ∆ and κ > 0 such that kθ(e)(ν)k ≤ κkek (e ∈ E). sup ν∈∆\T Lemma 3.6. Let σ : ∆θ → Ω be a map satisfying θ(I E addition, that F is a Banach C0(∆)-normed module. (a) Nθ,σ is an open subset of ∆. (b) If σ is injective, then Uθ is a finite set consisting of isolated points of ∆. σ(ν)) ⊆ K F ν (ν ∈ ∆θ). Suppose, in If, in addition, Uθ 6= ∆, then F = F∆\Uθ ⊕Lν∈Uθ ΞF ν and θ0 := Pθ,σ ◦ θEΩ\σ(Uθ) : EΩ\σ(Uθ ) → F∆\Uθ AUTOMATIC CONTINUITY AND C0(Ω)-LINEARITY 9 is a bounded linear map (where Pθ,σ : F → F∆\Uθ is the canonical projection) such that (3.3) (e ∈ EΩ\σ(Uθ ), ϕ ∈ C0(Ω \ σ(Uθ))) θ0(e · ϕ) = θ0(e) · ϕ ◦ σ (note that the value of σ on Zθ can be set arbitrarily). Proof: Notice, first of all, that as F is C0(Ω)-convex, one can regard θ = θ. (a) As ∆θ is open in ∆ and Nθ,σ ⊆ Uθ ⊆ ∆θ, it suffices to show that Nθ,σ is open in ∆θ. By Lemma 3.3(a), κ := sup ν /∈Uθ sup kek≤1 kθ(e)(ν)k < ∞. Let {νi}i∈I be a net in ∆θ \ Nθ,σ converging to ν0 ∈ ∆θ, and e be an arbitrary element in K E σ(ν0). By Lemma 3.1, we know that σ(νi) → σ(ν0). Now, we consider two cases separately. The first case is when {σ(νi)}i∈I is finite. In this case, by passing to subnet, we can assume that σ(νi) = σ(ν0) (i ∈ I). As e(σ(ν0)) = 0 and νi /∈ Nθ,σ, we have θ(e)(νi) = 0 which gives θ(e)(ν0) = 0, and so, θ(e) ∈ K F ν0. The second case (of {σ(νi)}i∈I being infinite) can be subdivided into two cases. More precisely, if there exists i0 ∈ I such that νj ∈ Uθ for every j ≥ i0, then we can assume that {σ(νi)}i∈I ⊆ σ(Uθ) which is a finite set, and the above implies that θ(e) ∈ K F ν0. Otherwise, {i ∈ I : νi /∈ Uθ} is cofinal, and by passing to a subnet, we may assume that νi /∈ Uθ (i ∈ I). For any ǫ > 0, pick V ∈ NΩ(σ(ν0)) and eV ∈ K E V with keV − ek < ǫ. When i is large enough, σ(νi) ∈ V and eV (σ(νi)) = 0. Thus, kθ(e)(νi)k = kθ(e − eV )(νi)k ≤ κǫ. By the continuity of the norm function on ΞF , we have kθ(e)(ν0)k ≤ κǫ which implies that θ(e)(ν0) = 0. (b) By the hypothesis and Lemma 3.3(a), one knows that Uθ is finite. Without loss of generality, we assume ∆ 6= Uθ. Let (3.4) κ := sup ν∈∆\Uθ sup kek≤1 kθ(e)(ν)k < ∞. Suppose on the contrary that there is ν0 ∈ Uθ which is not an isolated point in ∆. As Uθ is finite, there is a net {νi} in ∆ \ Uθ such that νi → ν0. By the definition of Uθ, there is e ∈ E with kek ≤ 1 and kθ(e)(ν0)k > κ + 1. However, this will contradict the continuity of θ(e) (because of (3.4)). Now, as Uθ is a finite set consisting of isolated points in ∆ and F is the space of C0-sections on ΞF , we see that By Lemma 3.3(b) and the argument of Lemma 3.3(c) (more precisely, (3.2)), one easily check that θ0 will satisfy (3.3). On the other hand, the boundedness θ0 follows from (3.4). (cid:3) F = K F Uθ ⊕Mν∈Uθ ΞF ν . 10 CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG Observe that in Lemmas 3.3(c) and 3.6(b), one can replace the injectivity of σ with the condition that σ−1(ω) is at most finite for any ω ∈ Ω. The following is our second nearly automatically boundedness result that applies, in particular, when F is a Hilbert C0(∆)-module. Theorem 3.7. Let Ω and ∆ be two locally compact Hausdorff spaces. Let E be an essential Banach C0(Ω)-module, let F be an essential Banach C0(∆)-normed module, and let θ : E → F be a C-linear map (not assumed to be bounded). Suppose that σ : ∆θ → Ω is an injection satisfying θ(I E (a) If ∆ contains no isolated point, then θ is bounded. (b) If σ sends isolated points in ∆θ to isolated points in Ω, then Nθ,σ = ∅ and there exist a finite set T consisting of isolated points of ∆, a bounded linear map θ0 : EΩ\σ(T ) → F∆\T as well as linear maps θν : ΞE σ(ν), ν for all ν ∈ T such that E = EΩ\σ(T ) ⊕Lν∈T ΞE σ(ν)) ⊆ K F ν (ν ∈ ∆). F = F∆\T ⊕Lν∈T ΞF ν and θ = θ0 ⊕Lν∈T θν. σ(ν) → ΞF (a) This follows directly from Lemma 3.6(b). Proof: (b) The first conclusion follows from Lemma 3.3(c) and the second conclusion follows from Lemma 3.6(b) (notice that we have a sharper conclusion here since Nθ,σ = ∅). (cid:3) 4. Applications to local linear mappings In the section, we will consider the case when ∆ = Ω, σ = id, and the C-linear map θ is a local map in the sense that θ(e) · ϕ = 0 whenever e ∈ E and ϕ ∈ C0(Ω) satisfying e · ϕ = 0. It is obvious that any C0(Ω)-module map is local. Remark 4.1. Suppose that θ is local. Let U, V ⊆ Ω be open sets with the closure of V being a compact subset of U, and consider λ ∈ UΩ(V, U). For any e ∈ K E U and any ǫ > 0, there exists ϕ ∈ KU with ke − eϕk < ǫ. Thus, eλ = 0 which implies that θ(e)λ = 0 and θ(e) = θ(e)(1 − λ) ∈ K F V . This shows that σ = id will satisfy the hypothesis in all the results in Section 3. The following theorem (which follows directly from the results in Section 3 as well as Remark 4.1) is our main result concerning local linear maps. Theorem 4.2. Let Ω be a locally compact Hausdorff space. Suppose that E and F are essential Banach C0(Ω)-modules, and θ : E → F is a local C-linear map (not assumed to be bounded). (a) θ is a C0(Ω)-module map and there exist a finite subset T ⊆ ∆ and κ > 0 such that supν∈∆\T kθ(e)(ν)k ≤ κkek (e ∈ E). AUTOMATIC CONTINUITY AND C0(Ω)-LINEARITY 11 (b) If, in addition, F is a Banach C0(Ω)-normed module, then θ is a C0(Ω)-module map and there exist a finite set T consisting of isolated points of Ω, a bounded linear map θ0 : EΩ\T → FΩ\T as well as a linear map θν : ΞE ν for each ν ∈ T such that ν → ΞF E = EΩ\T ⊕Lν∈T ΞE ν , F = FΩ\T ⊕Lν∈T ΞF ν and θ = θ0 ⊕Lν∈T θν. It is natural to ask if one can relax the assumption of F being C0(Ω)-normed to C0(Ω)-convex in the second statement of Theorem 4.2 (in particular, whether it is true that every C0(Ω)-module map from an essential Banach C0(Ω)-module to an essential Banach C0(Ω)-convex module is automatically bounded provided that Ω contains no isolated point). Unfortunately, it is not the case as can be seen by the following simple example. Example 4.3. Let E := C([0, 1]) ⊕∞ X and F := C([0, 1]) ⊕∞ Y , where X and Y are two infinite dimensional Banach spaces. Then E is an essential Banach C([0, 1])-convex module under the multiplication: (e, x)·ϕ = (e·ϕ, xϕ(0)) (e, ϕ ∈ C([0, 1]); x ∈ X). In the same way, F is an essential Banach C([0, 1])-convex module. Suppose that R : X → Y is an unbounded linear map and θ : E → F is given by θ(e, x) = (e, R(x)) (e ∈ C([0, 1]); x ∈ X). Then θ is a C([0, 1])-module map which is not bounded (as its restriction on X is R). In this case, we have Uθ = {0}. Corollary 4.4. Let Ω be a locally compact Hausdorff space. Any local C-linear θ from an essential Banach C0(Ω)-module to a Hilbert C0(Ω)-module is a C0(Ω)-module map. Moreover, if Ω contains no isolated point, then any such θ is automatically bounded. Remark 4.5. Let LC0(Ω)(E; C0(Ω)) (respectively, BC0(Ω)(E; C0(Ω))) be the "algebraic dual" (respectively, "topological dual") of E, i.e. the collection of all C0(Ω)-module maps (respectively, bounded C0(Ω)-module maps) from E into C0(Ω). An application of Corol- lary 4.4 is that the algebraic dual and the topological dual of E coincide in many cases: If Ω is a locally compact Hausdorff space having no isolated point and E is an essential Banach C0(Ω)-module, then BC0(Ω)(E; C0(Ω)) = LC0(Ω)(E; C0(Ω)). Corollary 4.6. Let Ξ and Λ be respectively an (H)-Banach bundle and an (F)-Banach bundle over the same base space Ω. If ρ : Ξ → Λ is a fibrewise linear map covering id (without any boundedness nor continuity assumption) such that ρ ◦ e ∈ Γ0(Λ) for every e ∈ Γ0(Ξ), then there exists a finite subset S ⊆ Ω consisting of isolated points such that ρ restricts to a bounded Banach bundle map ρ0 : ΞΩ\S → ΛΩ\S. Let X be a Banach space. We denote by ℓ∞(X) and c0(X) the set of all bounded sequences and the set of all c0-sequences in X, respectively. We recall that ℓ∞ ∼= C(βN) where βN is the Stone-Cech compactification of N (which can be identified with the collection of all ultrafilters on N). 12 CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG Proposition 4.7. Let X and Y be Banach spaces, and let θk : X → Y (k ∈ N ∪ {∞}) be linear maps (not assumed to be bounded). For any sequence {xk}k∈N in X, we put θ({xk}k∈N) := {θk(xk)}k∈N. (a) If θ(c0(X)) ⊆ c0(Y ), then there exists n0 ∈ N such that supn≥n0 kθnk < ∞. (b) If limk→∞ θk(xk) = θ∞(x) for any {xk}k∈N ∈ ℓ∞(X) with limk→∞ xk = x, then θ∞ is bounded, and there is n0 ∈ N such that supn≥n0 kθnk < ∞. (c) Suppose that θ(ℓ∞(X)) ⊆ ℓ∞(Y ), and limF θk(xk) = 0 for every {xk}k∈N ∈ ℓ∞(X) and every ultrafilter F on N with limF xk = 0. Then there exist F1, ..., Fn ∈ βN with supF6=F1,...,Fn kθFk < ∞ (where θF : Ξℓ∞(X) is the induced map). In particular, supn≥n0 kθnk < ∞ for some n0 ∈ N. → Ξℓ∞(Y ) F F (a) Let E = c0(X) and F = c0(Y ). Then θ is a C0(N)-module map and we can Proof: apply Theorem 4.2. (b) Let E = C(N∞, X) and F = C(N∞, Y ). Then θ ⊕θ∞ is a well defined C(N∞)-module map from E into F and Theorem 4.2 implies this part. (c) Let E = ℓ∞(X) and F = ℓ∞(Y ). Then E and F are unital Banach C(βN)-modules. For any ultrafilter F ∈ βN, one has K E F = {(xn) ∈ E : lim F xn = 0} and K F F = {(yn) ∈ F : lim yn = 0}. F The first hypothesis shows that θ(E) ⊆ F and the second one tells us that θ(K E On the other hand, if n ∈ N and Fn := {U ⊆ N : n ∈ U}, then F ) ⊆ K F F . Fn = {(xk) ∈ ℓ∞(X) : xn = 0} and so, θFn = θn. Now, this part follows from Theorem 4.2. K E (cid:3) Remark 4.8. Note that if F is a free ultrafilter on N, then Ξℓ∞(X) can be identified with the ultrapowers X F and Y F of X and Y (over F) respectively. One can interpret Proposition 4.7(c) as follows: and Ξℓ∞(Y ) F F If the sequence {θn} as in Proposition 4.7 induces canonically a map θ : ℓ∞(X) → ℓ∞(Y ) as well as a map θF : X F → Y F for every free ultrafilter F (none of them assumed to be bounded), then for all but a finite number of ultrafilters F, the map θF is bounded and they have a common bound. It can be shown easily that the converse of the above is also true (but we left it to the readers to check the details): If the sequence {θn} is as in Proposition 4.7 and there exists n0 ∈ N with supn≥n0 kθnk < ∞, then {θn} induces canonically a map from ℓ∞(X) to ℓ∞(Y ) as well as a map from X F to Y F for every free ultrafilter F. AUTOMATIC CONTINUITY AND C0(Ω)-LINEARITY 13 Another important point in Theorem 4.2 is the automatic C0(Ω)-linearity. In fact, it can be shown that for every C ∗-algebra A, any bounded local linear map from a Banach right A-module into a Hilbert A-module is automatically A-linear (see Proposition A.1 in the Appendix). Theorem 4.2 tells us that if A is commutative, then one can relax the assumption of the range space to a Banach A-convex module and one can remove the boundedness assumption. Another application of this theorem is that if A is a finite dimensional C ∗-algebras, then every local linear map between any two Banach right A-modules is A-linear. Corollary 4.9. Let A be a finite dimensional C ∗-algebra. Suppose that E and F are unital Banach right A-modules. If θ : E → F is a local C-linear map in the sense of Proposition A.1 (not assumed to be bounded), then θ is an A-module map. Proof: Pick any x ∈ E and a ∈ Asa. Let Aa := C ∗(a, 1). By Remark 2.5(c), both E and F are unital Banach Aa-convex modules. Thus, Theorem 4.2 tell us that θ is a Aa-module map. In particular, θ(xa) = θ(x)a. (cid:3) Remark 4.10. (a) Suppose that A is a unital C ∗-algebra and F is a unital Banach right A-convex module in the sense kxa + y(1 − a)k ≤ max{kxk, kyk} for x, y ∈ F and a ∈ A+ with a ≤ 1. Then, by the argument of Corollary 4.9, all local linear maps from any unital Banach right A-module into F are automatically A-linear. (b) If one can show that for every compact subset Ω ⊆ R and every essential Banach C(Ω)-module F , the map ∼: F → F is injective, then using the argument of Corollary 4.9, one can show that for each C ∗-algebra A, all local linear maps between any two Banach right A-modules are A-module maps (without assuming that θ is bounded). However, we do not know if it is true. 5. Applications to separating mappings In this section, we consider Ω and ∆ to be possibly different spaces. In this case, one cannot define local property any more, but one has a weaker natural property called separating. More precisely, θ is said to be separating if θ(e) · θ(g) = 0, whenever e, g ∈ E satisfying e · g = 0. In the case when E = C0(Ω) and F = C0(∆), this coincides with the well-known notion of disjointness preserving (see e.g. [1, 5, 18, 14, 12, 15]). Lemma 5.1. If θ is separating, there is a continuous map σ : ∆θ → Ω∞ such that θ(I E ν (ν ∈ ∆θ). σ(ν)) ⊆ I F 14 CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG Proof: Set Sν := {ω ∈ Ω∞ : θ(I E ω ) ⊆ I F ν } (ν ∈ ∆θ). Uω with θ(eω) /∈ I F Suppose there is ν ∈ ∆θ with Sν = ∅. Then for each ω ∈ Ω∞, there exist Uω ∈ NΩ∞(ω) and eω ∈ K E i=1 be a partition of unity subordinate to {Uωi}n obtain θ(gϕi)θ(eωi) = 0, which implies that θ(gϕi)(ν) = 0 (because of Remark 2.5(a) and the fact that θ(eωi) /∈ I F i=1. Take any g ∈ E. From fgϕifeωi = 0, we i=1 be a finite subcover of {Uω}ω∈Ω∞ and {ϕi}n ν ). Consequently, ν . Let {Uωi}n θ(g)(ν) = θ(gϕi)(ν) = 0, nXi=1 and we arrive in the contradiction that ν ∈ Zθ. Suppose there is ν ∈ ∆θ with Sν containing two distinct points ω1 and ω2. Let U, V ∈ NΩ∞(ω1) with V ⊆ IntΩ∞(U) and ω2 /∈ U. For any ϕ ∈ UΩ∞(V, U) and e ∈ E, we have e(1 − ϕ) ∈ I E ω2 which implies that ω1 and eϕ ∈ I E θ(e) = θ(e(1 − ϕ)) + θ(eϕ) ∈ I F ν . This gives the contradiction that ν ∈ Zθ. Therefore, we can define σ(ν) to be the only ν . Now, the continuity of σ follows from (cid:3) point in Sν, and it is clear that θ(cid:16)I E σ(ν)(cid:17) ⊆ I F Lemma 3.1. Corollary 5.2. Let Ξ be an (H)-Banach bundle over Ω, let Λ be an (F)-Banach bundle over ∆, and let ρ : Ξ → Λ be a map (not assumed to be bounded nor continuous). Suppose that σ : ∆ → Ω is an injection sending isolated points in ∆ to isolated points in Ω such that e 7→ ρ ◦ e ◦ σ defines a linear map θ : Γ0(Ξ) → Γ0(Λ). Then there exists a finite set T consisting of isolated points of ∆ such that the restriction of ρ induces a bounded Banach bundle map ρ0 : ΞΩ\σ(T ) → Λ∆\T (covering σ∆\T ). Moreover, σ is continuous on ∆ \ Zρ,σ where Zρ,σ := {ν ∈ ∆ : ρ(e(σ(ν))) = 0 for all e ∈ E}. Proof: The first conclusion follows from Theorem 3.7. To see the second conclusion, we note that θ is separating and we can apply Lemma 5.1 (observe that Zρ,σ = Zθ). (cid:3) Theorem 5.3. Let Ω and ∆ be two locally compact Hausdorff spaces, and let E be a full essential Banach C0(Ω)-module (see Remark 3.2(b)) and F be a full essential Banach C0(∆)-normed module. Suppose that θ : E → F is a bijective C-linear map (not assumed to be bounded) such that it is biseparating in the sense that both θ and θ−1 are separating. (a) There exists a homeomorphism σ : ∆ → Ω satisfying θ(e · ϕ) = θ(e) · ϕ ◦ σ (e ∈ E; ϕ ∈ C0(Ω)). (b) There exists isolated points ν1, ..., νn ∈ ∆ such that the restriction of θ induces a Banach space isomorphism θ0 : EΩθ → F∆θ, where ∆θ := ∆\{ν1, ..., νn} and Ωθ := σ(∆θ). AUTOMATIC CONTINUITY AND C0(Ω)-LINEARITY 15 Proof: (a) If e ∈ E with e = 0, then θ(e) = θ(e) = 0 (as θ is separating and F is well. The fullness of E and F as well as the surjectivity of θ and θ−1 ensure that Zθ = ∅ and Zθ−1 = ∅. Therefore, by Lemma 5.1, we have two continuous maps C0(∆)-convex), which gives e = 0 (as θ is injective). Hence, one can regard fθ−1 = θ−1 as such that θ−1(cid:16)I F ω (ω ∈ Ω) and θ(cid:16)I E ν ∈ ∆0 := σ−1(Ω) and ω ∈ Ω0 := τ −1(∆), we have ν (ν ∈ ∆). Consequently, for any τ (ω)(cid:17) ⊆ I E σ(ν)(cid:17) ⊆ I F τ : Ω → ∆∞ and σ : ∆ → Ω∞ σ(τ (ω)) = ω and τ (σ(ν)) = ν (because I E σ(τ (ω)) ⊆ I E ω , I F τ (σ(ν)) ⊆ I F ν , and E as well as F are full). ν ∈ ∆ \ Nθ,σ (Nθ,σ as in Lemma 3.3(b)) with σ(ν) = ∞, then F = θ(cid:0)K E contradicts the fullness of F . Thus, ∞(cid:1) ⊆ K F If there exists ν , which ∆ \ Nθ,σ ⊆ ∆0. On the other hand, as ∆0 ∩ Nθ,σ is a finite set (by Lemma 3.3(a)&(b) and the fact that σ is injective on ∆0) and is open in ∆ (by Lemma 3.6(a)), we see that ∆0 ∩ Nθ,σ consists of isolated points of ∆. Thus, σ(∆0 ∩ Nθ,σ) consists of isolated points of Ω0 (as σ restricts to a homeomorphism from ∆0 to Ω0). We want to show that ∆0 ∩ Nθ,σ = ∅. Suppose on the contrary that there is ν ∈ ∆0 ∩ Nθ,σ. We know that σ(ν) (6= ∞) is a non-isolated point of Ω∞ (by Lemma 3.3(b)). Therefore, there exists a net {ωi}i∈I in Ω \ {σ(ν)} converging to σ(ν). If {i ∈ I : ωi ∈ Ω0} is cofinal, then there is a net in Ω0 \ {σ(ν)} converging to σ(ν), which contradicts σ(ν) being an isolated point in Ω0. Otherwise, ωi ∈ τ −1(∞) eventually, which gives the contradiction that ν = ∞ (note that τ (ωi) → ν as ν ∈ ∆0). Consequently, ∆ \ Nθ,σ = ∆0. Suppose that Nθ,σ 6= ∅ and ν ∈ Nθ,σ. Since Nθ,σ is an open subset of ∆ (by Lemma 3.6(a)), there exists V ∈ N∆(ν) with V ⊆ Nθ,σ. Take any f ∈ F such that f (ν) 6= 0 (by the fullness of F ) and f vanishes outside V . Thus, f ∈ I F ∞ (as V is compact) and so, θ−1(f )(ω) = 0 for any ω ∈ τ −1(∞). On the other hand, for any ω ∈ Ω0, one has τ (ω) ∈ ∆0 and so, f ∈ I F τ (ω) (as f vanishes on the open set ∆0 containing τ (ω)) which implies that θ−1(f )(ω) = 0. Hence θ−1(f ) = 0 which contradicts the injectivity of θ−1. Therefore, Nθ,σ = ∅. Now, part (a) follows from Lemma 3.3(c). (b) This follows directly from Theorem 3.7(b). (cid:3) One can apply the above to the case when F is a full Hilbert C0(∆)-module. Another direct application of Theorem 5.3 is the following theorem which extends and enriches a result of Chan [8] (by removing the boundedness assumption on θ), as well as results 16 CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG concerning the product bundle cases discussed in [4, 13]. Notice that if (Ω, {Ξx}, E) is a continuous fields of Banach spaces over a locally compact Hausdorff space Ω (as defined in [9, 11]), then E is a full essential Banach C0(Ω)-normed module. Theorem 5.4. Let (Ω, {Ξx}, E) and (∆, {Λy}, F ) be continuous fields of Banach spaces over locally compact Hausdorff spaces Ω and ∆ respectively. Let θ : E → F be a bi- jective linear map such that both θ and its inverse θ−1 are separating. Then there is a homeomorphism σ : ∆ → Ω and a bijective linear operator Hν : Ξσ(ν) → Λν such that θ(f )(ν) = Hν(f (σ(ν))) (f ∈ E, ν ∈ ∆). Moreover, at most finitely many Hν are unbounded, and this can happen only when ν is an isolated point in ∆. In particular, if Ω (or ∆) contains no isolated point, then θ is automatically bounded. Appendix A. Bounded local linear maps are A-linear Proposition A.1. Let A be a C ∗-algebra, and let θ be a bounded linear map from a Banach right A-modules E into a Hilbert A-module F . Then θ is a right A-module map if and only if θ is local (in the sense that θ(e)a = 0 whenever e ∈ E and a ∈ A with ea = 0). Proof. Suppose θ is local. Observe, first of all, that E∗∗ and F ∗∗ are unital Banach A∗∗- modules, and the bidual map θ∗∗ : E∗∗ → F ∗∗ is a bounded weak-weak-continuous linear map. Fix x ∈ E and a ∈ A+, and let Φ : C(σ(a))∗∗ → A∗∗ be the map induced by the canonical normal ∗-homomorphism Ψ : M(A)∗∗ → A∗∗. Pick α, β ∈ R+ with α < β, and define p := Φ(χσ(a)∩(α,β)). Let {fn} and {gn} be two bounded sequences in C(σ(a))+ such that fngn = 0, as well as fn ↑ χσ(a)∩(α,β) and gn ↓ χσ(a)\(α,β) pointwisely. Note that as Ψ(A) ⊆ A, we have an := Φ(fn) ∈ A, and we can write bn := Φ(gn) as cn + γn1 (where cn ∈ A and γn ∈ C). Fix n ∈ N. Since an and cn commute, there is a locally compact Hausdorff space Ω with C ∗(an, cn) ∼= C0(Ω). By considering ∼= C ∗(1, an, cn)+, one can find a net {di}i∈I in C0(Ω)+ ⊆ A+ such that bn ∈ C(Ω∞)+ di ≤ bn (i ∈ I) and di → bn pointwisely. As 0 ≤ di ≤ bn and anbn = 0 in C(Ω∞), one knows that andi = 0. Now, the relation θ(xan)di = 0 and θ(xdi)an = 0 imply that θ∗∗(xan)bn = 0 and θ∗∗(xbn)an = 0. Since the multiplication in the bidual of the linking algebra of F is jointly weak-continuous on bounded subsets, we see that AUTOMATIC CONTINUITY AND C0(Ω)-LINEARITY 17 θ∗∗(xp)(1 − p) = 0 and θ∗∗(x(1 − p))p = 0, which implies that θ∗∗(xp) = θ∗∗(x)p. Finally, there exists rk ∈ R and αk, βk ∈ R+ such that αk ≤ βk and a(t) − → 0. Thus, by the weak-continuity again, we get θ∗∗(xa) = θ∗∗(x)a as required. (cid:3) sup t∈σ(a)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) MXk=1 rkχσ(a)∩(αk ,βk)(t)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) References [1] Yu. A. Abramovich, Multiplicative representation of the operators preserving disjointness, Indag. Math. 45 (1983), 265 -- 279. [2] E. Albrecht and M. M. Neumann, Automatic continuity of generalized local linear operators, Manuscripta Math. 32 (1980), 263 -- 294. [3] J. Araujo, Linear biseparating maps between spaces of vector-valued differentiable functions and automatic continuity, Adv. Math. 187 (2004), no. 2, 488 -- 520. [4] J. Araujo and K. Jarosz, Automatic continuity of biseparating maps, Studia Math. 155 (2003), no. 3, 231 -- 239. [5] W. Arendt, Spectral properties of Lamperti operators, Indiana Univ. Math. J. 32 (1983), 199 -- 215. [6] W. Arendt and S. Thomaschewski, Local operators and forms, Positivity 9 (2005), 357 -- 367. [7] E. Beckenstein, L. Narici, and A. R. Todd, Automatic continuity of linear maps on spaces of con- tinuous functions, Manuscripta Math. 62 (1988), 257 -- 275. [8] J. T. Chan, Operators with the disjoint support property, J. Operator Theory 24 (1990), 383 -- 391. [9] J. Dixmier, C*-algebras, North-Holland publishing company, Amsterdam -- New York -- Oxford, 1977. [10] M. J. Dupr´e and R. M. Gillette, Banach bundles, Banach modules and automorphisms of C ∗- algebras, Research Notes in Mathematics 92, Pitman (1983). [11] J. M. G. Fell, The structure of algebras of operator fields, Acta Math., 106 (1961), 233 -- 280. [12] J. J. Font and S. Hern´andez, On separating maps between locally compact spaces, Arch. Math. (Basel) 63 (1994), 158 -- 165. [13] H.-L. Gau, J.-S. Jeang and N.-C. Wong, Biseparating linear maps between continuous vector-valued function spaces, J. Australian Math. Soc., Series A, 74 (2003), no. 1, 101 -- 111. [14] K. Jarosz, Automatic continuity of separating linear isomorphisms, Canad. Math. Bull. 33 (1990), 139 -- 144. [15] J.-S. Jeang and N.-C. Wong, Weighted composition operators of C0(X)'s, J. Math. Anal. Appl. 201 (1996), 981 -- 993. [16] R. Kantrowitz and M. M. Neumann, Disjointness preserving and local operators on algebras of differentiable functions, Glasg. Math. J. 43 (2001), 295-309. [17] R. Narasimhan, Analysis on real and complex manifolds, Advanced Studies in Pure Mathematics 1, North-Holland Publishing Co., Amsterdam 1968 [18] B. de Pagter, A note on disjointness preserving operators, Proc. Amer. Math. Soc. 90 (1984), 543 -- 550. [19] J. Peetre, R´ectification `a l'article "Une caract´erisation abstraite des op´erateurs diff´erentiels", Math. Scand. 8 (1960), 116 -- 120. [20] R. G. Swan, Vector bundles and projective modules, Trans. Amer. Math. Soc. 105 (1962), 264 -- 277. 18 CHI-WAI LEUNG, CHI-KEUNG NG AND NGAI-CHING WONG (Chi-Wai Leung) Department of Mathematics, The Chinese University of Hong Kong, Hong Kong. E-mail address: [email protected] (Chi-Keung Ng) Chern Institute of Mathematics and LPMC, Nankai University, Tianjin 300071, China. E-mail address: [email protected] (Ngai-Ching Wong) Department of Applied Mathematics, National Sun Yat-sen Univer- sity, Kaohsiung, 80424, Taiwan. E-mail address: [email protected]
1512.04260	1	1512	2015-12-14T11:18:56	Fredholm operators on $C^*$-algebras	[ "math.OA", "math.FA" ]	The aim of this note is to generalize the notion of Fredholm operator to an arbitrary $C^$-algebra. Namely, we define "finite type" elements in an axiomatic way, and also we define Fredholm type element $a$ as such element of a given $C^$-algebra for which there are finite type elements $p$ and $q$ such that $(1-q)a(1-p)$ is "invertible". We derive index theorem for such operators. In applications we show that classical Fredholm operators on a Hilbert space, Fredholm operators in the sense of Breuer, Atiyah and Singer on a properly infinite von Neumann algebra, and Fredholm operators on Hilbert $C^$-modules over an unital $C^$-algebra in the sense of Mishchenko and Fomenko are special cases of our theory.	math.OA	math	FREDHOLM OPERATORS ON C∗-ALGEBRAS DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Abstract. The aim of this note is to generalize the notion of Fredholm op- erator to an arbitrary C ∗-algebra. Namely, we define "finite type" elements in an axiomatic way, and also we define Fredholm type element a as such element of a given C ∗-algebra for which there are finite type elements p and q such that (1 − q)a(1 − p) is "invertible". We derive index theorem for such operators. In applications we show that classical Fredholm operators on a Hilbert space, Fredholm operators in the sense of Breuer, Atiyah and Singer on a properly infinite von Neumann algebra, and Fredholm operators on Hilbert C ∗-modules over an unital C ∗-algebra in the sense of Mishchenko and Fomenko are special cases of our theory. 1. Introduction Fredholm operators have been investigated for many years. Initially, they was considered as those operators acting on a Banach space with finite dimensional kernel and cokernel. Their index is defined as a difference of dimensions of the kernel and cokernel. The most important properties of Fredholm operators are the following 1. The index theorem. It asserts that given two Fredholm operators T and S, the operator T S is also Fredholm and ind(T S) = ind T + ind S 2. Theorem of Atkinson. It asserts that the operator T is Fredholm if and only if it is invertible modulo compact operators, or equivalently if and only if its image in the Calkin algebra B(X)/C(X) is invertible, where C(X) denote the algebra of compact operators. 3. Perturbation theorem. If T is a Fredholm operator, then the operator T + K is Fredholm as well, provided that K is compact. In this case ind(T + K) = ind T . 4. Continuity of index. The Fredholm operators form an open set in the norm topology and the index is constant on each connected component of this set. The generalization of Fredholm theory to the level of von Neumann algebras was initially done by Breuer [4, 5], but it became famous after Atiyah's work [2]. In order to generalize earlier result of him and Singer, the well known index theorem, to noncompact manifolds he considered the operators with kernel and cokernel that don't belong to finite dimensional subspaces, but affiliated to some von Neumann algebra. He defined the dimension of such subspaces as the trace of the correspond- ing projection in the appropriate von Neumann algebra, see [2]. Soon after that 2010 Mathematics Subject Classification. 47A53, 46L08, 46L80. Key words and phrases. C ∗-algebra, Fredholm operators, K group, index. The authors was supported in part by the Ministry of education and science, Republic of Serbia, Grant #174034. 1 2 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Mischenko and Fomenko introduced the notion of Fredholm operator in the frame- work of Hilbert C∗ modules, see [9]. It was repeated with a different approach by Mingo, see [8] and also important references therein. Several decades ago, there are, also, some attempts to establish axiomatic Fredholm theory in the framework of von Neumann algebras G (relative to ideal I). Fredholm operators are defined, in this case, as invertible elements in the quotient space G/I, and the index is defined as an element from C(Ω) (the set of all continuous function on Ω, where C(Ω) is isomorphic to the center Z of G), see [6], [11] and references therein. There are, also, other attempts in generalization of Fredholm theory, for instance [1]. The aim of this note is to single out the properties of finite rank operators which ensure the development of Fredholm theory. In other words we shall introduce ax- iomatic foundation of Fredholm theory. We shall deal within a framework of a unital C∗-algebra A and its faithful representation ρ : A → B(H) as the subalgebra ρ(A) of the algebra of all bounded operators on some Hilbert space H. Further, we prove that the standard Fredholm operators, Fredholm operators in the sense of Atiyah and Singer on II∞ factors, and Fredholm operators in the sense of Mishchenko and Fomenko are special cases of our theory. Throughout this paper we shall assume that a given C∗-algebra is always unital, even if this is not explicitly mentioned. By ker (coker) we shall denote the ker- nel(cokernel) of some operator. If X is a closed subspace of some Hilbert space H, PX will denote the orthogonal projection on X. Sometimes we shall omit the word orthogonal, i.e, projection will mean orthogonal projection. The letter 1 will stand for unit element in an abstract algebra, whereas I will stand for identity operator on some Hilbert space. Similarly, we use small letters a, b, t, etc. to denote ele- ments of an abstract algebra, whereas we use capital letters A, B, T , etc. to denote operators on a concrete Hilbert space. and a ∈ F implies a∗ ∈ F ; Definition 1.1. Let A be an unital C∗-algebra, and let F ⊆ A be a subalgebra which satisfies the following conditions: (i) F is a selfadjoint ideal in A, i.e. for all a ∈ A, b ∈ F there holds ab, ba ∈ F (ii) There is an approximate unit pα for F , consisting of projections; (iii) If p, q ∈ F are projections, then there exists v ∈ A, such that vv∗ = q and Such a family we shall call finite type elements. In further, we shall denote it v∗v⊥p, i.e. v∗v + p is a projection as well; by F . Definition 1.2. Let A be an unital C∗-algebra, and let F ⊆ A be an algebra of finite type elements. In the set Proj(F ) we define the equivalence relation: p ∼ q ⇔ ∃v ∈ A vv∗ = p, v∗v = q, i.e. Murray - von Neumann equivalence. The set S(F ) = Proj(F )/ ∼ is a commu- tative semigroup with respect to addition, and the set K(F ) = G(S(F )), where G denotes the Grothendic functor, is a commutative group. FREDHOLM OPERATORS ON C ∗-ALGEBRAS 3 2. Results We devide this section into four subsections. First three of them, Well known Hilbert space Lemmata, Almost invertibility and Approximate units technique con- tain introductory material necessary for the last one, Index and its properties, which contains the main results. Well known Hilbert space Lemmata. First, we list, and also prove, three elementary statements concerning operators on a Hilbert space. The first Lemma establishes that two projections, close enough to each other, are unitarily equivalent, and also that obtained unitary is close to the identity operator. Lemma 2.1. Let P , Q ∈ B(H), let H be a Hilbert space and let P − Q < 1. Then there are partial isometries V , W such that (2.1) V V ∗ = P, V ∗V = Q, W W ∗ = I − P, Moreover, there is a unitary U such that W ∗W = I − Q. U ∗P U = Q, U ∗(I − P )U = I − Q. In addition, V , W , U ∈ C∗(I, P, Q) - a unital C∗-algebra generated by P and Q. Finally, if P − Q < 1/2, we have the estimate I − U ≤ CP − Q, (2.2) where C is an absolute constant, i.e. it does not depend neither on P nor on Q. Proof. The existence of V ∈ B(H) with properties (2.1) was proved in [13, §105, page 268]. Also, it is given that V = P (I + P (Q − P )P )−1/2Q ∈ C∗(I, P, Q). Since (I − P )− (I − Q) = P − Q, we can apply previous reasoning to I − P and I − Q to obtain W with required properties. Clearly, U = V + W is a unitary we are looking for. Finally, to obtain (2.2) we have Q − V = Q − P (I + P (Q − P )P )−1/2Q ≤ Q − P (I + P (Q − P )P )−1/2. Let αn be the coefficients in Taylor expansion of the function t 7→ (1 + t)−1/2. Then (I + P (Q − P )P )−1/2 = I +P∞ Q − V ≤ Q − P − n=1 αn(P (Q − P )P )n and hence Xn=1 αn(P (Q − P )P )n ≤ Q − P + ∞ Xn=1 ∞ αnQ − Pn = Q − P(1 + ∞ Xn=1 αnQ − Pn−1) ≤ C1Q − P, where C1 = 1 +P∞ n=1 αn/2n−1 < +∞. Thus I − U = I − Q + Q − W − V ≤ Q − V + I − Q − W ≤ 2C1Q − P. (Indeed, since C1 does not depend on P , Q, applying the previous reasoning to the projections I − P , I − Q we have I − Q − W ≤ C1(I − Q) − (I − P ).) (cid:3) The second Lemma characterizes "almost orthogonal" projections. 4 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Lemma 2.2. Let P , Q ∈ B(H) be projections, and let for all ξ ∈ P (H) and all η ∈ Q(H) holds hξ, ηi ≤ cξη. Then P Q ≤ c. Proof. Let ξ, η ∈ H. Then P ξ ∈ P (H), Qη ∈ Q(H), and hQP ξ, ηi = hP ξ, Qηi ≤ cP ξQη ≤ cξη. Therefore QP ≤ c and also P Q = (QP )∗ ≤ c. (cid:3) The last Lemma in this subsection deals with the adjoint of an isomorphism between two different subspaces. Lemma 2.3. Let H1, H2 be Hilbert spaces, K ≤ H1, L ≤ H2. Also, let T : H1 → H2 be a bounded mapping that maps K bijectively to L and TK⊥ = 0. Then T ∗ maps bijectively L to K, and T ∗L⊥ = 0. Proof. Since ker T ∗ = (ran T )⊥ = L⊥, it follows T ∗L⊥ = 0. In particular, T ∗ is injective on L. Also, ran T ∗ = (ker T )⊥ = K, i.e. ran T ∗ is dense in K. Since T maps bijectively K to L (K, L closed), by open mapping theorem, T is bounded below on K, i.e. there is c > 0 such that T ξ ≥ cξ for all ξ ∈ K. Let η ∈ L. Then T ∗η ∈ K and T ∗η = sup ξ∈K ξ=1 hξ, T ∗ηi = sup ξ∈K ξ=1 hT ξ, ηi = sup ξ∈K ξ=1 c sup ξ∈K ξ=1 (cid:12)(cid:12)(cid:12)(cid:12) (cid:28) T ξ T ξ , η(cid:29)(cid:12)(cid:12)(cid:12)(cid:12) (cid:28) T ξ T ξ T ξ (cid:12)(cid:12)(cid:12)(cid:12) ≥ , η(cid:29)(cid:12)(cid:12)(cid:12)(cid:12) = c sup η′∈L η′=1 hη′, ηi = cη. The last equality is due to T (K) = L. Therefore, T ∗ is topologically injective on L. Hence its range is closed and we (cid:3) are done. Almost invertibility. In this subsection, we introduce the notion of almost invert- ibility, i.e. invertibility up to a pair of projections, following definitions of Fredholm operators in the framework of Hilbert C∗-modules [7, Definition 2.7.4], and [15, Chapter 17]. Definition 2.1. Let a ∈ A and let p, q ∈ F . We say that a is invertible up to pair (p, q) if the element a′ = (1 − q)a(1 − p) is invertible, i.e., if there is some b ∈ A with b = (1 − p)b(1 − q) (and immediately bq = 0, pb = 0, b = (1 − p)b = b(1 − q)) such that We refer to such b as almost inverse of a, or (p, q)-inverse of a. a′b = 1 − q, ba′ = 1 − p. First, we establish that almost invertible elements make an open set, and that almost inverse of a (in a certain sense) continuously changes with respect to a, as well as with respect to p and q Lemma 2.4. Let a be invertible up to (p, q) and let b be (p, q)-inverse of a. FREDHOLM OPERATORS ON C ∗-ALGEBRAS 5 a) The element a + c is also invertible up to (p, q) for every c ∈ A, which satisfies c < b−1. If b1 denote (p, q)-inverse of a + c, then (2.3) b . b1 ≤ 1 − bc b) If p − p′, q − q′ < min{1/2, 1/4Cab}, where C is the constant from (2.2) then a is also invertible up to (p′, q′). Moreover, if b′ is (p′, q′)-inverse of a, then b′ ≤ 2b. Proof. a) Let b be (p, q)-inverse of a. Then (1 − q)(a + c)(1 − p) = a′ + c′ = a′(1 − p) + (1 − q)c = a′(1 − p) + a′bc′ = a′(1 − p + bc′), where c′ = (1 − q)c(1 − p). The element bc′ belongs to (1− p)A(1− p) and its norm is bc′ ≤ bc < 1, for c < b−1. Therefore, 1−p+bc′ is invertible in the corner algebra (1−p)A(1−p). Denote its inverse by t. We have (1−q)(a+c)(1−p)tb = a′(1−p+bc′)tb = a′b = 1−q and tb(1−q)(a+c)(1−p) = tba′(1−p+bc′) = t(1−p)(1−p+bc′) = 1−p. Therefore b1 = tb is (p, q)-inverse of a + c, for c < b−1. Let us prove (2.3). We have ∞ t = (1 − p) + Xn=1 (−1)n(bc′)n ≤ 1 + ∞ Xn=1 bnc′n = 1 1 − bc′ . b) By Lemma 2.1 there is a unitary u such that u∗q′u = q. Then, we have Since c′ ≤ c and b1 = tb, (2.3) follows. u∗(1− q′)a(1− p) = (1− q)a(1− p)− u∗(1− q′)(u− 1)a(1− p) = (1− q)a(1− p)− c. Note that c ∈ (1 − q)A(1 − p), as well as c < u − 1a < Cq − q′a < 1/(4b). By the previous part a), we have that there is (p, q)-inverse of u∗(1 − q′)a(1 − p), say b1. It is easy to check that b1u∗ is (p, q′)-inverse of a, and also b1u∗ ≤ b1 < b 1 − bc ≤ 4b 3 . In a similar way we can substitute p with p′ to obtain that a is invertible up to (p′, q′). Denote its (p, q)-inverse by b′. Also we have the inequality b′ ≤ b1 1 − b1c′ ≤ 4b/3 1 − 4bc′/3 ≤ 2b. (cid:3) The next Lemma will be used in the sequel many times. It allows us to transfer a projection from the right side of an element to its left side, if the considered element is invertible on this projection. In the framework of a representation, it means that we can carry a projection from the domain of an operator to its codomain. The transferred projection is equivalent to the initial one. Lemma 2.5. Let A be a unital C∗-algebra, let a ∈ A, let p, r ∈ A be projections such that a is left invertible up to p, i.e. there is b ∈ A such that ba(1 − p) = 1 − p. Also, let r ≤ 1 − p. Then: (i) ar has a polar decomposition in A, i.e. ar = var, (2.4) Further, s = v∗v ∈ A is the minimal projection in A, such that sar = ar. (Obvi- ously, s ∼ r, and if r ∈ F , then s ∈ F as well.) Moreover, a is invertible up to (1 − r, 1 − s). v,ar ∈ A, r = vv∗. 6 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ ρ (ii) If A ֒→ B(H) is any faithful representation of A, then L = ρ(ar)(H) is In particular projection on ρ a closed subspace, and ρ(s) is the projection on L. L = ρ(ar)(H) belongs to ρ(A). ֒→ B(H) be some faithful representation of C∗-algebra A and let Proof. Let A ρ(r)(H) = K. Denote A = ρ(a), B = ρ(b), P = ρ(p), R = ρ(r). Obviously, AR has a polar decomposition AR = V AR in B(H). Since the partial isometry v is uniquely determined by (2.4), we have only to prove that V ∈ ρ(A). ξ = BA(I − P )ξ ≤ BA(I − P )ξ = k−1Aξ, Let H1 = (I − P )(H). For ξ ∈ H1 we have where k = B−1. (We used (I − P )ξ = ξ.) Thus, Aξ ≥ kξ, ξ ∈ H1, i.e. A is injective and bounded below on H1, and consequently on K = R(H) ≤ H1 (r ≤ 1 − p). Therefore, L = AR(H) = A(K) is a closed subspace. (This proves the first claim in (ii).) As AR maps bijectively K to L, (AR)∗ maps bijectively L to K - Lemma 2.3, and hence, RA∗AR = (AR)∗AR is an isomorphism of K. Denote its inverse by T : K → K. Let T : H → H be the extension of T , defined to be 0 on K ⊥. Trivially, T ≥ 0, and √T exists. Also, √T maps K to K and √TK⊥ = 0. We claim that V = AR√T . Indeed, both V and AR√T annihilates K ⊥, whereas on K, √T and AR are inverses to each other, since AR = ((AR)∗AR)1/2 and T is inverse for (AR)∗AR. Thus, it remains to prove that T ∈ ρ(A). As it is easy to check, (AR)∗AR + I − R and T + I − R are inverses to each other. Since (AR)∗AR + I − R ∈ ρ(A), its inverse also belongs to ρ(A). Namely, by Theorem 11.29 from [12] (or Theorem 2.1.11 from [10]) the spectrum of an element is the same with respect to algebra as well as with respect to subalgebra. Apply this conclusion to 0 as the point of spectrum to obtain T + I − R ∈ ρ(A) and immediately, T ∈ ρ(A). Therefore T = ρ(t) for some t ∈ A. Hence V = ρ(v), where v = ar√t and ar has a polar decomposition in A. To finish the proof, note that sar = ar, since s = v∗v and S = ρ(s) is the projection on the subspace L = AR(H). If for some other projection s1, there holds s1ar = ar then must be S1(H) ≥ L and hence S1 ≥ S implying s1 ≥ s. (cid:3) The following Lemma (as many others), in fact, deals with 2 × 2 matrices. It can be freely reformulated as follows: If [ai,j]2 i,j=1 is invertible, a11 is invertible and if this matrix is either triangular (a12 = 0) or close to triangular (a12 small) then a22 is also invertible. However, we do not use matrix notation (now and in the sequel) in order to avoid confusions such as with respect to which pair(s) of projections a matrix is formed. Lemma 2.6. Let a, p, q ∈ A, let p, q be projections such that a is invertible up to (p, q) and let r1, r2, s1, s2 be projections such that 1 − p = r1 + r2, 1 − q = s1 + s2. If a is invertible up to (1 − r1, 1 − s1) and if s2ar1 < b−1 or s1ar2 < b−1, where b is (p, q) inverse of a, then a is invertible up to (1 − r2, 1 − s2). Proof. 1◦ case - let s2ar1 = 0. Decompose (1 − q)a(1 − p) and b as (2.5) (1 − q)a(1 − p) = (s1 + s2)a(r1 + r2) = s1ar1 + s1ar2 + s2ar2. (2.6) b = (1 − p)b(1 − q) = (r1 + r2)b(s1 + s2) = r1bs1 + r1bs2 + r2bs1 + r2bs2. FREDHOLM OPERATORS ON C ∗-ALGEBRAS 7 Multiplying (2.5) by (2.6) from right we get s1 + s2 = 1 − q = s1ar1bs1 + s1ar1bs2 + s1ar2bs1 + s1ar2bs2 + s2ar2bs1 + s2ar2bs2. Multiplying the last equality by s2 from both, left and right side we obtain: (2.7) s2 = s2ar2r2bs2. Multiplying (2.5) and (2.6) in reverse order we get (2.8) r1 + r2 = 1 − p = r1bs1ar1 + r1bs1ar2 + r1bs2ar2 + r2bs1ar1 + r2bs1ar2 + r2bs2ar2. Multiply (2.8) by r2 from the left, and by r1 from the right to obtain r2bs1ar1 = 0. The element a has an (1 − r1, 1 − s1) inverse, say b′. It holds s1ar1b′ = s1, and we have r2bs1 = r2bs1ar1b′ = 0b′ = 0. Finally, multiply (2.8) by r2 from both, left and right side to obtain r2 = r2bs1ar2 + r2bs2ar2 = 0ar2 + r2bs2ar2 = r2bs2s2ar2. This together with (2.7) means that r2bs2 is (1− r2, 1− s2)-inverse for a. Therefore, by Lemma 2.5 applying to r2 ≤ r1 + r2 we found that r2 ∼ s2. 2◦ case - general. Consider a = a − s2ar1. Obviously, s1ar1 = s1ar1, s2ar2 = s2ar2 and s2ar1 = 0. Since s2ar1 < b−1, by Lemma 2.4 a), a is invertible up to (p, q) and we can apply the previous case. If s1ar2 < b−1, apply the presented proof to a∗. The next Lemma is a refinement of Lemma 2.5. (cid:3) Lemma 2.7. Let A be a unital C∗-algebra, let p, q, r ∈ A be projections such that a is invertible up to (p, q), and r ≤ 1 − p, and let s be the projection obtained in Lemma 2.5. If qa(1 − p) = 0, then s ≤ 1 − q, and 1 − q − s ∼ 1 − p − r. Proof. Since a is invertible up to (p, q), a is left invertible up to p. So, the projection s is well defined. If qa(1 − p) = 0, then (1 − q)a(1 − p) = a(1 − p) which implies ar = a(1 − p)r = (1 − q)a(1 − p)r = (1 − q)ar. So, s ≤ 1 − q by minimality of s. Denote r1 = 1 − p − r. Apply the previous part of the proof to r1 ≤ 1 − p to obtain the minimal s1 ≤ 1 − q such that s1ar1 = ar1 and r1 ∼ s1. Denote s2 = 1 − q − s. Projections s1 and s2 might not coincide, but they must be equivalent. More precisely s2 ∼ r1 ∼ s1. To show this it is enough to prove that s2ar1 is "invertible". However, from sar = ar we get s2ar = (1 − q)ar − sar = (1 − q)a(1 − p)r − sar = a(1 − p)r − qa(1 − p)r − sar = ar − 0 − sar = 0. Thus, it is sufficient to apply Lemma 2.6. (cid:3) The last statement in this subsection ensures that almost invertible elements can be triangularized. More precisely, if a is invertible up to (p, q) then we can substitute p or q (one of them - not simultaneously) by an equivalent projection such that almost invertibility is not harmed and such that with respect to new projections the considered element has a triangular form. Proposition 2.8. Let A be a unital C∗-algebra, let p, q ∈ A and let a be invertible up to (p, q). a) Then there is a projection q′ ∈ A, q′ ∼ q such that a is invertible up to (p, q′) b) Also, there is another projection p′ ∈ A, p′ ∼ p such that a is invertible up to and q′a(1 − p) = 0; (p′, q) and (1 − q)ap′ = 0. 8 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Proof. a) Let b be the (p, q)-inverse of a. Then (2.9) (1 − q)a(1 − p)b = 1 − q, b(1 − q)a(1 − p) = 1 − p, (1 − p)b(1 − q) = b, pb = bq = 0, b = (1 − p)b, b = b(1 − q). Let u = 1 + qab. This element has its inverse, u−1 = 1 − qab, which can be easily checked. Consider the element a1 = u−1a = (1 − qab)a = a − qaba. Using b(1 − q) = b and b(1 − q)a(1 − p) = 1 − p (see (2.9)), we have (2.10) qa1(1−p) = q(a−qaba)(1−p) = qa(1−p)−qab(1−q)a(1−p) = qa(1−p)−qa(1−p) = 0. The element u−1∗ is invertible together with u and we can apply Lemma 2.5 to obtain a minimal projection q′ ∈ A such that q′ ∼ q, u−1∗ is invertible up to (1 − q, 1 − q′) and q′u−1∗q = u−1∗q. The last equality implies qu−1(1 − q′) = 0, (2.11) which ensures, by Lemma 2.6 that u−1 is invertible up to (1 − q′, 1 − q), as well as up to (q′, q). qu−1 = qu−1q′, Let us, first prove that q′a(1 − p) = 0. Indeed, by (2.10) and (2.11) we have 0 = qa1(1 − p) = qu−1a(1 − p) = qu−1q′a(1 − p). (2.12) Let t is (1− q′, 1− q)-inverse for u−1. We have tqu−1q′ = q′. Multiply the equation 2.12 from the left by t to obtain 0 = tqu−1q′a(1 − p) = q′a(1 − p). Finally, let us prove that a is invertible up to (p, q′). Note that (1− q)a1(1− p) = (1 − q)(a − qaba)(1 − p) = (1 − q)a(1 − p). Next, by (2.11) we get qu−1(1 − q′) = 0 and (1 − q)u−1(1 − q′) = u−1(1 − q′) and hence (1 − q′)a(1 − p) = a(1 − p) = uu−1a(1 − p) = ua1(1 − p) = u(1 − q)a1(1 − p) = u(1 − q)a(1 − p). Now, it is easy to check that bu−1(1 − q′) is (p, q′)-inverse of a. Indeed (1 − q′)a(1 − p) · bu−1(1 − q′) = u(1 − q)a(1 − p)bu−1(1 − q′) = = u(1 − q)u−1(1 − q′) = uu−1(1 − q′) = 1 − q′ and bu−1(1 − q′) · (1 − q′)a(1 − p) = bu−1u(1 − q)a(1 − p) = 1 − p. b) It is enough to apply the previous conclusion to a∗. (cid:3) Approximate units technique. For the development of the abstract Fredholm theory, we required in Definition 1.1 an ideal with an approximate unit consisting of projections. In this subsection we derive some properties of such an ideal. In the next two Lemmata we obtain that an approximate unit absorbs any other projections and that any finite projection is an element of some approximate unit. Lemma 2.9. Let pα ∈ F be an approximate unit, and let p ∈ F . Then there is some α0 and p′ such that p ∼ p′ ≤ pα0. Moreover, α0 can be chosen such that p − p′ is arbitrarily small. Proof. Since pα is an approximate unit, we have p − pαp → 0, as α → ∞. Therefore, there is α0 such that p − pα0 p ≤ δ < 1, p − ppα0p ≤ δ. implying (2.13) FREDHOLM OPERATORS ON C ∗-ALGEBRAS 9 First, we obtain that ppα0p is invertible element of the corner algebra pAp. Now, pick a faithful representation ρ of A on some Hilbert space H. Denote the images of ρ by the corresponding capital letters, P = ρ(p), Pα0 = ρ(pα0) etc. Let K = P (H). Since pα0 is invertible up to (1 − p, 1 − p) we can apply Lemma 2.5 (i) to conclude that there is p′ ∈ A, p′ ∼ p. By the part (ii) of the same Lemma, we get p′ ≤ pα0 (P ′ is the projection on the range of Pα0 P . The last is, obviously, subspace of the range of Pα0 .) It remains to prove that p − p′ can be arbitrarily small. To do this, let us prove that pα0 does not change norm of ξ ∈ K too much. Indeed, for ξ ∈ K, using (2.13) we have pα0 ξ = pα0pξ = pξ − (p − pα0 p)ξ ≤ (1 + δ)ξ, and also, (2.14) pα0 ξ = pα0pξ = pξ−(p−pα0p)ξ ≥ ξ−(p−pα0p)ξ ≥ (1−δ)ξ. Now, we want to prove that 1− p′ and p (and similarly 1− p and p′) are "almost Let η = (1 − p)η and ζ = p′ζ. Then ζ = pα0ξ for some ξ = pξ ∈ K, and orthogonal". ζ ≥ (1 − δ)ξ by (2.14). We have hζ, ηi = hpα0 pξ, (1 − p)ηi = − h(p − pα0 p)ξ, (1 − p)ηi , and hence hζ, ηi ≤ p − pα0pξη ≤ Therefore, by Lemma 2.2 δ 1 − δζ,η. (2.15) p′(1 − p) < δ 1 − δ . Now, let ζ = (1 − p′)ζ and η = pη. Then, ζ ⊥ pα0pη (since pα0pη ∈ p′H) and therefore hζ, ηi = h(1 − p′)ζ, pηi = h(1 − p′)ζ, (p − pα0p)ηi ≤ ≤ p − pα0 pζη ≤ δζη, from which we conclude again by Lemma 2.2 (2.16) p − p′p = (p − p′p)∗ = p − pp′ ≤ δ. From (2.15) and (2.16) we get p − p′ ≤ p − pp′ + p′ − pp′ ≤ δ which can be arbitrarily small. 2 − δ 1 − δ , (cid:3) Lemma 2.10. Let F be an algebra of finite type elements. For every projection p ∈ F there is an approximate unit pα in F such that for all α there holds p ≤ pα. Proof. Let p ∈ F and let pα be an approximate unit. By Lemma 2.9, for α large enough, we have p ∼ p′ ≤ pα and p − p′ < 1. By Lemma 2.1 there is a unitary u α = u∗pαu is an approximate unit that contains p. (cid:3) such that p′ = u∗pu. Then p′ 10 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ The following Proposition plays the key role in this paper. It ensures that we can transfer an approximate unit from the right side of an almost invertible element to its left side, retaining some triangular properties. Briefly, if such almost invertible element is upper triangular, with upper left entry invertible, then this entry itself has a triangular form (but lower - not upper) with respect to an approximate unit from the right and its corresponding approximate unit from the left. Proposition 2.11. Let p, q ∈ F , let a be invertible up to (p, q), and let qa(1 − p) = 0. Further, let pα ≥ p be an approximate unit for F . Then there exists an approximate unit qα ∈ F , such that qα − q ∼ pα − p, a is invertible up to (pα, qα), and (2.17) for all α. (1 − qα)a(pα − p) = 0, Proof. a) Since a′ = (1 − q)a(1 − p) is invertible, and since pα − p ≤ 1 − p, there is (by Lemma 2.5) a minimal projection q′ α such that q′ αa(pα − p) = a(pα − p) and a is invertible up to (1 − (pα − p), 1 − q′ α). By qa(1 − p) = 0 and by Lemma 2.7 we have q′ α. We have α ≤ 1 − q. Set qα = q + q′ (2.18) (qα − q)a(pα − p) = a(pα − p) and as a consequence (2.17). Indeed, using qa(1 − p) = 0, we find qa = qap and hence qapα = qappα = qap. Thus (qα − q)a(pα − p) = qαapα − qαap− qapα + qap = qαapα − qαap = qαa(pα − p). Thus, (2.18) becomes qαa(pα − p) = a(pα − p) which is equivalent to (2.17). Since pα − p ≤ 1 − p we have qa(pα − p) = 0, and from (2.17) we have 0 = (1 − qα)a(pα − p) = (1 − qα)a(pα − p) + qa(pα − p) = (1 − (qα − q))a(pα − p). By Lemma 2.6 a is invertible up to (pα, qα). Also, q′ α ∼ pα − p ∈ F and therefore qα = q′ α + q ∈ F , as well. The first claim is proved. Let us prove that qα is a left approximate unit for F . Let a′ = (1 − q)a(1 − p), and let f ∈ F . Since qa(1 − p) = 0, there holds a′ = a − (1 − q)ap− qap. Therefore (1 − qα)a′ = (1 − qα)a − (1 − qa)ap, since qα ≥ q. Now, by (2.17), we have (1 − qα)a′f = (1 − qα)a(1 − p)f = (1 − qα)a(1 − pα)f → 0 in norm topology for any f ∈ F , since (1 − qα) is norm bounded and pα is an approximate unit. Thus qα is a left approximate unit for a′F . However, any f ∈ F can be expressed as f = (1 − q)f + qf = a′bf + qf . Using q ≤ qα we get (1 − qα)q = 0 and therefore (1 − qα)f = (1 − qα)a′bf → 0, in norm. To finish the proof, note that the mapping pα 7→ qα preserves order. Indeed, if pβ ≥ pα, we have (qα − q)a(pα − p) = a(pα − p), (qβ − q)a(pβ − p) = a(pβ − p), where qα−q, qβ−q have minimal property. Multiply the second equality by (pα−p) and using (pβ − p)(pα − p) = pα − p we find (qβ − q)a(pα − p) = a(pα − p), which implies qβ ≥ qα by minimal property. Since F is a ∗-ideal, we have that qα is a right approximate unit, as well. (cid:3) FREDHOLM OPERATORS ON C ∗-ALGEBRAS 11 Index and its properties. First, we establish that the difference [p]− [q] will not change as long as a fixed a is invertible up to (p, q). We need such a result to define the index, exactly to be the mentioned difference in the sequel Definition. Proposition 2.12. Let a ∈ A be invertible up to (p, q), and also invertible up to (p′, q′), and let p, q, p′, q′ ∈ F . Then in K(F ) we have [p] − [q] = [p′] − [q′]. Proof. Due to the Proposition 2.8, we may assume that qa(1− p) = q′a(1− p′) = 0. Since the same proposition follows that assumption will not change classes [q] and [q′]. By Lemma 2.10 there is an approximate unit pα of F containing p. For all α we have pα = p + rα. By Lemma 2.9, choose α large enough, such that p′′ ∼ p′, p′′ − p′ < δ and pα = p′′ + r′ α, where δ < 1/2, δ < 1/(4Cab′), C is the absolute constant from (2.2) and b′ is (p′, q′)-inverse of a. The previous two displayed formulae yields (2.19) [p] + [rα] = [p′] + [r′ α] By Proposition 2.11, there is another approximate unit qα ≥ q such that qα−q ∼ pα − p and such that (2.17) holds. Let sα = qα − q. Then (2.20) qα = q + sα, rα ∼ sα. Enlarging α, if necessary, there is q′′ ≤ qα such that q′′ ∼ q′ and q′′ − q′ < δ and consequently Thus we have, also (2.21) qα = q′′ + s′ α. [q] + [sα] = [q′] + [s′ α]. By Lemma 2.4 b), using p′′− p′, q′′− q′ < 1/(4Cab′), we conclude that a is invertible up to (p′′, q′′). Also, if b′′ is (p′′, q′′)-inverse of a, then b′′ ≤ 2b′. Next, we want to estimate (1 − qα)a(pα − p′′). Using (2.17) we have (1 − qα)a(pα − p′′) = (1 − qα)a(pα − p + p − p′ + p′ − p′′) = (1 − qα)a(pα − p) + (1 − qα)a(p − p′) + (1 − qα)a(p′ − p′′) = (1 − qα)a(p − p′) + (1 − qα)a(p′ − p′′). Enlarging α, once again, we can assume that (1− qα)a(p− p′) < aδ. Thus, we have (1 − qα)a(pα − p′′) < 2aδ < 1/(2Cb′) ≤ 1/(Cb′′). Since C > 1, we can apply Lemma 2.6 to obtain qα − q′′ ∼ pα − p′′, i.e. (2.22) α ∼ s′ r′ α. Subtracting (2.19) and (2.21) we obtain [p] + [rα] − [q] − [sα] = [p′] + [r′ α] − [q′] − [s′ α] which finishes the proof, because [rα] − [sα] = 0 by (2.20), and [r′ (2.22). α] − [s′ α] = 0 by (cid:3) 12 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Remark 2.1. Note that in all previous proofs, we obtain Murray von Neumann equivalency, whereas, at this moment, we include the cancellation law in K group, which produces that it can be [p] = [q] though p and q might not be Murray - von Neumann equivalent. Namely, the Grothendick functor expand initial equivalence relation, if the underlying semigroup does not satisfy the cancellation law. This fact is known in K theory as "stable equivalency". Definition 2.2. Let F be finite type elements. We say that a ∈ A is of Fredholm type (or abstract Fredholm element) if there are p, q ∈ F such that a is invertible up to (p, q). The index of the element a (or abstract index) is the element of the group K(F ) defined by ind(a) = ([p], [q]) ∈ K(F ), or less formally Note that the index is well defined due to the Proposition 2.12. ind(a) = [p] − [q]. Now, we proceed deriving the properties of the abstract index, defined in the previous Definition. Proposition 2.13. The set of Fredholm type elements is open in A and the index is a locally constant function. Proof. It follows immediately from Lemma 2.4. Proposition 2.14. a) Let a ∈ A be of Fredholm type, and let f ∈ F . Then a + f is also of Fredholm type, and ind(a + f ) = ind a. b) If f ∈ F , then 1 + f is of Fredholm type, and ind(1 + f ) = 0. Moreover, there is p ∈ F such that 1 + f is invertible up to (p, p). Proof. a) Let p, q ∈ F be projections such that a is invertible up to (p, q). By Proposition 2.11 there are approximate units pα ≥ p, qα ≥ q such that a is invertible up to (pα, qα). Choose α large enough such that f ≤ bα−1, where bα is (pα, qα)- inverse of a. Then, (cid:3) (1 − qα)(a + f )(1 − pα) = (1 − qα)a(1 − pα) + (1 − qα)f (1 − pα) and (1 − qα)f (1 − pα) ≤ f ≤ bα−1. By Lemma 2.4, a + f is also invertible up to (pα, qα) and hence ind(a + f ) = [pα] − [qα] = ind a; b) In this special case, a = 1 and we can choose qα = pα. (cid:3) Proposition 2.15. a) If a is of Fredholm type, then a is invertible modulo F ; b) Conversely, if a is invertible modulo F , then a is of Fredholm type. Proof. a) Assume that a is invertible up to (p, q). Let b be (p, q)-inverse of a. Then ab = (1 − q)ab + qab = (1 − q)a(1 − p)b + qab = 1 − q + qab ∈ 1 + F , and also ba = ba(1 − p) + bap = b(1 − q)a(1 − p) + bap = 1 − p + bap ∈ 1 + F . b) Let ab1 = 1 + f1, b2a = 1 + f2, f1, f2 ∈ F . By the previous Proposition, It follows that a∗ is left there is p ∈ F such that ab1 is invertible up to (p, p). invertible up to p (its left inverse is c∗(1 − p)b∗ 1, where c ∈ (1 − p)A(1 − p) is (p, p)- inverse of ab1). Therefore, by Lemma 2.5, there is a projection 1− r ∈ A such that (1 − r)a∗(1 − p) = a∗(1 − p), or equivalently, (2.23) Furthermore, a is invertible up to (p, r). It remains to prove r ∈ F . (1 − p)a = (1 − p)a(1 − r). FREDHOLM OPERATORS ON C ∗-ALGEBRAS 13 Considering (1 − p)a instead of a, we find that b2(1 − p)a = b2a − b2pa = 1 + f2 − b2pa = 1 + f ′ 2 ∈ 1 + F . Again, by the previous Proposition, there is a q ∈ F such that b2(1−p)a is invertible up to (q, q). By Proposition 2.8, there is a projection q′ ∈ F such that b2(1− p)a is invertible up to (q′, q) and such that (1 − q)b2(1 − p)aq′ = 0. The last is equivalent to (1 − q)b2(1 − p)a(1 − q′) = (1 − q)b2(1 − p)a. If t is (q′, q)-inverse of b2(1 − p)a, then (1 − q)r = t(1 − q)b2(1 − p)a(1 − q′)r = t(1 − q)b2(1 − p)ar = 0, since from (2.23) we easily find (1 − p)ar = 0. Thus, r = qr ∈ F , since F is an ideal. (cid:3) Theorem 2.16 (index theorem). Let A be a C∗-algebra, and let F ⊆ A be an algebra of finite type elements. If t1 and t2 are Fredholm type elements then t1t2 is of Fredholm type as well. Moreover there holds ind(t1t2) = ind t1 + ind t2. In other words, If we denote the set of all Fredholm type elements by Fred(F ), then Fred(F ) is a semigroup (with unit) with respect to multiplication, and the mapping ind is a homomorphism from (Fred(F ),·) to (K(F ), +). Proof. Let p1, q1, p2, q2 ∈ F be projections such that t1 is invertible up to (p1, q1) and t2 is invertible up to (p2, q2). By Proposition 2.11, there are approximate units pα ≥ p2, qα ≥ q2 such that t2 is also invertible up to (pα, qα) and (2.24) By Lemma 2.9 there is an α and p′ ∼ p1 such that p′ ≤ qα and p1 − p′ < 1/2. By Lemma 2.4 b), t1 is invertible up to (p′, q1), and therefore ind t1 = [p′] − [q1] Next, by Proposition 2.8 there is a projection q′ ∼ q1 such that t1 is invertible up to (p′, q′) and ind t2 = [p2] − [q2] = [pα] − [qα]. (2.25) It follows (2.26) q′t1(1 − p′) = 0. ind t1 = [p′] − [q′]. Let r = qα − p′. We have r ≤ 1 − p′, and by Lemma 2.5, there is a minimal projection s such that st1r = t1r. (2.27) Also r ∼ s and t1 is invertible up to (1 − r, 1 − s). By (2.25) and Lemma 2.7 we conclude s ≤ 1 − q′. In particular, 1 − q′ − s is a projection and 1 − q′ − s ≤ 1 − s. So, from (2.27) we conclude that (1 − s)t1r = 0 and hence (1 − q′ − s)t1r = (1 − q′ − s)(1 − s)t1r = 0. Therefore, by Lemma 2.6 we obtain that t1 is invertible up to (qα, q′ + s). Now, it is easy to derive that t1(1 − qα)t2 is invertible up to (pα, q′ + s). Indeed, if v1 is (qα, q′ + s)-inverse of t1, and v2 is (pα, qα)-inverse of t2, a straightforward 14 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ calculation yields that v2v1 is (pα, q′ + s)-inverse of t1(1 − qα)t2. Thus t1(1 − qα)t2 is of Fredholm type and ind(t1(1 − qα)t2) = [pα] − [q′] − [s] = [pα] − [qα] + [qα] − [q′] − [r] = [pα] − [qα] + [p′] − [q′] = ind t1 + ind t2. Since t1t2 and t1(1 − qα)t2 differ by t1qαt2 ∈ F , by Proposition 2.14 a) t1t2 is also of Fredholm type and its index is the same as that of t1(1 − qα)t2. The proof is complete. (cid:3) 3. Applications In this section we apply results obtained in the previous section in three different ways. Namely, we shall prove corollaries concerning the ordinary Fredholm opera- tors, the Fredholm operators in the sense of Atiyah and Singer over the II∞ factors and Fredholm operators on Hilbert C∗-modules over a C∗-algebra. All of these corollaries have been known for a long time. Nevertheless, we shall derive them in order to show that results obtained in the previous section are a generalization of all of them. Classic Fredholm operators on a Hilbert space. Corollary 3.1. Let A be the full algebra of all bounded operators on some infinite dimensional Hilbert space H, and let F be the ideal of all compact operators. Then the couple (A,F ) satisfy the conditions (i)−(iii) of Definition 1.1. Hence, ordinary Fredholm operators are the special case of Fredholm type elements defined in this note. Proof. (i) The ideal of all compact operators is a selfadjoint ideal in B(H). (ii) As it is well known, the set of all finite rank projections is an approximate unit for compact operators. Even more, if the Hilbert space H is separable, there is a countable approximate unit for F . In more details if Pn is the projection on the space generated by {e1, . . . , en}, where {ej} is an countable complete orthonormal system, then {Pn n ∈ N} is a countable approximate unit; (iii) Any compact projection P ∈ F is a finite rank projection. Then H ⊖ P H ∼= H. Let V : H → H⊖P H be an isomorphism. Then V QV ∗ is a projection equivalent to Q such that V QV ∗(H) ⊥ P (H). Therefore, V QV ∗ + P is projection. Fredholm operators on a von Neumann algebra. Our second application is devoted to Fredholm operators in the sense of Breuer [4, 5] on a properly infinite von Neumann algebra. First, we give the Breuer definition Definition 3.1. Let A be a von Neumann algebra, let Proj(A) be the set of all projections belonging to A, and let Proj0(A) be the set of all finite projections in A (i.e. those projections that are not Murray von Neumann equivalent to any its proper subprojection). The operator T ∈ A is said to be A-Fredholm if (i) Pker T ∈ Proj0(A), where Pker T is the projection to the subspace ker T and (ii) There is a projection E ∈ Proj0(A) such that ran(I − E) ⊆ ran T . The second condition ensures that Pker T ∗ also belongs to Proj0(A). (cid:3) The index of an A-Fredholm operator T is defined as ind T = dim(ker T ) − dim(ker T ∗) ∈ I(A). (3.1) FREDHOLM OPERATORS ON C ∗-ALGEBRAS 15 Here, I(A) is the so called index group of a von Neumann algebra A defined as the Grothendieck group of the commutative monoid of all representations of the commutant A′ generated by representations of the form A′ ∋ S 7→ ES = πE(S) for some E ∈ Proj0(A) [4, Section 2]. For a subspace L, its dimension dim L is defined as the class [πPL ] ∈ I(A) of the representation πPL , where PL is the projection to L. Before proving that the Breuer's A-Fredholm operators is a special case of ab- stract Fredholm operators, we list some well known facts concerning von Neumann algebras. Lemma 3.2. The set Proj(A) = {p ∈ A p is a projection} is a complete lattice. In particular, if p, q ∈ Proj(A), then p ∨ q (the least upper bound of the set {p, q}) belongs to Proj(A). Proof. The proof can be found in [14, Proposition V.1.1] or [3, I.9.2.1(ii)]. (cid:3) The operator T ∈ A is called finite if the projection to the closure of its range Pran T ∈ Proj0(A). The set of all finite operators is denoted by m0, and its norm closure is denoted by m. Lemma 3.3. The set m is a selfadjoint, two sided ideal of A. Also, m is generated (as a closed selfadjoint two sided ideal) by Proj0(A). Proof. This was proved in the Section 3 of [4]. (cid:3) Lemma 3.4. If p, q ∈ Proj0(A) then p ∨ q ∈ Proj0(A) as well. Proof. The proof can be found in [14, Proposition V.1.6] or [3, III.1.1.3]. (cid:3) Lemma 3.5. Let T ∈ A and let T = UT be its polar decomposition. Then both U , T ∈ A. Proof. The proof can be found in [14, Proposition II.3.14] or [3, I.9.2.1(iii)]. (cid:3) Lemma 3.6. Let 0 ≤ T ∈ A. Then all spectral projections of T also belongs to A. Proof. The proof can be found in [3, I.9.2.1(iv)]. (cid:3) Lemma 3.7. Let A be a von Neumann algebra which is not of finite type. Then the operator T ∈ A is A-Fredholm if and only if T is invertible modulo m. Proof. This is [5, Theorem 1]. (cid:3) Lemma 3.8. Let A be a von Neumann algebra which is not of finite type, and let T be A-Fredholm operator. Then T is A-Fredholm as well. Proof. Let T = V T be the polar decomposition of T . By Lemma 3.5, T ∈ A, and by Lemma 3.7 there is S ∈ A such that ST , T S ∈ 1 + m, i.e. S is m-inverse of T . Then, as it easy to see, SV is a left m-inverse of T, whereas V ∗S∗ is a right inverse of T. Therefore, T is invertible modulo m, and again by Lemma 3.7, T is A-Fredholm. Lemma 3.9. Let T ≥ 0 be an A-Fredholm operator, and let ET be its spectral measure. Then, there is δ > 0 such that ET ([0, δ)) ∈ Proj0(A). (cid:3) 16 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ 1 Proof. By Definition 3.1, there is a closed subspace L ⊆ ran T , which complement is finite (i.e. I − PL ∈ Proj0(A)). Consider the restriction T1 = Tker T ⊥ , and define L1 = T −1 (L). Since both L1 and L are closed, by open mapping theorem T1 realizes a topological isomorphism between L1 and L. Hence, there is δ > 0 such that for all ξ ∈ L1 there holds T ξ > δξ. However, for ξ ∈ ran ET ([0, δ)) we have T ξ ≤ δξ, i.e. L1 ∩ ran ET ([0, δ)) = {0}. Thus, it suffices to prove that L⊥ 1 is finite. Consider L2 = (ker T )⊥ ⊖ L1. Obviously, T (L2) ∩ L = {0}, implying that (I − PL)T maps L2 injectively to a subspace of L⊥. By Lemma 3.5 applied to (I − PL)T PL2 there is a partial isometry W ∈ A that connects L2 and the closure L3 of (I − PL)T (L2) which is a subspace of L⊥. Since L⊥ is finite, both L3 ⊆ L⊥ and L2 ∼ L3 are also finite. Since by Definition 3.1 ker T is finite as well, it follows that L⊥ 1 is finite. The proof is complete, because ET ([0, δ)) ≤ PL⊥ (cid:3) . 1 Corollary 3.10. Let A be a properly infinite von Neumann algebra acting on a Hilbert space H, and let m be defined as above. Then the couple (A, m) satisfies the conditions (i) − (iii) of Definition 1.1. Moreover, abstract Fredholm elements are generalized Fredholm operators in the sense of Breuer. Proof. By Lemma 3.3, m is an ∗-ideal which proves (i) in Definition 1.1. Let us prove that the set Proj0(A) is an approximate unit. First, this is a directed Indeed, let p, q ∈ Proj0(A). Since projection in a von Neumann algebra set. makes a complete lattice (Lemma 3.2), the projection p ∨ q ∈ A. By Lemma 3.4 p ∨ q ∈ Proj0(A) as well. Let T ∈ m, and let T = UT be its polar decomposition. By Lemma 3.5 U , T ∈ A, as well as any spectral projection of T also belongs to A. Pick ε > 0, and denote Pε = ET (ε, +∞), where ET stands for the spectral measure of T. Then Pε ≤ Pran T implying Pε ∈ Proj0(A). Also, T(I−Pε) ≤ ε, since the function λ 7→ λ(1 − χ(ε,+∞)) is bounded by ε. For all P ≥ Pε we have I − P ≤ I − Pε and hence, UT(I − P ) = UT(I − Pε)(I − P ) ≤ ε. We proved (ii) in Definition 1.1. Since A is properly infinite, by [5, Lemma 8], for P , Q ∈ Proj0(A), there holds I−P ∼ I, i.e. there is a a partial isometry V such that V ∗V = I, and V V ∗ = I −P . Then V Q is a partial isometry, required in the property (iii) od Definition 1.1. Using Proposition 2.15 and Lemma 3.7 we see that abstract Fredholm operators with respect to (A, m) coincide with A-Fredholm operators in the sense of Breuer. Let us prove that their indices are equal. Let T be Fredholm operator. By Lemma 3.9, there is a δ > 0 such that P = ET [0, δ) ∈ Proj0(A). The operator T is bounded below on the space ET [δ,∞)H = (I − P )H, hence T = V T is also bounded below on (I − P )H. Let I − Q be the projection on the closure of T (I − P )H. Then, obviously, T is invertible up to (P, Q). We have P = Pker T ⊕ ET (0, δ) = Pker T ⊕ P1, and N1 = P1H is invariant for T. Since V acts as a partial isometry on N , T maps N1 some subspace N2 ∼ N1. Thus, P (H) = ker T ⊕ PN1 and Q = ker T ∗ ⊕ PN2 , and PN1 ∼ PN2 and the result follows. (cid:3) Remark 3.1. The condition (iii) in Definition 1.1 is not substantial. If it is sup- pressed, we have not a semigroup, but a conditional addition. However, we can also FREDHOLM OPERATORS ON C ∗-ALGEBRAS 17 form a corresponding K-group, starting with a free group with Proj0(A) as genera- tors and then using appropriate identifications. Obtained group will be isomorphic to I(A), see [4, Section 2] Fredholm operators on the standard Hilbert module. The last corollary is devoted to Fredholm theory on the standard Hilbert C∗-module l2(B) over a unital C∗-algebra B. We list some definitions and known facts. For references and further details the reader is referred to [7]. Let B be a unital C∗-algebra. We define the right Hilbert module l2(B) as l2(B) = {(an)∞ n=1 an ∈ B,Xn han, ani converges in norm} equipped with the right action of B (an)b = (anb) and with the inner product h·,·i : l2(B) × l2(B) → B given by ∞ h(an), (bn)i = a∗ nbn. Xn=1 denote the algebra of all operators from B(l2(B)) that have the adjoint. Let B(l2(B)) denote the set of all bounded operators on l2(B), and let Ba(l2(B)) Let C0(l2(B)) denote the algebra generated by all operators of the form θx,y(z) = xhy, zi . Its closure we will denote by C(l2(B)). Although such operators might not map bounded into relatively compact sets it is common to call them compact operators. Next, we quote definitions and statements given by Mingo [8] and Mischenko and Fomenko [9]. Definition 3.2. T ∈ Ba(l2(B)) is Fredholm in the sense of Mingo, if it is invertible modulo compact operators. [8, §1.1. Definition]. Proposition 3.11. T is Fredholm in the sense of Mingo if and only if there is a compact perturbation S of T (i.e. T − S is compact) such that ker S and ker S∗ are finitely generated projective. The difference (3.2) [ker S] − [ker S∗] ∈ K0(B) does not depend on the choice of S [8, §1.4.]. Definition 3.3. The index of a Fredholm operator T is defined to be the difference (3.2) where S is any compact perturbation of T [8, §1.4.]. Definition 3.4. T is Fredholm in the sense of Mischenko and Fomenko (MF sense in further) if (i) T is adjointable; (ii) there are decompositions of the domain l2(B) = M1 ⊕N1 and of the codomain l2(B) = M2 ⊕N2, where N1, N2 are finitely generated projective submodules and T with respect to such decompositions has a matrix form T = (cid:18)T1 The index of T is ind T = [N1] − [N2] [7, Definition 2.7.8]. T2(cid:19), with T1 is an isomorphism. [7, Definition 2.7.4]. 0 0 Proposition 3.12. If T is invertible modulo compacts then T is Fredholm in MF sense. [7, Theorem 2.7.14] 18 DRAGOLJUB J. KEČKIĆ AND ZLATKO LAZOVIĆ Note that the other implication trivially holds. Namely, if T is Fredholm in MF sense, then (cid:18)T −1 1 0 0 0(cid:19) is the inverse of T modulo compacts. We also need the following Lemma. n, y′′ y′ n ∈ M, y′′ n ∈ N . n = P xn + y′′ (I − P )xn = y′ n → 0, since M and N are complemented. However, from (3.3) we n ∈ N . Thus both sides are equal to 0, since n → 0 which gives a contradiction with Lemma 3.13. Let l2(B) = M ⊕N , where N is finitely generated and projective and M is complemented. Next, let P be the orthogonal projection to N . Then I − P is bounded below on M. Proof. Suppose this is not true, i.e. that there is a sequence xn ∈ M, xn = 1, (I − P )xn → 0. Decompose (I − P )xn as n + y′′ (3.3) n, Then, y′ obtain M ∋ xn − y′ the sum M ⊕N is direct. Hence xn = y′ xn = 1. Corollary 3.14. Let A = l2(B) be the standard Hilbert module over some unital C∗-algebra B and let F = C(l2(B)), the set of all compact operators on l2(B). Then the couple (A,F ) satisfies conditions in Definition 1.1. Moreover, the set of abstract Fredholm operators, Fredholom operators in the sense of Mingo, and in the MF sense coincide. Finally all three indices are equal. Proof. (i) It is well known that F is a closed ideal. with the unit being the i-th entry. (ii) Let l2(B) possesses the standard basis (ei), i ∈ N, where ei = (0, ..., 0, 1, 0, ...) Let Pn ∈ F be sequence of projections on Bn, Pn(x1, x2, . . . ) = (x1, . . . , xn, 0, . . . ) (cid:3) and let K ∈ C(l2(B)). By [7, Proposition 2.2.1] we have K − KPn → 0, n → ∞. Therefore, Pn is an approximate unit. (iii) Let P , Q ∈ F be projections. By [15, Theorem 15.4.2], P (A) and Q(A) are isomorphic (as modules) to some direct summand in Bn and Bm, respectively, for some m, n ∈ N. Hence, there is partial isometry V (x1, x2, ...) = em+1x1 + em+2x2 + ... + em+nxn. V : l2(B) → l2(B), Then P1 = V P V ∗ is orthogonal to Q. Consider the operator U = V P . We have U U ∗ = V P V ∗ = P1 and U ∗U = P V ∗V P = P, since V ∗V = Pe1,...,en ≥ P . Also (U U ∗ + Q)(U U ∗ + Q) = V P V ∗V P V ∗ + Q2 = V P V ∗ + Q implying that U U ∗ + Q is a projection. Thus, property (iii) is proved. Proposition 2.15, Proposition 3.12 together with a comment after it and Defini- tion 3.2 proves that all three kind of Fredholm operators (abstract, in the sense of Mingo and in MF sense) coincide. Let us prove that their indices are equal. Let T be a Fredholm operator, and let T be invertible up to (P, Q). Then T ′ = (I − Q)T (I − P ) is a compact perturbation of T (its difference is QT (I − P ) + (I − Q)T P + QT P ), and ker T ′ = P (l2(B)), ker T ′∗ = Q(l2(B)). Hence, abstract index is equal to Mingo's index. T2(cid:19) with respect to decompo- Let T be a Fredholm operator, and let T = (cid:18)T1 sitions l2(B) = M1 ⊕N1 = M2 ⊕N2 of the domain and the codomain, respectively. 0 0 FREDHOLM OPERATORS ON C ∗-ALGEBRAS 19 By [7, Theorem 2.7.6], M1, N1, M2, N2 can be chosen such that all of them are complemented, N1 and N2 are finitely generated projective and also M1 = N ⊥ 1 . (In truth, the last equality is not emphasized in the statement, but it follows from the proof.) In the same proof we can found that TM1 is bounded below. Denote by P and Q the orthogonal projections to N1 and N2. By Lemma 3.13, I − Q is bounded below on the space M2 = T (M1) = ran T (I − P ). This ensures that 1 onto N ⊥ (I − Q)T (I − P ) is an isomorphism from M1 = N ⊥ 2 . Thus the abstract index is [P ] − [Q] = [N1] − [N2], i.e. it is equal to MF index. (cid:3) References [1] T. Alvarez and M. Onieva. Generalized Fredholm operators. Arch. Math., 44:270 -- 277, 1985. [2] M. F. Atiyah. Elliptic operators, discrete groups and von Neumann algebras. Asterisque, 32 -- 33:43 -- 72, 1976. [3] B. Blackadar. Operator Algebras - Theory of C ∗-algebras and von Neumann algebras, volume 122 of Encyclopaedea of Mathematical Sciences. Springer, Berlin, Heidelberg, 2006. [4] Manfred Breuer. Fredholm theories in von neumann algebras I. Math. Annalen, 178(3):243 -- 254, 1968. [5] Manfred Breuer. Fredholm theories in von neumann algebras II. Math. Annalen, 180(4):313 -- 325, 1969. [6] L. A. Coburn, R. G. Douglas, D. G. Schaeffer, and I. M. Singer. C ∗-algebras of operators on a half-space, ii Index theory. Arch. Publications mathematiques de l'I.H.É.S., 40:68 -- 79, 1971. [7] V. M. Manuilov and E. V. Troitsky. Hilbert C ∗-modules. Translations of mathematical mono- graphs Vol. 226. AMS, Providence, Rhode Island, 2005. [8] James A. Mingo. K-theory and multipliers of stable C ∗-algebras. Trans. Amer. Math. Soc., 299(1):397 -- 411, 1987. [9] A. S. Mishchenko and A. T. Fomenko. The index of elliptic operators over C ∗-algebras. Math. USSR Izv., 15(1):87 -- 112, 1980. [10] G. J. Murphy. C ∗-algebras and operator theory. Academic press, London, 1990. [11] Catherine L. Olsen. Index theory in von Neumann algebras. Mem. Amer. Math. Soc., 47(294):1 -- 71, 1984. [12] W. Rudin. Functional Analysis. McGraw-Hill Book Company, 1973. [13] B. Sz.-Nagy and F. Riesz. Functional analysis. Frederick Ungar Publishing Co., New York, 1955. [14] M. Takesaki. Theory of Operator Algebras I, volume 124 of Encyclopaedea of Mathematical Sciences. Springer, Berlin, Heidelberg, etc., 2001. [15] N. E. Wegge-Olsen. K-theory and C ∗-algebras - A friendly approach. Oxford University Press, Oxford, New York, Tokyo, 1993. University of Belgrade, Faculty of Mathematics, Studentski trg 16-18, 11000 Beograd, Serbia E-mail address: [email protected] University of Belgrade, Faculty of Mathematics, Studentski trg 16-18, 11000 Beograd, Serbia E-mail address: [email protected]

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.